Hi Bernd,
I believe that all omp threads are created in detached state,
so pthread_join should be undefined on them, just tsan*thinks*
otherwise?
When I look further on the libgomp sources, I see there
are two completely different implementations of the
mutexes, barriers, etc.
One using posix functions which should be visible to tsan
and another one using raw linux futex calls which should be
invisible to tsan, by default the linux system calls are
used, which explains why tsan seems to be unaware of the
actual synchronization in this example.
Have you tried to build with --enable-linux-futex=no ?
I just did that, and it appears that the thread sanitizer
no longer detects any issues even without my patch
(at least on powerpc64-unknown-linux-gnu, gcc110),
which showed the behavior before.
So, what is the correct way forward?
Modifying apparently correct code in libgfortran isn't the
answer, I think.
Having false positives when users specify -fopenmp -fsanitize=thread
is also not nice - this makes debugging multithreaded applications
hard.
Would it be possible to build a posix-thread-only version
of libgomp to be linked in when -fsanitize=thread is
supplied? Jakub, any suggestions?
Regards
Thomas
