Re: Boost serialization (bug?) issue with MSVC, with vector<shared_ptr> in several libraries

From: Robert Ramey (ramey_at_[hidden])
Date: 2019-11-22 00:36:10

On 11/21/19 3:06 PM, Nicolas Guillot via Boost wrote:
>>> Thanks a lot for your help. And if any clarification is needed, I'm
>>> available of course !
> 
>> Please do the following:
> 
>> a) double check closed issues on git on the serialization library. I
>> recently fixed a bug which might be related. It might be helpful to
>> look at this and see if it matches your case.
> 
>> b) Open up a github issue with all the information you've included in
>> this post.
> 
>> c) I'll take a look at it.
> 
>> Robert Ramey
> 
> Hi Robert and thanks for your answer.
> 
> I didn't know boost serialization project was on git hub. I only looked for
> a similar bug fix at boost website. And no release note since boost 1.58,

I fix bugs as they are found.  I make other small enhancements all the 
time.  I'm remiss in not updating the release notes.

> wahoo! no bug !
> 
> Please can you advice me a little bit: I guess the issue related to my
> problem is the issue 154 (
> https://github.com/boostorg/serialization/issues/154).

That's a good possibility
> 
> What do you advise: take the code from git, or only the bug fix? Is the
> code source from git compatible with last version of boost (1.71.0)?

I advise downloading the whole library - latest version and re-building 
it from scratch.  For reasons I don't want to into now - this is harder 
than it should be.  But it's the only way that you know that you have 
the latest package.

I also recommend that you run the test suite on your local system(s). 
This is an incredibly exhaustive series of tests which almost guarentees 
that you have a bug free combination of compiler, linker, other 
libraries etc.

This seems like it would be a lot of work - it is.  But it's better than 
playing whack-a-mole trying to understand problems via experimentation. 
These days any bug still existing is very, very, very, obscure and 
almost impossible to find.  It's very hard even to make a useful test 
case. And if you can reproduce it, it's almost impossible to fix. So I 
recommend downloading the whole of boost 1.71 and building everything 
from scratch.

> If you advice to take only the commit that fix this issue, which is it?

I don't advise this, but if you want to do it this way, you'll just have 
to read the comments as I don't remember.

> Now it's 0:00 in my part of the world, but tomorrow for sure I'll do the
> test and keep you inform.

Instead, how about doing what I suggest and informing me if it doesn't work.

> Thanks for your help, and congratulation for this library.

You're welcome, I'm happy to be of service.

Robert Ramey
> 
> 
> On Thu, Nov 21, 2019 at 5:51 PM Nicolas Guillot <nicolas.guillot_at_[hidden]>
> wrote:
> 
>> We use boost serialization in ours projects, both on linux (gcc) and
>> Windows.
>>
>> We encounter a problem only on Windows (for "every" versions of Windows (7
>> and 10) and of MSVC: Visual 2010 and 2015, ie. MSVC++ 10.0 _MSC_VER == 1600
>> or MSVC++ 14.0 _MSC_VER == 1900).
>>
>> *Here is the problem summarized*, but I need to give more details after:
>> When I serialize a class
>>
>>     - which have base class
>>     - and have 2 vector of shared_ptr of ClassB, with the same contents,
>>     and B inherits from the same base class
>>
>> the serialization works, but the deserialization doesn't ("input stream
>> error").
>>
>> The following image is here : https://i.stack.imgur.com/osoiF.png (if
>> it's not displayed)
>>
>> [image: enter image description here]
>> <https://i.stack.imgur.com/osoiF.png>
>>
>> Yes, it's very strange:
>>
>>     - if I remove the base class, it works.
>>     - if I inline the serialization code (empty function) of the base
>>     class, it works.
>>     - if the 2 vectors don't have the same content, it works !!
>>     -
>>
>>     if I register the ClassBase in the serialize method of ClassTest, it
>>     works. I.e.:
>>
>>     // in ClassTest:template <class Archive>void serialize(Archive & ar, const unsigned int version) {
>>         ar.template register_type<ClassBase>();
>>         ...}
>>
>>
>> *Some details*
>>
>> *When it works* on linux, and when it works on Windows with ClassBase
>> type registration in the archive, we have the same xml output.
>> It's worth noting that:
>>
>>     - boost_serialization is version 17 on Windows and 9 on linux (I
>>     didn't tried to update boost on linux yet).
>>     - On BaseClass xml tag, the tracking_level="1", so we have object_id.
>>
>> *And when it doesn't work* (once again: the serialization of ClassTest is
>> OK, but cannot deserialized "input stream error"):
>>
>>     - The main difference is in the serialization of ClassTest: the
>>     ClassBase part has no tracking_level.
>>     - However it's not the only cause, because if the vectors are
>>     differents (same 2 first items, and a third item in one vector),
>>     tracking_level of ClassBase in ClassTest part is "0".
>>
>> Last thing, if it can give you some clues: while we debugged our
>> application, we noticed that the deserialization framework tries to
>> deserialize a class into another.
>> It makes me think the serialization framework is confused about the classe
>> types identifications.
>> Could it be a matter of dll linking? Of global static shared storage ?
>>
>> I've posted a visual studio solution here
>> <https://1drv.ms/u/s!Aoy4ax0S3ZS5gkxv1UIvRUd32oVO?e=ACVi4c>.
>>
>> Thanks a lot for your help. And if any clarification is needed, I'm
>> available of course !
>>
> 
> _______________________________________________
> Unsubscribe & other changes: http://lists.boost.org/mailman/listinfo.cgi/boost
>