Re: [Boost-users] serialization - portability

From: Robert Ramey (ramey_at_[hidden])
Date: 2007-08-19 14:37:44

When the collections were first made, I made the decision that,
for efficiency purposes, that collections of intrinsic types would
have a lower implemenation level - no class information.  This
should be faster and I never expected the need for versioning.

So in the standard .../serialization/vector.hpp there is something
which sets the implementation level for these types.  It would
seem that "long long" matches to one type on a 64 bit machine
and to a different type on a 32 bit machine,  and that one
of them maps to a type in one in the list of intrinsics.

I'd have to think about the best way to handle it.

Robert Ramey

Stefaan wrote:
> Hi!
>
> I'm having an issue with archive portability.  As I'm quite new to the
> internals of boost, I hope my attempt at pinpointing the problem is
> sensible:
>
> #include <iostream>
> #include <boost/serialization/vector.hpp>
>
> int main()
> {
> std::cerr << boost::serialization::implementation_level <std::vector
> <long long> >::value << std::endl;
> return 0;
> }
>
> built on a x86_64 platform prints out "3", on a i386 platform it
> prints out "2". I tested this both with boost 1.33.1 and 1.34.1. It is
> my guess that it causes the serialization of the used type to result
> in a larger archive on x86_64 than on i386.  (Which is the original
> problem that led me to discover this)
> So my question is: is this the result of a bug in boost? If not, then
> what is the explanation of this difference?
>
> Using linux-2.6 with gcc.
>
> Thanks,
> Stefaan