Squashed 'third_party/protobuf/' content from commit e35e248

Change-Id: I6cbe123d09fe50fdcad0e51466665daeee7433c7
git-subtree-dir: third_party/protobuf
git-subtree-split: e35e24800fb8d694bdeea5fd63dc7d1b14d68723
diff --git a/python/google/protobuf/internal/__init__.py b/python/google/protobuf/internal/__init__.py
new file mode 100755
index 0000000..e69de29
--- /dev/null
+++ b/python/google/protobuf/internal/__init__.py
diff --git a/python/google/protobuf/internal/_parameterized.py b/python/google/protobuf/internal/_parameterized.py
new file mode 100755
index 0000000..dea3f19
--- /dev/null
+++ b/python/google/protobuf/internal/_parameterized.py
@@ -0,0 +1,443 @@
+#! /usr/bin/env python
+#
+# Protocol Buffers - Google's data interchange format
+# Copyright 2008 Google Inc.  All rights reserved.
+# https://developers.google.com/protocol-buffers/
+#
+# Redistribution and use in source and binary forms, with or without
+# modification, are permitted provided that the following conditions are
+# met:
+#
+#     * Redistributions of source code must retain the above copyright
+# notice, this list of conditions and the following disclaimer.
+#     * Redistributions in binary form must reproduce the above
+# copyright notice, this list of conditions and the following disclaimer
+# in the documentation and/or other materials provided with the
+# distribution.
+#     * Neither the name of Google Inc. nor the names of its
+# contributors may be used to endorse or promote products derived from
+# this software without specific prior written permission.
+#
+# THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+# "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
+# LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
+# A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
+# OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
+# SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
+# LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
+# DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
+# THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
+# (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+# OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+
+"""Adds support for parameterized tests to Python's unittest TestCase class.
+
+A parameterized test is a method in a test case that is invoked with different
+argument tuples.
+
+A simple example:
+
+  class AdditionExample(parameterized.ParameterizedTestCase):
+    @parameterized.Parameters(
+       (1, 2, 3),
+       (4, 5, 9),
+       (1, 1, 3))
+    def testAddition(self, op1, op2, result):
+      self.assertEqual(result, op1 + op2)
+
+
+Each invocation is a separate test case and properly isolated just
+like a normal test method, with its own setUp/tearDown cycle. In the
+example above, there are three separate testcases, one of which will
+fail due to an assertion error (1 + 1 != 3).
+
+Parameters for invididual test cases can be tuples (with positional parameters)
+or dictionaries (with named parameters):
+
+  class AdditionExample(parameterized.ParameterizedTestCase):
+    @parameterized.Parameters(
+       {'op1': 1, 'op2': 2, 'result': 3},
+       {'op1': 4, 'op2': 5, 'result': 9},
+    )
+    def testAddition(self, op1, op2, result):
+      self.assertEqual(result, op1 + op2)
+
+If a parameterized test fails, the error message will show the
+original test name (which is modified internally) and the arguments
+for the specific invocation, which are part of the string returned by
+the shortDescription() method on test cases.
+
+The id method of the test, used internally by the unittest framework,
+is also modified to show the arguments. To make sure that test names
+stay the same across several invocations, object representations like
+
+  >>> class Foo(object):
+  ...  pass
+  >>> repr(Foo())
+  '<__main__.Foo object at 0x23d8610>'
+
+are turned into '<__main__.Foo>'. For even more descriptive names,
+especially in test logs, you can use the NamedParameters decorator. In
+this case, only tuples are supported, and the first parameters has to
+be a string (or an object that returns an apt name when converted via
+str()):
+
+  class NamedExample(parameterized.ParameterizedTestCase):
+    @parameterized.NamedParameters(
+       ('Normal', 'aa', 'aaa', True),
+       ('EmptyPrefix', '', 'abc', True),
+       ('BothEmpty', '', '', True))
+    def testStartsWith(self, prefix, string, result):
+      self.assertEqual(result, strings.startswith(prefix))
+
+Named tests also have the benefit that they can be run individually
+from the command line:
+
+  $ testmodule.py NamedExample.testStartsWithNormal
+  .
+  --------------------------------------------------------------------
+  Ran 1 test in 0.000s
+
+  OK
+
+Parameterized Classes
+=====================
+If invocation arguments are shared across test methods in a single
+ParameterizedTestCase class, instead of decorating all test methods
+individually, the class itself can be decorated:
+
+  @parameterized.Parameters(
+    (1, 2, 3)
+    (4, 5, 9))
+  class ArithmeticTest(parameterized.ParameterizedTestCase):
+    def testAdd(self, arg1, arg2, result):
+      self.assertEqual(arg1 + arg2, result)
+
+    def testSubtract(self, arg2, arg2, result):
+      self.assertEqual(result - arg1, arg2)
+
+Inputs from Iterables
+=====================
+If parameters should be shared across several test cases, or are dynamically
+created from other sources, a single non-tuple iterable can be passed into
+the decorator. This iterable will be used to obtain the test cases:
+
+  class AdditionExample(parameterized.ParameterizedTestCase):
+    @parameterized.Parameters(
+      c.op1, c.op2, c.result for c in testcases
+    )
+    def testAddition(self, op1, op2, result):
+      self.assertEqual(result, op1 + op2)
+
+
+Single-Argument Test Methods
+============================
+If a test method takes only one argument, the single argument does not need to
+be wrapped into a tuple:
+
+  class NegativeNumberExample(parameterized.ParameterizedTestCase):
+    @parameterized.Parameters(
+       -1, -3, -4, -5
+    )
+    def testIsNegative(self, arg):
+      self.assertTrue(IsNegative(arg))
+"""
+
+__author__ = 'tmarek@google.com (Torsten Marek)'
+
+import collections
+import functools
+import re
+import types
+try:
+  import unittest2 as unittest
+except ImportError:
+  import unittest
+import uuid
+
+import six
+
+ADDR_RE = re.compile(r'\<([a-zA-Z0-9_\-\.]+) object at 0x[a-fA-F0-9]+\>')
+_SEPARATOR = uuid.uuid1().hex
+_FIRST_ARG = object()
+_ARGUMENT_REPR = object()
+
+
+def _CleanRepr(obj):
+  return ADDR_RE.sub(r'<\1>', repr(obj))
+
+
+# Helper function formerly from the unittest module, removed from it in
+# Python 2.7.
+def _StrClass(cls):
+  return '%s.%s' % (cls.__module__, cls.__name__)
+
+
+def _NonStringIterable(obj):
+  return (isinstance(obj, collections.Iterable) and not
+          isinstance(obj, six.string_types))
+
+
+def _FormatParameterList(testcase_params):
+  if isinstance(testcase_params, collections.Mapping):
+    return ', '.join('%s=%s' % (argname, _CleanRepr(value))
+                     for argname, value in testcase_params.items())
+  elif _NonStringIterable(testcase_params):
+    return ', '.join(map(_CleanRepr, testcase_params))
+  else:
+    return _FormatParameterList((testcase_params,))
+
+
+class _ParameterizedTestIter(object):
+  """Callable and iterable class for producing new test cases."""
+
+  def __init__(self, test_method, testcases, naming_type):
+    """Returns concrete test functions for a test and a list of parameters.
+
+    The naming_type is used to determine the name of the concrete
+    functions as reported by the unittest framework. If naming_type is
+    _FIRST_ARG, the testcases must be tuples, and the first element must
+    have a string representation that is a valid Python identifier.
+
+    Args:
+      test_method: The decorated test method.
+      testcases: (list of tuple/dict) A list of parameter
+                 tuples/dicts for individual test invocations.
+      naming_type: The test naming type, either _NAMED or _ARGUMENT_REPR.
+    """
+    self._test_method = test_method
+    self.testcases = testcases
+    self._naming_type = naming_type
+
+  def __call__(self, *args, **kwargs):
+    raise RuntimeError('You appear to be running a parameterized test case '
+                       'without having inherited from parameterized.'
+                       'ParameterizedTestCase. This is bad because none of '
+                       'your test cases are actually being run.')
+
+  def __iter__(self):
+    test_method = self._test_method
+    naming_type = self._naming_type
+
+    def MakeBoundParamTest(testcase_params):
+      @functools.wraps(test_method)
+      def BoundParamTest(self):
+        if isinstance(testcase_params, collections.Mapping):
+          test_method(self, **testcase_params)
+        elif _NonStringIterable(testcase_params):
+          test_method(self, *testcase_params)
+        else:
+          test_method(self, testcase_params)
+
+      if naming_type is _FIRST_ARG:
+        # Signal the metaclass that the name of the test function is unique
+        # and descriptive.
+        BoundParamTest.__x_use_name__ = True
+        BoundParamTest.__name__ += str(testcase_params[0])
+        testcase_params = testcase_params[1:]
+      elif naming_type is _ARGUMENT_REPR:
+        # __x_extra_id__ is used to pass naming information to the __new__
+        # method of TestGeneratorMetaclass.
+        # The metaclass will make sure to create a unique, but nondescriptive
+        # name for this test.
+        BoundParamTest.__x_extra_id__ = '(%s)' % (
+            _FormatParameterList(testcase_params),)
+      else:
+        raise RuntimeError('%s is not a valid naming type.' % (naming_type,))
+
+      BoundParamTest.__doc__ = '%s(%s)' % (
+          BoundParamTest.__name__, _FormatParameterList(testcase_params))
+      if test_method.__doc__:
+        BoundParamTest.__doc__ += '\n%s' % (test_method.__doc__,)
+      return BoundParamTest
+    return (MakeBoundParamTest(c) for c in self.testcases)
+
+
+def _IsSingletonList(testcases):
+  """True iff testcases contains only a single non-tuple element."""
+  return len(testcases) == 1 and not isinstance(testcases[0], tuple)
+
+
+def _ModifyClass(class_object, testcases, naming_type):
+  assert not getattr(class_object, '_id_suffix', None), (
+      'Cannot add parameters to %s,'
+      ' which already has parameterized methods.' % (class_object,))
+  class_object._id_suffix = id_suffix = {}
+  # We change the size of __dict__ while we iterate over it, 
+  # which Python 3.x will complain about, so use copy().
+  for name, obj in class_object.__dict__.copy().items():
+    if (name.startswith(unittest.TestLoader.testMethodPrefix)
+        and isinstance(obj, types.FunctionType)):
+      delattr(class_object, name)
+      methods = {}
+      _UpdateClassDictForParamTestCase(
+          methods, id_suffix, name,
+          _ParameterizedTestIter(obj, testcases, naming_type))
+      for name, meth in methods.items():
+        setattr(class_object, name, meth)
+
+
+def _ParameterDecorator(naming_type, testcases):
+  """Implementation of the parameterization decorators.
+
+  Args:
+    naming_type: The naming type.
+    testcases: Testcase parameters.
+
+  Returns:
+    A function for modifying the decorated object.
+  """
+  def _Apply(obj):
+    if isinstance(obj, type):
+      _ModifyClass(
+          obj,
+          list(testcases) if not isinstance(testcases, collections.Sequence)
+          else testcases,
+          naming_type)
+      return obj
+    else:
+      return _ParameterizedTestIter(obj, testcases, naming_type)
+
+  if _IsSingletonList(testcases):
+    assert _NonStringIterable(testcases[0]), (
+        'Single parameter argument must be a non-string iterable')
+    testcases = testcases[0]
+
+  return _Apply
+
+
+def Parameters(*testcases):
+  """A decorator for creating parameterized tests.
+
+  See the module docstring for a usage example.
+  Args:
+    *testcases: Parameters for the decorated method, either a single
+                iterable, or a list of tuples/dicts/objects (for tests
+                with only one argument).
+
+  Returns:
+     A test generator to be handled by TestGeneratorMetaclass.
+  """
+  return _ParameterDecorator(_ARGUMENT_REPR, testcases)
+
+
+def NamedParameters(*testcases):
+  """A decorator for creating parameterized tests.
+
+  See the module docstring for a usage example. The first element of
+  each parameter tuple should be a string and will be appended to the
+  name of the test method.
+
+  Args:
+    *testcases: Parameters for the decorated method, either a single
+                iterable, or a list of tuples.
+
+  Returns:
+     A test generator to be handled by TestGeneratorMetaclass.
+  """
+  return _ParameterDecorator(_FIRST_ARG, testcases)
+
+
+class TestGeneratorMetaclass(type):
+  """Metaclass for test cases with test generators.
+
+  A test generator is an iterable in a testcase that produces callables. These
+  callables must be single-argument methods. These methods are injected into
+  the class namespace and the original iterable is removed. If the name of the
+  iterable conforms to the test pattern, the injected methods will be picked
+  up as tests by the unittest framework.
+
+  In general, it is supposed to be used in conjuction with the
+  Parameters decorator.
+  """
+
+  def __new__(mcs, class_name, bases, dct):
+    dct['_id_suffix'] = id_suffix = {}
+    for name, obj in dct.items():
+      if (name.startswith(unittest.TestLoader.testMethodPrefix) and
+          _NonStringIterable(obj)):
+        iterator = iter(obj)
+        dct.pop(name)
+        _UpdateClassDictForParamTestCase(dct, id_suffix, name, iterator)
+
+    return type.__new__(mcs, class_name, bases, dct)
+
+
+def _UpdateClassDictForParamTestCase(dct, id_suffix, name, iterator):
+  """Adds individual test cases to a dictionary.
+
+  Args:
+    dct: The target dictionary.
+    id_suffix: The dictionary for mapping names to test IDs.
+    name: The original name of the test case.
+    iterator: The iterator generating the individual test cases.
+  """
+  for idx, func in enumerate(iterator):
+    assert callable(func), 'Test generators must yield callables, got %r' % (
+        func,)
+    if getattr(func, '__x_use_name__', False):
+      new_name = func.__name__
+    else:
+      new_name = '%s%s%d' % (name, _SEPARATOR, idx)
+    assert new_name not in dct, (
+        'Name of parameterized test case "%s" not unique' % (new_name,))
+    dct[new_name] = func
+    id_suffix[new_name] = getattr(func, '__x_extra_id__', '')
+
+
+class ParameterizedTestCase(unittest.TestCase):
+  """Base class for test cases using the Parameters decorator."""
+  __metaclass__ = TestGeneratorMetaclass
+
+  def _OriginalName(self):
+    return self._testMethodName.split(_SEPARATOR)[0]
+
+  def __str__(self):
+    return '%s (%s)' % (self._OriginalName(), _StrClass(self.__class__))
+
+  def id(self):  # pylint: disable=invalid-name
+    """Returns the descriptive ID of the test.
+
+    This is used internally by the unittesting framework to get a name
+    for the test to be used in reports.
+
+    Returns:
+      The test id.
+    """
+    return '%s.%s%s' % (_StrClass(self.__class__),
+                        self._OriginalName(),
+                        self._id_suffix.get(self._testMethodName, ''))
+
+
+def CoopParameterizedTestCase(other_base_class):
+  """Returns a new base class with a cooperative metaclass base.
+
+  This enables the ParameterizedTestCase to be used in combination
+  with other base classes that have custom metaclasses, such as
+  mox.MoxTestBase.
+
+  Only works with metaclasses that do not override type.__new__.
+
+  Example:
+
+    import google3
+    import mox
+
+    from google3.testing.pybase import parameterized
+
+    class ExampleTest(parameterized.CoopParameterizedTestCase(mox.MoxTestBase)):
+      ...
+
+  Args:
+    other_base_class: (class) A test case base class.
+
+  Returns:
+    A new class object.
+  """
+  metaclass = type(
+      'CoopMetaclass',
+      (other_base_class.__metaclass__,
+       TestGeneratorMetaclass), {})
+  return metaclass(
+      'CoopParameterizedTestCase',
+      (other_base_class, ParameterizedTestCase), {})
diff --git a/python/google/protobuf/internal/any_test.proto b/python/google/protobuf/internal/any_test.proto
new file mode 100644
index 0000000..cd641ca
--- /dev/null
+++ b/python/google/protobuf/internal/any_test.proto
@@ -0,0 +1,42 @@
+// Protocol Buffers - Google's data interchange format
+// Copyright 2008 Google Inc.  All rights reserved.
+// https://developers.google.com/protocol-buffers/
+//
+// Redistribution and use in source and binary forms, with or without
+// modification, are permitted provided that the following conditions are
+// met:
+//
+//     * Redistributions of source code must retain the above copyright
+// notice, this list of conditions and the following disclaimer.
+//     * Redistributions in binary form must reproduce the above
+// copyright notice, this list of conditions and the following disclaimer
+// in the documentation and/or other materials provided with the
+// distribution.
+//     * Neither the name of Google Inc. nor the names of its
+// contributors may be used to endorse or promote products derived from
+// this software without specific prior written permission.
+//
+// THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+// "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
+// LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
+// A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
+// OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
+// SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
+// LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
+// DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
+// THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
+// (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+// OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+
+// Author: jieluo@google.com (Jie Luo)
+
+syntax = "proto3";
+
+package google.protobuf.internal;
+
+import "google/protobuf/any.proto";
+
+message TestAny {
+  google.protobuf.Any value = 1;
+  int32 int_value = 2;
+}
diff --git a/python/google/protobuf/internal/api_implementation.cc b/python/google/protobuf/internal/api_implementation.cc
new file mode 100644
index 0000000..6db12e8
--- /dev/null
+++ b/python/google/protobuf/internal/api_implementation.cc
@@ -0,0 +1,129 @@
+// Protocol Buffers - Google's data interchange format
+// Copyright 2008 Google Inc.  All rights reserved.
+// https://developers.google.com/protocol-buffers/
+//
+// Redistribution and use in source and binary forms, with or without
+// modification, are permitted provided that the following conditions are
+// met:
+//
+//     * Redistributions of source code must retain the above copyright
+// notice, this list of conditions and the following disclaimer.
+//     * Redistributions in binary form must reproduce the above
+// copyright notice, this list of conditions and the following disclaimer
+// in the documentation and/or other materials provided with the
+// distribution.
+//     * Neither the name of Google Inc. nor the names of its
+// contributors may be used to endorse or promote products derived from
+// this software without specific prior written permission.
+//
+// THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+// "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
+// LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
+// A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
+// OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
+// SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
+// LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
+// DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
+// THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
+// (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+// OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+
+#include <Python.h>
+
+namespace google {
+namespace protobuf {
+namespace python {
+
+// Version constant.
+// This is either 0 for python, 1 for CPP V1, 2 for CPP V2.
+//
+// 0 is default and is equivalent to
+//   PROTOCOL_BUFFERS_PYTHON_IMPLEMENTATION=python
+//
+// 1 is set with -DPYTHON_PROTO2_CPP_IMPL_V1 and is equivalent to
+//   PROTOCOL_BUFFERS_PYTHON_IMPLEMENTATION=cpp
+// and
+//   PROTOCOL_BUFFERS_PYTHON_IMPLEMENTATION_VERSION=1
+//
+// 2 is set with -DPYTHON_PROTO2_CPP_IMPL_V2 and is equivalent to
+//   PROTOCOL_BUFFERS_PYTHON_IMPLEMENTATION=cpp
+// and
+//   PROTOCOL_BUFFERS_PYTHON_IMPLEMENTATION_VERSION=2
+#ifdef PYTHON_PROTO2_CPP_IMPL_V1
+#error "PYTHON_PROTO2_CPP_IMPL_V1 is no longer supported."
+#else
+#ifdef PYTHON_PROTO2_CPP_IMPL_V2
+static int kImplVersion = 2;
+#else
+#ifdef PYTHON_PROTO2_PYTHON_IMPL
+static int kImplVersion = 0;
+#else
+
+static int kImplVersion = -1;  // -1 means "Unspecified by compiler flags".
+
+#endif  // PYTHON_PROTO2_PYTHON_IMPL
+#endif  // PYTHON_PROTO2_CPP_IMPL_V2
+#endif  // PYTHON_PROTO2_CPP_IMPL_V1
+
+static const char* kImplVersionName = "api_version";
+
+static const char* kModuleName = "_api_implementation";
+static const char kModuleDocstring[] =
+"_api_implementation is a module that exposes compile-time constants that\n"
+"determine the default API implementation to use for Python proto2.\n"
+"\n"
+"It complements api_implementation.py by setting defaults using compile-time\n"
+"constants defined in C, such that one can set defaults at compilation\n"
+"(e.g. with blaze flag --copt=-DPYTHON_PROTO2_CPP_IMPL_V2).";
+
+#if PY_MAJOR_VERSION >= 3
+static struct PyModuleDef _module = {
+  PyModuleDef_HEAD_INIT,
+  kModuleName,
+  kModuleDocstring,
+  -1,
+  NULL,
+  NULL,
+  NULL,
+  NULL,
+  NULL
+};
+#define INITFUNC PyInit__api_implementation
+#define INITFUNC_ERRORVAL NULL
+#else
+#define INITFUNC init_api_implementation
+#define INITFUNC_ERRORVAL
+#endif
+
+extern "C" {
+  PyMODINIT_FUNC INITFUNC() {
+#if PY_MAJOR_VERSION >= 3
+    PyObject *module = PyModule_Create(&_module);
+#else
+    PyObject *module = Py_InitModule3(
+        const_cast<char*>(kModuleName),
+        NULL,
+        const_cast<char*>(kModuleDocstring));
+#endif
+    if (module == NULL) {
+      return INITFUNC_ERRORVAL;
+    }
+
+    // Adds the module variable "api_version".
+    if (PyModule_AddIntConstant(
+        module,
+        const_cast<char*>(kImplVersionName),
+        kImplVersion))
+#if PY_MAJOR_VERSION < 3
+      return;
+#else
+      { Py_DECREF(module); return NULL; }
+
+    return module;
+#endif
+  }
+}
+
+}  // namespace python
+}  // namespace protobuf
+}  // namespace google
diff --git a/python/google/protobuf/internal/api_implementation.py b/python/google/protobuf/internal/api_implementation.py
new file mode 100755
index 0000000..ffcf751
--- /dev/null
+++ b/python/google/protobuf/internal/api_implementation.py
@@ -0,0 +1,107 @@
+# Protocol Buffers - Google's data interchange format
+# Copyright 2008 Google Inc.  All rights reserved.
+# https://developers.google.com/protocol-buffers/
+#
+# Redistribution and use in source and binary forms, with or without
+# modification, are permitted provided that the following conditions are
+# met:
+#
+#     * Redistributions of source code must retain the above copyright
+# notice, this list of conditions and the following disclaimer.
+#     * Redistributions in binary form must reproduce the above
+# copyright notice, this list of conditions and the following disclaimer
+# in the documentation and/or other materials provided with the
+# distribution.
+#     * Neither the name of Google Inc. nor the names of its
+# contributors may be used to endorse or promote products derived from
+# this software without specific prior written permission.
+#
+# THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+# "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
+# LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
+# A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
+# OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
+# SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
+# LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
+# DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
+# THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
+# (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+# OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+
+"""Determine which implementation of the protobuf API is used in this process.
+"""
+
+import os
+import sys
+
+try:
+  # pylint: disable=g-import-not-at-top
+  from google.protobuf.internal import _api_implementation
+  # The compile-time constants in the _api_implementation module can be used to
+  # switch to a certain implementation of the Python API at build time.
+  _api_version = _api_implementation.api_version
+  _proto_extension_modules_exist_in_build = True
+except ImportError:
+  _api_version = -1  # Unspecified by compiler flags.
+  _proto_extension_modules_exist_in_build = False
+
+if _api_version == 1:
+  raise ValueError('api_version=1 is no longer supported.')
+if _api_version < 0:  # Still unspecified?
+  try:
+    # The presence of this module in a build allows the proto implementation to
+    # be upgraded merely via build deps rather than a compiler flag or the
+    # runtime environment variable.
+    # pylint: disable=g-import-not-at-top
+    from google.protobuf import _use_fast_cpp_protos
+    # Work around a known issue in the classic bootstrap .par import hook.
+    if not _use_fast_cpp_protos:
+      raise ImportError('_use_fast_cpp_protos import succeeded but was None')
+    del _use_fast_cpp_protos
+    _api_version = 2
+  except ImportError:
+    if _proto_extension_modules_exist_in_build:
+      if sys.version_info[0] >= 3:  # Python 3 defaults to C++ impl v2.
+        _api_version = 2
+      # TODO(b/17427486): Make Python 2 default to C++ impl v2.
+
+_default_implementation_type = (
+    'python' if _api_version <= 0 else 'cpp')
+
+# This environment variable can be used to switch to a certain implementation
+# of the Python API, overriding the compile-time constants in the
+# _api_implementation module. Right now only 'python' and 'cpp' are valid
+# values. Any other value will be ignored.
+_implementation_type = os.getenv('PROTOCOL_BUFFERS_PYTHON_IMPLEMENTATION',
+                                 _default_implementation_type)
+
+if _implementation_type != 'python':
+  _implementation_type = 'cpp'
+
+# This environment variable can be used to switch between the two
+# 'cpp' implementations, overriding the compile-time constants in the
+# _api_implementation module. Right now only '2' is supported. Any other
+# value will cause an error to be raised.
+_implementation_version_str = os.getenv(
+    'PROTOCOL_BUFFERS_PYTHON_IMPLEMENTATION_VERSION', '2')
+
+if _implementation_version_str != '2':
+  raise ValueError(
+      'unsupported PROTOCOL_BUFFERS_PYTHON_IMPLEMENTATION_VERSION: "' +
+      _implementation_version_str + '" (supported versions: 2)'
+      )
+
+_implementation_version = int(_implementation_version_str)
+
+
+# Usage of this function is discouraged. Clients shouldn't care which
+# implementation of the API is in use. Note that there is no guarantee
+# that differences between APIs will be maintained.
+# Please don't use this function if possible.
+def Type():
+  return _implementation_type
+
+
+# See comment on 'Type' above.
+def Version():
+  return _implementation_version
diff --git a/python/google/protobuf/internal/containers.py b/python/google/protobuf/internal/containers.py
new file mode 100755
index 0000000..97cdd84
--- /dev/null
+++ b/python/google/protobuf/internal/containers.py
@@ -0,0 +1,611 @@
+# Protocol Buffers - Google's data interchange format
+# Copyright 2008 Google Inc.  All rights reserved.
+# https://developers.google.com/protocol-buffers/
+#
+# Redistribution and use in source and binary forms, with or without
+# modification, are permitted provided that the following conditions are
+# met:
+#
+#     * Redistributions of source code must retain the above copyright
+# notice, this list of conditions and the following disclaimer.
+#     * Redistributions in binary form must reproduce the above
+# copyright notice, this list of conditions and the following disclaimer
+# in the documentation and/or other materials provided with the
+# distribution.
+#     * Neither the name of Google Inc. nor the names of its
+# contributors may be used to endorse or promote products derived from
+# this software without specific prior written permission.
+#
+# THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+# "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
+# LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
+# A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
+# OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
+# SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
+# LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
+# DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
+# THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
+# (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+# OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+
+"""Contains container classes to represent different protocol buffer types.
+
+This file defines container classes which represent categories of protocol
+buffer field types which need extra maintenance. Currently these categories
+are:
+  - Repeated scalar fields - These are all repeated fields which aren't
+    composite (e.g. they are of simple types like int32, string, etc).
+  - Repeated composite fields - Repeated fields which are composite. This
+    includes groups and nested messages.
+"""
+
+__author__ = 'petar@google.com (Petar Petrov)'
+
+import collections
+import sys
+
+if sys.version_info[0] < 3:
+  # We would use collections.MutableMapping all the time, but in Python 2 it
+  # doesn't define __slots__.  This causes two significant problems:
+  #
+  # 1. we can't disallow arbitrary attribute assignment, even if our derived
+  #    classes *do* define __slots__.
+  #
+  # 2. we can't safely derive a C type from it without __slots__ defined (the
+  #    interpreter expects to find a dict at tp_dictoffset, which we can't
+  #    robustly provide.  And we don't want an instance dict anyway.
+  #
+  # So this is the Python 2.7 definition of Mapping/MutableMapping functions
+  # verbatim, except that:
+  # 1. We declare __slots__.
+  # 2. We don't declare this as a virtual base class.  The classes defined
+  #    in collections are the interesting base classes, not us.
+  #
+  # Note: deriving from object is critical.  It is the only thing that makes
+  # this a true type, allowing us to derive from it in C++ cleanly and making
+  # __slots__ properly disallow arbitrary element assignment.
+
+  class Mapping(object):
+    __slots__ = ()
+
+    def get(self, key, default=None):
+      try:
+        return self[key]
+      except KeyError:
+        return default
+
+    def __contains__(self, key):
+      try:
+        self[key]
+      except KeyError:
+        return False
+      else:
+        return True
+
+    def iterkeys(self):
+      return iter(self)
+
+    def itervalues(self):
+      for key in self:
+        yield self[key]
+
+    def iteritems(self):
+      for key in self:
+        yield (key, self[key])
+
+    def keys(self):
+      return list(self)
+
+    def items(self):
+      return [(key, self[key]) for key in self]
+
+    def values(self):
+      return [self[key] for key in self]
+
+    # Mappings are not hashable by default, but subclasses can change this
+    __hash__ = None
+
+    def __eq__(self, other):
+      if not isinstance(other, collections.Mapping):
+        return NotImplemented
+      return dict(self.items()) == dict(other.items())
+
+    def __ne__(self, other):
+      return not (self == other)
+
+  class MutableMapping(Mapping):
+    __slots__ = ()
+
+    __marker = object()
+
+    def pop(self, key, default=__marker):
+      try:
+        value = self[key]
+      except KeyError:
+        if default is self.__marker:
+          raise
+        return default
+      else:
+        del self[key]
+        return value
+
+    def popitem(self):
+      try:
+        key = next(iter(self))
+      except StopIteration:
+        raise KeyError
+      value = self[key]
+      del self[key]
+      return key, value
+
+    def clear(self):
+      try:
+        while True:
+          self.popitem()
+      except KeyError:
+        pass
+
+    def update(*args, **kwds):
+      if len(args) > 2:
+        raise TypeError("update() takes at most 2 positional "
+                        "arguments ({} given)".format(len(args)))
+      elif not args:
+        raise TypeError("update() takes at least 1 argument (0 given)")
+      self = args[0]
+      other = args[1] if len(args) >= 2 else ()
+
+      if isinstance(other, Mapping):
+        for key in other:
+          self[key] = other[key]
+      elif hasattr(other, "keys"):
+        for key in other.keys():
+          self[key] = other[key]
+      else:
+        for key, value in other:
+          self[key] = value
+      for key, value in kwds.items():
+        self[key] = value
+
+    def setdefault(self, key, default=None):
+      try:
+        return self[key]
+      except KeyError:
+        self[key] = default
+      return default
+
+  collections.Mapping.register(Mapping)
+  collections.MutableMapping.register(MutableMapping)
+
+else:
+  # In Python 3 we can just use MutableMapping directly, because it defines
+  # __slots__.
+  MutableMapping = collections.MutableMapping
+
+
+class BaseContainer(object):
+
+  """Base container class."""
+
+  # Minimizes memory usage and disallows assignment to other attributes.
+  __slots__ = ['_message_listener', '_values']
+
+  def __init__(self, message_listener):
+    """
+    Args:
+      message_listener: A MessageListener implementation.
+        The RepeatedScalarFieldContainer will call this object's
+        Modified() method when it is modified.
+    """
+    self._message_listener = message_listener
+    self._values = []
+
+  def __getitem__(self, key):
+    """Retrieves item by the specified key."""
+    return self._values[key]
+
+  def __len__(self):
+    """Returns the number of elements in the container."""
+    return len(self._values)
+
+  def __ne__(self, other):
+    """Checks if another instance isn't equal to this one."""
+    # The concrete classes should define __eq__.
+    return not self == other
+
+  def __hash__(self):
+    raise TypeError('unhashable object')
+
+  def __repr__(self):
+    return repr(self._values)
+
+  def sort(self, *args, **kwargs):
+    # Continue to support the old sort_function keyword argument.
+    # This is expected to be a rare occurrence, so use LBYL to avoid
+    # the overhead of actually catching KeyError.
+    if 'sort_function' in kwargs:
+      kwargs['cmp'] = kwargs.pop('sort_function')
+    self._values.sort(*args, **kwargs)
+
+
+class RepeatedScalarFieldContainer(BaseContainer):
+
+  """Simple, type-checked, list-like container for holding repeated scalars."""
+
+  # Disallows assignment to other attributes.
+  __slots__ = ['_type_checker']
+
+  def __init__(self, message_listener, type_checker):
+    """
+    Args:
+      message_listener: A MessageListener implementation.
+        The RepeatedScalarFieldContainer will call this object's
+        Modified() method when it is modified.
+      type_checker: A type_checkers.ValueChecker instance to run on elements
+        inserted into this container.
+    """
+    super(RepeatedScalarFieldContainer, self).__init__(message_listener)
+    self._type_checker = type_checker
+
+  def append(self, value):
+    """Appends an item to the list. Similar to list.append()."""
+    self._values.append(self._type_checker.CheckValue(value))
+    if not self._message_listener.dirty:
+      self._message_listener.Modified()
+
+  def insert(self, key, value):
+    """Inserts the item at the specified position. Similar to list.insert()."""
+    self._values.insert(key, self._type_checker.CheckValue(value))
+    if not self._message_listener.dirty:
+      self._message_listener.Modified()
+
+  def extend(self, elem_seq):
+    """Extends by appending the given iterable. Similar to list.extend()."""
+
+    if elem_seq is None:
+      return
+    try:
+      elem_seq_iter = iter(elem_seq)
+    except TypeError:
+      if not elem_seq:
+        # silently ignore falsy inputs :-/.
+        # TODO(ptucker): Deprecate this behavior. b/18413862
+        return
+      raise
+
+    new_values = [self._type_checker.CheckValue(elem) for elem in elem_seq_iter]
+    if new_values:
+      self._values.extend(new_values)
+      self._message_listener.Modified()
+
+  def MergeFrom(self, other):
+    """Appends the contents of another repeated field of the same type to this
+    one. We do not check the types of the individual fields.
+    """
+    self._values.extend(other._values)
+    self._message_listener.Modified()
+
+  def remove(self, elem):
+    """Removes an item from the list. Similar to list.remove()."""
+    self._values.remove(elem)
+    self._message_listener.Modified()
+
+  def pop(self, key=-1):
+    """Removes and returns an item at a given index. Similar to list.pop()."""
+    value = self._values[key]
+    self.__delitem__(key)
+    return value
+
+  def __setitem__(self, key, value):
+    """Sets the item on the specified position."""
+    if isinstance(key, slice):  # PY3
+      if key.step is not None:
+        raise ValueError('Extended slices not supported')
+      self.__setslice__(key.start, key.stop, value)
+    else:
+      self._values[key] = self._type_checker.CheckValue(value)
+      self._message_listener.Modified()
+
+  def __getslice__(self, start, stop):
+    """Retrieves the subset of items from between the specified indices."""
+    return self._values[start:stop]
+
+  def __setslice__(self, start, stop, values):
+    """Sets the subset of items from between the specified indices."""
+    new_values = []
+    for value in values:
+      new_values.append(self._type_checker.CheckValue(value))
+    self._values[start:stop] = new_values
+    self._message_listener.Modified()
+
+  def __delitem__(self, key):
+    """Deletes the item at the specified position."""
+    del self._values[key]
+    self._message_listener.Modified()
+
+  def __delslice__(self, start, stop):
+    """Deletes the subset of items from between the specified indices."""
+    del self._values[start:stop]
+    self._message_listener.Modified()
+
+  def __eq__(self, other):
+    """Compares the current instance with another one."""
+    if self is other:
+      return True
+    # Special case for the same type which should be common and fast.
+    if isinstance(other, self.__class__):
+      return other._values == self._values
+    # We are presumably comparing against some other sequence type.
+    return other == self._values
+
+collections.MutableSequence.register(BaseContainer)
+
+
+class RepeatedCompositeFieldContainer(BaseContainer):
+
+  """Simple, list-like container for holding repeated composite fields."""
+
+  # Disallows assignment to other attributes.
+  __slots__ = ['_message_descriptor']
+
+  def __init__(self, message_listener, message_descriptor):
+    """
+    Note that we pass in a descriptor instead of the generated directly,
+    since at the time we construct a _RepeatedCompositeFieldContainer we
+    haven't yet necessarily initialized the type that will be contained in the
+    container.
+
+    Args:
+      message_listener: A MessageListener implementation.
+        The RepeatedCompositeFieldContainer will call this object's
+        Modified() method when it is modified.
+      message_descriptor: A Descriptor instance describing the protocol type
+        that should be present in this container.  We'll use the
+        _concrete_class field of this descriptor when the client calls add().
+    """
+    super(RepeatedCompositeFieldContainer, self).__init__(message_listener)
+    self._message_descriptor = message_descriptor
+
+  def add(self, **kwargs):
+    """Adds a new element at the end of the list and returns it. Keyword
+    arguments may be used to initialize the element.
+    """
+    new_element = self._message_descriptor._concrete_class(**kwargs)
+    new_element._SetListener(self._message_listener)
+    self._values.append(new_element)
+    if not self._message_listener.dirty:
+      self._message_listener.Modified()
+    return new_element
+
+  def extend(self, elem_seq):
+    """Extends by appending the given sequence of elements of the same type
+    as this one, copying each individual message.
+    """
+    message_class = self._message_descriptor._concrete_class
+    listener = self._message_listener
+    values = self._values
+    for message in elem_seq:
+      new_element = message_class()
+      new_element._SetListener(listener)
+      new_element.MergeFrom(message)
+      values.append(new_element)
+    listener.Modified()
+
+  def MergeFrom(self, other):
+    """Appends the contents of another repeated field of the same type to this
+    one, copying each individual message.
+    """
+    self.extend(other._values)
+
+  def remove(self, elem):
+    """Removes an item from the list. Similar to list.remove()."""
+    self._values.remove(elem)
+    self._message_listener.Modified()
+
+  def pop(self, key=-1):
+    """Removes and returns an item at a given index. Similar to list.pop()."""
+    value = self._values[key]
+    self.__delitem__(key)
+    return value
+
+  def __getslice__(self, start, stop):
+    """Retrieves the subset of items from between the specified indices."""
+    return self._values[start:stop]
+
+  def __delitem__(self, key):
+    """Deletes the item at the specified position."""
+    del self._values[key]
+    self._message_listener.Modified()
+
+  def __delslice__(self, start, stop):
+    """Deletes the subset of items from between the specified indices."""
+    del self._values[start:stop]
+    self._message_listener.Modified()
+
+  def __eq__(self, other):
+    """Compares the current instance with another one."""
+    if self is other:
+      return True
+    if not isinstance(other, self.__class__):
+      raise TypeError('Can only compare repeated composite fields against '
+                      'other repeated composite fields.')
+    return self._values == other._values
+
+
+class ScalarMap(MutableMapping):
+
+  """Simple, type-checked, dict-like container for holding repeated scalars."""
+
+  # Disallows assignment to other attributes.
+  __slots__ = ['_key_checker', '_value_checker', '_values', '_message_listener']
+
+  def __init__(self, message_listener, key_checker, value_checker):
+    """
+    Args:
+      message_listener: A MessageListener implementation.
+        The ScalarMap will call this object's Modified() method when it
+        is modified.
+      key_checker: A type_checkers.ValueChecker instance to run on keys
+        inserted into this container.
+      value_checker: A type_checkers.ValueChecker instance to run on values
+        inserted into this container.
+    """
+    self._message_listener = message_listener
+    self._key_checker = key_checker
+    self._value_checker = value_checker
+    self._values = {}
+
+  def __getitem__(self, key):
+    try:
+      return self._values[key]
+    except KeyError:
+      key = self._key_checker.CheckValue(key)
+      val = self._value_checker.DefaultValue()
+      self._values[key] = val
+      return val
+
+  def __contains__(self, item):
+    # We check the key's type to match the strong-typing flavor of the API.
+    # Also this makes it easier to match the behavior of the C++ implementation.
+    self._key_checker.CheckValue(item)
+    return item in self._values
+
+  # We need to override this explicitly, because our defaultdict-like behavior
+  # will make the default implementation (from our base class) always insert
+  # the key.
+  def get(self, key, default=None):
+    if key in self:
+      return self[key]
+    else:
+      return default
+
+  def __setitem__(self, key, value):
+    checked_key = self._key_checker.CheckValue(key)
+    checked_value = self._value_checker.CheckValue(value)
+    self._values[checked_key] = checked_value
+    self._message_listener.Modified()
+
+  def __delitem__(self, key):
+    del self._values[key]
+    self._message_listener.Modified()
+
+  def __len__(self):
+    return len(self._values)
+
+  def __iter__(self):
+    return iter(self._values)
+
+  def __repr__(self):
+    return repr(self._values)
+
+  def MergeFrom(self, other):
+    self._values.update(other._values)
+    self._message_listener.Modified()
+
+  def InvalidateIterators(self):
+    # It appears that the only way to reliably invalidate iterators to
+    # self._values is to ensure that its size changes.
+    original = self._values
+    self._values = original.copy()
+    original[None] = None
+
+  # This is defined in the abstract base, but we can do it much more cheaply.
+  def clear(self):
+    self._values.clear()
+    self._message_listener.Modified()
+
+
+class MessageMap(MutableMapping):
+
+  """Simple, type-checked, dict-like container for with submessage values."""
+
+  # Disallows assignment to other attributes.
+  __slots__ = ['_key_checker', '_values', '_message_listener',
+               '_message_descriptor']
+
+  def __init__(self, message_listener, message_descriptor, key_checker):
+    """
+    Args:
+      message_listener: A MessageListener implementation.
+        The ScalarMap will call this object's Modified() method when it
+        is modified.
+      key_checker: A type_checkers.ValueChecker instance to run on keys
+        inserted into this container.
+      value_checker: A type_checkers.ValueChecker instance to run on values
+        inserted into this container.
+    """
+    self._message_listener = message_listener
+    self._message_descriptor = message_descriptor
+    self._key_checker = key_checker
+    self._values = {}
+
+  def __getitem__(self, key):
+    try:
+      return self._values[key]
+    except KeyError:
+      key = self._key_checker.CheckValue(key)
+      new_element = self._message_descriptor._concrete_class()
+      new_element._SetListener(self._message_listener)
+      self._values[key] = new_element
+      self._message_listener.Modified()
+
+      return new_element
+
+  def get_or_create(self, key):
+    """get_or_create() is an alias for getitem (ie. map[key]).
+
+    Args:
+      key: The key to get or create in the map.
+
+    This is useful in cases where you want to be explicit that the call is
+    mutating the map.  This can avoid lint errors for statements like this
+    that otherwise would appear to be pointless statements:
+
+      msg.my_map[key]
+    """
+    return self[key]
+
+  # We need to override this explicitly, because our defaultdict-like behavior
+  # will make the default implementation (from our base class) always insert
+  # the key.
+  def get(self, key, default=None):
+    if key in self:
+      return self[key]
+    else:
+      return default
+
+  def __contains__(self, item):
+    return item in self._values
+
+  def __setitem__(self, key, value):
+    raise ValueError('May not set values directly, call my_map[key].foo = 5')
+
+  def __delitem__(self, key):
+    del self._values[key]
+    self._message_listener.Modified()
+
+  def __len__(self):
+    return len(self._values)
+
+  def __iter__(self):
+    return iter(self._values)
+
+  def __repr__(self):
+    return repr(self._values)
+
+  def MergeFrom(self, other):
+    for key in other:
+      self[key].MergeFrom(other[key])
+    # self._message_listener.Modified() not required here, because
+    # mutations to submessages already propagate.
+
+  def InvalidateIterators(self):
+    # It appears that the only way to reliably invalidate iterators to
+    # self._values is to ensure that its size changes.
+    original = self._values
+    self._values = original.copy()
+    original[None] = None
+
+  # This is defined in the abstract base, but we can do it much more cheaply.
+  def clear(self):
+    self._values.clear()
+    self._message_listener.Modified()
diff --git a/python/google/protobuf/internal/decoder.py b/python/google/protobuf/internal/decoder.py
new file mode 100755
index 0000000..31869e4
--- /dev/null
+++ b/python/google/protobuf/internal/decoder.py
@@ -0,0 +1,854 @@
+# Protocol Buffers - Google's data interchange format
+# Copyright 2008 Google Inc.  All rights reserved.
+# https://developers.google.com/protocol-buffers/
+#
+# Redistribution and use in source and binary forms, with or without
+# modification, are permitted provided that the following conditions are
+# met:
+#
+#     * Redistributions of source code must retain the above copyright
+# notice, this list of conditions and the following disclaimer.
+#     * Redistributions in binary form must reproduce the above
+# copyright notice, this list of conditions and the following disclaimer
+# in the documentation and/or other materials provided with the
+# distribution.
+#     * Neither the name of Google Inc. nor the names of its
+# contributors may be used to endorse or promote products derived from
+# this software without specific prior written permission.
+#
+# THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+# "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
+# LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
+# A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
+# OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
+# SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
+# LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
+# DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
+# THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
+# (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+# OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+
+"""Code for decoding protocol buffer primitives.
+
+This code is very similar to encoder.py -- read the docs for that module first.
+
+A "decoder" is a function with the signature:
+  Decode(buffer, pos, end, message, field_dict)
+The arguments are:
+  buffer:     The string containing the encoded message.
+  pos:        The current position in the string.
+  end:        The position in the string where the current message ends.  May be
+              less than len(buffer) if we're reading a sub-message.
+  message:    The message object into which we're parsing.
+  field_dict: message._fields (avoids a hashtable lookup).
+The decoder reads the field and stores it into field_dict, returning the new
+buffer position.  A decoder for a repeated field may proactively decode all of
+the elements of that field, if they appear consecutively.
+
+Note that decoders may throw any of the following:
+  IndexError:  Indicates a truncated message.
+  struct.error:  Unpacking of a fixed-width field failed.
+  message.DecodeError:  Other errors.
+
+Decoders are expected to raise an exception if they are called with pos > end.
+This allows callers to be lax about bounds checking:  it's fineto read past
+"end" as long as you are sure that someone else will notice and throw an
+exception later on.
+
+Something up the call stack is expected to catch IndexError and struct.error
+and convert them to message.DecodeError.
+
+Decoders are constructed using decoder constructors with the signature:
+  MakeDecoder(field_number, is_repeated, is_packed, key, new_default)
+The arguments are:
+  field_number:  The field number of the field we want to decode.
+  is_repeated:   Is the field a repeated field? (bool)
+  is_packed:     Is the field a packed field? (bool)
+  key:           The key to use when looking up the field within field_dict.
+                 (This is actually the FieldDescriptor but nothing in this
+                 file should depend on that.)
+  new_default:   A function which takes a message object as a parameter and
+                 returns a new instance of the default value for this field.
+                 (This is called for repeated fields and sub-messages, when an
+                 instance does not already exist.)
+
+As with encoders, we define a decoder constructor for every type of field.
+Then, for every field of every message class we construct an actual decoder.
+That decoder goes into a dict indexed by tag, so when we decode a message
+we repeatedly read a tag, look up the corresponding decoder, and invoke it.
+"""
+
+__author__ = 'kenton@google.com (Kenton Varda)'
+
+import struct
+
+import six
+
+if six.PY3:
+  long = int
+
+from google.protobuf.internal import encoder
+from google.protobuf.internal import wire_format
+from google.protobuf import message
+
+
+# This will overflow and thus become IEEE-754 "infinity".  We would use
+# "float('inf')" but it doesn't work on Windows pre-Python-2.6.
+_POS_INF = 1e10000
+_NEG_INF = -_POS_INF
+_NAN = _POS_INF * 0
+
+
+# This is not for optimization, but rather to avoid conflicts with local
+# variables named "message".
+_DecodeError = message.DecodeError
+
+
+def _VarintDecoder(mask, result_type):
+  """Return an encoder for a basic varint value (does not include tag).
+
+  Decoded values will be bitwise-anded with the given mask before being
+  returned, e.g. to limit them to 32 bits.  The returned decoder does not
+  take the usual "end" parameter -- the caller is expected to do bounds checking
+  after the fact (often the caller can defer such checking until later).  The
+  decoder returns a (value, new_pos) pair.
+  """
+
+  def DecodeVarint(buffer, pos):
+    result = 0
+    shift = 0
+    while 1:
+      b = six.indexbytes(buffer, pos)
+      result |= ((b & 0x7f) << shift)
+      pos += 1
+      if not (b & 0x80):
+        result &= mask
+        result = result_type(result)
+        return (result, pos)
+      shift += 7
+      if shift >= 64:
+        raise _DecodeError('Too many bytes when decoding varint.')
+  return DecodeVarint
+
+
+def _SignedVarintDecoder(mask, result_type):
+  """Like _VarintDecoder() but decodes signed values."""
+
+  def DecodeVarint(buffer, pos):
+    result = 0
+    shift = 0
+    while 1:
+      b = six.indexbytes(buffer, pos)
+      result |= ((b & 0x7f) << shift)
+      pos += 1
+      if not (b & 0x80):
+        if result > 0x7fffffffffffffff:
+          result -= (1 << 64)
+          result |= ~mask
+        else:
+          result &= mask
+        result = result_type(result)
+        return (result, pos)
+      shift += 7
+      if shift >= 64:
+        raise _DecodeError('Too many bytes when decoding varint.')
+  return DecodeVarint
+
+# We force 32-bit values to int and 64-bit values to long to make
+# alternate implementations where the distinction is more significant
+# (e.g. the C++ implementation) simpler.
+
+_DecodeVarint = _VarintDecoder((1 << 64) - 1, long)
+_DecodeSignedVarint = _SignedVarintDecoder((1 << 64) - 1, long)
+
+# Use these versions for values which must be limited to 32 bits.
+_DecodeVarint32 = _VarintDecoder((1 << 32) - 1, int)
+_DecodeSignedVarint32 = _SignedVarintDecoder((1 << 32) - 1, int)
+
+
+def ReadTag(buffer, pos):
+  """Read a tag from the buffer, and return a (tag_bytes, new_pos) tuple.
+
+  We return the raw bytes of the tag rather than decoding them.  The raw
+  bytes can then be used to look up the proper decoder.  This effectively allows
+  us to trade some work that would be done in pure-python (decoding a varint)
+  for work that is done in C (searching for a byte string in a hash table).
+  In a low-level language it would be much cheaper to decode the varint and
+  use that, but not in Python.
+  """
+
+  start = pos
+  while six.indexbytes(buffer, pos) & 0x80:
+    pos += 1
+  pos += 1
+  return (buffer[start:pos], pos)
+
+
+# --------------------------------------------------------------------
+
+
+def _SimpleDecoder(wire_type, decode_value):
+  """Return a constructor for a decoder for fields of a particular type.
+
+  Args:
+      wire_type:  The field's wire type.
+      decode_value:  A function which decodes an individual value, e.g.
+        _DecodeVarint()
+  """
+
+  def SpecificDecoder(field_number, is_repeated, is_packed, key, new_default):
+    if is_packed:
+      local_DecodeVarint = _DecodeVarint
+      def DecodePackedField(buffer, pos, end, message, field_dict):
+        value = field_dict.get(key)
+        if value is None:
+          value = field_dict.setdefault(key, new_default(message))
+        (endpoint, pos) = local_DecodeVarint(buffer, pos)
+        endpoint += pos
+        if endpoint > end:
+          raise _DecodeError('Truncated message.')
+        while pos < endpoint:
+          (element, pos) = decode_value(buffer, pos)
+          value.append(element)
+        if pos > endpoint:
+          del value[-1]   # Discard corrupt value.
+          raise _DecodeError('Packed element was truncated.')
+        return pos
+      return DecodePackedField
+    elif is_repeated:
+      tag_bytes = encoder.TagBytes(field_number, wire_type)
+      tag_len = len(tag_bytes)
+      def DecodeRepeatedField(buffer, pos, end, message, field_dict):
+        value = field_dict.get(key)
+        if value is None:
+          value = field_dict.setdefault(key, new_default(message))
+        while 1:
+          (element, new_pos) = decode_value(buffer, pos)
+          value.append(element)
+          # Predict that the next tag is another copy of the same repeated
+          # field.
+          pos = new_pos + tag_len
+          if buffer[new_pos:pos] != tag_bytes or new_pos >= end:
+            # Prediction failed.  Return.
+            if new_pos > end:
+              raise _DecodeError('Truncated message.')
+            return new_pos
+      return DecodeRepeatedField
+    else:
+      def DecodeField(buffer, pos, end, message, field_dict):
+        (field_dict[key], pos) = decode_value(buffer, pos)
+        if pos > end:
+          del field_dict[key]  # Discard corrupt value.
+          raise _DecodeError('Truncated message.')
+        return pos
+      return DecodeField
+
+  return SpecificDecoder
+
+
+def _ModifiedDecoder(wire_type, decode_value, modify_value):
+  """Like SimpleDecoder but additionally invokes modify_value on every value
+  before storing it.  Usually modify_value is ZigZagDecode.
+  """
+
+  # Reusing _SimpleDecoder is slightly slower than copying a bunch of code, but
+  # not enough to make a significant difference.
+
+  def InnerDecode(buffer, pos):
+    (result, new_pos) = decode_value(buffer, pos)
+    return (modify_value(result), new_pos)
+  return _SimpleDecoder(wire_type, InnerDecode)
+
+
+def _StructPackDecoder(wire_type, format):
+  """Return a constructor for a decoder for a fixed-width field.
+
+  Args:
+      wire_type:  The field's wire type.
+      format:  The format string to pass to struct.unpack().
+  """
+
+  value_size = struct.calcsize(format)
+  local_unpack = struct.unpack
+
+  # Reusing _SimpleDecoder is slightly slower than copying a bunch of code, but
+  # not enough to make a significant difference.
+
+  # Note that we expect someone up-stack to catch struct.error and convert
+  # it to _DecodeError -- this way we don't have to set up exception-
+  # handling blocks every time we parse one value.
+
+  def InnerDecode(buffer, pos):
+    new_pos = pos + value_size
+    result = local_unpack(format, buffer[pos:new_pos])[0]
+    return (result, new_pos)
+  return _SimpleDecoder(wire_type, InnerDecode)
+
+
+def _FloatDecoder():
+  """Returns a decoder for a float field.
+
+  This code works around a bug in struct.unpack for non-finite 32-bit
+  floating-point values.
+  """
+
+  local_unpack = struct.unpack
+
+  def InnerDecode(buffer, pos):
+    # We expect a 32-bit value in little-endian byte order.  Bit 1 is the sign
+    # bit, bits 2-9 represent the exponent, and bits 10-32 are the significand.
+    new_pos = pos + 4
+    float_bytes = buffer[pos:new_pos]
+
+    # If this value has all its exponent bits set, then it's non-finite.
+    # In Python 2.4, struct.unpack will convert it to a finite 64-bit value.
+    # To avoid that, we parse it specially.
+    if (float_bytes[3:4] in b'\x7F\xFF' and float_bytes[2:3] >= b'\x80'):
+      # If at least one significand bit is set...
+      if float_bytes[0:3] != b'\x00\x00\x80':
+        return (_NAN, new_pos)
+      # If sign bit is set...
+      if float_bytes[3:4] == b'\xFF':
+        return (_NEG_INF, new_pos)
+      return (_POS_INF, new_pos)
+
+    # Note that we expect someone up-stack to catch struct.error and convert
+    # it to _DecodeError -- this way we don't have to set up exception-
+    # handling blocks every time we parse one value.
+    result = local_unpack('<f', float_bytes)[0]
+    return (result, new_pos)
+  return _SimpleDecoder(wire_format.WIRETYPE_FIXED32, InnerDecode)
+
+
+def _DoubleDecoder():
+  """Returns a decoder for a double field.
+
+  This code works around a bug in struct.unpack for not-a-number.
+  """
+
+  local_unpack = struct.unpack
+
+  def InnerDecode(buffer, pos):
+    # We expect a 64-bit value in little-endian byte order.  Bit 1 is the sign
+    # bit, bits 2-12 represent the exponent, and bits 13-64 are the significand.
+    new_pos = pos + 8
+    double_bytes = buffer[pos:new_pos]
+
+    # If this value has all its exponent bits set and at least one significand
+    # bit set, it's not a number.  In Python 2.4, struct.unpack will treat it
+    # as inf or -inf.  To avoid that, we treat it specially.
+    if ((double_bytes[7:8] in b'\x7F\xFF')
+        and (double_bytes[6:7] >= b'\xF0')
+        and (double_bytes[0:7] != b'\x00\x00\x00\x00\x00\x00\xF0')):
+      return (_NAN, new_pos)
+
+    # Note that we expect someone up-stack to catch struct.error and convert
+    # it to _DecodeError -- this way we don't have to set up exception-
+    # handling blocks every time we parse one value.
+    result = local_unpack('<d', double_bytes)[0]
+    return (result, new_pos)
+  return _SimpleDecoder(wire_format.WIRETYPE_FIXED64, InnerDecode)
+
+
+def EnumDecoder(field_number, is_repeated, is_packed, key, new_default):
+  enum_type = key.enum_type
+  if is_packed:
+    local_DecodeVarint = _DecodeVarint
+    def DecodePackedField(buffer, pos, end, message, field_dict):
+      value = field_dict.get(key)
+      if value is None:
+        value = field_dict.setdefault(key, new_default(message))
+      (endpoint, pos) = local_DecodeVarint(buffer, pos)
+      endpoint += pos
+      if endpoint > end:
+        raise _DecodeError('Truncated message.')
+      while pos < endpoint:
+        value_start_pos = pos
+        (element, pos) = _DecodeSignedVarint32(buffer, pos)
+        if element in enum_type.values_by_number:
+          value.append(element)
+        else:
+          if not message._unknown_fields:
+            message._unknown_fields = []
+          tag_bytes = encoder.TagBytes(field_number,
+                                       wire_format.WIRETYPE_VARINT)
+          message._unknown_fields.append(
+              (tag_bytes, buffer[value_start_pos:pos]))
+      if pos > endpoint:
+        if element in enum_type.values_by_number:
+          del value[-1]   # Discard corrupt value.
+        else:
+          del message._unknown_fields[-1]
+        raise _DecodeError('Packed element was truncated.')
+      return pos
+    return DecodePackedField
+  elif is_repeated:
+    tag_bytes = encoder.TagBytes(field_number, wire_format.WIRETYPE_VARINT)
+    tag_len = len(tag_bytes)
+    def DecodeRepeatedField(buffer, pos, end, message, field_dict):
+      value = field_dict.get(key)
+      if value is None:
+        value = field_dict.setdefault(key, new_default(message))
+      while 1:
+        (element, new_pos) = _DecodeSignedVarint32(buffer, pos)
+        if element in enum_type.values_by_number:
+          value.append(element)
+        else:
+          if not message._unknown_fields:
+            message._unknown_fields = []
+          message._unknown_fields.append(
+              (tag_bytes, buffer[pos:new_pos]))
+        # Predict that the next tag is another copy of the same repeated
+        # field.
+        pos = new_pos + tag_len
+        if buffer[new_pos:pos] != tag_bytes or new_pos >= end:
+          # Prediction failed.  Return.
+          if new_pos > end:
+            raise _DecodeError('Truncated message.')
+          return new_pos
+    return DecodeRepeatedField
+  else:
+    def DecodeField(buffer, pos, end, message, field_dict):
+      value_start_pos = pos
+      (enum_value, pos) = _DecodeSignedVarint32(buffer, pos)
+      if pos > end:
+        raise _DecodeError('Truncated message.')
+      if enum_value in enum_type.values_by_number:
+        field_dict[key] = enum_value
+      else:
+        if not message._unknown_fields:
+          message._unknown_fields = []
+        tag_bytes = encoder.TagBytes(field_number,
+                                     wire_format.WIRETYPE_VARINT)
+        message._unknown_fields.append(
+          (tag_bytes, buffer[value_start_pos:pos]))
+      return pos
+    return DecodeField
+
+
+# --------------------------------------------------------------------
+
+
+Int32Decoder = _SimpleDecoder(
+    wire_format.WIRETYPE_VARINT, _DecodeSignedVarint32)
+
+Int64Decoder = _SimpleDecoder(
+    wire_format.WIRETYPE_VARINT, _DecodeSignedVarint)
+
+UInt32Decoder = _SimpleDecoder(wire_format.WIRETYPE_VARINT, _DecodeVarint32)
+UInt64Decoder = _SimpleDecoder(wire_format.WIRETYPE_VARINT, _DecodeVarint)
+
+SInt32Decoder = _ModifiedDecoder(
+    wire_format.WIRETYPE_VARINT, _DecodeVarint32, wire_format.ZigZagDecode)
+SInt64Decoder = _ModifiedDecoder(
+    wire_format.WIRETYPE_VARINT, _DecodeVarint, wire_format.ZigZagDecode)
+
+# Note that Python conveniently guarantees that when using the '<' prefix on
+# formats, they will also have the same size across all platforms (as opposed
+# to without the prefix, where their sizes depend on the C compiler's basic
+# type sizes).
+Fixed32Decoder  = _StructPackDecoder(wire_format.WIRETYPE_FIXED32, '<I')
+Fixed64Decoder  = _StructPackDecoder(wire_format.WIRETYPE_FIXED64, '<Q')
+SFixed32Decoder = _StructPackDecoder(wire_format.WIRETYPE_FIXED32, '<i')
+SFixed64Decoder = _StructPackDecoder(wire_format.WIRETYPE_FIXED64, '<q')
+FloatDecoder = _FloatDecoder()
+DoubleDecoder = _DoubleDecoder()
+
+BoolDecoder = _ModifiedDecoder(
+    wire_format.WIRETYPE_VARINT, _DecodeVarint, bool)
+
+
+def StringDecoder(field_number, is_repeated, is_packed, key, new_default):
+  """Returns a decoder for a string field."""
+
+  local_DecodeVarint = _DecodeVarint
+  local_unicode = six.text_type
+
+  def _ConvertToUnicode(byte_str):
+    try:
+      return local_unicode(byte_str, 'utf-8')
+    except UnicodeDecodeError as e:
+      # add more information to the error message and re-raise it.
+      e.reason = '%s in field: %s' % (e, key.full_name)
+      raise
+
+  assert not is_packed
+  if is_repeated:
+    tag_bytes = encoder.TagBytes(field_number,
+                                 wire_format.WIRETYPE_LENGTH_DELIMITED)
+    tag_len = len(tag_bytes)
+    def DecodeRepeatedField(buffer, pos, end, message, field_dict):
+      value = field_dict.get(key)
+      if value is None:
+        value = field_dict.setdefault(key, new_default(message))
+      while 1:
+        (size, pos) = local_DecodeVarint(buffer, pos)
+        new_pos = pos + size
+        if new_pos > end:
+          raise _DecodeError('Truncated string.')
+        value.append(_ConvertToUnicode(buffer[pos:new_pos]))
+        # Predict that the next tag is another copy of the same repeated field.
+        pos = new_pos + tag_len
+        if buffer[new_pos:pos] != tag_bytes or new_pos == end:
+          # Prediction failed.  Return.
+          return new_pos
+    return DecodeRepeatedField
+  else:
+    def DecodeField(buffer, pos, end, message, field_dict):
+      (size, pos) = local_DecodeVarint(buffer, pos)
+      new_pos = pos + size
+      if new_pos > end:
+        raise _DecodeError('Truncated string.')
+      field_dict[key] = _ConvertToUnicode(buffer[pos:new_pos])
+      return new_pos
+    return DecodeField
+
+
+def BytesDecoder(field_number, is_repeated, is_packed, key, new_default):
+  """Returns a decoder for a bytes field."""
+
+  local_DecodeVarint = _DecodeVarint
+
+  assert not is_packed
+  if is_repeated:
+    tag_bytes = encoder.TagBytes(field_number,
+                                 wire_format.WIRETYPE_LENGTH_DELIMITED)
+    tag_len = len(tag_bytes)
+    def DecodeRepeatedField(buffer, pos, end, message, field_dict):
+      value = field_dict.get(key)
+      if value is None:
+        value = field_dict.setdefault(key, new_default(message))
+      while 1:
+        (size, pos) = local_DecodeVarint(buffer, pos)
+        new_pos = pos + size
+        if new_pos > end:
+          raise _DecodeError('Truncated string.')
+        value.append(buffer[pos:new_pos])
+        # Predict that the next tag is another copy of the same repeated field.
+        pos = new_pos + tag_len
+        if buffer[new_pos:pos] != tag_bytes or new_pos == end:
+          # Prediction failed.  Return.
+          return new_pos
+    return DecodeRepeatedField
+  else:
+    def DecodeField(buffer, pos, end, message, field_dict):
+      (size, pos) = local_DecodeVarint(buffer, pos)
+      new_pos = pos + size
+      if new_pos > end:
+        raise _DecodeError('Truncated string.')
+      field_dict[key] = buffer[pos:new_pos]
+      return new_pos
+    return DecodeField
+
+
+def GroupDecoder(field_number, is_repeated, is_packed, key, new_default):
+  """Returns a decoder for a group field."""
+
+  end_tag_bytes = encoder.TagBytes(field_number,
+                                   wire_format.WIRETYPE_END_GROUP)
+  end_tag_len = len(end_tag_bytes)
+
+  assert not is_packed
+  if is_repeated:
+    tag_bytes = encoder.TagBytes(field_number,
+                                 wire_format.WIRETYPE_START_GROUP)
+    tag_len = len(tag_bytes)
+    def DecodeRepeatedField(buffer, pos, end, message, field_dict):
+      value = field_dict.get(key)
+      if value is None:
+        value = field_dict.setdefault(key, new_default(message))
+      while 1:
+        value = field_dict.get(key)
+        if value is None:
+          value = field_dict.setdefault(key, new_default(message))
+        # Read sub-message.
+        pos = value.add()._InternalParse(buffer, pos, end)
+        # Read end tag.
+        new_pos = pos+end_tag_len
+        if buffer[pos:new_pos] != end_tag_bytes or new_pos > end:
+          raise _DecodeError('Missing group end tag.')
+        # Predict that the next tag is another copy of the same repeated field.
+        pos = new_pos + tag_len
+        if buffer[new_pos:pos] != tag_bytes or new_pos == end:
+          # Prediction failed.  Return.
+          return new_pos
+    return DecodeRepeatedField
+  else:
+    def DecodeField(buffer, pos, end, message, field_dict):
+      value = field_dict.get(key)
+      if value is None:
+        value = field_dict.setdefault(key, new_default(message))
+      # Read sub-message.
+      pos = value._InternalParse(buffer, pos, end)
+      # Read end tag.
+      new_pos = pos+end_tag_len
+      if buffer[pos:new_pos] != end_tag_bytes or new_pos > end:
+        raise _DecodeError('Missing group end tag.')
+      return new_pos
+    return DecodeField
+
+
+def MessageDecoder(field_number, is_repeated, is_packed, key, new_default):
+  """Returns a decoder for a message field."""
+
+  local_DecodeVarint = _DecodeVarint
+
+  assert not is_packed
+  if is_repeated:
+    tag_bytes = encoder.TagBytes(field_number,
+                                 wire_format.WIRETYPE_LENGTH_DELIMITED)
+    tag_len = len(tag_bytes)
+    def DecodeRepeatedField(buffer, pos, end, message, field_dict):
+      value = field_dict.get(key)
+      if value is None:
+        value = field_dict.setdefault(key, new_default(message))
+      while 1:
+        # Read length.
+        (size, pos) = local_DecodeVarint(buffer, pos)
+        new_pos = pos + size
+        if new_pos > end:
+          raise _DecodeError('Truncated message.')
+        # Read sub-message.
+        if value.add()._InternalParse(buffer, pos, new_pos) != new_pos:
+          # The only reason _InternalParse would return early is if it
+          # encountered an end-group tag.
+          raise _DecodeError('Unexpected end-group tag.')
+        # Predict that the next tag is another copy of the same repeated field.
+        pos = new_pos + tag_len
+        if buffer[new_pos:pos] != tag_bytes or new_pos == end:
+          # Prediction failed.  Return.
+          return new_pos
+    return DecodeRepeatedField
+  else:
+    def DecodeField(buffer, pos, end, message, field_dict):
+      value = field_dict.get(key)
+      if value is None:
+        value = field_dict.setdefault(key, new_default(message))
+      # Read length.
+      (size, pos) = local_DecodeVarint(buffer, pos)
+      new_pos = pos + size
+      if new_pos > end:
+        raise _DecodeError('Truncated message.')
+      # Read sub-message.
+      if value._InternalParse(buffer, pos, new_pos) != new_pos:
+        # The only reason _InternalParse would return early is if it encountered
+        # an end-group tag.
+        raise _DecodeError('Unexpected end-group tag.')
+      return new_pos
+    return DecodeField
+
+
+# --------------------------------------------------------------------
+
+MESSAGE_SET_ITEM_TAG = encoder.TagBytes(1, wire_format.WIRETYPE_START_GROUP)
+
+def MessageSetItemDecoder(extensions_by_number):
+  """Returns a decoder for a MessageSet item.
+
+  The parameter is the _extensions_by_number map for the message class.
+
+  The message set message looks like this:
+    message MessageSet {
+      repeated group Item = 1 {
+        required int32 type_id = 2;
+        required string message = 3;
+      }
+    }
+  """
+
+  type_id_tag_bytes = encoder.TagBytes(2, wire_format.WIRETYPE_VARINT)
+  message_tag_bytes = encoder.TagBytes(3, wire_format.WIRETYPE_LENGTH_DELIMITED)
+  item_end_tag_bytes = encoder.TagBytes(1, wire_format.WIRETYPE_END_GROUP)
+
+  local_ReadTag = ReadTag
+  local_DecodeVarint = _DecodeVarint
+  local_SkipField = SkipField
+
+  def DecodeItem(buffer, pos, end, message, field_dict):
+    message_set_item_start = pos
+    type_id = -1
+    message_start = -1
+    message_end = -1
+
+    # Technically, type_id and message can appear in any order, so we need
+    # a little loop here.
+    while 1:
+      (tag_bytes, pos) = local_ReadTag(buffer, pos)
+      if tag_bytes == type_id_tag_bytes:
+        (type_id, pos) = local_DecodeVarint(buffer, pos)
+      elif tag_bytes == message_tag_bytes:
+        (size, message_start) = local_DecodeVarint(buffer, pos)
+        pos = message_end = message_start + size
+      elif tag_bytes == item_end_tag_bytes:
+        break
+      else:
+        pos = SkipField(buffer, pos, end, tag_bytes)
+        if pos == -1:
+          raise _DecodeError('Missing group end tag.')
+
+    if pos > end:
+      raise _DecodeError('Truncated message.')
+
+    if type_id == -1:
+      raise _DecodeError('MessageSet item missing type_id.')
+    if message_start == -1:
+      raise _DecodeError('MessageSet item missing message.')
+
+    extension = extensions_by_number.get(type_id)
+    if extension is not None:
+      value = field_dict.get(extension)
+      if value is None:
+        value = field_dict.setdefault(
+            extension, extension.message_type._concrete_class())
+      if value._InternalParse(buffer, message_start,message_end) != message_end:
+        # The only reason _InternalParse would return early is if it encountered
+        # an end-group tag.
+        raise _DecodeError('Unexpected end-group tag.')
+    else:
+      if not message._unknown_fields:
+        message._unknown_fields = []
+      message._unknown_fields.append((MESSAGE_SET_ITEM_TAG,
+                                      buffer[message_set_item_start:pos]))
+
+    return pos
+
+  return DecodeItem
+
+# --------------------------------------------------------------------
+
+def MapDecoder(field_descriptor, new_default, is_message_map):
+  """Returns a decoder for a map field."""
+
+  key = field_descriptor
+  tag_bytes = encoder.TagBytes(field_descriptor.number,
+                               wire_format.WIRETYPE_LENGTH_DELIMITED)
+  tag_len = len(tag_bytes)
+  local_DecodeVarint = _DecodeVarint
+  # Can't read _concrete_class yet; might not be initialized.
+  message_type = field_descriptor.message_type
+
+  def DecodeMap(buffer, pos, end, message, field_dict):
+    submsg = message_type._concrete_class()
+    value = field_dict.get(key)
+    if value is None:
+      value = field_dict.setdefault(key, new_default(message))
+    while 1:
+      # Read length.
+      (size, pos) = local_DecodeVarint(buffer, pos)
+      new_pos = pos + size
+      if new_pos > end:
+        raise _DecodeError('Truncated message.')
+      # Read sub-message.
+      submsg.Clear()
+      if submsg._InternalParse(buffer, pos, new_pos) != new_pos:
+        # The only reason _InternalParse would return early is if it
+        # encountered an end-group tag.
+        raise _DecodeError('Unexpected end-group tag.')
+
+      if is_message_map:
+        value[submsg.key].MergeFrom(submsg.value)
+      else:
+        value[submsg.key] = submsg.value
+
+      # Predict that the next tag is another copy of the same repeated field.
+      pos = new_pos + tag_len
+      if buffer[new_pos:pos] != tag_bytes or new_pos == end:
+        # Prediction failed.  Return.
+        return new_pos
+
+  return DecodeMap
+
+# --------------------------------------------------------------------
+# Optimization is not as heavy here because calls to SkipField() are rare,
+# except for handling end-group tags.
+
+def _SkipVarint(buffer, pos, end):
+  """Skip a varint value.  Returns the new position."""
+  # Previously ord(buffer[pos]) raised IndexError when pos is out of range.
+  # With this code, ord(b'') raises TypeError.  Both are handled in
+  # python_message.py to generate a 'Truncated message' error.
+  while ord(buffer[pos:pos+1]) & 0x80:
+    pos += 1
+  pos += 1
+  if pos > end:
+    raise _DecodeError('Truncated message.')
+  return pos
+
+def _SkipFixed64(buffer, pos, end):
+  """Skip a fixed64 value.  Returns the new position."""
+
+  pos += 8
+  if pos > end:
+    raise _DecodeError('Truncated message.')
+  return pos
+
+def _SkipLengthDelimited(buffer, pos, end):
+  """Skip a length-delimited value.  Returns the new position."""
+
+  (size, pos) = _DecodeVarint(buffer, pos)
+  pos += size
+  if pos > end:
+    raise _DecodeError('Truncated message.')
+  return pos
+
+def _SkipGroup(buffer, pos, end):
+  """Skip sub-group.  Returns the new position."""
+
+  while 1:
+    (tag_bytes, pos) = ReadTag(buffer, pos)
+    new_pos = SkipField(buffer, pos, end, tag_bytes)
+    if new_pos == -1:
+      return pos
+    pos = new_pos
+
+def _EndGroup(buffer, pos, end):
+  """Skipping an END_GROUP tag returns -1 to tell the parent loop to break."""
+
+  return -1
+
+def _SkipFixed32(buffer, pos, end):
+  """Skip a fixed32 value.  Returns the new position."""
+
+  pos += 4
+  if pos > end:
+    raise _DecodeError('Truncated message.')
+  return pos
+
+def _RaiseInvalidWireType(buffer, pos, end):
+  """Skip function for unknown wire types.  Raises an exception."""
+
+  raise _DecodeError('Tag had invalid wire type.')
+
+def _FieldSkipper():
+  """Constructs the SkipField function."""
+
+  WIRETYPE_TO_SKIPPER = [
+      _SkipVarint,
+      _SkipFixed64,
+      _SkipLengthDelimited,
+      _SkipGroup,
+      _EndGroup,
+      _SkipFixed32,
+      _RaiseInvalidWireType,
+      _RaiseInvalidWireType,
+      ]
+
+  wiretype_mask = wire_format.TAG_TYPE_MASK
+
+  def SkipField(buffer, pos, end, tag_bytes):
+    """Skips a field with the specified tag.
+
+    |pos| should point to the byte immediately after the tag.
+
+    Returns:
+        The new position (after the tag value), or -1 if the tag is an end-group
+        tag (in which case the calling loop should break).
+    """
+
+    # The wire type is always in the first byte since varints are little-endian.
+    wire_type = ord(tag_bytes[0:1]) & wiretype_mask
+    return WIRETYPE_TO_SKIPPER[wire_type](buffer, pos, end)
+
+  return SkipField
+
+SkipField = _FieldSkipper()
diff --git a/python/google/protobuf/internal/descriptor_database_test.py b/python/google/protobuf/internal/descriptor_database_test.py
new file mode 100644
index 0000000..1baff7d
--- /dev/null
+++ b/python/google/protobuf/internal/descriptor_database_test.py
@@ -0,0 +1,68 @@
+#! /usr/bin/env python
+#
+# Protocol Buffers - Google's data interchange format
+# Copyright 2008 Google Inc.  All rights reserved.
+# https://developers.google.com/protocol-buffers/
+#
+# Redistribution and use in source and binary forms, with or without
+# modification, are permitted provided that the following conditions are
+# met:
+#
+#     * Redistributions of source code must retain the above copyright
+# notice, this list of conditions and the following disclaimer.
+#     * Redistributions in binary form must reproduce the above
+# copyright notice, this list of conditions and the following disclaimer
+# in the documentation and/or other materials provided with the
+# distribution.
+#     * Neither the name of Google Inc. nor the names of its
+# contributors may be used to endorse or promote products derived from
+# this software without specific prior written permission.
+#
+# THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+# "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
+# LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
+# A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
+# OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
+# SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
+# LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
+# DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
+# THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
+# (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+# OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+
+"""Tests for google.protobuf.descriptor_database."""
+
+__author__ = 'matthewtoia@google.com (Matt Toia)'
+
+try:
+  import unittest2 as unittest
+except ImportError:
+  import unittest
+from google.protobuf import descriptor_pb2
+from google.protobuf.internal import factory_test2_pb2
+from google.protobuf import descriptor_database
+
+
+class DescriptorDatabaseTest(unittest.TestCase):
+
+  def testAdd(self):
+    db = descriptor_database.DescriptorDatabase()
+    file_desc_proto = descriptor_pb2.FileDescriptorProto.FromString(
+        factory_test2_pb2.DESCRIPTOR.serialized_pb)
+    db.Add(file_desc_proto)
+
+    self.assertEqual(file_desc_proto, db.FindFileByName(
+        'google/protobuf/internal/factory_test2.proto'))
+    self.assertEqual(file_desc_proto, db.FindFileContainingSymbol(
+        'google.protobuf.python.internal.Factory2Message'))
+    self.assertEqual(file_desc_proto, db.FindFileContainingSymbol(
+        'google.protobuf.python.internal.Factory2Message.NestedFactory2Message'))
+    self.assertEqual(file_desc_proto, db.FindFileContainingSymbol(
+        'google.protobuf.python.internal.Factory2Enum'))
+    self.assertEqual(file_desc_proto, db.FindFileContainingSymbol(
+        'google.protobuf.python.internal.Factory2Message.NestedFactory2Enum'))
+    self.assertEqual(file_desc_proto, db.FindFileContainingSymbol(
+        'google.protobuf.python.internal.MessageWithNestedEnumOnly.NestedEnum'))
+
+if __name__ == '__main__':
+  unittest.main()
diff --git a/python/google/protobuf/internal/descriptor_pool_test.py b/python/google/protobuf/internal/descriptor_pool_test.py
new file mode 100644
index 0000000..f1d6bf9
--- /dev/null
+++ b/python/google/protobuf/internal/descriptor_pool_test.py
@@ -0,0 +1,763 @@
+#! /usr/bin/env python
+#
+# Protocol Buffers - Google's data interchange format
+# Copyright 2008 Google Inc.  All rights reserved.
+# https://developers.google.com/protocol-buffers/
+#
+# Redistribution and use in source and binary forms, with or without
+# modification, are permitted provided that the following conditions are
+# met:
+#
+#     * Redistributions of source code must retain the above copyright
+# notice, this list of conditions and the following disclaimer.
+#     * Redistributions in binary form must reproduce the above
+# copyright notice, this list of conditions and the following disclaimer
+# in the documentation and/or other materials provided with the
+# distribution.
+#     * Neither the name of Google Inc. nor the names of its
+# contributors may be used to endorse or promote products derived from
+# this software without specific prior written permission.
+#
+# THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+# "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
+# LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
+# A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
+# OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
+# SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
+# LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
+# DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
+# THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
+# (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+# OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+
+"""Tests for google.protobuf.descriptor_pool."""
+
+__author__ = 'matthewtoia@google.com (Matt Toia)'
+
+import os
+
+try:
+  import unittest2 as unittest
+except ImportError:
+  import unittest
+from google.protobuf import unittest_import_pb2
+from google.protobuf import unittest_import_public_pb2
+from google.protobuf import unittest_pb2
+from google.protobuf import descriptor_pb2
+from google.protobuf.internal import api_implementation
+from google.protobuf.internal import descriptor_pool_test1_pb2
+from google.protobuf.internal import descriptor_pool_test2_pb2
+from google.protobuf.internal import factory_test1_pb2
+from google.protobuf.internal import factory_test2_pb2
+from google.protobuf.internal import test_util
+from google.protobuf import descriptor
+from google.protobuf import descriptor_database
+from google.protobuf import descriptor_pool
+from google.protobuf import message_factory
+from google.protobuf import symbol_database
+
+
+class DescriptorPoolTest(unittest.TestCase):
+
+  def CreatePool(self):
+    return descriptor_pool.DescriptorPool()
+
+  def setUp(self):
+    self.pool = self.CreatePool()
+    self.factory_test1_fd = descriptor_pb2.FileDescriptorProto.FromString(
+        factory_test1_pb2.DESCRIPTOR.serialized_pb)
+    self.factory_test2_fd = descriptor_pb2.FileDescriptorProto.FromString(
+        factory_test2_pb2.DESCRIPTOR.serialized_pb)
+    self.pool.Add(self.factory_test1_fd)
+    self.pool.Add(self.factory_test2_fd)
+
+  def testFindFileByName(self):
+    name1 = 'google/protobuf/internal/factory_test1.proto'
+    file_desc1 = self.pool.FindFileByName(name1)
+    self.assertIsInstance(file_desc1, descriptor.FileDescriptor)
+    self.assertEqual(name1, file_desc1.name)
+    self.assertEqual('google.protobuf.python.internal', file_desc1.package)
+    self.assertIn('Factory1Message', file_desc1.message_types_by_name)
+
+    name2 = 'google/protobuf/internal/factory_test2.proto'
+    file_desc2 = self.pool.FindFileByName(name2)
+    self.assertIsInstance(file_desc2, descriptor.FileDescriptor)
+    self.assertEqual(name2, file_desc2.name)
+    self.assertEqual('google.protobuf.python.internal', file_desc2.package)
+    self.assertIn('Factory2Message', file_desc2.message_types_by_name)
+
+  def testFindFileByNameFailure(self):
+    with self.assertRaises(KeyError):
+      self.pool.FindFileByName('Does not exist')
+
+  def testFindFileContainingSymbol(self):
+    file_desc1 = self.pool.FindFileContainingSymbol(
+        'google.protobuf.python.internal.Factory1Message')
+    self.assertIsInstance(file_desc1, descriptor.FileDescriptor)
+    self.assertEqual('google/protobuf/internal/factory_test1.proto',
+                     file_desc1.name)
+    self.assertEqual('google.protobuf.python.internal', file_desc1.package)
+    self.assertIn('Factory1Message', file_desc1.message_types_by_name)
+
+    file_desc2 = self.pool.FindFileContainingSymbol(
+        'google.protobuf.python.internal.Factory2Message')
+    self.assertIsInstance(file_desc2, descriptor.FileDescriptor)
+    self.assertEqual('google/protobuf/internal/factory_test2.proto',
+                     file_desc2.name)
+    self.assertEqual('google.protobuf.python.internal', file_desc2.package)
+    self.assertIn('Factory2Message', file_desc2.message_types_by_name)
+
+  def testFindFileContainingSymbolFailure(self):
+    with self.assertRaises(KeyError):
+      self.pool.FindFileContainingSymbol('Does not exist')
+
+  def testFindMessageTypeByName(self):
+    msg1 = self.pool.FindMessageTypeByName(
+        'google.protobuf.python.internal.Factory1Message')
+    self.assertIsInstance(msg1, descriptor.Descriptor)
+    self.assertEqual('Factory1Message', msg1.name)
+    self.assertEqual('google.protobuf.python.internal.Factory1Message',
+                     msg1.full_name)
+    self.assertEqual(None, msg1.containing_type)
+
+    nested_msg1 = msg1.nested_types[0]
+    self.assertEqual('NestedFactory1Message', nested_msg1.name)
+    self.assertEqual(msg1, nested_msg1.containing_type)
+
+    nested_enum1 = msg1.enum_types[0]
+    self.assertEqual('NestedFactory1Enum', nested_enum1.name)
+    self.assertEqual(msg1, nested_enum1.containing_type)
+
+    self.assertEqual(nested_msg1, msg1.fields_by_name[
+        'nested_factory_1_message'].message_type)
+    self.assertEqual(nested_enum1, msg1.fields_by_name[
+        'nested_factory_1_enum'].enum_type)
+
+    msg2 = self.pool.FindMessageTypeByName(
+        'google.protobuf.python.internal.Factory2Message')
+    self.assertIsInstance(msg2, descriptor.Descriptor)
+    self.assertEqual('Factory2Message', msg2.name)
+    self.assertEqual('google.protobuf.python.internal.Factory2Message',
+                     msg2.full_name)
+    self.assertIsNone(msg2.containing_type)
+
+    nested_msg2 = msg2.nested_types[0]
+    self.assertEqual('NestedFactory2Message', nested_msg2.name)
+    self.assertEqual(msg2, nested_msg2.containing_type)
+
+    nested_enum2 = msg2.enum_types[0]
+    self.assertEqual('NestedFactory2Enum', nested_enum2.name)
+    self.assertEqual(msg2, nested_enum2.containing_type)
+
+    self.assertEqual(nested_msg2, msg2.fields_by_name[
+        'nested_factory_2_message'].message_type)
+    self.assertEqual(nested_enum2, msg2.fields_by_name[
+        'nested_factory_2_enum'].enum_type)
+
+    self.assertTrue(msg2.fields_by_name['int_with_default'].has_default_value)
+    self.assertEqual(
+        1776, msg2.fields_by_name['int_with_default'].default_value)
+
+    self.assertTrue(
+        msg2.fields_by_name['double_with_default'].has_default_value)
+    self.assertEqual(
+        9.99, msg2.fields_by_name['double_with_default'].default_value)
+
+    self.assertTrue(
+        msg2.fields_by_name['string_with_default'].has_default_value)
+    self.assertEqual(
+        'hello world', msg2.fields_by_name['string_with_default'].default_value)
+
+    self.assertTrue(msg2.fields_by_name['bool_with_default'].has_default_value)
+    self.assertFalse(msg2.fields_by_name['bool_with_default'].default_value)
+
+    self.assertTrue(msg2.fields_by_name['enum_with_default'].has_default_value)
+    self.assertEqual(
+        1, msg2.fields_by_name['enum_with_default'].default_value)
+
+    msg3 = self.pool.FindMessageTypeByName(
+        'google.protobuf.python.internal.Factory2Message.NestedFactory2Message')
+    self.assertEqual(nested_msg2, msg3)
+
+    self.assertTrue(msg2.fields_by_name['bytes_with_default'].has_default_value)
+    self.assertEqual(
+        b'a\xfb\x00c',
+        msg2.fields_by_name['bytes_with_default'].default_value)
+
+    self.assertEqual(1, len(msg2.oneofs))
+    self.assertEqual(1, len(msg2.oneofs_by_name))
+    self.assertEqual(2, len(msg2.oneofs[0].fields))
+    for name in ['oneof_int', 'oneof_string']:
+      self.assertEqual(msg2.oneofs[0],
+                       msg2.fields_by_name[name].containing_oneof)
+      self.assertIn(msg2.fields_by_name[name], msg2.oneofs[0].fields)
+
+  def testFindMessageTypeByNameFailure(self):
+    with self.assertRaises(KeyError):
+      self.pool.FindMessageTypeByName('Does not exist')
+
+  def testFindEnumTypeByName(self):
+    enum1 = self.pool.FindEnumTypeByName(
+        'google.protobuf.python.internal.Factory1Enum')
+    self.assertIsInstance(enum1, descriptor.EnumDescriptor)
+    self.assertEqual(0, enum1.values_by_name['FACTORY_1_VALUE_0'].number)
+    self.assertEqual(1, enum1.values_by_name['FACTORY_1_VALUE_1'].number)
+
+    nested_enum1 = self.pool.FindEnumTypeByName(
+        'google.protobuf.python.internal.Factory1Message.NestedFactory1Enum')
+    self.assertIsInstance(nested_enum1, descriptor.EnumDescriptor)
+    self.assertEqual(
+        0, nested_enum1.values_by_name['NESTED_FACTORY_1_VALUE_0'].number)
+    self.assertEqual(
+        1, nested_enum1.values_by_name['NESTED_FACTORY_1_VALUE_1'].number)
+
+    enum2 = self.pool.FindEnumTypeByName(
+        'google.protobuf.python.internal.Factory2Enum')
+    self.assertIsInstance(enum2, descriptor.EnumDescriptor)
+    self.assertEqual(0, enum2.values_by_name['FACTORY_2_VALUE_0'].number)
+    self.assertEqual(1, enum2.values_by_name['FACTORY_2_VALUE_1'].number)
+
+    nested_enum2 = self.pool.FindEnumTypeByName(
+        'google.protobuf.python.internal.Factory2Message.NestedFactory2Enum')
+    self.assertIsInstance(nested_enum2, descriptor.EnumDescriptor)
+    self.assertEqual(
+        0, nested_enum2.values_by_name['NESTED_FACTORY_2_VALUE_0'].number)
+    self.assertEqual(
+        1, nested_enum2.values_by_name['NESTED_FACTORY_2_VALUE_1'].number)
+
+  def testFindEnumTypeByNameFailure(self):
+    with self.assertRaises(KeyError):
+      self.pool.FindEnumTypeByName('Does not exist')
+
+  def testFindFieldByName(self):
+    field = self.pool.FindFieldByName(
+        'google.protobuf.python.internal.Factory1Message.list_value')
+    self.assertEqual(field.name, 'list_value')
+    self.assertEqual(field.label, field.LABEL_REPEATED)
+    with self.assertRaises(KeyError):
+      self.pool.FindFieldByName('Does not exist')
+
+  def testFindExtensionByName(self):
+    # An extension defined in a message.
+    extension = self.pool.FindExtensionByName(
+        'google.protobuf.python.internal.Factory2Message.one_more_field')
+    self.assertEqual(extension.name, 'one_more_field')
+    # An extension defined at file scope.
+    extension = self.pool.FindExtensionByName(
+        'google.protobuf.python.internal.another_field')
+    self.assertEqual(extension.name, 'another_field')
+    self.assertEqual(extension.number, 1002)
+    with self.assertRaises(KeyError):
+      self.pool.FindFieldByName('Does not exist')
+
+  def testExtensionsAreNotFields(self):
+    with self.assertRaises(KeyError):
+      self.pool.FindFieldByName('google.protobuf.python.internal.another_field')
+    with self.assertRaises(KeyError):
+      self.pool.FindFieldByName(
+          'google.protobuf.python.internal.Factory2Message.one_more_field')
+    with self.assertRaises(KeyError):
+      self.pool.FindExtensionByName(
+          'google.protobuf.python.internal.Factory1Message.list_value')
+
+  def testUserDefinedDB(self):
+    db = descriptor_database.DescriptorDatabase()
+    self.pool = descriptor_pool.DescriptorPool(db)
+    db.Add(self.factory_test1_fd)
+    db.Add(self.factory_test2_fd)
+    self.testFindMessageTypeByName()
+
+  def testAddSerializedFile(self):
+    self.pool = descriptor_pool.DescriptorPool()
+    self.pool.AddSerializedFile(self.factory_test1_fd.SerializeToString())
+    self.pool.AddSerializedFile(self.factory_test2_fd.SerializeToString())
+    self.testFindMessageTypeByName()
+
+  def testComplexNesting(self):
+    test1_desc = descriptor_pb2.FileDescriptorProto.FromString(
+        descriptor_pool_test1_pb2.DESCRIPTOR.serialized_pb)
+    test2_desc = descriptor_pb2.FileDescriptorProto.FromString(
+        descriptor_pool_test2_pb2.DESCRIPTOR.serialized_pb)
+    self.pool.Add(test1_desc)
+    self.pool.Add(test2_desc)
+    TEST1_FILE.CheckFile(self, self.pool)
+    TEST2_FILE.CheckFile(self, self.pool)
+
+
+  def testEnumDefaultValue(self):
+    """Test the default value of enums which don't start at zero."""
+    def _CheckDefaultValue(file_descriptor):
+      default_value = (file_descriptor
+                       .message_types_by_name['DescriptorPoolTest1']
+                       .fields_by_name['nested_enum']
+                       .default_value)
+      self.assertEqual(default_value,
+                       descriptor_pool_test1_pb2.DescriptorPoolTest1.BETA)
+    # First check what the generated descriptor contains.
+    _CheckDefaultValue(descriptor_pool_test1_pb2.DESCRIPTOR)
+    # Then check the generated pool. Normally this is the same descriptor.
+    file_descriptor = symbol_database.Default().pool.FindFileByName(
+        'google/protobuf/internal/descriptor_pool_test1.proto')
+    self.assertIs(file_descriptor, descriptor_pool_test1_pb2.DESCRIPTOR)
+    _CheckDefaultValue(file_descriptor)
+
+    # Then check the dynamic pool and its internal DescriptorDatabase.
+    descriptor_proto = descriptor_pb2.FileDescriptorProto.FromString(
+        descriptor_pool_test1_pb2.DESCRIPTOR.serialized_pb)
+    self.pool.Add(descriptor_proto)
+    # And do the same check as above
+    file_descriptor = self.pool.FindFileByName(
+        'google/protobuf/internal/descriptor_pool_test1.proto')
+    _CheckDefaultValue(file_descriptor)
+
+  def testDefaultValueForCustomMessages(self):
+    """Check the value returned by non-existent fields."""
+    def _CheckValueAndType(value, expected_value, expected_type):
+      self.assertEqual(value, expected_value)
+      self.assertIsInstance(value, expected_type)
+
+    def _CheckDefaultValues(msg):
+      try:
+        int64 = long
+      except NameError:  # Python3
+        int64 = int
+      try:
+        unicode_type = unicode
+      except NameError:  # Python3
+        unicode_type = str
+      _CheckValueAndType(msg.optional_int32, 0, int)
+      _CheckValueAndType(msg.optional_uint64, 0, (int64, int))
+      _CheckValueAndType(msg.optional_float, 0, (float, int))
+      _CheckValueAndType(msg.optional_double, 0, (float, int))
+      _CheckValueAndType(msg.optional_bool, False, bool)
+      _CheckValueAndType(msg.optional_string, u'', unicode_type)
+      _CheckValueAndType(msg.optional_bytes, b'', bytes)
+      _CheckValueAndType(msg.optional_nested_enum, msg.FOO, int)
+    # First for the generated message
+    _CheckDefaultValues(unittest_pb2.TestAllTypes())
+    # Then for a message built with from the DescriptorPool.
+    pool = descriptor_pool.DescriptorPool()
+    pool.Add(descriptor_pb2.FileDescriptorProto.FromString(
+        unittest_import_public_pb2.DESCRIPTOR.serialized_pb))
+    pool.Add(descriptor_pb2.FileDescriptorProto.FromString(
+        unittest_import_pb2.DESCRIPTOR.serialized_pb))
+    pool.Add(descriptor_pb2.FileDescriptorProto.FromString(
+        unittest_pb2.DESCRIPTOR.serialized_pb))
+    message_class = message_factory.MessageFactory(pool).GetPrototype(
+        pool.FindMessageTypeByName(
+            unittest_pb2.TestAllTypes.DESCRIPTOR.full_name))
+    _CheckDefaultValues(message_class())
+
+
+@unittest.skipIf(api_implementation.Type() != 'cpp',
+                            'explicit tests of the C++ implementation')
+class CppDescriptorPoolTest(DescriptorPoolTest):
+  # TODO(amauryfa): remove when descriptor_pool.DescriptorPool() creates true
+  # C++ descriptor pool object for C++ implementation.
+
+  def CreatePool(self):
+    # pylint: disable=g-import-not-at-top
+    from google.protobuf.pyext import _message
+    return _message.DescriptorPool()
+
+
+class ProtoFile(object):
+
+  def __init__(self, name, package, messages, dependencies=None):
+    self.name = name
+    self.package = package
+    self.messages = messages
+    self.dependencies = dependencies or []
+
+  def CheckFile(self, test, pool):
+    file_desc = pool.FindFileByName(self.name)
+    test.assertEqual(self.name, file_desc.name)
+    test.assertEqual(self.package, file_desc.package)
+    dependencies_names = [f.name for f in file_desc.dependencies]
+    test.assertEqual(self.dependencies, dependencies_names)
+    for name, msg_type in self.messages.items():
+      msg_type.CheckType(test, None, name, file_desc)
+
+
+class EnumType(object):
+
+  def __init__(self, values):
+    self.values = values
+
+  def CheckType(self, test, msg_desc, name, file_desc):
+    enum_desc = msg_desc.enum_types_by_name[name]
+    test.assertEqual(name, enum_desc.name)
+    expected_enum_full_name = '.'.join([msg_desc.full_name, name])
+    test.assertEqual(expected_enum_full_name, enum_desc.full_name)
+    test.assertEqual(msg_desc, enum_desc.containing_type)
+    test.assertEqual(file_desc, enum_desc.file)
+    for index, (value, number) in enumerate(self.values):
+      value_desc = enum_desc.values_by_name[value]
+      test.assertEqual(value, value_desc.name)
+      test.assertEqual(index, value_desc.index)
+      test.assertEqual(number, value_desc.number)
+      test.assertEqual(enum_desc, value_desc.type)
+      test.assertIn(value, msg_desc.enum_values_by_name)
+
+
+class MessageType(object):
+
+  def __init__(self, type_dict, field_list, is_extendable=False,
+               extensions=None):
+    self.type_dict = type_dict
+    self.field_list = field_list
+    self.is_extendable = is_extendable
+    self.extensions = extensions or []
+
+  def CheckType(self, test, containing_type_desc, name, file_desc):
+    if containing_type_desc is None:
+      desc = file_desc.message_types_by_name[name]
+      expected_full_name = '.'.join([file_desc.package, name])
+    else:
+      desc = containing_type_desc.nested_types_by_name[name]
+      expected_full_name = '.'.join([containing_type_desc.full_name, name])
+
+    test.assertEqual(name, desc.name)
+    test.assertEqual(expected_full_name, desc.full_name)
+    test.assertEqual(containing_type_desc, desc.containing_type)
+    test.assertEqual(desc.file, file_desc)
+    test.assertEqual(self.is_extendable, desc.is_extendable)
+    for name, subtype in self.type_dict.items():
+      subtype.CheckType(test, desc, name, file_desc)
+
+    for index, (name, field) in enumerate(self.field_list):
+      field.CheckField(test, desc, name, index)
+
+    for index, (name, field) in enumerate(self.extensions):
+      field.CheckField(test, desc, name, index)
+
+
+class EnumField(object):
+
+  def __init__(self, number, type_name, default_value):
+    self.number = number
+    self.type_name = type_name
+    self.default_value = default_value
+
+  def CheckField(self, test, msg_desc, name, index):
+    field_desc = msg_desc.fields_by_name[name]
+    enum_desc = msg_desc.enum_types_by_name[self.type_name]
+    test.assertEqual(name, field_desc.name)
+    expected_field_full_name = '.'.join([msg_desc.full_name, name])
+    test.assertEqual(expected_field_full_name, field_desc.full_name)
+    test.assertEqual(index, field_desc.index)
+    test.assertEqual(self.number, field_desc.number)
+    test.assertEqual(descriptor.FieldDescriptor.TYPE_ENUM, field_desc.type)
+    test.assertEqual(descriptor.FieldDescriptor.CPPTYPE_ENUM,
+                     field_desc.cpp_type)
+    test.assertTrue(field_desc.has_default_value)
+    test.assertEqual(enum_desc.values_by_name[self.default_value].number,
+                     field_desc.default_value)
+    test.assertEqual(msg_desc, field_desc.containing_type)
+    test.assertEqual(enum_desc, field_desc.enum_type)
+
+
+class MessageField(object):
+
+  def __init__(self, number, type_name):
+    self.number = number
+    self.type_name = type_name
+
+  def CheckField(self, test, msg_desc, name, index):
+    field_desc = msg_desc.fields_by_name[name]
+    field_type_desc = msg_desc.nested_types_by_name[self.type_name]
+    test.assertEqual(name, field_desc.name)
+    expected_field_full_name = '.'.join([msg_desc.full_name, name])
+    test.assertEqual(expected_field_full_name, field_desc.full_name)
+    test.assertEqual(index, field_desc.index)
+    test.assertEqual(self.number, field_desc.number)
+    test.assertEqual(descriptor.FieldDescriptor.TYPE_MESSAGE, field_desc.type)
+    test.assertEqual(descriptor.FieldDescriptor.CPPTYPE_MESSAGE,
+                     field_desc.cpp_type)
+    test.assertFalse(field_desc.has_default_value)
+    test.assertEqual(msg_desc, field_desc.containing_type)
+    test.assertEqual(field_type_desc, field_desc.message_type)
+
+
+class StringField(object):
+
+  def __init__(self, number, default_value):
+    self.number = number
+    self.default_value = default_value
+
+  def CheckField(self, test, msg_desc, name, index):
+    field_desc = msg_desc.fields_by_name[name]
+    test.assertEqual(name, field_desc.name)
+    expected_field_full_name = '.'.join([msg_desc.full_name, name])
+    test.assertEqual(expected_field_full_name, field_desc.full_name)
+    test.assertEqual(index, field_desc.index)
+    test.assertEqual(self.number, field_desc.number)
+    test.assertEqual(descriptor.FieldDescriptor.TYPE_STRING, field_desc.type)
+    test.assertEqual(descriptor.FieldDescriptor.CPPTYPE_STRING,
+                     field_desc.cpp_type)
+    test.assertTrue(field_desc.has_default_value)
+    test.assertEqual(self.default_value, field_desc.default_value)
+
+
+class ExtensionField(object):
+
+  def __init__(self, number, extended_type):
+    self.number = number
+    self.extended_type = extended_type
+
+  def CheckField(self, test, msg_desc, name, index):
+    field_desc = msg_desc.extensions_by_name[name]
+    test.assertEqual(name, field_desc.name)
+    expected_field_full_name = '.'.join([msg_desc.full_name, name])
+    test.assertEqual(expected_field_full_name, field_desc.full_name)
+    test.assertEqual(self.number, field_desc.number)
+    test.assertEqual(index, field_desc.index)
+    test.assertEqual(descriptor.FieldDescriptor.TYPE_MESSAGE, field_desc.type)
+    test.assertEqual(descriptor.FieldDescriptor.CPPTYPE_MESSAGE,
+                     field_desc.cpp_type)
+    test.assertFalse(field_desc.has_default_value)
+    test.assertTrue(field_desc.is_extension)
+    test.assertEqual(msg_desc, field_desc.extension_scope)
+    test.assertEqual(msg_desc, field_desc.message_type)
+    test.assertEqual(self.extended_type, field_desc.containing_type.name)
+
+
+class AddDescriptorTest(unittest.TestCase):
+
+  def _TestMessage(self, prefix):
+    pool = descriptor_pool.DescriptorPool()
+    pool.AddDescriptor(unittest_pb2.TestAllTypes.DESCRIPTOR)
+    self.assertEqual(
+        'protobuf_unittest.TestAllTypes',
+        pool.FindMessageTypeByName(
+            prefix + 'protobuf_unittest.TestAllTypes').full_name)
+
+    # AddDescriptor is not recursive.
+    with self.assertRaises(KeyError):
+      pool.FindMessageTypeByName(
+          prefix + 'protobuf_unittest.TestAllTypes.NestedMessage')
+
+    pool.AddDescriptor(unittest_pb2.TestAllTypes.NestedMessage.DESCRIPTOR)
+    self.assertEqual(
+        'protobuf_unittest.TestAllTypes.NestedMessage',
+        pool.FindMessageTypeByName(
+            prefix + 'protobuf_unittest.TestAllTypes.NestedMessage').full_name)
+
+    # Files are implicitly also indexed when messages are added.
+    self.assertEqual(
+        'google/protobuf/unittest.proto',
+        pool.FindFileByName(
+            'google/protobuf/unittest.proto').name)
+
+    self.assertEqual(
+        'google/protobuf/unittest.proto',
+        pool.FindFileContainingSymbol(
+            prefix + 'protobuf_unittest.TestAllTypes.NestedMessage').name)
+
+  @unittest.skipIf(api_implementation.Type() == 'cpp',
+                    'With the cpp implementation, Add() must be called first')
+  def testMessage(self):
+    self._TestMessage('')
+    self._TestMessage('.')
+
+  def _TestEnum(self, prefix):
+    pool = descriptor_pool.DescriptorPool()
+    pool.AddEnumDescriptor(unittest_pb2.ForeignEnum.DESCRIPTOR)
+    self.assertEqual(
+        'protobuf_unittest.ForeignEnum',
+        pool.FindEnumTypeByName(
+            prefix + 'protobuf_unittest.ForeignEnum').full_name)
+
+    # AddEnumDescriptor is not recursive.
+    with self.assertRaises(KeyError):
+      pool.FindEnumTypeByName(
+          prefix + 'protobuf_unittest.ForeignEnum.NestedEnum')
+
+    pool.AddEnumDescriptor(unittest_pb2.TestAllTypes.NestedEnum.DESCRIPTOR)
+    self.assertEqual(
+        'protobuf_unittest.TestAllTypes.NestedEnum',
+        pool.FindEnumTypeByName(
+            prefix + 'protobuf_unittest.TestAllTypes.NestedEnum').full_name)
+
+    # Files are implicitly also indexed when enums are added.
+    self.assertEqual(
+        'google/protobuf/unittest.proto',
+        pool.FindFileByName(
+            'google/protobuf/unittest.proto').name)
+
+    self.assertEqual(
+        'google/protobuf/unittest.proto',
+        pool.FindFileContainingSymbol(
+            prefix + 'protobuf_unittest.TestAllTypes.NestedEnum').name)
+
+  @unittest.skipIf(api_implementation.Type() == 'cpp',
+                    'With the cpp implementation, Add() must be called first')
+  def testEnum(self):
+    self._TestEnum('')
+    self._TestEnum('.')
+
+  @unittest.skipIf(api_implementation.Type() == 'cpp',
+                    'With the cpp implementation, Add() must be called first')
+  def testFile(self):
+    pool = descriptor_pool.DescriptorPool()
+    pool.AddFileDescriptor(unittest_pb2.DESCRIPTOR)
+    self.assertEqual(
+        'google/protobuf/unittest.proto',
+        pool.FindFileByName(
+            'google/protobuf/unittest.proto').name)
+
+    # AddFileDescriptor is not recursive; messages and enums within files must
+    # be explicitly registered.
+    with self.assertRaises(KeyError):
+      pool.FindFileContainingSymbol(
+          'protobuf_unittest.TestAllTypes')
+
+  def _GetDescriptorPoolClass(self):
+    # Test with both implementations of descriptor pools.
+    if api_implementation.Type() == 'cpp':
+      # pylint: disable=g-import-not-at-top
+      from google.protobuf.pyext import _message
+      return _message.DescriptorPool
+    else:
+      return descriptor_pool.DescriptorPool
+
+  def testEmptyDescriptorPool(self):
+    # Check that an empty DescriptorPool() contains no message.
+    pool = self._GetDescriptorPoolClass()()
+    proto_file_name = descriptor_pb2.DESCRIPTOR.name
+    self.assertRaises(KeyError, pool.FindFileByName, proto_file_name)
+    # Add the above file to the pool
+    file_descriptor = descriptor_pb2.FileDescriptorProto()
+    descriptor_pb2.DESCRIPTOR.CopyToProto(file_descriptor)
+    pool.Add(file_descriptor)
+    # Now it exists.
+    self.assertTrue(pool.FindFileByName(proto_file_name))
+
+  def testCustomDescriptorPool(self):
+    # Create a new pool, and add a file descriptor.
+    pool = self._GetDescriptorPoolClass()()
+    file_desc = descriptor_pb2.FileDescriptorProto(
+        name='some/file.proto', package='package')
+    file_desc.message_type.add(name='Message')
+    pool.Add(file_desc)
+    self.assertEqual(pool.FindFileByName('some/file.proto').name,
+                     'some/file.proto')
+    self.assertEqual(pool.FindMessageTypeByName('package.Message').name,
+                     'Message')
+
+
+@unittest.skipIf(
+    api_implementation.Type() != 'cpp',
+    'default_pool is only supported by the C++ implementation')
+class DefaultPoolTest(unittest.TestCase):
+
+  def testFindMethods(self):
+    # pylint: disable=g-import-not-at-top
+    from google.protobuf.pyext import _message
+    pool = _message.default_pool
+    self.assertIs(
+        pool.FindFileByName('google/protobuf/unittest.proto'),
+        unittest_pb2.DESCRIPTOR)
+    self.assertIs(
+        pool.FindMessageTypeByName('protobuf_unittest.TestAllTypes'),
+        unittest_pb2.TestAllTypes.DESCRIPTOR)
+    self.assertIs(
+        pool.FindFieldByName('protobuf_unittest.TestAllTypes.optional_int32'),
+        unittest_pb2.TestAllTypes.DESCRIPTOR.fields_by_name['optional_int32'])
+    self.assertIs(
+        pool.FindExtensionByName('protobuf_unittest.optional_int32_extension'),
+        unittest_pb2.DESCRIPTOR.extensions_by_name['optional_int32_extension'])
+    self.assertIs(
+        pool.FindEnumTypeByName('protobuf_unittest.ForeignEnum'),
+        unittest_pb2.ForeignEnum.DESCRIPTOR)
+    self.assertIs(
+        pool.FindOneofByName('protobuf_unittest.TestAllTypes.oneof_field'),
+        unittest_pb2.TestAllTypes.DESCRIPTOR.oneofs_by_name['oneof_field'])
+
+  def testAddFileDescriptor(self):
+    # pylint: disable=g-import-not-at-top
+    from google.protobuf.pyext import _message
+    pool = _message.default_pool
+    file_desc = descriptor_pb2.FileDescriptorProto(name='some/file.proto')
+    pool.Add(file_desc)
+    pool.AddSerializedFile(file_desc.SerializeToString())
+
+
+TEST1_FILE = ProtoFile(
+    'google/protobuf/internal/descriptor_pool_test1.proto',
+    'google.protobuf.python.internal',
+    {
+        'DescriptorPoolTest1': MessageType({
+            'NestedEnum': EnumType([('ALPHA', 1), ('BETA', 2)]),
+            'NestedMessage': MessageType({
+                'NestedEnum': EnumType([('EPSILON', 5), ('ZETA', 6)]),
+                'DeepNestedMessage': MessageType({
+                    'NestedEnum': EnumType([('ETA', 7), ('THETA', 8)]),
+                }, [
+                    ('nested_enum', EnumField(1, 'NestedEnum', 'ETA')),
+                    ('nested_field', StringField(2, 'theta')),
+                ]),
+            }, [
+                ('nested_enum', EnumField(1, 'NestedEnum', 'ZETA')),
+                ('nested_field', StringField(2, 'beta')),
+                ('deep_nested_message', MessageField(3, 'DeepNestedMessage')),
+            ])
+        }, [
+            ('nested_enum', EnumField(1, 'NestedEnum', 'BETA')),
+            ('nested_message', MessageField(2, 'NestedMessage')),
+        ], is_extendable=True),
+
+        'DescriptorPoolTest2': MessageType({
+            'NestedEnum': EnumType([('GAMMA', 3), ('DELTA', 4)]),
+            'NestedMessage': MessageType({
+                'NestedEnum': EnumType([('IOTA', 9), ('KAPPA', 10)]),
+                'DeepNestedMessage': MessageType({
+                    'NestedEnum': EnumType([('LAMBDA', 11), ('MU', 12)]),
+                }, [
+                    ('nested_enum', EnumField(1, 'NestedEnum', 'MU')),
+                    ('nested_field', StringField(2, 'lambda')),
+                ]),
+            }, [
+                ('nested_enum', EnumField(1, 'NestedEnum', 'IOTA')),
+                ('nested_field', StringField(2, 'delta')),
+                ('deep_nested_message', MessageField(3, 'DeepNestedMessage')),
+            ])
+        }, [
+            ('nested_enum', EnumField(1, 'NestedEnum', 'GAMMA')),
+            ('nested_message', MessageField(2, 'NestedMessage')),
+        ]),
+    })
+
+
+TEST2_FILE = ProtoFile(
+    'google/protobuf/internal/descriptor_pool_test2.proto',
+    'google.protobuf.python.internal',
+    {
+        'DescriptorPoolTest3': MessageType({
+            'NestedEnum': EnumType([('NU', 13), ('XI', 14)]),
+            'NestedMessage': MessageType({
+                'NestedEnum': EnumType([('OMICRON', 15), ('PI', 16)]),
+                'DeepNestedMessage': MessageType({
+                    'NestedEnum': EnumType([('RHO', 17), ('SIGMA', 18)]),
+                }, [
+                    ('nested_enum', EnumField(1, 'NestedEnum', 'RHO')),
+                    ('nested_field', StringField(2, 'sigma')),
+                ]),
+            }, [
+                ('nested_enum', EnumField(1, 'NestedEnum', 'PI')),
+                ('nested_field', StringField(2, 'nu')),
+                ('deep_nested_message', MessageField(3, 'DeepNestedMessage')),
+            ])
+        }, [
+            ('nested_enum', EnumField(1, 'NestedEnum', 'XI')),
+            ('nested_message', MessageField(2, 'NestedMessage')),
+        ], extensions=[
+            ('descriptor_pool_test',
+             ExtensionField(1001, 'DescriptorPoolTest1')),
+        ]),
+    },
+    dependencies=['google/protobuf/internal/descriptor_pool_test1.proto'])
+
+
+if __name__ == '__main__':
+  unittest.main()
diff --git a/python/google/protobuf/internal/descriptor_pool_test1.proto b/python/google/protobuf/internal/descriptor_pool_test1.proto
new file mode 100644
index 0000000..00816b7
--- /dev/null
+++ b/python/google/protobuf/internal/descriptor_pool_test1.proto
@@ -0,0 +1,96 @@
+// Protocol Buffers - Google's data interchange format
+// Copyright 2008 Google Inc.  All rights reserved.
+// https://developers.google.com/protocol-buffers/
+//
+// Redistribution and use in source and binary forms, with or without
+// modification, are permitted provided that the following conditions are
+// met:
+//
+//     * Redistributions of source code must retain the above copyright
+// notice, this list of conditions and the following disclaimer.
+//     * Redistributions in binary form must reproduce the above
+// copyright notice, this list of conditions and the following disclaimer
+// in the documentation and/or other materials provided with the
+// distribution.
+//     * Neither the name of Google Inc. nor the names of its
+// contributors may be used to endorse or promote products derived from
+// this software without specific prior written permission.
+//
+// THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+// "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
+// LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
+// A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
+// OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
+// SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
+// LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
+// DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
+// THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
+// (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+// OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+
+syntax = "proto2";
+
+package google.protobuf.python.internal;
+
+
+message DescriptorPoolTest1 {
+  extensions 1000 to max;
+
+  enum NestedEnum {
+    ALPHA = 1;
+    BETA = 2;
+  }
+
+  optional NestedEnum nested_enum = 1 [default = BETA];
+
+  message NestedMessage {
+    enum NestedEnum {
+      EPSILON = 5;
+      ZETA = 6;
+    }
+    optional NestedEnum nested_enum = 1 [default = ZETA];
+    optional string nested_field = 2 [default = "beta"];
+    optional DeepNestedMessage deep_nested_message = 3;
+
+    message DeepNestedMessage {
+      enum NestedEnum {
+        ETA = 7;
+        THETA = 8;
+      }
+      optional NestedEnum nested_enum = 1 [default = ETA];
+      optional string nested_field = 2 [default = "theta"];
+    }
+  }
+
+  optional NestedMessage nested_message = 2;
+}
+
+message DescriptorPoolTest2 {
+  enum NestedEnum {
+    GAMMA = 3;
+    DELTA = 4;
+  }
+
+  optional NestedEnum nested_enum = 1 [default = GAMMA];
+
+  message NestedMessage {
+    enum NestedEnum {
+      IOTA = 9;
+      KAPPA = 10;
+    }
+    optional NestedEnum nested_enum = 1 [default = IOTA];
+    optional string nested_field = 2 [default = "delta"];
+    optional DeepNestedMessage deep_nested_message = 3;
+
+    message DeepNestedMessage {
+      enum NestedEnum {
+        LAMBDA = 11;
+        MU = 12;
+      }
+      optional NestedEnum nested_enum = 1 [default = MU];
+      optional string nested_field = 2 [default = "lambda"];
+    }
+  }
+
+  optional NestedMessage nested_message = 2;
+}
diff --git a/python/google/protobuf/internal/descriptor_pool_test2.proto b/python/google/protobuf/internal/descriptor_pool_test2.proto
new file mode 100644
index 0000000..e3fa660
--- /dev/null
+++ b/python/google/protobuf/internal/descriptor_pool_test2.proto
@@ -0,0 +1,72 @@
+// Protocol Buffers - Google's data interchange format
+// Copyright 2008 Google Inc.  All rights reserved.
+// https://developers.google.com/protocol-buffers/
+//
+// Redistribution and use in source and binary forms, with or without
+// modification, are permitted provided that the following conditions are
+// met:
+//
+//     * Redistributions of source code must retain the above copyright
+// notice, this list of conditions and the following disclaimer.
+//     * Redistributions in binary form must reproduce the above
+// copyright notice, this list of conditions and the following disclaimer
+// in the documentation and/or other materials provided with the
+// distribution.
+//     * Neither the name of Google Inc. nor the names of its
+// contributors may be used to endorse or promote products derived from
+// this software without specific prior written permission.
+//
+// THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+// "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
+// LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
+// A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
+// OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
+// SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
+// LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
+// DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
+// THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
+// (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+// OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+
+syntax = "proto2";
+
+package google.protobuf.python.internal;
+
+import "google/protobuf/internal/descriptor_pool_test1.proto";
+
+
+message DescriptorPoolTest3 {
+
+  extend DescriptorPoolTest1 {
+    optional DescriptorPoolTest3 descriptor_pool_test = 1001;
+  }
+
+  enum NestedEnum {
+    NU = 13;
+    XI = 14;
+  }
+
+  optional NestedEnum nested_enum = 1 [default = XI];
+
+  message NestedMessage {
+    enum NestedEnum {
+      OMICRON = 15;
+      PI = 16;
+    }
+    optional NestedEnum nested_enum = 1 [default = PI];
+    optional string nested_field = 2 [default = "nu"];
+    optional DeepNestedMessage deep_nested_message = 3;
+
+    message DeepNestedMessage {
+      enum NestedEnum {
+        RHO = 17;
+        SIGMA = 18;
+      }
+      optional NestedEnum nested_enum = 1 [default = RHO];
+      optional string nested_field = 2 [default = "sigma"];
+    }
+  }
+
+  optional NestedMessage nested_message = 2;
+}
+
diff --git a/python/google/protobuf/internal/descriptor_test.py b/python/google/protobuf/internal/descriptor_test.py
new file mode 100755
index 0000000..fee09a5
--- /dev/null
+++ b/python/google/protobuf/internal/descriptor_test.py
@@ -0,0 +1,821 @@
+#! /usr/bin/env python
+#
+# Protocol Buffers - Google's data interchange format
+# Copyright 2008 Google Inc.  All rights reserved.
+# https://developers.google.com/protocol-buffers/
+#
+# Redistribution and use in source and binary forms, with or without
+# modification, are permitted provided that the following conditions are
+# met:
+#
+#     * Redistributions of source code must retain the above copyright
+# notice, this list of conditions and the following disclaimer.
+#     * Redistributions in binary form must reproduce the above
+# copyright notice, this list of conditions and the following disclaimer
+# in the documentation and/or other materials provided with the
+# distribution.
+#     * Neither the name of Google Inc. nor the names of its
+# contributors may be used to endorse or promote products derived from
+# this software without specific prior written permission.
+#
+# THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+# "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
+# LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
+# A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
+# OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
+# SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
+# LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
+# DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
+# THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
+# (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+# OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+
+"""Unittest for google.protobuf.internal.descriptor."""
+
+__author__ = 'robinson@google.com (Will Robinson)'
+
+import sys
+
+try:
+  import unittest2 as unittest
+except ImportError:
+  import unittest
+from google.protobuf import unittest_custom_options_pb2
+from google.protobuf import unittest_import_pb2
+from google.protobuf import unittest_pb2
+from google.protobuf import descriptor_pb2
+from google.protobuf.internal import api_implementation
+from google.protobuf.internal import test_util
+from google.protobuf import descriptor
+from google.protobuf import descriptor_pool
+from google.protobuf import symbol_database
+from google.protobuf import text_format
+
+
+TEST_EMPTY_MESSAGE_DESCRIPTOR_ASCII = """
+name: 'TestEmptyMessage'
+"""
+
+
+class DescriptorTest(unittest.TestCase):
+
+  def setUp(self):
+    file_proto = descriptor_pb2.FileDescriptorProto(
+        name='some/filename/some.proto',
+        package='protobuf_unittest')
+    message_proto = file_proto.message_type.add(
+        name='NestedMessage')
+    message_proto.field.add(
+        name='bb',
+        number=1,
+        type=descriptor_pb2.FieldDescriptorProto.TYPE_INT32,
+        label=descriptor_pb2.FieldDescriptorProto.LABEL_OPTIONAL)
+    enum_proto = message_proto.enum_type.add(
+        name='ForeignEnum')
+    enum_proto.value.add(name='FOREIGN_FOO', number=4)
+    enum_proto.value.add(name='FOREIGN_BAR', number=5)
+    enum_proto.value.add(name='FOREIGN_BAZ', number=6)
+
+    self.pool = self.GetDescriptorPool()
+    self.pool.Add(file_proto)
+    self.my_file = self.pool.FindFileByName(file_proto.name)
+    self.my_message = self.my_file.message_types_by_name[message_proto.name]
+    self.my_enum = self.my_message.enum_types_by_name[enum_proto.name]
+
+    self.my_method = descriptor.MethodDescriptor(
+        name='Bar',
+        full_name='protobuf_unittest.TestService.Bar',
+        index=0,
+        containing_service=None,
+        input_type=None,
+        output_type=None)
+    self.my_service = descriptor.ServiceDescriptor(
+        name='TestServiceWithOptions',
+        full_name='protobuf_unittest.TestServiceWithOptions',
+        file=self.my_file,
+        index=0,
+        methods=[
+            self.my_method
+        ])
+
+  def GetDescriptorPool(self):
+    return symbol_database.Default().pool
+
+  def testEnumValueName(self):
+    self.assertEqual(self.my_message.EnumValueName('ForeignEnum', 4),
+                     'FOREIGN_FOO')
+
+    self.assertEqual(
+        self.my_message.enum_types_by_name[
+            'ForeignEnum'].values_by_number[4].name,
+        self.my_message.EnumValueName('ForeignEnum', 4))
+
+  def testEnumFixups(self):
+    self.assertEqual(self.my_enum, self.my_enum.values[0].type)
+
+  def testContainingTypeFixups(self):
+    self.assertEqual(self.my_message, self.my_message.fields[0].containing_type)
+    self.assertEqual(self.my_message, self.my_enum.containing_type)
+
+  def testContainingServiceFixups(self):
+    self.assertEqual(self.my_service, self.my_method.containing_service)
+
+  def testGetOptions(self):
+    self.assertEqual(self.my_enum.GetOptions(),
+                     descriptor_pb2.EnumOptions())
+    self.assertEqual(self.my_enum.values[0].GetOptions(),
+                     descriptor_pb2.EnumValueOptions())
+    self.assertEqual(self.my_message.GetOptions(),
+                     descriptor_pb2.MessageOptions())
+    self.assertEqual(self.my_message.fields[0].GetOptions(),
+                     descriptor_pb2.FieldOptions())
+    self.assertEqual(self.my_method.GetOptions(),
+                     descriptor_pb2.MethodOptions())
+    self.assertEqual(self.my_service.GetOptions(),
+                     descriptor_pb2.ServiceOptions())
+
+  def testSimpleCustomOptions(self):
+    file_descriptor = unittest_custom_options_pb2.DESCRIPTOR
+    message_descriptor =\
+        unittest_custom_options_pb2.TestMessageWithCustomOptions.DESCRIPTOR
+    field_descriptor = message_descriptor.fields_by_name["field1"]
+    enum_descriptor = message_descriptor.enum_types_by_name["AnEnum"]
+    enum_value_descriptor =\
+        message_descriptor.enum_values_by_name["ANENUM_VAL2"]
+    service_descriptor =\
+        unittest_custom_options_pb2.TestServiceWithCustomOptions.DESCRIPTOR
+    method_descriptor = service_descriptor.FindMethodByName("Foo")
+
+    file_options = file_descriptor.GetOptions()
+    file_opt1 = unittest_custom_options_pb2.file_opt1
+    self.assertEqual(9876543210, file_options.Extensions[file_opt1])
+    message_options = message_descriptor.GetOptions()
+    message_opt1 = unittest_custom_options_pb2.message_opt1
+    self.assertEqual(-56, message_options.Extensions[message_opt1])
+    field_options = field_descriptor.GetOptions()
+    field_opt1 = unittest_custom_options_pb2.field_opt1
+    self.assertEqual(8765432109, field_options.Extensions[field_opt1])
+    field_opt2 = unittest_custom_options_pb2.field_opt2
+    self.assertEqual(42, field_options.Extensions[field_opt2])
+    enum_options = enum_descriptor.GetOptions()
+    enum_opt1 = unittest_custom_options_pb2.enum_opt1
+    self.assertEqual(-789, enum_options.Extensions[enum_opt1])
+    enum_value_options = enum_value_descriptor.GetOptions()
+    enum_value_opt1 = unittest_custom_options_pb2.enum_value_opt1
+    self.assertEqual(123, enum_value_options.Extensions[enum_value_opt1])
+
+    service_options = service_descriptor.GetOptions()
+    service_opt1 = unittest_custom_options_pb2.service_opt1
+    self.assertEqual(-9876543210, service_options.Extensions[service_opt1])
+    method_options = method_descriptor.GetOptions()
+    method_opt1 = unittest_custom_options_pb2.method_opt1
+    self.assertEqual(unittest_custom_options_pb2.METHODOPT1_VAL2,
+                     method_options.Extensions[method_opt1])
+
+    message_descriptor = (
+        unittest_custom_options_pb2.DummyMessageContainingEnum.DESCRIPTOR)
+    self.assertTrue(file_descriptor.has_options)
+    self.assertFalse(message_descriptor.has_options)
+
+  def testDifferentCustomOptionTypes(self):
+    kint32min = -2**31
+    kint64min = -2**63
+    kint32max = 2**31 - 1
+    kint64max = 2**63 - 1
+    kuint32max = 2**32 - 1
+    kuint64max = 2**64 - 1
+
+    message_descriptor =\
+        unittest_custom_options_pb2.CustomOptionMinIntegerValues.DESCRIPTOR
+    message_options = message_descriptor.GetOptions()
+    self.assertEqual(False, message_options.Extensions[
+        unittest_custom_options_pb2.bool_opt])
+    self.assertEqual(kint32min, message_options.Extensions[
+        unittest_custom_options_pb2.int32_opt])
+    self.assertEqual(kint64min, message_options.Extensions[
+        unittest_custom_options_pb2.int64_opt])
+    self.assertEqual(0, message_options.Extensions[
+        unittest_custom_options_pb2.uint32_opt])
+    self.assertEqual(0, message_options.Extensions[
+        unittest_custom_options_pb2.uint64_opt])
+    self.assertEqual(kint32min, message_options.Extensions[
+        unittest_custom_options_pb2.sint32_opt])
+    self.assertEqual(kint64min, message_options.Extensions[
+        unittest_custom_options_pb2.sint64_opt])
+    self.assertEqual(0, message_options.Extensions[
+        unittest_custom_options_pb2.fixed32_opt])
+    self.assertEqual(0, message_options.Extensions[
+        unittest_custom_options_pb2.fixed64_opt])
+    self.assertEqual(kint32min, message_options.Extensions[
+        unittest_custom_options_pb2.sfixed32_opt])
+    self.assertEqual(kint64min, message_options.Extensions[
+        unittest_custom_options_pb2.sfixed64_opt])
+
+    message_descriptor =\
+        unittest_custom_options_pb2.CustomOptionMaxIntegerValues.DESCRIPTOR
+    message_options = message_descriptor.GetOptions()
+    self.assertEqual(True, message_options.Extensions[
+        unittest_custom_options_pb2.bool_opt])
+    self.assertEqual(kint32max, message_options.Extensions[
+        unittest_custom_options_pb2.int32_opt])
+    self.assertEqual(kint64max, message_options.Extensions[
+        unittest_custom_options_pb2.int64_opt])
+    self.assertEqual(kuint32max, message_options.Extensions[
+        unittest_custom_options_pb2.uint32_opt])
+    self.assertEqual(kuint64max, message_options.Extensions[
+        unittest_custom_options_pb2.uint64_opt])
+    self.assertEqual(kint32max, message_options.Extensions[
+        unittest_custom_options_pb2.sint32_opt])
+    self.assertEqual(kint64max, message_options.Extensions[
+        unittest_custom_options_pb2.sint64_opt])
+    self.assertEqual(kuint32max, message_options.Extensions[
+        unittest_custom_options_pb2.fixed32_opt])
+    self.assertEqual(kuint64max, message_options.Extensions[
+        unittest_custom_options_pb2.fixed64_opt])
+    self.assertEqual(kint32max, message_options.Extensions[
+        unittest_custom_options_pb2.sfixed32_opt])
+    self.assertEqual(kint64max, message_options.Extensions[
+        unittest_custom_options_pb2.sfixed64_opt])
+
+    message_descriptor =\
+        unittest_custom_options_pb2.CustomOptionOtherValues.DESCRIPTOR
+    message_options = message_descriptor.GetOptions()
+    self.assertEqual(-100, message_options.Extensions[
+        unittest_custom_options_pb2.int32_opt])
+    self.assertAlmostEqual(12.3456789, message_options.Extensions[
+        unittest_custom_options_pb2.float_opt], 6)
+    self.assertAlmostEqual(1.234567890123456789, message_options.Extensions[
+        unittest_custom_options_pb2.double_opt])
+    self.assertEqual("Hello, \"World\"", message_options.Extensions[
+        unittest_custom_options_pb2.string_opt])
+    self.assertEqual(b"Hello\0World", message_options.Extensions[
+        unittest_custom_options_pb2.bytes_opt])
+    dummy_enum = unittest_custom_options_pb2.DummyMessageContainingEnum
+    self.assertEqual(
+        dummy_enum.TEST_OPTION_ENUM_TYPE2,
+        message_options.Extensions[unittest_custom_options_pb2.enum_opt])
+
+    message_descriptor =\
+        unittest_custom_options_pb2.SettingRealsFromPositiveInts.DESCRIPTOR
+    message_options = message_descriptor.GetOptions()
+    self.assertAlmostEqual(12, message_options.Extensions[
+        unittest_custom_options_pb2.float_opt], 6)
+    self.assertAlmostEqual(154, message_options.Extensions[
+        unittest_custom_options_pb2.double_opt])
+
+    message_descriptor =\
+        unittest_custom_options_pb2.SettingRealsFromNegativeInts.DESCRIPTOR
+    message_options = message_descriptor.GetOptions()
+    self.assertAlmostEqual(-12, message_options.Extensions[
+        unittest_custom_options_pb2.float_opt], 6)
+    self.assertAlmostEqual(-154, message_options.Extensions[
+        unittest_custom_options_pb2.double_opt])
+
+  def testComplexExtensionOptions(self):
+    descriptor =\
+        unittest_custom_options_pb2.VariousComplexOptions.DESCRIPTOR
+    options = descriptor.GetOptions()
+    self.assertEqual(42, options.Extensions[
+        unittest_custom_options_pb2.complex_opt1].foo)
+    self.assertEqual(324, options.Extensions[
+        unittest_custom_options_pb2.complex_opt1].Extensions[
+            unittest_custom_options_pb2.quux])
+    self.assertEqual(876, options.Extensions[
+        unittest_custom_options_pb2.complex_opt1].Extensions[
+            unittest_custom_options_pb2.corge].qux)
+    self.assertEqual(987, options.Extensions[
+        unittest_custom_options_pb2.complex_opt2].baz)
+    self.assertEqual(654, options.Extensions[
+        unittest_custom_options_pb2.complex_opt2].Extensions[
+            unittest_custom_options_pb2.grault])
+    self.assertEqual(743, options.Extensions[
+        unittest_custom_options_pb2.complex_opt2].bar.foo)
+    self.assertEqual(1999, options.Extensions[
+        unittest_custom_options_pb2.complex_opt2].bar.Extensions[
+            unittest_custom_options_pb2.quux])
+    self.assertEqual(2008, options.Extensions[
+        unittest_custom_options_pb2.complex_opt2].bar.Extensions[
+            unittest_custom_options_pb2.corge].qux)
+    self.assertEqual(741, options.Extensions[
+        unittest_custom_options_pb2.complex_opt2].Extensions[
+            unittest_custom_options_pb2.garply].foo)
+    self.assertEqual(1998, options.Extensions[
+        unittest_custom_options_pb2.complex_opt2].Extensions[
+            unittest_custom_options_pb2.garply].Extensions[
+                unittest_custom_options_pb2.quux])
+    self.assertEqual(2121, options.Extensions[
+        unittest_custom_options_pb2.complex_opt2].Extensions[
+            unittest_custom_options_pb2.garply].Extensions[
+                unittest_custom_options_pb2.corge].qux)
+    self.assertEqual(1971, options.Extensions[
+        unittest_custom_options_pb2.ComplexOptionType2
+        .ComplexOptionType4.complex_opt4].waldo)
+    self.assertEqual(321, options.Extensions[
+        unittest_custom_options_pb2.complex_opt2].fred.waldo)
+    self.assertEqual(9, options.Extensions[
+        unittest_custom_options_pb2.complex_opt3].qux)
+    self.assertEqual(22, options.Extensions[
+        unittest_custom_options_pb2.complex_opt3].complexoptiontype5.plugh)
+    self.assertEqual(24, options.Extensions[
+        unittest_custom_options_pb2.complexopt6].xyzzy)
+
+  # Check that aggregate options were parsed and saved correctly in
+  # the appropriate descriptors.
+  def testAggregateOptions(self):
+    file_descriptor = unittest_custom_options_pb2.DESCRIPTOR
+    message_descriptor =\
+        unittest_custom_options_pb2.AggregateMessage.DESCRIPTOR
+    field_descriptor = message_descriptor.fields_by_name["fieldname"]
+    enum_descriptor = unittest_custom_options_pb2.AggregateEnum.DESCRIPTOR
+    enum_value_descriptor = enum_descriptor.values_by_name["VALUE"]
+    service_descriptor =\
+        unittest_custom_options_pb2.AggregateService.DESCRIPTOR
+    method_descriptor = service_descriptor.FindMethodByName("Method")
+
+    # Tests for the different types of data embedded in fileopt
+    file_options = file_descriptor.GetOptions().Extensions[
+        unittest_custom_options_pb2.fileopt]
+    self.assertEqual(100, file_options.i)
+    self.assertEqual("FileAnnotation", file_options.s)
+    self.assertEqual("NestedFileAnnotation", file_options.sub.s)
+    self.assertEqual("FileExtensionAnnotation", file_options.file.Extensions[
+        unittest_custom_options_pb2.fileopt].s)
+    self.assertEqual("EmbeddedMessageSetElement", file_options.mset.Extensions[
+        unittest_custom_options_pb2.AggregateMessageSetElement
+        .message_set_extension].s)
+
+    # Simple tests for all the other types of annotations
+    self.assertEqual(
+        "MessageAnnotation",
+        message_descriptor.GetOptions().Extensions[
+            unittest_custom_options_pb2.msgopt].s)
+    self.assertEqual(
+        "FieldAnnotation",
+        field_descriptor.GetOptions().Extensions[
+            unittest_custom_options_pb2.fieldopt].s)
+    self.assertEqual(
+        "EnumAnnotation",
+        enum_descriptor.GetOptions().Extensions[
+            unittest_custom_options_pb2.enumopt].s)
+    self.assertEqual(
+        "EnumValueAnnotation",
+        enum_value_descriptor.GetOptions().Extensions[
+            unittest_custom_options_pb2.enumvalopt].s)
+    self.assertEqual(
+        "ServiceAnnotation",
+        service_descriptor.GetOptions().Extensions[
+            unittest_custom_options_pb2.serviceopt].s)
+    self.assertEqual(
+        "MethodAnnotation",
+        method_descriptor.GetOptions().Extensions[
+            unittest_custom_options_pb2.methodopt].s)
+
+  def testNestedOptions(self):
+    nested_message =\
+        unittest_custom_options_pb2.NestedOptionType.NestedMessage.DESCRIPTOR
+    self.assertEqual(1001, nested_message.GetOptions().Extensions[
+        unittest_custom_options_pb2.message_opt1])
+    nested_field = nested_message.fields_by_name["nested_field"]
+    self.assertEqual(1002, nested_field.GetOptions().Extensions[
+        unittest_custom_options_pb2.field_opt1])
+    outer_message =\
+        unittest_custom_options_pb2.NestedOptionType.DESCRIPTOR
+    nested_enum = outer_message.enum_types_by_name["NestedEnum"]
+    self.assertEqual(1003, nested_enum.GetOptions().Extensions[
+        unittest_custom_options_pb2.enum_opt1])
+    nested_enum_value = outer_message.enum_values_by_name["NESTED_ENUM_VALUE"]
+    self.assertEqual(1004, nested_enum_value.GetOptions().Extensions[
+        unittest_custom_options_pb2.enum_value_opt1])
+    nested_extension = outer_message.extensions_by_name["nested_extension"]
+    self.assertEqual(1005, nested_extension.GetOptions().Extensions[
+        unittest_custom_options_pb2.field_opt2])
+
+  def testFileDescriptorReferences(self):
+    self.assertEqual(self.my_enum.file, self.my_file)
+    self.assertEqual(self.my_message.file, self.my_file)
+
+  def testFileDescriptor(self):
+    self.assertEqual(self.my_file.name, 'some/filename/some.proto')
+    self.assertEqual(self.my_file.package, 'protobuf_unittest')
+    self.assertEqual(self.my_file.pool, self.pool)
+    # Generated modules also belong to the default pool.
+    self.assertEqual(unittest_pb2.DESCRIPTOR.pool, descriptor_pool.Default())
+
+  @unittest.skipIf(
+      api_implementation.Type() != 'cpp' or api_implementation.Version() != 2,
+      'Immutability of descriptors is only enforced in v2 implementation')
+  def testImmutableCppDescriptor(self):
+    message_descriptor = unittest_pb2.TestAllTypes.DESCRIPTOR
+    with self.assertRaises(AttributeError):
+      message_descriptor.fields_by_name = None
+    with self.assertRaises(TypeError):
+      message_descriptor.fields_by_name['Another'] = None
+    with self.assertRaises(TypeError):
+      message_descriptor.fields.append(None)
+
+
+class NewDescriptorTest(DescriptorTest):
+  """Redo the same tests as above, but with a separate DescriptorPool."""
+
+  def GetDescriptorPool(self):
+    return descriptor_pool.DescriptorPool()
+
+
+class GeneratedDescriptorTest(unittest.TestCase):
+  """Tests for the properties of descriptors in generated code."""
+
+  def CheckMessageDescriptor(self, message_descriptor):
+    # Basic properties
+    self.assertEqual(message_descriptor.name, 'TestAllTypes')
+    self.assertEqual(message_descriptor.full_name,
+                     'protobuf_unittest.TestAllTypes')
+    # Test equality and hashability
+    self.assertEqual(message_descriptor, message_descriptor)
+    self.assertEqual(message_descriptor.fields[0].containing_type,
+                     message_descriptor)
+    self.assertIn(message_descriptor, [message_descriptor])
+    self.assertIn(message_descriptor, {message_descriptor: None})
+    # Test field containers
+    self.CheckDescriptorSequence(message_descriptor.fields)
+    self.CheckDescriptorMapping(message_descriptor.fields_by_name)
+    self.CheckDescriptorMapping(message_descriptor.fields_by_number)
+    self.CheckDescriptorMapping(message_descriptor.fields_by_camelcase_name)
+
+  def CheckFieldDescriptor(self, field_descriptor):
+    # Basic properties
+    self.assertEqual(field_descriptor.name, 'optional_int32')
+    self.assertEqual(field_descriptor.camelcase_name, 'optionalInt32')
+    self.assertEqual(field_descriptor.full_name,
+                     'protobuf_unittest.TestAllTypes.optional_int32')
+    self.assertEqual(field_descriptor.containing_type.name, 'TestAllTypes')
+    # Test equality and hashability
+    self.assertEqual(field_descriptor, field_descriptor)
+    self.assertEqual(
+        field_descriptor.containing_type.fields_by_name['optional_int32'],
+        field_descriptor)
+    self.assertEqual(
+        field_descriptor.containing_type.fields_by_camelcase_name[
+            'optionalInt32'],
+        field_descriptor)
+    self.assertIn(field_descriptor, [field_descriptor])
+    self.assertIn(field_descriptor, {field_descriptor: None})
+
+  def CheckDescriptorSequence(self, sequence):
+    # Verifies that a property like 'messageDescriptor.fields' has all the
+    # properties of an immutable abc.Sequence.
+    self.assertGreater(len(sequence), 0)  # Sized
+    self.assertEqual(len(sequence), len(list(sequence)))  # Iterable
+    item = sequence[0]
+    self.assertEqual(item, sequence[0])
+    self.assertIn(item, sequence)  # Container
+    self.assertEqual(sequence.index(item), 0)
+    self.assertEqual(sequence.count(item), 1)
+    reversed_iterator = reversed(sequence)
+    self.assertEqual(list(reversed_iterator), list(sequence)[::-1])
+    self.assertRaises(StopIteration, next, reversed_iterator)
+
+  def CheckDescriptorMapping(self, mapping):
+    # Verifies that a property like 'messageDescriptor.fields' has all the
+    # properties of an immutable abc.Mapping.
+    self.assertGreater(len(mapping), 0)  # Sized
+    self.assertEqual(len(mapping), len(list(mapping)))  # Iterable
+    if sys.version_info >= (3,):
+      key, item = next(iter(mapping.items()))
+    else:
+      key, item = mapping.items()[0]
+    self.assertIn(key, mapping)  # Container
+    self.assertEqual(mapping.get(key), item)
+    # keys(), iterkeys() &co
+    item = (next(iter(mapping.keys())), next(iter(mapping.values())))
+    self.assertEqual(item, next(iter(mapping.items())))
+    if sys.version_info < (3,):
+      def CheckItems(seq, iterator):
+        self.assertEqual(next(iterator), seq[0])
+        self.assertEqual(list(iterator), seq[1:])
+      CheckItems(mapping.keys(), mapping.iterkeys())
+      CheckItems(mapping.values(), mapping.itervalues())
+      CheckItems(mapping.items(), mapping.iteritems())
+
+  def testDescriptor(self):
+    message_descriptor = unittest_pb2.TestAllTypes.DESCRIPTOR
+    self.CheckMessageDescriptor(message_descriptor)
+    field_descriptor = message_descriptor.fields_by_name['optional_int32']
+    self.CheckFieldDescriptor(field_descriptor)
+    field_descriptor = message_descriptor.fields_by_camelcase_name[
+        'optionalInt32']
+    self.CheckFieldDescriptor(field_descriptor)
+
+  def testCppDescriptorContainer(self):
+    # Check that the collection is still valid even if the parent disappeared.
+    enum = unittest_pb2.TestAllTypes.DESCRIPTOR.enum_types_by_name['NestedEnum']
+    values = enum.values
+    del enum
+    self.assertEqual('FOO', values[0].name)
+
+  def testCppDescriptorContainer_Iterator(self):
+    # Same test with the iterator
+    enum = unittest_pb2.TestAllTypes.DESCRIPTOR.enum_types_by_name['NestedEnum']
+    values_iter = iter(enum.values)
+    del enum
+    self.assertEqual('FOO', next(values_iter).name)
+
+
+class DescriptorCopyToProtoTest(unittest.TestCase):
+  """Tests for CopyTo functions of Descriptor."""
+
+  def _AssertProtoEqual(self, actual_proto, expected_class, expected_ascii):
+    expected_proto = expected_class()
+    text_format.Merge(expected_ascii, expected_proto)
+
+    self.assertEqual(
+        actual_proto, expected_proto,
+        'Not equal,\nActual:\n%s\nExpected:\n%s\n'
+        % (str(actual_proto), str(expected_proto)))
+
+  def _InternalTestCopyToProto(self, desc, expected_proto_class,
+                               expected_proto_ascii):
+    actual = expected_proto_class()
+    desc.CopyToProto(actual)
+    self._AssertProtoEqual(
+        actual, expected_proto_class, expected_proto_ascii)
+
+  def testCopyToProto_EmptyMessage(self):
+    self._InternalTestCopyToProto(
+        unittest_pb2.TestEmptyMessage.DESCRIPTOR,
+        descriptor_pb2.DescriptorProto,
+        TEST_EMPTY_MESSAGE_DESCRIPTOR_ASCII)
+
+  def testCopyToProto_NestedMessage(self):
+    TEST_NESTED_MESSAGE_ASCII = """
+      name: 'NestedMessage'
+      field: <
+        name: 'bb'
+        number: 1
+        label: 1  # Optional
+        type: 5  # TYPE_INT32
+      >
+      """
+
+    self._InternalTestCopyToProto(
+        unittest_pb2.TestAllTypes.NestedMessage.DESCRIPTOR,
+        descriptor_pb2.DescriptorProto,
+        TEST_NESTED_MESSAGE_ASCII)
+
+  def testCopyToProto_ForeignNestedMessage(self):
+    TEST_FOREIGN_NESTED_ASCII = """
+      name: 'TestForeignNested'
+      field: <
+        name: 'foreign_nested'
+        number: 1
+        label: 1  # Optional
+        type: 11  # TYPE_MESSAGE
+        type_name: '.protobuf_unittest.TestAllTypes.NestedMessage'
+      >
+      """
+
+    self._InternalTestCopyToProto(
+        unittest_pb2.TestForeignNested.DESCRIPTOR,
+        descriptor_pb2.DescriptorProto,
+        TEST_FOREIGN_NESTED_ASCII)
+
+  def testCopyToProto_ForeignEnum(self):
+    TEST_FOREIGN_ENUM_ASCII = """
+      name: 'ForeignEnum'
+      value: <
+        name: 'FOREIGN_FOO'
+        number: 4
+      >
+      value: <
+        name: 'FOREIGN_BAR'
+        number: 5
+      >
+      value: <
+        name: 'FOREIGN_BAZ'
+        number: 6
+      >
+      """
+
+    self._InternalTestCopyToProto(
+        unittest_pb2._FOREIGNENUM,
+        descriptor_pb2.EnumDescriptorProto,
+        TEST_FOREIGN_ENUM_ASCII)
+
+  def testCopyToProto_Options(self):
+    TEST_DEPRECATED_FIELDS_ASCII = """
+      name: 'TestDeprecatedFields'
+      field: <
+        name: 'deprecated_int32'
+        number: 1
+        label: 1  # Optional
+        type: 5  # TYPE_INT32
+        options: <
+          deprecated: true
+        >
+      >
+      """
+
+    self._InternalTestCopyToProto(
+        unittest_pb2.TestDeprecatedFields.DESCRIPTOR,
+        descriptor_pb2.DescriptorProto,
+        TEST_DEPRECATED_FIELDS_ASCII)
+
+  def testCopyToProto_AllExtensions(self):
+    TEST_EMPTY_MESSAGE_WITH_EXTENSIONS_ASCII = """
+      name: 'TestEmptyMessageWithExtensions'
+      extension_range: <
+        start: 1
+        end: 536870912
+      >
+      """
+
+    self._InternalTestCopyToProto(
+        unittest_pb2.TestEmptyMessageWithExtensions.DESCRIPTOR,
+        descriptor_pb2.DescriptorProto,
+        TEST_EMPTY_MESSAGE_WITH_EXTENSIONS_ASCII)
+
+  def testCopyToProto_SeveralExtensions(self):
+    TEST_MESSAGE_WITH_SEVERAL_EXTENSIONS_ASCII = """
+      name: 'TestMultipleExtensionRanges'
+      extension_range: <
+        start: 42
+        end: 43
+      >
+      extension_range: <
+        start: 4143
+        end: 4244
+      >
+      extension_range: <
+        start: 65536
+        end: 536870912
+      >
+      """
+
+    self._InternalTestCopyToProto(
+        unittest_pb2.TestMultipleExtensionRanges.DESCRIPTOR,
+        descriptor_pb2.DescriptorProto,
+        TEST_MESSAGE_WITH_SEVERAL_EXTENSIONS_ASCII)
+
+  # Disable this test so we can make changes to the proto file.
+  # TODO(xiaofeng): Enable this test after cl/55530659 is submitted.
+  #
+  # def testCopyToProto_FileDescriptor(self):
+  #   UNITTEST_IMPORT_FILE_DESCRIPTOR_ASCII = ("""
+  #     name: 'google/protobuf/unittest_import.proto'
+  #     package: 'protobuf_unittest_import'
+  #     dependency: 'google/protobuf/unittest_import_public.proto'
+  #     message_type: <
+  #       name: 'ImportMessage'
+  #       field: <
+  #         name: 'd'
+  #         number: 1
+  #         label: 1  # Optional
+  #         type: 5  # TYPE_INT32
+  #       >
+  #     >
+  #     """ +
+  #     """enum_type: <
+  #       name: 'ImportEnum'
+  #       value: <
+  #         name: 'IMPORT_FOO'
+  #         number: 7
+  #       >
+  #       value: <
+  #         name: 'IMPORT_BAR'
+  #         number: 8
+  #       >
+  #       value: <
+  #         name: 'IMPORT_BAZ'
+  #         number: 9
+  #       >
+  #     >
+  #     options: <
+  #       java_package: 'com.google.protobuf.test'
+  #       optimize_for: 1  # SPEED
+  #     >
+  #     public_dependency: 0
+  #  """)
+  #  self._InternalTestCopyToProto(
+  #      unittest_import_pb2.DESCRIPTOR,
+  #      descriptor_pb2.FileDescriptorProto,
+  #      UNITTEST_IMPORT_FILE_DESCRIPTOR_ASCII)
+
+  def testCopyToProto_ServiceDescriptor(self):
+    TEST_SERVICE_ASCII = """
+      name: 'TestService'
+      method: <
+        name: 'Foo'
+        input_type: '.protobuf_unittest.FooRequest'
+        output_type: '.protobuf_unittest.FooResponse'
+      >
+      method: <
+        name: 'Bar'
+        input_type: '.protobuf_unittest.BarRequest'
+        output_type: '.protobuf_unittest.BarResponse'
+      >
+      """
+    # TODO(rocking): enable this test after the proto descriptor change is
+    # checked in.
+    #self._InternalTestCopyToProto(
+    #    unittest_pb2.TestService.DESCRIPTOR,
+    #    descriptor_pb2.ServiceDescriptorProto,
+    #    TEST_SERVICE_ASCII)
+
+
+class MakeDescriptorTest(unittest.TestCase):
+
+  def testMakeDescriptorWithNestedFields(self):
+    file_descriptor_proto = descriptor_pb2.FileDescriptorProto()
+    file_descriptor_proto.name = 'Foo2'
+    message_type = file_descriptor_proto.message_type.add()
+    message_type.name = file_descriptor_proto.name
+    nested_type = message_type.nested_type.add()
+    nested_type.name = 'Sub'
+    enum_type = nested_type.enum_type.add()
+    enum_type.name = 'FOO'
+    enum_type_val = enum_type.value.add()
+    enum_type_val.name = 'BAR'
+    enum_type_val.number = 3
+    field = message_type.field.add()
+    field.number = 1
+    field.name = 'uint64_field'
+    field.label = descriptor.FieldDescriptor.LABEL_REQUIRED
+    field.type = descriptor.FieldDescriptor.TYPE_UINT64
+    field = message_type.field.add()
+    field.number = 2
+    field.name = 'nested_message_field'
+    field.label = descriptor.FieldDescriptor.LABEL_REQUIRED
+    field.type = descriptor.FieldDescriptor.TYPE_MESSAGE
+    field.type_name = 'Sub'
+    enum_field = nested_type.field.add()
+    enum_field.number = 2
+    enum_field.name = 'bar_field'
+    enum_field.label = descriptor.FieldDescriptor.LABEL_REQUIRED
+    enum_field.type = descriptor.FieldDescriptor.TYPE_ENUM
+    enum_field.type_name = 'Foo2.Sub.FOO'
+
+    result = descriptor.MakeDescriptor(message_type)
+    self.assertEqual(result.fields[0].cpp_type,
+                     descriptor.FieldDescriptor.CPPTYPE_UINT64)
+    self.assertEqual(result.fields[1].cpp_type,
+                     descriptor.FieldDescriptor.CPPTYPE_MESSAGE)
+    self.assertEqual(result.fields[1].message_type.containing_type,
+                     result)
+    self.assertEqual(result.nested_types[0].fields[0].full_name,
+                     'Foo2.Sub.bar_field')
+    self.assertEqual(result.nested_types[0].fields[0].enum_type,
+                     result.nested_types[0].enum_types[0])
+
+  def testMakeDescriptorWithUnsignedIntField(self):
+    file_descriptor_proto = descriptor_pb2.FileDescriptorProto()
+    file_descriptor_proto.name = 'Foo'
+    message_type = file_descriptor_proto.message_type.add()
+    message_type.name = file_descriptor_proto.name
+    enum_type = message_type.enum_type.add()
+    enum_type.name = 'FOO'
+    enum_type_val = enum_type.value.add()
+    enum_type_val.name = 'BAR'
+    enum_type_val.number = 3
+    field = message_type.field.add()
+    field.number = 1
+    field.name = 'uint64_field'
+    field.label = descriptor.FieldDescriptor.LABEL_REQUIRED
+    field.type = descriptor.FieldDescriptor.TYPE_UINT64
+    enum_field = message_type.field.add()
+    enum_field.number = 2
+    enum_field.name = 'bar_field'
+    enum_field.label = descriptor.FieldDescriptor.LABEL_REQUIRED
+    enum_field.type = descriptor.FieldDescriptor.TYPE_ENUM
+    enum_field.type_name = 'Foo.FOO'
+
+    result = descriptor.MakeDescriptor(message_type)
+    self.assertEqual(result.fields[0].cpp_type,
+                     descriptor.FieldDescriptor.CPPTYPE_UINT64)
+
+
+  def testMakeDescriptorWithOptions(self):
+    descriptor_proto = descriptor_pb2.DescriptorProto()
+    aggregate_message = unittest_custom_options_pb2.AggregateMessage
+    aggregate_message.DESCRIPTOR.CopyToProto(descriptor_proto)
+    reformed_descriptor = descriptor.MakeDescriptor(descriptor_proto)
+
+    options = reformed_descriptor.GetOptions()
+    self.assertEqual(101,
+                      options.Extensions[unittest_custom_options_pb2.msgopt].i)
+
+  def testCamelcaseName(self):
+    descriptor_proto = descriptor_pb2.DescriptorProto()
+    descriptor_proto.name = 'Bar'
+    names = ['foo_foo', 'FooBar', 'fooBaz', 'fooFoo', 'foobar']
+    camelcase_names = ['fooFoo', 'fooBar', 'fooBaz', 'fooFoo', 'foobar']
+    for index in range(len(names)):
+      field = descriptor_proto.field.add()
+      field.number = index + 1
+      field.name = names[index]
+    result = descriptor.MakeDescriptor(descriptor_proto)
+    for index in range(len(camelcase_names)):
+      self.assertEqual(result.fields[index].camelcase_name,
+                       camelcase_names[index])
+
+
+if __name__ == '__main__':
+  unittest.main()
diff --git a/python/google/protobuf/internal/encoder.py b/python/google/protobuf/internal/encoder.py
new file mode 100755
index 0000000..48ef2df
--- /dev/null
+++ b/python/google/protobuf/internal/encoder.py
@@ -0,0 +1,823 @@
+# Protocol Buffers - Google's data interchange format
+# Copyright 2008 Google Inc.  All rights reserved.
+# https://developers.google.com/protocol-buffers/
+#
+# Redistribution and use in source and binary forms, with or without
+# modification, are permitted provided that the following conditions are
+# met:
+#
+#     * Redistributions of source code must retain the above copyright
+# notice, this list of conditions and the following disclaimer.
+#     * Redistributions in binary form must reproduce the above
+# copyright notice, this list of conditions and the following disclaimer
+# in the documentation and/or other materials provided with the
+# distribution.
+#     * Neither the name of Google Inc. nor the names of its
+# contributors may be used to endorse or promote products derived from
+# this software without specific prior written permission.
+#
+# THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+# "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
+# LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
+# A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
+# OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
+# SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
+# LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
+# DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
+# THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
+# (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+# OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+
+"""Code for encoding protocol message primitives.
+
+Contains the logic for encoding every logical protocol field type
+into one of the 5 physical wire types.
+
+This code is designed to push the Python interpreter's performance to the
+limits.
+
+The basic idea is that at startup time, for every field (i.e. every
+FieldDescriptor) we construct two functions:  a "sizer" and an "encoder".  The
+sizer takes a value of this field's type and computes its byte size.  The
+encoder takes a writer function and a value.  It encodes the value into byte
+strings and invokes the writer function to write those strings.  Typically the
+writer function is the write() method of a BytesIO.
+
+We try to do as much work as possible when constructing the writer and the
+sizer rather than when calling them.  In particular:
+* We copy any needed global functions to local variables, so that we do not need
+  to do costly global table lookups at runtime.
+* Similarly, we try to do any attribute lookups at startup time if possible.
+* Every field's tag is encoded to bytes at startup, since it can't change at
+  runtime.
+* Whatever component of the field size we can compute at startup, we do.
+* We *avoid* sharing code if doing so would make the code slower and not sharing
+  does not burden us too much.  For example, encoders for repeated fields do
+  not just call the encoders for singular fields in a loop because this would
+  add an extra function call overhead for every loop iteration; instead, we
+  manually inline the single-value encoder into the loop.
+* If a Python function lacks a return statement, Python actually generates
+  instructions to pop the result of the last statement off the stack, push
+  None onto the stack, and then return that.  If we really don't care what
+  value is returned, then we can save two instructions by returning the
+  result of the last statement.  It looks funny but it helps.
+* We assume that type and bounds checking has happened at a higher level.
+"""
+
+__author__ = 'kenton@google.com (Kenton Varda)'
+
+import struct
+
+import six
+
+from google.protobuf.internal import wire_format
+
+
+# This will overflow and thus become IEEE-754 "infinity".  We would use
+# "float('inf')" but it doesn't work on Windows pre-Python-2.6.
+_POS_INF = 1e10000
+_NEG_INF = -_POS_INF
+
+
+def _VarintSize(value):
+  """Compute the size of a varint value."""
+  if value <= 0x7f: return 1
+  if value <= 0x3fff: return 2
+  if value <= 0x1fffff: return 3
+  if value <= 0xfffffff: return 4
+  if value <= 0x7ffffffff: return 5
+  if value <= 0x3ffffffffff: return 6
+  if value <= 0x1ffffffffffff: return 7
+  if value <= 0xffffffffffffff: return 8
+  if value <= 0x7fffffffffffffff: return 9
+  return 10
+
+
+def _SignedVarintSize(value):
+  """Compute the size of a signed varint value."""
+  if value < 0: return 10
+  if value <= 0x7f: return 1
+  if value <= 0x3fff: return 2
+  if value <= 0x1fffff: return 3
+  if value <= 0xfffffff: return 4
+  if value <= 0x7ffffffff: return 5
+  if value <= 0x3ffffffffff: return 6
+  if value <= 0x1ffffffffffff: return 7
+  if value <= 0xffffffffffffff: return 8
+  if value <= 0x7fffffffffffffff: return 9
+  return 10
+
+
+def _TagSize(field_number):
+  """Returns the number of bytes required to serialize a tag with this field
+  number."""
+  # Just pass in type 0, since the type won't affect the tag+type size.
+  return _VarintSize(wire_format.PackTag(field_number, 0))
+
+
+# --------------------------------------------------------------------
+# In this section we define some generic sizers.  Each of these functions
+# takes parameters specific to a particular field type, e.g. int32 or fixed64.
+# It returns another function which in turn takes parameters specific to a
+# particular field, e.g. the field number and whether it is repeated or packed.
+# Look at the next section to see how these are used.
+
+
+def _SimpleSizer(compute_value_size):
+  """A sizer which uses the function compute_value_size to compute the size of
+  each value.  Typically compute_value_size is _VarintSize."""
+
+  def SpecificSizer(field_number, is_repeated, is_packed):
+    tag_size = _TagSize(field_number)
+    if is_packed:
+      local_VarintSize = _VarintSize
+      def PackedFieldSize(value):
+        result = 0
+        for element in value:
+          result += compute_value_size(element)
+        return result + local_VarintSize(result) + tag_size
+      return PackedFieldSize
+    elif is_repeated:
+      def RepeatedFieldSize(value):
+        result = tag_size * len(value)
+        for element in value:
+          result += compute_value_size(element)
+        return result
+      return RepeatedFieldSize
+    else:
+      def FieldSize(value):
+        return tag_size + compute_value_size(value)
+      return FieldSize
+
+  return SpecificSizer
+
+
+def _ModifiedSizer(compute_value_size, modify_value):
+  """Like SimpleSizer, but modify_value is invoked on each value before it is
+  passed to compute_value_size.  modify_value is typically ZigZagEncode."""
+
+  def SpecificSizer(field_number, is_repeated, is_packed):
+    tag_size = _TagSize(field_number)
+    if is_packed:
+      local_VarintSize = _VarintSize
+      def PackedFieldSize(value):
+        result = 0
+        for element in value:
+          result += compute_value_size(modify_value(element))
+        return result + local_VarintSize(result) + tag_size
+      return PackedFieldSize
+    elif is_repeated:
+      def RepeatedFieldSize(value):
+        result = tag_size * len(value)
+        for element in value:
+          result += compute_value_size(modify_value(element))
+        return result
+      return RepeatedFieldSize
+    else:
+      def FieldSize(value):
+        return tag_size + compute_value_size(modify_value(value))
+      return FieldSize
+
+  return SpecificSizer
+
+
+def _FixedSizer(value_size):
+  """Like _SimpleSizer except for a fixed-size field.  The input is the size
+  of one value."""
+
+  def SpecificSizer(field_number, is_repeated, is_packed):
+    tag_size = _TagSize(field_number)
+    if is_packed:
+      local_VarintSize = _VarintSize
+      def PackedFieldSize(value):
+        result = len(value) * value_size
+        return result + local_VarintSize(result) + tag_size
+      return PackedFieldSize
+    elif is_repeated:
+      element_size = value_size + tag_size
+      def RepeatedFieldSize(value):
+        return len(value) * element_size
+      return RepeatedFieldSize
+    else:
+      field_size = value_size + tag_size
+      def FieldSize(value):
+        return field_size
+      return FieldSize
+
+  return SpecificSizer
+
+
+# ====================================================================
+# Here we declare a sizer constructor for each field type.  Each "sizer
+# constructor" is a function that takes (field_number, is_repeated, is_packed)
+# as parameters and returns a sizer, which in turn takes a field value as
+# a parameter and returns its encoded size.
+
+
+Int32Sizer = Int64Sizer = EnumSizer = _SimpleSizer(_SignedVarintSize)
+
+UInt32Sizer = UInt64Sizer = _SimpleSizer(_VarintSize)
+
+SInt32Sizer = SInt64Sizer = _ModifiedSizer(
+    _SignedVarintSize, wire_format.ZigZagEncode)
+
+Fixed32Sizer = SFixed32Sizer = FloatSizer  = _FixedSizer(4)
+Fixed64Sizer = SFixed64Sizer = DoubleSizer = _FixedSizer(8)
+
+BoolSizer = _FixedSizer(1)
+
+
+def StringSizer(field_number, is_repeated, is_packed):
+  """Returns a sizer for a string field."""
+
+  tag_size = _TagSize(field_number)
+  local_VarintSize = _VarintSize
+  local_len = len
+  assert not is_packed
+  if is_repeated:
+    def RepeatedFieldSize(value):
+      result = tag_size * len(value)
+      for element in value:
+        l = local_len(element.encode('utf-8'))
+        result += local_VarintSize(l) + l
+      return result
+    return RepeatedFieldSize
+  else:
+    def FieldSize(value):
+      l = local_len(value.encode('utf-8'))
+      return tag_size + local_VarintSize(l) + l
+    return FieldSize
+
+
+def BytesSizer(field_number, is_repeated, is_packed):
+  """Returns a sizer for a bytes field."""
+
+  tag_size = _TagSize(field_number)
+  local_VarintSize = _VarintSize
+  local_len = len
+  assert not is_packed
+  if is_repeated:
+    def RepeatedFieldSize(value):
+      result = tag_size * len(value)
+      for element in value:
+        l = local_len(element)
+        result += local_VarintSize(l) + l
+      return result
+    return RepeatedFieldSize
+  else:
+    def FieldSize(value):
+      l = local_len(value)
+      return tag_size + local_VarintSize(l) + l
+    return FieldSize
+
+
+def GroupSizer(field_number, is_repeated, is_packed):
+  """Returns a sizer for a group field."""
+
+  tag_size = _TagSize(field_number) * 2
+  assert not is_packed
+  if is_repeated:
+    def RepeatedFieldSize(value):
+      result = tag_size * len(value)
+      for element in value:
+        result += element.ByteSize()
+      return result
+    return RepeatedFieldSize
+  else:
+    def FieldSize(value):
+      return tag_size + value.ByteSize()
+    return FieldSize
+
+
+def MessageSizer(field_number, is_repeated, is_packed):
+  """Returns a sizer for a message field."""
+
+  tag_size = _TagSize(field_number)
+  local_VarintSize = _VarintSize
+  assert not is_packed
+  if is_repeated:
+    def RepeatedFieldSize(value):
+      result = tag_size * len(value)
+      for element in value:
+        l = element.ByteSize()
+        result += local_VarintSize(l) + l
+      return result
+    return RepeatedFieldSize
+  else:
+    def FieldSize(value):
+      l = value.ByteSize()
+      return tag_size + local_VarintSize(l) + l
+    return FieldSize
+
+
+# --------------------------------------------------------------------
+# MessageSet is special: it needs custom logic to compute its size properly.
+
+
+def MessageSetItemSizer(field_number):
+  """Returns a sizer for extensions of MessageSet.
+
+  The message set message looks like this:
+    message MessageSet {
+      repeated group Item = 1 {
+        required int32 type_id = 2;
+        required string message = 3;
+      }
+    }
+  """
+  static_size = (_TagSize(1) * 2 + _TagSize(2) + _VarintSize(field_number) +
+                 _TagSize(3))
+  local_VarintSize = _VarintSize
+
+  def FieldSize(value):
+    l = value.ByteSize()
+    return static_size + local_VarintSize(l) + l
+
+  return FieldSize
+
+
+# --------------------------------------------------------------------
+# Map is special: it needs custom logic to compute its size properly.
+
+
+def MapSizer(field_descriptor):
+  """Returns a sizer for a map field."""
+
+  # Can't look at field_descriptor.message_type._concrete_class because it may
+  # not have been initialized yet.
+  message_type = field_descriptor.message_type
+  message_sizer = MessageSizer(field_descriptor.number, False, False)
+
+  def FieldSize(map_value):
+    total = 0
+    for key in map_value:
+      value = map_value[key]
+      # It's wasteful to create the messages and throw them away one second
+      # later since we'll do the same for the actual encode.  But there's not an
+      # obvious way to avoid this within the current design without tons of code
+      # duplication.
+      entry_msg = message_type._concrete_class(key=key, value=value)
+      total += message_sizer(entry_msg)
+    return total
+
+  return FieldSize
+
+# ====================================================================
+# Encoders!
+
+
+def _VarintEncoder():
+  """Return an encoder for a basic varint value (does not include tag)."""
+
+  def EncodeVarint(write, value):
+    bits = value & 0x7f
+    value >>= 7
+    while value:
+      write(six.int2byte(0x80|bits))
+      bits = value & 0x7f
+      value >>= 7
+    return write(six.int2byte(bits))
+
+  return EncodeVarint
+
+
+def _SignedVarintEncoder():
+  """Return an encoder for a basic signed varint value (does not include
+  tag)."""
+
+  def EncodeSignedVarint(write, value):
+    if value < 0:
+      value += (1 << 64)
+    bits = value & 0x7f
+    value >>= 7
+    while value:
+      write(six.int2byte(0x80|bits))
+      bits = value & 0x7f
+      value >>= 7
+    return write(six.int2byte(bits))
+
+  return EncodeSignedVarint
+
+
+_EncodeVarint = _VarintEncoder()
+_EncodeSignedVarint = _SignedVarintEncoder()
+
+
+def _VarintBytes(value):
+  """Encode the given integer as a varint and return the bytes.  This is only
+  called at startup time so it doesn't need to be fast."""
+
+  pieces = []
+  _EncodeVarint(pieces.append, value)
+  return b"".join(pieces)
+
+
+def TagBytes(field_number, wire_type):
+  """Encode the given tag and return the bytes.  Only called at startup."""
+
+  return _VarintBytes(wire_format.PackTag(field_number, wire_type))
+
+# --------------------------------------------------------------------
+# As with sizers (see above), we have a number of common encoder
+# implementations.
+
+
+def _SimpleEncoder(wire_type, encode_value, compute_value_size):
+  """Return a constructor for an encoder for fields of a particular type.
+
+  Args:
+      wire_type:  The field's wire type, for encoding tags.
+      encode_value:  A function which encodes an individual value, e.g.
+        _EncodeVarint().
+      compute_value_size:  A function which computes the size of an individual
+        value, e.g. _VarintSize().
+  """
+
+  def SpecificEncoder(field_number, is_repeated, is_packed):
+    if is_packed:
+      tag_bytes = TagBytes(field_number, wire_format.WIRETYPE_LENGTH_DELIMITED)
+      local_EncodeVarint = _EncodeVarint
+      def EncodePackedField(write, value):
+        write(tag_bytes)
+        size = 0
+        for element in value:
+          size += compute_value_size(element)
+        local_EncodeVarint(write, size)
+        for element in value:
+          encode_value(write, element)
+      return EncodePackedField
+    elif is_repeated:
+      tag_bytes = TagBytes(field_number, wire_type)
+      def EncodeRepeatedField(write, value):
+        for element in value:
+          write(tag_bytes)
+          encode_value(write, element)
+      return EncodeRepeatedField
+    else:
+      tag_bytes = TagBytes(field_number, wire_type)
+      def EncodeField(write, value):
+        write(tag_bytes)
+        return encode_value(write, value)
+      return EncodeField
+
+  return SpecificEncoder
+
+
+def _ModifiedEncoder(wire_type, encode_value, compute_value_size, modify_value):
+  """Like SimpleEncoder but additionally invokes modify_value on every value
+  before passing it to encode_value.  Usually modify_value is ZigZagEncode."""
+
+  def SpecificEncoder(field_number, is_repeated, is_packed):
+    if is_packed:
+      tag_bytes = TagBytes(field_number, wire_format.WIRETYPE_LENGTH_DELIMITED)
+      local_EncodeVarint = _EncodeVarint
+      def EncodePackedField(write, value):
+        write(tag_bytes)
+        size = 0
+        for element in value:
+          size += compute_value_size(modify_value(element))
+        local_EncodeVarint(write, size)
+        for element in value:
+          encode_value(write, modify_value(element))
+      return EncodePackedField
+    elif is_repeated:
+      tag_bytes = TagBytes(field_number, wire_type)
+      def EncodeRepeatedField(write, value):
+        for element in value:
+          write(tag_bytes)
+          encode_value(write, modify_value(element))
+      return EncodeRepeatedField
+    else:
+      tag_bytes = TagBytes(field_number, wire_type)
+      def EncodeField(write, value):
+        write(tag_bytes)
+        return encode_value(write, modify_value(value))
+      return EncodeField
+
+  return SpecificEncoder
+
+
+def _StructPackEncoder(wire_type, format):
+  """Return a constructor for an encoder for a fixed-width field.
+
+  Args:
+      wire_type:  The field's wire type, for encoding tags.
+      format:  The format string to pass to struct.pack().
+  """
+
+  value_size = struct.calcsize(format)
+
+  def SpecificEncoder(field_number, is_repeated, is_packed):
+    local_struct_pack = struct.pack
+    if is_packed:
+      tag_bytes = TagBytes(field_number, wire_format.WIRETYPE_LENGTH_DELIMITED)
+      local_EncodeVarint = _EncodeVarint
+      def EncodePackedField(write, value):
+        write(tag_bytes)
+        local_EncodeVarint(write, len(value) * value_size)
+        for element in value:
+          write(local_struct_pack(format, element))
+      return EncodePackedField
+    elif is_repeated:
+      tag_bytes = TagBytes(field_number, wire_type)
+      def EncodeRepeatedField(write, value):
+        for element in value:
+          write(tag_bytes)
+          write(local_struct_pack(format, element))
+      return EncodeRepeatedField
+    else:
+      tag_bytes = TagBytes(field_number, wire_type)
+      def EncodeField(write, value):
+        write(tag_bytes)
+        return write(local_struct_pack(format, value))
+      return EncodeField
+
+  return SpecificEncoder
+
+
+def _FloatingPointEncoder(wire_type, format):
+  """Return a constructor for an encoder for float fields.
+
+  This is like StructPackEncoder, but catches errors that may be due to
+  passing non-finite floating-point values to struct.pack, and makes a
+  second attempt to encode those values.
+
+  Args:
+      wire_type:  The field's wire type, for encoding tags.
+      format:  The format string to pass to struct.pack().
+  """
+
+  value_size = struct.calcsize(format)
+  if value_size == 4:
+    def EncodeNonFiniteOrRaise(write, value):
+      # Remember that the serialized form uses little-endian byte order.
+      if value == _POS_INF:
+        write(b'\x00\x00\x80\x7F')
+      elif value == _NEG_INF:
+        write(b'\x00\x00\x80\xFF')
+      elif value != value:           # NaN
+        write(b'\x00\x00\xC0\x7F')
+      else:
+        raise
+  elif value_size == 8:
+    def EncodeNonFiniteOrRaise(write, value):
+      if value == _POS_INF:
+        write(b'\x00\x00\x00\x00\x00\x00\xF0\x7F')
+      elif value == _NEG_INF:
+        write(b'\x00\x00\x00\x00\x00\x00\xF0\xFF')
+      elif value != value:                         # NaN
+        write(b'\x00\x00\x00\x00\x00\x00\xF8\x7F')
+      else:
+        raise
+  else:
+    raise ValueError('Can\'t encode floating-point values that are '
+                     '%d bytes long (only 4 or 8)' % value_size)
+
+  def SpecificEncoder(field_number, is_repeated, is_packed):
+    local_struct_pack = struct.pack
+    if is_packed:
+      tag_bytes = TagBytes(field_number, wire_format.WIRETYPE_LENGTH_DELIMITED)
+      local_EncodeVarint = _EncodeVarint
+      def EncodePackedField(write, value):
+        write(tag_bytes)
+        local_EncodeVarint(write, len(value) * value_size)
+        for element in value:
+          # This try/except block is going to be faster than any code that
+          # we could write to check whether element is finite.
+          try:
+            write(local_struct_pack(format, element))
+          except SystemError:
+            EncodeNonFiniteOrRaise(write, element)
+      return EncodePackedField
+    elif is_repeated:
+      tag_bytes = TagBytes(field_number, wire_type)
+      def EncodeRepeatedField(write, value):
+        for element in value:
+          write(tag_bytes)
+          try:
+            write(local_struct_pack(format, element))
+          except SystemError:
+            EncodeNonFiniteOrRaise(write, element)
+      return EncodeRepeatedField
+    else:
+      tag_bytes = TagBytes(field_number, wire_type)
+      def EncodeField(write, value):
+        write(tag_bytes)
+        try:
+          write(local_struct_pack(format, value))
+        except SystemError:
+          EncodeNonFiniteOrRaise(write, value)
+      return EncodeField
+
+  return SpecificEncoder
+
+
+# ====================================================================
+# Here we declare an encoder constructor for each field type.  These work
+# very similarly to sizer constructors, described earlier.
+
+
+Int32Encoder = Int64Encoder = EnumEncoder = _SimpleEncoder(
+    wire_format.WIRETYPE_VARINT, _EncodeSignedVarint, _SignedVarintSize)
+
+UInt32Encoder = UInt64Encoder = _SimpleEncoder(
+    wire_format.WIRETYPE_VARINT, _EncodeVarint, _VarintSize)
+
+SInt32Encoder = SInt64Encoder = _ModifiedEncoder(
+    wire_format.WIRETYPE_VARINT, _EncodeVarint, _VarintSize,
+    wire_format.ZigZagEncode)
+
+# Note that Python conveniently guarantees that when using the '<' prefix on
+# formats, they will also have the same size across all platforms (as opposed
+# to without the prefix, where their sizes depend on the C compiler's basic
+# type sizes).
+Fixed32Encoder  = _StructPackEncoder(wire_format.WIRETYPE_FIXED32, '<I')
+Fixed64Encoder  = _StructPackEncoder(wire_format.WIRETYPE_FIXED64, '<Q')
+SFixed32Encoder = _StructPackEncoder(wire_format.WIRETYPE_FIXED32, '<i')
+SFixed64Encoder = _StructPackEncoder(wire_format.WIRETYPE_FIXED64, '<q')
+FloatEncoder    = _FloatingPointEncoder(wire_format.WIRETYPE_FIXED32, '<f')
+DoubleEncoder   = _FloatingPointEncoder(wire_format.WIRETYPE_FIXED64, '<d')
+
+
+def BoolEncoder(field_number, is_repeated, is_packed):
+  """Returns an encoder for a boolean field."""
+
+  false_byte = b'\x00'
+  true_byte = b'\x01'
+  if is_packed:
+    tag_bytes = TagBytes(field_number, wire_format.WIRETYPE_LENGTH_DELIMITED)
+    local_EncodeVarint = _EncodeVarint
+    def EncodePackedField(write, value):
+      write(tag_bytes)
+      local_EncodeVarint(write, len(value))
+      for element in value:
+        if element:
+          write(true_byte)
+        else:
+          write(false_byte)
+    return EncodePackedField
+  elif is_repeated:
+    tag_bytes = TagBytes(field_number, wire_format.WIRETYPE_VARINT)
+    def EncodeRepeatedField(write, value):
+      for element in value:
+        write(tag_bytes)
+        if element:
+          write(true_byte)
+        else:
+          write(false_byte)
+    return EncodeRepeatedField
+  else:
+    tag_bytes = TagBytes(field_number, wire_format.WIRETYPE_VARINT)
+    def EncodeField(write, value):
+      write(tag_bytes)
+      if value:
+        return write(true_byte)
+      return write(false_byte)
+    return EncodeField
+
+
+def StringEncoder(field_number, is_repeated, is_packed):
+  """Returns an encoder for a string field."""
+
+  tag = TagBytes(field_number, wire_format.WIRETYPE_LENGTH_DELIMITED)
+  local_EncodeVarint = _EncodeVarint
+  local_len = len
+  assert not is_packed
+  if is_repeated:
+    def EncodeRepeatedField(write, value):
+      for element in value:
+        encoded = element.encode('utf-8')
+        write(tag)
+        local_EncodeVarint(write, local_len(encoded))
+        write(encoded)
+    return EncodeRepeatedField
+  else:
+    def EncodeField(write, value):
+      encoded = value.encode('utf-8')
+      write(tag)
+      local_EncodeVarint(write, local_len(encoded))
+      return write(encoded)
+    return EncodeField
+
+
+def BytesEncoder(field_number, is_repeated, is_packed):
+  """Returns an encoder for a bytes field."""
+
+  tag = TagBytes(field_number, wire_format.WIRETYPE_LENGTH_DELIMITED)
+  local_EncodeVarint = _EncodeVarint
+  local_len = len
+  assert not is_packed
+  if is_repeated:
+    def EncodeRepeatedField(write, value):
+      for element in value:
+        write(tag)
+        local_EncodeVarint(write, local_len(element))
+        write(element)
+    return EncodeRepeatedField
+  else:
+    def EncodeField(write, value):
+      write(tag)
+      local_EncodeVarint(write, local_len(value))
+      return write(value)
+    return EncodeField
+
+
+def GroupEncoder(field_number, is_repeated, is_packed):
+  """Returns an encoder for a group field."""
+
+  start_tag = TagBytes(field_number, wire_format.WIRETYPE_START_GROUP)
+  end_tag = TagBytes(field_number, wire_format.WIRETYPE_END_GROUP)
+  assert not is_packed
+  if is_repeated:
+    def EncodeRepeatedField(write, value):
+      for element in value:
+        write(start_tag)
+        element._InternalSerialize(write)
+        write(end_tag)
+    return EncodeRepeatedField
+  else:
+    def EncodeField(write, value):
+      write(start_tag)
+      value._InternalSerialize(write)
+      return write(end_tag)
+    return EncodeField
+
+
+def MessageEncoder(field_number, is_repeated, is_packed):
+  """Returns an encoder for a message field."""
+
+  tag = TagBytes(field_number, wire_format.WIRETYPE_LENGTH_DELIMITED)
+  local_EncodeVarint = _EncodeVarint
+  assert not is_packed
+  if is_repeated:
+    def EncodeRepeatedField(write, value):
+      for element in value:
+        write(tag)
+        local_EncodeVarint(write, element.ByteSize())
+        element._InternalSerialize(write)
+    return EncodeRepeatedField
+  else:
+    def EncodeField(write, value):
+      write(tag)
+      local_EncodeVarint(write, value.ByteSize())
+      return value._InternalSerialize(write)
+    return EncodeField
+
+
+# --------------------------------------------------------------------
+# As before, MessageSet is special.
+
+
+def MessageSetItemEncoder(field_number):
+  """Encoder for extensions of MessageSet.
+
+  The message set message looks like this:
+    message MessageSet {
+      repeated group Item = 1 {
+        required int32 type_id = 2;
+        required string message = 3;
+      }
+    }
+  """
+  start_bytes = b"".join([
+      TagBytes(1, wire_format.WIRETYPE_START_GROUP),
+      TagBytes(2, wire_format.WIRETYPE_VARINT),
+      _VarintBytes(field_number),
+      TagBytes(3, wire_format.WIRETYPE_LENGTH_DELIMITED)])
+  end_bytes = TagBytes(1, wire_format.WIRETYPE_END_GROUP)
+  local_EncodeVarint = _EncodeVarint
+
+  def EncodeField(write, value):
+    write(start_bytes)
+    local_EncodeVarint(write, value.ByteSize())
+    value._InternalSerialize(write)
+    return write(end_bytes)
+
+  return EncodeField
+
+
+# --------------------------------------------------------------------
+# As before, Map is special.
+
+
+def MapEncoder(field_descriptor):
+  """Encoder for extensions of MessageSet.
+
+  Maps always have a wire format like this:
+    message MapEntry {
+      key_type key = 1;
+      value_type value = 2;
+    }
+    repeated MapEntry map = N;
+  """
+  # Can't look at field_descriptor.message_type._concrete_class because it may
+  # not have been initialized yet.
+  message_type = field_descriptor.message_type
+  encode_message = MessageEncoder(field_descriptor.number, False, False)
+
+  def EncodeField(write, value):
+    for key in value:
+      entry_msg = message_type._concrete_class(key=key, value=value[key])
+      encode_message(write, entry_msg)
+
+  return EncodeField
diff --git a/python/google/protobuf/internal/enum_type_wrapper.py b/python/google/protobuf/internal/enum_type_wrapper.py
new file mode 100644
index 0000000..1cffe35
--- /dev/null
+++ b/python/google/protobuf/internal/enum_type_wrapper.py
@@ -0,0 +1,89 @@
+# Protocol Buffers - Google's data interchange format
+# Copyright 2008 Google Inc.  All rights reserved.
+# https://developers.google.com/protocol-buffers/
+#
+# Redistribution and use in source and binary forms, with or without
+# modification, are permitted provided that the following conditions are
+# met:
+#
+#     * Redistributions of source code must retain the above copyright
+# notice, this list of conditions and the following disclaimer.
+#     * Redistributions in binary form must reproduce the above
+# copyright notice, this list of conditions and the following disclaimer
+# in the documentation and/or other materials provided with the
+# distribution.
+#     * Neither the name of Google Inc. nor the names of its
+# contributors may be used to endorse or promote products derived from
+# this software without specific prior written permission.
+#
+# THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+# "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
+# LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
+# A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
+# OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
+# SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
+# LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
+# DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
+# THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
+# (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+# OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+
+"""A simple wrapper around enum types to expose utility functions.
+
+Instances are created as properties with the same name as the enum they wrap
+on proto classes.  For usage, see:
+  reflection_test.py
+"""
+
+__author__ = 'rabsatt@google.com (Kevin Rabsatt)'
+
+
+class EnumTypeWrapper(object):
+  """A utility for finding the names of enum values."""
+
+  DESCRIPTOR = None
+
+  def __init__(self, enum_type):
+    """Inits EnumTypeWrapper with an EnumDescriptor."""
+    self._enum_type = enum_type
+    self.DESCRIPTOR = enum_type;
+
+  def Name(self, number):
+    """Returns a string containing the name of an enum value."""
+    if number in self._enum_type.values_by_number:
+      return self._enum_type.values_by_number[number].name
+    raise ValueError('Enum %s has no name defined for value %d' % (
+        self._enum_type.name, number))
+
+  def Value(self, name):
+    """Returns the value coresponding to the given enum name."""
+    if name in self._enum_type.values_by_name:
+      return self._enum_type.values_by_name[name].number
+    raise ValueError('Enum %s has no value defined for name %s' % (
+        self._enum_type.name, name))
+
+  def keys(self):
+    """Return a list of the string names in the enum.
+
+    These are returned in the order they were defined in the .proto file.
+    """
+
+    return [value_descriptor.name
+            for value_descriptor in self._enum_type.values]
+
+  def values(self):
+    """Return a list of the integer values in the enum.
+
+    These are returned in the order they were defined in the .proto file.
+    """
+
+    return [value_descriptor.number
+            for value_descriptor in self._enum_type.values]
+
+  def items(self):
+    """Return a list of the (name, value) pairs of the enum.
+
+    These are returned in the order they were defined in the .proto file.
+    """
+    return [(value_descriptor.name, value_descriptor.number)
+            for value_descriptor in self._enum_type.values]
diff --git a/python/google/protobuf/internal/factory_test1.proto b/python/google/protobuf/internal/factory_test1.proto
new file mode 100644
index 0000000..d2fbbee
--- /dev/null
+++ b/python/google/protobuf/internal/factory_test1.proto
@@ -0,0 +1,58 @@
+// Protocol Buffers - Google's data interchange format
+// Copyright 2008 Google Inc.  All rights reserved.
+// https://developers.google.com/protocol-buffers/
+//
+// Redistribution and use in source and binary forms, with or without
+// modification, are permitted provided that the following conditions are
+// met:
+//
+//     * Redistributions of source code must retain the above copyright
+// notice, this list of conditions and the following disclaimer.
+//     * Redistributions in binary form must reproduce the above
+// copyright notice, this list of conditions and the following disclaimer
+// in the documentation and/or other materials provided with the
+// distribution.
+//     * Neither the name of Google Inc. nor the names of its
+// contributors may be used to endorse or promote products derived from
+// this software without specific prior written permission.
+//
+// THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+// "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
+// LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
+// A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
+// OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
+// SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
+// LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
+// DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
+// THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
+// (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+// OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+
+// Author: matthewtoia@google.com (Matt Toia)
+
+syntax = "proto2";
+
+package google.protobuf.python.internal;
+
+
+enum Factory1Enum {
+  FACTORY_1_VALUE_0 = 0;
+  FACTORY_1_VALUE_1 = 1;
+}
+
+message Factory1Message {
+  optional Factory1Enum factory_1_enum = 1;
+  enum NestedFactory1Enum {
+    NESTED_FACTORY_1_VALUE_0 = 0;
+    NESTED_FACTORY_1_VALUE_1 = 1;
+  }
+  optional NestedFactory1Enum nested_factory_1_enum = 2;
+  message NestedFactory1Message {
+    optional string value = 1;
+  }
+  optional NestedFactory1Message nested_factory_1_message = 3;
+  optional int32 scalar_value = 4;
+  repeated string list_value = 5;
+
+  extensions 1000 to max;
+}
diff --git a/python/google/protobuf/internal/factory_test2.proto b/python/google/protobuf/internal/factory_test2.proto
new file mode 100644
index 0000000..bb1b54a
--- /dev/null
+++ b/python/google/protobuf/internal/factory_test2.proto
@@ -0,0 +1,99 @@
+// Protocol Buffers - Google's data interchange format
+// Copyright 2008 Google Inc.  All rights reserved.
+// https://developers.google.com/protocol-buffers/
+//
+// Redistribution and use in source and binary forms, with or without
+// modification, are permitted provided that the following conditions are
+// met:
+//
+//     * Redistributions of source code must retain the above copyright
+// notice, this list of conditions and the following disclaimer.
+//     * Redistributions in binary form must reproduce the above
+// copyright notice, this list of conditions and the following disclaimer
+// in the documentation and/or other materials provided with the
+// distribution.
+//     * Neither the name of Google Inc. nor the names of its
+// contributors may be used to endorse or promote products derived from
+// this software without specific prior written permission.
+//
+// THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+// "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
+// LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
+// A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
+// OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
+// SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
+// LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
+// DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
+// THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
+// (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+// OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+
+// Author: matthewtoia@google.com (Matt Toia)
+
+syntax = "proto2";
+
+package google.protobuf.python.internal;
+
+import "google/protobuf/internal/factory_test1.proto";
+
+
+enum Factory2Enum {
+  FACTORY_2_VALUE_0 = 0;
+  FACTORY_2_VALUE_1 = 1;
+}
+
+message Factory2Message {
+  required int32 mandatory = 1;
+  optional Factory2Enum factory_2_enum = 2;
+  enum NestedFactory2Enum {
+    NESTED_FACTORY_2_VALUE_0 = 0;
+    NESTED_FACTORY_2_VALUE_1 = 1;
+  }
+  optional NestedFactory2Enum nested_factory_2_enum = 3;
+  message NestedFactory2Message {
+    optional string value = 1;
+  }
+  optional NestedFactory2Message nested_factory_2_message = 4;
+  optional Factory1Message factory_1_message = 5;
+  optional Factory1Enum factory_1_enum = 6;
+  optional Factory1Message.NestedFactory1Enum nested_factory_1_enum = 7;
+  optional Factory1Message.NestedFactory1Message nested_factory_1_message = 8;
+  optional Factory2Message circular_message = 9;
+  optional string scalar_value = 10;
+  repeated string list_value = 11;
+  repeated group Grouped = 12 {
+    optional string part_1 = 13;
+    optional string part_2 = 14;
+  }
+  optional LoopMessage loop = 15;
+  optional int32 int_with_default = 16 [default = 1776];
+  optional double double_with_default = 17 [default = 9.99];
+  optional string string_with_default = 18 [default = "hello world"];
+  optional bool bool_with_default = 19 [default = false];
+  optional Factory2Enum enum_with_default = 20 [default = FACTORY_2_VALUE_1];
+  optional bytes bytes_with_default = 21 [default = "a\373\000c"];
+
+
+  extend Factory1Message {
+    optional string one_more_field = 1001;
+  }
+
+  oneof oneof_field {
+    int32 oneof_int = 22;
+    string oneof_string = 23;
+  }
+}
+
+message LoopMessage {
+  optional Factory2Message loop = 1;
+}
+
+message MessageWithNestedEnumOnly {
+  enum NestedEnum {
+    NESTED_MESSAGE_ENUM_0 = 0;
+  }
+}
+
+extend Factory1Message {
+  optional string another_field = 1002;
+}
diff --git a/python/google/protobuf/internal/generator_test.py b/python/google/protobuf/internal/generator_test.py
new file mode 100755
index 0000000..9956da5
--- /dev/null
+++ b/python/google/protobuf/internal/generator_test.py
@@ -0,0 +1,347 @@
+#! /usr/bin/env python
+#
+# Protocol Buffers - Google's data interchange format
+# Copyright 2008 Google Inc.  All rights reserved.
+# https://developers.google.com/protocol-buffers/
+#
+# Redistribution and use in source and binary forms, with or without
+# modification, are permitted provided that the following conditions are
+# met:
+#
+#     * Redistributions of source code must retain the above copyright
+# notice, this list of conditions and the following disclaimer.
+#     * Redistributions in binary form must reproduce the above
+# copyright notice, this list of conditions and the following disclaimer
+# in the documentation and/or other materials provided with the
+# distribution.
+#     * Neither the name of Google Inc. nor the names of its
+# contributors may be used to endorse or promote products derived from
+# this software without specific prior written permission.
+#
+# THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+# "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
+# LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
+# A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
+# OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
+# SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
+# LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
+# DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
+# THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
+# (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+# OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+
+# TODO(robinson): Flesh this out considerably.  We focused on reflection_test.py
+# first, since it's testing the subtler code, and since it provides decent
+# indirect testing of the protocol compiler output.
+
+"""Unittest that directly tests the output of the pure-Python protocol
+compiler.  See //google/protobuf/internal/reflection_test.py for a test which
+further ensures that we can use Python protocol message objects as we expect.
+"""
+
+__author__ = 'robinson@google.com (Will Robinson)'
+
+try:
+  import unittest2 as unittest
+except ImportError:
+  import unittest
+from google.protobuf.internal import test_bad_identifiers_pb2
+from google.protobuf import unittest_custom_options_pb2
+from google.protobuf import unittest_import_pb2
+from google.protobuf import unittest_import_public_pb2
+from google.protobuf import unittest_mset_pb2
+from google.protobuf import unittest_mset_wire_format_pb2
+from google.protobuf import unittest_no_generic_services_pb2
+from google.protobuf import unittest_pb2
+from google.protobuf import service
+from google.protobuf import symbol_database
+
+MAX_EXTENSION = 536870912
+
+
+class GeneratorTest(unittest.TestCase):
+
+  def testNestedMessageDescriptor(self):
+    field_name = 'optional_nested_message'
+    proto_type = unittest_pb2.TestAllTypes
+    self.assertEqual(
+        proto_type.NestedMessage.DESCRIPTOR,
+        proto_type.DESCRIPTOR.fields_by_name[field_name].message_type)
+
+  def testEnums(self):
+    # We test only module-level enums here.
+    # TODO(robinson): Examine descriptors directly to check
+    # enum descriptor output.
+    self.assertEqual(4, unittest_pb2.FOREIGN_FOO)
+    self.assertEqual(5, unittest_pb2.FOREIGN_BAR)
+    self.assertEqual(6, unittest_pb2.FOREIGN_BAZ)
+
+    proto = unittest_pb2.TestAllTypes()
+    self.assertEqual(1, proto.FOO)
+    self.assertEqual(1, unittest_pb2.TestAllTypes.FOO)
+    self.assertEqual(2, proto.BAR)
+    self.assertEqual(2, unittest_pb2.TestAllTypes.BAR)
+    self.assertEqual(3, proto.BAZ)
+    self.assertEqual(3, unittest_pb2.TestAllTypes.BAZ)
+
+  def testExtremeDefaultValues(self):
+    message = unittest_pb2.TestExtremeDefaultValues()
+
+    # Python pre-2.6 does not have isinf() or isnan() functions, so we have
+    # to provide our own.
+    def isnan(val):
+      # NaN is never equal to itself.
+      return val != val
+    def isinf(val):
+      # Infinity times zero equals NaN.
+      return not isnan(val) and isnan(val * 0)
+
+    self.assertTrue(isinf(message.inf_double))
+    self.assertTrue(message.inf_double > 0)
+    self.assertTrue(isinf(message.neg_inf_double))
+    self.assertTrue(message.neg_inf_double < 0)
+    self.assertTrue(isnan(message.nan_double))
+
+    self.assertTrue(isinf(message.inf_float))
+    self.assertTrue(message.inf_float > 0)
+    self.assertTrue(isinf(message.neg_inf_float))
+    self.assertTrue(message.neg_inf_float < 0)
+    self.assertTrue(isnan(message.nan_float))
+    self.assertEqual("? ? ?? ?? ??? ??/ ??-", message.cpp_trigraph)
+
+  def testHasDefaultValues(self):
+    desc = unittest_pb2.TestAllTypes.DESCRIPTOR
+
+    expected_has_default_by_name = {
+        'optional_int32': False,
+        'repeated_int32': False,
+        'optional_nested_message': False,
+        'default_int32': True,
+    }
+
+    has_default_by_name = dict(
+        [(f.name, f.has_default_value)
+         for f in desc.fields
+         if f.name in expected_has_default_by_name])
+    self.assertEqual(expected_has_default_by_name, has_default_by_name)
+
+  def testContainingTypeBehaviorForExtensions(self):
+    self.assertEqual(unittest_pb2.optional_int32_extension.containing_type,
+                     unittest_pb2.TestAllExtensions.DESCRIPTOR)
+    self.assertEqual(unittest_pb2.TestRequired.single.containing_type,
+                     unittest_pb2.TestAllExtensions.DESCRIPTOR)
+
+  def testExtensionScope(self):
+    self.assertEqual(unittest_pb2.optional_int32_extension.extension_scope,
+                     None)
+    self.assertEqual(unittest_pb2.TestRequired.single.extension_scope,
+                     unittest_pb2.TestRequired.DESCRIPTOR)
+
+  def testIsExtension(self):
+    self.assertTrue(unittest_pb2.optional_int32_extension.is_extension)
+    self.assertTrue(unittest_pb2.TestRequired.single.is_extension)
+
+    message_descriptor = unittest_pb2.TestRequired.DESCRIPTOR
+    non_extension_descriptor = message_descriptor.fields_by_name['a']
+    self.assertTrue(not non_extension_descriptor.is_extension)
+
+  def testOptions(self):
+    proto = unittest_mset_wire_format_pb2.TestMessageSet()
+    self.assertTrue(proto.DESCRIPTOR.GetOptions().message_set_wire_format)
+
+  def testMessageWithCustomOptions(self):
+    proto = unittest_custom_options_pb2.TestMessageWithCustomOptions()
+    enum_options = proto.DESCRIPTOR.enum_types_by_name['AnEnum'].GetOptions()
+    self.assertTrue(enum_options is not None)
+    # TODO(gps): We really should test for the presence of the enum_opt1
+    # extension and for its value to be set to -789.
+
+  def testNestedTypes(self):
+    self.assertEqual(
+        set(unittest_pb2.TestAllTypes.DESCRIPTOR.nested_types),
+        set([
+            unittest_pb2.TestAllTypes.NestedMessage.DESCRIPTOR,
+            unittest_pb2.TestAllTypes.OptionalGroup.DESCRIPTOR,
+            unittest_pb2.TestAllTypes.RepeatedGroup.DESCRIPTOR,
+        ]))
+    self.assertEqual(unittest_pb2.TestEmptyMessage.DESCRIPTOR.nested_types, [])
+    self.assertEqual(
+        unittest_pb2.TestAllTypes.NestedMessage.DESCRIPTOR.nested_types, [])
+
+  def testContainingType(self):
+    self.assertTrue(
+        unittest_pb2.TestEmptyMessage.DESCRIPTOR.containing_type is None)
+    self.assertTrue(
+        unittest_pb2.TestAllTypes.DESCRIPTOR.containing_type is None)
+    self.assertEqual(
+        unittest_pb2.TestAllTypes.NestedMessage.DESCRIPTOR.containing_type,
+        unittest_pb2.TestAllTypes.DESCRIPTOR)
+    self.assertEqual(
+        unittest_pb2.TestAllTypes.NestedMessage.DESCRIPTOR.containing_type,
+        unittest_pb2.TestAllTypes.DESCRIPTOR)
+    self.assertEqual(
+        unittest_pb2.TestAllTypes.RepeatedGroup.DESCRIPTOR.containing_type,
+        unittest_pb2.TestAllTypes.DESCRIPTOR)
+
+  def testContainingTypeInEnumDescriptor(self):
+    self.assertTrue(unittest_pb2._FOREIGNENUM.containing_type is None)
+    self.assertEqual(unittest_pb2._TESTALLTYPES_NESTEDENUM.containing_type,
+                     unittest_pb2.TestAllTypes.DESCRIPTOR)
+
+  def testPackage(self):
+    self.assertEqual(
+        unittest_pb2.TestAllTypes.DESCRIPTOR.file.package,
+        'protobuf_unittest')
+    desc = unittest_pb2.TestAllTypes.NestedMessage.DESCRIPTOR
+    self.assertEqual(desc.file.package, 'protobuf_unittest')
+    self.assertEqual(
+        unittest_import_pb2.ImportMessage.DESCRIPTOR.file.package,
+        'protobuf_unittest_import')
+
+    self.assertEqual(
+        unittest_pb2._FOREIGNENUM.file.package, 'protobuf_unittest')
+    self.assertEqual(
+        unittest_pb2._TESTALLTYPES_NESTEDENUM.file.package,
+        'protobuf_unittest')
+    self.assertEqual(
+        unittest_import_pb2._IMPORTENUM.file.package,
+        'protobuf_unittest_import')
+
+  def testExtensionRange(self):
+    self.assertEqual(
+        unittest_pb2.TestAllTypes.DESCRIPTOR.extension_ranges, [])
+    self.assertEqual(
+        unittest_pb2.TestAllExtensions.DESCRIPTOR.extension_ranges,
+        [(1, MAX_EXTENSION)])
+    self.assertEqual(
+        unittest_pb2.TestMultipleExtensionRanges.DESCRIPTOR.extension_ranges,
+        [(42, 43), (4143, 4244), (65536, MAX_EXTENSION)])
+
+  def testFileDescriptor(self):
+    self.assertEqual(unittest_pb2.DESCRIPTOR.name,
+                     'google/protobuf/unittest.proto')
+    self.assertEqual(unittest_pb2.DESCRIPTOR.package, 'protobuf_unittest')
+    self.assertFalse(unittest_pb2.DESCRIPTOR.serialized_pb is None)
+    self.assertEqual(unittest_pb2.DESCRIPTOR.dependencies,
+                     [unittest_import_pb2.DESCRIPTOR])
+    self.assertEqual(unittest_import_pb2.DESCRIPTOR.dependencies,
+                     [unittest_import_public_pb2.DESCRIPTOR])
+
+  def testNoGenericServices(self):
+    self.assertTrue(hasattr(unittest_no_generic_services_pb2, "TestMessage"))
+    self.assertTrue(hasattr(unittest_no_generic_services_pb2, "FOO"))
+    self.assertTrue(hasattr(unittest_no_generic_services_pb2, "test_extension"))
+
+    # Make sure unittest_no_generic_services_pb2 has no services subclassing
+    # Proto2 Service class.
+    if hasattr(unittest_no_generic_services_pb2, "TestService"):
+      self.assertFalse(issubclass(unittest_no_generic_services_pb2.TestService,
+                                  service.Service))
+
+  def testMessageTypesByName(self):
+    file_type = unittest_pb2.DESCRIPTOR
+    self.assertEqual(
+        unittest_pb2._TESTALLTYPES,
+        file_type.message_types_by_name[unittest_pb2._TESTALLTYPES.name])
+
+    # Nested messages shouldn't be included in the message_types_by_name
+    # dictionary (like in the C++ API).
+    self.assertFalse(
+        unittest_pb2._TESTALLTYPES_NESTEDMESSAGE.name in
+        file_type.message_types_by_name)
+
+  def testEnumTypesByName(self):
+    file_type = unittest_pb2.DESCRIPTOR
+    self.assertEqual(
+        unittest_pb2._FOREIGNENUM,
+        file_type.enum_types_by_name[unittest_pb2._FOREIGNENUM.name])
+
+  def testExtensionsByName(self):
+    file_type = unittest_pb2.DESCRIPTOR
+    self.assertEqual(
+        unittest_pb2.my_extension_string,
+        file_type.extensions_by_name[unittest_pb2.my_extension_string.name])
+
+  def testPublicImports(self):
+    # Test public imports as embedded message.
+    all_type_proto = unittest_pb2.TestAllTypes()
+    self.assertEqual(0, all_type_proto.optional_public_import_message.e)
+
+    # PublicImportMessage is actually defined in unittest_import_public_pb2
+    # module, and is public imported by unittest_import_pb2 module.
+    public_import_proto = unittest_import_pb2.PublicImportMessage()
+    self.assertEqual(0, public_import_proto.e)
+    self.assertTrue(unittest_import_public_pb2.PublicImportMessage is
+                    unittest_import_pb2.PublicImportMessage)
+
+  def testBadIdentifiers(self):
+    # We're just testing that the code was imported without problems.
+    message = test_bad_identifiers_pb2.TestBadIdentifiers()
+    self.assertEqual(message.Extensions[test_bad_identifiers_pb2.message],
+                     "foo")
+    self.assertEqual(message.Extensions[test_bad_identifiers_pb2.descriptor],
+                     "bar")
+    self.assertEqual(message.Extensions[test_bad_identifiers_pb2.reflection],
+                     "baz")
+    self.assertEqual(message.Extensions[test_bad_identifiers_pb2.service],
+                     "qux")
+
+  def testOneof(self):
+    desc = unittest_pb2.TestAllTypes.DESCRIPTOR
+    self.assertEqual(1, len(desc.oneofs))
+    self.assertEqual('oneof_field', desc.oneofs[0].name)
+    self.assertEqual(0, desc.oneofs[0].index)
+    self.assertIs(desc, desc.oneofs[0].containing_type)
+    self.assertIs(desc.oneofs[0], desc.oneofs_by_name['oneof_field'])
+    nested_names = set(['oneof_uint32', 'oneof_nested_message',
+                        'oneof_string', 'oneof_bytes'])
+    self.assertEqual(
+        nested_names,
+        set([field.name for field in desc.oneofs[0].fields]))
+    for field_name, field_desc in desc.fields_by_name.items():
+      if field_name in nested_names:
+        self.assertIs(desc.oneofs[0], field_desc.containing_oneof)
+      else:
+        self.assertIsNone(field_desc.containing_oneof)
+
+
+class SymbolDatabaseRegistrationTest(unittest.TestCase):
+  """Checks that messages, enums and files are correctly registered."""
+
+  def testGetSymbol(self):
+    self.assertEqual(
+        unittest_pb2.TestAllTypes, symbol_database.Default().GetSymbol(
+            'protobuf_unittest.TestAllTypes'))
+    self.assertEqual(
+        unittest_pb2.TestAllTypes.NestedMessage,
+        symbol_database.Default().GetSymbol(
+            'protobuf_unittest.TestAllTypes.NestedMessage'))
+    with self.assertRaises(KeyError):
+      symbol_database.Default().GetSymbol('protobuf_unittest.NestedMessage')
+    self.assertEqual(
+        unittest_pb2.TestAllTypes.OptionalGroup,
+        symbol_database.Default().GetSymbol(
+            'protobuf_unittest.TestAllTypes.OptionalGroup'))
+    self.assertEqual(
+        unittest_pb2.TestAllTypes.RepeatedGroup,
+        symbol_database.Default().GetSymbol(
+            'protobuf_unittest.TestAllTypes.RepeatedGroup'))
+
+  def testEnums(self):
+    self.assertEqual(
+        'protobuf_unittest.ForeignEnum',
+        symbol_database.Default().pool.FindEnumTypeByName(
+            'protobuf_unittest.ForeignEnum').full_name)
+    self.assertEqual(
+        'protobuf_unittest.TestAllTypes.NestedEnum',
+        symbol_database.Default().pool.FindEnumTypeByName(
+            'protobuf_unittest.TestAllTypes.NestedEnum').full_name)
+
+  def testFindFileByName(self):
+    self.assertEqual(
+        'google/protobuf/unittest.proto',
+        symbol_database.Default().pool.FindFileByName(
+            'google/protobuf/unittest.proto').name)
+
+if __name__ == '__main__':
+  unittest.main()
diff --git a/python/google/protobuf/internal/import_test_package/__init__.py b/python/google/protobuf/internal/import_test_package/__init__.py
new file mode 100644
index 0000000..5121dd0
--- /dev/null
+++ b/python/google/protobuf/internal/import_test_package/__init__.py
@@ -0,0 +1,33 @@
+# Protocol Buffers - Google's data interchange format
+# Copyright 2008 Google Inc.  All rights reserved.
+# https://developers.google.com/protocol-buffers/
+#
+# Redistribution and use in source and binary forms, with or without
+# modification, are permitted provided that the following conditions are
+# met:
+#
+#     * Redistributions of source code must retain the above copyright
+# notice, this list of conditions and the following disclaimer.
+#     * Redistributions in binary form must reproduce the above
+# copyright notice, this list of conditions and the following disclaimer
+# in the documentation and/or other materials provided with the
+# distribution.
+#     * Neither the name of Google Inc. nor the names of its
+# contributors may be used to endorse or promote products derived from
+# this software without specific prior written permission.
+#
+# THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+# "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
+# LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
+# A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
+# OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
+# SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
+# LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
+# DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
+# THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
+# (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+# OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+
+"""Sample module importing a nested proto from itself."""
+
+from google.protobuf.internal.import_test_package import outer_pb2 as myproto
diff --git a/python/google/protobuf/internal/import_test_package/inner.proto b/python/google/protobuf/internal/import_test_package/inner.proto
new file mode 100644
index 0000000..2887c12
--- /dev/null
+++ b/python/google/protobuf/internal/import_test_package/inner.proto
@@ -0,0 +1,37 @@
+// Protocol Buffers - Google's data interchange format
+// Copyright 2008 Google Inc.  All rights reserved.
+// https://developers.google.com/protocol-buffers/
+//
+// Redistribution and use in source and binary forms, with or without
+// modification, are permitted provided that the following conditions are
+// met:
+//
+//     * Redistributions of source code must retain the above copyright
+// notice, this list of conditions and the following disclaimer.
+//     * Redistributions in binary form must reproduce the above
+// copyright notice, this list of conditions and the following disclaimer
+// in the documentation and/or other materials provided with the
+// distribution.
+//     * Neither the name of Google Inc. nor the names of its
+// contributors may be used to endorse or promote products derived from
+// this software without specific prior written permission.
+//
+// THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+// "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
+// LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
+// A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
+// OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
+// SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
+// LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
+// DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
+// THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
+// (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+// OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+
+syntax = "proto2";
+
+package google.protobuf.python.internal.import_test_package;
+
+message Inner {
+  optional int32 value = 1 [default = 57];
+}
diff --git a/python/google/protobuf/internal/import_test_package/outer.proto b/python/google/protobuf/internal/import_test_package/outer.proto
new file mode 100644
index 0000000..a27fb5c
--- /dev/null
+++ b/python/google/protobuf/internal/import_test_package/outer.proto
@@ -0,0 +1,39 @@
+// Protocol Buffers - Google's data interchange format
+// Copyright 2008 Google Inc.  All rights reserved.
+// https://developers.google.com/protocol-buffers/
+//
+// Redistribution and use in source and binary forms, with or without
+// modification, are permitted provided that the following conditions are
+// met:
+//
+//     * Redistributions of source code must retain the above copyright
+// notice, this list of conditions and the following disclaimer.
+//     * Redistributions in binary form must reproduce the above
+// copyright notice, this list of conditions and the following disclaimer
+// in the documentation and/or other materials provided with the
+// distribution.
+//     * Neither the name of Google Inc. nor the names of its
+// contributors may be used to endorse or promote products derived from
+// this software without specific prior written permission.
+//
+// THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+// "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
+// LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
+// A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
+// OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
+// SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
+// LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
+// DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
+// THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
+// (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+// OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+
+syntax = "proto2";
+
+package google.protobuf.python.internal.import_test_package;
+
+import "google/protobuf/internal/import_test_package/inner.proto";
+
+message Outer {
+  optional Inner inner = 1;
+}
diff --git a/python/google/protobuf/internal/json_format_test.py b/python/google/protobuf/internal/json_format_test.py
new file mode 100644
index 0000000..49e96a4
--- /dev/null
+++ b/python/google/protobuf/internal/json_format_test.py
@@ -0,0 +1,768 @@
+#! /usr/bin/env python
+#
+# Protocol Buffers - Google's data interchange format
+# Copyright 2008 Google Inc.  All rights reserved.
+# https://developers.google.com/protocol-buffers/
+#
+# Redistribution and use in source and binary forms, with or without
+# modification, are permitted provided that the following conditions are
+# met:
+#
+#     * Redistributions of source code must retain the above copyright
+# notice, this list of conditions and the following disclaimer.
+#     * Redistributions in binary form must reproduce the above
+# copyright notice, this list of conditions and the following disclaimer
+# in the documentation and/or other materials provided with the
+# distribution.
+#     * Neither the name of Google Inc. nor the names of its
+# contributors may be used to endorse or promote products derived from
+# this software without specific prior written permission.
+#
+# THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+# "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
+# LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
+# A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
+# OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
+# SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
+# LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
+# DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
+# THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
+# (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+# OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+
+"""Test for google.protobuf.json_format."""
+
+__author__ = 'jieluo@google.com (Jie Luo)'
+
+import json
+import math
+import sys
+
+try:
+  import unittest2 as unittest
+except ImportError:
+  import unittest
+from google.protobuf import any_pb2
+from google.protobuf import duration_pb2
+from google.protobuf import field_mask_pb2
+from google.protobuf import struct_pb2
+from google.protobuf import timestamp_pb2
+from google.protobuf import wrappers_pb2
+from google.protobuf.internal import well_known_types
+from google.protobuf import json_format
+from google.protobuf.util import json_format_proto3_pb2
+
+
+class JsonFormatBase(unittest.TestCase):
+
+  def FillAllFields(self, message):
+    message.int32_value = 20
+    message.int64_value = -20
+    message.uint32_value = 3120987654
+    message.uint64_value = 12345678900
+    message.float_value = float('-inf')
+    message.double_value = 3.1415
+    message.bool_value = True
+    message.string_value = 'foo'
+    message.bytes_value = b'bar'
+    message.message_value.value = 10
+    message.enum_value = json_format_proto3_pb2.BAR
+    # Repeated
+    message.repeated_int32_value.append(0x7FFFFFFF)
+    message.repeated_int32_value.append(-2147483648)
+    message.repeated_int64_value.append(9007199254740992)
+    message.repeated_int64_value.append(-9007199254740992)
+    message.repeated_uint32_value.append(0xFFFFFFF)
+    message.repeated_uint32_value.append(0x7FFFFFF)
+    message.repeated_uint64_value.append(9007199254740992)
+    message.repeated_uint64_value.append(9007199254740991)
+    message.repeated_float_value.append(0)
+
+    message.repeated_double_value.append(1E-15)
+    message.repeated_double_value.append(float('inf'))
+    message.repeated_bool_value.append(True)
+    message.repeated_bool_value.append(False)
+    message.repeated_string_value.append('Few symbols!#$,;')
+    message.repeated_string_value.append('bar')
+    message.repeated_bytes_value.append(b'foo')
+    message.repeated_bytes_value.append(b'bar')
+    message.repeated_message_value.add().value = 10
+    message.repeated_message_value.add().value = 11
+    message.repeated_enum_value.append(json_format_proto3_pb2.FOO)
+    message.repeated_enum_value.append(json_format_proto3_pb2.BAR)
+    self.message = message
+
+  def CheckParseBack(self, message, parsed_message):
+    json_format.Parse(json_format.MessageToJson(message),
+                      parsed_message)
+    self.assertEqual(message, parsed_message)
+
+  def CheckError(self, text, error_message):
+    message = json_format_proto3_pb2.TestMessage()
+    self.assertRaisesRegexp(
+        json_format.ParseError,
+        error_message,
+        json_format.Parse, text, message)
+
+
+class JsonFormatTest(JsonFormatBase):
+
+  def testEmptyMessageToJson(self):
+    message = json_format_proto3_pb2.TestMessage()
+    self.assertEqual(json_format.MessageToJson(message),
+                     '{}')
+    parsed_message = json_format_proto3_pb2.TestMessage()
+    self.CheckParseBack(message, parsed_message)
+
+  def testPartialMessageToJson(self):
+    message = json_format_proto3_pb2.TestMessage(
+        string_value='test',
+        repeated_int32_value=[89, 4])
+    self.assertEqual(json.loads(json_format.MessageToJson(message)),
+                     json.loads('{"stringValue": "test", '
+                                '"repeatedInt32Value": [89, 4]}'))
+    parsed_message = json_format_proto3_pb2.TestMessage()
+    self.CheckParseBack(message, parsed_message)
+
+  def testAllFieldsToJson(self):
+    message = json_format_proto3_pb2.TestMessage()
+    text = ('{"int32Value": 20, '
+            '"int64Value": "-20", '
+            '"uint32Value": 3120987654,'
+            '"uint64Value": "12345678900",'
+            '"floatValue": "-Infinity",'
+            '"doubleValue": 3.1415,'
+            '"boolValue": true,'
+            '"stringValue": "foo",'
+            '"bytesValue": "YmFy",'
+            '"messageValue": {"value": 10},'
+            '"enumValue": "BAR",'
+            '"repeatedInt32Value": [2147483647, -2147483648],'
+            '"repeatedInt64Value": ["9007199254740992", "-9007199254740992"],'
+            '"repeatedUint32Value": [268435455, 134217727],'
+            '"repeatedUint64Value": ["9007199254740992", "9007199254740991"],'
+            '"repeatedFloatValue": [0],'
+            '"repeatedDoubleValue": [1e-15, "Infinity"],'
+            '"repeatedBoolValue": [true, false],'
+            '"repeatedStringValue": ["Few symbols!#$,;", "bar"],'
+            '"repeatedBytesValue": ["Zm9v", "YmFy"],'
+            '"repeatedMessageValue": [{"value": 10}, {"value": 11}],'
+            '"repeatedEnumValue": ["FOO", "BAR"]'
+            '}')
+    self.FillAllFields(message)
+    self.assertEqual(
+        json.loads(json_format.MessageToJson(message)),
+        json.loads(text))
+    parsed_message = json_format_proto3_pb2.TestMessage()
+    json_format.Parse(text, parsed_message)
+    self.assertEqual(message, parsed_message)
+
+  def testJsonEscapeString(self):
+    message = json_format_proto3_pb2.TestMessage()
+    if sys.version_info[0] < 3:
+      message.string_value = '&\n<\"\r>\b\t\f\\\001/\xe2\x80\xa8\xe2\x80\xa9'
+    else:
+      message.string_value = '&\n<\"\r>\b\t\f\\\001/'
+      message.string_value += (b'\xe2\x80\xa8\xe2\x80\xa9').decode('utf-8')
+    self.assertEqual(
+        json_format.MessageToJson(message),
+        '{\n  "stringValue": '
+        '"&\\n<\\\"\\r>\\b\\t\\f\\\\\\u0001/\\u2028\\u2029"\n}')
+    parsed_message = json_format_proto3_pb2.TestMessage()
+    self.CheckParseBack(message, parsed_message)
+    text = u'{"int32Value": "\u0031"}'
+    json_format.Parse(text, message)
+    self.assertEqual(message.int32_value, 1)
+
+  def testAlwaysSeriliaze(self):
+    message = json_format_proto3_pb2.TestMessage(
+        string_value='foo')
+    self.assertEqual(
+        json.loads(json_format.MessageToJson(message, True)),
+        json.loads('{'
+                   '"repeatedStringValue": [],'
+                   '"stringValue": "foo",'
+                   '"repeatedBoolValue": [],'
+                   '"repeatedUint32Value": [],'
+                   '"repeatedInt32Value": [],'
+                   '"enumValue": "FOO",'
+                   '"int32Value": 0,'
+                   '"floatValue": 0,'
+                   '"int64Value": "0",'
+                   '"uint32Value": 0,'
+                   '"repeatedBytesValue": [],'
+                   '"repeatedUint64Value": [],'
+                   '"repeatedDoubleValue": [],'
+                   '"bytesValue": "",'
+                   '"boolValue": false,'
+                   '"repeatedEnumValue": [],'
+                   '"uint64Value": "0",'
+                   '"doubleValue": 0,'
+                   '"repeatedFloatValue": [],'
+                   '"repeatedInt64Value": [],'
+                   '"repeatedMessageValue": []}'))
+    parsed_message = json_format_proto3_pb2.TestMessage()
+    self.CheckParseBack(message, parsed_message)
+
+  def testMapFields(self):
+    message = json_format_proto3_pb2.TestMap()
+    message.bool_map[True] = 1
+    message.bool_map[False] = 2
+    message.int32_map[1] = 2
+    message.int32_map[2] = 3
+    message.int64_map[1] = 2
+    message.int64_map[2] = 3
+    message.uint32_map[1] = 2
+    message.uint32_map[2] = 3
+    message.uint64_map[1] = 2
+    message.uint64_map[2] = 3
+    message.string_map['1'] = 2
+    message.string_map['null'] = 3
+    self.assertEqual(
+        json.loads(json_format.MessageToJson(message, True)),
+        json.loads('{'
+                   '"boolMap": {"false": 2, "true": 1},'
+                   '"int32Map": {"1": 2, "2": 3},'
+                   '"int64Map": {"1": 2, "2": 3},'
+                   '"uint32Map": {"1": 2, "2": 3},'
+                   '"uint64Map": {"1": 2, "2": 3},'
+                   '"stringMap": {"1": 2, "null": 3}'
+                   '}'))
+    parsed_message = json_format_proto3_pb2.TestMap()
+    self.CheckParseBack(message, parsed_message)
+
+  def testOneofFields(self):
+    message = json_format_proto3_pb2.TestOneof()
+    # Always print does not affect oneof fields.
+    self.assertEqual(
+        json_format.MessageToJson(message, True),
+        '{}')
+    message.oneof_int32_value = 0
+    self.assertEqual(
+        json_format.MessageToJson(message, True),
+        '{\n'
+        '  "oneofInt32Value": 0\n'
+        '}')
+    parsed_message = json_format_proto3_pb2.TestOneof()
+    self.CheckParseBack(message, parsed_message)
+
+  def testTimestampMessage(self):
+    message = json_format_proto3_pb2.TestTimestamp()
+    message.value.seconds = 0
+    message.value.nanos = 0
+    message.repeated_value.add().seconds = 20
+    message.repeated_value[0].nanos = 1
+    message.repeated_value.add().seconds = 0
+    message.repeated_value[1].nanos = 10000
+    message.repeated_value.add().seconds = 100000000
+    message.repeated_value[2].nanos = 0
+    # Maximum time
+    message.repeated_value.add().seconds = 253402300799
+    message.repeated_value[3].nanos = 999999999
+    # Minimum time
+    message.repeated_value.add().seconds = -62135596800
+    message.repeated_value[4].nanos = 0
+    self.assertEqual(
+        json.loads(json_format.MessageToJson(message, True)),
+        json.loads('{'
+                   '"value": "1970-01-01T00:00:00Z",'
+                   '"repeatedValue": ['
+                   '  "1970-01-01T00:00:20.000000001Z",'
+                   '  "1970-01-01T00:00:00.000010Z",'
+                   '  "1973-03-03T09:46:40Z",'
+                   '  "9999-12-31T23:59:59.999999999Z",'
+                   '  "0001-01-01T00:00:00Z"'
+                   ']'
+                   '}'))
+    parsed_message = json_format_proto3_pb2.TestTimestamp()
+    self.CheckParseBack(message, parsed_message)
+    text = (r'{"value": "1970-01-01T00:00:00.01+08:00",'
+            r'"repeatedValue":['
+            r'  "1970-01-01T00:00:00.01+08:30",'
+            r'  "1970-01-01T00:00:00.01-01:23"]}')
+    json_format.Parse(text, parsed_message)
+    self.assertEqual(parsed_message.value.seconds, -8 * 3600)
+    self.assertEqual(parsed_message.value.nanos, 10000000)
+    self.assertEqual(parsed_message.repeated_value[0].seconds, -8.5 * 3600)
+    self.assertEqual(parsed_message.repeated_value[1].seconds, 3600 + 23 * 60)
+
+  def testDurationMessage(self):
+    message = json_format_proto3_pb2.TestDuration()
+    message.value.seconds = 1
+    message.repeated_value.add().seconds = 0
+    message.repeated_value[0].nanos = 10
+    message.repeated_value.add().seconds = -1
+    message.repeated_value[1].nanos = -1000
+    message.repeated_value.add().seconds = 10
+    message.repeated_value[2].nanos = 11000000
+    message.repeated_value.add().seconds = -315576000000
+    message.repeated_value.add().seconds = 315576000000
+    self.assertEqual(
+        json.loads(json_format.MessageToJson(message, True)),
+        json.loads('{'
+                   '"value": "1s",'
+                   '"repeatedValue": ['
+                   '  "0.000000010s",'
+                   '  "-1.000001s",'
+                   '  "10.011s",'
+                   '  "-315576000000s",'
+                   '  "315576000000s"'
+                   ']'
+                   '}'))
+    parsed_message = json_format_proto3_pb2.TestDuration()
+    self.CheckParseBack(message, parsed_message)
+
+  def testFieldMaskMessage(self):
+    message = json_format_proto3_pb2.TestFieldMask()
+    message.value.paths.append('foo.bar')
+    message.value.paths.append('bar')
+    self.assertEqual(
+        json_format.MessageToJson(message, True),
+        '{\n'
+        '  "value": "foo.bar,bar"\n'
+        '}')
+    parsed_message = json_format_proto3_pb2.TestFieldMask()
+    self.CheckParseBack(message, parsed_message)
+
+  def testWrapperMessage(self):
+    message = json_format_proto3_pb2.TestWrapper()
+    message.bool_value.value = False
+    message.int32_value.value = 0
+    message.string_value.value = ''
+    message.bytes_value.value = b''
+    message.repeated_bool_value.add().value = True
+    message.repeated_bool_value.add().value = False
+    message.repeated_int32_value.add()
+    self.assertEqual(
+        json.loads(json_format.MessageToJson(message, True)),
+        json.loads('{\n'
+                   '  "int32Value": 0,'
+                   '  "boolValue": false,'
+                   '  "stringValue": "",'
+                   '  "bytesValue": "",'
+                   '  "repeatedBoolValue": [true, false],'
+                   '  "repeatedInt32Value": [0],'
+                   '  "repeatedUint32Value": [],'
+                   '  "repeatedFloatValue": [],'
+                   '  "repeatedDoubleValue": [],'
+                   '  "repeatedBytesValue": [],'
+                   '  "repeatedInt64Value": [],'
+                   '  "repeatedUint64Value": [],'
+                   '  "repeatedStringValue": []'
+                   '}'))
+    parsed_message = json_format_proto3_pb2.TestWrapper()
+    self.CheckParseBack(message, parsed_message)
+
+  def testStructMessage(self):
+    message = json_format_proto3_pb2.TestStruct()
+    message.value['name'] = 'Jim'
+    message.value['age'] = 10
+    message.value['attend'] = True
+    message.value['email'] = None
+    message.value.get_or_create_struct('address')['city'] = 'SFO'
+    message.value['address']['house_number'] = 1024
+    struct_list = message.value.get_or_create_list('list')
+    struct_list.extend([6, 'seven', True, False, None])
+    struct_list.add_struct()['subkey2'] = 9
+    message.repeated_value.add()['age'] = 11
+    message.repeated_value.add()
+    self.assertEqual(
+        json.loads(json_format.MessageToJson(message, False)),
+        json.loads(
+            '{'
+            '  "value": {'
+            '    "address": {'
+            '      "city": "SFO", '
+            '      "house_number": 1024'
+            '    }, '
+            '    "age": 10, '
+            '    "name": "Jim", '
+            '    "attend": true, '
+            '    "email": null, '
+            '    "list": [6, "seven", true, false, null, {"subkey2": 9}]'
+            '  },'
+            '  "repeatedValue": [{"age": 11}, {}]'
+            '}'))
+    parsed_message = json_format_proto3_pb2.TestStruct()
+    self.CheckParseBack(message, parsed_message)
+
+  def testValueMessage(self):
+    message = json_format_proto3_pb2.TestValue()
+    message.value.string_value = 'hello'
+    message.repeated_value.add().number_value = 11.1
+    message.repeated_value.add().bool_value = False
+    message.repeated_value.add().null_value = 0
+    self.assertEqual(
+        json.loads(json_format.MessageToJson(message, False)),
+        json.loads(
+            '{'
+            '  "value": "hello",'
+            '  "repeatedValue": [11.1, false, null]'
+            '}'))
+    parsed_message = json_format_proto3_pb2.TestValue()
+    self.CheckParseBack(message, parsed_message)
+    # Can't parse back if the Value message is not set.
+    message.repeated_value.add()
+    self.assertEqual(
+        json.loads(json_format.MessageToJson(message, False)),
+        json.loads(
+            '{'
+            '  "value": "hello",'
+            '  "repeatedValue": [11.1, false, null, null]'
+            '}'))
+
+  def testListValueMessage(self):
+    message = json_format_proto3_pb2.TestListValue()
+    message.value.values.add().number_value = 11.1
+    message.value.values.add().null_value = 0
+    message.value.values.add().bool_value = True
+    message.value.values.add().string_value = 'hello'
+    message.value.values.add().struct_value['name'] = 'Jim'
+    message.repeated_value.add().values.add().number_value = 1
+    message.repeated_value.add()
+    self.assertEqual(
+        json.loads(json_format.MessageToJson(message, False)),
+        json.loads(
+            '{"value": [11.1, null, true, "hello", {"name": "Jim"}]\n,'
+            '"repeatedValue": [[1], []]}'))
+    parsed_message = json_format_proto3_pb2.TestListValue()
+    self.CheckParseBack(message, parsed_message)
+
+  def testAnyMessage(self):
+    message = json_format_proto3_pb2.TestAny()
+    value1 = json_format_proto3_pb2.MessageType()
+    value2 = json_format_proto3_pb2.MessageType()
+    value1.value = 1234
+    value2.value = 5678
+    message.value.Pack(value1)
+    message.repeated_value.add().Pack(value1)
+    message.repeated_value.add().Pack(value2)
+    message.repeated_value.add()
+    self.assertEqual(
+        json.loads(json_format.MessageToJson(message, True)),
+        json.loads(
+            '{\n'
+            '  "repeatedValue": [ {\n'
+            '    "@type": "type.googleapis.com/proto3.MessageType",\n'
+            '    "value": 1234\n'
+            '  }, {\n'
+            '    "@type": "type.googleapis.com/proto3.MessageType",\n'
+            '    "value": 5678\n'
+            '  },\n'
+            '  {}],\n'
+            '  "value": {\n'
+            '    "@type": "type.googleapis.com/proto3.MessageType",\n'
+            '    "value": 1234\n'
+            '  }\n'
+            '}\n'))
+    parsed_message = json_format_proto3_pb2.TestAny()
+    self.CheckParseBack(message, parsed_message)
+
+  def testWellKnownInAnyMessage(self):
+    message = any_pb2.Any()
+    int32_value = wrappers_pb2.Int32Value()
+    int32_value.value = 1234
+    message.Pack(int32_value)
+    self.assertEqual(
+        json.loads(json_format.MessageToJson(message, True)),
+        json.loads(
+            '{\n'
+            '  "@type": \"type.googleapis.com/google.protobuf.Int32Value\",\n'
+            '  "value": 1234\n'
+            '}\n'))
+    parsed_message = any_pb2.Any()
+    self.CheckParseBack(message, parsed_message)
+
+    timestamp = timestamp_pb2.Timestamp()
+    message.Pack(timestamp)
+    self.assertEqual(
+        json.loads(json_format.MessageToJson(message, True)),
+        json.loads(
+            '{\n'
+            '  "@type": "type.googleapis.com/google.protobuf.Timestamp",\n'
+            '  "value": "1970-01-01T00:00:00Z"\n'
+            '}\n'))
+    self.CheckParseBack(message, parsed_message)
+
+    duration = duration_pb2.Duration()
+    duration.seconds = 1
+    message.Pack(duration)
+    self.assertEqual(
+        json.loads(json_format.MessageToJson(message, True)),
+        json.loads(
+            '{\n'
+            '  "@type": "type.googleapis.com/google.protobuf.Duration",\n'
+            '  "value": "1s"\n'
+            '}\n'))
+    self.CheckParseBack(message, parsed_message)
+
+    field_mask = field_mask_pb2.FieldMask()
+    field_mask.paths.append('foo.bar')
+    field_mask.paths.append('bar')
+    message.Pack(field_mask)
+    self.assertEqual(
+        json.loads(json_format.MessageToJson(message, True)),
+        json.loads(
+            '{\n'
+            '  "@type": "type.googleapis.com/google.protobuf.FieldMask",\n'
+            '  "value": "foo.bar,bar"\n'
+            '}\n'))
+    self.CheckParseBack(message, parsed_message)
+
+    struct_message = struct_pb2.Struct()
+    struct_message['name'] = 'Jim'
+    message.Pack(struct_message)
+    self.assertEqual(
+        json.loads(json_format.MessageToJson(message, True)),
+        json.loads(
+            '{\n'
+            '  "@type": "type.googleapis.com/google.protobuf.Struct",\n'
+            '  "value": {"name": "Jim"}\n'
+            '}\n'))
+    self.CheckParseBack(message, parsed_message)
+
+    nested_any = any_pb2.Any()
+    int32_value.value = 5678
+    nested_any.Pack(int32_value)
+    message.Pack(nested_any)
+    self.assertEqual(
+        json.loads(json_format.MessageToJson(message, True)),
+        json.loads(
+            '{\n'
+            '  "@type": "type.googleapis.com/google.protobuf.Any",\n'
+            '  "value": {\n'
+            '    "@type": "type.googleapis.com/google.protobuf.Int32Value",\n'
+            '    "value": 5678\n'
+            '  }\n'
+            '}\n'))
+    self.CheckParseBack(message, parsed_message)
+
+  def testParseNull(self):
+    message = json_format_proto3_pb2.TestMessage()
+    parsed_message = json_format_proto3_pb2.TestMessage()
+    self.FillAllFields(parsed_message)
+    json_format.Parse('{"int32Value": null, '
+                      '"int64Value": null, '
+                      '"uint32Value": null,'
+                      '"uint64Value": null,'
+                      '"floatValue": null,'
+                      '"doubleValue": null,'
+                      '"boolValue": null,'
+                      '"stringValue": null,'
+                      '"bytesValue": null,'
+                      '"messageValue": null,'
+                      '"enumValue": null,'
+                      '"repeatedInt32Value": null,'
+                      '"repeatedInt64Value": null,'
+                      '"repeatedUint32Value": null,'
+                      '"repeatedUint64Value": null,'
+                      '"repeatedFloatValue": null,'
+                      '"repeatedDoubleValue": null,'
+                      '"repeatedBoolValue": null,'
+                      '"repeatedStringValue": null,'
+                      '"repeatedBytesValue": null,'
+                      '"repeatedMessageValue": null,'
+                      '"repeatedEnumValue": null'
+                      '}',
+                      parsed_message)
+    self.assertEqual(message, parsed_message)
+    self.assertRaisesRegexp(
+        json_format.ParseError,
+        'Failed to parse repeatedInt32Value field: '
+        'null is not allowed to be used as an element in a repeated field.',
+        json_format.Parse,
+        '{"repeatedInt32Value":[1, null]}',
+        parsed_message)
+
+  def testNanFloat(self):
+    message = json_format_proto3_pb2.TestMessage()
+    message.float_value = float('nan')
+    text = '{\n  "floatValue": "NaN"\n}'
+    self.assertEqual(json_format.MessageToJson(message), text)
+    parsed_message = json_format_proto3_pb2.TestMessage()
+    json_format.Parse(text, parsed_message)
+    self.assertTrue(math.isnan(parsed_message.float_value))
+
+  def testParseEmptyText(self):
+    self.CheckError('',
+                    r'Failed to load JSON: (Expecting value)|(No JSON).')
+
+  def testParseBadEnumValue(self):
+    self.CheckError(
+        '{"enumValue": 1}',
+        'Enum value must be a string literal with double quotes. '
+        'Type "proto3.EnumType" has no value named 1.')
+    self.CheckError(
+        '{"enumValue": "baz"}',
+        'Enum value must be a string literal with double quotes. '
+        'Type "proto3.EnumType" has no value named baz.')
+
+  def testParseBadIdentifer(self):
+    self.CheckError('{int32Value: 1}',
+                    (r'Failed to load JSON: Expecting property name'
+                     r'( enclosed in double quotes)?: line 1'))
+    self.CheckError('{"unknownName": 1}',
+                    'Message type "proto3.TestMessage" has no field named '
+                    '"unknownName".')
+
+  def testDuplicateField(self):
+    # Duplicate key check is not supported for python2.6
+    if sys.version_info < (2, 7):
+      return
+    self.CheckError('{"int32Value": 1,\n"int32Value":2}',
+                    'Failed to load JSON: duplicate key int32Value.')
+
+  def testInvalidBoolValue(self):
+    self.CheckError('{"boolValue": 1}',
+                    'Failed to parse boolValue field: '
+                    'Expected true or false without quotes.')
+    self.CheckError('{"boolValue": "true"}',
+                    'Failed to parse boolValue field: '
+                    'Expected true or false without quotes.')
+
+  def testInvalidIntegerValue(self):
+    message = json_format_proto3_pb2.TestMessage()
+    text = '{"int32Value": 0x12345}'
+    self.assertRaises(json_format.ParseError,
+                      json_format.Parse, text, message)
+    self.CheckError('{"int32Value": 012345}',
+                    (r'Failed to load JSON: Expecting \'?,\'? delimiter: '
+                     r'line 1.'))
+    self.CheckError('{"int32Value": 1.0}',
+                    'Failed to parse int32Value field: '
+                    'Couldn\'t parse integer: 1.0.')
+    self.CheckError('{"int32Value": " 1 "}',
+                    'Failed to parse int32Value field: '
+                    'Couldn\'t parse integer: " 1 ".')
+    self.CheckError('{"int32Value": "1 "}',
+                    'Failed to parse int32Value field: '
+                    'Couldn\'t parse integer: "1 ".')
+    self.CheckError('{"int32Value": 12345678901234567890}',
+                    'Failed to parse int32Value field: Value out of range: '
+                    '12345678901234567890.')
+    self.CheckError('{"int32Value": 1e5}',
+                    'Failed to parse int32Value field: '
+                    'Couldn\'t parse integer: 100000.0.')
+    self.CheckError('{"uint32Value": -1}',
+                    'Failed to parse uint32Value field: '
+                    'Value out of range: -1.')
+
+  def testInvalidFloatValue(self):
+    self.CheckError('{"floatValue": "nan"}',
+                    'Failed to parse floatValue field: Couldn\'t '
+                    'parse float "nan", use "NaN" instead.')
+
+  def testInvalidBytesValue(self):
+    self.CheckError('{"bytesValue": "AQI"}',
+                    'Failed to parse bytesValue field: Incorrect padding.')
+    self.CheckError('{"bytesValue": "AQI*"}',
+                    'Failed to parse bytesValue field: Incorrect padding.')
+
+  def testInvalidMap(self):
+    message = json_format_proto3_pb2.TestMap()
+    text = '{"int32Map": {"null": 2, "2": 3}}'
+    self.assertRaisesRegexp(
+        json_format.ParseError,
+        'Failed to parse int32Map field: invalid literal',
+        json_format.Parse, text, message)
+    text = '{"int32Map": {1: 2, "2": 3}}'
+    self.assertRaisesRegexp(
+        json_format.ParseError,
+        (r'Failed to load JSON: Expecting property name'
+         r'( enclosed in double quotes)?: line 1'),
+        json_format.Parse, text, message)
+    text = '{"boolMap": {"null": 1}}'
+    self.assertRaisesRegexp(
+        json_format.ParseError,
+        'Failed to parse boolMap field: Expected "true" or "false", not null.',
+        json_format.Parse, text, message)
+    if sys.version_info < (2, 7):
+      return
+    text = r'{"stringMap": {"a": 3, "\u0061": 2}}'
+    self.assertRaisesRegexp(
+        json_format.ParseError,
+        'Failed to load JSON: duplicate key a',
+        json_format.Parse, text, message)
+
+  def testInvalidTimestamp(self):
+    message = json_format_proto3_pb2.TestTimestamp()
+    text = '{"value": "10000-01-01T00:00:00.00Z"}'
+    self.assertRaisesRegexp(
+        json_format.ParseError,
+        'time data \'10000-01-01T00:00:00\' does not match'
+        ' format \'%Y-%m-%dT%H:%M:%S\'.',
+        json_format.Parse, text, message)
+    text = '{"value": "1970-01-01T00:00:00.0123456789012Z"}'
+    self.assertRaisesRegexp(
+        well_known_types.ParseError,
+        'nanos 0123456789012 more than 9 fractional digits.',
+        json_format.Parse, text, message)
+    text = '{"value": "1972-01-01T01:00:00.01+08"}'
+    self.assertRaisesRegexp(
+        well_known_types.ParseError,
+        (r'Invalid timezone offset value: \+08.'),
+        json_format.Parse, text, message)
+    # Time smaller than minimum time.
+    text = '{"value": "0000-01-01T00:00:00Z"}'
+    self.assertRaisesRegexp(
+        json_format.ParseError,
+        'Failed to parse value field: year is out of range.',
+        json_format.Parse, text, message)
+    # Time bigger than maxinum time.
+    message.value.seconds = 253402300800
+    self.assertRaisesRegexp(
+        OverflowError,
+        'date value out of range',
+        json_format.MessageToJson, message)
+
+  def testInvalidOneof(self):
+    message = json_format_proto3_pb2.TestOneof()
+    text = '{"oneofInt32Value": 1, "oneofStringValue": "2"}'
+    self.assertRaisesRegexp(
+        json_format.ParseError,
+        'Message type "proto3.TestOneof"'
+        ' should not have multiple "oneof_value" oneof fields.',
+        json_format.Parse, text, message)
+
+  def testInvalidListValue(self):
+    message = json_format_proto3_pb2.TestListValue()
+    text = '{"value": 1234}'
+    self.assertRaisesRegexp(
+        json_format.ParseError,
+        r'Failed to parse value field: ListValue must be in \[\] which is 1234',
+        json_format.Parse, text, message)
+
+  def testInvalidStruct(self):
+    message = json_format_proto3_pb2.TestStruct()
+    text = '{"value": 1234}'
+    self.assertRaisesRegexp(
+        json_format.ParseError,
+        'Failed to parse value field: Struct must be in a dict which is 1234',
+        json_format.Parse, text, message)
+
+  def testInvalidAny(self):
+    message = any_pb2.Any()
+    text = '{"@type": "type.googleapis.com/google.protobuf.Int32Value"}'
+    self.assertRaisesRegexp(
+        KeyError,
+        'value',
+        json_format.Parse, text, message)
+    text = '{"value": 1234}'
+    self.assertRaisesRegexp(
+        json_format.ParseError,
+        '@type is missing when parsing any message.',
+        json_format.Parse, text, message)
+    text = '{"@type": "type.googleapis.com/MessageNotExist", "value": 1234}'
+    self.assertRaisesRegexp(
+        TypeError,
+        'Can not find message descriptor by type_url: '
+        'type.googleapis.com/MessageNotExist.',
+        json_format.Parse, text, message)
+    # Only last part is to be used.
+    text = (r'{"@type": "incorrect.googleapis.com/google.protobuf.Int32Value",'
+            r'"value": 1234}')
+    json_format.Parse(text, message)
+
+
+if __name__ == '__main__':
+  unittest.main()
diff --git a/python/google/protobuf/internal/message_factory_test.py b/python/google/protobuf/internal/message_factory_test.py
new file mode 100644
index 0000000..2fbe5ea
--- /dev/null
+++ b/python/google/protobuf/internal/message_factory_test.py
@@ -0,0 +1,135 @@
+#! /usr/bin/env python
+#
+# Protocol Buffers - Google's data interchange format
+# Copyright 2008 Google Inc.  All rights reserved.
+# https://developers.google.com/protocol-buffers/
+#
+# Redistribution and use in source and binary forms, with or without
+# modification, are permitted provided that the following conditions are
+# met:
+#
+#     * Redistributions of source code must retain the above copyright
+# notice, this list of conditions and the following disclaimer.
+#     * Redistributions in binary form must reproduce the above
+# copyright notice, this list of conditions and the following disclaimer
+# in the documentation and/or other materials provided with the
+# distribution.
+#     * Neither the name of Google Inc. nor the names of its
+# contributors may be used to endorse or promote products derived from
+# this software without specific prior written permission.
+#
+# THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+# "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
+# LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
+# A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
+# OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
+# SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
+# LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
+# DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
+# THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
+# (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+# OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+
+"""Tests for google.protobuf.message_factory."""
+
+__author__ = 'matthewtoia@google.com (Matt Toia)'
+
+try:
+  import unittest2 as unittest
+except ImportError:
+  import unittest
+from google.protobuf import descriptor_pb2
+from google.protobuf.internal import factory_test1_pb2
+from google.protobuf.internal import factory_test2_pb2
+from google.protobuf import descriptor_database
+from google.protobuf import descriptor_pool
+from google.protobuf import message_factory
+
+
+class MessageFactoryTest(unittest.TestCase):
+
+  def setUp(self):
+    self.factory_test1_fd = descriptor_pb2.FileDescriptorProto.FromString(
+        factory_test1_pb2.DESCRIPTOR.serialized_pb)
+    self.factory_test2_fd = descriptor_pb2.FileDescriptorProto.FromString(
+        factory_test2_pb2.DESCRIPTOR.serialized_pb)
+
+  def _ExerciseDynamicClass(self, cls):
+    msg = cls()
+    msg.mandatory = 42
+    msg.nested_factory_2_enum = 0
+    msg.nested_factory_2_message.value = 'nested message value'
+    msg.factory_1_message.factory_1_enum = 1
+    msg.factory_1_message.nested_factory_1_enum = 0
+    msg.factory_1_message.nested_factory_1_message.value = (
+        'nested message value')
+    msg.factory_1_message.scalar_value = 22
+    msg.factory_1_message.list_value.extend([u'one', u'two', u'three'])
+    msg.factory_1_message.list_value.append(u'four')
+    msg.factory_1_enum = 1
+    msg.nested_factory_1_enum = 0
+    msg.nested_factory_1_message.value = 'nested message value'
+    msg.circular_message.mandatory = 1
+    msg.circular_message.circular_message.mandatory = 2
+    msg.circular_message.scalar_value = 'one deep'
+    msg.scalar_value = 'zero deep'
+    msg.list_value.extend([u'four', u'three', u'two'])
+    msg.list_value.append(u'one')
+    msg.grouped.add()
+    msg.grouped[0].part_1 = 'hello'
+    msg.grouped[0].part_2 = 'world'
+    msg.grouped.add(part_1='testing', part_2='123')
+    msg.loop.loop.mandatory = 2
+    msg.loop.loop.loop.loop.mandatory = 4
+    serialized = msg.SerializeToString()
+    converted = factory_test2_pb2.Factory2Message.FromString(serialized)
+    reserialized = converted.SerializeToString()
+    self.assertEqual(serialized, reserialized)
+    result = cls.FromString(reserialized)
+    self.assertEqual(msg, result)
+
+  def testGetPrototype(self):
+    db = descriptor_database.DescriptorDatabase()
+    pool = descriptor_pool.DescriptorPool(db)
+    db.Add(self.factory_test1_fd)
+    db.Add(self.factory_test2_fd)
+    factory = message_factory.MessageFactory()
+    cls = factory.GetPrototype(pool.FindMessageTypeByName(
+        'google.protobuf.python.internal.Factory2Message'))
+    self.assertFalse(cls is factory_test2_pb2.Factory2Message)
+    self._ExerciseDynamicClass(cls)
+    cls2 = factory.GetPrototype(pool.FindMessageTypeByName(
+        'google.protobuf.python.internal.Factory2Message'))
+    self.assertTrue(cls is cls2)
+
+  def testGetMessages(self):
+    # performed twice because multiple calls with the same input must be allowed
+    for _ in range(2):
+      messages = message_factory.GetMessages([self.factory_test1_fd,
+                                              self.factory_test2_fd])
+      self.assertTrue(
+          set(['google.protobuf.python.internal.Factory2Message',
+               'google.protobuf.python.internal.Factory1Message'],
+             ).issubset(set(messages.keys())))
+      self._ExerciseDynamicClass(
+          messages['google.protobuf.python.internal.Factory2Message'])
+      self.assertTrue(
+          set(['google.protobuf.python.internal.Factory2Message.one_more_field',
+               'google.protobuf.python.internal.another_field'],
+             ).issubset(
+                 set(messages['google.protobuf.python.internal.Factory1Message']
+                     ._extensions_by_name.keys())))
+      factory_msg1 = messages['google.protobuf.python.internal.Factory1Message']
+      msg1 = messages['google.protobuf.python.internal.Factory1Message']()
+      ext1 = factory_msg1._extensions_by_name[
+          'google.protobuf.python.internal.Factory2Message.one_more_field']
+      ext2 = factory_msg1._extensions_by_name[
+          'google.protobuf.python.internal.another_field']
+      msg1.Extensions[ext1] = 'test1'
+      msg1.Extensions[ext2] = 'test2'
+      self.assertEqual('test1', msg1.Extensions[ext1])
+      self.assertEqual('test2', msg1.Extensions[ext2])
+
+
+if __name__ == '__main__':
+  unittest.main()
diff --git a/python/google/protobuf/internal/message_listener.py b/python/google/protobuf/internal/message_listener.py
new file mode 100755
index 0000000..0fc255a
--- /dev/null
+++ b/python/google/protobuf/internal/message_listener.py
@@ -0,0 +1,78 @@
+# Protocol Buffers - Google's data interchange format
+# Copyright 2008 Google Inc.  All rights reserved.
+# https://developers.google.com/protocol-buffers/
+#
+# Redistribution and use in source and binary forms, with or without
+# modification, are permitted provided that the following conditions are
+# met:
+#
+#     * Redistributions of source code must retain the above copyright
+# notice, this list of conditions and the following disclaimer.
+#     * Redistributions in binary form must reproduce the above
+# copyright notice, this list of conditions and the following disclaimer
+# in the documentation and/or other materials provided with the
+# distribution.
+#     * Neither the name of Google Inc. nor the names of its
+# contributors may be used to endorse or promote products derived from
+# this software without specific prior written permission.
+#
+# THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+# "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
+# LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
+# A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
+# OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
+# SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
+# LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
+# DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
+# THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
+# (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+# OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+
+"""Defines a listener interface for observing certain
+state transitions on Message objects.
+
+Also defines a null implementation of this interface.
+"""
+
+__author__ = 'robinson@google.com (Will Robinson)'
+
+
+class MessageListener(object):
+
+  """Listens for modifications made to a message.  Meant to be registered via
+  Message._SetListener().
+
+  Attributes:
+    dirty:  If True, then calling Modified() would be a no-op.  This can be
+            used to avoid these calls entirely in the common case.
+  """
+
+  def Modified(self):
+    """Called every time the message is modified in such a way that the parent
+    message may need to be updated.  This currently means either:
+    (a) The message was modified for the first time, so the parent message
+        should henceforth mark the message as present.
+    (b) The message's cached byte size became dirty -- i.e. the message was
+        modified for the first time after a previous call to ByteSize().
+        Therefore the parent should also mark its byte size as dirty.
+    Note that (a) implies (b), since new objects start out with a client cached
+    size (zero).  However, we document (a) explicitly because it is important.
+
+    Modified() will *only* be called in response to one of these two events --
+    not every time the sub-message is modified.
+
+    Note that if the listener's |dirty| attribute is true, then calling
+    Modified at the moment would be a no-op, so it can be skipped.  Performance-
+    sensitive callers should check this attribute directly before calling since
+    it will be true most of the time.
+    """
+
+    raise NotImplementedError
+
+
+class NullMessageListener(object):
+
+  """No-op MessageListener implementation."""
+
+  def Modified(self):
+    pass
diff --git a/python/google/protobuf/internal/message_set_extensions.proto b/python/google/protobuf/internal/message_set_extensions.proto
new file mode 100644
index 0000000..14e5f19
--- /dev/null
+++ b/python/google/protobuf/internal/message_set_extensions.proto
@@ -0,0 +1,74 @@
+// Protocol Buffers - Google's data interchange format
+// Copyright 2008 Google Inc.  All rights reserved.
+// https://developers.google.com/protocol-buffers/
+//
+// Redistribution and use in source and binary forms, with or without
+// modification, are permitted provided that the following conditions are
+// met:
+//
+//     * Redistributions of source code must retain the above copyright
+// notice, this list of conditions and the following disclaimer.
+//     * Redistributions in binary form must reproduce the above
+// copyright notice, this list of conditions and the following disclaimer
+// in the documentation and/or other materials provided with the
+// distribution.
+//     * Neither the name of Google Inc. nor the names of its
+// contributors may be used to endorse or promote products derived from
+// this software without specific prior written permission.
+//
+// THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+// "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
+// LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
+// A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
+// OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
+// SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
+// LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
+// DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
+// THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
+// (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+// OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+
+// This file contains messages that extend MessageSet.
+
+syntax = "proto2";
+package google.protobuf.internal;
+
+
+// A message with message_set_wire_format.
+message TestMessageSet {
+  option message_set_wire_format = true;
+  extensions 4 to max;
+}
+
+message TestMessageSetExtension1 {
+  extend TestMessageSet {
+    optional TestMessageSetExtension1 message_set_extension = 98418603;
+  }
+  optional int32 i = 15;
+}
+
+message TestMessageSetExtension2 {
+  extend TestMessageSet {
+    optional TestMessageSetExtension2 message_set_extension = 98418634;
+  }
+  optional string str = 25;
+}
+
+message TestMessageSetExtension3 {
+  optional string text = 35;
+}
+
+extend TestMessageSet {
+  optional TestMessageSetExtension3 message_set_extension3 = 98418655;
+}
+
+// This message was used to generate
+// //net/proto2/python/internal/testdata/message_set_message, but is commented
+// out since it must not actually exist in code, to simulate an "unknown"
+// extension.
+// message TestMessageSetUnknownExtension {
+//   extend TestMessageSet {
+//     optional TestMessageSetUnknownExtension message_set_extension = 56141421;
+//   }
+//   optional int64 a = 1;
+// }
diff --git a/python/google/protobuf/internal/message_test.py b/python/google/protobuf/internal/message_test.py
new file mode 100755
index 0000000..d03f2d2
--- /dev/null
+++ b/python/google/protobuf/internal/message_test.py
@@ -0,0 +1,1780 @@
+#! /usr/bin/env python
+#
+# Protocol Buffers - Google's data interchange format
+# Copyright 2008 Google Inc.  All rights reserved.
+# https://developers.google.com/protocol-buffers/
+#
+# Redistribution and use in source and binary forms, with or without
+# modification, are permitted provided that the following conditions are
+# met:
+#
+#     * Redistributions of source code must retain the above copyright
+# notice, this list of conditions and the following disclaimer.
+#     * Redistributions in binary form must reproduce the above
+# copyright notice, this list of conditions and the following disclaimer
+# in the documentation and/or other materials provided with the
+# distribution.
+#     * Neither the name of Google Inc. nor the names of its
+# contributors may be used to endorse or promote products derived from
+# this software without specific prior written permission.
+#
+# THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+# "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
+# LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
+# A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
+# OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
+# SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
+# LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
+# DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
+# THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
+# (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+# OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+
+"""Tests python protocol buffers against the golden message.
+
+Note that the golden messages exercise every known field type, thus this
+test ends up exercising and verifying nearly all of the parsing and
+serialization code in the whole library.
+
+TODO(kenton):  Merge with wire_format_test?  It doesn't make a whole lot of
+sense to call this a test of the "message" module, which only declares an
+abstract interface.
+"""
+
+__author__ = 'gps@google.com (Gregory P. Smith)'
+
+
+import collections
+import copy
+import math
+import operator
+import pickle
+import six
+import sys
+
+try:
+  import unittest2 as unittest
+except ImportError:
+  import unittest
+from google.protobuf.internal import _parameterized
+from google.protobuf import map_unittest_pb2
+from google.protobuf import unittest_pb2
+from google.protobuf import unittest_proto3_arena_pb2
+from google.protobuf.internal import any_test_pb2
+from google.protobuf.internal import api_implementation
+from google.protobuf.internal import packed_field_test_pb2
+from google.protobuf.internal import test_util
+from google.protobuf import message
+
+if six.PY3:
+  long = int
+
+# Python pre-2.6 does not have isinf() or isnan() functions, so we have
+# to provide our own.
+def isnan(val):
+  # NaN is never equal to itself.
+  return val != val
+def isinf(val):
+  # Infinity times zero equals NaN.
+  return not isnan(val) and isnan(val * 0)
+def IsPosInf(val):
+  return isinf(val) and (val > 0)
+def IsNegInf(val):
+  return isinf(val) and (val < 0)
+
+
+@_parameterized.Parameters(
+    (unittest_pb2),
+    (unittest_proto3_arena_pb2))
+class MessageTest(unittest.TestCase):
+
+  def testBadUtf8String(self, message_module):
+    if api_implementation.Type() != 'python':
+      self.skipTest("Skipping testBadUtf8String, currently only the python "
+                    "api implementation raises UnicodeDecodeError when a "
+                    "string field contains bad utf-8.")
+    bad_utf8_data = test_util.GoldenFileData('bad_utf8_string')
+    with self.assertRaises(UnicodeDecodeError) as context:
+      message_module.TestAllTypes.FromString(bad_utf8_data)
+    self.assertIn('TestAllTypes.optional_string', str(context.exception))
+
+  def testGoldenMessage(self, message_module):
+    # Proto3 doesn't have the "default_foo" members or foreign enums,
+    # and doesn't preserve unknown fields, so for proto3 we use a golden
+    # message that doesn't have these fields set.
+    if message_module is unittest_pb2:
+      golden_data = test_util.GoldenFileData(
+          'golden_message_oneof_implemented')
+    else:
+      golden_data = test_util.GoldenFileData('golden_message_proto3')
+
+    golden_message = message_module.TestAllTypes()
+    golden_message.ParseFromString(golden_data)
+    if message_module is unittest_pb2:
+      test_util.ExpectAllFieldsSet(self, golden_message)
+    self.assertEqual(golden_data, golden_message.SerializeToString())
+    golden_copy = copy.deepcopy(golden_message)
+    self.assertEqual(golden_data, golden_copy.SerializeToString())
+
+  def testGoldenPackedMessage(self, message_module):
+    golden_data = test_util.GoldenFileData('golden_packed_fields_message')
+    golden_message = message_module.TestPackedTypes()
+    golden_message.ParseFromString(golden_data)
+    all_set = message_module.TestPackedTypes()
+    test_util.SetAllPackedFields(all_set)
+    self.assertEqual(all_set, golden_message)
+    self.assertEqual(golden_data, all_set.SerializeToString())
+    golden_copy = copy.deepcopy(golden_message)
+    self.assertEqual(golden_data, golden_copy.SerializeToString())
+
+  def testPickleSupport(self, message_module):
+    golden_data = test_util.GoldenFileData('golden_message')
+    golden_message = message_module.TestAllTypes()
+    golden_message.ParseFromString(golden_data)
+    pickled_message = pickle.dumps(golden_message)
+
+    unpickled_message = pickle.loads(pickled_message)
+    self.assertEqual(unpickled_message, golden_message)
+
+  def testPositiveInfinity(self, message_module):
+    if message_module is unittest_pb2:
+      golden_data = (b'\x5D\x00\x00\x80\x7F'
+                     b'\x61\x00\x00\x00\x00\x00\x00\xF0\x7F'
+                     b'\xCD\x02\x00\x00\x80\x7F'
+                     b'\xD1\x02\x00\x00\x00\x00\x00\x00\xF0\x7F')
+    else:
+      golden_data = (b'\x5D\x00\x00\x80\x7F'
+                     b'\x61\x00\x00\x00\x00\x00\x00\xF0\x7F'
+                     b'\xCA\x02\x04\x00\x00\x80\x7F'
+                     b'\xD2\x02\x08\x00\x00\x00\x00\x00\x00\xF0\x7F')
+
+    golden_message = message_module.TestAllTypes()
+    golden_message.ParseFromString(golden_data)
+    self.assertTrue(IsPosInf(golden_message.optional_float))
+    self.assertTrue(IsPosInf(golden_message.optional_double))
+    self.assertTrue(IsPosInf(golden_message.repeated_float[0]))
+    self.assertTrue(IsPosInf(golden_message.repeated_double[0]))
+    self.assertEqual(golden_data, golden_message.SerializeToString())
+
+  def testNegativeInfinity(self, message_module):
+    if message_module is unittest_pb2:
+      golden_data = (b'\x5D\x00\x00\x80\xFF'
+                     b'\x61\x00\x00\x00\x00\x00\x00\xF0\xFF'
+                     b'\xCD\x02\x00\x00\x80\xFF'
+                     b'\xD1\x02\x00\x00\x00\x00\x00\x00\xF0\xFF')
+    else:
+      golden_data = (b'\x5D\x00\x00\x80\xFF'
+                     b'\x61\x00\x00\x00\x00\x00\x00\xF0\xFF'
+                     b'\xCA\x02\x04\x00\x00\x80\xFF'
+                     b'\xD2\x02\x08\x00\x00\x00\x00\x00\x00\xF0\xFF')
+
+    golden_message = message_module.TestAllTypes()
+    golden_message.ParseFromString(golden_data)
+    self.assertTrue(IsNegInf(golden_message.optional_float))
+    self.assertTrue(IsNegInf(golden_message.optional_double))
+    self.assertTrue(IsNegInf(golden_message.repeated_float[0]))
+    self.assertTrue(IsNegInf(golden_message.repeated_double[0]))
+    self.assertEqual(golden_data, golden_message.SerializeToString())
+
+  def testNotANumber(self, message_module):
+    golden_data = (b'\x5D\x00\x00\xC0\x7F'
+                   b'\x61\x00\x00\x00\x00\x00\x00\xF8\x7F'
+                   b'\xCD\x02\x00\x00\xC0\x7F'
+                   b'\xD1\x02\x00\x00\x00\x00\x00\x00\xF8\x7F')
+    golden_message = message_module.TestAllTypes()
+    golden_message.ParseFromString(golden_data)
+    self.assertTrue(isnan(golden_message.optional_float))
+    self.assertTrue(isnan(golden_message.optional_double))
+    self.assertTrue(isnan(golden_message.repeated_float[0]))
+    self.assertTrue(isnan(golden_message.repeated_double[0]))
+
+    # The protocol buffer may serialize to any one of multiple different
+    # representations of a NaN.  Rather than verify a specific representation,
+    # verify the serialized string can be converted into a correctly
+    # behaving protocol buffer.
+    serialized = golden_message.SerializeToString()
+    message = message_module.TestAllTypes()
+    message.ParseFromString(serialized)
+    self.assertTrue(isnan(message.optional_float))
+    self.assertTrue(isnan(message.optional_double))
+    self.assertTrue(isnan(message.repeated_float[0]))
+    self.assertTrue(isnan(message.repeated_double[0]))
+
+  def testPositiveInfinityPacked(self, message_module):
+    golden_data = (b'\xA2\x06\x04\x00\x00\x80\x7F'
+                   b'\xAA\x06\x08\x00\x00\x00\x00\x00\x00\xF0\x7F')
+    golden_message = message_module.TestPackedTypes()
+    golden_message.ParseFromString(golden_data)
+    self.assertTrue(IsPosInf(golden_message.packed_float[0]))
+    self.assertTrue(IsPosInf(golden_message.packed_double[0]))
+    self.assertEqual(golden_data, golden_message.SerializeToString())
+
+  def testNegativeInfinityPacked(self, message_module):
+    golden_data = (b'\xA2\x06\x04\x00\x00\x80\xFF'
+                   b'\xAA\x06\x08\x00\x00\x00\x00\x00\x00\xF0\xFF')
+    golden_message = message_module.TestPackedTypes()
+    golden_message.ParseFromString(golden_data)
+    self.assertTrue(IsNegInf(golden_message.packed_float[0]))
+    self.assertTrue(IsNegInf(golden_message.packed_double[0]))
+    self.assertEqual(golden_data, golden_message.SerializeToString())
+
+  def testNotANumberPacked(self, message_module):
+    golden_data = (b'\xA2\x06\x04\x00\x00\xC0\x7F'
+                   b'\xAA\x06\x08\x00\x00\x00\x00\x00\x00\xF8\x7F')
+    golden_message = message_module.TestPackedTypes()
+    golden_message.ParseFromString(golden_data)
+    self.assertTrue(isnan(golden_message.packed_float[0]))
+    self.assertTrue(isnan(golden_message.packed_double[0]))
+
+    serialized = golden_message.SerializeToString()
+    message = message_module.TestPackedTypes()
+    message.ParseFromString(serialized)
+    self.assertTrue(isnan(message.packed_float[0]))
+    self.assertTrue(isnan(message.packed_double[0]))
+
+  def testExtremeFloatValues(self, message_module):
+    message = message_module.TestAllTypes()
+
+    # Most positive exponent, no significand bits set.
+    kMostPosExponentNoSigBits = math.pow(2, 127)
+    message.optional_float = kMostPosExponentNoSigBits
+    message.ParseFromString(message.SerializeToString())
+    self.assertTrue(message.optional_float == kMostPosExponentNoSigBits)
+
+    # Most positive exponent, one significand bit set.
+    kMostPosExponentOneSigBit = 1.5 * math.pow(2, 127)
+    message.optional_float = kMostPosExponentOneSigBit
+    message.ParseFromString(message.SerializeToString())
+    self.assertTrue(message.optional_float == kMostPosExponentOneSigBit)
+
+    # Repeat last two cases with values of same magnitude, but negative.
+    message.optional_float = -kMostPosExponentNoSigBits
+    message.ParseFromString(message.SerializeToString())
+    self.assertTrue(message.optional_float == -kMostPosExponentNoSigBits)
+
+    message.optional_float = -kMostPosExponentOneSigBit
+    message.ParseFromString(message.SerializeToString())
+    self.assertTrue(message.optional_float == -kMostPosExponentOneSigBit)
+
+    # Most negative exponent, no significand bits set.
+    kMostNegExponentNoSigBits = math.pow(2, -127)
+    message.optional_float = kMostNegExponentNoSigBits
+    message.ParseFromString(message.SerializeToString())
+    self.assertTrue(message.optional_float == kMostNegExponentNoSigBits)
+
+    # Most negative exponent, one significand bit set.
+    kMostNegExponentOneSigBit = 1.5 * math.pow(2, -127)
+    message.optional_float = kMostNegExponentOneSigBit
+    message.ParseFromString(message.SerializeToString())
+    self.assertTrue(message.optional_float == kMostNegExponentOneSigBit)
+
+    # Repeat last two cases with values of the same magnitude, but negative.
+    message.optional_float = -kMostNegExponentNoSigBits
+    message.ParseFromString(message.SerializeToString())
+    self.assertTrue(message.optional_float == -kMostNegExponentNoSigBits)
+
+    message.optional_float = -kMostNegExponentOneSigBit
+    message.ParseFromString(message.SerializeToString())
+    self.assertTrue(message.optional_float == -kMostNegExponentOneSigBit)
+
+  def testExtremeDoubleValues(self, message_module):
+    message = message_module.TestAllTypes()
+
+    # Most positive exponent, no significand bits set.
+    kMostPosExponentNoSigBits = math.pow(2, 1023)
+    message.optional_double = kMostPosExponentNoSigBits
+    message.ParseFromString(message.SerializeToString())
+    self.assertTrue(message.optional_double == kMostPosExponentNoSigBits)
+
+    # Most positive exponent, one significand bit set.
+    kMostPosExponentOneSigBit = 1.5 * math.pow(2, 1023)
+    message.optional_double = kMostPosExponentOneSigBit
+    message.ParseFromString(message.SerializeToString())
+    self.assertTrue(message.optional_double == kMostPosExponentOneSigBit)
+
+    # Repeat last two cases with values of same magnitude, but negative.
+    message.optional_double = -kMostPosExponentNoSigBits
+    message.ParseFromString(message.SerializeToString())
+    self.assertTrue(message.optional_double == -kMostPosExponentNoSigBits)
+
+    message.optional_double = -kMostPosExponentOneSigBit
+    message.ParseFromString(message.SerializeToString())
+    self.assertTrue(message.optional_double == -kMostPosExponentOneSigBit)
+
+    # Most negative exponent, no significand bits set.
+    kMostNegExponentNoSigBits = math.pow(2, -1023)
+    message.optional_double = kMostNegExponentNoSigBits
+    message.ParseFromString(message.SerializeToString())
+    self.assertTrue(message.optional_double == kMostNegExponentNoSigBits)
+
+    # Most negative exponent, one significand bit set.
+    kMostNegExponentOneSigBit = 1.5 * math.pow(2, -1023)
+    message.optional_double = kMostNegExponentOneSigBit
+    message.ParseFromString(message.SerializeToString())
+    self.assertTrue(message.optional_double == kMostNegExponentOneSigBit)
+
+    # Repeat last two cases with values of the same magnitude, but negative.
+    message.optional_double = -kMostNegExponentNoSigBits
+    message.ParseFromString(message.SerializeToString())
+    self.assertTrue(message.optional_double == -kMostNegExponentNoSigBits)
+
+    message.optional_double = -kMostNegExponentOneSigBit
+    message.ParseFromString(message.SerializeToString())
+    self.assertTrue(message.optional_double == -kMostNegExponentOneSigBit)
+
+  def testFloatPrinting(self, message_module):
+    message = message_module.TestAllTypes()
+    message.optional_float = 2.0
+    self.assertEqual(str(message), 'optional_float: 2.0\n')
+
+  def testHighPrecisionFloatPrinting(self, message_module):
+    message = message_module.TestAllTypes()
+    message.optional_double = 0.12345678912345678
+    if sys.version_info >= (3,):
+      self.assertEqual(str(message), 'optional_double: 0.12345678912345678\n')
+    else:
+      self.assertEqual(str(message), 'optional_double: 0.123456789123\n')
+
+  def testUnknownFieldPrinting(self, message_module):
+    populated = message_module.TestAllTypes()
+    test_util.SetAllNonLazyFields(populated)
+    empty = message_module.TestEmptyMessage()
+    empty.ParseFromString(populated.SerializeToString())
+    self.assertEqual(str(empty), '')
+
+  def testRepeatedNestedFieldIteration(self, message_module):
+    msg = message_module.TestAllTypes()
+    msg.repeated_nested_message.add(bb=1)
+    msg.repeated_nested_message.add(bb=2)
+    msg.repeated_nested_message.add(bb=3)
+    msg.repeated_nested_message.add(bb=4)
+
+    self.assertEqual([1, 2, 3, 4],
+                     [m.bb for m in msg.repeated_nested_message])
+    self.assertEqual([4, 3, 2, 1],
+                     [m.bb for m in reversed(msg.repeated_nested_message)])
+    self.assertEqual([4, 3, 2, 1],
+                     [m.bb for m in msg.repeated_nested_message[::-1]])
+
+  def testSortingRepeatedScalarFieldsDefaultComparator(self, message_module):
+    """Check some different types with the default comparator."""
+    message = message_module.TestAllTypes()
+
+    # TODO(mattp): would testing more scalar types strengthen test?
+    message.repeated_int32.append(1)
+    message.repeated_int32.append(3)
+    message.repeated_int32.append(2)
+    message.repeated_int32.sort()
+    self.assertEqual(message.repeated_int32[0], 1)
+    self.assertEqual(message.repeated_int32[1], 2)
+    self.assertEqual(message.repeated_int32[2], 3)
+
+    message.repeated_float.append(1.1)
+    message.repeated_float.append(1.3)
+    message.repeated_float.append(1.2)
+    message.repeated_float.sort()
+    self.assertAlmostEqual(message.repeated_float[0], 1.1)
+    self.assertAlmostEqual(message.repeated_float[1], 1.2)
+    self.assertAlmostEqual(message.repeated_float[2], 1.3)
+
+    message.repeated_string.append('a')
+    message.repeated_string.append('c')
+    message.repeated_string.append('b')
+    message.repeated_string.sort()
+    self.assertEqual(message.repeated_string[0], 'a')
+    self.assertEqual(message.repeated_string[1], 'b')
+    self.assertEqual(message.repeated_string[2], 'c')
+
+    message.repeated_bytes.append(b'a')
+    message.repeated_bytes.append(b'c')
+    message.repeated_bytes.append(b'b')
+    message.repeated_bytes.sort()
+    self.assertEqual(message.repeated_bytes[0], b'a')
+    self.assertEqual(message.repeated_bytes[1], b'b')
+    self.assertEqual(message.repeated_bytes[2], b'c')
+
+  def testSortingRepeatedScalarFieldsCustomComparator(self, message_module):
+    """Check some different types with custom comparator."""
+    message = message_module.TestAllTypes()
+
+    message.repeated_int32.append(-3)
+    message.repeated_int32.append(-2)
+    message.repeated_int32.append(-1)
+    message.repeated_int32.sort(key=abs)
+    self.assertEqual(message.repeated_int32[0], -1)
+    self.assertEqual(message.repeated_int32[1], -2)
+    self.assertEqual(message.repeated_int32[2], -3)
+
+    message.repeated_string.append('aaa')
+    message.repeated_string.append('bb')
+    message.repeated_string.append('c')
+    message.repeated_string.sort(key=len)
+    self.assertEqual(message.repeated_string[0], 'c')
+    self.assertEqual(message.repeated_string[1], 'bb')
+    self.assertEqual(message.repeated_string[2], 'aaa')
+
+  def testSortingRepeatedCompositeFieldsCustomComparator(self, message_module):
+    """Check passing a custom comparator to sort a repeated composite field."""
+    message = message_module.TestAllTypes()
+
+    message.repeated_nested_message.add().bb = 1
+    message.repeated_nested_message.add().bb = 3
+    message.repeated_nested_message.add().bb = 2
+    message.repeated_nested_message.add().bb = 6
+    message.repeated_nested_message.add().bb = 5
+    message.repeated_nested_message.add().bb = 4
+    message.repeated_nested_message.sort(key=operator.attrgetter('bb'))
+    self.assertEqual(message.repeated_nested_message[0].bb, 1)
+    self.assertEqual(message.repeated_nested_message[1].bb, 2)
+    self.assertEqual(message.repeated_nested_message[2].bb, 3)
+    self.assertEqual(message.repeated_nested_message[3].bb, 4)
+    self.assertEqual(message.repeated_nested_message[4].bb, 5)
+    self.assertEqual(message.repeated_nested_message[5].bb, 6)
+
+  def testSortingRepeatedCompositeFieldsStable(self, message_module):
+    """Check passing a custom comparator to sort a repeated composite field."""
+    message = message_module.TestAllTypes()
+
+    message.repeated_nested_message.add().bb = 21
+    message.repeated_nested_message.add().bb = 20
+    message.repeated_nested_message.add().bb = 13
+    message.repeated_nested_message.add().bb = 33
+    message.repeated_nested_message.add().bb = 11
+    message.repeated_nested_message.add().bb = 24
+    message.repeated_nested_message.add().bb = 10
+    message.repeated_nested_message.sort(key=lambda z: z.bb // 10)
+    self.assertEqual(
+        [13, 11, 10, 21, 20, 24, 33],
+        [n.bb for n in message.repeated_nested_message])
+
+    # Make sure that for the C++ implementation, the underlying fields
+    # are actually reordered.
+    pb = message.SerializeToString()
+    message.Clear()
+    message.MergeFromString(pb)
+    self.assertEqual(
+        [13, 11, 10, 21, 20, 24, 33],
+        [n.bb for n in message.repeated_nested_message])
+
+  def testRepeatedCompositeFieldSortArguments(self, message_module):
+    """Check sorting a repeated composite field using list.sort() arguments."""
+    message = message_module.TestAllTypes()
+
+    get_bb = operator.attrgetter('bb')
+    cmp_bb = lambda a, b: cmp(a.bb, b.bb)
+    message.repeated_nested_message.add().bb = 1
+    message.repeated_nested_message.add().bb = 3
+    message.repeated_nested_message.add().bb = 2
+    message.repeated_nested_message.add().bb = 6
+    message.repeated_nested_message.add().bb = 5
+    message.repeated_nested_message.add().bb = 4
+    message.repeated_nested_message.sort(key=get_bb)
+    self.assertEqual([k.bb for k in message.repeated_nested_message],
+                     [1, 2, 3, 4, 5, 6])
+    message.repeated_nested_message.sort(key=get_bb, reverse=True)
+    self.assertEqual([k.bb for k in message.repeated_nested_message],
+                     [6, 5, 4, 3, 2, 1])
+    if sys.version_info >= (3,): return  # No cmp sorting in PY3.
+    message.repeated_nested_message.sort(sort_function=cmp_bb)
+    self.assertEqual([k.bb for k in message.repeated_nested_message],
+                     [1, 2, 3, 4, 5, 6])
+    message.repeated_nested_message.sort(cmp=cmp_bb, reverse=True)
+    self.assertEqual([k.bb for k in message.repeated_nested_message],
+                     [6, 5, 4, 3, 2, 1])
+
+  def testRepeatedScalarFieldSortArguments(self, message_module):
+    """Check sorting a scalar field using list.sort() arguments."""
+    message = message_module.TestAllTypes()
+
+    message.repeated_int32.append(-3)
+    message.repeated_int32.append(-2)
+    message.repeated_int32.append(-1)
+    message.repeated_int32.sort(key=abs)
+    self.assertEqual(list(message.repeated_int32), [-1, -2, -3])
+    message.repeated_int32.sort(key=abs, reverse=True)
+    self.assertEqual(list(message.repeated_int32), [-3, -2, -1])
+    if sys.version_info < (3,):  # No cmp sorting in PY3.
+      abs_cmp = lambda a, b: cmp(abs(a), abs(b))
+      message.repeated_int32.sort(sort_function=abs_cmp)
+      self.assertEqual(list(message.repeated_int32), [-1, -2, -3])
+      message.repeated_int32.sort(cmp=abs_cmp, reverse=True)
+      self.assertEqual(list(message.repeated_int32), [-3, -2, -1])
+
+    message.repeated_string.append('aaa')
+    message.repeated_string.append('bb')
+    message.repeated_string.append('c')
+    message.repeated_string.sort(key=len)
+    self.assertEqual(list(message.repeated_string), ['c', 'bb', 'aaa'])
+    message.repeated_string.sort(key=len, reverse=True)
+    self.assertEqual(list(message.repeated_string), ['aaa', 'bb', 'c'])
+    if sys.version_info < (3,):  # No cmp sorting in PY3.
+      len_cmp = lambda a, b: cmp(len(a), len(b))
+      message.repeated_string.sort(sort_function=len_cmp)
+      self.assertEqual(list(message.repeated_string), ['c', 'bb', 'aaa'])
+      message.repeated_string.sort(cmp=len_cmp, reverse=True)
+      self.assertEqual(list(message.repeated_string), ['aaa', 'bb', 'c'])
+
+  def testRepeatedFieldsComparable(self, message_module):
+    m1 = message_module.TestAllTypes()
+    m2 = message_module.TestAllTypes()
+    m1.repeated_int32.append(0)
+    m1.repeated_int32.append(1)
+    m1.repeated_int32.append(2)
+    m2.repeated_int32.append(0)
+    m2.repeated_int32.append(1)
+    m2.repeated_int32.append(2)
+    m1.repeated_nested_message.add().bb = 1
+    m1.repeated_nested_message.add().bb = 2
+    m1.repeated_nested_message.add().bb = 3
+    m2.repeated_nested_message.add().bb = 1
+    m2.repeated_nested_message.add().bb = 2
+    m2.repeated_nested_message.add().bb = 3
+
+    if sys.version_info >= (3,): return  # No cmp() in PY3.
+
+    # These comparisons should not raise errors.
+    _ = m1 < m2
+    _ = m1.repeated_nested_message < m2.repeated_nested_message
+
+    # Make sure cmp always works. If it wasn't defined, these would be
+    # id() comparisons and would all fail.
+    self.assertEqual(cmp(m1, m2), 0)
+    self.assertEqual(cmp(m1.repeated_int32, m2.repeated_int32), 0)
+    self.assertEqual(cmp(m1.repeated_int32, [0, 1, 2]), 0)
+    self.assertEqual(cmp(m1.repeated_nested_message,
+                         m2.repeated_nested_message), 0)
+    with self.assertRaises(TypeError):
+      # Can't compare repeated composite containers to lists.
+      cmp(m1.repeated_nested_message, m2.repeated_nested_message[:])
+
+    # TODO(anuraag): Implement extensiondict comparison in C++ and then add test
+
+  def testRepeatedFieldsAreSequences(self, message_module):
+    m = message_module.TestAllTypes()
+    self.assertIsInstance(m.repeated_int32, collections.MutableSequence)
+    self.assertIsInstance(m.repeated_nested_message,
+                          collections.MutableSequence)
+
+  def ensureNestedMessageExists(self, msg, attribute):
+    """Make sure that a nested message object exists.
+
+    As soon as a nested message attribute is accessed, it will be present in the
+    _fields dict, without being marked as actually being set.
+    """
+    getattr(msg, attribute)
+    self.assertFalse(msg.HasField(attribute))
+
+  def testOneofGetCaseNonexistingField(self, message_module):
+    m = message_module.TestAllTypes()
+    self.assertRaises(ValueError, m.WhichOneof, 'no_such_oneof_field')
+
+  def testOneofDefaultValues(self, message_module):
+    m = message_module.TestAllTypes()
+    self.assertIs(None, m.WhichOneof('oneof_field'))
+    self.assertFalse(m.HasField('oneof_uint32'))
+
+    # Oneof is set even when setting it to a default value.
+    m.oneof_uint32 = 0
+    self.assertEqual('oneof_uint32', m.WhichOneof('oneof_field'))
+    self.assertTrue(m.HasField('oneof_uint32'))
+    self.assertFalse(m.HasField('oneof_string'))
+
+    m.oneof_string = ""
+    self.assertEqual('oneof_string', m.WhichOneof('oneof_field'))
+    self.assertTrue(m.HasField('oneof_string'))
+    self.assertFalse(m.HasField('oneof_uint32'))
+
+  def testOneofSemantics(self, message_module):
+    m = message_module.TestAllTypes()
+    self.assertIs(None, m.WhichOneof('oneof_field'))
+
+    m.oneof_uint32 = 11
+    self.assertEqual('oneof_uint32', m.WhichOneof('oneof_field'))
+    self.assertTrue(m.HasField('oneof_uint32'))
+
+    m.oneof_string = u'foo'
+    self.assertEqual('oneof_string', m.WhichOneof('oneof_field'))
+    self.assertFalse(m.HasField('oneof_uint32'))
+    self.assertTrue(m.HasField('oneof_string'))
+
+    # Read nested message accessor without accessing submessage.
+    m.oneof_nested_message
+    self.assertEqual('oneof_string', m.WhichOneof('oneof_field'))
+    self.assertTrue(m.HasField('oneof_string'))
+    self.assertFalse(m.HasField('oneof_nested_message'))
+
+    # Read accessor of nested message without accessing submessage.
+    m.oneof_nested_message.bb
+    self.assertEqual('oneof_string', m.WhichOneof('oneof_field'))
+    self.assertTrue(m.HasField('oneof_string'))
+    self.assertFalse(m.HasField('oneof_nested_message'))
+
+    m.oneof_nested_message.bb = 11
+    self.assertEqual('oneof_nested_message', m.WhichOneof('oneof_field'))
+    self.assertFalse(m.HasField('oneof_string'))
+    self.assertTrue(m.HasField('oneof_nested_message'))
+
+    m.oneof_bytes = b'bb'
+    self.assertEqual('oneof_bytes', m.WhichOneof('oneof_field'))
+    self.assertFalse(m.HasField('oneof_nested_message'))
+    self.assertTrue(m.HasField('oneof_bytes'))
+
+  def testOneofCompositeFieldReadAccess(self, message_module):
+    m = message_module.TestAllTypes()
+    m.oneof_uint32 = 11
+
+    self.ensureNestedMessageExists(m, 'oneof_nested_message')
+    self.assertEqual('oneof_uint32', m.WhichOneof('oneof_field'))
+    self.assertEqual(11, m.oneof_uint32)
+
+  def testOneofWhichOneof(self, message_module):
+    m = message_module.TestAllTypes()
+    self.assertIs(None, m.WhichOneof('oneof_field'))
+    if message_module is unittest_pb2:
+      self.assertFalse(m.HasField('oneof_field'))
+
+    m.oneof_uint32 = 11
+    self.assertEqual('oneof_uint32', m.WhichOneof('oneof_field'))
+    if message_module is unittest_pb2:
+      self.assertTrue(m.HasField('oneof_field'))
+
+    m.oneof_bytes = b'bb'
+    self.assertEqual('oneof_bytes', m.WhichOneof('oneof_field'))
+
+    m.ClearField('oneof_bytes')
+    self.assertIs(None, m.WhichOneof('oneof_field'))
+    if message_module is unittest_pb2:
+      self.assertFalse(m.HasField('oneof_field'))
+
+  def testOneofClearField(self, message_module):
+    m = message_module.TestAllTypes()
+    m.oneof_uint32 = 11
+    m.ClearField('oneof_field')
+    if message_module is unittest_pb2:
+      self.assertFalse(m.HasField('oneof_field'))
+    self.assertFalse(m.HasField('oneof_uint32'))
+    self.assertIs(None, m.WhichOneof('oneof_field'))
+
+  def testOneofClearSetField(self, message_module):
+    m = message_module.TestAllTypes()
+    m.oneof_uint32 = 11
+    m.ClearField('oneof_uint32')
+    if message_module is unittest_pb2:
+      self.assertFalse(m.HasField('oneof_field'))
+    self.assertFalse(m.HasField('oneof_uint32'))
+    self.assertIs(None, m.WhichOneof('oneof_field'))
+
+  def testOneofClearUnsetField(self, message_module):
+    m = message_module.TestAllTypes()
+    m.oneof_uint32 = 11
+    self.ensureNestedMessageExists(m, 'oneof_nested_message')
+    m.ClearField('oneof_nested_message')
+    self.assertEqual(11, m.oneof_uint32)
+    if message_module is unittest_pb2:
+      self.assertTrue(m.HasField('oneof_field'))
+    self.assertTrue(m.HasField('oneof_uint32'))
+    self.assertEqual('oneof_uint32', m.WhichOneof('oneof_field'))
+
+  def testOneofDeserialize(self, message_module):
+    m = message_module.TestAllTypes()
+    m.oneof_uint32 = 11
+    m2 = message_module.TestAllTypes()
+    m2.ParseFromString(m.SerializeToString())
+    self.assertEqual('oneof_uint32', m2.WhichOneof('oneof_field'))
+
+  def testOneofCopyFrom(self, message_module):
+    m = message_module.TestAllTypes()
+    m.oneof_uint32 = 11
+    m2 = message_module.TestAllTypes()
+    m2.CopyFrom(m)
+    self.assertEqual('oneof_uint32', m2.WhichOneof('oneof_field'))
+
+  def testOneofNestedMergeFrom(self, message_module):
+    m = message_module.NestedTestAllTypes()
+    m.payload.oneof_uint32 = 11
+    m2 = message_module.NestedTestAllTypes()
+    m2.payload.oneof_bytes = b'bb'
+    m2.child.payload.oneof_bytes = b'bb'
+    m2.MergeFrom(m)
+    self.assertEqual('oneof_uint32', m2.payload.WhichOneof('oneof_field'))
+    self.assertEqual('oneof_bytes', m2.child.payload.WhichOneof('oneof_field'))
+
+  def testOneofMessageMergeFrom(self, message_module):
+    m = message_module.NestedTestAllTypes()
+    m.payload.oneof_nested_message.bb = 11
+    m.child.payload.oneof_nested_message.bb = 12
+    m2 = message_module.NestedTestAllTypes()
+    m2.payload.oneof_uint32 = 13
+    m2.MergeFrom(m)
+    self.assertEqual('oneof_nested_message',
+                     m2.payload.WhichOneof('oneof_field'))
+    self.assertEqual('oneof_nested_message',
+                     m2.child.payload.WhichOneof('oneof_field'))
+
+  def testOneofNestedMessageInit(self, message_module):
+    m = message_module.TestAllTypes(
+        oneof_nested_message=message_module.TestAllTypes.NestedMessage())
+    self.assertEqual('oneof_nested_message', m.WhichOneof('oneof_field'))
+
+  def testOneofClear(self, message_module):
+    m = message_module.TestAllTypes()
+    m.oneof_uint32 = 11
+    m.Clear()
+    self.assertIsNone(m.WhichOneof('oneof_field'))
+    m.oneof_bytes = b'bb'
+    self.assertEqual('oneof_bytes', m.WhichOneof('oneof_field'))
+
+  def testAssignByteStringToUnicodeField(self, message_module):
+    """Assigning a byte string to a string field should result
+    in the value being converted to a Unicode string."""
+    m = message_module.TestAllTypes()
+    m.optional_string = str('')
+    self.assertIsInstance(m.optional_string, six.text_type)
+
+  def testLongValuedSlice(self, message_module):
+    """It should be possible to use long-valued indicies in slices
+
+    This didn't used to work in the v2 C++ implementation.
+    """
+    m = message_module.TestAllTypes()
+
+    # Repeated scalar
+    m.repeated_int32.append(1)
+    sl = m.repeated_int32[long(0):long(len(m.repeated_int32))]
+    self.assertEqual(len(m.repeated_int32), len(sl))
+
+    # Repeated composite
+    m.repeated_nested_message.add().bb = 3
+    sl = m.repeated_nested_message[long(0):long(len(m.repeated_nested_message))]
+    self.assertEqual(len(m.repeated_nested_message), len(sl))
+
+  def testExtendShouldNotSwallowExceptions(self, message_module):
+    """This didn't use to work in the v2 C++ implementation."""
+    m = message_module.TestAllTypes()
+    with self.assertRaises(NameError) as _:
+      m.repeated_int32.extend(a for i in range(10))  # pylint: disable=undefined-variable
+    with self.assertRaises(NameError) as _:
+      m.repeated_nested_enum.extend(
+          a for i in range(10))  # pylint: disable=undefined-variable
+
+  FALSY_VALUES = [None, False, 0, 0.0, b'', u'', bytearray(), [], {}, set()]
+
+  def testExtendInt32WithNothing(self, message_module):
+    """Test no-ops extending repeated int32 fields."""
+    m = message_module.TestAllTypes()
+    self.assertSequenceEqual([], m.repeated_int32)
+
+    # TODO(ptucker): Deprecate this behavior. b/18413862
+    for falsy_value in MessageTest.FALSY_VALUES:
+      m.repeated_int32.extend(falsy_value)
+      self.assertSequenceEqual([], m.repeated_int32)
+
+    m.repeated_int32.extend([])
+    self.assertSequenceEqual([], m.repeated_int32)
+
+  def testExtendFloatWithNothing(self, message_module):
+    """Test no-ops extending repeated float fields."""
+    m = message_module.TestAllTypes()
+    self.assertSequenceEqual([], m.repeated_float)
+
+    # TODO(ptucker): Deprecate this behavior. b/18413862
+    for falsy_value in MessageTest.FALSY_VALUES:
+      m.repeated_float.extend(falsy_value)
+      self.assertSequenceEqual([], m.repeated_float)
+
+    m.repeated_float.extend([])
+    self.assertSequenceEqual([], m.repeated_float)
+
+  def testExtendStringWithNothing(self, message_module):
+    """Test no-ops extending repeated string fields."""
+    m = message_module.TestAllTypes()
+    self.assertSequenceEqual([], m.repeated_string)
+
+    # TODO(ptucker): Deprecate this behavior. b/18413862
+    for falsy_value in MessageTest.FALSY_VALUES:
+      m.repeated_string.extend(falsy_value)
+      self.assertSequenceEqual([], m.repeated_string)
+
+    m.repeated_string.extend([])
+    self.assertSequenceEqual([], m.repeated_string)
+
+  def testExtendInt32WithPythonList(self, message_module):
+    """Test extending repeated int32 fields with python lists."""
+    m = message_module.TestAllTypes()
+    self.assertSequenceEqual([], m.repeated_int32)
+    m.repeated_int32.extend([0])
+    self.assertSequenceEqual([0], m.repeated_int32)
+    m.repeated_int32.extend([1, 2])
+    self.assertSequenceEqual([0, 1, 2], m.repeated_int32)
+    m.repeated_int32.extend([3, 4])
+    self.assertSequenceEqual([0, 1, 2, 3, 4], m.repeated_int32)
+
+  def testExtendFloatWithPythonList(self, message_module):
+    """Test extending repeated float fields with python lists."""
+    m = message_module.TestAllTypes()
+    self.assertSequenceEqual([], m.repeated_float)
+    m.repeated_float.extend([0.0])
+    self.assertSequenceEqual([0.0], m.repeated_float)
+    m.repeated_float.extend([1.0, 2.0])
+    self.assertSequenceEqual([0.0, 1.0, 2.0], m.repeated_float)
+    m.repeated_float.extend([3.0, 4.0])
+    self.assertSequenceEqual([0.0, 1.0, 2.0, 3.0, 4.0], m.repeated_float)
+
+  def testExtendStringWithPythonList(self, message_module):
+    """Test extending repeated string fields with python lists."""
+    m = message_module.TestAllTypes()
+    self.assertSequenceEqual([], m.repeated_string)
+    m.repeated_string.extend([''])
+    self.assertSequenceEqual([''], m.repeated_string)
+    m.repeated_string.extend(['11', '22'])
+    self.assertSequenceEqual(['', '11', '22'], m.repeated_string)
+    m.repeated_string.extend(['33', '44'])
+    self.assertSequenceEqual(['', '11', '22', '33', '44'], m.repeated_string)
+
+  def testExtendStringWithString(self, message_module):
+    """Test extending repeated string fields with characters from a string."""
+    m = message_module.TestAllTypes()
+    self.assertSequenceEqual([], m.repeated_string)
+    m.repeated_string.extend('abc')
+    self.assertSequenceEqual(['a', 'b', 'c'], m.repeated_string)
+
+  class TestIterable(object):
+    """This iterable object mimics the behavior of numpy.array.
+
+    __nonzero__ fails for length > 1, and returns bool(item[0]) for length == 1.
+
+    """
+
+    def __init__(self, values=None):
+      self._list = values or []
+
+    def __nonzero__(self):
+      size = len(self._list)
+      if size == 0:
+        return False
+      if size == 1:
+        return bool(self._list[0])
+      raise ValueError('Truth value is ambiguous.')
+
+    def __len__(self):
+      return len(self._list)
+
+    def __iter__(self):
+      return self._list.__iter__()
+
+  def testExtendInt32WithIterable(self, message_module):
+    """Test extending repeated int32 fields with iterable."""
+    m = message_module.TestAllTypes()
+    self.assertSequenceEqual([], m.repeated_int32)
+    m.repeated_int32.extend(MessageTest.TestIterable([]))
+    self.assertSequenceEqual([], m.repeated_int32)
+    m.repeated_int32.extend(MessageTest.TestIterable([0]))
+    self.assertSequenceEqual([0], m.repeated_int32)
+    m.repeated_int32.extend(MessageTest.TestIterable([1, 2]))
+    self.assertSequenceEqual([0, 1, 2], m.repeated_int32)
+    m.repeated_int32.extend(MessageTest.TestIterable([3, 4]))
+    self.assertSequenceEqual([0, 1, 2, 3, 4], m.repeated_int32)
+
+  def testExtendFloatWithIterable(self, message_module):
+    """Test extending repeated float fields with iterable."""
+    m = message_module.TestAllTypes()
+    self.assertSequenceEqual([], m.repeated_float)
+    m.repeated_float.extend(MessageTest.TestIterable([]))
+    self.assertSequenceEqual([], m.repeated_float)
+    m.repeated_float.extend(MessageTest.TestIterable([0.0]))
+    self.assertSequenceEqual([0.0], m.repeated_float)
+    m.repeated_float.extend(MessageTest.TestIterable([1.0, 2.0]))
+    self.assertSequenceEqual([0.0, 1.0, 2.0], m.repeated_float)
+    m.repeated_float.extend(MessageTest.TestIterable([3.0, 4.0]))
+    self.assertSequenceEqual([0.0, 1.0, 2.0, 3.0, 4.0], m.repeated_float)
+
+  def testExtendStringWithIterable(self, message_module):
+    """Test extending repeated string fields with iterable."""
+    m = message_module.TestAllTypes()
+    self.assertSequenceEqual([], m.repeated_string)
+    m.repeated_string.extend(MessageTest.TestIterable([]))
+    self.assertSequenceEqual([], m.repeated_string)
+    m.repeated_string.extend(MessageTest.TestIterable(['']))
+    self.assertSequenceEqual([''], m.repeated_string)
+    m.repeated_string.extend(MessageTest.TestIterable(['1', '2']))
+    self.assertSequenceEqual(['', '1', '2'], m.repeated_string)
+    m.repeated_string.extend(MessageTest.TestIterable(['3', '4']))
+    self.assertSequenceEqual(['', '1', '2', '3', '4'], m.repeated_string)
+
+  def testPickleRepeatedScalarContainer(self, message_module):
+    # TODO(tibell): The pure-Python implementation support pickling of
+    #   scalar containers in *some* cases. For now the cpp2 version
+    #   throws an exception to avoid a segfault. Investigate if we
+    #   want to support pickling of these fields.
+    #
+    # For more information see: https://b2.corp.google.com/u/0/issues/18677897
+    if (api_implementation.Type() != 'cpp' or
+        api_implementation.Version() == 2):
+      return
+    m = message_module.TestAllTypes()
+    with self.assertRaises(pickle.PickleError) as _:
+      pickle.dumps(m.repeated_int32, pickle.HIGHEST_PROTOCOL)
+
+  def testSortEmptyRepeatedCompositeContainer(self, message_module):
+    """Exercise a scenario that has led to segfaults in the past.
+    """
+    m = message_module.TestAllTypes()
+    m.repeated_nested_message.sort()
+
+  def testHasFieldOnRepeatedField(self, message_module):
+    """Using HasField on a repeated field should raise an exception.
+    """
+    m = message_module.TestAllTypes()
+    with self.assertRaises(ValueError) as _:
+      m.HasField('repeated_int32')
+
+  def testRepeatedScalarFieldPop(self, message_module):
+    m = message_module.TestAllTypes()
+    with self.assertRaises(IndexError) as _:
+      m.repeated_int32.pop()
+    m.repeated_int32.extend(range(5))
+    self.assertEqual(4, m.repeated_int32.pop())
+    self.assertEqual(0, m.repeated_int32.pop(0))
+    self.assertEqual(2, m.repeated_int32.pop(1))
+    self.assertEqual([1, 3], m.repeated_int32)
+
+  def testRepeatedCompositeFieldPop(self, message_module):
+    m = message_module.TestAllTypes()
+    with self.assertRaises(IndexError) as _:
+      m.repeated_nested_message.pop()
+    for i in range(5):
+      n = m.repeated_nested_message.add()
+      n.bb = i
+    self.assertEqual(4, m.repeated_nested_message.pop().bb)
+    self.assertEqual(0, m.repeated_nested_message.pop(0).bb)
+    self.assertEqual(2, m.repeated_nested_message.pop(1).bb)
+    self.assertEqual([1, 3], [n.bb for n in m.repeated_nested_message])
+
+
+# Class to test proto2-only features (required, extensions, etc.)
+class Proto2Test(unittest.TestCase):
+
+  def testFieldPresence(self):
+    message = unittest_pb2.TestAllTypes()
+
+    self.assertFalse(message.HasField("optional_int32"))
+    self.assertFalse(message.HasField("optional_bool"))
+    self.assertFalse(message.HasField("optional_nested_message"))
+
+    with self.assertRaises(ValueError):
+      message.HasField("field_doesnt_exist")
+
+    with self.assertRaises(ValueError):
+      message.HasField("repeated_int32")
+    with self.assertRaises(ValueError):
+      message.HasField("repeated_nested_message")
+
+    self.assertEqual(0, message.optional_int32)
+    self.assertEqual(False, message.optional_bool)
+    self.assertEqual(0, message.optional_nested_message.bb)
+
+    # Fields are set even when setting the values to default values.
+    message.optional_int32 = 0
+    message.optional_bool = False
+    message.optional_nested_message.bb = 0
+    self.assertTrue(message.HasField("optional_int32"))
+    self.assertTrue(message.HasField("optional_bool"))
+    self.assertTrue(message.HasField("optional_nested_message"))
+
+    # Set the fields to non-default values.
+    message.optional_int32 = 5
+    message.optional_bool = True
+    message.optional_nested_message.bb = 15
+
+    self.assertTrue(message.HasField("optional_int32"))
+    self.assertTrue(message.HasField("optional_bool"))
+    self.assertTrue(message.HasField("optional_nested_message"))
+
+    # Clearing the fields unsets them and resets their value to default.
+    message.ClearField("optional_int32")
+    message.ClearField("optional_bool")
+    message.ClearField("optional_nested_message")
+
+    self.assertFalse(message.HasField("optional_int32"))
+    self.assertFalse(message.HasField("optional_bool"))
+    self.assertFalse(message.HasField("optional_nested_message"))
+    self.assertEqual(0, message.optional_int32)
+    self.assertEqual(False, message.optional_bool)
+    self.assertEqual(0, message.optional_nested_message.bb)
+
+  # TODO(tibell): The C++ implementations actually allows assignment
+  # of unknown enum values to *scalar* fields (but not repeated
+  # fields). Once checked enum fields becomes the default in the
+  # Python implementation, the C++ implementation should follow suit.
+  def testAssignInvalidEnum(self):
+    """It should not be possible to assign an invalid enum number to an
+    enum field."""
+    m = unittest_pb2.TestAllTypes()
+
+    with self.assertRaises(ValueError) as _:
+      m.optional_nested_enum = 1234567
+    self.assertRaises(ValueError, m.repeated_nested_enum.append, 1234567)
+
+  def testGoldenExtensions(self):
+    golden_data = test_util.GoldenFileData('golden_message')
+    golden_message = unittest_pb2.TestAllExtensions()
+    golden_message.ParseFromString(golden_data)
+    all_set = unittest_pb2.TestAllExtensions()
+    test_util.SetAllExtensions(all_set)
+    self.assertEqual(all_set, golden_message)
+    self.assertEqual(golden_data, golden_message.SerializeToString())
+    golden_copy = copy.deepcopy(golden_message)
+    self.assertEqual(golden_data, golden_copy.SerializeToString())
+
+  def testGoldenPackedExtensions(self):
+    golden_data = test_util.GoldenFileData('golden_packed_fields_message')
+    golden_message = unittest_pb2.TestPackedExtensions()
+    golden_message.ParseFromString(golden_data)
+    all_set = unittest_pb2.TestPackedExtensions()
+    test_util.SetAllPackedExtensions(all_set)
+    self.assertEqual(all_set, golden_message)
+    self.assertEqual(golden_data, all_set.SerializeToString())
+    golden_copy = copy.deepcopy(golden_message)
+    self.assertEqual(golden_data, golden_copy.SerializeToString())
+
+  def testPickleIncompleteProto(self):
+    golden_message = unittest_pb2.TestRequired(a=1)
+    pickled_message = pickle.dumps(golden_message)
+
+    unpickled_message = pickle.loads(pickled_message)
+    self.assertEqual(unpickled_message, golden_message)
+    self.assertEqual(unpickled_message.a, 1)
+    # This is still an incomplete proto - so serializing should fail
+    self.assertRaises(message.EncodeError, unpickled_message.SerializeToString)
+
+
+  # TODO(haberman): this isn't really a proto2-specific test except that this
+  # message has a required field in it.  Should probably be factored out so
+  # that we can test the other parts with proto3.
+  def testParsingMerge(self):
+    """Check the merge behavior when a required or optional field appears
+    multiple times in the input."""
+    messages = [
+        unittest_pb2.TestAllTypes(),
+        unittest_pb2.TestAllTypes(),
+        unittest_pb2.TestAllTypes() ]
+    messages[0].optional_int32 = 1
+    messages[1].optional_int64 = 2
+    messages[2].optional_int32 = 3
+    messages[2].optional_string = 'hello'
+
+    merged_message = unittest_pb2.TestAllTypes()
+    merged_message.optional_int32 = 3
+    merged_message.optional_int64 = 2
+    merged_message.optional_string = 'hello'
+
+    generator = unittest_pb2.TestParsingMerge.RepeatedFieldsGenerator()
+    generator.field1.extend(messages)
+    generator.field2.extend(messages)
+    generator.field3.extend(messages)
+    generator.ext1.extend(messages)
+    generator.ext2.extend(messages)
+    generator.group1.add().field1.MergeFrom(messages[0])
+    generator.group1.add().field1.MergeFrom(messages[1])
+    generator.group1.add().field1.MergeFrom(messages[2])
+    generator.group2.add().field1.MergeFrom(messages[0])
+    generator.group2.add().field1.MergeFrom(messages[1])
+    generator.group2.add().field1.MergeFrom(messages[2])
+
+    data = generator.SerializeToString()
+    parsing_merge = unittest_pb2.TestParsingMerge()
+    parsing_merge.ParseFromString(data)
+
+    # Required and optional fields should be merged.
+    self.assertEqual(parsing_merge.required_all_types, merged_message)
+    self.assertEqual(parsing_merge.optional_all_types, merged_message)
+    self.assertEqual(parsing_merge.optionalgroup.optional_group_all_types,
+                     merged_message)
+    self.assertEqual(parsing_merge.Extensions[
+                     unittest_pb2.TestParsingMerge.optional_ext],
+                     merged_message)
+
+    # Repeated fields should not be merged.
+    self.assertEqual(len(parsing_merge.repeated_all_types), 3)
+    self.assertEqual(len(parsing_merge.repeatedgroup), 3)
+    self.assertEqual(len(parsing_merge.Extensions[
+        unittest_pb2.TestParsingMerge.repeated_ext]), 3)
+
+  def testPythonicInit(self):
+    message = unittest_pb2.TestAllTypes(
+        optional_int32=100,
+        optional_fixed32=200,
+        optional_float=300.5,
+        optional_bytes=b'x',
+        optionalgroup={'a': 400},
+        optional_nested_message={'bb': 500},
+        optional_nested_enum='BAZ',
+        repeatedgroup=[{'a': 600},
+                       {'a': 700}],
+        repeated_nested_enum=['FOO', unittest_pb2.TestAllTypes.BAR],
+        default_int32=800,
+        oneof_string='y')
+    self.assertIsInstance(message, unittest_pb2.TestAllTypes)
+    self.assertEqual(100, message.optional_int32)
+    self.assertEqual(200, message.optional_fixed32)
+    self.assertEqual(300.5, message.optional_float)
+    self.assertEqual(b'x', message.optional_bytes)
+    self.assertEqual(400, message.optionalgroup.a)
+    self.assertIsInstance(message.optional_nested_message, unittest_pb2.TestAllTypes.NestedMessage)
+    self.assertEqual(500, message.optional_nested_message.bb)
+    self.assertEqual(unittest_pb2.TestAllTypes.BAZ,
+                     message.optional_nested_enum)
+    self.assertEqual(2, len(message.repeatedgroup))
+    self.assertEqual(600, message.repeatedgroup[0].a)
+    self.assertEqual(700, message.repeatedgroup[1].a)
+    self.assertEqual(2, len(message.repeated_nested_enum))
+    self.assertEqual(unittest_pb2.TestAllTypes.FOO,
+                     message.repeated_nested_enum[0])
+    self.assertEqual(unittest_pb2.TestAllTypes.BAR,
+                     message.repeated_nested_enum[1])
+    self.assertEqual(800, message.default_int32)
+    self.assertEqual('y', message.oneof_string)
+    self.assertFalse(message.HasField('optional_int64'))
+    self.assertEqual(0, len(message.repeated_float))
+    self.assertEqual(42, message.default_int64)
+
+    message = unittest_pb2.TestAllTypes(optional_nested_enum=u'BAZ')
+    self.assertEqual(unittest_pb2.TestAllTypes.BAZ,
+                     message.optional_nested_enum)
+
+    with self.assertRaises(ValueError):
+      unittest_pb2.TestAllTypes(
+          optional_nested_message={'INVALID_NESTED_FIELD': 17})
+
+    with self.assertRaises(TypeError):
+      unittest_pb2.TestAllTypes(
+          optional_nested_message={'bb': 'INVALID_VALUE_TYPE'})
+
+    with self.assertRaises(ValueError):
+      unittest_pb2.TestAllTypes(optional_nested_enum='INVALID_LABEL')
+
+    with self.assertRaises(ValueError):
+      unittest_pb2.TestAllTypes(repeated_nested_enum='FOO')
+
+
+# Class to test proto3-only features/behavior (updated field presence & enums)
+class Proto3Test(unittest.TestCase):
+
+  # Utility method for comparing equality with a map.
+  def assertMapIterEquals(self, map_iter, dict_value):
+    # Avoid mutating caller's copy.
+    dict_value = dict(dict_value)
+
+    for k, v in map_iter:
+      self.assertEqual(v, dict_value[k])
+      del dict_value[k]
+
+    self.assertEqual({}, dict_value)
+
+  def testFieldPresence(self):
+    message = unittest_proto3_arena_pb2.TestAllTypes()
+
+    # We can't test presence of non-repeated, non-submessage fields.
+    with self.assertRaises(ValueError):
+      message.HasField('optional_int32')
+    with self.assertRaises(ValueError):
+      message.HasField('optional_float')
+    with self.assertRaises(ValueError):
+      message.HasField('optional_string')
+    with self.assertRaises(ValueError):
+      message.HasField('optional_bool')
+
+    # But we can still test presence of submessage fields.
+    self.assertFalse(message.HasField('optional_nested_message'))
+
+    # As with proto2, we can't test presence of fields that don't exist, or
+    # repeated fields.
+    with self.assertRaises(ValueError):
+      message.HasField('field_doesnt_exist')
+
+    with self.assertRaises(ValueError):
+      message.HasField('repeated_int32')
+    with self.assertRaises(ValueError):
+      message.HasField('repeated_nested_message')
+
+    # Fields should default to their type-specific default.
+    self.assertEqual(0, message.optional_int32)
+    self.assertEqual(0, message.optional_float)
+    self.assertEqual('', message.optional_string)
+    self.assertEqual(False, message.optional_bool)
+    self.assertEqual(0, message.optional_nested_message.bb)
+
+    # Setting a submessage should still return proper presence information.
+    message.optional_nested_message.bb = 0
+    self.assertTrue(message.HasField('optional_nested_message'))
+
+    # Set the fields to non-default values.
+    message.optional_int32 = 5
+    message.optional_float = 1.1
+    message.optional_string = 'abc'
+    message.optional_bool = True
+    message.optional_nested_message.bb = 15
+
+    # Clearing the fields unsets them and resets their value to default.
+    message.ClearField('optional_int32')
+    message.ClearField('optional_float')
+    message.ClearField('optional_string')
+    message.ClearField('optional_bool')
+    message.ClearField('optional_nested_message')
+
+    self.assertEqual(0, message.optional_int32)
+    self.assertEqual(0, message.optional_float)
+    self.assertEqual('', message.optional_string)
+    self.assertEqual(False, message.optional_bool)
+    self.assertEqual(0, message.optional_nested_message.bb)
+
+  def testAssignUnknownEnum(self):
+    """Assigning an unknown enum value is allowed and preserves the value."""
+    m = unittest_proto3_arena_pb2.TestAllTypes()
+
+    m.optional_nested_enum = 1234567
+    self.assertEqual(1234567, m.optional_nested_enum)
+    m.repeated_nested_enum.append(22334455)
+    self.assertEqual(22334455, m.repeated_nested_enum[0])
+    # Assignment is a different code path than append for the C++ impl.
+    m.repeated_nested_enum[0] = 7654321
+    self.assertEqual(7654321, m.repeated_nested_enum[0])
+    serialized = m.SerializeToString()
+
+    m2 = unittest_proto3_arena_pb2.TestAllTypes()
+    m2.ParseFromString(serialized)
+    self.assertEqual(1234567, m2.optional_nested_enum)
+    self.assertEqual(7654321, m2.repeated_nested_enum[0])
+
+  # Map isn't really a proto3-only feature. But there is no proto2 equivalent
+  # of google/protobuf/map_unittest.proto right now, so it's not easy to
+  # test both with the same test like we do for the other proto2/proto3 tests.
+  # (google/protobuf/map_protobuf_unittest.proto is very different in the set
+  # of messages and fields it contains).
+  def testScalarMapDefaults(self):
+    msg = map_unittest_pb2.TestMap()
+
+    # Scalars start out unset.
+    self.assertFalse(-123 in msg.map_int32_int32)
+    self.assertFalse(-2**33 in msg.map_int64_int64)
+    self.assertFalse(123 in msg.map_uint32_uint32)
+    self.assertFalse(2**33 in msg.map_uint64_uint64)
+    self.assertFalse('abc' in msg.map_string_string)
+    self.assertFalse(888 in msg.map_int32_enum)
+
+    # Accessing an unset key returns the default.
+    self.assertEqual(0, msg.map_int32_int32[-123])
+    self.assertEqual(0, msg.map_int64_int64[-2**33])
+    self.assertEqual(0, msg.map_uint32_uint32[123])
+    self.assertEqual(0, msg.map_uint64_uint64[2**33])
+    self.assertEqual('', msg.map_string_string['abc'])
+    self.assertEqual(0, msg.map_int32_enum[888])
+
+    # It also sets the value in the map
+    self.assertTrue(-123 in msg.map_int32_int32)
+    self.assertTrue(-2**33 in msg.map_int64_int64)
+    self.assertTrue(123 in msg.map_uint32_uint32)
+    self.assertTrue(2**33 in msg.map_uint64_uint64)
+    self.assertTrue('abc' in msg.map_string_string)
+    self.assertTrue(888 in msg.map_int32_enum)
+
+    self.assertIsInstance(msg.map_string_string['abc'], six.text_type)
+
+    # Accessing an unset key still throws TypeError if the type of the key
+    # is incorrect.
+    with self.assertRaises(TypeError):
+      msg.map_string_string[123]
+
+    with self.assertRaises(TypeError):
+      123 in msg.map_string_string
+
+  def testMapGet(self):
+    # Need to test that get() properly returns the default, even though the dict
+    # has defaultdict-like semantics.
+    msg = map_unittest_pb2.TestMap()
+
+    self.assertIsNone(msg.map_int32_int32.get(5))
+    self.assertEqual(10, msg.map_int32_int32.get(5, 10))
+    self.assertIsNone(msg.map_int32_int32.get(5))
+
+    msg.map_int32_int32[5] = 15
+    self.assertEqual(15, msg.map_int32_int32.get(5))
+
+    self.assertIsNone(msg.map_int32_foreign_message.get(5))
+    self.assertEqual(10, msg.map_int32_foreign_message.get(5, 10))
+
+    submsg = msg.map_int32_foreign_message[5]
+    self.assertIs(submsg, msg.map_int32_foreign_message.get(5))
+
+  def testScalarMap(self):
+    msg = map_unittest_pb2.TestMap()
+
+    self.assertEqual(0, len(msg.map_int32_int32))
+    self.assertFalse(5 in msg.map_int32_int32)
+
+    msg.map_int32_int32[-123] = -456
+    msg.map_int64_int64[-2**33] = -2**34
+    msg.map_uint32_uint32[123] = 456
+    msg.map_uint64_uint64[2**33] = 2**34
+    msg.map_string_string['abc'] = '123'
+    msg.map_int32_enum[888] = 2
+
+    self.assertEqual([], msg.FindInitializationErrors())
+
+    self.assertEqual(1, len(msg.map_string_string))
+
+    # Bad key.
+    with self.assertRaises(TypeError):
+      msg.map_string_string[123] = '123'
+
+    # Verify that trying to assign a bad key doesn't actually add a member to
+    # the map.
+    self.assertEqual(1, len(msg.map_string_string))
+
+    # Bad value.
+    with self.assertRaises(TypeError):
+      msg.map_string_string['123'] = 123
+
+    serialized = msg.SerializeToString()
+    msg2 = map_unittest_pb2.TestMap()
+    msg2.ParseFromString(serialized)
+
+    # Bad key.
+    with self.assertRaises(TypeError):
+      msg2.map_string_string[123] = '123'
+
+    # Bad value.
+    with self.assertRaises(TypeError):
+      msg2.map_string_string['123'] = 123
+
+    self.assertEqual(-456, msg2.map_int32_int32[-123])
+    self.assertEqual(-2**34, msg2.map_int64_int64[-2**33])
+    self.assertEqual(456, msg2.map_uint32_uint32[123])
+    self.assertEqual(2**34, msg2.map_uint64_uint64[2**33])
+    self.assertEqual('123', msg2.map_string_string['abc'])
+    self.assertEqual(2, msg2.map_int32_enum[888])
+
+  def testStringUnicodeConversionInMap(self):
+    msg = map_unittest_pb2.TestMap()
+
+    unicode_obj = u'\u1234'
+    bytes_obj = unicode_obj.encode('utf8')
+
+    msg.map_string_string[bytes_obj] = bytes_obj
+
+    (key, value) = list(msg.map_string_string.items())[0]
+
+    self.assertEqual(key, unicode_obj)
+    self.assertEqual(value, unicode_obj)
+
+    self.assertIsInstance(key, six.text_type)
+    self.assertIsInstance(value, six.text_type)
+
+  def testMessageMap(self):
+    msg = map_unittest_pb2.TestMap()
+
+    self.assertEqual(0, len(msg.map_int32_foreign_message))
+    self.assertFalse(5 in msg.map_int32_foreign_message)
+
+    msg.map_int32_foreign_message[123]
+    # get_or_create() is an alias for getitem.
+    msg.map_int32_foreign_message.get_or_create(-456)
+
+    self.assertEqual(2, len(msg.map_int32_foreign_message))
+    self.assertIn(123, msg.map_int32_foreign_message)
+    self.assertIn(-456, msg.map_int32_foreign_message)
+    self.assertEqual(2, len(msg.map_int32_foreign_message))
+
+    # Bad key.
+    with self.assertRaises(TypeError):
+      msg.map_int32_foreign_message['123']
+
+    # Can't assign directly to submessage.
+    with self.assertRaises(ValueError):
+      msg.map_int32_foreign_message[999] = msg.map_int32_foreign_message[123]
+
+    # Verify that trying to assign a bad key doesn't actually add a member to
+    # the map.
+    self.assertEqual(2, len(msg.map_int32_foreign_message))
+
+    serialized = msg.SerializeToString()
+    msg2 = map_unittest_pb2.TestMap()
+    msg2.ParseFromString(serialized)
+
+    self.assertEqual(2, len(msg2.map_int32_foreign_message))
+    self.assertIn(123, msg2.map_int32_foreign_message)
+    self.assertIn(-456, msg2.map_int32_foreign_message)
+    self.assertEqual(2, len(msg2.map_int32_foreign_message))
+
+  def testMergeFrom(self):
+    msg = map_unittest_pb2.TestMap()
+    msg.map_int32_int32[12] = 34
+    msg.map_int32_int32[56] = 78
+    msg.map_int64_int64[22] = 33
+    msg.map_int32_foreign_message[111].c = 5
+    msg.map_int32_foreign_message[222].c = 10
+
+    msg2 = map_unittest_pb2.TestMap()
+    msg2.map_int32_int32[12] = 55
+    msg2.map_int64_int64[88] = 99
+    msg2.map_int32_foreign_message[222].c = 15
+
+    msg2.MergeFrom(msg)
+
+    self.assertEqual(34, msg2.map_int32_int32[12])
+    self.assertEqual(78, msg2.map_int32_int32[56])
+    self.assertEqual(33, msg2.map_int64_int64[22])
+    self.assertEqual(99, msg2.map_int64_int64[88])
+    self.assertEqual(5, msg2.map_int32_foreign_message[111].c)
+    self.assertEqual(10, msg2.map_int32_foreign_message[222].c)
+
+    # Verify that there is only one entry per key, even though the MergeFrom
+    # may have internally created multiple entries for a single key in the
+    # list representation.
+    as_dict = {}
+    for key in msg2.map_int32_foreign_message:
+      self.assertFalse(key in as_dict)
+      as_dict[key] = msg2.map_int32_foreign_message[key].c
+
+    self.assertEqual({111: 5, 222: 10}, as_dict)
+
+    # Special case: test that delete of item really removes the item, even if
+    # there might have physically been duplicate keys due to the previous merge.
+    # This is only a special case for the C++ implementation which stores the
+    # map as an array.
+    del msg2.map_int32_int32[12]
+    self.assertFalse(12 in msg2.map_int32_int32)
+
+    del msg2.map_int32_foreign_message[222]
+    self.assertFalse(222 in msg2.map_int32_foreign_message)
+
+  def testIntegerMapWithLongs(self):
+    msg = map_unittest_pb2.TestMap()
+    msg.map_int32_int32[long(-123)] = long(-456)
+    msg.map_int64_int64[long(-2**33)] = long(-2**34)
+    msg.map_uint32_uint32[long(123)] = long(456)
+    msg.map_uint64_uint64[long(2**33)] = long(2**34)
+
+    serialized = msg.SerializeToString()
+    msg2 = map_unittest_pb2.TestMap()
+    msg2.ParseFromString(serialized)
+
+    self.assertEqual(-456, msg2.map_int32_int32[-123])
+    self.assertEqual(-2**34, msg2.map_int64_int64[-2**33])
+    self.assertEqual(456, msg2.map_uint32_uint32[123])
+    self.assertEqual(2**34, msg2.map_uint64_uint64[2**33])
+
+  def testMapAssignmentCausesPresence(self):
+    msg = map_unittest_pb2.TestMapSubmessage()
+    msg.test_map.map_int32_int32[123] = 456
+
+    serialized = msg.SerializeToString()
+    msg2 = map_unittest_pb2.TestMapSubmessage()
+    msg2.ParseFromString(serialized)
+
+    self.assertEqual(msg, msg2)
+
+    # Now test that various mutations of the map properly invalidate the
+    # cached size of the submessage.
+    msg.test_map.map_int32_int32[888] = 999
+    serialized = msg.SerializeToString()
+    msg2.ParseFromString(serialized)
+    self.assertEqual(msg, msg2)
+
+    msg.test_map.map_int32_int32.clear()
+    serialized = msg.SerializeToString()
+    msg2.ParseFromString(serialized)
+    self.assertEqual(msg, msg2)
+
+  def testMapAssignmentCausesPresenceForSubmessages(self):
+    msg = map_unittest_pb2.TestMapSubmessage()
+    msg.test_map.map_int32_foreign_message[123].c = 5
+
+    serialized = msg.SerializeToString()
+    msg2 = map_unittest_pb2.TestMapSubmessage()
+    msg2.ParseFromString(serialized)
+
+    self.assertEqual(msg, msg2)
+
+    # Now test that various mutations of the map properly invalidate the
+    # cached size of the submessage.
+    msg.test_map.map_int32_foreign_message[888].c = 7
+    serialized = msg.SerializeToString()
+    msg2.ParseFromString(serialized)
+    self.assertEqual(msg, msg2)
+
+    msg.test_map.map_int32_foreign_message[888].MergeFrom(
+        msg.test_map.map_int32_foreign_message[123])
+    serialized = msg.SerializeToString()
+    msg2.ParseFromString(serialized)
+    self.assertEqual(msg, msg2)
+
+    msg.test_map.map_int32_foreign_message.clear()
+    serialized = msg.SerializeToString()
+    msg2.ParseFromString(serialized)
+    self.assertEqual(msg, msg2)
+
+  def testModifyMapWhileIterating(self):
+    msg = map_unittest_pb2.TestMap()
+
+    string_string_iter = iter(msg.map_string_string)
+    int32_foreign_iter = iter(msg.map_int32_foreign_message)
+
+    msg.map_string_string['abc'] = '123'
+    msg.map_int32_foreign_message[5].c = 5
+
+    with self.assertRaises(RuntimeError):
+      for key in string_string_iter:
+        pass
+
+    with self.assertRaises(RuntimeError):
+      for key in int32_foreign_iter:
+        pass
+
+  def testSubmessageMap(self):
+    msg = map_unittest_pb2.TestMap()
+
+    submsg = msg.map_int32_foreign_message[111]
+    self.assertIs(submsg, msg.map_int32_foreign_message[111])
+    self.assertIsInstance(submsg, unittest_pb2.ForeignMessage)
+
+    submsg.c = 5
+
+    serialized = msg.SerializeToString()
+    msg2 = map_unittest_pb2.TestMap()
+    msg2.ParseFromString(serialized)
+
+    self.assertEqual(5, msg2.map_int32_foreign_message[111].c)
+
+    # Doesn't allow direct submessage assignment.
+    with self.assertRaises(ValueError):
+      msg.map_int32_foreign_message[88] = unittest_pb2.ForeignMessage()
+
+  def testMapIteration(self):
+    msg = map_unittest_pb2.TestMap()
+
+    for k, v in msg.map_int32_int32.items():
+      # Should not be reached.
+      self.assertTrue(False)
+
+    msg.map_int32_int32[2] = 4
+    msg.map_int32_int32[3] = 6
+    msg.map_int32_int32[4] = 8
+    self.assertEqual(3, len(msg.map_int32_int32))
+
+    matching_dict = {2: 4, 3: 6, 4: 8}
+    self.assertMapIterEquals(msg.map_int32_int32.items(), matching_dict)
+
+  def testMapIterationClearMessage(self):
+    # Iterator needs to work even if message and map are deleted.
+    msg = map_unittest_pb2.TestMap()
+
+    msg.map_int32_int32[2] = 4
+    msg.map_int32_int32[3] = 6
+    msg.map_int32_int32[4] = 8
+
+    it = msg.map_int32_int32.items()
+    del msg
+
+    matching_dict = {2: 4, 3: 6, 4: 8}
+    self.assertMapIterEquals(it, matching_dict)
+
+  def testMapConstruction(self):
+    msg = map_unittest_pb2.TestMap(map_int32_int32={1: 2, 3: 4})
+    self.assertEqual(2, msg.map_int32_int32[1])
+    self.assertEqual(4, msg.map_int32_int32[3])
+
+    msg = map_unittest_pb2.TestMap(
+        map_int32_foreign_message={3: unittest_pb2.ForeignMessage(c=5)})
+    self.assertEqual(5, msg.map_int32_foreign_message[3].c)
+
+  def testMapValidAfterFieldCleared(self):
+    # Map needs to work even if field is cleared.
+    # For the C++ implementation this tests the correctness of
+    # ScalarMapContainer::Release()
+    msg = map_unittest_pb2.TestMap()
+    int32_map = msg.map_int32_int32
+
+    int32_map[2] = 4
+    int32_map[3] = 6
+    int32_map[4] = 8
+
+    msg.ClearField('map_int32_int32')
+    self.assertEqual(b'', msg.SerializeToString())
+    matching_dict = {2: 4, 3: 6, 4: 8}
+    self.assertMapIterEquals(int32_map.items(), matching_dict)
+
+  def testMessageMapValidAfterFieldCleared(self):
+    # Map needs to work even if field is cleared.
+    # For the C++ implementation this tests the correctness of
+    # ScalarMapContainer::Release()
+    msg = map_unittest_pb2.TestMap()
+    int32_foreign_message = msg.map_int32_foreign_message
+
+    int32_foreign_message[2].c = 5
+
+    msg.ClearField('map_int32_foreign_message')
+    self.assertEqual(b'', msg.SerializeToString())
+    self.assertTrue(2 in int32_foreign_message.keys())
+
+  def testMapIterInvalidatedByClearField(self):
+    # Map iterator is invalidated when field is cleared.
+    # But this case does need to not crash the interpreter.
+    # For the C++ implementation this tests the correctness of
+    # ScalarMapContainer::Release()
+    msg = map_unittest_pb2.TestMap()
+
+    it = iter(msg.map_int32_int32)
+
+    msg.ClearField('map_int32_int32')
+    with self.assertRaises(RuntimeError):
+      for _ in it:
+        pass
+
+    it = iter(msg.map_int32_foreign_message)
+    msg.ClearField('map_int32_foreign_message')
+    with self.assertRaises(RuntimeError):
+      for _ in it:
+        pass
+
+  def testMapDelete(self):
+    msg = map_unittest_pb2.TestMap()
+
+    self.assertEqual(0, len(msg.map_int32_int32))
+
+    msg.map_int32_int32[4] = 6
+    self.assertEqual(1, len(msg.map_int32_int32))
+
+    with self.assertRaises(KeyError):
+      del msg.map_int32_int32[88]
+
+    del msg.map_int32_int32[4]
+    self.assertEqual(0, len(msg.map_int32_int32))
+
+  def testMapsAreMapping(self):
+    msg = map_unittest_pb2.TestMap()
+    self.assertIsInstance(msg.map_int32_int32, collections.Mapping)
+    self.assertIsInstance(msg.map_int32_int32, collections.MutableMapping)
+    self.assertIsInstance(msg.map_int32_foreign_message, collections.Mapping)
+    self.assertIsInstance(msg.map_int32_foreign_message,
+                          collections.MutableMapping)
+
+  def testMapFindInitializationErrorsSmokeTest(self):
+    msg = map_unittest_pb2.TestMap()
+    msg.map_string_string['abc'] = '123'
+    msg.map_int32_int32[35] = 64
+    msg.map_string_foreign_message['foo'].c = 5
+    self.assertEqual(0, len(msg.FindInitializationErrors()))
+
+  def testAnyMessage(self):
+    # Creates and sets message.
+    msg = any_test_pb2.TestAny()
+    msg_descriptor = msg.DESCRIPTOR
+    all_types = unittest_pb2.TestAllTypes()
+    all_descriptor = all_types.DESCRIPTOR
+    all_types.repeated_string.append(u'\u00fc\ua71f')
+    # Packs to Any.
+    msg.value.Pack(all_types)
+    self.assertEqual(msg.value.type_url,
+                     'type.googleapis.com/%s' % all_descriptor.full_name)
+    self.assertEqual(msg.value.value,
+                     all_types.SerializeToString())
+    # Tests Is() method.
+    self.assertTrue(msg.value.Is(all_descriptor))
+    self.assertFalse(msg.value.Is(msg_descriptor))
+    # Unpacks Any.
+    unpacked_message = unittest_pb2.TestAllTypes()
+    self.assertTrue(msg.value.Unpack(unpacked_message))
+    self.assertEqual(all_types, unpacked_message)
+    # Unpacks to different type.
+    self.assertFalse(msg.value.Unpack(msg))
+    # Only Any messages have Pack method.
+    try:
+      msg.Pack(all_types)
+    except AttributeError:
+      pass
+    else:
+      raise AttributeError('%s should not have Pack method.' %
+                           msg_descriptor.full_name)
+
+
+
+class ValidTypeNamesTest(unittest.TestCase):
+
+  def assertImportFromName(self, msg, base_name):
+    # Parse <type 'module.class_name'> to extra 'some.name' as a string.
+    tp_name = str(type(msg)).split("'")[1]
+    valid_names = ('Repeated%sContainer' % base_name,
+                   'Repeated%sFieldContainer' % base_name)
+    self.assertTrue(any(tp_name.endswith(v) for v in valid_names),
+                    '%r does end with any of %r' % (tp_name, valid_names))
+
+    parts = tp_name.split('.')
+    class_name = parts[-1]
+    module_name = '.'.join(parts[:-1])
+    __import__(module_name, fromlist=[class_name])
+
+  def testTypeNamesCanBeImported(self):
+    # If import doesn't work, pickling won't work either.
+    pb = unittest_pb2.TestAllTypes()
+    self.assertImportFromName(pb.repeated_int32, 'Scalar')
+    self.assertImportFromName(pb.repeated_nested_message, 'Composite')
+
+class PackedFieldTest(unittest.TestCase):
+
+  def setMessage(self, message):
+    message.repeated_int32.append(1)
+    message.repeated_int64.append(1)
+    message.repeated_uint32.append(1)
+    message.repeated_uint64.append(1)
+    message.repeated_sint32.append(1)
+    message.repeated_sint64.append(1)
+    message.repeated_fixed32.append(1)
+    message.repeated_fixed64.append(1)
+    message.repeated_sfixed32.append(1)
+    message.repeated_sfixed64.append(1)
+    message.repeated_float.append(1.0)
+    message.repeated_double.append(1.0)
+    message.repeated_bool.append(True)
+    message.repeated_nested_enum.append(1)
+
+  def testPackedFields(self):
+    message = packed_field_test_pb2.TestPackedTypes()
+    self.setMessage(message)
+    golden_data = (b'\x0A\x01\x01'
+                   b'\x12\x01\x01'
+                   b'\x1A\x01\x01'
+                   b'\x22\x01\x01'
+                   b'\x2A\x01\x02'
+                   b'\x32\x01\x02'
+                   b'\x3A\x04\x01\x00\x00\x00'
+                   b'\x42\x08\x01\x00\x00\x00\x00\x00\x00\x00'
+                   b'\x4A\x04\x01\x00\x00\x00'
+                   b'\x52\x08\x01\x00\x00\x00\x00\x00\x00\x00'
+                   b'\x5A\x04\x00\x00\x80\x3f'
+                   b'\x62\x08\x00\x00\x00\x00\x00\x00\xf0\x3f'
+                   b'\x6A\x01\x01'
+                   b'\x72\x01\x01')
+    self.assertEqual(golden_data, message.SerializeToString())
+
+  def testUnpackedFields(self):
+    message = packed_field_test_pb2.TestUnpackedTypes()
+    self.setMessage(message)
+    golden_data = (b'\x08\x01'
+                   b'\x10\x01'
+                   b'\x18\x01'
+                   b'\x20\x01'
+                   b'\x28\x02'
+                   b'\x30\x02'
+                   b'\x3D\x01\x00\x00\x00'
+                   b'\x41\x01\x00\x00\x00\x00\x00\x00\x00'
+                   b'\x4D\x01\x00\x00\x00'
+                   b'\x51\x01\x00\x00\x00\x00\x00\x00\x00'
+                   b'\x5D\x00\x00\x80\x3f'
+                   b'\x61\x00\x00\x00\x00\x00\x00\xf0\x3f'
+                   b'\x68\x01'
+                   b'\x70\x01')
+    self.assertEqual(golden_data, message.SerializeToString())
+
+if __name__ == '__main__':
+  unittest.main()
diff --git a/python/google/protobuf/internal/missing_enum_values.proto b/python/google/protobuf/internal/missing_enum_values.proto
new file mode 100644
index 0000000..1850be5
--- /dev/null
+++ b/python/google/protobuf/internal/missing_enum_values.proto
@@ -0,0 +1,56 @@
+// Protocol Buffers - Google's data interchange format
+// Copyright 2008 Google Inc.  All rights reserved.
+// https://developers.google.com/protocol-buffers/
+//
+// Redistribution and use in source and binary forms, with or without
+// modification, are permitted provided that the following conditions are
+// met:
+//
+//     * Redistributions of source code must retain the above copyright
+// notice, this list of conditions and the following disclaimer.
+//     * Redistributions in binary form must reproduce the above
+// copyright notice, this list of conditions and the following disclaimer
+// in the documentation and/or other materials provided with the
+// distribution.
+//     * Neither the name of Google Inc. nor the names of its
+// contributors may be used to endorse or promote products derived from
+// this software without specific prior written permission.
+//
+// THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+// "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
+// LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
+// A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
+// OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
+// SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
+// LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
+// DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
+// THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
+// (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+// OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+
+syntax = "proto2";
+
+package google.protobuf.python.internal;
+
+message TestEnumValues {
+  enum NestedEnum {
+    ZERO = 0;
+    ONE = 1;
+  }
+  optional NestedEnum optional_nested_enum = 1;
+  repeated NestedEnum repeated_nested_enum = 2;
+  repeated NestedEnum packed_nested_enum = 3 [packed = true];
+}
+
+message TestMissingEnumValues {
+  enum NestedEnum {
+    TWO = 2;
+  }
+  optional NestedEnum optional_nested_enum = 1;
+  repeated NestedEnum repeated_nested_enum = 2;
+  repeated NestedEnum packed_nested_enum = 3 [packed = true];
+}
+
+message JustString {
+  required string dummy = 1;
+}
diff --git a/python/google/protobuf/internal/more_extensions.proto b/python/google/protobuf/internal/more_extensions.proto
new file mode 100644
index 0000000..78f1467
--- /dev/null
+++ b/python/google/protobuf/internal/more_extensions.proto
@@ -0,0 +1,59 @@
+// Protocol Buffers - Google's data interchange format
+// Copyright 2008 Google Inc.  All rights reserved.
+// https://developers.google.com/protocol-buffers/
+//
+// Redistribution and use in source and binary forms, with or without
+// modification, are permitted provided that the following conditions are
+// met:
+//
+//     * Redistributions of source code must retain the above copyright
+// notice, this list of conditions and the following disclaimer.
+//     * Redistributions in binary form must reproduce the above
+// copyright notice, this list of conditions and the following disclaimer
+// in the documentation and/or other materials provided with the
+// distribution.
+//     * Neither the name of Google Inc. nor the names of its
+// contributors may be used to endorse or promote products derived from
+// this software without specific prior written permission.
+//
+// THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+// "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
+// LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
+// A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
+// OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
+// SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
+// LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
+// DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
+// THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
+// (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+// OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+
+// Author: robinson@google.com (Will Robinson)
+
+syntax = "proto2";
+
+package google.protobuf.internal;
+
+
+message TopLevelMessage {
+  optional ExtendedMessage submessage = 1;
+}
+
+
+message ExtendedMessage {
+  extensions 1 to max;
+}
+
+
+message ForeignMessage {
+  optional int32 foreign_message_int = 1;
+}
+
+
+extend ExtendedMessage {
+  optional int32 optional_int_extension = 1;
+  optional ForeignMessage optional_message_extension = 2;
+
+  repeated int32 repeated_int_extension = 3;
+  repeated ForeignMessage repeated_message_extension = 4;
+}
diff --git a/python/google/protobuf/internal/more_extensions_dynamic.proto b/python/google/protobuf/internal/more_extensions_dynamic.proto
new file mode 100644
index 0000000..11f85ef
--- /dev/null
+++ b/python/google/protobuf/internal/more_extensions_dynamic.proto
@@ -0,0 +1,50 @@
+// Protocol Buffers - Google's data interchange format
+// Copyright 2008 Google Inc.  All rights reserved.
+// https://developers.google.com/protocol-buffers/
+//
+// Redistribution and use in source and binary forms, with or without
+// modification, are permitted provided that the following conditions are
+// met:
+//
+//     * Redistributions of source code must retain the above copyright
+// notice, this list of conditions and the following disclaimer.
+//     * Redistributions in binary form must reproduce the above
+// copyright notice, this list of conditions and the following disclaimer
+// in the documentation and/or other materials provided with the
+// distribution.
+//     * Neither the name of Google Inc. nor the names of its
+// contributors may be used to endorse or promote products derived from
+// this software without specific prior written permission.
+//
+// THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+// "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
+// LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
+// A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
+// OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
+// SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
+// LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
+// DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
+// THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
+// (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+// OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+
+// Author: jasonh@google.com (Jason Hsueh)
+//
+// This file is used to test a corner case in the CPP implementation where the
+// generated C++ type is available for the extendee, but the extension is
+// defined in a file whose C++ type is not in the binary.
+
+syntax = "proto2";
+
+import "google/protobuf/internal/more_extensions.proto";
+
+package google.protobuf.internal;
+
+message DynamicMessageType {
+  optional int32 a = 1;
+}
+
+extend ExtendedMessage {
+  optional int32 dynamic_int32_extension = 100;
+  optional DynamicMessageType dynamic_message_extension = 101;
+}
diff --git a/python/google/protobuf/internal/more_messages.proto b/python/google/protobuf/internal/more_messages.proto
new file mode 100644
index 0000000..2c6ab9e
--- /dev/null
+++ b/python/google/protobuf/internal/more_messages.proto
@@ -0,0 +1,52 @@
+// Protocol Buffers - Google's data interchange format
+// Copyright 2008 Google Inc.  All rights reserved.
+// https://developers.google.com/protocol-buffers/
+//
+// Redistribution and use in source and binary forms, with or without
+// modification, are permitted provided that the following conditions are
+// met:
+//
+//     * Redistributions of source code must retain the above copyright
+// notice, this list of conditions and the following disclaimer.
+//     * Redistributions in binary form must reproduce the above
+// copyright notice, this list of conditions and the following disclaimer
+// in the documentation and/or other materials provided with the
+// distribution.
+//     * Neither the name of Google Inc. nor the names of its
+// contributors may be used to endorse or promote products derived from
+// this software without specific prior written permission.
+//
+// THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+// "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
+// LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
+// A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
+// OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
+// SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
+// LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
+// DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
+// THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
+// (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+// OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+
+// Author: robinson@google.com (Will Robinson)
+
+syntax = "proto2";
+
+package google.protobuf.internal;
+
+// A message where tag numbers are listed out of order, to allow us to test our
+// canonicalization of serialized output, which should always be in tag order.
+// We also mix in some extensions for extra fun.
+message OutOfOrderFields {
+  optional   sint32 optional_sint32   =  5;
+  extensions 4 to 4;
+  optional   uint32 optional_uint32   =  3;
+  extensions 2 to 2;
+  optional    int32 optional_int32    =  1;
+};
+
+
+extend OutOfOrderFields {
+  optional   uint64 optional_uint64   =  4;
+  optional    int64 optional_int64    =  2;
+}
diff --git a/python/google/protobuf/internal/packed_field_test.proto b/python/google/protobuf/internal/packed_field_test.proto
new file mode 100644
index 0000000..0dfdc10
--- /dev/null
+++ b/python/google/protobuf/internal/packed_field_test.proto
@@ -0,0 +1,73 @@
+// Protocol Buffers - Google's data interchange format
+// Copyright 2008 Google Inc.  All rights reserved.
+// https://developers.google.com/protocol-buffers/
+//
+// Redistribution and use in source and binary forms, with or without
+// modification, are permitted provided that the following conditions are
+// met:
+//
+//     * Redistributions of source code must retain the above copyright
+// notice, this list of conditions and the following disclaimer.
+//     * Redistributions in binary form must reproduce the above
+// copyright notice, this list of conditions and the following disclaimer
+// in the documentation and/or other materials provided with the
+// distribution.
+//     * Neither the name of Google Inc. nor the names of its
+// contributors may be used to endorse or promote products derived from
+// this software without specific prior written permission.
+//
+// THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+// "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
+// LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
+// A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
+// OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
+// SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
+// LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
+// DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
+// THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
+// (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+// OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+
+syntax = "proto3";
+
+package google.protobuf.python.internal;
+
+message TestPackedTypes {
+  enum NestedEnum {
+    FOO = 0;
+    BAR = 1;
+    BAZ = 2;
+  }
+
+  repeated    int32 repeated_int32    =  1;
+  repeated    int64 repeated_int64    =  2;
+  repeated   uint32 repeated_uint32   =  3;
+  repeated   uint64 repeated_uint64   =  4;
+  repeated   sint32 repeated_sint32   =  5;
+  repeated   sint64 repeated_sint64   =  6;
+  repeated  fixed32 repeated_fixed32  =  7;
+  repeated  fixed64 repeated_fixed64  =  8;
+  repeated sfixed32 repeated_sfixed32 =  9;
+  repeated sfixed64 repeated_sfixed64 = 10;
+  repeated    float repeated_float    = 11;
+  repeated   double repeated_double   = 12;
+  repeated     bool repeated_bool     = 13;
+  repeated NestedEnum repeated_nested_enum = 14;
+}
+
+message TestUnpackedTypes {
+  repeated    int32 repeated_int32    =  1 [packed = false];
+  repeated    int64 repeated_int64    =  2 [packed = false];
+  repeated   uint32 repeated_uint32   =  3 [packed = false];
+  repeated   uint64 repeated_uint64   =  4 [packed = false];
+  repeated   sint32 repeated_sint32   =  5 [packed = false];
+  repeated   sint64 repeated_sint64   =  6 [packed = false];
+  repeated  fixed32 repeated_fixed32  =  7 [packed = false];
+  repeated  fixed64 repeated_fixed64  =  8 [packed = false];
+  repeated sfixed32 repeated_sfixed32 =  9 [packed = false];
+  repeated sfixed64 repeated_sfixed64 = 10 [packed = false];
+  repeated    float repeated_float    = 11 [packed = false];
+  repeated   double repeated_double   = 12 [packed = false];
+  repeated     bool repeated_bool     = 13 [packed = false];
+  repeated TestPackedTypes.NestedEnum repeated_nested_enum = 14 [packed = false];
+}
diff --git a/python/google/protobuf/internal/proto_builder_test.py b/python/google/protobuf/internal/proto_builder_test.py
new file mode 100644
index 0000000..822ad89
--- /dev/null
+++ b/python/google/protobuf/internal/proto_builder_test.py
@@ -0,0 +1,95 @@
+#! /usr/bin/env python
+#
+# Protocol Buffers - Google's data interchange format
+# Copyright 2008 Google Inc.  All rights reserved.
+# https://developers.google.com/protocol-buffers/
+#
+# Redistribution and use in source and binary forms, with or without
+# modification, are permitted provided that the following conditions are
+# met:
+#
+#     * Redistributions of source code must retain the above copyright
+# notice, this list of conditions and the following disclaimer.
+#     * Redistributions in binary form must reproduce the above
+# copyright notice, this list of conditions and the following disclaimer
+# in the documentation and/or other materials provided with the
+# distribution.
+#     * Neither the name of Google Inc. nor the names of its
+# contributors may be used to endorse or promote products derived from
+# this software without specific prior written permission.
+#
+# THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+# "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
+# LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
+# A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
+# OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
+# SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
+# LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
+# DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
+# THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
+# (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+# OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+
+"""Tests for google.protobuf.proto_builder."""
+
+try:
+    from collections import OrderedDict
+except ImportError:
+    from ordereddict import OrderedDict  #PY26
+try:
+  import unittest2 as unittest
+except ImportError:
+  import unittest
+from google.protobuf import descriptor_pb2
+from google.protobuf import descriptor_pool
+from google.protobuf import proto_builder
+from google.protobuf import text_format
+
+
+class ProtoBuilderTest(unittest.TestCase):
+
+  def setUp(self):
+    self.ordered_fields = OrderedDict([
+        ('foo', descriptor_pb2.FieldDescriptorProto.TYPE_INT64),
+        ('bar', descriptor_pb2.FieldDescriptorProto.TYPE_STRING),
+        ])
+    self._fields = dict(self.ordered_fields)
+
+  def testMakeSimpleProtoClass(self):
+    """Test that we can create a proto class."""
+    proto_cls = proto_builder.MakeSimpleProtoClass(
+        self._fields,
+        full_name='net.proto2.python.public.proto_builder_test.Test')
+    proto = proto_cls()
+    proto.foo = 12345
+    proto.bar = 'asdf'
+    self.assertMultiLineEqual(
+        'bar: "asdf"\nfoo: 12345\n', text_format.MessageToString(proto))
+
+  def testOrderedFields(self):
+    """Test that the field order is maintained when given an OrderedDict."""
+    proto_cls = proto_builder.MakeSimpleProtoClass(
+        self.ordered_fields,
+        full_name='net.proto2.python.public.proto_builder_test.OrderedTest')
+    proto = proto_cls()
+    proto.foo = 12345
+    proto.bar = 'asdf'
+    self.assertMultiLineEqual(
+        'foo: 12345\nbar: "asdf"\n', text_format.MessageToString(proto))
+
+  def testMakeSameProtoClassTwice(self):
+    """Test that the DescriptorPool is used."""
+    pool = descriptor_pool.DescriptorPool()
+    proto_cls1 = proto_builder.MakeSimpleProtoClass(
+        self._fields,
+        full_name='net.proto2.python.public.proto_builder_test.Test',
+        pool=pool)
+    proto_cls2 = proto_builder.MakeSimpleProtoClass(
+        self._fields,
+        full_name='net.proto2.python.public.proto_builder_test.Test',
+        pool=pool)
+    self.assertIs(proto_cls1.DESCRIPTOR, proto_cls2.DESCRIPTOR)
+
+
+if __name__ == '__main__':
+  unittest.main()
diff --git a/python/google/protobuf/internal/python_message.py b/python/google/protobuf/internal/python_message.py
new file mode 100755
index 0000000..87f6066
--- /dev/null
+++ b/python/google/protobuf/internal/python_message.py
@@ -0,0 +1,1520 @@
+# Protocol Buffers - Google's data interchange format
+# Copyright 2008 Google Inc.  All rights reserved.
+# https://developers.google.com/protocol-buffers/
+#
+# Redistribution and use in source and binary forms, with or without
+# modification, are permitted provided that the following conditions are
+# met:
+#
+#     * Redistributions of source code must retain the above copyright
+# notice, this list of conditions and the following disclaimer.
+#     * Redistributions in binary form must reproduce the above
+# copyright notice, this list of conditions and the following disclaimer
+# in the documentation and/or other materials provided with the
+# distribution.
+#     * Neither the name of Google Inc. nor the names of its
+# contributors may be used to endorse or promote products derived from
+# this software without specific prior written permission.
+#
+# THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+# "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
+# LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
+# A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
+# OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
+# SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
+# LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
+# DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
+# THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
+# (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+# OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+
+# This code is meant to work on Python 2.4 and above only.
+#
+# TODO(robinson): Helpers for verbose, common checks like seeing if a
+# descriptor's cpp_type is CPPTYPE_MESSAGE.
+
+"""Contains a metaclass and helper functions used to create
+protocol message classes from Descriptor objects at runtime.
+
+Recall that a metaclass is the "type" of a class.
+(A class is to a metaclass what an instance is to a class.)
+
+In this case, we use the GeneratedProtocolMessageType metaclass
+to inject all the useful functionality into the classes
+output by the protocol compiler at compile-time.
+
+The upshot of all this is that the real implementation
+details for ALL pure-Python protocol buffers are *here in
+this file*.
+"""
+
+__author__ = 'robinson@google.com (Will Robinson)'
+
+from io import BytesIO
+import sys
+import struct
+import weakref
+
+import six
+import six.moves.copyreg as copyreg
+
+# We use "as" to avoid name collisions with variables.
+from google.protobuf.internal import containers
+from google.protobuf.internal import decoder
+from google.protobuf.internal import encoder
+from google.protobuf.internal import enum_type_wrapper
+from google.protobuf.internal import message_listener as message_listener_mod
+from google.protobuf.internal import type_checkers
+from google.protobuf.internal import well_known_types
+from google.protobuf.internal import wire_format
+from google.protobuf import descriptor as descriptor_mod
+from google.protobuf import message as message_mod
+from google.protobuf import symbol_database
+from google.protobuf import text_format
+
+_FieldDescriptor = descriptor_mod.FieldDescriptor
+_AnyFullTypeName = 'google.protobuf.Any'
+
+
+class GeneratedProtocolMessageType(type):
+
+  """Metaclass for protocol message classes created at runtime from Descriptors.
+
+  We add implementations for all methods described in the Message class.  We
+  also create properties to allow getting/setting all fields in the protocol
+  message.  Finally, we create slots to prevent users from accidentally
+  "setting" nonexistent fields in the protocol message, which then wouldn't get
+  serialized / deserialized properly.
+
+  The protocol compiler currently uses this metaclass to create protocol
+  message classes at runtime.  Clients can also manually create their own
+  classes at runtime, as in this example:
+
+  mydescriptor = Descriptor(.....)
+  class MyProtoClass(Message):
+    __metaclass__ = GeneratedProtocolMessageType
+    DESCRIPTOR = mydescriptor
+  myproto_instance = MyProtoClass()
+  myproto.foo_field = 23
+  ...
+
+  The above example will not work for nested types. If you wish to include them,
+  use reflection.MakeClass() instead of manually instantiating the class in
+  order to create the appropriate class structure.
+  """
+
+  # Must be consistent with the protocol-compiler code in
+  # proto2/compiler/internal/generator.*.
+  _DESCRIPTOR_KEY = 'DESCRIPTOR'
+
+  def __new__(cls, name, bases, dictionary):
+    """Custom allocation for runtime-generated class types.
+
+    We override __new__ because this is apparently the only place
+    where we can meaningfully set __slots__ on the class we're creating(?).
+    (The interplay between metaclasses and slots is not very well-documented).
+
+    Args:
+      name: Name of the class (ignored, but required by the
+        metaclass protocol).
+      bases: Base classes of the class we're constructing.
+        (Should be message.Message).  We ignore this field, but
+        it's required by the metaclass protocol
+      dictionary: The class dictionary of the class we're
+        constructing.  dictionary[_DESCRIPTOR_KEY] must contain
+        a Descriptor object describing this protocol message
+        type.
+
+    Returns:
+      Newly-allocated class.
+    """
+    descriptor = dictionary[GeneratedProtocolMessageType._DESCRIPTOR_KEY]
+    if descriptor.full_name in well_known_types.WKTBASES:
+      bases += (well_known_types.WKTBASES[descriptor.full_name],)
+    _AddClassAttributesForNestedExtensions(descriptor, dictionary)
+    _AddSlots(descriptor, dictionary)
+
+    superclass = super(GeneratedProtocolMessageType, cls)
+    new_class = superclass.__new__(cls, name, bases, dictionary)
+    return new_class
+
+  def __init__(cls, name, bases, dictionary):
+    """Here we perform the majority of our work on the class.
+    We add enum getters, an __init__ method, implementations
+    of all Message methods, and properties for all fields
+    in the protocol type.
+
+    Args:
+      name: Name of the class (ignored, but required by the
+        metaclass protocol).
+      bases: Base classes of the class we're constructing.
+        (Should be message.Message).  We ignore this field, but
+        it's required by the metaclass protocol
+      dictionary: The class dictionary of the class we're
+        constructing.  dictionary[_DESCRIPTOR_KEY] must contain
+        a Descriptor object describing this protocol message
+        type.
+    """
+    descriptor = dictionary[GeneratedProtocolMessageType._DESCRIPTOR_KEY]
+    cls._decoders_by_tag = {}
+    cls._extensions_by_name = {}
+    cls._extensions_by_number = {}
+    if (descriptor.has_options and
+        descriptor.GetOptions().message_set_wire_format):
+      cls._decoders_by_tag[decoder.MESSAGE_SET_ITEM_TAG] = (
+          decoder.MessageSetItemDecoder(cls._extensions_by_number), None)
+
+    # Attach stuff to each FieldDescriptor for quick lookup later on.
+    for field in descriptor.fields:
+      _AttachFieldHelpers(cls, field)
+
+    descriptor._concrete_class = cls  # pylint: disable=protected-access
+    _AddEnumValues(descriptor, cls)
+    _AddInitMethod(descriptor, cls)
+    _AddPropertiesForFields(descriptor, cls)
+    _AddPropertiesForExtensions(descriptor, cls)
+    _AddStaticMethods(cls)
+    _AddMessageMethods(descriptor, cls)
+    _AddPrivateHelperMethods(descriptor, cls)
+    copyreg.pickle(cls, lambda obj: (cls, (), obj.__getstate__()))
+
+    superclass = super(GeneratedProtocolMessageType, cls)
+    superclass.__init__(name, bases, dictionary)
+
+
+# Stateless helpers for GeneratedProtocolMessageType below.
+# Outside clients should not access these directly.
+#
+# I opted not to make any of these methods on the metaclass, to make it more
+# clear that I'm not really using any state there and to keep clients from
+# thinking that they have direct access to these construction helpers.
+
+
+def _PropertyName(proto_field_name):
+  """Returns the name of the public property attribute which
+  clients can use to get and (in some cases) set the value
+  of a protocol message field.
+
+  Args:
+    proto_field_name: The protocol message field name, exactly
+      as it appears (or would appear) in a .proto file.
+  """
+  # TODO(robinson): Escape Python keywords (e.g., yield), and test this support.
+  # nnorwitz makes my day by writing:
+  # """
+  # FYI.  See the keyword module in the stdlib. This could be as simple as:
+  #
+  # if keyword.iskeyword(proto_field_name):
+  #   return proto_field_name + "_"
+  # return proto_field_name
+  # """
+  # Kenton says:  The above is a BAD IDEA.  People rely on being able to use
+  #   getattr() and setattr() to reflectively manipulate field values.  If we
+  #   rename the properties, then every such user has to also make sure to apply
+  #   the same transformation.  Note that currently if you name a field "yield",
+  #   you can still access it just fine using getattr/setattr -- it's not even
+  #   that cumbersome to do so.
+  # TODO(kenton):  Remove this method entirely if/when everyone agrees with my
+  #   position.
+  return proto_field_name
+
+
+def _VerifyExtensionHandle(message, extension_handle):
+  """Verify that the given extension handle is valid."""
+
+  if not isinstance(extension_handle, _FieldDescriptor):
+    raise KeyError('HasExtension() expects an extension handle, got: %s' %
+                   extension_handle)
+
+  if not extension_handle.is_extension:
+    raise KeyError('"%s" is not an extension.' % extension_handle.full_name)
+
+  if not extension_handle.containing_type:
+    raise KeyError('"%s" is missing a containing_type.'
+                   % extension_handle.full_name)
+
+  if extension_handle.containing_type is not message.DESCRIPTOR:
+    raise KeyError('Extension "%s" extends message type "%s", but this '
+                   'message is of type "%s".' %
+                   (extension_handle.full_name,
+                    extension_handle.containing_type.full_name,
+                    message.DESCRIPTOR.full_name))
+
+
+def _AddSlots(message_descriptor, dictionary):
+  """Adds a __slots__ entry to dictionary, containing the names of all valid
+  attributes for this message type.
+
+  Args:
+    message_descriptor: A Descriptor instance describing this message type.
+    dictionary: Class dictionary to which we'll add a '__slots__' entry.
+  """
+  dictionary['__slots__'] = ['_cached_byte_size',
+                             '_cached_byte_size_dirty',
+                             '_fields',
+                             '_unknown_fields',
+                             '_is_present_in_parent',
+                             '_listener',
+                             '_listener_for_children',
+                             '__weakref__',
+                             '_oneofs']
+
+
+def _IsMessageSetExtension(field):
+  return (field.is_extension and
+          field.containing_type.has_options and
+          field.containing_type.GetOptions().message_set_wire_format and
+          field.type == _FieldDescriptor.TYPE_MESSAGE and
+          field.label == _FieldDescriptor.LABEL_OPTIONAL)
+
+
+def _IsMapField(field):
+  return (field.type == _FieldDescriptor.TYPE_MESSAGE and
+          field.message_type.has_options and
+          field.message_type.GetOptions().map_entry)
+
+
+def _IsMessageMapField(field):
+  value_type = field.message_type.fields_by_name["value"]
+  return value_type.cpp_type == _FieldDescriptor.CPPTYPE_MESSAGE
+
+
+def _AttachFieldHelpers(cls, field_descriptor):
+  is_repeated = (field_descriptor.label == _FieldDescriptor.LABEL_REPEATED)
+  is_packable = (is_repeated and
+                 wire_format.IsTypePackable(field_descriptor.type))
+  if not is_packable:
+    is_packed = False
+  elif field_descriptor.containing_type.syntax == "proto2":
+    is_packed = (field_descriptor.has_options and
+                field_descriptor.GetOptions().packed)
+  else:
+    has_packed_false = (field_descriptor.has_options and
+                        field_descriptor.GetOptions().HasField("packed") and
+                        field_descriptor.GetOptions().packed == False)
+    is_packed = not has_packed_false
+  is_map_entry = _IsMapField(field_descriptor)
+
+  if is_map_entry:
+    field_encoder = encoder.MapEncoder(field_descriptor)
+    sizer = encoder.MapSizer(field_descriptor)
+  elif _IsMessageSetExtension(field_descriptor):
+    field_encoder = encoder.MessageSetItemEncoder(field_descriptor.number)
+    sizer = encoder.MessageSetItemSizer(field_descriptor.number)
+  else:
+    field_encoder = type_checkers.TYPE_TO_ENCODER[field_descriptor.type](
+        field_descriptor.number, is_repeated, is_packed)
+    sizer = type_checkers.TYPE_TO_SIZER[field_descriptor.type](
+        field_descriptor.number, is_repeated, is_packed)
+
+  field_descriptor._encoder = field_encoder
+  field_descriptor._sizer = sizer
+  field_descriptor._default_constructor = _DefaultValueConstructorForField(
+      field_descriptor)
+
+  def AddDecoder(wiretype, is_packed):
+    tag_bytes = encoder.TagBytes(field_descriptor.number, wiretype)
+    decode_type = field_descriptor.type
+    if (decode_type == _FieldDescriptor.TYPE_ENUM and
+        type_checkers.SupportsOpenEnums(field_descriptor)):
+      decode_type = _FieldDescriptor.TYPE_INT32
+
+    oneof_descriptor = None
+    if field_descriptor.containing_oneof is not None:
+      oneof_descriptor = field_descriptor
+
+    if is_map_entry:
+      is_message_map = _IsMessageMapField(field_descriptor)
+
+      field_decoder = decoder.MapDecoder(
+          field_descriptor, _GetInitializeDefaultForMap(field_descriptor),
+          is_message_map)
+    else:
+      field_decoder = type_checkers.TYPE_TO_DECODER[decode_type](
+              field_descriptor.number, is_repeated, is_packed,
+              field_descriptor, field_descriptor._default_constructor)
+
+    cls._decoders_by_tag[tag_bytes] = (field_decoder, oneof_descriptor)
+
+  AddDecoder(type_checkers.FIELD_TYPE_TO_WIRE_TYPE[field_descriptor.type],
+             False)
+
+  if is_repeated and wire_format.IsTypePackable(field_descriptor.type):
+    # To support wire compatibility of adding packed = true, add a decoder for
+    # packed values regardless of the field's options.
+    AddDecoder(wire_format.WIRETYPE_LENGTH_DELIMITED, True)
+
+
+def _AddClassAttributesForNestedExtensions(descriptor, dictionary):
+  extension_dict = descriptor.extensions_by_name
+  for extension_name, extension_field in extension_dict.items():
+    assert extension_name not in dictionary
+    dictionary[extension_name] = extension_field
+
+
+def _AddEnumValues(descriptor, cls):
+  """Sets class-level attributes for all enum fields defined in this message.
+
+  Also exporting a class-level object that can name enum values.
+
+  Args:
+    descriptor: Descriptor object for this message type.
+    cls: Class we're constructing for this message type.
+  """
+  for enum_type in descriptor.enum_types:
+    setattr(cls, enum_type.name, enum_type_wrapper.EnumTypeWrapper(enum_type))
+    for enum_value in enum_type.values:
+      setattr(cls, enum_value.name, enum_value.number)
+
+
+def _GetInitializeDefaultForMap(field):
+  if field.label != _FieldDescriptor.LABEL_REPEATED:
+    raise ValueError('map_entry set on non-repeated field %s' % (
+        field.name))
+  fields_by_name = field.message_type.fields_by_name
+  key_checker = type_checkers.GetTypeChecker(fields_by_name['key'])
+
+  value_field = fields_by_name['value']
+  if _IsMessageMapField(field):
+    def MakeMessageMapDefault(message):
+      return containers.MessageMap(
+          message._listener_for_children, value_field.message_type, key_checker)
+    return MakeMessageMapDefault
+  else:
+    value_checker = type_checkers.GetTypeChecker(value_field)
+    def MakePrimitiveMapDefault(message):
+      return containers.ScalarMap(
+          message._listener_for_children, key_checker, value_checker)
+    return MakePrimitiveMapDefault
+
+def _DefaultValueConstructorForField(field):
+  """Returns a function which returns a default value for a field.
+
+  Args:
+    field: FieldDescriptor object for this field.
+
+  The returned function has one argument:
+    message: Message instance containing this field, or a weakref proxy
+      of same.
+
+  That function in turn returns a default value for this field.  The default
+    value may refer back to |message| via a weak reference.
+  """
+
+  if _IsMapField(field):
+    return _GetInitializeDefaultForMap(field)
+
+  if field.label == _FieldDescriptor.LABEL_REPEATED:
+    if field.has_default_value and field.default_value != []:
+      raise ValueError('Repeated field default value not empty list: %s' % (
+          field.default_value))
+    if field.cpp_type == _FieldDescriptor.CPPTYPE_MESSAGE:
+      # We can't look at _concrete_class yet since it might not have
+      # been set.  (Depends on order in which we initialize the classes).
+      message_type = field.message_type
+      def MakeRepeatedMessageDefault(message):
+        return containers.RepeatedCompositeFieldContainer(
+            message._listener_for_children, field.message_type)
+      return MakeRepeatedMessageDefault
+    else:
+      type_checker = type_checkers.GetTypeChecker(field)
+      def MakeRepeatedScalarDefault(message):
+        return containers.RepeatedScalarFieldContainer(
+            message._listener_for_children, type_checker)
+      return MakeRepeatedScalarDefault
+
+  if field.cpp_type == _FieldDescriptor.CPPTYPE_MESSAGE:
+    # _concrete_class may not yet be initialized.
+    message_type = field.message_type
+    def MakeSubMessageDefault(message):
+      result = message_type._concrete_class()
+      result._SetListener(
+          _OneofListener(message, field)
+          if field.containing_oneof is not None
+          else message._listener_for_children)
+      return result
+    return MakeSubMessageDefault
+
+  def MakeScalarDefault(message):
+    # TODO(protobuf-team): This may be broken since there may not be
+    # default_value.  Combine with has_default_value somehow.
+    return field.default_value
+  return MakeScalarDefault
+
+
+def _ReraiseTypeErrorWithFieldName(message_name, field_name):
+  """Re-raise the currently-handled TypeError with the field name added."""
+  exc = sys.exc_info()[1]
+  if len(exc.args) == 1 and type(exc) is TypeError:
+    # simple TypeError; add field name to exception message
+    exc = TypeError('%s for field %s.%s' % (str(exc), message_name, field_name))
+
+  # re-raise possibly-amended exception with original traceback:
+  six.reraise(type(exc), exc, sys.exc_info()[2])
+
+
+def _AddInitMethod(message_descriptor, cls):
+  """Adds an __init__ method to cls."""
+
+  def _GetIntegerEnumValue(enum_type, value):
+    """Convert a string or integer enum value to an integer.
+
+    If the value is a string, it is converted to the enum value in
+    enum_type with the same name.  If the value is not a string, it's
+    returned as-is.  (No conversion or bounds-checking is done.)
+    """
+    if isinstance(value, six.string_types):
+      try:
+        return enum_type.values_by_name[value].number
+      except KeyError:
+        raise ValueError('Enum type %s: unknown label "%s"' % (
+            enum_type.full_name, value))
+    return value
+
+  def init(self, **kwargs):
+    self._cached_byte_size = 0
+    self._cached_byte_size_dirty = len(kwargs) > 0
+    self._fields = {}
+    # Contains a mapping from oneof field descriptors to the descriptor
+    # of the currently set field in that oneof field.
+    self._oneofs = {}
+
+    # _unknown_fields is () when empty for efficiency, and will be turned into
+    # a list if fields are added.
+    self._unknown_fields = ()
+    self._is_present_in_parent = False
+    self._listener = message_listener_mod.NullMessageListener()
+    self._listener_for_children = _Listener(self)
+    for field_name, field_value in kwargs.items():
+      field = _GetFieldByName(message_descriptor, field_name)
+      if field is None:
+        raise TypeError("%s() got an unexpected keyword argument '%s'" %
+                        (message_descriptor.name, field_name))
+      if field.label == _FieldDescriptor.LABEL_REPEATED:
+        copy = field._default_constructor(self)
+        if field.cpp_type == _FieldDescriptor.CPPTYPE_MESSAGE:  # Composite
+          if _IsMapField(field):
+            if _IsMessageMapField(field):
+              for key in field_value:
+                copy[key].MergeFrom(field_value[key])
+            else:
+              copy.update(field_value)
+          else:
+            for val in field_value:
+              if isinstance(val, dict):
+                copy.add(**val)
+              else:
+                copy.add().MergeFrom(val)
+        else:  # Scalar
+          if field.cpp_type == _FieldDescriptor.CPPTYPE_ENUM:
+            field_value = [_GetIntegerEnumValue(field.enum_type, val)
+                           for val in field_value]
+          copy.extend(field_value)
+        self._fields[field] = copy
+      elif field.cpp_type == _FieldDescriptor.CPPTYPE_MESSAGE:
+        copy = field._default_constructor(self)
+        new_val = field_value
+        if isinstance(field_value, dict):
+          new_val = field.message_type._concrete_class(**field_value)
+        try:
+          copy.MergeFrom(new_val)
+        except TypeError:
+          _ReraiseTypeErrorWithFieldName(message_descriptor.name, field_name)
+        self._fields[field] = copy
+      else:
+        if field.cpp_type == _FieldDescriptor.CPPTYPE_ENUM:
+          field_value = _GetIntegerEnumValue(field.enum_type, field_value)
+        try:
+          setattr(self, field_name, field_value)
+        except TypeError:
+          _ReraiseTypeErrorWithFieldName(message_descriptor.name, field_name)
+
+  init.__module__ = None
+  init.__doc__ = None
+  cls.__init__ = init
+
+
+def _GetFieldByName(message_descriptor, field_name):
+  """Returns a field descriptor by field name.
+
+  Args:
+    message_descriptor: A Descriptor describing all fields in message.
+    field_name: The name of the field to retrieve.
+  Returns:
+    The field descriptor associated with the field name.
+  """
+  try:
+    return message_descriptor.fields_by_name[field_name]
+  except KeyError:
+    raise ValueError('Protocol message %s has no "%s" field.' %
+                     (message_descriptor.name, field_name))
+
+
+def _AddPropertiesForFields(descriptor, cls):
+  """Adds properties for all fields in this protocol message type."""
+  for field in descriptor.fields:
+    _AddPropertiesForField(field, cls)
+
+  if descriptor.is_extendable:
+    # _ExtensionDict is just an adaptor with no state so we allocate a new one
+    # every time it is accessed.
+    cls.Extensions = property(lambda self: _ExtensionDict(self))
+
+
+def _AddPropertiesForField(field, cls):
+  """Adds a public property for a protocol message field.
+  Clients can use this property to get and (in the case
+  of non-repeated scalar fields) directly set the value
+  of a protocol message field.
+
+  Args:
+    field: A FieldDescriptor for this field.
+    cls: The class we're constructing.
+  """
+  # Catch it if we add other types that we should
+  # handle specially here.
+  assert _FieldDescriptor.MAX_CPPTYPE == 10
+
+  constant_name = field.name.upper() + "_FIELD_NUMBER"
+  setattr(cls, constant_name, field.number)
+
+  if field.label == _FieldDescriptor.LABEL_REPEATED:
+    _AddPropertiesForRepeatedField(field, cls)
+  elif field.cpp_type == _FieldDescriptor.CPPTYPE_MESSAGE:
+    _AddPropertiesForNonRepeatedCompositeField(field, cls)
+  else:
+    _AddPropertiesForNonRepeatedScalarField(field, cls)
+
+
+def _AddPropertiesForRepeatedField(field, cls):
+  """Adds a public property for a "repeated" protocol message field.  Clients
+  can use this property to get the value of the field, which will be either a
+  _RepeatedScalarFieldContainer or _RepeatedCompositeFieldContainer (see
+  below).
+
+  Note that when clients add values to these containers, we perform
+  type-checking in the case of repeated scalar fields, and we also set any
+  necessary "has" bits as a side-effect.
+
+  Args:
+    field: A FieldDescriptor for this field.
+    cls: The class we're constructing.
+  """
+  proto_field_name = field.name
+  property_name = _PropertyName(proto_field_name)
+
+  def getter(self):
+    field_value = self._fields.get(field)
+    if field_value is None:
+      # Construct a new object to represent this field.
+      field_value = field._default_constructor(self)
+
+      # Atomically check if another thread has preempted us and, if not, swap
+      # in the new object we just created.  If someone has preempted us, we
+      # take that object and discard ours.
+      # WARNING:  We are relying on setdefault() being atomic.  This is true
+      #   in CPython but we haven't investigated others.  This warning appears
+      #   in several other locations in this file.
+      field_value = self._fields.setdefault(field, field_value)
+    return field_value
+  getter.__module__ = None
+  getter.__doc__ = 'Getter for %s.' % proto_field_name
+
+  # We define a setter just so we can throw an exception with a more
+  # helpful error message.
+  def setter(self, new_value):
+    raise AttributeError('Assignment not allowed to repeated field '
+                         '"%s" in protocol message object.' % proto_field_name)
+
+  doc = 'Magic attribute generated for "%s" proto field.' % proto_field_name
+  setattr(cls, property_name, property(getter, setter, doc=doc))
+
+
+def _AddPropertiesForNonRepeatedScalarField(field, cls):
+  """Adds a public property for a nonrepeated, scalar protocol message field.
+  Clients can use this property to get and directly set the value of the field.
+  Note that when the client sets the value of a field by using this property,
+  all necessary "has" bits are set as a side-effect, and we also perform
+  type-checking.
+
+  Args:
+    field: A FieldDescriptor for this field.
+    cls: The class we're constructing.
+  """
+  proto_field_name = field.name
+  property_name = _PropertyName(proto_field_name)
+  type_checker = type_checkers.GetTypeChecker(field)
+  default_value = field.default_value
+  valid_values = set()
+  is_proto3 = field.containing_type.syntax == "proto3"
+
+  def getter(self):
+    # TODO(protobuf-team): This may be broken since there may not be
+    # default_value.  Combine with has_default_value somehow.
+    return self._fields.get(field, default_value)
+  getter.__module__ = None
+  getter.__doc__ = 'Getter for %s.' % proto_field_name
+
+  clear_when_set_to_default = is_proto3 and not field.containing_oneof
+
+  def field_setter(self, new_value):
+    # pylint: disable=protected-access
+    # Testing the value for truthiness captures all of the proto3 defaults
+    # (0, 0.0, enum 0, and False).
+    new_value = type_checker.CheckValue(new_value)
+    if clear_when_set_to_default and not new_value:
+      self._fields.pop(field, None)
+    else:
+      self._fields[field] = new_value
+    # Check _cached_byte_size_dirty inline to improve performance, since scalar
+    # setters are called frequently.
+    if not self._cached_byte_size_dirty:
+      self._Modified()
+
+  if field.containing_oneof:
+    def setter(self, new_value):
+      field_setter(self, new_value)
+      self._UpdateOneofState(field)
+  else:
+    setter = field_setter
+
+  setter.__module__ = None
+  setter.__doc__ = 'Setter for %s.' % proto_field_name
+
+  # Add a property to encapsulate the getter/setter.
+  doc = 'Magic attribute generated for "%s" proto field.' % proto_field_name
+  setattr(cls, property_name, property(getter, setter, doc=doc))
+
+
+def _AddPropertiesForNonRepeatedCompositeField(field, cls):
+  """Adds a public property for a nonrepeated, composite protocol message field.
+  A composite field is a "group" or "message" field.
+
+  Clients can use this property to get the value of the field, but cannot
+  assign to the property directly.
+
+  Args:
+    field: A FieldDescriptor for this field.
+    cls: The class we're constructing.
+  """
+  # TODO(robinson): Remove duplication with similar method
+  # for non-repeated scalars.
+  proto_field_name = field.name
+  property_name = _PropertyName(proto_field_name)
+
+  def getter(self):
+    field_value = self._fields.get(field)
+    if field_value is None:
+      # Construct a new object to represent this field.
+      field_value = field._default_constructor(self)
+
+      # Atomically check if another thread has preempted us and, if not, swap
+      # in the new object we just created.  If someone has preempted us, we
+      # take that object and discard ours.
+      # WARNING:  We are relying on setdefault() being atomic.  This is true
+      #   in CPython but we haven't investigated others.  This warning appears
+      #   in several other locations in this file.
+      field_value = self._fields.setdefault(field, field_value)
+    return field_value
+  getter.__module__ = None
+  getter.__doc__ = 'Getter for %s.' % proto_field_name
+
+  # We define a setter just so we can throw an exception with a more
+  # helpful error message.
+  def setter(self, new_value):
+    raise AttributeError('Assignment not allowed to composite field '
+                         '"%s" in protocol message object.' % proto_field_name)
+
+  # Add a property to encapsulate the getter.
+  doc = 'Magic attribute generated for "%s" proto field.' % proto_field_name
+  setattr(cls, property_name, property(getter, setter, doc=doc))
+
+
+def _AddPropertiesForExtensions(descriptor, cls):
+  """Adds properties for all fields in this protocol message type."""
+  extension_dict = descriptor.extensions_by_name
+  for extension_name, extension_field in extension_dict.items():
+    constant_name = extension_name.upper() + "_FIELD_NUMBER"
+    setattr(cls, constant_name, extension_field.number)
+
+
+def _AddStaticMethods(cls):
+  # TODO(robinson): This probably needs to be thread-safe(?)
+  def RegisterExtension(extension_handle):
+    extension_handle.containing_type = cls.DESCRIPTOR
+    _AttachFieldHelpers(cls, extension_handle)
+
+    # Try to insert our extension, failing if an extension with the same number
+    # already exists.
+    actual_handle = cls._extensions_by_number.setdefault(
+        extension_handle.number, extension_handle)
+    if actual_handle is not extension_handle:
+      raise AssertionError(
+          'Extensions "%s" and "%s" both try to extend message type "%s" with '
+          'field number %d.' %
+          (extension_handle.full_name, actual_handle.full_name,
+           cls.DESCRIPTOR.full_name, extension_handle.number))
+
+    cls._extensions_by_name[extension_handle.full_name] = extension_handle
+
+    handle = extension_handle  # avoid line wrapping
+    if _IsMessageSetExtension(handle):
+      # MessageSet extension.  Also register under type name.
+      cls._extensions_by_name[
+          extension_handle.message_type.full_name] = extension_handle
+
+  cls.RegisterExtension = staticmethod(RegisterExtension)
+
+  def FromString(s):
+    message = cls()
+    message.MergeFromString(s)
+    return message
+  cls.FromString = staticmethod(FromString)
+
+
+def _IsPresent(item):
+  """Given a (FieldDescriptor, value) tuple from _fields, return true if the
+  value should be included in the list returned by ListFields()."""
+
+  if item[0].label == _FieldDescriptor.LABEL_REPEATED:
+    return bool(item[1])
+  elif item[0].cpp_type == _FieldDescriptor.CPPTYPE_MESSAGE:
+    return item[1]._is_present_in_parent
+  else:
+    return True
+
+
+def _AddListFieldsMethod(message_descriptor, cls):
+  """Helper for _AddMessageMethods()."""
+
+  def ListFields(self):
+    all_fields = [item for item in self._fields.items() if _IsPresent(item)]
+    all_fields.sort(key = lambda item: item[0].number)
+    return all_fields
+
+  cls.ListFields = ListFields
+
+_Proto3HasError = 'Protocol message has no non-repeated submessage field "%s"'
+_Proto2HasError = 'Protocol message has no non-repeated field "%s"'
+
+def _AddHasFieldMethod(message_descriptor, cls):
+  """Helper for _AddMessageMethods()."""
+
+  is_proto3 = (message_descriptor.syntax == "proto3")
+  error_msg = _Proto3HasError if is_proto3 else _Proto2HasError
+
+  hassable_fields = {}
+  for field in message_descriptor.fields:
+    if field.label == _FieldDescriptor.LABEL_REPEATED:
+      continue
+    # For proto3, only submessages and fields inside a oneof have presence.
+    if (is_proto3 and field.cpp_type != _FieldDescriptor.CPPTYPE_MESSAGE and
+        not field.containing_oneof):
+      continue
+    hassable_fields[field.name] = field
+
+  if not is_proto3:
+    # Fields inside oneofs are never repeated (enforced by the compiler).
+    for oneof in message_descriptor.oneofs:
+      hassable_fields[oneof.name] = oneof
+
+  def HasField(self, field_name):
+    try:
+      field = hassable_fields[field_name]
+    except KeyError:
+      raise ValueError(error_msg % field_name)
+
+    if isinstance(field, descriptor_mod.OneofDescriptor):
+      try:
+        return HasField(self, self._oneofs[field].name)
+      except KeyError:
+        return False
+    else:
+      if field.cpp_type == _FieldDescriptor.CPPTYPE_MESSAGE:
+        value = self._fields.get(field)
+        return value is not None and value._is_present_in_parent
+      else:
+        return field in self._fields
+
+  cls.HasField = HasField
+
+
+def _AddClearFieldMethod(message_descriptor, cls):
+  """Helper for _AddMessageMethods()."""
+  def ClearField(self, field_name):
+    try:
+      field = message_descriptor.fields_by_name[field_name]
+    except KeyError:
+      try:
+        field = message_descriptor.oneofs_by_name[field_name]
+        if field in self._oneofs:
+          field = self._oneofs[field]
+        else:
+          return
+      except KeyError:
+        raise ValueError('Protocol message %s() has no "%s" field.' %
+                         (message_descriptor.name, field_name))
+
+    if field in self._fields:
+      # To match the C++ implementation, we need to invalidate iterators
+      # for map fields when ClearField() happens.
+      if hasattr(self._fields[field], 'InvalidateIterators'):
+        self._fields[field].InvalidateIterators()
+
+      # Note:  If the field is a sub-message, its listener will still point
+      #   at us.  That's fine, because the worst than can happen is that it
+      #   will call _Modified() and invalidate our byte size.  Big deal.
+      del self._fields[field]
+
+      if self._oneofs.get(field.containing_oneof, None) is field:
+        del self._oneofs[field.containing_oneof]
+
+    # Always call _Modified() -- even if nothing was changed, this is
+    # a mutating method, and thus calling it should cause the field to become
+    # present in the parent message.
+    self._Modified()
+
+  cls.ClearField = ClearField
+
+
+def _AddClearExtensionMethod(cls):
+  """Helper for _AddMessageMethods()."""
+  def ClearExtension(self, extension_handle):
+    _VerifyExtensionHandle(self, extension_handle)
+
+    # Similar to ClearField(), above.
+    if extension_handle in self._fields:
+      del self._fields[extension_handle]
+    self._Modified()
+  cls.ClearExtension = ClearExtension
+
+
+def _AddClearMethod(message_descriptor, cls):
+  """Helper for _AddMessageMethods()."""
+  def Clear(self):
+    # Clear fields.
+    self._fields = {}
+    self._unknown_fields = ()
+    self._oneofs = {}
+    self._Modified()
+  cls.Clear = Clear
+
+
+def _AddHasExtensionMethod(cls):
+  """Helper for _AddMessageMethods()."""
+  def HasExtension(self, extension_handle):
+    _VerifyExtensionHandle(self, extension_handle)
+    if extension_handle.label == _FieldDescriptor.LABEL_REPEATED:
+      raise KeyError('"%s" is repeated.' % extension_handle.full_name)
+
+    if extension_handle.cpp_type == _FieldDescriptor.CPPTYPE_MESSAGE:
+      value = self._fields.get(extension_handle)
+      return value is not None and value._is_present_in_parent
+    else:
+      return extension_handle in self._fields
+  cls.HasExtension = HasExtension
+
+def _InternalUnpackAny(msg):
+  """Unpacks Any message and returns the unpacked message.
+
+  This internal method is differnt from public Any Unpack method which takes
+  the target message as argument. _InternalUnpackAny method does not have
+  target message type and need to find the message type in descriptor pool.
+
+  Args:
+    msg: An Any message to be unpacked.
+
+  Returns:
+    The unpacked message.
+  """
+  type_url = msg.type_url
+  db = symbol_database.Default()
+
+  if not type_url:
+    return None
+
+  # TODO(haberman): For now we just strip the hostname.  Better logic will be
+  # required.
+  type_name = type_url.split("/")[-1]
+  descriptor = db.pool.FindMessageTypeByName(type_name)
+
+  if descriptor is None:
+    return None
+
+  message_class = db.GetPrototype(descriptor)
+  message = message_class()
+
+  message.ParseFromString(msg.value)
+  return message
+
+def _AddEqualsMethod(message_descriptor, cls):
+  """Helper for _AddMessageMethods()."""
+  def __eq__(self, other):
+    if (not isinstance(other, message_mod.Message) or
+        other.DESCRIPTOR != self.DESCRIPTOR):
+      return False
+
+    if self is other:
+      return True
+
+    if self.DESCRIPTOR.full_name == _AnyFullTypeName:
+      any_a = _InternalUnpackAny(self)
+      any_b = _InternalUnpackAny(other)
+      if any_a and any_b:
+        return any_a == any_b
+
+    if not self.ListFields() == other.ListFields():
+      return False
+
+    # Sort unknown fields because their order shouldn't affect equality test.
+    unknown_fields = list(self._unknown_fields)
+    unknown_fields.sort()
+    other_unknown_fields = list(other._unknown_fields)
+    other_unknown_fields.sort()
+
+    return unknown_fields == other_unknown_fields
+
+  cls.__eq__ = __eq__
+
+
+def _AddStrMethod(message_descriptor, cls):
+  """Helper for _AddMessageMethods()."""
+  def __str__(self):
+    return text_format.MessageToString(self)
+  cls.__str__ = __str__
+
+
+def _AddReprMethod(message_descriptor, cls):
+  """Helper for _AddMessageMethods()."""
+  def __repr__(self):
+    return text_format.MessageToString(self)
+  cls.__repr__ = __repr__
+
+
+def _AddUnicodeMethod(unused_message_descriptor, cls):
+  """Helper for _AddMessageMethods()."""
+
+  def __unicode__(self):
+    return text_format.MessageToString(self, as_utf8=True).decode('utf-8')
+  cls.__unicode__ = __unicode__
+
+
+def _AddSetListenerMethod(cls):
+  """Helper for _AddMessageMethods()."""
+  def SetListener(self, listener):
+    if listener is None:
+      self._listener = message_listener_mod.NullMessageListener()
+    else:
+      self._listener = listener
+  cls._SetListener = SetListener
+
+
+def _BytesForNonRepeatedElement(value, field_number, field_type):
+  """Returns the number of bytes needed to serialize a non-repeated element.
+  The returned byte count includes space for tag information and any
+  other additional space associated with serializing value.
+
+  Args:
+    value: Value we're serializing.
+    field_number: Field number of this value.  (Since the field number
+      is stored as part of a varint-encoded tag, this has an impact
+      on the total bytes required to serialize the value).
+    field_type: The type of the field.  One of the TYPE_* constants
+      within FieldDescriptor.
+  """
+  try:
+    fn = type_checkers.TYPE_TO_BYTE_SIZE_FN[field_type]
+    return fn(field_number, value)
+  except KeyError:
+    raise message_mod.EncodeError('Unrecognized field type: %d' % field_type)
+
+
+def _AddByteSizeMethod(message_descriptor, cls):
+  """Helper for _AddMessageMethods()."""
+
+  def ByteSize(self):
+    if not self._cached_byte_size_dirty:
+      return self._cached_byte_size
+
+    size = 0
+    for field_descriptor, field_value in self.ListFields():
+      size += field_descriptor._sizer(field_value)
+
+    for tag_bytes, value_bytes in self._unknown_fields:
+      size += len(tag_bytes) + len(value_bytes)
+
+    self._cached_byte_size = size
+    self._cached_byte_size_dirty = False
+    self._listener_for_children.dirty = False
+    return size
+
+  cls.ByteSize = ByteSize
+
+
+def _AddSerializeToStringMethod(message_descriptor, cls):
+  """Helper for _AddMessageMethods()."""
+
+  def SerializeToString(self):
+    # Check if the message has all of its required fields set.
+    errors = []
+    if not self.IsInitialized():
+      raise message_mod.EncodeError(
+          'Message %s is missing required fields: %s' % (
+          self.DESCRIPTOR.full_name, ','.join(self.FindInitializationErrors())))
+    return self.SerializePartialToString()
+  cls.SerializeToString = SerializeToString
+
+
+def _AddSerializePartialToStringMethod(message_descriptor, cls):
+  """Helper for _AddMessageMethods()."""
+
+  def SerializePartialToString(self):
+    out = BytesIO()
+    self._InternalSerialize(out.write)
+    return out.getvalue()
+  cls.SerializePartialToString = SerializePartialToString
+
+  def InternalSerialize(self, write_bytes):
+    for field_descriptor, field_value in self.ListFields():
+      field_descriptor._encoder(write_bytes, field_value)
+    for tag_bytes, value_bytes in self._unknown_fields:
+      write_bytes(tag_bytes)
+      write_bytes(value_bytes)
+  cls._InternalSerialize = InternalSerialize
+
+
+def _AddMergeFromStringMethod(message_descriptor, cls):
+  """Helper for _AddMessageMethods()."""
+  def MergeFromString(self, serialized):
+    length = len(serialized)
+    try:
+      if self._InternalParse(serialized, 0, length) != length:
+        # The only reason _InternalParse would return early is if it
+        # encountered an end-group tag.
+        raise message_mod.DecodeError('Unexpected end-group tag.')
+    except (IndexError, TypeError):
+      # Now ord(buf[p:p+1]) == ord('') gets TypeError.
+      raise message_mod.DecodeError('Truncated message.')
+    except struct.error as e:
+      raise message_mod.DecodeError(e)
+    return length   # Return this for legacy reasons.
+  cls.MergeFromString = MergeFromString
+
+  local_ReadTag = decoder.ReadTag
+  local_SkipField = decoder.SkipField
+  decoders_by_tag = cls._decoders_by_tag
+  is_proto3 = message_descriptor.syntax == "proto3"
+
+  def InternalParse(self, buffer, pos, end):
+    self._Modified()
+    field_dict = self._fields
+    unknown_field_list = self._unknown_fields
+    while pos != end:
+      (tag_bytes, new_pos) = local_ReadTag(buffer, pos)
+      field_decoder, field_desc = decoders_by_tag.get(tag_bytes, (None, None))
+      if field_decoder is None:
+        value_start_pos = new_pos
+        new_pos = local_SkipField(buffer, new_pos, end, tag_bytes)
+        if new_pos == -1:
+          return pos
+        if not is_proto3:
+          if not unknown_field_list:
+            unknown_field_list = self._unknown_fields = []
+          unknown_field_list.append(
+              (tag_bytes, buffer[value_start_pos:new_pos]))
+        pos = new_pos
+      else:
+        pos = field_decoder(buffer, new_pos, end, self, field_dict)
+        if field_desc:
+          self._UpdateOneofState(field_desc)
+    return pos
+  cls._InternalParse = InternalParse
+
+
+def _AddIsInitializedMethod(message_descriptor, cls):
+  """Adds the IsInitialized and FindInitializationError methods to the
+  protocol message class."""
+
+  required_fields = [field for field in message_descriptor.fields
+                           if field.label == _FieldDescriptor.LABEL_REQUIRED]
+
+  def IsInitialized(self, errors=None):
+    """Checks if all required fields of a message are set.
+
+    Args:
+      errors:  A list which, if provided, will be populated with the field
+               paths of all missing required fields.
+
+    Returns:
+      True iff the specified message has all required fields set.
+    """
+
+    # Performance is critical so we avoid HasField() and ListFields().
+
+    for field in required_fields:
+      if (field not in self._fields or
+          (field.cpp_type == _FieldDescriptor.CPPTYPE_MESSAGE and
+           not self._fields[field]._is_present_in_parent)):
+        if errors is not None:
+          errors.extend(self.FindInitializationErrors())
+        return False
+
+    for field, value in list(self._fields.items()):  # dict can change size!
+      if field.cpp_type == _FieldDescriptor.CPPTYPE_MESSAGE:
+        if field.label == _FieldDescriptor.LABEL_REPEATED:
+          if (field.message_type.has_options and
+              field.message_type.GetOptions().map_entry):
+            continue
+          for element in value:
+            if not element.IsInitialized():
+              if errors is not None:
+                errors.extend(self.FindInitializationErrors())
+              return False
+        elif value._is_present_in_parent and not value.IsInitialized():
+          if errors is not None:
+            errors.extend(self.FindInitializationErrors())
+          return False
+
+    return True
+
+  cls.IsInitialized = IsInitialized
+
+  def FindInitializationErrors(self):
+    """Finds required fields which are not initialized.
+
+    Returns:
+      A list of strings.  Each string is a path to an uninitialized field from
+      the top-level message, e.g. "foo.bar[5].baz".
+    """
+
+    errors = []  # simplify things
+
+    for field in required_fields:
+      if not self.HasField(field.name):
+        errors.append(field.name)
+
+    for field, value in self.ListFields():
+      if field.cpp_type == _FieldDescriptor.CPPTYPE_MESSAGE:
+        if field.is_extension:
+          name = "(%s)" % field.full_name
+        else:
+          name = field.name
+
+        if _IsMapField(field):
+          if _IsMessageMapField(field):
+            for key in value:
+              element = value[key]
+              prefix = "%s[%s]." % (name, key)
+              sub_errors = element.FindInitializationErrors()
+              errors += [prefix + error for error in sub_errors]
+          else:
+            # ScalarMaps can't have any initialization errors.
+            pass
+        elif field.label == _FieldDescriptor.LABEL_REPEATED:
+          for i in range(len(value)):
+            element = value[i]
+            prefix = "%s[%d]." % (name, i)
+            sub_errors = element.FindInitializationErrors()
+            errors += [prefix + error for error in sub_errors]
+        else:
+          prefix = name + "."
+          sub_errors = value.FindInitializationErrors()
+          errors += [prefix + error for error in sub_errors]
+
+    return errors
+
+  cls.FindInitializationErrors = FindInitializationErrors
+
+
+def _AddMergeFromMethod(cls):
+  LABEL_REPEATED = _FieldDescriptor.LABEL_REPEATED
+  CPPTYPE_MESSAGE = _FieldDescriptor.CPPTYPE_MESSAGE
+
+  def MergeFrom(self, msg):
+    if not isinstance(msg, cls):
+      raise TypeError(
+          "Parameter to MergeFrom() must be instance of same class: "
+          "expected %s got %s." % (cls.__name__, type(msg).__name__))
+
+    assert msg is not self
+    self._Modified()
+
+    fields = self._fields
+
+    for field, value in msg._fields.items():
+      if field.label == LABEL_REPEATED:
+        field_value = fields.get(field)
+        if field_value is None:
+          # Construct a new object to represent this field.
+          field_value = field._default_constructor(self)
+          fields[field] = field_value
+        field_value.MergeFrom(value)
+      elif field.cpp_type == CPPTYPE_MESSAGE:
+        if value._is_present_in_parent:
+          field_value = fields.get(field)
+          if field_value is None:
+            # Construct a new object to represent this field.
+            field_value = field._default_constructor(self)
+            fields[field] = field_value
+          field_value.MergeFrom(value)
+      else:
+        self._fields[field] = value
+        if field.containing_oneof:
+          self._UpdateOneofState(field)
+
+    if msg._unknown_fields:
+      if not self._unknown_fields:
+        self._unknown_fields = []
+      self._unknown_fields.extend(msg._unknown_fields)
+
+  cls.MergeFrom = MergeFrom
+
+
+def _AddWhichOneofMethod(message_descriptor, cls):
+  def WhichOneof(self, oneof_name):
+    """Returns the name of the currently set field inside a oneof, or None."""
+    try:
+      field = message_descriptor.oneofs_by_name[oneof_name]
+    except KeyError:
+      raise ValueError(
+          'Protocol message has no oneof "%s" field.' % oneof_name)
+
+    nested_field = self._oneofs.get(field, None)
+    if nested_field is not None and self.HasField(nested_field.name):
+      return nested_field.name
+    else:
+      return None
+
+  cls.WhichOneof = WhichOneof
+
+
+def _AddMessageMethods(message_descriptor, cls):
+  """Adds implementations of all Message methods to cls."""
+  _AddListFieldsMethod(message_descriptor, cls)
+  _AddHasFieldMethod(message_descriptor, cls)
+  _AddClearFieldMethod(message_descriptor, cls)
+  if message_descriptor.is_extendable:
+    _AddClearExtensionMethod(cls)
+    _AddHasExtensionMethod(cls)
+  _AddClearMethod(message_descriptor, cls)
+  _AddEqualsMethod(message_descriptor, cls)
+  _AddStrMethod(message_descriptor, cls)
+  _AddReprMethod(message_descriptor, cls)
+  _AddUnicodeMethod(message_descriptor, cls)
+  _AddSetListenerMethod(cls)
+  _AddByteSizeMethod(message_descriptor, cls)
+  _AddSerializeToStringMethod(message_descriptor, cls)
+  _AddSerializePartialToStringMethod(message_descriptor, cls)
+  _AddMergeFromStringMethod(message_descriptor, cls)
+  _AddIsInitializedMethod(message_descriptor, cls)
+  _AddMergeFromMethod(cls)
+  _AddWhichOneofMethod(message_descriptor, cls)
+
+
+def _AddPrivateHelperMethods(message_descriptor, cls):
+  """Adds implementation of private helper methods to cls."""
+
+  def Modified(self):
+    """Sets the _cached_byte_size_dirty bit to true,
+    and propagates this to our listener iff this was a state change.
+    """
+
+    # Note:  Some callers check _cached_byte_size_dirty before calling
+    #   _Modified() as an extra optimization.  So, if this method is ever
+    #   changed such that it does stuff even when _cached_byte_size_dirty is
+    #   already true, the callers need to be updated.
+    if not self._cached_byte_size_dirty:
+      self._cached_byte_size_dirty = True
+      self._listener_for_children.dirty = True
+      self._is_present_in_parent = True
+      self._listener.Modified()
+
+  def _UpdateOneofState(self, field):
+    """Sets field as the active field in its containing oneof.
+
+    Will also delete currently active field in the oneof, if it is different
+    from the argument. Does not mark the message as modified.
+    """
+    other_field = self._oneofs.setdefault(field.containing_oneof, field)
+    if other_field is not field:
+      del self._fields[other_field]
+      self._oneofs[field.containing_oneof] = field
+
+  cls._Modified = Modified
+  cls.SetInParent = Modified
+  cls._UpdateOneofState = _UpdateOneofState
+
+
+class _Listener(object):
+
+  """MessageListener implementation that a parent message registers with its
+  child message.
+
+  In order to support semantics like:
+
+    foo.bar.baz.qux = 23
+    assert foo.HasField('bar')
+
+  ...child objects must have back references to their parents.
+  This helper class is at the heart of this support.
+  """
+
+  def __init__(self, parent_message):
+    """Args:
+      parent_message: The message whose _Modified() method we should call when
+        we receive Modified() messages.
+    """
+    # This listener establishes a back reference from a child (contained) object
+    # to its parent (containing) object.  We make this a weak reference to avoid
+    # creating cyclic garbage when the client finishes with the 'parent' object
+    # in the tree.
+    if isinstance(parent_message, weakref.ProxyType):
+      self._parent_message_weakref = parent_message
+    else:
+      self._parent_message_weakref = weakref.proxy(parent_message)
+
+    # As an optimization, we also indicate directly on the listener whether
+    # or not the parent message is dirty.  This way we can avoid traversing
+    # up the tree in the common case.
+    self.dirty = False
+
+  def Modified(self):
+    if self.dirty:
+      return
+    try:
+      # Propagate the signal to our parents iff this is the first field set.
+      self._parent_message_weakref._Modified()
+    except ReferenceError:
+      # We can get here if a client has kept a reference to a child object,
+      # and is now setting a field on it, but the child's parent has been
+      # garbage-collected.  This is not an error.
+      pass
+
+
+class _OneofListener(_Listener):
+  """Special listener implementation for setting composite oneof fields."""
+
+  def __init__(self, parent_message, field):
+    """Args:
+      parent_message: The message whose _Modified() method we should call when
+        we receive Modified() messages.
+      field: The descriptor of the field being set in the parent message.
+    """
+    super(_OneofListener, self).__init__(parent_message)
+    self._field = field
+
+  def Modified(self):
+    """Also updates the state of the containing oneof in the parent message."""
+    try:
+      self._parent_message_weakref._UpdateOneofState(self._field)
+      super(_OneofListener, self).Modified()
+    except ReferenceError:
+      pass
+
+
+# TODO(robinson): Move elsewhere?  This file is getting pretty ridiculous...
+# TODO(robinson): Unify error handling of "unknown extension" crap.
+# TODO(robinson): Support iteritems()-style iteration over all
+# extensions with the "has" bits turned on?
+class _ExtensionDict(object):
+
+  """Dict-like container for supporting an indexable "Extensions"
+  field on proto instances.
+
+  Note that in all cases we expect extension handles to be
+  FieldDescriptors.
+  """
+
+  def __init__(self, extended_message):
+    """extended_message: Message instance for which we are the Extensions dict.
+    """
+
+    self._extended_message = extended_message
+
+  def __getitem__(self, extension_handle):
+    """Returns the current value of the given extension handle."""
+
+    _VerifyExtensionHandle(self._extended_message, extension_handle)
+
+    result = self._extended_message._fields.get(extension_handle)
+    if result is not None:
+      return result
+
+    if extension_handle.label == _FieldDescriptor.LABEL_REPEATED:
+      result = extension_handle._default_constructor(self._extended_message)
+    elif extension_handle.cpp_type == _FieldDescriptor.CPPTYPE_MESSAGE:
+      result = extension_handle.message_type._concrete_class()
+      try:
+        result._SetListener(self._extended_message._listener_for_children)
+      except ReferenceError:
+        pass
+    else:
+      # Singular scalar -- just return the default without inserting into the
+      # dict.
+      return extension_handle.default_value
+
+    # Atomically check if another thread has preempted us and, if not, swap
+    # in the new object we just created.  If someone has preempted us, we
+    # take that object and discard ours.
+    # WARNING:  We are relying on setdefault() being atomic.  This is true
+    #   in CPython but we haven't investigated others.  This warning appears
+    #   in several other locations in this file.
+    result = self._extended_message._fields.setdefault(
+        extension_handle, result)
+
+    return result
+
+  def __eq__(self, other):
+    if not isinstance(other, self.__class__):
+      return False
+
+    my_fields = self._extended_message.ListFields()
+    other_fields = other._extended_message.ListFields()
+
+    # Get rid of non-extension fields.
+    my_fields    = [ field for field in my_fields    if field.is_extension ]
+    other_fields = [ field for field in other_fields if field.is_extension ]
+
+    return my_fields == other_fields
+
+  def __ne__(self, other):
+    return not self == other
+
+  def __hash__(self):
+    raise TypeError('unhashable object')
+
+  # Note that this is only meaningful for non-repeated, scalar extension
+  # fields.  Note also that we may have to call _Modified() when we do
+  # successfully set a field this way, to set any necssary "has" bits in the
+  # ancestors of the extended message.
+  def __setitem__(self, extension_handle, value):
+    """If extension_handle specifies a non-repeated, scalar extension
+    field, sets the value of that field.
+    """
+
+    _VerifyExtensionHandle(self._extended_message, extension_handle)
+
+    if (extension_handle.label == _FieldDescriptor.LABEL_REPEATED or
+        extension_handle.cpp_type == _FieldDescriptor.CPPTYPE_MESSAGE):
+      raise TypeError(
+          'Cannot assign to extension "%s" because it is a repeated or '
+          'composite type.' % extension_handle.full_name)
+
+    # It's slightly wasteful to lookup the type checker each time,
+    # but we expect this to be a vanishingly uncommon case anyway.
+    type_checker = type_checkers.GetTypeChecker(extension_handle)
+    # pylint: disable=protected-access
+    self._extended_message._fields[extension_handle] = (
+        type_checker.CheckValue(value))
+    self._extended_message._Modified()
+
+  def _FindExtensionByName(self, name):
+    """Tries to find a known extension with the specified name.
+
+    Args:
+      name: Extension full name.
+
+    Returns:
+      Extension field descriptor.
+    """
+    return self._extended_message._extensions_by_name.get(name, None)
diff --git a/python/google/protobuf/internal/reflection_test.py b/python/google/protobuf/internal/reflection_test.py
new file mode 100755
index 0000000..752f2f5
--- /dev/null
+++ b/python/google/protobuf/internal/reflection_test.py
@@ -0,0 +1,2976 @@
+#! /usr/bin/env python
+# -*- coding: utf-8 -*-
+#
+# Protocol Buffers - Google's data interchange format
+# Copyright 2008 Google Inc.  All rights reserved.
+# https://developers.google.com/protocol-buffers/
+#
+# Redistribution and use in source and binary forms, with or without
+# modification, are permitted provided that the following conditions are
+# met:
+#
+#     * Redistributions of source code must retain the above copyright
+# notice, this list of conditions and the following disclaimer.
+#     * Redistributions in binary form must reproduce the above
+# copyright notice, this list of conditions and the following disclaimer
+# in the documentation and/or other materials provided with the
+# distribution.
+#     * Neither the name of Google Inc. nor the names of its
+# contributors may be used to endorse or promote products derived from
+# this software without specific prior written permission.
+#
+# THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+# "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
+# LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
+# A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
+# OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
+# SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
+# LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
+# DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
+# THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
+# (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+# OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+
+"""Unittest for reflection.py, which also indirectly tests the output of the
+pure-Python protocol compiler.
+"""
+
+import copy
+import gc
+import operator
+import six
+import struct
+
+try:
+  import unittest2 as unittest
+except ImportError:
+  import unittest
+from google.protobuf import unittest_import_pb2
+from google.protobuf import unittest_mset_pb2
+from google.protobuf import unittest_pb2
+from google.protobuf import descriptor_pb2
+from google.protobuf import descriptor
+from google.protobuf import message
+from google.protobuf import reflection
+from google.protobuf import text_format
+from google.protobuf.internal import api_implementation
+from google.protobuf.internal import more_extensions_pb2
+from google.protobuf.internal import more_messages_pb2
+from google.protobuf.internal import message_set_extensions_pb2
+from google.protobuf.internal import wire_format
+from google.protobuf.internal import test_util
+from google.protobuf.internal import decoder
+
+
+class _MiniDecoder(object):
+  """Decodes a stream of values from a string.
+
+  Once upon a time we actually had a class called decoder.Decoder.  Then we
+  got rid of it during a redesign that made decoding much, much faster overall.
+  But a couple tests in this file used it to check that the serialized form of
+  a message was correct.  So, this class implements just the methods that were
+  used by said tests, so that we don't have to rewrite the tests.
+  """
+
+  def __init__(self, bytes):
+    self._bytes = bytes
+    self._pos = 0
+
+  def ReadVarint(self):
+    result, self._pos = decoder._DecodeVarint(self._bytes, self._pos)
+    return result
+
+  ReadInt32 = ReadVarint
+  ReadInt64 = ReadVarint
+  ReadUInt32 = ReadVarint
+  ReadUInt64 = ReadVarint
+
+  def ReadSInt64(self):
+    return wire_format.ZigZagDecode(self.ReadVarint())
+
+  ReadSInt32 = ReadSInt64
+
+  def ReadFieldNumberAndWireType(self):
+    return wire_format.UnpackTag(self.ReadVarint())
+
+  def ReadFloat(self):
+    result = struct.unpack("<f", self._bytes[self._pos:self._pos+4])[0]
+    self._pos += 4
+    return result
+
+  def ReadDouble(self):
+    result = struct.unpack("<d", self._bytes[self._pos:self._pos+8])[0]
+    self._pos += 8
+    return result
+
+  def EndOfStream(self):
+    return self._pos == len(self._bytes)
+
+
+class ReflectionTest(unittest.TestCase):
+
+  def assertListsEqual(self, values, others):
+    self.assertEqual(len(values), len(others))
+    for i in range(len(values)):
+      self.assertEqual(values[i], others[i])
+
+  def testScalarConstructor(self):
+    # Constructor with only scalar types should succeed.
+    proto = unittest_pb2.TestAllTypes(
+        optional_int32=24,
+        optional_double=54.321,
+        optional_string='optional_string')
+
+    self.assertEqual(24, proto.optional_int32)
+    self.assertEqual(54.321, proto.optional_double)
+    self.assertEqual('optional_string', proto.optional_string)
+
+  def testRepeatedScalarConstructor(self):
+    # Constructor with only repeated scalar types should succeed.
+    proto = unittest_pb2.TestAllTypes(
+        repeated_int32=[1, 2, 3, 4],
+        repeated_double=[1.23, 54.321],
+        repeated_bool=[True, False, False],
+        repeated_string=["optional_string"])
+
+    self.assertEqual([1, 2, 3, 4], list(proto.repeated_int32))
+    self.assertEqual([1.23, 54.321], list(proto.repeated_double))
+    self.assertEqual([True, False, False], list(proto.repeated_bool))
+    self.assertEqual(["optional_string"], list(proto.repeated_string))
+
+  def testRepeatedCompositeConstructor(self):
+    # Constructor with only repeated composite types should succeed.
+    proto = unittest_pb2.TestAllTypes(
+        repeated_nested_message=[
+            unittest_pb2.TestAllTypes.NestedMessage(
+                bb=unittest_pb2.TestAllTypes.FOO),
+            unittest_pb2.TestAllTypes.NestedMessage(
+                bb=unittest_pb2.TestAllTypes.BAR)],
+        repeated_foreign_message=[
+            unittest_pb2.ForeignMessage(c=-43),
+            unittest_pb2.ForeignMessage(c=45324),
+            unittest_pb2.ForeignMessage(c=12)],
+        repeatedgroup=[
+            unittest_pb2.TestAllTypes.RepeatedGroup(),
+            unittest_pb2.TestAllTypes.RepeatedGroup(a=1),
+            unittest_pb2.TestAllTypes.RepeatedGroup(a=2)])
+
+    self.assertEqual(
+        [unittest_pb2.TestAllTypes.NestedMessage(
+            bb=unittest_pb2.TestAllTypes.FOO),
+         unittest_pb2.TestAllTypes.NestedMessage(
+             bb=unittest_pb2.TestAllTypes.BAR)],
+        list(proto.repeated_nested_message))
+    self.assertEqual(
+        [unittest_pb2.ForeignMessage(c=-43),
+         unittest_pb2.ForeignMessage(c=45324),
+         unittest_pb2.ForeignMessage(c=12)],
+        list(proto.repeated_foreign_message))
+    self.assertEqual(
+        [unittest_pb2.TestAllTypes.RepeatedGroup(),
+         unittest_pb2.TestAllTypes.RepeatedGroup(a=1),
+         unittest_pb2.TestAllTypes.RepeatedGroup(a=2)],
+        list(proto.repeatedgroup))
+
+  def testMixedConstructor(self):
+    # Constructor with only mixed types should succeed.
+    proto = unittest_pb2.TestAllTypes(
+        optional_int32=24,
+        optional_string='optional_string',
+        repeated_double=[1.23, 54.321],
+        repeated_bool=[True, False, False],
+        repeated_nested_message=[
+            unittest_pb2.TestAllTypes.NestedMessage(
+                bb=unittest_pb2.TestAllTypes.FOO),
+            unittest_pb2.TestAllTypes.NestedMessage(
+                bb=unittest_pb2.TestAllTypes.BAR)],
+        repeated_foreign_message=[
+            unittest_pb2.ForeignMessage(c=-43),
+            unittest_pb2.ForeignMessage(c=45324),
+            unittest_pb2.ForeignMessage(c=12)])
+
+    self.assertEqual(24, proto.optional_int32)
+    self.assertEqual('optional_string', proto.optional_string)
+    self.assertEqual([1.23, 54.321], list(proto.repeated_double))
+    self.assertEqual([True, False, False], list(proto.repeated_bool))
+    self.assertEqual(
+        [unittest_pb2.TestAllTypes.NestedMessage(
+            bb=unittest_pb2.TestAllTypes.FOO),
+         unittest_pb2.TestAllTypes.NestedMessage(
+             bb=unittest_pb2.TestAllTypes.BAR)],
+        list(proto.repeated_nested_message))
+    self.assertEqual(
+        [unittest_pb2.ForeignMessage(c=-43),
+         unittest_pb2.ForeignMessage(c=45324),
+         unittest_pb2.ForeignMessage(c=12)],
+        list(proto.repeated_foreign_message))
+
+  def testConstructorTypeError(self):
+    self.assertRaises(
+        TypeError, unittest_pb2.TestAllTypes, optional_int32="foo")
+    self.assertRaises(
+        TypeError, unittest_pb2.TestAllTypes, optional_string=1234)
+    self.assertRaises(
+        TypeError, unittest_pb2.TestAllTypes, optional_nested_message=1234)
+    self.assertRaises(
+        TypeError, unittest_pb2.TestAllTypes, repeated_int32=1234)
+    self.assertRaises(
+        TypeError, unittest_pb2.TestAllTypes, repeated_int32=["foo"])
+    self.assertRaises(
+        TypeError, unittest_pb2.TestAllTypes, repeated_string=1234)
+    self.assertRaises(
+        TypeError, unittest_pb2.TestAllTypes, repeated_string=[1234])
+    self.assertRaises(
+        TypeError, unittest_pb2.TestAllTypes, repeated_nested_message=1234)
+    self.assertRaises(
+        TypeError, unittest_pb2.TestAllTypes, repeated_nested_message=[1234])
+
+  def testConstructorInvalidatesCachedByteSize(self):
+    message = unittest_pb2.TestAllTypes(optional_int32 = 12)
+    self.assertEqual(2, message.ByteSize())
+
+    message = unittest_pb2.TestAllTypes(
+        optional_nested_message = unittest_pb2.TestAllTypes.NestedMessage())
+    self.assertEqual(3, message.ByteSize())
+
+    message = unittest_pb2.TestAllTypes(repeated_int32 = [12])
+    self.assertEqual(3, message.ByteSize())
+
+    message = unittest_pb2.TestAllTypes(
+        repeated_nested_message = [unittest_pb2.TestAllTypes.NestedMessage()])
+    self.assertEqual(3, message.ByteSize())
+
+  def testSimpleHasBits(self):
+    # Test a scalar.
+    proto = unittest_pb2.TestAllTypes()
+    self.assertTrue(not proto.HasField('optional_int32'))
+    self.assertEqual(0, proto.optional_int32)
+    # HasField() shouldn't be true if all we've done is
+    # read the default value.
+    self.assertTrue(not proto.HasField('optional_int32'))
+    proto.optional_int32 = 1
+    # Setting a value however *should* set the "has" bit.
+    self.assertTrue(proto.HasField('optional_int32'))
+    proto.ClearField('optional_int32')
+    # And clearing that value should unset the "has" bit.
+    self.assertTrue(not proto.HasField('optional_int32'))
+
+  def testHasBitsWithSinglyNestedScalar(self):
+    # Helper used to test foreign messages and groups.
+    #
+    # composite_field_name should be the name of a non-repeated
+    # composite (i.e., foreign or group) field in TestAllTypes,
+    # and scalar_field_name should be the name of an integer-valued
+    # scalar field within that composite.
+    #
+    # I never thought I'd miss C++ macros and templates so much. :(
+    # This helper is semantically just:
+    #
+    #   assert proto.composite_field.scalar_field == 0
+    #   assert not proto.composite_field.HasField('scalar_field')
+    #   assert not proto.HasField('composite_field')
+    #
+    #   proto.composite_field.scalar_field = 10
+    #   old_composite_field = proto.composite_field
+    #
+    #   assert proto.composite_field.scalar_field == 10
+    #   assert proto.composite_field.HasField('scalar_field')
+    #   assert proto.HasField('composite_field')
+    #
+    #   proto.ClearField('composite_field')
+    #
+    #   assert not proto.composite_field.HasField('scalar_field')
+    #   assert not proto.HasField('composite_field')
+    #   assert proto.composite_field.scalar_field == 0
+    #
+    #   # Now ensure that ClearField('composite_field') disconnected
+    #   # the old field object from the object tree...
+    #   assert old_composite_field is not proto.composite_field
+    #   old_composite_field.scalar_field = 20
+    #   assert not proto.composite_field.HasField('scalar_field')
+    #   assert not proto.HasField('composite_field')
+    def TestCompositeHasBits(composite_field_name, scalar_field_name):
+      proto = unittest_pb2.TestAllTypes()
+      # First, check that we can get the scalar value, and see that it's the
+      # default (0), but that proto.HasField('omposite') and
+      # proto.composite.HasField('scalar') will still return False.
+      composite_field = getattr(proto, composite_field_name)
+      original_scalar_value = getattr(composite_field, scalar_field_name)
+      self.assertEqual(0, original_scalar_value)
+      # Assert that the composite object does not "have" the scalar.
+      self.assertTrue(not composite_field.HasField(scalar_field_name))
+      # Assert that proto does not "have" the composite field.
+      self.assertTrue(not proto.HasField(composite_field_name))
+
+      # Now set the scalar within the composite field.  Ensure that the setting
+      # is reflected, and that proto.HasField('composite') and
+      # proto.composite.HasField('scalar') now both return True.
+      new_val = 20
+      setattr(composite_field, scalar_field_name, new_val)
+      self.assertEqual(new_val, getattr(composite_field, scalar_field_name))
+      # Hold on to a reference to the current composite_field object.
+      old_composite_field = composite_field
+      # Assert that the has methods now return true.
+      self.assertTrue(composite_field.HasField(scalar_field_name))
+      self.assertTrue(proto.HasField(composite_field_name))
+
+      # Now call the clear method...
+      proto.ClearField(composite_field_name)
+
+      # ...and ensure that the "has" bits are all back to False...
+      composite_field = getattr(proto, composite_field_name)
+      self.assertTrue(not composite_field.HasField(scalar_field_name))
+      self.assertTrue(not proto.HasField(composite_field_name))
+      # ...and ensure that the scalar field has returned to its default.
+      self.assertEqual(0, getattr(composite_field, scalar_field_name))
+
+      self.assertTrue(old_composite_field is not composite_field)
+      setattr(old_composite_field, scalar_field_name, new_val)
+      self.assertTrue(not composite_field.HasField(scalar_field_name))
+      self.assertTrue(not proto.HasField(composite_field_name))
+      self.assertEqual(0, getattr(composite_field, scalar_field_name))
+
+    # Test simple, single-level nesting when we set a scalar.
+    TestCompositeHasBits('optionalgroup', 'a')
+    TestCompositeHasBits('optional_nested_message', 'bb')
+    TestCompositeHasBits('optional_foreign_message', 'c')
+    TestCompositeHasBits('optional_import_message', 'd')
+
+  def testReferencesToNestedMessage(self):
+    proto = unittest_pb2.TestAllTypes()
+    nested = proto.optional_nested_message
+    del proto
+    # A previous version had a bug where this would raise an exception when
+    # hitting a now-dead weak reference.
+    nested.bb = 23
+
+  def testDisconnectingNestedMessageBeforeSettingField(self):
+    proto = unittest_pb2.TestAllTypes()
+    nested = proto.optional_nested_message
+    proto.ClearField('optional_nested_message')  # Should disconnect from parent
+    self.assertTrue(nested is not proto.optional_nested_message)
+    nested.bb = 23
+    self.assertTrue(not proto.HasField('optional_nested_message'))
+    self.assertEqual(0, proto.optional_nested_message.bb)
+
+  def testGetDefaultMessageAfterDisconnectingDefaultMessage(self):
+    proto = unittest_pb2.TestAllTypes()
+    nested = proto.optional_nested_message
+    proto.ClearField('optional_nested_message')
+    del proto
+    del nested
+    # Force a garbage collect so that the underlying CMessages are freed along
+    # with the Messages they point to. This is to make sure we're not deleting
+    # default message instances.
+    gc.collect()
+    proto = unittest_pb2.TestAllTypes()
+    nested = proto.optional_nested_message
+
+  def testDisconnectingNestedMessageAfterSettingField(self):
+    proto = unittest_pb2.TestAllTypes()
+    nested = proto.optional_nested_message
+    nested.bb = 5
+    self.assertTrue(proto.HasField('optional_nested_message'))
+    proto.ClearField('optional_nested_message')  # Should disconnect from parent
+    self.assertEqual(5, nested.bb)
+    self.assertEqual(0, proto.optional_nested_message.bb)
+    self.assertTrue(nested is not proto.optional_nested_message)
+    nested.bb = 23
+    self.assertTrue(not proto.HasField('optional_nested_message'))
+    self.assertEqual(0, proto.optional_nested_message.bb)
+
+  def testDisconnectingNestedMessageBeforeGettingField(self):
+    proto = unittest_pb2.TestAllTypes()
+    self.assertTrue(not proto.HasField('optional_nested_message'))
+    proto.ClearField('optional_nested_message')
+    self.assertTrue(not proto.HasField('optional_nested_message'))
+
+  def testDisconnectingNestedMessageAfterMerge(self):
+    # This test exercises the code path that does not use ReleaseMessage().
+    # The underlying fear is that if we use ReleaseMessage() incorrectly,
+    # we will have memory leaks.  It's hard to check that that doesn't happen,
+    # but at least we can exercise that code path to make sure it works.
+    proto1 = unittest_pb2.TestAllTypes()
+    proto2 = unittest_pb2.TestAllTypes()
+    proto2.optional_nested_message.bb = 5
+    proto1.MergeFrom(proto2)
+    self.assertTrue(proto1.HasField('optional_nested_message'))
+    proto1.ClearField('optional_nested_message')
+    self.assertTrue(not proto1.HasField('optional_nested_message'))
+
+  def testDisconnectingLazyNestedMessage(self):
+    # This test exercises releasing a nested message that is lazy. This test
+    # only exercises real code in the C++ implementation as Python does not
+    # support lazy parsing, but the current C++ implementation results in
+    # memory corruption and a crash.
+    if api_implementation.Type() != 'python':
+      return
+    proto = unittest_pb2.TestAllTypes()
+    proto.optional_lazy_message.bb = 5
+    proto.ClearField('optional_lazy_message')
+    del proto
+    gc.collect()
+
+  def testHasBitsWhenModifyingRepeatedFields(self):
+    # Test nesting when we add an element to a repeated field in a submessage.
+    proto = unittest_pb2.TestNestedMessageHasBits()
+    proto.optional_nested_message.nestedmessage_repeated_int32.append(5)
+    self.assertEqual(
+        [5], proto.optional_nested_message.nestedmessage_repeated_int32)
+    self.assertTrue(proto.HasField('optional_nested_message'))
+
+    # Do the same test, but with a repeated composite field within the
+    # submessage.
+    proto.ClearField('optional_nested_message')
+    self.assertTrue(not proto.HasField('optional_nested_message'))
+    proto.optional_nested_message.nestedmessage_repeated_foreignmessage.add()
+    self.assertTrue(proto.HasField('optional_nested_message'))
+
+  def testHasBitsForManyLevelsOfNesting(self):
+    # Test nesting many levels deep.
+    recursive_proto = unittest_pb2.TestMutualRecursionA()
+    self.assertTrue(not recursive_proto.HasField('bb'))
+    self.assertEqual(0, recursive_proto.bb.a.bb.a.bb.optional_int32)
+    self.assertTrue(not recursive_proto.HasField('bb'))
+    recursive_proto.bb.a.bb.a.bb.optional_int32 = 5
+    self.assertEqual(5, recursive_proto.bb.a.bb.a.bb.optional_int32)
+    self.assertTrue(recursive_proto.HasField('bb'))
+    self.assertTrue(recursive_proto.bb.HasField('a'))
+    self.assertTrue(recursive_proto.bb.a.HasField('bb'))
+    self.assertTrue(recursive_proto.bb.a.bb.HasField('a'))
+    self.assertTrue(recursive_proto.bb.a.bb.a.HasField('bb'))
+    self.assertTrue(not recursive_proto.bb.a.bb.a.bb.HasField('a'))
+    self.assertTrue(recursive_proto.bb.a.bb.a.bb.HasField('optional_int32'))
+
+  def testSingularListFields(self):
+    proto = unittest_pb2.TestAllTypes()
+    proto.optional_fixed32 = 1
+    proto.optional_int32 = 5
+    proto.optional_string = 'foo'
+    # Access sub-message but don't set it yet.
+    nested_message = proto.optional_nested_message
+    self.assertEqual(
+      [ (proto.DESCRIPTOR.fields_by_name['optional_int32'  ], 5),
+        (proto.DESCRIPTOR.fields_by_name['optional_fixed32'], 1),
+        (proto.DESCRIPTOR.fields_by_name['optional_string' ], 'foo') ],
+      proto.ListFields())
+
+    proto.optional_nested_message.bb = 123
+    self.assertEqual(
+      [ (proto.DESCRIPTOR.fields_by_name['optional_int32'  ], 5),
+        (proto.DESCRIPTOR.fields_by_name['optional_fixed32'], 1),
+        (proto.DESCRIPTOR.fields_by_name['optional_string' ], 'foo'),
+        (proto.DESCRIPTOR.fields_by_name['optional_nested_message' ],
+             nested_message) ],
+      proto.ListFields())
+
+  def testRepeatedListFields(self):
+    proto = unittest_pb2.TestAllTypes()
+    proto.repeated_fixed32.append(1)
+    proto.repeated_int32.append(5)
+    proto.repeated_int32.append(11)
+    proto.repeated_string.extend(['foo', 'bar'])
+    proto.repeated_string.extend([])
+    proto.repeated_string.append('baz')
+    proto.repeated_string.extend(str(x) for x in range(2))
+    proto.optional_int32 = 21
+    proto.repeated_bool  # Access but don't set anything; should not be listed.
+    self.assertEqual(
+      [ (proto.DESCRIPTOR.fields_by_name['optional_int32'  ], 21),
+        (proto.DESCRIPTOR.fields_by_name['repeated_int32'  ], [5, 11]),
+        (proto.DESCRIPTOR.fields_by_name['repeated_fixed32'], [1]),
+        (proto.DESCRIPTOR.fields_by_name['repeated_string' ],
+          ['foo', 'bar', 'baz', '0', '1']) ],
+      proto.ListFields())
+
+  def testSingularListExtensions(self):
+    proto = unittest_pb2.TestAllExtensions()
+    proto.Extensions[unittest_pb2.optional_fixed32_extension] = 1
+    proto.Extensions[unittest_pb2.optional_int32_extension  ] = 5
+    proto.Extensions[unittest_pb2.optional_string_extension ] = 'foo'
+    self.assertEqual(
+      [ (unittest_pb2.optional_int32_extension  , 5),
+        (unittest_pb2.optional_fixed32_extension, 1),
+        (unittest_pb2.optional_string_extension , 'foo') ],
+      proto.ListFields())
+
+  def testRepeatedListExtensions(self):
+    proto = unittest_pb2.TestAllExtensions()
+    proto.Extensions[unittest_pb2.repeated_fixed32_extension].append(1)
+    proto.Extensions[unittest_pb2.repeated_int32_extension  ].append(5)
+    proto.Extensions[unittest_pb2.repeated_int32_extension  ].append(11)
+    proto.Extensions[unittest_pb2.repeated_string_extension ].append('foo')
+    proto.Extensions[unittest_pb2.repeated_string_extension ].append('bar')
+    proto.Extensions[unittest_pb2.repeated_string_extension ].append('baz')
+    proto.Extensions[unittest_pb2.optional_int32_extension  ] = 21
+    self.assertEqual(
+      [ (unittest_pb2.optional_int32_extension  , 21),
+        (unittest_pb2.repeated_int32_extension  , [5, 11]),
+        (unittest_pb2.repeated_fixed32_extension, [1]),
+        (unittest_pb2.repeated_string_extension , ['foo', 'bar', 'baz']) ],
+      proto.ListFields())
+
+  def testListFieldsAndExtensions(self):
+    proto = unittest_pb2.TestFieldOrderings()
+    test_util.SetAllFieldsAndExtensions(proto)
+    unittest_pb2.my_extension_int
+    self.assertEqual(
+      [ (proto.DESCRIPTOR.fields_by_name['my_int'   ], 1),
+        (unittest_pb2.my_extension_int               , 23),
+        (proto.DESCRIPTOR.fields_by_name['my_string'], 'foo'),
+        (unittest_pb2.my_extension_string            , 'bar'),
+        (proto.DESCRIPTOR.fields_by_name['my_float' ], 1.0) ],
+      proto.ListFields())
+
+  def testDefaultValues(self):
+    proto = unittest_pb2.TestAllTypes()
+    self.assertEqual(0, proto.optional_int32)
+    self.assertEqual(0, proto.optional_int64)
+    self.assertEqual(0, proto.optional_uint32)
+    self.assertEqual(0, proto.optional_uint64)
+    self.assertEqual(0, proto.optional_sint32)
+    self.assertEqual(0, proto.optional_sint64)
+    self.assertEqual(0, proto.optional_fixed32)
+    self.assertEqual(0, proto.optional_fixed64)
+    self.assertEqual(0, proto.optional_sfixed32)
+    self.assertEqual(0, proto.optional_sfixed64)
+    self.assertEqual(0.0, proto.optional_float)
+    self.assertEqual(0.0, proto.optional_double)
+    self.assertEqual(False, proto.optional_bool)
+    self.assertEqual('', proto.optional_string)
+    self.assertEqual(b'', proto.optional_bytes)
+
+    self.assertEqual(41, proto.default_int32)
+    self.assertEqual(42, proto.default_int64)
+    self.assertEqual(43, proto.default_uint32)
+    self.assertEqual(44, proto.default_uint64)
+    self.assertEqual(-45, proto.default_sint32)
+    self.assertEqual(46, proto.default_sint64)
+    self.assertEqual(47, proto.default_fixed32)
+    self.assertEqual(48, proto.default_fixed64)
+    self.assertEqual(49, proto.default_sfixed32)
+    self.assertEqual(-50, proto.default_sfixed64)
+    self.assertEqual(51.5, proto.default_float)
+    self.assertEqual(52e3, proto.default_double)
+    self.assertEqual(True, proto.default_bool)
+    self.assertEqual('hello', proto.default_string)
+    self.assertEqual(b'world', proto.default_bytes)
+    self.assertEqual(unittest_pb2.TestAllTypes.BAR, proto.default_nested_enum)
+    self.assertEqual(unittest_pb2.FOREIGN_BAR, proto.default_foreign_enum)
+    self.assertEqual(unittest_import_pb2.IMPORT_BAR,
+                     proto.default_import_enum)
+
+    proto = unittest_pb2.TestExtremeDefaultValues()
+    self.assertEqual(u'\u1234', proto.utf8_string)
+
+  def testHasFieldWithUnknownFieldName(self):
+    proto = unittest_pb2.TestAllTypes()
+    self.assertRaises(ValueError, proto.HasField, 'nonexistent_field')
+
+  def testClearFieldWithUnknownFieldName(self):
+    proto = unittest_pb2.TestAllTypes()
+    self.assertRaises(ValueError, proto.ClearField, 'nonexistent_field')
+
+  def testClearRemovesChildren(self):
+    # Make sure there aren't any implementation bugs that are only partially
+    # clearing the message (which can happen in the more complex C++
+    # implementation which has parallel message lists).
+    proto = unittest_pb2.TestRequiredForeign()
+    for i in range(10):
+      proto.repeated_message.add()
+    proto2 = unittest_pb2.TestRequiredForeign()
+    proto.CopyFrom(proto2)
+    self.assertRaises(IndexError, lambda: proto.repeated_message[5])
+
+  def testDisallowedAssignments(self):
+    # It's illegal to assign values directly to repeated fields
+    # or to nonrepeated composite fields.  Ensure that this fails.
+    proto = unittest_pb2.TestAllTypes()
+    # Repeated fields.
+    self.assertRaises(AttributeError, setattr, proto, 'repeated_int32', 10)
+    # Lists shouldn't work, either.
+    self.assertRaises(AttributeError, setattr, proto, 'repeated_int32', [10])
+    # Composite fields.
+    self.assertRaises(AttributeError, setattr, proto,
+                      'optional_nested_message', 23)
+    # Assignment to a repeated nested message field without specifying
+    # the index in the array of nested messages.
+    self.assertRaises(AttributeError, setattr, proto.repeated_nested_message,
+                      'bb', 34)
+    # Assignment to an attribute of a repeated field.
+    self.assertRaises(AttributeError, setattr, proto.repeated_float,
+                      'some_attribute', 34)
+    # proto.nonexistent_field = 23 should fail as well.
+    self.assertRaises(AttributeError, setattr, proto, 'nonexistent_field', 23)
+
+  def testSingleScalarTypeSafety(self):
+    proto = unittest_pb2.TestAllTypes()
+    self.assertRaises(TypeError, setattr, proto, 'optional_int32', 1.1)
+    self.assertRaises(TypeError, setattr, proto, 'optional_int32', 'foo')
+    self.assertRaises(TypeError, setattr, proto, 'optional_string', 10)
+    self.assertRaises(TypeError, setattr, proto, 'optional_bytes', 10)
+
+  def testIntegerTypes(self):
+    def TestGetAndDeserialize(field_name, value, expected_type):
+      proto = unittest_pb2.TestAllTypes()
+      setattr(proto, field_name, value)
+      self.assertIsInstance(getattr(proto, field_name), expected_type)
+      proto2 = unittest_pb2.TestAllTypes()
+      proto2.ParseFromString(proto.SerializeToString())
+      self.assertIsInstance(getattr(proto2, field_name), expected_type)
+
+    TestGetAndDeserialize('optional_int32', 1, int)
+    TestGetAndDeserialize('optional_int32', 1 << 30, int)
+    TestGetAndDeserialize('optional_uint32', 1 << 30, int)
+    try:
+      integer_64 = long
+    except NameError: # Python3
+      integer_64 = int
+    if struct.calcsize('L') == 4:
+      # Python only has signed ints, so 32-bit python can't fit an uint32
+      # in an int.
+      TestGetAndDeserialize('optional_uint32', 1 << 31, long)
+    else:
+      # 64-bit python can fit uint32 inside an int
+      TestGetAndDeserialize('optional_uint32', 1 << 31, int)
+    TestGetAndDeserialize('optional_int64', 1 << 30, integer_64)
+    TestGetAndDeserialize('optional_int64', 1 << 60, integer_64)
+    TestGetAndDeserialize('optional_uint64', 1 << 30, integer_64)
+    TestGetAndDeserialize('optional_uint64', 1 << 60, integer_64)
+
+  def testSingleScalarBoundsChecking(self):
+    def TestMinAndMaxIntegers(field_name, expected_min, expected_max):
+      pb = unittest_pb2.TestAllTypes()
+      setattr(pb, field_name, expected_min)
+      self.assertEqual(expected_min, getattr(pb, field_name))
+      setattr(pb, field_name, expected_max)
+      self.assertEqual(expected_max, getattr(pb, field_name))
+      self.assertRaises(ValueError, setattr, pb, field_name, expected_min - 1)
+      self.assertRaises(ValueError, setattr, pb, field_name, expected_max + 1)
+
+    TestMinAndMaxIntegers('optional_int32', -(1 << 31), (1 << 31) - 1)
+    TestMinAndMaxIntegers('optional_uint32', 0, 0xffffffff)
+    TestMinAndMaxIntegers('optional_int64', -(1 << 63), (1 << 63) - 1)
+    TestMinAndMaxIntegers('optional_uint64', 0, 0xffffffffffffffff)
+
+    pb = unittest_pb2.TestAllTypes()
+    pb.optional_nested_enum = 1
+    self.assertEqual(1, pb.optional_nested_enum)
+
+  def testRepeatedScalarTypeSafety(self):
+    proto = unittest_pb2.TestAllTypes()
+    self.assertRaises(TypeError, proto.repeated_int32.append, 1.1)
+    self.assertRaises(TypeError, proto.repeated_int32.append, 'foo')
+    self.assertRaises(TypeError, proto.repeated_string, 10)
+    self.assertRaises(TypeError, proto.repeated_bytes, 10)
+
+    proto.repeated_int32.append(10)
+    proto.repeated_int32[0] = 23
+    self.assertRaises(IndexError, proto.repeated_int32.__setitem__, 500, 23)
+    self.assertRaises(TypeError, proto.repeated_int32.__setitem__, 0, 'abc')
+
+    # Repeated enums tests.
+    #proto.repeated_nested_enum.append(0)
+
+  def testSingleScalarGettersAndSetters(self):
+    proto = unittest_pb2.TestAllTypes()
+    self.assertEqual(0, proto.optional_int32)
+    proto.optional_int32 = 1
+    self.assertEqual(1, proto.optional_int32)
+
+    proto.optional_uint64 = 0xffffffffffff
+    self.assertEqual(0xffffffffffff, proto.optional_uint64)
+    proto.optional_uint64 = 0xffffffffffffffff
+    self.assertEqual(0xffffffffffffffff, proto.optional_uint64)
+    # TODO(robinson): Test all other scalar field types.
+
+  def testSingleScalarClearField(self):
+    proto = unittest_pb2.TestAllTypes()
+    # Should be allowed to clear something that's not there (a no-op).
+    proto.ClearField('optional_int32')
+    proto.optional_int32 = 1
+    self.assertTrue(proto.HasField('optional_int32'))
+    proto.ClearField('optional_int32')
+    self.assertEqual(0, proto.optional_int32)
+    self.assertTrue(not proto.HasField('optional_int32'))
+    # TODO(robinson): Test all other scalar field types.
+
+  def testEnums(self):
+    proto = unittest_pb2.TestAllTypes()
+    self.assertEqual(1, proto.FOO)
+    self.assertEqual(1, unittest_pb2.TestAllTypes.FOO)
+    self.assertEqual(2, proto.BAR)
+    self.assertEqual(2, unittest_pb2.TestAllTypes.BAR)
+    self.assertEqual(3, proto.BAZ)
+    self.assertEqual(3, unittest_pb2.TestAllTypes.BAZ)
+
+  def testEnum_Name(self):
+    self.assertEqual('FOREIGN_FOO',
+                     unittest_pb2.ForeignEnum.Name(unittest_pb2.FOREIGN_FOO))
+    self.assertEqual('FOREIGN_BAR',
+                     unittest_pb2.ForeignEnum.Name(unittest_pb2.FOREIGN_BAR))
+    self.assertEqual('FOREIGN_BAZ',
+                     unittest_pb2.ForeignEnum.Name(unittest_pb2.FOREIGN_BAZ))
+    self.assertRaises(ValueError,
+                      unittest_pb2.ForeignEnum.Name, 11312)
+
+    proto = unittest_pb2.TestAllTypes()
+    self.assertEqual('FOO',
+                     proto.NestedEnum.Name(proto.FOO))
+    self.assertEqual('FOO',
+                     unittest_pb2.TestAllTypes.NestedEnum.Name(proto.FOO))
+    self.assertEqual('BAR',
+                     proto.NestedEnum.Name(proto.BAR))
+    self.assertEqual('BAR',
+                     unittest_pb2.TestAllTypes.NestedEnum.Name(proto.BAR))
+    self.assertEqual('BAZ',
+                     proto.NestedEnum.Name(proto.BAZ))
+    self.assertEqual('BAZ',
+                     unittest_pb2.TestAllTypes.NestedEnum.Name(proto.BAZ))
+    self.assertRaises(ValueError,
+                      proto.NestedEnum.Name, 11312)
+    self.assertRaises(ValueError,
+                      unittest_pb2.TestAllTypes.NestedEnum.Name, 11312)
+
+  def testEnum_Value(self):
+    self.assertEqual(unittest_pb2.FOREIGN_FOO,
+                     unittest_pb2.ForeignEnum.Value('FOREIGN_FOO'))
+    self.assertEqual(unittest_pb2.FOREIGN_BAR,
+                     unittest_pb2.ForeignEnum.Value('FOREIGN_BAR'))
+    self.assertEqual(unittest_pb2.FOREIGN_BAZ,
+                     unittest_pb2.ForeignEnum.Value('FOREIGN_BAZ'))
+    self.assertRaises(ValueError,
+                      unittest_pb2.ForeignEnum.Value, 'FO')
+
+    proto = unittest_pb2.TestAllTypes()
+    self.assertEqual(proto.FOO,
+                     proto.NestedEnum.Value('FOO'))
+    self.assertEqual(proto.FOO,
+                     unittest_pb2.TestAllTypes.NestedEnum.Value('FOO'))
+    self.assertEqual(proto.BAR,
+                     proto.NestedEnum.Value('BAR'))
+    self.assertEqual(proto.BAR,
+                     unittest_pb2.TestAllTypes.NestedEnum.Value('BAR'))
+    self.assertEqual(proto.BAZ,
+                     proto.NestedEnum.Value('BAZ'))
+    self.assertEqual(proto.BAZ,
+                     unittest_pb2.TestAllTypes.NestedEnum.Value('BAZ'))
+    self.assertRaises(ValueError,
+                      proto.NestedEnum.Value, 'Foo')
+    self.assertRaises(ValueError,
+                      unittest_pb2.TestAllTypes.NestedEnum.Value, 'Foo')
+
+  def testEnum_KeysAndValues(self):
+    self.assertEqual(['FOREIGN_FOO', 'FOREIGN_BAR', 'FOREIGN_BAZ'],
+                     list(unittest_pb2.ForeignEnum.keys()))
+    self.assertEqual([4, 5, 6],
+                     list(unittest_pb2.ForeignEnum.values()))
+    self.assertEqual([('FOREIGN_FOO', 4), ('FOREIGN_BAR', 5),
+                      ('FOREIGN_BAZ', 6)],
+                     list(unittest_pb2.ForeignEnum.items()))
+
+    proto = unittest_pb2.TestAllTypes()
+    self.assertEqual(['FOO', 'BAR', 'BAZ', 'NEG'], list(proto.NestedEnum.keys()))
+    self.assertEqual([1, 2, 3, -1], list(proto.NestedEnum.values()))
+    self.assertEqual([('FOO', 1), ('BAR', 2), ('BAZ', 3), ('NEG', -1)],
+                     list(proto.NestedEnum.items()))
+
+  def testRepeatedScalars(self):
+    proto = unittest_pb2.TestAllTypes()
+
+    self.assertTrue(not proto.repeated_int32)
+    self.assertEqual(0, len(proto.repeated_int32))
+    proto.repeated_int32.append(5)
+    proto.repeated_int32.append(10)
+    proto.repeated_int32.append(15)
+    self.assertTrue(proto.repeated_int32)
+    self.assertEqual(3, len(proto.repeated_int32))
+
+    self.assertEqual([5, 10, 15], proto.repeated_int32)
+
+    # Test single retrieval.
+    self.assertEqual(5, proto.repeated_int32[0])
+    self.assertEqual(15, proto.repeated_int32[-1])
+    # Test out-of-bounds indices.
+    self.assertRaises(IndexError, proto.repeated_int32.__getitem__, 1234)
+    self.assertRaises(IndexError, proto.repeated_int32.__getitem__, -1234)
+    # Test incorrect types passed to __getitem__.
+    self.assertRaises(TypeError, proto.repeated_int32.__getitem__, 'foo')
+    self.assertRaises(TypeError, proto.repeated_int32.__getitem__, None)
+
+    # Test single assignment.
+    proto.repeated_int32[1] = 20
+    self.assertEqual([5, 20, 15], proto.repeated_int32)
+
+    # Test insertion.
+    proto.repeated_int32.insert(1, 25)
+    self.assertEqual([5, 25, 20, 15], proto.repeated_int32)
+
+    # Test slice retrieval.
+    proto.repeated_int32.append(30)
+    self.assertEqual([25, 20, 15], proto.repeated_int32[1:4])
+    self.assertEqual([5, 25, 20, 15, 30], proto.repeated_int32[:])
+
+    # Test slice assignment with an iterator
+    proto.repeated_int32[1:4] = (i for i in range(3))
+    self.assertEqual([5, 0, 1, 2, 30], proto.repeated_int32)
+
+    # Test slice assignment.
+    proto.repeated_int32[1:4] = [35, 40, 45]
+    self.assertEqual([5, 35, 40, 45, 30], proto.repeated_int32)
+
+    # Test that we can use the field as an iterator.
+    result = []
+    for i in proto.repeated_int32:
+      result.append(i)
+    self.assertEqual([5, 35, 40, 45, 30], result)
+
+    # Test single deletion.
+    del proto.repeated_int32[2]
+    self.assertEqual([5, 35, 45, 30], proto.repeated_int32)
+
+    # Test slice deletion.
+    del proto.repeated_int32[2:]
+    self.assertEqual([5, 35], proto.repeated_int32)
+
+    # Test extending.
+    proto.repeated_int32.extend([3, 13])
+    self.assertEqual([5, 35, 3, 13], proto.repeated_int32)
+
+    # Test clearing.
+    proto.ClearField('repeated_int32')
+    self.assertTrue(not proto.repeated_int32)
+    self.assertEqual(0, len(proto.repeated_int32))
+
+    proto.repeated_int32.append(1)
+    self.assertEqual(1, proto.repeated_int32[-1])
+    # Test assignment to a negative index.
+    proto.repeated_int32[-1] = 2
+    self.assertEqual(2, proto.repeated_int32[-1])
+
+    # Test deletion at negative indices.
+    proto.repeated_int32[:] = [0, 1, 2, 3]
+    del proto.repeated_int32[-1]
+    self.assertEqual([0, 1, 2], proto.repeated_int32)
+
+    del proto.repeated_int32[-2]
+    self.assertEqual([0, 2], proto.repeated_int32)
+
+    self.assertRaises(IndexError, proto.repeated_int32.__delitem__, -3)
+    self.assertRaises(IndexError, proto.repeated_int32.__delitem__, 300)
+
+    del proto.repeated_int32[-2:-1]
+    self.assertEqual([2], proto.repeated_int32)
+
+    del proto.repeated_int32[100:10000]
+    self.assertEqual([2], proto.repeated_int32)
+
+  def testRepeatedScalarsRemove(self):
+    proto = unittest_pb2.TestAllTypes()
+
+    self.assertTrue(not proto.repeated_int32)
+    self.assertEqual(0, len(proto.repeated_int32))
+    proto.repeated_int32.append(5)
+    proto.repeated_int32.append(10)
+    proto.repeated_int32.append(5)
+    proto.repeated_int32.append(5)
+
+    self.assertEqual(4, len(proto.repeated_int32))
+    proto.repeated_int32.remove(5)
+    self.assertEqual(3, len(proto.repeated_int32))
+    self.assertEqual(10, proto.repeated_int32[0])
+    self.assertEqual(5, proto.repeated_int32[1])
+    self.assertEqual(5, proto.repeated_int32[2])
+
+    proto.repeated_int32.remove(5)
+    self.assertEqual(2, len(proto.repeated_int32))
+    self.assertEqual(10, proto.repeated_int32[0])
+    self.assertEqual(5, proto.repeated_int32[1])
+
+    proto.repeated_int32.remove(10)
+    self.assertEqual(1, len(proto.repeated_int32))
+    self.assertEqual(5, proto.repeated_int32[0])
+
+    # Remove a non-existent element.
+    self.assertRaises(ValueError, proto.repeated_int32.remove, 123)
+
+  def testRepeatedComposites(self):
+    proto = unittest_pb2.TestAllTypes()
+    self.assertTrue(not proto.repeated_nested_message)
+    self.assertEqual(0, len(proto.repeated_nested_message))
+    m0 = proto.repeated_nested_message.add()
+    m1 = proto.repeated_nested_message.add()
+    self.assertTrue(proto.repeated_nested_message)
+    self.assertEqual(2, len(proto.repeated_nested_message))
+    self.assertListsEqual([m0, m1], proto.repeated_nested_message)
+    self.assertIsInstance(m0, unittest_pb2.TestAllTypes.NestedMessage)
+
+    # Test out-of-bounds indices.
+    self.assertRaises(IndexError, proto.repeated_nested_message.__getitem__,
+                      1234)
+    self.assertRaises(IndexError, proto.repeated_nested_message.__getitem__,
+                      -1234)
+
+    # Test incorrect types passed to __getitem__.
+    self.assertRaises(TypeError, proto.repeated_nested_message.__getitem__,
+                      'foo')
+    self.assertRaises(TypeError, proto.repeated_nested_message.__getitem__,
+                      None)
+
+    # Test slice retrieval.
+    m2 = proto.repeated_nested_message.add()
+    m3 = proto.repeated_nested_message.add()
+    m4 = proto.repeated_nested_message.add()
+    self.assertListsEqual(
+        [m1, m2, m3], proto.repeated_nested_message[1:4])
+    self.assertListsEqual(
+        [m0, m1, m2, m3, m4], proto.repeated_nested_message[:])
+    self.assertListsEqual(
+        [m0, m1], proto.repeated_nested_message[:2])
+    self.assertListsEqual(
+        [m2, m3, m4], proto.repeated_nested_message[2:])
+    self.assertEqual(
+        m0, proto.repeated_nested_message[0])
+    self.assertListsEqual(
+        [m0], proto.repeated_nested_message[:1])
+
+    # Test that we can use the field as an iterator.
+    result = []
+    for i in proto.repeated_nested_message:
+      result.append(i)
+    self.assertListsEqual([m0, m1, m2, m3, m4], result)
+
+    # Test single deletion.
+    del proto.repeated_nested_message[2]
+    self.assertListsEqual([m0, m1, m3, m4], proto.repeated_nested_message)
+
+    # Test slice deletion.
+    del proto.repeated_nested_message[2:]
+    self.assertListsEqual([m0, m1], proto.repeated_nested_message)
+
+    # Test extending.
+    n1 = unittest_pb2.TestAllTypes.NestedMessage(bb=1)
+    n2 = unittest_pb2.TestAllTypes.NestedMessage(bb=2)
+    proto.repeated_nested_message.extend([n1,n2])
+    self.assertEqual(4, len(proto.repeated_nested_message))
+    self.assertEqual(n1, proto.repeated_nested_message[2])
+    self.assertEqual(n2, proto.repeated_nested_message[3])
+
+    # Test clearing.
+    proto.ClearField('repeated_nested_message')
+    self.assertTrue(not proto.repeated_nested_message)
+    self.assertEqual(0, len(proto.repeated_nested_message))
+
+    # Test constructing an element while adding it.
+    proto.repeated_nested_message.add(bb=23)
+    self.assertEqual(1, len(proto.repeated_nested_message))
+    self.assertEqual(23, proto.repeated_nested_message[0].bb)
+
+  def testRepeatedCompositeRemove(self):
+    proto = unittest_pb2.TestAllTypes()
+
+    self.assertEqual(0, len(proto.repeated_nested_message))
+    m0 = proto.repeated_nested_message.add()
+    # Need to set some differentiating variable so m0 != m1 != m2:
+    m0.bb = len(proto.repeated_nested_message)
+    m1 = proto.repeated_nested_message.add()
+    m1.bb = len(proto.repeated_nested_message)
+    self.assertTrue(m0 != m1)
+    m2 = proto.repeated_nested_message.add()
+    m2.bb = len(proto.repeated_nested_message)
+    self.assertListsEqual([m0, m1, m2], proto.repeated_nested_message)
+
+    self.assertEqual(3, len(proto.repeated_nested_message))
+    proto.repeated_nested_message.remove(m0)
+    self.assertEqual(2, len(proto.repeated_nested_message))
+    self.assertEqual(m1, proto.repeated_nested_message[0])
+    self.assertEqual(m2, proto.repeated_nested_message[1])
+
+    # Removing m0 again or removing None should raise error
+    self.assertRaises(ValueError, proto.repeated_nested_message.remove, m0)
+    self.assertRaises(ValueError, proto.repeated_nested_message.remove, None)
+    self.assertEqual(2, len(proto.repeated_nested_message))
+
+    proto.repeated_nested_message.remove(m2)
+    self.assertEqual(1, len(proto.repeated_nested_message))
+    self.assertEqual(m1, proto.repeated_nested_message[0])
+
+  def testHandWrittenReflection(self):
+    # Hand written extensions are only supported by the pure-Python
+    # implementation of the API.
+    if api_implementation.Type() != 'python':
+      return
+
+    FieldDescriptor = descriptor.FieldDescriptor
+    foo_field_descriptor = FieldDescriptor(
+        name='foo_field', full_name='MyProto.foo_field',
+        index=0, number=1, type=FieldDescriptor.TYPE_INT64,
+        cpp_type=FieldDescriptor.CPPTYPE_INT64,
+        label=FieldDescriptor.LABEL_OPTIONAL, default_value=0,
+        containing_type=None, message_type=None, enum_type=None,
+        is_extension=False, extension_scope=None,
+        options=descriptor_pb2.FieldOptions())
+    mydescriptor = descriptor.Descriptor(
+        name='MyProto', full_name='MyProto', filename='ignored',
+        containing_type=None, nested_types=[], enum_types=[],
+        fields=[foo_field_descriptor], extensions=[],
+        options=descriptor_pb2.MessageOptions())
+    class MyProtoClass(six.with_metaclass(reflection.GeneratedProtocolMessageType, message.Message)):
+      DESCRIPTOR = mydescriptor
+    myproto_instance = MyProtoClass()
+    self.assertEqual(0, myproto_instance.foo_field)
+    self.assertTrue(not myproto_instance.HasField('foo_field'))
+    myproto_instance.foo_field = 23
+    self.assertEqual(23, myproto_instance.foo_field)
+    self.assertTrue(myproto_instance.HasField('foo_field'))
+
+  def testDescriptorProtoSupport(self):
+    # Hand written descriptors/reflection are only supported by the pure-Python
+    # implementation of the API.
+    if api_implementation.Type() != 'python':
+      return
+
+    def AddDescriptorField(proto, field_name, field_type):
+      AddDescriptorField.field_index += 1
+      new_field = proto.field.add()
+      new_field.name = field_name
+      new_field.type = field_type
+      new_field.number = AddDescriptorField.field_index
+      new_field.label = descriptor_pb2.FieldDescriptorProto.LABEL_OPTIONAL
+
+    AddDescriptorField.field_index = 0
+
+    desc_proto = descriptor_pb2.DescriptorProto()
+    desc_proto.name = 'Car'
+    fdp = descriptor_pb2.FieldDescriptorProto
+    AddDescriptorField(desc_proto, 'name', fdp.TYPE_STRING)
+    AddDescriptorField(desc_proto, 'year', fdp.TYPE_INT64)
+    AddDescriptorField(desc_proto, 'automatic', fdp.TYPE_BOOL)
+    AddDescriptorField(desc_proto, 'price', fdp.TYPE_DOUBLE)
+    # Add a repeated field
+    AddDescriptorField.field_index += 1
+    new_field = desc_proto.field.add()
+    new_field.name = 'owners'
+    new_field.type = fdp.TYPE_STRING
+    new_field.number = AddDescriptorField.field_index
+    new_field.label = descriptor_pb2.FieldDescriptorProto.LABEL_REPEATED
+
+    desc = descriptor.MakeDescriptor(desc_proto)
+    self.assertTrue('name' in desc.fields_by_name)
+    self.assertTrue('year' in desc.fields_by_name)
+    self.assertTrue('automatic' in desc.fields_by_name)
+    self.assertTrue('price' in desc.fields_by_name)
+    self.assertTrue('owners' in desc.fields_by_name)
+
+    class CarMessage(six.with_metaclass(reflection.GeneratedProtocolMessageType, message.Message)):
+      DESCRIPTOR = desc
+
+    prius = CarMessage()
+    prius.name = 'prius'
+    prius.year = 2010
+    prius.automatic = True
+    prius.price = 25134.75
+    prius.owners.extend(['bob', 'susan'])
+
+    serialized_prius = prius.SerializeToString()
+    new_prius = reflection.ParseMessage(desc, serialized_prius)
+    self.assertTrue(new_prius is not prius)
+    self.assertEqual(prius, new_prius)
+
+    # these are unnecessary assuming message equality works as advertised but
+    # explicitly check to be safe since we're mucking about in metaclass foo
+    self.assertEqual(prius.name, new_prius.name)
+    self.assertEqual(prius.year, new_prius.year)
+    self.assertEqual(prius.automatic, new_prius.automatic)
+    self.assertEqual(prius.price, new_prius.price)
+    self.assertEqual(prius.owners, new_prius.owners)
+
+  def testTopLevelExtensionsForOptionalScalar(self):
+    extendee_proto = unittest_pb2.TestAllExtensions()
+    extension = unittest_pb2.optional_int32_extension
+    self.assertTrue(not extendee_proto.HasExtension(extension))
+    self.assertEqual(0, extendee_proto.Extensions[extension])
+    # As with normal scalar fields, just doing a read doesn't actually set the
+    # "has" bit.
+    self.assertTrue(not extendee_proto.HasExtension(extension))
+    # Actually set the thing.
+    extendee_proto.Extensions[extension] = 23
+    self.assertEqual(23, extendee_proto.Extensions[extension])
+    self.assertTrue(extendee_proto.HasExtension(extension))
+    # Ensure that clearing works as well.
+    extendee_proto.ClearExtension(extension)
+    self.assertEqual(0, extendee_proto.Extensions[extension])
+    self.assertTrue(not extendee_proto.HasExtension(extension))
+
+  def testTopLevelExtensionsForRepeatedScalar(self):
+    extendee_proto = unittest_pb2.TestAllExtensions()
+    extension = unittest_pb2.repeated_string_extension
+    self.assertEqual(0, len(extendee_proto.Extensions[extension]))
+    extendee_proto.Extensions[extension].append('foo')
+    self.assertEqual(['foo'], extendee_proto.Extensions[extension])
+    string_list = extendee_proto.Extensions[extension]
+    extendee_proto.ClearExtension(extension)
+    self.assertEqual(0, len(extendee_proto.Extensions[extension]))
+    self.assertTrue(string_list is not extendee_proto.Extensions[extension])
+    # Shouldn't be allowed to do Extensions[extension] = 'a'
+    self.assertRaises(TypeError, operator.setitem, extendee_proto.Extensions,
+                      extension, 'a')
+
+  def testTopLevelExtensionsForOptionalMessage(self):
+    extendee_proto = unittest_pb2.TestAllExtensions()
+    extension = unittest_pb2.optional_foreign_message_extension
+    self.assertTrue(not extendee_proto.HasExtension(extension))
+    self.assertEqual(0, extendee_proto.Extensions[extension].c)
+    # As with normal (non-extension) fields, merely reading from the
+    # thing shouldn't set the "has" bit.
+    self.assertTrue(not extendee_proto.HasExtension(extension))
+    extendee_proto.Extensions[extension].c = 23
+    self.assertEqual(23, extendee_proto.Extensions[extension].c)
+    self.assertTrue(extendee_proto.HasExtension(extension))
+    # Save a reference here.
+    foreign_message = extendee_proto.Extensions[extension]
+    extendee_proto.ClearExtension(extension)
+    self.assertTrue(foreign_message is not extendee_proto.Extensions[extension])
+    # Setting a field on foreign_message now shouldn't set
+    # any "has" bits on extendee_proto.
+    foreign_message.c = 42
+    self.assertEqual(42, foreign_message.c)
+    self.assertTrue(foreign_message.HasField('c'))
+    self.assertTrue(not extendee_proto.HasExtension(extension))
+    # Shouldn't be allowed to do Extensions[extension] = 'a'
+    self.assertRaises(TypeError, operator.setitem, extendee_proto.Extensions,
+                      extension, 'a')
+
+  def testTopLevelExtensionsForRepeatedMessage(self):
+    extendee_proto = unittest_pb2.TestAllExtensions()
+    extension = unittest_pb2.repeatedgroup_extension
+    self.assertEqual(0, len(extendee_proto.Extensions[extension]))
+    group = extendee_proto.Extensions[extension].add()
+    group.a = 23
+    self.assertEqual(23, extendee_proto.Extensions[extension][0].a)
+    group.a = 42
+    self.assertEqual(42, extendee_proto.Extensions[extension][0].a)
+    group_list = extendee_proto.Extensions[extension]
+    extendee_proto.ClearExtension(extension)
+    self.assertEqual(0, len(extendee_proto.Extensions[extension]))
+    self.assertTrue(group_list is not extendee_proto.Extensions[extension])
+    # Shouldn't be allowed to do Extensions[extension] = 'a'
+    self.assertRaises(TypeError, operator.setitem, extendee_proto.Extensions,
+                      extension, 'a')
+
+  def testNestedExtensions(self):
+    extendee_proto = unittest_pb2.TestAllExtensions()
+    extension = unittest_pb2.TestRequired.single
+
+    # We just test the non-repeated case.
+    self.assertTrue(not extendee_proto.HasExtension(extension))
+    required = extendee_proto.Extensions[extension]
+    self.assertEqual(0, required.a)
+    self.assertTrue(not extendee_proto.HasExtension(extension))
+    required.a = 23
+    self.assertEqual(23, extendee_proto.Extensions[extension].a)
+    self.assertTrue(extendee_proto.HasExtension(extension))
+    extendee_proto.ClearExtension(extension)
+    self.assertTrue(required is not extendee_proto.Extensions[extension])
+    self.assertTrue(not extendee_proto.HasExtension(extension))
+
+  def testRegisteredExtensions(self):
+    self.assertTrue('protobuf_unittest.optional_int32_extension' in
+                    unittest_pb2.TestAllExtensions._extensions_by_name)
+    self.assertTrue(1 in unittest_pb2.TestAllExtensions._extensions_by_number)
+    # Make sure extensions haven't been registered into types that shouldn't
+    # have any.
+    self.assertEqual(0, len(unittest_pb2.TestAllTypes._extensions_by_name))
+
+  # If message A directly contains message B, and
+  # a.HasField('b') is currently False, then mutating any
+  # extension in B should change a.HasField('b') to True
+  # (and so on up the object tree).
+  def testHasBitsForAncestorsOfExtendedMessage(self):
+    # Optional scalar extension.
+    toplevel = more_extensions_pb2.TopLevelMessage()
+    self.assertTrue(not toplevel.HasField('submessage'))
+    self.assertEqual(0, toplevel.submessage.Extensions[
+        more_extensions_pb2.optional_int_extension])
+    self.assertTrue(not toplevel.HasField('submessage'))
+    toplevel.submessage.Extensions[
+        more_extensions_pb2.optional_int_extension] = 23
+    self.assertEqual(23, toplevel.submessage.Extensions[
+        more_extensions_pb2.optional_int_extension])
+    self.assertTrue(toplevel.HasField('submessage'))
+
+    # Repeated scalar extension.
+    toplevel = more_extensions_pb2.TopLevelMessage()
+    self.assertTrue(not toplevel.HasField('submessage'))
+    self.assertEqual([], toplevel.submessage.Extensions[
+        more_extensions_pb2.repeated_int_extension])
+    self.assertTrue(not toplevel.HasField('submessage'))
+    toplevel.submessage.Extensions[
+        more_extensions_pb2.repeated_int_extension].append(23)
+    self.assertEqual([23], toplevel.submessage.Extensions[
+        more_extensions_pb2.repeated_int_extension])
+    self.assertTrue(toplevel.HasField('submessage'))
+
+    # Optional message extension.
+    toplevel = more_extensions_pb2.TopLevelMessage()
+    self.assertTrue(not toplevel.HasField('submessage'))
+    self.assertEqual(0, toplevel.submessage.Extensions[
+        more_extensions_pb2.optional_message_extension].foreign_message_int)
+    self.assertTrue(not toplevel.HasField('submessage'))
+    toplevel.submessage.Extensions[
+        more_extensions_pb2.optional_message_extension].foreign_message_int = 23
+    self.assertEqual(23, toplevel.submessage.Extensions[
+        more_extensions_pb2.optional_message_extension].foreign_message_int)
+    self.assertTrue(toplevel.HasField('submessage'))
+
+    # Repeated message extension.
+    toplevel = more_extensions_pb2.TopLevelMessage()
+    self.assertTrue(not toplevel.HasField('submessage'))
+    self.assertEqual(0, len(toplevel.submessage.Extensions[
+        more_extensions_pb2.repeated_message_extension]))
+    self.assertTrue(not toplevel.HasField('submessage'))
+    foreign = toplevel.submessage.Extensions[
+        more_extensions_pb2.repeated_message_extension].add()
+    self.assertEqual(foreign, toplevel.submessage.Extensions[
+        more_extensions_pb2.repeated_message_extension][0])
+    self.assertTrue(toplevel.HasField('submessage'))
+
+  def testDisconnectionAfterClearingEmptyMessage(self):
+    toplevel = more_extensions_pb2.TopLevelMessage()
+    extendee_proto = toplevel.submessage
+    extension = more_extensions_pb2.optional_message_extension
+    extension_proto = extendee_proto.Extensions[extension]
+    extendee_proto.ClearExtension(extension)
+    extension_proto.foreign_message_int = 23
+
+    self.assertTrue(extension_proto is not extendee_proto.Extensions[extension])
+
+  def testExtensionFailureModes(self):
+    extendee_proto = unittest_pb2.TestAllExtensions()
+
+    # Try non-extension-handle arguments to HasExtension,
+    # ClearExtension(), and Extensions[]...
+    self.assertRaises(KeyError, extendee_proto.HasExtension, 1234)
+    self.assertRaises(KeyError, extendee_proto.ClearExtension, 1234)
+    self.assertRaises(KeyError, extendee_proto.Extensions.__getitem__, 1234)
+    self.assertRaises(KeyError, extendee_proto.Extensions.__setitem__, 1234, 5)
+
+    # Try something that *is* an extension handle, just not for
+    # this message...
+    for unknown_handle in (more_extensions_pb2.optional_int_extension,
+                           more_extensions_pb2.optional_message_extension,
+                           more_extensions_pb2.repeated_int_extension,
+                           more_extensions_pb2.repeated_message_extension):
+      self.assertRaises(KeyError, extendee_proto.HasExtension,
+                        unknown_handle)
+      self.assertRaises(KeyError, extendee_proto.ClearExtension,
+                        unknown_handle)
+      self.assertRaises(KeyError, extendee_proto.Extensions.__getitem__,
+                        unknown_handle)
+      self.assertRaises(KeyError, extendee_proto.Extensions.__setitem__,
+                        unknown_handle, 5)
+
+    # Try call HasExtension() with a valid handle, but for a
+    # *repeated* field.  (Just as with non-extension repeated
+    # fields, Has*() isn't supported for extension repeated fields).
+    self.assertRaises(KeyError, extendee_proto.HasExtension,
+                      unittest_pb2.repeated_string_extension)
+
+  def testStaticParseFrom(self):
+    proto1 = unittest_pb2.TestAllTypes()
+    test_util.SetAllFields(proto1)
+
+    string1 = proto1.SerializeToString()
+    proto2 = unittest_pb2.TestAllTypes.FromString(string1)
+
+    # Messages should be equal.
+    self.assertEqual(proto2, proto1)
+
+  def testMergeFromSingularField(self):
+    # Test merge with just a singular field.
+    proto1 = unittest_pb2.TestAllTypes()
+    proto1.optional_int32 = 1
+
+    proto2 = unittest_pb2.TestAllTypes()
+    # This shouldn't get overwritten.
+    proto2.optional_string = 'value'
+
+    proto2.MergeFrom(proto1)
+    self.assertEqual(1, proto2.optional_int32)
+    self.assertEqual('value', proto2.optional_string)
+
+  def testMergeFromRepeatedField(self):
+    # Test merge with just a repeated field.
+    proto1 = unittest_pb2.TestAllTypes()
+    proto1.repeated_int32.append(1)
+    proto1.repeated_int32.append(2)
+
+    proto2 = unittest_pb2.TestAllTypes()
+    proto2.repeated_int32.append(0)
+    proto2.MergeFrom(proto1)
+
+    self.assertEqual(0, proto2.repeated_int32[0])
+    self.assertEqual(1, proto2.repeated_int32[1])
+    self.assertEqual(2, proto2.repeated_int32[2])
+
+  def testMergeFromOptionalGroup(self):
+    # Test merge with an optional group.
+    proto1 = unittest_pb2.TestAllTypes()
+    proto1.optionalgroup.a = 12
+    proto2 = unittest_pb2.TestAllTypes()
+    proto2.MergeFrom(proto1)
+    self.assertEqual(12, proto2.optionalgroup.a)
+
+  def testMergeFromRepeatedNestedMessage(self):
+    # Test merge with a repeated nested message.
+    proto1 = unittest_pb2.TestAllTypes()
+    m = proto1.repeated_nested_message.add()
+    m.bb = 123
+    m = proto1.repeated_nested_message.add()
+    m.bb = 321
+
+    proto2 = unittest_pb2.TestAllTypes()
+    m = proto2.repeated_nested_message.add()
+    m.bb = 999
+    proto2.MergeFrom(proto1)
+    self.assertEqual(999, proto2.repeated_nested_message[0].bb)
+    self.assertEqual(123, proto2.repeated_nested_message[1].bb)
+    self.assertEqual(321, proto2.repeated_nested_message[2].bb)
+
+    proto3 = unittest_pb2.TestAllTypes()
+    proto3.repeated_nested_message.MergeFrom(proto2.repeated_nested_message)
+    self.assertEqual(999, proto3.repeated_nested_message[0].bb)
+    self.assertEqual(123, proto3.repeated_nested_message[1].bb)
+    self.assertEqual(321, proto3.repeated_nested_message[2].bb)
+
+  def testMergeFromAllFields(self):
+    # With all fields set.
+    proto1 = unittest_pb2.TestAllTypes()
+    test_util.SetAllFields(proto1)
+    proto2 = unittest_pb2.TestAllTypes()
+    proto2.MergeFrom(proto1)
+
+    # Messages should be equal.
+    self.assertEqual(proto2, proto1)
+
+    # Serialized string should be equal too.
+    string1 = proto1.SerializeToString()
+    string2 = proto2.SerializeToString()
+    self.assertEqual(string1, string2)
+
+  def testMergeFromExtensionsSingular(self):
+    proto1 = unittest_pb2.TestAllExtensions()
+    proto1.Extensions[unittest_pb2.optional_int32_extension] = 1
+
+    proto2 = unittest_pb2.TestAllExtensions()
+    proto2.MergeFrom(proto1)
+    self.assertEqual(
+        1, proto2.Extensions[unittest_pb2.optional_int32_extension])
+
+  def testMergeFromExtensionsRepeated(self):
+    proto1 = unittest_pb2.TestAllExtensions()
+    proto1.Extensions[unittest_pb2.repeated_int32_extension].append(1)
+    proto1.Extensions[unittest_pb2.repeated_int32_extension].append(2)
+
+    proto2 = unittest_pb2.TestAllExtensions()
+    proto2.Extensions[unittest_pb2.repeated_int32_extension].append(0)
+    proto2.MergeFrom(proto1)
+    self.assertEqual(
+        3, len(proto2.Extensions[unittest_pb2.repeated_int32_extension]))
+    self.assertEqual(
+        0, proto2.Extensions[unittest_pb2.repeated_int32_extension][0])
+    self.assertEqual(
+        1, proto2.Extensions[unittest_pb2.repeated_int32_extension][1])
+    self.assertEqual(
+        2, proto2.Extensions[unittest_pb2.repeated_int32_extension][2])
+
+  def testMergeFromExtensionsNestedMessage(self):
+    proto1 = unittest_pb2.TestAllExtensions()
+    ext1 = proto1.Extensions[
+        unittest_pb2.repeated_nested_message_extension]
+    m = ext1.add()
+    m.bb = 222
+    m = ext1.add()
+    m.bb = 333
+
+    proto2 = unittest_pb2.TestAllExtensions()
+    ext2 = proto2.Extensions[
+        unittest_pb2.repeated_nested_message_extension]
+    m = ext2.add()
+    m.bb = 111
+
+    proto2.MergeFrom(proto1)
+    ext2 = proto2.Extensions[
+        unittest_pb2.repeated_nested_message_extension]
+    self.assertEqual(3, len(ext2))
+    self.assertEqual(111, ext2[0].bb)
+    self.assertEqual(222, ext2[1].bb)
+    self.assertEqual(333, ext2[2].bb)
+
+  def testMergeFromBug(self):
+    message1 = unittest_pb2.TestAllTypes()
+    message2 = unittest_pb2.TestAllTypes()
+
+    # Cause optional_nested_message to be instantiated within message1, even
+    # though it is not considered to be "present".
+    message1.optional_nested_message
+    self.assertFalse(message1.HasField('optional_nested_message'))
+
+    # Merge into message2.  This should not instantiate the field is message2.
+    message2.MergeFrom(message1)
+    self.assertFalse(message2.HasField('optional_nested_message'))
+
+  def testCopyFromSingularField(self):
+    # Test copy with just a singular field.
+    proto1 = unittest_pb2.TestAllTypes()
+    proto1.optional_int32 = 1
+    proto1.optional_string = 'important-text'
+
+    proto2 = unittest_pb2.TestAllTypes()
+    proto2.optional_string = 'value'
+
+    proto2.CopyFrom(proto1)
+    self.assertEqual(1, proto2.optional_int32)
+    self.assertEqual('important-text', proto2.optional_string)
+
+  def testCopyFromRepeatedField(self):
+    # Test copy with a repeated field.
+    proto1 = unittest_pb2.TestAllTypes()
+    proto1.repeated_int32.append(1)
+    proto1.repeated_int32.append(2)
+
+    proto2 = unittest_pb2.TestAllTypes()
+    proto2.repeated_int32.append(0)
+    proto2.CopyFrom(proto1)
+
+    self.assertEqual(1, proto2.repeated_int32[0])
+    self.assertEqual(2, proto2.repeated_int32[1])
+
+  def testCopyFromAllFields(self):
+    # With all fields set.
+    proto1 = unittest_pb2.TestAllTypes()
+    test_util.SetAllFields(proto1)
+    proto2 = unittest_pb2.TestAllTypes()
+    proto2.CopyFrom(proto1)
+
+    # Messages should be equal.
+    self.assertEqual(proto2, proto1)
+
+    # Serialized string should be equal too.
+    string1 = proto1.SerializeToString()
+    string2 = proto2.SerializeToString()
+    self.assertEqual(string1, string2)
+
+  def testCopyFromSelf(self):
+    proto1 = unittest_pb2.TestAllTypes()
+    proto1.repeated_int32.append(1)
+    proto1.optional_int32 = 2
+    proto1.optional_string = 'important-text'
+
+    proto1.CopyFrom(proto1)
+    self.assertEqual(1, proto1.repeated_int32[0])
+    self.assertEqual(2, proto1.optional_int32)
+    self.assertEqual('important-text', proto1.optional_string)
+
+  def testCopyFromBadType(self):
+    # The python implementation doesn't raise an exception in this
+    # case. In theory it should.
+    if api_implementation.Type() == 'python':
+      return
+    proto1 = unittest_pb2.TestAllTypes()
+    proto2 = unittest_pb2.TestAllExtensions()
+    self.assertRaises(TypeError, proto1.CopyFrom, proto2)
+
+  def testDeepCopy(self):
+    proto1 = unittest_pb2.TestAllTypes()
+    proto1.optional_int32 = 1
+    proto2 = copy.deepcopy(proto1)
+    self.assertEqual(1, proto2.optional_int32)
+
+    proto1.repeated_int32.append(2)
+    proto1.repeated_int32.append(3)
+    container = copy.deepcopy(proto1.repeated_int32)
+    self.assertEqual([2, 3], container)
+
+    # TODO(anuraag): Implement deepcopy for repeated composite / extension dict
+
+  def testClear(self):
+    proto = unittest_pb2.TestAllTypes()
+    # C++ implementation does not support lazy fields right now so leave it
+    # out for now.
+    if api_implementation.Type() == 'python':
+      test_util.SetAllFields(proto)
+    else:
+      test_util.SetAllNonLazyFields(proto)
+    # Clear the message.
+    proto.Clear()
+    self.assertEqual(proto.ByteSize(), 0)
+    empty_proto = unittest_pb2.TestAllTypes()
+    self.assertEqual(proto, empty_proto)
+
+    # Test if extensions which were set are cleared.
+    proto = unittest_pb2.TestAllExtensions()
+    test_util.SetAllExtensions(proto)
+    # Clear the message.
+    proto.Clear()
+    self.assertEqual(proto.ByteSize(), 0)
+    empty_proto = unittest_pb2.TestAllExtensions()
+    self.assertEqual(proto, empty_proto)
+
+  def testDisconnectingBeforeClear(self):
+    proto = unittest_pb2.TestAllTypes()
+    nested = proto.optional_nested_message
+    proto.Clear()
+    self.assertTrue(nested is not proto.optional_nested_message)
+    nested.bb = 23
+    self.assertTrue(not proto.HasField('optional_nested_message'))
+    self.assertEqual(0, proto.optional_nested_message.bb)
+
+    proto = unittest_pb2.TestAllTypes()
+    nested = proto.optional_nested_message
+    nested.bb = 5
+    foreign = proto.optional_foreign_message
+    foreign.c = 6
+
+    proto.Clear()
+    self.assertTrue(nested is not proto.optional_nested_message)
+    self.assertTrue(foreign is not proto.optional_foreign_message)
+    self.assertEqual(5, nested.bb)
+    self.assertEqual(6, foreign.c)
+    nested.bb = 15
+    foreign.c = 16
+    self.assertFalse(proto.HasField('optional_nested_message'))
+    self.assertEqual(0, proto.optional_nested_message.bb)
+    self.assertFalse(proto.HasField('optional_foreign_message'))
+    self.assertEqual(0, proto.optional_foreign_message.c)
+
+  def testOneOf(self):
+    proto = unittest_pb2.TestAllTypes()
+    proto.oneof_uint32 = 10
+    proto.oneof_nested_message.bb = 11
+    self.assertEqual(11, proto.oneof_nested_message.bb)
+    self.assertFalse(proto.HasField('oneof_uint32'))
+    nested = proto.oneof_nested_message
+    proto.oneof_string = 'abc'
+    self.assertEqual('abc', proto.oneof_string)
+    self.assertEqual(11, nested.bb)
+    self.assertFalse(proto.HasField('oneof_nested_message'))
+
+  def assertInitialized(self, proto):
+    self.assertTrue(proto.IsInitialized())
+    # Neither method should raise an exception.
+    proto.SerializeToString()
+    proto.SerializePartialToString()
+
+  def assertNotInitialized(self, proto):
+    self.assertFalse(proto.IsInitialized())
+    self.assertRaises(message.EncodeError, proto.SerializeToString)
+    # "Partial" serialization doesn't care if message is uninitialized.
+    proto.SerializePartialToString()
+
+  def testIsInitialized(self):
+    # Trivial cases - all optional fields and extensions.
+    proto = unittest_pb2.TestAllTypes()
+    self.assertInitialized(proto)
+    proto = unittest_pb2.TestAllExtensions()
+    self.assertInitialized(proto)
+
+    # The case of uninitialized required fields.
+    proto = unittest_pb2.TestRequired()
+    self.assertNotInitialized(proto)
+    proto.a = proto.b = proto.c = 2
+    self.assertInitialized(proto)
+
+    # The case of uninitialized submessage.
+    proto = unittest_pb2.TestRequiredForeign()
+    self.assertInitialized(proto)
+    proto.optional_message.a = 1
+    self.assertNotInitialized(proto)
+    proto.optional_message.b = 0
+    proto.optional_message.c = 0
+    self.assertInitialized(proto)
+
+    # Uninitialized repeated submessage.
+    message1 = proto.repeated_message.add()
+    self.assertNotInitialized(proto)
+    message1.a = message1.b = message1.c = 0
+    self.assertInitialized(proto)
+
+    # Uninitialized repeated group in an extension.
+    proto = unittest_pb2.TestAllExtensions()
+    extension = unittest_pb2.TestRequired.multi
+    message1 = proto.Extensions[extension].add()
+    message2 = proto.Extensions[extension].add()
+    self.assertNotInitialized(proto)
+    message1.a = 1
+    message1.b = 1
+    message1.c = 1
+    self.assertNotInitialized(proto)
+    message2.a = 2
+    message2.b = 2
+    message2.c = 2
+    self.assertInitialized(proto)
+
+    # Uninitialized nonrepeated message in an extension.
+    proto = unittest_pb2.TestAllExtensions()
+    extension = unittest_pb2.TestRequired.single
+    proto.Extensions[extension].a = 1
+    self.assertNotInitialized(proto)
+    proto.Extensions[extension].b = 2
+    proto.Extensions[extension].c = 3
+    self.assertInitialized(proto)
+
+    # Try passing an errors list.
+    errors = []
+    proto = unittest_pb2.TestRequired()
+    self.assertFalse(proto.IsInitialized(errors))
+    self.assertEqual(errors, ['a', 'b', 'c'])
+
+  @unittest.skipIf(
+      api_implementation.Type() != 'cpp' or api_implementation.Version() != 2,
+      'Errors are only available from the most recent C++ implementation.')
+  def testFileDescriptorErrors(self):
+    file_name = 'test_file_descriptor_errors.proto'
+    package_name = 'test_file_descriptor_errors.proto'
+    file_descriptor_proto = descriptor_pb2.FileDescriptorProto()
+    file_descriptor_proto.name = file_name
+    file_descriptor_proto.package = package_name
+    m1 = file_descriptor_proto.message_type.add()
+    m1.name = 'msg1'
+    # Compiles the proto into the C++ descriptor pool
+    descriptor.FileDescriptor(
+        file_name,
+        package_name,
+        serialized_pb=file_descriptor_proto.SerializeToString())
+    # Add a FileDescriptorProto that has duplicate symbols
+    another_file_name = 'another_test_file_descriptor_errors.proto'
+    file_descriptor_proto.name = another_file_name
+    m2 = file_descriptor_proto.message_type.add()
+    m2.name = 'msg2'
+    with self.assertRaises(TypeError) as cm:
+      descriptor.FileDescriptor(
+          another_file_name,
+          package_name,
+          serialized_pb=file_descriptor_proto.SerializeToString())
+      self.assertTrue(hasattr(cm, 'exception'), '%s not raised' %
+                      getattr(cm.expected, '__name__', cm.expected))
+      self.assertIn('test_file_descriptor_errors.proto', str(cm.exception))
+      # Error message will say something about this definition being a
+      # duplicate, though we don't check the message exactly to avoid a
+      # dependency on the C++ logging code.
+      self.assertIn('test_file_descriptor_errors.msg1', str(cm.exception))
+
+  def testStringUTF8Encoding(self):
+    proto = unittest_pb2.TestAllTypes()
+
+    # Assignment of a unicode object to a field of type 'bytes' is not allowed.
+    self.assertRaises(TypeError,
+                      setattr, proto, 'optional_bytes', u'unicode object')
+
+    # Check that the default value is of python's 'unicode' type.
+    self.assertEqual(type(proto.optional_string), six.text_type)
+
+    proto.optional_string = six.text_type('Testing')
+    self.assertEqual(proto.optional_string, str('Testing'))
+
+    # Assign a value of type 'str' which can be encoded in UTF-8.
+    proto.optional_string = str('Testing')
+    self.assertEqual(proto.optional_string, six.text_type('Testing'))
+
+    # Try to assign a 'bytes' object which contains non-UTF-8.
+    self.assertRaises(ValueError,
+                      setattr, proto, 'optional_string', b'a\x80a')
+    # No exception: Assign already encoded UTF-8 bytes to a string field.
+    utf8_bytes = u'Тест'.encode('utf-8')
+    proto.optional_string = utf8_bytes
+    # No exception: Assign the a non-ascii unicode object.
+    proto.optional_string = u'Тест'
+    # No exception thrown (normal str assignment containing ASCII).
+    proto.optional_string = 'abc'
+
+  def testStringUTF8Serialization(self):
+    proto = message_set_extensions_pb2.TestMessageSet()
+    extension_message = message_set_extensions_pb2.TestMessageSetExtension2
+    extension = extension_message.message_set_extension
+
+    test_utf8 = u'Тест'
+    test_utf8_bytes = test_utf8.encode('utf-8')
+
+    # 'Test' in another language, using UTF-8 charset.
+    proto.Extensions[extension].str = test_utf8
+
+    # Serialize using the MessageSet wire format (this is specified in the
+    # .proto file).
+    serialized = proto.SerializeToString()
+
+    # Check byte size.
+    self.assertEqual(proto.ByteSize(), len(serialized))
+
+    raw = unittest_mset_pb2.RawMessageSet()
+    bytes_read = raw.MergeFromString(serialized)
+    self.assertEqual(len(serialized), bytes_read)
+
+    message2 = message_set_extensions_pb2.TestMessageSetExtension2()
+
+    self.assertEqual(1, len(raw.item))
+    # Check that the type_id is the same as the tag ID in the .proto file.
+    self.assertEqual(raw.item[0].type_id, 98418634)
+
+    # Check the actual bytes on the wire.
+    self.assertTrue(raw.item[0].message.endswith(test_utf8_bytes))
+    bytes_read = message2.MergeFromString(raw.item[0].message)
+    self.assertEqual(len(raw.item[0].message), bytes_read)
+
+    self.assertEqual(type(message2.str), six.text_type)
+    self.assertEqual(message2.str, test_utf8)
+
+    # The pure Python API throws an exception on MergeFromString(),
+    # if any of the string fields of the message can't be UTF-8 decoded.
+    # The C++ implementation of the API has no way to check that on
+    # MergeFromString and thus has no way to throw the exception.
+    #
+    # The pure Python API always returns objects of type 'unicode' (UTF-8
+    # encoded), or 'bytes' (in 7 bit ASCII).
+    badbytes = raw.item[0].message.replace(
+        test_utf8_bytes, len(test_utf8_bytes) * b'\xff')
+
+    unicode_decode_failed = False
+    try:
+      message2.MergeFromString(badbytes)
+    except UnicodeDecodeError:
+      unicode_decode_failed = True
+    string_field = message2.str
+    self.assertTrue(unicode_decode_failed or type(string_field) is bytes)
+
+  def testBytesInTextFormat(self):
+    proto = unittest_pb2.TestAllTypes(optional_bytes=b'\x00\x7f\x80\xff')
+    self.assertEqual(u'optional_bytes: "\\000\\177\\200\\377"\n',
+                     six.text_type(proto))
+
+  def testEmptyNestedMessage(self):
+    proto = unittest_pb2.TestAllTypes()
+    proto.optional_nested_message.MergeFrom(
+        unittest_pb2.TestAllTypes.NestedMessage())
+    self.assertTrue(proto.HasField('optional_nested_message'))
+
+    proto = unittest_pb2.TestAllTypes()
+    proto.optional_nested_message.CopyFrom(
+        unittest_pb2.TestAllTypes.NestedMessage())
+    self.assertTrue(proto.HasField('optional_nested_message'))
+
+    proto = unittest_pb2.TestAllTypes()
+    bytes_read = proto.optional_nested_message.MergeFromString(b'')
+    self.assertEqual(0, bytes_read)
+    self.assertTrue(proto.HasField('optional_nested_message'))
+
+    proto = unittest_pb2.TestAllTypes()
+    proto.optional_nested_message.ParseFromString(b'')
+    self.assertTrue(proto.HasField('optional_nested_message'))
+
+    serialized = proto.SerializeToString()
+    proto2 = unittest_pb2.TestAllTypes()
+    self.assertEqual(
+        len(serialized),
+        proto2.MergeFromString(serialized))
+    self.assertTrue(proto2.HasField('optional_nested_message'))
+
+  def testSetInParent(self):
+    proto = unittest_pb2.TestAllTypes()
+    self.assertFalse(proto.HasField('optionalgroup'))
+    proto.optionalgroup.SetInParent()
+    self.assertTrue(proto.HasField('optionalgroup'))
+
+  def testPackageInitializationImport(self):
+    """Test that we can import nested messages from their __init__.py.
+
+    Such setup is not trivial since at the time of processing of __init__.py one
+    can't refer to its submodules by name in code, so expressions like
+    google.protobuf.internal.import_test_package.inner_pb2
+    don't work. They do work in imports, so we have assign an alias at import
+    and then use that alias in generated code.
+    """
+    # We import here since it's the import that used to fail, and we want
+    # the failure to have the right context.
+    # pylint: disable=g-import-not-at-top
+    from google.protobuf.internal import import_test_package
+    # pylint: enable=g-import-not-at-top
+    msg = import_test_package.myproto.Outer()
+    # Just check the default value.
+    self.assertEqual(57, msg.inner.value)
+
+#  Since we had so many tests for protocol buffer equality, we broke these out
+#  into separate TestCase classes.
+
+
+class TestAllTypesEqualityTest(unittest.TestCase):
+
+  def setUp(self):
+    self.first_proto = unittest_pb2.TestAllTypes()
+    self.second_proto = unittest_pb2.TestAllTypes()
+
+  def testNotHashable(self):
+    self.assertRaises(TypeError, hash, self.first_proto)
+
+  def testSelfEquality(self):
+    self.assertEqual(self.first_proto, self.first_proto)
+
+  def testEmptyProtosEqual(self):
+    self.assertEqual(self.first_proto, self.second_proto)
+
+
+class FullProtosEqualityTest(unittest.TestCase):
+
+  """Equality tests using completely-full protos as a starting point."""
+
+  def setUp(self):
+    self.first_proto = unittest_pb2.TestAllTypes()
+    self.second_proto = unittest_pb2.TestAllTypes()
+    test_util.SetAllFields(self.first_proto)
+    test_util.SetAllFields(self.second_proto)
+
+  def testNotHashable(self):
+    self.assertRaises(TypeError, hash, self.first_proto)
+
+  def testNoneNotEqual(self):
+    self.assertNotEqual(self.first_proto, None)
+    self.assertNotEqual(None, self.second_proto)
+
+  def testNotEqualToOtherMessage(self):
+    third_proto = unittest_pb2.TestRequired()
+    self.assertNotEqual(self.first_proto, third_proto)
+    self.assertNotEqual(third_proto, self.second_proto)
+
+  def testAllFieldsFilledEquality(self):
+    self.assertEqual(self.first_proto, self.second_proto)
+
+  def testNonRepeatedScalar(self):
+    # Nonrepeated scalar field change should cause inequality.
+    self.first_proto.optional_int32 += 1
+    self.assertNotEqual(self.first_proto, self.second_proto)
+    # ...as should clearing a field.
+    self.first_proto.ClearField('optional_int32')
+    self.assertNotEqual(self.first_proto, self.second_proto)
+
+  def testNonRepeatedComposite(self):
+    # Change a nonrepeated composite field.
+    self.first_proto.optional_nested_message.bb += 1
+    self.assertNotEqual(self.first_proto, self.second_proto)
+    self.first_proto.optional_nested_message.bb -= 1
+    self.assertEqual(self.first_proto, self.second_proto)
+    # Clear a field in the nested message.
+    self.first_proto.optional_nested_message.ClearField('bb')
+    self.assertNotEqual(self.first_proto, self.second_proto)
+    self.first_proto.optional_nested_message.bb = (
+        self.second_proto.optional_nested_message.bb)
+    self.assertEqual(self.first_proto, self.second_proto)
+    # Remove the nested message entirely.
+    self.first_proto.ClearField('optional_nested_message')
+    self.assertNotEqual(self.first_proto, self.second_proto)
+
+  def testRepeatedScalar(self):
+    # Change a repeated scalar field.
+    self.first_proto.repeated_int32.append(5)
+    self.assertNotEqual(self.first_proto, self.second_proto)
+    self.first_proto.ClearField('repeated_int32')
+    self.assertNotEqual(self.first_proto, self.second_proto)
+
+  def testRepeatedComposite(self):
+    # Change value within a repeated composite field.
+    self.first_proto.repeated_nested_message[0].bb += 1
+    self.assertNotEqual(self.first_proto, self.second_proto)
+    self.first_proto.repeated_nested_message[0].bb -= 1
+    self.assertEqual(self.first_proto, self.second_proto)
+    # Add a value to a repeated composite field.
+    self.first_proto.repeated_nested_message.add()
+    self.assertNotEqual(self.first_proto, self.second_proto)
+    self.second_proto.repeated_nested_message.add()
+    self.assertEqual(self.first_proto, self.second_proto)
+
+  def testNonRepeatedScalarHasBits(self):
+    # Ensure that we test "has" bits as well as value for
+    # nonrepeated scalar field.
+    self.first_proto.ClearField('optional_int32')
+    self.second_proto.optional_int32 = 0
+    self.assertNotEqual(self.first_proto, self.second_proto)
+
+  def testNonRepeatedCompositeHasBits(self):
+    # Ensure that we test "has" bits as well as value for
+    # nonrepeated composite field.
+    self.first_proto.ClearField('optional_nested_message')
+    self.second_proto.optional_nested_message.ClearField('bb')
+    self.assertNotEqual(self.first_proto, self.second_proto)
+    self.first_proto.optional_nested_message.bb = 0
+    self.first_proto.optional_nested_message.ClearField('bb')
+    self.assertEqual(self.first_proto, self.second_proto)
+
+
+class ExtensionEqualityTest(unittest.TestCase):
+
+  def testExtensionEquality(self):
+    first_proto = unittest_pb2.TestAllExtensions()
+    second_proto = unittest_pb2.TestAllExtensions()
+    self.assertEqual(first_proto, second_proto)
+    test_util.SetAllExtensions(first_proto)
+    self.assertNotEqual(first_proto, second_proto)
+    test_util.SetAllExtensions(second_proto)
+    self.assertEqual(first_proto, second_proto)
+
+    # Ensure that we check value equality.
+    first_proto.Extensions[unittest_pb2.optional_int32_extension] += 1
+    self.assertNotEqual(first_proto, second_proto)
+    first_proto.Extensions[unittest_pb2.optional_int32_extension] -= 1
+    self.assertEqual(first_proto, second_proto)
+
+    # Ensure that we also look at "has" bits.
+    first_proto.ClearExtension(unittest_pb2.optional_int32_extension)
+    second_proto.Extensions[unittest_pb2.optional_int32_extension] = 0
+    self.assertNotEqual(first_proto, second_proto)
+    first_proto.Extensions[unittest_pb2.optional_int32_extension] = 0
+    self.assertEqual(first_proto, second_proto)
+
+    # Ensure that differences in cached values
+    # don't matter if "has" bits are both false.
+    first_proto = unittest_pb2.TestAllExtensions()
+    second_proto = unittest_pb2.TestAllExtensions()
+    self.assertEqual(
+        0, first_proto.Extensions[unittest_pb2.optional_int32_extension])
+    self.assertEqual(first_proto, second_proto)
+
+
+class MutualRecursionEqualityTest(unittest.TestCase):
+
+  def testEqualityWithMutualRecursion(self):
+    first_proto = unittest_pb2.TestMutualRecursionA()
+    second_proto = unittest_pb2.TestMutualRecursionA()
+    self.assertEqual(first_proto, second_proto)
+    first_proto.bb.a.bb.optional_int32 = 23
+    self.assertNotEqual(first_proto, second_proto)
+    second_proto.bb.a.bb.optional_int32 = 23
+    self.assertEqual(first_proto, second_proto)
+
+
+class ByteSizeTest(unittest.TestCase):
+
+  def setUp(self):
+    self.proto = unittest_pb2.TestAllTypes()
+    self.extended_proto = more_extensions_pb2.ExtendedMessage()
+    self.packed_proto = unittest_pb2.TestPackedTypes()
+    self.packed_extended_proto = unittest_pb2.TestPackedExtensions()
+
+  def Size(self):
+    return self.proto.ByteSize()
+
+  def testEmptyMessage(self):
+    self.assertEqual(0, self.proto.ByteSize())
+
+  def testSizedOnKwargs(self):
+    # Use a separate message to ensure testing right after creation.
+    proto = unittest_pb2.TestAllTypes()
+    self.assertEqual(0, proto.ByteSize())
+    proto_kwargs = unittest_pb2.TestAllTypes(optional_int64 = 1)
+    # One byte for the tag, one to encode varint 1.
+    self.assertEqual(2, proto_kwargs.ByteSize())
+
+  def testVarints(self):
+    def Test(i, expected_varint_size):
+      self.proto.Clear()
+      self.proto.optional_int64 = i
+      # Add one to the varint size for the tag info
+      # for tag 1.
+      self.assertEqual(expected_varint_size + 1, self.Size())
+    Test(0, 1)
+    Test(1, 1)
+    for i, num_bytes in zip(range(7, 63, 7), range(1, 10000)):
+      Test((1 << i) - 1, num_bytes)
+    Test(-1, 10)
+    Test(-2, 10)
+    Test(-(1 << 63), 10)
+
+  def testStrings(self):
+    self.proto.optional_string = ''
+    # Need one byte for tag info (tag #14), and one byte for length.
+    self.assertEqual(2, self.Size())
+
+    self.proto.optional_string = 'abc'
+    # Need one byte for tag info (tag #14), and one byte for length.
+    self.assertEqual(2 + len(self.proto.optional_string), self.Size())
+
+    self.proto.optional_string = 'x' * 128
+    # Need one byte for tag info (tag #14), and TWO bytes for length.
+    self.assertEqual(3 + len(self.proto.optional_string), self.Size())
+
+  def testOtherNumerics(self):
+    self.proto.optional_fixed32 = 1234
+    # One byte for tag and 4 bytes for fixed32.
+    self.assertEqual(5, self.Size())
+    self.proto = unittest_pb2.TestAllTypes()
+
+    self.proto.optional_fixed64 = 1234
+    # One byte for tag and 8 bytes for fixed64.
+    self.assertEqual(9, self.Size())
+    self.proto = unittest_pb2.TestAllTypes()
+
+    self.proto.optional_float = 1.234
+    # One byte for tag and 4 bytes for float.
+    self.assertEqual(5, self.Size())
+    self.proto = unittest_pb2.TestAllTypes()
+
+    self.proto.optional_double = 1.234
+    # One byte for tag and 8 bytes for float.
+    self.assertEqual(9, self.Size())
+    self.proto = unittest_pb2.TestAllTypes()
+
+    self.proto.optional_sint32 = 64
+    # One byte for tag and 2 bytes for zig-zag-encoded 64.
+    self.assertEqual(3, self.Size())
+    self.proto = unittest_pb2.TestAllTypes()
+
+  def testComposites(self):
+    # 3 bytes.
+    self.proto.optional_nested_message.bb = (1 << 14)
+    # Plus one byte for bb tag.
+    # Plus 1 byte for optional_nested_message serialized size.
+    # Plus two bytes for optional_nested_message tag.
+    self.assertEqual(3 + 1 + 1 + 2, self.Size())
+
+  def testGroups(self):
+    # 4 bytes.
+    self.proto.optionalgroup.a = (1 << 21)
+    # Plus two bytes for |a| tag.
+    # Plus 2 * two bytes for START_GROUP and END_GROUP tags.
+    self.assertEqual(4 + 2 + 2*2, self.Size())
+
+  def testRepeatedScalars(self):
+    self.proto.repeated_int32.append(10)  # 1 byte.
+    self.proto.repeated_int32.append(128)  # 2 bytes.
+    # Also need 2 bytes for each entry for tag.
+    self.assertEqual(1 + 2 + 2*2, self.Size())
+
+  def testRepeatedScalarsExtend(self):
+    self.proto.repeated_int32.extend([10, 128])  # 3 bytes.
+    # Also need 2 bytes for each entry for tag.
+    self.assertEqual(1 + 2 + 2*2, self.Size())
+
+  def testRepeatedScalarsRemove(self):
+    self.proto.repeated_int32.append(10)  # 1 byte.
+    self.proto.repeated_int32.append(128)  # 2 bytes.
+    # Also need 2 bytes for each entry for tag.
+    self.assertEqual(1 + 2 + 2*2, self.Size())
+    self.proto.repeated_int32.remove(128)
+    self.assertEqual(1 + 2, self.Size())
+
+  def testRepeatedComposites(self):
+    # Empty message.  2 bytes tag plus 1 byte length.
+    foreign_message_0 = self.proto.repeated_nested_message.add()
+    # 2 bytes tag plus 1 byte length plus 1 byte bb tag 1 byte int.
+    foreign_message_1 = self.proto.repeated_nested_message.add()
+    foreign_message_1.bb = 7
+    self.assertEqual(2 + 1 + 2 + 1 + 1 + 1, self.Size())
+
+  def testRepeatedCompositesDelete(self):
+    # Empty message.  2 bytes tag plus 1 byte length.
+    foreign_message_0 = self.proto.repeated_nested_message.add()
+    # 2 bytes tag plus 1 byte length plus 1 byte bb tag 1 byte int.
+    foreign_message_1 = self.proto.repeated_nested_message.add()
+    foreign_message_1.bb = 9
+    self.assertEqual(2 + 1 + 2 + 1 + 1 + 1, self.Size())
+
+    # 2 bytes tag plus 1 byte length plus 1 byte bb tag 1 byte int.
+    del self.proto.repeated_nested_message[0]
+    self.assertEqual(2 + 1 + 1 + 1, self.Size())
+
+    # Now add a new message.
+    foreign_message_2 = self.proto.repeated_nested_message.add()
+    foreign_message_2.bb = 12
+
+    # 2 bytes tag plus 1 byte length plus 1 byte bb tag 1 byte int.
+    # 2 bytes tag plus 1 byte length plus 1 byte bb tag 1 byte int.
+    self.assertEqual(2 + 1 + 1 + 1 + 2 + 1 + 1 + 1, self.Size())
+
+    # 2 bytes tag plus 1 byte length plus 1 byte bb tag 1 byte int.
+    del self.proto.repeated_nested_message[1]
+    self.assertEqual(2 + 1 + 1 + 1, self.Size())
+
+    del self.proto.repeated_nested_message[0]
+    self.assertEqual(0, self.Size())
+
+  def testRepeatedGroups(self):
+    # 2-byte START_GROUP plus 2-byte END_GROUP.
+    group_0 = self.proto.repeatedgroup.add()
+    # 2-byte START_GROUP plus 2-byte |a| tag + 1-byte |a|
+    # plus 2-byte END_GROUP.
+    group_1 = self.proto.repeatedgroup.add()
+    group_1.a =  7
+    self.assertEqual(2 + 2 + 2 + 2 + 1 + 2, self.Size())
+
+  def testExtensions(self):
+    proto = unittest_pb2.TestAllExtensions()
+    self.assertEqual(0, proto.ByteSize())
+    extension = unittest_pb2.optional_int32_extension  # Field #1, 1 byte.
+    proto.Extensions[extension] = 23
+    # 1 byte for tag, 1 byte for value.
+    self.assertEqual(2, proto.ByteSize())
+
+  def testCacheInvalidationForNonrepeatedScalar(self):
+    # Test non-extension.
+    self.proto.optional_int32 = 1
+    self.assertEqual(2, self.proto.ByteSize())
+    self.proto.optional_int32 = 128
+    self.assertEqual(3, self.proto.ByteSize())
+    self.proto.ClearField('optional_int32')
+    self.assertEqual(0, self.proto.ByteSize())
+
+    # Test within extension.
+    extension = more_extensions_pb2.optional_int_extension
+    self.extended_proto.Extensions[extension] = 1
+    self.assertEqual(2, self.extended_proto.ByteSize())
+    self.extended_proto.Extensions[extension] = 128
+    self.assertEqual(3, self.extended_proto.ByteSize())
+    self.extended_proto.ClearExtension(extension)
+    self.assertEqual(0, self.extended_proto.ByteSize())
+
+  def testCacheInvalidationForRepeatedScalar(self):
+    # Test non-extension.
+    self.proto.repeated_int32.append(1)
+    self.assertEqual(3, self.proto.ByteSize())
+    self.proto.repeated_int32.append(1)
+    self.assertEqual(6, self.proto.ByteSize())
+    self.proto.repeated_int32[1] = 128
+    self.assertEqual(7, self.proto.ByteSize())
+    self.proto.ClearField('repeated_int32')
+    self.assertEqual(0, self.proto.ByteSize())
+
+    # Test within extension.
+    extension = more_extensions_pb2.repeated_int_extension
+    repeated = self.extended_proto.Extensions[extension]
+    repeated.append(1)
+    self.assertEqual(2, self.extended_proto.ByteSize())
+    repeated.append(1)
+    self.assertEqual(4, self.extended_proto.ByteSize())
+    repeated[1] = 128
+    self.assertEqual(5, self.extended_proto.ByteSize())
+    self.extended_proto.ClearExtension(extension)
+    self.assertEqual(0, self.extended_proto.ByteSize())
+
+  def testCacheInvalidationForNonrepeatedMessage(self):
+    # Test non-extension.
+    self.proto.optional_foreign_message.c = 1
+    self.assertEqual(5, self.proto.ByteSize())
+    self.proto.optional_foreign_message.c = 128
+    self.assertEqual(6, self.proto.ByteSize())
+    self.proto.optional_foreign_message.ClearField('c')
+    self.assertEqual(3, self.proto.ByteSize())
+    self.proto.ClearField('optional_foreign_message')
+    self.assertEqual(0, self.proto.ByteSize())
+
+    if api_implementation.Type() == 'python':
+      # This is only possible in pure-Python implementation of the API.
+      child = self.proto.optional_foreign_message
+      self.proto.ClearField('optional_foreign_message')
+      child.c = 128
+      self.assertEqual(0, self.proto.ByteSize())
+
+    # Test within extension.
+    extension = more_extensions_pb2.optional_message_extension
+    child = self.extended_proto.Extensions[extension]
+    self.assertEqual(0, self.extended_proto.ByteSize())
+    child.foreign_message_int = 1
+    self.assertEqual(4, self.extended_proto.ByteSize())
+    child.foreign_message_int = 128
+    self.assertEqual(5, self.extended_proto.ByteSize())
+    self.extended_proto.ClearExtension(extension)
+    self.assertEqual(0, self.extended_proto.ByteSize())
+
+  def testCacheInvalidationForRepeatedMessage(self):
+    # Test non-extension.
+    child0 = self.proto.repeated_foreign_message.add()
+    self.assertEqual(3, self.proto.ByteSize())
+    self.proto.repeated_foreign_message.add()
+    self.assertEqual(6, self.proto.ByteSize())
+    child0.c = 1
+    self.assertEqual(8, self.proto.ByteSize())
+    self.proto.ClearField('repeated_foreign_message')
+    self.assertEqual(0, self.proto.ByteSize())
+
+    # Test within extension.
+    extension = more_extensions_pb2.repeated_message_extension
+    child_list = self.extended_proto.Extensions[extension]
+    child0 = child_list.add()
+    self.assertEqual(2, self.extended_proto.ByteSize())
+    child_list.add()
+    self.assertEqual(4, self.extended_proto.ByteSize())
+    child0.foreign_message_int = 1
+    self.assertEqual(6, self.extended_proto.ByteSize())
+    child0.ClearField('foreign_message_int')
+    self.assertEqual(4, self.extended_proto.ByteSize())
+    self.extended_proto.ClearExtension(extension)
+    self.assertEqual(0, self.extended_proto.ByteSize())
+
+  def testPackedRepeatedScalars(self):
+    self.assertEqual(0, self.packed_proto.ByteSize())
+
+    self.packed_proto.packed_int32.append(10)   # 1 byte.
+    self.packed_proto.packed_int32.append(128)  # 2 bytes.
+    # The tag is 2 bytes (the field number is 90), and the varint
+    # storing the length is 1 byte.
+    int_size = 1 + 2 + 3
+    self.assertEqual(int_size, self.packed_proto.ByteSize())
+
+    self.packed_proto.packed_double.append(4.2)   # 8 bytes
+    self.packed_proto.packed_double.append(3.25)  # 8 bytes
+    # 2 more tag bytes, 1 more length byte.
+    double_size = 8 + 8 + 3
+    self.assertEqual(int_size+double_size, self.packed_proto.ByteSize())
+
+    self.packed_proto.ClearField('packed_int32')
+    self.assertEqual(double_size, self.packed_proto.ByteSize())
+
+  def testPackedExtensions(self):
+    self.assertEqual(0, self.packed_extended_proto.ByteSize())
+    extension = self.packed_extended_proto.Extensions[
+        unittest_pb2.packed_fixed32_extension]
+    extension.extend([1, 2, 3, 4])   # 16 bytes
+    # Tag is 3 bytes.
+    self.assertEqual(19, self.packed_extended_proto.ByteSize())
+
+
+# Issues to be sure to cover include:
+#   * Handling of unrecognized tags ("uninterpreted_bytes").
+#   * Handling of MessageSets.
+#   * Consistent ordering of tags in the wire format,
+#     including ordering between extensions and non-extension
+#     fields.
+#   * Consistent serialization of negative numbers, especially
+#     negative int32s.
+#   * Handling of empty submessages (with and without "has"
+#     bits set).
+
+class SerializationTest(unittest.TestCase):
+
+  def testSerializeEmtpyMessage(self):
+    first_proto = unittest_pb2.TestAllTypes()
+    second_proto = unittest_pb2.TestAllTypes()
+    serialized = first_proto.SerializeToString()
+    self.assertEqual(first_proto.ByteSize(), len(serialized))
+    self.assertEqual(
+        len(serialized),
+        second_proto.MergeFromString(serialized))
+    self.assertEqual(first_proto, second_proto)
+
+  def testSerializeAllFields(self):
+    first_proto = unittest_pb2.TestAllTypes()
+    second_proto = unittest_pb2.TestAllTypes()
+    test_util.SetAllFields(first_proto)
+    serialized = first_proto.SerializeToString()
+    self.assertEqual(first_proto.ByteSize(), len(serialized))
+    self.assertEqual(
+        len(serialized),
+        second_proto.MergeFromString(serialized))
+    self.assertEqual(first_proto, second_proto)
+
+  def testSerializeAllExtensions(self):
+    first_proto = unittest_pb2.TestAllExtensions()
+    second_proto = unittest_pb2.TestAllExtensions()
+    test_util.SetAllExtensions(first_proto)
+    serialized = first_proto.SerializeToString()
+    self.assertEqual(
+        len(serialized),
+        second_proto.MergeFromString(serialized))
+    self.assertEqual(first_proto, second_proto)
+
+  def testSerializeWithOptionalGroup(self):
+    first_proto = unittest_pb2.TestAllTypes()
+    second_proto = unittest_pb2.TestAllTypes()
+    first_proto.optionalgroup.a = 242
+    serialized = first_proto.SerializeToString()
+    self.assertEqual(
+        len(serialized),
+        second_proto.MergeFromString(serialized))
+    self.assertEqual(first_proto, second_proto)
+
+  def testSerializeNegativeValues(self):
+    first_proto = unittest_pb2.TestAllTypes()
+
+    first_proto.optional_int32 = -1
+    first_proto.optional_int64 = -(2 << 40)
+    first_proto.optional_sint32 = -3
+    first_proto.optional_sint64 = -(4 << 40)
+    first_proto.optional_sfixed32 = -5
+    first_proto.optional_sfixed64 = -(6 << 40)
+
+    second_proto = unittest_pb2.TestAllTypes.FromString(
+        first_proto.SerializeToString())
+
+    self.assertEqual(first_proto, second_proto)
+
+  def testParseTruncated(self):
+    # This test is only applicable for the Python implementation of the API.
+    if api_implementation.Type() != 'python':
+      return
+
+    first_proto = unittest_pb2.TestAllTypes()
+    test_util.SetAllFields(first_proto)
+    serialized = first_proto.SerializeToString()
+
+    for truncation_point in range(len(serialized) + 1):
+      try:
+        second_proto = unittest_pb2.TestAllTypes()
+        unknown_fields = unittest_pb2.TestEmptyMessage()
+        pos = second_proto._InternalParse(serialized, 0, truncation_point)
+        # If we didn't raise an error then we read exactly the amount expected.
+        self.assertEqual(truncation_point, pos)
+
+        # Parsing to unknown fields should not throw if parsing to known fields
+        # did not.
+        try:
+          pos2 = unknown_fields._InternalParse(serialized, 0, truncation_point)
+          self.assertEqual(truncation_point, pos2)
+        except message.DecodeError:
+          self.fail('Parsing unknown fields failed when parsing known fields '
+                    'did not.')
+      except message.DecodeError:
+        # Parsing unknown fields should also fail.
+        self.assertRaises(message.DecodeError, unknown_fields._InternalParse,
+                          serialized, 0, truncation_point)
+
+  def testCanonicalSerializationOrder(self):
+    proto = more_messages_pb2.OutOfOrderFields()
+    # These are also their tag numbers.  Even though we're setting these in
+    # reverse-tag order AND they're listed in reverse tag-order in the .proto
+    # file, they should nonetheless be serialized in tag order.
+    proto.optional_sint32 = 5
+    proto.Extensions[more_messages_pb2.optional_uint64] = 4
+    proto.optional_uint32 = 3
+    proto.Extensions[more_messages_pb2.optional_int64] = 2
+    proto.optional_int32 = 1
+    serialized = proto.SerializeToString()
+    self.assertEqual(proto.ByteSize(), len(serialized))
+    d = _MiniDecoder(serialized)
+    ReadTag = d.ReadFieldNumberAndWireType
+    self.assertEqual((1, wire_format.WIRETYPE_VARINT), ReadTag())
+    self.assertEqual(1, d.ReadInt32())
+    self.assertEqual((2, wire_format.WIRETYPE_VARINT), ReadTag())
+    self.assertEqual(2, d.ReadInt64())
+    self.assertEqual((3, wire_format.WIRETYPE_VARINT), ReadTag())
+    self.assertEqual(3, d.ReadUInt32())
+    self.assertEqual((4, wire_format.WIRETYPE_VARINT), ReadTag())
+    self.assertEqual(4, d.ReadUInt64())
+    self.assertEqual((5, wire_format.WIRETYPE_VARINT), ReadTag())
+    self.assertEqual(5, d.ReadSInt32())
+
+  def testCanonicalSerializationOrderSameAsCpp(self):
+    # Copy of the same test we use for C++.
+    proto = unittest_pb2.TestFieldOrderings()
+    test_util.SetAllFieldsAndExtensions(proto)
+    serialized = proto.SerializeToString()
+    test_util.ExpectAllFieldsAndExtensionsInOrder(serialized)
+
+  def testMergeFromStringWhenFieldsAlreadySet(self):
+    first_proto = unittest_pb2.TestAllTypes()
+    first_proto.repeated_string.append('foobar')
+    first_proto.optional_int32 = 23
+    first_proto.optional_nested_message.bb = 42
+    serialized = first_proto.SerializeToString()
+
+    second_proto = unittest_pb2.TestAllTypes()
+    second_proto.repeated_string.append('baz')
+    second_proto.optional_int32 = 100
+    second_proto.optional_nested_message.bb = 999
+
+    bytes_parsed = second_proto.MergeFromString(serialized)
+    self.assertEqual(len(serialized), bytes_parsed)
+
+    # Ensure that we append to repeated fields.
+    self.assertEqual(['baz', 'foobar'], list(second_proto.repeated_string))
+    # Ensure that we overwrite nonrepeatd scalars.
+    self.assertEqual(23, second_proto.optional_int32)
+    # Ensure that we recursively call MergeFromString() on
+    # submessages.
+    self.assertEqual(42, second_proto.optional_nested_message.bb)
+
+  def testMessageSetWireFormat(self):
+    proto = message_set_extensions_pb2.TestMessageSet()
+    extension_message1 = message_set_extensions_pb2.TestMessageSetExtension1
+    extension_message2 = message_set_extensions_pb2.TestMessageSetExtension2
+    extension1 = extension_message1.message_set_extension
+    extension2 = extension_message2.message_set_extension
+    extension3 = message_set_extensions_pb2.message_set_extension3
+    proto.Extensions[extension1].i = 123
+    proto.Extensions[extension2].str = 'foo'
+    proto.Extensions[extension3].text = 'bar'
+
+    # Serialize using the MessageSet wire format (this is specified in the
+    # .proto file).
+    serialized = proto.SerializeToString()
+
+    raw = unittest_mset_pb2.RawMessageSet()
+    self.assertEqual(False,
+                     raw.DESCRIPTOR.GetOptions().message_set_wire_format)
+    self.assertEqual(
+        len(serialized),
+        raw.MergeFromString(serialized))
+    self.assertEqual(3, len(raw.item))
+
+    message1 = message_set_extensions_pb2.TestMessageSetExtension1()
+    self.assertEqual(
+        len(raw.item[0].message),
+        message1.MergeFromString(raw.item[0].message))
+    self.assertEqual(123, message1.i)
+
+    message2 = message_set_extensions_pb2.TestMessageSetExtension2()
+    self.assertEqual(
+        len(raw.item[1].message),
+        message2.MergeFromString(raw.item[1].message))
+    self.assertEqual('foo', message2.str)
+
+    message3 = message_set_extensions_pb2.TestMessageSetExtension3()
+    self.assertEqual(
+        len(raw.item[2].message),
+        message3.MergeFromString(raw.item[2].message))
+    self.assertEqual('bar', message3.text)
+
+    # Deserialize using the MessageSet wire format.
+    proto2 = message_set_extensions_pb2.TestMessageSet()
+    self.assertEqual(
+        len(serialized),
+        proto2.MergeFromString(serialized))
+    self.assertEqual(123, proto2.Extensions[extension1].i)
+    self.assertEqual('foo', proto2.Extensions[extension2].str)
+    self.assertEqual('bar', proto2.Extensions[extension3].text)
+
+    # Check byte size.
+    self.assertEqual(proto2.ByteSize(), len(serialized))
+    self.assertEqual(proto.ByteSize(), len(serialized))
+
+  def testMessageSetWireFormatUnknownExtension(self):
+    # Create a message using the message set wire format with an unknown
+    # message.
+    raw = unittest_mset_pb2.RawMessageSet()
+
+    # Add an item.
+    item = raw.item.add()
+    item.type_id = 98418603
+    extension_message1 = message_set_extensions_pb2.TestMessageSetExtension1
+    message1 = message_set_extensions_pb2.TestMessageSetExtension1()
+    message1.i = 12345
+    item.message = message1.SerializeToString()
+
+    # Add a second, unknown extension.
+    item = raw.item.add()
+    item.type_id = 98418604
+    extension_message1 = message_set_extensions_pb2.TestMessageSetExtension1
+    message1 = message_set_extensions_pb2.TestMessageSetExtension1()
+    message1.i = 12346
+    item.message = message1.SerializeToString()
+
+    # Add another unknown extension.
+    item = raw.item.add()
+    item.type_id = 98418605
+    message1 = message_set_extensions_pb2.TestMessageSetExtension2()
+    message1.str = 'foo'
+    item.message = message1.SerializeToString()
+
+    serialized = raw.SerializeToString()
+
+    # Parse message using the message set wire format.
+    proto = message_set_extensions_pb2.TestMessageSet()
+    self.assertEqual(
+        len(serialized),
+        proto.MergeFromString(serialized))
+
+    # Check that the message parsed well.
+    extension_message1 = message_set_extensions_pb2.TestMessageSetExtension1
+    extension1 = extension_message1.message_set_extension
+    self.assertEqual(12345, proto.Extensions[extension1].i)
+
+  def testUnknownFields(self):
+    proto = unittest_pb2.TestAllTypes()
+    test_util.SetAllFields(proto)
+
+    serialized = proto.SerializeToString()
+
+    # The empty message should be parsable with all of the fields
+    # unknown.
+    proto2 = unittest_pb2.TestEmptyMessage()
+
+    # Parsing this message should succeed.
+    self.assertEqual(
+        len(serialized),
+        proto2.MergeFromString(serialized))
+
+    # Now test with a int64 field set.
+    proto = unittest_pb2.TestAllTypes()
+    proto.optional_int64 = 0x0fffffffffffffff
+    serialized = proto.SerializeToString()
+    # The empty message should be parsable with all of the fields
+    # unknown.
+    proto2 = unittest_pb2.TestEmptyMessage()
+    # Parsing this message should succeed.
+    self.assertEqual(
+        len(serialized),
+        proto2.MergeFromString(serialized))
+
+  def _CheckRaises(self, exc_class, callable_obj, exception):
+    """This method checks if the excpetion type and message are as expected."""
+    try:
+      callable_obj()
+    except exc_class as ex:
+      # Check if the exception message is the right one.
+      self.assertEqual(exception, str(ex))
+      return
+    else:
+      raise self.failureException('%s not raised' % str(exc_class))
+
+  def testSerializeUninitialized(self):
+    proto = unittest_pb2.TestRequired()
+    self._CheckRaises(
+        message.EncodeError,
+        proto.SerializeToString,
+        'Message protobuf_unittest.TestRequired is missing required fields: '
+        'a,b,c')
+    # Shouldn't raise exceptions.
+    partial = proto.SerializePartialToString()
+
+    proto2 = unittest_pb2.TestRequired()
+    self.assertFalse(proto2.HasField('a'))
+    # proto2 ParseFromString does not check that required fields are set.
+    proto2.ParseFromString(partial)
+    self.assertFalse(proto2.HasField('a'))
+
+    proto.a = 1
+    self._CheckRaises(
+        message.EncodeError,
+        proto.SerializeToString,
+        'Message protobuf_unittest.TestRequired is missing required fields: b,c')
+    # Shouldn't raise exceptions.
+    partial = proto.SerializePartialToString()
+
+    proto.b = 2
+    self._CheckRaises(
+        message.EncodeError,
+        proto.SerializeToString,
+        'Message protobuf_unittest.TestRequired is missing required fields: c')
+    # Shouldn't raise exceptions.
+    partial = proto.SerializePartialToString()
+
+    proto.c = 3
+    serialized = proto.SerializeToString()
+    # Shouldn't raise exceptions.
+    partial = proto.SerializePartialToString()
+
+    proto2 = unittest_pb2.TestRequired()
+    self.assertEqual(
+        len(serialized),
+        proto2.MergeFromString(serialized))
+    self.assertEqual(1, proto2.a)
+    self.assertEqual(2, proto2.b)
+    self.assertEqual(3, proto2.c)
+    self.assertEqual(
+        len(partial),
+        proto2.MergeFromString(partial))
+    self.assertEqual(1, proto2.a)
+    self.assertEqual(2, proto2.b)
+    self.assertEqual(3, proto2.c)
+
+  def testSerializeUninitializedSubMessage(self):
+    proto = unittest_pb2.TestRequiredForeign()
+
+    # Sub-message doesn't exist yet, so this succeeds.
+    proto.SerializeToString()
+
+    proto.optional_message.a = 1
+    self._CheckRaises(
+        message.EncodeError,
+        proto.SerializeToString,
+        'Message protobuf_unittest.TestRequiredForeign '
+        'is missing required fields: '
+        'optional_message.b,optional_message.c')
+
+    proto.optional_message.b = 2
+    proto.optional_message.c = 3
+    proto.SerializeToString()
+
+    proto.repeated_message.add().a = 1
+    proto.repeated_message.add().b = 2
+    self._CheckRaises(
+        message.EncodeError,
+        proto.SerializeToString,
+        'Message protobuf_unittest.TestRequiredForeign is missing required fields: '
+        'repeated_message[0].b,repeated_message[0].c,'
+        'repeated_message[1].a,repeated_message[1].c')
+
+    proto.repeated_message[0].b = 2
+    proto.repeated_message[0].c = 3
+    proto.repeated_message[1].a = 1
+    proto.repeated_message[1].c = 3
+    proto.SerializeToString()
+
+  def testSerializeAllPackedFields(self):
+    first_proto = unittest_pb2.TestPackedTypes()
+    second_proto = unittest_pb2.TestPackedTypes()
+    test_util.SetAllPackedFields(first_proto)
+    serialized = first_proto.SerializeToString()
+    self.assertEqual(first_proto.ByteSize(), len(serialized))
+    bytes_read = second_proto.MergeFromString(serialized)
+    self.assertEqual(second_proto.ByteSize(), bytes_read)
+    self.assertEqual(first_proto, second_proto)
+
+  def testSerializeAllPackedExtensions(self):
+    first_proto = unittest_pb2.TestPackedExtensions()
+    second_proto = unittest_pb2.TestPackedExtensions()
+    test_util.SetAllPackedExtensions(first_proto)
+    serialized = first_proto.SerializeToString()
+    bytes_read = second_proto.MergeFromString(serialized)
+    self.assertEqual(second_proto.ByteSize(), bytes_read)
+    self.assertEqual(first_proto, second_proto)
+
+  def testMergePackedFromStringWhenSomeFieldsAlreadySet(self):
+    first_proto = unittest_pb2.TestPackedTypes()
+    first_proto.packed_int32.extend([1, 2])
+    first_proto.packed_double.append(3.0)
+    serialized = first_proto.SerializeToString()
+
+    second_proto = unittest_pb2.TestPackedTypes()
+    second_proto.packed_int32.append(3)
+    second_proto.packed_double.extend([1.0, 2.0])
+    second_proto.packed_sint32.append(4)
+
+    self.assertEqual(
+        len(serialized),
+        second_proto.MergeFromString(serialized))
+    self.assertEqual([3, 1, 2], second_proto.packed_int32)
+    self.assertEqual([1.0, 2.0, 3.0], second_proto.packed_double)
+    self.assertEqual([4], second_proto.packed_sint32)
+
+  def testPackedFieldsWireFormat(self):
+    proto = unittest_pb2.TestPackedTypes()
+    proto.packed_int32.extend([1, 2, 150, 3])  # 1 + 1 + 2 + 1 bytes
+    proto.packed_double.extend([1.0, 1000.0])  # 8 + 8 bytes
+    proto.packed_float.append(2.0)             # 4 bytes, will be before double
+    serialized = proto.SerializeToString()
+    self.assertEqual(proto.ByteSize(), len(serialized))
+    d = _MiniDecoder(serialized)
+    ReadTag = d.ReadFieldNumberAndWireType
+    self.assertEqual((90, wire_format.WIRETYPE_LENGTH_DELIMITED), ReadTag())
+    self.assertEqual(1+1+1+2, d.ReadInt32())
+    self.assertEqual(1, d.ReadInt32())
+    self.assertEqual(2, d.ReadInt32())
+    self.assertEqual(150, d.ReadInt32())
+    self.assertEqual(3, d.ReadInt32())
+    self.assertEqual((100, wire_format.WIRETYPE_LENGTH_DELIMITED), ReadTag())
+    self.assertEqual(4, d.ReadInt32())
+    self.assertEqual(2.0, d.ReadFloat())
+    self.assertEqual((101, wire_format.WIRETYPE_LENGTH_DELIMITED), ReadTag())
+    self.assertEqual(8+8, d.ReadInt32())
+    self.assertEqual(1.0, d.ReadDouble())
+    self.assertEqual(1000.0, d.ReadDouble())
+    self.assertTrue(d.EndOfStream())
+
+  def testParsePackedFromUnpacked(self):
+    unpacked = unittest_pb2.TestUnpackedTypes()
+    test_util.SetAllUnpackedFields(unpacked)
+    packed = unittest_pb2.TestPackedTypes()
+    serialized = unpacked.SerializeToString()
+    self.assertEqual(
+        len(serialized),
+        packed.MergeFromString(serialized))
+    expected = unittest_pb2.TestPackedTypes()
+    test_util.SetAllPackedFields(expected)
+    self.assertEqual(expected, packed)
+
+  def testParseUnpackedFromPacked(self):
+    packed = unittest_pb2.TestPackedTypes()
+    test_util.SetAllPackedFields(packed)
+    unpacked = unittest_pb2.TestUnpackedTypes()
+    serialized = packed.SerializeToString()
+    self.assertEqual(
+        len(serialized),
+        unpacked.MergeFromString(serialized))
+    expected = unittest_pb2.TestUnpackedTypes()
+    test_util.SetAllUnpackedFields(expected)
+    self.assertEqual(expected, unpacked)
+
+  def testFieldNumbers(self):
+    proto = unittest_pb2.TestAllTypes()
+    self.assertEqual(unittest_pb2.TestAllTypes.NestedMessage.BB_FIELD_NUMBER, 1)
+    self.assertEqual(unittest_pb2.TestAllTypes.OPTIONAL_INT32_FIELD_NUMBER, 1)
+    self.assertEqual(unittest_pb2.TestAllTypes.OPTIONALGROUP_FIELD_NUMBER, 16)
+    self.assertEqual(
+      unittest_pb2.TestAllTypes.OPTIONAL_NESTED_MESSAGE_FIELD_NUMBER, 18)
+    self.assertEqual(
+      unittest_pb2.TestAllTypes.OPTIONAL_NESTED_ENUM_FIELD_NUMBER, 21)
+    self.assertEqual(unittest_pb2.TestAllTypes.REPEATED_INT32_FIELD_NUMBER, 31)
+    self.assertEqual(unittest_pb2.TestAllTypes.REPEATEDGROUP_FIELD_NUMBER, 46)
+    self.assertEqual(
+      unittest_pb2.TestAllTypes.REPEATED_NESTED_MESSAGE_FIELD_NUMBER, 48)
+    self.assertEqual(
+      unittest_pb2.TestAllTypes.REPEATED_NESTED_ENUM_FIELD_NUMBER, 51)
+
+  def testExtensionFieldNumbers(self):
+    self.assertEqual(unittest_pb2.TestRequired.single.number, 1000)
+    self.assertEqual(unittest_pb2.TestRequired.SINGLE_FIELD_NUMBER, 1000)
+    self.assertEqual(unittest_pb2.TestRequired.multi.number, 1001)
+    self.assertEqual(unittest_pb2.TestRequired.MULTI_FIELD_NUMBER, 1001)
+    self.assertEqual(unittest_pb2.optional_int32_extension.number, 1)
+    self.assertEqual(unittest_pb2.OPTIONAL_INT32_EXTENSION_FIELD_NUMBER, 1)
+    self.assertEqual(unittest_pb2.optionalgroup_extension.number, 16)
+    self.assertEqual(unittest_pb2.OPTIONALGROUP_EXTENSION_FIELD_NUMBER, 16)
+    self.assertEqual(unittest_pb2.optional_nested_message_extension.number, 18)
+    self.assertEqual(
+      unittest_pb2.OPTIONAL_NESTED_MESSAGE_EXTENSION_FIELD_NUMBER, 18)
+    self.assertEqual(unittest_pb2.optional_nested_enum_extension.number, 21)
+    self.assertEqual(unittest_pb2.OPTIONAL_NESTED_ENUM_EXTENSION_FIELD_NUMBER,
+      21)
+    self.assertEqual(unittest_pb2.repeated_int32_extension.number, 31)
+    self.assertEqual(unittest_pb2.REPEATED_INT32_EXTENSION_FIELD_NUMBER, 31)
+    self.assertEqual(unittest_pb2.repeatedgroup_extension.number, 46)
+    self.assertEqual(unittest_pb2.REPEATEDGROUP_EXTENSION_FIELD_NUMBER, 46)
+    self.assertEqual(unittest_pb2.repeated_nested_message_extension.number, 48)
+    self.assertEqual(
+      unittest_pb2.REPEATED_NESTED_MESSAGE_EXTENSION_FIELD_NUMBER, 48)
+    self.assertEqual(unittest_pb2.repeated_nested_enum_extension.number, 51)
+    self.assertEqual(unittest_pb2.REPEATED_NESTED_ENUM_EXTENSION_FIELD_NUMBER,
+      51)
+
+  def testInitKwargs(self):
+    proto = unittest_pb2.TestAllTypes(
+        optional_int32=1,
+        optional_string='foo',
+        optional_bool=True,
+        optional_bytes=b'bar',
+        optional_nested_message=unittest_pb2.TestAllTypes.NestedMessage(bb=1),
+        optional_foreign_message=unittest_pb2.ForeignMessage(c=1),
+        optional_nested_enum=unittest_pb2.TestAllTypes.FOO,
+        optional_foreign_enum=unittest_pb2.FOREIGN_FOO,
+        repeated_int32=[1, 2, 3])
+    self.assertTrue(proto.IsInitialized())
+    self.assertTrue(proto.HasField('optional_int32'))
+    self.assertTrue(proto.HasField('optional_string'))
+    self.assertTrue(proto.HasField('optional_bool'))
+    self.assertTrue(proto.HasField('optional_bytes'))
+    self.assertTrue(proto.HasField('optional_nested_message'))
+    self.assertTrue(proto.HasField('optional_foreign_message'))
+    self.assertTrue(proto.HasField('optional_nested_enum'))
+    self.assertTrue(proto.HasField('optional_foreign_enum'))
+    self.assertEqual(1, proto.optional_int32)
+    self.assertEqual('foo', proto.optional_string)
+    self.assertEqual(True, proto.optional_bool)
+    self.assertEqual(b'bar', proto.optional_bytes)
+    self.assertEqual(1, proto.optional_nested_message.bb)
+    self.assertEqual(1, proto.optional_foreign_message.c)
+    self.assertEqual(unittest_pb2.TestAllTypes.FOO,
+                     proto.optional_nested_enum)
+    self.assertEqual(unittest_pb2.FOREIGN_FOO, proto.optional_foreign_enum)
+    self.assertEqual([1, 2, 3], proto.repeated_int32)
+
+  def testInitArgsUnknownFieldName(self):
+    def InitalizeEmptyMessageWithExtraKeywordArg():
+      unused_proto = unittest_pb2.TestEmptyMessage(unknown='unknown')
+    self._CheckRaises(
+        ValueError,
+        InitalizeEmptyMessageWithExtraKeywordArg,
+        'Protocol message TestEmptyMessage has no "unknown" field.')
+
+  def testInitRequiredKwargs(self):
+    proto = unittest_pb2.TestRequired(a=1, b=1, c=1)
+    self.assertTrue(proto.IsInitialized())
+    self.assertTrue(proto.HasField('a'))
+    self.assertTrue(proto.HasField('b'))
+    self.assertTrue(proto.HasField('c'))
+    self.assertTrue(not proto.HasField('dummy2'))
+    self.assertEqual(1, proto.a)
+    self.assertEqual(1, proto.b)
+    self.assertEqual(1, proto.c)
+
+  def testInitRequiredForeignKwargs(self):
+    proto = unittest_pb2.TestRequiredForeign(
+        optional_message=unittest_pb2.TestRequired(a=1, b=1, c=1))
+    self.assertTrue(proto.IsInitialized())
+    self.assertTrue(proto.HasField('optional_message'))
+    self.assertTrue(proto.optional_message.IsInitialized())
+    self.assertTrue(proto.optional_message.HasField('a'))
+    self.assertTrue(proto.optional_message.HasField('b'))
+    self.assertTrue(proto.optional_message.HasField('c'))
+    self.assertTrue(not proto.optional_message.HasField('dummy2'))
+    self.assertEqual(unittest_pb2.TestRequired(a=1, b=1, c=1),
+                     proto.optional_message)
+    self.assertEqual(1, proto.optional_message.a)
+    self.assertEqual(1, proto.optional_message.b)
+    self.assertEqual(1, proto.optional_message.c)
+
+  def testInitRepeatedKwargs(self):
+    proto = unittest_pb2.TestAllTypes(repeated_int32=[1, 2, 3])
+    self.assertTrue(proto.IsInitialized())
+    self.assertEqual(1, proto.repeated_int32[0])
+    self.assertEqual(2, proto.repeated_int32[1])
+    self.assertEqual(3, proto.repeated_int32[2])
+
+
+class OptionsTest(unittest.TestCase):
+
+  def testMessageOptions(self):
+    proto = message_set_extensions_pb2.TestMessageSet()
+    self.assertEqual(True,
+                     proto.DESCRIPTOR.GetOptions().message_set_wire_format)
+    proto = unittest_pb2.TestAllTypes()
+    self.assertEqual(False,
+                     proto.DESCRIPTOR.GetOptions().message_set_wire_format)
+
+  def testPackedOptions(self):
+    proto = unittest_pb2.TestAllTypes()
+    proto.optional_int32 = 1
+    proto.optional_double = 3.0
+    for field_descriptor, _ in proto.ListFields():
+      self.assertEqual(False, field_descriptor.GetOptions().packed)
+
+    proto = unittest_pb2.TestPackedTypes()
+    proto.packed_int32.append(1)
+    proto.packed_double.append(3.0)
+    for field_descriptor, _ in proto.ListFields():
+      self.assertEqual(True, field_descriptor.GetOptions().packed)
+      self.assertEqual(descriptor.FieldDescriptor.LABEL_REPEATED,
+                       field_descriptor.label)
+
+
+
+class ClassAPITest(unittest.TestCase):
+
+  @unittest.skipIf(
+      api_implementation.Type() == 'cpp' and api_implementation.Version() == 2,
+      'C++ implementation requires a call to MakeDescriptor()')
+  def testMakeClassWithNestedDescriptor(self):
+    leaf_desc = descriptor.Descriptor('leaf', 'package.parent.child.leaf', '',
+                                      containing_type=None, fields=[],
+                                      nested_types=[], enum_types=[],
+                                      extensions=[])
+    child_desc = descriptor.Descriptor('child', 'package.parent.child', '',
+                                       containing_type=None, fields=[],
+                                       nested_types=[leaf_desc], enum_types=[],
+                                       extensions=[])
+    sibling_desc = descriptor.Descriptor('sibling', 'package.parent.sibling',
+                                         '', containing_type=None, fields=[],
+                                         nested_types=[], enum_types=[],
+                                         extensions=[])
+    parent_desc = descriptor.Descriptor('parent', 'package.parent', '',
+                                        containing_type=None, fields=[],
+                                        nested_types=[child_desc, sibling_desc],
+                                        enum_types=[], extensions=[])
+    message_class = reflection.MakeClass(parent_desc)
+    self.assertIn('child', message_class.__dict__)
+    self.assertIn('sibling', message_class.__dict__)
+    self.assertIn('leaf', message_class.child.__dict__)
+
+  def _GetSerializedFileDescriptor(self, name):
+    """Get a serialized representation of a test FileDescriptorProto.
+
+    Args:
+      name: All calls to this must use a unique message name, to avoid
+          collisions in the cpp descriptor pool.
+    Returns:
+      A string containing the serialized form of a test FileDescriptorProto.
+    """
+    file_descriptor_str = (
+        'message_type {'
+        '  name: "' + name + '"'
+        '  field {'
+        '    name: "flat"'
+        '    number: 1'
+        '    label: LABEL_REPEATED'
+        '    type: TYPE_UINT32'
+        '  }'
+        '  field {'
+        '    name: "bar"'
+        '    number: 2'
+        '    label: LABEL_OPTIONAL'
+        '    type: TYPE_MESSAGE'
+        '    type_name: "Bar"'
+        '  }'
+        '  nested_type {'
+        '    name: "Bar"'
+        '    field {'
+        '      name: "baz"'
+        '      number: 3'
+        '      label: LABEL_OPTIONAL'
+        '      type: TYPE_MESSAGE'
+        '      type_name: "Baz"'
+        '    }'
+        '    nested_type {'
+        '      name: "Baz"'
+        '      enum_type {'
+        '        name: "deep_enum"'
+        '        value {'
+        '          name: "VALUE_A"'
+        '          number: 0'
+        '        }'
+        '      }'
+        '      field {'
+        '        name: "deep"'
+        '        number: 4'
+        '        label: LABEL_OPTIONAL'
+        '        type: TYPE_UINT32'
+        '      }'
+        '    }'
+        '  }'
+        '}')
+    file_descriptor = descriptor_pb2.FileDescriptorProto()
+    text_format.Merge(file_descriptor_str, file_descriptor)
+    return file_descriptor.SerializeToString()
+
+  def testParsingFlatClassWithExplicitClassDeclaration(self):
+    """Test that the generated class can parse a flat message."""
+    # TODO(xiaofeng): This test fails with cpp implemetnation in the call
+    # of six.with_metaclass(). The other two callsites of with_metaclass
+    # in this file are both excluded from cpp test, so it might be expected
+    # to fail. Need someone more familiar with the python code to take a
+    # look at this.
+    if api_implementation.Type() != 'python':
+      return
+    file_descriptor = descriptor_pb2.FileDescriptorProto()
+    file_descriptor.ParseFromString(self._GetSerializedFileDescriptor('A'))
+    msg_descriptor = descriptor.MakeDescriptor(
+        file_descriptor.message_type[0])
+
+    class MessageClass(six.with_metaclass(reflection.GeneratedProtocolMessageType, message.Message)):
+      DESCRIPTOR = msg_descriptor
+    msg = MessageClass()
+    msg_str = (
+        'flat: 0 '
+        'flat: 1 '
+        'flat: 2 ')
+    text_format.Merge(msg_str, msg)
+    self.assertEqual(msg.flat, [0, 1, 2])
+
+  def testParsingFlatClass(self):
+    """Test that the generated class can parse a flat message."""
+    file_descriptor = descriptor_pb2.FileDescriptorProto()
+    file_descriptor.ParseFromString(self._GetSerializedFileDescriptor('B'))
+    msg_descriptor = descriptor.MakeDescriptor(
+        file_descriptor.message_type[0])
+    msg_class = reflection.MakeClass(msg_descriptor)
+    msg = msg_class()
+    msg_str = (
+        'flat: 0 '
+        'flat: 1 '
+        'flat: 2 ')
+    text_format.Merge(msg_str, msg)
+    self.assertEqual(msg.flat, [0, 1, 2])
+
+  def testParsingNestedClass(self):
+    """Test that the generated class can parse a nested message."""
+    file_descriptor = descriptor_pb2.FileDescriptorProto()
+    file_descriptor.ParseFromString(self._GetSerializedFileDescriptor('C'))
+    msg_descriptor = descriptor.MakeDescriptor(
+        file_descriptor.message_type[0])
+    msg_class = reflection.MakeClass(msg_descriptor)
+    msg = msg_class()
+    msg_str = (
+        'bar {'
+        '  baz {'
+        '    deep: 4'
+        '  }'
+        '}')
+    text_format.Merge(msg_str, msg)
+    self.assertEqual(msg.bar.baz.deep, 4)
+
+if __name__ == '__main__':
+  unittest.main()
diff --git a/python/google/protobuf/internal/service_reflection_test.py b/python/google/protobuf/internal/service_reflection_test.py
new file mode 100755
index 0000000..98614b7
--- /dev/null
+++ b/python/google/protobuf/internal/service_reflection_test.py
@@ -0,0 +1,138 @@
+#! /usr/bin/env python
+#
+# Protocol Buffers - Google's data interchange format
+# Copyright 2008 Google Inc.  All rights reserved.
+# https://developers.google.com/protocol-buffers/
+#
+# Redistribution and use in source and binary forms, with or without
+# modification, are permitted provided that the following conditions are
+# met:
+#
+#     * Redistributions of source code must retain the above copyright
+# notice, this list of conditions and the following disclaimer.
+#     * Redistributions in binary form must reproduce the above
+# copyright notice, this list of conditions and the following disclaimer
+# in the documentation and/or other materials provided with the
+# distribution.
+#     * Neither the name of Google Inc. nor the names of its
+# contributors may be used to endorse or promote products derived from
+# this software without specific prior written permission.
+#
+# THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+# "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
+# LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
+# A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
+# OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
+# SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
+# LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
+# DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
+# THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
+# (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+# OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+
+"""Tests for google.protobuf.internal.service_reflection."""
+
+__author__ = 'petar@google.com (Petar Petrov)'
+
+try:
+  import unittest2 as unittest
+except ImportError:
+  import unittest
+from google.protobuf import unittest_pb2
+from google.protobuf import service_reflection
+from google.protobuf import service
+
+
+class FooUnitTest(unittest.TestCase):
+
+  def testService(self):
+    class MockRpcChannel(service.RpcChannel):
+      def CallMethod(self, method, controller, request, response, callback):
+        self.method = method
+        self.controller = controller
+        self.request = request
+        callback(response)
+
+    class MockRpcController(service.RpcController):
+      def SetFailed(self, msg):
+        self.failure_message = msg
+
+    self.callback_response = None
+
+    class MyService(unittest_pb2.TestService):
+      pass
+
+    self.callback_response = None
+
+    def MyCallback(response):
+      self.callback_response = response
+
+    rpc_controller = MockRpcController()
+    channel = MockRpcChannel()
+    srvc = MyService()
+    srvc.Foo(rpc_controller, unittest_pb2.FooRequest(), MyCallback)
+    self.assertEqual('Method Foo not implemented.',
+                     rpc_controller.failure_message)
+    self.assertEqual(None, self.callback_response)
+
+    rpc_controller.failure_message = None
+
+    service_descriptor = unittest_pb2.TestService.GetDescriptor()
+    srvc.CallMethod(service_descriptor.methods[1], rpc_controller,
+                    unittest_pb2.BarRequest(), MyCallback)
+    self.assertEqual('Method Bar not implemented.',
+                     rpc_controller.failure_message)
+    self.assertEqual(None, self.callback_response)
+
+    class MyServiceImpl(unittest_pb2.TestService):
+      def Foo(self, rpc_controller, request, done):
+        self.foo_called = True
+      def Bar(self, rpc_controller, request, done):
+        self.bar_called = True
+
+    srvc = MyServiceImpl()
+    rpc_controller.failure_message = None
+    srvc.Foo(rpc_controller, unittest_pb2.FooRequest(), MyCallback)
+    self.assertEqual(None, rpc_controller.failure_message)
+    self.assertEqual(True, srvc.foo_called)
+
+    rpc_controller.failure_message = None
+    srvc.CallMethod(service_descriptor.methods[1], rpc_controller,
+                    unittest_pb2.BarRequest(), MyCallback)
+    self.assertEqual(None, rpc_controller.failure_message)
+    self.assertEqual(True, srvc.bar_called)
+
+  def testServiceStub(self):
+    class MockRpcChannel(service.RpcChannel):
+      def CallMethod(self, method, controller, request,
+                     response_class, callback):
+        self.method = method
+        self.controller = controller
+        self.request = request
+        callback(response_class())
+
+    self.callback_response = None
+
+    def MyCallback(response):
+      self.callback_response = response
+
+    channel = MockRpcChannel()
+    stub = unittest_pb2.TestService_Stub(channel)
+    rpc_controller = 'controller'
+    request = 'request'
+
+    # GetDescriptor now static, still works as instance method for compatibility
+    self.assertEqual(unittest_pb2.TestService_Stub.GetDescriptor(),
+                     stub.GetDescriptor())
+
+    # Invoke method.
+    stub.Foo(rpc_controller, request, MyCallback)
+
+    self.assertIsInstance(self.callback_response, unittest_pb2.FooResponse)
+    self.assertEqual(request, channel.request)
+    self.assertEqual(rpc_controller, channel.controller)
+    self.assertEqual(stub.GetDescriptor().methods[0], channel.method)
+
+
+if __name__ == '__main__':
+  unittest.main()
diff --git a/python/google/protobuf/internal/symbol_database_test.py b/python/google/protobuf/internal/symbol_database_test.py
new file mode 100644
index 0000000..0cb935a
--- /dev/null
+++ b/python/google/protobuf/internal/symbol_database_test.py
@@ -0,0 +1,128 @@
+#! /usr/bin/env python
+#
+# Protocol Buffers - Google's data interchange format
+# Copyright 2008 Google Inc.  All rights reserved.
+# https://developers.google.com/protocol-buffers/
+#
+# Redistribution and use in source and binary forms, with or without
+# modification, are permitted provided that the following conditions are
+# met:
+#
+#     * Redistributions of source code must retain the above copyright
+# notice, this list of conditions and the following disclaimer.
+#     * Redistributions in binary form must reproduce the above
+# copyright notice, this list of conditions and the following disclaimer
+# in the documentation and/or other materials provided with the
+# distribution.
+#     * Neither the name of Google Inc. nor the names of its
+# contributors may be used to endorse or promote products derived from
+# this software without specific prior written permission.
+#
+# THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+# "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
+# LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
+# A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
+# OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
+# SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
+# LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
+# DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
+# THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
+# (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+# OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+
+"""Tests for google.protobuf.symbol_database."""
+
+try:
+  import unittest2 as unittest
+except ImportError:
+  import unittest
+from google.protobuf import unittest_pb2
+from google.protobuf import descriptor
+from google.protobuf import symbol_database
+
+class SymbolDatabaseTest(unittest.TestCase):
+
+  def _Database(self):
+    # TODO(b/17734095): Remove this difference when the C++ implementation
+    # supports multiple databases.
+    if descriptor._USE_C_DESCRIPTORS:
+      return symbol_database.Default()
+    else:
+      db = symbol_database.SymbolDatabase()
+      # Register representative types from unittest_pb2.
+      db.RegisterFileDescriptor(unittest_pb2.DESCRIPTOR)
+      db.RegisterMessage(unittest_pb2.TestAllTypes)
+      db.RegisterMessage(unittest_pb2.TestAllTypes.NestedMessage)
+      db.RegisterMessage(unittest_pb2.TestAllTypes.OptionalGroup)
+      db.RegisterMessage(unittest_pb2.TestAllTypes.RepeatedGroup)
+      db.RegisterEnumDescriptor(unittest_pb2.ForeignEnum.DESCRIPTOR)
+      db.RegisterEnumDescriptor(unittest_pb2.TestAllTypes.NestedEnum.DESCRIPTOR)
+      return db
+
+  def testGetPrototype(self):
+    instance = self._Database().GetPrototype(
+        unittest_pb2.TestAllTypes.DESCRIPTOR)
+    self.assertTrue(instance is unittest_pb2.TestAllTypes)
+
+  def testGetMessages(self):
+    messages = self._Database().GetMessages(
+        ['google/protobuf/unittest.proto'])
+    self.assertTrue(
+        unittest_pb2.TestAllTypes is
+        messages['protobuf_unittest.TestAllTypes'])
+
+  def testGetSymbol(self):
+    self.assertEqual(
+        unittest_pb2.TestAllTypes, self._Database().GetSymbol(
+            'protobuf_unittest.TestAllTypes'))
+    self.assertEqual(
+        unittest_pb2.TestAllTypes.NestedMessage, self._Database().GetSymbol(
+            'protobuf_unittest.TestAllTypes.NestedMessage'))
+    self.assertEqual(
+        unittest_pb2.TestAllTypes.OptionalGroup, self._Database().GetSymbol(
+            'protobuf_unittest.TestAllTypes.OptionalGroup'))
+    self.assertEqual(
+        unittest_pb2.TestAllTypes.RepeatedGroup, self._Database().GetSymbol(
+            'protobuf_unittest.TestAllTypes.RepeatedGroup'))
+
+  def testEnums(self):
+    # Check registration of types in the pool.
+    self.assertEqual(
+        'protobuf_unittest.ForeignEnum',
+        self._Database().pool.FindEnumTypeByName(
+            'protobuf_unittest.ForeignEnum').full_name)
+    self.assertEqual(
+        'protobuf_unittest.TestAllTypes.NestedEnum',
+        self._Database().pool.FindEnumTypeByName(
+            'protobuf_unittest.TestAllTypes.NestedEnum').full_name)
+
+  def testFindMessageTypeByName(self):
+    self.assertEqual(
+        'protobuf_unittest.TestAllTypes',
+        self._Database().pool.FindMessageTypeByName(
+            'protobuf_unittest.TestAllTypes').full_name)
+    self.assertEqual(
+        'protobuf_unittest.TestAllTypes.NestedMessage',
+        self._Database().pool.FindMessageTypeByName(
+            'protobuf_unittest.TestAllTypes.NestedMessage').full_name)
+
+  def testFindFindContainingSymbol(self):
+    # Lookup based on either enum or message.
+    self.assertEqual(
+        'google/protobuf/unittest.proto',
+        self._Database().pool.FindFileContainingSymbol(
+            'protobuf_unittest.TestAllTypes.NestedEnum').name)
+    self.assertEqual(
+        'google/protobuf/unittest.proto',
+        self._Database().pool.FindFileContainingSymbol(
+            'protobuf_unittest.TestAllTypes').name)
+
+  def testFindFileByName(self):
+    self.assertEqual(
+        'google/protobuf/unittest.proto',
+        self._Database().pool.FindFileByName(
+            'google/protobuf/unittest.proto').name)
+
+
+if __name__ == '__main__':
+  unittest.main()
diff --git a/python/google/protobuf/internal/test_bad_identifiers.proto b/python/google/protobuf/internal/test_bad_identifiers.proto
new file mode 100644
index 0000000..c4860ea
--- /dev/null
+++ b/python/google/protobuf/internal/test_bad_identifiers.proto
@@ -0,0 +1,53 @@
+// Protocol Buffers - Google's data interchange format
+// Copyright 2008 Google Inc.  All rights reserved.
+// https://developers.google.com/protocol-buffers/
+//
+// Redistribution and use in source and binary forms, with or without
+// modification, are permitted provided that the following conditions are
+// met:
+//
+//     * Redistributions of source code must retain the above copyright
+// notice, this list of conditions and the following disclaimer.
+//     * Redistributions in binary form must reproduce the above
+// copyright notice, this list of conditions and the following disclaimer
+// in the documentation and/or other materials provided with the
+// distribution.
+//     * Neither the name of Google Inc. nor the names of its
+// contributors may be used to endorse or promote products derived from
+// this software without specific prior written permission.
+//
+// THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+// "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
+// LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
+// A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
+// OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
+// SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
+// LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
+// DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
+// THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
+// (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+// OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+
+// Author: kenton@google.com (Kenton Varda)
+
+syntax = "proto2";
+
+package protobuf_unittest;
+
+option py_generic_services = true;
+
+message TestBadIdentifiers {
+  extensions 100 to max;
+}
+
+// Make sure these reasonable extension names don't conflict with internal
+// variables.
+extend TestBadIdentifiers {
+  optional string message = 100 [default="foo"];
+  optional string descriptor = 101 [default="bar"];
+  optional string reflection = 102 [default="baz"];
+  optional string service = 103 [default="qux"];
+}
+
+message AnotherMessage {}
+service AnotherService {}
diff --git a/python/google/protobuf/internal/test_util.py b/python/google/protobuf/internal/test_util.py
new file mode 100755
index 0000000..ac88fa8
--- /dev/null
+++ b/python/google/protobuf/internal/test_util.py
@@ -0,0 +1,694 @@
+# Protocol Buffers - Google's data interchange format
+# Copyright 2008 Google Inc.  All rights reserved.
+# https://developers.google.com/protocol-buffers/
+#
+# Redistribution and use in source and binary forms, with or without
+# modification, are permitted provided that the following conditions are
+# met:
+#
+#     * Redistributions of source code must retain the above copyright
+# notice, this list of conditions and the following disclaimer.
+#     * Redistributions in binary form must reproduce the above
+# copyright notice, this list of conditions and the following disclaimer
+# in the documentation and/or other materials provided with the
+# distribution.
+#     * Neither the name of Google Inc. nor the names of its
+# contributors may be used to endorse or promote products derived from
+# this software without specific prior written permission.
+#
+# THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+# "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
+# LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
+# A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
+# OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
+# SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
+# LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
+# DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
+# THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
+# (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+# OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+
+"""Utilities for Python proto2 tests.
+
+This is intentionally modeled on C++ code in
+//google/protobuf/test_util.*.
+"""
+
+__author__ = 'robinson@google.com (Will Robinson)'
+
+import os.path
+
+from google.protobuf import unittest_import_pb2
+from google.protobuf import unittest_pb2
+from google.protobuf import descriptor_pb2
+
+# Tests whether the given TestAllTypes message is proto2 or not.
+# This is used to gate several fields/features that only exist
+# for the proto2 version of the message.
+def IsProto2(message):
+  return message.DESCRIPTOR.syntax == "proto2"
+
+def SetAllNonLazyFields(message):
+  """Sets every non-lazy field in the message to a unique value.
+
+  Args:
+    message: A TestAllTypes instance.
+  """
+
+  #
+  # Optional fields.
+  #
+
+  message.optional_int32    = 101
+  message.optional_int64    = 102
+  message.optional_uint32   = 103
+  message.optional_uint64   = 104
+  message.optional_sint32   = 105
+  message.optional_sint64   = 106
+  message.optional_fixed32  = 107
+  message.optional_fixed64  = 108
+  message.optional_sfixed32 = 109
+  message.optional_sfixed64 = 110
+  message.optional_float    = 111
+  message.optional_double   = 112
+  message.optional_bool     = True
+  message.optional_string   = u'115'
+  message.optional_bytes    = b'116'
+
+  if IsProto2(message):
+    message.optionalgroup.a = 117
+  message.optional_nested_message.bb = 118
+  message.optional_foreign_message.c = 119
+  message.optional_import_message.d = 120
+  message.optional_public_import_message.e = 126
+
+  message.optional_nested_enum = unittest_pb2.TestAllTypes.BAZ
+  message.optional_foreign_enum = unittest_pb2.FOREIGN_BAZ
+  if IsProto2(message):
+    message.optional_import_enum = unittest_import_pb2.IMPORT_BAZ
+
+  message.optional_string_piece = u'124'
+  message.optional_cord = u'125'
+
+  #
+  # Repeated fields.
+  #
+
+  message.repeated_int32.append(201)
+  message.repeated_int64.append(202)
+  message.repeated_uint32.append(203)
+  message.repeated_uint64.append(204)
+  message.repeated_sint32.append(205)
+  message.repeated_sint64.append(206)
+  message.repeated_fixed32.append(207)
+  message.repeated_fixed64.append(208)
+  message.repeated_sfixed32.append(209)
+  message.repeated_sfixed64.append(210)
+  message.repeated_float.append(211)
+  message.repeated_double.append(212)
+  message.repeated_bool.append(True)
+  message.repeated_string.append(u'215')
+  message.repeated_bytes.append(b'216')
+
+  if IsProto2(message):
+    message.repeatedgroup.add().a = 217
+  message.repeated_nested_message.add().bb = 218
+  message.repeated_foreign_message.add().c = 219
+  message.repeated_import_message.add().d = 220
+  message.repeated_lazy_message.add().bb = 227
+
+  message.repeated_nested_enum.append(unittest_pb2.TestAllTypes.BAR)
+  message.repeated_foreign_enum.append(unittest_pb2.FOREIGN_BAR)
+  if IsProto2(message):
+    message.repeated_import_enum.append(unittest_import_pb2.IMPORT_BAR)
+
+  message.repeated_string_piece.append(u'224')
+  message.repeated_cord.append(u'225')
+
+  # Add a second one of each field.
+  message.repeated_int32.append(301)
+  message.repeated_int64.append(302)
+  message.repeated_uint32.append(303)
+  message.repeated_uint64.append(304)
+  message.repeated_sint32.append(305)
+  message.repeated_sint64.append(306)
+  message.repeated_fixed32.append(307)
+  message.repeated_fixed64.append(308)
+  message.repeated_sfixed32.append(309)
+  message.repeated_sfixed64.append(310)
+  message.repeated_float.append(311)
+  message.repeated_double.append(312)
+  message.repeated_bool.append(False)
+  message.repeated_string.append(u'315')
+  message.repeated_bytes.append(b'316')
+
+  if IsProto2(message):
+    message.repeatedgroup.add().a = 317
+  message.repeated_nested_message.add().bb = 318
+  message.repeated_foreign_message.add().c = 319
+  message.repeated_import_message.add().d = 320
+  message.repeated_lazy_message.add().bb = 327
+
+  message.repeated_nested_enum.append(unittest_pb2.TestAllTypes.BAZ)
+  message.repeated_foreign_enum.append(unittest_pb2.FOREIGN_BAZ)
+  if IsProto2(message):
+    message.repeated_import_enum.append(unittest_import_pb2.IMPORT_BAZ)
+
+  message.repeated_string_piece.append(u'324')
+  message.repeated_cord.append(u'325')
+
+  #
+  # Fields that have defaults.
+  #
+
+  if IsProto2(message):
+    message.default_int32 = 401
+    message.default_int64 = 402
+    message.default_uint32 = 403
+    message.default_uint64 = 404
+    message.default_sint32 = 405
+    message.default_sint64 = 406
+    message.default_fixed32 = 407
+    message.default_fixed64 = 408
+    message.default_sfixed32 = 409
+    message.default_sfixed64 = 410
+    message.default_float = 411
+    message.default_double = 412
+    message.default_bool = False
+    message.default_string = '415'
+    message.default_bytes = b'416'
+
+    message.default_nested_enum = unittest_pb2.TestAllTypes.FOO
+    message.default_foreign_enum = unittest_pb2.FOREIGN_FOO
+    message.default_import_enum = unittest_import_pb2.IMPORT_FOO
+
+    message.default_string_piece = '424'
+    message.default_cord = '425'
+
+  message.oneof_uint32 = 601
+  message.oneof_nested_message.bb = 602
+  message.oneof_string = '603'
+  message.oneof_bytes = b'604'
+
+
+def SetAllFields(message):
+  SetAllNonLazyFields(message)
+  message.optional_lazy_message.bb = 127
+
+
+def SetAllExtensions(message):
+  """Sets every extension in the message to a unique value.
+
+  Args:
+    message: A unittest_pb2.TestAllExtensions instance.
+  """
+
+  extensions = message.Extensions
+  pb2 = unittest_pb2
+  import_pb2 = unittest_import_pb2
+
+  #
+  # Optional fields.
+  #
+
+  extensions[pb2.optional_int32_extension] = 101
+  extensions[pb2.optional_int64_extension] = 102
+  extensions[pb2.optional_uint32_extension] = 103
+  extensions[pb2.optional_uint64_extension] = 104
+  extensions[pb2.optional_sint32_extension] = 105
+  extensions[pb2.optional_sint64_extension] = 106
+  extensions[pb2.optional_fixed32_extension] = 107
+  extensions[pb2.optional_fixed64_extension] = 108
+  extensions[pb2.optional_sfixed32_extension] = 109
+  extensions[pb2.optional_sfixed64_extension] = 110
+  extensions[pb2.optional_float_extension] = 111
+  extensions[pb2.optional_double_extension] = 112
+  extensions[pb2.optional_bool_extension] = True
+  extensions[pb2.optional_string_extension] = u'115'
+  extensions[pb2.optional_bytes_extension] = b'116'
+
+  extensions[pb2.optionalgroup_extension].a = 117
+  extensions[pb2.optional_nested_message_extension].bb = 118
+  extensions[pb2.optional_foreign_message_extension].c = 119
+  extensions[pb2.optional_import_message_extension].d = 120
+  extensions[pb2.optional_public_import_message_extension].e = 126
+  extensions[pb2.optional_lazy_message_extension].bb = 127
+
+  extensions[pb2.optional_nested_enum_extension] = pb2.TestAllTypes.BAZ
+  extensions[pb2.optional_nested_enum_extension] = pb2.TestAllTypes.BAZ
+  extensions[pb2.optional_foreign_enum_extension] = pb2.FOREIGN_BAZ
+  extensions[pb2.optional_import_enum_extension] = import_pb2.IMPORT_BAZ
+
+  extensions[pb2.optional_string_piece_extension] = u'124'
+  extensions[pb2.optional_cord_extension] = u'125'
+
+  #
+  # Repeated fields.
+  #
+
+  extensions[pb2.repeated_int32_extension].append(201)
+  extensions[pb2.repeated_int64_extension].append(202)
+  extensions[pb2.repeated_uint32_extension].append(203)
+  extensions[pb2.repeated_uint64_extension].append(204)
+  extensions[pb2.repeated_sint32_extension].append(205)
+  extensions[pb2.repeated_sint64_extension].append(206)
+  extensions[pb2.repeated_fixed32_extension].append(207)
+  extensions[pb2.repeated_fixed64_extension].append(208)
+  extensions[pb2.repeated_sfixed32_extension].append(209)
+  extensions[pb2.repeated_sfixed64_extension].append(210)
+  extensions[pb2.repeated_float_extension].append(211)
+  extensions[pb2.repeated_double_extension].append(212)
+  extensions[pb2.repeated_bool_extension].append(True)
+  extensions[pb2.repeated_string_extension].append(u'215')
+  extensions[pb2.repeated_bytes_extension].append(b'216')
+
+  extensions[pb2.repeatedgroup_extension].add().a = 217
+  extensions[pb2.repeated_nested_message_extension].add().bb = 218
+  extensions[pb2.repeated_foreign_message_extension].add().c = 219
+  extensions[pb2.repeated_import_message_extension].add().d = 220
+  extensions[pb2.repeated_lazy_message_extension].add().bb = 227
+
+  extensions[pb2.repeated_nested_enum_extension].append(pb2.TestAllTypes.BAR)
+  extensions[pb2.repeated_foreign_enum_extension].append(pb2.FOREIGN_BAR)
+  extensions[pb2.repeated_import_enum_extension].append(import_pb2.IMPORT_BAR)
+
+  extensions[pb2.repeated_string_piece_extension].append(u'224')
+  extensions[pb2.repeated_cord_extension].append(u'225')
+
+  # Append a second one of each field.
+  extensions[pb2.repeated_int32_extension].append(301)
+  extensions[pb2.repeated_int64_extension].append(302)
+  extensions[pb2.repeated_uint32_extension].append(303)
+  extensions[pb2.repeated_uint64_extension].append(304)
+  extensions[pb2.repeated_sint32_extension].append(305)
+  extensions[pb2.repeated_sint64_extension].append(306)
+  extensions[pb2.repeated_fixed32_extension].append(307)
+  extensions[pb2.repeated_fixed64_extension].append(308)
+  extensions[pb2.repeated_sfixed32_extension].append(309)
+  extensions[pb2.repeated_sfixed64_extension].append(310)
+  extensions[pb2.repeated_float_extension].append(311)
+  extensions[pb2.repeated_double_extension].append(312)
+  extensions[pb2.repeated_bool_extension].append(False)
+  extensions[pb2.repeated_string_extension].append(u'315')
+  extensions[pb2.repeated_bytes_extension].append(b'316')
+
+  extensions[pb2.repeatedgroup_extension].add().a = 317
+  extensions[pb2.repeated_nested_message_extension].add().bb = 318
+  extensions[pb2.repeated_foreign_message_extension].add().c = 319
+  extensions[pb2.repeated_import_message_extension].add().d = 320
+  extensions[pb2.repeated_lazy_message_extension].add().bb = 327
+
+  extensions[pb2.repeated_nested_enum_extension].append(pb2.TestAllTypes.BAZ)
+  extensions[pb2.repeated_foreign_enum_extension].append(pb2.FOREIGN_BAZ)
+  extensions[pb2.repeated_import_enum_extension].append(import_pb2.IMPORT_BAZ)
+
+  extensions[pb2.repeated_string_piece_extension].append(u'324')
+  extensions[pb2.repeated_cord_extension].append(u'325')
+
+  #
+  # Fields with defaults.
+  #
+
+  extensions[pb2.default_int32_extension] = 401
+  extensions[pb2.default_int64_extension] = 402
+  extensions[pb2.default_uint32_extension] = 403
+  extensions[pb2.default_uint64_extension] = 404
+  extensions[pb2.default_sint32_extension] = 405
+  extensions[pb2.default_sint64_extension] = 406
+  extensions[pb2.default_fixed32_extension] = 407
+  extensions[pb2.default_fixed64_extension] = 408
+  extensions[pb2.default_sfixed32_extension] = 409
+  extensions[pb2.default_sfixed64_extension] = 410
+  extensions[pb2.default_float_extension] = 411
+  extensions[pb2.default_double_extension] = 412
+  extensions[pb2.default_bool_extension] = False
+  extensions[pb2.default_string_extension] = u'415'
+  extensions[pb2.default_bytes_extension] = b'416'
+
+  extensions[pb2.default_nested_enum_extension] = pb2.TestAllTypes.FOO
+  extensions[pb2.default_foreign_enum_extension] = pb2.FOREIGN_FOO
+  extensions[pb2.default_import_enum_extension] = import_pb2.IMPORT_FOO
+
+  extensions[pb2.default_string_piece_extension] = u'424'
+  extensions[pb2.default_cord_extension] = '425'
+
+  extensions[pb2.oneof_uint32_extension] = 601
+  extensions[pb2.oneof_nested_message_extension].bb = 602
+  extensions[pb2.oneof_string_extension] = u'603'
+  extensions[pb2.oneof_bytes_extension] = b'604'
+
+
+def SetAllFieldsAndExtensions(message):
+  """Sets every field and extension in the message to a unique value.
+
+  Args:
+    message: A unittest_pb2.TestAllExtensions message.
+  """
+  message.my_int = 1
+  message.my_string = 'foo'
+  message.my_float = 1.0
+  message.Extensions[unittest_pb2.my_extension_int] = 23
+  message.Extensions[unittest_pb2.my_extension_string] = 'bar'
+
+
+def ExpectAllFieldsAndExtensionsInOrder(serialized):
+  """Ensures that serialized is the serialization we expect for a message
+  filled with SetAllFieldsAndExtensions().  (Specifically, ensures that the
+  serialization is in canonical, tag-number order).
+  """
+  my_extension_int = unittest_pb2.my_extension_int
+  my_extension_string = unittest_pb2.my_extension_string
+  expected_strings = []
+  message = unittest_pb2.TestFieldOrderings()
+  message.my_int = 1  # Field 1.
+  expected_strings.append(message.SerializeToString())
+  message.Clear()
+  message.Extensions[my_extension_int] = 23  # Field 5.
+  expected_strings.append(message.SerializeToString())
+  message.Clear()
+  message.my_string = 'foo'  # Field 11.
+  expected_strings.append(message.SerializeToString())
+  message.Clear()
+  message.Extensions[my_extension_string] = 'bar'  # Field 50.
+  expected_strings.append(message.SerializeToString())
+  message.Clear()
+  message.my_float = 1.0
+  expected_strings.append(message.SerializeToString())
+  message.Clear()
+  expected = b''.join(expected_strings)
+
+  if expected != serialized:
+    raise ValueError('Expected %r, found %r' % (expected, serialized))
+
+
+def ExpectAllFieldsSet(test_case, message):
+  """Check all fields for correct values have after Set*Fields() is called."""
+  test_case.assertTrue(message.HasField('optional_int32'))
+  test_case.assertTrue(message.HasField('optional_int64'))
+  test_case.assertTrue(message.HasField('optional_uint32'))
+  test_case.assertTrue(message.HasField('optional_uint64'))
+  test_case.assertTrue(message.HasField('optional_sint32'))
+  test_case.assertTrue(message.HasField('optional_sint64'))
+  test_case.assertTrue(message.HasField('optional_fixed32'))
+  test_case.assertTrue(message.HasField('optional_fixed64'))
+  test_case.assertTrue(message.HasField('optional_sfixed32'))
+  test_case.assertTrue(message.HasField('optional_sfixed64'))
+  test_case.assertTrue(message.HasField('optional_float'))
+  test_case.assertTrue(message.HasField('optional_double'))
+  test_case.assertTrue(message.HasField('optional_bool'))
+  test_case.assertTrue(message.HasField('optional_string'))
+  test_case.assertTrue(message.HasField('optional_bytes'))
+
+  if IsProto2(message):
+    test_case.assertTrue(message.HasField('optionalgroup'))
+  test_case.assertTrue(message.HasField('optional_nested_message'))
+  test_case.assertTrue(message.HasField('optional_foreign_message'))
+  test_case.assertTrue(message.HasField('optional_import_message'))
+
+  test_case.assertTrue(message.optionalgroup.HasField('a'))
+  test_case.assertTrue(message.optional_nested_message.HasField('bb'))
+  test_case.assertTrue(message.optional_foreign_message.HasField('c'))
+  test_case.assertTrue(message.optional_import_message.HasField('d'))
+
+  test_case.assertTrue(message.HasField('optional_nested_enum'))
+  test_case.assertTrue(message.HasField('optional_foreign_enum'))
+  if IsProto2(message):
+    test_case.assertTrue(message.HasField('optional_import_enum'))
+
+  test_case.assertTrue(message.HasField('optional_string_piece'))
+  test_case.assertTrue(message.HasField('optional_cord'))
+
+  test_case.assertEqual(101, message.optional_int32)
+  test_case.assertEqual(102, message.optional_int64)
+  test_case.assertEqual(103, message.optional_uint32)
+  test_case.assertEqual(104, message.optional_uint64)
+  test_case.assertEqual(105, message.optional_sint32)
+  test_case.assertEqual(106, message.optional_sint64)
+  test_case.assertEqual(107, message.optional_fixed32)
+  test_case.assertEqual(108, message.optional_fixed64)
+  test_case.assertEqual(109, message.optional_sfixed32)
+  test_case.assertEqual(110, message.optional_sfixed64)
+  test_case.assertEqual(111, message.optional_float)
+  test_case.assertEqual(112, message.optional_double)
+  test_case.assertEqual(True, message.optional_bool)
+  test_case.assertEqual('115', message.optional_string)
+  test_case.assertEqual(b'116', message.optional_bytes)
+
+  if IsProto2(message):
+    test_case.assertEqual(117, message.optionalgroup.a)
+  test_case.assertEqual(118, message.optional_nested_message.bb)
+  test_case.assertEqual(119, message.optional_foreign_message.c)
+  test_case.assertEqual(120, message.optional_import_message.d)
+  test_case.assertEqual(126, message.optional_public_import_message.e)
+  test_case.assertEqual(127, message.optional_lazy_message.bb)
+
+  test_case.assertEqual(unittest_pb2.TestAllTypes.BAZ,
+                        message.optional_nested_enum)
+  test_case.assertEqual(unittest_pb2.FOREIGN_BAZ,
+                        message.optional_foreign_enum)
+  if IsProto2(message):
+    test_case.assertEqual(unittest_import_pb2.IMPORT_BAZ,
+                          message.optional_import_enum)
+
+  # -----------------------------------------------------------------
+
+  test_case.assertEqual(2, len(message.repeated_int32))
+  test_case.assertEqual(2, len(message.repeated_int64))
+  test_case.assertEqual(2, len(message.repeated_uint32))
+  test_case.assertEqual(2, len(message.repeated_uint64))
+  test_case.assertEqual(2, len(message.repeated_sint32))
+  test_case.assertEqual(2, len(message.repeated_sint64))
+  test_case.assertEqual(2, len(message.repeated_fixed32))
+  test_case.assertEqual(2, len(message.repeated_fixed64))
+  test_case.assertEqual(2, len(message.repeated_sfixed32))
+  test_case.assertEqual(2, len(message.repeated_sfixed64))
+  test_case.assertEqual(2, len(message.repeated_float))
+  test_case.assertEqual(2, len(message.repeated_double))
+  test_case.assertEqual(2, len(message.repeated_bool))
+  test_case.assertEqual(2, len(message.repeated_string))
+  test_case.assertEqual(2, len(message.repeated_bytes))
+
+  if IsProto2(message):
+    test_case.assertEqual(2, len(message.repeatedgroup))
+  test_case.assertEqual(2, len(message.repeated_nested_message))
+  test_case.assertEqual(2, len(message.repeated_foreign_message))
+  test_case.assertEqual(2, len(message.repeated_import_message))
+  test_case.assertEqual(2, len(message.repeated_nested_enum))
+  test_case.assertEqual(2, len(message.repeated_foreign_enum))
+  if IsProto2(message):
+    test_case.assertEqual(2, len(message.repeated_import_enum))
+
+  test_case.assertEqual(2, len(message.repeated_string_piece))
+  test_case.assertEqual(2, len(message.repeated_cord))
+
+  test_case.assertEqual(201, message.repeated_int32[0])
+  test_case.assertEqual(202, message.repeated_int64[0])
+  test_case.assertEqual(203, message.repeated_uint32[0])
+  test_case.assertEqual(204, message.repeated_uint64[0])
+  test_case.assertEqual(205, message.repeated_sint32[0])
+  test_case.assertEqual(206, message.repeated_sint64[0])
+  test_case.assertEqual(207, message.repeated_fixed32[0])
+  test_case.assertEqual(208, message.repeated_fixed64[0])
+  test_case.assertEqual(209, message.repeated_sfixed32[0])
+  test_case.assertEqual(210, message.repeated_sfixed64[0])
+  test_case.assertEqual(211, message.repeated_float[0])
+  test_case.assertEqual(212, message.repeated_double[0])
+  test_case.assertEqual(True, message.repeated_bool[0])
+  test_case.assertEqual('215', message.repeated_string[0])
+  test_case.assertEqual(b'216', message.repeated_bytes[0])
+
+  if IsProto2(message):
+    test_case.assertEqual(217, message.repeatedgroup[0].a)
+  test_case.assertEqual(218, message.repeated_nested_message[0].bb)
+  test_case.assertEqual(219, message.repeated_foreign_message[0].c)
+  test_case.assertEqual(220, message.repeated_import_message[0].d)
+  test_case.assertEqual(227, message.repeated_lazy_message[0].bb)
+
+  test_case.assertEqual(unittest_pb2.TestAllTypes.BAR,
+                        message.repeated_nested_enum[0])
+  test_case.assertEqual(unittest_pb2.FOREIGN_BAR,
+                        message.repeated_foreign_enum[0])
+  if IsProto2(message):
+    test_case.assertEqual(unittest_import_pb2.IMPORT_BAR,
+                          message.repeated_import_enum[0])
+
+  test_case.assertEqual(301, message.repeated_int32[1])
+  test_case.assertEqual(302, message.repeated_int64[1])
+  test_case.assertEqual(303, message.repeated_uint32[1])
+  test_case.assertEqual(304, message.repeated_uint64[1])
+  test_case.assertEqual(305, message.repeated_sint32[1])
+  test_case.assertEqual(306, message.repeated_sint64[1])
+  test_case.assertEqual(307, message.repeated_fixed32[1])
+  test_case.assertEqual(308, message.repeated_fixed64[1])
+  test_case.assertEqual(309, message.repeated_sfixed32[1])
+  test_case.assertEqual(310, message.repeated_sfixed64[1])
+  test_case.assertEqual(311, message.repeated_float[1])
+  test_case.assertEqual(312, message.repeated_double[1])
+  test_case.assertEqual(False, message.repeated_bool[1])
+  test_case.assertEqual('315', message.repeated_string[1])
+  test_case.assertEqual(b'316', message.repeated_bytes[1])
+
+  if IsProto2(message):
+    test_case.assertEqual(317, message.repeatedgroup[1].a)
+  test_case.assertEqual(318, message.repeated_nested_message[1].bb)
+  test_case.assertEqual(319, message.repeated_foreign_message[1].c)
+  test_case.assertEqual(320, message.repeated_import_message[1].d)
+  test_case.assertEqual(327, message.repeated_lazy_message[1].bb)
+
+  test_case.assertEqual(unittest_pb2.TestAllTypes.BAZ,
+                        message.repeated_nested_enum[1])
+  test_case.assertEqual(unittest_pb2.FOREIGN_BAZ,
+                        message.repeated_foreign_enum[1])
+  if IsProto2(message):
+    test_case.assertEqual(unittest_import_pb2.IMPORT_BAZ,
+                          message.repeated_import_enum[1])
+
+  # -----------------------------------------------------------------
+
+  if IsProto2(message):
+    test_case.assertTrue(message.HasField('default_int32'))
+    test_case.assertTrue(message.HasField('default_int64'))
+    test_case.assertTrue(message.HasField('default_uint32'))
+    test_case.assertTrue(message.HasField('default_uint64'))
+    test_case.assertTrue(message.HasField('default_sint32'))
+    test_case.assertTrue(message.HasField('default_sint64'))
+    test_case.assertTrue(message.HasField('default_fixed32'))
+    test_case.assertTrue(message.HasField('default_fixed64'))
+    test_case.assertTrue(message.HasField('default_sfixed32'))
+    test_case.assertTrue(message.HasField('default_sfixed64'))
+    test_case.assertTrue(message.HasField('default_float'))
+    test_case.assertTrue(message.HasField('default_double'))
+    test_case.assertTrue(message.HasField('default_bool'))
+    test_case.assertTrue(message.HasField('default_string'))
+    test_case.assertTrue(message.HasField('default_bytes'))
+
+    test_case.assertTrue(message.HasField('default_nested_enum'))
+    test_case.assertTrue(message.HasField('default_foreign_enum'))
+    test_case.assertTrue(message.HasField('default_import_enum'))
+
+    test_case.assertEqual(401, message.default_int32)
+    test_case.assertEqual(402, message.default_int64)
+    test_case.assertEqual(403, message.default_uint32)
+    test_case.assertEqual(404, message.default_uint64)
+    test_case.assertEqual(405, message.default_sint32)
+    test_case.assertEqual(406, message.default_sint64)
+    test_case.assertEqual(407, message.default_fixed32)
+    test_case.assertEqual(408, message.default_fixed64)
+    test_case.assertEqual(409, message.default_sfixed32)
+    test_case.assertEqual(410, message.default_sfixed64)
+    test_case.assertEqual(411, message.default_float)
+    test_case.assertEqual(412, message.default_double)
+    test_case.assertEqual(False, message.default_bool)
+    test_case.assertEqual('415', message.default_string)
+    test_case.assertEqual(b'416', message.default_bytes)
+
+    test_case.assertEqual(unittest_pb2.TestAllTypes.FOO,
+                          message.default_nested_enum)
+    test_case.assertEqual(unittest_pb2.FOREIGN_FOO,
+                          message.default_foreign_enum)
+    test_case.assertEqual(unittest_import_pb2.IMPORT_FOO,
+                          message.default_import_enum)
+
+
+def GoldenFile(filename):
+  """Finds the given golden file and returns a file object representing it."""
+
+  # Search up the directory tree looking for the C++ protobuf source code.
+  path = '.'
+  while os.path.exists(path):
+    if os.path.exists(os.path.join(path, 'src/google/protobuf')):
+      # Found it.  Load the golden file from the testdata directory.
+      full_path = os.path.join(path, 'src/google/protobuf/testdata', filename)
+      return open(full_path, 'rb')
+    path = os.path.join(path, '..')
+
+  # Search internally.
+  path = '.'
+  full_path = os.path.join(path, 'third_party/py/google/protobuf/testdata',
+                           filename)
+  if os.path.exists(full_path):
+    # Found it.  Load the golden file from the testdata directory.
+    return open(full_path, 'rb')
+
+  raise RuntimeError(
+      'Could not find golden files.  This test must be run from within the '
+      'protobuf source package so that it can read test data files from the '
+      'C++ source tree.')
+
+
+def GoldenFileData(filename):
+  """Finds the given golden file and returns its contents."""
+  with GoldenFile(filename) as f:
+    return f.read()
+
+
+def SetAllPackedFields(message):
+  """Sets every field in the message to a unique value.
+
+  Args:
+    message: A TestPackedTypes instance.
+  """
+  message.packed_int32.extend([601, 701])
+  message.packed_int64.extend([602, 702])
+  message.packed_uint32.extend([603, 703])
+  message.packed_uint64.extend([604, 704])
+  message.packed_sint32.extend([605, 705])
+  message.packed_sint64.extend([606, 706])
+  message.packed_fixed32.extend([607, 707])
+  message.packed_fixed64.extend([608, 708])
+  message.packed_sfixed32.extend([609, 709])
+  message.packed_sfixed64.extend([610, 710])
+  message.packed_float.extend([611.0, 711.0])
+  message.packed_double.extend([612.0, 712.0])
+  message.packed_bool.extend([True, False])
+  message.packed_enum.extend([unittest_pb2.FOREIGN_BAR,
+                              unittest_pb2.FOREIGN_BAZ])
+
+
+def SetAllPackedExtensions(message):
+  """Sets every extension in the message to a unique value.
+
+  Args:
+    message: A unittest_pb2.TestPackedExtensions instance.
+  """
+  extensions = message.Extensions
+  pb2 = unittest_pb2
+
+  extensions[pb2.packed_int32_extension].extend([601, 701])
+  extensions[pb2.packed_int64_extension].extend([602, 702])
+  extensions[pb2.packed_uint32_extension].extend([603, 703])
+  extensions[pb2.packed_uint64_extension].extend([604, 704])
+  extensions[pb2.packed_sint32_extension].extend([605, 705])
+  extensions[pb2.packed_sint64_extension].extend([606, 706])
+  extensions[pb2.packed_fixed32_extension].extend([607, 707])
+  extensions[pb2.packed_fixed64_extension].extend([608, 708])
+  extensions[pb2.packed_sfixed32_extension].extend([609, 709])
+  extensions[pb2.packed_sfixed64_extension].extend([610, 710])
+  extensions[pb2.packed_float_extension].extend([611.0, 711.0])
+  extensions[pb2.packed_double_extension].extend([612.0, 712.0])
+  extensions[pb2.packed_bool_extension].extend([True, False])
+  extensions[pb2.packed_enum_extension].extend([unittest_pb2.FOREIGN_BAR,
+                                                unittest_pb2.FOREIGN_BAZ])
+
+
+def SetAllUnpackedFields(message):
+  """Sets every field in the message to a unique value.
+
+  Args:
+    message: A unittest_pb2.TestUnpackedTypes instance.
+  """
+  message.unpacked_int32.extend([601, 701])
+  message.unpacked_int64.extend([602, 702])
+  message.unpacked_uint32.extend([603, 703])
+  message.unpacked_uint64.extend([604, 704])
+  message.unpacked_sint32.extend([605, 705])
+  message.unpacked_sint64.extend([606, 706])
+  message.unpacked_fixed32.extend([607, 707])
+  message.unpacked_fixed64.extend([608, 708])
+  message.unpacked_sfixed32.extend([609, 709])
+  message.unpacked_sfixed64.extend([610, 710])
+  message.unpacked_float.extend([611.0, 711.0])
+  message.unpacked_double.extend([612.0, 712.0])
+  message.unpacked_bool.extend([True, False])
+  message.unpacked_enum.extend([unittest_pb2.FOREIGN_BAR,
+                                unittest_pb2.FOREIGN_BAZ])
diff --git a/python/google/protobuf/internal/text_encoding_test.py b/python/google/protobuf/internal/text_encoding_test.py
new file mode 100755
index 0000000..338a287
--- /dev/null
+++ b/python/google/protobuf/internal/text_encoding_test.py
@@ -0,0 +1,71 @@
+#! /usr/bin/env python
+#
+# Protocol Buffers - Google's data interchange format
+# Copyright 2008 Google Inc.  All rights reserved.
+# https://developers.google.com/protocol-buffers/
+#
+# Redistribution and use in source and binary forms, with or without
+# modification, are permitted provided that the following conditions are
+# met:
+#
+#     * Redistributions of source code must retain the above copyright
+# notice, this list of conditions and the following disclaimer.
+#     * Redistributions in binary form must reproduce the above
+# copyright notice, this list of conditions and the following disclaimer
+# in the documentation and/or other materials provided with the
+# distribution.
+#     * Neither the name of Google Inc. nor the names of its
+# contributors may be used to endorse or promote products derived from
+# this software without specific prior written permission.
+#
+# THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+# "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
+# LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
+# A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
+# OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
+# SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
+# LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
+# DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
+# THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
+# (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+# OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+
+"""Tests for google.protobuf.text_encoding."""
+
+try:
+  import unittest2 as unittest
+except ImportError:
+  import unittest
+from google.protobuf import text_encoding
+
+TEST_VALUES = [
+    ("foo\\rbar\\nbaz\\t",
+     "foo\\rbar\\nbaz\\t",
+     b"foo\rbar\nbaz\t"),
+    ("\\'full of \\\"sound\\\" and \\\"fury\\\"\\'",
+     "\\'full of \\\"sound\\\" and \\\"fury\\\"\\'",
+     b"'full of \"sound\" and \"fury\"'"),
+    ("signi\\\\fying\\\\ nothing\\\\",
+     "signi\\\\fying\\\\ nothing\\\\",
+     b"signi\\fying\\ nothing\\"),
+    ("\\010\\t\\n\\013\\014\\r",
+     "\x08\\t\\n\x0b\x0c\\r",
+     b"\010\011\012\013\014\015")]
+
+
+class TextEncodingTestCase(unittest.TestCase):
+  def testCEscape(self):
+    for escaped, escaped_utf8, unescaped in TEST_VALUES:
+      self.assertEqual(escaped,
+                        text_encoding.CEscape(unescaped, as_utf8=False))
+      self.assertEqual(escaped_utf8,
+                        text_encoding.CEscape(unescaped, as_utf8=True))
+
+  def testCUnescape(self):
+    for escaped, escaped_utf8, unescaped in TEST_VALUES:
+      self.assertEqual(unescaped, text_encoding.CUnescape(escaped))
+      self.assertEqual(unescaped, text_encoding.CUnescape(escaped_utf8))
+
+
+if __name__ == "__main__":
+  unittest.main()
diff --git a/python/google/protobuf/internal/text_format_test.py b/python/google/protobuf/internal/text_format_test.py
new file mode 100755
index 0000000..0e14556
--- /dev/null
+++ b/python/google/protobuf/internal/text_format_test.py
@@ -0,0 +1,1039 @@
+#! /usr/bin/env python
+#
+# Protocol Buffers - Google's data interchange format
+# Copyright 2008 Google Inc.  All rights reserved.
+# https://developers.google.com/protocol-buffers/
+#
+# Redistribution and use in source and binary forms, with or without
+# modification, are permitted provided that the following conditions are
+# met:
+#
+#     * Redistributions of source code must retain the above copyright
+# notice, this list of conditions and the following disclaimer.
+#     * Redistributions in binary form must reproduce the above
+# copyright notice, this list of conditions and the following disclaimer
+# in the documentation and/or other materials provided with the
+# distribution.
+#     * Neither the name of Google Inc. nor the names of its
+# contributors may be used to endorse or promote products derived from
+# this software without specific prior written permission.
+#
+# THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+# "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
+# LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
+# A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
+# OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
+# SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
+# LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
+# DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
+# THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
+# (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+# OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+
+"""Test for google.protobuf.text_format."""
+
+__author__ = 'kenton@google.com (Kenton Varda)'
+
+
+import re
+import six
+import string
+
+try:
+  import unittest2 as unittest
+except ImportError:
+  import unittest
+from google.protobuf.internal import _parameterized
+
+from google.protobuf import map_unittest_pb2
+from google.protobuf import unittest_mset_pb2
+from google.protobuf import unittest_pb2
+from google.protobuf import unittest_proto3_arena_pb2
+from google.protobuf.internal import api_implementation
+from google.protobuf.internal import test_util
+from google.protobuf.internal import message_set_extensions_pb2
+from google.protobuf import text_format
+
+
+# Low-level nuts-n-bolts tests.
+class SimpleTextFormatTests(unittest.TestCase):
+
+  # The members of _QUOTES are formatted into a regexp template that
+  # expects single characters.  Therefore it's an error (in addition to being
+  # non-sensical in the first place) to try to specify a "quote mark" that is
+  # more than one character.
+  def TestQuoteMarksAreSingleChars(self):
+    for quote in text_format._QUOTES:
+      self.assertEqual(1, len(quote))
+
+
+# Base class with some common functionality.
+class TextFormatBase(unittest.TestCase):
+
+  def ReadGolden(self, golden_filename):
+    with test_util.GoldenFile(golden_filename) as f:
+      return (f.readlines() if str is bytes else  # PY3
+              [golden_line.decode('utf-8') for golden_line in f])
+
+  def CompareToGoldenFile(self, text, golden_filename):
+    golden_lines = self.ReadGolden(golden_filename)
+    self.assertMultiLineEqual(text, ''.join(golden_lines))
+
+  def CompareToGoldenText(self, text, golden_text):
+    self.assertEqual(text, golden_text)
+
+  def RemoveRedundantZeros(self, text):
+    # Some platforms print 1e+5 as 1e+005.  This is fine, but we need to remove
+    # these zeros in order to match the golden file.
+    text = text.replace('e+0','e+').replace('e+0','e+') \
+               .replace('e-0','e-').replace('e-0','e-')
+    # Floating point fields are printed with .0 suffix even if they are
+    # actualy integer numbers.
+    text = re.compile('\.0$', re.MULTILINE).sub('', text)
+    return text
+
+
+@_parameterized.Parameters(
+    (unittest_pb2),
+    (unittest_proto3_arena_pb2))
+class TextFormatTest(TextFormatBase):
+
+  def testPrintExotic(self, message_module):
+    message = message_module.TestAllTypes()
+    message.repeated_int64.append(-9223372036854775808)
+    message.repeated_uint64.append(18446744073709551615)
+    message.repeated_double.append(123.456)
+    message.repeated_double.append(1.23e22)
+    message.repeated_double.append(1.23e-18)
+    message.repeated_string.append('\000\001\a\b\f\n\r\t\v\\\'"')
+    message.repeated_string.append(u'\u00fc\ua71f')
+    self.CompareToGoldenText(
+        self.RemoveRedundantZeros(text_format.MessageToString(message)),
+        'repeated_int64: -9223372036854775808\n'
+        'repeated_uint64: 18446744073709551615\n'
+        'repeated_double: 123.456\n'
+        'repeated_double: 1.23e+22\n'
+        'repeated_double: 1.23e-18\n'
+        'repeated_string:'
+        ' "\\000\\001\\007\\010\\014\\n\\r\\t\\013\\\\\\\'\\""\n'
+        'repeated_string: "\\303\\274\\352\\234\\237"\n')
+
+  def testPrintExoticUnicodeSubclass(self, message_module):
+    class UnicodeSub(six.text_type):
+      pass
+    message = message_module.TestAllTypes()
+    message.repeated_string.append(UnicodeSub(u'\u00fc\ua71f'))
+    self.CompareToGoldenText(
+        text_format.MessageToString(message),
+        'repeated_string: "\\303\\274\\352\\234\\237"\n')
+
+  def testPrintNestedMessageAsOneLine(self, message_module):
+    message = message_module.TestAllTypes()
+    msg = message.repeated_nested_message.add()
+    msg.bb = 42
+    self.CompareToGoldenText(
+        text_format.MessageToString(message, as_one_line=True),
+        'repeated_nested_message { bb: 42 }')
+
+  def testPrintRepeatedFieldsAsOneLine(self, message_module):
+    message = message_module.TestAllTypes()
+    message.repeated_int32.append(1)
+    message.repeated_int32.append(1)
+    message.repeated_int32.append(3)
+    message.repeated_string.append('Google')
+    message.repeated_string.append('Zurich')
+    self.CompareToGoldenText(
+        text_format.MessageToString(message, as_one_line=True),
+        'repeated_int32: 1 repeated_int32: 1 repeated_int32: 3 '
+        'repeated_string: "Google" repeated_string: "Zurich"')
+
+  def testPrintNestedNewLineInStringAsOneLine(self, message_module):
+    message = message_module.TestAllTypes()
+    message.optional_string = 'a\nnew\nline'
+    self.CompareToGoldenText(
+        text_format.MessageToString(message, as_one_line=True),
+        'optional_string: "a\\nnew\\nline"')
+
+  def testPrintExoticAsOneLine(self, message_module):
+    message = message_module.TestAllTypes()
+    message.repeated_int64.append(-9223372036854775808)
+    message.repeated_uint64.append(18446744073709551615)
+    message.repeated_double.append(123.456)
+    message.repeated_double.append(1.23e22)
+    message.repeated_double.append(1.23e-18)
+    message.repeated_string.append('\000\001\a\b\f\n\r\t\v\\\'"')
+    message.repeated_string.append(u'\u00fc\ua71f')
+    self.CompareToGoldenText(
+        self.RemoveRedundantZeros(
+            text_format.MessageToString(message, as_one_line=True)),
+        'repeated_int64: -9223372036854775808'
+        ' repeated_uint64: 18446744073709551615'
+        ' repeated_double: 123.456'
+        ' repeated_double: 1.23e+22'
+        ' repeated_double: 1.23e-18'
+        ' repeated_string: '
+        '"\\000\\001\\007\\010\\014\\n\\r\\t\\013\\\\\\\'\\""'
+        ' repeated_string: "\\303\\274\\352\\234\\237"')
+
+  def testRoundTripExoticAsOneLine(self, message_module):
+    message = message_module.TestAllTypes()
+    message.repeated_int64.append(-9223372036854775808)
+    message.repeated_uint64.append(18446744073709551615)
+    message.repeated_double.append(123.456)
+    message.repeated_double.append(1.23e22)
+    message.repeated_double.append(1.23e-18)
+    message.repeated_string.append('\000\001\a\b\f\n\r\t\v\\\'"')
+    message.repeated_string.append(u'\u00fc\ua71f')
+
+    # Test as_utf8 = False.
+    wire_text = text_format.MessageToString(
+        message, as_one_line=True, as_utf8=False)
+    parsed_message = message_module.TestAllTypes()
+    r = text_format.Parse(wire_text, parsed_message)
+    self.assertIs(r, parsed_message)
+    self.assertEqual(message, parsed_message)
+
+    # Test as_utf8 = True.
+    wire_text = text_format.MessageToString(
+        message, as_one_line=True, as_utf8=True)
+    parsed_message = message_module.TestAllTypes()
+    r = text_format.Parse(wire_text, parsed_message)
+    self.assertIs(r, parsed_message)
+    self.assertEqual(message, parsed_message,
+                      '\n%s != %s' % (message, parsed_message))
+
+  def testPrintRawUtf8String(self, message_module):
+    message = message_module.TestAllTypes()
+    message.repeated_string.append(u'\u00fc\ua71f')
+    text = text_format.MessageToString(message, as_utf8=True)
+    self.CompareToGoldenText(text, 'repeated_string: "\303\274\352\234\237"\n')
+    parsed_message = message_module.TestAllTypes()
+    text_format.Parse(text, parsed_message)
+    self.assertEqual(message, parsed_message,
+                      '\n%s != %s' % (message, parsed_message))
+
+  def testPrintFloatFormat(self, message_module):
+    # Check that float_format argument is passed to sub-message formatting.
+    message = message_module.NestedTestAllTypes()
+    # We use 1.25 as it is a round number in binary.  The proto 32-bit float
+    # will not gain additional imprecise digits as a 64-bit Python float and
+    # show up in its str.  32-bit 1.2 is noisy when extended to 64-bit:
+    #  >>> struct.unpack('f', struct.pack('f', 1.2))[0]
+    #  1.2000000476837158
+    #  >>> struct.unpack('f', struct.pack('f', 1.25))[0]
+    #  1.25
+    message.payload.optional_float = 1.25
+    # Check rounding at 15 significant digits
+    message.payload.optional_double = -.000003456789012345678
+    # Check no decimal point.
+    message.payload.repeated_float.append(-5642)
+    # Check no trailing zeros.
+    message.payload.repeated_double.append(.000078900)
+    formatted_fields = ['optional_float: 1.25',
+                        'optional_double: -3.45678901234568e-6',
+                        'repeated_float: -5642',
+                        'repeated_double: 7.89e-5']
+    text_message = text_format.MessageToString(message, float_format='.15g')
+    self.CompareToGoldenText(
+        self.RemoveRedundantZeros(text_message),
+        'payload {{\n  {0}\n  {1}\n  {2}\n  {3}\n}}\n'.format(*formatted_fields))
+    # as_one_line=True is a separate code branch where float_format is passed.
+    text_message = text_format.MessageToString(message, as_one_line=True,
+                                               float_format='.15g')
+    self.CompareToGoldenText(
+        self.RemoveRedundantZeros(text_message),
+        'payload {{ {0} {1} {2} {3} }}'.format(*formatted_fields))
+
+  def testMessageToString(self, message_module):
+    message = message_module.ForeignMessage()
+    message.c = 123
+    self.assertEqual('c: 123\n', str(message))
+
+  def testParseAllFields(self, message_module):
+    message = message_module.TestAllTypes()
+    test_util.SetAllFields(message)
+    ascii_text = text_format.MessageToString(message)
+
+    parsed_message = message_module.TestAllTypes()
+    text_format.Parse(ascii_text, parsed_message)
+    self.assertEqual(message, parsed_message)
+    if message_module is unittest_pb2:
+      test_util.ExpectAllFieldsSet(self, message)
+
+  def testParseExotic(self, message_module):
+    message = message_module.TestAllTypes()
+    text = ('repeated_int64: -9223372036854775808\n'
+            'repeated_uint64: 18446744073709551615\n'
+            'repeated_double: 123.456\n'
+            'repeated_double: 1.23e+22\n'
+            'repeated_double: 1.23e-18\n'
+            'repeated_string: \n'
+            '"\\000\\001\\007\\010\\014\\n\\r\\t\\013\\\\\\\'\\""\n'
+            'repeated_string: "foo" \'corge\' "grault"\n'
+            'repeated_string: "\\303\\274\\352\\234\\237"\n'
+            'repeated_string: "\\xc3\\xbc"\n'
+            'repeated_string: "\xc3\xbc"\n')
+    text_format.Parse(text, message)
+
+    self.assertEqual(-9223372036854775808, message.repeated_int64[0])
+    self.assertEqual(18446744073709551615, message.repeated_uint64[0])
+    self.assertEqual(123.456, message.repeated_double[0])
+    self.assertEqual(1.23e22, message.repeated_double[1])
+    self.assertEqual(1.23e-18, message.repeated_double[2])
+    self.assertEqual(
+        '\000\001\a\b\f\n\r\t\v\\\'"', message.repeated_string[0])
+    self.assertEqual('foocorgegrault', message.repeated_string[1])
+    self.assertEqual(u'\u00fc\ua71f', message.repeated_string[2])
+    self.assertEqual(u'\u00fc', message.repeated_string[3])
+
+  def testParseTrailingCommas(self, message_module):
+    message = message_module.TestAllTypes()
+    text = ('repeated_int64: 100;\n'
+            'repeated_int64: 200;\n'
+            'repeated_int64: 300,\n'
+            'repeated_string: "one",\n'
+            'repeated_string: "two";\n')
+    text_format.Parse(text, message)
+
+    self.assertEqual(100, message.repeated_int64[0])
+    self.assertEqual(200, message.repeated_int64[1])
+    self.assertEqual(300, message.repeated_int64[2])
+    self.assertEqual(u'one', message.repeated_string[0])
+    self.assertEqual(u'two', message.repeated_string[1])
+
+  def testParseRepeatedScalarShortFormat(self, message_module):
+    message = message_module.TestAllTypes()
+    text = ('repeated_int64: [100, 200];\n'
+            'repeated_int64: 300,\n'
+            'repeated_string: ["one", "two"];\n')
+    text_format.Parse(text, message)
+
+    self.assertEqual(100, message.repeated_int64[0])
+    self.assertEqual(200, message.repeated_int64[1])
+    self.assertEqual(300, message.repeated_int64[2])
+    self.assertEqual(u'one', message.repeated_string[0])
+    self.assertEqual(u'two', message.repeated_string[1])
+
+  def testParseEmptyText(self, message_module):
+    message = message_module.TestAllTypes()
+    text = ''
+    text_format.Parse(text, message)
+    self.assertEqual(message_module.TestAllTypes(), message)
+
+  def testParseInvalidUtf8(self, message_module):
+    message = message_module.TestAllTypes()
+    text = 'repeated_string: "\\xc3\\xc3"'
+    self.assertRaises(text_format.ParseError, text_format.Parse, text, message)
+
+  def testParseSingleWord(self, message_module):
+    message = message_module.TestAllTypes()
+    text = 'foo'
+    six.assertRaisesRegex(self,
+        text_format.ParseError,
+        (r'1:1 : Message type "\w+.TestAllTypes" has no field named '
+         r'"foo".'),
+        text_format.Parse, text, message)
+
+  def testParseUnknownField(self, message_module):
+    message = message_module.TestAllTypes()
+    text = 'unknown_field: 8\n'
+    six.assertRaisesRegex(self,
+        text_format.ParseError,
+        (r'1:1 : Message type "\w+.TestAllTypes" has no field named '
+         r'"unknown_field".'),
+        text_format.Parse, text, message)
+
+  def testParseBadEnumValue(self, message_module):
+    message = message_module.TestAllTypes()
+    text = 'optional_nested_enum: BARR'
+    six.assertRaisesRegex(self,
+        text_format.ParseError,
+        (r'1:23 : Enum type "\w+.TestAllTypes.NestedEnum" '
+         r'has no value named BARR.'),
+        text_format.Parse, text, message)
+
+    message = message_module.TestAllTypes()
+    text = 'optional_nested_enum: 100'
+    six.assertRaisesRegex(self,
+        text_format.ParseError,
+        (r'1:23 : Enum type "\w+.TestAllTypes.NestedEnum" '
+         r'has no value with number 100.'),
+        text_format.Parse, text, message)
+
+  def testParseBadIntValue(self, message_module):
+    message = message_module.TestAllTypes()
+    text = 'optional_int32: bork'
+    six.assertRaisesRegex(self,
+        text_format.ParseError,
+        ('1:17 : Couldn\'t parse integer: bork'),
+        text_format.Parse, text, message)
+
+  def testParseStringFieldUnescape(self, message_module):
+    message = message_module.TestAllTypes()
+    text = r'''repeated_string: "\xf\x62"
+               repeated_string: "\\xf\\x62"
+               repeated_string: "\\\xf\\\x62"
+               repeated_string: "\\\\xf\\\\x62"
+               repeated_string: "\\\\\xf\\\\\x62"
+               repeated_string: "\x5cx20"'''
+    text_format.Parse(text, message)
+
+    SLASH = '\\'
+    self.assertEqual('\x0fb', message.repeated_string[0])
+    self.assertEqual(SLASH + 'xf' + SLASH + 'x62', message.repeated_string[1])
+    self.assertEqual(SLASH + '\x0f' + SLASH + 'b', message.repeated_string[2])
+    self.assertEqual(SLASH + SLASH + 'xf' + SLASH + SLASH + 'x62',
+                     message.repeated_string[3])
+    self.assertEqual(SLASH + SLASH + '\x0f' + SLASH + SLASH + 'b',
+                     message.repeated_string[4])
+    self.assertEqual(SLASH + 'x20', message.repeated_string[5])
+
+  def testMergeDuplicateScalars(self, message_module):
+    message = message_module.TestAllTypes()
+    text = ('optional_int32: 42 '
+            'optional_int32: 67')
+    r = text_format.Merge(text, message)
+    self.assertIs(r, message)
+    self.assertEqual(67, message.optional_int32)
+
+  def testMergeDuplicateNestedMessageScalars(self, message_module):
+    message = message_module.TestAllTypes()
+    text = ('optional_nested_message { bb: 1 } '
+            'optional_nested_message { bb: 2 }')
+    r = text_format.Merge(text, message)
+    self.assertTrue(r is message)
+    self.assertEqual(2, message.optional_nested_message.bb)
+
+  def testParseOneof(self, message_module):
+    m = message_module.TestAllTypes()
+    m.oneof_uint32 = 11
+    m2 = message_module.TestAllTypes()
+    text_format.Parse(text_format.MessageToString(m), m2)
+    self.assertEqual('oneof_uint32', m2.WhichOneof('oneof_field'))
+
+
+# These are tests that aren't fundamentally specific to proto2, but are at
+# the moment because of differences between the proto2 and proto3 test schemas.
+# Ideally the schemas would be made more similar so these tests could pass.
+class OnlyWorksWithProto2RightNowTests(TextFormatBase):
+
+  def testPrintAllFieldsPointy(self):
+    message = unittest_pb2.TestAllTypes()
+    test_util.SetAllFields(message)
+    self.CompareToGoldenFile(
+        self.RemoveRedundantZeros(
+            text_format.MessageToString(message, pointy_brackets=True)),
+        'text_format_unittest_data_pointy_oneof.txt')
+
+  def testParseGolden(self):
+    golden_text = '\n'.join(self.ReadGolden('text_format_unittest_data.txt'))
+    parsed_message = unittest_pb2.TestAllTypes()
+    r = text_format.Parse(golden_text, parsed_message)
+    self.assertIs(r, parsed_message)
+
+    message = unittest_pb2.TestAllTypes()
+    test_util.SetAllFields(message)
+    self.assertEqual(message, parsed_message)
+
+  def testPrintAllFields(self):
+    message = unittest_pb2.TestAllTypes()
+    test_util.SetAllFields(message)
+    self.CompareToGoldenFile(
+        self.RemoveRedundantZeros(text_format.MessageToString(message)),
+        'text_format_unittest_data_oneof_implemented.txt')
+
+  def testPrintAllFieldsPointy(self):
+    message = unittest_pb2.TestAllTypes()
+    test_util.SetAllFields(message)
+    self.CompareToGoldenFile(
+        self.RemoveRedundantZeros(
+            text_format.MessageToString(message, pointy_brackets=True)),
+        'text_format_unittest_data_pointy_oneof.txt')
+
+  def testPrintInIndexOrder(self):
+    message = unittest_pb2.TestFieldOrderings()
+    message.my_string = '115'
+    message.my_int = 101
+    message.my_float = 111
+    message.optional_nested_message.oo = 0
+    message.optional_nested_message.bb = 1
+    self.CompareToGoldenText(
+        self.RemoveRedundantZeros(text_format.MessageToString(
+            message, use_index_order=True)),
+        'my_string: \"115\"\nmy_int: 101\nmy_float: 111\n'
+        'optional_nested_message {\n  oo: 0\n  bb: 1\n}\n')
+    self.CompareToGoldenText(
+        self.RemoveRedundantZeros(text_format.MessageToString(
+            message)),
+        'my_int: 101\nmy_string: \"115\"\nmy_float: 111\n'
+        'optional_nested_message {\n  bb: 1\n  oo: 0\n}\n')
+
+  def testMergeLinesGolden(self):
+    opened = self.ReadGolden('text_format_unittest_data.txt')
+    parsed_message = unittest_pb2.TestAllTypes()
+    r = text_format.MergeLines(opened, parsed_message)
+    self.assertIs(r, parsed_message)
+
+    message = unittest_pb2.TestAllTypes()
+    test_util.SetAllFields(message)
+    self.assertEqual(message, parsed_message)
+
+  def testParseLinesGolden(self):
+    opened = self.ReadGolden('text_format_unittest_data.txt')
+    parsed_message = unittest_pb2.TestAllTypes()
+    r = text_format.ParseLines(opened, parsed_message)
+    self.assertIs(r, parsed_message)
+
+    message = unittest_pb2.TestAllTypes()
+    test_util.SetAllFields(message)
+    self.assertEqual(message, parsed_message)
+
+  def testPrintMap(self):
+    message = map_unittest_pb2.TestMap()
+
+    message.map_int32_int32[-123] = -456
+    message.map_int64_int64[-2**33] = -2**34
+    message.map_uint32_uint32[123] = 456
+    message.map_uint64_uint64[2**33] = 2**34
+    message.map_string_string["abc"] = "123"
+    message.map_int32_foreign_message[111].c = 5
+
+    # Maps are serialized to text format using their underlying repeated
+    # representation.
+    self.CompareToGoldenText(
+        text_format.MessageToString(message),
+        'map_int32_int32 {\n'
+        '  key: -123\n'
+        '  value: -456\n'
+        '}\n'
+        'map_int64_int64 {\n'
+        '  key: -8589934592\n'
+        '  value: -17179869184\n'
+        '}\n'
+        'map_uint32_uint32 {\n'
+        '  key: 123\n'
+        '  value: 456\n'
+        '}\n'
+        'map_uint64_uint64 {\n'
+        '  key: 8589934592\n'
+        '  value: 17179869184\n'
+        '}\n'
+        'map_string_string {\n'
+        '  key: "abc"\n'
+        '  value: "123"\n'
+        '}\n'
+        'map_int32_foreign_message {\n'
+        '  key: 111\n'
+        '  value {\n'
+        '    c: 5\n'
+        '  }\n'
+        '}\n')
+
+  def testMapOrderEnforcement(self):
+    message = map_unittest_pb2.TestMap()
+    for letter in string.ascii_uppercase[13:26]:
+      message.map_string_string[letter] = 'dummy'
+    for letter in reversed(string.ascii_uppercase[0:13]):
+      message.map_string_string[letter] = 'dummy'
+    golden = ''.join((
+        'map_string_string {\n  key: "%c"\n  value: "dummy"\n}\n' % (letter,)
+        for letter in string.ascii_uppercase))
+    self.CompareToGoldenText(text_format.MessageToString(message), golden)
+
+  def testMapOrderSemantics(self):
+    golden_lines = self.ReadGolden('map_test_data.txt')
+    # The C++ implementation emits defaulted-value fields, while the Python
+    # implementation does not.  Adjusting for this is awkward, but it is
+    # valuable to test against a common golden file.
+    line_blacklist = ('  key: 0\n',
+                      '  value: 0\n',
+                      '  key: false\n',
+                      '  value: false\n')
+    golden_lines = [line for line in golden_lines if line not in line_blacklist]
+
+    message = map_unittest_pb2.TestMap()
+    text_format.ParseLines(golden_lines, message)
+    candidate = text_format.MessageToString(message)
+    # The Python implementation emits "1.0" for the double value that the C++
+    # implementation emits as "1".
+    candidate = candidate.replace('1.0', '1', 2)
+    self.assertMultiLineEqual(candidate, ''.join(golden_lines))
+
+
+# Tests of proto2-only features (MessageSet, extensions, etc.).
+class Proto2Tests(TextFormatBase):
+
+  def testPrintMessageSet(self):
+    message = unittest_mset_pb2.TestMessageSetContainer()
+    ext1 = unittest_mset_pb2.TestMessageSetExtension1.message_set_extension
+    ext2 = unittest_mset_pb2.TestMessageSetExtension2.message_set_extension
+    message.message_set.Extensions[ext1].i = 23
+    message.message_set.Extensions[ext2].str = 'foo'
+    self.CompareToGoldenText(
+        text_format.MessageToString(message),
+        'message_set {\n'
+        '  [protobuf_unittest.TestMessageSetExtension1] {\n'
+        '    i: 23\n'
+        '  }\n'
+        '  [protobuf_unittest.TestMessageSetExtension2] {\n'
+        '    str: \"foo\"\n'
+        '  }\n'
+        '}\n')
+
+    message = message_set_extensions_pb2.TestMessageSet()
+    ext = message_set_extensions_pb2.message_set_extension3
+    message.Extensions[ext].text = 'bar'
+    self.CompareToGoldenText(
+        text_format.MessageToString(message),
+        '[google.protobuf.internal.TestMessageSetExtension3] {\n'
+        '  text: \"bar\"\n'
+        '}\n')
+
+  def testPrintMessageSetAsOneLine(self):
+    message = unittest_mset_pb2.TestMessageSetContainer()
+    ext1 = unittest_mset_pb2.TestMessageSetExtension1.message_set_extension
+    ext2 = unittest_mset_pb2.TestMessageSetExtension2.message_set_extension
+    message.message_set.Extensions[ext1].i = 23
+    message.message_set.Extensions[ext2].str = 'foo'
+    self.CompareToGoldenText(
+        text_format.MessageToString(message, as_one_line=True),
+        'message_set {'
+        ' [protobuf_unittest.TestMessageSetExtension1] {'
+        ' i: 23'
+        ' }'
+        ' [protobuf_unittest.TestMessageSetExtension2] {'
+        ' str: \"foo\"'
+        ' }'
+        ' }')
+
+  def testParseMessageSet(self):
+    message = unittest_pb2.TestAllTypes()
+    text = ('repeated_uint64: 1\n'
+            'repeated_uint64: 2\n')
+    text_format.Parse(text, message)
+    self.assertEqual(1, message.repeated_uint64[0])
+    self.assertEqual(2, message.repeated_uint64[1])
+
+    message = unittest_mset_pb2.TestMessageSetContainer()
+    text = ('message_set {\n'
+            '  [protobuf_unittest.TestMessageSetExtension1] {\n'
+            '    i: 23\n'
+            '  }\n'
+            '  [protobuf_unittest.TestMessageSetExtension2] {\n'
+            '    str: \"foo\"\n'
+            '  }\n'
+            '}\n')
+    text_format.Parse(text, message)
+    ext1 = unittest_mset_pb2.TestMessageSetExtension1.message_set_extension
+    ext2 = unittest_mset_pb2.TestMessageSetExtension2.message_set_extension
+    self.assertEqual(23, message.message_set.Extensions[ext1].i)
+    self.assertEqual('foo', message.message_set.Extensions[ext2].str)
+
+  def testPrintAllExtensions(self):
+    message = unittest_pb2.TestAllExtensions()
+    test_util.SetAllExtensions(message)
+    self.CompareToGoldenFile(
+        self.RemoveRedundantZeros(text_format.MessageToString(message)),
+        'text_format_unittest_extensions_data.txt')
+
+  def testPrintAllExtensionsPointy(self):
+    message = unittest_pb2.TestAllExtensions()
+    test_util.SetAllExtensions(message)
+    self.CompareToGoldenFile(
+        self.RemoveRedundantZeros(text_format.MessageToString(
+            message, pointy_brackets=True)),
+        'text_format_unittest_extensions_data_pointy.txt')
+
+  def testParseGoldenExtensions(self):
+    golden_text = '\n'.join(self.ReadGolden(
+        'text_format_unittest_extensions_data.txt'))
+    parsed_message = unittest_pb2.TestAllExtensions()
+    text_format.Parse(golden_text, parsed_message)
+
+    message = unittest_pb2.TestAllExtensions()
+    test_util.SetAllExtensions(message)
+    self.assertEqual(message, parsed_message)
+
+  def testParseAllExtensions(self):
+    message = unittest_pb2.TestAllExtensions()
+    test_util.SetAllExtensions(message)
+    ascii_text = text_format.MessageToString(message)
+
+    parsed_message = unittest_pb2.TestAllExtensions()
+    text_format.Parse(ascii_text, parsed_message)
+    self.assertEqual(message, parsed_message)
+
+  def testParseAllowedUnknownExtension(self):
+    # Skip over unknown extension correctly.
+    message = unittest_mset_pb2.TestMessageSetContainer()
+    text = ('message_set {\n'
+            '  [unknown_extension] {\n'
+            '    i: 23\n'
+            '    [nested_unknown_ext]: {\n'
+            '      i: 23\n'
+            '      test: "test_string"\n'
+            '      floaty_float: -0.315\n'
+            '      num: -inf\n'
+            '      multiline_str: "abc"\n'
+            '          "def"\n'
+            '          "xyz."\n'
+            '      [nested_unknown_ext]: <\n'
+            '        i: 23\n'
+            '        i: 24\n'
+            '        pointfloat: .3\n'
+            '        test: "test_string"\n'
+            '        floaty_float: -0.315\n'
+            '        num: -inf\n'
+            '        long_string: "test" "test2" \n'
+            '      >\n'
+            '    }\n'
+            '  }\n'
+            '  [unknown_extension]: 5\n'
+            '}\n')
+    text_format.Parse(text, message, allow_unknown_extension=True)
+    golden = 'message_set {\n}\n'
+    self.CompareToGoldenText(text_format.MessageToString(message), golden)
+
+    # Catch parse errors in unknown extension.
+    message = unittest_mset_pb2.TestMessageSetContainer()
+    malformed = ('message_set {\n'
+                 '  [unknown_extension] {\n'
+                 '    i:\n'  # Missing value.
+                 '  }\n'
+                 '}\n')
+    six.assertRaisesRegex(self,
+                          text_format.ParseError,
+                          'Invalid field value: }',
+                          text_format.Parse, malformed, message,
+                          allow_unknown_extension=True)
+
+    message = unittest_mset_pb2.TestMessageSetContainer()
+    malformed = ('message_set {\n'
+                 '  [unknown_extension] {\n'
+                 '    str: "malformed string\n'  # Missing closing quote.
+                 '  }\n'
+                 '}\n')
+    six.assertRaisesRegex(self,
+                          text_format.ParseError,
+                          'Invalid field value: "',
+                          text_format.Parse, malformed, message,
+                          allow_unknown_extension=True)
+
+    message = unittest_mset_pb2.TestMessageSetContainer()
+    malformed = ('message_set {\n'
+                 '  [unknown_extension] {\n'
+                 '    str: "malformed\n multiline\n string\n'
+                 '  }\n'
+                 '}\n')
+    six.assertRaisesRegex(self,
+                          text_format.ParseError,
+                          'Invalid field value: "',
+                          text_format.Parse, malformed, message,
+                          allow_unknown_extension=True)
+
+    message = unittest_mset_pb2.TestMessageSetContainer()
+    malformed = ('message_set {\n'
+                 '  [malformed_extension] <\n'
+                 '    i: -5\n'
+                 '  \n'  # Missing '>' here.
+                 '}\n')
+    six.assertRaisesRegex(self,
+                          text_format.ParseError,
+                          '5:1 : Expected ">".',
+                          text_format.Parse, malformed, message,
+                          allow_unknown_extension=True)
+
+    # Don't allow unknown fields with allow_unknown_extension=True.
+    message = unittest_mset_pb2.TestMessageSetContainer()
+    malformed = ('message_set {\n'
+                 '  unknown_field: true\n'
+                 '  \n'  # Missing '>' here.
+                 '}\n')
+    six.assertRaisesRegex(self,
+                          text_format.ParseError,
+                          ('2:3 : Message type '
+                           '"proto2_wireformat_unittest.TestMessageSet" has no'
+                           ' field named "unknown_field".'),
+                          text_format.Parse, malformed, message,
+                          allow_unknown_extension=True)
+
+    # Parse known extension correcty.
+    message = unittest_mset_pb2.TestMessageSetContainer()
+    text = ('message_set {\n'
+            '  [protobuf_unittest.TestMessageSetExtension1] {\n'
+            '    i: 23\n'
+            '  }\n'
+            '  [protobuf_unittest.TestMessageSetExtension2] {\n'
+            '    str: \"foo\"\n'
+            '  }\n'
+            '}\n')
+    text_format.Parse(text, message, allow_unknown_extension=True)
+    ext1 = unittest_mset_pb2.TestMessageSetExtension1.message_set_extension
+    ext2 = unittest_mset_pb2.TestMessageSetExtension2.message_set_extension
+    self.assertEqual(23, message.message_set.Extensions[ext1].i)
+    self.assertEqual('foo', message.message_set.Extensions[ext2].str)
+
+  def testParseBadExtension(self):
+    message = unittest_pb2.TestAllExtensions()
+    text = '[unknown_extension]: 8\n'
+    six.assertRaisesRegex(self,
+        text_format.ParseError,
+        '1:2 : Extension "unknown_extension" not registered.',
+        text_format.Parse, text, message)
+    message = unittest_pb2.TestAllTypes()
+    six.assertRaisesRegex(self,
+        text_format.ParseError,
+        ('1:2 : Message type "protobuf_unittest.TestAllTypes" does not have '
+         'extensions.'),
+        text_format.Parse, text, message)
+
+  def testMergeDuplicateExtensionScalars(self):
+    message = unittest_pb2.TestAllExtensions()
+    text = ('[protobuf_unittest.optional_int32_extension]: 42 '
+            '[protobuf_unittest.optional_int32_extension]: 67')
+    text_format.Merge(text, message)
+    self.assertEqual(
+        67,
+        message.Extensions[unittest_pb2.optional_int32_extension])
+
+  def testParseDuplicateExtensionScalars(self):
+    message = unittest_pb2.TestAllExtensions()
+    text = ('[protobuf_unittest.optional_int32_extension]: 42 '
+            '[protobuf_unittest.optional_int32_extension]: 67')
+    six.assertRaisesRegex(self,
+        text_format.ParseError,
+        ('1:96 : Message type "protobuf_unittest.TestAllExtensions" '
+         'should not have multiple '
+         '"protobuf_unittest.optional_int32_extension" extensions.'),
+        text_format.Parse, text, message)
+
+  def testParseDuplicateNestedMessageScalars(self):
+    message = unittest_pb2.TestAllTypes()
+    text = ('optional_nested_message { bb: 1 } '
+            'optional_nested_message { bb: 2 }')
+    six.assertRaisesRegex(self,
+        text_format.ParseError,
+        ('1:65 : Message type "protobuf_unittest.TestAllTypes.NestedMessage" '
+         'should not have multiple "bb" fields.'),
+        text_format.Parse, text, message)
+
+  def testParseDuplicateScalars(self):
+    message = unittest_pb2.TestAllTypes()
+    text = ('optional_int32: 42 '
+            'optional_int32: 67')
+    six.assertRaisesRegex(self,
+        text_format.ParseError,
+        ('1:36 : Message type "protobuf_unittest.TestAllTypes" should not '
+         'have multiple "optional_int32" fields.'),
+        text_format.Parse, text, message)
+
+  def testParseGroupNotClosed(self):
+    message = unittest_pb2.TestAllTypes()
+    text = 'RepeatedGroup: <'
+    six.assertRaisesRegex(self,
+        text_format.ParseError, '1:16 : Expected ">".',
+        text_format.Parse, text, message)
+    text = 'RepeatedGroup: {'
+    six.assertRaisesRegex(self,
+        text_format.ParseError, '1:16 : Expected "}".',
+        text_format.Parse, text, message)
+
+  def testParseEmptyGroup(self):
+    message = unittest_pb2.TestAllTypes()
+    text = 'OptionalGroup: {}'
+    text_format.Parse(text, message)
+    self.assertTrue(message.HasField('optionalgroup'))
+
+    message.Clear()
+
+    message = unittest_pb2.TestAllTypes()
+    text = 'OptionalGroup: <>'
+    text_format.Parse(text, message)
+    self.assertTrue(message.HasField('optionalgroup'))
+
+  # Maps aren't really proto2-only, but our test schema only has maps for
+  # proto2.
+  def testParseMap(self):
+    text = ('map_int32_int32 {\n'
+            '  key: -123\n'
+            '  value: -456\n'
+            '}\n'
+            'map_int64_int64 {\n'
+            '  key: -8589934592\n'
+            '  value: -17179869184\n'
+            '}\n'
+            'map_uint32_uint32 {\n'
+            '  key: 123\n'
+            '  value: 456\n'
+            '}\n'
+            'map_uint64_uint64 {\n'
+            '  key: 8589934592\n'
+            '  value: 17179869184\n'
+            '}\n'
+            'map_string_string {\n'
+            '  key: "abc"\n'
+            '  value: "123"\n'
+            '}\n'
+            'map_int32_foreign_message {\n'
+            '  key: 111\n'
+            '  value {\n'
+            '    c: 5\n'
+            '  }\n'
+            '}\n')
+    message = map_unittest_pb2.TestMap()
+    text_format.Parse(text, message)
+
+    self.assertEqual(-456, message.map_int32_int32[-123])
+    self.assertEqual(-2**34, message.map_int64_int64[-2**33])
+    self.assertEqual(456, message.map_uint32_uint32[123])
+    self.assertEqual(2**34, message.map_uint64_uint64[2**33])
+    self.assertEqual("123", message.map_string_string["abc"])
+    self.assertEqual(5, message.map_int32_foreign_message[111].c)
+
+
+class TokenizerTest(unittest.TestCase):
+
+  def testSimpleTokenCases(self):
+    text = ('identifier1:"string1"\n     \n\n'
+            'identifier2 : \n \n123  \n  identifier3 :\'string\'\n'
+            'identifiER_4 : 1.1e+2 ID5:-0.23 ID6:\'aaaa\\\'bbbb\'\n'
+            'ID7 : "aa\\"bb"\n\n\n\n ID8: {A:inf B:-inf C:true D:false}\n'
+            'ID9: 22 ID10: -111111111111111111 ID11: -22\n'
+            'ID12: 2222222222222222222 ID13: 1.23456f ID14: 1.2e+2f '
+            'false_bool:  0 true_BOOL:t \n true_bool1:  1 false_BOOL1:f ')
+    tokenizer = text_format._Tokenizer(text.splitlines())
+    methods = [(tokenizer.ConsumeIdentifier, 'identifier1'),
+               ':',
+               (tokenizer.ConsumeString, 'string1'),
+               (tokenizer.ConsumeIdentifier, 'identifier2'),
+               ':',
+               (tokenizer.ConsumeInt32, 123),
+               (tokenizer.ConsumeIdentifier, 'identifier3'),
+               ':',
+               (tokenizer.ConsumeString, 'string'),
+               (tokenizer.ConsumeIdentifier, 'identifiER_4'),
+               ':',
+               (tokenizer.ConsumeFloat, 1.1e+2),
+               (tokenizer.ConsumeIdentifier, 'ID5'),
+               ':',
+               (tokenizer.ConsumeFloat, -0.23),
+               (tokenizer.ConsumeIdentifier, 'ID6'),
+               ':',
+               (tokenizer.ConsumeString, 'aaaa\'bbbb'),
+               (tokenizer.ConsumeIdentifier, 'ID7'),
+               ':',
+               (tokenizer.ConsumeString, 'aa\"bb'),
+               (tokenizer.ConsumeIdentifier, 'ID8'),
+               ':',
+               '{',
+               (tokenizer.ConsumeIdentifier, 'A'),
+               ':',
+               (tokenizer.ConsumeFloat, float('inf')),
+               (tokenizer.ConsumeIdentifier, 'B'),
+               ':',
+               (tokenizer.ConsumeFloat, -float('inf')),
+               (tokenizer.ConsumeIdentifier, 'C'),
+               ':',
+               (tokenizer.ConsumeBool, True),
+               (tokenizer.ConsumeIdentifier, 'D'),
+               ':',
+               (tokenizer.ConsumeBool, False),
+               '}',
+               (tokenizer.ConsumeIdentifier, 'ID9'),
+               ':',
+               (tokenizer.ConsumeUint32, 22),
+               (tokenizer.ConsumeIdentifier, 'ID10'),
+               ':',
+               (tokenizer.ConsumeInt64, -111111111111111111),
+               (tokenizer.ConsumeIdentifier, 'ID11'),
+               ':',
+               (tokenizer.ConsumeInt32, -22),
+               (tokenizer.ConsumeIdentifier, 'ID12'),
+               ':',
+               (tokenizer.ConsumeUint64, 2222222222222222222),
+               (tokenizer.ConsumeIdentifier, 'ID13'),
+               ':',
+               (tokenizer.ConsumeFloat, 1.23456),
+               (tokenizer.ConsumeIdentifier, 'ID14'),
+               ':',
+               (tokenizer.ConsumeFloat, 1.2e+2),
+               (tokenizer.ConsumeIdentifier, 'false_bool'),
+               ':',
+               (tokenizer.ConsumeBool, False),
+               (tokenizer.ConsumeIdentifier, 'true_BOOL'),
+               ':',
+               (tokenizer.ConsumeBool, True),
+               (tokenizer.ConsumeIdentifier, 'true_bool1'),
+               ':',
+               (tokenizer.ConsumeBool, True),
+               (tokenizer.ConsumeIdentifier, 'false_BOOL1'),
+               ':',
+               (tokenizer.ConsumeBool, False)]
+
+    i = 0
+    while not tokenizer.AtEnd():
+      m = methods[i]
+      if type(m) == str:
+        token = tokenizer.token
+        self.assertEqual(token, m)
+        tokenizer.NextToken()
+      else:
+        self.assertEqual(m[1], m[0]())
+      i += 1
+
+  def testConsumeIntegers(self):
+    # This test only tests the failures in the integer parsing methods as well
+    # as the '0' special cases.
+    int64_max = (1 << 63) - 1
+    uint32_max = (1 << 32) - 1
+    text = '-1 %d %d' % (uint32_max + 1, int64_max + 1)
+    tokenizer = text_format._Tokenizer(text.splitlines())
+    self.assertRaises(text_format.ParseError, tokenizer.ConsumeUint32)
+    self.assertRaises(text_format.ParseError, tokenizer.ConsumeUint64)
+    self.assertEqual(-1, tokenizer.ConsumeInt32())
+
+    self.assertRaises(text_format.ParseError, tokenizer.ConsumeUint32)
+    self.assertRaises(text_format.ParseError, tokenizer.ConsumeInt32)
+    self.assertEqual(uint32_max + 1, tokenizer.ConsumeInt64())
+
+    self.assertRaises(text_format.ParseError, tokenizer.ConsumeInt64)
+    self.assertEqual(int64_max + 1, tokenizer.ConsumeUint64())
+    self.assertTrue(tokenizer.AtEnd())
+
+    text = '-0 -0 0 0'
+    tokenizer = text_format._Tokenizer(text.splitlines())
+    self.assertEqual(0, tokenizer.ConsumeUint32())
+    self.assertEqual(0, tokenizer.ConsumeUint64())
+    self.assertEqual(0, tokenizer.ConsumeUint32())
+    self.assertEqual(0, tokenizer.ConsumeUint64())
+    self.assertTrue(tokenizer.AtEnd())
+
+  def testConsumeByteString(self):
+    text = '"string1\''
+    tokenizer = text_format._Tokenizer(text.splitlines())
+    self.assertRaises(text_format.ParseError, tokenizer.ConsumeByteString)
+
+    text = 'string1"'
+    tokenizer = text_format._Tokenizer(text.splitlines())
+    self.assertRaises(text_format.ParseError, tokenizer.ConsumeByteString)
+
+    text = '\n"\\xt"'
+    tokenizer = text_format._Tokenizer(text.splitlines())
+    self.assertRaises(text_format.ParseError, tokenizer.ConsumeByteString)
+
+    text = '\n"\\"'
+    tokenizer = text_format._Tokenizer(text.splitlines())
+    self.assertRaises(text_format.ParseError, tokenizer.ConsumeByteString)
+
+    text = '\n"\\x"'
+    tokenizer = text_format._Tokenizer(text.splitlines())
+    self.assertRaises(text_format.ParseError, tokenizer.ConsumeByteString)
+
+  def testConsumeBool(self):
+    text = 'not-a-bool'
+    tokenizer = text_format._Tokenizer(text.splitlines())
+    self.assertRaises(text_format.ParseError, tokenizer.ConsumeBool)
+
+
+if __name__ == '__main__':
+  unittest.main()
diff --git a/python/google/protobuf/internal/type_checkers.py b/python/google/protobuf/internal/type_checkers.py
new file mode 100755
index 0000000..f30ca6a
--- /dev/null
+++ b/python/google/protobuf/internal/type_checkers.py
@@ -0,0 +1,341 @@
+# Protocol Buffers - Google's data interchange format
+# Copyright 2008 Google Inc.  All rights reserved.
+# https://developers.google.com/protocol-buffers/
+#
+# Redistribution and use in source and binary forms, with or without
+# modification, are permitted provided that the following conditions are
+# met:
+#
+#     * Redistributions of source code must retain the above copyright
+# notice, this list of conditions and the following disclaimer.
+#     * Redistributions in binary form must reproduce the above
+# copyright notice, this list of conditions and the following disclaimer
+# in the documentation and/or other materials provided with the
+# distribution.
+#     * Neither the name of Google Inc. nor the names of its
+# contributors may be used to endorse or promote products derived from
+# this software without specific prior written permission.
+#
+# THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+# "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
+# LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
+# A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
+# OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
+# SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
+# LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
+# DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
+# THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
+# (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+# OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+
+"""Provides type checking routines.
+
+This module defines type checking utilities in the forms of dictionaries:
+
+VALUE_CHECKERS: A dictionary of field types and a value validation object.
+TYPE_TO_BYTE_SIZE_FN: A dictionary with field types and a size computing
+  function.
+TYPE_TO_SERIALIZE_METHOD: A dictionary with field types and serialization
+  function.
+FIELD_TYPE_TO_WIRE_TYPE: A dictionary with field typed and their
+  coresponding wire types.
+TYPE_TO_DESERIALIZE_METHOD: A dictionary with field types and deserialization
+  function.
+"""
+
+__author__ = 'robinson@google.com (Will Robinson)'
+
+import six
+
+if six.PY3:
+  long = int
+
+from google.protobuf.internal import api_implementation
+from google.protobuf.internal import decoder
+from google.protobuf.internal import encoder
+from google.protobuf.internal import wire_format
+from google.protobuf import descriptor
+
+_FieldDescriptor = descriptor.FieldDescriptor
+
+def SupportsOpenEnums(field_descriptor):
+  return field_descriptor.containing_type.syntax == "proto3"
+
+def GetTypeChecker(field):
+  """Returns a type checker for a message field of the specified types.
+
+  Args:
+    field: FieldDescriptor object for this field.
+
+  Returns:
+    An instance of TypeChecker which can be used to verify the types
+    of values assigned to a field of the specified type.
+  """
+  if (field.cpp_type == _FieldDescriptor.CPPTYPE_STRING and
+      field.type == _FieldDescriptor.TYPE_STRING):
+    return UnicodeValueChecker()
+  if field.cpp_type == _FieldDescriptor.CPPTYPE_ENUM:
+    if SupportsOpenEnums(field):
+      # When open enums are supported, any int32 can be assigned.
+      return _VALUE_CHECKERS[_FieldDescriptor.CPPTYPE_INT32]
+    else:
+      return EnumValueChecker(field.enum_type)
+  return _VALUE_CHECKERS[field.cpp_type]
+
+
+# None of the typecheckers below make any attempt to guard against people
+# subclassing builtin types and doing weird things.  We're not trying to
+# protect against malicious clients here, just people accidentally shooting
+# themselves in the foot in obvious ways.
+
+class TypeChecker(object):
+
+  """Type checker used to catch type errors as early as possible
+  when the client is setting scalar fields in protocol messages.
+  """
+
+  def __init__(self, *acceptable_types):
+    self._acceptable_types = acceptable_types
+
+  def CheckValue(self, proposed_value):
+    """Type check the provided value and return it.
+
+    The returned value might have been normalized to another type.
+    """
+    if not isinstance(proposed_value, self._acceptable_types):
+      message = ('%.1024r has type %s, but expected one of: %s' %
+                 (proposed_value, type(proposed_value), self._acceptable_types))
+      raise TypeError(message)
+    return proposed_value
+
+
+# IntValueChecker and its subclasses perform integer type-checks
+# and bounds-checks.
+class IntValueChecker(object):
+
+  """Checker used for integer fields.  Performs type-check and range check."""
+
+  def CheckValue(self, proposed_value):
+    if not isinstance(proposed_value, six.integer_types):
+      message = ('%.1024r has type %s, but expected one of: %s' %
+                 (proposed_value, type(proposed_value), six.integer_types))
+      raise TypeError(message)
+    if not self._MIN <= proposed_value <= self._MAX:
+      raise ValueError('Value out of range: %d' % proposed_value)
+    # We force 32-bit values to int and 64-bit values to long to make
+    # alternate implementations where the distinction is more significant
+    # (e.g. the C++ implementation) simpler.
+    proposed_value = self._TYPE(proposed_value)
+    return proposed_value
+
+  def DefaultValue(self):
+    return 0
+
+
+class EnumValueChecker(object):
+
+  """Checker used for enum fields.  Performs type-check and range check."""
+
+  def __init__(self, enum_type):
+    self._enum_type = enum_type
+
+  def CheckValue(self, proposed_value):
+    if not isinstance(proposed_value, six.integer_types):
+      message = ('%.1024r has type %s, but expected one of: %s' %
+                 (proposed_value, type(proposed_value), six.integer_types))
+      raise TypeError(message)
+    if proposed_value not in self._enum_type.values_by_number:
+      raise ValueError('Unknown enum value: %d' % proposed_value)
+    return proposed_value
+
+  def DefaultValue(self):
+    return self._enum_type.values[0].number
+
+
+class UnicodeValueChecker(object):
+
+  """Checker used for string fields.
+
+  Always returns a unicode value, even if the input is of type str.
+  """
+
+  def CheckValue(self, proposed_value):
+    if not isinstance(proposed_value, (bytes, six.text_type)):
+      message = ('%.1024r has type %s, but expected one of: %s' %
+                 (proposed_value, type(proposed_value), (bytes, six.text_type)))
+      raise TypeError(message)
+
+    # If the value is of type 'bytes' make sure that it is valid UTF-8 data.
+    if isinstance(proposed_value, bytes):
+      try:
+        proposed_value = proposed_value.decode('utf-8')
+      except UnicodeDecodeError:
+        raise ValueError('%.1024r has type bytes, but isn\'t valid UTF-8 '
+                         'encoding. Non-UTF-8 strings must be converted to '
+                         'unicode objects before being added.' %
+                         (proposed_value))
+    return proposed_value
+
+  def DefaultValue(self):
+    return u""
+
+
+class Int32ValueChecker(IntValueChecker):
+  # We're sure to use ints instead of longs here since comparison may be more
+  # efficient.
+  _MIN = -2147483648
+  _MAX = 2147483647
+  _TYPE = int
+
+
+class Uint32ValueChecker(IntValueChecker):
+  _MIN = 0
+  _MAX = (1 << 32) - 1
+  _TYPE = int
+
+
+class Int64ValueChecker(IntValueChecker):
+  _MIN = -(1 << 63)
+  _MAX = (1 << 63) - 1
+  _TYPE = long
+
+
+class Uint64ValueChecker(IntValueChecker):
+  _MIN = 0
+  _MAX = (1 << 64) - 1
+  _TYPE = long
+
+
+# Type-checkers for all scalar CPPTYPEs.
+_VALUE_CHECKERS = {
+    _FieldDescriptor.CPPTYPE_INT32: Int32ValueChecker(),
+    _FieldDescriptor.CPPTYPE_INT64: Int64ValueChecker(),
+    _FieldDescriptor.CPPTYPE_UINT32: Uint32ValueChecker(),
+    _FieldDescriptor.CPPTYPE_UINT64: Uint64ValueChecker(),
+    _FieldDescriptor.CPPTYPE_DOUBLE: TypeChecker(
+        float, int, long),
+    _FieldDescriptor.CPPTYPE_FLOAT: TypeChecker(
+        float, int, long),
+    _FieldDescriptor.CPPTYPE_BOOL: TypeChecker(bool, int),
+    _FieldDescriptor.CPPTYPE_STRING: TypeChecker(bytes),
+    }
+
+
+# Map from field type to a function F, such that F(field_num, value)
+# gives the total byte size for a value of the given type.  This
+# byte size includes tag information and any other additional space
+# associated with serializing "value".
+TYPE_TO_BYTE_SIZE_FN = {
+    _FieldDescriptor.TYPE_DOUBLE: wire_format.DoubleByteSize,
+    _FieldDescriptor.TYPE_FLOAT: wire_format.FloatByteSize,
+    _FieldDescriptor.TYPE_INT64: wire_format.Int64ByteSize,
+    _FieldDescriptor.TYPE_UINT64: wire_format.UInt64ByteSize,
+    _FieldDescriptor.TYPE_INT32: wire_format.Int32ByteSize,
+    _FieldDescriptor.TYPE_FIXED64: wire_format.Fixed64ByteSize,
+    _FieldDescriptor.TYPE_FIXED32: wire_format.Fixed32ByteSize,
+    _FieldDescriptor.TYPE_BOOL: wire_format.BoolByteSize,
+    _FieldDescriptor.TYPE_STRING: wire_format.StringByteSize,
+    _FieldDescriptor.TYPE_GROUP: wire_format.GroupByteSize,
+    _FieldDescriptor.TYPE_MESSAGE: wire_format.MessageByteSize,
+    _FieldDescriptor.TYPE_BYTES: wire_format.BytesByteSize,
+    _FieldDescriptor.TYPE_UINT32: wire_format.UInt32ByteSize,
+    _FieldDescriptor.TYPE_ENUM: wire_format.EnumByteSize,
+    _FieldDescriptor.TYPE_SFIXED32: wire_format.SFixed32ByteSize,
+    _FieldDescriptor.TYPE_SFIXED64: wire_format.SFixed64ByteSize,
+    _FieldDescriptor.TYPE_SINT32: wire_format.SInt32ByteSize,
+    _FieldDescriptor.TYPE_SINT64: wire_format.SInt64ByteSize
+    }
+
+
+# Maps from field types to encoder constructors.
+TYPE_TO_ENCODER = {
+    _FieldDescriptor.TYPE_DOUBLE: encoder.DoubleEncoder,
+    _FieldDescriptor.TYPE_FLOAT: encoder.FloatEncoder,
+    _FieldDescriptor.TYPE_INT64: encoder.Int64Encoder,
+    _FieldDescriptor.TYPE_UINT64: encoder.UInt64Encoder,
+    _FieldDescriptor.TYPE_INT32: encoder.Int32Encoder,
+    _FieldDescriptor.TYPE_FIXED64: encoder.Fixed64Encoder,
+    _FieldDescriptor.TYPE_FIXED32: encoder.Fixed32Encoder,
+    _FieldDescriptor.TYPE_BOOL: encoder.BoolEncoder,
+    _FieldDescriptor.TYPE_STRING: encoder.StringEncoder,
+    _FieldDescriptor.TYPE_GROUP: encoder.GroupEncoder,
+    _FieldDescriptor.TYPE_MESSAGE: encoder.MessageEncoder,
+    _FieldDescriptor.TYPE_BYTES: encoder.BytesEncoder,
+    _FieldDescriptor.TYPE_UINT32: encoder.UInt32Encoder,
+    _FieldDescriptor.TYPE_ENUM: encoder.EnumEncoder,
+    _FieldDescriptor.TYPE_SFIXED32: encoder.SFixed32Encoder,
+    _FieldDescriptor.TYPE_SFIXED64: encoder.SFixed64Encoder,
+    _FieldDescriptor.TYPE_SINT32: encoder.SInt32Encoder,
+    _FieldDescriptor.TYPE_SINT64: encoder.SInt64Encoder,
+    }
+
+
+# Maps from field types to sizer constructors.
+TYPE_TO_SIZER = {
+    _FieldDescriptor.TYPE_DOUBLE: encoder.DoubleSizer,
+    _FieldDescriptor.TYPE_FLOAT: encoder.FloatSizer,
+    _FieldDescriptor.TYPE_INT64: encoder.Int64Sizer,
+    _FieldDescriptor.TYPE_UINT64: encoder.UInt64Sizer,
+    _FieldDescriptor.TYPE_INT32: encoder.Int32Sizer,
+    _FieldDescriptor.TYPE_FIXED64: encoder.Fixed64Sizer,
+    _FieldDescriptor.TYPE_FIXED32: encoder.Fixed32Sizer,
+    _FieldDescriptor.TYPE_BOOL: encoder.BoolSizer,
+    _FieldDescriptor.TYPE_STRING: encoder.StringSizer,
+    _FieldDescriptor.TYPE_GROUP: encoder.GroupSizer,
+    _FieldDescriptor.TYPE_MESSAGE: encoder.MessageSizer,
+    _FieldDescriptor.TYPE_BYTES: encoder.BytesSizer,
+    _FieldDescriptor.TYPE_UINT32: encoder.UInt32Sizer,
+    _FieldDescriptor.TYPE_ENUM: encoder.EnumSizer,
+    _FieldDescriptor.TYPE_SFIXED32: encoder.SFixed32Sizer,
+    _FieldDescriptor.TYPE_SFIXED64: encoder.SFixed64Sizer,
+    _FieldDescriptor.TYPE_SINT32: encoder.SInt32Sizer,
+    _FieldDescriptor.TYPE_SINT64: encoder.SInt64Sizer,
+    }
+
+
+# Maps from field type to a decoder constructor.
+TYPE_TO_DECODER = {
+    _FieldDescriptor.TYPE_DOUBLE: decoder.DoubleDecoder,
+    _FieldDescriptor.TYPE_FLOAT: decoder.FloatDecoder,
+    _FieldDescriptor.TYPE_INT64: decoder.Int64Decoder,
+    _FieldDescriptor.TYPE_UINT64: decoder.UInt64Decoder,
+    _FieldDescriptor.TYPE_INT32: decoder.Int32Decoder,
+    _FieldDescriptor.TYPE_FIXED64: decoder.Fixed64Decoder,
+    _FieldDescriptor.TYPE_FIXED32: decoder.Fixed32Decoder,
+    _FieldDescriptor.TYPE_BOOL: decoder.BoolDecoder,
+    _FieldDescriptor.TYPE_STRING: decoder.StringDecoder,
+    _FieldDescriptor.TYPE_GROUP: decoder.GroupDecoder,
+    _FieldDescriptor.TYPE_MESSAGE: decoder.MessageDecoder,
+    _FieldDescriptor.TYPE_BYTES: decoder.BytesDecoder,
+    _FieldDescriptor.TYPE_UINT32: decoder.UInt32Decoder,
+    _FieldDescriptor.TYPE_ENUM: decoder.EnumDecoder,
+    _FieldDescriptor.TYPE_SFIXED32: decoder.SFixed32Decoder,
+    _FieldDescriptor.TYPE_SFIXED64: decoder.SFixed64Decoder,
+    _FieldDescriptor.TYPE_SINT32: decoder.SInt32Decoder,
+    _FieldDescriptor.TYPE_SINT64: decoder.SInt64Decoder,
+    }
+
+# Maps from field type to expected wiretype.
+FIELD_TYPE_TO_WIRE_TYPE = {
+    _FieldDescriptor.TYPE_DOUBLE: wire_format.WIRETYPE_FIXED64,
+    _FieldDescriptor.TYPE_FLOAT: wire_format.WIRETYPE_FIXED32,
+    _FieldDescriptor.TYPE_INT64: wire_format.WIRETYPE_VARINT,
+    _FieldDescriptor.TYPE_UINT64: wire_format.WIRETYPE_VARINT,
+    _FieldDescriptor.TYPE_INT32: wire_format.WIRETYPE_VARINT,
+    _FieldDescriptor.TYPE_FIXED64: wire_format.WIRETYPE_FIXED64,
+    _FieldDescriptor.TYPE_FIXED32: wire_format.WIRETYPE_FIXED32,
+    _FieldDescriptor.TYPE_BOOL: wire_format.WIRETYPE_VARINT,
+    _FieldDescriptor.TYPE_STRING:
+      wire_format.WIRETYPE_LENGTH_DELIMITED,
+    _FieldDescriptor.TYPE_GROUP: wire_format.WIRETYPE_START_GROUP,
+    _FieldDescriptor.TYPE_MESSAGE:
+      wire_format.WIRETYPE_LENGTH_DELIMITED,
+    _FieldDescriptor.TYPE_BYTES:
+      wire_format.WIRETYPE_LENGTH_DELIMITED,
+    _FieldDescriptor.TYPE_UINT32: wire_format.WIRETYPE_VARINT,
+    _FieldDescriptor.TYPE_ENUM: wire_format.WIRETYPE_VARINT,
+    _FieldDescriptor.TYPE_SFIXED32: wire_format.WIRETYPE_FIXED32,
+    _FieldDescriptor.TYPE_SFIXED64: wire_format.WIRETYPE_FIXED64,
+    _FieldDescriptor.TYPE_SINT32: wire_format.WIRETYPE_VARINT,
+    _FieldDescriptor.TYPE_SINT64: wire_format.WIRETYPE_VARINT,
+    }
diff --git a/python/google/protobuf/internal/unknown_fields_test.py b/python/google/protobuf/internal/unknown_fields_test.py
new file mode 100755
index 0000000..9685b8b
--- /dev/null
+++ b/python/google/protobuf/internal/unknown_fields_test.py
@@ -0,0 +1,300 @@
+#! /usr/bin/env python
+# -*- coding: utf-8 -*-
+#
+# Protocol Buffers - Google's data interchange format
+# Copyright 2008 Google Inc.  All rights reserved.
+# https://developers.google.com/protocol-buffers/
+#
+# Redistribution and use in source and binary forms, with or without
+# modification, are permitted provided that the following conditions are
+# met:
+#
+#     * Redistributions of source code must retain the above copyright
+# notice, this list of conditions and the following disclaimer.
+#     * Redistributions in binary form must reproduce the above
+# copyright notice, this list of conditions and the following disclaimer
+# in the documentation and/or other materials provided with the
+# distribution.
+#     * Neither the name of Google Inc. nor the names of its
+# contributors may be used to endorse or promote products derived from
+# this software without specific prior written permission.
+#
+# THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+# "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
+# LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
+# A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
+# OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
+# SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
+# LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
+# DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
+# THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
+# (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+# OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+
+"""Test for preservation of unknown fields in the pure Python implementation."""
+
+__author__ = 'bohdank@google.com (Bohdan Koval)'
+
+try:
+  import unittest2 as unittest
+except ImportError:
+  import unittest
+from google.protobuf import unittest_mset_pb2
+from google.protobuf import unittest_pb2
+from google.protobuf import unittest_proto3_arena_pb2
+from google.protobuf.internal import api_implementation
+from google.protobuf.internal import encoder
+from google.protobuf.internal import message_set_extensions_pb2
+from google.protobuf.internal import missing_enum_values_pb2
+from google.protobuf.internal import test_util
+from google.protobuf.internal import type_checkers
+
+
+def SkipIfCppImplementation(func):
+  return unittest.skipIf(
+      api_implementation.Type() == 'cpp' and api_implementation.Version() == 2,
+      'C++ implementation does not expose unknown fields to Python')(func)
+
+
+class UnknownFieldsTest(unittest.TestCase):
+
+  def setUp(self):
+    self.descriptor = unittest_pb2.TestAllTypes.DESCRIPTOR
+    self.all_fields = unittest_pb2.TestAllTypes()
+    test_util.SetAllFields(self.all_fields)
+    self.all_fields_data = self.all_fields.SerializeToString()
+    self.empty_message = unittest_pb2.TestEmptyMessage()
+    self.empty_message.ParseFromString(self.all_fields_data)
+
+  def testSerialize(self):
+    data = self.empty_message.SerializeToString()
+
+    # Don't use assertEqual because we don't want to dump raw binary data to
+    # stdout.
+    self.assertTrue(data == self.all_fields_data)
+
+  def testSerializeProto3(self):
+    # Verify that proto3 doesn't preserve unknown fields.
+    message = unittest_proto3_arena_pb2.TestEmptyMessage()
+    message.ParseFromString(self.all_fields_data)
+    self.assertEqual(0, len(message.SerializeToString()))
+
+  def testByteSize(self):
+    self.assertEqual(self.all_fields.ByteSize(), self.empty_message.ByteSize())
+
+  def testListFields(self):
+    # Make sure ListFields doesn't return unknown fields.
+    self.assertEqual(0, len(self.empty_message.ListFields()))
+
+  def testSerializeMessageSetWireFormatUnknownExtension(self):
+    # Create a message using the message set wire format with an unknown
+    # message.
+    raw = unittest_mset_pb2.RawMessageSet()
+
+    # Add an unknown extension.
+    item = raw.item.add()
+    item.type_id = 98418603
+    message1 = message_set_extensions_pb2.TestMessageSetExtension1()
+    message1.i = 12345
+    item.message = message1.SerializeToString()
+
+    serialized = raw.SerializeToString()
+
+    # Parse message using the message set wire format.
+    proto = message_set_extensions_pb2.TestMessageSet()
+    proto.MergeFromString(serialized)
+
+    # Verify that the unknown extension is serialized unchanged
+    reserialized = proto.SerializeToString()
+    new_raw = unittest_mset_pb2.RawMessageSet()
+    new_raw.MergeFromString(reserialized)
+    self.assertEqual(raw, new_raw)
+
+  def testEquals(self):
+    message = unittest_pb2.TestEmptyMessage()
+    message.ParseFromString(self.all_fields_data)
+    self.assertEqual(self.empty_message, message)
+
+    self.all_fields.ClearField('optional_string')
+    message.ParseFromString(self.all_fields.SerializeToString())
+    self.assertNotEqual(self.empty_message, message)
+
+
+class UnknownFieldsAccessorsTest(unittest.TestCase):
+
+  def setUp(self):
+    self.descriptor = unittest_pb2.TestAllTypes.DESCRIPTOR
+    self.all_fields = unittest_pb2.TestAllTypes()
+    test_util.SetAllFields(self.all_fields)
+    self.all_fields_data = self.all_fields.SerializeToString()
+    self.empty_message = unittest_pb2.TestEmptyMessage()
+    self.empty_message.ParseFromString(self.all_fields_data)
+    if api_implementation.Type() != 'cpp':
+      # _unknown_fields is an implementation detail.
+      self.unknown_fields = self.empty_message._unknown_fields
+
+  # All the tests that use GetField() check an implementation detail of the
+  # Python implementation, which stores unknown fields as serialized strings.
+  # These tests are skipped by the C++ implementation: it's enough to check that
+  # the message is correctly serialized.
+
+  def GetField(self, name):
+    field_descriptor = self.descriptor.fields_by_name[name]
+    wire_type = type_checkers.FIELD_TYPE_TO_WIRE_TYPE[field_descriptor.type]
+    field_tag = encoder.TagBytes(field_descriptor.number, wire_type)
+    result_dict = {}
+    for tag_bytes, value in self.unknown_fields:
+      if tag_bytes == field_tag:
+        decoder = unittest_pb2.TestAllTypes._decoders_by_tag[tag_bytes][0]
+        decoder(value, 0, len(value), self.all_fields, result_dict)
+    return result_dict[field_descriptor]
+
+  @SkipIfCppImplementation
+  def testEnum(self):
+    value = self.GetField('optional_nested_enum')
+    self.assertEqual(self.all_fields.optional_nested_enum, value)
+
+  @SkipIfCppImplementation
+  def testRepeatedEnum(self):
+    value = self.GetField('repeated_nested_enum')
+    self.assertEqual(self.all_fields.repeated_nested_enum, value)
+
+  @SkipIfCppImplementation
+  def testVarint(self):
+    value = self.GetField('optional_int32')
+    self.assertEqual(self.all_fields.optional_int32, value)
+
+  @SkipIfCppImplementation
+  def testFixed32(self):
+    value = self.GetField('optional_fixed32')
+    self.assertEqual(self.all_fields.optional_fixed32, value)
+
+  @SkipIfCppImplementation
+  def testFixed64(self):
+    value = self.GetField('optional_fixed64')
+    self.assertEqual(self.all_fields.optional_fixed64, value)
+
+  @SkipIfCppImplementation
+  def testLengthDelimited(self):
+    value = self.GetField('optional_string')
+    self.assertEqual(self.all_fields.optional_string, value)
+
+  @SkipIfCppImplementation
+  def testGroup(self):
+    value = self.GetField('optionalgroup')
+    self.assertEqual(self.all_fields.optionalgroup, value)
+
+  def testCopyFrom(self):
+    message = unittest_pb2.TestEmptyMessage()
+    message.CopyFrom(self.empty_message)
+    self.assertEqual(message.SerializeToString(), self.all_fields_data)
+
+  def testMergeFrom(self):
+    message = unittest_pb2.TestAllTypes()
+    message.optional_int32 = 1
+    message.optional_uint32 = 2
+    source = unittest_pb2.TestEmptyMessage()
+    source.ParseFromString(message.SerializeToString())
+
+    message.ClearField('optional_int32')
+    message.optional_int64 = 3
+    message.optional_uint32 = 4
+    destination = unittest_pb2.TestEmptyMessage()
+    destination.ParseFromString(message.SerializeToString())
+
+    destination.MergeFrom(source)
+    # Check that the fields where correctly merged, even stored in the unknown
+    # fields set.
+    message.ParseFromString(destination.SerializeToString())
+    self.assertEqual(message.optional_int32, 1)
+    self.assertEqual(message.optional_uint32, 2)
+    self.assertEqual(message.optional_int64, 3)
+
+  def testClear(self):
+    self.empty_message.Clear()
+    # All cleared, even unknown fields.
+    self.assertEqual(self.empty_message.SerializeToString(), b'')
+
+  def testUnknownExtensions(self):
+    message = unittest_pb2.TestEmptyMessageWithExtensions()
+    message.ParseFromString(self.all_fields_data)
+    self.assertEqual(message.SerializeToString(), self.all_fields_data)
+
+
+class UnknownEnumValuesTest(unittest.TestCase):
+
+  def setUp(self):
+    self.descriptor = missing_enum_values_pb2.TestEnumValues.DESCRIPTOR
+
+    self.message = missing_enum_values_pb2.TestEnumValues()
+    self.message.optional_nested_enum = (
+      missing_enum_values_pb2.TestEnumValues.ZERO)
+    self.message.repeated_nested_enum.extend([
+      missing_enum_values_pb2.TestEnumValues.ZERO,
+      missing_enum_values_pb2.TestEnumValues.ONE,
+      ])
+    self.message.packed_nested_enum.extend([
+      missing_enum_values_pb2.TestEnumValues.ZERO,
+      missing_enum_values_pb2.TestEnumValues.ONE,
+      ])
+    self.message_data = self.message.SerializeToString()
+    self.missing_message = missing_enum_values_pb2.TestMissingEnumValues()
+    self.missing_message.ParseFromString(self.message_data)
+    if api_implementation.Type() != 'cpp':
+      # _unknown_fields is an implementation detail.
+      self.unknown_fields = self.missing_message._unknown_fields
+
+  # All the tests that use GetField() check an implementation detail of the
+  # Python implementation, which stores unknown fields as serialized strings.
+  # These tests are skipped by the C++ implementation: it's enough to check that
+  # the message is correctly serialized.
+
+  def GetField(self, name):
+    field_descriptor = self.descriptor.fields_by_name[name]
+    wire_type = type_checkers.FIELD_TYPE_TO_WIRE_TYPE[field_descriptor.type]
+    field_tag = encoder.TagBytes(field_descriptor.number, wire_type)
+    result_dict = {}
+    for tag_bytes, value in self.unknown_fields:
+      if tag_bytes == field_tag:
+        decoder = missing_enum_values_pb2.TestEnumValues._decoders_by_tag[
+          tag_bytes][0]
+        decoder(value, 0, len(value), self.message, result_dict)
+    return result_dict[field_descriptor]
+
+  def testUnknownParseMismatchEnumValue(self):
+    just_string = missing_enum_values_pb2.JustString()
+    just_string.dummy = 'blah'
+
+    missing = missing_enum_values_pb2.TestEnumValues()
+    # The parse is invalid, storing the string proto into the set of
+    # unknown fields.
+    missing.ParseFromString(just_string.SerializeToString())
+
+    # Fetching the enum field shouldn't crash, instead returning the
+    # default value.
+    self.assertEqual(missing.optional_nested_enum, 0)
+
+  @SkipIfCppImplementation
+  def testUnknownEnumValue(self):
+    self.assertFalse(self.missing_message.HasField('optional_nested_enum'))
+    value = self.GetField('optional_nested_enum')
+    self.assertEqual(self.message.optional_nested_enum, value)
+
+  @SkipIfCppImplementation
+  def testUnknownRepeatedEnumValue(self):
+    value = self.GetField('repeated_nested_enum')
+    self.assertEqual(self.message.repeated_nested_enum, value)
+
+  @SkipIfCppImplementation
+  def testUnknownPackedEnumValue(self):
+    value = self.GetField('packed_nested_enum')
+    self.assertEqual(self.message.packed_nested_enum, value)
+
+  def testRoundTrip(self):
+    new_message = missing_enum_values_pb2.TestEnumValues()
+    new_message.ParseFromString(self.missing_message.SerializeToString())
+    self.assertEqual(self.message, new_message)
+
+
+if __name__ == '__main__':
+  unittest.main()
diff --git a/python/google/protobuf/internal/well_known_types.py b/python/google/protobuf/internal/well_known_types.py
new file mode 100644
index 0000000..d35fcc5
--- /dev/null
+++ b/python/google/protobuf/internal/well_known_types.py
@@ -0,0 +1,720 @@
+# Protocol Buffers - Google's data interchange format
+# Copyright 2008 Google Inc.  All rights reserved.
+# https://developers.google.com/protocol-buffers/
+#
+# Redistribution and use in source and binary forms, with or without
+# modification, are permitted provided that the following conditions are
+# met:
+#
+#     * Redistributions of source code must retain the above copyright
+# notice, this list of conditions and the following disclaimer.
+#     * Redistributions in binary form must reproduce the above
+# copyright notice, this list of conditions and the following disclaimer
+# in the documentation and/or other materials provided with the
+# distribution.
+#     * Neither the name of Google Inc. nor the names of its
+# contributors may be used to endorse or promote products derived from
+# this software without specific prior written permission.
+#
+# THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+# "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
+# LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
+# A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
+# OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
+# SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
+# LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
+# DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
+# THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
+# (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+# OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+
+"""Contains well known classes.
+
+This files defines well known classes which need extra maintenance including:
+  - Any
+  - Duration
+  - FieldMask
+  - Struct
+  - Timestamp
+"""
+
+__author__ = 'jieluo@google.com (Jie Luo)'
+
+from datetime import datetime
+from datetime import timedelta
+import six
+
+from google.protobuf.descriptor import FieldDescriptor
+
+_TIMESTAMPFOMAT = '%Y-%m-%dT%H:%M:%S'
+_NANOS_PER_SECOND = 1000000000
+_NANOS_PER_MILLISECOND = 1000000
+_NANOS_PER_MICROSECOND = 1000
+_MILLIS_PER_SECOND = 1000
+_MICROS_PER_SECOND = 1000000
+_SECONDS_PER_DAY = 24 * 3600
+
+
+class Error(Exception):
+  """Top-level module error."""
+
+
+class ParseError(Error):
+  """Thrown in case of parsing error."""
+
+
+class Any(object):
+  """Class for Any Message type."""
+
+  def Pack(self, msg, type_url_prefix='type.googleapis.com/'):
+    """Packs the specified message into current Any message."""
+    if len(type_url_prefix) < 1 or type_url_prefix[-1] != '/':
+      self.type_url = '%s/%s' % (type_url_prefix, msg.DESCRIPTOR.full_name)
+    else:
+      self.type_url = '%s%s' % (type_url_prefix, msg.DESCRIPTOR.full_name)
+    self.value = msg.SerializeToString()
+
+  def Unpack(self, msg):
+    """Unpacks the current Any message into specified message."""
+    descriptor = msg.DESCRIPTOR
+    if not self.Is(descriptor):
+      return False
+    msg.ParseFromString(self.value)
+    return True
+
+  def Is(self, descriptor):
+    """Checks if this Any represents the given protobuf type."""
+    # Only last part is to be used: b/25630112
+    return self.type_url.split('/')[-1] == descriptor.full_name
+
+
+class Timestamp(object):
+  """Class for Timestamp message type."""
+
+  def ToJsonString(self):
+    """Converts Timestamp to RFC 3339 date string format.
+
+    Returns:
+      A string converted from timestamp. The string is always Z-normalized
+      and uses 3, 6 or 9 fractional digits as required to represent the
+      exact time. Example of the return format: '1972-01-01T10:00:20.021Z'
+    """
+    nanos = self.nanos % _NANOS_PER_SECOND
+    total_sec = self.seconds + (self.nanos - nanos) // _NANOS_PER_SECOND
+    seconds = total_sec % _SECONDS_PER_DAY
+    days = (total_sec - seconds) // _SECONDS_PER_DAY
+    dt = datetime(1970, 1, 1) + timedelta(days, seconds)
+
+    result = dt.isoformat()
+    if (nanos % 1e9) == 0:
+      # If there are 0 fractional digits, the fractional
+      # point '.' should be omitted when serializing.
+      return result + 'Z'
+    if (nanos % 1e6) == 0:
+      # Serialize 3 fractional digits.
+      return result + '.%03dZ' % (nanos / 1e6)
+    if (nanos % 1e3) == 0:
+      # Serialize 6 fractional digits.
+      return result + '.%06dZ' % (nanos / 1e3)
+    # Serialize 9 fractional digits.
+    return result + '.%09dZ' % nanos
+
+  def FromJsonString(self, value):
+    """Parse a RFC 3339 date string format to Timestamp.
+
+    Args:
+      value: A date string. Any fractional digits (or none) and any offset are
+          accepted as long as they fit into nano-seconds precision.
+          Example of accepted format: '1972-01-01T10:00:20.021-05:00'
+
+    Raises:
+      ParseError: On parsing problems.
+    """
+    timezone_offset = value.find('Z')
+    if timezone_offset == -1:
+      timezone_offset = value.find('+')
+    if timezone_offset == -1:
+      timezone_offset = value.rfind('-')
+    if timezone_offset == -1:
+      raise ParseError(
+          'Failed to parse timestamp: missing valid timezone offset.')
+    time_value = value[0:timezone_offset]
+    # Parse datetime and nanos.
+    point_position = time_value.find('.')
+    if point_position == -1:
+      second_value = time_value
+      nano_value = ''
+    else:
+      second_value = time_value[:point_position]
+      nano_value = time_value[point_position + 1:]
+    date_object = datetime.strptime(second_value, _TIMESTAMPFOMAT)
+    td = date_object - datetime(1970, 1, 1)
+    seconds = td.seconds + td.days * _SECONDS_PER_DAY
+    if len(nano_value) > 9:
+      raise ParseError(
+          'Failed to parse Timestamp: nanos {0} more than '
+          '9 fractional digits.'.format(nano_value))
+    if nano_value:
+      nanos = round(float('0.' + nano_value) * 1e9)
+    else:
+      nanos = 0
+    # Parse timezone offsets.
+    if value[timezone_offset] == 'Z':
+      if len(value) != timezone_offset + 1:
+        raise ParseError('Failed to parse timestamp: invalid trailing'
+                         ' data {0}.'.format(value))
+    else:
+      timezone = value[timezone_offset:]
+      pos = timezone.find(':')
+      if pos == -1:
+        raise ParseError(
+            'Invalid timezone offset value: {0}.'.format(timezone))
+      if timezone[0] == '+':
+        seconds -= (int(timezone[1:pos])*60+int(timezone[pos+1:]))*60
+      else:
+        seconds += (int(timezone[1:pos])*60+int(timezone[pos+1:]))*60
+    # Set seconds and nanos
+    self.seconds = int(seconds)
+    self.nanos = int(nanos)
+
+  def GetCurrentTime(self):
+    """Get the current UTC into Timestamp."""
+    self.FromDatetime(datetime.utcnow())
+
+  def ToNanoseconds(self):
+    """Converts Timestamp to nanoseconds since epoch."""
+    return self.seconds * _NANOS_PER_SECOND + self.nanos
+
+  def ToMicroseconds(self):
+    """Converts Timestamp to microseconds since epoch."""
+    return (self.seconds * _MICROS_PER_SECOND +
+            self.nanos // _NANOS_PER_MICROSECOND)
+
+  def ToMilliseconds(self):
+    """Converts Timestamp to milliseconds since epoch."""
+    return (self.seconds * _MILLIS_PER_SECOND +
+            self.nanos // _NANOS_PER_MILLISECOND)
+
+  def ToSeconds(self):
+    """Converts Timestamp to seconds since epoch."""
+    return self.seconds
+
+  def FromNanoseconds(self, nanos):
+    """Converts nanoseconds since epoch to Timestamp."""
+    self.seconds = nanos // _NANOS_PER_SECOND
+    self.nanos = nanos % _NANOS_PER_SECOND
+
+  def FromMicroseconds(self, micros):
+    """Converts microseconds since epoch to Timestamp."""
+    self.seconds = micros // _MICROS_PER_SECOND
+    self.nanos = (micros % _MICROS_PER_SECOND) * _NANOS_PER_MICROSECOND
+
+  def FromMilliseconds(self, millis):
+    """Converts milliseconds since epoch to Timestamp."""
+    self.seconds = millis // _MILLIS_PER_SECOND
+    self.nanos = (millis % _MILLIS_PER_SECOND) * _NANOS_PER_MILLISECOND
+
+  def FromSeconds(self, seconds):
+    """Converts seconds since epoch to Timestamp."""
+    self.seconds = seconds
+    self.nanos = 0
+
+  def ToDatetime(self):
+    """Converts Timestamp to datetime."""
+    return datetime.utcfromtimestamp(
+        self.seconds + self.nanos / float(_NANOS_PER_SECOND))
+
+  def FromDatetime(self, dt):
+    """Converts datetime to Timestamp."""
+    td = dt - datetime(1970, 1, 1)
+    self.seconds = td.seconds + td.days * _SECONDS_PER_DAY
+    self.nanos = td.microseconds * _NANOS_PER_MICROSECOND
+
+
+class Duration(object):
+  """Class for Duration message type."""
+
+  def ToJsonString(self):
+    """Converts Duration to string format.
+
+    Returns:
+      A string converted from self. The string format will contains
+      3, 6, or 9 fractional digits depending on the precision required to
+      represent the exact Duration value. For example: "1s", "1.010s",
+      "1.000000100s", "-3.100s"
+    """
+    if self.seconds < 0 or self.nanos < 0:
+      result = '-'
+      seconds = - self.seconds + int((0 - self.nanos) // 1e9)
+      nanos = (0 - self.nanos) % 1e9
+    else:
+      result = ''
+      seconds = self.seconds + int(self.nanos // 1e9)
+      nanos = self.nanos % 1e9
+    result += '%d' % seconds
+    if (nanos % 1e9) == 0:
+      # If there are 0 fractional digits, the fractional
+      # point '.' should be omitted when serializing.
+      return result + 's'
+    if (nanos % 1e6) == 0:
+      # Serialize 3 fractional digits.
+      return result + '.%03ds' % (nanos / 1e6)
+    if (nanos % 1e3) == 0:
+      # Serialize 6 fractional digits.
+      return result + '.%06ds' % (nanos / 1e3)
+    # Serialize 9 fractional digits.
+    return result + '.%09ds' % nanos
+
+  def FromJsonString(self, value):
+    """Converts a string to Duration.
+
+    Args:
+      value: A string to be converted. The string must end with 's'. Any
+          fractional digits (or none) are accepted as long as they fit into
+          precision. For example: "1s", "1.01s", "1.0000001s", "-3.100s
+
+    Raises:
+      ParseError: On parsing problems.
+    """
+    if len(value) < 1 or value[-1] != 's':
+      raise ParseError(
+          'Duration must end with letter "s": {0}.'.format(value))
+    try:
+      pos = value.find('.')
+      if pos == -1:
+        self.seconds = int(value[:-1])
+        self.nanos = 0
+      else:
+        self.seconds = int(value[:pos])
+        if value[0] == '-':
+          self.nanos = int(round(float('-0{0}'.format(value[pos: -1])) *1e9))
+        else:
+          self.nanos = int(round(float('0{0}'.format(value[pos: -1])) *1e9))
+    except ValueError:
+      raise ParseError(
+          'Couldn\'t parse duration: {0}.'.format(value))
+
+  def ToNanoseconds(self):
+    """Converts a Duration to nanoseconds."""
+    return self.seconds * _NANOS_PER_SECOND + self.nanos
+
+  def ToMicroseconds(self):
+    """Converts a Duration to microseconds."""
+    micros = _RoundTowardZero(self.nanos, _NANOS_PER_MICROSECOND)
+    return self.seconds * _MICROS_PER_SECOND + micros
+
+  def ToMilliseconds(self):
+    """Converts a Duration to milliseconds."""
+    millis = _RoundTowardZero(self.nanos, _NANOS_PER_MILLISECOND)
+    return self.seconds * _MILLIS_PER_SECOND + millis
+
+  def ToSeconds(self):
+    """Converts a Duration to seconds."""
+    return self.seconds
+
+  def FromNanoseconds(self, nanos):
+    """Converts nanoseconds to Duration."""
+    self._NormalizeDuration(nanos // _NANOS_PER_SECOND,
+                            nanos % _NANOS_PER_SECOND)
+
+  def FromMicroseconds(self, micros):
+    """Converts microseconds to Duration."""
+    self._NormalizeDuration(
+        micros // _MICROS_PER_SECOND,
+        (micros % _MICROS_PER_SECOND) * _NANOS_PER_MICROSECOND)
+
+  def FromMilliseconds(self, millis):
+    """Converts milliseconds to Duration."""
+    self._NormalizeDuration(
+        millis // _MILLIS_PER_SECOND,
+        (millis % _MILLIS_PER_SECOND) * _NANOS_PER_MILLISECOND)
+
+  def FromSeconds(self, seconds):
+    """Converts seconds to Duration."""
+    self.seconds = seconds
+    self.nanos = 0
+
+  def ToTimedelta(self):
+    """Converts Duration to timedelta."""
+    return timedelta(
+        seconds=self.seconds, microseconds=_RoundTowardZero(
+            self.nanos, _NANOS_PER_MICROSECOND))
+
+  def FromTimedelta(self, td):
+    """Convertd timedelta to Duration."""
+    self._NormalizeDuration(td.seconds + td.days * _SECONDS_PER_DAY,
+                            td.microseconds * _NANOS_PER_MICROSECOND)
+
+  def _NormalizeDuration(self, seconds, nanos):
+    """Set Duration by seconds and nonas."""
+    # Force nanos to be negative if the duration is negative.
+    if seconds < 0 and nanos > 0:
+      seconds += 1
+      nanos -= _NANOS_PER_SECOND
+    self.seconds = seconds
+    self.nanos = nanos
+
+
+def _RoundTowardZero(value, divider):
+  """Truncates the remainder part after division."""
+  # For some languanges, the sign of the remainder is implementation
+  # dependent if any of the operands is negative. Here we enforce
+  # "rounded toward zero" semantics. For example, for (-5) / 2 an
+  # implementation may give -3 as the result with the remainder being
+  # 1. This function ensures we always return -2 (closer to zero).
+  result = value // divider
+  remainder = value % divider
+  if result < 0 and remainder > 0:
+    return result + 1
+  else:
+    return result
+
+
+class FieldMask(object):
+  """Class for FieldMask message type."""
+
+  def ToJsonString(self):
+    """Converts FieldMask to string according to proto3 JSON spec."""
+    return ','.join(self.paths)
+
+  def FromJsonString(self, value):
+    """Converts string to FieldMask according to proto3 JSON spec."""
+    self.Clear()
+    for path in value.split(','):
+      self.paths.append(path)
+
+  def IsValidForDescriptor(self, message_descriptor):
+    """Checks whether the FieldMask is valid for Message Descriptor."""
+    for path in self.paths:
+      if not _IsValidPath(message_descriptor, path):
+        return False
+    return True
+
+  def AllFieldsFromDescriptor(self, message_descriptor):
+    """Gets all direct fields of Message Descriptor to FieldMask."""
+    self.Clear()
+    for field in message_descriptor.fields:
+      self.paths.append(field.name)
+
+  def CanonicalFormFromMask(self, mask):
+    """Converts a FieldMask to the canonical form.
+
+    Removes paths that are covered by another path. For example,
+    "foo.bar" is covered by "foo" and will be removed if "foo"
+    is also in the FieldMask. Then sorts all paths in alphabetical order.
+
+    Args:
+      mask: The original FieldMask to be converted.
+    """
+    tree = _FieldMaskTree(mask)
+    tree.ToFieldMask(self)
+
+  def Union(self, mask1, mask2):
+    """Merges mask1 and mask2 into this FieldMask."""
+    _CheckFieldMaskMessage(mask1)
+    _CheckFieldMaskMessage(mask2)
+    tree = _FieldMaskTree(mask1)
+    tree.MergeFromFieldMask(mask2)
+    tree.ToFieldMask(self)
+
+  def Intersect(self, mask1, mask2):
+    """Intersects mask1 and mask2 into this FieldMask."""
+    _CheckFieldMaskMessage(mask1)
+    _CheckFieldMaskMessage(mask2)
+    tree = _FieldMaskTree(mask1)
+    intersection = _FieldMaskTree()
+    for path in mask2.paths:
+      tree.IntersectPath(path, intersection)
+    intersection.ToFieldMask(self)
+
+  def MergeMessage(
+      self, source, destination,
+      replace_message_field=False, replace_repeated_field=False):
+    """Merges fields specified in FieldMask from source to destination.
+
+    Args:
+      source: Source message.
+      destination: The destination message to be merged into.
+      replace_message_field: Replace message field if True. Merge message
+          field if False.
+      replace_repeated_field: Replace repeated field if True. Append
+          elements of repeated field if False.
+    """
+    tree = _FieldMaskTree(self)
+    tree.MergeMessage(
+        source, destination, replace_message_field, replace_repeated_field)
+
+
+def _IsValidPath(message_descriptor, path):
+  """Checks whether the path is valid for Message Descriptor."""
+  parts = path.split('.')
+  last = parts.pop()
+  for name in parts:
+    field = message_descriptor.fields_by_name[name]
+    if (field is None or
+        field.label == FieldDescriptor.LABEL_REPEATED or
+        field.type != FieldDescriptor.TYPE_MESSAGE):
+      return False
+    message_descriptor = field.message_type
+  return last in message_descriptor.fields_by_name
+
+
+def _CheckFieldMaskMessage(message):
+  """Raises ValueError if message is not a FieldMask."""
+  message_descriptor = message.DESCRIPTOR
+  if (message_descriptor.name != 'FieldMask' or
+      message_descriptor.file.name != 'google/protobuf/field_mask.proto'):
+    raise ValueError('Message {0} is not a FieldMask.'.format(
+        message_descriptor.full_name))
+
+
+class _FieldMaskTree(object):
+  """Represents a FieldMask in a tree structure.
+
+  For example, given a FieldMask "foo.bar,foo.baz,bar.baz",
+  the FieldMaskTree will be:
+      [_root] -+- foo -+- bar
+            |       |
+            |       +- baz
+            |
+            +- bar --- baz
+  In the tree, each leaf node represents a field path.
+  """
+
+  def __init__(self, field_mask=None):
+    """Initializes the tree by FieldMask."""
+    self._root = {}
+    if field_mask:
+      self.MergeFromFieldMask(field_mask)
+
+  def MergeFromFieldMask(self, field_mask):
+    """Merges a FieldMask to the tree."""
+    for path in field_mask.paths:
+      self.AddPath(path)
+
+  def AddPath(self, path):
+    """Adds a field path into the tree.
+
+    If the field path to add is a sub-path of an existing field path
+    in the tree (i.e., a leaf node), it means the tree already matches
+    the given path so nothing will be added to the tree. If the path
+    matches an existing non-leaf node in the tree, that non-leaf node
+    will be turned into a leaf node with all its children removed because
+    the path matches all the node's children. Otherwise, a new path will
+    be added.
+
+    Args:
+      path: The field path to add.
+    """
+    node = self._root
+    for name in path.split('.'):
+      if name not in node:
+        node[name] = {}
+      elif not node[name]:
+        # Pre-existing empty node implies we already have this entire tree.
+        return
+      node = node[name]
+    # Remove any sub-trees we might have had.
+    node.clear()
+
+  def ToFieldMask(self, field_mask):
+    """Converts the tree to a FieldMask."""
+    field_mask.Clear()
+    _AddFieldPaths(self._root, '', field_mask)
+
+  def IntersectPath(self, path, intersection):
+    """Calculates the intersection part of a field path with this tree.
+
+    Args:
+      path: The field path to calculates.
+      intersection: The out tree to record the intersection part.
+    """
+    node = self._root
+    for name in path.split('.'):
+      if name not in node:
+        return
+      elif not node[name]:
+        intersection.AddPath(path)
+        return
+      node = node[name]
+    intersection.AddLeafNodes(path, node)
+
+  def AddLeafNodes(self, prefix, node):
+    """Adds leaf nodes begin with prefix to this tree."""
+    if not node:
+      self.AddPath(prefix)
+    for name in node:
+      child_path = prefix + '.' + name
+      self.AddLeafNodes(child_path, node[name])
+
+  def MergeMessage(
+      self, source, destination,
+      replace_message, replace_repeated):
+    """Merge all fields specified by this tree from source to destination."""
+    _MergeMessage(
+        self._root, source, destination, replace_message, replace_repeated)
+
+
+def _StrConvert(value):
+  """Converts value to str if it is not."""
+  # This file is imported by c extension and some methods like ClearField
+  # requires string for the field name. py2/py3 has different text
+  # type and may use unicode.
+  if not isinstance(value, str):
+    return value.encode('utf-8')
+  return value
+
+
+def _MergeMessage(
+    node, source, destination, replace_message, replace_repeated):
+  """Merge all fields specified by a sub-tree from source to destination."""
+  source_descriptor = source.DESCRIPTOR
+  for name in node:
+    child = node[name]
+    field = source_descriptor.fields_by_name[name]
+    if field is None:
+      raise ValueError('Error: Can\'t find field {0} in message {1}.'.format(
+          name, source_descriptor.full_name))
+    if child:
+      # Sub-paths are only allowed for singular message fields.
+      if (field.label == FieldDescriptor.LABEL_REPEATED or
+          field.cpp_type != FieldDescriptor.CPPTYPE_MESSAGE):
+        raise ValueError('Error: Field {0} in message {1} is not a singular '
+                         'message field and cannot have sub-fields.'.format(
+                             name, source_descriptor.full_name))
+      _MergeMessage(
+          child, getattr(source, name), getattr(destination, name),
+          replace_message, replace_repeated)
+      continue
+    if field.label == FieldDescriptor.LABEL_REPEATED:
+      if replace_repeated:
+        destination.ClearField(_StrConvert(name))
+      repeated_source = getattr(source, name)
+      repeated_destination = getattr(destination, name)
+      if field.cpp_type == FieldDescriptor.CPPTYPE_MESSAGE:
+        for item in repeated_source:
+          repeated_destination.add().MergeFrom(item)
+      else:
+        repeated_destination.extend(repeated_source)
+    else:
+      if field.cpp_type == FieldDescriptor.CPPTYPE_MESSAGE:
+        if replace_message:
+          destination.ClearField(_StrConvert(name))
+        if source.HasField(name):
+          getattr(destination, name).MergeFrom(getattr(source, name))
+      else:
+        setattr(destination, name, getattr(source, name))
+
+
+def _AddFieldPaths(node, prefix, field_mask):
+  """Adds the field paths descended from node to field_mask."""
+  if not node:
+    field_mask.paths.append(prefix)
+    return
+  for name in sorted(node):
+    if prefix:
+      child_path = prefix + '.' + name
+    else:
+      child_path = name
+    _AddFieldPaths(node[name], child_path, field_mask)
+
+
+_INT_OR_FLOAT = six.integer_types + (float,)
+
+
+def _SetStructValue(struct_value, value):
+  if value is None:
+    struct_value.null_value = 0
+  elif isinstance(value, bool):
+    # Note: this check must come before the number check because in Python
+    # True and False are also considered numbers.
+    struct_value.bool_value = value
+  elif isinstance(value, six.string_types):
+    struct_value.string_value = value
+  elif isinstance(value, _INT_OR_FLOAT):
+    struct_value.number_value = value
+  else:
+    raise ValueError('Unexpected type')
+
+
+def _GetStructValue(struct_value):
+  which = struct_value.WhichOneof('kind')
+  if which == 'struct_value':
+    return struct_value.struct_value
+  elif which == 'null_value':
+    return None
+  elif which == 'number_value':
+    return struct_value.number_value
+  elif which == 'string_value':
+    return struct_value.string_value
+  elif which == 'bool_value':
+    return struct_value.bool_value
+  elif which == 'list_value':
+    return struct_value.list_value
+  elif which is None:
+    raise ValueError('Value not set')
+
+
+class Struct(object):
+  """Class for Struct message type."""
+
+  __slots__ = []
+
+  def __getitem__(self, key):
+    return _GetStructValue(self.fields[key])
+
+  def __setitem__(self, key, value):
+    _SetStructValue(self.fields[key], value)
+
+  def get_or_create_list(self, key):
+    """Returns a list for this key, creating if it didn't exist already."""
+    return self.fields[key].list_value
+
+  def get_or_create_struct(self, key):
+    """Returns a struct for this key, creating if it didn't exist already."""
+    return self.fields[key].struct_value
+
+  # TODO(haberman): allow constructing/merging from dict.
+
+
+class ListValue(object):
+  """Class for ListValue message type."""
+
+  def __len__(self):
+    return len(self.values)
+
+  def append(self, value):
+    _SetStructValue(self.values.add(), value)
+
+  def extend(self, elem_seq):
+    for value in elem_seq:
+      self.append(value)
+
+  def __getitem__(self, index):
+    """Retrieves item by the specified index."""
+    return _GetStructValue(self.values.__getitem__(index))
+
+  def __setitem__(self, index, value):
+    _SetStructValue(self.values.__getitem__(index), value)
+
+  def items(self):
+    for i in range(len(self)):
+      yield self[i]
+
+  def add_struct(self):
+    """Appends and returns a struct value as the next value in the list."""
+    return self.values.add().struct_value
+
+  def add_list(self):
+    """Appends and returns a list value as the next value in the list."""
+    return self.values.add().list_value
+
+
+WKTBASES = {
+    'google.protobuf.Any': Any,
+    'google.protobuf.Duration': Duration,
+    'google.protobuf.FieldMask': FieldMask,
+    'google.protobuf.ListValue': ListValue,
+    'google.protobuf.Struct': Struct,
+    'google.protobuf.Timestamp': Timestamp,
+}
diff --git a/python/google/protobuf/internal/well_known_types_test.py b/python/google/protobuf/internal/well_known_types_test.py
new file mode 100644
index 0000000..6acbee2
--- /dev/null
+++ b/python/google/protobuf/internal/well_known_types_test.py
@@ -0,0 +1,636 @@
+#! /usr/bin/env python
+#
+# Protocol Buffers - Google's data interchange format
+# Copyright 2008 Google Inc.  All rights reserved.
+# https://developers.google.com/protocol-buffers/
+#
+# Redistribution and use in source and binary forms, with or without
+# modification, are permitted provided that the following conditions are
+# met:
+#
+#     * Redistributions of source code must retain the above copyright
+# notice, this list of conditions and the following disclaimer.
+#     * Redistributions in binary form must reproduce the above
+# copyright notice, this list of conditions and the following disclaimer
+# in the documentation and/or other materials provided with the
+# distribution.
+#     * Neither the name of Google Inc. nor the names of its
+# contributors may be used to endorse or promote products derived from
+# this software without specific prior written permission.
+#
+# THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+# "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
+# LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
+# A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
+# OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
+# SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
+# LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
+# DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
+# THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
+# (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+# OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+
+"""Test for google.protobuf.internal.well_known_types."""
+
+__author__ = 'jieluo@google.com (Jie Luo)'
+
+from datetime import datetime
+
+try:
+  import unittest2 as unittest
+except ImportError:
+  import unittest
+
+from google.protobuf import any_pb2
+from google.protobuf import duration_pb2
+from google.protobuf import field_mask_pb2
+from google.protobuf import struct_pb2
+from google.protobuf import timestamp_pb2
+from google.protobuf import unittest_pb2
+from google.protobuf.internal import any_test_pb2
+from google.protobuf.internal import test_util
+from google.protobuf.internal import well_known_types
+from google.protobuf import descriptor
+from google.protobuf import text_format
+
+
+class TimeUtilTestBase(unittest.TestCase):
+
+  def CheckTimestampConversion(self, message, text):
+    self.assertEqual(text, message.ToJsonString())
+    parsed_message = timestamp_pb2.Timestamp()
+    parsed_message.FromJsonString(text)
+    self.assertEqual(message, parsed_message)
+
+  def CheckDurationConversion(self, message, text):
+    self.assertEqual(text, message.ToJsonString())
+    parsed_message = duration_pb2.Duration()
+    parsed_message.FromJsonString(text)
+    self.assertEqual(message, parsed_message)
+
+
+class TimeUtilTest(TimeUtilTestBase):
+
+  def testTimestampSerializeAndParse(self):
+    message = timestamp_pb2.Timestamp()
+    # Generated output should contain 3, 6, or 9 fractional digits.
+    message.seconds = 0
+    message.nanos = 0
+    self.CheckTimestampConversion(message, '1970-01-01T00:00:00Z')
+    message.nanos = 10000000
+    self.CheckTimestampConversion(message, '1970-01-01T00:00:00.010Z')
+    message.nanos = 10000
+    self.CheckTimestampConversion(message, '1970-01-01T00:00:00.000010Z')
+    message.nanos = 10
+    self.CheckTimestampConversion(message, '1970-01-01T00:00:00.000000010Z')
+    # Test min timestamps.
+    message.seconds = -62135596800
+    message.nanos = 0
+    self.CheckTimestampConversion(message, '0001-01-01T00:00:00Z')
+    # Test max timestamps.
+    message.seconds = 253402300799
+    message.nanos = 999999999
+    self.CheckTimestampConversion(message, '9999-12-31T23:59:59.999999999Z')
+    # Test negative timestamps.
+    message.seconds = -1
+    self.CheckTimestampConversion(message, '1969-12-31T23:59:59.999999999Z')
+
+    # Parsing accepts an fractional digits as long as they fit into nano
+    # precision.
+    message.FromJsonString('1970-01-01T00:00:00.1Z')
+    self.assertEqual(0, message.seconds)
+    self.assertEqual(100000000, message.nanos)
+    # Parsing accpets offsets.
+    message.FromJsonString('1970-01-01T00:00:00-08:00')
+    self.assertEqual(8 * 3600, message.seconds)
+    self.assertEqual(0, message.nanos)
+
+  def testDurationSerializeAndParse(self):
+    message = duration_pb2.Duration()
+    # Generated output should contain 3, 6, or 9 fractional digits.
+    message.seconds = 0
+    message.nanos = 0
+    self.CheckDurationConversion(message, '0s')
+    message.nanos = 10000000
+    self.CheckDurationConversion(message, '0.010s')
+    message.nanos = 10000
+    self.CheckDurationConversion(message, '0.000010s')
+    message.nanos = 10
+    self.CheckDurationConversion(message, '0.000000010s')
+
+    # Test min and max
+    message.seconds = 315576000000
+    message.nanos = 999999999
+    self.CheckDurationConversion(message, '315576000000.999999999s')
+    message.seconds = -315576000000
+    message.nanos = -999999999
+    self.CheckDurationConversion(message, '-315576000000.999999999s')
+
+    # Parsing accepts an fractional digits as long as they fit into nano
+    # precision.
+    message.FromJsonString('0.1s')
+    self.assertEqual(100000000, message.nanos)
+    message.FromJsonString('0.0000001s')
+    self.assertEqual(100, message.nanos)
+
+  def testTimestampIntegerConversion(self):
+    message = timestamp_pb2.Timestamp()
+    message.FromNanoseconds(1)
+    self.assertEqual('1970-01-01T00:00:00.000000001Z',
+                     message.ToJsonString())
+    self.assertEqual(1, message.ToNanoseconds())
+
+    message.FromNanoseconds(-1)
+    self.assertEqual('1969-12-31T23:59:59.999999999Z',
+                     message.ToJsonString())
+    self.assertEqual(-1, message.ToNanoseconds())
+
+    message.FromMicroseconds(1)
+    self.assertEqual('1970-01-01T00:00:00.000001Z',
+                     message.ToJsonString())
+    self.assertEqual(1, message.ToMicroseconds())
+
+    message.FromMicroseconds(-1)
+    self.assertEqual('1969-12-31T23:59:59.999999Z',
+                     message.ToJsonString())
+    self.assertEqual(-1, message.ToMicroseconds())
+
+    message.FromMilliseconds(1)
+    self.assertEqual('1970-01-01T00:00:00.001Z',
+                     message.ToJsonString())
+    self.assertEqual(1, message.ToMilliseconds())
+
+    message.FromMilliseconds(-1)
+    self.assertEqual('1969-12-31T23:59:59.999Z',
+                     message.ToJsonString())
+    self.assertEqual(-1, message.ToMilliseconds())
+
+    message.FromSeconds(1)
+    self.assertEqual('1970-01-01T00:00:01Z',
+                     message.ToJsonString())
+    self.assertEqual(1, message.ToSeconds())
+
+    message.FromSeconds(-1)
+    self.assertEqual('1969-12-31T23:59:59Z',
+                     message.ToJsonString())
+    self.assertEqual(-1, message.ToSeconds())
+
+    message.FromNanoseconds(1999)
+    self.assertEqual(1, message.ToMicroseconds())
+    # For negative values, Timestamp will be rounded down.
+    # For example, "1969-12-31T23:59:59.5Z" (i.e., -0.5s) rounded to seconds
+    # will be "1969-12-31T23:59:59Z" (i.e., -1s) rather than
+    # "1970-01-01T00:00:00Z" (i.e., 0s).
+    message.FromNanoseconds(-1999)
+    self.assertEqual(-2, message.ToMicroseconds())
+
+  def testDurationIntegerConversion(self):
+    message = duration_pb2.Duration()
+    message.FromNanoseconds(1)
+    self.assertEqual('0.000000001s',
+                     message.ToJsonString())
+    self.assertEqual(1, message.ToNanoseconds())
+
+    message.FromNanoseconds(-1)
+    self.assertEqual('-0.000000001s',
+                     message.ToJsonString())
+    self.assertEqual(-1, message.ToNanoseconds())
+
+    message.FromMicroseconds(1)
+    self.assertEqual('0.000001s',
+                     message.ToJsonString())
+    self.assertEqual(1, message.ToMicroseconds())
+
+    message.FromMicroseconds(-1)
+    self.assertEqual('-0.000001s',
+                     message.ToJsonString())
+    self.assertEqual(-1, message.ToMicroseconds())
+
+    message.FromMilliseconds(1)
+    self.assertEqual('0.001s',
+                     message.ToJsonString())
+    self.assertEqual(1, message.ToMilliseconds())
+
+    message.FromMilliseconds(-1)
+    self.assertEqual('-0.001s',
+                     message.ToJsonString())
+    self.assertEqual(-1, message.ToMilliseconds())
+
+    message.FromSeconds(1)
+    self.assertEqual('1s', message.ToJsonString())
+    self.assertEqual(1, message.ToSeconds())
+
+    message.FromSeconds(-1)
+    self.assertEqual('-1s',
+                     message.ToJsonString())
+    self.assertEqual(-1, message.ToSeconds())
+
+    # Test truncation behavior.
+    message.FromNanoseconds(1999)
+    self.assertEqual(1, message.ToMicroseconds())
+
+    # For negative values, Duration will be rounded towards 0.
+    message.FromNanoseconds(-1999)
+    self.assertEqual(-1, message.ToMicroseconds())
+
+  def testDatetimeConverison(self):
+    message = timestamp_pb2.Timestamp()
+    dt = datetime(1970, 1, 1)
+    message.FromDatetime(dt)
+    self.assertEqual(dt, message.ToDatetime())
+
+    message.FromMilliseconds(1999)
+    self.assertEqual(datetime(1970, 1, 1, 0, 0, 1, 999000),
+                     message.ToDatetime())
+
+  def testTimedeltaConversion(self):
+    message = duration_pb2.Duration()
+    message.FromNanoseconds(1999999999)
+    td = message.ToTimedelta()
+    self.assertEqual(1, td.seconds)
+    self.assertEqual(999999, td.microseconds)
+
+    message.FromNanoseconds(-1999999999)
+    td = message.ToTimedelta()
+    self.assertEqual(-1, td.days)
+    self.assertEqual(86398, td.seconds)
+    self.assertEqual(1, td.microseconds)
+
+    message.FromMicroseconds(-1)
+    td = message.ToTimedelta()
+    self.assertEqual(-1, td.days)
+    self.assertEqual(86399, td.seconds)
+    self.assertEqual(999999, td.microseconds)
+    converted_message = duration_pb2.Duration()
+    converted_message.FromTimedelta(td)
+    self.assertEqual(message, converted_message)
+
+  def testInvalidTimestamp(self):
+    message = timestamp_pb2.Timestamp()
+    self.assertRaisesRegexp(
+        ValueError,
+        'time data \'10000-01-01T00:00:00\' does not match'
+        ' format \'%Y-%m-%dT%H:%M:%S\'',
+        message.FromJsonString, '10000-01-01T00:00:00.00Z')
+    self.assertRaisesRegexp(
+        well_known_types.ParseError,
+        'nanos 0123456789012 more than 9 fractional digits.',
+        message.FromJsonString,
+        '1970-01-01T00:00:00.0123456789012Z')
+    self.assertRaisesRegexp(
+        well_known_types.ParseError,
+        (r'Invalid timezone offset value: \+08.'),
+        message.FromJsonString,
+        '1972-01-01T01:00:00.01+08',)
+    self.assertRaisesRegexp(
+        ValueError,
+        'year is out of range',
+        message.FromJsonString,
+        '0000-01-01T00:00:00Z')
+    message.seconds = 253402300800
+    self.assertRaisesRegexp(
+        OverflowError,
+        'date value out of range',
+        message.ToJsonString)
+
+  def testInvalidDuration(self):
+    message = duration_pb2.Duration()
+    self.assertRaisesRegexp(
+        well_known_types.ParseError,
+        'Duration must end with letter "s": 1.',
+        message.FromJsonString, '1')
+    self.assertRaisesRegexp(
+        well_known_types.ParseError,
+        'Couldn\'t parse duration: 1...2s.',
+        message.FromJsonString, '1...2s')
+
+
+class FieldMaskTest(unittest.TestCase):
+
+  def testStringFormat(self):
+    mask = field_mask_pb2.FieldMask()
+    self.assertEqual('', mask.ToJsonString())
+    mask.paths.append('foo')
+    self.assertEqual('foo', mask.ToJsonString())
+    mask.paths.append('bar')
+    self.assertEqual('foo,bar', mask.ToJsonString())
+
+    mask.FromJsonString('')
+    self.assertEqual('', mask.ToJsonString())
+    mask.FromJsonString('foo')
+    self.assertEqual(['foo'], mask.paths)
+    mask.FromJsonString('foo,bar')
+    self.assertEqual(['foo', 'bar'], mask.paths)
+
+  def testDescriptorToFieldMask(self):
+    mask = field_mask_pb2.FieldMask()
+    msg_descriptor = unittest_pb2.TestAllTypes.DESCRIPTOR
+    mask.AllFieldsFromDescriptor(msg_descriptor)
+    self.assertEqual(75, len(mask.paths))
+    self.assertTrue(mask.IsValidForDescriptor(msg_descriptor))
+    for field in msg_descriptor.fields:
+      self.assertTrue(field.name in mask.paths)
+    mask.paths.append('optional_nested_message.bb')
+    self.assertTrue(mask.IsValidForDescriptor(msg_descriptor))
+    mask.paths.append('repeated_nested_message.bb')
+    self.assertFalse(mask.IsValidForDescriptor(msg_descriptor))
+
+  def testCanonicalFrom(self):
+    mask = field_mask_pb2.FieldMask()
+    out_mask = field_mask_pb2.FieldMask()
+    # Paths will be sorted.
+    mask.FromJsonString('baz.quz,bar,foo')
+    out_mask.CanonicalFormFromMask(mask)
+    self.assertEqual('bar,baz.quz,foo', out_mask.ToJsonString())
+    # Duplicated paths will be removed.
+    mask.FromJsonString('foo,bar,foo')
+    out_mask.CanonicalFormFromMask(mask)
+    self.assertEqual('bar,foo', out_mask.ToJsonString())
+    # Sub-paths of other paths will be removed.
+    mask.FromJsonString('foo.b1,bar.b1,foo.b2,bar')
+    out_mask.CanonicalFormFromMask(mask)
+    self.assertEqual('bar,foo.b1,foo.b2', out_mask.ToJsonString())
+
+    # Test more deeply nested cases.
+    mask.FromJsonString(
+        'foo.bar.baz1,foo.bar.baz2.quz,foo.bar.baz2')
+    out_mask.CanonicalFormFromMask(mask)
+    self.assertEqual('foo.bar.baz1,foo.bar.baz2',
+                     out_mask.ToJsonString())
+    mask.FromJsonString(
+        'foo.bar.baz1,foo.bar.baz2,foo.bar.baz2.quz')
+    out_mask.CanonicalFormFromMask(mask)
+    self.assertEqual('foo.bar.baz1,foo.bar.baz2',
+                     out_mask.ToJsonString())
+    mask.FromJsonString(
+        'foo.bar.baz1,foo.bar.baz2,foo.bar.baz2.quz,foo.bar')
+    out_mask.CanonicalFormFromMask(mask)
+    self.assertEqual('foo.bar', out_mask.ToJsonString())
+    mask.FromJsonString(
+        'foo.bar.baz1,foo.bar.baz2,foo.bar.baz2.quz,foo')
+    out_mask.CanonicalFormFromMask(mask)
+    self.assertEqual('foo', out_mask.ToJsonString())
+
+  def testUnion(self):
+    mask1 = field_mask_pb2.FieldMask()
+    mask2 = field_mask_pb2.FieldMask()
+    out_mask = field_mask_pb2.FieldMask()
+    mask1.FromJsonString('foo,baz')
+    mask2.FromJsonString('bar,quz')
+    out_mask.Union(mask1, mask2)
+    self.assertEqual('bar,baz,foo,quz', out_mask.ToJsonString())
+    # Overlap with duplicated paths.
+    mask1.FromJsonString('foo,baz.bb')
+    mask2.FromJsonString('baz.bb,quz')
+    out_mask.Union(mask1, mask2)
+    self.assertEqual('baz.bb,foo,quz', out_mask.ToJsonString())
+    # Overlap with paths covering some other paths.
+    mask1.FromJsonString('foo.bar.baz,quz')
+    mask2.FromJsonString('foo.bar,bar')
+    out_mask.Union(mask1, mask2)
+    self.assertEqual('bar,foo.bar,quz', out_mask.ToJsonString())
+
+  def testIntersect(self):
+    mask1 = field_mask_pb2.FieldMask()
+    mask2 = field_mask_pb2.FieldMask()
+    out_mask = field_mask_pb2.FieldMask()
+    # Test cases without overlapping.
+    mask1.FromJsonString('foo,baz')
+    mask2.FromJsonString('bar,quz')
+    out_mask.Intersect(mask1, mask2)
+    self.assertEqual('', out_mask.ToJsonString())
+    # Overlap with duplicated paths.
+    mask1.FromJsonString('foo,baz.bb')
+    mask2.FromJsonString('baz.bb,quz')
+    out_mask.Intersect(mask1, mask2)
+    self.assertEqual('baz.bb', out_mask.ToJsonString())
+    # Overlap with paths covering some other paths.
+    mask1.FromJsonString('foo.bar.baz,quz')
+    mask2.FromJsonString('foo.bar,bar')
+    out_mask.Intersect(mask1, mask2)
+    self.assertEqual('foo.bar.baz', out_mask.ToJsonString())
+    mask1.FromJsonString('foo.bar,bar')
+    mask2.FromJsonString('foo.bar.baz,quz')
+    out_mask.Intersect(mask1, mask2)
+    self.assertEqual('foo.bar.baz', out_mask.ToJsonString())
+
+  def testMergeMessage(self):
+    # Test merge one field.
+    src = unittest_pb2.TestAllTypes()
+    test_util.SetAllFields(src)
+    for field in src.DESCRIPTOR.fields:
+      if field.containing_oneof:
+        continue
+      field_name = field.name
+      dst = unittest_pb2.TestAllTypes()
+      # Only set one path to mask.
+      mask = field_mask_pb2.FieldMask()
+      mask.paths.append(field_name)
+      mask.MergeMessage(src, dst)
+      # The expected result message.
+      msg = unittest_pb2.TestAllTypes()
+      if field.label == descriptor.FieldDescriptor.LABEL_REPEATED:
+        repeated_src = getattr(src, field_name)
+        repeated_msg = getattr(msg, field_name)
+        if field.cpp_type == descriptor.FieldDescriptor.CPPTYPE_MESSAGE:
+          for item in repeated_src:
+            repeated_msg.add().CopyFrom(item)
+        else:
+          repeated_msg.extend(repeated_src)
+      elif field.cpp_type == descriptor.FieldDescriptor.CPPTYPE_MESSAGE:
+        getattr(msg, field_name).CopyFrom(getattr(src, field_name))
+      else:
+        setattr(msg, field_name, getattr(src, field_name))
+      # Only field specified in mask is merged.
+      self.assertEqual(msg, dst)
+
+    # Test merge nested fields.
+    nested_src = unittest_pb2.NestedTestAllTypes()
+    nested_dst = unittest_pb2.NestedTestAllTypes()
+    nested_src.child.payload.optional_int32 = 1234
+    nested_src.child.child.payload.optional_int32 = 5678
+    mask = field_mask_pb2.FieldMask()
+    mask.FromJsonString('child.payload')
+    mask.MergeMessage(nested_src, nested_dst)
+    self.assertEqual(1234, nested_dst.child.payload.optional_int32)
+    self.assertEqual(0, nested_dst.child.child.payload.optional_int32)
+
+    mask.FromJsonString('child.child.payload')
+    mask.MergeMessage(nested_src, nested_dst)
+    self.assertEqual(1234, nested_dst.child.payload.optional_int32)
+    self.assertEqual(5678, nested_dst.child.child.payload.optional_int32)
+
+    nested_dst.Clear()
+    mask.FromJsonString('child.child.payload')
+    mask.MergeMessage(nested_src, nested_dst)
+    self.assertEqual(0, nested_dst.child.payload.optional_int32)
+    self.assertEqual(5678, nested_dst.child.child.payload.optional_int32)
+
+    nested_dst.Clear()
+    mask.FromJsonString('child')
+    mask.MergeMessage(nested_src, nested_dst)
+    self.assertEqual(1234, nested_dst.child.payload.optional_int32)
+    self.assertEqual(5678, nested_dst.child.child.payload.optional_int32)
+
+    # Test MergeOptions.
+    nested_dst.Clear()
+    nested_dst.child.payload.optional_int64 = 4321
+    # Message fields will be merged by default.
+    mask.FromJsonString('child.payload')
+    mask.MergeMessage(nested_src, nested_dst)
+    self.assertEqual(1234, nested_dst.child.payload.optional_int32)
+    self.assertEqual(4321, nested_dst.child.payload.optional_int64)
+    # Change the behavior to replace message fields.
+    mask.FromJsonString('child.payload')
+    mask.MergeMessage(nested_src, nested_dst, True, False)
+    self.assertEqual(1234, nested_dst.child.payload.optional_int32)
+    self.assertEqual(0, nested_dst.child.payload.optional_int64)
+
+    # By default, fields missing in source are not cleared in destination.
+    nested_dst.payload.optional_int32 = 1234
+    self.assertTrue(nested_dst.HasField('payload'))
+    mask.FromJsonString('payload')
+    mask.MergeMessage(nested_src, nested_dst)
+    self.assertTrue(nested_dst.HasField('payload'))
+    # But they are cleared when replacing message fields.
+    nested_dst.Clear()
+    nested_dst.payload.optional_int32 = 1234
+    mask.FromJsonString('payload')
+    mask.MergeMessage(nested_src, nested_dst, True, False)
+    self.assertFalse(nested_dst.HasField('payload'))
+
+    nested_src.payload.repeated_int32.append(1234)
+    nested_dst.payload.repeated_int32.append(5678)
+    # Repeated fields will be appended by default.
+    mask.FromJsonString('payload.repeated_int32')
+    mask.MergeMessage(nested_src, nested_dst)
+    self.assertEqual(2, len(nested_dst.payload.repeated_int32))
+    self.assertEqual(5678, nested_dst.payload.repeated_int32[0])
+    self.assertEqual(1234, nested_dst.payload.repeated_int32[1])
+    # Change the behavior to replace repeated fields.
+    mask.FromJsonString('payload.repeated_int32')
+    mask.MergeMessage(nested_src, nested_dst, False, True)
+    self.assertEqual(1, len(nested_dst.payload.repeated_int32))
+    self.assertEqual(1234, nested_dst.payload.repeated_int32[0])
+
+
+class StructTest(unittest.TestCase):
+
+  def testStruct(self):
+    struct = struct_pb2.Struct()
+    struct_class = struct.__class__
+
+    struct['key1'] = 5
+    struct['key2'] = 'abc'
+    struct['key3'] = True
+    struct.get_or_create_struct('key4')['subkey'] = 11.0
+    struct_list = struct.get_or_create_list('key5')
+    struct_list.extend([6, 'seven', True, False, None])
+    struct_list.add_struct()['subkey2'] = 9
+
+    self.assertTrue(isinstance(struct, well_known_types.Struct))
+    self.assertEquals(5, struct['key1'])
+    self.assertEquals('abc', struct['key2'])
+    self.assertIs(True, struct['key3'])
+    self.assertEquals(11, struct['key4']['subkey'])
+    inner_struct = struct_class()
+    inner_struct['subkey2'] = 9
+    self.assertEquals([6, 'seven', True, False, None, inner_struct],
+                      list(struct['key5'].items()))
+
+    serialized = struct.SerializeToString()
+
+    struct2 = struct_pb2.Struct()
+    struct2.ParseFromString(serialized)
+
+    self.assertEquals(struct, struct2)
+
+    self.assertTrue(isinstance(struct2, well_known_types.Struct))
+    self.assertEquals(5, struct2['key1'])
+    self.assertEquals('abc', struct2['key2'])
+    self.assertIs(True, struct2['key3'])
+    self.assertEquals(11, struct2['key4']['subkey'])
+    self.assertEquals([6, 'seven', True, False, None, inner_struct],
+                      list(struct2['key5'].items()))
+
+    struct_list = struct2['key5']
+    self.assertEquals(6, struct_list[0])
+    self.assertEquals('seven', struct_list[1])
+    self.assertEquals(True, struct_list[2])
+    self.assertEquals(False, struct_list[3])
+    self.assertEquals(None, struct_list[4])
+    self.assertEquals(inner_struct, struct_list[5])
+
+    struct_list[1] = 7
+    self.assertEquals(7, struct_list[1])
+
+    struct_list.add_list().extend([1, 'two', True, False, None])
+    self.assertEquals([1, 'two', True, False, None],
+                      list(struct_list[6].items()))
+
+    text_serialized = str(struct)
+    struct3 = struct_pb2.Struct()
+    text_format.Merge(text_serialized, struct3)
+    self.assertEquals(struct, struct3)
+
+    struct.get_or_create_struct('key3')['replace'] = 12
+    self.assertEquals(12, struct['key3']['replace'])
+
+
+class AnyTest(unittest.TestCase):
+
+  def testAnyMessage(self):
+    # Creates and sets message.
+    msg = any_test_pb2.TestAny()
+    msg_descriptor = msg.DESCRIPTOR
+    all_types = unittest_pb2.TestAllTypes()
+    all_descriptor = all_types.DESCRIPTOR
+    all_types.repeated_string.append(u'\u00fc\ua71f')
+    # Packs to Any.
+    msg.value.Pack(all_types)
+    self.assertEqual(msg.value.type_url,
+                     'type.googleapis.com/%s' % all_descriptor.full_name)
+    self.assertEqual(msg.value.value,
+                     all_types.SerializeToString())
+    # Tests Is() method.
+    self.assertTrue(msg.value.Is(all_descriptor))
+    self.assertFalse(msg.value.Is(msg_descriptor))
+    # Unpacks Any.
+    unpacked_message = unittest_pb2.TestAllTypes()
+    self.assertTrue(msg.value.Unpack(unpacked_message))
+    self.assertEqual(all_types, unpacked_message)
+    # Unpacks to different type.
+    self.assertFalse(msg.value.Unpack(msg))
+    # Only Any messages have Pack method.
+    try:
+      msg.Pack(all_types)
+    except AttributeError:
+      pass
+    else:
+      raise AttributeError('%s should not have Pack method.' %
+                           msg_descriptor.full_name)
+
+  def testPackWithCustomTypeUrl(self):
+    submessage = any_test_pb2.TestAny()
+    submessage.int_value = 12345
+    msg = any_pb2.Any()
+    # Pack with a custom type URL prefix.
+    msg.Pack(submessage, 'type.myservice.com')
+    self.assertEqual(msg.type_url,
+                     'type.myservice.com/%s' % submessage.DESCRIPTOR.full_name)
+    # Pack with a custom type URL prefix ending with '/'.
+    msg.Pack(submessage, 'type.myservice.com/')
+    self.assertEqual(msg.type_url,
+                     'type.myservice.com/%s' % submessage.DESCRIPTOR.full_name)
+    # Pack with an empty type URL prefix.
+    msg.Pack(submessage, '')
+    self.assertEqual(msg.type_url,
+                     '/%s' % submessage.DESCRIPTOR.full_name)
+    # Test unpacking the type.
+    unpacked_message = any_test_pb2.TestAny()
+    self.assertTrue(msg.Unpack(unpacked_message))
+    self.assertEqual(submessage, unpacked_message)
+
+
+if __name__ == '__main__':
+  unittest.main()
diff --git a/python/google/protobuf/internal/wire_format.py b/python/google/protobuf/internal/wire_format.py
new file mode 100755
index 0000000..883f525
--- /dev/null
+++ b/python/google/protobuf/internal/wire_format.py
@@ -0,0 +1,268 @@
+# Protocol Buffers - Google's data interchange format
+# Copyright 2008 Google Inc.  All rights reserved.
+# https://developers.google.com/protocol-buffers/
+#
+# Redistribution and use in source and binary forms, with or without
+# modification, are permitted provided that the following conditions are
+# met:
+#
+#     * Redistributions of source code must retain the above copyright
+# notice, this list of conditions and the following disclaimer.
+#     * Redistributions in binary form must reproduce the above
+# copyright notice, this list of conditions and the following disclaimer
+# in the documentation and/or other materials provided with the
+# distribution.
+#     * Neither the name of Google Inc. nor the names of its
+# contributors may be used to endorse or promote products derived from
+# this software without specific prior written permission.
+#
+# THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+# "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
+# LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
+# A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
+# OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
+# SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
+# LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
+# DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
+# THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
+# (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+# OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+
+"""Constants and static functions to support protocol buffer wire format."""
+
+__author__ = 'robinson@google.com (Will Robinson)'
+
+import struct
+from google.protobuf import descriptor
+from google.protobuf import message
+
+
+TAG_TYPE_BITS = 3  # Number of bits used to hold type info in a proto tag.
+TAG_TYPE_MASK = (1 << TAG_TYPE_BITS) - 1  # 0x7
+
+# These numbers identify the wire type of a protocol buffer value.
+# We use the least-significant TAG_TYPE_BITS bits of the varint-encoded
+# tag-and-type to store one of these WIRETYPE_* constants.
+# These values must match WireType enum in google/protobuf/wire_format.h.
+WIRETYPE_VARINT = 0
+WIRETYPE_FIXED64 = 1
+WIRETYPE_LENGTH_DELIMITED = 2
+WIRETYPE_START_GROUP = 3
+WIRETYPE_END_GROUP = 4
+WIRETYPE_FIXED32 = 5
+_WIRETYPE_MAX = 5
+
+
+# Bounds for various integer types.
+INT32_MAX = int((1 << 31) - 1)
+INT32_MIN = int(-(1 << 31))
+UINT32_MAX = (1 << 32) - 1
+
+INT64_MAX = (1 << 63) - 1
+INT64_MIN = -(1 << 63)
+UINT64_MAX = (1 << 64) - 1
+
+# "struct" format strings that will encode/decode the specified formats.
+FORMAT_UINT32_LITTLE_ENDIAN = '<I'
+FORMAT_UINT64_LITTLE_ENDIAN = '<Q'
+FORMAT_FLOAT_LITTLE_ENDIAN = '<f'
+FORMAT_DOUBLE_LITTLE_ENDIAN = '<d'
+
+
+# We'll have to provide alternate implementations of AppendLittleEndian*() on
+# any architectures where these checks fail.
+if struct.calcsize(FORMAT_UINT32_LITTLE_ENDIAN) != 4:
+  raise AssertionError('Format "I" is not a 32-bit number.')
+if struct.calcsize(FORMAT_UINT64_LITTLE_ENDIAN) != 8:
+  raise AssertionError('Format "Q" is not a 64-bit number.')
+
+
+def PackTag(field_number, wire_type):
+  """Returns an unsigned 32-bit integer that encodes the field number and
+  wire type information in standard protocol message wire format.
+
+  Args:
+    field_number: Expected to be an integer in the range [1, 1 << 29)
+    wire_type: One of the WIRETYPE_* constants.
+  """
+  if not 0 <= wire_type <= _WIRETYPE_MAX:
+    raise message.EncodeError('Unknown wire type: %d' % wire_type)
+  return (field_number << TAG_TYPE_BITS) | wire_type
+
+
+def UnpackTag(tag):
+  """The inverse of PackTag().  Given an unsigned 32-bit number,
+  returns a (field_number, wire_type) tuple.
+  """
+  return (tag >> TAG_TYPE_BITS), (tag & TAG_TYPE_MASK)
+
+
+def ZigZagEncode(value):
+  """ZigZag Transform:  Encodes signed integers so that they can be
+  effectively used with varint encoding.  See wire_format.h for
+  more details.
+  """
+  if value >= 0:
+    return value << 1
+  return (value << 1) ^ (~0)
+
+
+def ZigZagDecode(value):
+  """Inverse of ZigZagEncode()."""
+  if not value & 0x1:
+    return value >> 1
+  return (value >> 1) ^ (~0)
+
+
+
+# The *ByteSize() functions below return the number of bytes required to
+# serialize "field number + type" information and then serialize the value.
+
+
+def Int32ByteSize(field_number, int32):
+  return Int64ByteSize(field_number, int32)
+
+
+def Int32ByteSizeNoTag(int32):
+  return _VarUInt64ByteSizeNoTag(0xffffffffffffffff & int32)
+
+
+def Int64ByteSize(field_number, int64):
+  # Have to convert to uint before calling UInt64ByteSize().
+  return UInt64ByteSize(field_number, 0xffffffffffffffff & int64)
+
+
+def UInt32ByteSize(field_number, uint32):
+  return UInt64ByteSize(field_number, uint32)
+
+
+def UInt64ByteSize(field_number, uint64):
+  return TagByteSize(field_number) + _VarUInt64ByteSizeNoTag(uint64)
+
+
+def SInt32ByteSize(field_number, int32):
+  return UInt32ByteSize(field_number, ZigZagEncode(int32))
+
+
+def SInt64ByteSize(field_number, int64):
+  return UInt64ByteSize(field_number, ZigZagEncode(int64))
+
+
+def Fixed32ByteSize(field_number, fixed32):
+  return TagByteSize(field_number) + 4
+
+
+def Fixed64ByteSize(field_number, fixed64):
+  return TagByteSize(field_number) + 8
+
+
+def SFixed32ByteSize(field_number, sfixed32):
+  return TagByteSize(field_number) + 4
+
+
+def SFixed64ByteSize(field_number, sfixed64):
+  return TagByteSize(field_number) + 8
+
+
+def FloatByteSize(field_number, flt):
+  return TagByteSize(field_number) + 4
+
+
+def DoubleByteSize(field_number, double):
+  return TagByteSize(field_number) + 8
+
+
+def BoolByteSize(field_number, b):
+  return TagByteSize(field_number) + 1
+
+
+def EnumByteSize(field_number, enum):
+  return UInt32ByteSize(field_number, enum)
+
+
+def StringByteSize(field_number, string):
+  return BytesByteSize(field_number, string.encode('utf-8'))
+
+
+def BytesByteSize(field_number, b):
+  return (TagByteSize(field_number)
+          + _VarUInt64ByteSizeNoTag(len(b))
+          + len(b))
+
+
+def GroupByteSize(field_number, message):
+  return (2 * TagByteSize(field_number)  # START and END group.
+          + message.ByteSize())
+
+
+def MessageByteSize(field_number, message):
+  return (TagByteSize(field_number)
+          + _VarUInt64ByteSizeNoTag(message.ByteSize())
+          + message.ByteSize())
+
+
+def MessageSetItemByteSize(field_number, msg):
+  # First compute the sizes of the tags.
+  # There are 2 tags for the beginning and ending of the repeated group, that
+  # is field number 1, one with field number 2 (type_id) and one with field
+  # number 3 (message).
+  total_size = (2 * TagByteSize(1) + TagByteSize(2) + TagByteSize(3))
+
+  # Add the number of bytes for type_id.
+  total_size += _VarUInt64ByteSizeNoTag(field_number)
+
+  message_size = msg.ByteSize()
+
+  # The number of bytes for encoding the length of the message.
+  total_size += _VarUInt64ByteSizeNoTag(message_size)
+
+  # The size of the message.
+  total_size += message_size
+  return total_size
+
+
+def TagByteSize(field_number):
+  """Returns the bytes required to serialize a tag with this field number."""
+  # Just pass in type 0, since the type won't affect the tag+type size.
+  return _VarUInt64ByteSizeNoTag(PackTag(field_number, 0))
+
+
+# Private helper function for the *ByteSize() functions above.
+
+def _VarUInt64ByteSizeNoTag(uint64):
+  """Returns the number of bytes required to serialize a single varint
+  using boundary value comparisons. (unrolled loop optimization -WPierce)
+  uint64 must be unsigned.
+  """
+  if uint64 <= 0x7f: return 1
+  if uint64 <= 0x3fff: return 2
+  if uint64 <= 0x1fffff: return 3
+  if uint64 <= 0xfffffff: return 4
+  if uint64 <= 0x7ffffffff: return 5
+  if uint64 <= 0x3ffffffffff: return 6
+  if uint64 <= 0x1ffffffffffff: return 7
+  if uint64 <= 0xffffffffffffff: return 8
+  if uint64 <= 0x7fffffffffffffff: return 9
+  if uint64 > UINT64_MAX:
+    raise message.EncodeError('Value out of range: %d' % uint64)
+  return 10
+
+
+NON_PACKABLE_TYPES = (
+  descriptor.FieldDescriptor.TYPE_STRING,
+  descriptor.FieldDescriptor.TYPE_GROUP,
+  descriptor.FieldDescriptor.TYPE_MESSAGE,
+  descriptor.FieldDescriptor.TYPE_BYTES
+)
+
+
+def IsTypePackable(field_type):
+  """Return true iff packable = true is valid for fields of this type.
+
+  Args:
+    field_type: a FieldDescriptor::Type value.
+
+  Returns:
+    True iff fields of this type are packable.
+  """
+  return field_type not in NON_PACKABLE_TYPES
diff --git a/python/google/protobuf/internal/wire_format_test.py b/python/google/protobuf/internal/wire_format_test.py
new file mode 100755
index 0000000..f659d18
--- /dev/null
+++ b/python/google/protobuf/internal/wire_format_test.py
@@ -0,0 +1,256 @@
+#! /usr/bin/env python
+#
+# Protocol Buffers - Google's data interchange format
+# Copyright 2008 Google Inc.  All rights reserved.
+# https://developers.google.com/protocol-buffers/
+#
+# Redistribution and use in source and binary forms, with or without
+# modification, are permitted provided that the following conditions are
+# met:
+#
+#     * Redistributions of source code must retain the above copyright
+# notice, this list of conditions and the following disclaimer.
+#     * Redistributions in binary form must reproduce the above
+# copyright notice, this list of conditions and the following disclaimer
+# in the documentation and/or other materials provided with the
+# distribution.
+#     * Neither the name of Google Inc. nor the names of its
+# contributors may be used to endorse or promote products derived from
+# this software without specific prior written permission.
+#
+# THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+# "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
+# LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
+# A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
+# OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
+# SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
+# LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
+# DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
+# THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
+# (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+# OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+
+"""Test for google.protobuf.internal.wire_format."""
+
+__author__ = 'robinson@google.com (Will Robinson)'
+
+try:
+  import unittest2 as unittest
+except ImportError:
+  import unittest
+from google.protobuf import message
+from google.protobuf.internal import wire_format
+
+
+class WireFormatTest(unittest.TestCase):
+
+  def testPackTag(self):
+    field_number = 0xabc
+    tag_type = 2
+    self.assertEqual((field_number << 3) | tag_type,
+                     wire_format.PackTag(field_number, tag_type))
+    PackTag = wire_format.PackTag
+    # Number too high.
+    self.assertRaises(message.EncodeError, PackTag, field_number, 6)
+    # Number too low.
+    self.assertRaises(message.EncodeError, PackTag, field_number, -1)
+
+  def testUnpackTag(self):
+    # Test field numbers that will require various varint sizes.
+    for expected_field_number in (1, 15, 16, 2047, 2048):
+      for expected_wire_type in range(6):  # Highest-numbered wiretype is 5.
+        field_number, wire_type = wire_format.UnpackTag(
+            wire_format.PackTag(expected_field_number, expected_wire_type))
+        self.assertEqual(expected_field_number, field_number)
+        self.assertEqual(expected_wire_type, wire_type)
+
+    self.assertRaises(TypeError, wire_format.UnpackTag, None)
+    self.assertRaises(TypeError, wire_format.UnpackTag, 'abc')
+    self.assertRaises(TypeError, wire_format.UnpackTag, 0.0)
+    self.assertRaises(TypeError, wire_format.UnpackTag, object())
+
+  def testZigZagEncode(self):
+    Z = wire_format.ZigZagEncode
+    self.assertEqual(0, Z(0))
+    self.assertEqual(1, Z(-1))
+    self.assertEqual(2, Z(1))
+    self.assertEqual(3, Z(-2))
+    self.assertEqual(4, Z(2))
+    self.assertEqual(0xfffffffe, Z(0x7fffffff))
+    self.assertEqual(0xffffffff, Z(-0x80000000))
+    self.assertEqual(0xfffffffffffffffe, Z(0x7fffffffffffffff))
+    self.assertEqual(0xffffffffffffffff, Z(-0x8000000000000000))
+
+    self.assertRaises(TypeError, Z, None)
+    self.assertRaises(TypeError, Z, 'abcd')
+    self.assertRaises(TypeError, Z, 0.0)
+    self.assertRaises(TypeError, Z, object())
+
+  def testZigZagDecode(self):
+    Z = wire_format.ZigZagDecode
+    self.assertEqual(0, Z(0))
+    self.assertEqual(-1, Z(1))
+    self.assertEqual(1, Z(2))
+    self.assertEqual(-2, Z(3))
+    self.assertEqual(2, Z(4))
+    self.assertEqual(0x7fffffff, Z(0xfffffffe))
+    self.assertEqual(-0x80000000, Z(0xffffffff))
+    self.assertEqual(0x7fffffffffffffff, Z(0xfffffffffffffffe))
+    self.assertEqual(-0x8000000000000000, Z(0xffffffffffffffff))
+
+    self.assertRaises(TypeError, Z, None)
+    self.assertRaises(TypeError, Z, 'abcd')
+    self.assertRaises(TypeError, Z, 0.0)
+    self.assertRaises(TypeError, Z, object())
+
+  def NumericByteSizeTestHelper(self, byte_size_fn, value, expected_value_size):
+    # Use field numbers that cause various byte sizes for the tag information.
+    for field_number, tag_bytes in ((15, 1), (16, 2), (2047, 2), (2048, 3)):
+      expected_size = expected_value_size + tag_bytes
+      actual_size = byte_size_fn(field_number, value)
+      self.assertEqual(expected_size, actual_size,
+                       'byte_size_fn: %s, field_number: %d, value: %r\n'
+                       'Expected: %d, Actual: %d'% (
+          byte_size_fn, field_number, value, expected_size, actual_size))
+
+  def testByteSizeFunctions(self):
+    # Test all numeric *ByteSize() functions.
+    NUMERIC_ARGS = [
+        # Int32ByteSize().
+        [wire_format.Int32ByteSize, 0, 1],
+        [wire_format.Int32ByteSize, 127, 1],
+        [wire_format.Int32ByteSize, 128, 2],
+        [wire_format.Int32ByteSize, -1, 10],
+        # Int64ByteSize().
+        [wire_format.Int64ByteSize, 0, 1],
+        [wire_format.Int64ByteSize, 127, 1],
+        [wire_format.Int64ByteSize, 128, 2],
+        [wire_format.Int64ByteSize, -1, 10],
+        # UInt32ByteSize().
+        [wire_format.UInt32ByteSize, 0, 1],
+        [wire_format.UInt32ByteSize, 127, 1],
+        [wire_format.UInt32ByteSize, 128, 2],
+        [wire_format.UInt32ByteSize, wire_format.UINT32_MAX, 5],
+        # UInt64ByteSize().
+        [wire_format.UInt64ByteSize, 0, 1],
+        [wire_format.UInt64ByteSize, 127, 1],
+        [wire_format.UInt64ByteSize, 128, 2],
+        [wire_format.UInt64ByteSize, wire_format.UINT64_MAX, 10],
+        # SInt32ByteSize().
+        [wire_format.SInt32ByteSize, 0, 1],
+        [wire_format.SInt32ByteSize, -1, 1],
+        [wire_format.SInt32ByteSize, 1, 1],
+        [wire_format.SInt32ByteSize, -63, 1],
+        [wire_format.SInt32ByteSize, 63, 1],
+        [wire_format.SInt32ByteSize, -64, 1],
+        [wire_format.SInt32ByteSize, 64, 2],
+        # SInt64ByteSize().
+        [wire_format.SInt64ByteSize, 0, 1],
+        [wire_format.SInt64ByteSize, -1, 1],
+        [wire_format.SInt64ByteSize, 1, 1],
+        [wire_format.SInt64ByteSize, -63, 1],
+        [wire_format.SInt64ByteSize, 63, 1],
+        [wire_format.SInt64ByteSize, -64, 1],
+        [wire_format.SInt64ByteSize, 64, 2],
+        # Fixed32ByteSize().
+        [wire_format.Fixed32ByteSize, 0, 4],
+        [wire_format.Fixed32ByteSize, wire_format.UINT32_MAX, 4],
+        # Fixed64ByteSize().
+        [wire_format.Fixed64ByteSize, 0, 8],
+        [wire_format.Fixed64ByteSize, wire_format.UINT64_MAX, 8],
+        # SFixed32ByteSize().
+        [wire_format.SFixed32ByteSize, 0, 4],
+        [wire_format.SFixed32ByteSize, wire_format.INT32_MIN, 4],
+        [wire_format.SFixed32ByteSize, wire_format.INT32_MAX, 4],
+        # SFixed64ByteSize().
+        [wire_format.SFixed64ByteSize, 0, 8],
+        [wire_format.SFixed64ByteSize, wire_format.INT64_MIN, 8],
+        [wire_format.SFixed64ByteSize, wire_format.INT64_MAX, 8],
+        # FloatByteSize().
+        [wire_format.FloatByteSize, 0.0, 4],
+        [wire_format.FloatByteSize, 1000000000.0, 4],
+        [wire_format.FloatByteSize, -1000000000.0, 4],
+        # DoubleByteSize().
+        [wire_format.DoubleByteSize, 0.0, 8],
+        [wire_format.DoubleByteSize, 1000000000.0, 8],
+        [wire_format.DoubleByteSize, -1000000000.0, 8],
+        # BoolByteSize().
+        [wire_format.BoolByteSize, False, 1],
+        [wire_format.BoolByteSize, True, 1],
+        # EnumByteSize().
+        [wire_format.EnumByteSize, 0, 1],
+        [wire_format.EnumByteSize, 127, 1],
+        [wire_format.EnumByteSize, 128, 2],
+        [wire_format.EnumByteSize, wire_format.UINT32_MAX, 5],
+        ]
+    for args in NUMERIC_ARGS:
+      self.NumericByteSizeTestHelper(*args)
+
+    # Test strings and bytes.
+    for byte_size_fn in (wire_format.StringByteSize, wire_format.BytesByteSize):
+      # 1 byte for tag, 1 byte for length, 3 bytes for contents.
+      self.assertEqual(5, byte_size_fn(10, 'abc'))
+      # 2 bytes for tag, 1 byte for length, 3 bytes for contents.
+      self.assertEqual(6, byte_size_fn(16, 'abc'))
+      # 2 bytes for tag, 2 bytes for length, 128 bytes for contents.
+      self.assertEqual(132, byte_size_fn(16, 'a' * 128))
+
+    # Test UTF-8 string byte size calculation.
+    # 1 byte for tag, 1 byte for length, 8 bytes for content.
+    self.assertEqual(10, wire_format.StringByteSize(
+        5, b'\xd0\xa2\xd0\xb5\xd1\x81\xd1\x82'.decode('utf-8')))
+
+    class MockMessage(object):
+      def __init__(self, byte_size):
+        self.byte_size = byte_size
+      def ByteSize(self):
+        return self.byte_size
+
+    message_byte_size = 10
+    mock_message = MockMessage(byte_size=message_byte_size)
+    # Test groups.
+    # (2 * 1) bytes for begin and end tags, plus message_byte_size.
+    self.assertEqual(2 + message_byte_size,
+                     wire_format.GroupByteSize(1, mock_message))
+    # (2 * 2) bytes for begin and end tags, plus message_byte_size.
+    self.assertEqual(4 + message_byte_size,
+                     wire_format.GroupByteSize(16, mock_message))
+
+    # Test messages.
+    # 1 byte for tag, plus 1 byte for length, plus contents.
+    self.assertEqual(2 + mock_message.byte_size,
+                     wire_format.MessageByteSize(1, mock_message))
+    # 2 bytes for tag, plus 1 byte for length, plus contents.
+    self.assertEqual(3 + mock_message.byte_size,
+                     wire_format.MessageByteSize(16, mock_message))
+    # 2 bytes for tag, plus 2 bytes for length, plus contents.
+    mock_message.byte_size = 128
+    self.assertEqual(4 + mock_message.byte_size,
+                     wire_format.MessageByteSize(16, mock_message))
+
+
+    # Test message set item byte size.
+    # 4 bytes for tags, plus 1 byte for length, plus 1 byte for type_id,
+    # plus contents.
+    mock_message.byte_size = 10
+    self.assertEqual(mock_message.byte_size + 6,
+                     wire_format.MessageSetItemByteSize(1, mock_message))
+
+    # 4 bytes for tags, plus 2 bytes for length, plus 1 byte for type_id,
+    # plus contents.
+    mock_message.byte_size = 128
+    self.assertEqual(mock_message.byte_size + 7,
+                     wire_format.MessageSetItemByteSize(1, mock_message))
+
+    # 4 bytes for tags, plus 2 bytes for length, plus 2 byte for type_id,
+    # plus contents.
+    self.assertEqual(mock_message.byte_size + 8,
+                     wire_format.MessageSetItemByteSize(128, mock_message))
+
+    # Too-long varint.
+    self.assertRaises(message.EncodeError,
+                      wire_format.UInt64ByteSize, 1, 1 << 128)
+
+
+if __name__ == '__main__':
+  unittest.main()