Re: [boost] [1.39.0] Beta candidate

Subject: Re: [boost] [1.39.0] Beta candidate
From: Robert Ramey (ramey_at_[hidden])
Date: 2009-04-25 22:13:32

troy d. straszheim wrote:
> Ruediger Berlich wrote:
>> Hi Troy,
>>
>> gdb reports:
>>
>> Program received signal SIGSEGV, Segmentation fault.
>> 0x00007ffff60274b8 in
>> boost::serialization::void_cast_detail::void_caster::recursive_unregister()
>> const ()
>>    from
>> /opt/boost139/lib/libboost_serialization-gcc43-mt-1_39.so.1.39.0

FYI - one of the most significant changes from 1.38 to 1.39 is a change
to void_cast system.  Pre 1.38 void_cast was not thread safe.  1.38
made void_cast (and the whole library) threadsaft.  But the required
void_cast implementation was very, very slow.  The changes in
1.39 correct the speed problem while still keeping the library
thread safe at the cost of making void_cast ever more subtle
and intricate.

It would be fairly easy to test this by running the serializaiton tests
and verifying that only those tests which invoke void cast are
the ones that fail.

>>
>> More (albeit IMHO not very helpful) details from the backtrace can
>> be found at the end of this post. Somehow looks like an infinite
>> recursion ? Note that I can of course not rule out a problem in my own 
>> code.
>> However, this "feels" like the pre-1.38 problem in that it does not
>> affect the runtime behaviour of my programs (nor of the library's
>> tests) and that the segfault happens after the last statement in
>> main(). And as I said, the code runs fine with 1.38 .
>>
>
> Hm.  I can't reproduce it.  I'm not getting the error with the release
> beta.
>
> One thing I stumbled on when looking at this last time was the
>  following gcc flag:
>
> ----
>
>   -fuse-cxa-atexit
>
>     Register destructors for objects with static storage duration with
> the __cxa_atexit function rather than the atexit function. This option
> is required for fully standards-compliant handling of static
> destructors, but will only work if your C library supports
> __cxa_atexit.

You could be on to something here. One thing that I discovered
is that on some platforms, the order of static destructors is
not necessarily the reverse order of the corresponding static
constructors.  To address this, we've added some extra
if statements which in theory should never be necessary but
seems in fact that they might be.

If you could indicate the line of source code, we might be
able to apply this fix.

Robert Ramey

>
> ----
>
> and as I recall, '-fno-use-cxa-atexit' was on by default on the linux
> distribution I was on (might have been ubuntu 8.04).   I'm still
> curious as to when that flag could possibly be useful.
>
> -t
>
> _______________________________________________
> Unsubscribe & other changes:
> http://lists.boost.org/mailman/listinfo.cgi/boost