https://gcc.gnu.org/bugzilla/show_bug.cgi?id=68080
--- Comment #3 from Martin Liška <marxin at gcc dot gnu.org> --- Well, the test-case is very interesting as it exposes 2 issues: 1) Majority of instrumented profiling code is not thread-safe, for instance edge profiler: PROF_edge_counter_11 = __gcov0._ZL20__gthread_mutex_lockP15pthread_mutex_t[0]; PROF_edge_counter_12 = PROF_edge_counter_11 + 1; __gcov0._ZL20__gthread_mutex_lockP15pthread_mutex_t[0] = PROF_edge_counter_12; 2) As the test-case does not properly join created threads and original thread reaches gcov_exit, we get messed profile. One thread is writing to disk while the others are still modifying counters. I'm going to discuss the aforementioned problem on GCC mailing list.
