Re: [Boost-bugs] [Boost C++ Libraries] #3747: Serialization code speculate in order of static initialization

Subject: Re: [Boost-bugs] [Boost C++ Libraries] #3747: Serialization code speculate in order of static initialization
From: Boost C++ Libraries (noreply_at_[hidden])
Date: 2010-03-30 07:44:38

#3747: Serialization code speculate in order of static initialization
--------------------------------------------------+-------------------------
  Reporter:  Runar Undheim <r.undheim@…>          |       Owner:  ramey                     
      Type:  Bugs                                 |      Status:  reopened                  
 Milestone:  Boost 1.43.0                         |   Component:  serialization             
   Version:  Boost 1.41.0                         |    Severity:  Showstopper               
Resolution:                                       |    Keywords:  serialization static order
--------------------------------------------------+-------------------------

Comment(by Runar Undheim <r.undheim@…>):

 ''What is the problem?''[[BR]]
 The BOOST_CLASS_EXPORT does not register classes correctly. I got an
 exception from basic_oarchive.cpp line 351 (boost 1.42)
 (archive_exception(archive_exception::unregistered_class)) during export.
 This message only come when CLASS_EXPORT_IN_MAIN is 0. I got a pointer to
 serialization::extended_type_info (variable eti), but the m_key is NULL.

 ''Why should static initialization be in some particular order?''[[BR]]
 It should not, the order is decided by the compiler. But why does the code
 work when the static initialization is in one particular order and not in
 an other order? The serialization code should handle any order.

 ''I unzipped the files. and tried to compile and got ...'' [[BR]]
 I am sorry, but I work with VisualC++ 2005 where abstract is a legal
 keyword. But you could just remove the abstract behind "class Base". It
 does not matter on how the code is working. I think it may be impossible
 to reproduce this on other compilers, because it is the order of the
 static initialization that cause the problem. I have only tested with
 Visual C++ 2008 and Visual C++ 2010 RC, but I got the same problem there.

 I set a break point in constructor to extended_type_info. When
 CLASS_EXPORT_IN_MAIN is 0 the key is always NULL pointer. When
 CLASS_EXPORT_IN_MAIN is 1 I got a valid key.

 {{{
     extended_type_info_typeid() :
         typeid_system::extended_type_info_typeid_0(get_key())
     {
         type_register(typeid(T));
         key_register();
     }

     const char * get_key() const {
         return boost::serialization::guid<T>();
     }

 }}}

 The extended_type_info_typeid::get_key function speculate in that the
 boost::serialization::guid<T>() is registered.

 So this static function:
 {{{
 boost::serialization::singleton<boost::archive::detail::oserializer<boost::archive::text_oarchive,Main>
 >::instance''(void)
 }}}

 speculate that this function is called before:

 {{{
 boost::archive::detail::`anonymous namespace'::init_guid<Main>::g''(void)
 }}}

 This is not always the case. At least not with Visual Studio C++ 2005 and
 newer. See order_of_static_registration_fail.txt for order of static
 initialization when the code fail and see
 order_of_static_registration_ok.txt for order when everything works nice.

 Runar Undheim

-- 
Ticket URL: <https://svn.boost.org/trac/boost/ticket/3747#comment:12>
Boost C++ Libraries <http://www.boost.org/>
Boost provides free peer-reviewed portable C++ source libraries.