libgomp.fortran/recursion1.f90 aborted at random on Linux/Core i7. There are --- implicit none integer :: i,s s=0 !$omp parallel do private(i) shared(s) do i=1,10 call sub(i) end do !$omp end parallel do if (s/=55) call abort() contains subroutine sub (n) integer :: n s = s + n print '(A,i3)',"loop =",n end subroutine end --- 's' isn't always 55. It has the value of the last running thread. Does OpenMP guarantees that 's' will be 55?
There is no such guarantee, the testcase contains data race. It can increment a shared variable in multiple threads using non-atomic instructions. One fix would be to use !$omp atomic s = s + n instead of just s = s + n Or put s = s + n into !$omp single, or use some other kind of locking.
A patch is posted at http://gcc.gnu.org/ml/gcc-patches/2010-01/msg00141.html
Subject: Bug 42602 Author: hjl Date: Mon Jan 4 14:28:30 2010 New Revision: 155615 URL: http://gcc.gnu.org/viewcvs?root=gcc&view=rev&rev=155615 Log: Make 's' atomic 2010-01-04 H.J. Lu <hongjiu.lu@intel.com> PR libgomp/42602 * libgomp.fortran/recursion1.f90 (sub): Make 's' atomic. Modified: trunk/libgomp/ChangeLog trunk/libgomp/testsuite/libgomp.fortran/recursion1.f90
Now I got At line 14 of file /export/gnu/import/svn/gcc-test/src-trunk/libgomp/testsuite/libgomp.fortran/recursion1.f90 Fortran runtime error: Recursive call to nonrecursive procedure 'sub'
(In reply to comment #4) > Now I got > > At line 14 of file > /export/gnu/import/svn/gcc-test/src-trunk/libgomp/testsuite/libgomp.fortran/recursion1.f90 > Fortran runtime error: Recursive call to nonrecursive procedure 'sub' > This is PR 42517. A patch is posted at http://gcc.gnu.org/ml/gcc-patches/2010-01/msg00187.html