Boost C++ Libraries: Ticket #8642: Global locale prevents from using Boost.Filesystem in global constructors and destructors

status changed

Beman Dawes — Sat, 02 Aug 2014 16:17:53 GMT

status new → assigned

Beman Dawes — Wed, 29 Oct 2014 13:40:53 GMT

Andy,

Could you please supply a simple test program that exhibits this problem?

Even if it is fixed now, I'd like to include something in the test suite that would detect any future regression.

Thanks,

--Beman

Andrey Semashev — Wed, 29 Oct 2014 14:45:19 GMT

I don't have a code sample at hand; there were reports about Boost.Log (e.g. #9119) which I debugged to this root problem. I can try and write a self-contained test later this week.

daniel kruegler — Fri, 31 Oct 2014 22:16:50 GMT

The following example should demonstrate the problem for every currently supported platform:

#include "boost/filesystem/path.hpp"
namespace fs = boost::filesystem;
template<class>
struct conv_char;
template<>
struct conv_char<wchar_t>
{
  typedef char value_type;
  static const value_type* file_name() { return "abcdefghijklmnopqrst.uvw"; }
};
template<>
struct conv_char<char>
{
  typedef wchar_t value_type;
  static const value_type* file_name() { return L"abcdefghijklmnopqrst.uvw"; }
};
typedef conv_char<fs::path::value_type> path_conv_char;
struct A
{
  A()
  {
  }
  ~A()
  {
    fs::path pa(path_conv_char::file_name());
  }
  static A& instance()
  {
    static A result;
    return result;
  }
};
struct B
{
  B()
  {
  }
  ~B()
  {
    A::instance();
    fs::path pb(path_conv_char::file_name());
  }
} b;
int main()
{
}

The code ensures that we perform a code conversion in all cases and takes advantage of the life-cycle rules of objects with static storage duration. It gives me a seg fault using gcc and a similar crash for Visual Studio in the destructor of A. The example provokes that the locale produced by path_locale() (within path.cpp) is already destroyed when pa becomes constructed, because by path_locale() holds another function local static.

cc set

Fri, 31 Oct 2014 23:15:40 GMT

cc daniel.kruegler@… added

attachment set

Andrey Semashev — Tue, 04 Nov 2014 18:38:56 GMT

attachment → ticket8642.cpp

A testcase to reproduce the problem

attachment set

Andrey Semashev — Tue, 04 Nov 2014 18:39:14 GMT

attachment → ticket8642.valgrind.log

Valgrind log

Andrey Semashev — Tue, 04 Nov 2014 18:40:40 GMT

I attached a test case which works similarly to Boost.Log. I also attached a valgrind log.

cc changed

raad@… — Mon, 27 Jun 2016 13:12:56 GMT

cc raad@… added

bchretien — Tue, 07 Mar 2017 09:17:41 GMT

Replying to andysem:

I attached a test case which works similarly to Boost.Log. I also attached a valgrind log.

Note: to test the repro code from ticket8642.cpp, remember to set your locale (e.g. LANG=en_US ./ticket8642).