Possible data race on boost::detail::get_current_thread_data()

[sochsenk]

I'm using boost 1.70.0

I'm not sure if this is a known issue or a false positive. I got helgrind issues (possible data races) reported when running the following code snippet:

   constexpr unsigned NR_THREADS = 3;
   boost::barrier startBarrier(NR_THREADS);
   boost::barrier endBarrier(NR_THREADS);

   auto accumulate = [&]() {
      unsigned sum = 0;
      for (unsigned j = 0; j < 100; ++j) {

         startBarrier.wait();

         for (unsigned i = 0; i < 1000; ++i) {
            sum += i;
         }

         endBarrier.wait();
      }
      return sum;
   };

   std::vector< boost::unique_future< unsigned > > results;
   std::array< std::thread, NR_THREADS > threads;
   for (unsigned t = 0; t < NR_THREADS; ++t) {

      boost::packaged_task< unsigned > task{accumulate};
      results.push_back(task.get_future());
      threads[t] = std::thread(std::move(task));   
   }

   for (auto&& r : results) {
      std::cout << r.get() << "\n";
   }

   for (auto&& t : threads) {
      t.join();
   }

Note that I'm using std::thread but I got similar issues with boost::thread Besides other issues, it seems waiting for the future (r.get()) and the barrier waits, both seem to access boost::detail::get_current_thread_data(). Part of the helgrind report:

==32289== Possible data race during read of size 4 at 0x6C5BBA8 by thread #1
==32289== Locks held: 1, at address 0x8CFB630
==32289==    at 0x6A52CCA: boost::thread_detail::enter_once_region(boost::once_flag&) (in ...)
==32289==    by 0x6A47E8F: boost::detail::get_current_thread_data() (in /...)
==32289==    by 0x4AE5CD: boost::condition_variable::wait(boost::unique_lock<boost::mutex>&) (in ...)
==32289==    by 0x4AF0EA: boost::detail::shared_state_base::wait_internal(boost::unique_lock<boost::mutex>&, bool) (in ...)
==32289== 
==32289== This conflicts with a previous write of size 4 by thread #2
==32289== Locks held: 2, at addresses 0x6C5BC80 0x1FFEFFEEB0
==32289==    at 0x6A52DA0: boost::thread_detail::commit_once_region(boost::once_flag&) (in ...)
==32289==    by 0x6A47EB2: boost::detail::get_current_thread_data() (in ...)
==32289==    by 0x4AE5CD: boost::condition_variable::wait(boost::unique_lock<boost::mutex>&) (in ...)
==32289==    by 0x4AE82A: boost::barrier::wait() (in ...)

[sochsenk]

I'm still analyzing, but my guess is there are boost::atomics somewhere below and Helgrind doesn't understand atomics.

[sochsenk]

I think it is a false positive due to Helgrinds inability to handle atomics. threadsanitizer couldn't find anything.