Add protobuf.bzl and Python protobuf sources at 48ebb29a8ec118bf6b9ee39f6be42b57321c099a to third_party/protobuf.

86 files changed, 34767 insertions, 0 deletions

diff --git a/third_party/protobuf/protobuf.bzl b/third_party/protobuf/protobuf.bzl
new file mode 100644
index 0000000000..71eaba22f0
--- /dev/null
+++ b/third_party/protobuf/protobuf.bzl
@@ -0,0 +1,279 @@
+# -*- mode: python; -*- PYTHON-PREPROCESSING-REQUIRED
+
+def _GetPath(ctx, path):
+  if ctx.label.workspace_root:
+    return ctx.label.workspace_root + '/' + path
+  else:
+    return path
+
+def _GenDir(ctx):
+  if not ctx.attr.includes:
+    return ctx.label.workspace_root
+  if not ctx.attr.includes[0]:
+    return _GetPath(ctx, ctx.label.package)
+  if not ctx.label.package:
+    return _GetPath(ctx, ctx.attr.includes[0])
+  return _GetPath(ctx, ctx.label.package + '/' + ctx.attr.includes[0])
+
+def _CcOuts(srcs, use_grpc_plugin=False):
+  ret = [s[:-len(".proto")] + ".pb.h" for s in srcs] + \
+        [s[:-len(".proto")] + ".pb.cc" for s in srcs]
+  if use_grpc_plugin:
+    ret += [s[:-len(".proto")] + ".grpc.pb.h" for s in srcs] + \
+           [s[:-len(".proto")] + ".grpc.pb.cc" for s in srcs]
+  return ret
+
+def _PyOuts(srcs):
+  return [s[:-len(".proto")] + "_pb2.py" for s in srcs]
+
+def _RelativeOutputPath(path, include):
+  if include == None:
+    return path
+
+  if not path.startswith(include):
+    fail("Include path %s isn't part of the path %s." % (include, path))
+
+  if include and include[-1] != '/':
+    include = include + '/'
+
+  path = path[len(include):]
+
+  if not path.startswith(PACKAGE_NAME):
+    fail("The package %s is not within the path %s" % (PACKAGE_NAME, path))
+
+  if not PACKAGE_NAME:
+    return path
+
+  return path[len(PACKAGE_NAME)+1:]
+
+def _proto_gen_impl(ctx):
+  """General implementation for generating protos"""
+  srcs = ctx.files.srcs
+  deps = []
+  deps += ctx.files.srcs
+  gen_dir = _GenDir(ctx)
+  if gen_dir:
+    import_flags = ["-I" + gen_dir]
+  else:
+    import_flags = ["-I."]
+
+  for dep in ctx.attr.deps:
+    import_flags += dep.proto.import_flags
+    deps += dep.proto.deps
+
+  args = []
+  if ctx.attr.gen_cc:
+    args += ["--cpp_out=" + ctx.var["GENDIR"] + "/" + gen_dir]
+  if ctx.attr.gen_py:
+    args += ["--python_out=" + ctx.var["GENDIR"] + "/" + gen_dir]
+
+  if ctx.executable.grpc_cpp_plugin:
+    args += ["--plugin=protoc-gen-grpc=" + ctx.executable.grpc_cpp_plugin.path]
+    args += ["--grpc_out=" + ctx.var["GENDIR"] + "/" + gen_dir]
+
+  if args:
+    ctx.action(
+        inputs=srcs + deps,
+        outputs=ctx.outputs.outs,
+        arguments=args + import_flags + [s.path for s in srcs],
+        executable=ctx.executable.protoc,
+    )
+
+  return struct(
+      proto=struct(
+          srcs=srcs,
+          import_flags=import_flags,
+          deps=deps,
+      ),
+  )
+
+_proto_gen = rule(
+    attrs = {
+        "srcs": attr.label_list(allow_files = True),
+        "deps": attr.label_list(providers = ["proto"]),
+        "includes": attr.string_list(),
+        "protoc": attr.label(
+            cfg = HOST_CFG,
+            executable = True,
+            single_file = True,
+            mandatory = True,
+        ),
+        "grpc_cpp_plugin": attr.label(
+            cfg = HOST_CFG,
+            executable = True,
+            single_file = True,
+        ),
+        "gen_cc": attr.bool(),
+        "gen_py": attr.bool(),
+        "outs": attr.output_list(),
+    },
+    output_to_genfiles = True,
+    implementation = _proto_gen_impl,
+)
+
+def cc_proto_library(
+        name,
+        srcs=[],
+        deps=[],
+        cc_libs=[],
+        include=None,
+        protoc="//:protoc",
+        internal_bootstrap_hack=False,
+        use_grpc_plugin=False,
+        default_runtime="//:protobuf",
+        **kargs):
+  """Bazel rule to create a C++ protobuf library from proto source files
+
+  NOTE: the rule is only an internal workaround to generate protos. The
+  interface may change and the rule may be removed when bazel has introduced
+  the native rule.
+
+  Args:
+    name: the name of the cc_proto_library.
+    srcs: the .proto files of the cc_proto_library.
+    deps: a list of dependency labels; must be cc_proto_library.
+    cc_libs: a list of other cc_library targets depended by the generated
+        cc_library.
+    include: a string indicating the include path of the .proto files.
+    protoc: the label of the protocol compiler to generate the sources.
+    internal_bootstrap_hack: a flag indicate the cc_proto_library is used only
+        for bootstraping. When it is set to True, no files will be generated.
+        The rule will simply be a provider for .proto files, so that other
+        cc_proto_library can depend on it.
+    use_grpc_plugin: a flag to indicate whether to call the grpc C++ plugin
+        when processing the proto files.
+    default_runtime: the implicitly default runtime which will be depended on by
+        the generated cc_library target.
+    **kargs: other keyword arguments that are passed to cc_library.
+
+  """
+
+  includes = []
+  if include != None:
+    includes = [include]
+
+  if internal_bootstrap_hack:
+    # For pre-checked-in generated files, we add the internal_bootstrap_hack
+    # which will skip the codegen action.
+    _proto_gen(
+        name=name + "_genproto",
+        srcs=srcs,
+        deps=[s + "_genproto" for s in deps],
+        includes=includes,
+        protoc=protoc,
+        visibility=["//visibility:public"],
+    )
+    # An empty cc_library to make rule dependency consistent.
+    native.cc_library(
+        name=name,
+        **kargs)
+    return
+
+  grpc_cpp_plugin = None
+  if use_grpc_plugin:
+    grpc_cpp_plugin = "//external:grpc_cpp_plugin"
+
+  outs = _CcOuts(srcs, use_grpc_plugin)
+
+  _proto_gen(
+      name=name + "_genproto",
+      srcs=srcs,
+      deps=[s + "_genproto" for s in deps],
+      includes=includes,
+      protoc=protoc,
+      grpc_cpp_plugin=grpc_cpp_plugin,
+      gen_cc=1,
+      outs=outs,
+      visibility=["//visibility:public"],
+  )
+
+  if default_runtime and not default_runtime in cc_libs:
+    cc_libs += [default_runtime]
+  if use_grpc_plugin:
+    cc_libs += ["//external:grpc_lib"]
+
+  native.cc_library(
+      name=name,
+      srcs=outs,
+      deps=cc_libs + deps,
+      includes=includes,
+      **kargs)
+
+def py_proto_library(
+        name,
+        srcs=[],
+        deps=[],
+        py_libs=[],
+        py_extra_srcs=[],
+        include=None,
+        default_runtime="//:protobuf_python",
+        protoc="//:protoc",
+        **kargs):
+  """Bazel rule to create a Python protobuf library from proto source files
+
+  NOTE: the rule is only an internal workaround to generate protos. The
+  interface may change and the rule may be removed when bazel has introduced
+  the native rule.
+
+  Args:
+    name: the name of the py_proto_library.
+    srcs: the .proto files of the py_proto_library.
+    deps: a list of dependency labels; must be py_proto_library.
+    py_libs: a list of other py_library targets depended by the generated
+        py_library.
+    py_extra_srcs: extra source files that will be added to the output
+        py_library. This attribute is used for internal bootstrapping.
+    include: a string indicating the include path of the .proto files.
+    default_runtime: the implicitly default runtime which will be depended on by
+        the generated py_library target.
+    protoc: the label of the protocol compiler to generate the sources.
+    **kargs: other keyword arguments that are passed to cc_library.
+
+  """
+  outs = _PyOuts(srcs)
+
+  includes = []
+  if include != None:
+    includes = [include]
+
+  _proto_gen(
+      name=name + "_genproto",
+      srcs=srcs,
+      deps=[s + "_genproto" for s in deps],
+      includes=includes,
+      protoc=protoc,
+      gen_py=1,
+      outs=outs,
+      visibility=["//visibility:public"],
+  )
+
+  if default_runtime and not default_runtime in py_libs + deps:
+    py_libs += [default_runtime]
+
+  native.py_library(
+      name=name,
+      srcs=outs+py_extra_srcs,
+      deps=py_libs+deps,
+      imports=includes,
+      **kargs)
+
+def internal_protobuf_py_tests(
+    name,
+    modules=[],
+    **kargs):
+  """Bazel rules to create batch tests for protobuf internal.
+
+  Args:
+    name: the name of the rule.
+    modules: a list of modules for tests. The macro will create a py_test for
+        each of the parameter with the source "google/protobuf/%s.py"
+    kargs: extra parameters that will be passed into the py_test.
+
+  """
+  for m in modules:
+    s = "python/google/protobuf/internal/%s.py" % m
+    native.py_test(
+        name="py_%s" % m,
+        srcs=[s],
+        main=s,
+        **kargs)
diff --git a/third_party/protobuf/python/google/__init__.py b/third_party/protobuf/python/google/__init__.py
new file mode 100755
index 0000000000..5585614122
--- /dev/null
+++ b/third_party/protobuf/python/google/__init__.py
@@ -0,0 +1,4 @@
+try:
+  __import__('pkg_resources').declare_namespace(__name__)
+except ImportError:
+  __path__ = __import__('pkgutil').extend_path(__path__, __name__)
diff --git a/third_party/protobuf/python/google/protobuf/__init__.py b/third_party/protobuf/python/google/protobuf/__init__.py
new file mode 100755
index 0000000000..fcb1734e47
--- /dev/null
+++ b/third_party/protobuf/python/google/protobuf/__init__.py
@@ -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.
+
+# Copyright 2007 Google Inc. All Rights Reserved.
+
+__version__ = '3.0.0b2.post2'
+
+if __name__ != '__main__':
+  try:
+    __import__('pkg_resources').declare_namespace(__name__)
+  except ImportError:
+    __path__ = __import__('pkgutil').extend_path(__path__, __name__)
diff --git a/third_party/protobuf/python/google/protobuf/descriptor.py b/third_party/protobuf/python/google/protobuf/descriptor.py
new file mode 100755
index 0000000000..5f613c880b
--- /dev/null
+++ b/third_party/protobuf/python/google/protobuf/descriptor.py
@@ -0,0 +1,966 @@
+# 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.
+
+"""Descriptors essentially contain exactly the information found in a .proto
+file, in types that make this information accessible in Python.
+"""
+
+__author__ = 'robinson@google.com (Will Robinson)'
+
+import six
+
+from google.protobuf.internal import api_implementation
+
+_USE_C_DESCRIPTORS = False
+if api_implementation.Type() == 'cpp':
+  # Used by MakeDescriptor in cpp mode
+  import os
+  import uuid
+  from google.protobuf.pyext import _message
+  _USE_C_DESCRIPTORS = getattr(_message, '_USE_C_DESCRIPTORS', False)
+
+
+class Error(Exception):
+  """Base error for this module."""
+
+
+class TypeTransformationError(Error):
+  """Error transforming between python proto type and corresponding C++ type."""
+
+
+if _USE_C_DESCRIPTORS:
+  # This metaclass allows to override the behavior of code like
+  #     isinstance(my_descriptor, FieldDescriptor)
+  # and make it return True when the descriptor is an instance of the extension
+  # type written in C++.
+  class DescriptorMetaclass(type):
+    def __instancecheck__(cls, obj):
+      if super(DescriptorMetaclass, cls).__instancecheck__(obj):
+        return True
+      if isinstance(obj, cls._C_DESCRIPTOR_CLASS):
+        return True
+      return False
+else:
+  # The standard metaclass; nothing changes.
+  DescriptorMetaclass = type
+
+
+class DescriptorBase(six.with_metaclass(DescriptorMetaclass)):
+
+  """Descriptors base class.
+
+  This class is the base of all descriptor classes. It provides common options
+  related functionality.
+
+  Attributes:
+    has_options:  True if the descriptor has non-default options.  Usually it
+        is not necessary to read this -- just call GetOptions() which will
+        happily return the default instance.  However, it's sometimes useful
+        for efficiency, and also useful inside the protobuf implementation to
+        avoid some bootstrapping issues.
+  """
+
+  if _USE_C_DESCRIPTORS:
+    # The class, or tuple of classes, that are considered as "virtual
+    # subclasses" of this descriptor class.
+    _C_DESCRIPTOR_CLASS = ()
+
+  def __init__(self, options, options_class_name):
+    """Initialize the descriptor given its options message and the name of the
+    class of the options message. The name of the class is required in case
+    the options message is None and has to be created.
+    """
+    self._options = options
+    self._options_class_name = options_class_name
+
+    # Does this descriptor have non-default options?
+    self.has_options = options is not None
+
+  def _SetOptions(self, options, options_class_name):
+    """Sets the descriptor's options
+
+    This function is used in generated proto2 files to update descriptor
+    options. It must not be used outside proto2.
+    """
+    self._options = options
+    self._options_class_name = options_class_name
+
+    # Does this descriptor have non-default options?
+    self.has_options = options is not None
+
+  def GetOptions(self):
+    """Retrieves descriptor options.
+
+    This method returns the options set or creates the default options for the
+    descriptor.
+    """
+    if self._options:
+      return self._options
+    from google.protobuf import descriptor_pb2
+    try:
+      options_class = getattr(descriptor_pb2, self._options_class_name)
+    except AttributeError:
+      raise RuntimeError('Unknown options class name %s!' %
+                         (self._options_class_name))
+    self._options = options_class()
+    return self._options
+
+
+class _NestedDescriptorBase(DescriptorBase):
+  """Common class for descriptors that can be nested."""
+
+  def __init__(self, options, options_class_name, name, full_name,
+               file, containing_type, serialized_start=None,
+               serialized_end=None):
+    """Constructor.
+
+    Args:
+      options: Protocol message options or None
+        to use default message options.
+      options_class_name: (str) The class name of the above options.
+
+      name: (str) Name of this protocol message type.
+      full_name: (str) Fully-qualified name of this protocol message type,
+        which will include protocol "package" name and the name of any
+        enclosing types.
+      file: (FileDescriptor) Reference to file info.
+      containing_type: if provided, this is a nested descriptor, with this
+        descriptor as parent, otherwise None.
+      serialized_start: The start index (inclusive) in block in the
+        file.serialized_pb that describes this descriptor.
+      serialized_end: The end index (exclusive) in block in the
+        file.serialized_pb that describes this descriptor.
+    """
+    super(_NestedDescriptorBase, self).__init__(
+        options, options_class_name)
+
+    self.name = name
+    # TODO(falk): Add function to calculate full_name instead of having it in
+    #             memory?
+    self.full_name = full_name
+    self.file = file
+    self.containing_type = containing_type
+
+    self._serialized_start = serialized_start
+    self._serialized_end = serialized_end
+
+  def GetTopLevelContainingType(self):
+    """Returns the root if this is a nested type, or itself if its the root."""
+    desc = self
+    while desc.containing_type is not None:
+      desc = desc.containing_type
+    return desc
+
+  def CopyToProto(self, proto):
+    """Copies this to the matching proto in descriptor_pb2.
+
+    Args:
+      proto: An empty proto instance from descriptor_pb2.
+
+    Raises:
+      Error: If self couldnt be serialized, due to to few constructor arguments.
+    """
+    if (self.file is not None and
+        self._serialized_start is not None and
+        self._serialized_end is not None):
+      proto.ParseFromString(self.file.serialized_pb[
+          self._serialized_start:self._serialized_end])
+    else:
+      raise Error('Descriptor does not contain serialization.')
+
+
+class Descriptor(_NestedDescriptorBase):
+
+  """Descriptor for a protocol message type.
+
+  A Descriptor instance has the following attributes:
+
+    name: (str) Name of this protocol message type.
+    full_name: (str) Fully-qualified name of this protocol message type,
+      which will include protocol "package" name and the name of any
+      enclosing types.
+
+    containing_type: (Descriptor) Reference to the descriptor of the
+      type containing us, or None if this is top-level.
+
+    fields: (list of FieldDescriptors) Field descriptors for all
+      fields in this type.
+    fields_by_number: (dict int -> FieldDescriptor) Same FieldDescriptor
+      objects as in |fields|, but indexed by "number" attribute in each
+      FieldDescriptor.
+    fields_by_name: (dict str -> FieldDescriptor) Same FieldDescriptor
+      objects as in |fields|, but indexed by "name" attribute in each
+      FieldDescriptor.
+    fields_by_camelcase_name: (dict str -> FieldDescriptor) Same
+      FieldDescriptor objects as in |fields|, but indexed by
+      "camelcase_name" attribute in each FieldDescriptor.
+
+    nested_types: (list of Descriptors) Descriptor references
+      for all protocol message types nested within this one.
+    nested_types_by_name: (dict str -> Descriptor) Same Descriptor
+      objects as in |nested_types|, but indexed by "name" attribute
+      in each Descriptor.
+
+    enum_types: (list of EnumDescriptors) EnumDescriptor references
+      for all enums contained within this type.
+    enum_types_by_name: (dict str ->EnumDescriptor) Same EnumDescriptor
+      objects as in |enum_types|, but indexed by "name" attribute
+      in each EnumDescriptor.
+    enum_values_by_name: (dict str -> EnumValueDescriptor) Dict mapping
+      from enum value name to EnumValueDescriptor for that value.
+
+    extensions: (list of FieldDescriptor) All extensions defined directly
+      within this message type (NOT within a nested type).
+    extensions_by_name: (dict, string -> FieldDescriptor) Same FieldDescriptor
+      objects as |extensions|, but indexed by "name" attribute of each
+      FieldDescriptor.
+
+    is_extendable:  Does this type define any extension ranges?
+
+    oneofs: (list of OneofDescriptor) The list of descriptors for oneof fields
+      in this message.
+    oneofs_by_name: (dict str -> OneofDescriptor) Same objects as in |oneofs|,
+      but indexed by "name" attribute.
+
+    file: (FileDescriptor) Reference to file descriptor.
+  """
+
+  if _USE_C_DESCRIPTORS:
+    _C_DESCRIPTOR_CLASS = _message.Descriptor
+
+    def __new__(cls, name, full_name, filename, containing_type, fields,
+                nested_types, enum_types, extensions, options=None,
+                is_extendable=True, extension_ranges=None, oneofs=None,
+                file=None, serialized_start=None, serialized_end=None,
+                syntax=None):
+      _message.Message._CheckCalledFromGeneratedFile()
+      return _message.default_pool.FindMessageTypeByName(full_name)
+
+  # NOTE(tmarek): The file argument redefining a builtin is nothing we can
+  # fix right now since we don't know how many clients already rely on the
+  # name of the argument.
+  def __init__(self, name, full_name, filename, containing_type, fields,
+               nested_types, enum_types, extensions, options=None,
+               is_extendable=True, extension_ranges=None, oneofs=None,
+               file=None, serialized_start=None, serialized_end=None,
+               syntax=None):  # pylint:disable=redefined-builtin
+    """Arguments to __init__() are as described in the description
+    of Descriptor fields above.
+
+    Note that filename is an obsolete argument, that is not used anymore.
+    Please use file.name to access this as an attribute.
+    """
+    super(Descriptor, self).__init__(
+        options, 'MessageOptions', name, full_name, file,
+        containing_type, serialized_start=serialized_start,
+        serialized_end=serialized_end)
+
+    # We have fields in addition to fields_by_name and fields_by_number,
+    # so that:
+    #   1. Clients can index fields by "order in which they're listed."
+    #   2. Clients can easily iterate over all fields with the terse
+    #      syntax: for f in descriptor.fields: ...
+    self.fields = fields
+    for field in self.fields:
+      field.containing_type = self
+    self.fields_by_number = dict((f.number, f) for f in fields)
+    self.fields_by_name = dict((f.name, f) for f in fields)
+    self._fields_by_camelcase_name = None
+
+    self.nested_types = nested_types
+    for nested_type in nested_types:
+      nested_type.containing_type = self
+    self.nested_types_by_name = dict((t.name, t) for t in nested_types)
+
+    self.enum_types = enum_types
+    for enum_type in self.enum_types:
+      enum_type.containing_type = self
+    self.enum_types_by_name = dict((t.name, t) for t in enum_types)
+    self.enum_values_by_name = dict(
+        (v.name, v) for t in enum_types for v in t.values)
+
+    self.extensions = extensions
+    for extension in self.extensions:
+      extension.extension_scope = self
+    self.extensions_by_name = dict((f.name, f) for f in extensions)
+    self.is_extendable = is_extendable
+    self.extension_ranges = extension_ranges
+    self.oneofs = oneofs if oneofs is not None else []
+    self.oneofs_by_name = dict((o.name, o) for o in self.oneofs)
+    for oneof in self.oneofs:
+      oneof.containing_type = self
+    self.syntax = syntax or "proto2"
+
+  @property
+  def fields_by_camelcase_name(self):
+    if self._fields_by_camelcase_name is None:
+      self._fields_by_camelcase_name = dict(
+          (f.camelcase_name, f) for f in self.fields)
+    return self._fields_by_camelcase_name
+
+  def EnumValueName(self, enum, value):
+    """Returns the string name of an enum value.
+
+    This is just a small helper method to simplify a common operation.
+
+    Args:
+      enum: string name of the Enum.
+      value: int, value of the enum.
+
+    Returns:
+      string name of the enum value.
+
+    Raises:
+      KeyError if either the Enum doesn't exist or the value is not a valid
+        value for the enum.
+    """
+    return self.enum_types_by_name[enum].values_by_number[value].name
+
+  def CopyToProto(self, proto):
+    """Copies this to a descriptor_pb2.DescriptorProto.
+
+    Args:
+      proto: An empty descriptor_pb2.DescriptorProto.
+    """
+    # This function is overriden to give a better doc comment.
+    super(Descriptor, self).CopyToProto(proto)
+
+
+# TODO(robinson): We should have aggressive checking here,
+# for example:
+#   * If you specify a repeated field, you should not be allowed
+#     to specify a default value.
+#   * [Other examples here as needed].
+#
+# TODO(robinson): for this and other *Descriptor classes, we
+# might also want to lock things down aggressively (e.g.,
+# prevent clients from setting the attributes).  Having
+# stronger invariants here in general will reduce the number
+# of runtime checks we must do in reflection.py...
+class FieldDescriptor(DescriptorBase):
+
+  """Descriptor for a single field in a .proto file.
+
+  A FieldDescriptor instance has the following attributes:
+
+    name: (str) Name of this field, exactly as it appears in .proto.
+    full_name: (str) Name of this field, including containing scope.  This is
+      particularly relevant for extensions.
+    camelcase_name: (str) Camelcase name of this field.
+    index: (int) Dense, 0-indexed index giving the order that this
+      field textually appears within its message in the .proto file.
+    number: (int) Tag number declared for this field in the .proto file.
+
+    type: (One of the TYPE_* constants below) Declared type.
+    cpp_type: (One of the CPPTYPE_* constants below) C++ type used to
+      represent this field.
+
+    label: (One of the LABEL_* constants below) Tells whether this
+      field is optional, required, or repeated.
+    has_default_value: (bool) True if this field has a default value defined,
+      otherwise false.
+    default_value: (Varies) Default value of this field.  Only
+      meaningful for non-repeated scalar fields.  Repeated fields
+      should always set this to [], and non-repeated composite
+      fields should always set this to None.
+
+    containing_type: (Descriptor) Descriptor of the protocol message
+      type that contains this field.  Set by the Descriptor constructor
+      if we're passed into one.
+      Somewhat confusingly, for extension fields, this is the
+      descriptor of the EXTENDED message, not the descriptor
+      of the message containing this field.  (See is_extension and
+      extension_scope below).
+    message_type: (Descriptor) If a composite field, a descriptor
+      of the message type contained in this field.  Otherwise, this is None.
+    enum_type: (EnumDescriptor) If this field contains an enum, a
+      descriptor of that enum.  Otherwise, this is None.
+
+    is_extension: True iff this describes an extension field.
+    extension_scope: (Descriptor) Only meaningful if is_extension is True.
+      Gives the message that immediately contains this extension field.
+      Will be None iff we're a top-level (file-level) extension field.
+
+    options: (descriptor_pb2.FieldOptions) Protocol message field options or
+      None to use default field options.
+
+    containing_oneof: (OneofDescriptor) If the field is a member of a oneof
+      union, contains its descriptor. Otherwise, None.
+  """
+
+  # Must be consistent with C++ FieldDescriptor::Type enum in
+  # descriptor.h.
+  #
+  # TODO(robinson): Find a way to eliminate this repetition.
+  TYPE_DOUBLE         = 1
+  TYPE_FLOAT          = 2
+  TYPE_INT64          = 3
+  TYPE_UINT64         = 4
+  TYPE_INT32          = 5
+  TYPE_FIXED64        = 6
+  TYPE_FIXED32        = 7
+  TYPE_BOOL           = 8
+  TYPE_STRING         = 9
+  TYPE_GROUP          = 10
+  TYPE_MESSAGE        = 11
+  TYPE_BYTES          = 12
+  TYPE_UINT32         = 13
+  TYPE_ENUM           = 14
+  TYPE_SFIXED32       = 15
+  TYPE_SFIXED64       = 16
+  TYPE_SINT32         = 17
+  TYPE_SINT64         = 18
+  MAX_TYPE            = 18
+
+  # Must be consistent with C++ FieldDescriptor::CppType enum in
+  # descriptor.h.
+  #
+  # TODO(robinson): Find a way to eliminate this repetition.
+  CPPTYPE_INT32       = 1
+  CPPTYPE_INT64       = 2
+  CPPTYPE_UINT32      = 3
+  CPPTYPE_UINT64      = 4
+  CPPTYPE_DOUBLE      = 5
+  CPPTYPE_FLOAT       = 6
+  CPPTYPE_BOOL        = 7
+  CPPTYPE_ENUM        = 8
+  CPPTYPE_STRING      = 9
+  CPPTYPE_MESSAGE     = 10
+  MAX_CPPTYPE         = 10
+
+  _PYTHON_TO_CPP_PROTO_TYPE_MAP = {
+      TYPE_DOUBLE: CPPTYPE_DOUBLE,
+      TYPE_FLOAT: CPPTYPE_FLOAT,
+      TYPE_ENUM: CPPTYPE_ENUM,
+      TYPE_INT64: CPPTYPE_INT64,
+      TYPE_SINT64: CPPTYPE_INT64,
+      TYPE_SFIXED64: CPPTYPE_INT64,
+      TYPE_UINT64: CPPTYPE_UINT64,
+      TYPE_FIXED64: CPPTYPE_UINT64,
+      TYPE_INT32: CPPTYPE_INT32,
+      TYPE_SFIXED32: CPPTYPE_INT32,
+      TYPE_SINT32: CPPTYPE_INT32,
+      TYPE_UINT32: CPPTYPE_UINT32,
+      TYPE_FIXED32: CPPTYPE_UINT32,
+      TYPE_BYTES: CPPTYPE_STRING,
+      TYPE_STRING: CPPTYPE_STRING,
+      TYPE_BOOL: CPPTYPE_BOOL,
+      TYPE_MESSAGE: CPPTYPE_MESSAGE,
+      TYPE_GROUP: CPPTYPE_MESSAGE
+      }
+
+  # Must be consistent with C++ FieldDescriptor::Label enum in
+  # descriptor.h.
+  #
+  # TODO(robinson): Find a way to eliminate this repetition.
+  LABEL_OPTIONAL      = 1
+  LABEL_REQUIRED      = 2
+  LABEL_REPEATED      = 3
+  MAX_LABEL           = 3
+
+  # Must be consistent with C++ constants kMaxNumber, kFirstReservedNumber,
+  # and kLastReservedNumber in descriptor.h
+  MAX_FIELD_NUMBER = (1 << 29) - 1
+  FIRST_RESERVED_FIELD_NUMBER = 19000
+  LAST_RESERVED_FIELD_NUMBER = 19999
+
+  if _USE_C_DESCRIPTORS:
+    _C_DESCRIPTOR_CLASS = _message.FieldDescriptor
+
+    def __new__(cls, name, full_name, index, number, type, cpp_type, label,
+                default_value, message_type, enum_type, containing_type,
+                is_extension, extension_scope, options=None,
+                has_default_value=True, containing_oneof=None):
+      _message.Message._CheckCalledFromGeneratedFile()
+      if is_extension:
+        return _message.default_pool.FindExtensionByName(full_name)
+      else:
+        return _message.default_pool.FindFieldByName(full_name)
+
+  def __init__(self, name, full_name, index, number, type, cpp_type, label,
+               default_value, message_type, enum_type, containing_type,
+               is_extension, extension_scope, options=None,
+               has_default_value=True, containing_oneof=None):
+    """The arguments are as described in the description of FieldDescriptor
+    attributes above.
+
+    Note that containing_type may be None, and may be set later if necessary
+    (to deal with circular references between message types, for example).
+    Likewise for extension_scope.
+    """
+    super(FieldDescriptor, self).__init__(options, 'FieldOptions')
+    self.name = name
+    self.full_name = full_name
+    self._camelcase_name = None
+    self.index = index
+    self.number = number
+    self.type = type
+    self.cpp_type = cpp_type
+    self.label = label
+    self.has_default_value = has_default_value
+    self.default_value = default_value
+    self.containing_type = containing_type
+    self.message_type = message_type
+    self.enum_type = enum_type
+    self.is_extension = is_extension
+    self.extension_scope = extension_scope
+    self.containing_oneof = containing_oneof
+    if api_implementation.Type() == 'cpp':
+      if is_extension:
+        self._cdescriptor = _message.default_pool.FindExtensionByName(full_name)
+      else:
+        self._cdescriptor = _message.default_pool.FindFieldByName(full_name)
+    else:
+      self._cdescriptor = None
+
+  @property
+  def camelcase_name(self):
+    if self._camelcase_name is None:
+      self._camelcase_name = _ToCamelCase(self.name)
+    return self._camelcase_name
+
+  @staticmethod
+  def ProtoTypeToCppProtoType(proto_type):
+    """Converts from a Python proto type to a C++ Proto Type.
+
+    The Python ProtocolBuffer classes specify both the 'Python' datatype and the
+    'C++' datatype - and they're not the same. This helper method should
+    translate from one to another.
+
+    Args:
+      proto_type: the Python proto type (descriptor.FieldDescriptor.TYPE_*)
+    Returns:
+      descriptor.FieldDescriptor.CPPTYPE_*, the C++ type.
+    Raises:
+      TypeTransformationError: when the Python proto type isn't known.
+    """
+    try:
+      return FieldDescriptor._PYTHON_TO_CPP_PROTO_TYPE_MAP[proto_type]
+    except KeyError:
+      raise TypeTransformationError('Unknown proto_type: %s' % proto_type)
+
+
+class EnumDescriptor(_NestedDescriptorBase):
+
+  """Descriptor for an enum defined in a .proto file.
+
+  An EnumDescriptor instance has the following attributes:
+
+    name: (str) Name of the enum type.
+    full_name: (str) Full name of the type, including package name
+      and any enclosing type(s).
+
+    values: (list of EnumValueDescriptors) List of the values
+      in this enum.
+    values_by_name: (dict str -> EnumValueDescriptor) Same as |values|,
+      but indexed by the "name" field of each EnumValueDescriptor.
+    values_by_number: (dict int -> EnumValueDescriptor) Same as |values|,
+      but indexed by the "number" field of each EnumValueDescriptor.
+    containing_type: (Descriptor) Descriptor of the immediate containing
+      type of this enum, or None if this is an enum defined at the
+      top level in a .proto file.  Set by Descriptor's constructor
+      if we're passed into one.
+    file: (FileDescriptor) Reference to file descriptor.
+    options: (descriptor_pb2.EnumOptions) Enum options message or
+      None to use default enum options.
+  """
+
+  if _USE_C_DESCRIPTORS:
+    _C_DESCRIPTOR_CLASS = _message.EnumDescriptor
+
+    def __new__(cls, name, full_name, filename, values,
+                containing_type=None, options=None, file=None,
+                serialized_start=None, serialized_end=None):
+      _message.Message._CheckCalledFromGeneratedFile()
+      return _message.default_pool.FindEnumTypeByName(full_name)
+
+  def __init__(self, name, full_name, filename, values,
+               containing_type=None, options=None, file=None,
+               serialized_start=None, serialized_end=None):
+    """Arguments are as described in the attribute description above.
+
+    Note that filename is an obsolete argument, that is not used anymore.
+    Please use file.name to access this as an attribute.
+    """
+    super(EnumDescriptor, self).__init__(
+        options, 'EnumOptions', name, full_name, file,
+        containing_type, serialized_start=serialized_start,
+        serialized_end=serialized_end)
+
+    self.values = values
+    for value in self.values:
+      value.type = self
+    self.values_by_name = dict((v.name, v) for v in values)
+    self.values_by_number = dict((v.number, v) for v in values)
+
+  def CopyToProto(self, proto):
+    """Copies this to a descriptor_pb2.EnumDescriptorProto.
+
+    Args:
+      proto: An empty descriptor_pb2.EnumDescriptorProto.
+    """
+    # This function is overriden to give a better doc comment.
+    super(EnumDescriptor, self).CopyToProto(proto)
+
+
+class EnumValueDescriptor(DescriptorBase):
+
+  """Descriptor for a single value within an enum.
+
+    name: (str) Name of this value.
+    index: (int) Dense, 0-indexed index giving the order that this
+      value appears textually within its enum in the .proto file.
+    number: (int) Actual number assigned to this enum value.
+    type: (EnumDescriptor) EnumDescriptor to which this value
+      belongs.  Set by EnumDescriptor's constructor if we're
+      passed into one.
+    options: (descriptor_pb2.EnumValueOptions) Enum value options message or
+      None to use default enum value options options.
+  """
+
+  if _USE_C_DESCRIPTORS:
+    _C_DESCRIPTOR_CLASS = _message.EnumValueDescriptor
+
+    def __new__(cls, name, index, number, type=None, options=None):
+      _message.Message._CheckCalledFromGeneratedFile()
+      # There is no way we can build a complete EnumValueDescriptor with the
+      # given parameters (the name of the Enum is not known, for example).
+      # Fortunately generated files just pass it to the EnumDescriptor()
+      # constructor, which will ignore it, so returning None is good enough.
+      return None
+
+  def __init__(self, name, index, number, type=None, options=None):
+    """Arguments are as described in the attribute description above."""
+    super(EnumValueDescriptor, self).__init__(options, 'EnumValueOptions')
+    self.name = name
+    self.index = index
+    self.number = number
+    self.type = type
+
+
+class OneofDescriptor(object):
+  """Descriptor for a oneof field.
+
+    name: (str) Name of the oneof field.
+    full_name: (str) Full name of the oneof field, including package name.
+    index: (int) 0-based index giving the order of the oneof field inside
+      its containing type.
+    containing_type: (Descriptor) Descriptor of the protocol message
+      type that contains this field.  Set by the Descriptor constructor
+      if we're passed into one.
+    fields: (list of FieldDescriptor) The list of field descriptors this
+      oneof can contain.
+  """
+
+  if _USE_C_DESCRIPTORS:
+    _C_DESCRIPTOR_CLASS = _message.OneofDescriptor
+
+    def __new__(cls, name, full_name, index, containing_type, fields):
+      _message.Message._CheckCalledFromGeneratedFile()
+      return _message.default_pool.FindOneofByName(full_name)
+
+  def __init__(self, name, full_name, index, containing_type, fields):
+    """Arguments are as described in the attribute description above."""
+    self.name = name
+    self.full_name = full_name
+    self.index = index
+    self.containing_type = containing_type
+    self.fields = fields
+
+
+class ServiceDescriptor(_NestedDescriptorBase):
+
+  """Descriptor for a service.
+
+    name: (str) Name of the service.
+    full_name: (str) Full name of the service, including package name.
+    index: (int) 0-indexed index giving the order that this services
+      definition appears withing the .proto file.
+    methods: (list of MethodDescriptor) List of methods provided by this
+      service.
+    options: (descriptor_pb2.ServiceOptions) Service options message or
+      None to use default service options.
+    file: (FileDescriptor) Reference to file info.
+  """
+
+  def __init__(self, name, full_name, index, methods, options=None, file=None,
+               serialized_start=None, serialized_end=None):
+    super(ServiceDescriptor, self).__init__(
+        options, 'ServiceOptions', name, full_name, file,
+        None, serialized_start=serialized_start,
+        serialized_end=serialized_end)
+    self.index = index
+    self.methods = methods
+    # Set the containing service for each method in this service.
+    for method in self.methods:
+      method.containing_service = self
+
+  def FindMethodByName(self, name):
+    """Searches for the specified method, and returns its descriptor."""
+    for method in self.methods:
+      if name == method.name:
+        return method
+    return None
+
+  def CopyToProto(self, proto):
+    """Copies this to a descriptor_pb2.ServiceDescriptorProto.
+
+    Args:
+      proto: An empty descriptor_pb2.ServiceDescriptorProto.
+    """
+    # This function is overriden to give a better doc comment.
+    super(ServiceDescriptor, self).CopyToProto(proto)
+
+
+class MethodDescriptor(DescriptorBase):
+
+  """Descriptor for a method in a service.
+
+  name: (str) Name of the method within the service.
+  full_name: (str) Full name of method.
+  index: (int) 0-indexed index of the method inside the service.
+  containing_service: (ServiceDescriptor) The service that contains this
+    method.
+  input_type: The descriptor of the message that this method accepts.
+  output_type: The descriptor of the message that this method returns.
+  options: (descriptor_pb2.MethodOptions) Method options message or
+    None to use default method options.
+  """
+
+  def __init__(self, name, full_name, index, containing_service,
+               input_type, output_type, options=None):
+    """The arguments are as described in the description of MethodDescriptor
+    attributes above.
+
+    Note that containing_service may be None, and may be set later if necessary.
+    """
+    super(MethodDescriptor, self).__init__(options, 'MethodOptions')
+    self.name = name
+    self.full_name = full_name
+    self.index = index
+    self.containing_service = containing_service
+    self.input_type = input_type
+    self.output_type = output_type
+
+
+class FileDescriptor(DescriptorBase):
+  """Descriptor for a file. Mimics the descriptor_pb2.FileDescriptorProto.
+
+  Note that enum_types_by_name, extensions_by_name, and dependencies
+  fields are only set by the message_factory module, and not by the
+  generated proto code.
+
+  name: name of file, relative to root of source tree.
+  package: name of the package
+  syntax: string indicating syntax of the file (can be "proto2" or "proto3")
+  serialized_pb: (str) Byte string of serialized
+    descriptor_pb2.FileDescriptorProto.
+  dependencies: List of other FileDescriptors this FileDescriptor depends on.
+  message_types_by_name: Dict of message names of their descriptors.
+  enum_types_by_name: Dict of enum names and their descriptors.
+  extensions_by_name: Dict of extension names and their descriptors.
+  pool: the DescriptorPool this descriptor belongs to.  When not passed to the
+    constructor, the global default pool is used.
+  """
+
+  if _USE_C_DESCRIPTORS:
+    _C_DESCRIPTOR_CLASS = _message.FileDescriptor
+
+    def __new__(cls, name, package, options=None, serialized_pb=None,
+                dependencies=None, syntax=None, pool=None):
+      # FileDescriptor() is called from various places, not only from generated
+      # files, to register dynamic proto files and messages.
+      if serialized_pb:
+        # TODO(amauryfa): use the pool passed as argument. This will work only
+        # for C++-implemented DescriptorPools.
+        return _message.default_pool.AddSerializedFile(serialized_pb)
+      else:
+        return super(FileDescriptor, cls).__new__(cls)
+
+  def __init__(self, name, package, options=None, serialized_pb=None,
+               dependencies=None, syntax=None, pool=None):
+    """Constructor."""
+    super(FileDescriptor, self).__init__(options, 'FileOptions')
+
+    if pool is None:
+      from google.protobuf import descriptor_pool
+      pool = descriptor_pool.Default()
+    self.pool = pool
+    self.message_types_by_name = {}
+    self.name = name
+    self.package = package
+    self.syntax = syntax or "proto2"
+    self.serialized_pb = serialized_pb
+
+    self.enum_types_by_name = {}
+    self.extensions_by_name = {}
+    self.dependencies = (dependencies or [])
+
+    if (api_implementation.Type() == 'cpp' and
+        self.serialized_pb is not None):
+      _message.default_pool.AddSerializedFile(self.serialized_pb)
+
+  def CopyToProto(self, proto):
+    """Copies this to a descriptor_pb2.FileDescriptorProto.
+
+    Args:
+      proto: An empty descriptor_pb2.FileDescriptorProto.
+    """
+    proto.ParseFromString(self.serialized_pb)
+
+
+def _ParseOptions(message, string):
+  """Parses serialized options.
+
+  This helper function is used to parse serialized options in generated
+  proto2 files. It must not be used outside proto2.
+  """
+  message.ParseFromString(string)
+  return message
+
+
+def _ToCamelCase(name):
+  """Converts name to camel-case and returns it."""
+  capitalize_next = False
+  result = []
+
+  for c in name:
+    if c == '_':
+      if result:
+        capitalize_next = True
+    elif capitalize_next:
+      result.append(c.upper())
+      capitalize_next = False
+    else:
+      result += c
+
+  # Lower-case the first letter.
+  if result and result[0].isupper():
+    result[0] = result[0].lower()
+  return ''.join(result)
+
+
+def MakeDescriptor(desc_proto, package='', build_file_if_cpp=True,
+                   syntax=None):
+  """Make a protobuf Descriptor given a DescriptorProto protobuf.
+
+  Handles nested descriptors. Note that this is limited to the scope of defining
+  a message inside of another message. Composite fields can currently only be
+  resolved if the message is defined in the same scope as the field.
+
+  Args:
+    desc_proto: The descriptor_pb2.DescriptorProto protobuf message.
+    package: Optional package name for the new message Descriptor (string).
+    build_file_if_cpp: Update the C++ descriptor pool if api matches.
+                       Set to False on recursion, so no duplicates are created.
+    syntax: The syntax/semantics that should be used.  Set to "proto3" to get
+            proto3 field presence semantics.
+  Returns:
+    A Descriptor for protobuf messages.
+  """
+  if api_implementation.Type() == 'cpp' and build_file_if_cpp:
+    # The C++ implementation requires all descriptors to be backed by the same
+    # definition in the C++ descriptor pool. To do this, we build a
+    # FileDescriptorProto with the same definition as this descriptor and build
+    # it into the pool.
+    from google.protobuf import descriptor_pb2
+    file_descriptor_proto = descriptor_pb2.FileDescriptorProto()
+    file_descriptor_proto.message_type.add().MergeFrom(desc_proto)
+
+    # Generate a random name for this proto file to prevent conflicts with any
+    # imported ones. We need to specify a file name so the descriptor pool
+    # accepts our FileDescriptorProto, but it is not important what that file
+    # name is actually set to.
+    proto_name = str(uuid.uuid4())
+
+    if package:
+      file_descriptor_proto.name = os.path.join(package.replace('.', '/'),
+                                                proto_name + '.proto')
+      file_descriptor_proto.package = package
+    else:
+      file_descriptor_proto.name = proto_name + '.proto'
+
+    _message.default_pool.Add(file_descriptor_proto)
+    result = _message.default_pool.FindFileByName(file_descriptor_proto.name)
+
+    if _USE_C_DESCRIPTORS:
+      return result.message_types_by_name[desc_proto.name]
+
+  full_message_name = [desc_proto.name]
+  if package: full_message_name.insert(0, package)
+
+  # Create Descriptors for enum types
+  enum_types = {}
+  for enum_proto in desc_proto.enum_type:
+    full_name = '.'.join(full_message_name + [enum_proto.name])
+    enum_desc = EnumDescriptor(
+      enum_proto.name, full_name, None, [
+          EnumValueDescriptor(enum_val.name, ii, enum_val.number)
+          for ii, enum_val in enumerate(enum_proto.value)])
+    enum_types[full_name] = enum_desc
+
+  # Create Descriptors for nested types
+  nested_types = {}
+  for nested_proto in desc_proto.nested_type:
+    full_name = '.'.join(full_message_name + [nested_proto.name])
+    # Nested types are just those defined inside of the message, not all types
+    # used by fields in the message, so no loops are possible here.
+    nested_desc = MakeDescriptor(nested_proto,
+                                 package='.'.join(full_message_name),
+                                 build_file_if_cpp=False,
+                                 syntax=syntax)
+    nested_types[full_name] = nested_desc
+
+  fields = []
+  for field_proto in desc_proto.field:
+    full_name = '.'.join(full_message_name + [field_proto.name])
+    enum_desc = None
+    nested_desc = None
+    if field_proto.HasField('type_name'):
+      type_name = field_proto.type_name
+      full_type_name = '.'.join(full_message_name +
+                                [type_name[type_name.rfind('.')+1:]])
+      if full_type_name in nested_types:
+        nested_desc = nested_types[full_type_name]
+      elif full_type_name in enum_types:
+        enum_desc = enum_types[full_type_name]
+      # Else type_name references a non-local type, which isn't implemented
+    field = FieldDescriptor(
+        field_proto.name, full_name, field_proto.number - 1,
+        field_proto.number, field_proto.type,
+        FieldDescriptor.ProtoTypeToCppProtoType(field_proto.type),
+        field_proto.label, None, nested_desc, enum_desc, None, False, None,
+        options=field_proto.options, has_default_value=False)
+    fields.append(field)
+
+  desc_name = '.'.join(full_message_name)
+  return Descriptor(desc_proto.name, desc_name, None, None, fields,
+                    list(nested_types.values()), list(enum_types.values()), [],
+                    options=desc_proto.options)
diff --git a/third_party/protobuf/python/google/protobuf/descriptor_database.py b/third_party/protobuf/python/google/protobuf/descriptor_database.py
new file mode 100644
index 0000000000..1333f9966e
--- /dev/null
+++ b/third_party/protobuf/python/google/protobuf/descriptor_database.py
@@ -0,0 +1,141 @@
+# 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 a container for DescriptorProtos."""
+
+__author__ = 'matthewtoia@google.com (Matt Toia)'
+
+
+class Error(Exception):
+  pass
+
+
+class DescriptorDatabaseConflictingDefinitionError(Error):
+  """Raised when a proto is added with the same name & different descriptor."""
+
+
+class DescriptorDatabase(object):
+  """A container accepting FileDescriptorProtos and maps DescriptorProtos."""
+
+  def __init__(self):
+    self._file_desc_protos_by_file = {}
+    self._file_desc_protos_by_symbol = {}
+
+  def Add(self, file_desc_proto):
+    """Adds the FileDescriptorProto and its types to this database.
+
+    Args:
+      file_desc_proto: The FileDescriptorProto to add.
+    Raises:
+      DescriptorDatabaseException: if an attempt is made to add a proto
+        with the same name but different definition than an exisiting
+        proto in the database.
+    """
+    proto_name = file_desc_proto.name
+    if proto_name not in self._file_desc_protos_by_file:
+      self._file_desc_protos_by_file[proto_name] = file_desc_proto
+    elif self._file_desc_protos_by_file[proto_name] != file_desc_proto:
+      raise DescriptorDatabaseConflictingDefinitionError(
+          '%s already added, but with different descriptor.' % proto_name)
+
+    # Add the top-level Message, Enum and Extension descriptors to the index.
+    package = file_desc_proto.package
+    for message in file_desc_proto.message_type:
+      self._file_desc_protos_by_symbol.update(
+          (name, file_desc_proto) for name in _ExtractSymbols(message, package))
+    for enum in file_desc_proto.enum_type:
+      self._file_desc_protos_by_symbol[
+          '.'.join((package, enum.name))] = file_desc_proto
+    for extension in file_desc_proto.extension:
+      self._file_desc_protos_by_symbol[
+          '.'.join((package, extension.name))] = file_desc_proto
+
+  def FindFileByName(self, name):
+    """Finds the file descriptor proto by file name.
+
+    Typically the file name is a relative path ending to a .proto file. The
+    proto with the given name will have to have been added to this database
+    using the Add method or else an error will be raised.
+
+    Args:
+      name: The file name to find.
+
+    Returns:
+      The file descriptor proto matching the name.
+
+    Raises:
+      KeyError if no file by the given name was added.
+    """
+
+    return self._file_desc_protos_by_file[name]
+
+  def FindFileContainingSymbol(self, symbol):
+    """Finds the file descriptor proto containing the specified symbol.
+
+    The symbol should be a fully qualified name including the file descriptor's
+    package and any containing messages. Some examples:
+
+    'some.package.name.Message'
+    'some.package.name.Message.NestedEnum'
+
+    The file descriptor proto containing the specified symbol must be added to
+    this database using the Add method or else an error will be raised.
+
+    Args:
+      symbol: The fully qualified symbol name.
+
+    Returns:
+      The file descriptor proto containing the symbol.
+
+    Raises:
+      KeyError if no file contains the specified symbol.
+    """
+
+    return self._file_desc_protos_by_symbol[symbol]
+
+
+def _ExtractSymbols(desc_proto, package):
+  """Pulls out all the symbols from a descriptor proto.
+
+  Args:
+    desc_proto: The proto to extract symbols from.
+    package: The package containing the descriptor type.
+
+  Yields:
+    The fully qualified name found in the descriptor.
+  """
+
+  message_name = '.'.join((package, desc_proto.name))
+  yield message_name
+  for nested_type in desc_proto.nested_type:
+    for symbol in _ExtractSymbols(nested_type, message_name):
+      yield symbol
+  for enum_type in desc_proto.enum_type:
+    yield '.'.join((message_name, enum_type.name))
diff --git a/third_party/protobuf/python/google/protobuf/descriptor_pool.py b/third_party/protobuf/python/google/protobuf/descriptor_pool.py
new file mode 100644
index 0000000000..3e80795c86
--- /dev/null
+++ b/third_party/protobuf/python/google/protobuf/descriptor_pool.py
@@ -0,0 +1,749 @@
+# 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 DescriptorPool to use as a container for proto2 descriptors.
+
+The DescriptorPool is used in conjection with a DescriptorDatabase to maintain
+a collection of protocol buffer descriptors for use when dynamically creating
+message types at runtime.
+
+For most applications protocol buffers should be used via modules generated by
+the protocol buffer compiler tool. This should only be used when the type of
+protocol buffers used in an application or library cannot be predetermined.
+
+Below is a straightforward example on how to use this class:
+
+  pool = DescriptorPool()
+  file_descriptor_protos = [ ... ]
+  for file_descriptor_proto in file_descriptor_protos:
+    pool.Add(file_descriptor_proto)
+  my_message_descriptor = pool.FindMessageTypeByName('some.package.MessageType')
+
+The message descriptor can be used in conjunction with the message_factory
+module in order to create a protocol buffer class that can be encoded and
+decoded.
+
+If you want to get a Python class for the specified proto, use the
+helper functions inside google.protobuf.message_factory
+directly instead of this class.
+"""
+
+__author__ = 'matthewtoia@google.com (Matt Toia)'
+
+from google.protobuf import descriptor
+from google.protobuf import descriptor_database
+from google.protobuf import text_encoding
+
+
+_USE_C_DESCRIPTORS = descriptor._USE_C_DESCRIPTORS
+
+
+def _NormalizeFullyQualifiedName(name):
+  """Remove leading period from fully-qualified type name.
+
+  Due to b/13860351 in descriptor_database.py, types in the root namespace are
+  generated with a leading period. This function removes that prefix.
+
+  Args:
+    name: A str, the fully-qualified symbol name.
+
+  Returns:
+    A str, the normalized fully-qualified symbol name.
+  """
+  return name.lstrip('.')
+
+
+class DescriptorPool(object):
+  """A collection of protobufs dynamically constructed by descriptor protos."""
+
+  if _USE_C_DESCRIPTORS:
+
+    def __new__(cls, descriptor_db=None):
+      # pylint: disable=protected-access
+      return descriptor._message.DescriptorPool(descriptor_db)
+
+  def __init__(self, descriptor_db=None):
+    """Initializes a Pool of proto buffs.
+
+    The descriptor_db argument to the constructor is provided to allow
+    specialized file descriptor proto lookup code to be triggered on demand. An
+    example would be an implementation which will read and compile a file
+    specified in a call to FindFileByName() and not require the call to Add()
+    at all. Results from this database will be cached internally here as well.
+
+    Args:
+      descriptor_db: A secondary source of file descriptors.
+    """
+
+    self._internal_db = descriptor_database.DescriptorDatabase()
+    self._descriptor_db = descriptor_db
+    self._descriptors = {}
+    self._enum_descriptors = {}
+    self._file_descriptors = {}
+
+  def Add(self, file_desc_proto):
+    """Adds the FileDescriptorProto and its types to this pool.
+
+    Args:
+      file_desc_proto: The FileDescriptorProto to add.
+    """
+
+    self._internal_db.Add(file_desc_proto)
+
+  def AddSerializedFile(self, serialized_file_desc_proto):
+    """Adds the FileDescriptorProto and its types to this pool.
+
+    Args:
+      serialized_file_desc_proto: A bytes string, serialization of the
+        FileDescriptorProto to add.
+    """
+
+    # pylint: disable=g-import-not-at-top
+    from google.protobuf import descriptor_pb2
+    file_desc_proto = descriptor_pb2.FileDescriptorProto.FromString(
+        serialized_file_desc_proto)
+    self.Add(file_desc_proto)
+
+  def AddDescriptor(self, desc):
+    """Adds a Descriptor to the pool, non-recursively.
+
+    If the Descriptor contains nested messages or enums, the caller must
+    explicitly register them. This method also registers the FileDescriptor
+    associated with the message.
+
+    Args:
+      desc: A Descriptor.
+    """
+    if not isinstance(desc, descriptor.Descriptor):
+      raise TypeError('Expected instance of descriptor.Descriptor.')
+
+    self._descriptors[desc.full_name] = desc
+    self.AddFileDescriptor(desc.file)
+
+  def AddEnumDescriptor(self, enum_desc):
+    """Adds an EnumDescriptor to the pool.
+
+    This method also registers the FileDescriptor associated with the message.
+
+    Args:
+      enum_desc: An EnumDescriptor.
+    """
+
+    if not isinstance(enum_desc, descriptor.EnumDescriptor):
+      raise TypeError('Expected instance of descriptor.EnumDescriptor.')
+
+    self._enum_descriptors[enum_desc.full_name] = enum_desc
+    self.AddFileDescriptor(enum_desc.file)
+
+  def AddFileDescriptor(self, file_desc):
+    """Adds a FileDescriptor to the pool, non-recursively.
+
+    If the FileDescriptor contains messages or enums, the caller must explicitly
+    register them.
+
+    Args:
+      file_desc: A FileDescriptor.
+    """
+
+    if not isinstance(file_desc, descriptor.FileDescriptor):
+      raise TypeError('Expected instance of descriptor.FileDescriptor.')
+    self._file_descriptors[file_desc.name] = file_desc
+
+  def FindFileByName(self, file_name):
+    """Gets a FileDescriptor by file name.
+
+    Args:
+      file_name: The path to the file to get a descriptor for.
+
+    Returns:
+      A FileDescriptor for the named file.
+
+    Raises:
+      KeyError: if the file can not be found in the pool.
+    """
+
+    try:
+      return self._file_descriptors[file_name]
+    except KeyError:
+      pass
+
+    try:
+      file_proto = self._internal_db.FindFileByName(file_name)
+    except KeyError as error:
+      if self._descriptor_db:
+        file_proto = self._descriptor_db.FindFileByName(file_name)
+      else:
+        raise error
+    if not file_proto:
+      raise KeyError('Cannot find a file named %s' % file_name)
+    return self._ConvertFileProtoToFileDescriptor(file_proto)
+
+  def FindFileContainingSymbol(self, symbol):
+    """Gets the FileDescriptor for the file containing the specified symbol.
+
+    Args:
+      symbol: The name of the symbol to search for.
+
+    Returns:
+      A FileDescriptor that contains the specified symbol.
+
+    Raises:
+      KeyError: if the file can not be found in the pool.
+    """
+
+    symbol = _NormalizeFullyQualifiedName(symbol)
+    try:
+      return self._descriptors[symbol].file
+    except KeyError:
+      pass
+
+    try:
+      return self._enum_descriptors[symbol].file
+    except KeyError:
+      pass
+
+    try:
+      file_proto = self._internal_db.FindFileContainingSymbol(symbol)
+    except KeyError as error:
+      if self._descriptor_db:
+        file_proto = self._descriptor_db.FindFileContainingSymbol(symbol)
+      else:
+        raise error
+    if not file_proto:
+      raise KeyError('Cannot find a file containing %s' % symbol)
+    return self._ConvertFileProtoToFileDescriptor(file_proto)
+
+  def FindMessageTypeByName(self, full_name):
+    """Loads the named descriptor from the pool.
+
+    Args:
+      full_name: The full name of the descriptor to load.
+
+    Returns:
+      The descriptor for the named type.
+    """
+
+    full_name = _NormalizeFullyQualifiedName(full_name)
+    if full_name not in self._descriptors:
+      self.FindFileContainingSymbol(full_name)
+    return self._descriptors[full_name]
+
+  def FindEnumTypeByName(self, full_name):
+    """Loads the named enum descriptor from the pool.
+
+    Args:
+      full_name: The full name of the enum descriptor to load.
+
+    Returns:
+      The enum descriptor for the named type.
+    """
+
+    full_name = _NormalizeFullyQualifiedName(full_name)
+    if full_name not in self._enum_descriptors:
+      self.FindFileContainingSymbol(full_name)
+    return self._enum_descriptors[full_name]
+
+  def FindFieldByName(self, full_name):
+    """Loads the named field descriptor from the pool.
+
+    Args:
+      full_name: The full name of the field descriptor to load.
+
+    Returns:
+      The field descriptor for the named field.
+    """
+    full_name = _NormalizeFullyQualifiedName(full_name)
+    message_name, _, field_name = full_name.rpartition('.')
+    message_descriptor = self.FindMessageTypeByName(message_name)
+    return message_descriptor.fields_by_name[field_name]
+
+  def FindExtensionByName(self, full_name):
+    """Loads the named extension descriptor from the pool.
+
+    Args:
+      full_name: The full name of the extension descriptor to load.
+
+    Returns:
+      A FieldDescriptor, describing the named extension.
+    """
+    full_name = _NormalizeFullyQualifiedName(full_name)
+    message_name, _, extension_name = full_name.rpartition('.')
+    try:
+      # Most extensions are nested inside a message.
+      scope = self.FindMessageTypeByName(message_name)
+    except KeyError:
+      # Some extensions are defined at file scope.
+      scope = self.FindFileContainingSymbol(full_name)
+    return scope.extensions_by_name[extension_name]
+
+  def _ConvertFileProtoToFileDescriptor(self, file_proto):
+    """Creates a FileDescriptor from a proto or returns a cached copy.
+
+    This method also has the side effect of loading all the symbols found in
+    the file into the appropriate dictionaries in the pool.
+
+    Args:
+      file_proto: The proto to convert.
+
+    Returns:
+      A FileDescriptor matching the passed in proto.
+    """
+
+    if file_proto.name not in self._file_descriptors:
+      built_deps = list(self._GetDeps(file_proto.dependency))
+      direct_deps = [self.FindFileByName(n) for n in file_proto.dependency]
+
+      file_descriptor = descriptor.FileDescriptor(
+          pool=self,
+          name=file_proto.name,
+          package=file_proto.package,
+          syntax=file_proto.syntax,
+          options=file_proto.options,
+          serialized_pb=file_proto.SerializeToString(),
+          dependencies=direct_deps)
+      if _USE_C_DESCRIPTORS:
+        # When using C++ descriptors, all objects defined in the file were added
+        # to the C++ database when the FileDescriptor was built above.
+        # Just add them to this descriptor pool.
+        def _AddMessageDescriptor(message_desc):
+          self._descriptors[message_desc.full_name] = message_desc
+          for nested in message_desc.nested_types:
+            _AddMessageDescriptor(nested)
+          for enum_type in message_desc.enum_types:
+            _AddEnumDescriptor(enum_type)
+        def _AddEnumDescriptor(enum_desc):
+          self._enum_descriptors[enum_desc.full_name] = enum_desc
+        for message_type in file_descriptor.message_types_by_name.values():
+          _AddMessageDescriptor(message_type)
+        for enum_type in file_descriptor.enum_types_by_name.values():
+          _AddEnumDescriptor(enum_type)
+      else:
+        scope = {}
+
+        # This loop extracts all the message and enum types from all the
+        # dependencies of the file_proto. This is necessary to create the
+        # scope of available message types when defining the passed in
+        # file proto.
+        for dependency in built_deps:
+          scope.update(self._ExtractSymbols(
+              dependency.message_types_by_name.values()))
+          scope.update((_PrefixWithDot(enum.full_name), enum)
+                       for enum in dependency.enum_types_by_name.values())
+
+        for message_type in file_proto.message_type:
+          message_desc = self._ConvertMessageDescriptor(
+              message_type, file_proto.package, file_descriptor, scope,
+              file_proto.syntax)
+          file_descriptor.message_types_by_name[message_desc.name] = (
+              message_desc)
+
+        for enum_type in file_proto.enum_type:
+          file_descriptor.enum_types_by_name[enum_type.name] = (
+              self._ConvertEnumDescriptor(enum_type, file_proto.package,
+                                          file_descriptor, None, scope))
+
+        for index, extension_proto in enumerate(file_proto.extension):
+          extension_desc = self._MakeFieldDescriptor(
+              extension_proto, file_proto.package, index, is_extension=True)
+          extension_desc.containing_type = self._GetTypeFromScope(
+              file_descriptor.package, extension_proto.extendee, scope)
+          self._SetFieldType(extension_proto, extension_desc,
+                            file_descriptor.package, scope)
+          file_descriptor.extensions_by_name[extension_desc.name] = (
+              extension_desc)
+
+        for desc_proto in file_proto.message_type:
+          self._SetAllFieldTypes(file_proto.package, desc_proto, scope)
+
+        if file_proto.package:
+          desc_proto_prefix = _PrefixWithDot(file_proto.package)
+        else:
+          desc_proto_prefix = ''
+
+        for desc_proto in file_proto.message_type:
+          desc = self._GetTypeFromScope(
+              desc_proto_prefix, desc_proto.name, scope)
+          file_descriptor.message_types_by_name[desc_proto.name] = desc
+
+      self.Add(file_proto)
+      self._file_descriptors[file_proto.name] = file_descriptor
+
+    return self._file_descriptors[file_proto.name]
+
+  def _ConvertMessageDescriptor(self, desc_proto, package=None, file_desc=None,
+                                scope=None, syntax=None):
+    """Adds the proto to the pool in the specified package.
+
+    Args:
+      desc_proto: The descriptor_pb2.DescriptorProto protobuf message.
+      package: The package the proto should be located in.
+      file_desc: The file containing this message.
+      scope: Dict mapping short and full symbols to message and enum types.
+
+    Returns:
+      The added descriptor.
+    """
+
+    if package:
+      desc_name = '.'.join((package, desc_proto.name))
+    else:
+      desc_name = desc_proto.name
+
+    if file_desc is None:
+      file_name = None
+    else:
+      file_name = file_desc.name
+
+    if scope is None:
+      scope = {}
+
+    nested = [
+        self._ConvertMessageDescriptor(
+            nested, desc_name, file_desc, scope, syntax)
+        for nested in desc_proto.nested_type]
+    enums = [
+        self._ConvertEnumDescriptor(enum, desc_name, file_desc, None, scope)
+        for enum in desc_proto.enum_type]
+    fields = [self._MakeFieldDescriptor(field, desc_name, index)
+              for index, field in enumerate(desc_proto.field)]
+    extensions = [
+        self._MakeFieldDescriptor(extension, desc_name, index,
+                                  is_extension=True)
+        for index, extension in enumerate(desc_proto.extension)]
+    oneofs = [
+        descriptor.OneofDescriptor(desc.name, '.'.join((desc_name, desc.name)),
+                                   index, None, [])
+        for index, desc in enumerate(desc_proto.oneof_decl)]
+    extension_ranges = [(r.start, r.end) for r in desc_proto.extension_range]
+    if extension_ranges:
+      is_extendable = True
+    else:
+      is_extendable = False
+    desc = descriptor.Descriptor(
+        name=desc_proto.name,
+        full_name=desc_name,
+        filename=file_name,
+        containing_type=None,
+        fields=fields,
+        oneofs=oneofs,
+        nested_types=nested,
+        enum_types=enums,
+        extensions=extensions,
+        options=desc_proto.options,
+        is_extendable=is_extendable,
+        extension_ranges=extension_ranges,
+        file=file_desc,
+        serialized_start=None,
+        serialized_end=None,
+        syntax=syntax)
+    for nested in desc.nested_types:
+      nested.containing_type = desc
+    for enum in desc.enum_types:
+      enum.containing_type = desc
+    for field_index, field_desc in enumerate(desc_proto.field):
+      if field_desc.HasField('oneof_index'):
+        oneof_index = field_desc.oneof_index
+        oneofs[oneof_index].fields.append(fields[field_index])
+        fields[field_index].containing_oneof = oneofs[oneof_index]
+
+    scope[_PrefixWithDot(desc_name)] = desc
+    self._descriptors[desc_name] = desc
+    return desc
+
+  def _ConvertEnumDescriptor(self, enum_proto, package=None, file_desc=None,
+                             containing_type=None, scope=None):
+    """Make a protobuf EnumDescriptor given an EnumDescriptorProto protobuf.
+
+    Args:
+      enum_proto: The descriptor_pb2.EnumDescriptorProto protobuf message.
+      package: Optional package name for the new message EnumDescriptor.
+      file_desc: The file containing the enum descriptor.
+      containing_type: The type containing this enum.
+      scope: Scope containing available types.
+
+    Returns:
+      The added descriptor
+    """
+
+    if package:
+      enum_name = '.'.join((package, enum_proto.name))
+    else:
+      enum_name = enum_proto.name
+
+    if file_desc is None:
+      file_name = None
+    else:
+      file_name = file_desc.name
+
+    values = [self._MakeEnumValueDescriptor(value, index)
+              for index, value in enumerate(enum_proto.value)]
+    desc = descriptor.EnumDescriptor(name=enum_proto.name,
+                                     full_name=enum_name,
+                                     filename=file_name,
+                                     file=file_desc,
+                                     values=values,
+                                     containing_type=containing_type,
+                                     options=enum_proto.options)
+    scope['.%s' % enum_name] = desc
+    self._enum_descriptors[enum_name] = desc
+    return desc
+
+  def _MakeFieldDescriptor(self, field_proto, message_name, index,
+                           is_extension=False):
+    """Creates a field descriptor from a FieldDescriptorProto.
+
+    For message and enum type fields, this method will do a look up
+    in the pool for the appropriate descriptor for that type. If it
+    is unavailable, it will fall back to the _source function to
+    create it. If this type is still unavailable, construction will
+    fail.
+
+    Args:
+      field_proto: The proto describing the field.
+      message_name: The name of the containing message.
+      index: Index of the field
+      is_extension: Indication that this field is for an extension.
+
+    Returns:
+      An initialized FieldDescriptor object
+    """
+
+    if message_name:
+      full_name = '.'.join((message_name, field_proto.name))
+    else:
+      full_name = field_proto.name
+
+    return descriptor.FieldDescriptor(
+        name=field_proto.name,
+        full_name=full_name,
+        index=index,
+        number=field_proto.number,
+        type=field_proto.type,
+        cpp_type=None,
+        message_type=None,
+        enum_type=None,
+        containing_type=None,
+        label=field_proto.label,
+        has_default_value=False,
+        default_value=None,
+        is_extension=is_extension,
+        extension_scope=None,
+        options=field_proto.options)
+
+  def _SetAllFieldTypes(self, package, desc_proto, scope):
+    """Sets all the descriptor's fields's types.
+
+    This method also sets the containing types on any extensions.
+
+    Args:
+      package: The current package of desc_proto.
+      desc_proto: The message descriptor to update.
+      scope: Enclosing scope of available types.
+    """
+
+    package = _PrefixWithDot(package)
+
+    main_desc = self._GetTypeFromScope(package, desc_proto.name, scope)
+
+    if package == '.':
+      nested_package = _PrefixWithDot(desc_proto.name)
+    else:
+      nested_package = '.'.join([package, desc_proto.name])
+
+    for field_proto, field_desc in zip(desc_proto.field, main_desc.fields):
+      self._SetFieldType(field_proto, field_desc, nested_package, scope)
+
+    for extension_proto, extension_desc in (
+        zip(desc_proto.extension, main_desc.extensions)):
+      extension_desc.containing_type = self._GetTypeFromScope(
+          nested_package, extension_proto.extendee, scope)
+      self._SetFieldType(extension_proto, extension_desc, nested_package, scope)
+
+    for nested_type in desc_proto.nested_type:
+      self._SetAllFieldTypes(nested_package, nested_type, scope)
+
+  def _SetFieldType(self, field_proto, field_desc, package, scope):
+    """Sets the field's type, cpp_type, message_type and enum_type.
+
+    Args:
+      field_proto: Data about the field in proto format.
+      field_desc: The descriptor to modiy.
+      package: The package the field's container is in.
+      scope: Enclosing scope of available types.
+    """
+    if field_proto.type_name:
+      desc = self._GetTypeFromScope(package, field_proto.type_name, scope)
+    else:
+      desc = None
+
+    if not field_proto.HasField('type'):
+      if isinstance(desc, descriptor.Descriptor):
+        field_proto.type = descriptor.FieldDescriptor.TYPE_MESSAGE
+      else:
+        field_proto.type = descriptor.FieldDescriptor.TYPE_ENUM
+
+    field_desc.cpp_type = descriptor.FieldDescriptor.ProtoTypeToCppProtoType(
+        field_proto.type)
+
+    if (field_proto.type == descriptor.FieldDescriptor.TYPE_MESSAGE
+        or field_proto.type == descriptor.FieldDescriptor.TYPE_GROUP):
+      field_desc.message_type = desc
+
+    if field_proto.type == descriptor.FieldDescriptor.TYPE_ENUM:
+      field_desc.enum_type = desc
+
+    if field_proto.label == descriptor.FieldDescriptor.LABEL_REPEATED:
+      field_desc.has_default_value = False
+      field_desc.default_value = []
+    elif field_proto.HasField('default_value'):
+      field_desc.has_default_value = True
+      if (field_proto.type == descriptor.FieldDescriptor.TYPE_DOUBLE or
+          field_proto.type == descriptor.FieldDescriptor.TYPE_FLOAT):
+        field_desc.default_value = float(field_proto.default_value)
+      elif field_proto.type == descriptor.FieldDescriptor.TYPE_STRING:
+        field_desc.default_value = field_proto.default_value
+      elif field_proto.type == descriptor.FieldDescriptor.TYPE_BOOL:
+        field_desc.default_value = field_proto.default_value.lower() == 'true'
+      elif field_proto.type == descriptor.FieldDescriptor.TYPE_ENUM:
+        field_desc.default_value = field_desc.enum_type.values_by_name[
+            field_proto.default_value].number
+      elif field_proto.type == descriptor.FieldDescriptor.TYPE_BYTES:
+        field_desc.default_value = text_encoding.CUnescape(
+            field_proto.default_value)
+      else:
+        # All other types are of the "int" type.
+        field_desc.default_value = int(field_proto.default_value)
+    else:
+      field_desc.has_default_value = False
+      if (field_proto.type == descriptor.FieldDescriptor.TYPE_DOUBLE or
+          field_proto.type == descriptor.FieldDescriptor.TYPE_FLOAT):
+        field_desc.default_value = 0.0
+      elif field_proto.type == descriptor.FieldDescriptor.TYPE_STRING:
+        field_desc.default_value = u''
+      elif field_proto.type == descriptor.FieldDescriptor.TYPE_BOOL:
+        field_desc.default_value = False
+      elif field_proto.type == descriptor.FieldDescriptor.TYPE_ENUM:
+        field_desc.default_value = field_desc.enum_type.values[0].number
+      elif field_proto.type == descriptor.FieldDescriptor.TYPE_BYTES:
+        field_desc.default_value = b''
+      else:
+        # All other types are of the "int" type.
+        field_desc.default_value = 0
+
+    field_desc.type = field_proto.type
+
+  def _MakeEnumValueDescriptor(self, value_proto, index):
+    """Creates a enum value descriptor object from a enum value proto.
+
+    Args:
+      value_proto: The proto describing the enum value.
+      index: The index of the enum value.
+
+    Returns:
+      An initialized EnumValueDescriptor object.
+    """
+
+    return descriptor.EnumValueDescriptor(
+        name=value_proto.name,
+        index=index,
+        number=value_proto.number,
+        options=value_proto.options,
+        type=None)
+
+  def _ExtractSymbols(self, descriptors):
+    """Pulls out all the symbols from descriptor protos.
+
+    Args:
+      descriptors: The messages to extract descriptors from.
+    Yields:
+      A two element tuple of the type name and descriptor object.
+    """
+
+    for desc in descriptors:
+      yield (_PrefixWithDot(desc.full_name), desc)
+      for symbol in self._ExtractSymbols(desc.nested_types):
+        yield symbol
+      for enum in desc.enum_types:
+        yield (_PrefixWithDot(enum.full_name), enum)
+
+  def _GetDeps(self, dependencies):
+    """Recursively finds dependencies for file protos.
+
+    Args:
+      dependencies: The names of the files being depended on.
+
+    Yields:
+      Each direct and indirect dependency.
+    """
+
+    for dependency in dependencies:
+      dep_desc = self.FindFileByName(dependency)
+      yield dep_desc
+      for parent_dep in dep_desc.dependencies:
+        yield parent_dep
+
+  def _GetTypeFromScope(self, package, type_name, scope):
+    """Finds a given type name in the current scope.
+
+    Args:
+      package: The package the proto should be located in.
+      type_name: The name of the type to be found in the scope.
+      scope: Dict mapping short and full symbols to message and enum types.
+
+    Returns:
+      The descriptor for the requested type.
+    """
+    if type_name not in scope:
+      components = _PrefixWithDot(package).split('.')
+      while components:
+        possible_match = '.'.join(components + [type_name])
+        if possible_match in scope:
+          type_name = possible_match
+          break
+        else:
+          components.pop(-1)
+    return scope[type_name]
+
+
+def _PrefixWithDot(name):
+  return name if name.startswith('.') else '.%s' % name
+
+
+if _USE_C_DESCRIPTORS:
+  # TODO(amauryfa): This pool could be constructed from Python code, when we
+  # support a flag like 'use_cpp_generated_pool=True'.
+  # pylint: disable=protected-access
+  _DEFAULT = descriptor._message.default_pool
+else:
+  _DEFAULT = DescriptorPool()
+
+
+def Default():
+  return _DEFAULT
diff --git a/third_party/protobuf/python/google/protobuf/internal/__init__.py b/third_party/protobuf/python/google/protobuf/internal/__init__.py
new file mode 100755
index 0000000000..e69de29bb2
--- /dev/null
+++ b/third_party/protobuf/python/google/protobuf/internal/__init__.py
diff --git a/third_party/protobuf/python/google/protobuf/internal/_parameterized.py b/third_party/protobuf/python/google/protobuf/internal/_parameterized.py
new file mode 100755
index 0000000000..dea3f19975
--- /dev/null
+++ b/third_party/protobuf/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/third_party/protobuf/python/google/protobuf/internal/any_test.proto b/third_party/protobuf/python/google/protobuf/internal/any_test.proto
new file mode 100644
index 0000000000..cd641ca0b8
--- /dev/null
+++ b/third_party/protobuf/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/third_party/protobuf/python/google/protobuf/internal/api_implementation.cc b/third_party/protobuf/python/google/protobuf/internal/api_implementation.cc
new file mode 100644
index 0000000000..6db12e8dc6
--- /dev/null
+++ b/third_party/protobuf/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/third_party/protobuf/python/google/protobuf/internal/api_implementation.py b/third_party/protobuf/python/google/protobuf/internal/api_implementation.py
new file mode 100755
index 0000000000..ffcf751167
--- /dev/null
+++ b/third_party/protobuf/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/third_party/protobuf/python/google/protobuf/internal/containers.py b/third_party/protobuf/python/google/protobuf/internal/containers.py
new file mode 100755
index 0000000000..97cdd848e3
--- /dev/null
+++ b/third_party/protobuf/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/third_party/protobuf/python/google/protobuf/internal/decoder.py b/third_party/protobuf/python/google/protobuf/internal/decoder.py
new file mode 100755
index 0000000000..31869e4575
--- /dev/null
+++ b/third_party/protobuf/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/third_party/protobuf/python/google/protobuf/internal/descriptor_database_test.py b/third_party/protobuf/python/google/protobuf/internal/descriptor_database_test.py
new file mode 100644
index 0000000000..1baff7d1d1
--- /dev/null
+++ b/third_party/protobuf/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/third_party/protobuf/python/google/protobuf/internal/descriptor_pool_test.py b/third_party/protobuf/python/google/protobuf/internal/descriptor_pool_test.py
new file mode 100644
index 0000000000..f1d6bf9900
--- /dev/null
+++ b/third_party/protobuf/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/third_party/protobuf/python/google/protobuf/internal/descriptor_pool_test1.proto b/third_party/protobuf/python/google/protobuf/internal/descriptor_pool_test1.proto
new file mode 100644
index 0000000000..00816b78ec
--- /dev/null
+++ b/third_party/protobuf/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/third_party/protobuf/python/google/protobuf/internal/descriptor_pool_test2.proto b/third_party/protobuf/python/google/protobuf/internal/descriptor_pool_test2.proto
new file mode 100644
index 0000000000..e3fa660ce1
--- /dev/null
+++ b/third_party/protobuf/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/third_party/protobuf/python/google/protobuf/internal/descriptor_test.py b/third_party/protobuf/python/google/protobuf/internal/descriptor_test.py
new file mode 100755
index 0000000000..fee09a5627
--- /dev/null
+++ b/third_party/protobuf/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/third_party/protobuf/python/google/protobuf/internal/encoder.py b/third_party/protobuf/python/google/protobuf/internal/encoder.py
new file mode 100755
index 0000000000..48ef2df31c
--- /dev/null
+++ b/third_party/protobuf/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/third_party/protobuf/python/google/protobuf/internal/enum_type_wrapper.py b/third_party/protobuf/python/google/protobuf/internal/enum_type_wrapper.py
new file mode 100644
index 0000000000..1cffe35295
--- /dev/null
+++ b/third_party/protobuf/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/third_party/protobuf/python/google/protobuf/internal/factory_test1.proto b/third_party/protobuf/python/google/protobuf/internal/factory_test1.proto
new file mode 100644
index 0000000000..d2fbbeecf1
--- /dev/null
+++ b/third_party/protobuf/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/third_party/protobuf/python/google/protobuf/internal/factory_test2.proto b/third_party/protobuf/python/google/protobuf/internal/factory_test2.proto
new file mode 100644
index 0000000000..bb1b54ada2
--- /dev/null
+++ b/third_party/protobuf/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/third_party/protobuf/python/google/protobuf/internal/generator_test.py b/third_party/protobuf/python/google/protobuf/internal/generator_test.py
new file mode 100755
index 0000000000..9956da59b4
--- /dev/null
+++ b/third_party/protobuf/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/third_party/protobuf/python/google/protobuf/internal/import_test_package/__init__.py b/third_party/protobuf/python/google/protobuf/internal/import_test_package/__init__.py
new file mode 100644
index 0000000000..5121dd0ec5
--- /dev/null
+++ b/third_party/protobuf/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/third_party/protobuf/python/google/protobuf/internal/import_test_package/inner.proto b/third_party/protobuf/python/google/protobuf/internal/import_test_package/inner.proto
new file mode 100644
index 0000000000..2887c1230e
--- /dev/null
+++ b/third_party/protobuf/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/third_party/protobuf/python/google/protobuf/internal/import_test_package/outer.proto b/third_party/protobuf/python/google/protobuf/internal/import_test_package/outer.proto
new file mode 100644
index 0000000000..a27fb5c8f4
--- /dev/null
+++ b/third_party/protobuf/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/third_party/protobuf/python/google/protobuf/internal/json_format_test.py b/third_party/protobuf/python/google/protobuf/internal/json_format_test.py
new file mode 100644
index 0000000000..49e96a46a8
--- /dev/null
+++ b/third_party/protobuf/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/third_party/protobuf/python/google/protobuf/internal/message_factory_test.py b/third_party/protobuf/python/google/protobuf/internal/message_factory_test.py
new file mode 100644
index 0000000000..2fbe5ea7ac
--- /dev/null
+++ b/third_party/protobuf/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/third_party/protobuf/python/google/protobuf/internal/message_listener.py b/third_party/protobuf/python/google/protobuf/internal/message_listener.py
new file mode 100755
index 0000000000..0fc255a774
--- /dev/null
+++ b/third_party/protobuf/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/third_party/protobuf/python/google/protobuf/internal/message_set_extensions.proto b/third_party/protobuf/python/google/protobuf/internal/message_set_extensions.proto
new file mode 100644
index 0000000000..14e5f19375
--- /dev/null
+++ b/third_party/protobuf/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/third_party/protobuf/python/google/protobuf/internal/message_test.py b/third_party/protobuf/python/google/protobuf/internal/message_test.py
new file mode 100755
index 0000000000..d03f2d25db
--- /dev/null
+++ b/third_party/protobuf/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/third_party/protobuf/python/google/protobuf/internal/missing_enum_values.proto b/third_party/protobuf/python/google/protobuf/internal/missing_enum_values.proto
new file mode 100644
index 0000000000..1850be5bb7
--- /dev/null
+++ b/third_party/protobuf/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/third_party/protobuf/python/google/protobuf/internal/more_extensions.proto b/third_party/protobuf/python/google/protobuf/internal/more_extensions.proto
new file mode 100644
index 0000000000..78f1467361
--- /dev/null
+++ b/third_party/protobuf/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/third_party/protobuf/python/google/protobuf/internal/more_extensions_dynamic.proto b/third_party/protobuf/python/google/protobuf/internal/more_extensions_dynamic.proto
new file mode 100644
index 0000000000..11f85ef60c
--- /dev/null
+++ b/third_party/protobuf/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/third_party/protobuf/python/google/protobuf/internal/more_messages.proto b/third_party/protobuf/python/google/protobuf/internal/more_messages.proto
new file mode 100644
index 0000000000..2c6ab9efdf
--- /dev/null
+++ b/third_party/protobuf/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/third_party/protobuf/python/google/protobuf/internal/packed_field_test.proto b/third_party/protobuf/python/google/protobuf/internal/packed_field_test.proto
new file mode 100644
index 0000000000..0dfdc10a87
--- /dev/null
+++ b/third_party/protobuf/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/third_party/protobuf/python/google/protobuf/internal/proto_builder_test.py b/third_party/protobuf/python/google/protobuf/internal/proto_builder_test.py
new file mode 100644
index 0000000000..822ad895ed
--- /dev/null
+++ b/third_party/protobuf/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/third_party/protobuf/python/google/protobuf/internal/python_message.py b/third_party/protobuf/python/google/protobuf/internal/python_message.py
new file mode 100755
index 0000000000..87f60666ab
--- /dev/null
+++ b/third_party/protobuf/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/third_party/protobuf/python/google/protobuf/internal/reflection_test.py b/third_party/protobuf/python/google/protobuf/internal/reflection_test.py
new file mode 100755
index 0000000000..752f2f5d90
--- /dev/null
+++ b/third_party/protobuf/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/third_party/protobuf/python/google/protobuf/internal/service_reflection_test.py b/third_party/protobuf/python/google/protobuf/internal/service_reflection_test.py
new file mode 100755
index 0000000000..98614b77f6
--- /dev/null
+++ b/third_party/protobuf/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/third_party/protobuf/python/google/protobuf/internal/symbol_database_test.py b/third_party/protobuf/python/google/protobuf/internal/symbol_database_test.py
new file mode 100644
index 0000000000..0cb935a880
--- /dev/null
+++ b/third_party/protobuf/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/third_party/protobuf/python/google/protobuf/internal/test_bad_identifiers.proto b/third_party/protobuf/python/google/protobuf/internal/test_bad_identifiers.proto
new file mode 100644
index 0000000000..c4860ea88a
--- /dev/null
+++ b/third_party/protobuf/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/third_party/protobuf/python/google/protobuf/internal/test_util.py b/third_party/protobuf/python/google/protobuf/internal/test_util.py
new file mode 100755
index 0000000000..ac88fa81d8
--- /dev/null
+++ b/third_party/protobuf/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/third_party/protobuf/python/google/protobuf/internal/text_encoding_test.py b/third_party/protobuf/python/google/protobuf/internal/text_encoding_test.py
new file mode 100755
index 0000000000..338a287b1b
--- /dev/null
+++ b/third_party/protobuf/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/third_party/protobuf/python/google/protobuf/internal/text_format_test.py b/third_party/protobuf/python/google/protobuf/internal/text_format_test.py
new file mode 100755
index 0000000000..0e14556c6f
--- /dev/null
+++ b/third_party/protobuf/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/third_party/protobuf/python/google/protobuf/internal/type_checkers.py b/third_party/protobuf/python/google/protobuf/internal/type_checkers.py
new file mode 100755
index 0000000000..f30ca6a80e
--- /dev/null
+++ b/third_party/protobuf/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/third_party/protobuf/python/google/protobuf/internal/unknown_fields_test.py b/third_party/protobuf/python/google/protobuf/internal/unknown_fields_test.py
new file mode 100755
index 0000000000..9685b8b4f0
--- /dev/null
+++ b/third_party/protobuf/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/third_party/protobuf/python/google/protobuf/internal/well_known_types.py b/third_party/protobuf/python/google/protobuf/internal/well_known_types.py
new file mode 100644
index 0000000000..d35fcc5f45
--- /dev/null
+++ b/third_party/protobuf/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/third_party/protobuf/python/google/protobuf/internal/well_known_types_test.py b/third_party/protobuf/python/google/protobuf/internal/well_known_types_test.py
new file mode 100644
index 0000000000..6acbee227c
--- /dev/null
+++ b/third_party/protobuf/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/third_party/protobuf/python/google/protobuf/internal/wire_format.py b/third_party/protobuf/python/google/protobuf/internal/wire_format.py
new file mode 100755
index 0000000000..883f525585
--- /dev/null
+++ b/third_party/protobuf/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/third_party/protobuf/python/google/protobuf/internal/wire_format_test.py b/third_party/protobuf/python/google/protobuf/internal/wire_format_test.py
new file mode 100755
index 0000000000..f659d18e42
--- /dev/null
+++ b/third_party/protobuf/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()
diff --git a/third_party/protobuf/python/google/protobuf/json_format.py b/third_party/protobuf/python/google/protobuf/json_format.py
new file mode 100644
index 0000000000..23382bdb05
--- /dev/null
+++ b/third_party/protobuf/python/google/protobuf/json_format.py
@@ -0,0 +1,645 @@
+# 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 routines for printing protocol messages in JSON format.
+
+Simple usage example:
+
+  # Create a proto object and serialize it to a json format string.
+  message = my_proto_pb2.MyMessage(foo='bar')
+  json_string = json_format.MessageToJson(message)
+
+  # Parse a json format string to proto object.
+  message = json_format.Parse(json_string, my_proto_pb2.MyMessage())
+"""
+
+__author__ = 'jieluo@google.com (Jie Luo)'
+
+import base64
+import json
+import math
+import six
+import sys
+
+from google.protobuf import descriptor
+from google.protobuf import symbol_database
+
+_TIMESTAMPFOMAT = '%Y-%m-%dT%H:%M:%S'
+_INT_TYPES = frozenset([descriptor.FieldDescriptor.CPPTYPE_INT32,
+                        descriptor.FieldDescriptor.CPPTYPE_UINT32,
+                        descriptor.FieldDescriptor.CPPTYPE_INT64,
+                        descriptor.FieldDescriptor.CPPTYPE_UINT64])
+_INT64_TYPES = frozenset([descriptor.FieldDescriptor.CPPTYPE_INT64,
+                          descriptor.FieldDescriptor.CPPTYPE_UINT64])
+_FLOAT_TYPES = frozenset([descriptor.FieldDescriptor.CPPTYPE_FLOAT,
+                          descriptor.FieldDescriptor.CPPTYPE_DOUBLE])
+_INFINITY = 'Infinity'
+_NEG_INFINITY = '-Infinity'
+_NAN = 'NaN'
+
+
+class Error(Exception):
+  """Top-level module error for json_format."""
+
+
+class SerializeToJsonError(Error):
+  """Thrown if serialization to JSON fails."""
+
+
+class ParseError(Error):
+  """Thrown in case of parsing error."""
+
+
+def MessageToJson(message, including_default_value_fields=False):
+  """Converts protobuf message to JSON format.
+
+  Args:
+    message: The protocol buffers message instance to serialize.
+    including_default_value_fields: If True, singular primitive fields,
+        repeated fields, and map fields will always be serialized.  If
+        False, only serialize non-empty fields.  Singular message fields
+        and oneof fields are not affected by this option.
+
+  Returns:
+    A string containing the JSON formatted protocol buffer message.
+  """
+  js = _MessageToJsonObject(message, including_default_value_fields)
+  return json.dumps(js, indent=2)
+
+
+def _MessageToJsonObject(message, including_default_value_fields):
+  """Converts message to an object according to Proto3 JSON Specification."""
+  message_descriptor = message.DESCRIPTOR
+  full_name = message_descriptor.full_name
+  if _IsWrapperMessage(message_descriptor):
+    return _WrapperMessageToJsonObject(message)
+  if full_name in _WKTJSONMETHODS:
+    return _WKTJSONMETHODS[full_name][0](
+        message, including_default_value_fields)
+  js = {}
+  return _RegularMessageToJsonObject(
+      message, js, including_default_value_fields)
+
+
+def _IsMapEntry(field):
+  return (field.type == descriptor.FieldDescriptor.TYPE_MESSAGE and
+          field.message_type.has_options and
+          field.message_type.GetOptions().map_entry)
+
+
+def _RegularMessageToJsonObject(message, js, including_default_value_fields):
+  """Converts normal message according to Proto3 JSON Specification."""
+  fields = message.ListFields()
+  include_default = including_default_value_fields
+
+  try:
+    for field, value in fields:
+      name = field.camelcase_name
+      if _IsMapEntry(field):
+        # Convert a map field.
+        v_field = field.message_type.fields_by_name['value']
+        js_map = {}
+        for key in value:
+          if isinstance(key, bool):
+            if key:
+              recorded_key = 'true'
+            else:
+              recorded_key = 'false'
+          else:
+            recorded_key = key
+          js_map[recorded_key] = _FieldToJsonObject(
+              v_field, value[key], including_default_value_fields)
+        js[name] = js_map
+      elif field.label == descriptor.FieldDescriptor.LABEL_REPEATED:
+        # Convert a repeated field.
+        js[name] = [_FieldToJsonObject(field, k, include_default)
+                    for k in value]
+      else:
+        js[name] = _FieldToJsonObject(field, value, include_default)
+
+    # Serialize default value if including_default_value_fields is True.
+    if including_default_value_fields:
+      message_descriptor = message.DESCRIPTOR
+      for field in message_descriptor.fields:
+        # Singular message fields and oneof fields will not be affected.
+        if ((field.label != descriptor.FieldDescriptor.LABEL_REPEATED and
+             field.cpp_type == descriptor.FieldDescriptor.CPPTYPE_MESSAGE) or
+            field.containing_oneof):
+          continue
+        name = field.camelcase_name
+        if name in js:
+          # Skip the field which has been serailized already.
+          continue
+        if _IsMapEntry(field):
+          js[name] = {}
+        elif field.label == descriptor.FieldDescriptor.LABEL_REPEATED:
+          js[name] = []
+        else:
+          js[name] = _FieldToJsonObject(field, field.default_value)
+
+  except ValueError as e:
+    raise SerializeToJsonError(
+        'Failed to serialize {0} field: {1}.'.format(field.name, e))
+
+  return js
+
+
+def _FieldToJsonObject(
+    field, value, including_default_value_fields=False):
+  """Converts field value according to Proto3 JSON Specification."""
+  if field.cpp_type == descriptor.FieldDescriptor.CPPTYPE_MESSAGE:
+    return _MessageToJsonObject(value, including_default_value_fields)
+  elif field.cpp_type == descriptor.FieldDescriptor.CPPTYPE_ENUM:
+    enum_value = field.enum_type.values_by_number.get(value, None)
+    if enum_value is not None:
+      return enum_value.name
+    else:
+      raise SerializeToJsonError('Enum field contains an integer value '
+                                 'which can not mapped to an enum value.')
+  elif field.cpp_type == descriptor.FieldDescriptor.CPPTYPE_STRING:
+    if field.type == descriptor.FieldDescriptor.TYPE_BYTES:
+      # Use base64 Data encoding for bytes
+      return base64.b64encode(value).decode('utf-8')
+    else:
+      return value
+  elif field.cpp_type == descriptor.FieldDescriptor.CPPTYPE_BOOL:
+    return bool(value)
+  elif field.cpp_type in _INT64_TYPES:
+    return str(value)
+  elif field.cpp_type in _FLOAT_TYPES:
+    if math.isinf(value):
+      if value < 0.0:
+        return _NEG_INFINITY
+      else:
+        return _INFINITY
+    if math.isnan(value):
+      return _NAN
+  return value
+
+
+def _AnyMessageToJsonObject(message, including_default):
+  """Converts Any message according to Proto3 JSON Specification."""
+  if not message.ListFields():
+    return {}
+  js = {}
+  type_url = message.type_url
+  js['@type'] = type_url
+  sub_message = _CreateMessageFromTypeUrl(type_url)
+  sub_message.ParseFromString(message.value)
+  message_descriptor = sub_message.DESCRIPTOR
+  full_name = message_descriptor.full_name
+  if _IsWrapperMessage(message_descriptor):
+    js['value'] = _WrapperMessageToJsonObject(sub_message)
+    return js
+  if full_name in _WKTJSONMETHODS:
+    js['value'] = _WKTJSONMETHODS[full_name][0](sub_message, including_default)
+    return js
+  return _RegularMessageToJsonObject(sub_message, js, including_default)
+
+
+def _CreateMessageFromTypeUrl(type_url):
+  # TODO(jieluo): Should add a way that users can register the type resolver
+  # instead of the default one.
+  db = symbol_database.Default()
+  type_name = type_url.split('/')[-1]
+  try:
+    message_descriptor = db.pool.FindMessageTypeByName(type_name)
+  except KeyError:
+    raise TypeError(
+        'Can not find message descriptor by type_url: {0}.'.format(type_url))
+  message_class = db.GetPrototype(message_descriptor)
+  return message_class()
+
+
+def _GenericMessageToJsonObject(message, unused_including_default):
+  """Converts message by ToJsonString according to Proto3 JSON Specification."""
+  # Duration, Timestamp and FieldMask have ToJsonString method to do the
+  # convert. Users can also call the method directly.
+  return message.ToJsonString()
+
+
+def _ValueMessageToJsonObject(message, unused_including_default=False):
+  """Converts Value message according to Proto3 JSON Specification."""
+  which = message.WhichOneof('kind')
+  # If the Value message is not set treat as null_value when serialize
+  # to JSON. The parse back result will be different from original message.
+  if which is None or which == 'null_value':
+    return None
+  if which == 'list_value':
+    return _ListValueMessageToJsonObject(message.list_value)
+  if which == 'struct_value':
+    value = message.struct_value
+  else:
+    value = getattr(message, which)
+  oneof_descriptor = message.DESCRIPTOR.fields_by_name[which]
+  return _FieldToJsonObject(oneof_descriptor, value)
+
+
+def _ListValueMessageToJsonObject(message, unused_including_default=False):
+  """Converts ListValue message according to Proto3 JSON Specification."""
+  return [_ValueMessageToJsonObject(value)
+          for value in message.values]
+
+
+def _StructMessageToJsonObject(message, unused_including_default=False):
+  """Converts Struct message according to Proto3 JSON Specification."""
+  fields = message.fields
+  js = {}
+  for key in fields.keys():
+    js[key] = _ValueMessageToJsonObject(fields[key])
+  return js
+
+
+def _IsWrapperMessage(message_descriptor):
+  return message_descriptor.file.name == 'google/protobuf/wrappers.proto'
+
+
+def _WrapperMessageToJsonObject(message):
+  return _FieldToJsonObject(
+      message.DESCRIPTOR.fields_by_name['value'], message.value)
+
+
+def _DuplicateChecker(js):
+  result = {}
+  for name, value in js:
+    if name in result:
+      raise ParseError('Failed to load JSON: duplicate key {0}.'.format(name))
+    result[name] = value
+  return result
+
+
+def Parse(text, message):
+  """Parses a JSON representation of a protocol message into a message.
+
+  Args:
+    text: Message JSON representation.
+    message: A protocol beffer message to merge into.
+
+  Returns:
+    The same message passed as argument.
+
+  Raises::
+    ParseError: On JSON parsing problems.
+  """
+  if not isinstance(text, six.text_type): text = text.decode('utf-8')
+  try:
+    if sys.version_info < (2, 7):
+      # object_pair_hook is not supported before python2.7
+      js = json.loads(text)
+    else:
+      js = json.loads(text, object_pairs_hook=_DuplicateChecker)
+  except ValueError as e:
+    raise ParseError('Failed to load JSON: {0}.'.format(str(e)))
+  _ConvertMessage(js, message)
+  return message
+
+
+def _ConvertFieldValuePair(js, message):
+  """Convert field value pairs into regular message.
+
+  Args:
+    js: A JSON object to convert the field value pairs.
+    message: A regular protocol message to record the data.
+
+  Raises:
+    ParseError: In case of problems converting.
+  """
+  names = []
+  message_descriptor = message.DESCRIPTOR
+  for name in js:
+    try:
+      field = message_descriptor.fields_by_camelcase_name.get(name, None)
+      if not field:
+        raise ParseError(
+            'Message type "{0}" has no field named "{1}".'.format(
+                message_descriptor.full_name, name))
+      if name in names:
+        raise ParseError(
+            'Message type "{0}" should not have multiple "{1}" fields.'.format(
+                message.DESCRIPTOR.full_name, name))
+      names.append(name)
+      # Check no other oneof field is parsed.
+      if field.containing_oneof is not None:
+        oneof_name = field.containing_oneof.name
+        if oneof_name in names:
+          raise ParseError('Message type "{0}" should not have multiple "{1}" '
+                           'oneof fields.'.format(
+                               message.DESCRIPTOR.full_name, oneof_name))
+        names.append(oneof_name)
+
+      value = js[name]
+      if value is None:
+        message.ClearField(field.name)
+        continue
+
+      # Parse field value.
+      if _IsMapEntry(field):
+        message.ClearField(field.name)
+        _ConvertMapFieldValue(value, message, field)
+      elif field.label == descriptor.FieldDescriptor.LABEL_REPEATED:
+        message.ClearField(field.name)
+        if not isinstance(value, list):
+          raise ParseError('repeated field {0} must be in [] which is '
+                           '{1}.'.format(name, value))
+        if field.cpp_type == descriptor.FieldDescriptor.CPPTYPE_MESSAGE:
+          # Repeated message field.
+          for item in value:
+            sub_message = getattr(message, field.name).add()
+            # None is a null_value in Value.
+            if (item is None and
+                sub_message.DESCRIPTOR.full_name != 'google.protobuf.Value'):
+              raise ParseError('null is not allowed to be used as an element'
+                               ' in a repeated field.')
+            _ConvertMessage(item, sub_message)
+        else:
+          # Repeated scalar field.
+          for item in value:
+            if item is None:
+              raise ParseError('null is not allowed to be used as an element'
+                               ' in a repeated field.')
+            getattr(message, field.name).append(
+                _ConvertScalarFieldValue(item, field))
+      elif field.cpp_type == descriptor.FieldDescriptor.CPPTYPE_MESSAGE:
+        sub_message = getattr(message, field.name)
+        _ConvertMessage(value, sub_message)
+      else:
+        setattr(message, field.name, _ConvertScalarFieldValue(value, field))
+    except ParseError as e:
+      if field and field.containing_oneof is None:
+        raise ParseError('Failed to parse {0} field: {1}'.format(name, e))
+      else:
+        raise ParseError(str(e))
+    except ValueError as e:
+      raise ParseError('Failed to parse {0} field: {1}.'.format(name, e))
+    except TypeError as e:
+      raise ParseError('Failed to parse {0} field: {1}.'.format(name, e))
+
+
+def _ConvertMessage(value, message):
+  """Convert a JSON object into a message.
+
+  Args:
+    value: A JSON object.
+    message: A WKT or regular protocol message to record the data.
+
+  Raises:
+    ParseError: In case of convert problems.
+  """
+  message_descriptor = message.DESCRIPTOR
+  full_name = message_descriptor.full_name
+  if _IsWrapperMessage(message_descriptor):
+    _ConvertWrapperMessage(value, message)
+  elif full_name in _WKTJSONMETHODS:
+    _WKTJSONMETHODS[full_name][1](value, message)
+  else:
+    _ConvertFieldValuePair(value, message)
+
+
+def _ConvertAnyMessage(value, message):
+  """Convert a JSON representation into Any message."""
+  if isinstance(value, dict) and not value:
+    return
+  try:
+    type_url = value['@type']
+  except KeyError:
+    raise ParseError('@type is missing when parsing any message.')
+
+  sub_message = _CreateMessageFromTypeUrl(type_url)
+  message_descriptor = sub_message.DESCRIPTOR
+  full_name = message_descriptor.full_name
+  if _IsWrapperMessage(message_descriptor):
+    _ConvertWrapperMessage(value['value'], sub_message)
+  elif full_name in _WKTJSONMETHODS:
+    _WKTJSONMETHODS[full_name][1](value['value'], sub_message)
+  else:
+    del value['@type']
+    _ConvertFieldValuePair(value, sub_message)
+  # Sets Any message
+  message.value = sub_message.SerializeToString()
+  message.type_url = type_url
+
+
+def _ConvertGenericMessage(value, message):
+  """Convert a JSON representation into message with FromJsonString."""
+  # Durantion, Timestamp, FieldMask have FromJsonString method to do the
+  # convert. Users can also call the method directly.
+  message.FromJsonString(value)
+
+
+_INT_OR_FLOAT = six.integer_types + (float,)
+
+
+def _ConvertValueMessage(value, message):
+  """Convert a JSON representation into Value message."""
+  if isinstance(value, dict):
+    _ConvertStructMessage(value, message.struct_value)
+  elif isinstance(value, list):
+    _ConvertListValueMessage(value, message.list_value)
+  elif value is None:
+    message.null_value = 0
+  elif isinstance(value, bool):
+    message.bool_value = value
+  elif isinstance(value, six.string_types):
+    message.string_value = value
+  elif isinstance(value, _INT_OR_FLOAT):
+    message.number_value = value
+  else:
+    raise ParseError('Unexpected type for Value message.')
+
+
+def _ConvertListValueMessage(value, message):
+  """Convert a JSON representation into ListValue message."""
+  if not isinstance(value, list):
+    raise ParseError(
+        'ListValue must be in [] which is {0}.'.format(value))
+  message.ClearField('values')
+  for item in value:
+    _ConvertValueMessage(item, message.values.add())
+
+
+def _ConvertStructMessage(value, message):
+  """Convert a JSON representation into Struct message."""
+  if not isinstance(value, dict):
+    raise ParseError(
+        'Struct must be in a dict which is {0}.'.format(value))
+  for key in value:
+    _ConvertValueMessage(value[key], message.fields[key])
+  return
+
+
+def _ConvertWrapperMessage(value, message):
+  """Convert a JSON representation into Wrapper message."""
+  field = message.DESCRIPTOR.fields_by_name['value']
+  setattr(message, 'value', _ConvertScalarFieldValue(value, field))
+
+
+def _ConvertMapFieldValue(value, message, field):
+  """Convert map field value for a message map field.
+
+  Args:
+    value: A JSON object to convert the map field value.
+    message: A protocol message to record the converted data.
+    field: The descriptor of the map field to be converted.
+
+  Raises:
+    ParseError: In case of convert problems.
+  """
+  if not isinstance(value, dict):
+    raise ParseError(
+        'Map field {0} must be in a dict which is {1}.'.format(
+            field.name, value))
+  key_field = field.message_type.fields_by_name['key']
+  value_field = field.message_type.fields_by_name['value']
+  for key in value:
+    key_value = _ConvertScalarFieldValue(key, key_field, True)
+    if value_field.cpp_type == descriptor.FieldDescriptor.CPPTYPE_MESSAGE:
+      _ConvertMessage(value[key], getattr(message, field.name)[key_value])
+    else:
+      getattr(message, field.name)[key_value] = _ConvertScalarFieldValue(
+          value[key], value_field)
+
+
+def _ConvertScalarFieldValue(value, field, require_str=False):
+  """Convert a single scalar field value.
+
+  Args:
+    value: A scalar value to convert the scalar field value.
+    field: The descriptor of the field to convert.
+    require_str: If True, the field value must be a str.
+
+  Returns:
+    The converted scalar field value
+
+  Raises:
+    ParseError: In case of convert problems.
+  """
+  if field.cpp_type in _INT_TYPES:
+    return _ConvertInteger(value)
+  elif field.cpp_type in _FLOAT_TYPES:
+    return _ConvertFloat(value)
+  elif field.cpp_type == descriptor.FieldDescriptor.CPPTYPE_BOOL:
+    return _ConvertBool(value, require_str)
+  elif field.cpp_type == descriptor.FieldDescriptor.CPPTYPE_STRING:
+    if field.type == descriptor.FieldDescriptor.TYPE_BYTES:
+      return base64.b64decode(value)
+    else:
+      return value
+  elif field.cpp_type == descriptor.FieldDescriptor.CPPTYPE_ENUM:
+    # Convert an enum value.
+    enum_value = field.enum_type.values_by_name.get(value, None)
+    if enum_value is None:
+      raise ParseError(
+          'Enum value must be a string literal with double quotes. '
+          'Type "{0}" has no value named {1}.'.format(
+              field.enum_type.full_name, value))
+    return enum_value.number
+
+
+def _ConvertInteger(value):
+  """Convert an integer.
+
+  Args:
+    value: A scalar value to convert.
+
+  Returns:
+    The integer value.
+
+  Raises:
+    ParseError: If an integer couldn't be consumed.
+  """
+  if isinstance(value, float):
+    raise ParseError('Couldn\'t parse integer: {0}.'.format(value))
+
+  if isinstance(value, six.text_type) and value.find(' ') != -1:
+    raise ParseError('Couldn\'t parse integer: "{0}".'.format(value))
+
+  return int(value)
+
+
+def _ConvertFloat(value):
+  """Convert an floating point number."""
+  if value == 'nan':
+    raise ParseError('Couldn\'t parse float "nan", use "NaN" instead.')
+  try:
+    # Assume Python compatible syntax.
+    return float(value)
+  except ValueError:
+    # Check alternative spellings.
+    if value == _NEG_INFINITY:
+      return float('-inf')
+    elif value == _INFINITY:
+      return float('inf')
+    elif value == _NAN:
+      return float('nan')
+    else:
+      raise ParseError('Couldn\'t parse float: {0}.'.format(value))
+
+
+def _ConvertBool(value, require_str):
+  """Convert a boolean value.
+
+  Args:
+    value: A scalar value to convert.
+    require_str: If True, value must be a str.
+
+  Returns:
+    The bool parsed.
+
+  Raises:
+    ParseError: If a boolean value couldn't be consumed.
+  """
+  if require_str:
+    if value == 'true':
+      return True
+    elif value == 'false':
+      return False
+    else:
+      raise ParseError('Expected "true" or "false", not {0}.'.format(value))
+
+  if not isinstance(value, bool):
+    raise ParseError('Expected true or false without quotes.')
+  return value
+
+_WKTJSONMETHODS = {
+    'google.protobuf.Any': [_AnyMessageToJsonObject,
+                            _ConvertAnyMessage],
+    'google.protobuf.Duration': [_GenericMessageToJsonObject,
+                                 _ConvertGenericMessage],
+    'google.protobuf.FieldMask': [_GenericMessageToJsonObject,
+                                  _ConvertGenericMessage],
+    'google.protobuf.ListValue': [_ListValueMessageToJsonObject,
+                                  _ConvertListValueMessage],
+    'google.protobuf.Struct': [_StructMessageToJsonObject,
+                               _ConvertStructMessage],
+    'google.protobuf.Timestamp': [_GenericMessageToJsonObject,
+                                  _ConvertGenericMessage],
+    'google.protobuf.Value': [_ValueMessageToJsonObject,
+                              _ConvertValueMessage]
+}
diff --git a/third_party/protobuf/python/google/protobuf/message.py b/third_party/protobuf/python/google/protobuf/message.py
new file mode 100755
index 0000000000..de2f5697e2
--- /dev/null
+++ b/third_party/protobuf/python/google/protobuf/message.py
@@ -0,0 +1,292 @@
+# 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): We should just make these methods all "pure-virtual" and move
+# all implementation out, into reflection.py for now.
+
+
+"""Contains an abstract base class for protocol messages."""
+
+__author__ = 'robinson@google.com (Will Robinson)'
+
+class Error(Exception): pass
+class DecodeError(Error): pass
+class EncodeError(Error): pass
+
+
+class Message(object):
+
+  """Abstract base class for protocol messages.
+
+  Protocol message classes are almost always generated by the protocol
+  compiler.  These generated types subclass Message and implement the methods
+  shown below.
+
+  TODO(robinson): Link to an HTML document here.
+
+  TODO(robinson): Document that instances of this class will also
+  have an Extensions attribute with __getitem__ and __setitem__.
+  Again, not sure how to best convey this.
+
+  TODO(robinson): Document that the class must also have a static
+    RegisterExtension(extension_field) method.
+    Not sure how to best express at this point.
+  """
+
+  # TODO(robinson): Document these fields and methods.
+
+  __slots__ = []
+
+  DESCRIPTOR = None
+
+  def __deepcopy__(self, memo=None):
+    clone = type(self)()
+    clone.MergeFrom(self)
+    return clone
+
+  def __eq__(self, other_msg):
+    """Recursively compares two messages by value and structure."""
+    raise NotImplementedError
+
+  def __ne__(self, other_msg):
+    # Can't just say self != other_msg, since that would infinitely recurse. :)
+    return not self == other_msg
+
+  def __hash__(self):
+    raise TypeError('unhashable object')
+
+  def __str__(self):
+    """Outputs a human-readable representation of the message."""
+    raise NotImplementedError
+
+  def __unicode__(self):
+    """Outputs a human-readable representation of the message."""
+    raise NotImplementedError
+
+  def MergeFrom(self, other_msg):
+    """Merges the contents of the specified message into current message.
+
+    This method merges the contents of the specified message into the current
+    message. Singular fields that are set in the specified message overwrite
+    the corresponding fields in the current message. Repeated fields are
+    appended. Singular sub-messages and groups are recursively merged.
+
+    Args:
+      other_msg: Message to merge into the current message.
+    """
+    raise NotImplementedError
+
+  def CopyFrom(self, other_msg):
+    """Copies the content of the specified message into the current message.
+
+    The method clears the current message and then merges the specified
+    message using MergeFrom.
+
+    Args:
+      other_msg: Message to copy into the current one.
+    """
+    if self is other_msg:
+      return
+    self.Clear()
+    self.MergeFrom(other_msg)
+
+  def Clear(self):
+    """Clears all data that was set in the message."""
+    raise NotImplementedError
+
+  def SetInParent(self):
+    """Mark this as present in the parent.
+
+    This normally happens automatically when you assign a field of a
+    sub-message, but sometimes you want to make the sub-message
+    present while keeping it empty.  If you find yourself using this,
+    you may want to reconsider your design."""
+    raise NotImplementedError
+
+  def IsInitialized(self):
+    """Checks if the message is initialized.
+
+    Returns:
+      The method returns True if the message is initialized (i.e. all of its
+      required fields are set).
+    """
+    raise NotImplementedError
+
+  # TODO(robinson): MergeFromString() should probably return None and be
+  # implemented in terms of a helper that returns the # of bytes read.  Our
+  # deserialization routines would use the helper when recursively
+  # deserializing, but the end user would almost always just want the no-return
+  # MergeFromString().
+
+  def MergeFromString(self, serialized):
+    """Merges serialized protocol buffer data into this message.
+
+    When we find a field in |serialized| that is already present
+    in this message:
+      - If it's a "repeated" field, we append to the end of our list.
+      - Else, if it's a scalar, we overwrite our field.
+      - Else, (it's a nonrepeated composite), we recursively merge
+        into the existing composite.
+
+    TODO(robinson): Document handling of unknown fields.
+
+    Args:
+      serialized: Any object that allows us to call buffer(serialized)
+        to access a string of bytes using the buffer interface.
+
+    TODO(robinson): When we switch to a helper, this will return None.
+
+    Returns:
+      The number of bytes read from |serialized|.
+      For non-group messages, this will always be len(serialized),
+      but for messages which are actually groups, this will
+      generally be less than len(serialized), since we must
+      stop when we reach an END_GROUP tag.  Note that if
+      we *do* stop because of an END_GROUP tag, the number
+      of bytes returned does not include the bytes
+      for the END_GROUP tag information.
+    """
+    raise NotImplementedError
+
+  def ParseFromString(self, serialized):
+    """Parse serialized protocol buffer data into this message.
+
+    Like MergeFromString(), except we clear the object first and
+    do not return the value that MergeFromString returns.
+    """
+    self.Clear()
+    self.MergeFromString(serialized)
+
+  def SerializeToString(self):
+    """Serializes the protocol message to a binary string.
+
+    Returns:
+      A binary string representation of the message if all of the required
+      fields in the message are set (i.e. the message is initialized).
+
+    Raises:
+      message.EncodeError if the message isn't initialized.
+    """
+    raise NotImplementedError
+
+  def SerializePartialToString(self):
+    """Serializes the protocol message to a binary string.
+
+    This method is similar to SerializeToString but doesn't check if the
+    message is initialized.
+
+    Returns:
+      A string representation of the partial message.
+    """
+    raise NotImplementedError
+
+  # TODO(robinson): Decide whether we like these better
+  # than auto-generated has_foo() and clear_foo() methods
+  # on the instances themselves.  This way is less consistent
+  # with C++, but it makes reflection-type access easier and
+  # reduces the number of magically autogenerated things.
+  #
+  # TODO(robinson): Be sure to document (and test) exactly
+  # which field names are accepted here.  Are we case-sensitive?
+  # What do we do with fields that share names with Python keywords
+  # like 'lambda' and 'yield'?
+  #
+  # nnorwitz says:
+  # """
+  # Typically (in python), an underscore is appended to names that are
+  # keywords. So they would become lambda_ or yield_.
+  # """
+  def ListFields(self):
+    """Returns a list of (FieldDescriptor, value) tuples for all
+    fields in the message which are not empty.  A singular field is non-empty
+    if HasField() would return true, and a repeated field is non-empty if
+    it contains at least one element.  The fields are ordered by field
+    number"""
+    raise NotImplementedError
+
+  def HasField(self, field_name):
+    """Checks if a certain field is set for the message, or if any field inside
+    a oneof group is set.  Note that if the field_name is not defined in the
+    message descriptor, ValueError will be raised."""
+    raise NotImplementedError
+
+  def ClearField(self, field_name):
+    """Clears the contents of a given field, or the field set inside a oneof
+    group.  If the name neither refers to a defined field or oneof group,
+    ValueError is raised."""
+    raise NotImplementedError
+
+  def WhichOneof(self, oneof_group):
+    """Returns the name of the field that is set inside a oneof group, or
+    None if no field is set.  If no group with the given name exists, ValueError
+    will be raised."""
+    raise NotImplementedError
+
+  def HasExtension(self, extension_handle):
+    raise NotImplementedError
+
+  def ClearExtension(self, extension_handle):
+    raise NotImplementedError
+
+  def ByteSize(self):
+    """Returns the serialized size of this message.
+    Recursively calls ByteSize() on all contained messages.
+    """
+    raise NotImplementedError
+
+  def _SetListener(self, message_listener):
+    """Internal method used by the protocol message implementation.
+    Clients should not call this directly.
+
+    Sets a listener that this message will call on certain state transitions.
+
+    The purpose of this method is to register back-edges from children to
+    parents at runtime, for the purpose of setting "has" bits and
+    byte-size-dirty bits in the parent and ancestor objects whenever a child or
+    descendant object is modified.
+
+    If the client wants to disconnect this Message from the object tree, she
+    explicitly sets callback to None.
+
+    If message_listener is None, unregisters any existing listener.  Otherwise,
+    message_listener must implement the MessageListener interface in
+    internal/message_listener.py, and we discard any listener registered
+    via a previous _SetListener() call.
+    """
+    raise NotImplementedError
+
+  def __getstate__(self):
+    """Support the pickle protocol."""
+    return dict(serialized=self.SerializePartialToString())
+
+  def __setstate__(self, state):
+    """Support the pickle protocol."""
+    self.__init__()
+    self.ParseFromString(state['serialized'])
diff --git a/third_party/protobuf/python/google/protobuf/message_factory.py b/third_party/protobuf/python/google/protobuf/message_factory.py
new file mode 100644
index 0000000000..1b059d130b
--- /dev/null
+++ b/third_party/protobuf/python/google/protobuf/message_factory.py
@@ -0,0 +1,147 @@
+# 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 a factory class for generating dynamic messages.
+
+The easiest way to use this class is if you have access to the FileDescriptor
+protos containing the messages you want to create you can just do the following:
+
+message_classes = message_factory.GetMessages(iterable_of_file_descriptors)
+my_proto_instance = message_classes['some.proto.package.MessageName']()
+"""
+
+__author__ = 'matthewtoia@google.com (Matt Toia)'
+
+from google.protobuf import descriptor_pool
+from google.protobuf import message
+from google.protobuf import reflection
+
+
+class MessageFactory(object):
+  """Factory for creating Proto2 messages from descriptors in a pool."""
+
+  def __init__(self, pool=None):
+    """Initializes a new factory."""
+    self.pool = pool or descriptor_pool.DescriptorPool()
+
+    # local cache of all classes built from protobuf descriptors
+    self._classes = {}
+
+  def GetPrototype(self, descriptor):
+    """Builds a proto2 message class based on the passed in descriptor.
+
+    Passing a descriptor with a fully qualified name matching a previous
+    invocation will cause the same class to be returned.
+
+    Args:
+      descriptor: The descriptor to build from.
+
+    Returns:
+      A class describing the passed in descriptor.
+    """
+    if descriptor.full_name not in self._classes:
+      descriptor_name = descriptor.name
+      if str is bytes:  # PY2
+        descriptor_name = descriptor.name.encode('ascii', 'ignore')
+      result_class = reflection.GeneratedProtocolMessageType(
+          descriptor_name,
+          (message.Message,),
+          {'DESCRIPTOR': descriptor, '__module__': None})
+          # If module not set, it wrongly points to the reflection.py module.
+      self._classes[descriptor.full_name] = result_class
+      for field in descriptor.fields:
+        if field.message_type:
+          self.GetPrototype(field.message_type)
+      for extension in result_class.DESCRIPTOR.extensions:
+        if extension.containing_type.full_name not in self._classes:
+          self.GetPrototype(extension.containing_type)
+        extended_class = self._classes[extension.containing_type.full_name]
+        extended_class.RegisterExtension(extension)
+    return self._classes[descriptor.full_name]
+
+  def GetMessages(self, files):
+    """Gets all the messages from a specified file.
+
+    This will find and resolve dependencies, failing if the descriptor
+    pool cannot satisfy them.
+
+    Args:
+      files: The file names to extract messages from.
+
+    Returns:
+      A dictionary mapping proto names to the message classes. This will include
+      any dependent messages as well as any messages defined in the same file as
+      a specified message.
+    """
+    result = {}
+    for file_name in files:
+      file_desc = self.pool.FindFileByName(file_name)
+      for name, msg in file_desc.message_types_by_name.items():
+        if file_desc.package:
+          full_name = '.'.join([file_desc.package, name])
+        else:
+          full_name = msg.name
+        result[full_name] = self.GetPrototype(
+            self.pool.FindMessageTypeByName(full_name))
+
+      # While the extension FieldDescriptors are created by the descriptor pool,
+      # the python classes created in the factory need them to be registered
+      # explicitly, which is done below.
+      #
+      # The call to RegisterExtension will specifically check if the
+      # extension was already registered on the object and either
+      # ignore the registration if the original was the same, or raise
+      # an error if they were different.
+
+      for name, extension in file_desc.extensions_by_name.items():
+        if extension.containing_type.full_name not in self._classes:
+          self.GetPrototype(extension.containing_type)
+        extended_class = self._classes[extension.containing_type.full_name]
+        extended_class.RegisterExtension(extension)
+    return result
+
+
+_FACTORY = MessageFactory()
+
+
+def GetMessages(file_protos):
+  """Builds a dictionary of all the messages available in a set of files.
+
+  Args:
+    file_protos: A sequence of file protos to build messages out of.
+
+  Returns:
+    A dictionary mapping proto names to the message classes. This will include
+    any dependent messages as well as any messages defined in the same file as
+    a specified message.
+  """
+  for file_proto in file_protos:
+    _FACTORY.pool.Add(file_proto)
+  return _FACTORY.GetMessages([file_proto.name for file_proto in file_protos])
diff --git a/third_party/protobuf/python/google/protobuf/proto_builder.py b/third_party/protobuf/python/google/protobuf/proto_builder.py
new file mode 100644
index 0000000000..736caed385
--- /dev/null
+++ b/third_party/protobuf/python/google/protobuf/proto_builder.py
@@ -0,0 +1,130 @@
+# 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.
+
+"""Dynamic Protobuf class creator."""
+
+try:
+    from collections import OrderedDict
+except ImportError:
+    from ordereddict import OrderedDict  #PY26
+import hashlib
+import os
+
+from google.protobuf import descriptor_pb2
+from google.protobuf import message_factory
+
+
+def _GetMessageFromFactory(factory, full_name):
+  """Get a proto class from the MessageFactory by name.
+
+  Args:
+    factory: a MessageFactory instance.
+    full_name: str, the fully qualified name of the proto type.
+  Returns:
+    A class, for the type identified by full_name.
+  Raises:
+    KeyError, if the proto is not found in the factory's descriptor pool.
+  """
+  proto_descriptor = factory.pool.FindMessageTypeByName(full_name)
+  proto_cls = factory.GetPrototype(proto_descriptor)
+  return proto_cls
+
+
+def MakeSimpleProtoClass(fields, full_name=None, pool=None):
+  """Create a Protobuf class whose fields are basic types.
+
+  Note: this doesn't validate field names!
+
+  Args:
+    fields: dict of {name: field_type} mappings for each field in the proto. If
+        this is an OrderedDict the order will be maintained, otherwise the
+        fields will be sorted by name.
+    full_name: optional str, the fully-qualified name of the proto type.
+    pool: optional DescriptorPool instance.
+  Returns:
+    a class, the new protobuf class with a FileDescriptor.
+  """
+  factory = message_factory.MessageFactory(pool=pool)
+
+  if full_name is not None:
+    try:
+      proto_cls = _GetMessageFromFactory(factory, full_name)
+      return proto_cls
+    except KeyError:
+      # The factory's DescriptorPool doesn't know about this class yet.
+      pass
+
+  # Get a list of (name, field_type) tuples from the fields dict. If fields was
+  # an OrderedDict we keep the order, but otherwise we sort the field to ensure
+  # consistent ordering.
+  field_items = fields.items()
+  if not isinstance(fields, OrderedDict):
+    field_items = sorted(field_items)
+
+  # Use a consistent file name that is unlikely to conflict with any imported
+  # proto files.
+  fields_hash = hashlib.sha1()
+  for f_name, f_type in field_items:
+    fields_hash.update(f_name.encode('utf-8'))
+    fields_hash.update(str(f_type).encode('utf-8'))
+  proto_file_name = fields_hash.hexdigest() + '.proto'
+
+  # If the proto is anonymous, use the same hash to name it.
+  if full_name is None:
+    full_name = ('net.proto2.python.public.proto_builder.AnonymousProto_' +
+                 fields_hash.hexdigest())
+    try:
+      proto_cls = _GetMessageFromFactory(factory, full_name)
+      return proto_cls
+    except KeyError:
+      # The factory's DescriptorPool doesn't know about this class yet.
+      pass
+
+  # This is the first time we see this proto: add a new descriptor to the pool.
+  factory.pool.Add(
+      _MakeFileDescriptorProto(proto_file_name, full_name, field_items))
+  return _GetMessageFromFactory(factory, full_name)
+
+
+def _MakeFileDescriptorProto(proto_file_name, full_name, field_items):
+  """Populate FileDescriptorProto for MessageFactory's DescriptorPool."""
+  package, name = full_name.rsplit('.', 1)
+  file_proto = descriptor_pb2.FileDescriptorProto()
+  file_proto.name = os.path.join(package.replace('.', '/'), proto_file_name)
+  file_proto.package = package
+  desc_proto = file_proto.message_type.add()
+  desc_proto.name = name
+  for f_number, (f_name, f_type) in enumerate(field_items, 1):
+    field_proto = desc_proto.field.add()
+    field_proto.name = f_name
+    field_proto.number = f_number
+    field_proto.label = descriptor_pb2.FieldDescriptorProto.LABEL_OPTIONAL
+    field_proto.type = f_type
+  return file_proto
diff --git a/third_party/protobuf/python/google/protobuf/pyext/README b/third_party/protobuf/python/google/protobuf/pyext/README
new file mode 100644
index 0000000000..6d61cb45bf
--- /dev/null
+++ b/third_party/protobuf/python/google/protobuf/pyext/README
@@ -0,0 +1,6 @@
+This is the 'v2' C++ implementation for python proto2.
+
+It is active when:
+
+PROTOCOL_BUFFERS_PYTHON_IMPLEMENTATION=cpp
+PROTOCOL_BUFFERS_PYTHON_IMPLEMENTATION_VERSION=2
diff --git a/third_party/protobuf/python/google/protobuf/pyext/__init__.py b/third_party/protobuf/python/google/protobuf/pyext/__init__.py
new file mode 100644
index 0000000000..5585614122
--- /dev/null
+++ b/third_party/protobuf/python/google/protobuf/pyext/__init__.py
@@ -0,0 +1,4 @@
+try:
+  __import__('pkg_resources').declare_namespace(__name__)
+except ImportError:
+  __path__ = __import__('pkgutil').extend_path(__path__, __name__)
diff --git a/third_party/protobuf/python/google/protobuf/pyext/cpp_message.py b/third_party/protobuf/python/google/protobuf/pyext/cpp_message.py
new file mode 100644
index 0000000000..b215211ee5
--- /dev/null
+++ b/third_party/protobuf/python/google/protobuf/pyext/cpp_message.py
@@ -0,0 +1,65 @@
+# 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.
+
+"""Protocol message implementation hooks for C++ implementation.
+
+Contains helper functions used to create protocol message classes from
+Descriptor objects at runtime backed by the protocol buffer C++ API.
+"""
+
+__author__ = 'tibell@google.com (Johan Tibell)'
+
+from google.protobuf.pyext import _message
+
+
+class GeneratedProtocolMessageType(_message.MessageMeta):
+
+  """Metaclass for protocol message classes created at runtime from Descriptors.
+
+  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'
diff --git a/third_party/protobuf/python/google/protobuf/pyext/descriptor.cc b/third_party/protobuf/python/google/protobuf/pyext/descriptor.cc
new file mode 100644
index 0000000000..a875a7be86
--- /dev/null
+++ b/third_party/protobuf/python/google/protobuf/pyext/descriptor.cc
@@ -0,0 +1,1583 @@
+// 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: petar@google.com (Petar Petrov)
+
+#include <Python.h>
+#include <frameobject.h>
+#include <string>
+
+#include <google/protobuf/io/coded_stream.h>
+#include <google/protobuf/descriptor.pb.h>
+#include <google/protobuf/dynamic_message.h>
+#include <google/protobuf/pyext/descriptor.h>
+#include <google/protobuf/pyext/descriptor_containers.h>
+#include <google/protobuf/pyext/descriptor_pool.h>
+#include <google/protobuf/pyext/message.h>
+#include <google/protobuf/pyext/scoped_pyobject_ptr.h>
+
+#if PY_MAJOR_VERSION >= 3
+  #define PyString_FromStringAndSize PyUnicode_FromStringAndSize
+  #define PyString_Check PyUnicode_Check
+  #define PyString_InternFromString PyUnicode_InternFromString
+  #define PyInt_FromLong PyLong_FromLong
+  #define PyInt_FromSize_t PyLong_FromSize_t
+  #if PY_VERSION_HEX < 0x03030000
+    #error "Python 3.0 - 3.2 are not supported."
+  #endif
+  #define PyString_AsStringAndSize(ob, charpp, sizep) \
+    (PyUnicode_Check(ob)? \
+       ((*(charpp) = PyUnicode_AsUTF8AndSize(ob, (sizep))) == NULL? -1: 0): \
+       PyBytes_AsStringAndSize(ob, (charpp), (sizep)))
+#endif
+
+namespace google {
+namespace protobuf {
+namespace python {
+
+// Store interned descriptors, so that the same C++ descriptor yields the same
+// Python object. Objects are not immortal: this map does not own the
+// references, and items are deleted when the last reference to the object is
+// released.
+// This is enough to support the "is" operator on live objects.
+// All descriptors are stored here.
+hash_map<const void*, PyObject*> interned_descriptors;
+
+PyObject* PyString_FromCppString(const string& str) {
+  return PyString_FromStringAndSize(str.c_str(), str.size());
+}
+
+// Check that the calling Python code is the global scope of a _pb2.py module.
+// This function is used to support the current code generated by the proto
+// compiler, which creates descriptors, then update some properties.
+// For example:
+//   message_descriptor = Descriptor(
+//       name='Message',
+//       fields = [FieldDescriptor(name='field')]
+//   message_descriptor.fields[0].containing_type = message_descriptor
+//
+// This code is still executed, but the descriptors now have no other storage
+// than the (const) C++ pointer, and are immutable.
+// So we let this code pass, by simply ignoring the new value.
+//
+// From user code, descriptors still look immutable.
+//
+// TODO(amauryfa): Change the proto2 compiler to remove the assignments, and
+// remove this hack.
+bool _CalledFromGeneratedFile(int stacklevel) {
+  PyThreadState *state = PyThreadState_GET();
+  if (state == NULL) {
+    return false;
+  }
+  PyFrameObject* frame = state->frame;
+  if (frame == NULL) {
+    return false;
+  }
+  while (stacklevel-- > 0) {
+    frame = frame->f_back;
+    if (frame == NULL) {
+      return false;
+    }
+  }
+  if (frame->f_globals != frame->f_locals) {
+    // Not at global module scope
+    return false;
+  }
+
+  if (frame->f_code->co_filename == NULL) {
+    return false;
+  }
+  char* filename;
+  Py_ssize_t filename_size;
+  if (PyString_AsStringAndSize(frame->f_code->co_filename,
+                               &filename, &filename_size) < 0) {
+    // filename is not a string.
+    PyErr_Clear();
+    return false;
+  }
+  if (filename_size < 7) {
+    // filename is too short.
+    return false;
+  }
+  if (strcmp(&filename[filename_size - 7], "_pb2.py") != 0) {
+    // Filename is not ending with _pb2.
+    return false;
+  }
+  return true;
+}
+
+// If the calling code is not a _pb2.py file, raise AttributeError.
+// To be used in attribute setters.
+static int CheckCalledFromGeneratedFile(const char* attr_name) {
+  if (_CalledFromGeneratedFile(0)) {
+    return 0;
+  }
+  PyErr_Format(PyExc_AttributeError,
+               "attribute is not writable: %s", attr_name);
+  return -1;
+}
+
+
+#ifndef PyVarObject_HEAD_INIT
+#define PyVarObject_HEAD_INIT(type, size) PyObject_HEAD_INIT(type) size,
+#endif
+#ifndef Py_TYPE
+#define Py_TYPE(ob) (((PyObject*)(ob))->ob_type)
+#endif
+
+
+// Helper functions for descriptor objects.
+
+// A set of templates to retrieve the C++ FileDescriptor of any descriptor.
+template<class DescriptorClass>
+const FileDescriptor* GetFileDescriptor(const DescriptorClass* descriptor) {
+  return descriptor->file();
+}
+template<>
+const FileDescriptor* GetFileDescriptor(const FileDescriptor* descriptor) {
+  return descriptor;
+}
+template<>
+const FileDescriptor* GetFileDescriptor(const EnumValueDescriptor* descriptor) {
+  return descriptor->type()->file();
+}
+template<>
+const FileDescriptor* GetFileDescriptor(const OneofDescriptor* descriptor) {
+  return descriptor->containing_type()->file();
+}
+
+// Converts options into a Python protobuf, and cache the result.
+//
+// This is a bit tricky because options can contain extension fields defined in
+// the same proto file. In this case the options parsed from the serialized_pb
+// have unkown fields, and we need to parse them again.
+//
+// Always returns a new reference.
+template<class DescriptorClass>
+static PyObject* GetOrBuildOptions(const DescriptorClass *descriptor) {
+  // Options (and their extensions) are completely resolved in the proto file
+  // containing the descriptor.
+  PyDescriptorPool* pool = GetDescriptorPool_FromPool(
+      GetFileDescriptor(descriptor)->pool());
+
+  hash_map<const void*, PyObject*>* descriptor_options =
+      pool->descriptor_options;
+  // First search in the cache.
+  if (descriptor_options->find(descriptor) != descriptor_options->end()) {
+    PyObject *value = (*descriptor_options)[descriptor];
+    Py_INCREF(value);
+    return value;
+  }
+
+  // Build the Options object: get its Python class, and make a copy of the C++
+  // read-only instance.
+  const Message& options(descriptor->options());
+  const Descriptor *message_type = options.GetDescriptor();
+  PyObject* message_class(cdescriptor_pool::GetMessageClass(
+      pool, message_type));
+  if (message_class == NULL) {
+    // The Options message was not found in the current DescriptorPool.
+    // In this case, there cannot be extensions to these options, and we can
+    // try to use the basic pool instead.
+    PyErr_Clear();
+    message_class = cdescriptor_pool::GetMessageClass(
+      GetDefaultDescriptorPool(), message_type);
+  }
+  if (message_class == NULL) {
+    PyErr_Format(PyExc_TypeError, "Could not retrieve class for Options: %s",
+                 message_type->full_name().c_str());
+    return NULL;
+  }
+  ScopedPyObjectPtr value(PyEval_CallObject(message_class, NULL));
+  if (value == NULL) {
+    return NULL;
+  }
+  if (!PyObject_TypeCheck(value.get(), &CMessage_Type)) {
+      PyErr_Format(PyExc_TypeError, "Invalid class for %s: %s",
+                   message_type->full_name().c_str(),
+                   Py_TYPE(value.get())->tp_name);
+      return NULL;
+  }
+  CMessage* cmsg = reinterpret_cast<CMessage*>(value.get());
+
+  const Reflection* reflection = options.GetReflection();
+  const UnknownFieldSet& unknown_fields(reflection->GetUnknownFields(options));
+  if (unknown_fields.empty()) {
+    cmsg->message->CopyFrom(options);
+  } else {
+    // Reparse options string!  XXX call cmessage::MergeFromString
+    string serialized;
+    options.SerializeToString(&serialized);
+    io::CodedInputStream input(
+        reinterpret_cast<const uint8*>(serialized.c_str()), serialized.size());
+    input.SetExtensionRegistry(pool->pool, pool->message_factory);
+    bool success = cmsg->message->MergePartialFromCodedStream(&input);
+    if (!success) {
+      PyErr_Format(PyExc_ValueError, "Error parsing Options message");
+      return NULL;
+    }
+  }
+
+  // Cache the result.
+  Py_INCREF(value.get());
+  (*pool->descriptor_options)[descriptor] = value.get();
+
+  return value.release();
+}
+
+// Copy the C++ descriptor to a Python message.
+// The Python message is an instance of descriptor_pb2.DescriptorProto
+// or similar.
+template<class DescriptorProtoClass, class DescriptorClass>
+static PyObject* CopyToPythonProto(const DescriptorClass *descriptor,
+                                   PyObject *target) {
+  const Descriptor* self_descriptor =
+      DescriptorProtoClass::default_instance().GetDescriptor();
+  CMessage* message = reinterpret_cast<CMessage*>(target);
+  if (!PyObject_TypeCheck(target, &CMessage_Type) ||
+      message->message->GetDescriptor() != self_descriptor) {
+    PyErr_Format(PyExc_TypeError, "Not a %s message",
+                 self_descriptor->full_name().c_str());
+    return NULL;
+  }
+  cmessage::AssureWritable(message);
+  DescriptorProtoClass* descriptor_message =
+      static_cast<DescriptorProtoClass*>(message->message);
+  descriptor->CopyTo(descriptor_message);
+  Py_RETURN_NONE;
+}
+
+// All Descriptors classes share the same memory layout.
+typedef struct PyBaseDescriptor {
+  PyObject_HEAD
+
+  // Pointer to the C++ proto2 descriptor.
+  // Like all descriptors, it is owned by the global DescriptorPool.
+  const void* descriptor;
+
+  // Owned reference to the DescriptorPool, to ensure it is kept alive.
+  PyDescriptorPool* pool;
+} PyBaseDescriptor;
+
+
+// FileDescriptor structure "inherits" from the base descriptor.
+typedef struct PyFileDescriptor {
+  PyBaseDescriptor base;
+
+  // The cached version of serialized pb. Either NULL, or a Bytes string.
+  // We own the reference.
+  PyObject *serialized_pb;
+} PyFileDescriptor;
+
+
+namespace descriptor {
+
+// Creates or retrieve a Python descriptor of the specified type.
+// Objects are interned: the same descriptor will return the same object if it
+// was kept alive.
+// 'was_created' is an optional pointer to a bool, and is set to true if a new
+// object was allocated.
+// Always return a new reference.
+template<class DescriptorClass>
+PyObject* NewInternedDescriptor(PyTypeObject* type,
+                                const DescriptorClass* descriptor,
+                                bool* was_created) {
+  if (was_created) {
+    *was_created = false;
+  }
+  if (descriptor == NULL) {
+    PyErr_BadInternalCall();
+    return NULL;
+  }
+
+  // See if the object is in the map of interned descriptors
+  hash_map<const void*, PyObject*>::iterator it =
+      interned_descriptors.find(descriptor);
+  if (it != interned_descriptors.end()) {
+    GOOGLE_DCHECK(Py_TYPE(it->second) == type);
+    Py_INCREF(it->second);
+    return it->second;
+  }
+  // Create a new descriptor object
+  PyBaseDescriptor* py_descriptor = PyObject_New(
+      PyBaseDescriptor, type);
+  if (py_descriptor == NULL) {
+    return NULL;
+  }
+  py_descriptor->descriptor = descriptor;
+
+  // and cache it.
+  interned_descriptors.insert(
+      std::make_pair(descriptor, reinterpret_cast<PyObject*>(py_descriptor)));
+
+  // Ensures that the DescriptorPool stays alive.
+  PyDescriptorPool* pool = GetDescriptorPool_FromPool(
+      GetFileDescriptor(descriptor)->pool());
+  if (pool == NULL) {
+    // Don't DECREF, the object is not fully initialized.
+    PyObject_Del(py_descriptor);
+    return NULL;
+  }
+  Py_INCREF(pool);
+  py_descriptor->pool = pool;
+
+  if (was_created) {
+    *was_created = true;
+  }
+  return reinterpret_cast<PyObject*>(py_descriptor);
+}
+
+static void Dealloc(PyBaseDescriptor* self) {
+  // Remove from interned dictionary
+  interned_descriptors.erase(self->descriptor);
+  Py_CLEAR(self->pool);
+  Py_TYPE(self)->tp_free(reinterpret_cast<PyObject*>(self));
+}
+
+static PyGetSetDef Getters[] = {
+  {NULL}
+};
+
+PyTypeObject PyBaseDescriptor_Type = {
+  PyVarObject_HEAD_INIT(&PyType_Type, 0)
+  FULL_MODULE_NAME ".DescriptorBase",   // tp_name
+  sizeof(PyBaseDescriptor),             // tp_basicsize
+  0,                                    // tp_itemsize
+  (destructor)Dealloc,                  // tp_dealloc
+  0,                                    // tp_print
+  0,                                    // tp_getattr
+  0,                                    // tp_setattr
+  0,                                    // tp_compare
+  0,                                    // tp_repr
+  0,                                    // tp_as_number
+  0,                                    // tp_as_sequence
+  0,                                    // tp_as_mapping
+  0,                                    // tp_hash
+  0,                                    // tp_call
+  0,                                    // tp_str
+  0,                                    // tp_getattro
+  0,                                    // tp_setattro
+  0,                                    // tp_as_buffer
+  Py_TPFLAGS_DEFAULT,                   // tp_flags
+  "Descriptors base class",             // tp_doc
+  0,                                    // tp_traverse
+  0,                                    // tp_clear
+  0,                                    // tp_richcompare
+  0,                                    // tp_weaklistoffset
+  0,                                    // tp_iter
+  0,                                    // tp_iternext
+  0,                                    // tp_methods
+  0,                                    // tp_members
+  Getters,                              // tp_getset
+};
+
+}  // namespace descriptor
+
+const void* PyDescriptor_AsVoidPtr(PyObject* obj) {
+  if (!PyObject_TypeCheck(obj, &descriptor::PyBaseDescriptor_Type)) {
+    PyErr_SetString(PyExc_TypeError, "Not a BaseDescriptor");
+    return NULL;
+  }
+  return reinterpret_cast<PyBaseDescriptor*>(obj)->descriptor;
+}
+
+namespace message_descriptor {
+
+// Unchecked accessor to the C++ pointer.
+static const Descriptor* _GetDescriptor(PyBaseDescriptor* self) {
+  return reinterpret_cast<const Descriptor*>(self->descriptor);
+}
+
+static PyObject* GetName(PyBaseDescriptor* self, void *closure) {
+  return PyString_FromCppString(_GetDescriptor(self)->name());
+}
+
+static PyObject* GetFullName(PyBaseDescriptor* self, void *closure) {
+  return PyString_FromCppString(_GetDescriptor(self)->full_name());
+}
+
+static PyObject* GetFile(PyBaseDescriptor *self, void *closure) {
+  return PyFileDescriptor_FromDescriptor(_GetDescriptor(self)->file());
+}
+
+static PyObject* GetConcreteClass(PyBaseDescriptor* self, void *closure) {
+  // Retuns the canonical class for the given descriptor.
+  // This is the class that was registered with the primary descriptor pool
+  // which contains this descriptor.
+  // This might not be the one you expect! For example the returned object does
+  // not know about extensions defined in a custom pool.
+  PyObject* concrete_class(cdescriptor_pool::GetMessageClass(
+      GetDescriptorPool_FromPool(_GetDescriptor(self)->file()->pool()),
+      _GetDescriptor(self)));
+  Py_XINCREF(concrete_class);
+  return concrete_class;
+}
+
+static PyObject* GetFieldsByName(PyBaseDescriptor* self, void *closure) {
+  return NewMessageFieldsByName(_GetDescriptor(self));
+}
+
+static PyObject* GetFieldsByCamelcaseName(PyBaseDescriptor* self,
+                                          void *closure) {
+  return NewMessageFieldsByCamelcaseName(_GetDescriptor(self));
+}
+
+static PyObject* GetFieldsByNumber(PyBaseDescriptor* self, void *closure) {
+  return NewMessageFieldsByNumber(_GetDescriptor(self));
+}
+
+static PyObject* GetFieldsSeq(PyBaseDescriptor* self, void *closure) {
+  return NewMessageFieldsSeq(_GetDescriptor(self));
+}
+
+static PyObject* GetNestedTypesByName(PyBaseDescriptor* self, void *closure) {
+  return NewMessageNestedTypesByName(_GetDescriptor(self));
+}
+
+static PyObject* GetNestedTypesSeq(PyBaseDescriptor* self, void *closure) {
+  return NewMessageNestedTypesSeq(_GetDescriptor(self));
+}
+
+static PyObject* GetExtensionsByName(PyBaseDescriptor* self, void *closure) {
+  return NewMessageExtensionsByName(_GetDescriptor(self));
+}
+
+static PyObject* GetExtensions(PyBaseDescriptor* self, void *closure) {
+  return NewMessageExtensionsSeq(_GetDescriptor(self));
+}
+
+static PyObject* GetEnumsSeq(PyBaseDescriptor* self, void *closure) {
+  return NewMessageEnumsSeq(_GetDescriptor(self));
+}
+
+static PyObject* GetEnumTypesByName(PyBaseDescriptor* self, void *closure) {
+  return NewMessageEnumsByName(_GetDescriptor(self));
+}
+
+static PyObject* GetEnumValuesByName(PyBaseDescriptor* self, void *closure) {
+  return NewMessageEnumValuesByName(_GetDescriptor(self));
+}
+
+static PyObject* GetOneofsByName(PyBaseDescriptor* self, void *closure) {
+  return NewMessageOneofsByName(_GetDescriptor(self));
+}
+
+static PyObject* GetOneofsSeq(PyBaseDescriptor* self, void *closure) {
+  return NewMessageOneofsSeq(_GetDescriptor(self));
+}
+
+static PyObject* IsExtendable(PyBaseDescriptor *self, void *closure) {
+  if (_GetDescriptor(self)->extension_range_count() > 0) {
+    Py_RETURN_TRUE;
+  } else {
+    Py_RETURN_FALSE;
+  }
+}
+
+static PyObject* GetExtensionRanges(PyBaseDescriptor *self, void *closure) {
+  const Descriptor* descriptor = _GetDescriptor(self);
+  PyObject* range_list = PyList_New(descriptor->extension_range_count());
+
+  for (int i = 0; i < descriptor->extension_range_count(); i++) {
+    const Descriptor::ExtensionRange* range = descriptor->extension_range(i);
+    PyObject* start = PyInt_FromLong(range->start);
+    PyObject* end = PyInt_FromLong(range->end);
+    PyList_SetItem(range_list, i, PyTuple_Pack(2, start, end));
+  }
+
+  return range_list;
+}
+
+static PyObject* GetContainingType(PyBaseDescriptor *self, void *closure) {
+  const Descriptor* containing_type =
+      _GetDescriptor(self)->containing_type();
+  if (containing_type) {
+    return PyMessageDescriptor_FromDescriptor(containing_type);
+  } else {
+    Py_RETURN_NONE;
+  }
+}
+
+static int SetContainingType(PyBaseDescriptor *self, PyObject *value,
+                             void *closure) {
+  return CheckCalledFromGeneratedFile("containing_type");
+}
+
+static PyObject* GetHasOptions(PyBaseDescriptor *self, void *closure) {
+  const MessageOptions& options(_GetDescriptor(self)->options());
+  if (&options != &MessageOptions::default_instance()) {
+    Py_RETURN_TRUE;
+  } else {
+    Py_RETURN_FALSE;
+  }
+}
+static int SetHasOptions(PyBaseDescriptor *self, PyObject *value,
+                             void *closure) {
+  return CheckCalledFromGeneratedFile("has_options");
+}
+
+static PyObject* GetOptions(PyBaseDescriptor *self) {
+  return GetOrBuildOptions(_GetDescriptor(self));
+}
+
+static int SetOptions(PyBaseDescriptor *self, PyObject *value,
+                      void *closure) {
+  return CheckCalledFromGeneratedFile("_options");
+}
+
+static PyObject* CopyToProto(PyBaseDescriptor *self, PyObject *target) {
+  return CopyToPythonProto<DescriptorProto>(_GetDescriptor(self), target);
+}
+
+static PyObject* EnumValueName(PyBaseDescriptor *self, PyObject *args) {
+  const char *enum_name;
+  int number;
+  if (!PyArg_ParseTuple(args, "si", &enum_name, &number))
+    return NULL;
+  const EnumDescriptor *enum_type =
+      _GetDescriptor(self)->FindEnumTypeByName(enum_name);
+  if (enum_type == NULL) {
+    PyErr_SetString(PyExc_KeyError, enum_name);
+    return NULL;
+  }
+  const EnumValueDescriptor *enum_value =
+      enum_type->FindValueByNumber(number);
+  if (enum_value == NULL) {
+    PyErr_Format(PyExc_KeyError, "%d", number);
+    return NULL;
+  }
+  return PyString_FromCppString(enum_value->name());
+}
+
+static PyObject* GetSyntax(PyBaseDescriptor *self, void *closure) {
+  return PyString_InternFromString(
+      FileDescriptor::SyntaxName(_GetDescriptor(self)->file()->syntax()));
+}
+
+static PyGetSetDef Getters[] = {
+  { "name", (getter)GetName, NULL, "Last name"},
+  { "full_name", (getter)GetFullName, NULL, "Full name"},
+  { "_concrete_class", (getter)GetConcreteClass, NULL, "concrete class"},
+  { "file", (getter)GetFile, NULL, "File descriptor"},
+
+  { "fields", (getter)GetFieldsSeq, NULL, "Fields sequence"},
+  { "fields_by_name", (getter)GetFieldsByName, NULL, "Fields by name"},
+  { "fields_by_camelcase_name", (getter)GetFieldsByCamelcaseName, NULL,
+    "Fields by camelCase name"},
+  { "fields_by_number", (getter)GetFieldsByNumber, NULL, "Fields by number"},
+  { "nested_types", (getter)GetNestedTypesSeq, NULL, "Nested types sequence"},
+  { "nested_types_by_name", (getter)GetNestedTypesByName, NULL,
+    "Nested types by name"},
+  { "extensions", (getter)GetExtensions, NULL, "Extensions Sequence"},
+  { "extensions_by_name", (getter)GetExtensionsByName, NULL,
+    "Extensions by name"},
+  { "extension_ranges", (getter)GetExtensionRanges, NULL, "Extension ranges"},
+  { "enum_types", (getter)GetEnumsSeq, NULL, "Enum sequence"},
+  { "enum_types_by_name", (getter)GetEnumTypesByName, NULL,
+    "Enum types by name"},
+  { "enum_values_by_name", (getter)GetEnumValuesByName, NULL,
+    "Enum values by name"},
+  { "oneofs_by_name", (getter)GetOneofsByName, NULL, "Oneofs by name"},
+  { "oneofs", (getter)GetOneofsSeq, NULL, "Oneofs by name"},
+  { "containing_type", (getter)GetContainingType, (setter)SetContainingType,
+    "Containing type"},
+  { "is_extendable", (getter)IsExtendable, (setter)NULL},
+  { "has_options", (getter)GetHasOptions, (setter)SetHasOptions, "Has Options"},
+  { "_options", (getter)NULL, (setter)SetOptions, "Options"},
+  { "syntax", (getter)GetSyntax, (setter)NULL, "Syntax"},
+  {NULL}
+};
+
+static PyMethodDef Methods[] = {
+  { "GetOptions", (PyCFunction)GetOptions, METH_NOARGS, },
+  { "CopyToProto", (PyCFunction)CopyToProto, METH_O, },
+  { "EnumValueName", (PyCFunction)EnumValueName, METH_VARARGS, },
+  {NULL}
+};
+
+}  // namespace message_descriptor
+
+PyTypeObject PyMessageDescriptor_Type = {
+  PyVarObject_HEAD_INIT(&PyType_Type, 0)
+  FULL_MODULE_NAME ".MessageDescriptor",  // tp_name
+  sizeof(PyBaseDescriptor),             // tp_basicsize
+  0,                                    // tp_itemsize
+  0,                                    // tp_dealloc
+  0,                                    // tp_print
+  0,                                    // tp_getattr
+  0,                                    // tp_setattr
+  0,                                    // tp_compare
+  0,                                    // tp_repr
+  0,                                    // tp_as_number
+  0,                                    // tp_as_sequence
+  0,                                    // tp_as_mapping
+  0,                                    // tp_hash
+  0,                                    // tp_call
+  0,                                    // tp_str
+  0,                                    // tp_getattro
+  0,                                    // tp_setattro
+  0,                                    // tp_as_buffer
+  Py_TPFLAGS_DEFAULT,                   // tp_flags
+  "A Message Descriptor",               // tp_doc
+  0,                                    // tp_traverse
+  0,                                    // tp_clear
+  0,                                    // tp_richcompare
+  0,                                    // tp_weaklistoffset
+  0,                                    // tp_iter
+  0,                                    // tp_iternext
+  message_descriptor::Methods,          // tp_methods
+  0,                                    // tp_members
+  message_descriptor::Getters,          // tp_getset
+  &descriptor::PyBaseDescriptor_Type,   // tp_base
+};
+
+PyObject* PyMessageDescriptor_FromDescriptor(
+    const Descriptor* message_descriptor) {
+  return descriptor::NewInternedDescriptor(
+      &PyMessageDescriptor_Type, message_descriptor, NULL);
+}
+
+const Descriptor* PyMessageDescriptor_AsDescriptor(PyObject* obj) {
+  if (!PyObject_TypeCheck(obj, &PyMessageDescriptor_Type)) {
+    PyErr_SetString(PyExc_TypeError, "Not a MessageDescriptor");
+    return NULL;
+  }
+  return reinterpret_cast<const Descriptor*>(
+      reinterpret_cast<PyBaseDescriptor*>(obj)->descriptor);
+}
+
+namespace field_descriptor {
+
+// Unchecked accessor to the C++ pointer.
+static const FieldDescriptor* _GetDescriptor(
+    PyBaseDescriptor *self) {
+  return reinterpret_cast<const FieldDescriptor*>(self->descriptor);
+}
+
+static PyObject* GetFullName(PyBaseDescriptor* self, void *closure) {
+  return PyString_FromCppString(_GetDescriptor(self)->full_name());
+}
+
+static PyObject* GetName(PyBaseDescriptor *self, void *closure) {
+  return PyString_FromCppString(_GetDescriptor(self)->name());
+}
+
+static PyObject* GetCamelcaseName(PyBaseDescriptor* self, void *closure) {
+  return PyString_FromCppString(_GetDescriptor(self)->camelcase_name());
+}
+
+static PyObject* GetType(PyBaseDescriptor *self, void *closure) {
+  return PyInt_FromLong(_GetDescriptor(self)->type());
+}
+
+static PyObject* GetCppType(PyBaseDescriptor *self, void *closure) {
+  return PyInt_FromLong(_GetDescriptor(self)->cpp_type());
+}
+
+static PyObject* GetLabel(PyBaseDescriptor *self, void *closure) {
+  return PyInt_FromLong(_GetDescriptor(self)->label());
+}
+
+static PyObject* GetNumber(PyBaseDescriptor *self, void *closure) {
+  return PyInt_FromLong(_GetDescriptor(self)->number());
+}
+
+static PyObject* GetIndex(PyBaseDescriptor *self, void *closure) {
+  return PyInt_FromLong(_GetDescriptor(self)->index());
+}
+
+static PyObject* GetID(PyBaseDescriptor *self, void *closure) {
+  return PyLong_FromVoidPtr(self);
+}
+
+static PyObject* IsExtension(PyBaseDescriptor *self, void *closure) {
+  return PyBool_FromLong(_GetDescriptor(self)->is_extension());
+}
+
+static PyObject* HasDefaultValue(PyBaseDescriptor *self, void *closure) {
+  return PyBool_FromLong(_GetDescriptor(self)->has_default_value());
+}
+
+static PyObject* GetDefaultValue(PyBaseDescriptor *self, void *closure) {
+  PyObject *result;
+
+  switch (_GetDescriptor(self)->cpp_type()) {
+    case FieldDescriptor::CPPTYPE_INT32: {
+      int32 value = _GetDescriptor(self)->default_value_int32();
+      result = PyInt_FromLong(value);
+      break;
+    }
+    case FieldDescriptor::CPPTYPE_INT64: {
+      int64 value = _GetDescriptor(self)->default_value_int64();
+      result = PyLong_FromLongLong(value);
+      break;
+    }
+    case FieldDescriptor::CPPTYPE_UINT32: {
+      uint32 value = _GetDescriptor(self)->default_value_uint32();
+      result = PyInt_FromSize_t(value);
+      break;
+    }
+    case FieldDescriptor::CPPTYPE_UINT64: {
+      uint64 value = _GetDescriptor(self)->default_value_uint64();
+      result = PyLong_FromUnsignedLongLong(value);
+      break;
+    }
+    case FieldDescriptor::CPPTYPE_FLOAT: {
+      float value = _GetDescriptor(self)->default_value_float();
+      result = PyFloat_FromDouble(value);
+      break;
+    }
+    case FieldDescriptor::CPPTYPE_DOUBLE: {
+      double value = _GetDescriptor(self)->default_value_double();
+      result = PyFloat_FromDouble(value);
+      break;
+    }
+    case FieldDescriptor::CPPTYPE_BOOL: {
+      bool value = _GetDescriptor(self)->default_value_bool();
+      result = PyBool_FromLong(value);
+      break;
+    }
+    case FieldDescriptor::CPPTYPE_STRING: {
+      string value = _GetDescriptor(self)->default_value_string();
+      result = ToStringObject(_GetDescriptor(self), value);
+      break;
+    }
+    case FieldDescriptor::CPPTYPE_ENUM: {
+      const EnumValueDescriptor* value =
+          _GetDescriptor(self)->default_value_enum();
+      result = PyInt_FromLong(value->number());
+      break;
+    }
+    default:
+      PyErr_Format(PyExc_NotImplementedError, "default value for %s",
+                   _GetDescriptor(self)->full_name().c_str());
+      return NULL;
+  }
+  return result;
+}
+
+static PyObject* GetCDescriptor(PyObject *self, void *closure) {
+  Py_INCREF(self);
+  return self;
+}
+
+static PyObject *GetEnumType(PyBaseDescriptor *self, void *closure) {
+  const EnumDescriptor* enum_type = _GetDescriptor(self)->enum_type();
+  if (enum_type) {
+    return PyEnumDescriptor_FromDescriptor(enum_type);
+  } else {
+    Py_RETURN_NONE;
+  }
+}
+
+static int SetEnumType(PyBaseDescriptor *self, PyObject *value, void *closure) {
+  return CheckCalledFromGeneratedFile("enum_type");
+}
+
+static PyObject *GetMessageType(PyBaseDescriptor *self, void *closure) {
+  const Descriptor* message_type = _GetDescriptor(self)->message_type();
+  if (message_type) {
+    return PyMessageDescriptor_FromDescriptor(message_type);
+  } else {
+    Py_RETURN_NONE;
+  }
+}
+
+static int SetMessageType(PyBaseDescriptor *self, PyObject *value,
+                          void *closure) {
+  return CheckCalledFromGeneratedFile("message_type");
+}
+
+static PyObject* GetContainingType(PyBaseDescriptor *self, void *closure) {
+  const Descriptor* containing_type =
+      _GetDescriptor(self)->containing_type();
+  if (containing_type) {
+    return PyMessageDescriptor_FromDescriptor(containing_type);
+  } else {
+    Py_RETURN_NONE;
+  }
+}
+
+static int SetContainingType(PyBaseDescriptor *self, PyObject *value,
+                             void *closure) {
+  return CheckCalledFromGeneratedFile("containing_type");
+}
+
+static PyObject* GetExtensionScope(PyBaseDescriptor *self, void *closure) {
+  const Descriptor* extension_scope =
+      _GetDescriptor(self)->extension_scope();
+  if (extension_scope) {
+    return PyMessageDescriptor_FromDescriptor(extension_scope);
+  } else {
+    Py_RETURN_NONE;
+  }
+}
+
+static PyObject* GetContainingOneof(PyBaseDescriptor *self, void *closure) {
+  const OneofDescriptor* containing_oneof =
+      _GetDescriptor(self)->containing_oneof();
+  if (containing_oneof) {
+    return PyOneofDescriptor_FromDescriptor(containing_oneof);
+  } else {
+    Py_RETURN_NONE;
+  }
+}
+
+static int SetContainingOneof(PyBaseDescriptor *self, PyObject *value,
+                              void *closure) {
+  return CheckCalledFromGeneratedFile("containing_oneof");
+}
+
+static PyObject* GetHasOptions(PyBaseDescriptor *self, void *closure) {
+  const FieldOptions& options(_GetDescriptor(self)->options());
+  if (&options != &FieldOptions::default_instance()) {
+    Py_RETURN_TRUE;
+  } else {
+    Py_RETURN_FALSE;
+  }
+}
+static int SetHasOptions(PyBaseDescriptor *self, PyObject *value,
+                         void *closure) {
+  return CheckCalledFromGeneratedFile("has_options");
+}
+
+static PyObject* GetOptions(PyBaseDescriptor *self) {
+  return GetOrBuildOptions(_GetDescriptor(self));
+}
+
+static int SetOptions(PyBaseDescriptor *self, PyObject *value,
+                      void *closure) {
+  return CheckCalledFromGeneratedFile("_options");
+}
+
+
+static PyGetSetDef Getters[] = {
+  { "full_name", (getter)GetFullName, NULL, "Full name"},
+  { "name", (getter)GetName, NULL, "Unqualified name"},
+  { "camelcase_name", (getter)GetCamelcaseName, NULL, "Camelcase name"},
+  { "type", (getter)GetType, NULL, "C++ Type"},
+  { "cpp_type", (getter)GetCppType, NULL, "C++ Type"},
+  { "label", (getter)GetLabel, NULL, "Label"},
+  { "number", (getter)GetNumber, NULL, "Number"},
+  { "index", (getter)GetIndex, NULL, "Index"},
+  { "default_value", (getter)GetDefaultValue, NULL, "Default Value"},
+  { "has_default_value", (getter)HasDefaultValue},
+  { "is_extension", (getter)IsExtension, NULL, "ID"},
+  { "id", (getter)GetID, NULL, "ID"},
+  { "_cdescriptor", (getter)GetCDescriptor, NULL, "HAACK REMOVE ME"},
+
+  { "message_type", (getter)GetMessageType, (setter)SetMessageType,
+    "Message type"},
+  { "enum_type", (getter)GetEnumType, (setter)SetEnumType, "Enum type"},
+  { "containing_type", (getter)GetContainingType, (setter)SetContainingType,
+    "Containing type"},
+  { "extension_scope", (getter)GetExtensionScope, (setter)NULL,
+    "Extension scope"},
+  { "containing_oneof", (getter)GetContainingOneof, (setter)SetContainingOneof,
+    "Containing oneof"},
+  { "has_options", (getter)GetHasOptions, (setter)SetHasOptions, "Has Options"},
+  { "_options", (getter)NULL, (setter)SetOptions, "Options"},
+  {NULL}
+};
+
+static PyMethodDef Methods[] = {
+  { "GetOptions", (PyCFunction)GetOptions, METH_NOARGS, },
+  {NULL}
+};
+
+}  // namespace field_descriptor
+
+PyTypeObject PyFieldDescriptor_Type = {
+  PyVarObject_HEAD_INIT(&PyType_Type, 0)
+  FULL_MODULE_NAME ".FieldDescriptor",  // tp_name
+  sizeof(PyBaseDescriptor),             // tp_basicsize
+  0,                                    // tp_itemsize
+  0,                                    // tp_dealloc
+  0,                                    // tp_print
+  0,                                    // tp_getattr
+  0,                                    // tp_setattr
+  0,                                    // tp_compare
+  0,                                    // tp_repr
+  0,                                    // tp_as_number
+  0,                                    // tp_as_sequence
+  0,                                    // tp_as_mapping
+  0,                                    // tp_hash
+  0,                                    // tp_call
+  0,                                    // tp_str
+  0,                                    // tp_getattro
+  0,                                    // tp_setattro
+  0,                                    // tp_as_buffer
+  Py_TPFLAGS_DEFAULT,                   // tp_flags
+  "A Field Descriptor",                 // tp_doc
+  0,                                    // tp_traverse
+  0,                                    // tp_clear
+  0,                                    // tp_richcompare
+  0,                                    // tp_weaklistoffset
+  0,                                    // tp_iter
+  0,                                    // tp_iternext
+  field_descriptor::Methods,            // tp_methods
+  0,                                    // tp_members
+  field_descriptor::Getters,            // tp_getset
+  &descriptor::PyBaseDescriptor_Type,   // tp_base
+};
+
+PyObject* PyFieldDescriptor_FromDescriptor(
+    const FieldDescriptor* field_descriptor) {
+  return descriptor::NewInternedDescriptor(
+      &PyFieldDescriptor_Type, field_descriptor, NULL);
+}
+
+const FieldDescriptor* PyFieldDescriptor_AsDescriptor(PyObject* obj) {
+  if (!PyObject_TypeCheck(obj, &PyFieldDescriptor_Type)) {
+    PyErr_SetString(PyExc_TypeError, "Not a FieldDescriptor");
+    return NULL;
+  }
+  return reinterpret_cast<const FieldDescriptor*>(
+      reinterpret_cast<PyBaseDescriptor*>(obj)->descriptor);
+}
+
+namespace enum_descriptor {
+
+// Unchecked accessor to the C++ pointer.
+static const EnumDescriptor* _GetDescriptor(
+    PyBaseDescriptor *self) {
+  return reinterpret_cast<const EnumDescriptor*>(self->descriptor);
+}
+
+static PyObject* GetFullName(PyBaseDescriptor* self, void *closure) {
+  return PyString_FromCppString(_GetDescriptor(self)->full_name());
+}
+
+static PyObject* GetName(PyBaseDescriptor *self, void *closure) {
+  return PyString_FromCppString(_GetDescriptor(self)->name());
+}
+
+static PyObject* GetFile(PyBaseDescriptor *self, void *closure) {
+  return PyFileDescriptor_FromDescriptor(_GetDescriptor(self)->file());
+}
+
+static PyObject* GetEnumvaluesByName(PyBaseDescriptor* self, void *closure) {
+  return NewEnumValuesByName(_GetDescriptor(self));
+}
+
+static PyObject* GetEnumvaluesByNumber(PyBaseDescriptor* self, void *closure) {
+  return NewEnumValuesByNumber(_GetDescriptor(self));
+}
+
+static PyObject* GetEnumvaluesSeq(PyBaseDescriptor* self, void *closure) {
+  return NewEnumValuesSeq(_GetDescriptor(self));
+}
+
+static PyObject* GetContainingType(PyBaseDescriptor *self, void *closure) {
+  const Descriptor* containing_type =
+      _GetDescriptor(self)->containing_type();
+  if (containing_type) {
+    return PyMessageDescriptor_FromDescriptor(containing_type);
+  } else {
+    Py_RETURN_NONE;
+  }
+}
+
+static int SetContainingType(PyBaseDescriptor *self, PyObject *value,
+                             void *closure) {
+  return CheckCalledFromGeneratedFile("containing_type");
+}
+
+
+static PyObject* GetHasOptions(PyBaseDescriptor *self, void *closure) {
+  const EnumOptions& options(_GetDescriptor(self)->options());
+  if (&options != &EnumOptions::default_instance()) {
+    Py_RETURN_TRUE;
+  } else {
+    Py_RETURN_FALSE;
+  }
+}
+static int SetHasOptions(PyBaseDescriptor *self, PyObject *value,
+                         void *closure) {
+  return CheckCalledFromGeneratedFile("has_options");
+}
+
+static PyObject* GetOptions(PyBaseDescriptor *self) {
+  return GetOrBuildOptions(_GetDescriptor(self));
+}
+
+static int SetOptions(PyBaseDescriptor *self, PyObject *value,
+                      void *closure) {
+  return CheckCalledFromGeneratedFile("_options");
+}
+
+static PyObject* CopyToProto(PyBaseDescriptor *self, PyObject *target) {
+  return CopyToPythonProto<EnumDescriptorProto>(_GetDescriptor(self), target);
+}
+
+static PyMethodDef Methods[] = {
+  { "GetOptions", (PyCFunction)GetOptions, METH_NOARGS, },
+  { "CopyToProto", (PyCFunction)CopyToProto, METH_O, },
+  {NULL}
+};
+
+static PyGetSetDef Getters[] = {
+  { "full_name", (getter)GetFullName, NULL, "Full name"},
+  { "name", (getter)GetName, NULL, "last name"},
+  { "file", (getter)GetFile, NULL, "File descriptor"},
+  { "values", (getter)GetEnumvaluesSeq, NULL, "values"},
+  { "values_by_name", (getter)GetEnumvaluesByName, NULL,
+    "Enum values by name"},
+  { "values_by_number", (getter)GetEnumvaluesByNumber, NULL,
+    "Enum values by number"},
+
+  { "containing_type", (getter)GetContainingType, (setter)SetContainingType,
+    "Containing type"},
+  { "has_options", (getter)GetHasOptions, (setter)SetHasOptions, "Has Options"},
+  { "_options", (getter)NULL, (setter)SetOptions, "Options"},
+  {NULL}
+};
+
+}  // namespace enum_descriptor
+
+PyTypeObject PyEnumDescriptor_Type = {
+  PyVarObject_HEAD_INIT(&PyType_Type, 0)
+  FULL_MODULE_NAME ".EnumDescriptor",   // tp_name
+  sizeof(PyBaseDescriptor),             // tp_basicsize
+  0,                                    // tp_itemsize
+  0,                                    // tp_dealloc
+  0,                                    // tp_print
+  0,                                    // tp_getattr
+  0,                                    // tp_setattr
+  0,                                    // tp_compare
+  0,                                    // tp_repr
+  0,                                    // tp_as_number
+  0,                                    // tp_as_sequence
+  0,                                    // tp_as_mapping
+  0,                                    // tp_hash
+  0,                                    // tp_call
+  0,                                    // tp_str
+  0,                                    // tp_getattro
+  0,                                    // tp_setattro
+  0,                                    // tp_as_buffer
+  Py_TPFLAGS_DEFAULT,                   // tp_flags
+  "A Enum Descriptor",                  // tp_doc
+  0,                                    // tp_traverse
+  0,                                    // tp_clear
+  0,                                    // tp_richcompare
+  0,                                    // tp_weaklistoffset
+  0,                                    // tp_iter
+  0,                                    // tp_iternext
+  enum_descriptor::Methods,             // tp_getset
+  0,                                    // tp_members
+  enum_descriptor::Getters,             // tp_getset
+  &descriptor::PyBaseDescriptor_Type,   // tp_base
+};
+
+PyObject* PyEnumDescriptor_FromDescriptor(
+    const EnumDescriptor* enum_descriptor) {
+  return descriptor::NewInternedDescriptor(
+      &PyEnumDescriptor_Type, enum_descriptor, NULL);
+}
+
+const EnumDescriptor* PyEnumDescriptor_AsDescriptor(PyObject* obj) {
+  if (!PyObject_TypeCheck(obj, &PyEnumDescriptor_Type)) {
+    PyErr_SetString(PyExc_TypeError, "Not an EnumDescriptor");
+    return NULL;
+  }
+  return reinterpret_cast<const EnumDescriptor*>(
+      reinterpret_cast<PyBaseDescriptor*>(obj)->descriptor);
+}
+
+namespace enumvalue_descriptor {
+
+// Unchecked accessor to the C++ pointer.
+static const EnumValueDescriptor* _GetDescriptor(
+    PyBaseDescriptor *self) {
+  return reinterpret_cast<const EnumValueDescriptor*>(self->descriptor);
+}
+
+static PyObject* GetName(PyBaseDescriptor *self, void *closure) {
+  return PyString_FromCppString(_GetDescriptor(self)->name());
+}
+
+static PyObject* GetNumber(PyBaseDescriptor *self, void *closure) {
+  return PyInt_FromLong(_GetDescriptor(self)->number());
+}
+
+static PyObject* GetIndex(PyBaseDescriptor *self, void *closure) {
+  return PyInt_FromLong(_GetDescriptor(self)->index());
+}
+
+static PyObject* GetType(PyBaseDescriptor *self, void *closure) {
+  return PyEnumDescriptor_FromDescriptor(_GetDescriptor(self)->type());
+}
+
+static PyObject* GetHasOptions(PyBaseDescriptor *self, void *closure) {
+  const EnumValueOptions& options(_GetDescriptor(self)->options());
+  if (&options != &EnumValueOptions::default_instance()) {
+    Py_RETURN_TRUE;
+  } else {
+    Py_RETURN_FALSE;
+  }
+}
+static int SetHasOptions(PyBaseDescriptor *self, PyObject *value,
+                         void *closure) {
+  return CheckCalledFromGeneratedFile("has_options");
+}
+
+static PyObject* GetOptions(PyBaseDescriptor *self) {
+  return GetOrBuildOptions(_GetDescriptor(self));
+}
+
+static int SetOptions(PyBaseDescriptor *self, PyObject *value,
+                      void *closure) {
+  return CheckCalledFromGeneratedFile("_options");
+}
+
+
+static PyGetSetDef Getters[] = {
+  { "name", (getter)GetName, NULL, "name"},
+  { "number", (getter)GetNumber, NULL, "number"},
+  { "index", (getter)GetIndex, NULL, "index"},
+  { "type", (getter)GetType, NULL, "index"},
+
+  { "has_options", (getter)GetHasOptions, (setter)SetHasOptions, "Has Options"},
+  { "_options", (getter)NULL, (setter)SetOptions, "Options"},
+  {NULL}
+};
+
+static PyMethodDef Methods[] = {
+  { "GetOptions", (PyCFunction)GetOptions, METH_NOARGS, },
+  {NULL}
+};
+
+}  // namespace enumvalue_descriptor
+
+PyTypeObject PyEnumValueDescriptor_Type = {
+  PyVarObject_HEAD_INIT(&PyType_Type, 0)
+  FULL_MODULE_NAME ".EnumValueDescriptor",  // tp_name
+  sizeof(PyBaseDescriptor),             // tp_basicsize
+  0,                                    // tp_itemsize
+  0,                                    // tp_dealloc
+  0,                                    // tp_print
+  0,                                    // tp_getattr
+  0,                                    // tp_setattr
+  0,                                    // tp_compare
+  0,                                    // tp_repr
+  0,                                    // tp_as_number
+  0,                                    // tp_as_sequence
+  0,                                    // tp_as_mapping
+  0,                                    // tp_hash
+  0,                                    // tp_call
+  0,                                    // tp_str
+  0,                                    // tp_getattro
+  0,                                    // tp_setattro
+  0,                                    // tp_as_buffer
+  Py_TPFLAGS_DEFAULT,                   // tp_flags
+  "A EnumValue Descriptor",             // tp_doc
+  0,                                    // tp_traverse
+  0,                                    // tp_clear
+  0,                                    // tp_richcompare
+  0,                                    // tp_weaklistoffset
+  0,                                    // tp_iter
+  0,                                    // tp_iternext
+  enumvalue_descriptor::Methods,        // tp_methods
+  0,                                    // tp_members
+  enumvalue_descriptor::Getters,        // tp_getset
+  &descriptor::PyBaseDescriptor_Type,   // tp_base
+};
+
+PyObject* PyEnumValueDescriptor_FromDescriptor(
+    const EnumValueDescriptor* enumvalue_descriptor) {
+  return descriptor::NewInternedDescriptor(
+      &PyEnumValueDescriptor_Type, enumvalue_descriptor, NULL);
+}
+
+namespace file_descriptor {
+
+// Unchecked accessor to the C++ pointer.
+static const FileDescriptor* _GetDescriptor(PyFileDescriptor *self) {
+  return reinterpret_cast<const FileDescriptor*>(self->base.descriptor);
+}
+
+static void Dealloc(PyFileDescriptor* self) {
+  Py_XDECREF(self->serialized_pb);
+  descriptor::Dealloc(&self->base);
+}
+
+static PyObject* GetPool(PyFileDescriptor *self, void *closure) {
+  PyObject* pool = reinterpret_cast<PyObject*>(
+      GetDescriptorPool_FromPool(_GetDescriptor(self)->pool()));
+  Py_XINCREF(pool);
+  return pool;
+}
+
+static PyObject* GetName(PyFileDescriptor *self, void *closure) {
+  return PyString_FromCppString(_GetDescriptor(self)->name());
+}
+
+static PyObject* GetPackage(PyFileDescriptor *self, void *closure) {
+  return PyString_FromCppString(_GetDescriptor(self)->package());
+}
+
+static PyObject* GetSerializedPb(PyFileDescriptor *self, void *closure) {
+  PyObject *serialized_pb = self->serialized_pb;
+  if (serialized_pb != NULL) {
+    Py_INCREF(serialized_pb);
+    return serialized_pb;
+  }
+  FileDescriptorProto file_proto;
+  _GetDescriptor(self)->CopyTo(&file_proto);
+  string contents;
+  file_proto.SerializePartialToString(&contents);
+  self->serialized_pb = PyBytes_FromStringAndSize(
+      contents.c_str(), contents.size());
+  if (self->serialized_pb == NULL) {
+    return NULL;
+  }
+  Py_INCREF(self->serialized_pb);
+  return self->serialized_pb;
+}
+
+static PyObject* GetMessageTypesByName(PyFileDescriptor* self, void *closure) {
+  return NewFileMessageTypesByName(_GetDescriptor(self));
+}
+
+static PyObject* GetEnumTypesByName(PyFileDescriptor* self, void *closure) {
+  return NewFileEnumTypesByName(_GetDescriptor(self));
+}
+
+static PyObject* GetExtensionsByName(PyFileDescriptor* self, void *closure) {
+  return NewFileExtensionsByName(_GetDescriptor(self));
+}
+
+static PyObject* GetDependencies(PyFileDescriptor* self, void *closure) {
+  return NewFileDependencies(_GetDescriptor(self));
+}
+
+static PyObject* GetPublicDependencies(PyFileDescriptor* self, void *closure) {
+  return NewFilePublicDependencies(_GetDescriptor(self));
+}
+
+static PyObject* GetHasOptions(PyFileDescriptor *self, void *closure) {
+  const FileOptions& options(_GetDescriptor(self)->options());
+  if (&options != &FileOptions::default_instance()) {
+    Py_RETURN_TRUE;
+  } else {
+    Py_RETURN_FALSE;
+  }
+}
+static int SetHasOptions(PyFileDescriptor *self, PyObject *value,
+                         void *closure) {
+  return CheckCalledFromGeneratedFile("has_options");
+}
+
+static PyObject* GetOptions(PyFileDescriptor *self) {
+  return GetOrBuildOptions(_GetDescriptor(self));
+}
+
+static int SetOptions(PyFileDescriptor *self, PyObject *value,
+                      void *closure) {
+  return CheckCalledFromGeneratedFile("_options");
+}
+
+static PyObject* GetSyntax(PyFileDescriptor *self, void *closure) {
+  return PyString_InternFromString(
+      FileDescriptor::SyntaxName(_GetDescriptor(self)->syntax()));
+}
+
+static PyObject* CopyToProto(PyFileDescriptor *self, PyObject *target) {
+  return CopyToPythonProto<FileDescriptorProto>(_GetDescriptor(self), target);
+}
+
+static PyGetSetDef Getters[] = {
+  { "pool", (getter)GetPool, NULL, "pool"},
+  { "name", (getter)GetName, NULL, "name"},
+  { "package", (getter)GetPackage, NULL, "package"},
+  { "serialized_pb", (getter)GetSerializedPb},
+  { "message_types_by_name", (getter)GetMessageTypesByName, NULL,
+    "Messages by name"},
+  { "enum_types_by_name", (getter)GetEnumTypesByName, NULL, "Enums by name"},
+  { "extensions_by_name", (getter)GetExtensionsByName, NULL,
+    "Extensions by name"},
+  { "dependencies", (getter)GetDependencies, NULL, "Dependencies"},
+  { "public_dependencies", (getter)GetPublicDependencies, NULL, "Dependencies"},
+
+  { "has_options", (getter)GetHasOptions, (setter)SetHasOptions, "Has Options"},
+  { "_options", (getter)NULL, (setter)SetOptions, "Options"},
+  { "syntax", (getter)GetSyntax, (setter)NULL, "Syntax"},
+  {NULL}
+};
+
+static PyMethodDef Methods[] = {
+  { "GetOptions", (PyCFunction)GetOptions, METH_NOARGS, },
+  { "CopyToProto", (PyCFunction)CopyToProto, METH_O, },
+  {NULL}
+};
+
+}  // namespace file_descriptor
+
+PyTypeObject PyFileDescriptor_Type = {
+  PyVarObject_HEAD_INIT(&PyType_Type, 0)
+  FULL_MODULE_NAME ".FileDescriptor",   // tp_name
+  sizeof(PyFileDescriptor),             // tp_basicsize
+  0,                                    // tp_itemsize
+  (destructor)file_descriptor::Dealloc,  // tp_dealloc
+  0,                                    // tp_print
+  0,                                    // tp_getattr
+  0,                                    // tp_setattr
+  0,                                    // tp_compare
+  0,                                    // tp_repr
+  0,                                    // tp_as_number
+  0,                                    // tp_as_sequence
+  0,                                    // tp_as_mapping
+  0,                                    // tp_hash
+  0,                                    // tp_call
+  0,                                    // tp_str
+  0,                                    // tp_getattro
+  0,                                    // tp_setattro
+  0,                                    // tp_as_buffer
+  Py_TPFLAGS_DEFAULT,                   // tp_flags
+  "A File Descriptor",                  // tp_doc
+  0,                                    // tp_traverse
+  0,                                    // tp_clear
+  0,                                    // tp_richcompare
+  0,                                    // tp_weaklistoffset
+  0,                                    // tp_iter
+  0,                                    // tp_iternext
+  file_descriptor::Methods,             // tp_methods
+  0,                                    // tp_members
+  file_descriptor::Getters,             // tp_getset
+  &descriptor::PyBaseDescriptor_Type,   // tp_base
+  0,                                    // tp_dict
+  0,                                    // tp_descr_get
+  0,                                    // tp_descr_set
+  0,                                    // tp_dictoffset
+  0,                                    // tp_init
+  0,                                    // tp_alloc
+  0,                                    // tp_new
+  PyObject_Del,                         // tp_free
+};
+
+PyObject* PyFileDescriptor_FromDescriptor(
+    const FileDescriptor* file_descriptor) {
+  return PyFileDescriptor_FromDescriptorWithSerializedPb(file_descriptor,
+                                                         NULL);
+}
+
+PyObject* PyFileDescriptor_FromDescriptorWithSerializedPb(
+    const FileDescriptor* file_descriptor, PyObject *serialized_pb) {
+  bool was_created;
+  PyObject* py_descriptor = descriptor::NewInternedDescriptor(
+      &PyFileDescriptor_Type, file_descriptor, &was_created);
+  if (py_descriptor == NULL) {
+    return NULL;
+  }
+  if (was_created) {
+    PyFileDescriptor* cfile_descriptor =
+        reinterpret_cast<PyFileDescriptor*>(py_descriptor);
+    Py_XINCREF(serialized_pb);
+    cfile_descriptor->serialized_pb = serialized_pb;
+  }
+  // TODO(amauryfa): In the case of a cached object, check that serialized_pb
+  // is the same as before.
+
+  return py_descriptor;
+}
+
+const FileDescriptor* PyFileDescriptor_AsDescriptor(PyObject* obj) {
+  if (!PyObject_TypeCheck(obj, &PyFileDescriptor_Type)) {
+    PyErr_SetString(PyExc_TypeError, "Not a FileDescriptor");
+    return NULL;
+  }
+  return reinterpret_cast<const FileDescriptor*>(
+      reinterpret_cast<PyBaseDescriptor*>(obj)->descriptor);
+}
+
+namespace oneof_descriptor {
+
+// Unchecked accessor to the C++ pointer.
+static const OneofDescriptor* _GetDescriptor(
+    PyBaseDescriptor *self) {
+  return reinterpret_cast<const OneofDescriptor*>(self->descriptor);
+}
+
+static PyObject* GetName(PyBaseDescriptor* self, void *closure) {
+  return PyString_FromCppString(_GetDescriptor(self)->name());
+}
+
+static PyObject* GetFullName(PyBaseDescriptor* self, void *closure) {
+  return PyString_FromCppString(_GetDescriptor(self)->full_name());
+}
+
+static PyObject* GetIndex(PyBaseDescriptor *self, void *closure) {
+  return PyInt_FromLong(_GetDescriptor(self)->index());
+}
+
+static PyObject* GetFields(PyBaseDescriptor* self, void *closure) {
+  return NewOneofFieldsSeq(_GetDescriptor(self));
+}
+
+static PyObject* GetContainingType(PyBaseDescriptor *self, void *closure) {
+  const Descriptor* containing_type =
+      _GetDescriptor(self)->containing_type();
+  if (containing_type) {
+    return PyMessageDescriptor_FromDescriptor(containing_type);
+  } else {
+    Py_RETURN_NONE;
+  }
+}
+
+static PyGetSetDef Getters[] = {
+  { "name", (getter)GetName, NULL, "Name"},
+  { "full_name", (getter)GetFullName, NULL, "Full name"},
+  { "index", (getter)GetIndex, NULL, "Index"},
+
+  { "containing_type", (getter)GetContainingType, NULL, "Containing type"},
+  { "fields", (getter)GetFields, NULL, "Fields"},
+  {NULL}
+};
+
+}  // namespace oneof_descriptor
+
+PyTypeObject PyOneofDescriptor_Type = {
+  PyVarObject_HEAD_INIT(&PyType_Type, 0)
+  FULL_MODULE_NAME ".OneofDescriptor",  // tp_name
+  sizeof(PyBaseDescriptor),             // tp_basicsize
+  0,                                    // tp_itemsize
+  0,                                    // tp_dealloc
+  0,                                    // tp_print
+  0,                                    // tp_getattr
+  0,                                    // tp_setattr
+  0,                                    // tp_compare
+  0,                                    // tp_repr
+  0,                                    // tp_as_number
+  0,                                    // tp_as_sequence
+  0,                                    // tp_as_mapping
+  0,                                    // tp_hash
+  0,                                    // tp_call
+  0,                                    // tp_str
+  0,                                    // tp_getattro
+  0,                                    // tp_setattro
+  0,                                    // tp_as_buffer
+  Py_TPFLAGS_DEFAULT,                   // tp_flags
+  "A Oneof Descriptor",                 // tp_doc
+  0,                                    // tp_traverse
+  0,                                    // tp_clear
+  0,                                    // tp_richcompare
+  0,                                    // tp_weaklistoffset
+  0,                                    // tp_iter
+  0,                                    // tp_iternext
+  0,                                    // tp_methods
+  0,                                    // tp_members
+  oneof_descriptor::Getters,            // tp_getset
+  &descriptor::PyBaseDescriptor_Type,   // tp_base
+};
+
+PyObject* PyOneofDescriptor_FromDescriptor(
+    const OneofDescriptor* oneof_descriptor) {
+  return descriptor::NewInternedDescriptor(
+      &PyOneofDescriptor_Type, oneof_descriptor, NULL);
+}
+
+// Add a enum values to a type dictionary.
+static bool AddEnumValues(PyTypeObject *type,
+                          const EnumDescriptor* enum_descriptor) {
+  for (int i = 0; i < enum_descriptor->value_count(); ++i) {
+    const EnumValueDescriptor* value = enum_descriptor->value(i);
+    ScopedPyObjectPtr obj(PyInt_FromLong(value->number()));
+    if (obj == NULL) {
+      return false;
+    }
+    if (PyDict_SetItemString(type->tp_dict, value->name().c_str(), obj.get()) <
+        0) {
+      return false;
+    }
+  }
+  return true;
+}
+
+static bool AddIntConstant(PyTypeObject *type, const char* name, int value) {
+  ScopedPyObjectPtr obj(PyInt_FromLong(value));
+  if (PyDict_SetItemString(type->tp_dict, name, obj.get()) < 0) {
+    return false;
+  }
+  return true;
+}
+
+
+bool InitDescriptor() {
+  if (PyType_Ready(&PyMessageDescriptor_Type) < 0)
+    return false;
+
+  if (PyType_Ready(&PyFieldDescriptor_Type) < 0)
+    return false;
+
+  if (!AddEnumValues(&PyFieldDescriptor_Type,
+                     FieldDescriptorProto::Label_descriptor())) {
+    return false;
+  }
+  if (!AddEnumValues(&PyFieldDescriptor_Type,
+                     FieldDescriptorProto::Type_descriptor())) {
+    return false;
+  }
+#define ADD_FIELDDESC_CONSTANT(NAME) AddIntConstant( \
+    &PyFieldDescriptor_Type, #NAME, FieldDescriptor::NAME)
+  if (!ADD_FIELDDESC_CONSTANT(CPPTYPE_INT32) ||
+      !ADD_FIELDDESC_CONSTANT(CPPTYPE_INT64) ||
+      !ADD_FIELDDESC_CONSTANT(CPPTYPE_UINT32) ||
+      !ADD_FIELDDESC_CONSTANT(CPPTYPE_UINT64) ||
+      !ADD_FIELDDESC_CONSTANT(CPPTYPE_DOUBLE) ||
+      !ADD_FIELDDESC_CONSTANT(CPPTYPE_FLOAT) ||
+      !ADD_FIELDDESC_CONSTANT(CPPTYPE_BOOL) ||
+      !ADD_FIELDDESC_CONSTANT(CPPTYPE_ENUM) ||
+      !ADD_FIELDDESC_CONSTANT(CPPTYPE_STRING) ||
+      !ADD_FIELDDESC_CONSTANT(CPPTYPE_MESSAGE)) {
+    return false;
+  }
+#undef ADD_FIELDDESC_CONSTANT
+
+  if (PyType_Ready(&PyEnumDescriptor_Type) < 0)
+    return false;
+
+  if (PyType_Ready(&PyEnumValueDescriptor_Type) < 0)
+    return false;
+
+  if (PyType_Ready(&PyFileDescriptor_Type) < 0)
+    return false;
+
+  if (PyType_Ready(&PyOneofDescriptor_Type) < 0)
+    return false;
+
+  if (!InitDescriptorMappingTypes())
+    return false;
+
+  return true;
+}
+
+}  // namespace python
+}  // namespace protobuf
+}  // namespace google
diff --git a/third_party/protobuf/python/google/protobuf/pyext/descriptor.h b/third_party/protobuf/python/google/protobuf/pyext/descriptor.h
new file mode 100644
index 0000000000..eb99df182e
--- /dev/null
+++ b/third_party/protobuf/python/google/protobuf/pyext/descriptor.h
@@ -0,0 +1,97 @@
+// 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: petar@google.com (Petar Petrov)
+
+#ifndef GOOGLE_PROTOBUF_PYTHON_CPP_DESCRIPTOR_H__
+#define GOOGLE_PROTOBUF_PYTHON_CPP_DESCRIPTOR_H__
+
+#include <Python.h>
+
+#include <google/protobuf/descriptor.h>
+
+namespace google {
+namespace protobuf {
+namespace python {
+
+extern PyTypeObject PyMessageDescriptor_Type;
+extern PyTypeObject PyFieldDescriptor_Type;
+extern PyTypeObject PyEnumDescriptor_Type;
+extern PyTypeObject PyEnumValueDescriptor_Type;
+extern PyTypeObject PyFileDescriptor_Type;
+extern PyTypeObject PyOneofDescriptor_Type;
+
+// Wraps a Descriptor in a Python object.
+// The C++ pointer is usually borrowed from the global DescriptorPool.
+// In any case, it must stay alive as long as the Python object.
+// Returns a new reference.
+PyObject* PyMessageDescriptor_FromDescriptor(const Descriptor* descriptor);
+PyObject* PyFieldDescriptor_FromDescriptor(const FieldDescriptor* descriptor);
+PyObject* PyEnumDescriptor_FromDescriptor(const EnumDescriptor* descriptor);
+PyObject* PyEnumValueDescriptor_FromDescriptor(
+    const EnumValueDescriptor* descriptor);
+PyObject* PyOneofDescriptor_FromDescriptor(const OneofDescriptor* descriptor);
+PyObject* PyFileDescriptor_FromDescriptor(
+    const FileDescriptor* file_descriptor);
+
+// Alternate constructor of PyFileDescriptor, used when we already have a
+// serialized FileDescriptorProto that can be cached.
+// Returns a new reference.
+PyObject* PyFileDescriptor_FromDescriptorWithSerializedPb(
+    const FileDescriptor* file_descriptor, PyObject* serialized_pb);
+
+// Return the C++ descriptor pointer.
+// This function checks the parameter type; on error, return NULL with a Python
+// exception set.
+const Descriptor* PyMessageDescriptor_AsDescriptor(PyObject* obj);
+const FieldDescriptor* PyFieldDescriptor_AsDescriptor(PyObject* obj);
+const EnumDescriptor* PyEnumDescriptor_AsDescriptor(PyObject* obj);
+const FileDescriptor* PyFileDescriptor_AsDescriptor(PyObject* obj);
+
+// Returns the raw C++ pointer.
+const void* PyDescriptor_AsVoidPtr(PyObject* obj);
+
+// Check that the calling Python code is the global scope of a _pb2.py module.
+// This function is used to support the current code generated by the proto
+// compiler, which insists on modifying descriptors after they have been
+// created.
+//
+// stacklevel indicates which Python frame should be the _pb2.py module.
+//
+// Don't use this function outside descriptor classes.
+bool _CalledFromGeneratedFile(int stacklevel);
+
+bool InitDescriptor();
+
+}  // namespace python
+}  // namespace protobuf
+
+}  // namespace google
+#endif  // GOOGLE_PROTOBUF_PYTHON_CPP_DESCRIPTOR_H__
diff --git a/third_party/protobuf/python/google/protobuf/pyext/descriptor_containers.cc b/third_party/protobuf/python/google/protobuf/pyext/descriptor_containers.cc
new file mode 100644
index 0000000000..e505d8122c
--- /dev/null
+++ b/third_party/protobuf/python/google/protobuf/pyext/descriptor_containers.cc
@@ -0,0 +1,1652 @@
+// 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.
+
+// Mappings and Sequences of descriptors.
+// Used by Descriptor.fields_by_name, EnumDescriptor.values...
+//
+// They avoid the allocation of a full dictionary or a full list: they simply
+// store a pointer to the parent descriptor, use the C++ Descriptor methods (see
+// google/protobuf/descriptor.h) to retrieve other descriptors, and create
+// Python objects on the fly.
+//
+// The containers fully conform to abc.Mapping and abc.Sequence, and behave just
+// like read-only dictionaries and lists.
+//
+// Because the interface of C++ Descriptors is quite regular, this file actually
+// defines only three types, the exact behavior of a container is controlled by
+// a DescriptorContainerDef structure, which contains functions that uses the
+// public Descriptor API.
+//
+// Note: This DescriptorContainerDef is similar to the "virtual methods table"
+// that a C++ compiler generates for a class. We have to make it explicit
+// because the Python API is based on C, and does not play well with C++
+// inheritance.
+
+#include <Python.h>
+
+#include <google/protobuf/descriptor.h>
+#include <google/protobuf/pyext/descriptor_containers.h>
+#include <google/protobuf/pyext/descriptor_pool.h>
+#include <google/protobuf/pyext/descriptor.h>
+#include <google/protobuf/pyext/scoped_pyobject_ptr.h>
+
+#if PY_MAJOR_VERSION >= 3
+  #define PyString_FromStringAndSize PyUnicode_FromStringAndSize
+  #define PyString_FromFormat PyUnicode_FromFormat
+  #define PyInt_FromLong PyLong_FromLong
+  #if PY_VERSION_HEX < 0x03030000
+    #error "Python 3.0 - 3.2 are not supported."
+  #endif
+  #define PyString_AsStringAndSize(ob, charpp, sizep) \
+    (PyUnicode_Check(ob)? \
+       ((*(charpp) = PyUnicode_AsUTF8AndSize(ob, (sizep))) == NULL? -1: 0): \
+       PyBytes_AsStringAndSize(ob, (charpp), (sizep)))
+#endif
+
+namespace google {
+namespace protobuf {
+namespace python {
+
+struct PyContainer;
+
+typedef int (*CountMethod)(PyContainer* self);
+typedef const void* (*GetByIndexMethod)(PyContainer* self, int index);
+typedef const void* (*GetByNameMethod)(PyContainer* self, const string& name);
+typedef const void* (*GetByCamelcaseNameMethod)(PyContainer* self,
+                                                const string& name);
+typedef const void* (*GetByNumberMethod)(PyContainer* self, int index);
+typedef PyObject* (*NewObjectFromItemMethod)(const void* descriptor);
+typedef const string& (*GetItemNameMethod)(const void* descriptor);
+typedef const string& (*GetItemCamelcaseNameMethod)(const void* descriptor);
+typedef int (*GetItemNumberMethod)(const void* descriptor);
+typedef int (*GetItemIndexMethod)(const void* descriptor);
+
+struct DescriptorContainerDef {
+  const char* mapping_name;
+  // Returns the number of items in the container.
+  CountMethod count_fn;
+  // Retrieve item by index (usually the order of declaration in the proto file)
+  // Used by sequences, but also iterators. 0 <= index < Count().
+  GetByIndexMethod get_by_index_fn;
+  // Retrieve item by name (usually a call to some 'FindByName' method).
+  // Used by "by_name" mappings.
+  GetByNameMethod get_by_name_fn;
+  // Retrieve item by camelcase name (usually a call to some
+  // 'FindByCamelcaseName' method). Used by "by_camelcase_name" mappings.
+  GetByCamelcaseNameMethod get_by_camelcase_name_fn;
+  // Retrieve item by declared number (field tag, or enum value).
+  // Used by "by_number" mappings.
+  GetByNumberMethod get_by_number_fn;
+  // Converts a item C++ descriptor to a Python object. Returns a new reference.
+  NewObjectFromItemMethod new_object_from_item_fn;
+  // Retrieve the name of an item. Used by iterators on "by_name" mappings.
+  GetItemNameMethod get_item_name_fn;
+  // Retrieve the camelcase name of an item. Used by iterators on
+  // "by_camelcase_name" mappings.
+  GetItemCamelcaseNameMethod get_item_camelcase_name_fn;
+  // Retrieve the number of an item. Used by iterators on "by_number" mappings.
+  GetItemNumberMethod get_item_number_fn;
+  // Retrieve the index of an item for the container type.
+  // Used by "__contains__".
+  // If not set, "x in sequence" will do a linear search.
+  GetItemIndexMethod get_item_index_fn;
+};
+
+struct PyContainer {
+  PyObject_HEAD
+
+  // The proto2 descriptor this container belongs to the global DescriptorPool.
+  const void* descriptor;
+
+  // A pointer to a static structure with function pointers that control the
+  // behavior of the container. Very similar to the table of virtual functions
+  // of a C++ class.
+  const DescriptorContainerDef* container_def;
+
+  // The kind of container: list, or dict by name or value.
+  enum ContainerKind {
+    KIND_SEQUENCE,
+    KIND_BYNAME,
+    KIND_BYCAMELCASENAME,
+    KIND_BYNUMBER,
+  } kind;
+};
+
+struct PyContainerIterator {
+  PyObject_HEAD
+
+  // The container we are iterating over. Own a reference.
+  PyContainer* container;
+
+  // The current index in the iterator.
+  int index;
+
+  // The kind of container: list, or dict by name or value.
+  enum IterKind {
+    KIND_ITERKEY,
+    KIND_ITERVALUE,
+    KIND_ITERITEM,
+    KIND_ITERVALUE_REVERSED,  // For sequences
+  } kind;
+};
+
+namespace descriptor {
+
+// Returns the C++ item descriptor for a given Python key.
+// When the descriptor is found, return true and set *item.
+// When the descriptor is not found, return true, but set *item to NULL.
+// On error, returns false with an exception set.
+static bool _GetItemByKey(PyContainer* self, PyObject* key, const void** item) {
+  switch (self->kind) {
+    case PyContainer::KIND_BYNAME:
+      {
+        char* name;
+        Py_ssize_t name_size;
+        if (PyString_AsStringAndSize(key, &name, &name_size) < 0) {
+          if (PyErr_ExceptionMatches(PyExc_TypeError)) {
+            // Not a string, cannot be in the container.
+            PyErr_Clear();
+            *item = NULL;
+            return true;
+          }
+          return false;
+        }
+        *item = self->container_def->get_by_name_fn(
+            self, string(name, name_size));
+        return true;
+      }
+    case PyContainer::KIND_BYCAMELCASENAME:
+      {
+        char* camelcase_name;
+        Py_ssize_t name_size;
+        if (PyString_AsStringAndSize(key, &camelcase_name, &name_size) < 0) {
+          if (PyErr_ExceptionMatches(PyExc_TypeError)) {
+            // Not a string, cannot be in the container.
+            PyErr_Clear();
+            *item = NULL;
+            return true;
+          }
+          return false;
+        }
+        *item = self->container_def->get_by_camelcase_name_fn(
+            self, string(camelcase_name, name_size));
+        return true;
+      }
+    case PyContainer::KIND_BYNUMBER:
+      {
+        Py_ssize_t number = PyNumber_AsSsize_t(key, NULL);
+        if (number == -1 && PyErr_Occurred()) {
+          if (PyErr_ExceptionMatches(PyExc_TypeError)) {
+            // Not a number, cannot be in the container.
+            PyErr_Clear();
+            *item = NULL;
+            return true;
+          }
+          return false;
+        }
+        *item = self->container_def->get_by_number_fn(self, number);
+        return true;
+      }
+    default:
+      PyErr_SetNone(PyExc_NotImplementedError);
+      return false;
+  }
+}
+
+// Returns the key of the object at the given index.
+// Used when iterating over mappings.
+static PyObject* _NewKey_ByIndex(PyContainer* self, Py_ssize_t index) {
+  const void* item = self->container_def->get_by_index_fn(self, index);
+  switch (self->kind) {
+    case PyContainer::KIND_BYNAME:
+      {
+        const string& name(self->container_def->get_item_name_fn(item));
+        return PyString_FromStringAndSize(name.c_str(), name.size());
+      }
+    case PyContainer::KIND_BYCAMELCASENAME:
+      {
+        const string& name(
+            self->container_def->get_item_camelcase_name_fn(item));
+        return PyString_FromStringAndSize(name.c_str(), name.size());
+      }
+    case PyContainer::KIND_BYNUMBER:
+      {
+        int value = self->container_def->get_item_number_fn(item);
+        return PyInt_FromLong(value);
+      }
+    default:
+      PyErr_SetNone(PyExc_NotImplementedError);
+      return NULL;
+  }
+}
+
+// Returns the object at the given index.
+// Also used when iterating over mappings.
+static PyObject* _NewObj_ByIndex(PyContainer* self, Py_ssize_t index) {
+  return self->container_def->new_object_from_item_fn(
+      self->container_def->get_by_index_fn(self, index));
+}
+
+static Py_ssize_t Length(PyContainer* self) {
+  return self->container_def->count_fn(self);
+}
+
+// The DescriptorMapping type.
+
+static PyObject* Subscript(PyContainer* self, PyObject* key) {
+  const void* item = NULL;
+  if (!_GetItemByKey(self, key, &item)) {
+    return NULL;
+  }
+  if (!item) {
+    PyErr_SetObject(PyExc_KeyError, key);
+    return NULL;
+  }
+  return self->container_def->new_object_from_item_fn(item);
+}
+
+static int AssSubscript(PyContainer* self, PyObject* key, PyObject* value) {
+  if (_CalledFromGeneratedFile(0)) {
+    return 0;
+  }
+  PyErr_Format(PyExc_TypeError,
+               "'%.200s' object does not support item assignment",
+               Py_TYPE(self)->tp_name);
+  return -1;
+}
+
+static PyMappingMethods MappingMappingMethods = {
+  (lenfunc)Length,               // mp_length
+  (binaryfunc)Subscript,         // mp_subscript
+  (objobjargproc)AssSubscript,   // mp_ass_subscript
+};
+
+static int Contains(PyContainer* self, PyObject* key) {
+  const void* item = NULL;
+  if (!_GetItemByKey(self, key, &item)) {
+    return -1;
+  }
+  if (item) {
+    return 1;
+  } else {
+    return 0;
+  }
+}
+
+static PyObject* ContainerRepr(PyContainer* self) {
+  const char* kind = "";
+  switch (self->kind) {
+    case PyContainer::KIND_SEQUENCE:
+      kind = "sequence";
+      break;
+    case PyContainer::KIND_BYNAME:
+      kind = "mapping by name";
+      break;
+    case PyContainer::KIND_BYCAMELCASENAME:
+      kind = "mapping by camelCase name";
+      break;
+    case PyContainer::KIND_BYNUMBER:
+      kind = "mapping by number";
+      break;
+  }
+  return PyString_FromFormat(
+      "<%s %s>", self->container_def->mapping_name, kind);
+}
+
+extern PyTypeObject DescriptorMapping_Type;
+extern PyTypeObject DescriptorSequence_Type;
+
+// A sequence container can only be equal to another sequence container, or (for
+// backward compatibility) to a list containing the same items.
+// Returns 1 if equal, 0 if unequal, -1 on error.
+static int DescriptorSequence_Equal(PyContainer* self, PyObject* other) {
+  // Check the identity of C++ pointers.
+  if (PyObject_TypeCheck(other, &DescriptorSequence_Type)) {
+    PyContainer* other_container = reinterpret_cast<PyContainer*>(other);
+    if (self->descriptor == other_container->descriptor &&
+        self->container_def == other_container->container_def &&
+        self->kind == other_container->kind) {
+      return 1;
+    } else {
+      return 0;
+    }
+  }
+
+  // If other is a list
+  if (PyList_Check(other)) {
+    // return list(self) == other
+    int size = Length(self);
+    if (size != PyList_Size(other)) {
+      return false;
+    }
+    for (int index = 0; index < size; index++) {
+      ScopedPyObjectPtr value1(_NewObj_ByIndex(self, index));
+      if (value1 == NULL) {
+        return -1;
+      }
+      PyObject* value2 = PyList_GetItem(other, index);
+      if (value2 == NULL) {
+        return -1;
+      }
+      int cmp = PyObject_RichCompareBool(value1.get(), value2, Py_EQ);
+      if (cmp != 1)  // error or not equal
+          return cmp;
+    }
+    // All items were found and equal
+    return 1;
+  }
+
+  // Any other object is different.
+  return 0;
+}
+
+// A mapping container can only be equal to another mapping container, or (for
+// backward compatibility) to a dict containing the same items.
+// Returns 1 if equal, 0 if unequal, -1 on error.
+static int DescriptorMapping_Equal(PyContainer* self, PyObject* other) {
+  // Check the identity of C++ pointers.
+  if (PyObject_TypeCheck(other, &DescriptorMapping_Type)) {
+    PyContainer* other_container = reinterpret_cast<PyContainer*>(other);
+    if (self->descriptor == other_container->descriptor &&
+        self->container_def == other_container->container_def &&
+        self->kind == other_container->kind) {
+      return 1;
+    } else {
+      return 0;
+    }
+  }
+
+  // If other is a dict
+  if (PyDict_Check(other)) {
+    // equivalent to dict(self.items()) == other
+    int size = Length(self);
+    if (size != PyDict_Size(other)) {
+      return false;
+    }
+    for (int index = 0; index < size; index++) {
+      ScopedPyObjectPtr key(_NewKey_ByIndex(self, index));
+      if (key == NULL) {
+        return -1;
+      }
+      ScopedPyObjectPtr value1(_NewObj_ByIndex(self, index));
+      if (value1 == NULL) {
+        return -1;
+      }
+      PyObject* value2 = PyDict_GetItem(other, key.get());
+      if (value2 == NULL) {
+        // Not found in the other dictionary
+        return 0;
+      }
+      int cmp = PyObject_RichCompareBool(value1.get(), value2, Py_EQ);
+      if (cmp != 1)  // error or not equal
+          return cmp;
+    }
+    // All items were found and equal
+    return 1;
+  }
+
+  // Any other object is different.
+  return 0;
+}
+
+static PyObject* RichCompare(PyContainer* self, PyObject* other, int opid) {
+  if (opid != Py_EQ && opid != Py_NE) {
+    Py_INCREF(Py_NotImplemented);
+    return Py_NotImplemented;
+  }
+
+  int result;
+
+  if (self->kind == PyContainer::KIND_SEQUENCE) {
+    result = DescriptorSequence_Equal(self, other);
+  } else {
+    result = DescriptorMapping_Equal(self, other);
+  }
+  if (result < 0) {
+    return NULL;
+  }
+  if (result ^ (opid == Py_NE)) {
+    Py_RETURN_TRUE;
+  } else {
+    Py_RETURN_FALSE;
+  }
+}
+
+static PySequenceMethods MappingSequenceMethods = {
+    0,                      // sq_length
+    0,                      // sq_concat
+    0,                      // sq_repeat
+    0,                      // sq_item
+    0,                      // sq_slice
+    0,                      // sq_ass_item
+    0,                      // sq_ass_slice
+    (objobjproc)Contains,   // sq_contains
+};
+
+static PyObject* Get(PyContainer* self, PyObject* args) {
+  PyObject* key;
+  PyObject* default_value = Py_None;
+  if (!PyArg_UnpackTuple(args, "get", 1, 2, &key, &default_value)) {
+    return NULL;
+  }
+
+  const void* item;
+  if (!_GetItemByKey(self, key, &item)) {
+    return NULL;
+  }
+  if (item == NULL) {
+    Py_INCREF(default_value);
+    return default_value;
+  }
+  return self->container_def->new_object_from_item_fn(item);
+}
+
+static PyObject* Keys(PyContainer* self, PyObject* args) {
+  Py_ssize_t count = Length(self);
+  ScopedPyObjectPtr list(PyList_New(count));
+  if (list == NULL) {
+    return NULL;
+  }
+  for (Py_ssize_t index = 0; index < count; ++index) {
+    PyObject* key = _NewKey_ByIndex(self, index);
+    if (key == NULL) {
+      return NULL;
+    }
+    PyList_SET_ITEM(list.get(), index, key);
+  }
+  return list.release();
+}
+
+static PyObject* Values(PyContainer* self, PyObject* args) {
+  Py_ssize_t count = Length(self);
+  ScopedPyObjectPtr list(PyList_New(count));
+  if (list == NULL) {
+    return NULL;
+  }
+  for (Py_ssize_t index = 0; index < count; ++index) {
+    PyObject* value = _NewObj_ByIndex(self, index);
+    if (value == NULL) {
+      return NULL;
+    }
+    PyList_SET_ITEM(list.get(), index, value);
+  }
+  return list.release();
+}
+
+static PyObject* Items(PyContainer* self, PyObject* args) {
+  Py_ssize_t count = Length(self);
+  ScopedPyObjectPtr list(PyList_New(count));
+  if (list == NULL) {
+    return NULL;
+  }
+  for (Py_ssize_t index = 0; index < count; ++index) {
+    ScopedPyObjectPtr obj(PyTuple_New(2));
+    if (obj == NULL) {
+      return NULL;
+    }
+    PyObject* key = _NewKey_ByIndex(self, index);
+    if (key == NULL) {
+      return NULL;
+    }
+    PyTuple_SET_ITEM(obj.get(), 0, key);
+    PyObject* value = _NewObj_ByIndex(self, index);
+    if (value == NULL) {
+      return NULL;
+    }
+    PyTuple_SET_ITEM(obj.get(), 1, value);
+    PyList_SET_ITEM(list.get(), index, obj.release());
+  }
+  return list.release();
+}
+
+static PyObject* NewContainerIterator(PyContainer* mapping,
+                                      PyContainerIterator::IterKind kind);
+
+static PyObject* Iter(PyContainer* self) {
+  return NewContainerIterator(self, PyContainerIterator::KIND_ITERKEY);
+}
+static PyObject* IterKeys(PyContainer* self, PyObject* args) {
+  return NewContainerIterator(self, PyContainerIterator::KIND_ITERKEY);
+}
+static PyObject* IterValues(PyContainer* self, PyObject* args) {
+  return NewContainerIterator(self, PyContainerIterator::KIND_ITERVALUE);
+}
+static PyObject* IterItems(PyContainer* self, PyObject* args) {
+  return NewContainerIterator(self, PyContainerIterator::KIND_ITERITEM);
+}
+
+static PyMethodDef MappingMethods[] = {
+  { "get", (PyCFunction)Get, METH_VARARGS, },
+  { "keys", (PyCFunction)Keys, METH_NOARGS, },
+  { "values", (PyCFunction)Values, METH_NOARGS, },
+  { "items", (PyCFunction)Items, METH_NOARGS, },
+  { "iterkeys", (PyCFunction)IterKeys, METH_NOARGS, },
+  { "itervalues", (PyCFunction)IterValues, METH_NOARGS, },
+  { "iteritems", (PyCFunction)IterItems, METH_NOARGS, },
+  {NULL}
+};
+
+PyTypeObject DescriptorMapping_Type = {
+  PyVarObject_HEAD_INIT(&PyType_Type, 0)
+  "DescriptorMapping",                  // tp_name
+  sizeof(PyContainer),                  // tp_basicsize
+  0,                                    // tp_itemsize
+  0,                                    // tp_dealloc
+  0,                                    // tp_print
+  0,                                    // tp_getattr
+  0,                                    // tp_setattr
+  0,                                    // tp_compare
+  (reprfunc)ContainerRepr,              // tp_repr
+  0,                                    // tp_as_number
+  &MappingSequenceMethods,              // tp_as_sequence
+  &MappingMappingMethods,               // tp_as_mapping
+  0,                                    // tp_hash
+  0,                                    // tp_call
+  0,                                    // tp_str
+  0,                                    // tp_getattro
+  0,                                    // tp_setattro
+  0,                                    // tp_as_buffer
+  Py_TPFLAGS_DEFAULT,                   // tp_flags
+  0,                                    // tp_doc
+  0,                                    // tp_traverse
+  0,                                    // tp_clear
+  (richcmpfunc)RichCompare,             // tp_richcompare
+  0,                                    // tp_weaklistoffset
+  (getiterfunc)Iter,                    // tp_iter
+  0,                                    // tp_iternext
+  MappingMethods,                       // tp_methods
+  0,                                    // tp_members
+  0,                                    // tp_getset
+  0,                                    // tp_base
+  0,                                    // tp_dict
+  0,                                    // tp_descr_get
+  0,                                    // tp_descr_set
+  0,                                    // tp_dictoffset
+  0,                                    // tp_init
+  0,                                    // tp_alloc
+  0,                                    // tp_new
+  0,                                    // tp_free
+};
+
+// The DescriptorSequence type.
+
+static PyObject* GetItem(PyContainer* self, Py_ssize_t index) {
+  if (index < 0) {
+    index += Length(self);
+  }
+  if (index < 0 || index >= Length(self)) {
+    PyErr_SetString(PyExc_IndexError, "index out of range");
+    return NULL;
+  }
+  return _NewObj_ByIndex(self, index);
+}
+
+// Returns the position of the item in the sequence, of -1 if not found.
+// This function never fails.
+int Find(PyContainer* self, PyObject* item) {
+  // The item can only be in one position: item.index.
+  // Check that self[item.index] == item, it's faster than a linear search.
+  //
+  // This assumes that sequences are only defined by syntax of the .proto file:
+  // a specific item belongs to only one sequence, depending on its position in
+  // the .proto file definition.
+  const void* descriptor_ptr = PyDescriptor_AsVoidPtr(item);
+  if (descriptor_ptr == NULL) {
+    // Not a descriptor, it cannot be in the list.
+    return -1;
+  }
+  if (self->container_def->get_item_index_fn) {
+    int index = self->container_def->get_item_index_fn(descriptor_ptr);
+    if (index < 0 || index >= Length(self)) {
+      // This index is not from this collection.
+      return -1;
+    }
+    if (self->container_def->get_by_index_fn(self, index) != descriptor_ptr) {
+      // The descriptor at this index is not the same.
+      return -1;
+    }
+    // self[item.index] == item, so return the index.
+    return index;
+  } else {
+    // Fall back to linear search.
+    int length = Length(self);
+    for (int index=0; index < length; index++) {
+      if (self->container_def->get_by_index_fn(self, index) == descriptor_ptr) {
+        return index;
+      }
+    }
+    // Not found
+    return -1;
+  }
+}
+
+// Implements list.index(): the position of the item is in the sequence.
+static PyObject* Index(PyContainer* self, PyObject* item) {
+  int position = Find(self, item);
+  if (position < 0) {
+    // Not found
+    PyErr_SetNone(PyExc_ValueError);
+    return NULL;
+  } else {
+    return PyInt_FromLong(position);
+  }
+}
+// Implements "list.__contains__()": is the object in the sequence.
+static int SeqContains(PyContainer* self, PyObject* item) {
+  int position = Find(self, item);
+  if (position < 0) {
+    return 0;
+  } else {
+    return 1;
+  }
+}
+
+// Implements list.count(): number of occurrences of the item in the sequence.
+// An item can only appear once in a sequence. If it exists, return 1.
+static PyObject* Count(PyContainer* self, PyObject* item) {
+  int position = Find(self, item);
+  if (position < 0) {
+    return PyInt_FromLong(0);
+  } else {
+    return PyInt_FromLong(1);
+  }
+}
+
+static PyObject* Append(PyContainer* self, PyObject* args) {
+  if (_CalledFromGeneratedFile(0)) {
+    Py_RETURN_NONE;
+  }
+  PyErr_Format(PyExc_TypeError,
+               "'%.200s' object is not a mutable sequence",
+               Py_TYPE(self)->tp_name);
+  return NULL;
+}
+
+static PyObject* Reversed(PyContainer* self, PyObject* args) {
+  return NewContainerIterator(self,
+                              PyContainerIterator::KIND_ITERVALUE_REVERSED);
+}
+
+static PyMethodDef SeqMethods[] = {
+  { "index", (PyCFunction)Index, METH_O, },
+  { "count", (PyCFunction)Count, METH_O, },
+  { "append", (PyCFunction)Append, METH_O, },
+  { "__reversed__", (PyCFunction)Reversed, METH_NOARGS, },
+  {NULL}
+};
+
+static PySequenceMethods SeqSequenceMethods = {
+    (lenfunc)Length,          // sq_length
+    0,                        // sq_concat
+    0,                        // sq_repeat
+    (ssizeargfunc)GetItem,    // sq_item
+    0,                        // sq_slice
+    0,                        // sq_ass_item
+    0,                        // sq_ass_slice
+    (objobjproc)SeqContains,  // sq_contains
+};
+
+PyTypeObject DescriptorSequence_Type = {
+  PyVarObject_HEAD_INIT(&PyType_Type, 0)
+  "DescriptorSequence",                 // tp_name
+  sizeof(PyContainer),                  // tp_basicsize
+  0,                                    // tp_itemsize
+  0,                                    // tp_dealloc
+  0,                                    // tp_print
+  0,                                    // tp_getattr
+  0,                                    // tp_setattr
+  0,                                    // tp_compare
+  (reprfunc)ContainerRepr,              // tp_repr
+  0,                                    // tp_as_number
+  &SeqSequenceMethods,                  // tp_as_sequence
+  0,                                    // tp_as_mapping
+  0,                                    // tp_hash
+  0,                                    // tp_call
+  0,                                    // tp_str
+  0,                                    // tp_getattro
+  0,                                    // tp_setattro
+  0,                                    // tp_as_buffer
+  Py_TPFLAGS_DEFAULT,                   // tp_flags
+  0,                                    // tp_doc
+  0,                                    // tp_traverse
+  0,                                    // tp_clear
+  (richcmpfunc)RichCompare,             // tp_richcompare
+  0,                                    // tp_weaklistoffset
+  0,                                    // tp_iter
+  0,                                    // tp_iternext
+  SeqMethods,                           // tp_methods
+  0,                                    // tp_members
+  0,                                    // tp_getset
+  0,                                    // tp_base
+  0,                                    // tp_dict
+  0,                                    // tp_descr_get
+  0,                                    // tp_descr_set
+  0,                                    // tp_dictoffset
+  0,                                    // tp_init
+  0,                                    // tp_alloc
+  0,                                    // tp_new
+  0,                                    // tp_free
+};
+
+static PyObject* NewMappingByName(
+    DescriptorContainerDef* container_def, const void* descriptor) {
+  PyContainer* self = PyObject_New(PyContainer, &DescriptorMapping_Type);
+  if (self == NULL) {
+    return NULL;
+  }
+  self->descriptor = descriptor;
+  self->container_def = container_def;
+  self->kind = PyContainer::KIND_BYNAME;
+  return reinterpret_cast<PyObject*>(self);
+}
+
+static PyObject* NewMappingByCamelcaseName(
+    DescriptorContainerDef* container_def, const void* descriptor) {
+  PyContainer* self = PyObject_New(PyContainer, &DescriptorMapping_Type);
+  if (self == NULL) {
+    return NULL;
+  }
+  self->descriptor = descriptor;
+  self->container_def = container_def;
+  self->kind = PyContainer::KIND_BYCAMELCASENAME;
+  return reinterpret_cast<PyObject*>(self);
+}
+
+static PyObject* NewMappingByNumber(
+    DescriptorContainerDef* container_def, const void* descriptor) {
+  if (container_def->get_by_number_fn == NULL ||
+      container_def->get_item_number_fn == NULL) {
+    PyErr_SetNone(PyExc_NotImplementedError);
+    return NULL;
+  }
+  PyContainer* self = PyObject_New(PyContainer, &DescriptorMapping_Type);
+  if (self == NULL) {
+    return NULL;
+  }
+  self->descriptor = descriptor;
+  self->container_def = container_def;
+  self->kind = PyContainer::KIND_BYNUMBER;
+  return reinterpret_cast<PyObject*>(self);
+}
+
+static PyObject* NewSequence(
+    DescriptorContainerDef* container_def, const void* descriptor) {
+  PyContainer* self = PyObject_New(PyContainer, &DescriptorSequence_Type);
+  if (self == NULL) {
+    return NULL;
+  }
+  self->descriptor = descriptor;
+  self->container_def = container_def;
+  self->kind = PyContainer::KIND_SEQUENCE;
+  return reinterpret_cast<PyObject*>(self);
+}
+
+// Implement iterators over PyContainers.
+
+static void Iterator_Dealloc(PyContainerIterator* self) {
+  Py_CLEAR(self->container);
+  Py_TYPE(self)->tp_free(reinterpret_cast<PyObject*>(self));
+}
+
+static PyObject* Iterator_Next(PyContainerIterator* self) {
+  int count = self->container->container_def->count_fn(self->container);
+  if (self->index >= count) {
+    // Return NULL with no exception to indicate the end.
+    return NULL;
+  }
+  int index = self->index;
+  self->index += 1;
+  switch (self->kind) {
+    case PyContainerIterator::KIND_ITERKEY:
+      return _NewKey_ByIndex(self->container, index);
+    case PyContainerIterator::KIND_ITERVALUE:
+      return _NewObj_ByIndex(self->container, index);
+    case PyContainerIterator::KIND_ITERVALUE_REVERSED:
+      return _NewObj_ByIndex(self->container, count - index - 1);
+    case PyContainerIterator::KIND_ITERITEM:
+      {
+        PyObject* obj = PyTuple_New(2);
+        if (obj == NULL) {
+          return NULL;
+        }
+        PyObject* key = _NewKey_ByIndex(self->container, index);
+        if (key == NULL) {
+          Py_DECREF(obj);
+          return NULL;
+        }
+        PyTuple_SET_ITEM(obj, 0, key);
+        PyObject* value = _NewObj_ByIndex(self->container, index);
+        if (value == NULL) {
+          Py_DECREF(obj);
+          return NULL;
+        }
+        PyTuple_SET_ITEM(obj, 1, value);
+        return obj;
+      }
+    default:
+      PyErr_SetNone(PyExc_NotImplementedError);
+      return NULL;
+  }
+}
+
+static PyTypeObject ContainerIterator_Type = {
+  PyVarObject_HEAD_INIT(&PyType_Type, 0)
+  "DescriptorContainerIterator",        // tp_name
+  sizeof(PyContainerIterator),          // tp_basicsize
+  0,                                    // tp_itemsize
+  (destructor)Iterator_Dealloc,         // tp_dealloc
+  0,                                    // tp_print
+  0,                                    // tp_getattr
+  0,                                    // tp_setattr
+  0,                                    // tp_compare
+  0,                                    // tp_repr
+  0,                                    // tp_as_number
+  0,                                    // tp_as_sequence
+  0,                                    // tp_as_mapping
+  0,                                    // tp_hash
+  0,                                    // tp_call
+  0,                                    // tp_str
+  0,                                    // tp_getattro
+  0,                                    // tp_setattro
+  0,                                    // tp_as_buffer
+  Py_TPFLAGS_DEFAULT,                   // tp_flags
+  0,                                    // tp_doc
+  0,                                    // tp_traverse
+  0,                                    // tp_clear
+  0,                                    // tp_richcompare
+  0,                                    // tp_weaklistoffset
+  PyObject_SelfIter,                    // tp_iter
+  (iternextfunc)Iterator_Next,          // tp_iternext
+  0,                                    // tp_methods
+  0,                                    // tp_members
+  0,                                    // tp_getset
+  0,                                    // tp_base
+  0,                                    // tp_dict
+  0,                                    // tp_descr_get
+  0,                                    // tp_descr_set
+  0,                                    // tp_dictoffset
+  0,                                    // tp_init
+  0,                                    // tp_alloc
+  0,                                    // tp_new
+  0,                                    // tp_free
+};
+
+static PyObject* NewContainerIterator(PyContainer* container,
+                                      PyContainerIterator::IterKind kind) {
+  PyContainerIterator* self = PyObject_New(PyContainerIterator,
+                                           &ContainerIterator_Type);
+  if (self == NULL) {
+    return NULL;
+  }
+  Py_INCREF(container);
+  self->container = container;
+  self->kind = kind;
+  self->index = 0;
+
+  return reinterpret_cast<PyObject*>(self);
+}
+
+}  // namespace descriptor
+
+// Now define the real collections!
+
+namespace message_descriptor {
+
+typedef const Descriptor* ParentDescriptor;
+
+static ParentDescriptor GetDescriptor(PyContainer* self) {
+  return reinterpret_cast<ParentDescriptor>(self->descriptor);
+}
+
+namespace fields {
+
+typedef const FieldDescriptor* ItemDescriptor;
+
+static int Count(PyContainer* self) {
+  return GetDescriptor(self)->field_count();
+}
+
+static ItemDescriptor GetByName(PyContainer* self, const string& name) {
+  return GetDescriptor(self)->FindFieldByName(name);
+}
+
+static ItemDescriptor GetByCamelcaseName(PyContainer* self,
+                                         const string& name) {
+  return GetDescriptor(self)->FindFieldByCamelcaseName(name);
+}
+
+static ItemDescriptor GetByNumber(PyContainer* self, int number) {
+  return GetDescriptor(self)->FindFieldByNumber(number);
+}
+
+static ItemDescriptor GetByIndex(PyContainer* self, int index) {
+  return GetDescriptor(self)->field(index);
+}
+
+static PyObject* NewObjectFromItem(ItemDescriptor item) {
+  return PyFieldDescriptor_FromDescriptor(item);
+}
+
+static const string& GetItemName(ItemDescriptor item) {
+  return item->name();
+}
+
+static const string& GetItemCamelcaseName(ItemDescriptor item) {
+  return item->camelcase_name();
+}
+
+static int GetItemNumber(ItemDescriptor item) {
+  return item->number();
+}
+
+static int GetItemIndex(ItemDescriptor item) {
+  return item->index();
+}
+
+static DescriptorContainerDef ContainerDef = {
+  "MessageFields",
+  (CountMethod)Count,
+  (GetByIndexMethod)GetByIndex,
+  (GetByNameMethod)GetByName,
+  (GetByCamelcaseNameMethod)GetByCamelcaseName,
+  (GetByNumberMethod)GetByNumber,
+  (NewObjectFromItemMethod)NewObjectFromItem,
+  (GetItemNameMethod)GetItemName,
+  (GetItemCamelcaseNameMethod)GetItemCamelcaseName,
+  (GetItemNumberMethod)GetItemNumber,
+  (GetItemIndexMethod)GetItemIndex,
+};
+
+}  // namespace fields
+
+PyObject* NewMessageFieldsByName(ParentDescriptor descriptor) {
+  return descriptor::NewMappingByName(&fields::ContainerDef, descriptor);
+}
+
+PyObject* NewMessageFieldsByCamelcaseName(ParentDescriptor descriptor) {
+  return descriptor::NewMappingByCamelcaseName(&fields::ContainerDef,
+                                               descriptor);
+}
+
+PyObject* NewMessageFieldsByNumber(ParentDescriptor descriptor) {
+  return descriptor::NewMappingByNumber(&fields::ContainerDef, descriptor);
+}
+
+PyObject* NewMessageFieldsSeq(ParentDescriptor descriptor) {
+  return descriptor::NewSequence(&fields::ContainerDef, descriptor);
+}
+
+namespace nested_types {
+
+typedef const Descriptor* ItemDescriptor;
+
+static int Count(PyContainer* self) {
+  return GetDescriptor(self)->nested_type_count();
+}
+
+static ItemDescriptor GetByName(PyContainer* self, const string& name) {
+  return GetDescriptor(self)->FindNestedTypeByName(name);
+}
+
+static ItemDescriptor GetByIndex(PyContainer* self, int index) {
+  return GetDescriptor(self)->nested_type(index);
+}
+
+static PyObject* NewObjectFromItem(ItemDescriptor item) {
+  return PyMessageDescriptor_FromDescriptor(item);
+}
+
+static const string& GetItemName(ItemDescriptor item) {
+  return item->name();
+}
+
+static int GetItemIndex(ItemDescriptor item) {
+  return item->index();
+}
+
+static DescriptorContainerDef ContainerDef = {
+  "MessageNestedTypes",
+  (CountMethod)Count,
+  (GetByIndexMethod)GetByIndex,
+  (GetByNameMethod)GetByName,
+  (GetByCamelcaseNameMethod)NULL,
+  (GetByNumberMethod)NULL,
+  (NewObjectFromItemMethod)NewObjectFromItem,
+  (GetItemNameMethod)GetItemName,
+  (GetItemCamelcaseNameMethod)NULL,
+  (GetItemNumberMethod)NULL,
+  (GetItemIndexMethod)GetItemIndex,
+};
+
+}  // namespace nested_types
+
+PyObject* NewMessageNestedTypesSeq(ParentDescriptor descriptor) {
+  return descriptor::NewSequence(&nested_types::ContainerDef, descriptor);
+}
+
+PyObject* NewMessageNestedTypesByName(ParentDescriptor descriptor) {
+  return descriptor::NewMappingByName(&nested_types::ContainerDef, descriptor);
+}
+
+namespace enums {
+
+typedef const EnumDescriptor* ItemDescriptor;
+
+static int Count(PyContainer* self) {
+  return GetDescriptor(self)->enum_type_count();
+}
+
+static ItemDescriptor GetByName(PyContainer* self, const string& name) {
+  return GetDescriptor(self)->FindEnumTypeByName(name);
+}
+
+static ItemDescriptor GetByIndex(PyContainer* self, int index) {
+  return GetDescriptor(self)->enum_type(index);
+}
+
+static PyObject* NewObjectFromItem(ItemDescriptor item) {
+  return PyEnumDescriptor_FromDescriptor(item);
+}
+
+static const string& GetItemName(ItemDescriptor item) {
+  return item->name();
+}
+
+static int GetItemIndex(ItemDescriptor item) {
+  return item->index();
+}
+
+static DescriptorContainerDef ContainerDef = {
+  "MessageNestedEnums",
+  (CountMethod)Count,
+  (GetByIndexMethod)GetByIndex,
+  (GetByNameMethod)GetByName,
+  (GetByCamelcaseNameMethod)NULL,
+  (GetByNumberMethod)NULL,
+  (NewObjectFromItemMethod)NewObjectFromItem,
+  (GetItemNameMethod)GetItemName,
+  (GetItemCamelcaseNameMethod)NULL,
+  (GetItemNumberMethod)NULL,
+  (GetItemIndexMethod)GetItemIndex,
+};
+
+}  // namespace enums
+
+PyObject* NewMessageEnumsByName(ParentDescriptor descriptor) {
+  return descriptor::NewMappingByName(&enums::ContainerDef, descriptor);
+}
+
+PyObject* NewMessageEnumsSeq(ParentDescriptor descriptor) {
+  return descriptor::NewSequence(&enums::ContainerDef, descriptor);
+}
+
+namespace enumvalues {
+
+// This is the "enum_values_by_name" mapping, which collects values from all
+// enum types in a message.
+//
+// Note that the behavior of the C++ descriptor is different: it will search and
+// return the first value that matches the name, whereas the Python
+// implementation retrieves the last one.
+
+typedef const EnumValueDescriptor* ItemDescriptor;
+
+static int Count(PyContainer* self) {
+  int count = 0;
+  for (int i = 0; i < GetDescriptor(self)->enum_type_count(); ++i) {
+    count += GetDescriptor(self)->enum_type(i)->value_count();
+  }
+  return count;
+}
+
+static ItemDescriptor GetByName(PyContainer* self, const string& name) {
+  return GetDescriptor(self)->FindEnumValueByName(name);
+}
+
+static ItemDescriptor GetByIndex(PyContainer* self, int index) {
+  // This is not optimal, but the number of enums *types* in a given message
+  // is small.  This function is only used when iterating over the mapping.
+  const EnumDescriptor* enum_type = NULL;
+  int enum_type_count = GetDescriptor(self)->enum_type_count();
+  for (int i = 0; i < enum_type_count; ++i) {
+    enum_type = GetDescriptor(self)->enum_type(i);
+    int enum_value_count = enum_type->value_count();
+    if (index < enum_value_count) {
+      // Found it!
+      break;
+    }
+    index -= enum_value_count;
+  }
+  // The next statement cannot overflow, because this function is only called by
+  // internal iterators which ensure that 0 <= index < Count().
+  return enum_type->value(index);
+}
+
+static PyObject* NewObjectFromItem(ItemDescriptor item) {
+  return PyEnumValueDescriptor_FromDescriptor(item);
+}
+
+static const string& GetItemName(ItemDescriptor item) {
+  return item->name();
+}
+
+static DescriptorContainerDef ContainerDef = {
+  "MessageEnumValues",
+  (CountMethod)Count,
+  (GetByIndexMethod)GetByIndex,
+  (GetByNameMethod)GetByName,
+  (GetByCamelcaseNameMethod)NULL,
+  (GetByNumberMethod)NULL,
+  (NewObjectFromItemMethod)NewObjectFromItem,
+  (GetItemNameMethod)GetItemName,
+  (GetItemCamelcaseNameMethod)NULL,
+  (GetItemNumberMethod)NULL,
+  (GetItemIndexMethod)NULL,
+};
+
+}  // namespace enumvalues
+
+PyObject* NewMessageEnumValuesByName(ParentDescriptor descriptor) {
+  return descriptor::NewMappingByName(&enumvalues::ContainerDef, descriptor);
+}
+
+namespace extensions {
+
+typedef const FieldDescriptor* ItemDescriptor;
+
+static int Count(PyContainer* self) {
+  return GetDescriptor(self)->extension_count();
+}
+
+static ItemDescriptor GetByName(PyContainer* self, const string& name) {
+  return GetDescriptor(self)->FindExtensionByName(name);
+}
+
+static ItemDescriptor GetByIndex(PyContainer* self, int index) {
+  return GetDescriptor(self)->extension(index);
+}
+
+static PyObject* NewObjectFromItem(ItemDescriptor item) {
+  return PyFieldDescriptor_FromDescriptor(item);
+}
+
+static const string& GetItemName(ItemDescriptor item) {
+  return item->name();
+}
+
+static int GetItemIndex(ItemDescriptor item) {
+  return item->index();
+}
+
+static DescriptorContainerDef ContainerDef = {
+  "MessageExtensions",
+  (CountMethod)Count,
+  (GetByIndexMethod)GetByIndex,
+  (GetByNameMethod)GetByName,
+  (GetByCamelcaseNameMethod)NULL,
+  (GetByNumberMethod)NULL,
+  (NewObjectFromItemMethod)NewObjectFromItem,
+  (GetItemNameMethod)GetItemName,
+  (GetItemCamelcaseNameMethod)NULL,
+  (GetItemNumberMethod)NULL,
+  (GetItemIndexMethod)GetItemIndex,
+};
+
+}  // namespace extensions
+
+PyObject* NewMessageExtensionsByName(ParentDescriptor descriptor) {
+  return descriptor::NewMappingByName(&extensions::ContainerDef, descriptor);
+}
+
+PyObject* NewMessageExtensionsSeq(ParentDescriptor descriptor) {
+  return descriptor::NewSequence(&extensions::ContainerDef, descriptor);
+}
+
+namespace oneofs {
+
+typedef const OneofDescriptor* ItemDescriptor;
+
+static int Count(PyContainer* self) {
+  return GetDescriptor(self)->oneof_decl_count();
+}
+
+static ItemDescriptor GetByName(PyContainer* self, const string& name) {
+  return GetDescriptor(self)->FindOneofByName(name);
+}
+
+static ItemDescriptor GetByIndex(PyContainer* self, int index) {
+  return GetDescriptor(self)->oneof_decl(index);
+}
+
+static PyObject* NewObjectFromItem(ItemDescriptor item) {
+  return PyOneofDescriptor_FromDescriptor(item);
+}
+
+static const string& GetItemName(ItemDescriptor item) {
+  return item->name();
+}
+
+static int GetItemIndex(ItemDescriptor item) {
+  return item->index();
+}
+
+static DescriptorContainerDef ContainerDef = {
+  "MessageOneofs",
+  (CountMethod)Count,
+  (GetByIndexMethod)GetByIndex,
+  (GetByNameMethod)GetByName,
+  (GetByCamelcaseNameMethod)NULL,
+  (GetByNumberMethod)NULL,
+  (NewObjectFromItemMethod)NewObjectFromItem,
+  (GetItemNameMethod)GetItemName,
+  (GetItemCamelcaseNameMethod)NULL,
+  (GetItemNumberMethod)NULL,
+  (GetItemIndexMethod)GetItemIndex,
+};
+
+}  // namespace oneofs
+
+PyObject* NewMessageOneofsByName(ParentDescriptor descriptor) {
+  return descriptor::NewMappingByName(&oneofs::ContainerDef, descriptor);
+}
+
+PyObject* NewMessageOneofsSeq(ParentDescriptor descriptor) {
+  return descriptor::NewSequence(&oneofs::ContainerDef, descriptor);
+}
+
+}  // namespace message_descriptor
+
+namespace enum_descriptor {
+
+typedef const EnumDescriptor* ParentDescriptor;
+
+static ParentDescriptor GetDescriptor(PyContainer* self) {
+  return reinterpret_cast<ParentDescriptor>(self->descriptor);
+}
+
+namespace enumvalues {
+
+typedef const EnumValueDescriptor* ItemDescriptor;
+
+static int Count(PyContainer* self) {
+  return GetDescriptor(self)->value_count();
+}
+
+static ItemDescriptor GetByIndex(PyContainer* self, int index) {
+  return GetDescriptor(self)->value(index);
+}
+
+static ItemDescriptor GetByName(PyContainer* self, const string& name) {
+  return GetDescriptor(self)->FindValueByName(name);
+}
+
+static ItemDescriptor GetByNumber(PyContainer* self, int number) {
+  return GetDescriptor(self)->FindValueByNumber(number);
+}
+
+static PyObject* NewObjectFromItem(ItemDescriptor item) {
+  return PyEnumValueDescriptor_FromDescriptor(item);
+}
+
+static const string& GetItemName(ItemDescriptor item) {
+  return item->name();
+}
+
+static int GetItemNumber(ItemDescriptor item) {
+  return item->number();
+}
+
+static int GetItemIndex(ItemDescriptor item) {
+  return item->index();
+}
+
+static DescriptorContainerDef ContainerDef = {
+  "EnumValues",
+  (CountMethod)Count,
+  (GetByIndexMethod)GetByIndex,
+  (GetByNameMethod)GetByName,
+  (GetByCamelcaseNameMethod)NULL,
+  (GetByNumberMethod)GetByNumber,
+  (NewObjectFromItemMethod)NewObjectFromItem,
+  (GetItemNameMethod)GetItemName,
+  (GetItemCamelcaseNameMethod)NULL,
+  (GetItemNumberMethod)GetItemNumber,
+  (GetItemIndexMethod)GetItemIndex,
+};
+
+}  // namespace enumvalues
+
+PyObject* NewEnumValuesByName(ParentDescriptor descriptor) {
+  return descriptor::NewMappingByName(&enumvalues::ContainerDef, descriptor);
+}
+
+PyObject* NewEnumValuesByNumber(ParentDescriptor descriptor) {
+  return descriptor::NewMappingByNumber(&enumvalues::ContainerDef, descriptor);
+}
+
+PyObject* NewEnumValuesSeq(ParentDescriptor descriptor) {
+  return descriptor::NewSequence(&enumvalues::ContainerDef, descriptor);
+}
+
+}  // namespace enum_descriptor
+
+namespace oneof_descriptor {
+
+typedef const OneofDescriptor* ParentDescriptor;
+
+static ParentDescriptor GetDescriptor(PyContainer* self) {
+  return reinterpret_cast<ParentDescriptor>(self->descriptor);
+}
+
+namespace fields {
+
+typedef const FieldDescriptor* ItemDescriptor;
+
+static int Count(PyContainer* self) {
+  return GetDescriptor(self)->field_count();
+}
+
+static ItemDescriptor GetByIndex(PyContainer* self, int index) {
+  return GetDescriptor(self)->field(index);
+}
+
+static PyObject* NewObjectFromItem(ItemDescriptor item) {
+  return PyFieldDescriptor_FromDescriptor(item);
+}
+
+static int GetItemIndex(ItemDescriptor item) {
+  return item->index_in_oneof();
+}
+
+static DescriptorContainerDef ContainerDef = {
+  "OneofFields",
+  (CountMethod)Count,
+  (GetByIndexMethod)GetByIndex,
+  (GetByNameMethod)NULL,
+  (GetByCamelcaseNameMethod)NULL,
+  (GetByNumberMethod)NULL,
+  (NewObjectFromItemMethod)NewObjectFromItem,
+  (GetItemNameMethod)NULL,
+  (GetItemCamelcaseNameMethod)NULL,
+  (GetItemNumberMethod)NULL,
+  (GetItemIndexMethod)GetItemIndex,
+};
+
+}  // namespace fields
+
+PyObject* NewOneofFieldsSeq(ParentDescriptor descriptor) {
+  return descriptor::NewSequence(&fields::ContainerDef, descriptor);
+}
+
+}  // namespace oneof_descriptor
+
+namespace file_descriptor {
+
+typedef const FileDescriptor* ParentDescriptor;
+
+static ParentDescriptor GetDescriptor(PyContainer* self) {
+  return reinterpret_cast<ParentDescriptor>(self->descriptor);
+}
+
+namespace messages {
+
+typedef const Descriptor* ItemDescriptor;
+
+static int Count(PyContainer* self) {
+  return GetDescriptor(self)->message_type_count();
+}
+
+static ItemDescriptor GetByName(PyContainer* self, const string& name) {
+  return GetDescriptor(self)->FindMessageTypeByName(name);
+}
+
+static ItemDescriptor GetByIndex(PyContainer* self, int index) {
+  return GetDescriptor(self)->message_type(index);
+}
+
+static PyObject* NewObjectFromItem(ItemDescriptor item) {
+  return PyMessageDescriptor_FromDescriptor(item);
+}
+
+static const string& GetItemName(ItemDescriptor item) {
+  return item->name();
+}
+
+static int GetItemIndex(ItemDescriptor item) {
+  return item->index();
+}
+
+static DescriptorContainerDef ContainerDef = {
+  "FileMessages",
+  (CountMethod)Count,
+  (GetByIndexMethod)GetByIndex,
+  (GetByNameMethod)GetByName,
+  (GetByCamelcaseNameMethod)NULL,
+  (GetByNumberMethod)NULL,
+  (NewObjectFromItemMethod)NewObjectFromItem,
+  (GetItemNameMethod)GetItemName,
+  (GetItemCamelcaseNameMethod)NULL,
+  (GetItemNumberMethod)NULL,
+  (GetItemIndexMethod)GetItemIndex,
+};
+
+}  // namespace messages
+
+PyObject* NewFileMessageTypesByName(const FileDescriptor* descriptor) {
+  return descriptor::NewMappingByName(&messages::ContainerDef, descriptor);
+}
+
+namespace enums {
+
+typedef const EnumDescriptor* ItemDescriptor;
+
+static int Count(PyContainer* self) {
+  return GetDescriptor(self)->enum_type_count();
+}
+
+static ItemDescriptor GetByName(PyContainer* self, const string& name) {
+  return GetDescriptor(self)->FindEnumTypeByName(name);
+}
+
+static ItemDescriptor GetByIndex(PyContainer* self, int index) {
+  return GetDescriptor(self)->enum_type(index);
+}
+
+static PyObject* NewObjectFromItem(ItemDescriptor item) {
+  return PyEnumDescriptor_FromDescriptor(item);
+}
+
+static const string& GetItemName(ItemDescriptor item) {
+  return item->name();
+}
+
+static int GetItemIndex(ItemDescriptor item) {
+  return item->index();
+}
+
+static DescriptorContainerDef ContainerDef = {
+  "FileEnums",
+  (CountMethod)Count,
+  (GetByIndexMethod)GetByIndex,
+  (GetByNameMethod)GetByName,
+  (GetByCamelcaseNameMethod)NULL,
+  (GetByNumberMethod)NULL,
+  (NewObjectFromItemMethod)NewObjectFromItem,
+  (GetItemNameMethod)GetItemName,
+  (GetItemCamelcaseNameMethod)NULL,
+  (GetItemNumberMethod)NULL,
+  (GetItemIndexMethod)GetItemIndex,
+};
+
+}  // namespace enums
+
+PyObject* NewFileEnumTypesByName(const FileDescriptor* descriptor) {
+  return descriptor::NewMappingByName(&enums::ContainerDef, descriptor);
+}
+
+namespace extensions {
+
+typedef const FieldDescriptor* ItemDescriptor;
+
+static int Count(PyContainer* self) {
+  return GetDescriptor(self)->extension_count();
+}
+
+static ItemDescriptor GetByName(PyContainer* self, const string& name) {
+  return GetDescriptor(self)->FindExtensionByName(name);
+}
+
+static ItemDescriptor GetByIndex(PyContainer* self, int index) {
+  return GetDescriptor(self)->extension(index);
+}
+
+static PyObject* NewObjectFromItem(ItemDescriptor item) {
+  return PyFieldDescriptor_FromDescriptor(item);
+}
+
+static const string& GetItemName(ItemDescriptor item) {
+  return item->name();
+}
+
+static int GetItemIndex(ItemDescriptor item) {
+  return item->index();
+}
+
+static DescriptorContainerDef ContainerDef = {
+  "FileExtensions",
+  (CountMethod)Count,
+  (GetByIndexMethod)GetByIndex,
+  (GetByNameMethod)GetByName,
+  (GetByCamelcaseNameMethod)NULL,
+  (GetByNumberMethod)NULL,
+  (NewObjectFromItemMethod)NewObjectFromItem,
+  (GetItemNameMethod)GetItemName,
+  (GetItemCamelcaseNameMethod)NULL,
+  (GetItemNumberMethod)NULL,
+  (GetItemIndexMethod)GetItemIndex,
+};
+
+}  // namespace extensions
+
+PyObject* NewFileExtensionsByName(const FileDescriptor* descriptor) {
+  return descriptor::NewMappingByName(&extensions::ContainerDef, descriptor);
+}
+
+namespace dependencies {
+
+typedef const FileDescriptor* ItemDescriptor;
+
+static int Count(PyContainer* self) {
+  return GetDescriptor(self)->dependency_count();
+}
+
+static ItemDescriptor GetByIndex(PyContainer* self, int index) {
+  return GetDescriptor(self)->dependency(index);
+}
+
+static PyObject* NewObjectFromItem(ItemDescriptor item) {
+  return PyFileDescriptor_FromDescriptor(item);
+}
+
+static DescriptorContainerDef ContainerDef = {
+  "FileDependencies",
+  (CountMethod)Count,
+  (GetByIndexMethod)GetByIndex,
+  (GetByNameMethod)NULL,
+  (GetByCamelcaseNameMethod)NULL,
+  (GetByNumberMethod)NULL,
+  (NewObjectFromItemMethod)NewObjectFromItem,
+  (GetItemNameMethod)NULL,
+  (GetItemCamelcaseNameMethod)NULL,
+  (GetItemNumberMethod)NULL,
+  (GetItemIndexMethod)NULL,
+};
+
+}  // namespace dependencies
+
+PyObject* NewFileDependencies(const FileDescriptor* descriptor) {
+  return descriptor::NewSequence(&dependencies::ContainerDef, descriptor);
+}
+
+namespace public_dependencies {
+
+typedef const FileDescriptor* ItemDescriptor;
+
+static int Count(PyContainer* self) {
+  return GetDescriptor(self)->public_dependency_count();
+}
+
+static ItemDescriptor GetByIndex(PyContainer* self, int index) {
+  return GetDescriptor(self)->public_dependency(index);
+}
+
+static PyObject* NewObjectFromItem(ItemDescriptor item) {
+  return PyFileDescriptor_FromDescriptor(item);
+}
+
+static DescriptorContainerDef ContainerDef = {
+  "FilePublicDependencies",
+  (CountMethod)Count,
+  (GetByIndexMethod)GetByIndex,
+  (GetByNameMethod)NULL,
+  (GetByCamelcaseNameMethod)NULL,
+  (GetByNumberMethod)NULL,
+  (NewObjectFromItemMethod)NewObjectFromItem,
+  (GetItemNameMethod)NULL,
+  (GetItemCamelcaseNameMethod)NULL,
+  (GetItemNumberMethod)NULL,
+  (GetItemIndexMethod)NULL,
+};
+
+}  // namespace public_dependencies
+
+PyObject* NewFilePublicDependencies(const FileDescriptor* descriptor) {
+  return descriptor::NewSequence(&public_dependencies::ContainerDef,
+                                 descriptor);
+}
+
+}  // namespace file_descriptor
+
+
+// Register all implementations
+
+bool InitDescriptorMappingTypes() {
+  if (PyType_Ready(&descriptor::DescriptorMapping_Type) < 0)
+    return false;
+  if (PyType_Ready(&descriptor::DescriptorSequence_Type) < 0)
+    return false;
+  if (PyType_Ready(&descriptor::ContainerIterator_Type) < 0)
+    return false;
+  return true;
+}
+
+}  // namespace python
+}  // namespace protobuf
+}  // namespace google
diff --git a/third_party/protobuf/python/google/protobuf/pyext/descriptor_containers.h b/third_party/protobuf/python/google/protobuf/pyext/descriptor_containers.h
new file mode 100644
index 0000000000..ce40747d57
--- /dev/null
+++ b/third_party/protobuf/python/google/protobuf/pyext/descriptor_containers.h
@@ -0,0 +1,101 @@
+// 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.
+
+#ifndef GOOGLE_PROTOBUF_PYTHON_CPP_DESCRIPTOR_CONTAINERS_H__
+#define GOOGLE_PROTOBUF_PYTHON_CPP_DESCRIPTOR_CONTAINERS_H__
+
+// Mappings and Sequences of descriptors.
+// They implement containers like fields_by_name, EnumDescriptor.values...
+// See descriptor_containers.cc for more description.
+#include <Python.h>
+
+namespace google {
+namespace protobuf {
+
+class Descriptor;
+class FileDescriptor;
+class EnumDescriptor;
+class OneofDescriptor;
+
+namespace python {
+
+// Initialize the various types and objects.
+bool InitDescriptorMappingTypes();
+
+// Each function below returns a Mapping, or a Sequence of descriptors.
+// They all return a new reference.
+
+namespace message_descriptor {
+PyObject* NewMessageFieldsByName(const Descriptor* descriptor);
+PyObject* NewMessageFieldsByCamelcaseName(const Descriptor* descriptor);
+PyObject* NewMessageFieldsByNumber(const Descriptor* descriptor);
+PyObject* NewMessageFieldsSeq(const Descriptor* descriptor);
+
+PyObject* NewMessageNestedTypesSeq(const Descriptor* descriptor);
+PyObject* NewMessageNestedTypesByName(const Descriptor* descriptor);
+
+PyObject* NewMessageEnumsByName(const Descriptor* descriptor);
+PyObject* NewMessageEnumsSeq(const Descriptor* descriptor);
+PyObject* NewMessageEnumValuesByName(const Descriptor* descriptor);
+
+PyObject* NewMessageExtensionsByName(const Descriptor* descriptor);
+PyObject* NewMessageExtensionsSeq(const Descriptor* descriptor);
+
+PyObject* NewMessageOneofsByName(const Descriptor* descriptor);
+PyObject* NewMessageOneofsSeq(const Descriptor* descriptor);
+}  // namespace message_descriptor
+
+namespace enum_descriptor {
+PyObject* NewEnumValuesByName(const EnumDescriptor* descriptor);
+PyObject* NewEnumValuesByNumber(const EnumDescriptor* descriptor);
+PyObject* NewEnumValuesSeq(const EnumDescriptor* descriptor);
+}  // namespace enum_descriptor
+
+namespace oneof_descriptor {
+PyObject* NewOneofFieldsSeq(const OneofDescriptor* descriptor);
+}  // namespace oneof_descriptor
+
+namespace file_descriptor {
+PyObject* NewFileMessageTypesByName(const FileDescriptor* descriptor);
+
+PyObject* NewFileEnumTypesByName(const FileDescriptor* descriptor);
+
+PyObject* NewFileExtensionsByName(const FileDescriptor* descriptor);
+
+PyObject* NewFileDependencies(const FileDescriptor* descriptor);
+PyObject* NewFilePublicDependencies(const FileDescriptor* descriptor);
+}  // namespace file_descriptor
+
+
+}  // namespace python
+}  // namespace protobuf
+
+}  // namespace google
+#endif  // GOOGLE_PROTOBUF_PYTHON_CPP_DESCRIPTOR_CONTAINERS_H__
diff --git a/third_party/protobuf/python/google/protobuf/pyext/descriptor_database.cc b/third_party/protobuf/python/google/protobuf/pyext/descriptor_database.cc
new file mode 100644
index 0000000000..514722b4c3
--- /dev/null
+++ b/third_party/protobuf/python/google/protobuf/pyext/descriptor_database.cc
@@ -0,0 +1,145 @@
+// 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 defines a C++ DescriptorDatabase, which wraps a Python Database
+// and delegate all its operations to Python methods.
+
+#include <google/protobuf/pyext/descriptor_database.h>
+
+#include <google/protobuf/stubs/logging.h>
+#include <google/protobuf/stubs/common.h>
+#include <google/protobuf/descriptor.pb.h>
+#include <google/protobuf/pyext/message.h>
+#include <google/protobuf/pyext/scoped_pyobject_ptr.h>
+
+namespace google {
+namespace protobuf {
+namespace python {
+
+PyDescriptorDatabase::PyDescriptorDatabase(PyObject* py_database)
+    : py_database_(py_database) {
+  Py_INCREF(py_database_);
+}
+
+PyDescriptorDatabase::~PyDescriptorDatabase() { Py_DECREF(py_database_); }
+
+// Convert a Python object to a FileDescriptorProto pointer.
+// Handles all kinds of Python errors, which are simply logged.
+static bool GetFileDescriptorProto(PyObject* py_descriptor,
+                                   FileDescriptorProto* output) {
+  if (py_descriptor == NULL) {
+    if (PyErr_ExceptionMatches(PyExc_KeyError)) {
+      // Expected error: item was simply not found.
+      PyErr_Clear();
+    } else {
+      GOOGLE_LOG(ERROR) << "DescriptorDatabase method raised an error";
+      PyErr_Print();
+    }
+    return false;
+  }
+  const Descriptor* filedescriptor_descriptor =
+      FileDescriptorProto::default_instance().GetDescriptor();
+  CMessage* message = reinterpret_cast<CMessage*>(py_descriptor);
+  if (PyObject_TypeCheck(py_descriptor, &CMessage_Type) &&
+      message->message->GetDescriptor() == filedescriptor_descriptor) {
+    // Fast path: Just use the pointer.
+    FileDescriptorProto* file_proto =
+        static_cast<FileDescriptorProto*>(message->message);
+    *output = *file_proto;
+    return true;
+  } else {
+    // Slow path: serialize the message. This allows to use databases which
+    // use a different implementation of FileDescriptorProto.
+    ScopedPyObjectPtr serialized_pb(
+        PyObject_CallMethod(py_descriptor, "SerializeToString", NULL));
+    if (serialized_pb == NULL) {
+      GOOGLE_LOG(ERROR)
+          << "DescriptorDatabase method did not return a FileDescriptorProto";
+      PyErr_Print();
+      return false;
+    }
+    char* str;
+    Py_ssize_t len;
+    if (PyBytes_AsStringAndSize(serialized_pb.get(), &str, &len) < 0) {
+      GOOGLE_LOG(ERROR)
+          << "DescriptorDatabase method did not return a FileDescriptorProto";
+      PyErr_Print();
+      return false;
+    }
+    FileDescriptorProto file_proto;
+    if (!file_proto.ParseFromArray(str, len)) {
+      GOOGLE_LOG(ERROR)
+          << "DescriptorDatabase method did not return a FileDescriptorProto";
+      return false;
+    }
+    *output = file_proto;
+    return true;
+  }
+}
+
+// Find a file by file name.
+bool PyDescriptorDatabase::FindFileByName(const string& filename,
+                                          FileDescriptorProto* output) {
+  ScopedPyObjectPtr py_descriptor(PyObject_CallMethod(
+      py_database_, "FindFileByName", "s#", filename.c_str(), filename.size()));
+  return GetFileDescriptorProto(py_descriptor.get(), output);
+}
+
+// Find the file that declares the given fully-qualified symbol name.
+bool PyDescriptorDatabase::FindFileContainingSymbol(
+    const string& symbol_name, FileDescriptorProto* output) {
+  ScopedPyObjectPtr py_descriptor(
+      PyObject_CallMethod(py_database_, "FindFileContainingSymbol", "s#",
+                          symbol_name.c_str(), symbol_name.size()));
+  return GetFileDescriptorProto(py_descriptor.get(), output);
+}
+
+// Find the file which defines an extension extending the given message type
+// with the given field number.
+// Python DescriptorDatabases are not required to implement this method.
+bool PyDescriptorDatabase::FindFileContainingExtension(
+    const string& containing_type, int field_number,
+    FileDescriptorProto* output) {
+  ScopedPyObjectPtr py_method(
+      PyObject_GetAttrString(py_database_, "FindFileContainingExtension"));
+  if (py_method == NULL) {
+    // This method is not implemented, returns without error.
+    PyErr_Clear();
+    return false;
+  }
+  ScopedPyObjectPtr py_descriptor(
+      PyObject_CallFunction(py_method.get(), "s#i", containing_type.c_str(),
+                            containing_type.size(), field_number));
+  return GetFileDescriptorProto(py_descriptor.get(), output);
+}
+
+}  // namespace python
+}  // namespace protobuf
+}  // namespace google
diff --git a/third_party/protobuf/python/google/protobuf/pyext/descriptor_database.h b/third_party/protobuf/python/google/protobuf/pyext/descriptor_database.h
new file mode 100644
index 0000000000..fc71c4bcb0
--- /dev/null
+++ b/third_party/protobuf/python/google/protobuf/pyext/descriptor_database.h
@@ -0,0 +1,75 @@
+// 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.
+
+#ifndef GOOGLE_PROTOBUF_PYTHON_CPP_DESCRIPTOR_DATABASE_H__
+#define GOOGLE_PROTOBUF_PYTHON_CPP_DESCRIPTOR_DATABASE_H__
+
+#include <Python.h>
+
+#include <google/protobuf/descriptor_database.h>
+
+namespace google {
+namespace protobuf {
+namespace python {
+
+class PyDescriptorDatabase : public DescriptorDatabase {
+ public:
+  explicit PyDescriptorDatabase(PyObject* py_database);
+  ~PyDescriptorDatabase();
+
+  // Implement the abstract interface. All these functions fill the output
+  // with a copy of FileDescriptorProto.
+
+  // Find a file by file name.
+  bool FindFileByName(const string& filename,
+                      FileDescriptorProto* output);
+
+  // Find the file that declares the given fully-qualified symbol name.
+  bool FindFileContainingSymbol(const string& symbol_name,
+                                FileDescriptorProto* output);
+
+  // Find the file which defines an extension extending the given message type
+  // with the given field number.
+  // Containing_type must be a fully-qualified type name.
+  // Python objects are not required to implement this method.
+  bool FindFileContainingExtension(const string& containing_type,
+                                   int field_number,
+                                   FileDescriptorProto* output);
+
+ private:
+  // The python object that implements the database. The reference is owned.
+  PyObject* py_database_;
+};
+
+}  // namespace python
+}  // namespace protobuf
+
+}  // namespace google
+#endif  // GOOGLE_PROTOBUF_PYTHON_CPP_DESCRIPTOR_DATABASE_H__
diff --git a/third_party/protobuf/python/google/protobuf/pyext/descriptor_pool.cc b/third_party/protobuf/python/google/protobuf/pyext/descriptor_pool.cc
new file mode 100644
index 0000000000..0bc76bc90a
--- /dev/null
+++ b/third_party/protobuf/python/google/protobuf/pyext/descriptor_pool.cc
@@ -0,0 +1,593 @@
+// 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.
+
+// Implements the DescriptorPool, which collects all descriptors.
+
+#include <Python.h>
+
+#include <google/protobuf/descriptor.pb.h>
+#include <google/protobuf/dynamic_message.h>
+#include <google/protobuf/pyext/descriptor.h>
+#include <google/protobuf/pyext/descriptor_database.h>
+#include <google/protobuf/pyext/descriptor_pool.h>
+#include <google/protobuf/pyext/message.h>
+#include <google/protobuf/pyext/scoped_pyobject_ptr.h>
+
+#if PY_MAJOR_VERSION >= 3
+  #define PyString_FromStringAndSize PyUnicode_FromStringAndSize
+  #if PY_VERSION_HEX < 0x03030000
+    #error "Python 3.0 - 3.2 are not supported."
+  #endif
+  #define PyString_AsStringAndSize(ob, charpp, sizep) \
+    (PyUnicode_Check(ob)? \
+       ((*(charpp) = PyUnicode_AsUTF8AndSize(ob, (sizep))) == NULL? -1: 0): \
+       PyBytes_AsStringAndSize(ob, (charpp), (sizep)))
+#endif
+
+namespace google {
+namespace protobuf {
+namespace python {
+
+// A map to cache Python Pools per C++ pointer.
+// Pointers are not owned here, and belong to the PyDescriptorPool.
+static hash_map<const DescriptorPool*, PyDescriptorPool*> descriptor_pool_map;
+
+namespace cdescriptor_pool {
+
+// Create a Python DescriptorPool object, but does not fill the "pool"
+// attribute.
+static PyDescriptorPool* _CreateDescriptorPool() {
+  PyDescriptorPool* cpool = PyObject_New(
+      PyDescriptorPool, &PyDescriptorPool_Type);
+  if (cpool == NULL) {
+    return NULL;
+  }
+
+  cpool->underlay = NULL;
+  cpool->database = NULL;
+
+  DynamicMessageFactory* message_factory = new DynamicMessageFactory();
+  // This option might be the default some day.
+  message_factory->SetDelegateToGeneratedFactory(true);
+  cpool->message_factory = message_factory;
+
+  // TODO(amauryfa): Rewrite the SymbolDatabase in C so that it uses the same
+  // storage.
+  cpool->classes_by_descriptor =
+      new PyDescriptorPool::ClassesByMessageMap();
+  cpool->descriptor_options =
+      new hash_map<const void*, PyObject *>();
+
+  return cpool;
+}
+
+// Create a Python DescriptorPool, using the given pool as an underlay:
+// new messages will be added to a custom pool, not to the underlay.
+//
+// Ownership of the underlay is not transferred, its pointer should
+// stay alive.
+static PyDescriptorPool* PyDescriptorPool_NewWithUnderlay(
+    const DescriptorPool* underlay) {
+  PyDescriptorPool* cpool = _CreateDescriptorPool();
+  if (cpool == NULL) {
+    return NULL;
+  }
+  cpool->pool = new DescriptorPool(underlay);
+  cpool->underlay = underlay;
+
+  if (!descriptor_pool_map.insert(
+      std::make_pair(cpool->pool, cpool)).second) {
+    // Should never happen -- would indicate an internal error / bug.
+    PyErr_SetString(PyExc_ValueError, "DescriptorPool already registered");
+    return NULL;
+  }
+
+  return cpool;
+}
+
+static PyDescriptorPool* PyDescriptorPool_NewWithDatabase(
+    DescriptorDatabase* database) {
+  PyDescriptorPool* cpool = _CreateDescriptorPool();
+  if (cpool == NULL) {
+    return NULL;
+  }
+  if (database != NULL) {
+    cpool->pool = new DescriptorPool(database);
+    cpool->database = database;
+  } else {
+    cpool->pool = new DescriptorPool();
+  }
+
+  if (!descriptor_pool_map.insert(std::make_pair(cpool->pool, cpool)).second) {
+    // Should never happen -- would indicate an internal error / bug.
+    PyErr_SetString(PyExc_ValueError, "DescriptorPool already registered");
+    return NULL;
+  }
+
+  return cpool;
+}
+
+// The public DescriptorPool constructor.
+static PyObject* New(PyTypeObject* type,
+                     PyObject* args, PyObject* kwargs) {
+  static char* kwlist[] = {"descriptor_db", 0};
+  PyObject* py_database = NULL;
+  if (!PyArg_ParseTupleAndKeywords(args, kwargs, "|O", kwlist, &py_database)) {
+    return NULL;
+  }
+  DescriptorDatabase* database = NULL;
+  if (py_database && py_database != Py_None) {
+    database = new PyDescriptorDatabase(py_database);
+  }
+  return reinterpret_cast<PyObject*>(
+      PyDescriptorPool_NewWithDatabase(database));
+}
+
+static void Dealloc(PyDescriptorPool* self) {
+  typedef PyDescriptorPool::ClassesByMessageMap::iterator iterator;
+  descriptor_pool_map.erase(self->pool);
+  for (iterator it = self->classes_by_descriptor->begin();
+       it != self->classes_by_descriptor->end(); ++it) {
+    Py_DECREF(it->second);
+  }
+  delete self->classes_by_descriptor;
+  for (hash_map<const void*, PyObject*>::iterator it =
+           self->descriptor_options->begin();
+       it != self->descriptor_options->end(); ++it) {
+    Py_DECREF(it->second);
+  }
+  delete self->descriptor_options;
+  delete self->message_factory;
+  delete self->database;
+  delete self->pool;
+  Py_TYPE(self)->tp_free(reinterpret_cast<PyObject*>(self));
+}
+
+PyObject* FindMessageByName(PyDescriptorPool* self, PyObject* arg) {
+  Py_ssize_t name_size;
+  char* name;
+  if (PyString_AsStringAndSize(arg, &name, &name_size) < 0) {
+    return NULL;
+  }
+
+  const Descriptor* message_descriptor =
+      self->pool->FindMessageTypeByName(string(name, name_size));
+
+  if (message_descriptor == NULL) {
+    PyErr_Format(PyExc_KeyError, "Couldn't find message %.200s", name);
+    return NULL;
+  }
+
+  return PyMessageDescriptor_FromDescriptor(message_descriptor);
+}
+
+// Add a message class to our database.
+int RegisterMessageClass(PyDescriptorPool* self,
+                         const Descriptor *message_descriptor,
+                         PyObject *message_class) {
+  Py_INCREF(message_class);
+  typedef PyDescriptorPool::ClassesByMessageMap::iterator iterator;
+  std::pair<iterator, bool> ret = self->classes_by_descriptor->insert(
+      std::make_pair(message_descriptor, message_class));
+  if (!ret.second) {
+    // Update case: DECREF the previous value.
+    Py_DECREF(ret.first->second);
+    ret.first->second = message_class;
+  }
+  return 0;
+}
+
+// Retrieve the message class added to our database.
+PyObject *GetMessageClass(PyDescriptorPool* self,
+                          const Descriptor *message_descriptor) {
+  typedef PyDescriptorPool::ClassesByMessageMap::iterator iterator;
+  iterator ret = self->classes_by_descriptor->find(message_descriptor);
+  if (ret == self->classes_by_descriptor->end()) {
+    PyErr_Format(PyExc_TypeError, "No message class registered for '%s'",
+                 message_descriptor->full_name().c_str());
+    return NULL;
+  } else {
+    return ret->second;
+  }
+}
+
+PyObject* FindFileByName(PyDescriptorPool* self, PyObject* arg) {
+  Py_ssize_t name_size;
+  char* name;
+  if (PyString_AsStringAndSize(arg, &name, &name_size) < 0) {
+    return NULL;
+  }
+
+  const FileDescriptor* file_descriptor =
+      self->pool->FindFileByName(string(name, name_size));
+  if (file_descriptor == NULL) {
+    PyErr_Format(PyExc_KeyError, "Couldn't find file %.200s",
+                 name);
+    return NULL;
+  }
+
+  return PyFileDescriptor_FromDescriptor(file_descriptor);
+}
+
+PyObject* FindFieldByName(PyDescriptorPool* self, PyObject* arg) {
+  Py_ssize_t name_size;
+  char* name;
+  if (PyString_AsStringAndSize(arg, &name, &name_size) < 0) {
+    return NULL;
+  }
+
+  const FieldDescriptor* field_descriptor =
+      self->pool->FindFieldByName(string(name, name_size));
+  if (field_descriptor == NULL) {
+    PyErr_Format(PyExc_KeyError, "Couldn't find field %.200s",
+                 name);
+    return NULL;
+  }
+
+  return PyFieldDescriptor_FromDescriptor(field_descriptor);
+}
+
+PyObject* FindExtensionByName(PyDescriptorPool* self, PyObject* arg) {
+  Py_ssize_t name_size;
+  char* name;
+  if (PyString_AsStringAndSize(arg, &name, &name_size) < 0) {
+    return NULL;
+  }
+
+  const FieldDescriptor* field_descriptor =
+      self->pool->FindExtensionByName(string(name, name_size));
+  if (field_descriptor == NULL) {
+    PyErr_Format(PyExc_KeyError, "Couldn't find extension field %.200s", name);
+    return NULL;
+  }
+
+  return PyFieldDescriptor_FromDescriptor(field_descriptor);
+}
+
+PyObject* FindEnumTypeByName(PyDescriptorPool* self, PyObject* arg) {
+  Py_ssize_t name_size;
+  char* name;
+  if (PyString_AsStringAndSize(arg, &name, &name_size) < 0) {
+    return NULL;
+  }
+
+  const EnumDescriptor* enum_descriptor =
+      self->pool->FindEnumTypeByName(string(name, name_size));
+  if (enum_descriptor == NULL) {
+    PyErr_Format(PyExc_KeyError, "Couldn't find enum %.200s", name);
+    return NULL;
+  }
+
+  return PyEnumDescriptor_FromDescriptor(enum_descriptor);
+}
+
+PyObject* FindOneofByName(PyDescriptorPool* self, PyObject* arg) {
+  Py_ssize_t name_size;
+  char* name;
+  if (PyString_AsStringAndSize(arg, &name, &name_size) < 0) {
+    return NULL;
+  }
+
+  const OneofDescriptor* oneof_descriptor =
+      self->pool->FindOneofByName(string(name, name_size));
+  if (oneof_descriptor == NULL) {
+    PyErr_Format(PyExc_KeyError, "Couldn't find oneof %.200s", name);
+    return NULL;
+  }
+
+  return PyOneofDescriptor_FromDescriptor(oneof_descriptor);
+}
+
+PyObject* FindFileContainingSymbol(PyDescriptorPool* self, PyObject* arg) {
+  Py_ssize_t name_size;
+  char* name;
+  if (PyString_AsStringAndSize(arg, &name, &name_size) < 0) {
+    return NULL;
+  }
+
+  const FileDescriptor* file_descriptor =
+      self->pool->FindFileContainingSymbol(string(name, name_size));
+  if (file_descriptor == NULL) {
+    PyErr_Format(PyExc_KeyError, "Couldn't find symbol %.200s", name);
+    return NULL;
+  }
+
+  return PyFileDescriptor_FromDescriptor(file_descriptor);
+}
+
+// These functions should not exist -- the only valid way to create
+// descriptors is to call Add() or AddSerializedFile().
+// But these AddDescriptor() functions were created in Python and some people
+// call them, so we support them for now for compatibility.
+// However we do check that the existing descriptor already exists in the pool,
+// which appears to always be true for existing calls -- but then why do people
+// call a function that will just be a no-op?
+// TODO(amauryfa): Need to investigate further.
+
+PyObject* AddFileDescriptor(PyDescriptorPool* self, PyObject* descriptor) {
+  const FileDescriptor* file_descriptor =
+      PyFileDescriptor_AsDescriptor(descriptor);
+  if (!file_descriptor) {
+    return NULL;
+  }
+  if (file_descriptor !=
+      self->pool->FindFileByName(file_descriptor->name())) {
+    PyErr_Format(PyExc_ValueError,
+                 "The file descriptor %s does not belong to this pool",
+                 file_descriptor->name().c_str());
+    return NULL;
+  }
+  Py_RETURN_NONE;
+}
+
+PyObject* AddDescriptor(PyDescriptorPool* self, PyObject* descriptor) {
+  const Descriptor* message_descriptor =
+      PyMessageDescriptor_AsDescriptor(descriptor);
+  if (!message_descriptor) {
+    return NULL;
+  }
+  if (message_descriptor !=
+      self->pool->FindMessageTypeByName(message_descriptor->full_name())) {
+    PyErr_Format(PyExc_ValueError,
+                 "The message descriptor %s does not belong to this pool",
+                 message_descriptor->full_name().c_str());
+    return NULL;
+  }
+  Py_RETURN_NONE;
+}
+
+PyObject* AddEnumDescriptor(PyDescriptorPool* self, PyObject* descriptor) {
+  const EnumDescriptor* enum_descriptor =
+      PyEnumDescriptor_AsDescriptor(descriptor);
+  if (!enum_descriptor) {
+    return NULL;
+  }
+  if (enum_descriptor !=
+      self->pool->FindEnumTypeByName(enum_descriptor->full_name())) {
+    PyErr_Format(PyExc_ValueError,
+                 "The enum descriptor %s does not belong to this pool",
+                 enum_descriptor->full_name().c_str());
+    return NULL;
+  }
+  Py_RETURN_NONE;
+}
+
+// The code below loads new Descriptors from a serialized FileDescriptorProto.
+
+
+// Collects errors that occur during proto file building to allow them to be
+// propagated in the python exception instead of only living in ERROR logs.
+class BuildFileErrorCollector : public DescriptorPool::ErrorCollector {
+ public:
+  BuildFileErrorCollector() : error_message(""), had_errors(false) {}
+
+  void AddError(const string& filename, const string& element_name,
+                const Message* descriptor, ErrorLocation location,
+                const string& message) {
+    // Replicates the logging behavior that happens in the C++ implementation
+    // when an error collector is not passed in.
+    if (!had_errors) {
+      error_message +=
+          ("Invalid proto descriptor for file \"" + filename + "\":\n");
+      had_errors = true;
+    }
+    // As this only happens on failure and will result in the program not
+    // running at all, no effort is made to optimize this string manipulation.
+    error_message += ("  " + element_name + ": " + message + "\n");
+  }
+
+  string error_message;
+  bool had_errors;
+};
+
+PyObject* AddSerializedFile(PyDescriptorPool* self, PyObject* serialized_pb) {
+  char* message_type;
+  Py_ssize_t message_len;
+
+  if (self->database != NULL) {
+    PyErr_SetString(
+        PyExc_ValueError,
+        "Cannot call Add on a DescriptorPool that uses a DescriptorDatabase. "
+        "Add your file to the underlying database.");
+    return NULL;
+  }
+
+  if (PyBytes_AsStringAndSize(serialized_pb, &message_type, &message_len) < 0) {
+    return NULL;
+  }
+
+  FileDescriptorProto file_proto;
+  if (!file_proto.ParseFromArray(message_type, message_len)) {
+    PyErr_SetString(PyExc_TypeError, "Couldn't parse file content!");
+    return NULL;
+  }
+
+  // If the file was already part of a C++ library, all its descriptors are in
+  // the underlying pool.  No need to do anything else.
+  const FileDescriptor* generated_file = NULL;
+  if (self->underlay) {
+    generated_file = self->underlay->FindFileByName(file_proto.name());
+  }
+  if (generated_file != NULL) {
+    return PyFileDescriptor_FromDescriptorWithSerializedPb(
+        generated_file, serialized_pb);
+  }
+
+  BuildFileErrorCollector error_collector;
+  const FileDescriptor* descriptor =
+      self->pool->BuildFileCollectingErrors(file_proto,
+                                            &error_collector);
+  if (descriptor == NULL) {
+    PyErr_Format(PyExc_TypeError,
+                 "Couldn't build proto file into descriptor pool!\n%s",
+                 error_collector.error_message.c_str());
+    return NULL;
+  }
+
+  return PyFileDescriptor_FromDescriptorWithSerializedPb(
+      descriptor, serialized_pb);
+}
+
+PyObject* Add(PyDescriptorPool* self, PyObject* file_descriptor_proto) {
+  ScopedPyObjectPtr serialized_pb(
+      PyObject_CallMethod(file_descriptor_proto, "SerializeToString", NULL));
+  if (serialized_pb == NULL) {
+    return NULL;
+  }
+  return AddSerializedFile(self, serialized_pb.get());
+}
+
+static PyMethodDef Methods[] = {
+  { "Add", (PyCFunction)Add, METH_O,
+    "Adds the FileDescriptorProto and its types to this pool." },
+  { "AddSerializedFile", (PyCFunction)AddSerializedFile, METH_O,
+    "Adds a serialized FileDescriptorProto to this pool." },
+
+  // TODO(amauryfa): Understand why the Python implementation differs from
+  // this one, ask users to use another API and deprecate these functions.
+  { "AddFileDescriptor", (PyCFunction)AddFileDescriptor, METH_O,
+    "No-op. Add() must have been called before." },
+  { "AddDescriptor", (PyCFunction)AddDescriptor, METH_O,
+    "No-op. Add() must have been called before." },
+  { "AddEnumDescriptor", (PyCFunction)AddEnumDescriptor, METH_O,
+    "No-op. Add() must have been called before." },
+
+  { "FindFileByName", (PyCFunction)FindFileByName, METH_O,
+    "Searches for a file descriptor by its .proto name." },
+  { "FindMessageTypeByName", (PyCFunction)FindMessageByName, METH_O,
+    "Searches for a message descriptor by full name." },
+  { "FindFieldByName", (PyCFunction)FindFieldByName, METH_O,
+    "Searches for a field descriptor by full name." },
+  { "FindExtensionByName", (PyCFunction)FindExtensionByName, METH_O,
+    "Searches for extension descriptor by full name." },
+  { "FindEnumTypeByName", (PyCFunction)FindEnumTypeByName, METH_O,
+    "Searches for enum type descriptor by full name." },
+  { "FindOneofByName", (PyCFunction)FindOneofByName, METH_O,
+    "Searches for oneof descriptor by full name." },
+
+  { "FindFileContainingSymbol", (PyCFunction)FindFileContainingSymbol, METH_O,
+    "Gets the FileDescriptor containing the specified symbol." },
+  {NULL}
+};
+
+}  // namespace cdescriptor_pool
+
+PyTypeObject PyDescriptorPool_Type = {
+  PyVarObject_HEAD_INIT(&PyType_Type, 0)
+  FULL_MODULE_NAME ".DescriptorPool",  // tp_name
+  sizeof(PyDescriptorPool),            // tp_basicsize
+  0,                                   // tp_itemsize
+  (destructor)cdescriptor_pool::Dealloc,  // tp_dealloc
+  0,                                   // tp_print
+  0,                                   // tp_getattr
+  0,                                   // tp_setattr
+  0,                                   // tp_compare
+  0,                                   // tp_repr
+  0,                                   // tp_as_number
+  0,                                   // tp_as_sequence
+  0,                                   // tp_as_mapping
+  0,                                   // tp_hash
+  0,                                   // tp_call
+  0,                                   // tp_str
+  0,                                   // tp_getattro
+  0,                                   // tp_setattro
+  0,                                   // tp_as_buffer
+  Py_TPFLAGS_DEFAULT,                  // tp_flags
+  "A Descriptor Pool",                 // tp_doc
+  0,                                   // tp_traverse
+  0,                                   // tp_clear
+  0,                                   // tp_richcompare
+  0,                                   // tp_weaklistoffset
+  0,                                   // tp_iter
+  0,                                   // tp_iternext
+  cdescriptor_pool::Methods,           // tp_methods
+  0,                                   // tp_members
+  0,                                   // tp_getset
+  0,                                   // tp_base
+  0,                                   // tp_dict
+  0,                                   // tp_descr_get
+  0,                                   // tp_descr_set
+  0,                                   // tp_dictoffset
+  0,                                   // tp_init
+  0,                                   // tp_alloc
+  cdescriptor_pool::New,               // tp_new
+  PyObject_Del,                        // tp_free
+};
+
+// This is the DescriptorPool which contains all the definitions from the
+// generated _pb2.py modules.
+static PyDescriptorPool* python_generated_pool = NULL;
+
+bool InitDescriptorPool() {
+  if (PyType_Ready(&PyDescriptorPool_Type) < 0)
+    return false;
+
+  // The Pool of messages declared in Python libraries.
+  // generated_pool() contains all messages already linked in C++ libraries, and
+  // is used as underlay.
+  python_generated_pool = cdescriptor_pool::PyDescriptorPool_NewWithUnderlay(
+      DescriptorPool::generated_pool());
+  if (python_generated_pool == NULL) {
+    return false;
+  }
+  // Register this pool to be found for C++-generated descriptors.
+  descriptor_pool_map.insert(
+      std::make_pair(DescriptorPool::generated_pool(),
+                     python_generated_pool));
+
+  return true;
+}
+
+// The default DescriptorPool used everywhere in this module.
+// Today it's the python_generated_pool.
+// TODO(amauryfa): Remove all usages of this function: the pool should be
+// derived from the context.
+PyDescriptorPool* GetDefaultDescriptorPool() {
+  return python_generated_pool;
+}
+
+PyDescriptorPool* GetDescriptorPool_FromPool(const DescriptorPool* pool) {
+  // Fast path for standard descriptors.
+  if (pool == python_generated_pool->pool ||
+      pool == DescriptorPool::generated_pool()) {
+    return python_generated_pool;
+  }
+  hash_map<const DescriptorPool*, PyDescriptorPool*>::iterator it =
+      descriptor_pool_map.find(pool);
+  if (it == descriptor_pool_map.end()) {
+    PyErr_SetString(PyExc_KeyError, "Unknown descriptor pool");
+    return NULL;
+  }
+  return it->second;
+}
+
+}  // namespace python
+}  // namespace protobuf
+}  // namespace google
diff --git a/third_party/protobuf/python/google/protobuf/pyext/descriptor_pool.h b/third_party/protobuf/python/google/protobuf/pyext/descriptor_pool.h
new file mode 100644
index 0000000000..16bc910c45
--- /dev/null
+++ b/third_party/protobuf/python/google/protobuf/pyext/descriptor_pool.h
@@ -0,0 +1,164 @@
+// 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.
+
+#ifndef GOOGLE_PROTOBUF_PYTHON_CPP_DESCRIPTOR_POOL_H__
+#define GOOGLE_PROTOBUF_PYTHON_CPP_DESCRIPTOR_POOL_H__
+
+#include <Python.h>
+
+#include <google/protobuf/stubs/hash.h>
+#include <google/protobuf/descriptor.h>
+
+namespace google {
+namespace protobuf {
+class MessageFactory;
+
+namespace python {
+
+// Wraps operations to the global DescriptorPool which contains information
+// about all messages and fields.
+//
+// There is normally one pool per process. We make it a Python object only
+// because it contains many Python references.
+// TODO(amauryfa): See whether such objects can appear in reference cycles, and
+// consider adding support for the cyclic GC.
+//
+// "Methods" that interacts with this DescriptorPool are in the cdescriptor_pool
+// namespace.
+typedef struct PyDescriptorPool {
+  PyObject_HEAD
+
+  // The C++ pool containing Descriptors.
+  DescriptorPool* pool;
+
+  // The C++ pool acting as an underlay. Can be NULL.
+  // This pointer is not owned and must stay alive.
+  const DescriptorPool* underlay;
+
+  // The C++ descriptor database used to fetch unknown protos. Can be NULL.
+  // This pointer is owned.
+  const DescriptorDatabase* database;
+
+  // DynamicMessageFactory used to create C++ instances of messages.
+  // This object cache the descriptors that were used, so the DescriptorPool
+  // needs to get rid of it before it can delete itself.
+  //
+  // Note: A C++ MessageFactory is different from the Python MessageFactory.
+  // The C++ one creates messages, when the Python one creates classes.
+  MessageFactory* message_factory;
+
+  // Make our own mapping to retrieve Python classes from C++ descriptors.
+  //
+  // Descriptor pointers stored here are owned by the DescriptorPool above.
+  // Python references to classes are owned by this PyDescriptorPool.
+  typedef hash_map<const Descriptor*, PyObject*> ClassesByMessageMap;
+  ClassesByMessageMap* classes_by_descriptor;
+
+  // Cache the options for any kind of descriptor.
+  // Descriptor pointers are owned by the DescriptorPool above.
+  // Python objects are owned by the map.
+  hash_map<const void*, PyObject*>* descriptor_options;
+} PyDescriptorPool;
+
+
+extern PyTypeObject PyDescriptorPool_Type;
+
+namespace cdescriptor_pool {
+
+// Looks up a message by name.
+// Returns a message Descriptor, or NULL if not found.
+const Descriptor* FindMessageTypeByName(PyDescriptorPool* self,
+                                        const string& name);
+
+// Registers a new Python class for the given message descriptor.
+// On error, returns -1 with a Python exception set.
+int RegisterMessageClass(PyDescriptorPool* self,
+                         const Descriptor* message_descriptor,
+                         PyObject* message_class);
+
+// Retrieves the Python class registered with the given message descriptor.
+//
+// Returns a *borrowed* reference if found, otherwise returns NULL with an
+// exception set.
+PyObject* GetMessageClass(PyDescriptorPool* self,
+                          const Descriptor* message_descriptor);
+
+// The functions below are also exposed as methods of the DescriptorPool type.
+
+// Looks up a message by name. Returns a PyMessageDescriptor corresponding to
+// the field on success, or NULL on failure.
+//
+// Returns a new reference.
+PyObject* FindMessageByName(PyDescriptorPool* self, PyObject* name);
+
+// Looks up a field by name. Returns a PyFieldDescriptor corresponding to
+// the field on success, or NULL on failure.
+//
+// Returns a new reference.
+PyObject* FindFieldByName(PyDescriptorPool* self, PyObject* name);
+
+// Looks up an extension by name. Returns a PyFieldDescriptor corresponding
+// to the field on success, or NULL on failure.
+//
+// Returns a new reference.
+PyObject* FindExtensionByName(PyDescriptorPool* self, PyObject* arg);
+
+// Looks up an enum type by name. Returns a PyEnumDescriptor corresponding
+// to the field on success, or NULL on failure.
+//
+// Returns a new reference.
+PyObject* FindEnumTypeByName(PyDescriptorPool* self, PyObject* arg);
+
+// Looks up a oneof by name. Returns a COneofDescriptor corresponding
+// to the oneof on success, or NULL on failure.
+//
+// Returns a new reference.
+PyObject* FindOneofByName(PyDescriptorPool* self, PyObject* arg);
+
+}  // namespace cdescriptor_pool
+
+// Retrieve the global descriptor pool owned by the _message module.
+// This is the one used by pb2.py generated modules.
+// Returns a *borrowed* reference.
+// "Default" pool used to register messages from _pb2.py modules.
+PyDescriptorPool* GetDefaultDescriptorPool();
+
+// Retrieve the python descriptor pool owning a C++ descriptor pool.
+// Returns a *borrowed* reference.
+PyDescriptorPool* GetDescriptorPool_FromPool(const DescriptorPool* pool);
+
+// Initialize objects used by this module.
+bool InitDescriptorPool();
+
+}  // namespace python
+}  // namespace protobuf
+
+}  // namespace google
+#endif  // GOOGLE_PROTOBUF_PYTHON_CPP_DESCRIPTOR_POOL_H__
diff --git a/third_party/protobuf/python/google/protobuf/pyext/extension_dict.cc b/third_party/protobuf/python/google/protobuf/pyext/extension_dict.cc
new file mode 100644
index 0000000000..555bd2937b
--- /dev/null
+++ b/third_party/protobuf/python/google/protobuf/pyext/extension_dict.cc
@@ -0,0 +1,320 @@
+// 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: anuraag@google.com (Anuraag Agrawal)
+// Author: tibell@google.com (Johan Tibell)
+
+#include <google/protobuf/pyext/extension_dict.h>
+
+#include <google/protobuf/stubs/logging.h>
+#include <google/protobuf/stubs/common.h>
+#include <google/protobuf/descriptor.h>
+#include <google/protobuf/dynamic_message.h>
+#include <google/protobuf/message.h>
+#include <google/protobuf/pyext/descriptor.h>
+#include <google/protobuf/pyext/descriptor_pool.h>
+#include <google/protobuf/pyext/message.h>
+#include <google/protobuf/pyext/repeated_composite_container.h>
+#include <google/protobuf/pyext/repeated_scalar_container.h>
+#include <google/protobuf/pyext/scoped_pyobject_ptr.h>
+#include <google/protobuf/stubs/shared_ptr.h>
+
+namespace google {
+namespace protobuf {
+namespace python {
+
+namespace extension_dict {
+
+PyObject* len(ExtensionDict* self) {
+#if PY_MAJOR_VERSION >= 3
+  return PyLong_FromLong(PyDict_Size(self->values));
+#else
+  return PyInt_FromLong(PyDict_Size(self->values));
+#endif
+}
+
+// TODO(tibell): Use VisitCompositeField.
+int ReleaseExtension(ExtensionDict* self,
+                     PyObject* extension,
+                     const FieldDescriptor* descriptor) {
+  if (descriptor->label() == FieldDescriptor::LABEL_REPEATED) {
+    if (descriptor->cpp_type() == FieldDescriptor::CPPTYPE_MESSAGE) {
+      if (repeated_composite_container::Release(
+              reinterpret_cast<RepeatedCompositeContainer*>(
+                  extension)) < 0) {
+        return -1;
+      }
+    } else {
+      if (repeated_scalar_container::Release(
+              reinterpret_cast<RepeatedScalarContainer*>(
+                  extension)) < 0) {
+        return -1;
+      }
+    }
+  } else if (descriptor->cpp_type() == FieldDescriptor::CPPTYPE_MESSAGE) {
+    if (cmessage::ReleaseSubMessage(
+            self->parent, descriptor,
+            reinterpret_cast<CMessage*>(extension)) < 0) {
+      return -1;
+    }
+  }
+
+  return 0;
+}
+
+PyObject* subscript(ExtensionDict* self, PyObject* key) {
+  const FieldDescriptor* descriptor = cmessage::GetExtensionDescriptor(key);
+  if (descriptor == NULL) {
+    return NULL;
+  }
+  if (!CheckFieldBelongsToMessage(descriptor, self->message)) {
+    return NULL;
+  }
+
+  if (descriptor->label() != FieldDescriptor::LABEL_REPEATED &&
+      descriptor->cpp_type() != FieldDescriptor::CPPTYPE_MESSAGE) {
+    return cmessage::InternalGetScalar(self->message, descriptor);
+  }
+
+  PyObject* value = PyDict_GetItem(self->values, key);
+  if (value != NULL) {
+    Py_INCREF(value);
+    return value;
+  }
+
+  if (self->parent == NULL) {
+    // We are in "detached" state. Don't allow further modifications.
+    // TODO(amauryfa): Support adding non-scalars to a detached extension dict.
+    // This probably requires to store the type of the main message.
+    PyErr_SetObject(PyExc_KeyError, key);
+    return NULL;
+  }
+
+  if (descriptor->label() != FieldDescriptor::LABEL_REPEATED &&
+      descriptor->cpp_type() == FieldDescriptor::CPPTYPE_MESSAGE) {
+    PyObject* sub_message = cmessage::InternalGetSubMessage(
+        self->parent, descriptor);
+    if (sub_message == NULL) {
+      return NULL;
+    }
+    PyDict_SetItem(self->values, key, sub_message);
+    return sub_message;
+  }
+
+  if (descriptor->label() == FieldDescriptor::LABEL_REPEATED) {
+    if (descriptor->cpp_type() == FieldDescriptor::CPPTYPE_MESSAGE) {
+      PyObject *message_class = cdescriptor_pool::GetMessageClass(
+          cmessage::GetDescriptorPoolForMessage(self->parent),
+          descriptor->message_type());
+      if (message_class == NULL) {
+        return NULL;
+      }
+      PyObject* py_container = repeated_composite_container::NewContainer(
+          self->parent, descriptor, message_class);
+      if (py_container == NULL) {
+        return NULL;
+      }
+      PyDict_SetItem(self->values, key, py_container);
+      return py_container;
+    } else {
+      PyObject* py_container = repeated_scalar_container::NewContainer(
+          self->parent, descriptor);
+      if (py_container == NULL) {
+        return NULL;
+      }
+      PyDict_SetItem(self->values, key, py_container);
+      return py_container;
+    }
+  }
+  PyErr_SetString(PyExc_ValueError, "control reached unexpected line");
+  return NULL;
+}
+
+int ass_subscript(ExtensionDict* self, PyObject* key, PyObject* value) {
+  const FieldDescriptor* descriptor = cmessage::GetExtensionDescriptor(key);
+  if (descriptor == NULL) {
+    return -1;
+  }
+  if (!CheckFieldBelongsToMessage(descriptor, self->message)) {
+    return -1;
+  }
+
+  if (descriptor->label() != FieldDescriptor::LABEL_OPTIONAL ||
+      descriptor->cpp_type() == FieldDescriptor::CPPTYPE_MESSAGE) {
+    PyErr_SetString(PyExc_TypeError, "Extension is repeated and/or composite "
+                    "type");
+    return -1;
+  }
+  if (self->parent) {
+    cmessage::AssureWritable(self->parent);
+    if (cmessage::InternalSetScalar(self->parent, descriptor, value) < 0) {
+      return -1;
+    }
+  }
+  // TODO(tibell): We shouldn't write scalars to the cache.
+  PyDict_SetItem(self->values, key, value);
+  return 0;
+}
+
+PyObject* ClearExtension(ExtensionDict* self, PyObject* extension) {
+  const FieldDescriptor* descriptor =
+      cmessage::GetExtensionDescriptor(extension);
+  if (descriptor == NULL) {
+    return NULL;
+  }
+  PyObject* value = PyDict_GetItem(self->values, extension);
+  if (self->parent) {
+    if (value != NULL) {
+      if (ReleaseExtension(self, value, descriptor) < 0) {
+        return NULL;
+      }
+    }
+    if (ScopedPyObjectPtr(cmessage::ClearFieldByDescriptor(
+            self->parent, descriptor)) == NULL) {
+      return NULL;
+    }
+  }
+  if (PyDict_DelItem(self->values, extension) < 0) {
+    PyErr_Clear();
+  }
+  Py_RETURN_NONE;
+}
+
+PyObject* HasExtension(ExtensionDict* self, PyObject* extension) {
+  const FieldDescriptor* descriptor =
+      cmessage::GetExtensionDescriptor(extension);
+  if (descriptor == NULL) {
+    return NULL;
+  }
+  if (self->parent) {
+    return cmessage::HasFieldByDescriptor(self->parent, descriptor);
+  } else {
+    int exists = PyDict_Contains(self->values, extension);
+    if (exists < 0) {
+      return NULL;
+    }
+    return PyBool_FromLong(exists);
+  }
+}
+
+PyObject* _FindExtensionByName(ExtensionDict* self, PyObject* name) {
+  ScopedPyObjectPtr extensions_by_name(PyObject_GetAttrString(
+      reinterpret_cast<PyObject*>(self->parent), "_extensions_by_name"));
+  if (extensions_by_name == NULL) {
+    return NULL;
+  }
+  PyObject* result = PyDict_GetItem(extensions_by_name.get(), name);
+  if (result == NULL) {
+    Py_RETURN_NONE;
+  } else {
+    Py_INCREF(result);
+    return result;
+  }
+}
+
+ExtensionDict* NewExtensionDict(CMessage *parent) {
+  ExtensionDict* self = reinterpret_cast<ExtensionDict*>(
+      PyType_GenericAlloc(&ExtensionDict_Type, 0));
+  if (self == NULL) {
+    return NULL;
+  }
+
+  self->parent = parent;  // Store a borrowed reference.
+  self->message = parent->message;
+  self->owner = parent->owner;
+  self->values = PyDict_New();
+  return self;
+}
+
+void dealloc(ExtensionDict* self) {
+  Py_CLEAR(self->values);
+  self->owner.reset();
+  Py_TYPE(self)->tp_free(reinterpret_cast<PyObject*>(self));
+}
+
+static PyMappingMethods MpMethods = {
+  (lenfunc)len,               /* mp_length */
+  (binaryfunc)subscript,      /* mp_subscript */
+  (objobjargproc)ass_subscript,/* mp_ass_subscript */
+};
+
+#define EDMETHOD(name, args, doc) { #name, (PyCFunction)name, args, doc }
+static PyMethodDef Methods[] = {
+  EDMETHOD(ClearExtension, METH_O, "Clears an extension from the object."),
+  EDMETHOD(HasExtension, METH_O, "Checks if the object has an extension."),
+  EDMETHOD(_FindExtensionByName, METH_O,
+           "Finds an extension by name."),
+  { NULL, NULL }
+};
+
+}  // namespace extension_dict
+
+PyTypeObject ExtensionDict_Type = {
+  PyVarObject_HEAD_INIT(&PyType_Type, 0)
+  FULL_MODULE_NAME ".ExtensionDict",   // tp_name
+  sizeof(ExtensionDict),               // tp_basicsize
+  0,                                   //  tp_itemsize
+  (destructor)extension_dict::dealloc,  //  tp_dealloc
+  0,                                   //  tp_print
+  0,                                   //  tp_getattr
+  0,                                   //  tp_setattr
+  0,                                   //  tp_compare
+  0,                                   //  tp_repr
+  0,                                   //  tp_as_number
+  0,                                   //  tp_as_sequence
+  &extension_dict::MpMethods,          //  tp_as_mapping
+  PyObject_HashNotImplemented,         //  tp_hash
+  0,                                   //  tp_call
+  0,                                   //  tp_str
+  0,                                   //  tp_getattro
+  0,                                   //  tp_setattro
+  0,                                   //  tp_as_buffer
+  Py_TPFLAGS_DEFAULT,                  //  tp_flags
+  "An extension dict",                 //  tp_doc
+  0,                                   //  tp_traverse
+  0,                                   //  tp_clear
+  0,                                   //  tp_richcompare
+  0,                                   //  tp_weaklistoffset
+  0,                                   //  tp_iter
+  0,                                   //  tp_iternext
+  extension_dict::Methods,             //  tp_methods
+  0,                                   //  tp_members
+  0,                                   //  tp_getset
+  0,                                   //  tp_base
+  0,                                   //  tp_dict
+  0,                                   //  tp_descr_get
+  0,                                   //  tp_descr_set
+  0,                                   //  tp_dictoffset
+  0,                                   //  tp_init
+};
+
+}  // namespace python
+}  // namespace protobuf
+}  // namespace google
diff --git a/third_party/protobuf/python/google/protobuf/pyext/extension_dict.h b/third_party/protobuf/python/google/protobuf/pyext/extension_dict.h
new file mode 100644
index 0000000000..d92cf956bd
--- /dev/null
+++ b/third_party/protobuf/python/google/protobuf/pyext/extension_dict.h
@@ -0,0 +1,131 @@
+// 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: anuraag@google.com (Anuraag Agrawal)
+// Author: tibell@google.com (Johan Tibell)
+
+#ifndef GOOGLE_PROTOBUF_PYTHON_CPP_EXTENSION_DICT_H__
+#define GOOGLE_PROTOBUF_PYTHON_CPP_EXTENSION_DICT_H__
+
+#include <Python.h>
+
+#include <memory>
+#ifndef _SHARED_PTR_H
+#include <google/protobuf/stubs/shared_ptr.h>
+#endif
+
+namespace google {
+namespace protobuf {
+
+class Message;
+class FieldDescriptor;
+
+#ifdef _SHARED_PTR_H
+using std::shared_ptr;
+#else
+using internal::shared_ptr;
+#endif
+
+namespace python {
+
+struct CMessage;
+
+typedef struct ExtensionDict {
+  PyObject_HEAD;
+
+  // This is the top-level C++ Message object that owns the whole
+  // proto tree.  Every Python container class holds a
+  // reference to it in order to keep it alive as long as there's a
+  // Python object that references any part of the tree.
+  shared_ptr<Message> owner;
+
+  // Weak reference to parent message. Used to make sure
+  // the parent is writable when an extension field is modified.
+  CMessage* parent;
+
+  // Pointer to the C++ Message that this ExtensionDict extends.
+  // Not owned by us.
+  Message* message;
+
+  // A dict of child messages, indexed by Extension descriptors.
+  // Similar to CMessage::composite_fields.
+  PyObject* values;
+} ExtensionDict;
+
+extern PyTypeObject ExtensionDict_Type;
+
+namespace extension_dict {
+
+// Builds an Extensions dict for a specific message.
+ExtensionDict* NewExtensionDict(CMessage *parent);
+
+// Gets the number of extension values in this ExtensionDict as a python object.
+//
+// Returns a new reference.
+PyObject* len(ExtensionDict* self);
+
+// Releases extensions referenced outside this dictionary to keep outside
+// references alive.
+//
+// Returns 0 on success, -1 on failure.
+int ReleaseExtension(ExtensionDict* self,
+                     PyObject* extension,
+                     const FieldDescriptor* descriptor);
+
+// Gets an extension from the dict for the given extension descriptor.
+//
+// Returns a new reference.
+PyObject* subscript(ExtensionDict* self, PyObject* key);
+
+// Assigns a value to an extension in the dict. Can only be used for singular
+// simple types.
+//
+// Returns 0 on success, -1 on failure.
+int ass_subscript(ExtensionDict* self, PyObject* key, PyObject* value);
+
+// Clears an extension from the dict. Will release the extension if there
+// is still an external reference left to it.
+//
+// Returns None on success.
+PyObject* ClearExtension(ExtensionDict* self,
+                                       PyObject* extension);
+
+// Gets an extension from the dict given the extension name as opposed to
+// descriptor.
+//
+// Returns a new reference.
+PyObject* _FindExtensionByName(ExtensionDict* self, PyObject* name);
+
+}  // namespace extension_dict
+}  // namespace python
+}  // namespace protobuf
+
+}  // namespace google
+#endif  // GOOGLE_PROTOBUF_PYTHON_CPP_EXTENSION_DICT_H__
diff --git a/third_party/protobuf/python/google/protobuf/pyext/map_container.cc b/third_party/protobuf/python/google/protobuf/pyext/map_container.cc
new file mode 100644
index 0000000000..df9138a4d8
--- /dev/null
+++ b/third_party/protobuf/python/google/protobuf/pyext/map_container.cc
@@ -0,0 +1,965 @@
+// 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: haberman@google.com (Josh Haberman)
+
+#include <google/protobuf/pyext/map_container.h>
+
+#include <google/protobuf/stubs/logging.h>
+#include <google/protobuf/stubs/common.h>
+#include <google/protobuf/stubs/scoped_ptr.h>
+#include <google/protobuf/map_field.h>
+#include <google/protobuf/map.h>
+#include <google/protobuf/message.h>
+#include <google/protobuf/pyext/message.h>
+#include <google/protobuf/pyext/scoped_pyobject_ptr.h>
+
+#if PY_MAJOR_VERSION >= 3
+  #define PyInt_FromLong PyLong_FromLong
+  #define PyInt_FromSize_t PyLong_FromSize_t
+#endif
+
+namespace google {
+namespace protobuf {
+namespace python {
+
+// Functions that need access to map reflection functionality.
+// They need to be contained in this class because it is friended.
+class MapReflectionFriend {
+ public:
+  // Methods that are in common between the map types.
+  static PyObject* Contains(PyObject* _self, PyObject* key);
+  static Py_ssize_t Length(PyObject* _self);
+  static PyObject* GetIterator(PyObject *_self);
+  static PyObject* IterNext(PyObject* _self);
+
+  // Methods that differ between the map types.
+  static PyObject* ScalarMapGetItem(PyObject* _self, PyObject* key);
+  static PyObject* MessageMapGetItem(PyObject* _self, PyObject* key);
+  static int ScalarMapSetItem(PyObject* _self, PyObject* key, PyObject* v);
+  static int MessageMapSetItem(PyObject* _self, PyObject* key, PyObject* v);
+};
+
+struct MapIterator {
+  PyObject_HEAD;
+
+  scoped_ptr< ::google::protobuf::MapIterator> iter;
+
+  // A pointer back to the container, so we can notice changes to the version.
+  // We own a ref on this.
+  MapContainer* container;
+
+  // We need to keep a ref on the Message* too, because
+  // MapIterator::~MapIterator() accesses it.  Normally this would be ok because
+  // the ref on container (above) would guarantee outlive semantics.  However in
+  // the case of ClearField(), InitializeAndCopyToParentContainer() resets the
+  // message pointer (and the owner) to a different message, a copy of the
+  // original.  But our iterator still points to the original, which could now
+  // get deleted before us.
+  //
+  // To prevent this, we ensure that the Message will always stay alive as long
+  // as this iterator does.  This is solely for the benefit of the MapIterator
+  // destructor -- we should never actually access the iterator in this state
+  // except to delete it.
+  shared_ptr<Message> owner;
+
+  // The version of the map when we took the iterator to it.
+  //
+  // We store this so that if the map is modified during iteration we can throw
+  // an error.
+  uint64 version;
+
+  // True if the container is empty.  We signal this separately to avoid calling
+  // any of the iteration methods, which are non-const.
+  bool empty;
+};
+
+Message* MapContainer::GetMutableMessage() {
+  cmessage::AssureWritable(parent);
+  return const_cast<Message*>(message);
+}
+
+// Consumes a reference on the Python string object.
+static bool PyStringToSTL(PyObject* py_string, string* stl_string) {
+  char *value;
+  Py_ssize_t value_len;
+
+  if (!py_string) {
+    return false;
+  }
+  if (PyBytes_AsStringAndSize(py_string, &value, &value_len) < 0) {
+    Py_DECREF(py_string);
+    return false;
+  } else {
+    stl_string->assign(value, value_len);
+    Py_DECREF(py_string);
+    return true;
+  }
+}
+
+static bool PythonToMapKey(PyObject* obj,
+                           const FieldDescriptor* field_descriptor,
+                           MapKey* key) {
+  switch (field_descriptor->cpp_type()) {
+    case FieldDescriptor::CPPTYPE_INT32: {
+      GOOGLE_CHECK_GET_INT32(obj, value, false);
+      key->SetInt32Value(value);
+      break;
+    }
+    case FieldDescriptor::CPPTYPE_INT64: {
+      GOOGLE_CHECK_GET_INT64(obj, value, false);
+      key->SetInt64Value(value);
+      break;
+    }
+    case FieldDescriptor::CPPTYPE_UINT32: {
+      GOOGLE_CHECK_GET_UINT32(obj, value, false);
+      key->SetUInt32Value(value);
+      break;
+    }
+    case FieldDescriptor::CPPTYPE_UINT64: {
+      GOOGLE_CHECK_GET_UINT64(obj, value, false);
+      key->SetUInt64Value(value);
+      break;
+    }
+    case FieldDescriptor::CPPTYPE_BOOL: {
+      GOOGLE_CHECK_GET_BOOL(obj, value, false);
+      key->SetBoolValue(value);
+      break;
+    }
+    case FieldDescriptor::CPPTYPE_STRING: {
+      string str;
+      if (!PyStringToSTL(CheckString(obj, field_descriptor), &str)) {
+        return false;
+      }
+      key->SetStringValue(str);
+      break;
+    }
+    default:
+      PyErr_Format(
+          PyExc_SystemError, "Type %d cannot be a map key",
+          field_descriptor->cpp_type());
+      return false;
+  }
+  return true;
+}
+
+static PyObject* MapKeyToPython(const FieldDescriptor* field_descriptor,
+                                const MapKey& key) {
+  switch (field_descriptor->cpp_type()) {
+    case FieldDescriptor::CPPTYPE_INT32:
+      return PyInt_FromLong(key.GetInt32Value());
+    case FieldDescriptor::CPPTYPE_INT64:
+      return PyLong_FromLongLong(key.GetInt64Value());
+    case FieldDescriptor::CPPTYPE_UINT32:
+      return PyInt_FromSize_t(key.GetUInt32Value());
+    case FieldDescriptor::CPPTYPE_UINT64:
+      return PyLong_FromUnsignedLongLong(key.GetUInt64Value());
+    case FieldDescriptor::CPPTYPE_BOOL:
+      return PyBool_FromLong(key.GetBoolValue());
+    case FieldDescriptor::CPPTYPE_STRING:
+      return ToStringObject(field_descriptor, key.GetStringValue());
+    default:
+      PyErr_Format(
+          PyExc_SystemError, "Couldn't convert type %d to value",
+          field_descriptor->cpp_type());
+      return NULL;
+  }
+}
+
+// This is only used for ScalarMap, so we don't need to handle the
+// CPPTYPE_MESSAGE case.
+PyObject* MapValueRefToPython(const FieldDescriptor* field_descriptor,
+                              MapValueRef* value) {
+  switch (field_descriptor->cpp_type()) {
+    case FieldDescriptor::CPPTYPE_INT32:
+      return PyInt_FromLong(value->GetInt32Value());
+    case FieldDescriptor::CPPTYPE_INT64:
+      return PyLong_FromLongLong(value->GetInt64Value());
+    case FieldDescriptor::CPPTYPE_UINT32:
+      return PyInt_FromSize_t(value->GetUInt32Value());
+    case FieldDescriptor::CPPTYPE_UINT64:
+      return PyLong_FromUnsignedLongLong(value->GetUInt64Value());
+    case FieldDescriptor::CPPTYPE_FLOAT:
+      return PyFloat_FromDouble(value->GetFloatValue());
+    case FieldDescriptor::CPPTYPE_DOUBLE:
+      return PyFloat_FromDouble(value->GetDoubleValue());
+    case FieldDescriptor::CPPTYPE_BOOL:
+      return PyBool_FromLong(value->GetBoolValue());
+    case FieldDescriptor::CPPTYPE_STRING:
+      return ToStringObject(field_descriptor, value->GetStringValue());
+    case FieldDescriptor::CPPTYPE_ENUM:
+      return PyInt_FromLong(value->GetEnumValue());
+    default:
+      PyErr_Format(
+          PyExc_SystemError, "Couldn't convert type %d to value",
+          field_descriptor->cpp_type());
+      return NULL;
+  }
+}
+
+// This is only used for ScalarMap, so we don't need to handle the
+// CPPTYPE_MESSAGE case.
+static bool PythonToMapValueRef(PyObject* obj,
+                                const FieldDescriptor* field_descriptor,
+                                bool allow_unknown_enum_values,
+                                MapValueRef* value_ref) {
+  switch (field_descriptor->cpp_type()) {
+    case FieldDescriptor::CPPTYPE_INT32: {
+      GOOGLE_CHECK_GET_INT32(obj, value, false);
+      value_ref->SetInt32Value(value);
+      return true;
+    }
+    case FieldDescriptor::CPPTYPE_INT64: {
+      GOOGLE_CHECK_GET_INT64(obj, value, false);
+      value_ref->SetInt64Value(value);
+      return true;
+    }
+    case FieldDescriptor::CPPTYPE_UINT32: {
+      GOOGLE_CHECK_GET_UINT32(obj, value, false);
+      value_ref->SetUInt32Value(value);
+      return true;
+    }
+    case FieldDescriptor::CPPTYPE_UINT64: {
+      GOOGLE_CHECK_GET_UINT64(obj, value, false);
+      value_ref->SetUInt64Value(value);
+      return true;
+    }
+    case FieldDescriptor::CPPTYPE_FLOAT: {
+      GOOGLE_CHECK_GET_FLOAT(obj, value, false);
+      value_ref->SetFloatValue(value);
+      return true;
+    }
+    case FieldDescriptor::CPPTYPE_DOUBLE: {
+      GOOGLE_CHECK_GET_DOUBLE(obj, value, false);
+      value_ref->SetDoubleValue(value);
+      return true;
+    }
+    case FieldDescriptor::CPPTYPE_BOOL: {
+      GOOGLE_CHECK_GET_BOOL(obj, value, false);
+      value_ref->SetBoolValue(value);
+      return true;;
+    }
+    case FieldDescriptor::CPPTYPE_STRING: {
+      string str;
+      if (!PyStringToSTL(CheckString(obj, field_descriptor), &str)) {
+        return false;
+      }
+      value_ref->SetStringValue(str);
+      return true;
+    }
+    case FieldDescriptor::CPPTYPE_ENUM: {
+      GOOGLE_CHECK_GET_INT32(obj, value, false);
+      if (allow_unknown_enum_values) {
+        value_ref->SetEnumValue(value);
+        return true;
+      } else {
+        const EnumDescriptor* enum_descriptor = field_descriptor->enum_type();
+        const EnumValueDescriptor* enum_value =
+            enum_descriptor->FindValueByNumber(value);
+        if (enum_value != NULL) {
+          value_ref->SetEnumValue(value);
+          return true;
+        } else {
+          PyErr_Format(PyExc_ValueError, "Unknown enum value: %d", value);
+          return false;
+        }
+      }
+      break;
+    }
+    default:
+      PyErr_Format(
+          PyExc_SystemError, "Setting value to a field of unknown type %d",
+          field_descriptor->cpp_type());
+      return false;
+  }
+}
+
+// Map methods common to ScalarMap and MessageMap //////////////////////////////
+
+static MapContainer* GetMap(PyObject* obj) {
+  return reinterpret_cast<MapContainer*>(obj);
+}
+
+Py_ssize_t MapReflectionFriend::Length(PyObject* _self) {
+  MapContainer* self = GetMap(_self);
+  const google::protobuf::Message* message = self->message;
+  return message->GetReflection()->MapSize(*message,
+                                           self->parent_field_descriptor);
+}
+
+PyObject* Clear(PyObject* _self) {
+  MapContainer* self = GetMap(_self);
+  Message* message = self->GetMutableMessage();
+  const Reflection* reflection = message->GetReflection();
+
+  reflection->ClearField(message, self->parent_field_descriptor);
+
+  Py_RETURN_NONE;
+}
+
+PyObject* MapReflectionFriend::Contains(PyObject* _self, PyObject* key) {
+  MapContainer* self = GetMap(_self);
+
+  const Message* message = self->message;
+  const Reflection* reflection = message->GetReflection();
+  MapKey map_key;
+
+  if (!PythonToMapKey(key, self->key_field_descriptor, &map_key)) {
+    return NULL;
+  }
+
+  if (reflection->ContainsMapKey(*message, self->parent_field_descriptor,
+                                 map_key)) {
+    Py_RETURN_TRUE;
+  } else {
+    Py_RETURN_FALSE;
+  }
+}
+
+// Initializes the underlying Message object of "to" so it becomes a new parent
+// repeated scalar, and copies all the values from "from" to it. A child scalar
+// container can be released by passing it as both from and to (e.g. making it
+// the recipient of the new parent message and copying the values from itself).
+static int InitializeAndCopyToParentContainer(MapContainer* from,
+                                              MapContainer* to) {
+  // For now we require from == to, re-evaluate if we want to support deep copy
+  // as in repeated_scalar_container.cc.
+  GOOGLE_DCHECK(from == to);
+  Message* new_message = from->message->New();
+
+  if (MapReflectionFriend::Length(reinterpret_cast<PyObject*>(from)) > 0) {
+    // A somewhat roundabout way of copying just one field from old_message to
+    // new_message.  This is the best we can do with what Reflection gives us.
+    Message* mutable_old = from->GetMutableMessage();
+    vector<const FieldDescriptor*> fields;
+    fields.push_back(from->parent_field_descriptor);
+
+    // Move the map field into the new message.
+    mutable_old->GetReflection()->SwapFields(mutable_old, new_message, fields);
+
+    // If/when we support from != to, this will be required also to copy the
+    // map field back into the existing message:
+    // mutable_old->MergeFrom(*new_message);
+  }
+
+  // If from == to this could delete old_message.
+  to->owner.reset(new_message);
+
+  to->parent = NULL;
+  to->parent_field_descriptor = from->parent_field_descriptor;
+  to->message = new_message;
+
+  // Invalidate iterators, since they point to the old copy of the field.
+  to->version++;
+
+  return 0;
+}
+
+int MapContainer::Release() {
+  return InitializeAndCopyToParentContainer(this, this);
+}
+
+
+// ScalarMap ///////////////////////////////////////////////////////////////////
+
+PyObject *NewScalarMapContainer(
+    CMessage* parent, const google::protobuf::FieldDescriptor* parent_field_descriptor) {
+  if (!CheckFieldBelongsToMessage(parent_field_descriptor, parent->message)) {
+    return NULL;
+  }
+
+#if PY_MAJOR_VERSION >= 3
+  ScopedPyObjectPtr obj(PyType_GenericAlloc(
+        reinterpret_cast<PyTypeObject *>(ScalarMapContainer_Type), 0));
+#else
+  ScopedPyObjectPtr obj(PyType_GenericAlloc(&ScalarMapContainer_Type, 0));
+#endif
+  if (obj.get() == NULL) {
+    return PyErr_Format(PyExc_RuntimeError,
+                        "Could not allocate new container.");
+  }
+
+  MapContainer* self = GetMap(obj.get());
+
+  self->message = parent->message;
+  self->parent = parent;
+  self->parent_field_descriptor = parent_field_descriptor;
+  self->owner = parent->owner;
+  self->version = 0;
+
+  self->key_field_descriptor =
+      parent_field_descriptor->message_type()->FindFieldByName("key");
+  self->value_field_descriptor =
+      parent_field_descriptor->message_type()->FindFieldByName("value");
+
+  if (self->key_field_descriptor == NULL ||
+      self->value_field_descriptor == NULL) {
+    return PyErr_Format(PyExc_KeyError,
+                        "Map entry descriptor did not have key/value fields");
+  }
+
+  return obj.release();
+}
+
+PyObject* MapReflectionFriend::ScalarMapGetItem(PyObject* _self,
+                                                PyObject* key) {
+  MapContainer* self = GetMap(_self);
+
+  Message* message = self->GetMutableMessage();
+  const Reflection* reflection = message->GetReflection();
+  MapKey map_key;
+  MapValueRef value;
+
+  if (!PythonToMapKey(key, self->key_field_descriptor, &map_key)) {
+    return NULL;
+  }
+
+  if (reflection->InsertOrLookupMapValue(message, self->parent_field_descriptor,
+                                         map_key, &value)) {
+    self->version++;
+  }
+
+  return MapValueRefToPython(self->value_field_descriptor, &value);
+}
+
+int MapReflectionFriend::ScalarMapSetItem(PyObject* _self, PyObject* key,
+                                          PyObject* v) {
+  MapContainer* self = GetMap(_self);
+
+  Message* message = self->GetMutableMessage();
+  const Reflection* reflection = message->GetReflection();
+  MapKey map_key;
+  MapValueRef value;
+
+  if (!PythonToMapKey(key, self->key_field_descriptor, &map_key)) {
+    return -1;
+  }
+
+  self->version++;
+
+  if (v) {
+    // Set item to v.
+    reflection->InsertOrLookupMapValue(message, self->parent_field_descriptor,
+                                       map_key, &value);
+
+    return PythonToMapValueRef(v, self->value_field_descriptor,
+                               reflection->SupportsUnknownEnumValues(), &value)
+               ? 0
+               : -1;
+  } else {
+    // Delete key from map.
+    if (reflection->DeleteMapValue(message, self->parent_field_descriptor,
+                                   map_key)) {
+      return 0;
+    } else {
+      PyErr_Format(PyExc_KeyError, "Key not present in map");
+      return -1;
+    }
+  }
+}
+
+static PyObject* ScalarMapGet(PyObject* self, PyObject* args) {
+  PyObject* key;
+  PyObject* default_value = NULL;
+  if (PyArg_ParseTuple(args, "O|O", &key, &default_value) < 0) {
+    return NULL;
+  }
+
+  ScopedPyObjectPtr is_present(MapReflectionFriend::Contains(self, key));
+  if (is_present.get() == NULL) {
+    return NULL;
+  }
+
+  if (PyObject_IsTrue(is_present.get())) {
+    return MapReflectionFriend::ScalarMapGetItem(self, key);
+  } else {
+    if (default_value != NULL) {
+      Py_INCREF(default_value);
+      return default_value;
+    } else {
+      Py_RETURN_NONE;
+    }
+  }
+}
+
+static void ScalarMapDealloc(PyObject* _self) {
+  MapContainer* self = GetMap(_self);
+  self->owner.reset();
+  Py_TYPE(_self)->tp_free(_self);
+}
+
+static PyMethodDef ScalarMapMethods[] = {
+  { "__contains__", MapReflectionFriend::Contains, METH_O,
+    "Tests whether a key is a member of the map." },
+  { "clear", (PyCFunction)Clear, METH_NOARGS,
+    "Removes all elements from the map." },
+  { "get", ScalarMapGet, METH_VARARGS,
+    "Gets the value for the given key if present, or otherwise a default" },
+  /*
+  { "__deepcopy__", (PyCFunction)DeepCopy, METH_VARARGS,
+    "Makes a deep copy of the class." },
+  { "__reduce__", (PyCFunction)Reduce, METH_NOARGS,
+    "Outputs picklable representation of the repeated field." },
+  */
+  {NULL, NULL},
+};
+
+#if PY_MAJOR_VERSION >= 3
+  static PyType_Slot ScalarMapContainer_Type_slots[] = {
+      {Py_tp_dealloc, (void *)ScalarMapDealloc},
+      {Py_mp_length, (void *)MapReflectionFriend::Length},
+      {Py_mp_subscript, (void *)MapReflectionFriend::ScalarMapGetItem},
+      {Py_mp_ass_subscript, (void *)MapReflectionFriend::ScalarMapSetItem},
+      {Py_tp_methods, (void *)ScalarMapMethods},
+      {Py_tp_iter, (void *)MapReflectionFriend::GetIterator},
+      {0, 0},
+  };
+
+  PyType_Spec ScalarMapContainer_Type_spec = {
+      FULL_MODULE_NAME ".ScalarMapContainer",
+      sizeof(MapContainer),
+      0,
+      Py_TPFLAGS_DEFAULT,
+      ScalarMapContainer_Type_slots
+  };
+  PyObject *ScalarMapContainer_Type;
+#else
+  static PyMappingMethods ScalarMapMappingMethods = {
+    MapReflectionFriend::Length,             // mp_length
+    MapReflectionFriend::ScalarMapGetItem,   // mp_subscript
+    MapReflectionFriend::ScalarMapSetItem,   // mp_ass_subscript
+  };
+
+  PyTypeObject ScalarMapContainer_Type = {
+    PyVarObject_HEAD_INIT(&PyType_Type, 0)
+    FULL_MODULE_NAME ".ScalarMapContainer",  //  tp_name
+    sizeof(MapContainer),                //  tp_basicsize
+    0,                                   //  tp_itemsize
+    ScalarMapDealloc,                    //  tp_dealloc
+    0,                                   //  tp_print
+    0,                                   //  tp_getattr
+    0,                                   //  tp_setattr
+    0,                                   //  tp_compare
+    0,                                   //  tp_repr
+    0,                                   //  tp_as_number
+    0,                                   //  tp_as_sequence
+    &ScalarMapMappingMethods,            //  tp_as_mapping
+    0,                                   //  tp_hash
+    0,                                   //  tp_call
+    0,                                   //  tp_str
+    0,                                   //  tp_getattro
+    0,                                   //  tp_setattro
+    0,                                   //  tp_as_buffer
+    Py_TPFLAGS_DEFAULT,                  //  tp_flags
+    "A scalar map container",            //  tp_doc
+    0,                                   //  tp_traverse
+    0,                                   //  tp_clear
+    0,                                   //  tp_richcompare
+    0,                                   //  tp_weaklistoffset
+    MapReflectionFriend::GetIterator,    //  tp_iter
+    0,                                   //  tp_iternext
+    ScalarMapMethods,                    //  tp_methods
+    0,                                   //  tp_members
+    0,                                   //  tp_getset
+    0,                                   //  tp_base
+    0,                                   //  tp_dict
+    0,                                   //  tp_descr_get
+    0,                                   //  tp_descr_set
+    0,                                   //  tp_dictoffset
+    0,                                   //  tp_init
+  };
+#endif
+
+
+// MessageMap //////////////////////////////////////////////////////////////////
+
+static MessageMapContainer* GetMessageMap(PyObject* obj) {
+  return reinterpret_cast<MessageMapContainer*>(obj);
+}
+
+static PyObject* GetCMessage(MessageMapContainer* self, Message* message) {
+  // Get or create the CMessage object corresponding to this message.
+  ScopedPyObjectPtr key(PyLong_FromVoidPtr(message));
+  PyObject* ret = PyDict_GetItem(self->message_dict, key.get());
+
+  if (ret == NULL) {
+    CMessage* cmsg = cmessage::NewEmptyMessage(self->subclass_init,
+                                               message->GetDescriptor());
+    ret = reinterpret_cast<PyObject*>(cmsg);
+
+    if (cmsg == NULL) {
+      return NULL;
+    }
+    cmsg->owner = self->owner;
+    cmsg->message = message;
+    cmsg->parent = self->parent;
+
+    if (PyDict_SetItem(self->message_dict, key.get(), ret) < 0) {
+      Py_DECREF(ret);
+      return NULL;
+    }
+  } else {
+    Py_INCREF(ret);
+  }
+
+  return ret;
+}
+
+PyObject* NewMessageMapContainer(
+    CMessage* parent, const google::protobuf::FieldDescriptor* parent_field_descriptor,
+    PyObject* concrete_class) {
+  if (!CheckFieldBelongsToMessage(parent_field_descriptor, parent->message)) {
+    return NULL;
+  }
+
+#if PY_MAJOR_VERSION >= 3
+  PyObject* obj = PyType_GenericAlloc(
+        reinterpret_cast<PyTypeObject *>(MessageMapContainer_Type), 0);
+#else
+  PyObject* obj = PyType_GenericAlloc(&MessageMapContainer_Type, 0);
+#endif
+  if (obj == NULL) {
+    return PyErr_Format(PyExc_RuntimeError,
+                        "Could not allocate new container.");
+  }
+
+  MessageMapContainer* self = GetMessageMap(obj);
+
+  self->message = parent->message;
+  self->parent = parent;
+  self->parent_field_descriptor = parent_field_descriptor;
+  self->owner = parent->owner;
+  self->version = 0;
+
+  self->key_field_descriptor =
+      parent_field_descriptor->message_type()->FindFieldByName("key");
+  self->value_field_descriptor =
+      parent_field_descriptor->message_type()->FindFieldByName("value");
+
+  self->message_dict = PyDict_New();
+  if (self->message_dict == NULL) {
+    return PyErr_Format(PyExc_RuntimeError,
+                        "Could not allocate message dict.");
+  }
+
+  Py_INCREF(concrete_class);
+  self->subclass_init = concrete_class;
+
+  if (self->key_field_descriptor == NULL ||
+      self->value_field_descriptor == NULL) {
+    Py_DECREF(obj);
+    return PyErr_Format(PyExc_KeyError,
+                        "Map entry descriptor did not have key/value fields");
+  }
+
+  return obj;
+}
+
+int MapReflectionFriend::MessageMapSetItem(PyObject* _self, PyObject* key,
+                                           PyObject* v) {
+  if (v) {
+    PyErr_Format(PyExc_ValueError,
+                 "Direct assignment of submessage not allowed");
+    return -1;
+  }
+
+  // Now we know that this is a delete, not a set.
+
+  MessageMapContainer* self = GetMessageMap(_self);
+  Message* message = self->GetMutableMessage();
+  const Reflection* reflection = message->GetReflection();
+  MapKey map_key;
+  MapValueRef value;
+
+  self->version++;
+
+  if (!PythonToMapKey(key, self->key_field_descriptor, &map_key)) {
+    return -1;
+  }
+
+  // Delete key from map.
+  if (reflection->DeleteMapValue(message, self->parent_field_descriptor,
+                                 map_key)) {
+    return 0;
+  } else {
+    PyErr_Format(PyExc_KeyError, "Key not present in map");
+    return -1;
+  }
+}
+
+PyObject* MapReflectionFriend::MessageMapGetItem(PyObject* _self,
+                                                 PyObject* key) {
+  MessageMapContainer* self = GetMessageMap(_self);
+
+  Message* message = self->GetMutableMessage();
+  const Reflection* reflection = message->GetReflection();
+  MapKey map_key;
+  MapValueRef value;
+
+  if (!PythonToMapKey(key, self->key_field_descriptor, &map_key)) {
+    return NULL;
+  }
+
+  if (reflection->InsertOrLookupMapValue(message, self->parent_field_descriptor,
+                                         map_key, &value)) {
+    self->version++;
+  }
+
+  return GetCMessage(self, value.MutableMessageValue());
+}
+
+PyObject* MessageMapGet(PyObject* self, PyObject* args) {
+  PyObject* key;
+  PyObject* default_value = NULL;
+  if (PyArg_ParseTuple(args, "O|O", &key, &default_value) < 0) {
+    return NULL;
+  }
+
+  ScopedPyObjectPtr is_present(MapReflectionFriend::Contains(self, key));
+  if (is_present.get() == NULL) {
+    return NULL;
+  }
+
+  if (PyObject_IsTrue(is_present.get())) {
+    return MapReflectionFriend::MessageMapGetItem(self, key);
+  } else {
+    if (default_value != NULL) {
+      Py_INCREF(default_value);
+      return default_value;
+    } else {
+      Py_RETURN_NONE;
+    }
+  }
+}
+
+static void MessageMapDealloc(PyObject* _self) {
+  MessageMapContainer* self = GetMessageMap(_self);
+  self->owner.reset();
+  Py_DECREF(self->message_dict);
+  Py_TYPE(_self)->tp_free(_self);
+}
+
+static PyMethodDef MessageMapMethods[] = {
+  { "__contains__", (PyCFunction)MapReflectionFriend::Contains, METH_O,
+    "Tests whether the map contains this element."},
+  { "clear", (PyCFunction)Clear, METH_NOARGS,
+    "Removes all elements from the map."},
+  { "get", MessageMapGet, METH_VARARGS,
+    "Gets the value for the given key if present, or otherwise a default" },
+  { "get_or_create", MapReflectionFriend::MessageMapGetItem, METH_O,
+    "Alias for getitem, useful to make explicit that the map is mutated." },
+  /*
+  { "__deepcopy__", (PyCFunction)DeepCopy, METH_VARARGS,
+    "Makes a deep copy of the class." },
+  { "__reduce__", (PyCFunction)Reduce, METH_NOARGS,
+    "Outputs picklable representation of the repeated field." },
+  */
+  {NULL, NULL},
+};
+
+#if PY_MAJOR_VERSION >= 3
+  static PyType_Slot MessageMapContainer_Type_slots[] = {
+      {Py_tp_dealloc, (void *)MessageMapDealloc},
+      {Py_mp_length, (void *)MapReflectionFriend::Length},
+      {Py_mp_subscript, (void *)MapReflectionFriend::MessageMapGetItem},
+      {Py_mp_ass_subscript, (void *)MapReflectionFriend::MessageMapSetItem},
+      {Py_tp_methods, (void *)MessageMapMethods},
+      {Py_tp_iter, (void *)MapReflectionFriend::GetIterator},
+      {0, 0}
+  };
+
+  PyType_Spec MessageMapContainer_Type_spec = {
+      FULL_MODULE_NAME ".MessageMapContainer",
+      sizeof(MessageMapContainer),
+      0,
+      Py_TPFLAGS_DEFAULT,
+      MessageMapContainer_Type_slots
+  };
+
+  PyObject *MessageMapContainer_Type;
+#else
+  static PyMappingMethods MessageMapMappingMethods = {
+    MapReflectionFriend::Length,              // mp_length
+    MapReflectionFriend::MessageMapGetItem,   // mp_subscript
+    MapReflectionFriend::MessageMapSetItem,   // mp_ass_subscript
+  };
+
+  PyTypeObject MessageMapContainer_Type = {
+    PyVarObject_HEAD_INIT(&PyType_Type, 0)
+    FULL_MODULE_NAME ".MessageMapContainer",  //  tp_name
+    sizeof(MessageMapContainer),         //  tp_basicsize
+    0,                                   //  tp_itemsize
+    MessageMapDealloc,                   //  tp_dealloc
+    0,                                   //  tp_print
+    0,                                   //  tp_getattr
+    0,                                   //  tp_setattr
+    0,                                   //  tp_compare
+    0,                                   //  tp_repr
+    0,                                   //  tp_as_number
+    0,                                   //  tp_as_sequence
+    &MessageMapMappingMethods,           //  tp_as_mapping
+    0,                                   //  tp_hash
+    0,                                   //  tp_call
+    0,                                   //  tp_str
+    0,                                   //  tp_getattro
+    0,                                   //  tp_setattro
+    0,                                   //  tp_as_buffer
+    Py_TPFLAGS_DEFAULT,                  //  tp_flags
+    "A map container for message",       //  tp_doc
+    0,                                   //  tp_traverse
+    0,                                   //  tp_clear
+    0,                                   //  tp_richcompare
+    0,                                   //  tp_weaklistoffset
+    MapReflectionFriend::GetIterator,    //  tp_iter
+    0,                                   //  tp_iternext
+    MessageMapMethods,                   //  tp_methods
+    0,                                   //  tp_members
+    0,                                   //  tp_getset
+    0,                                   //  tp_base
+    0,                                   //  tp_dict
+    0,                                   //  tp_descr_get
+    0,                                   //  tp_descr_set
+    0,                                   //  tp_dictoffset
+    0,                                   //  tp_init
+  };
+#endif
+
+// MapIterator /////////////////////////////////////////////////////////////////
+
+static MapIterator* GetIter(PyObject* obj) {
+  return reinterpret_cast<MapIterator*>(obj);
+}
+
+PyObject* MapReflectionFriend::GetIterator(PyObject *_self) {
+  MapContainer* self = GetMap(_self);
+
+  ScopedPyObjectPtr obj(PyType_GenericAlloc(&MapIterator_Type, 0));
+  if (obj == NULL) {
+    return PyErr_Format(PyExc_KeyError, "Could not allocate iterator");
+  }
+
+  MapIterator* iter = GetIter(obj.get());
+
+  Py_INCREF(self);
+  iter->container = self;
+  iter->version = self->version;
+  iter->owner = self->owner;
+
+  if (MapReflectionFriend::Length(_self) > 0) {
+    Message* message = self->GetMutableMessage();
+    const Reflection* reflection = message->GetReflection();
+
+    iter->iter.reset(new ::google::protobuf::MapIterator(
+        reflection->MapBegin(message, self->parent_field_descriptor)));
+  }
+
+  return obj.release();
+}
+
+PyObject* MapReflectionFriend::IterNext(PyObject* _self) {
+  MapIterator* self = GetIter(_self);
+
+  // This won't catch mutations to the map performed by MergeFrom(); no easy way
+  // to address that.
+  if (self->version != self->container->version) {
+    return PyErr_Format(PyExc_RuntimeError,
+                        "Map modified during iteration.");
+  }
+
+  if (self->iter.get() == NULL) {
+    return NULL;
+  }
+
+  Message* message = self->container->GetMutableMessage();
+  const Reflection* reflection = message->GetReflection();
+
+  if (*self->iter ==
+      reflection->MapEnd(message, self->container->parent_field_descriptor)) {
+    return NULL;
+  }
+
+  PyObject* ret = MapKeyToPython(self->container->key_field_descriptor,
+                                 self->iter->GetKey());
+
+  ++(*self->iter);
+
+  return ret;
+}
+
+static void DeallocMapIterator(PyObject* _self) {
+  MapIterator* self = GetIter(_self);
+  self->iter.reset();
+  self->owner.reset();
+  Py_XDECREF(self->container);
+  Py_TYPE(_self)->tp_free(_self);
+}
+
+PyTypeObject MapIterator_Type = {
+  PyVarObject_HEAD_INIT(&PyType_Type, 0)
+  FULL_MODULE_NAME ".MapIterator",     //  tp_name
+  sizeof(MapIterator),                 //  tp_basicsize
+  0,                                   //  tp_itemsize
+  DeallocMapIterator,                  //  tp_dealloc
+  0,                                   //  tp_print
+  0,                                   //  tp_getattr
+  0,                                   //  tp_setattr
+  0,                                   //  tp_compare
+  0,                                   //  tp_repr
+  0,                                   //  tp_as_number
+  0,                                   //  tp_as_sequence
+  0,                                   //  tp_as_mapping
+  0,                                   //  tp_hash
+  0,                                   //  tp_call
+  0,                                   //  tp_str
+  0,                                   //  tp_getattro
+  0,                                   //  tp_setattro
+  0,                                   //  tp_as_buffer
+  Py_TPFLAGS_DEFAULT,                  //  tp_flags
+  "A scalar map iterator",             //  tp_doc
+  0,                                   //  tp_traverse
+  0,                                   //  tp_clear
+  0,                                   //  tp_richcompare
+  0,                                   //  tp_weaklistoffset
+  PyObject_SelfIter,                   //  tp_iter
+  MapReflectionFriend::IterNext,       //  tp_iternext
+  0,                                   //  tp_methods
+  0,                                   //  tp_members
+  0,                                   //  tp_getset
+  0,                                   //  tp_base
+  0,                                   //  tp_dict
+  0,                                   //  tp_descr_get
+  0,                                   //  tp_descr_set
+  0,                                   //  tp_dictoffset
+  0,                                   //  tp_init
+};
+
+}  // namespace python
+}  // namespace protobuf
+}  // namespace google
diff --git a/third_party/protobuf/python/google/protobuf/pyext/map_container.h b/third_party/protobuf/python/google/protobuf/pyext/map_container.h
new file mode 100644
index 0000000000..ddf94be71b
--- /dev/null
+++ b/third_party/protobuf/python/google/protobuf/pyext/map_container.h
@@ -0,0 +1,141 @@
+// 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.
+
+#ifndef GOOGLE_PROTOBUF_PYTHON_CPP_MAP_CONTAINER_H__
+#define GOOGLE_PROTOBUF_PYTHON_CPP_MAP_CONTAINER_H__
+
+#include <Python.h>
+
+#include <memory>
+#ifndef _SHARED_PTR_H
+#include <google/protobuf/stubs/shared_ptr.h>
+#endif
+
+#include <google/protobuf/descriptor.h>
+#include <google/protobuf/message.h>
+
+namespace google {
+namespace protobuf {
+
+class Message;
+
+#ifdef _SHARED_PTR_H
+using std::shared_ptr;
+#else
+using internal::shared_ptr;
+#endif
+
+namespace python {
+
+struct CMessage;
+
+// This struct is used directly for ScalarMap, and is the base class of
+// MessageMapContainer, which is used for MessageMap.
+struct MapContainer {
+  PyObject_HEAD;
+
+  // This is the top-level C++ Message object that owns the whole
+  // proto tree.  Every Python MapContainer holds a
+  // reference to it in order to keep it alive as long as there's a
+  // Python object that references any part of the tree.
+  shared_ptr<Message> owner;
+
+  // Pointer to the C++ Message that contains this container.  The
+  // MapContainer does not own this pointer.
+  const Message* message;
+
+  // Use to get a mutable message when necessary.
+  Message* GetMutableMessage();
+
+  // Weak reference to a parent CMessage object (i.e. may be NULL.)
+  //
+  // Used to make sure all ancestors are also mutable when first
+  // modifying the container.
+  CMessage* parent;
+
+  // Pointer to the parent's descriptor that describes this
+  // field.  Used together with the parent's message when making a
+  // default message instance mutable.
+  // The pointer is owned by the global DescriptorPool.
+  const FieldDescriptor* parent_field_descriptor;
+  const FieldDescriptor* key_field_descriptor;
+  const FieldDescriptor* value_field_descriptor;
+
+  // We bump this whenever we perform a mutation, to invalidate existing
+  // iterators.
+  uint64 version;
+
+  // Releases the messages in the container to a new message.
+  //
+  // Returns 0 on success, -1 on failure.
+  int Release();
+
+  // Set the owner field of self and any children of self.
+  void SetOwner(const shared_ptr<Message>& new_owner) {
+    owner = new_owner;
+  }
+};
+
+struct MessageMapContainer : public MapContainer {
+  // A callable that is used to create new child messages.
+  PyObject* subclass_init;
+
+  // A dict mapping Message* -> CMessage.
+  PyObject* message_dict;
+};
+
+#if PY_MAJOR_VERSION >= 3
+  extern PyObject *MessageMapContainer_Type;
+  extern PyType_Spec MessageMapContainer_Type_spec;
+  extern PyObject *ScalarMapContainer_Type;
+  extern PyType_Spec ScalarMapContainer_Type_spec;
+#else
+  extern PyTypeObject MessageMapContainer_Type;
+  extern PyTypeObject ScalarMapContainer_Type;
+#endif
+
+extern PyTypeObject MapIterator_Type;  // Both map types use the same iterator.
+
+// Builds a MapContainer object, from a parent message and a
+// field descriptor.
+extern PyObject* NewScalarMapContainer(
+    CMessage* parent, const FieldDescriptor* parent_field_descriptor);
+
+// Builds a MessageMap object, from a parent message and a
+// field descriptor.
+extern PyObject* NewMessageMapContainer(
+    CMessage* parent, const FieldDescriptor* parent_field_descriptor,
+    PyObject* concrete_class);
+
+}  // namespace python
+}  // namespace protobuf
+
+}  // namespace google
+#endif  // GOOGLE_PROTOBUF_PYTHON_CPP_MAP_CONTAINER_H__
diff --git a/third_party/protobuf/python/google/protobuf/pyext/message.cc b/third_party/protobuf/python/google/protobuf/pyext/message.cc
new file mode 100644
index 0000000000..60ec9c1b96
--- /dev/null
+++ b/third_party/protobuf/python/google/protobuf/pyext/message.cc
@@ -0,0 +1,3091 @@
+// 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: anuraag@google.com (Anuraag Agrawal)
+// Author: tibell@google.com (Johan Tibell)
+
+#include <google/protobuf/pyext/message.h>
+
+#include <map>
+#include <memory>
+#ifndef _SHARED_PTR_H
+#include <google/protobuf/stubs/shared_ptr.h>
+#endif
+#include <string>
+#include <vector>
+#include <structmember.h>  // A Python header file.
+
+#ifndef PyVarObject_HEAD_INIT
+#define PyVarObject_HEAD_INIT(type, size) PyObject_HEAD_INIT(type) size,
+#endif
+#ifndef Py_TYPE
+#define Py_TYPE(ob) (((PyObject*)(ob))->ob_type)
+#endif
+#include <google/protobuf/descriptor.pb.h>
+#include <google/protobuf/stubs/common.h>
+#include <google/protobuf/stubs/logging.h>
+#include <google/protobuf/io/coded_stream.h>
+#include <google/protobuf/util/message_differencer.h>
+#include <google/protobuf/descriptor.h>
+#include <google/protobuf/message.h>
+#include <google/protobuf/text_format.h>
+#include <google/protobuf/unknown_field_set.h>
+#include <google/protobuf/pyext/descriptor.h>
+#include <google/protobuf/pyext/descriptor_pool.h>
+#include <google/protobuf/pyext/extension_dict.h>
+#include <google/protobuf/pyext/repeated_composite_container.h>
+#include <google/protobuf/pyext/repeated_scalar_container.h>
+#include <google/protobuf/pyext/map_container.h>
+#include <google/protobuf/pyext/scoped_pyobject_ptr.h>
+#include <google/protobuf/stubs/strutil.h>
+
+#if PY_MAJOR_VERSION >= 3
+  #define PyInt_Check PyLong_Check
+  #define PyInt_AsLong PyLong_AsLong
+  #define PyInt_FromLong PyLong_FromLong
+  #define PyInt_FromSize_t PyLong_FromSize_t
+  #define PyString_Check PyUnicode_Check
+  #define PyString_FromString PyUnicode_FromString
+  #define PyString_FromStringAndSize PyUnicode_FromStringAndSize
+  #if PY_VERSION_HEX < 0x03030000
+    #error "Python 3.0 - 3.2 are not supported."
+  #else
+  #define PyString_AsString(ob) \
+    (PyUnicode_Check(ob)? PyUnicode_AsUTF8(ob): PyBytes_AsString(ob))
+  #define PyString_AsStringAndSize(ob, charpp, sizep) \
+    (PyUnicode_Check(ob)? \
+       ((*(charpp) = PyUnicode_AsUTF8AndSize(ob, (sizep))) == NULL? -1: 0): \
+       PyBytes_AsStringAndSize(ob, (charpp), (sizep)))
+  #endif
+#endif
+
+namespace google {
+namespace protobuf {
+namespace python {
+
+static PyObject* kDESCRIPTOR;
+static PyObject* k_extensions_by_name;
+static PyObject* k_extensions_by_number;
+PyObject* EnumTypeWrapper_class;
+static PyObject* PythonMessage_class;
+static PyObject* kEmptyWeakref;
+static PyObject* WKT_classes = NULL;
+
+// Defines the Metaclass of all Message classes.
+// It allows us to cache some C++ pointers in the class object itself, they are
+// faster to extract than from the type's dictionary.
+
+struct PyMessageMeta {
+  // This is how CPython subclasses C structures: the base structure must be
+  // the first member of the object.
+  PyHeapTypeObject super;
+
+  // C++ descriptor of this message.
+  const Descriptor* message_descriptor;
+
+  // Owned reference, used to keep the pointer above alive.
+  PyObject* py_message_descriptor;
+
+  // The Python DescriptorPool used to create the class. It is needed to resolve
+  // fields descriptors, including extensions fields; its C++ MessageFactory is
+  // used to instantiate submessages.
+  // This can be different from DESCRIPTOR.file.pool, in the case of a custom
+  // DescriptorPool which defines new extensions.
+  // We own the reference, because it's important to keep the descriptors and
+  // factory alive.
+  PyDescriptorPool* py_descriptor_pool;
+};
+
+namespace message_meta {
+
+static int InsertEmptyWeakref(PyTypeObject* base);
+
+// Add the number of a field descriptor to the containing message class.
+// Equivalent to:
+//   _cls.<field>_FIELD_NUMBER = <number>
+static bool AddFieldNumberToClass(
+    PyObject* cls, const FieldDescriptor* field_descriptor) {
+  string constant_name = field_descriptor->name() + "_FIELD_NUMBER";
+  UpperString(&constant_name);
+  ScopedPyObjectPtr attr_name(PyString_FromStringAndSize(
+      constant_name.c_str(), constant_name.size()));
+  if (attr_name == NULL) {
+    return false;
+  }
+  ScopedPyObjectPtr number(PyInt_FromLong(field_descriptor->number()));
+  if (number == NULL) {
+    return false;
+  }
+  if (PyObject_SetAttr(cls, attr_name.get(), number.get()) == -1) {
+    return false;
+  }
+  return true;
+}
+
+
+// Finalize the creation of the Message class.
+static int AddDescriptors(PyObject* cls, const Descriptor* descriptor) {
+  // If there are extension_ranges, the message is "extendable", and extension
+  // classes will register themselves in this class.
+  if (descriptor->extension_range_count() > 0) {
+    ScopedPyObjectPtr by_name(PyDict_New());
+    if (PyObject_SetAttr(cls, k_extensions_by_name, by_name.get()) < 0) {
+      return -1;
+    }
+    ScopedPyObjectPtr by_number(PyDict_New());
+    if (PyObject_SetAttr(cls, k_extensions_by_number, by_number.get()) < 0) {
+      return -1;
+    }
+  }
+
+  // For each field set: cls.<field>_FIELD_NUMBER = <number>
+  for (int i = 0; i < descriptor->field_count(); ++i) {
+    if (!AddFieldNumberToClass(cls, descriptor->field(i))) {
+      return -1;
+    }
+  }
+
+  // For each enum set cls.<enum name> = EnumTypeWrapper(<enum descriptor>).
+  //
+  // The enum descriptor we get from
+  // <messagedescriptor>.enum_types_by_name[name]
+  // which was built previously.
+  for (int i = 0; i < descriptor->enum_type_count(); ++i) {
+    const EnumDescriptor* enum_descriptor = descriptor->enum_type(i);
+    ScopedPyObjectPtr enum_type(
+        PyEnumDescriptor_FromDescriptor(enum_descriptor));
+    if (enum_type == NULL) {
+      return -1;
+     }
+    // Add wrapped enum type to message class.
+    ScopedPyObjectPtr wrapped(PyObject_CallFunctionObjArgs(
+        EnumTypeWrapper_class, enum_type.get(), NULL));
+    if (wrapped == NULL) {
+      return -1;
+    }
+    if (PyObject_SetAttrString(
+            cls, enum_descriptor->name().c_str(), wrapped.get()) == -1) {
+      return -1;
+    }
+
+    // For each enum value add cls.<name> = <number>
+    for (int j = 0; j < enum_descriptor->value_count(); ++j) {
+      const EnumValueDescriptor* enum_value_descriptor =
+          enum_descriptor->value(j);
+      ScopedPyObjectPtr value_number(PyInt_FromLong(
+          enum_value_descriptor->number()));
+      if (value_number == NULL) {
+        return -1;
+      }
+      if (PyObject_SetAttrString(cls, enum_value_descriptor->name().c_str(),
+                                 value_number.get()) == -1) {
+        return -1;
+      }
+    }
+  }
+
+  // For each extension set cls.<extension name> = <extension descriptor>.
+  //
+  // Extension descriptors come from
+  // <message descriptor>.extensions_by_name[name]
+  // which was defined previously.
+  for (int i = 0; i < descriptor->extension_count(); ++i) {
+    const google::protobuf::FieldDescriptor* field = descriptor->extension(i);
+    ScopedPyObjectPtr extension_field(PyFieldDescriptor_FromDescriptor(field));
+    if (extension_field == NULL) {
+      return -1;
+    }
+
+    // Add the extension field to the message class.
+    if (PyObject_SetAttrString(
+            cls, field->name().c_str(), extension_field.get()) == -1) {
+      return -1;
+    }
+
+    // For each extension set cls.<extension name>_FIELD_NUMBER = <number>.
+    if (!AddFieldNumberToClass(cls, field)) {
+      return -1;
+    }
+  }
+
+  return 0;
+}
+
+static PyObject* New(PyTypeObject* type,
+                     PyObject* args, PyObject* kwargs) {
+  static char *kwlist[] = {"name", "bases", "dict", 0};
+  PyObject *bases, *dict;
+  const char* name;
+
+  // Check arguments: (name, bases, dict)
+  if (!PyArg_ParseTupleAndKeywords(args, kwargs, "sO!O!:type", kwlist,
+                                   &name,
+                                   &PyTuple_Type, &bases,
+                                   &PyDict_Type, &dict)) {
+    return NULL;
+  }
+
+  // Check bases: only (), or (message.Message,) are allowed
+  if (!(PyTuple_GET_SIZE(bases) == 0 ||
+        (PyTuple_GET_SIZE(bases) == 1 &&
+         PyTuple_GET_ITEM(bases, 0) == PythonMessage_class))) {
+    PyErr_SetString(PyExc_TypeError,
+                    "A Message class can only inherit from Message");
+    return NULL;
+  }
+
+  // Check dict['DESCRIPTOR']
+  PyObject* py_descriptor = PyDict_GetItem(dict, kDESCRIPTOR);
+  if (py_descriptor == NULL) {
+    PyErr_SetString(PyExc_TypeError, "Message class has no DESCRIPTOR");
+    return NULL;
+  }
+  if (!PyObject_TypeCheck(py_descriptor, &PyMessageDescriptor_Type)) {
+    PyErr_Format(PyExc_TypeError, "Expected a message Descriptor, got %s",
+                 py_descriptor->ob_type->tp_name);
+    return NULL;
+  }
+
+  // Build the arguments to the base metaclass.
+  // We change the __bases__ classes.
+  ScopedPyObjectPtr new_args;
+  const Descriptor* message_descriptor =
+      PyMessageDescriptor_AsDescriptor(py_descriptor);
+  if (message_descriptor == NULL) {
+    return NULL;
+  }
+
+  if (WKT_classes == NULL) {
+    ScopedPyObjectPtr well_known_types(PyImport_ImportModule(
+        "google.protobuf.internal.well_known_types"));
+    GOOGLE_DCHECK(well_known_types != NULL);
+
+    WKT_classes = PyObject_GetAttrString(well_known_types.get(), "WKTBASES");
+    GOOGLE_DCHECK(WKT_classes != NULL);
+  }
+
+  PyObject* well_known_class = PyDict_GetItemString(
+      WKT_classes, message_descriptor->full_name().c_str());
+  if (well_known_class == NULL) {
+    new_args.reset(Py_BuildValue("s(OO)O", name, &CMessage_Type,
+                                 PythonMessage_class, dict));
+  } else {
+    new_args.reset(Py_BuildValue("s(OOO)O", name, &CMessage_Type,
+                                 PythonMessage_class, well_known_class, dict));
+  }
+
+  if (new_args == NULL) {
+    return NULL;
+  }
+  // Call the base metaclass.
+  ScopedPyObjectPtr result(PyType_Type.tp_new(type, new_args.get(), NULL));
+  if (result == NULL) {
+    return NULL;
+  }
+  PyMessageMeta* newtype = reinterpret_cast<PyMessageMeta*>(result.get());
+
+  // Insert the empty weakref into the base classes.
+  if (InsertEmptyWeakref(
+          reinterpret_cast<PyTypeObject*>(PythonMessage_class)) < 0 ||
+      InsertEmptyWeakref(&CMessage_Type) < 0) {
+    return NULL;
+  }
+
+  // Cache the descriptor, both as Python object and as C++ pointer.
+  const Descriptor* descriptor =
+      PyMessageDescriptor_AsDescriptor(py_descriptor);
+  if (descriptor == NULL) {
+    return NULL;
+  }
+  Py_INCREF(py_descriptor);
+  newtype->py_message_descriptor = py_descriptor;
+  newtype->message_descriptor = descriptor;
+  // TODO(amauryfa): Don't always use the canonical pool of the descriptor,
+  // use the MessageFactory optionally passed in the class dict.
+  newtype->py_descriptor_pool = GetDescriptorPool_FromPool(
+      descriptor->file()->pool());
+  if (newtype->py_descriptor_pool == NULL) {
+    return NULL;
+  }
+  Py_INCREF(newtype->py_descriptor_pool);
+
+  // Add the message to the DescriptorPool.
+  if (cdescriptor_pool::RegisterMessageClass(newtype->py_descriptor_pool,
+                                             descriptor, result.get()) < 0) {
+    return NULL;
+  }
+
+  // Continue with type initialization: add other descriptors, enum values...
+  if (AddDescriptors(result.get(), descriptor) < 0) {
+    return NULL;
+  }
+  return result.release();
+}
+
+static void Dealloc(PyMessageMeta *self) {
+  Py_DECREF(self->py_message_descriptor);
+  Py_DECREF(self->py_descriptor_pool);
+  Py_TYPE(self)->tp_free(reinterpret_cast<PyObject*>(self));
+}
+
+
+// This function inserts and empty weakref at the end of the list of
+// subclasses for the main protocol buffer Message class.
+//
+// This eliminates a O(n^2) behaviour in the internal add_subclass
+// routine.
+static int InsertEmptyWeakref(PyTypeObject *base_type) {
+#if PY_MAJOR_VERSION >= 3
+  // Python 3.4 has already included the fix for the issue that this
+  // hack addresses. For further background and the fix please see
+  // https://bugs.python.org/issue17936.
+  return 0;
+#else
+  PyObject *subclasses = base_type->tp_subclasses;
+  if (subclasses && PyList_CheckExact(subclasses)) {
+    return PyList_Append(subclasses, kEmptyWeakref);
+  }
+  return 0;
+#endif  // PY_MAJOR_VERSION >= 3
+}
+
+}  // namespace message_meta
+
+PyTypeObject PyMessageMeta_Type = {
+  PyVarObject_HEAD_INIT(&PyType_Type, 0)
+  FULL_MODULE_NAME ".MessageMeta",     // tp_name
+  sizeof(PyMessageMeta),               // tp_basicsize
+  0,                                   // tp_itemsize
+  (destructor)message_meta::Dealloc,   // tp_dealloc
+  0,                                   // tp_print
+  0,                                   // tp_getattr
+  0,                                   // tp_setattr
+  0,                                   // tp_compare
+  0,                                   // tp_repr
+  0,                                   // tp_as_number
+  0,                                   // tp_as_sequence
+  0,                                   // tp_as_mapping
+  0,                                   // tp_hash
+  0,                                   // tp_call
+  0,                                   // tp_str
+  0,                                   // tp_getattro
+  0,                                   // tp_setattro
+  0,                                   // tp_as_buffer
+  Py_TPFLAGS_DEFAULT | Py_TPFLAGS_BASETYPE,  // tp_flags
+  "The metaclass of ProtocolMessages",  // tp_doc
+  0,                                   // tp_traverse
+  0,                                   // tp_clear
+  0,                                   // tp_richcompare
+  0,                                   // tp_weaklistoffset
+  0,                                   // tp_iter
+  0,                                   // tp_iternext
+  0,                                   // tp_methods
+  0,                                   // tp_members
+  0,                                   // tp_getset
+  0,                                   // tp_base
+  0,                                   // tp_dict
+  0,                                   // tp_descr_get
+  0,                                   // tp_descr_set
+  0,                                   // tp_dictoffset
+  0,                                   // tp_init
+  0,                                   // tp_alloc
+  message_meta::New,                   // tp_new
+};
+
+static PyMessageMeta* CheckMessageClass(PyTypeObject* cls) {
+  if (!PyObject_TypeCheck(cls, &PyMessageMeta_Type)) {
+    PyErr_Format(PyExc_TypeError, "Class %s is not a Message", cls->tp_name);
+    return NULL;
+  }
+  return reinterpret_cast<PyMessageMeta*>(cls);
+}
+
+static const Descriptor* GetMessageDescriptor(PyTypeObject* cls) {
+  PyMessageMeta* type = CheckMessageClass(cls);
+  if (type == NULL) {
+    return NULL;
+  }
+  return type->message_descriptor;
+}
+
+// Forward declarations
+namespace cmessage {
+int InternalReleaseFieldByDescriptor(
+    CMessage* self,
+    const FieldDescriptor* field_descriptor,
+    PyObject* composite_field);
+}  // namespace cmessage
+
+// ---------------------------------------------------------------------
+// Visiting the composite children of a CMessage
+
+struct ChildVisitor {
+  // Returns 0 on success, -1 on failure.
+  int VisitRepeatedCompositeContainer(RepeatedCompositeContainer* container) {
+    return 0;
+  }
+
+  // Returns 0 on success, -1 on failure.
+  int VisitRepeatedScalarContainer(RepeatedScalarContainer* container) {
+    return 0;
+  }
+
+  // Returns 0 on success, -1 on failure.
+  int VisitCMessage(CMessage* cmessage,
+                    const FieldDescriptor* field_descriptor) {
+    return 0;
+  }
+};
+
+// Apply a function to a composite field.  Does nothing if child is of
+// non-composite type.
+template<class Visitor>
+static int VisitCompositeField(const FieldDescriptor* descriptor,
+                               PyObject* child,
+                               Visitor visitor) {
+  if (descriptor->label() == FieldDescriptor::LABEL_REPEATED) {
+    if (descriptor->cpp_type() == FieldDescriptor::CPPTYPE_MESSAGE) {
+      if (descriptor->is_map()) {
+        MapContainer* container = reinterpret_cast<MapContainer*>(child);
+        if (visitor.VisitMapContainer(container) == -1) {
+          return -1;
+        }
+      } else {
+        RepeatedCompositeContainer* container =
+          reinterpret_cast<RepeatedCompositeContainer*>(child);
+        if (visitor.VisitRepeatedCompositeContainer(container) == -1)
+          return -1;
+      }
+    } else {
+      RepeatedScalarContainer* container =
+        reinterpret_cast<RepeatedScalarContainer*>(child);
+      if (visitor.VisitRepeatedScalarContainer(container) == -1)
+        return -1;
+    }
+  } else if (descriptor->cpp_type() == FieldDescriptor::CPPTYPE_MESSAGE) {
+    CMessage* cmsg = reinterpret_cast<CMessage*>(child);
+    if (visitor.VisitCMessage(cmsg, descriptor) == -1)
+      return -1;
+  }
+  // The ExtensionDict might contain non-composite fields, which we
+  // skip here.
+  return 0;
+}
+
+// Visit each composite field and extension field of this CMessage.
+// Returns -1 on error and 0 on success.
+template<class Visitor>
+int ForEachCompositeField(CMessage* self, Visitor visitor) {
+  Py_ssize_t pos = 0;
+  PyObject* key;
+  PyObject* field;
+
+  // Visit normal fields.
+  if (self->composite_fields) {
+    // Never use self->message in this function, it may be already freed.
+    const Descriptor* message_descriptor =
+        GetMessageDescriptor(Py_TYPE(self));
+    while (PyDict_Next(self->composite_fields, &pos, &key, &field)) {
+      Py_ssize_t key_str_size;
+      char *key_str_data;
+      if (PyString_AsStringAndSize(key, &key_str_data, &key_str_size) != 0)
+        return -1;
+      const string key_str(key_str_data, key_str_size);
+      const FieldDescriptor* descriptor =
+        message_descriptor->FindFieldByName(key_str);
+      if (descriptor != NULL) {
+        if (VisitCompositeField(descriptor, field, visitor) == -1)
+          return -1;
+      }
+    }
+  }
+
+  // Visit extension fields.
+  if (self->extensions != NULL) {
+    pos = 0;
+    while (PyDict_Next(self->extensions->values, &pos, &key, &field)) {
+      const FieldDescriptor* descriptor = cmessage::GetExtensionDescriptor(key);
+      if (descriptor == NULL)
+        return -1;
+      if (VisitCompositeField(descriptor, field, visitor) == -1)
+        return -1;
+    }
+  }
+
+  return 0;
+}
+
+// ---------------------------------------------------------------------
+
+// Constants used for integer type range checking.
+PyObject* kPythonZero;
+PyObject* kint32min_py;
+PyObject* kint32max_py;
+PyObject* kuint32max_py;
+PyObject* kint64min_py;
+PyObject* kint64max_py;
+PyObject* kuint64max_py;
+
+PyObject* EncodeError_class;
+PyObject* DecodeError_class;
+PyObject* PickleError_class;
+
+// Constant PyString values used for GetAttr/GetItem.
+static PyObject* k_cdescriptor;
+static PyObject* kfull_name;
+
+/* Is 64bit */
+void FormatTypeError(PyObject* arg, char* expected_types) {
+  PyObject* repr = PyObject_Repr(arg);
+  if (repr) {
+    PyErr_Format(PyExc_TypeError,
+                 "%.100s has type %.100s, but expected one of: %s",
+                 PyString_AsString(repr),
+                 Py_TYPE(arg)->tp_name,
+                 expected_types);
+    Py_DECREF(repr);
+  }
+}
+
+template<class T>
+bool CheckAndGetInteger(
+    PyObject* arg, T* value, PyObject* min, PyObject* max) {
+  bool is_long = PyLong_Check(arg);
+#if PY_MAJOR_VERSION < 3
+  if (!PyInt_Check(arg) && !is_long) {
+    FormatTypeError(arg, "int, long");
+    return false;
+  }
+  if (PyObject_Compare(min, arg) > 0 || PyObject_Compare(max, arg) < 0) {
+#else
+  if (!is_long) {
+    FormatTypeError(arg, "int");
+    return false;
+  }
+  if (PyObject_RichCompareBool(min, arg, Py_LE) != 1 ||
+      PyObject_RichCompareBool(max, arg, Py_GE) != 1) {
+#endif
+    if (!PyErr_Occurred()) {
+      PyObject *s = PyObject_Str(arg);
+      if (s) {
+        PyErr_Format(PyExc_ValueError,
+                     "Value out of range: %s",
+                     PyString_AsString(s));
+        Py_DECREF(s);
+      }
+    }
+    return false;
+  }
+#if PY_MAJOR_VERSION < 3
+  if (!is_long) {
+    *value = static_cast<T>(PyInt_AsLong(arg));
+  } else  // NOLINT
+#endif
+  {
+    if (min == kPythonZero) {
+      *value = static_cast<T>(PyLong_AsUnsignedLongLong(arg));
+    } else {
+      *value = static_cast<T>(PyLong_AsLongLong(arg));
+    }
+  }
+  return true;
+}
+
+// These are referenced by repeated_scalar_container, and must
+// be explicitly instantiated.
+template bool CheckAndGetInteger<int32>(
+    PyObject*, int32*, PyObject*, PyObject*);
+template bool CheckAndGetInteger<int64>(
+    PyObject*, int64*, PyObject*, PyObject*);
+template bool CheckAndGetInteger<uint32>(
+    PyObject*, uint32*, PyObject*, PyObject*);
+template bool CheckAndGetInteger<uint64>(
+    PyObject*, uint64*, PyObject*, PyObject*);
+
+bool CheckAndGetDouble(PyObject* arg, double* value) {
+  if (!PyInt_Check(arg) && !PyLong_Check(arg) &&
+      !PyFloat_Check(arg)) {
+    FormatTypeError(arg, "int, long, float");
+    return false;
+  }
+  *value = PyFloat_AsDouble(arg);
+  return true;
+}
+
+bool CheckAndGetFloat(PyObject* arg, float* value) {
+  double double_value;
+  if (!CheckAndGetDouble(arg, &double_value)) {
+    return false;
+  }
+  *value = static_cast<float>(double_value);
+  return true;
+}
+
+bool CheckAndGetBool(PyObject* arg, bool* value) {
+  if (!PyInt_Check(arg) && !PyBool_Check(arg) && !PyLong_Check(arg)) {
+    FormatTypeError(arg, "int, long, bool");
+    return false;
+  }
+  *value = static_cast<bool>(PyInt_AsLong(arg));
+  return true;
+}
+
+// Checks whether the given object (which must be "bytes" or "unicode") contains
+// valid UTF-8.
+bool IsValidUTF8(PyObject* obj) {
+  if (PyBytes_Check(obj)) {
+    PyObject* unicode = PyUnicode_FromEncodedObject(obj, "utf-8", NULL);
+
+    // Clear the error indicator; we report our own error when desired.
+    PyErr_Clear();
+
+    if (unicode) {
+      Py_DECREF(unicode);
+      return true;
+    } else {
+      return false;
+    }
+  } else {
+    // Unicode object, known to be valid UTF-8.
+    return true;
+  }
+}
+
+bool AllowInvalidUTF8(const FieldDescriptor* field) { return false; }
+
+PyObject* CheckString(PyObject* arg, const FieldDescriptor* descriptor) {
+  GOOGLE_DCHECK(descriptor->type() == FieldDescriptor::TYPE_STRING ||
+         descriptor->type() == FieldDescriptor::TYPE_BYTES);
+  if (descriptor->type() == FieldDescriptor::TYPE_STRING) {
+    if (!PyBytes_Check(arg) && !PyUnicode_Check(arg)) {
+      FormatTypeError(arg, "bytes, unicode");
+      return NULL;
+    }
+
+    if (!IsValidUTF8(arg) && !AllowInvalidUTF8(descriptor)) {
+      PyObject* repr = PyObject_Repr(arg);
+      PyErr_Format(PyExc_ValueError,
+                   "%s has type str, but isn't valid UTF-8 "
+                   "encoding. Non-UTF-8 strings must be converted to "
+                   "unicode objects before being added.",
+                   PyString_AsString(repr));
+      Py_DECREF(repr);
+      return NULL;
+    }
+  } else if (!PyBytes_Check(arg)) {
+    FormatTypeError(arg, "bytes");
+    return NULL;
+  }
+
+  PyObject* encoded_string = NULL;
+  if (descriptor->type() == FieldDescriptor::TYPE_STRING) {
+    if (PyBytes_Check(arg)) {
+      // The bytes were already validated as correctly encoded UTF-8 above.
+      encoded_string = arg;  // Already encoded.
+      Py_INCREF(encoded_string);
+    } else {
+      encoded_string = PyUnicode_AsEncodedObject(arg, "utf-8", NULL);
+    }
+  } else {
+    // In this case field type is "bytes".
+    encoded_string = arg;
+    Py_INCREF(encoded_string);
+  }
+
+  return encoded_string;
+}
+
+bool CheckAndSetString(
+    PyObject* arg, Message* message,
+    const FieldDescriptor* descriptor,
+    const Reflection* reflection,
+    bool append,
+    int index) {
+  ScopedPyObjectPtr encoded_string(CheckString(arg, descriptor));
+
+  if (encoded_string.get() == NULL) {
+    return false;
+  }
+
+  char* value;
+  Py_ssize_t value_len;
+  if (PyBytes_AsStringAndSize(encoded_string.get(), &value, &value_len) < 0) {
+    return false;
+  }
+
+  string value_string(value, value_len);
+  if (append) {
+    reflection->AddString(message, descriptor, value_string);
+  } else if (index < 0) {
+    reflection->SetString(message, descriptor, value_string);
+  } else {
+    reflection->SetRepeatedString(message, descriptor, index, value_string);
+  }
+  return true;
+}
+
+PyObject* ToStringObject(const FieldDescriptor* descriptor, string value) {
+  if (descriptor->type() != FieldDescriptor::TYPE_STRING) {
+    return PyBytes_FromStringAndSize(value.c_str(), value.length());
+  }
+
+  PyObject* result = PyUnicode_DecodeUTF8(value.c_str(), value.length(), NULL);
+  // If the string can't be decoded in UTF-8, just return a string object that
+  // contains the raw bytes. This can't happen if the value was assigned using
+  // the members of the Python message object, but can happen if the values were
+  // parsed from the wire (binary).
+  if (result == NULL) {
+    PyErr_Clear();
+    result = PyBytes_FromStringAndSize(value.c_str(), value.length());
+  }
+  return result;
+}
+
+bool CheckFieldBelongsToMessage(const FieldDescriptor* field_descriptor,
+                                const Message* message) {
+  if (message->GetDescriptor() == field_descriptor->containing_type()) {
+    return true;
+  }
+  PyErr_Format(PyExc_KeyError, "Field '%s' does not belong to message '%s'",
+               field_descriptor->full_name().c_str(),
+               message->GetDescriptor()->full_name().c_str());
+  return false;
+}
+
+namespace cmessage {
+
+PyDescriptorPool* GetDescriptorPoolForMessage(CMessage* message) {
+  // No need to check the type: the type of instances of CMessage is always
+  // an instance of PyMessageMeta. Let's prove it with a debug-only check.
+  GOOGLE_DCHECK(PyObject_TypeCheck(message, &CMessage_Type));
+  return reinterpret_cast<PyMessageMeta*>(Py_TYPE(message))->py_descriptor_pool;
+}
+
+MessageFactory* GetFactoryForMessage(CMessage* message) {
+  return GetDescriptorPoolForMessage(message)->message_factory;
+}
+
+static int MaybeReleaseOverlappingOneofField(
+    CMessage* cmessage,
+    const FieldDescriptor* field) {
+#ifdef GOOGLE_PROTOBUF_HAS_ONEOF
+  Message* message = cmessage->message;
+  const Reflection* reflection = message->GetReflection();
+  if (!field->containing_oneof() ||
+      !reflection->HasOneof(*message, field->containing_oneof()) ||
+      reflection->HasField(*message, field)) {
+    // No other field in this oneof, no need to release.
+    return 0;
+  }
+
+  const OneofDescriptor* oneof = field->containing_oneof();
+  const FieldDescriptor* existing_field =
+      reflection->GetOneofFieldDescriptor(*message, oneof);
+  if (existing_field->cpp_type() != FieldDescriptor::CPPTYPE_MESSAGE) {
+    // Non-message fields don't need to be released.
+    return 0;
+  }
+  const char* field_name = existing_field->name().c_str();
+  PyObject* child_message = cmessage->composite_fields ?
+      PyDict_GetItemString(cmessage->composite_fields, field_name) : NULL;
+  if (child_message == NULL) {
+    // No python reference to this field so no need to release.
+    return 0;
+  }
+
+  if (InternalReleaseFieldByDescriptor(
+          cmessage, existing_field, child_message) < 0) {
+    return -1;
+  }
+  return PyDict_DelItemString(cmessage->composite_fields, field_name);
+#else
+  return 0;
+#endif
+}
+
+// ---------------------------------------------------------------------
+// Making a message writable
+
+static Message* GetMutableMessage(
+    CMessage* parent,
+    const FieldDescriptor* parent_field) {
+  Message* parent_message = parent->message;
+  const Reflection* reflection = parent_message->GetReflection();
+  if (MaybeReleaseOverlappingOneofField(parent, parent_field) < 0) {
+    return NULL;
+  }
+  return reflection->MutableMessage(
+      parent_message, parent_field, GetFactoryForMessage(parent));
+}
+
+struct FixupMessageReference : public ChildVisitor {
+  // message must outlive this object.
+  explicit FixupMessageReference(Message* message) :
+      message_(message) {}
+
+  int VisitRepeatedCompositeContainer(RepeatedCompositeContainer* container) {
+    container->message = message_;
+    return 0;
+  }
+
+  int VisitRepeatedScalarContainer(RepeatedScalarContainer* container) {
+    container->message = message_;
+    return 0;
+  }
+
+  int VisitMapContainer(MapContainer* container) {
+    container->message = message_;
+    return 0;
+  }
+
+ private:
+  Message* message_;
+};
+
+int AssureWritable(CMessage* self) {
+  if (self == NULL || !self->read_only) {
+    return 0;
+  }
+
+  if (self->parent == NULL) {
+    // If parent is NULL but we are trying to modify a read-only message, this
+    // is a reference to a constant default instance that needs to be replaced
+    // with a mutable top-level message.
+    self->message = self->message->New();
+    self->owner.reset(self->message);
+    // Cascade the new owner to eventual children: even if this message is
+    // empty, some submessages or repeated containers might exist already.
+    SetOwner(self, self->owner);
+  } else {
+    // Otherwise, we need a mutable child message.
+    if (AssureWritable(self->parent) == -1)
+      return -1;
+
+    // Make self->message writable.
+    Message* mutable_message = GetMutableMessage(
+        self->parent,
+        self->parent_field_descriptor);
+    if (mutable_message == NULL) {
+      return -1;
+    }
+    self->message = mutable_message;
+  }
+  self->read_only = false;
+
+  // When a CMessage is made writable its Message pointer is updated
+  // to point to a new mutable Message.  When that happens we need to
+  // update any references to the old, read-only CMessage.  There are
+  // four places such references occur: RepeatedScalarContainer,
+  // RepeatedCompositeContainer, MapContainer, and ExtensionDict.
+  if (self->extensions != NULL)
+    self->extensions->message = self->message;
+  if (ForEachCompositeField(self, FixupMessageReference(self->message)) == -1)
+    return -1;
+
+  return 0;
+}
+
+// --- Globals:
+
+// Retrieve a C++ FieldDescriptor for a message attribute.
+// The C++ message must be valid.
+// TODO(amauryfa): This function should stay internal, because exception
+// handling is not consistent.
+static const FieldDescriptor* GetFieldDescriptor(
+    CMessage* self, PyObject* name) {
+  const Descriptor *message_descriptor = self->message->GetDescriptor();
+  char* field_name;
+  Py_ssize_t size;
+  if (PyString_AsStringAndSize(name, &field_name, &size) < 0) {
+    return NULL;
+  }
+  const FieldDescriptor *field_descriptor =
+      message_descriptor->FindFieldByName(string(field_name, size));
+  if (field_descriptor == NULL) {
+    // Note: No exception is set!
+    return NULL;
+  }
+  return field_descriptor;
+}
+
+// Retrieve a C++ FieldDescriptor for an extension handle.
+const FieldDescriptor* GetExtensionDescriptor(PyObject* extension) {
+  ScopedPyObjectPtr cdescriptor;
+  if (!PyObject_TypeCheck(extension, &PyFieldDescriptor_Type)) {
+    // Most callers consider extensions as a plain dictionary.  We should
+    // allow input which is not a field descriptor, and simply pretend it does
+    // not exist.
+    PyErr_SetObject(PyExc_KeyError, extension);
+    return NULL;
+  }
+  return PyFieldDescriptor_AsDescriptor(extension);
+}
+
+// If value is a string, convert it into an enum value based on the labels in
+// descriptor, otherwise simply return value.  Always returns a new reference.
+static PyObject* GetIntegerEnumValue(const FieldDescriptor& descriptor,
+                                     PyObject* value) {
+  if (PyString_Check(value) || PyUnicode_Check(value)) {
+    const EnumDescriptor* enum_descriptor = descriptor.enum_type();
+    if (enum_descriptor == NULL) {
+      PyErr_SetString(PyExc_TypeError, "not an enum field");
+      return NULL;
+    }
+    char* enum_label;
+    Py_ssize_t size;
+    if (PyString_AsStringAndSize(value, &enum_label, &size) < 0) {
+      return NULL;
+    }
+    const EnumValueDescriptor* enum_value_descriptor =
+        enum_descriptor->FindValueByName(string(enum_label, size));
+    if (enum_value_descriptor == NULL) {
+      PyErr_SetString(PyExc_ValueError, "unknown enum label");
+      return NULL;
+    }
+    return PyInt_FromLong(enum_value_descriptor->number());
+  }
+  Py_INCREF(value);
+  return value;
+}
+
+// If cmessage_list is not NULL, this function releases values into the
+// container CMessages instead of just removing. Repeated composite container
+// needs to do this to make sure CMessages stay alive if they're still
+// referenced after deletion. Repeated scalar container doesn't need to worry.
+int InternalDeleteRepeatedField(
+    CMessage* self,
+    const FieldDescriptor* field_descriptor,
+    PyObject* slice,
+    PyObject* cmessage_list) {
+  Message* message = self->message;
+  Py_ssize_t length, from, to, step, slice_length;
+  const Reflection* reflection = message->GetReflection();
+  int min, max;
+  length = reflection->FieldSize(*message, field_descriptor);
+
+  if (PyInt_Check(slice) || PyLong_Check(slice)) {
+    from = to = PyLong_AsLong(slice);
+    if (from < 0) {
+      from = to = length + from;
+    }
+    step = 1;
+    min = max = from;
+
+    // Range check.
+    if (from < 0 || from >= length) {
+      PyErr_Format(PyExc_IndexError, "list assignment index out of range");
+      return -1;
+    }
+  } else if (PySlice_Check(slice)) {
+    from = to = step = slice_length = 0;
+    PySlice_GetIndicesEx(
+#if PY_MAJOR_VERSION < 3
+        reinterpret_cast<PySliceObject*>(slice),
+#else
+        slice,
+#endif
+        length, &from, &to, &step, &slice_length);
+    if (from < to) {
+      min = from;
+      max = to - 1;
+    } else {
+      min = to + 1;
+      max = from;
+    }
+  } else {
+    PyErr_SetString(PyExc_TypeError, "list indices must be integers");
+    return -1;
+  }
+
+  Py_ssize_t i = from;
+  std::vector<bool> to_delete(length, false);
+  while (i >= min && i <= max) {
+    to_delete[i] = true;
+    i += step;
+  }
+
+  to = 0;
+  for (i = 0; i < length; ++i) {
+    if (!to_delete[i]) {
+      if (i != to) {
+        reflection->SwapElements(message, field_descriptor, i, to);
+        if (cmessage_list != NULL) {
+          // If a list of cmessages is passed in (i.e. from a repeated
+          // composite container), swap those as well to correspond to the
+          // swaps in the underlying message so they're in the right order
+          // when we start releasing.
+          PyObject* tmp = PyList_GET_ITEM(cmessage_list, i);
+          PyList_SET_ITEM(cmessage_list, i,
+                          PyList_GET_ITEM(cmessage_list, to));
+          PyList_SET_ITEM(cmessage_list, to, tmp);
+        }
+      }
+      ++to;
+    }
+  }
+
+  while (i > to) {
+    if (cmessage_list == NULL) {
+      reflection->RemoveLast(message, field_descriptor);
+    } else {
+      CMessage* last_cmessage = reinterpret_cast<CMessage*>(
+          PyList_GET_ITEM(cmessage_list, PyList_GET_SIZE(cmessage_list) - 1));
+      repeated_composite_container::ReleaseLastTo(
+          self, field_descriptor, last_cmessage);
+      if (PySequence_DelItem(cmessage_list, -1) < 0) {
+        return -1;
+      }
+    }
+    --i;
+  }
+
+  return 0;
+}
+
+// Initializes fields of a message. Used in constructors.
+int InitAttributes(CMessage* self, PyObject* kwargs) {
+  if (kwargs == NULL) {
+    return 0;
+  }
+
+  Py_ssize_t pos = 0;
+  PyObject* name;
+  PyObject* value;
+  while (PyDict_Next(kwargs, &pos, &name, &value)) {
+    if (!PyString_Check(name)) {
+      PyErr_SetString(PyExc_ValueError, "Field name must be a string");
+      return -1;
+    }
+    const FieldDescriptor* descriptor = GetFieldDescriptor(self, name);
+    if (descriptor == NULL) {
+      PyErr_Format(PyExc_ValueError, "Protocol message %s has no \"%s\" field.",
+                   self->message->GetDescriptor()->name().c_str(),
+                   PyString_AsString(name));
+      return -1;
+    }
+    if (descriptor->is_map()) {
+      ScopedPyObjectPtr map(GetAttr(self, name));
+      const FieldDescriptor* value_descriptor =
+          descriptor->message_type()->FindFieldByName("value");
+      if (value_descriptor->cpp_type() == FieldDescriptor::CPPTYPE_MESSAGE) {
+        Py_ssize_t map_pos = 0;
+        PyObject* map_key;
+        PyObject* map_value;
+        while (PyDict_Next(value, &map_pos, &map_key, &map_value)) {
+          ScopedPyObjectPtr function_return;
+          function_return.reset(PyObject_GetItem(map.get(), map_key));
+          if (function_return.get() == NULL) {
+            return -1;
+          }
+          ScopedPyObjectPtr ok(PyObject_CallMethod(
+              function_return.get(), "MergeFrom", "O", map_value));
+          if (ok.get() == NULL) {
+            return -1;
+          }
+        }
+      } else {
+        ScopedPyObjectPtr function_return;
+        function_return.reset(
+            PyObject_CallMethod(map.get(), "update", "O", value));
+        if (function_return.get() == NULL) {
+          return -1;
+        }
+      }
+    } else if (descriptor->label() == FieldDescriptor::LABEL_REPEATED) {
+      ScopedPyObjectPtr container(GetAttr(self, name));
+      if (container == NULL) {
+        return -1;
+      }
+      if (descriptor->cpp_type() == FieldDescriptor::CPPTYPE_MESSAGE) {
+        RepeatedCompositeContainer* rc_container =
+            reinterpret_cast<RepeatedCompositeContainer*>(container.get());
+        ScopedPyObjectPtr iter(PyObject_GetIter(value));
+        if (iter == NULL) {
+          PyErr_SetString(PyExc_TypeError, "Value must be iterable");
+          return -1;
+        }
+        ScopedPyObjectPtr next;
+        while ((next.reset(PyIter_Next(iter.get()))) != NULL) {
+          PyObject* kwargs = (PyDict_Check(next.get()) ? next.get() : NULL);
+          ScopedPyObjectPtr new_msg(
+              repeated_composite_container::Add(rc_container, NULL, kwargs));
+          if (new_msg == NULL) {
+            return -1;
+          }
+          if (kwargs == NULL) {
+            // next was not a dict, it's a message we need to merge
+            ScopedPyObjectPtr merged(MergeFrom(
+                reinterpret_cast<CMessage*>(new_msg.get()), next.get()));
+            if (merged.get() == NULL) {
+              return -1;
+            }
+          }
+        }
+        if (PyErr_Occurred()) {
+          // Check to see how PyIter_Next() exited.
+          return -1;
+        }
+      } else if (descriptor->cpp_type() == FieldDescriptor::CPPTYPE_ENUM) {
+        RepeatedScalarContainer* rs_container =
+            reinterpret_cast<RepeatedScalarContainer*>(container.get());
+        ScopedPyObjectPtr iter(PyObject_GetIter(value));
+        if (iter == NULL) {
+          PyErr_SetString(PyExc_TypeError, "Value must be iterable");
+          return -1;
+        }
+        ScopedPyObjectPtr next;
+        while ((next.reset(PyIter_Next(iter.get()))) != NULL) {
+          ScopedPyObjectPtr enum_value(
+              GetIntegerEnumValue(*descriptor, next.get()));
+          if (enum_value == NULL) {
+            return -1;
+          }
+          ScopedPyObjectPtr new_msg(repeated_scalar_container::Append(
+              rs_container, enum_value.get()));
+          if (new_msg == NULL) {
+            return -1;
+          }
+        }
+        if (PyErr_Occurred()) {
+          // Check to see how PyIter_Next() exited.
+          return -1;
+        }
+      } else {
+        if (ScopedPyObjectPtr(repeated_scalar_container::Extend(
+                reinterpret_cast<RepeatedScalarContainer*>(container.get()),
+                value)) ==
+            NULL) {
+          return -1;
+        }
+      }
+    } else if (descriptor->cpp_type() == FieldDescriptor::CPPTYPE_MESSAGE) {
+      ScopedPyObjectPtr message(GetAttr(self, name));
+      if (message == NULL) {
+        return -1;
+      }
+      CMessage* cmessage = reinterpret_cast<CMessage*>(message.get());
+      if (PyDict_Check(value)) {
+        if (InitAttributes(cmessage, value) < 0) {
+          return -1;
+        }
+      } else {
+        ScopedPyObjectPtr merged(MergeFrom(cmessage, value));
+        if (merged == NULL) {
+          return -1;
+        }
+      }
+    } else {
+      ScopedPyObjectPtr new_val;
+      if (descriptor->cpp_type() == FieldDescriptor::CPPTYPE_ENUM) {
+        new_val.reset(GetIntegerEnumValue(*descriptor, value));
+        if (new_val == NULL) {
+          return -1;
+        }
+      }
+      if (SetAttr(self, name, (new_val.get() == NULL) ? value : new_val.get()) <
+          0) {
+        return -1;
+      }
+    }
+  }
+  return 0;
+}
+
+// Allocates an incomplete Python Message: the caller must fill self->message,
+// self->owner and eventually self->parent.
+CMessage* NewEmptyMessage(PyObject* type, const Descriptor *descriptor) {
+  CMessage* self = reinterpret_cast<CMessage*>(
+      PyType_GenericAlloc(reinterpret_cast<PyTypeObject*>(type), 0));
+  if (self == NULL) {
+    return NULL;
+  }
+
+  self->message = NULL;
+  self->parent = NULL;
+  self->parent_field_descriptor = NULL;
+  self->read_only = false;
+  self->extensions = NULL;
+
+  self->composite_fields = NULL;
+
+  return self;
+}
+
+// The __new__ method of Message classes.
+// Creates a new C++ message and takes ownership.
+static PyObject* New(PyTypeObject* cls,
+                     PyObject* unused_args, PyObject* unused_kwargs) {
+  PyMessageMeta* type = CheckMessageClass(cls);
+  if (type == NULL) {
+    return NULL;
+  }
+  // Retrieve the message descriptor and the default instance (=prototype).
+  const Descriptor* message_descriptor = type->message_descriptor;
+  if (message_descriptor == NULL) {
+    return NULL;
+  }
+  const Message* default_message = type->py_descriptor_pool->message_factory
+                                   ->GetPrototype(message_descriptor);
+  if (default_message == NULL) {
+    PyErr_SetString(PyExc_TypeError, message_descriptor->full_name().c_str());
+    return NULL;
+  }
+
+  CMessage* self = NewEmptyMessage(reinterpret_cast<PyObject*>(type),
+                                   message_descriptor);
+  if (self == NULL) {
+    return NULL;
+  }
+  self->message = default_message->New();
+  self->owner.reset(self->message);
+  return reinterpret_cast<PyObject*>(self);
+}
+
+// The __init__ method of Message classes.
+// It initializes fields from keywords passed to the constructor.
+static int Init(CMessage* self, PyObject* args, PyObject* kwargs) {
+  if (PyTuple_Size(args) != 0) {
+    PyErr_SetString(PyExc_TypeError, "No positional arguments allowed");
+    return -1;
+  }
+
+  return InitAttributes(self, kwargs);
+}
+
+// ---------------------------------------------------------------------
+// Deallocating a CMessage
+//
+// Deallocating a CMessage requires that we clear any weak references
+// from children to the message being deallocated.
+
+// Clear the weak reference from the child to the parent.
+struct ClearWeakReferences : public ChildVisitor {
+  int VisitRepeatedCompositeContainer(RepeatedCompositeContainer* container) {
+    container->parent = NULL;
+    // The elements in the container have the same parent as the
+    // container itself, so NULL out that pointer as well.
+    const Py_ssize_t n = PyList_GET_SIZE(container->child_messages);
+    for (Py_ssize_t i = 0; i < n; ++i) {
+      CMessage* child_cmessage = reinterpret_cast<CMessage*>(
+          PyList_GET_ITEM(container->child_messages, i));
+      child_cmessage->parent = NULL;
+    }
+    return 0;
+  }
+
+  int VisitRepeatedScalarContainer(RepeatedScalarContainer* container) {
+    container->parent = NULL;
+    return 0;
+  }
+
+  int VisitMapContainer(MapContainer* container) {
+    container->parent = NULL;
+    return 0;
+  }
+
+  int VisitCMessage(CMessage* cmessage,
+                    const FieldDescriptor* field_descriptor) {
+    cmessage->parent = NULL;
+    return 0;
+  }
+};
+
+static void Dealloc(CMessage* self) {
+  // Null out all weak references from children to this message.
+  GOOGLE_CHECK_EQ(0, ForEachCompositeField(self, ClearWeakReferences()));
+  if (self->extensions) {
+    self->extensions->parent = NULL;
+  }
+
+  Py_CLEAR(self->extensions);
+  Py_CLEAR(self->composite_fields);
+  self->owner.reset();
+  Py_TYPE(self)->tp_free(reinterpret_cast<PyObject*>(self));
+}
+
+// ---------------------------------------------------------------------
+
+
+PyObject* IsInitialized(CMessage* self, PyObject* args) {
+  PyObject* errors = NULL;
+  if (PyArg_ParseTuple(args, "|O", &errors) < 0) {
+    return NULL;
+  }
+  if (self->message->IsInitialized()) {
+    Py_RETURN_TRUE;
+  }
+  if (errors != NULL) {
+    ScopedPyObjectPtr initialization_errors(
+        FindInitializationErrors(self));
+    if (initialization_errors == NULL) {
+      return NULL;
+    }
+    ScopedPyObjectPtr extend_name(PyString_FromString("extend"));
+    if (extend_name == NULL) {
+      return NULL;
+    }
+    ScopedPyObjectPtr result(PyObject_CallMethodObjArgs(
+        errors,
+        extend_name.get(),
+        initialization_errors.get(),
+        NULL));
+    if (result == NULL) {
+      return NULL;
+    }
+  }
+  Py_RETURN_FALSE;
+}
+
+PyObject* HasFieldByDescriptor(
+    CMessage* self, const FieldDescriptor* field_descriptor) {
+  Message* message = self->message;
+  if (!CheckFieldBelongsToMessage(field_descriptor, message)) {
+    return NULL;
+  }
+  if (field_descriptor->label() == FieldDescriptor::LABEL_REPEATED) {
+    PyErr_SetString(PyExc_KeyError,
+                    "Field is repeated. A singular method is required.");
+    return NULL;
+  }
+  bool has_field =
+      message->GetReflection()->HasField(*message, field_descriptor);
+  return PyBool_FromLong(has_field ? 1 : 0);
+}
+
+const FieldDescriptor* FindFieldWithOneofs(
+    const Message* message, const string& field_name, bool* in_oneof) {
+  *in_oneof = false;
+  const Descriptor* descriptor = message->GetDescriptor();
+  const FieldDescriptor* field_descriptor =
+      descriptor->FindFieldByName(field_name);
+  if (field_descriptor != NULL) {
+    return field_descriptor;
+  }
+  const OneofDescriptor* oneof_desc =
+      descriptor->FindOneofByName(field_name);
+  if (oneof_desc != NULL) {
+    *in_oneof = true;
+    return message->GetReflection()->GetOneofFieldDescriptor(*message,
+                                                             oneof_desc);
+  }
+  return NULL;
+}
+
+bool CheckHasPresence(const FieldDescriptor* field_descriptor, bool in_oneof) {
+  if (field_descriptor->label() == FieldDescriptor::LABEL_REPEATED) {
+    PyErr_Format(PyExc_ValueError,
+                 "Protocol message has no singular \"%s\" field.",
+                 field_descriptor->name().c_str());
+    return false;
+  }
+
+  if (field_descriptor->file()->syntax() == FileDescriptor::SYNTAX_PROTO3) {
+    // HasField() for a oneof *itself* isn't supported.
+    if (in_oneof) {
+      PyErr_Format(PyExc_ValueError,
+                   "Can't test oneof field \"%s\" for presence in proto3, use "
+                   "WhichOneof instead.",
+                   field_descriptor->containing_oneof()->name().c_str());
+      return false;
+    }
+
+    // ...but HasField() for fields *in* a oneof is supported.
+    if (field_descriptor->containing_oneof() != NULL) {
+      return true;
+    }
+
+    if (field_descriptor->cpp_type() != FieldDescriptor::CPPTYPE_MESSAGE) {
+      PyErr_Format(
+          PyExc_ValueError,
+          "Can't test non-submessage field \"%s\" for presence in proto3.",
+          field_descriptor->name().c_str());
+      return false;
+    }
+  }
+
+  return true;
+}
+
+PyObject* HasField(CMessage* self, PyObject* arg) {
+  char* field_name;
+  Py_ssize_t size;
+#if PY_MAJOR_VERSION < 3
+  if (PyString_AsStringAndSize(arg, &field_name, &size) < 0) {
+    return NULL;
+  }
+#else
+  field_name = PyUnicode_AsUTF8AndSize(arg, &size);
+  if (!field_name) {
+    return NULL;
+  }
+#endif
+
+  Message* message = self->message;
+  bool is_in_oneof;
+  const FieldDescriptor* field_descriptor =
+      FindFieldWithOneofs(message, string(field_name, size), &is_in_oneof);
+  if (field_descriptor == NULL) {
+    if (!is_in_oneof) {
+      PyErr_Format(PyExc_ValueError, "Unknown field %s.", field_name);
+      return NULL;
+    } else {
+      Py_RETURN_FALSE;
+    }
+  }
+
+  if (!CheckHasPresence(field_descriptor, is_in_oneof)) {
+    return NULL;
+  }
+
+  if (message->GetReflection()->HasField(*message, field_descriptor)) {
+    Py_RETURN_TRUE;
+  }
+  if (!message->GetReflection()->SupportsUnknownEnumValues() &&
+      field_descriptor->cpp_type() == FieldDescriptor::CPPTYPE_ENUM) {
+    // Special case: Python HasField() differs in semantics from C++
+    // slightly: we return HasField('enum_field') == true if there is
+    // an unknown enum value present. To implement this we have to
+    // look in the UnknownFieldSet.
+    const UnknownFieldSet& unknown_field_set =
+        message->GetReflection()->GetUnknownFields(*message);
+    for (int i = 0; i < unknown_field_set.field_count(); ++i) {
+      if (unknown_field_set.field(i).number() == field_descriptor->number()) {
+        Py_RETURN_TRUE;
+      }
+    }
+  }
+  Py_RETURN_FALSE;
+}
+
+PyObject* ClearExtension(CMessage* self, PyObject* extension) {
+  if (self->extensions != NULL) {
+    return extension_dict::ClearExtension(self->extensions, extension);
+  } else {
+    const FieldDescriptor* descriptor = GetExtensionDescriptor(extension);
+    if (descriptor == NULL) {
+      return NULL;
+    }
+    if (ScopedPyObjectPtr(ClearFieldByDescriptor(self, descriptor)) == NULL) {
+      return NULL;
+    }
+  }
+  Py_RETURN_NONE;
+}
+
+PyObject* HasExtension(CMessage* self, PyObject* extension) {
+  const FieldDescriptor* descriptor = GetExtensionDescriptor(extension);
+  if (descriptor == NULL) {
+    return NULL;
+  }
+  return HasFieldByDescriptor(self, descriptor);
+}
+
+// ---------------------------------------------------------------------
+// Releasing messages
+//
+// The Python API's ClearField() and Clear() methods behave
+// differently than their C++ counterparts.  While the C++ versions
+// clears the children the Python versions detaches the children,
+// without touching their content.  This impedance mismatch causes
+// some complexity in the implementation, which is captured in this
+// section.
+//
+// When a CMessage field is cleared we need to:
+//
+// * Release the Message used as the backing store for the CMessage
+//   from its parent.
+//
+// * Change the owner field of the released CMessage and all of its
+//   children to point to the newly released Message.
+//
+// * Clear the weak references from the released CMessage to the
+//   parent.
+//
+// When a RepeatedCompositeContainer field is cleared we need to:
+//
+// * Release all the Message used as the backing store for the
+//   CMessages stored in the container.
+//
+// * Change the owner field of all the released CMessage and all of
+//   their children to point to the newly released Messages.
+//
+// * Clear the weak references from the released container to the
+//   parent.
+
+struct SetOwnerVisitor : public ChildVisitor {
+  // new_owner must outlive this object.
+  explicit SetOwnerVisitor(const shared_ptr<Message>& new_owner)
+      : new_owner_(new_owner) {}
+
+  int VisitRepeatedCompositeContainer(RepeatedCompositeContainer* container) {
+    repeated_composite_container::SetOwner(container, new_owner_);
+    return 0;
+  }
+
+  int VisitRepeatedScalarContainer(RepeatedScalarContainer* container) {
+    repeated_scalar_container::SetOwner(container, new_owner_);
+    return 0;
+  }
+
+  int VisitMapContainer(MapContainer* container) {
+    container->SetOwner(new_owner_);
+    return 0;
+  }
+
+  int VisitCMessage(CMessage* cmessage,
+                    const FieldDescriptor* field_descriptor) {
+    return SetOwner(cmessage, new_owner_);
+  }
+
+ private:
+  const shared_ptr<Message>& new_owner_;
+};
+
+// Change the owner of this CMessage and all its children, recursively.
+int SetOwner(CMessage* self, const shared_ptr<Message>& new_owner) {
+  self->owner = new_owner;
+  if (ForEachCompositeField(self, SetOwnerVisitor(new_owner)) == -1)
+    return -1;
+  return 0;
+}
+
+// Releases the message specified by 'field' and returns the
+// pointer. If the field does not exist a new message is created using
+// 'descriptor'. The caller takes ownership of the returned pointer.
+Message* ReleaseMessage(CMessage* self,
+                        const Descriptor* descriptor,
+                        const FieldDescriptor* field_descriptor) {
+  MessageFactory* message_factory = GetFactoryForMessage(self);
+  Message* released_message = self->message->GetReflection()->ReleaseMessage(
+      self->message, field_descriptor, message_factory);
+  // ReleaseMessage will return NULL which differs from
+  // child_cmessage->message, if the field does not exist.  In this case,
+  // the latter points to the default instance via a const_cast<>, so we
+  // have to reset it to a new mutable object since we are taking ownership.
+  if (released_message == NULL) {
+    const Message* prototype = message_factory->GetPrototype(descriptor);
+    GOOGLE_DCHECK(prototype != NULL);
+    released_message = prototype->New();
+  }
+
+  return released_message;
+}
+
+int ReleaseSubMessage(CMessage* self,
+                      const FieldDescriptor* field_descriptor,
+                      CMessage* child_cmessage) {
+  // Release the Message
+  shared_ptr<Message> released_message(ReleaseMessage(
+      self, child_cmessage->message->GetDescriptor(), field_descriptor));
+  child_cmessage->message = released_message.get();
+  child_cmessage->owner.swap(released_message);
+  child_cmessage->parent = NULL;
+  child_cmessage->parent_field_descriptor = NULL;
+  child_cmessage->read_only = false;
+  return ForEachCompositeField(child_cmessage,
+                               SetOwnerVisitor(child_cmessage->owner));
+}
+
+struct ReleaseChild : public ChildVisitor {
+  // message must outlive this object.
+  explicit ReleaseChild(CMessage* parent) :
+      parent_(parent) {}
+
+  int VisitRepeatedCompositeContainer(RepeatedCompositeContainer* container) {
+    return repeated_composite_container::Release(
+        reinterpret_cast<RepeatedCompositeContainer*>(container));
+  }
+
+  int VisitRepeatedScalarContainer(RepeatedScalarContainer* container) {
+    return repeated_scalar_container::Release(
+        reinterpret_cast<RepeatedScalarContainer*>(container));
+  }
+
+  int VisitMapContainer(MapContainer* container) {
+    return reinterpret_cast<MapContainer*>(container)->Release();
+  }
+
+  int VisitCMessage(CMessage* cmessage,
+                    const FieldDescriptor* field_descriptor) {
+    return ReleaseSubMessage(parent_, field_descriptor,
+        reinterpret_cast<CMessage*>(cmessage));
+  }
+
+  CMessage* parent_;
+};
+
+int InternalReleaseFieldByDescriptor(
+    CMessage* self,
+    const FieldDescriptor* field_descriptor,
+    PyObject* composite_field) {
+  return VisitCompositeField(
+      field_descriptor,
+      composite_field,
+      ReleaseChild(self));
+}
+
+PyObject* ClearFieldByDescriptor(
+    CMessage* self,
+    const FieldDescriptor* descriptor) {
+  if (!CheckFieldBelongsToMessage(descriptor, self->message)) {
+    return NULL;
+  }
+  AssureWritable(self);
+  self->message->GetReflection()->ClearField(self->message, descriptor);
+  Py_RETURN_NONE;
+}
+
+PyObject* ClearField(CMessage* self, PyObject* arg) {
+  if (!PyString_Check(arg)) {
+    PyErr_SetString(PyExc_TypeError, "field name must be a string");
+    return NULL;
+  }
+#if PY_MAJOR_VERSION < 3
+  const char* field_name = PyString_AS_STRING(arg);
+  Py_ssize_t size = PyString_GET_SIZE(arg);
+#else
+  Py_ssize_t size;
+  const char* field_name = PyUnicode_AsUTF8AndSize(arg, &size);
+#endif
+  AssureWritable(self);
+  Message* message = self->message;
+  ScopedPyObjectPtr arg_in_oneof;
+  bool is_in_oneof;
+  const FieldDescriptor* field_descriptor =
+      FindFieldWithOneofs(message, string(field_name, size), &is_in_oneof);
+  if (field_descriptor == NULL) {
+    if (!is_in_oneof) {
+      PyErr_Format(PyExc_ValueError,
+                   "Protocol message has no \"%s\" field.", field_name);
+      return NULL;
+    } else {
+      Py_RETURN_NONE;
+    }
+  } else if (is_in_oneof) {
+    const string& name = field_descriptor->name();
+    arg_in_oneof.reset(PyString_FromStringAndSize(name.c_str(), name.size()));
+    arg = arg_in_oneof.get();
+  }
+
+  PyObject* composite_field = self->composite_fields ?
+      PyDict_GetItem(self->composite_fields, arg) : NULL;
+
+  // Only release the field if there's a possibility that there are
+  // references to it.
+  if (composite_field != NULL) {
+    if (InternalReleaseFieldByDescriptor(self, field_descriptor,
+                                         composite_field) < 0) {
+      return NULL;
+    }
+    PyDict_DelItem(self->composite_fields, arg);
+  }
+  message->GetReflection()->ClearField(message, field_descriptor);
+  if (field_descriptor->cpp_type() == FieldDescriptor::CPPTYPE_ENUM &&
+      !message->GetReflection()->SupportsUnknownEnumValues()) {
+    UnknownFieldSet* unknown_field_set =
+        message->GetReflection()->MutableUnknownFields(message);
+    unknown_field_set->DeleteByNumber(field_descriptor->number());
+  }
+
+  Py_RETURN_NONE;
+}
+
+PyObject* Clear(CMessage* self) {
+  AssureWritable(self);
+  if (ForEachCompositeField(self, ReleaseChild(self)) == -1)
+    return NULL;
+  Py_CLEAR(self->extensions);
+  if (self->composite_fields) {
+    PyDict_Clear(self->composite_fields);
+  }
+  self->message->Clear();
+  Py_RETURN_NONE;
+}
+
+// ---------------------------------------------------------------------
+
+static string GetMessageName(CMessage* self) {
+  if (self->parent_field_descriptor != NULL) {
+    return self->parent_field_descriptor->full_name();
+  } else {
+    return self->message->GetDescriptor()->full_name();
+  }
+}
+
+static PyObject* SerializeToString(CMessage* self, PyObject* args) {
+  if (!self->message->IsInitialized()) {
+    ScopedPyObjectPtr errors(FindInitializationErrors(self));
+    if (errors == NULL) {
+      return NULL;
+    }
+    ScopedPyObjectPtr comma(PyString_FromString(","));
+    if (comma == NULL) {
+      return NULL;
+    }
+    ScopedPyObjectPtr joined(
+        PyObject_CallMethod(comma.get(), "join", "O", errors.get()));
+    if (joined == NULL) {
+      return NULL;
+    }
+
+    // TODO(haberman): this is a (hopefully temporary) hack.  The unit testing
+    // infrastructure reloads all pure-Python modules for every test, but not
+    // C++ modules (because that's generally impossible:
+    // http://bugs.python.org/issue1144263).  But if we cache EncodeError, we'll
+    // return the EncodeError from a previous load of the module, which won't
+    // match a user's attempt to catch EncodeError.  So we have to look it up
+    // again every time.
+    ScopedPyObjectPtr message_module(PyImport_ImportModule(
+        "google.protobuf.message"));
+    if (message_module.get() == NULL) {
+      return NULL;
+    }
+
+    ScopedPyObjectPtr encode_error(
+        PyObject_GetAttrString(message_module.get(), "EncodeError"));
+    if (encode_error.get() == NULL) {
+      return NULL;
+    }
+    PyErr_Format(encode_error.get(),
+                 "Message %s is missing required fields: %s",
+                 GetMessageName(self).c_str(), PyString_AsString(joined.get()));
+    return NULL;
+  }
+  int size = self->message->ByteSize();
+  if (size <= 0) {
+    return PyBytes_FromString("");
+  }
+  PyObject* result = PyBytes_FromStringAndSize(NULL, size);
+  if (result == NULL) {
+    return NULL;
+  }
+  char* buffer = PyBytes_AS_STRING(result);
+  self->message->SerializeWithCachedSizesToArray(
+      reinterpret_cast<uint8*>(buffer));
+  return result;
+}
+
+static PyObject* SerializePartialToString(CMessage* self) {
+  string contents;
+  self->message->SerializePartialToString(&contents);
+  return PyBytes_FromStringAndSize(contents.c_str(), contents.size());
+}
+
+// Formats proto fields for ascii dumps using python formatting functions where
+// appropriate.
+class PythonFieldValuePrinter : public TextFormat::FieldValuePrinter {
+ public:
+  // Python has some differences from C++ when printing floating point numbers.
+  //
+  // 1) Trailing .0 is always printed.
+  // 2) (Python2) Output is rounded to 12 digits.
+  // 3) (Python3) The full precision of the double is preserved (and Python uses
+  //    David M. Gay's dtoa(), when the C++ code uses SimpleDtoa. There are some
+  //    differences, but they rarely happen)
+  //
+  // We override floating point printing with the C-API function for printing
+  // Python floats to ensure consistency.
+  string PrintFloat(float value) const { return PrintDouble(value); }
+  string PrintDouble(double value) const {
+    // This implementation is not highly optimized (it allocates two temporary
+    // Python objects) but it is simple and portable.  If this is shown to be a
+    // performance bottleneck, we can optimize it, but the results will likely
+    // be more complicated to accommodate the differing behavior of double
+    // formatting between Python 2 and Python 3.
+    //
+    // (Though a valid question is: do we really want to make out output
+    // dependent on the Python version?)
+    ScopedPyObjectPtr py_value(PyFloat_FromDouble(value));
+    if (!py_value.get()) {
+      return string();
+    }
+
+    ScopedPyObjectPtr py_str(PyObject_Str(py_value.get()));
+    if (!py_str.get()) {
+      return string();
+    }
+
+    return string(PyString_AsString(py_str.get()));
+  }
+};
+
+static PyObject* ToStr(CMessage* self) {
+  TextFormat::Printer printer;
+  // Passes ownership
+  printer.SetDefaultFieldValuePrinter(new PythonFieldValuePrinter());
+  printer.SetHideUnknownFields(true);
+  string output;
+  if (!printer.PrintToString(*self->message, &output)) {
+    PyErr_SetString(PyExc_ValueError, "Unable to convert message to str");
+    return NULL;
+  }
+  return PyString_FromString(output.c_str());
+}
+
+PyObject* MergeFrom(CMessage* self, PyObject* arg) {
+  CMessage* other_message;
+  if (!PyObject_TypeCheck(reinterpret_cast<PyObject *>(arg), &CMessage_Type)) {
+    PyErr_SetString(PyExc_TypeError, "Must be a message");
+    return NULL;
+  }
+
+  other_message = reinterpret_cast<CMessage*>(arg);
+  if (other_message->message->GetDescriptor() !=
+      self->message->GetDescriptor()) {
+    PyErr_Format(PyExc_TypeError,
+                 "Tried to merge from a message with a different type. "
+                 "to: %s, from: %s",
+                 self->message->GetDescriptor()->full_name().c_str(),
+                 other_message->message->GetDescriptor()->full_name().c_str());
+    return NULL;
+  }
+  AssureWritable(self);
+
+  // TODO(tibell): Message::MergeFrom might turn some child Messages
+  // into mutable messages, invalidating the message field in the
+  // corresponding CMessages.  We should run a FixupMessageReferences
+  // pass here.
+
+  self->message->MergeFrom(*other_message->message);
+  Py_RETURN_NONE;
+}
+
+static PyObject* CopyFrom(CMessage* self, PyObject* arg) {
+  CMessage* other_message;
+  if (!PyObject_TypeCheck(reinterpret_cast<PyObject *>(arg), &CMessage_Type)) {
+    PyErr_SetString(PyExc_TypeError, "Must be a message");
+    return NULL;
+  }
+
+  other_message = reinterpret_cast<CMessage*>(arg);
+
+  if (self == other_message) {
+    Py_RETURN_NONE;
+  }
+
+  if (other_message->message->GetDescriptor() !=
+      self->message->GetDescriptor()) {
+    PyErr_Format(PyExc_TypeError,
+                 "Tried to copy from a message with a different type. "
+                 "to: %s, from: %s",
+                 self->message->GetDescriptor()->full_name().c_str(),
+                 other_message->message->GetDescriptor()->full_name().c_str());
+    return NULL;
+  }
+
+  AssureWritable(self);
+
+  // CopyFrom on the message will not clean up self->composite_fields,
+  // which can leave us in an inconsistent state, so clear it out here.
+  (void)ScopedPyObjectPtr(Clear(self));
+
+  self->message->CopyFrom(*other_message->message);
+
+  Py_RETURN_NONE;
+}
+
+static PyObject* MergeFromString(CMessage* self, PyObject* arg) {
+  const void* data;
+  Py_ssize_t data_length;
+  if (PyObject_AsReadBuffer(arg, &data, &data_length) < 0) {
+    return NULL;
+  }
+
+  AssureWritable(self);
+  io::CodedInputStream input(
+      reinterpret_cast<const uint8*>(data), data_length);
+#if PROTOBUF_PYTHON_ALLOW_OVERSIZE_PROTOS
+  // Protobuf has a 64MB limit built in, this code will override this. Please do
+  // not enable this unless you fully understand the implications: protobufs
+  // must all be kept in memory at the same time, so if they grow too big you
+  // may get OOM errors. The protobuf APIs do not provide any tools for
+  // processing protobufs in chunks.  If you have protos this big you should
+  // break them up if it is at all convenient to do so.
+  input.SetTotalBytesLimit(INT_MAX, INT_MAX);
+#endif  // PROTOBUF_PYTHON_ALLOW_OVERSIZE_PROTOS
+  PyDescriptorPool* pool = GetDescriptorPoolForMessage(self);
+  input.SetExtensionRegistry(pool->pool, pool->message_factory);
+  bool success = self->message->MergePartialFromCodedStream(&input);
+  if (success) {
+    return PyInt_FromLong(input.CurrentPosition());
+  } else {
+    PyErr_Format(DecodeError_class, "Error parsing message");
+    return NULL;
+  }
+}
+
+static PyObject* ParseFromString(CMessage* self, PyObject* arg) {
+  if (ScopedPyObjectPtr(Clear(self)) == NULL) {
+    return NULL;
+  }
+  return MergeFromString(self, arg);
+}
+
+static PyObject* ByteSize(CMessage* self, PyObject* args) {
+  return PyLong_FromLong(self->message->ByteSize());
+}
+
+static PyObject* RegisterExtension(PyObject* cls,
+                                   PyObject* extension_handle) {
+  const FieldDescriptor* descriptor =
+      GetExtensionDescriptor(extension_handle);
+  if (descriptor == NULL) {
+    return NULL;
+  }
+
+  ScopedPyObjectPtr extensions_by_name(
+      PyObject_GetAttr(cls, k_extensions_by_name));
+  if (extensions_by_name == NULL) {
+    PyErr_SetString(PyExc_TypeError, "no extensions_by_name on class");
+    return NULL;
+  }
+  ScopedPyObjectPtr full_name(PyObject_GetAttr(extension_handle, kfull_name));
+  if (full_name == NULL) {
+    return NULL;
+  }
+
+  // If the extension was already registered, check that it is the same.
+  PyObject* existing_extension =
+      PyDict_GetItem(extensions_by_name.get(), full_name.get());
+  if (existing_extension != NULL) {
+    const FieldDescriptor* existing_extension_descriptor =
+        GetExtensionDescriptor(existing_extension);
+    if (existing_extension_descriptor != descriptor) {
+      PyErr_SetString(PyExc_ValueError, "Double registration of Extensions");
+      return NULL;
+    }
+    // Nothing else to do.
+    Py_RETURN_NONE;
+  }
+
+  if (PyDict_SetItem(extensions_by_name.get(), full_name.get(),
+                     extension_handle) < 0) {
+    return NULL;
+  }
+
+  // Also store a mapping from extension number to implementing class.
+  ScopedPyObjectPtr extensions_by_number(
+      PyObject_GetAttr(cls, k_extensions_by_number));
+  if (extensions_by_number == NULL) {
+    PyErr_SetString(PyExc_TypeError, "no extensions_by_number on class");
+    return NULL;
+  }
+  ScopedPyObjectPtr number(PyObject_GetAttrString(extension_handle, "number"));
+  if (number == NULL) {
+    return NULL;
+  }
+  if (PyDict_SetItem(extensions_by_number.get(), number.get(),
+                     extension_handle) < 0) {
+    return NULL;
+  }
+
+  // Check if it's a message set
+  if (descriptor->is_extension() &&
+      descriptor->containing_type()->options().message_set_wire_format() &&
+      descriptor->type() == FieldDescriptor::TYPE_MESSAGE &&
+      descriptor->label() == FieldDescriptor::LABEL_OPTIONAL) {
+    ScopedPyObjectPtr message_name(PyString_FromStringAndSize(
+        descriptor->message_type()->full_name().c_str(),
+        descriptor->message_type()->full_name().size()));
+    if (message_name == NULL) {
+      return NULL;
+    }
+    PyDict_SetItem(extensions_by_name.get(), message_name.get(),
+                   extension_handle);
+  }
+
+  Py_RETURN_NONE;
+}
+
+static PyObject* SetInParent(CMessage* self, PyObject* args) {
+  AssureWritable(self);
+  Py_RETURN_NONE;
+}
+
+static PyObject* WhichOneof(CMessage* self, PyObject* arg) {
+  Py_ssize_t name_size;
+  char *name_data;
+  if (PyString_AsStringAndSize(arg, &name_data, &name_size) < 0)
+    return NULL;
+  string oneof_name = string(name_data, name_size);
+  const OneofDescriptor* oneof_desc =
+      self->message->GetDescriptor()->FindOneofByName(oneof_name);
+  if (oneof_desc == NULL) {
+    PyErr_Format(PyExc_ValueError,
+                 "Protocol message has no oneof \"%s\" field.",
+                 oneof_name.c_str());
+    return NULL;
+  }
+  const FieldDescriptor* field_in_oneof =
+      self->message->GetReflection()->GetOneofFieldDescriptor(
+          *self->message, oneof_desc);
+  if (field_in_oneof == NULL) {
+    Py_RETURN_NONE;
+  } else {
+    const string& name = field_in_oneof->name();
+    return PyString_FromStringAndSize(name.c_str(), name.size());
+  }
+}
+
+static PyObject* GetExtensionDict(CMessage* self, void *closure);
+
+static PyObject* ListFields(CMessage* self) {
+  vector<const FieldDescriptor*> fields;
+  self->message->GetReflection()->ListFields(*self->message, &fields);
+
+  // Normally, the list will be exactly the size of the fields.
+  ScopedPyObjectPtr all_fields(PyList_New(fields.size()));
+  if (all_fields == NULL) {
+    return NULL;
+  }
+
+  // When there are unknown extensions, the py list will *not* contain
+  // the field information.  Thus the actual size of the py list will be
+  // smaller than the size of fields.  Set the actual size at the end.
+  Py_ssize_t actual_size = 0;
+  for (size_t i = 0; i < fields.size(); ++i) {
+    ScopedPyObjectPtr t(PyTuple_New(2));
+    if (t == NULL) {
+      return NULL;
+    }
+
+    if (fields[i]->is_extension()) {
+      ScopedPyObjectPtr extension_field(
+          PyFieldDescriptor_FromDescriptor(fields[i]));
+      if (extension_field == NULL) {
+        return NULL;
+      }
+      // With C++ descriptors, the field can always be retrieved, but for
+      // unknown extensions which have not been imported in Python code, there
+      // is no message class and we cannot retrieve the value.
+      // TODO(amauryfa): consider building the class on the fly!
+      if (fields[i]->message_type() != NULL &&
+          cdescriptor_pool::GetMessageClass(
+              GetDescriptorPoolForMessage(self),
+              fields[i]->message_type()) == NULL) {
+        PyErr_Clear();
+        continue;
+      }
+      ScopedPyObjectPtr extensions(GetExtensionDict(self, NULL));
+      if (extensions == NULL) {
+        return NULL;
+      }
+      // 'extension' reference later stolen by PyTuple_SET_ITEM.
+      PyObject* extension = PyObject_GetItem(
+          extensions.get(), extension_field.get());
+      if (extension == NULL) {
+        return NULL;
+      }
+      PyTuple_SET_ITEM(t.get(), 0, extension_field.release());
+      // Steals reference to 'extension'
+      PyTuple_SET_ITEM(t.get(), 1, extension);
+    } else {
+      // Normal field
+      const string& field_name = fields[i]->name();
+      ScopedPyObjectPtr py_field_name(PyString_FromStringAndSize(
+          field_name.c_str(), field_name.length()));
+      if (py_field_name == NULL) {
+        PyErr_SetString(PyExc_ValueError, "bad string");
+        return NULL;
+      }
+      ScopedPyObjectPtr field_descriptor(
+          PyFieldDescriptor_FromDescriptor(fields[i]));
+      if (field_descriptor == NULL) {
+        return NULL;
+      }
+
+      PyObject* field_value = GetAttr(self, py_field_name.get());
+      if (field_value == NULL) {
+        PyErr_SetObject(PyExc_ValueError, py_field_name.get());
+        return NULL;
+      }
+      PyTuple_SET_ITEM(t.get(), 0, field_descriptor.release());
+      PyTuple_SET_ITEM(t.get(), 1, field_value);
+    }
+    PyList_SET_ITEM(all_fields.get(), actual_size, t.release());
+    ++actual_size;
+  }
+  Py_SIZE(all_fields.get()) = actual_size;
+  return all_fields.release();
+}
+
+PyObject* FindInitializationErrors(CMessage* self) {
+  Message* message = self->message;
+  vector<string> errors;
+  message->FindInitializationErrors(&errors);
+
+  PyObject* error_list = PyList_New(errors.size());
+  if (error_list == NULL) {
+    return NULL;
+  }
+  for (size_t i = 0; i < errors.size(); ++i) {
+    const string& error = errors[i];
+    PyObject* error_string = PyString_FromStringAndSize(
+        error.c_str(), error.length());
+    if (error_string == NULL) {
+      Py_DECREF(error_list);
+      return NULL;
+    }
+    PyList_SET_ITEM(error_list, i, error_string);
+  }
+  return error_list;
+}
+
+static PyObject* RichCompare(CMessage* self, PyObject* other, int opid) {
+  // Only equality comparisons are implemented.
+  if (opid != Py_EQ && opid != Py_NE) {
+    Py_INCREF(Py_NotImplemented);
+    return Py_NotImplemented;
+  }
+  bool equals = true;
+  // If other is not a message, it cannot be equal.
+  if (!PyObject_TypeCheck(other, &CMessage_Type)) {
+    equals = false;
+  }
+  const google::protobuf::Message* other_message =
+      reinterpret_cast<CMessage*>(other)->message;
+  // If messages don't have the same descriptors, they are not equal.
+  if (equals &&
+      self->message->GetDescriptor() != other_message->GetDescriptor()) {
+    equals = false;
+  }
+  // Check the message contents.
+  if (equals && !google::protobuf::util::MessageDifferencer::Equals(
+          *self->message,
+          *reinterpret_cast<CMessage*>(other)->message)) {
+    equals = false;
+  }
+  if (equals ^ (opid == Py_EQ)) {
+    Py_RETURN_FALSE;
+  } else {
+    Py_RETURN_TRUE;
+  }
+}
+
+PyObject* InternalGetScalar(const Message* message,
+                            const FieldDescriptor* field_descriptor) {
+  const Reflection* reflection = message->GetReflection();
+
+  if (!CheckFieldBelongsToMessage(field_descriptor, message)) {
+    return NULL;
+  }
+
+  PyObject* result = NULL;
+  switch (field_descriptor->cpp_type()) {
+    case FieldDescriptor::CPPTYPE_INT32: {
+      int32 value = reflection->GetInt32(*message, field_descriptor);
+      result = PyInt_FromLong(value);
+      break;
+    }
+    case FieldDescriptor::CPPTYPE_INT64: {
+      int64 value = reflection->GetInt64(*message, field_descriptor);
+      result = PyLong_FromLongLong(value);
+      break;
+    }
+    case FieldDescriptor::CPPTYPE_UINT32: {
+      uint32 value = reflection->GetUInt32(*message, field_descriptor);
+      result = PyInt_FromSize_t(value);
+      break;
+    }
+    case FieldDescriptor::CPPTYPE_UINT64: {
+      uint64 value = reflection->GetUInt64(*message, field_descriptor);
+      result = PyLong_FromUnsignedLongLong(value);
+      break;
+    }
+    case FieldDescriptor::CPPTYPE_FLOAT: {
+      float value = reflection->GetFloat(*message, field_descriptor);
+      result = PyFloat_FromDouble(value);
+      break;
+    }
+    case FieldDescriptor::CPPTYPE_DOUBLE: {
+      double value = reflection->GetDouble(*message, field_descriptor);
+      result = PyFloat_FromDouble(value);
+      break;
+    }
+    case FieldDescriptor::CPPTYPE_BOOL: {
+      bool value = reflection->GetBool(*message, field_descriptor);
+      result = PyBool_FromLong(value);
+      break;
+    }
+    case FieldDescriptor::CPPTYPE_STRING: {
+      string value = reflection->GetString(*message, field_descriptor);
+      result = ToStringObject(field_descriptor, value);
+      break;
+    }
+    case FieldDescriptor::CPPTYPE_ENUM: {
+      if (!message->GetReflection()->SupportsUnknownEnumValues() &&
+          !message->GetReflection()->HasField(*message, field_descriptor)) {
+        // Look for the value in the unknown fields.
+        const UnknownFieldSet& unknown_field_set =
+            message->GetReflection()->GetUnknownFields(*message);
+        for (int i = 0; i < unknown_field_set.field_count(); ++i) {
+          if (unknown_field_set.field(i).number() ==
+              field_descriptor->number() &&
+              unknown_field_set.field(i).type() ==
+              google::protobuf::UnknownField::TYPE_VARINT) {
+            result = PyInt_FromLong(unknown_field_set.field(i).varint());
+            break;
+          }
+        }
+      }
+
+      if (result == NULL) {
+        const EnumValueDescriptor* enum_value =
+            message->GetReflection()->GetEnum(*message, field_descriptor);
+        result = PyInt_FromLong(enum_value->number());
+      }
+      break;
+    }
+    default:
+      PyErr_Format(
+          PyExc_SystemError, "Getting a value from a field of unknown type %d",
+          field_descriptor->cpp_type());
+  }
+
+  return result;
+}
+
+PyObject* InternalGetSubMessage(
+    CMessage* self, const FieldDescriptor* field_descriptor) {
+  const Reflection* reflection = self->message->GetReflection();
+  PyDescriptorPool* pool = GetDescriptorPoolForMessage(self);
+  const Message& sub_message = reflection->GetMessage(
+      *self->message, field_descriptor, pool->message_factory);
+
+  PyObject *message_class = cdescriptor_pool::GetMessageClass(
+      pool, field_descriptor->message_type());
+  if (message_class == NULL) {
+    return NULL;
+  }
+
+  CMessage* cmsg = cmessage::NewEmptyMessage(message_class,
+                                             sub_message.GetDescriptor());
+  if (cmsg == NULL) {
+    return NULL;
+  }
+
+  cmsg->owner = self->owner;
+  cmsg->parent = self;
+  cmsg->parent_field_descriptor = field_descriptor;
+  cmsg->read_only = !reflection->HasField(*self->message, field_descriptor);
+  cmsg->message = const_cast<Message*>(&sub_message);
+
+  return reinterpret_cast<PyObject*>(cmsg);
+}
+
+int InternalSetNonOneofScalar(
+    Message* message,
+    const FieldDescriptor* field_descriptor,
+    PyObject* arg) {
+  const Reflection* reflection = message->GetReflection();
+
+  if (!CheckFieldBelongsToMessage(field_descriptor, message)) {
+    return -1;
+  }
+
+  switch (field_descriptor->cpp_type()) {
+    case FieldDescriptor::CPPTYPE_INT32: {
+      GOOGLE_CHECK_GET_INT32(arg, value, -1);
+      reflection->SetInt32(message, field_descriptor, value);
+      break;
+    }
+    case FieldDescriptor::CPPTYPE_INT64: {
+      GOOGLE_CHECK_GET_INT64(arg, value, -1);
+      reflection->SetInt64(message, field_descriptor, value);
+      break;
+    }
+    case FieldDescriptor::CPPTYPE_UINT32: {
+      GOOGLE_CHECK_GET_UINT32(arg, value, -1);
+      reflection->SetUInt32(message, field_descriptor, value);
+      break;
+    }
+    case FieldDescriptor::CPPTYPE_UINT64: {
+      GOOGLE_CHECK_GET_UINT64(arg, value, -1);
+      reflection->SetUInt64(message, field_descriptor, value);
+      break;
+    }
+    case FieldDescriptor::CPPTYPE_FLOAT: {
+      GOOGLE_CHECK_GET_FLOAT(arg, value, -1);
+      reflection->SetFloat(message, field_descriptor, value);
+      break;
+    }
+    case FieldDescriptor::CPPTYPE_DOUBLE: {
+      GOOGLE_CHECK_GET_DOUBLE(arg, value, -1);
+      reflection->SetDouble(message, field_descriptor, value);
+      break;
+    }
+    case FieldDescriptor::CPPTYPE_BOOL: {
+      GOOGLE_CHECK_GET_BOOL(arg, value, -1);
+      reflection->SetBool(message, field_descriptor, value);
+      break;
+    }
+    case FieldDescriptor::CPPTYPE_STRING: {
+      if (!CheckAndSetString(
+          arg, message, field_descriptor, reflection, false, -1)) {
+        return -1;
+      }
+      break;
+    }
+    case FieldDescriptor::CPPTYPE_ENUM: {
+      GOOGLE_CHECK_GET_INT32(arg, value, -1);
+      if (reflection->SupportsUnknownEnumValues()) {
+        reflection->SetEnumValue(message, field_descriptor, value);
+      } else {
+        const EnumDescriptor* enum_descriptor = field_descriptor->enum_type();
+        const EnumValueDescriptor* enum_value =
+            enum_descriptor->FindValueByNumber(value);
+        if (enum_value != NULL) {
+          reflection->SetEnum(message, field_descriptor, enum_value);
+        } else {
+          PyErr_Format(PyExc_ValueError, "Unknown enum value: %d", value);
+          return -1;
+        }
+      }
+      break;
+    }
+    default:
+      PyErr_Format(
+          PyExc_SystemError, "Setting value to a field of unknown type %d",
+          field_descriptor->cpp_type());
+      return -1;
+  }
+
+  return 0;
+}
+
+int InternalSetScalar(
+    CMessage* self,
+    const FieldDescriptor* field_descriptor,
+    PyObject* arg) {
+  if (!CheckFieldBelongsToMessage(field_descriptor, self->message)) {
+    return -1;
+  }
+
+  if (MaybeReleaseOverlappingOneofField(self, field_descriptor) < 0) {
+    return -1;
+  }
+
+  return InternalSetNonOneofScalar(self->message, field_descriptor, arg);
+}
+
+PyObject* FromString(PyTypeObject* cls, PyObject* serialized) {
+  PyObject* py_cmsg = PyObject_CallObject(
+      reinterpret_cast<PyObject*>(cls), NULL);
+  if (py_cmsg == NULL) {
+    return NULL;
+  }
+  CMessage* cmsg = reinterpret_cast<CMessage*>(py_cmsg);
+
+  ScopedPyObjectPtr py_length(MergeFromString(cmsg, serialized));
+  if (py_length == NULL) {
+    Py_DECREF(py_cmsg);
+    return NULL;
+  }
+
+  return py_cmsg;
+}
+
+PyObject* DeepCopy(CMessage* self, PyObject* arg) {
+  PyObject* clone = PyObject_CallObject(
+      reinterpret_cast<PyObject*>(Py_TYPE(self)), NULL);
+  if (clone == NULL) {
+    return NULL;
+  }
+  if (!PyObject_TypeCheck(clone, &CMessage_Type)) {
+    Py_DECREF(clone);
+    return NULL;
+  }
+  if (ScopedPyObjectPtr(MergeFrom(
+          reinterpret_cast<CMessage*>(clone),
+          reinterpret_cast<PyObject*>(self))) == NULL) {
+    Py_DECREF(clone);
+    return NULL;
+  }
+  return clone;
+}
+
+PyObject* ToUnicode(CMessage* self) {
+  // Lazy import to prevent circular dependencies
+  ScopedPyObjectPtr text_format(
+      PyImport_ImportModule("google.protobuf.text_format"));
+  if (text_format == NULL) {
+    return NULL;
+  }
+  ScopedPyObjectPtr method_name(PyString_FromString("MessageToString"));
+  if (method_name == NULL) {
+    return NULL;
+  }
+  Py_INCREF(Py_True);
+  ScopedPyObjectPtr encoded(PyObject_CallMethodObjArgs(
+      text_format.get(), method_name.get(), self, Py_True, NULL));
+  Py_DECREF(Py_True);
+  if (encoded == NULL) {
+    return NULL;
+  }
+#if PY_MAJOR_VERSION < 3
+  PyObject* decoded = PyString_AsDecodedObject(encoded.get(), "utf-8", NULL);
+#else
+  PyObject* decoded = PyUnicode_FromEncodedObject(encoded.get(), "utf-8", NULL);
+#endif
+  if (decoded == NULL) {
+    return NULL;
+  }
+  return decoded;
+}
+
+PyObject* Reduce(CMessage* self) {
+  ScopedPyObjectPtr constructor(reinterpret_cast<PyObject*>(Py_TYPE(self)));
+  constructor.inc();
+  ScopedPyObjectPtr args(PyTuple_New(0));
+  if (args == NULL) {
+    return NULL;
+  }
+  ScopedPyObjectPtr state(PyDict_New());
+  if (state == NULL) {
+    return  NULL;
+  }
+  ScopedPyObjectPtr serialized(SerializePartialToString(self));
+  if (serialized == NULL) {
+    return NULL;
+  }
+  if (PyDict_SetItemString(state.get(), "serialized", serialized.get()) < 0) {
+    return NULL;
+  }
+  return Py_BuildValue("OOO", constructor.get(), args.get(), state.get());
+}
+
+PyObject* SetState(CMessage* self, PyObject* state) {
+  if (!PyDict_Check(state)) {
+    PyErr_SetString(PyExc_TypeError, "state not a dict");
+    return NULL;
+  }
+  PyObject* serialized = PyDict_GetItemString(state, "serialized");
+  if (serialized == NULL) {
+    return NULL;
+  }
+  if (ScopedPyObjectPtr(ParseFromString(self, serialized)) == NULL) {
+    return NULL;
+  }
+  Py_RETURN_NONE;
+}
+
+// CMessage static methods:
+PyObject* _CheckCalledFromGeneratedFile(PyObject* unused,
+                                        PyObject* unused_arg) {
+  if (!_CalledFromGeneratedFile(1)) {
+    PyErr_SetString(PyExc_TypeError,
+                    "Descriptors should not be created directly, "
+                    "but only retrieved from their parent.");
+    return NULL;
+  }
+  Py_RETURN_NONE;
+}
+
+static PyObject* GetExtensionDict(CMessage* self, void *closure) {
+  if (self->extensions)  {
+    Py_INCREF(self->extensions);
+    return reinterpret_cast<PyObject*>(self->extensions);
+  }
+
+  // If there are extension_ranges, the message is "extendable". Allocate a
+  // dictionary to store the extension fields.
+  const Descriptor* descriptor = GetMessageDescriptor(Py_TYPE(self));
+  if (descriptor->extension_range_count() > 0) {
+    ExtensionDict* extension_dict = extension_dict::NewExtensionDict(self);
+    if (extension_dict == NULL) {
+      return NULL;
+    }
+    self->extensions = extension_dict;
+    Py_INCREF(self->extensions);
+    return reinterpret_cast<PyObject*>(self->extensions);
+  }
+
+  PyErr_SetNone(PyExc_AttributeError);
+  return NULL;
+}
+
+static PyGetSetDef Getters[] = {
+  {"Extensions", (getter)GetExtensionDict, NULL, "Extension dict"},
+  {NULL}
+};
+
+static PyMethodDef Methods[] = {
+  { "__deepcopy__", (PyCFunction)DeepCopy, METH_VARARGS,
+    "Makes a deep copy of the class." },
+  { "__reduce__", (PyCFunction)Reduce, METH_NOARGS,
+    "Outputs picklable representation of the message." },
+  { "__setstate__", (PyCFunction)SetState, METH_O,
+    "Inputs picklable representation of the message." },
+  { "__unicode__", (PyCFunction)ToUnicode, METH_NOARGS,
+    "Outputs a unicode representation of the message." },
+  { "ByteSize", (PyCFunction)ByteSize, METH_NOARGS,
+    "Returns the size of the message in bytes." },
+  { "Clear", (PyCFunction)Clear, METH_NOARGS,
+    "Clears the message." },
+  { "ClearExtension", (PyCFunction)ClearExtension, METH_O,
+    "Clears a message field." },
+  { "ClearField", (PyCFunction)ClearField, METH_O,
+    "Clears a message field." },
+  { "CopyFrom", (PyCFunction)CopyFrom, METH_O,
+    "Copies a protocol message into the current message." },
+  { "FindInitializationErrors", (PyCFunction)FindInitializationErrors,
+    METH_NOARGS,
+    "Finds unset required fields." },
+  { "FromString", (PyCFunction)FromString, METH_O | METH_CLASS,
+    "Creates new method instance from given serialized data." },
+  { "HasExtension", (PyCFunction)HasExtension, METH_O,
+    "Checks if a message field is set." },
+  { "HasField", (PyCFunction)HasField, METH_O,
+    "Checks if a message field is set." },
+  { "IsInitialized", (PyCFunction)IsInitialized, METH_VARARGS,
+    "Checks if all required fields of a protocol message are set." },
+  { "ListFields", (PyCFunction)ListFields, METH_NOARGS,
+    "Lists all set fields of a message." },
+  { "MergeFrom", (PyCFunction)MergeFrom, METH_O,
+    "Merges a protocol message into the current message." },
+  { "MergeFromString", (PyCFunction)MergeFromString, METH_O,
+    "Merges a serialized message into the current message." },
+  { "ParseFromString", (PyCFunction)ParseFromString, METH_O,
+    "Parses a serialized message into the current message." },
+  { "RegisterExtension", (PyCFunction)RegisterExtension, METH_O | METH_CLASS,
+    "Registers an extension with the current message." },
+  { "SerializePartialToString", (PyCFunction)SerializePartialToString,
+    METH_NOARGS,
+    "Serializes the message to a string, even if it isn't initialized." },
+  { "SerializeToString", (PyCFunction)SerializeToString, METH_NOARGS,
+    "Serializes the message to a string, only for initialized messages." },
+  { "SetInParent", (PyCFunction)SetInParent, METH_NOARGS,
+    "Sets the has bit of the given field in its parent message." },
+  { "WhichOneof", (PyCFunction)WhichOneof, METH_O,
+    "Returns the name of the field set inside a oneof, "
+    "or None if no field is set." },
+
+  // Static Methods.
+  { "_CheckCalledFromGeneratedFile", (PyCFunction)_CheckCalledFromGeneratedFile,
+    METH_NOARGS | METH_STATIC,
+    "Raises TypeError if the caller is not in a _pb2.py file."},
+  { NULL, NULL}
+};
+
+static bool SetCompositeField(
+    CMessage* self, PyObject* name, PyObject* value) {
+  if (self->composite_fields == NULL) {
+    self->composite_fields = PyDict_New();
+    if (self->composite_fields == NULL) {
+      return false;
+    }
+  }
+  return PyDict_SetItem(self->composite_fields, name, value) == 0;
+}
+
+PyObject* GetAttr(CMessage* self, PyObject* name) {
+  PyObject* value = self->composite_fields ?
+      PyDict_GetItem(self->composite_fields, name) : NULL;
+  if (value != NULL) {
+    Py_INCREF(value);
+    return value;
+  }
+
+  const FieldDescriptor* field_descriptor = GetFieldDescriptor(self, name);
+  if (field_descriptor == NULL) {
+    return CMessage_Type.tp_base->tp_getattro(
+        reinterpret_cast<PyObject*>(self), name);
+  }
+
+  if (field_descriptor->is_map()) {
+    PyObject* py_container = NULL;
+    const Descriptor* entry_type = field_descriptor->message_type();
+    const FieldDescriptor* value_type = entry_type->FindFieldByName("value");
+    if (value_type->cpp_type() == FieldDescriptor::CPPTYPE_MESSAGE) {
+      PyObject* value_class = cdescriptor_pool::GetMessageClass(
+          GetDescriptorPoolForMessage(self), value_type->message_type());
+      if (value_class == NULL) {
+        return NULL;
+      }
+      py_container =
+          NewMessageMapContainer(self, field_descriptor, value_class);
+    } else {
+      py_container = NewScalarMapContainer(self, field_descriptor);
+    }
+    if (py_container == NULL) {
+      return NULL;
+    }
+    if (!SetCompositeField(self, name, py_container)) {
+      Py_DECREF(py_container);
+      return NULL;
+    }
+    return py_container;
+  }
+
+  if (field_descriptor->label() == FieldDescriptor::LABEL_REPEATED) {
+    PyObject* py_container = NULL;
+    if (field_descriptor->cpp_type() == FieldDescriptor::CPPTYPE_MESSAGE) {
+      PyObject *message_class = cdescriptor_pool::GetMessageClass(
+          GetDescriptorPoolForMessage(self), field_descriptor->message_type());
+      if (message_class == NULL) {
+        return NULL;
+      }
+      py_container = repeated_composite_container::NewContainer(
+          self, field_descriptor, message_class);
+    } else {
+      py_container = repeated_scalar_container::NewContainer(
+          self, field_descriptor);
+    }
+    if (py_container == NULL) {
+      return NULL;
+    }
+    if (!SetCompositeField(self, name, py_container)) {
+      Py_DECREF(py_container);
+      return NULL;
+    }
+    return py_container;
+  }
+
+  if (field_descriptor->cpp_type() == FieldDescriptor::CPPTYPE_MESSAGE) {
+    PyObject* sub_message = InternalGetSubMessage(self, field_descriptor);
+    if (sub_message == NULL) {
+      return NULL;
+    }
+    if (!SetCompositeField(self, name, sub_message)) {
+      Py_DECREF(sub_message);
+      return NULL;
+    }
+    return sub_message;
+  }
+
+  return InternalGetScalar(self->message, field_descriptor);
+}
+
+int SetAttr(CMessage* self, PyObject* name, PyObject* value) {
+  if (self->composite_fields && PyDict_Contains(self->composite_fields, name)) {
+    PyErr_SetString(PyExc_TypeError, "Can't set composite field");
+    return -1;
+  }
+
+  const FieldDescriptor* field_descriptor = GetFieldDescriptor(self, name);
+  if (field_descriptor != NULL) {
+    AssureWritable(self);
+    if (field_descriptor->label() == FieldDescriptor::LABEL_REPEATED) {
+      PyErr_Format(PyExc_AttributeError, "Assignment not allowed to repeated "
+                   "field \"%s\" in protocol message object.",
+                   field_descriptor->name().c_str());
+      return -1;
+    } else {
+      if (field_descriptor->cpp_type() == FieldDescriptor::CPPTYPE_MESSAGE) {
+        PyErr_Format(PyExc_AttributeError, "Assignment not allowed to "
+                     "field \"%s\" in protocol message object.",
+                     field_descriptor->name().c_str());
+        return -1;
+      } else {
+        return InternalSetScalar(self, field_descriptor, value);
+      }
+    }
+  }
+
+  PyErr_Format(PyExc_AttributeError,
+               "Assignment not allowed "
+               "(no field \"%s\"in protocol message object).",
+               PyString_AsString(name));
+  return -1;
+}
+
+}  // namespace cmessage
+
+PyTypeObject CMessage_Type = {
+  PyVarObject_HEAD_INIT(&PyMessageMeta_Type, 0)
+  FULL_MODULE_NAME ".CMessage",        // tp_name
+  sizeof(CMessage),                    // tp_basicsize
+  0,                                   //  tp_itemsize
+  (destructor)cmessage::Dealloc,       //  tp_dealloc
+  0,                                   //  tp_print
+  0,                                   //  tp_getattr
+  0,                                   //  tp_setattr
+  0,                                   //  tp_compare
+  0,                                   //  tp_repr
+  0,                                   //  tp_as_number
+  0,                                   //  tp_as_sequence
+  0,                                   //  tp_as_mapping
+  PyObject_HashNotImplemented,         //  tp_hash
+  0,                                   //  tp_call
+  (reprfunc)cmessage::ToStr,           //  tp_str
+  (getattrofunc)cmessage::GetAttr,     //  tp_getattro
+  (setattrofunc)cmessage::SetAttr,     //  tp_setattro
+  0,                                   //  tp_as_buffer
+  Py_TPFLAGS_DEFAULT | Py_TPFLAGS_BASETYPE,  //  tp_flags
+  "A ProtocolMessage",                 //  tp_doc
+  0,                                   //  tp_traverse
+  0,                                   //  tp_clear
+  (richcmpfunc)cmessage::RichCompare,  //  tp_richcompare
+  0,                                   //  tp_weaklistoffset
+  0,                                   //  tp_iter
+  0,                                   //  tp_iternext
+  cmessage::Methods,                   //  tp_methods
+  0,                                   //  tp_members
+  cmessage::Getters,                   //  tp_getset
+  0,                                   //  tp_base
+  0,                                   //  tp_dict
+  0,                                   //  tp_descr_get
+  0,                                   //  tp_descr_set
+  0,                                   //  tp_dictoffset
+  (initproc)cmessage::Init,            //  tp_init
+  0,                                   //  tp_alloc
+  cmessage::New,                       //  tp_new
+};
+
+// --- Exposing the C proto living inside Python proto to C code:
+
+const Message* (*GetCProtoInsidePyProtoPtr)(PyObject* msg);
+Message* (*MutableCProtoInsidePyProtoPtr)(PyObject* msg);
+
+static const Message* GetCProtoInsidePyProtoImpl(PyObject* msg) {
+  if (!PyObject_TypeCheck(msg, &CMessage_Type)) {
+    return NULL;
+  }
+  CMessage* cmsg = reinterpret_cast<CMessage*>(msg);
+  return cmsg->message;
+}
+
+static Message* MutableCProtoInsidePyProtoImpl(PyObject* msg) {
+  if (!PyObject_TypeCheck(msg, &CMessage_Type)) {
+    return NULL;
+  }
+  CMessage* cmsg = reinterpret_cast<CMessage*>(msg);
+  if ((cmsg->composite_fields && PyDict_Size(cmsg->composite_fields) != 0) ||
+      (cmsg->extensions != NULL &&
+       PyDict_Size(cmsg->extensions->values) != 0)) {
+    // There is currently no way of accurately syncing arbitrary changes to
+    // the underlying C++ message back to the CMessage (e.g. removed repeated
+    // composite containers). We only allow direct mutation of the underlying
+    // C++ message if there is no child data in the CMessage.
+    return NULL;
+  }
+  cmessage::AssureWritable(cmsg);
+  return cmsg->message;
+}
+
+static const char module_docstring[] =
+"python-proto2 is a module that can be used to enhance proto2 Python API\n"
+"performance.\n"
+"\n"
+"It provides access to the protocol buffers C++ reflection API that\n"
+"implements the basic protocol buffer functions.";
+
+void InitGlobals() {
+  // TODO(gps): Check all return values in this function for NULL and propagate
+  // the error (MemoryError) on up to result in an import failure.  These should
+  // also be freed and reset to NULL during finalization.
+  kPythonZero = PyInt_FromLong(0);
+  kint32min_py = PyInt_FromLong(kint32min);
+  kint32max_py = PyInt_FromLong(kint32max);
+  kuint32max_py = PyLong_FromLongLong(kuint32max);
+  kint64min_py = PyLong_FromLongLong(kint64min);
+  kint64max_py = PyLong_FromLongLong(kint64max);
+  kuint64max_py = PyLong_FromUnsignedLongLong(kuint64max);
+
+  kDESCRIPTOR = PyString_FromString("DESCRIPTOR");
+  k_cdescriptor = PyString_FromString("_cdescriptor");
+  kfull_name = PyString_FromString("full_name");
+  k_extensions_by_name = PyString_FromString("_extensions_by_name");
+  k_extensions_by_number = PyString_FromString("_extensions_by_number");
+
+  PyObject *dummy_obj = PySet_New(NULL);
+  kEmptyWeakref = PyWeakref_NewRef(dummy_obj, NULL);
+  Py_DECREF(dummy_obj);
+}
+
+bool InitProto2MessageModule(PyObject *m) {
+  // Initialize types and globals in descriptor.cc
+  if (!InitDescriptor()) {
+    return false;
+  }
+
+  // Initialize types and globals in descriptor_pool.cc
+  if (!InitDescriptorPool()) {
+    return false;
+  }
+
+  // Initialize constants defined in this file.
+  InitGlobals();
+
+  PyMessageMeta_Type.tp_base = &PyType_Type;
+  if (PyType_Ready(&PyMessageMeta_Type) < 0) {
+    return false;
+  }
+  PyModule_AddObject(m, "MessageMeta",
+                     reinterpret_cast<PyObject*>(&PyMessageMeta_Type));
+
+  if (PyType_Ready(&CMessage_Type) < 0) {
+    return false;
+  }
+
+  // DESCRIPTOR is set on each protocol buffer message class elsewhere, but set
+  // it here as well to document that subclasses need to set it.
+  PyDict_SetItem(CMessage_Type.tp_dict, kDESCRIPTOR, Py_None);
+  // Subclasses with message extensions will override _extensions_by_name and
+  // _extensions_by_number with fresh mutable dictionaries in AddDescriptors.
+  // All other classes can share this same immutable mapping.
+  ScopedPyObjectPtr empty_dict(PyDict_New());
+  if (empty_dict == NULL) {
+    return false;
+  }
+  ScopedPyObjectPtr immutable_dict(PyDictProxy_New(empty_dict.get()));
+  if (immutable_dict == NULL) {
+    return false;
+  }
+  if (PyDict_SetItem(CMessage_Type.tp_dict,
+                     k_extensions_by_name, immutable_dict.get()) < 0) {
+    return false;
+  }
+  if (PyDict_SetItem(CMessage_Type.tp_dict,
+                     k_extensions_by_number, immutable_dict.get()) < 0) {
+    return false;
+  }
+
+  PyModule_AddObject(m, "Message", reinterpret_cast<PyObject*>(&CMessage_Type));
+
+  // Initialize Repeated container types.
+  {
+    if (PyType_Ready(&RepeatedScalarContainer_Type) < 0) {
+      return false;
+    }
+
+    PyModule_AddObject(m, "RepeatedScalarContainer",
+                       reinterpret_cast<PyObject*>(
+                           &RepeatedScalarContainer_Type));
+
+    if (PyType_Ready(&RepeatedCompositeContainer_Type) < 0) {
+      return false;
+    }
+
+    PyModule_AddObject(
+        m, "RepeatedCompositeContainer",
+        reinterpret_cast<PyObject*>(
+            &RepeatedCompositeContainer_Type));
+
+    // Register them as collections.Sequence
+    ScopedPyObjectPtr collections(PyImport_ImportModule("collections"));
+    if (collections == NULL) {
+      return false;
+    }
+    ScopedPyObjectPtr mutable_sequence(
+        PyObject_GetAttrString(collections.get(), "MutableSequence"));
+    if (mutable_sequence == NULL) {
+      return false;
+    }
+    if (ScopedPyObjectPtr(
+            PyObject_CallMethod(mutable_sequence.get(), "register", "O",
+                                &RepeatedScalarContainer_Type)) == NULL) {
+      return false;
+    }
+    if (ScopedPyObjectPtr(
+            PyObject_CallMethod(mutable_sequence.get(), "register", "O",
+                                &RepeatedCompositeContainer_Type)) == NULL) {
+      return false;
+    }
+  }
+
+  // Initialize Map container types.
+  {
+    // ScalarMapContainer_Type derives from our MutableMapping type.
+    ScopedPyObjectPtr containers(PyImport_ImportModule(
+        "google.protobuf.internal.containers"));
+    if (containers == NULL) {
+      return false;
+    }
+
+    ScopedPyObjectPtr mutable_mapping(
+        PyObject_GetAttrString(containers.get(), "MutableMapping"));
+    if (mutable_mapping == NULL) {
+      return false;
+    }
+
+    if (!PyObject_TypeCheck(mutable_mapping.get(), &PyType_Type)) {
+      return false;
+    }
+
+    Py_INCREF(mutable_mapping.get());
+#if PY_MAJOR_VERSION >= 3
+    PyObject* bases = PyTuple_New(1);
+    PyTuple_SET_ITEM(bases, 0, mutable_mapping.get());
+
+    ScalarMapContainer_Type = 
+        PyType_FromSpecWithBases(&ScalarMapContainer_Type_spec, bases);
+    PyModule_AddObject(m, "ScalarMapContainer", ScalarMapContainer_Type);
+#else
+    ScalarMapContainer_Type.tp_base =
+        reinterpret_cast<PyTypeObject*>(mutable_mapping.get());
+
+    if (PyType_Ready(&ScalarMapContainer_Type) < 0) {
+      return false;
+    }
+
+    PyModule_AddObject(m, "ScalarMapContainer",
+                       reinterpret_cast<PyObject*>(&ScalarMapContainer_Type));
+#endif
+
+    if (PyType_Ready(&MapIterator_Type) < 0) {
+      return false;
+    }
+
+    PyModule_AddObject(m, "MapIterator",
+                       reinterpret_cast<PyObject*>(&MapIterator_Type));
+
+
+#if PY_MAJOR_VERSION >= 3
+    MessageMapContainer_Type = 
+        PyType_FromSpecWithBases(&MessageMapContainer_Type_spec, bases);
+    PyModule_AddObject(m, "MessageMapContainer", MessageMapContainer_Type);
+#else
+    Py_INCREF(mutable_mapping.get());
+    MessageMapContainer_Type.tp_base =
+        reinterpret_cast<PyTypeObject*>(mutable_mapping.get());
+
+    if (PyType_Ready(&MessageMapContainer_Type) < 0) {
+      return false;
+    }
+
+    PyModule_AddObject(m, "MessageMapContainer",
+                       reinterpret_cast<PyObject*>(&MessageMapContainer_Type));
+#endif
+  }
+
+  if (PyType_Ready(&ExtensionDict_Type) < 0) {
+    return false;
+  }
+  PyModule_AddObject(
+      m, "ExtensionDict",
+      reinterpret_cast<PyObject*>(&ExtensionDict_Type));
+
+  // Expose the DescriptorPool used to hold all descriptors added from generated
+  // pb2.py files.
+  // PyModule_AddObject steals a reference.
+  Py_INCREF(GetDefaultDescriptorPool());
+  PyModule_AddObject(m, "default_pool",
+                     reinterpret_cast<PyObject*>(GetDefaultDescriptorPool()));
+
+  PyModule_AddObject(m, "DescriptorPool", reinterpret_cast<PyObject*>(
+      &PyDescriptorPool_Type));
+
+  // This implementation provides full Descriptor types, we advertise it so that
+  // descriptor.py can use them in replacement of the Python classes.
+  PyModule_AddIntConstant(m, "_USE_C_DESCRIPTORS", 1);
+
+  PyModule_AddObject(m, "Descriptor", reinterpret_cast<PyObject*>(
+      &PyMessageDescriptor_Type));
+  PyModule_AddObject(m, "FieldDescriptor", reinterpret_cast<PyObject*>(
+      &PyFieldDescriptor_Type));
+  PyModule_AddObject(m, "EnumDescriptor", reinterpret_cast<PyObject*>(
+      &PyEnumDescriptor_Type));
+  PyModule_AddObject(m, "EnumValueDescriptor", reinterpret_cast<PyObject*>(
+      &PyEnumValueDescriptor_Type));
+  PyModule_AddObject(m, "FileDescriptor", reinterpret_cast<PyObject*>(
+      &PyFileDescriptor_Type));
+  PyModule_AddObject(m, "OneofDescriptor", reinterpret_cast<PyObject*>(
+      &PyOneofDescriptor_Type));
+
+  PyObject* enum_type_wrapper = PyImport_ImportModule(
+      "google.protobuf.internal.enum_type_wrapper");
+  if (enum_type_wrapper == NULL) {
+    return false;
+  }
+  EnumTypeWrapper_class =
+      PyObject_GetAttrString(enum_type_wrapper, "EnumTypeWrapper");
+  Py_DECREF(enum_type_wrapper);
+
+  PyObject* message_module = PyImport_ImportModule(
+      "google.protobuf.message");
+  if (message_module == NULL) {
+    return false;
+  }
+  EncodeError_class = PyObject_GetAttrString(message_module, "EncodeError");
+  DecodeError_class = PyObject_GetAttrString(message_module, "DecodeError");
+  PythonMessage_class = PyObject_GetAttrString(message_module, "Message");
+  Py_DECREF(message_module);
+
+  PyObject* pickle_module = PyImport_ImportModule("pickle");
+  if (pickle_module == NULL) {
+    return false;
+  }
+  PickleError_class = PyObject_GetAttrString(pickle_module, "PickleError");
+  Py_DECREF(pickle_module);
+
+  // Override {Get,Mutable}CProtoInsidePyProto.
+  GetCProtoInsidePyProtoPtr = GetCProtoInsidePyProtoImpl;
+  MutableCProtoInsidePyProtoPtr = MutableCProtoInsidePyProtoImpl;
+
+  return true;
+}
+
+}  // namespace python
+}  // namespace protobuf
+
+
+#if PY_MAJOR_VERSION >= 3
+static struct PyModuleDef _module = {
+  PyModuleDef_HEAD_INIT,
+  "_message",
+  google::protobuf::python::module_docstring,
+  -1,
+  NULL,
+  NULL,
+  NULL,
+  NULL,
+  NULL
+};
+#define INITFUNC PyInit__message
+#define INITFUNC_ERRORVAL NULL
+#else  // Python 2
+#define INITFUNC init_message
+#define INITFUNC_ERRORVAL
+#endif
+
+extern "C" {
+  PyMODINIT_FUNC INITFUNC(void) {
+    PyObject* m;
+#if PY_MAJOR_VERSION >= 3
+    m = PyModule_Create(&_module);
+#else
+    m = Py_InitModule3("_message", NULL, google::protobuf::python::module_docstring);
+#endif
+    if (m == NULL) {
+      return INITFUNC_ERRORVAL;
+    }
+
+    if (!google::protobuf::python::InitProto2MessageModule(m)) {
+      Py_DECREF(m);
+      return INITFUNC_ERRORVAL;
+    }
+
+#if PY_MAJOR_VERSION >= 3
+    return m;
+#endif
+  }
+}
+}  // namespace google
diff --git a/third_party/protobuf/python/google/protobuf/pyext/message.h b/third_party/protobuf/python/google/protobuf/pyext/message.h
new file mode 100644
index 0000000000..cc0012e95e
--- /dev/null
+++ b/third_party/protobuf/python/google/protobuf/pyext/message.h
@@ -0,0 +1,330 @@
+// 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: anuraag@google.com (Anuraag Agrawal)
+// Author: tibell@google.com (Johan Tibell)
+
+#ifndef GOOGLE_PROTOBUF_PYTHON_CPP_MESSAGE_H__
+#define GOOGLE_PROTOBUF_PYTHON_CPP_MESSAGE_H__
+
+#include <Python.h>
+
+#include <memory>
+#ifndef _SHARED_PTR_H
+#include <google/protobuf/stubs/shared_ptr.h>
+#endif
+#include <string>
+
+namespace google {
+namespace protobuf {
+
+class Message;
+class Reflection;
+class FieldDescriptor;
+class Descriptor;
+class DescriptorPool;
+class MessageFactory;
+
+#ifdef _SHARED_PTR_H
+using std::shared_ptr;
+using std::string;
+#else
+using internal::shared_ptr;
+#endif
+
+namespace python {
+
+struct ExtensionDict;
+struct PyDescriptorPool;
+
+typedef struct CMessage {
+  PyObject_HEAD;
+
+  // This is the top-level C++ Message object that owns the whole
+  // proto tree.  Every Python CMessage holds a reference to it in
+  // order to keep it alive as long as there's a Python object that
+  // references any part of the tree.
+  shared_ptr<Message> owner;
+
+  // Weak reference to a parent CMessage object. This is NULL for any top-level
+  // message and is set for any child message (i.e. a child submessage or a
+  // part of a repeated composite field).
+  //
+  // Used to make sure all ancestors are also mutable when first modifying
+  // a child submessage (in other words, turning a default message instance
+  // into a mutable one).
+  //
+  // If a submessage is released (becomes a new top-level message), this field
+  // MUST be set to NULL. The parent may get deallocated and further attempts
+  // to use this pointer will result in a crash.
+  struct CMessage* parent;
+
+  // Pointer to the parent's descriptor that describes this submessage.
+  // Used together with the parent's message when making a default message
+  // instance mutable.
+  // The pointer is owned by the global DescriptorPool.
+  const FieldDescriptor* parent_field_descriptor;
+
+  // Pointer to the C++ Message object for this CMessage.  The
+  // CMessage does not own this pointer.
+  Message* message;
+
+  // Indicates this submessage is pointing to a default instance of a message.
+  // Submessages are always first created as read only messages and are then
+  // made writable, at which point this field is set to false.
+  bool read_only;
+
+  // A reference to a Python dictionary containing CMessage,
+  // RepeatedCompositeContainer, and RepeatedScalarContainer
+  // objects. Used as a cache to make sure we don't have to make a
+  // Python wrapper for the C++ Message objects on every access, or
+  // deal with the synchronization nightmare that could create.
+  PyObject* composite_fields;
+
+  // A reference to the dictionary containing the message's extensions.
+  // Similar to composite_fields, acting as a cache, but also contains the
+  // required extension dict logic.
+  ExtensionDict* extensions;
+} CMessage;
+
+extern PyTypeObject CMessage_Type;
+
+namespace cmessage {
+
+// Internal function to create a new empty Message Python object, but with empty
+// pointers to the C++ objects.
+// The caller must fill self->message, self->owner and eventually self->parent.
+CMessage* NewEmptyMessage(PyObject* type, const Descriptor* descriptor);
+
+// Release a submessage from its proto tree, making it a new top-level messgae.
+// A new message will be created if this is a read-only default instance.
+//
+// Corresponds to reflection api method ReleaseMessage.
+int ReleaseSubMessage(CMessage* self,
+                      const FieldDescriptor* field_descriptor,
+                      CMessage* child_cmessage);
+
+// Retrieves the C++ descriptor of a Python Extension descriptor.
+// On error, return NULL with an exception set.
+const FieldDescriptor* GetExtensionDescriptor(PyObject* extension);
+
+// Initializes a new CMessage instance for a submessage. Only called once per
+// submessage as the result is cached in composite_fields.
+//
+// Corresponds to reflection api method GetMessage.
+PyObject* InternalGetSubMessage(
+    CMessage* self, const FieldDescriptor* field_descriptor);
+
+// Deletes a range of C++ submessages in a repeated field (following a
+// removal in a RepeatedCompositeContainer).
+//
+// Releases messages to the provided cmessage_list if it is not NULL rather
+// than just removing them from the underlying proto. This cmessage_list must
+// have a CMessage for each underlying submessage. The CMessages referred to
+// by slice will be removed from cmessage_list by this function.
+//
+// Corresponds to reflection api method RemoveLast.
+int InternalDeleteRepeatedField(CMessage* self,
+                                const FieldDescriptor* field_descriptor,
+                                PyObject* slice, PyObject* cmessage_list);
+
+// Sets the specified scalar value to the message.
+int InternalSetScalar(CMessage* self,
+                      const FieldDescriptor* field_descriptor,
+                      PyObject* value);
+
+// Sets the specified scalar value to the message.  Requires it is not a Oneof.
+int InternalSetNonOneofScalar(Message* message,
+                              const FieldDescriptor* field_descriptor,
+                              PyObject* arg);
+
+// Retrieves the specified scalar value from the message.
+//
+// Returns a new python reference.
+PyObject* InternalGetScalar(const Message* message,
+                            const FieldDescriptor* field_descriptor);
+
+// Clears the message, removing all contained data. Extension dictionary and
+// submessages are released first if there are remaining external references.
+//
+// Corresponds to message api method Clear.
+PyObject* Clear(CMessage* self);
+
+// Clears the data described by the given descriptor. Used to clear extensions
+// (which don't have names). Extension release is handled by ExtensionDict
+// class, not this function.
+// TODO(anuraag): Try to make this discrepancy in release semantics with
+//                ClearField less confusing.
+//
+// Corresponds to reflection api method ClearField.
+PyObject* ClearFieldByDescriptor(
+    CMessage* self, const FieldDescriptor* descriptor);
+
+// Clears the data for the given field name. The message is released if there
+// are any external references.
+//
+// Corresponds to reflection api method ClearField.
+PyObject* ClearField(CMessage* self, PyObject* arg);
+
+// Checks if the message has the field described by the descriptor. Used for
+// extensions (which have no name).
+//
+// Corresponds to reflection api method HasField
+PyObject* HasFieldByDescriptor(
+    CMessage* self, const FieldDescriptor* field_descriptor);
+
+// Checks if the message has the named field.
+//
+// Corresponds to reflection api method HasField.
+PyObject* HasField(CMessage* self, PyObject* arg);
+
+// Initializes values of fields on a newly constructed message.
+int InitAttributes(CMessage* self, PyObject* kwargs);
+
+PyObject* MergeFrom(CMessage* self, PyObject* arg);
+
+// Retrieves an attribute named 'name' from CMessage 'self'. Returns
+// the attribute value on success, or NULL on failure.
+//
+// Returns a new reference.
+PyObject* GetAttr(CMessage* self, PyObject* name);
+
+// Set the value of the attribute named 'name', for CMessage 'self',
+// to the value 'value'. Returns -1 on failure.
+int SetAttr(CMessage* self, PyObject* name, PyObject* value);
+
+PyObject* FindInitializationErrors(CMessage* self);
+
+// Set the owner field of self and any children of self, recursively.
+// Used when self is being released and thus has a new owner (the
+// released Message.)
+int SetOwner(CMessage* self, const shared_ptr<Message>& new_owner);
+
+int AssureWritable(CMessage* self);
+
+// Returns the "best" DescriptorPool for the given message.
+// This is often equivalent to message.DESCRIPTOR.pool, but not always, when
+// the message class was created from a MessageFactory using a custom pool which
+// uses the generated pool as an underlay.
+//
+// The returned pool is suitable for finding fields and building submessages,
+// even in the case of extensions.
+PyDescriptorPool* GetDescriptorPoolForMessage(CMessage* message);
+
+}  // namespace cmessage
+
+
+/* Is 64bit */
+#define IS_64BIT (SIZEOF_LONG == 8)
+
+#define FIELD_IS_REPEATED(field_descriptor)                 \
+    ((field_descriptor)->label() == FieldDescriptor::LABEL_REPEATED)
+
+#define GOOGLE_CHECK_GET_INT32(arg, value, err)                        \
+    int32 value;                                            \
+    if (!CheckAndGetInteger(arg, &value, kint32min_py, kint32max_py)) { \
+      return err;                                          \
+    }
+
+#define GOOGLE_CHECK_GET_INT64(arg, value, err)                        \
+    int64 value;                                            \
+    if (!CheckAndGetInteger(arg, &value, kint64min_py, kint64max_py)) { \
+      return err;                                          \
+    }
+
+#define GOOGLE_CHECK_GET_UINT32(arg, value, err)                       \
+    uint32 value;                                           \
+    if (!CheckAndGetInteger(arg, &value, kPythonZero, kuint32max_py)) { \
+      return err;                                          \
+    }
+
+#define GOOGLE_CHECK_GET_UINT64(arg, value, err)                       \
+    uint64 value;                                           \
+    if (!CheckAndGetInteger(arg, &value, kPythonZero, kuint64max_py)) { \
+      return err;                                          \
+    }
+
+#define GOOGLE_CHECK_GET_FLOAT(arg, value, err)                        \
+    float value;                                            \
+    if (!CheckAndGetFloat(arg, &value)) {                   \
+      return err;                                          \
+    }                                                       \
+
+#define GOOGLE_CHECK_GET_DOUBLE(arg, value, err)                       \
+    double value;                                           \
+    if (!CheckAndGetDouble(arg, &value)) {                  \
+      return err;                                          \
+    }
+
+#define GOOGLE_CHECK_GET_BOOL(arg, value, err)                         \
+    bool value;                                             \
+    if (!CheckAndGetBool(arg, &value)) {                    \
+      return err;                                          \
+    }
+
+
+extern PyObject* kPythonZero;
+extern PyObject* kint32min_py;
+extern PyObject* kint32max_py;
+extern PyObject* kuint32max_py;
+extern PyObject* kint64min_py;
+extern PyObject* kint64max_py;
+extern PyObject* kuint64max_py;
+
+#define FULL_MODULE_NAME "google.protobuf.pyext._message"
+
+void FormatTypeError(PyObject* arg, char* expected_types);
+template<class T>
+bool CheckAndGetInteger(
+    PyObject* arg, T* value, PyObject* min, PyObject* max);
+bool CheckAndGetDouble(PyObject* arg, double* value);
+bool CheckAndGetFloat(PyObject* arg, float* value);
+bool CheckAndGetBool(PyObject* arg, bool* value);
+PyObject* CheckString(PyObject* arg, const FieldDescriptor* descriptor);
+bool CheckAndSetString(
+    PyObject* arg, Message* message,
+    const FieldDescriptor* descriptor,
+    const Reflection* reflection,
+    bool append,
+    int index);
+PyObject* ToStringObject(const FieldDescriptor* descriptor, string value);
+
+// Check if the passed field descriptor belongs to the given message.
+// If not, return false and set a Python exception (a KeyError)
+bool CheckFieldBelongsToMessage(const FieldDescriptor* field_descriptor,
+                                const Message* message);
+
+extern PyObject* PickleError_class;
+
+}  // namespace python
+}  // namespace protobuf
+
+}  // namespace google
+#endif  // GOOGLE_PROTOBUF_PYTHON_CPP_MESSAGE_H__
diff --git a/third_party/protobuf/python/google/protobuf/pyext/proto2_api_test.proto b/third_party/protobuf/python/google/protobuf/pyext/proto2_api_test.proto
new file mode 100644
index 0000000000..18aecfb7d6
--- /dev/null
+++ b/third_party/protobuf/python/google/protobuf/pyext/proto2_api_test.proto
@@ -0,0 +1,40 @@
+// 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";
+
+import "google/protobuf/internal/cpp/proto1_api_test.proto";
+
+package google.protobuf.python.internal;
+
+message TestNestedProto1APIMessage {
+  optional int32 a = 1;
+  optional TestMessage.NestedMessage b = 2;
+}
diff --git a/third_party/protobuf/python/google/protobuf/pyext/python.proto b/third_party/protobuf/python/google/protobuf/pyext/python.proto
new file mode 100644
index 0000000000..cce645d71a
--- /dev/null
+++ b/third_party/protobuf/python/google/protobuf/pyext/python.proto
@@ -0,0 +1,68 @@
+// 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: tibell@google.com (Johan Tibell)
+//
+// These message definitions are used to exercises known corner cases
+// in the C++ implementation of the Python API.
+
+syntax = "proto2";
+
+package google.protobuf.python.internal;
+
+// Protos optimized for SPEED use a strict superset of the generated code
+// of equivalent ones optimized for CODE_SIZE, so we should optimize all our
+// tests for speed unless explicitly testing code size optimization.
+option optimize_for = SPEED;
+
+message TestAllTypes {
+  message NestedMessage {
+    optional int32 bb = 1;
+    optional ForeignMessage cc = 2;
+  }
+
+  repeated NestedMessage repeated_nested_message = 1;
+  optional NestedMessage optional_nested_message = 2;
+  optional int32 optional_int32 = 3;
+}
+
+message ForeignMessage {
+  optional int32 c = 1;
+  repeated int32 d = 2;
+}
+
+message TestAllExtensions {
+  extensions 1 to max;
+}
+
+extend TestAllExtensions {
+  optional TestAllTypes.NestedMessage optional_nested_message_extension = 1;
+  repeated TestAllTypes.NestedMessage repeated_nested_message_extension = 2;
+}
diff --git a/third_party/protobuf/python/google/protobuf/pyext/python_protobuf.h b/third_party/protobuf/python/google/protobuf/pyext/python_protobuf.h
new file mode 100644
index 0000000000..beb6e4604a
--- /dev/null
+++ b/third_party/protobuf/python/google/protobuf/pyext/python_protobuf.h
@@ -0,0 +1,57 @@
+// 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: qrczak@google.com (Marcin Kowalczyk)
+//
+// This module exposes the C proto inside the given Python proto, in
+// case the Python proto is implemented with a C proto.
+
+#ifndef GOOGLE_PROTOBUF_PYTHON_PYTHON_PROTOBUF_H__
+#define GOOGLE_PROTOBUF_PYTHON_PYTHON_PROTOBUF_H__
+
+#include <Python.h>
+
+namespace google {
+namespace protobuf {
+
+class Message;
+
+namespace python {
+
+// Return the pointer to the C proto inside the given Python proto,
+// or NULL when this is not a Python proto implemented with a C proto.
+const Message* GetCProtoInsidePyProto(PyObject* msg);
+Message* MutableCProtoInsidePyProto(PyObject* msg);
+
+}  // namespace python
+}  // namespace protobuf
+
+}  // namespace google
+#endif  // GOOGLE_PROTOBUF_PYTHON_PYTHON_PROTOBUF_H__
diff --git a/third_party/protobuf/python/google/protobuf/pyext/repeated_composite_container.cc b/third_party/protobuf/python/google/protobuf/pyext/repeated_composite_container.cc
new file mode 100644
index 0000000000..b01123b4c0
--- /dev/null
+++ b/third_party/protobuf/python/google/protobuf/pyext/repeated_composite_container.cc
@@ -0,0 +1,614 @@
+// 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: anuraag@google.com (Anuraag Agrawal)
+// Author: tibell@google.com (Johan Tibell)
+
+#include <google/protobuf/pyext/repeated_composite_container.h>
+
+#include <memory>
+#ifndef _SHARED_PTR_H
+#include <google/protobuf/stubs/shared_ptr.h>
+#endif
+
+#include <google/protobuf/stubs/logging.h>
+#include <google/protobuf/stubs/common.h>
+#include <google/protobuf/descriptor.h>
+#include <google/protobuf/dynamic_message.h>
+#include <google/protobuf/message.h>
+#include <google/protobuf/pyext/descriptor.h>
+#include <google/protobuf/pyext/descriptor_pool.h>
+#include <google/protobuf/pyext/message.h>
+#include <google/protobuf/pyext/scoped_pyobject_ptr.h>
+
+#if PY_MAJOR_VERSION >= 3
+  #define PyInt_Check PyLong_Check
+  #define PyInt_AsLong PyLong_AsLong
+  #define PyInt_FromLong PyLong_FromLong
+#endif
+
+namespace google {
+namespace protobuf {
+namespace python {
+
+namespace repeated_composite_container {
+
+// TODO(tibell): We might also want to check:
+//   GOOGLE_CHECK_NOTNULL((self)->owner.get());
+#define GOOGLE_CHECK_ATTACHED(self)             \
+  do {                                   \
+    GOOGLE_CHECK_NOTNULL((self)->message);      \
+    GOOGLE_CHECK_NOTNULL((self)->parent_field_descriptor); \
+  } while (0);
+
+#define GOOGLE_CHECK_RELEASED(self)             \
+  do {                                   \
+    GOOGLE_CHECK((self)->owner.get() == NULL);  \
+    GOOGLE_CHECK((self)->message == NULL);      \
+    GOOGLE_CHECK((self)->parent_field_descriptor == NULL); \
+    GOOGLE_CHECK((self)->parent == NULL);       \
+  } while (0);
+
+// ---------------------------------------------------------------------
+// len()
+
+static Py_ssize_t Length(RepeatedCompositeContainer* self) {
+  Message* message = self->message;
+  if (message != NULL) {
+    return message->GetReflection()->FieldSize(*message,
+                                               self->parent_field_descriptor);
+  } else {
+    // The container has been released (i.e. by a call to Clear() or
+    // ClearField() on the parent) and thus there's no message.
+    return PyList_GET_SIZE(self->child_messages);
+  }
+}
+
+// Returns 0 if successful; returns -1 and sets an exception if
+// unsuccessful.
+static int UpdateChildMessages(RepeatedCompositeContainer* self) {
+  if (self->message == NULL)
+    return 0;
+
+  // A MergeFrom on a parent message could have caused extra messages to be
+  // added in the underlying protobuf so add them to our list. They can never
+  // be removed in such a way so there's no need to worry about that.
+  Py_ssize_t message_length = Length(self);
+  Py_ssize_t child_length = PyList_GET_SIZE(self->child_messages);
+  Message* message = self->message;
+  const Reflection* reflection = message->GetReflection();
+  for (Py_ssize_t i = child_length; i < message_length; ++i) {
+    const Message& sub_message = reflection->GetRepeatedMessage(
+        *(self->message), self->parent_field_descriptor, i);
+    CMessage* cmsg = cmessage::NewEmptyMessage(self->subclass_init,
+                                               sub_message.GetDescriptor());
+    ScopedPyObjectPtr py_cmsg(reinterpret_cast<PyObject*>(cmsg));
+    if (cmsg == NULL) {
+      return -1;
+    }
+    cmsg->owner = self->owner;
+    cmsg->message = const_cast<Message*>(&sub_message);
+    cmsg->parent = self->parent;
+    if (PyList_Append(self->child_messages, py_cmsg.get()) < 0) {
+      return -1;
+    }
+  }
+  return 0;
+}
+
+// ---------------------------------------------------------------------
+// add()
+
+static PyObject* AddToAttached(RepeatedCompositeContainer* self,
+                               PyObject* args,
+                               PyObject* kwargs) {
+  GOOGLE_CHECK_ATTACHED(self);
+
+  if (UpdateChildMessages(self) < 0) {
+    return NULL;
+  }
+  if (cmessage::AssureWritable(self->parent) == -1)
+    return NULL;
+  Message* message = self->message;
+  Message* sub_message =
+      message->GetReflection()->AddMessage(message,
+                                           self->parent_field_descriptor);
+  CMessage* cmsg = cmessage::NewEmptyMessage(self->subclass_init,
+                                             sub_message->GetDescriptor());
+  if (cmsg == NULL)
+    return NULL;
+
+  cmsg->owner = self->owner;
+  cmsg->message = sub_message;
+  cmsg->parent = self->parent;
+  if (cmessage::InitAttributes(cmsg, kwargs) < 0) {
+    Py_DECREF(cmsg);
+    return NULL;
+  }
+
+  PyObject* py_cmsg = reinterpret_cast<PyObject*>(cmsg);
+  if (PyList_Append(self->child_messages, py_cmsg) < 0) {
+    Py_DECREF(py_cmsg);
+    return NULL;
+  }
+  return py_cmsg;
+}
+
+static PyObject* AddToReleased(RepeatedCompositeContainer* self,
+                               PyObject* args,
+                               PyObject* kwargs) {
+  GOOGLE_CHECK_RELEASED(self);
+
+  // Create a new Message detached from the rest.
+  PyObject* py_cmsg = PyEval_CallObjectWithKeywords(
+      self->subclass_init, NULL, kwargs);
+  if (py_cmsg == NULL)
+    return NULL;
+
+  if (PyList_Append(self->child_messages, py_cmsg) < 0) {
+    Py_DECREF(py_cmsg);
+    return NULL;
+  }
+  return py_cmsg;
+}
+
+PyObject* Add(RepeatedCompositeContainer* self,
+              PyObject* args,
+              PyObject* kwargs) {
+  if (self->message == NULL)
+    return AddToReleased(self, args, kwargs);
+  else
+    return AddToAttached(self, args, kwargs);
+}
+
+// ---------------------------------------------------------------------
+// extend()
+
+PyObject* Extend(RepeatedCompositeContainer* self, PyObject* value) {
+  cmessage::AssureWritable(self->parent);
+  if (UpdateChildMessages(self) < 0) {
+    return NULL;
+  }
+  ScopedPyObjectPtr iter(PyObject_GetIter(value));
+  if (iter == NULL) {
+    PyErr_SetString(PyExc_TypeError, "Value must be iterable");
+    return NULL;
+  }
+  ScopedPyObjectPtr next;
+  while ((next.reset(PyIter_Next(iter.get()))) != NULL) {
+    if (!PyObject_TypeCheck(next.get(), &CMessage_Type)) {
+      PyErr_SetString(PyExc_TypeError, "Not a cmessage");
+      return NULL;
+    }
+    ScopedPyObjectPtr new_message(Add(self, NULL, NULL));
+    if (new_message == NULL) {
+      return NULL;
+    }
+    CMessage* new_cmessage = reinterpret_cast<CMessage*>(new_message.get());
+    if (ScopedPyObjectPtr(cmessage::MergeFrom(new_cmessage, next.get())) ==
+        NULL) {
+      return NULL;
+    }
+  }
+  if (PyErr_Occurred()) {
+    return NULL;
+  }
+  Py_RETURN_NONE;
+}
+
+PyObject* MergeFrom(RepeatedCompositeContainer* self, PyObject* other) {
+  if (UpdateChildMessages(self) < 0) {
+    return NULL;
+  }
+  return Extend(self, other);
+}
+
+PyObject* Subscript(RepeatedCompositeContainer* self, PyObject* slice) {
+  if (UpdateChildMessages(self) < 0) {
+    return NULL;
+  }
+  // Just forward the call to the subscript-handling function of the
+  // list containing the child messages.
+  return PyObject_GetItem(self->child_messages, slice);
+}
+
+int AssignSubscript(RepeatedCompositeContainer* self,
+                    PyObject* slice,
+                    PyObject* value) {
+  if (UpdateChildMessages(self) < 0) {
+    return -1;
+  }
+  if (value != NULL) {
+    PyErr_SetString(PyExc_TypeError, "does not support assignment");
+    return -1;
+  }
+
+  // Delete from the underlying Message, if any.
+  if (self->parent != NULL) {
+    if (cmessage::InternalDeleteRepeatedField(self->parent,
+                                              self->parent_field_descriptor,
+                                              slice,
+                                              self->child_messages) < 0) {
+      return -1;
+    }
+  } else {
+    Py_ssize_t from;
+    Py_ssize_t to;
+    Py_ssize_t step;
+    Py_ssize_t length = Length(self);
+    Py_ssize_t slicelength;
+    if (PySlice_Check(slice)) {
+#if PY_MAJOR_VERSION >= 3
+      if (PySlice_GetIndicesEx(slice,
+#else
+      if (PySlice_GetIndicesEx(reinterpret_cast<PySliceObject*>(slice),
+#endif
+                               length, &from, &to, &step, &slicelength) == -1) {
+        return -1;
+      }
+      return PySequence_DelSlice(self->child_messages, from, to);
+    } else if (PyInt_Check(slice) || PyLong_Check(slice)) {
+      from = to = PyLong_AsLong(slice);
+      if (from < 0) {
+        from = to = length + from;
+      }
+      return PySequence_DelItem(self->child_messages, from);
+    }
+  }
+
+  return 0;
+}
+
+static PyObject* Remove(RepeatedCompositeContainer* self, PyObject* value) {
+  if (UpdateChildMessages(self) < 0) {
+    return NULL;
+  }
+  Py_ssize_t index = PySequence_Index(self->child_messages, value);
+  if (index == -1) {
+    return NULL;
+  }
+  ScopedPyObjectPtr py_index(PyLong_FromLong(index));
+  if (AssignSubscript(self, py_index.get(), NULL) < 0) {
+    return NULL;
+  }
+  Py_RETURN_NONE;
+}
+
+static PyObject* RichCompare(RepeatedCompositeContainer* self,
+                             PyObject* other,
+                             int opid) {
+  if (UpdateChildMessages(self) < 0) {
+    return NULL;
+  }
+  if (!PyObject_TypeCheck(other, &RepeatedCompositeContainer_Type)) {
+    PyErr_SetString(PyExc_TypeError,
+                    "Can only compare repeated composite fields "
+                    "against other repeated composite fields.");
+    return NULL;
+  }
+  if (opid == Py_EQ || opid == Py_NE) {
+    // TODO(anuraag): Don't make new lists just for this...
+    ScopedPyObjectPtr full_slice(PySlice_New(NULL, NULL, NULL));
+    if (full_slice == NULL) {
+      return NULL;
+    }
+    ScopedPyObjectPtr list(Subscript(self, full_slice.get()));
+    if (list == NULL) {
+      return NULL;
+    }
+    ScopedPyObjectPtr other_list(
+        Subscript(reinterpret_cast<RepeatedCompositeContainer*>(other),
+                  full_slice.get()));
+    if (other_list == NULL) {
+      return NULL;
+    }
+    return PyObject_RichCompare(list.get(), other_list.get(), opid);
+  } else {
+    Py_INCREF(Py_NotImplemented);
+    return Py_NotImplemented;
+  }
+}
+
+// ---------------------------------------------------------------------
+// sort()
+
+static void ReorderAttached(RepeatedCompositeContainer* self) {
+  Message* message = self->message;
+  const Reflection* reflection = message->GetReflection();
+  const FieldDescriptor* descriptor = self->parent_field_descriptor;
+  const Py_ssize_t length = Length(self);
+
+  // Since Python protobuf objects are never arena-allocated, adding and
+  // removing message pointers to the underlying array is just updating
+  // pointers.
+  for (Py_ssize_t i = 0; i < length; ++i)
+    reflection->ReleaseLast(message, descriptor);
+
+  for (Py_ssize_t i = 0; i < length; ++i) {
+    CMessage* py_cmsg = reinterpret_cast<CMessage*>(
+        PyList_GET_ITEM(self->child_messages, i));
+    reflection->AddAllocatedMessage(message, descriptor, py_cmsg->message);
+  }
+}
+
+// Returns 0 if successful; returns -1 and sets an exception if
+// unsuccessful.
+static int SortPythonMessages(RepeatedCompositeContainer* self,
+                               PyObject* args,
+                               PyObject* kwds) {
+  ScopedPyObjectPtr m(PyObject_GetAttrString(self->child_messages, "sort"));
+  if (m == NULL)
+    return -1;
+  if (PyObject_Call(m.get(), args, kwds) == NULL)
+    return -1;
+  if (self->message != NULL) {
+    ReorderAttached(self);
+  }
+  return 0;
+}
+
+static PyObject* Sort(RepeatedCompositeContainer* self,
+                      PyObject* args,
+                      PyObject* kwds) {
+  // Support the old sort_function argument for backwards
+  // compatibility.
+  if (kwds != NULL) {
+    PyObject* sort_func = PyDict_GetItemString(kwds, "sort_function");
+    if (sort_func != NULL) {
+      // Must set before deleting as sort_func is a borrowed reference
+      // and kwds might be the only thing keeping it alive.
+      PyDict_SetItemString(kwds, "cmp", sort_func);
+      PyDict_DelItemString(kwds, "sort_function");
+    }
+  }
+
+  if (UpdateChildMessages(self) < 0) {
+    return NULL;
+  }
+  if (SortPythonMessages(self, args, kwds) < 0) {
+    return NULL;
+  }
+  Py_RETURN_NONE;
+}
+
+// ---------------------------------------------------------------------
+
+static PyObject* Item(RepeatedCompositeContainer* self, Py_ssize_t index) {
+  if (UpdateChildMessages(self) < 0) {
+    return NULL;
+  }
+  Py_ssize_t length = Length(self);
+  if (index < 0) {
+    index = length + index;
+  }
+  PyObject* item = PyList_GetItem(self->child_messages, index);
+  if (item == NULL) {
+    return NULL;
+  }
+  Py_INCREF(item);
+  return item;
+}
+
+static PyObject* Pop(RepeatedCompositeContainer* self,
+                     PyObject* args) {
+  Py_ssize_t index = -1;
+  if (!PyArg_ParseTuple(args, "|n", &index)) {
+    return NULL;
+  }
+  PyObject* item = Item(self, index);
+  if (item == NULL) {
+    PyErr_Format(PyExc_IndexError,
+                 "list index (%zd) out of range",
+                 index);
+    return NULL;
+  }
+  ScopedPyObjectPtr py_index(PyLong_FromSsize_t(index));
+  if (AssignSubscript(self, py_index.get(), NULL) < 0) {
+    return NULL;
+  }
+  return item;
+}
+
+// Release field of parent message and transfer the ownership to target.
+void ReleaseLastTo(CMessage* parent,
+                   const FieldDescriptor* field,
+                   CMessage* target) {
+  GOOGLE_CHECK_NOTNULL(parent);
+  GOOGLE_CHECK_NOTNULL(field);
+  GOOGLE_CHECK_NOTNULL(target);
+
+  shared_ptr<Message> released_message(
+      parent->message->GetReflection()->ReleaseLast(parent->message, field));
+  // TODO(tibell): Deal with proto1.
+
+  target->parent = NULL;
+  target->parent_field_descriptor = NULL;
+  target->message = released_message.get();
+  target->read_only = false;
+  cmessage::SetOwner(target, released_message);
+}
+
+// Called to release a container using
+// ClearField('container_field_name') on the parent.
+int Release(RepeatedCompositeContainer* self) {
+  if (UpdateChildMessages(self) < 0) {
+    PyErr_WriteUnraisable(PyBytes_FromString("Failed to update released "
+                                             "messages"));
+    return -1;
+  }
+
+  Message* message = self->message;
+  const FieldDescriptor* field = self->parent_field_descriptor;
+
+  // The reflection API only lets us release the last message in a
+  // repeated field.  Therefore we iterate through the children
+  // starting with the last one.
+  const Py_ssize_t size = PyList_GET_SIZE(self->child_messages);
+  GOOGLE_DCHECK_EQ(size, message->GetReflection()->FieldSize(*message, field));
+  for (Py_ssize_t i = size - 1; i >= 0; --i) {
+    CMessage* child_cmessage = reinterpret_cast<CMessage*>(
+        PyList_GET_ITEM(self->child_messages, i));
+    ReleaseLastTo(self->parent, field, child_cmessage);
+  }
+
+  // Detach from containing message.
+  self->parent = NULL;
+  self->parent_field_descriptor = NULL;
+  self->message = NULL;
+  self->owner.reset();
+
+  return 0;
+}
+
+int SetOwner(RepeatedCompositeContainer* self,
+             const shared_ptr<Message>& new_owner) {
+  GOOGLE_CHECK_ATTACHED(self);
+
+  self->owner = new_owner;
+  const Py_ssize_t n = PyList_GET_SIZE(self->child_messages);
+  for (Py_ssize_t i = 0; i < n; ++i) {
+    PyObject* msg = PyList_GET_ITEM(self->child_messages, i);
+    if (cmessage::SetOwner(reinterpret_cast<CMessage*>(msg), new_owner) == -1) {
+      return -1;
+    }
+  }
+  return 0;
+}
+
+// The private constructor of RepeatedCompositeContainer objects.
+PyObject *NewContainer(
+    CMessage* parent,
+    const FieldDescriptor* parent_field_descriptor,
+    PyObject *concrete_class) {
+  if (!CheckFieldBelongsToMessage(parent_field_descriptor, parent->message)) {
+    return NULL;
+  }
+
+  RepeatedCompositeContainer* self =
+      reinterpret_cast<RepeatedCompositeContainer*>(
+          PyType_GenericAlloc(&RepeatedCompositeContainer_Type, 0));
+  if (self == NULL) {
+    return NULL;
+  }
+
+  self->message = parent->message;
+  self->parent = parent;
+  self->parent_field_descriptor = parent_field_descriptor;
+  self->owner = parent->owner;
+  Py_INCREF(concrete_class);
+  self->subclass_init = concrete_class;
+  self->child_messages = PyList_New(0);
+
+  return reinterpret_cast<PyObject*>(self);
+}
+
+static void Dealloc(RepeatedCompositeContainer* self) {
+  Py_CLEAR(self->child_messages);
+  Py_CLEAR(self->subclass_init);
+  // TODO(tibell): Do we need to call delete on these objects to make
+  // sure their destructors are called?
+  self->owner.reset();
+
+  Py_TYPE(self)->tp_free(reinterpret_cast<PyObject*>(self));
+}
+
+static PySequenceMethods SqMethods = {
+  (lenfunc)Length,        /* sq_length */
+  0, /* sq_concat */
+  0, /* sq_repeat */
+  (ssizeargfunc)Item /* sq_item */
+};
+
+static PyMappingMethods MpMethods = {
+  (lenfunc)Length,               /* mp_length */
+  (binaryfunc)Subscript,      /* mp_subscript */
+  (objobjargproc)AssignSubscript,/* mp_ass_subscript */
+};
+
+static PyMethodDef Methods[] = {
+  { "add", (PyCFunction) Add, METH_VARARGS | METH_KEYWORDS,
+    "Adds an object to the repeated container." },
+  { "extend", (PyCFunction) Extend, METH_O,
+    "Adds objects to the repeated container." },
+  { "pop", (PyCFunction)Pop, METH_VARARGS,
+    "Removes an object from the repeated container and returns it." },
+  { "remove", (PyCFunction) Remove, METH_O,
+    "Removes an object from the repeated container." },
+  { "sort", (PyCFunction) Sort, METH_VARARGS | METH_KEYWORDS,
+    "Sorts the repeated container." },
+  { "MergeFrom", (PyCFunction) MergeFrom, METH_O,
+    "Adds objects to the repeated container." },
+  { NULL, NULL }
+};
+
+}  // namespace repeated_composite_container
+
+PyTypeObject RepeatedCompositeContainer_Type = {
+  PyVarObject_HEAD_INIT(&PyType_Type, 0)
+  FULL_MODULE_NAME ".RepeatedCompositeContainer",  // tp_name
+  sizeof(RepeatedCompositeContainer),  // tp_basicsize
+  0,                                   //  tp_itemsize
+  (destructor)repeated_composite_container::Dealloc,  //  tp_dealloc
+  0,                                   //  tp_print
+  0,                                   //  tp_getattr
+  0,                                   //  tp_setattr
+  0,                                   //  tp_compare
+  0,                                   //  tp_repr
+  0,                                   //  tp_as_number
+  &repeated_composite_container::SqMethods,   //  tp_as_sequence
+  &repeated_composite_container::MpMethods,   //  tp_as_mapping
+  PyObject_HashNotImplemented,         //  tp_hash
+  0,                                   //  tp_call
+  0,                                   //  tp_str
+  0,                                   //  tp_getattro
+  0,                                   //  tp_setattro
+  0,                                   //  tp_as_buffer
+  Py_TPFLAGS_DEFAULT,                  //  tp_flags
+  "A Repeated scalar container",       //  tp_doc
+  0,                                   //  tp_traverse
+  0,                                   //  tp_clear
+  (richcmpfunc)repeated_composite_container::RichCompare,  //  tp_richcompare
+  0,                                   //  tp_weaklistoffset
+  0,                                   //  tp_iter
+  0,                                   //  tp_iternext
+  repeated_composite_container::Methods,   //  tp_methods
+  0,                                   //  tp_members
+  0,                                   //  tp_getset
+  0,                                   //  tp_base
+  0,                                   //  tp_dict
+  0,                                   //  tp_descr_get
+  0,                                   //  tp_descr_set
+  0,                                   //  tp_dictoffset
+  0,                                   //  tp_init
+};
+
+}  // namespace python
+}  // namespace protobuf
+}  // namespace google
diff --git a/third_party/protobuf/python/google/protobuf/pyext/repeated_composite_container.h b/third_party/protobuf/python/google/protobuf/pyext/repeated_composite_container.h
new file mode 100644
index 0000000000..58d37b0207
--- /dev/null
+++ b/third_party/protobuf/python/google/protobuf/pyext/repeated_composite_container.h
@@ -0,0 +1,178 @@
+// 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: anuraag@google.com (Anuraag Agrawal)
+// Author: tibell@google.com (Johan Tibell)
+
+#ifndef GOOGLE_PROTOBUF_PYTHON_CPP_REPEATED_COMPOSITE_CONTAINER_H__
+#define GOOGLE_PROTOBUF_PYTHON_CPP_REPEATED_COMPOSITE_CONTAINER_H__
+
+#include <Python.h>
+
+#include <memory>
+#ifndef _SHARED_PTR_H
+#include <google/protobuf/stubs/shared_ptr.h>
+#endif
+#include <string>
+#include <vector>
+
+namespace google {
+namespace protobuf {
+
+class FieldDescriptor;
+class Message;
+
+#ifdef _SHARED_PTR_H
+using std::shared_ptr;
+#else
+using internal::shared_ptr;
+#endif
+
+namespace python {
+
+struct CMessage;
+
+// A RepeatedCompositeContainer can be in one of two states: attached
+// or released.
+//
+// When in the attached state all modifications to the container are
+// done both on the 'message' and on the 'child_messages'
+// list.  In this state all Messages referred to by the children in
+// 'child_messages' are owner by the 'owner'.
+//
+// When in the released state 'message', 'owner', 'parent', and
+// 'parent_field_descriptor' are NULL.
+typedef struct RepeatedCompositeContainer {
+  PyObject_HEAD;
+
+  // This is the top-level C++ Message object that owns the whole
+  // proto tree.  Every Python RepeatedCompositeContainer holds a
+  // reference to it in order to keep it alive as long as there's a
+  // Python object that references any part of the tree.
+  shared_ptr<Message> owner;
+
+  // Weak reference to parent object. May be NULL. Used to make sure
+  // the parent is writable before modifying the
+  // RepeatedCompositeContainer.
+  CMessage* parent;
+
+  // A descriptor used to modify the underlying 'message'.
+  // The pointer is owned by the global DescriptorPool.
+  const FieldDescriptor* parent_field_descriptor;
+
+  // Pointer to the C++ Message that contains this container.  The
+  // RepeatedCompositeContainer does not own this pointer.
+  //
+  // If NULL, this message has been released from its parent (by
+  // calling Clear() or ClearField() on the parent.
+  Message* message;
+
+  // A callable that is used to create new child messages.
+  PyObject* subclass_init;
+
+  // A list of child messages.
+  PyObject* child_messages;
+} RepeatedCompositeContainer;
+
+extern PyTypeObject RepeatedCompositeContainer_Type;
+
+namespace repeated_composite_container {
+
+// Builds a RepeatedCompositeContainer object, from a parent message and a
+// field descriptor.
+PyObject *NewContainer(
+    CMessage* parent,
+    const FieldDescriptor* parent_field_descriptor,
+    PyObject *concrete_class);
+
+// Appends a new CMessage to the container and returns it.  The
+// CMessage is initialized using the content of kwargs.
+//
+// Returns a new reference if successful; returns NULL and sets an
+// exception if unsuccessful.
+PyObject* Add(RepeatedCompositeContainer* self,
+              PyObject* args,
+              PyObject* kwargs);
+
+// Appends all the CMessages in the input iterator to the container.
+//
+// Returns None if successful; returns NULL and sets an exception if
+// unsuccessful.
+PyObject* Extend(RepeatedCompositeContainer* self, PyObject* value);
+
+// Appends a new message to the container for each message in the
+// input iterator, merging each data element in. Equivalent to extend.
+//
+// Returns None if successful; returns NULL and sets an exception if
+// unsuccessful.
+PyObject* MergeFrom(RepeatedCompositeContainer* self, PyObject* other);
+
+// Accesses messages in the container.
+//
+// Returns a new reference to the message for an integer parameter.
+// Returns a new reference to a list of messages for a slice.
+PyObject* Subscript(RepeatedCompositeContainer* self, PyObject* slice);
+
+// Deletes items from the container (cannot be used for assignment).
+//
+// Returns 0 on success, -1 on failure.
+int AssignSubscript(RepeatedCompositeContainer* self,
+                    PyObject* slice,
+                    PyObject* value);
+
+// Releases the messages in the container to the given message.
+//
+// Returns 0 on success, -1 on failure.
+int ReleaseToMessage(RepeatedCompositeContainer* self, Message* new_message);
+
+// Releases the messages in the container to a new message.
+//
+// Returns 0 on success, -1 on failure.
+int Release(RepeatedCompositeContainer* self);
+
+// Returns 0 on success, -1 on failure.
+int SetOwner(RepeatedCompositeContainer* self,
+             const shared_ptr<Message>& new_owner);
+
+// Removes the last element of the repeated message field 'field' on
+// the Message 'parent', and transfers the ownership of the released
+// Message to 'target'.
+//
+// Corresponds to reflection api method ReleaseMessage.
+void ReleaseLastTo(CMessage* parent,
+                   const FieldDescriptor* field,
+                   CMessage* target);
+
+}  // namespace repeated_composite_container
+}  // namespace python
+}  // namespace protobuf
+
+}  // namespace google
+#endif  // GOOGLE_PROTOBUF_PYTHON_CPP_REPEATED_COMPOSITE_CONTAINER_H__
diff --git a/third_party/protobuf/python/google/protobuf/pyext/repeated_scalar_container.cc b/third_party/protobuf/python/google/protobuf/pyext/repeated_scalar_container.cc
new file mode 100644
index 0000000000..95da85f87b
--- /dev/null
+++ b/third_party/protobuf/python/google/protobuf/pyext/repeated_scalar_container.cc
@@ -0,0 +1,812 @@
+// 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: anuraag@google.com (Anuraag Agrawal)
+// Author: tibell@google.com (Johan Tibell)
+
+#include <google/protobuf/pyext/repeated_scalar_container.h>
+
+#include <memory>
+#ifndef _SHARED_PTR_H
+#include <google/protobuf/stubs/shared_ptr.h>
+#endif
+
+#include <google/protobuf/stubs/common.h>
+#include <google/protobuf/stubs/logging.h>
+#include <google/protobuf/descriptor.h>
+#include <google/protobuf/dynamic_message.h>
+#include <google/protobuf/message.h>
+#include <google/protobuf/pyext/descriptor.h>
+#include <google/protobuf/pyext/descriptor_pool.h>
+#include <google/protobuf/pyext/message.h>
+#include <google/protobuf/pyext/scoped_pyobject_ptr.h>
+
+#if PY_MAJOR_VERSION >= 3
+  #define PyInt_FromLong PyLong_FromLong
+  #if PY_VERSION_HEX < 0x03030000
+    #error "Python 3.0 - 3.2 are not supported."
+  #else
+  #define PyString_AsString(ob) \
+    (PyUnicode_Check(ob)? PyUnicode_AsUTF8(ob): PyBytes_AsString(ob))
+  #endif
+#endif
+
+namespace google {
+namespace protobuf {
+namespace python {
+
+namespace repeated_scalar_container {
+
+static int InternalAssignRepeatedField(
+    RepeatedScalarContainer* self, PyObject* list) {
+  self->message->GetReflection()->ClearField(self->message,
+                                             self->parent_field_descriptor);
+  for (Py_ssize_t i = 0; i < PyList_GET_SIZE(list); ++i) {
+    PyObject* value = PyList_GET_ITEM(list, i);
+    if (ScopedPyObjectPtr(Append(self, value)) == NULL) {
+      return -1;
+    }
+  }
+  return 0;
+}
+
+static Py_ssize_t Len(RepeatedScalarContainer* self) {
+  Message* message = self->message;
+  return message->GetReflection()->FieldSize(*message,
+                                             self->parent_field_descriptor);
+}
+
+static int AssignItem(RepeatedScalarContainer* self,
+                      Py_ssize_t index,
+                      PyObject* arg) {
+  cmessage::AssureWritable(self->parent);
+  Message* message = self->message;
+  const FieldDescriptor* field_descriptor = self->parent_field_descriptor;
+
+  const Reflection* reflection = message->GetReflection();
+  int field_size = reflection->FieldSize(*message, field_descriptor);
+  if (index < 0) {
+    index = field_size + index;
+  }
+  if (index < 0 || index >= field_size) {
+    PyErr_Format(PyExc_IndexError,
+                 "list assignment index (%d) out of range",
+                 static_cast<int>(index));
+    return -1;
+  }
+
+  if (arg == NULL) {
+    ScopedPyObjectPtr py_index(PyLong_FromLong(index));
+    return cmessage::InternalDeleteRepeatedField(self->parent, field_descriptor,
+                                                 py_index.get(), NULL);
+  }
+
+  if (PySequence_Check(arg) && !(PyBytes_Check(arg) || PyUnicode_Check(arg))) {
+    PyErr_SetString(PyExc_TypeError, "Value must be scalar");
+    return -1;
+  }
+
+  switch (field_descriptor->cpp_type()) {
+    case FieldDescriptor::CPPTYPE_INT32: {
+      GOOGLE_CHECK_GET_INT32(arg, value, -1);
+      reflection->SetRepeatedInt32(message, field_descriptor, index, value);
+      break;
+    }
+    case FieldDescriptor::CPPTYPE_INT64: {
+      GOOGLE_CHECK_GET_INT64(arg, value, -1);
+      reflection->SetRepeatedInt64(message, field_descriptor, index, value);
+      break;
+    }
+    case FieldDescriptor::CPPTYPE_UINT32: {
+      GOOGLE_CHECK_GET_UINT32(arg, value, -1);
+      reflection->SetRepeatedUInt32(message, field_descriptor, index, value);
+      break;
+    }
+    case FieldDescriptor::CPPTYPE_UINT64: {
+      GOOGLE_CHECK_GET_UINT64(arg, value, -1);
+      reflection->SetRepeatedUInt64(message, field_descriptor, index, value);
+      break;
+    }
+    case FieldDescriptor::CPPTYPE_FLOAT: {
+      GOOGLE_CHECK_GET_FLOAT(arg, value, -1);
+      reflection->SetRepeatedFloat(message, field_descriptor, index, value);
+      break;
+    }
+    case FieldDescriptor::CPPTYPE_DOUBLE: {
+      GOOGLE_CHECK_GET_DOUBLE(arg, value, -1);
+      reflection->SetRepeatedDouble(message, field_descriptor, index, value);
+      break;
+    }
+    case FieldDescriptor::CPPTYPE_BOOL: {
+      GOOGLE_CHECK_GET_BOOL(arg, value, -1);
+      reflection->SetRepeatedBool(message, field_descriptor, index, value);
+      break;
+    }
+    case FieldDescriptor::CPPTYPE_STRING: {
+      if (!CheckAndSetString(
+          arg, message, field_descriptor, reflection, false, index)) {
+        return -1;
+      }
+      break;
+    }
+    case FieldDescriptor::CPPTYPE_ENUM: {
+      GOOGLE_CHECK_GET_INT32(arg, value, -1);
+      if (reflection->SupportsUnknownEnumValues()) {
+        reflection->SetRepeatedEnumValue(message, field_descriptor, index,
+                                         value);
+      } else {
+        const EnumDescriptor* enum_descriptor = field_descriptor->enum_type();
+        const EnumValueDescriptor* enum_value =
+            enum_descriptor->FindValueByNumber(value);
+        if (enum_value != NULL) {
+          reflection->SetRepeatedEnum(message, field_descriptor, index,
+                                      enum_value);
+        } else {
+          ScopedPyObjectPtr s(PyObject_Str(arg));
+          if (s != NULL) {
+            PyErr_Format(PyExc_ValueError, "Unknown enum value: %s",
+                         PyString_AsString(s.get()));
+          }
+          return -1;
+        }
+      }
+      break;
+    }
+    default:
+      PyErr_Format(
+          PyExc_SystemError, "Adding value to a field of unknown type %d",
+          field_descriptor->cpp_type());
+      return -1;
+  }
+  return 0;
+}
+
+static PyObject* Item(RepeatedScalarContainer* self, Py_ssize_t index) {
+  Message* message = self->message;
+  const FieldDescriptor* field_descriptor = self->parent_field_descriptor;
+  const Reflection* reflection = message->GetReflection();
+
+  int field_size = reflection->FieldSize(*message, field_descriptor);
+  if (index < 0) {
+    index = field_size + index;
+  }
+  if (index < 0 || index >= field_size) {
+    PyErr_Format(PyExc_IndexError,
+                 "list index (%zd) out of range",
+                 index);
+    return NULL;
+  }
+
+  PyObject* result = NULL;
+  switch (field_descriptor->cpp_type()) {
+    case FieldDescriptor::CPPTYPE_INT32: {
+      int32 value = reflection->GetRepeatedInt32(
+          *message, field_descriptor, index);
+      result = PyInt_FromLong(value);
+      break;
+    }
+    case FieldDescriptor::CPPTYPE_INT64: {
+      int64 value = reflection->GetRepeatedInt64(
+          *message, field_descriptor, index);
+      result = PyLong_FromLongLong(value);
+      break;
+    }
+    case FieldDescriptor::CPPTYPE_UINT32: {
+      uint32 value = reflection->GetRepeatedUInt32(
+          *message, field_descriptor, index);
+      result = PyLong_FromLongLong(value);
+      break;
+    }
+    case FieldDescriptor::CPPTYPE_UINT64: {
+      uint64 value = reflection->GetRepeatedUInt64(
+          *message, field_descriptor, index);
+      result = PyLong_FromUnsignedLongLong(value);
+      break;
+    }
+    case FieldDescriptor::CPPTYPE_FLOAT: {
+      float value = reflection->GetRepeatedFloat(
+          *message, field_descriptor, index);
+      result = PyFloat_FromDouble(value);
+      break;
+    }
+    case FieldDescriptor::CPPTYPE_DOUBLE: {
+      double value = reflection->GetRepeatedDouble(
+          *message, field_descriptor, index);
+      result = PyFloat_FromDouble(value);
+      break;
+    }
+    case FieldDescriptor::CPPTYPE_BOOL: {
+      bool value = reflection->GetRepeatedBool(
+          *message, field_descriptor, index);
+      result = PyBool_FromLong(value ? 1 : 0);
+      break;
+    }
+    case FieldDescriptor::CPPTYPE_ENUM: {
+      const EnumValueDescriptor* enum_value =
+          message->GetReflection()->GetRepeatedEnum(
+              *message, field_descriptor, index);
+      result = PyInt_FromLong(enum_value->number());
+      break;
+    }
+    case FieldDescriptor::CPPTYPE_STRING: {
+      string value = reflection->GetRepeatedString(
+          *message, field_descriptor, index);
+      result = ToStringObject(field_descriptor, value);
+      break;
+    }
+    case FieldDescriptor::CPPTYPE_MESSAGE: {
+      PyObject* py_cmsg = PyObject_CallObject(reinterpret_cast<PyObject*>(
+          &CMessage_Type), NULL);
+      if (py_cmsg == NULL) {
+        return NULL;
+      }
+      CMessage* cmsg = reinterpret_cast<CMessage*>(py_cmsg);
+      const Message& msg = reflection->GetRepeatedMessage(
+          *message, field_descriptor, index);
+      cmsg->owner = self->owner;
+      cmsg->parent = self->parent;
+      cmsg->message = const_cast<Message*>(&msg);
+      cmsg->read_only = false;
+      result = reinterpret_cast<PyObject*>(py_cmsg);
+      break;
+    }
+    default:
+      PyErr_Format(
+          PyExc_SystemError,
+          "Getting value from a repeated field of unknown type %d",
+          field_descriptor->cpp_type());
+  }
+
+  return result;
+}
+
+static PyObject* Subscript(RepeatedScalarContainer* self, PyObject* slice) {
+  Py_ssize_t from;
+  Py_ssize_t to;
+  Py_ssize_t step;
+  Py_ssize_t length;
+  Py_ssize_t slicelength;
+  bool return_list = false;
+#if PY_MAJOR_VERSION < 3
+  if (PyInt_Check(slice)) {
+    from = to = PyInt_AsLong(slice);
+  } else  // NOLINT
+#endif
+  if (PyLong_Check(slice)) {
+    from = to = PyLong_AsLong(slice);
+  } else if (PySlice_Check(slice)) {
+    length = Len(self);
+#if PY_MAJOR_VERSION >= 3
+    if (PySlice_GetIndicesEx(slice,
+#else
+    if (PySlice_GetIndicesEx(reinterpret_cast<PySliceObject*>(slice),
+#endif
+                             length, &from, &to, &step, &slicelength) == -1) {
+      return NULL;
+    }
+    return_list = true;
+  } else {
+    PyErr_SetString(PyExc_TypeError, "list indices must be integers");
+    return NULL;
+  }
+
+  if (!return_list) {
+    return Item(self, from);
+  }
+
+  PyObject* list = PyList_New(0);
+  if (list == NULL) {
+    return NULL;
+  }
+  if (from <= to) {
+    if (step < 0) {
+      return list;
+    }
+    for (Py_ssize_t index = from; index < to; index += step) {
+      if (index < 0 || index >= length) {
+        break;
+      }
+      ScopedPyObjectPtr s(Item(self, index));
+      PyList_Append(list, s.get());
+    }
+  } else {
+    if (step > 0) {
+      return list;
+    }
+    for (Py_ssize_t index = from; index > to; index += step) {
+      if (index < 0 || index >= length) {
+        break;
+      }
+      ScopedPyObjectPtr s(Item(self, index));
+      PyList_Append(list, s.get());
+    }
+  }
+  return list;
+}
+
+PyObject* Append(RepeatedScalarContainer* self, PyObject* item) {
+  cmessage::AssureWritable(self->parent);
+  Message* message = self->message;
+  const FieldDescriptor* field_descriptor = self->parent_field_descriptor;
+
+  const Reflection* reflection = message->GetReflection();
+  switch (field_descriptor->cpp_type()) {
+    case FieldDescriptor::CPPTYPE_INT32: {
+      GOOGLE_CHECK_GET_INT32(item, value, NULL);
+      reflection->AddInt32(message, field_descriptor, value);
+      break;
+    }
+    case FieldDescriptor::CPPTYPE_INT64: {
+      GOOGLE_CHECK_GET_INT64(item, value, NULL);
+      reflection->AddInt64(message, field_descriptor, value);
+      break;
+    }
+    case FieldDescriptor::CPPTYPE_UINT32: {
+      GOOGLE_CHECK_GET_UINT32(item, value, NULL);
+      reflection->AddUInt32(message, field_descriptor, value);
+      break;
+    }
+    case FieldDescriptor::CPPTYPE_UINT64: {
+      GOOGLE_CHECK_GET_UINT64(item, value, NULL);
+      reflection->AddUInt64(message, field_descriptor, value);
+      break;
+    }
+    case FieldDescriptor::CPPTYPE_FLOAT: {
+      GOOGLE_CHECK_GET_FLOAT(item, value, NULL);
+      reflection->AddFloat(message, field_descriptor, value);
+      break;
+    }
+    case FieldDescriptor::CPPTYPE_DOUBLE: {
+      GOOGLE_CHECK_GET_DOUBLE(item, value, NULL);
+      reflection->AddDouble(message, field_descriptor, value);
+      break;
+    }
+    case FieldDescriptor::CPPTYPE_BOOL: {
+      GOOGLE_CHECK_GET_BOOL(item, value, NULL);
+      reflection->AddBool(message, field_descriptor, value);
+      break;
+    }
+    case FieldDescriptor::CPPTYPE_STRING: {
+      if (!CheckAndSetString(
+          item, message, field_descriptor, reflection, true, -1)) {
+        return NULL;
+      }
+      break;
+    }
+    case FieldDescriptor::CPPTYPE_ENUM: {
+      GOOGLE_CHECK_GET_INT32(item, value, NULL);
+      if (reflection->SupportsUnknownEnumValues()) {
+        reflection->AddEnumValue(message, field_descriptor, value);
+      } else {
+        const EnumDescriptor* enum_descriptor = field_descriptor->enum_type();
+        const EnumValueDescriptor* enum_value =
+            enum_descriptor->FindValueByNumber(value);
+        if (enum_value != NULL) {
+          reflection->AddEnum(message, field_descriptor, enum_value);
+        } else {
+          ScopedPyObjectPtr s(PyObject_Str(item));
+          if (s != NULL) {
+            PyErr_Format(PyExc_ValueError, "Unknown enum value: %s",
+                         PyString_AsString(s.get()));
+          }
+          return NULL;
+        }
+      }
+      break;
+    }
+    default:
+      PyErr_Format(
+          PyExc_SystemError, "Adding value to a field of unknown type %d",
+          field_descriptor->cpp_type());
+      return NULL;
+  }
+
+  Py_RETURN_NONE;
+}
+
+static int AssSubscript(RepeatedScalarContainer* self,
+                        PyObject* slice,
+                        PyObject* value) {
+  Py_ssize_t from;
+  Py_ssize_t to;
+  Py_ssize_t step;
+  Py_ssize_t length;
+  Py_ssize_t slicelength;
+  bool create_list = false;
+
+  cmessage::AssureWritable(self->parent);
+  Message* message = self->message;
+  const FieldDescriptor* field_descriptor =
+      self->parent_field_descriptor;
+
+#if PY_MAJOR_VERSION < 3
+  if (PyInt_Check(slice)) {
+    from = to = PyInt_AsLong(slice);
+  } else
+#endif
+  if (PyLong_Check(slice)) {
+    from = to = PyLong_AsLong(slice);
+  } else if (PySlice_Check(slice)) {
+    const Reflection* reflection = message->GetReflection();
+    length = reflection->FieldSize(*message, field_descriptor);
+#if PY_MAJOR_VERSION >= 3
+    if (PySlice_GetIndicesEx(slice,
+#else
+    if (PySlice_GetIndicesEx(reinterpret_cast<PySliceObject*>(slice),
+#endif
+                             length, &from, &to, &step, &slicelength) == -1) {
+      return -1;
+    }
+    create_list = true;
+  } else {
+    PyErr_SetString(PyExc_TypeError, "list indices must be integers");
+    return -1;
+  }
+
+  if (value == NULL) {
+    return cmessage::InternalDeleteRepeatedField(
+        self->parent, field_descriptor, slice, NULL);
+  }
+
+  if (!create_list) {
+    return AssignItem(self, from, value);
+  }
+
+  ScopedPyObjectPtr full_slice(PySlice_New(NULL, NULL, NULL));
+  if (full_slice == NULL) {
+    return -1;
+  }
+  ScopedPyObjectPtr new_list(Subscript(self, full_slice.get()));
+  if (new_list == NULL) {
+    return -1;
+  }
+  if (PySequence_SetSlice(new_list.get(), from, to, value) < 0) {
+    return -1;
+  }
+
+  return InternalAssignRepeatedField(self, new_list.get());
+}
+
+PyObject* Extend(RepeatedScalarContainer* self, PyObject* value) {
+  cmessage::AssureWritable(self->parent);
+
+  // TODO(ptucker): Deprecate this behavior. b/18413862
+  if (value == Py_None) {
+    Py_RETURN_NONE;
+  }
+  if ((Py_TYPE(value)->tp_as_sequence == NULL) && PyObject_Not(value)) {
+    Py_RETURN_NONE;
+  }
+
+  ScopedPyObjectPtr iter(PyObject_GetIter(value));
+  if (iter == NULL) {
+    PyErr_SetString(PyExc_TypeError, "Value must be iterable");
+    return NULL;
+  }
+  ScopedPyObjectPtr next;
+  while ((next.reset(PyIter_Next(iter.get()))) != NULL) {
+    if (ScopedPyObjectPtr(Append(self, next.get())) == NULL) {
+      return NULL;
+    }
+  }
+  if (PyErr_Occurred()) {
+    return NULL;
+  }
+  Py_RETURN_NONE;
+}
+
+static PyObject* Insert(RepeatedScalarContainer* self, PyObject* args) {
+  Py_ssize_t index;
+  PyObject* value;
+  if (!PyArg_ParseTuple(args, "lO", &index, &value)) {
+    return NULL;
+  }
+  ScopedPyObjectPtr full_slice(PySlice_New(NULL, NULL, NULL));
+  ScopedPyObjectPtr new_list(Subscript(self, full_slice.get()));
+  if (PyList_Insert(new_list.get(), index, value) < 0) {
+    return NULL;
+  }
+  int ret = InternalAssignRepeatedField(self, new_list.get());
+  if (ret < 0) {
+    return NULL;
+  }
+  Py_RETURN_NONE;
+}
+
+static PyObject* Remove(RepeatedScalarContainer* self, PyObject* value) {
+  Py_ssize_t match_index = -1;
+  for (Py_ssize_t i = 0; i < Len(self); ++i) {
+    ScopedPyObjectPtr elem(Item(self, i));
+    if (PyObject_RichCompareBool(elem.get(), value, Py_EQ)) {
+      match_index = i;
+      break;
+    }
+  }
+  if (match_index == -1) {
+    PyErr_SetString(PyExc_ValueError, "remove(x): x not in container");
+    return NULL;
+  }
+  if (AssignItem(self, match_index, NULL) < 0) {
+    return NULL;
+  }
+  Py_RETURN_NONE;
+}
+
+static PyObject* RichCompare(RepeatedScalarContainer* self,
+                             PyObject* other,
+                             int opid) {
+  if (opid != Py_EQ && opid != Py_NE) {
+    Py_INCREF(Py_NotImplemented);
+    return Py_NotImplemented;
+  }
+
+  // Copy the contents of this repeated scalar container, and other if it is
+  // also a repeated scalar container, into Python lists so we can delegate
+  // to the list's compare method.
+
+  ScopedPyObjectPtr full_slice(PySlice_New(NULL, NULL, NULL));
+  if (full_slice == NULL) {
+    return NULL;
+  }
+
+  ScopedPyObjectPtr other_list_deleter;
+  if (PyObject_TypeCheck(other, &RepeatedScalarContainer_Type)) {
+    other_list_deleter.reset(Subscript(
+        reinterpret_cast<RepeatedScalarContainer*>(other), full_slice.get()));
+    other = other_list_deleter.get();
+  }
+
+  ScopedPyObjectPtr list(Subscript(self, full_slice.get()));
+  if (list == NULL) {
+    return NULL;
+  }
+  return PyObject_RichCompare(list.get(), other, opid);
+}
+
+PyObject* Reduce(RepeatedScalarContainer* unused_self) {
+  PyErr_Format(
+      PickleError_class,
+      "can't pickle repeated message fields, convert to list first");
+  return NULL;
+}
+
+static PyObject* Sort(RepeatedScalarContainer* self,
+                      PyObject* args,
+                      PyObject* kwds) {
+  // Support the old sort_function argument for backwards
+  // compatibility.
+  if (kwds != NULL) {
+    PyObject* sort_func = PyDict_GetItemString(kwds, "sort_function");
+    if (sort_func != NULL) {
+      // Must set before deleting as sort_func is a borrowed reference
+      // and kwds might be the only thing keeping it alive.
+      if (PyDict_SetItemString(kwds, "cmp", sort_func) == -1)
+        return NULL;
+      if (PyDict_DelItemString(kwds, "sort_function") == -1)
+        return NULL;
+    }
+  }
+
+  ScopedPyObjectPtr full_slice(PySlice_New(NULL, NULL, NULL));
+  if (full_slice == NULL) {
+    return NULL;
+  }
+  ScopedPyObjectPtr list(Subscript(self, full_slice.get()));
+  if (list == NULL) {
+    return NULL;
+  }
+  ScopedPyObjectPtr m(PyObject_GetAttrString(list.get(), "sort"));
+  if (m == NULL) {
+    return NULL;
+  }
+  ScopedPyObjectPtr res(PyObject_Call(m.get(), args, kwds));
+  if (res == NULL) {
+    return NULL;
+  }
+  int ret = InternalAssignRepeatedField(self, list.get());
+  if (ret < 0) {
+    return NULL;
+  }
+  Py_RETURN_NONE;
+}
+
+static PyObject* Pop(RepeatedScalarContainer* self,
+                     PyObject* args) {
+  Py_ssize_t index = -1;
+  if (!PyArg_ParseTuple(args, "|n", &index)) {
+    return NULL;
+  }
+  PyObject* item = Item(self, index);
+  if (item == NULL) {
+    PyErr_Format(PyExc_IndexError,
+                 "list index (%zd) out of range",
+                 index);
+    return NULL;
+  }
+  if (AssignItem(self, index, NULL) < 0) {
+    return NULL;
+  }
+  return item;
+}
+
+// The private constructor of RepeatedScalarContainer objects.
+PyObject *NewContainer(
+    CMessage* parent, const FieldDescriptor* parent_field_descriptor) {
+  if (!CheckFieldBelongsToMessage(parent_field_descriptor, parent->message)) {
+    return NULL;
+  }
+
+  RepeatedScalarContainer* self = reinterpret_cast<RepeatedScalarContainer*>(
+      PyType_GenericAlloc(&RepeatedScalarContainer_Type, 0));
+  if (self == NULL) {
+    return NULL;
+  }
+
+  self->message = parent->message;
+  self->parent = parent;
+  self->parent_field_descriptor = parent_field_descriptor;
+  self->owner = parent->owner;
+
+  return reinterpret_cast<PyObject*>(self);
+}
+
+// Initializes the underlying Message object of "to" so it becomes a new parent
+// repeated scalar, and copies all the values from "from" to it. A child scalar
+// container can be released by passing it as both from and to (e.g. making it
+// the recipient of the new parent message and copying the values from itself).
+static int InitializeAndCopyToParentContainer(
+    RepeatedScalarContainer* from,
+    RepeatedScalarContainer* to) {
+  ScopedPyObjectPtr full_slice(PySlice_New(NULL, NULL, NULL));
+  if (full_slice == NULL) {
+    return -1;
+  }
+  ScopedPyObjectPtr values(Subscript(from, full_slice.get()));
+  if (values == NULL) {
+    return -1;
+  }
+  Message* new_message = from->message->New();
+  to->parent = NULL;
+  to->parent_field_descriptor = from->parent_field_descriptor;
+  to->message = new_message;
+  to->owner.reset(new_message);
+  if (InternalAssignRepeatedField(to, values.get()) < 0) {
+    return -1;
+  }
+  return 0;
+}
+
+int Release(RepeatedScalarContainer* self) {
+  return InitializeAndCopyToParentContainer(self, self);
+}
+
+PyObject* DeepCopy(RepeatedScalarContainer* self, PyObject* arg) {
+  RepeatedScalarContainer* clone = reinterpret_cast<RepeatedScalarContainer*>(
+      PyType_GenericAlloc(&RepeatedScalarContainer_Type, 0));
+  if (clone == NULL) {
+    return NULL;
+  }
+
+  if (InitializeAndCopyToParentContainer(self, clone) < 0) {
+    Py_DECREF(clone);
+    return NULL;
+  }
+  return reinterpret_cast<PyObject*>(clone);
+}
+
+static void Dealloc(RepeatedScalarContainer* self) {
+  self->owner.reset();
+  Py_TYPE(self)->tp_free(reinterpret_cast<PyObject*>(self));
+}
+
+void SetOwner(RepeatedScalarContainer* self,
+              const shared_ptr<Message>& new_owner) {
+  self->owner = new_owner;
+}
+
+static PySequenceMethods SqMethods = {
+  (lenfunc)Len,           /* sq_length */
+  0, /* sq_concat */
+  0, /* sq_repeat */
+  (ssizeargfunc)Item, /* sq_item */
+  0, /* sq_slice */
+  (ssizeobjargproc)AssignItem /* sq_ass_item */
+};
+
+static PyMappingMethods MpMethods = {
+  (lenfunc)Len,               /* mp_length */
+  (binaryfunc)Subscript,      /* mp_subscript */
+  (objobjargproc)AssSubscript, /* mp_ass_subscript */
+};
+
+static PyMethodDef Methods[] = {
+  { "__deepcopy__", (PyCFunction)DeepCopy, METH_VARARGS,
+    "Makes a deep copy of the class." },
+  { "__reduce__", (PyCFunction)Reduce, METH_NOARGS,
+    "Outputs picklable representation of the repeated field." },
+  { "append", (PyCFunction)Append, METH_O,
+    "Appends an object to the repeated container." },
+  { "extend", (PyCFunction)Extend, METH_O,
+    "Appends objects to the repeated container." },
+  { "insert", (PyCFunction)Insert, METH_VARARGS,
+    "Appends objects to the repeated container." },
+  { "pop", (PyCFunction)Pop, METH_VARARGS,
+    "Removes an object from the repeated container and returns it." },
+  { "remove", (PyCFunction)Remove, METH_O,
+    "Removes an object from the repeated container." },
+  { "sort", (PyCFunction)Sort, METH_VARARGS | METH_KEYWORDS,
+    "Sorts the repeated container."},
+  { NULL, NULL }
+};
+
+}  // namespace repeated_scalar_container
+
+PyTypeObject RepeatedScalarContainer_Type = {
+  PyVarObject_HEAD_INIT(&PyType_Type, 0)
+  FULL_MODULE_NAME ".RepeatedScalarContainer",  // tp_name
+  sizeof(RepeatedScalarContainer),     // tp_basicsize
+  0,                                   //  tp_itemsize
+  (destructor)repeated_scalar_container::Dealloc,  //  tp_dealloc
+  0,                                   //  tp_print
+  0,                                   //  tp_getattr
+  0,                                   //  tp_setattr
+  0,                                   //  tp_compare
+  0,                                   //  tp_repr
+  0,                                   //  tp_as_number
+  &repeated_scalar_container::SqMethods,   //  tp_as_sequence
+  &repeated_scalar_container::MpMethods,   //  tp_as_mapping
+  PyObject_HashNotImplemented,         //  tp_hash
+  0,                                   //  tp_call
+  0,                                   //  tp_str
+  0,                                   //  tp_getattro
+  0,                                   //  tp_setattro
+  0,                                   //  tp_as_buffer
+  Py_TPFLAGS_DEFAULT,                  //  tp_flags
+  "A Repeated scalar container",       //  tp_doc
+  0,                                   //  tp_traverse
+  0,                                   //  tp_clear
+  (richcmpfunc)repeated_scalar_container::RichCompare,  //  tp_richcompare
+  0,                                   //  tp_weaklistoffset
+  0,                                   //  tp_iter
+  0,                                   //  tp_iternext
+  repeated_scalar_container::Methods,      //  tp_methods
+  0,                                   //  tp_members
+  0,                                   //  tp_getset
+  0,                                   //  tp_base
+  0,                                   //  tp_dict
+  0,                                   //  tp_descr_get
+  0,                                   //  tp_descr_set
+  0,                                   //  tp_dictoffset
+  0,                                   //  tp_init
+};
+
+}  // namespace python
+}  // namespace protobuf
+}  // namespace google
diff --git a/third_party/protobuf/python/google/protobuf/pyext/repeated_scalar_container.h b/third_party/protobuf/python/google/protobuf/pyext/repeated_scalar_container.h
new file mode 100644
index 0000000000..555e621c9b
--- /dev/null
+++ b/third_party/protobuf/python/google/protobuf/pyext/repeated_scalar_container.h
@@ -0,0 +1,122 @@
+// 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: anuraag@google.com (Anuraag Agrawal)
+// Author: tibell@google.com (Johan Tibell)
+
+#ifndef GOOGLE_PROTOBUF_PYTHON_CPP_REPEATED_SCALAR_CONTAINER_H__
+#define GOOGLE_PROTOBUF_PYTHON_CPP_REPEATED_SCALAR_CONTAINER_H__
+
+#include <Python.h>
+
+#include <memory>
+#ifndef _SHARED_PTR_H
+#include <google/protobuf/stubs/shared_ptr.h>
+#endif
+
+#include <google/protobuf/descriptor.h>
+
+namespace google {
+namespace protobuf {
+
+class Message;
+
+#ifdef _SHARED_PTR_H
+using std::shared_ptr;
+#else
+using internal::shared_ptr;
+#endif
+
+namespace python {
+
+struct CMessage;
+
+typedef struct RepeatedScalarContainer {
+  PyObject_HEAD;
+
+  // This is the top-level C++ Message object that owns the whole
+  // proto tree.  Every Python RepeatedScalarContainer holds a
+  // reference to it in order to keep it alive as long as there's a
+  // Python object that references any part of the tree.
+  shared_ptr<Message> owner;
+
+  // Pointer to the C++ Message that contains this container.  The
+  // RepeatedScalarContainer does not own this pointer.
+  Message* message;
+
+  // Weak reference to a parent CMessage object (i.e. may be NULL.)
+  //
+  // Used to make sure all ancestors are also mutable when first
+  // modifying the container.
+  CMessage* parent;
+
+  // Pointer to the parent's descriptor that describes this
+  // field.  Used together with the parent's message when making a
+  // default message instance mutable.
+  // The pointer is owned by the global DescriptorPool.
+  const FieldDescriptor* parent_field_descriptor;
+} RepeatedScalarContainer;
+
+extern PyTypeObject RepeatedScalarContainer_Type;
+
+namespace repeated_scalar_container {
+
+// Builds a RepeatedScalarContainer object, from a parent message and a
+// field descriptor.
+extern PyObject *NewContainer(
+    CMessage* parent, const FieldDescriptor* parent_field_descriptor);
+
+// Appends the scalar 'item' to the end of the container 'self'.
+//
+// Returns None if successful; returns NULL and sets an exception if
+// unsuccessful.
+PyObject* Append(RepeatedScalarContainer* self, PyObject* item);
+
+// Releases the messages in the container to a new message.
+//
+// Returns 0 on success, -1 on failure.
+int Release(RepeatedScalarContainer* self);
+
+// Appends all the elements in the input iterator to the container.
+//
+// Returns None if successful; returns NULL and sets an exception if
+// unsuccessful.
+PyObject* Extend(RepeatedScalarContainer* self, PyObject* value);
+
+// Set the owner field of self and any children of self.
+void SetOwner(RepeatedScalarContainer* self,
+              const shared_ptr<Message>& new_owner);
+
+}  // namespace repeated_scalar_container
+}  // namespace python
+}  // namespace protobuf
+
+}  // namespace google
+#endif  // GOOGLE_PROTOBUF_PYTHON_CPP_REPEATED_SCALAR_CONTAINER_H__
diff --git a/third_party/protobuf/python/google/protobuf/pyext/scoped_pyobject_ptr.h b/third_party/protobuf/python/google/protobuf/pyext/scoped_pyobject_ptr.h
new file mode 100644
index 0000000000..a128cd4c61
--- /dev/null
+++ b/third_party/protobuf/python/google/protobuf/pyext/scoped_pyobject_ptr.h
@@ -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.
+
+// Author: tibell@google.com (Johan Tibell)
+
+#ifndef GOOGLE_PROTOBUF_PYTHON_CPP_SCOPED_PYOBJECT_PTR_H__
+#define GOOGLE_PROTOBUF_PYTHON_CPP_SCOPED_PYOBJECT_PTR_H__
+
+#include <google/protobuf/stubs/common.h>
+
+#include <Python.h>
+
+namespace google {
+class ScopedPyObjectPtr {
+ public:
+  // Constructor.  Defaults to initializing with NULL.
+  // There is no way to create an uninitialized ScopedPyObjectPtr.
+  explicit ScopedPyObjectPtr(PyObject* p = NULL) : ptr_(p) { }
+
+  // Destructor.  If there is a PyObject object, delete it.
+  ~ScopedPyObjectPtr() {
+    Py_XDECREF(ptr_);
+  }
+
+  // Reset.  Deletes the current owned object, if any.
+  // Then takes ownership of a new object, if given.
+  // This function must be called with a reference that you own.
+  //   this->reset(this->get()) is wrong!
+  //   this->reset(this->release()) is OK.
+  PyObject* reset(PyObject* p = NULL) {
+    Py_XDECREF(ptr_);
+    ptr_ = p;
+    return ptr_;
+  }
+
+  // Releases ownership of the object.
+  // The caller now owns the returned reference.
+  PyObject* release() {
+    PyObject* p = ptr_;
+    ptr_ = NULL;
+    return p;
+  }
+
+  PyObject* operator->() const  {
+    assert(ptr_ != NULL);
+    return ptr_;
+  }
+
+  PyObject* get() const { return ptr_; }
+
+  Py_ssize_t refcnt() const { return Py_REFCNT(ptr_); }
+
+  void inc() const { Py_INCREF(ptr_); }
+
+  // Comparison operators.
+  // These return whether a ScopedPyObjectPtr and a raw pointer
+  // refer to the same object, not just to two different but equal
+  // objects.
+  bool operator==(const PyObject* p) const { return ptr_ == p; }
+  bool operator!=(const PyObject* p) const { return ptr_ != p; }
+
+ private:
+  PyObject* ptr_;
+
+  GOOGLE_DISALLOW_EVIL_CONSTRUCTORS(ScopedPyObjectPtr);
+};
+
+}  // namespace google
+#endif  // GOOGLE_PROTOBUF_PYTHON_CPP_SCOPED_PYOBJECT_PTR_H__
diff --git a/third_party/protobuf/python/google/protobuf/reflection.py b/third_party/protobuf/python/google/protobuf/reflection.py
new file mode 100755
index 0000000000..0c757264f4
--- /dev/null
+++ b/third_party/protobuf/python/google/protobuf/reflection.py
@@ -0,0 +1,120 @@
+# 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.
+
+"""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 google.protobuf.internal import api_implementation
+from google.protobuf import message
+
+
+if api_implementation.Type() == 'cpp':
+  from google.protobuf.pyext import cpp_message as message_impl
+else:
+  from google.protobuf.internal import python_message as message_impl
+
+# The type of all Message classes.
+# Part of the public interface.
+#
+# Used by generated files, but clients can also use it at runtime:
+#   mydescriptor = pool.FindDescriptor(.....)
+#   class MyProtoClass(Message):
+#     __metaclass__ = GeneratedProtocolMessageType
+#     DESCRIPTOR = mydescriptor
+GeneratedProtocolMessageType = message_impl.GeneratedProtocolMessageType
+
+
+def ParseMessage(descriptor, byte_str):
+  """Generate a new Message instance from this Descriptor and a byte string.
+
+  Args:
+    descriptor: Protobuf Descriptor object
+    byte_str: Serialized protocol buffer byte string
+
+  Returns:
+    Newly created protobuf Message object.
+  """
+  result_class = MakeClass(descriptor)
+  new_msg = result_class()
+  new_msg.ParseFromString(byte_str)
+  return new_msg
+
+
+def MakeClass(descriptor):
+  """Construct a class object for a protobuf described by descriptor.
+
+  Composite descriptors are handled by defining the new class as a member of the
+  parent class, recursing as deep as necessary.
+  This is the dynamic equivalent to:
+
+  class Parent(message.Message):
+    __metaclass__ = GeneratedProtocolMessageType
+    DESCRIPTOR = descriptor
+    class Child(message.Message):
+      __metaclass__ = GeneratedProtocolMessageType
+      DESCRIPTOR = descriptor.nested_types[0]
+
+  Sample usage:
+    file_descriptor = descriptor_pb2.FileDescriptorProto()
+    file_descriptor.ParseFromString(proto2_string)
+    msg_descriptor = descriptor.MakeDescriptor(file_descriptor.message_type[0])
+    msg_class = reflection.MakeClass(msg_descriptor)
+    msg = msg_class()
+
+  Args:
+    descriptor: A descriptor.Descriptor object describing the protobuf.
+  Returns:
+    The Message class object described by the descriptor.
+  """
+  attributes = {}
+  for name, nested_type in descriptor.nested_types_by_name.items():
+    attributes[name] = MakeClass(nested_type)
+
+  attributes[GeneratedProtocolMessageType._DESCRIPTOR_KEY] = descriptor
+
+  return GeneratedProtocolMessageType(str(descriptor.name), (message.Message,),
+                                      attributes)
diff --git a/third_party/protobuf/python/google/protobuf/service.py b/third_party/protobuf/python/google/protobuf/service.py
new file mode 100755
index 0000000000..9e00de7042
--- /dev/null
+++ b/third_party/protobuf/python/google/protobuf/service.py
@@ -0,0 +1,226 @@
+# 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.
+
+"""DEPRECATED:  Declares the RPC service interfaces.
+
+This module declares the abstract interfaces underlying proto2 RPC
+services.  These are intended to be independent of any particular RPC
+implementation, so that proto2 services can be used on top of a variety
+of implementations.  Starting with version 2.3.0, RPC implementations should
+not try to build on these, but should instead provide code generator plugins
+which generate code specific to the particular RPC implementation.  This way
+the generated code can be more appropriate for the implementation in use
+and can avoid unnecessary layers of indirection.
+"""
+
+__author__ = 'petar@google.com (Petar Petrov)'
+
+
+class RpcException(Exception):
+  """Exception raised on failed blocking RPC method call."""
+  pass
+
+
+class Service(object):
+
+  """Abstract base interface for protocol-buffer-based RPC services.
+
+  Services themselves are abstract classes (implemented either by servers or as
+  stubs), but they subclass this base interface. The methods of this
+  interface can be used to call the methods of the service without knowing
+  its exact type at compile time (analogous to the Message interface).
+  """
+
+  def GetDescriptor():
+    """Retrieves this service's descriptor."""
+    raise NotImplementedError
+
+  def CallMethod(self, method_descriptor, rpc_controller,
+                 request, done):
+    """Calls a method of the service specified by method_descriptor.
+
+    If "done" is None then the call is blocking and the response
+    message will be returned directly.  Otherwise the call is asynchronous
+    and "done" will later be called with the response value.
+
+    In the blocking case, RpcException will be raised on error.
+
+    Preconditions:
+    * method_descriptor.service == GetDescriptor
+    * request is of the exact same classes as returned by
+      GetRequestClass(method).
+    * After the call has started, the request must not be modified.
+    * "rpc_controller" is of the correct type for the RPC implementation being
+      used by this Service.  For stubs, the "correct type" depends on the
+      RpcChannel which the stub is using.
+
+    Postconditions:
+    * "done" will be called when the method is complete.  This may be
+      before CallMethod() returns or it may be at some point in the future.
+    * If the RPC failed, the response value passed to "done" will be None.
+      Further details about the failure can be found by querying the
+      RpcController.
+    """
+    raise NotImplementedError
+
+  def GetRequestClass(self, method_descriptor):
+    """Returns the class of the request message for the specified method.
+
+    CallMethod() requires that the request is of a particular subclass of
+    Message. GetRequestClass() gets the default instance of this required
+    type.
+
+    Example:
+      method = service.GetDescriptor().FindMethodByName("Foo")
+      request = stub.GetRequestClass(method)()
+      request.ParseFromString(input)
+      service.CallMethod(method, request, callback)
+    """
+    raise NotImplementedError
+
+  def GetResponseClass(self, method_descriptor):
+    """Returns the class of the response message for the specified method.
+
+    This method isn't really needed, as the RpcChannel's CallMethod constructs
+    the response protocol message. It's provided anyway in case it is useful
+    for the caller to know the response type in advance.
+    """
+    raise NotImplementedError
+
+
+class RpcController(object):
+
+  """An RpcController mediates a single method call.
+
+  The primary purpose of the controller is to provide a way to manipulate
+  settings specific to the RPC implementation and to find out about RPC-level
+  errors. The methods provided by the RpcController interface are intended
+  to be a "least common denominator" set of features which we expect all
+  implementations to support.  Specific implementations may provide more
+  advanced features (e.g. deadline propagation).
+  """
+
+  # Client-side methods below
+
+  def Reset(self):
+    """Resets the RpcController to its initial state.
+
+    After the RpcController has been reset, it may be reused in
+    a new call. Must not be called while an RPC is in progress.
+    """
+    raise NotImplementedError
+
+  def Failed(self):
+    """Returns true if the call failed.
+
+    After a call has finished, returns true if the call failed.  The possible
+    reasons for failure depend on the RPC implementation.  Failed() must not
+    be called before a call has finished.  If Failed() returns true, the
+    contents of the response message are undefined.
+    """
+    raise NotImplementedError
+
+  def ErrorText(self):
+    """If Failed is true, returns a human-readable description of the error."""
+    raise NotImplementedError
+
+  def StartCancel(self):
+    """Initiate cancellation.
+
+    Advises the RPC system that the caller desires that the RPC call be
+    canceled.  The RPC system may cancel it immediately, may wait awhile and
+    then cancel it, or may not even cancel the call at all.  If the call is
+    canceled, the "done" callback will still be called and the RpcController
+    will indicate that the call failed at that time.
+    """
+    raise NotImplementedError
+
+  # Server-side methods below
+
+  def SetFailed(self, reason):
+    """Sets a failure reason.
+
+    Causes Failed() to return true on the client side.  "reason" will be
+    incorporated into the message returned by ErrorText().  If you find
+    you need to return machine-readable information about failures, you
+    should incorporate it into your response protocol buffer and should
+    NOT call SetFailed().
+    """
+    raise NotImplementedError
+
+  def IsCanceled(self):
+    """Checks if the client cancelled the RPC.
+
+    If true, indicates that the client canceled the RPC, so the server may
+    as well give up on replying to it.  The server should still call the
+    final "done" callback.
+    """
+    raise NotImplementedError
+
+  def NotifyOnCancel(self, callback):
+    """Sets a callback to invoke on cancel.
+
+    Asks that the given callback be called when the RPC is canceled.  The
+    callback will always be called exactly once.  If the RPC completes without
+    being canceled, the callback will be called after completion.  If the RPC
+    has already been canceled when NotifyOnCancel() is called, the callback
+    will be called immediately.
+
+    NotifyOnCancel() must be called no more than once per request.
+    """
+    raise NotImplementedError
+
+
+class RpcChannel(object):
+
+  """Abstract interface for an RPC channel.
+
+  An RpcChannel represents a communication line to a service which can be used
+  to call that service's methods.  The service may be running on another
+  machine. Normally, you should not use an RpcChannel directly, but instead
+  construct a stub {@link Service} wrapping it.  Example:
+
+  Example:
+    RpcChannel channel = rpcImpl.Channel("remotehost.example.com:1234")
+    RpcController controller = rpcImpl.Controller()
+    MyService service = MyService_Stub(channel)
+    service.MyMethod(controller, request, callback)
+  """
+
+  def CallMethod(self, method_descriptor, rpc_controller,
+                 request, response_class, done):
+    """Calls the method identified by the descriptor.
+
+    Call the given method of the remote service.  The signature of this
+    procedure looks the same as Service.CallMethod(), but the requirements
+    are less strict in one important way:  the request object doesn't have to
+    be of any specific class as long as its descriptor is method.input_type.
+    """
+    raise NotImplementedError
diff --git a/third_party/protobuf/python/google/protobuf/service_reflection.py b/third_party/protobuf/python/google/protobuf/service_reflection.py
new file mode 100755
index 0000000000..1c3636afe0
--- /dev/null
+++ b/third_party/protobuf/python/google/protobuf/service_reflection.py
@@ -0,0 +1,284 @@
+# 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 metaclasses used to create protocol service and service stub
+classes from ServiceDescriptor objects at runtime.
+
+The GeneratedServiceType and GeneratedServiceStubType metaclasses are used to
+inject all useful functionality into the classes output by the protocol
+compiler at compile-time.
+"""
+
+__author__ = 'petar@google.com (Petar Petrov)'
+
+
+class GeneratedServiceType(type):
+
+  """Metaclass for service classes created at runtime from ServiceDescriptors.
+
+  Implementations for all methods described in the Service class are added here
+  by this class. We also create properties to allow getting/setting all fields
+  in the protocol message.
+
+  The protocol compiler currently uses this metaclass to create protocol service
+  classes at runtime. Clients can also manually create their own classes at
+  runtime, as in this example:
+
+  mydescriptor = ServiceDescriptor(.....)
+  class MyProtoService(service.Service):
+    __metaclass__ = GeneratedServiceType
+    DESCRIPTOR = mydescriptor
+  myservice_instance = MyProtoService()
+  ...
+  """
+
+  _DESCRIPTOR_KEY = 'DESCRIPTOR'
+
+  def __init__(cls, name, bases, dictionary):
+    """Creates a message service class.
+
+    Args:
+      name: Name of the class (ignored, but required by the metaclass
+        protocol).
+      bases: Base classes of the class being constructed.
+      dictionary: The class dictionary of the class being constructed.
+        dictionary[_DESCRIPTOR_KEY] must contain a ServiceDescriptor object
+        describing this protocol service type.
+    """
+    # Don't do anything if this class doesn't have a descriptor. This happens
+    # when a service class is subclassed.
+    if GeneratedServiceType._DESCRIPTOR_KEY not in dictionary:
+      return
+    descriptor = dictionary[GeneratedServiceType._DESCRIPTOR_KEY]
+    service_builder = _ServiceBuilder(descriptor)
+    service_builder.BuildService(cls)
+
+
+class GeneratedServiceStubType(GeneratedServiceType):
+
+  """Metaclass for service stubs created at runtime from ServiceDescriptors.
+
+  This class has similar responsibilities as GeneratedServiceType, except that
+  it creates the service stub classes.
+  """
+
+  _DESCRIPTOR_KEY = 'DESCRIPTOR'
+
+  def __init__(cls, name, bases, dictionary):
+    """Creates a message service stub class.
+
+    Args:
+      name: Name of the class (ignored, here).
+      bases: Base classes of the class being constructed.
+      dictionary: The class dictionary of the class being constructed.
+        dictionary[_DESCRIPTOR_KEY] must contain a ServiceDescriptor object
+        describing this protocol service type.
+    """
+    super(GeneratedServiceStubType, cls).__init__(name, bases, dictionary)
+    # Don't do anything if this class doesn't have a descriptor. This happens
+    # when a service stub is subclassed.
+    if GeneratedServiceStubType._DESCRIPTOR_KEY not in dictionary:
+      return
+    descriptor = dictionary[GeneratedServiceStubType._DESCRIPTOR_KEY]
+    service_stub_builder = _ServiceStubBuilder(descriptor)
+    service_stub_builder.BuildServiceStub(cls)
+
+
+class _ServiceBuilder(object):
+
+  """This class constructs a protocol service class using a service descriptor.
+
+  Given a service descriptor, this class constructs a class that represents
+  the specified service descriptor. One service builder instance constructs
+  exactly one service class. That means all instances of that class share the
+  same builder.
+  """
+
+  def __init__(self, service_descriptor):
+    """Initializes an instance of the service class builder.
+
+    Args:
+      service_descriptor: ServiceDescriptor to use when constructing the
+        service class.
+    """
+    self.descriptor = service_descriptor
+
+  def BuildService(self, cls):
+    """Constructs the service class.
+
+    Args:
+      cls: The class that will be constructed.
+    """
+
+    # CallMethod needs to operate with an instance of the Service class. This
+    # internal wrapper function exists only to be able to pass the service
+    # instance to the method that does the real CallMethod work.
+    def _WrapCallMethod(srvc, method_descriptor,
+                        rpc_controller, request, callback):
+      return self._CallMethod(srvc, method_descriptor,
+                       rpc_controller, request, callback)
+    self.cls = cls
+    cls.CallMethod = _WrapCallMethod
+    cls.GetDescriptor = staticmethod(lambda: self.descriptor)
+    cls.GetDescriptor.__doc__ = "Returns the service descriptor."
+    cls.GetRequestClass = self._GetRequestClass
+    cls.GetResponseClass = self._GetResponseClass
+    for method in self.descriptor.methods:
+      setattr(cls, method.name, self._GenerateNonImplementedMethod(method))
+
+  def _CallMethod(self, srvc, method_descriptor,
+                  rpc_controller, request, callback):
+    """Calls the method described by a given method descriptor.
+
+    Args:
+      srvc: Instance of the service for which this method is called.
+      method_descriptor: Descriptor that represent the method to call.
+      rpc_controller: RPC controller to use for this method's execution.
+      request: Request protocol message.
+      callback: A callback to invoke after the method has completed.
+    """
+    if method_descriptor.containing_service != self.descriptor:
+      raise RuntimeError(
+          'CallMethod() given method descriptor for wrong service type.')
+    method = getattr(srvc, method_descriptor.name)
+    return method(rpc_controller, request, callback)
+
+  def _GetRequestClass(self, method_descriptor):
+    """Returns the class of the request protocol message.
+
+    Args:
+      method_descriptor: Descriptor of the method for which to return the
+        request protocol message class.
+
+    Returns:
+      A class that represents the input protocol message of the specified
+      method.
+    """
+    if method_descriptor.containing_service != self.descriptor:
+      raise RuntimeError(
+          'GetRequestClass() given method descriptor for wrong service type.')
+    return method_descriptor.input_type._concrete_class
+
+  def _GetResponseClass(self, method_descriptor):
+    """Returns the class of the response protocol message.
+
+    Args:
+      method_descriptor: Descriptor of the method for which to return the
+        response protocol message class.
+
+    Returns:
+      A class that represents the output protocol message of the specified
+      method.
+    """
+    if method_descriptor.containing_service != self.descriptor:
+      raise RuntimeError(
+          'GetResponseClass() given method descriptor for wrong service type.')
+    return method_descriptor.output_type._concrete_class
+
+  def _GenerateNonImplementedMethod(self, method):
+    """Generates and returns a method that can be set for a service methods.
+
+    Args:
+      method: Descriptor of the service method for which a method is to be
+        generated.
+
+    Returns:
+      A method that can be added to the service class.
+    """
+    return lambda inst, rpc_controller, request, callback: (
+        self._NonImplementedMethod(method.name, rpc_controller, callback))
+
+  def _NonImplementedMethod(self, method_name, rpc_controller, callback):
+    """The body of all methods in the generated service class.
+
+    Args:
+      method_name: Name of the method being executed.
+      rpc_controller: RPC controller used to execute this method.
+      callback: A callback which will be invoked when the method finishes.
+    """
+    rpc_controller.SetFailed('Method %s not implemented.' % method_name)
+    callback(None)
+
+
+class _ServiceStubBuilder(object):
+
+  """Constructs a protocol service stub class using a service descriptor.
+
+  Given a service descriptor, this class constructs a suitable stub class.
+  A stub is just a type-safe wrapper around an RpcChannel which emulates a
+  local implementation of the service.
+
+  One service stub builder instance constructs exactly one class. It means all
+  instances of that class share the same service stub builder.
+  """
+
+  def __init__(self, service_descriptor):
+    """Initializes an instance of the service stub class builder.
+
+    Args:
+      service_descriptor: ServiceDescriptor to use when constructing the
+        stub class.
+    """
+    self.descriptor = service_descriptor
+
+  def BuildServiceStub(self, cls):
+    """Constructs the stub class.
+
+    Args:
+      cls: The class that will be constructed.
+    """
+
+    def _ServiceStubInit(stub, rpc_channel):
+      stub.rpc_channel = rpc_channel
+    self.cls = cls
+    cls.__init__ = _ServiceStubInit
+    for method in self.descriptor.methods:
+      setattr(cls, method.name, self._GenerateStubMethod(method))
+
+  def _GenerateStubMethod(self, method):
+    return (lambda inst, rpc_controller, request, callback=None:
+        self._StubMethod(inst, method, rpc_controller, request, callback))
+
+  def _StubMethod(self, stub, method_descriptor,
+                  rpc_controller, request, callback):
+    """The body of all service methods in the generated stub class.
+
+    Args:
+      stub: Stub instance.
+      method_descriptor: Descriptor of the invoked method.
+      rpc_controller: Rpc controller to execute the method.
+      request: Request protocol message.
+      callback: A callback to execute when the method finishes.
+    Returns:
+      Response message (in case of blocking call).
+    """
+    return stub.rpc_channel.CallMethod(
+        method_descriptor, rpc_controller, request,
+        method_descriptor.output_type._concrete_class, callback)
diff --git a/third_party/protobuf/python/google/protobuf/symbol_database.py b/third_party/protobuf/python/google/protobuf/symbol_database.py
new file mode 100644
index 0000000000..87760f2630
--- /dev/null
+++ b/third_party/protobuf/python/google/protobuf/symbol_database.py
@@ -0,0 +1,185 @@
+# 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 database of Python protocol buffer generated symbols.
+
+SymbolDatabase makes it easy to create new instances of a registered type, given
+only the type's protocol buffer symbol name. Once all symbols are registered,
+they can be accessed using either the MessageFactory interface which
+SymbolDatabase exposes, or the DescriptorPool interface of the underlying
+pool.
+
+Example usage:
+
+  db = symbol_database.SymbolDatabase()
+
+  # Register symbols of interest, from one or multiple files.
+  db.RegisterFileDescriptor(my_proto_pb2.DESCRIPTOR)
+  db.RegisterMessage(my_proto_pb2.MyMessage)
+  db.RegisterEnumDescriptor(my_proto_pb2.MyEnum.DESCRIPTOR)
+
+  # The database can be used as a MessageFactory, to generate types based on
+  # their name:
+  types = db.GetMessages(['my_proto.proto'])
+  my_message_instance = types['MyMessage']()
+
+  # The database's underlying descriptor pool can be queried, so it's not
+  # necessary to know a type's filename to be able to generate it:
+  filename = db.pool.FindFileContainingSymbol('MyMessage')
+  my_message_instance = db.GetMessages([filename])['MyMessage']()
+
+  # This functionality is also provided directly via a convenience method:
+  my_message_instance = db.GetSymbol('MyMessage')()
+"""
+
+
+from google.protobuf import descriptor_pool
+
+
+class SymbolDatabase(object):
+  """A database of Python generated symbols.
+
+  SymbolDatabase also models message_factory.MessageFactory.
+
+  The symbol database can be used to keep a global registry of all protocol
+  buffer types used within a program.
+  """
+
+  def __init__(self, pool=None):
+    """Constructor."""
+
+    self._symbols = {}
+    self._symbols_by_file = {}
+    self.pool = pool or descriptor_pool.Default()
+
+  def RegisterMessage(self, message):
+    """Registers the given message type in the local database.
+
+    Args:
+      message: a message.Message, to be registered.
+
+    Returns:
+      The provided message.
+    """
+
+    desc = message.DESCRIPTOR
+    self._symbols[desc.full_name] = message
+    if desc.file.name not in self._symbols_by_file:
+      self._symbols_by_file[desc.file.name] = {}
+    self._symbols_by_file[desc.file.name][desc.full_name] = message
+    self.pool.AddDescriptor(desc)
+    return message
+
+  def RegisterEnumDescriptor(self, enum_descriptor):
+    """Registers the given enum descriptor in the local database.
+
+    Args:
+      enum_descriptor: a descriptor.EnumDescriptor.
+
+    Returns:
+      The provided descriptor.
+    """
+    self.pool.AddEnumDescriptor(enum_descriptor)
+    return enum_descriptor
+
+  def RegisterFileDescriptor(self, file_descriptor):
+    """Registers the given file descriptor in the local database.
+
+    Args:
+      file_descriptor: a descriptor.FileDescriptor.
+
+    Returns:
+      The provided descriptor.
+    """
+    self.pool.AddFileDescriptor(file_descriptor)
+
+  def GetSymbol(self, symbol):
+    """Tries to find a symbol in the local database.
+
+    Currently, this method only returns message.Message instances, however, if
+    may be extended in future to support other symbol types.
+
+    Args:
+      symbol: A str, a protocol buffer symbol.
+
+    Returns:
+      A Python class corresponding to the symbol.
+
+    Raises:
+      KeyError: if the symbol could not be found.
+    """
+
+    return self._symbols[symbol]
+
+  def GetPrototype(self, descriptor):
+    """Builds a proto2 message class based on the passed in descriptor.
+
+    Passing a descriptor with a fully qualified name matching a previous
+    invocation will cause the same class to be returned.
+
+    Args:
+      descriptor: The descriptor to build from.
+
+    Returns:
+      A class describing the passed in descriptor.
+    """
+
+    return self.GetSymbol(descriptor.full_name)
+
+  def GetMessages(self, files):
+    """Gets all the messages from a specified file.
+
+    This will find and resolve dependencies, failing if they are not registered
+    in the symbol database.
+
+
+    Args:
+      files: The file names to extract messages from.
+
+    Returns:
+      A dictionary mapping proto names to the message classes. This will include
+      any dependent messages as well as any messages defined in the same file as
+      a specified message.
+
+    Raises:
+      KeyError: if a file could not be found.
+    """
+
+    result = {}
+    for f in files:
+      result.update(self._symbols_by_file[f])
+    return result
+
+_DEFAULT = SymbolDatabase(pool=descriptor_pool.Default())
+
+
+def Default():
+  """Returns the default SymbolDatabase."""
+  return _DEFAULT
diff --git a/third_party/protobuf/python/google/protobuf/text_encoding.py b/third_party/protobuf/python/google/protobuf/text_encoding.py
new file mode 100644
index 0000000000..9899563825
--- /dev/null
+++ b/third_party/protobuf/python/google/protobuf/text_encoding.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.
+
+"""Encoding related utilities."""
+import re
+
+import six
+
+# Lookup table for utf8
+_cescape_utf8_to_str = [chr(i) for i in range(0, 256)]
+_cescape_utf8_to_str[9] = r'\t'  # optional escape
+_cescape_utf8_to_str[10] = r'\n'  # optional escape
+_cescape_utf8_to_str[13] = r'\r'  # optional escape
+_cescape_utf8_to_str[39] = r"\'"  # optional escape
+
+_cescape_utf8_to_str[34] = r'\"'  # necessary escape
+_cescape_utf8_to_str[92] = r'\\'  # necessary escape
+
+# Lookup table for non-utf8, with necessary escapes at (o >= 127 or o < 32)
+_cescape_byte_to_str = ([r'\%03o' % i for i in range(0, 32)] +
+                        [chr(i) for i in range(32, 127)] +
+                        [r'\%03o' % i for i in range(127, 256)])
+_cescape_byte_to_str[9] = r'\t'  # optional escape
+_cescape_byte_to_str[10] = r'\n'  # optional escape
+_cescape_byte_to_str[13] = r'\r'  # optional escape
+_cescape_byte_to_str[39] = r"\'"  # optional escape
+
+_cescape_byte_to_str[34] = r'\"'  # necessary escape
+_cescape_byte_to_str[92] = r'\\'  # necessary escape
+
+
+def CEscape(text, as_utf8):
+  """Escape a bytes string for use in an ascii protocol buffer.
+
+  text.encode('string_escape') does not seem to satisfy our needs as it
+  encodes unprintable characters using two-digit hex escapes whereas our
+  C++ unescaping function allows hex escapes to be any length.  So,
+  "\0011".encode('string_escape') ends up being "\\x011", which will be
+  decoded in C++ as a single-character string with char code 0x11.
+
+  Args:
+    text: A byte string to be escaped
+    as_utf8: Specifies if result should be returned in UTF-8 encoding
+  Returns:
+    Escaped string
+  """
+  # PY3 hack: make Ord work for str and bytes:
+  # //platforms/networking/data uses unicode here, hence basestring.
+  Ord = ord if isinstance(text, six.string_types) else lambda x: x
+  if as_utf8:
+    return ''.join(_cescape_utf8_to_str[Ord(c)] for c in text)
+  return ''.join(_cescape_byte_to_str[Ord(c)] for c in text)
+
+
+_CUNESCAPE_HEX = re.compile(r'(\\+)x([0-9a-fA-F])(?![0-9a-fA-F])')
+_cescape_highbit_to_str = ([chr(i) for i in range(0, 127)] +
+                           [r'\%03o' % i for i in range(127, 256)])
+
+
+def CUnescape(text):
+  """Unescape a text string with C-style escape sequences to UTF-8 bytes."""
+
+  def ReplaceHex(m):
+    # Only replace the match if the number of leading back slashes is odd. i.e.
+    # the slash itself is not escaped.
+    if len(m.group(1)) & 1:
+      return m.group(1) + 'x0' + m.group(2)
+    return m.group(0)
+
+  # This is required because the 'string_escape' encoding doesn't
+  # allow single-digit hex escapes (like '\xf').
+  result = _CUNESCAPE_HEX.sub(ReplaceHex, text)
+
+  if str is bytes:  # PY2
+    return result.decode('string_escape')
+  result = ''.join(_cescape_highbit_to_str[ord(c)] for c in result)
+  return (result.encode('ascii')  # Make it bytes to allow decode.
+          .decode('unicode_escape')
+          # Make it bytes again to return the proper type.
+          .encode('raw_unicode_escape'))
diff --git a/third_party/protobuf/python/google/protobuf/text_format.py b/third_party/protobuf/python/google/protobuf/text_format.py
new file mode 100755
index 0000000000..8d256076c2
--- /dev/null
+++ b/third_party/protobuf/python/google/protobuf/text_format.py
@@ -0,0 +1,1112 @@
+# 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 routines for printing protocol messages in text format.
+
+Simple usage example:
+
+  # Create a proto object and serialize it to a text proto string.
+  message = my_proto_pb2.MyMessage(foo='bar')
+  text_proto = text_format.MessageToString(message)
+
+  # Parse a text proto string.
+  message = text_format.Parse(text_proto, my_proto_pb2.MyMessage())
+"""
+
+__author__ = 'kenton@google.com (Kenton Varda)'
+
+import io
+import re
+
+import six
+
+if six.PY3:
+  long = int
+
+from google.protobuf.internal import type_checkers
+from google.protobuf import descriptor
+from google.protobuf import text_encoding
+
+__all__ = ['MessageToString', 'PrintMessage', 'PrintField',
+           'PrintFieldValue', 'Merge']
+
+
+_INTEGER_CHECKERS = (type_checkers.Uint32ValueChecker(),
+                     type_checkers.Int32ValueChecker(),
+                     type_checkers.Uint64ValueChecker(),
+                     type_checkers.Int64ValueChecker())
+_FLOAT_INFINITY = re.compile('-?inf(?:inity)?f?', re.IGNORECASE)
+_FLOAT_NAN = re.compile('nanf?', re.IGNORECASE)
+_FLOAT_TYPES = frozenset([descriptor.FieldDescriptor.CPPTYPE_FLOAT,
+                          descriptor.FieldDescriptor.CPPTYPE_DOUBLE])
+_QUOTES = frozenset(("'", '"'))
+
+
+class Error(Exception):
+  """Top-level module error for text_format."""
+
+
+class ParseError(Error):
+  """Thrown in case of text parsing error."""
+
+
+class TextWriter(object):
+  def __init__(self, as_utf8):
+    if six.PY2:
+      self._writer = io.BytesIO()
+    else:
+      self._writer = io.StringIO()
+
+  def write(self, val):
+    if six.PY2:
+      if isinstance(val, six.text_type):
+        val = val.encode('utf-8')
+    return self._writer.write(val)
+
+  def close(self):
+    return self._writer.close()
+
+  def getvalue(self):
+    return self._writer.getvalue()
+
+
+def MessageToString(message, as_utf8=False, as_one_line=False,
+                    pointy_brackets=False, use_index_order=False,
+                    float_format=None):
+  """Convert protobuf message to text format.
+
+  Floating point values can be formatted compactly with 15 digits of
+  precision (which is the most that IEEE 754 "double" can guarantee)
+  using float_format='.15g'. To ensure that converting to text and back to a
+  proto will result in an identical value, float_format='.17g' should be used.
+
+  Args:
+    message: The protocol buffers message.
+    as_utf8: Produce text output in UTF8 format.
+    as_one_line: Don't introduce newlines between fields.
+    pointy_brackets: If True, use angle brackets instead of curly braces for
+      nesting.
+    use_index_order: If True, print fields of a proto message using the order
+      defined in source code instead of the field number. By default, use the
+      field number order.
+    float_format: If set, use this to specify floating point number formatting
+      (per the "Format Specification Mini-Language"); otherwise, str() is used.
+
+  Returns:
+    A string of the text formatted protocol buffer message.
+  """
+  out = TextWriter(as_utf8)
+  PrintMessage(message, out, as_utf8=as_utf8, as_one_line=as_one_line,
+               pointy_brackets=pointy_brackets,
+               use_index_order=use_index_order,
+               float_format=float_format)
+  result = out.getvalue()
+  out.close()
+  if as_one_line:
+    return result.rstrip()
+  return result
+
+
+def _IsMapEntry(field):
+  return (field.type == descriptor.FieldDescriptor.TYPE_MESSAGE and
+          field.message_type.has_options and
+          field.message_type.GetOptions().map_entry)
+
+
+def PrintMessage(message, out, indent=0, as_utf8=False, as_one_line=False,
+                 pointy_brackets=False, use_index_order=False,
+                 float_format=None):
+  fields = message.ListFields()
+  if use_index_order:
+    fields.sort(key=lambda x: x[0].index)
+  for field, value in fields:
+    if _IsMapEntry(field):
+      for key in sorted(value):
+        # This is slow for maps with submessage entires because it copies the
+        # entire tree.  Unfortunately this would take significant refactoring
+        # of this file to work around.
+        #
+        # TODO(haberman): refactor and optimize if this becomes an issue.
+        entry_submsg = field.message_type._concrete_class(
+            key=key, value=value[key])
+        PrintField(field, entry_submsg, out, indent, as_utf8, as_one_line,
+                   pointy_brackets=pointy_brackets,
+                   use_index_order=use_index_order, float_format=float_format)
+    elif field.label == descriptor.FieldDescriptor.LABEL_REPEATED:
+      for element in value:
+        PrintField(field, element, out, indent, as_utf8, as_one_line,
+                   pointy_brackets=pointy_brackets,
+                   use_index_order=use_index_order,
+                   float_format=float_format)
+    else:
+      PrintField(field, value, out, indent, as_utf8, as_one_line,
+                 pointy_brackets=pointy_brackets,
+                 use_index_order=use_index_order,
+                 float_format=float_format)
+
+
+def PrintField(field, value, out, indent=0, as_utf8=False, as_one_line=False,
+               pointy_brackets=False, use_index_order=False, float_format=None):
+  """Print a single field name/value pair.  For repeated fields, the value
+  should be a single element.
+  """
+
+  out.write(' ' * indent)
+  if field.is_extension:
+    out.write('[')
+    if (field.containing_type.GetOptions().message_set_wire_format and
+        field.type == descriptor.FieldDescriptor.TYPE_MESSAGE and
+        field.label == descriptor.FieldDescriptor.LABEL_OPTIONAL):
+      out.write(field.message_type.full_name)
+    else:
+      out.write(field.full_name)
+    out.write(']')
+  elif field.type == descriptor.FieldDescriptor.TYPE_GROUP:
+    # For groups, use the capitalized name.
+    out.write(field.message_type.name)
+  else:
+    out.write(field.name)
+
+  if field.cpp_type != descriptor.FieldDescriptor.CPPTYPE_MESSAGE:
+    # The colon is optional in this case, but our cross-language golden files
+    # don't include it.
+    out.write(': ')
+
+  PrintFieldValue(field, value, out, indent, as_utf8, as_one_line,
+                  pointy_brackets=pointy_brackets,
+                  use_index_order=use_index_order,
+                  float_format=float_format)
+  if as_one_line:
+    out.write(' ')
+  else:
+    out.write('\n')
+
+
+def PrintFieldValue(field, value, out, indent=0, as_utf8=False,
+                    as_one_line=False, pointy_brackets=False,
+                    use_index_order=False,
+                    float_format=None):
+  """Print a single field value (not including name).  For repeated fields,
+  the value should be a single element."""
+
+  if pointy_brackets:
+    openb = '<'
+    closeb = '>'
+  else:
+    openb = '{'
+    closeb = '}'
+
+  if field.cpp_type == descriptor.FieldDescriptor.CPPTYPE_MESSAGE:
+    if as_one_line:
+      out.write(' %s ' % openb)
+      PrintMessage(value, out, indent, as_utf8, as_one_line,
+                   pointy_brackets=pointy_brackets,
+                   use_index_order=use_index_order,
+                   float_format=float_format)
+      out.write(closeb)
+    else:
+      out.write(' %s\n' % openb)
+      PrintMessage(value, out, indent + 2, as_utf8, as_one_line,
+                   pointy_brackets=pointy_brackets,
+                   use_index_order=use_index_order,
+                   float_format=float_format)
+      out.write(' ' * indent + closeb)
+  elif field.cpp_type == descriptor.FieldDescriptor.CPPTYPE_ENUM:
+    enum_value = field.enum_type.values_by_number.get(value, None)
+    if enum_value is not None:
+      out.write(enum_value.name)
+    else:
+      out.write(str(value))
+  elif field.cpp_type == descriptor.FieldDescriptor.CPPTYPE_STRING:
+    out.write('\"')
+    if isinstance(value, six.text_type):
+      out_value = value.encode('utf-8')
+    else:
+      out_value = value
+    if field.type == descriptor.FieldDescriptor.TYPE_BYTES:
+      # We need to escape non-UTF8 chars in TYPE_BYTES field.
+      out_as_utf8 = False
+    else:
+      out_as_utf8 = as_utf8
+    out.write(text_encoding.CEscape(out_value, out_as_utf8))
+    out.write('\"')
+  elif field.cpp_type == descriptor.FieldDescriptor.CPPTYPE_BOOL:
+    if value:
+      out.write('true')
+    else:
+      out.write('false')
+  elif field.cpp_type in _FLOAT_TYPES and float_format is not None:
+    out.write('{1:{0}}'.format(float_format, value))
+  else:
+    out.write(str(value))
+
+
+def Parse(text, message, allow_unknown_extension=False):
+  """Parses an text representation of a protocol message into a message.
+
+  Args:
+    text: Message text representation.
+    message: A protocol buffer message to merge into.
+    allow_unknown_extension: if True, skip over missing extensions and keep
+      parsing
+
+  Returns:
+    The same message passed as argument.
+
+  Raises:
+    ParseError: On text parsing problems.
+  """
+  if not isinstance(text, str):
+    text = text.decode('utf-8')
+  return ParseLines(text.split('\n'), message, allow_unknown_extension)
+
+
+def Merge(text, message, allow_unknown_extension=False):
+  """Parses an text representation of a protocol message into a message.
+
+  Like Parse(), but allows repeated values for a non-repeated field, and uses
+  the last one.
+
+  Args:
+    text: Message text representation.
+    message: A protocol buffer message to merge into.
+    allow_unknown_extension: if True, skip over missing extensions and keep
+      parsing
+
+  Returns:
+    The same message passed as argument.
+
+  Raises:
+    ParseError: On text parsing problems.
+  """
+  return MergeLines(text.split('\n'), message, allow_unknown_extension)
+
+
+def ParseLines(lines, message, allow_unknown_extension=False):
+  """Parses an text representation of a protocol message into a message.
+
+  Args:
+    lines: An iterable of lines of a message's text representation.
+    message: A protocol buffer message to merge into.
+    allow_unknown_extension: if True, skip over missing extensions and keep
+      parsing
+
+  Returns:
+    The same message passed as argument.
+
+  Raises:
+    ParseError: On text parsing problems.
+  """
+  _ParseOrMerge(lines, message, False, allow_unknown_extension)
+  return message
+
+
+def MergeLines(lines, message, allow_unknown_extension=False):
+  """Parses an text representation of a protocol message into a message.
+
+  Args:
+    lines: An iterable of lines of a message's text representation.
+    message: A protocol buffer message to merge into.
+    allow_unknown_extension: if True, skip over missing extensions and keep
+      parsing
+
+  Returns:
+    The same message passed as argument.
+
+  Raises:
+    ParseError: On text parsing problems.
+  """
+  _ParseOrMerge(lines, message, True, allow_unknown_extension)
+  return message
+
+
+def _ParseOrMerge(lines,
+                  message,
+                  allow_multiple_scalars,
+                  allow_unknown_extension=False):
+  """Converts an text representation of a protocol message into a message.
+
+  Args:
+    lines: Lines of a message's text representation.
+    message: A protocol buffer message to merge into.
+    allow_multiple_scalars: Determines if repeated values for a non-repeated
+      field are permitted, e.g., the string "foo: 1 foo: 2" for a
+      required/optional field named "foo".
+    allow_unknown_extension: if True, skip over missing extensions and keep
+      parsing
+
+  Raises:
+    ParseError: On text parsing problems.
+  """
+  tokenizer = _Tokenizer(lines)
+  while not tokenizer.AtEnd():
+    _MergeField(tokenizer, message, allow_multiple_scalars,
+                allow_unknown_extension)
+
+
+def _MergeField(tokenizer,
+                message,
+                allow_multiple_scalars,
+                allow_unknown_extension=False):
+  """Merges a single protocol message field into a message.
+
+  Args:
+    tokenizer: A tokenizer to parse the field name and values.
+    message: A protocol message to record the data.
+    allow_multiple_scalars: Determines if repeated values for a non-repeated
+      field are permitted, e.g., the string "foo: 1 foo: 2" for a
+      required/optional field named "foo".
+    allow_unknown_extension: if True, skip over missing extensions and keep
+      parsing
+
+  Raises:
+    ParseError: In case of text parsing problems.
+  """
+  message_descriptor = message.DESCRIPTOR
+  if (hasattr(message_descriptor, 'syntax') and
+      message_descriptor.syntax == 'proto3'):
+    # Proto3 doesn't represent presence so we can't test if multiple
+    # scalars have occurred.  We have to allow them.
+    allow_multiple_scalars = True
+  if tokenizer.TryConsume('['):
+    name = [tokenizer.ConsumeIdentifier()]
+    while tokenizer.TryConsume('.'):
+      name.append(tokenizer.ConsumeIdentifier())
+    name = '.'.join(name)
+
+    if not message_descriptor.is_extendable:
+      raise tokenizer.ParseErrorPreviousToken(
+          'Message type "%s" does not have extensions.' %
+          message_descriptor.full_name)
+    # pylint: disable=protected-access
+    field = message.Extensions._FindExtensionByName(name)
+    # pylint: enable=protected-access
+    if not field:
+      if allow_unknown_extension:
+        field = None
+      else:
+        raise tokenizer.ParseErrorPreviousToken(
+            'Extension "%s" not registered.' % name)
+    elif message_descriptor != field.containing_type:
+      raise tokenizer.ParseErrorPreviousToken(
+          'Extension "%s" does not extend message type "%s".' % (
+              name, message_descriptor.full_name))
+
+    tokenizer.Consume(']')
+
+  else:
+    name = tokenizer.ConsumeIdentifier()
+    field = message_descriptor.fields_by_name.get(name, None)
+
+    # Group names are expected to be capitalized as they appear in the
+    # .proto file, which actually matches their type names, not their field
+    # names.
+    if not field:
+      field = message_descriptor.fields_by_name.get(name.lower(), None)
+      if field and field.type != descriptor.FieldDescriptor.TYPE_GROUP:
+        field = None
+
+    if (field and field.type == descriptor.FieldDescriptor.TYPE_GROUP and
+        field.message_type.name != name):
+      field = None
+
+    if not field:
+      raise tokenizer.ParseErrorPreviousToken(
+          'Message type "%s" has no field named "%s".' % (
+              message_descriptor.full_name, name))
+
+  if field and field.cpp_type == descriptor.FieldDescriptor.CPPTYPE_MESSAGE:
+    is_map_entry = _IsMapEntry(field)
+    tokenizer.TryConsume(':')
+
+    if tokenizer.TryConsume('<'):
+      end_token = '>'
+    else:
+      tokenizer.Consume('{')
+      end_token = '}'
+
+    if field.label == descriptor.FieldDescriptor.LABEL_REPEATED:
+      if field.is_extension:
+        sub_message = message.Extensions[field].add()
+      elif is_map_entry:
+        sub_message = field.message_type._concrete_class()
+      else:
+        sub_message = getattr(message, field.name).add()
+    else:
+      if field.is_extension:
+        sub_message = message.Extensions[field]
+      else:
+        sub_message = getattr(message, field.name)
+      sub_message.SetInParent()
+
+    while not tokenizer.TryConsume(end_token):
+      if tokenizer.AtEnd():
+        raise tokenizer.ParseErrorPreviousToken('Expected "%s".' % (end_token))
+      _MergeField(tokenizer, sub_message, allow_multiple_scalars,
+                  allow_unknown_extension)
+
+    if is_map_entry:
+      value_cpptype = field.message_type.fields_by_name['value'].cpp_type
+      if value_cpptype == descriptor.FieldDescriptor.CPPTYPE_MESSAGE:
+        value = getattr(message, field.name)[sub_message.key]
+        value.MergeFrom(sub_message.value)
+      else:
+        getattr(message, field.name)[sub_message.key] = sub_message.value
+  elif field:
+    tokenizer.Consume(':')
+    if (field.label == descriptor.FieldDescriptor.LABEL_REPEATED and
+        tokenizer.TryConsume('[')):
+      # Short repeated format, e.g. "foo: [1, 2, 3]"
+      while True:
+        _MergeScalarField(tokenizer, message, field, allow_multiple_scalars)
+        if tokenizer.TryConsume(']'):
+          break
+        tokenizer.Consume(',')
+    else:
+      _MergeScalarField(tokenizer, message, field, allow_multiple_scalars)
+  else:  # Proto field is unknown.
+    assert allow_unknown_extension
+    _SkipFieldContents(tokenizer)
+
+  # For historical reasons, fields may optionally be separated by commas or
+  # semicolons.
+  if not tokenizer.TryConsume(','):
+    tokenizer.TryConsume(';')
+
+
+def _SkipFieldContents(tokenizer):
+  """Skips over contents (value or message) of a field.
+
+  Args:
+    tokenizer: A tokenizer to parse the field name and values.
+  """
+  # Try to guess the type of this field.
+  # If this field is not a message, there should be a ":" between the
+  # field name and the field value and also the field value should not
+  # start with "{" or "<" which indicates the beginning of a message body.
+  # If there is no ":" or there is a "{" or "<" after ":", this field has
+  # to be a message or the input is ill-formed.
+  if tokenizer.TryConsume(':') and not tokenizer.LookingAt(
+      '{') and not tokenizer.LookingAt('<'):
+    _SkipFieldValue(tokenizer)
+  else:
+    _SkipFieldMessage(tokenizer)
+
+
+def _SkipField(tokenizer):
+  """Skips over a complete field (name and value/message).
+
+  Args:
+    tokenizer: A tokenizer to parse the field name and values.
+  """
+  if tokenizer.TryConsume('['):
+    # Consume extension name.
+    tokenizer.ConsumeIdentifier()
+    while tokenizer.TryConsume('.'):
+      tokenizer.ConsumeIdentifier()
+    tokenizer.Consume(']')
+  else:
+    tokenizer.ConsumeIdentifier()
+
+  _SkipFieldContents(tokenizer)
+
+  # For historical reasons, fields may optionally be separated by commas or
+  # semicolons.
+  if not tokenizer.TryConsume(','):
+    tokenizer.TryConsume(';')
+
+
+def _SkipFieldMessage(tokenizer):
+  """Skips over a field message.
+
+  Args:
+    tokenizer: A tokenizer to parse the field name and values.
+  """
+
+  if tokenizer.TryConsume('<'):
+    delimiter = '>'
+  else:
+    tokenizer.Consume('{')
+    delimiter = '}'
+
+  while not tokenizer.LookingAt('>') and not tokenizer.LookingAt('}'):
+    _SkipField(tokenizer)
+
+  tokenizer.Consume(delimiter)
+
+
+def _SkipFieldValue(tokenizer):
+  """Skips over a field value.
+
+  Args:
+    tokenizer: A tokenizer to parse the field name and values.
+
+  Raises:
+    ParseError: In case an invalid field value is found.
+  """
+  # String tokens can come in multiple adjacent string literals.
+  # If we can consume one, consume as many as we can.
+  if tokenizer.TryConsumeString():
+    while tokenizer.TryConsumeString():
+      pass
+    return
+
+  if (not tokenizer.TryConsumeIdentifier() and
+      not tokenizer.TryConsumeInt64() and
+      not tokenizer.TryConsumeUint64() and
+      not tokenizer.TryConsumeFloat()):
+    raise ParseError('Invalid field value: ' + tokenizer.token)
+
+
+def _MergeScalarField(tokenizer, message, field, allow_multiple_scalars):
+  """Merges a single protocol message scalar field into a message.
+
+  Args:
+    tokenizer: A tokenizer to parse the field value.
+    message: A protocol message to record the data.
+    field: The descriptor of the field to be merged.
+    allow_multiple_scalars: Determines if repeated values for a non-repeated
+      field are permitted, e.g., the string "foo: 1 foo: 2" for a
+      required/optional field named "foo".
+
+  Raises:
+    ParseError: In case of text parsing problems.
+    RuntimeError: On runtime errors.
+  """
+  value = None
+
+  if field.type in (descriptor.FieldDescriptor.TYPE_INT32,
+                    descriptor.FieldDescriptor.TYPE_SINT32,
+                    descriptor.FieldDescriptor.TYPE_SFIXED32):
+    value = tokenizer.ConsumeInt32()
+  elif field.type in (descriptor.FieldDescriptor.TYPE_INT64,
+                      descriptor.FieldDescriptor.TYPE_SINT64,
+                      descriptor.FieldDescriptor.TYPE_SFIXED64):
+    value = tokenizer.ConsumeInt64()
+  elif field.type in (descriptor.FieldDescriptor.TYPE_UINT32,
+                      descriptor.FieldDescriptor.TYPE_FIXED32):
+    value = tokenizer.ConsumeUint32()
+  elif field.type in (descriptor.FieldDescriptor.TYPE_UINT64,
+                      descriptor.FieldDescriptor.TYPE_FIXED64):
+    value = tokenizer.ConsumeUint64()
+  elif field.type in (descriptor.FieldDescriptor.TYPE_FLOAT,
+                      descriptor.FieldDescriptor.TYPE_DOUBLE):
+    value = tokenizer.ConsumeFloat()
+  elif field.type == descriptor.FieldDescriptor.TYPE_BOOL:
+    value = tokenizer.ConsumeBool()
+  elif field.type == descriptor.FieldDescriptor.TYPE_STRING:
+    value = tokenizer.ConsumeString()
+  elif field.type == descriptor.FieldDescriptor.TYPE_BYTES:
+    value = tokenizer.ConsumeByteString()
+  elif field.type == descriptor.FieldDescriptor.TYPE_ENUM:
+    value = tokenizer.ConsumeEnum(field)
+  else:
+    raise RuntimeError('Unknown field type %d' % field.type)
+
+  if field.label == descriptor.FieldDescriptor.LABEL_REPEATED:
+    if field.is_extension:
+      message.Extensions[field].append(value)
+    else:
+      getattr(message, field.name).append(value)
+  else:
+    if field.is_extension:
+      if not allow_multiple_scalars and message.HasExtension(field):
+        raise tokenizer.ParseErrorPreviousToken(
+            'Message type "%s" should not have multiple "%s" extensions.' %
+            (message.DESCRIPTOR.full_name, field.full_name))
+      else:
+        message.Extensions[field] = value
+    else:
+      if not allow_multiple_scalars and message.HasField(field.name):
+        raise tokenizer.ParseErrorPreviousToken(
+            'Message type "%s" should not have multiple "%s" fields.' %
+            (message.DESCRIPTOR.full_name, field.name))
+      else:
+        setattr(message, field.name, value)
+
+
+class _Tokenizer(object):
+  """Protocol buffer text representation tokenizer.
+
+  This class handles the lower level string parsing by splitting it into
+  meaningful tokens.
+
+  It was directly ported from the Java protocol buffer API.
+  """
+
+  _WHITESPACE = re.compile('(\\s|(#.*$))+', re.MULTILINE)
+  _TOKEN = re.compile('|'.join([
+      r'[a-zA-Z_][0-9a-zA-Z_+-]*',             # an identifier
+      r'([0-9+-]|(\.[0-9]))[0-9a-zA-Z_.+-]*',  # a number
+  ] + [                                        # quoted str for each quote mark
+      r'{qt}([^{qt}\n\\]|\\.)*({qt}|\\?$)'.format(qt=mark) for mark in _QUOTES
+  ]))
+
+  _IDENTIFIER = re.compile(r'\w+')
+
+  def __init__(self, lines):
+    self._position = 0
+    self._line = -1
+    self._column = 0
+    self._token_start = None
+    self.token = ''
+    self._lines = iter(lines)
+    self._current_line = ''
+    self._previous_line = 0
+    self._previous_column = 0
+    self._more_lines = True
+    self._SkipWhitespace()
+    self.NextToken()
+
+  def LookingAt(self, token):
+    return self.token == token
+
+  def AtEnd(self):
+    """Checks the end of the text was reached.
+
+    Returns:
+      True iff the end was reached.
+    """
+    return not self.token
+
+  def _PopLine(self):
+    while len(self._current_line) <= self._column:
+      try:
+        self._current_line = next(self._lines)
+      except StopIteration:
+        self._current_line = ''
+        self._more_lines = False
+        return
+      else:
+        self._line += 1
+        self._column = 0
+
+  def _SkipWhitespace(self):
+    while True:
+      self._PopLine()
+      match = self._WHITESPACE.match(self._current_line, self._column)
+      if not match:
+        break
+      length = len(match.group(0))
+      self._column += length
+
+  def TryConsume(self, token):
+    """Tries to consume a given piece of text.
+
+    Args:
+      token: Text to consume.
+
+    Returns:
+      True iff the text was consumed.
+    """
+    if self.token == token:
+      self.NextToken()
+      return True
+    return False
+
+  def Consume(self, token):
+    """Consumes a piece of text.
+
+    Args:
+      token: Text to consume.
+
+    Raises:
+      ParseError: If the text couldn't be consumed.
+    """
+    if not self.TryConsume(token):
+      raise self._ParseError('Expected "%s".' % token)
+
+  def TryConsumeIdentifier(self):
+    try:
+      self.ConsumeIdentifier()
+      return True
+    except ParseError:
+      return False
+
+  def ConsumeIdentifier(self):
+    """Consumes protocol message field identifier.
+
+    Returns:
+      Identifier string.
+
+    Raises:
+      ParseError: If an identifier couldn't be consumed.
+    """
+    result = self.token
+    if not self._IDENTIFIER.match(result):
+      raise self._ParseError('Expected identifier.')
+    self.NextToken()
+    return result
+
+  def ConsumeInt32(self):
+    """Consumes a signed 32bit integer number.
+
+    Returns:
+      The integer parsed.
+
+    Raises:
+      ParseError: If a signed 32bit integer couldn't be consumed.
+    """
+    try:
+      result = ParseInteger(self.token, is_signed=True, is_long=False)
+    except ValueError as e:
+      raise self._ParseError(str(e))
+    self.NextToken()
+    return result
+
+  def ConsumeUint32(self):
+    """Consumes an unsigned 32bit integer number.
+
+    Returns:
+      The integer parsed.
+
+    Raises:
+      ParseError: If an unsigned 32bit integer couldn't be consumed.
+    """
+    try:
+      result = ParseInteger(self.token, is_signed=False, is_long=False)
+    except ValueError as e:
+      raise self._ParseError(str(e))
+    self.NextToken()
+    return result
+
+  def TryConsumeInt64(self):
+    try:
+      self.ConsumeInt64()
+      return True
+    except ParseError:
+      return False
+
+  def ConsumeInt64(self):
+    """Consumes a signed 64bit integer number.
+
+    Returns:
+      The integer parsed.
+
+    Raises:
+      ParseError: If a signed 64bit integer couldn't be consumed.
+    """
+    try:
+      result = ParseInteger(self.token, is_signed=True, is_long=True)
+    except ValueError as e:
+      raise self._ParseError(str(e))
+    self.NextToken()
+    return result
+
+  def TryConsumeUint64(self):
+    try:
+      self.ConsumeUint64()
+      return True
+    except ParseError:
+      return False
+
+  def ConsumeUint64(self):
+    """Consumes an unsigned 64bit integer number.
+
+    Returns:
+      The integer parsed.
+
+    Raises:
+      ParseError: If an unsigned 64bit integer couldn't be consumed.
+    """
+    try:
+      result = ParseInteger(self.token, is_signed=False, is_long=True)
+    except ValueError as e:
+      raise self._ParseError(str(e))
+    self.NextToken()
+    return result
+
+  def TryConsumeFloat(self):
+    try:
+      self.ConsumeFloat()
+      return True
+    except ParseError:
+      return False
+
+  def ConsumeFloat(self):
+    """Consumes an floating point number.
+
+    Returns:
+      The number parsed.
+
+    Raises:
+      ParseError: If a floating point number couldn't be consumed.
+    """
+    try:
+      result = ParseFloat(self.token)
+    except ValueError as e:
+      raise self._ParseError(str(e))
+    self.NextToken()
+    return result
+
+  def ConsumeBool(self):
+    """Consumes a boolean value.
+
+    Returns:
+      The bool parsed.
+
+    Raises:
+      ParseError: If a boolean value couldn't be consumed.
+    """
+    try:
+      result = ParseBool(self.token)
+    except ValueError as e:
+      raise self._ParseError(str(e))
+    self.NextToken()
+    return result
+
+  def TryConsumeString(self):
+    try:
+      self.ConsumeString()
+      return True
+    except ParseError:
+      return False
+
+  def ConsumeString(self):
+    """Consumes a string value.
+
+    Returns:
+      The string parsed.
+
+    Raises:
+      ParseError: If a string value couldn't be consumed.
+    """
+    the_bytes = self.ConsumeByteString()
+    try:
+      return six.text_type(the_bytes, 'utf-8')
+    except UnicodeDecodeError as e:
+      raise self._StringParseError(e)
+
+  def ConsumeByteString(self):
+    """Consumes a byte array value.
+
+    Returns:
+      The array parsed (as a string).
+
+    Raises:
+      ParseError: If a byte array value couldn't be consumed.
+    """
+    the_list = [self._ConsumeSingleByteString()]
+    while self.token and self.token[0] in _QUOTES:
+      the_list.append(self._ConsumeSingleByteString())
+    return b''.join(the_list)
+
+  def _ConsumeSingleByteString(self):
+    """Consume one token of a string literal.
+
+    String literals (whether bytes or text) can come in multiple adjacent
+    tokens which are automatically concatenated, like in C or Python.  This
+    method only consumes one token.
+
+    Returns:
+      The token parsed.
+    Raises:
+      ParseError: When the wrong format data is found.
+    """
+    text = self.token
+    if len(text) < 1 or text[0] not in _QUOTES:
+      raise self._ParseError('Expected string but found: %r' % (text,))
+
+    if len(text) < 2 or text[-1] != text[0]:
+      raise self._ParseError('String missing ending quote: %r' % (text,))
+
+    try:
+      result = text_encoding.CUnescape(text[1:-1])
+    except ValueError as e:
+      raise self._ParseError(str(e))
+    self.NextToken()
+    return result
+
+  def ConsumeEnum(self, field):
+    try:
+      result = ParseEnum(field, self.token)
+    except ValueError as e:
+      raise self._ParseError(str(e))
+    self.NextToken()
+    return result
+
+  def ParseErrorPreviousToken(self, message):
+    """Creates and *returns* a ParseError for the previously read token.
+
+    Args:
+      message: A message to set for the exception.
+
+    Returns:
+      A ParseError instance.
+    """
+    return ParseError('%d:%d : %s' % (
+        self._previous_line + 1, self._previous_column + 1, message))
+
+  def _ParseError(self, message):
+    """Creates and *returns* a ParseError for the current token."""
+    return ParseError('%d:%d : %s' % (
+        self._line + 1, self._column + 1, message))
+
+  def _StringParseError(self, e):
+    return self._ParseError('Couldn\'t parse string: ' + str(e))
+
+  def NextToken(self):
+    """Reads the next meaningful token."""
+    self._previous_line = self._line
+    self._previous_column = self._column
+
+    self._column += len(self.token)
+    self._SkipWhitespace()
+
+    if not self._more_lines:
+      self.token = ''
+      return
+
+    match = self._TOKEN.match(self._current_line, self._column)
+    if match:
+      token = match.group(0)
+      self.token = token
+    else:
+      self.token = self._current_line[self._column]
+
+
+def ParseInteger(text, is_signed=False, is_long=False):
+  """Parses an integer.
+
+  Args:
+    text: The text to parse.
+    is_signed: True if a signed integer must be parsed.
+    is_long: True if a long integer must be parsed.
+
+  Returns:
+    The integer value.
+
+  Raises:
+    ValueError: Thrown Iff the text is not a valid integer.
+  """
+  # Do the actual parsing. Exception handling is propagated to caller.
+  try:
+    # 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.
+    if is_long:
+      result = long(text, 0)
+    else:
+      result = int(text, 0)
+  except ValueError:
+    raise ValueError('Couldn\'t parse integer: %s' % text)
+
+  # Check if the integer is sane. Exceptions handled by callers.
+  checker = _INTEGER_CHECKERS[2 * int(is_long) + int(is_signed)]
+  checker.CheckValue(result)
+  return result
+
+
+def ParseFloat(text):
+  """Parse a floating point number.
+
+  Args:
+    text: Text to parse.
+
+  Returns:
+    The number parsed.
+
+  Raises:
+    ValueError: If a floating point number couldn't be parsed.
+  """
+  try:
+    # Assume Python compatible syntax.
+    return float(text)
+  except ValueError:
+    # Check alternative spellings.
+    if _FLOAT_INFINITY.match(text):
+      if text[0] == '-':
+        return float('-inf')
+      else:
+        return float('inf')
+    elif _FLOAT_NAN.match(text):
+      return float('nan')
+    else:
+      # assume '1.0f' format
+      try:
+        return float(text.rstrip('f'))
+      except ValueError:
+        raise ValueError('Couldn\'t parse float: %s' % text)
+
+
+def ParseBool(text):
+  """Parse a boolean value.
+
+  Args:
+    text: Text to parse.
+
+  Returns:
+    Boolean values parsed
+
+  Raises:
+    ValueError: If text is not a valid boolean.
+  """
+  if text in ('true', 't', '1'):
+    return True
+  elif text in ('false', 'f', '0'):
+    return False
+  else:
+    raise ValueError('Expected "true" or "false".')
+
+
+def ParseEnum(field, value):
+  """Parse an enum value.
+
+  The value can be specified by a number (the enum value), or by
+  a string literal (the enum name).
+
+  Args:
+    field: Enum field descriptor.
+    value: String value.
+
+  Returns:
+    Enum value number.
+
+  Raises:
+    ValueError: If the enum value could not be parsed.
+  """
+  enum_descriptor = field.enum_type
+  try:
+    number = int(value, 0)
+  except ValueError:
+    # Identifier.
+    enum_value = enum_descriptor.values_by_name.get(value, None)
+    if enum_value is None:
+      raise ValueError(
+          'Enum type "%s" has no value named %s.' % (
+              enum_descriptor.full_name, value))
+  else:
+    # Numeric value.
+    enum_value = enum_descriptor.values_by_number.get(number, None)
+    if enum_value is None:
+      raise ValueError(
+          'Enum type "%s" has no value with number %d.' % (
+              enum_descriptor.full_name, number))
+  return enum_value.number