http://gcc.gnu.org/bugzilla/show_bug.cgi?id=52540
--- Comment #10 from Kai Tietz <ktietz at gcc dot gnu.org> 2012-03-11 18:35:59 UTC --- (In reply to comment #9) > Then this is a very old issue, isn't it? Why nobody noticed it before? To be > clear, in debug-mode we don't use extern templates only for basic_string > (because _GLIBCXX_EXTERN_TEMPLATE becomes -1 when _GLIBCXX_DEBUG is defined), > all the other facilities remain the same still use extern tamplate. This is in > order to enable debug-mode checks also at -O0 for basic_string. Now, I don't > see why a few less extern template can make a difference on mingw in terms of > correctness of use_facet or anything else, but if that's really the case due > to > fundamental limitations somehwhere, we can trade it for the -O0 basic_string > checks on the affected systems, I have no problems with that (note that, > AFAICS, the problem will not go away with a new C++11 conforming string class, > unless we decide to not export from the .so parts of basic_string) I tested it by making sure that _GLIBCXX_EXTERN_TEMPLATE is defined to 1 even for _GLIBCXX_DEBUG (and PROFILE) mode. Indeed the test passes now without issue. The underlying issue here seems to be that DLL-code and non-external code aren't sharing all underlying data. For ELF this gets resolved, as ELF doesn't have final link for .so, but for PE it gets finally linked on DLL generation.
