http://gcc.gnu.org/bugzilla/show_bug.cgi?id=59546
--- Comment #1 from Jakub Jelinek <jakub at gcc dot gnu.org> --- i is uninitialized before the reduction, so if you add some number of threads to an uninitialized value, you can get pretty much anything. Also, there is no guarantee !$omp parallel will create exactly omp_get_max_threads () threads, that is just maximum for the number of threads, it can depend on env vars etc.
