On 8 May 2018 at 15:45, Marc Glisse <marc.gli...@inria.fr> wrote: > On Tue, 8 May 2018, Jonathan Wakely wrote: > >> On 8 May 2018 at 14:00, Jonathan Wakely wrote: >>> >>> On 8 May 2018 at 13:44, Stephan Bergmann wrote: >>>> >>>> I was recently bitten by the following issue (Linux, libstdc++ 8.0.1): A >>>> process loads two dynamic libraries A and B both using std::regex, and A >>>> is >>>> compiled without -D_GLIBCXX_DEBUG while B is compiled with >>>> -D_GLIBCXX_DEBUG. >>> >>> >>> This is only supported in very restricted cases. >>> >>>> B creates an instance of std::regex, which internally creates a >>>> std::shared_ptr<std::__detail::_NFA<std::__cxx11::regex_traits<char>>>, >>>> where _NFA has various members of std::__debug::vector type (but which >>>> isn't >>>> reflected in the mangled name of that _NFA instantiation itself). >>>> >>>> Now, when that instance of std::regex is destroyed again in library B, >>>> the >>>> >>>> std::shared_ptr<std::__detail::_NFA<std::__cxx11::regex_traits<char>>>::~shared_ptr >>>> destructor (and functions it in turn calls) that happens to get picked >>>> is >>>> the (inlined, and exported due to default visibility) instance from >>>> library >>>> A. And that assumes that that _NFA instantiation has members of >>>> non-debug >>>> std::vector type, which causes a crash. >>>> >>>> Should it be considered a bug that such mixture of debug and non-debug >>>> std::regex usage causes ODR violations? >>> >>> >>> Yes, but my frank response is "don't do that". >>> >>> The right fix here might be to ensure that _NFA always uses the >>> non-debug vector even in Debug Mode, but I'm fairly certain there are >>> other similar problems lurking. >> >> >> N.B. I think this discussion belongs on the libstdc++ list. > > > Would it make sense to use the abi_tag attribute to help with that? (I > didn't really think about it, maybe it doesn't)
Yes, I think we could add it conditionally in debug mode, so that types with members that are either std::xxx or __gnu_debug::xxx get a different mangled name in debug mode. For the regex _NFA type I don't think we want the debug mode checking, because users can't access it directly so any errors are in the libstdc++ implementation and we should have eliminated them ourselves, not be pushing detection of those logic errors into users' programs. For std::match_results (which derives from std::vector) it's possible for users to use invalid iterators obtained from a match_results, so Debug Mode can help. In that case we could decide whether to add the abi_tag, or always derive from _GLIBCXX_STD_C::vector (i.e. the non-debug mode one), or even provide an entire __gnu_debug::match_results type. > "don't do that" remains the most sensible answer. Yes, it's just asking for trouble.
