http://gcc.gnu.org/bugzilla/show_bug.cgi?id=48894
Summary: generic omp_get_ancestor_thread_num(l(i)) produces
incorrect output
Product: gcc
Version: 4.5.3
Status: UNCONFIRMED
Severity: minor
Priority: P3
Component: fortran
AssignedTo: unassig...@gcc.gnu.org
ReportedBy: lo...@cray.com
I wavered between "minor" and "trivial" on this one, but the OpenMP testers
came up with this end case. It seems very unlikely to arise in a real program,
but is also probably easy to fix in the library code.
> cat test1.f90
! derived from OpenMP test OMP3f/F03_3_2_17_1a.F90
! See - OpenMP 3.0 API spec, p. 131, lines 2-9.
program F03_3_2_17_1a
use omp_lib
implicit none
logical, parameter :: FALSE=.false.
logical, parameter :: TRUE=.true.
integer, parameter :: ZERO=0
integer, parameter :: TWO=2
integer*4, parameter :: ZERO4=0
integer, parameter :: MHUGE4=-huge(ZERO4)
integer :: level(10) = &
(/-huge(ZERO), -huge(ZERO)+3, MHUGE4, -1, 0, 1, 2, 3, 4, huge(ZERO)/)
integer :: i
integer, dimension(10) :: expected_ancestor_thread_num, &
actual_ancestor_thread_num = (/(-huge(i), i=1,10)/)
call omp_set_dynamic( FALSE )
call omp_set_nested( TRUE )
call omp_set_max_active_levels(TWO) ! one inactive parallel region
! if the implementation supports nested parallelism, then expected value of
! omp_get_ancestor_thread(2) == 1, else 0
if (omp_get_nested()) then
expected_ancestor_thread_num = (/-1, -1, -1, -1, 0, 2, 1, 0, -1, -1/)
else
expected_ancestor_thread_num = (/-1, -1, -1, -1, 0, 2, 0, 0, -1, -1/)
end if
!$omp parallel num_threads(3) ! nest level 1
if (omp_get_thread_num() == 2) then ! ancestor thread num = 2 at this level
!$omp parallel num_threads(2) ! nest level 2
if ((omp_get_nested() .and. omp_get_thread_num() == 1) .or. &
(omp_get_nested() .eqv. .false. .and. omp_get_thread_num() == 0))
then
!$omp parallel ! nest level 3
! inactive parallel region due to max-active-level-var == 2,
! so should be executed by only one thread
do i = 1, size(level)
actual_ancestor_thread_num(i)
=omp_get_ancestor_thread_num(level(i))
end do
!$omp end parallel
end if
!$omp end parallel
end if
!$omp end parallel
do i = 1, size(level)
print *, 'level(i)=', level(i)
if (actual_ancestor_thread_num(i) /= expected_ancestor_thread_num(i))
then
print *, "FAIL - omp_get_ancestor_thread_num(", level(i), ")", &
" == ", actual_ancestor_thread_num(i), " (expected ", &
expected_ancestor_thread_num(i), ")"
end if
end do
end program F03_3_2_17_1a
> gfortran -fopenmp -fdefault-integer-8 -fdefault-real-8 test1.f90
> ./a.out
level(i)= -9223372036854775807
FAIL - omp_get_ancestor_thread_num( -9223372036854775807 ) ==
2 (expected -1 )
level(i)= -9223372036854775804
level(i)= -2147483647
level(i)= -1
level(i)= 0
level(i)= 1
level(i)= 2
level(i)= 3
level(i)= 4
level(i)= 9223372036854775807
>
>From the OpenMP spec from page 131, lines 2-5:
"The omp_get_ancestor_thread_num routine returns the thread number of the
ancestor at a given nest level of the current thread or the thread number of
the current thread. If the requested nest level is outside the range of 0 and
the nest level of the current thread, as returned by the omp_get_level routine,
the routine returns -1."
The routine omp_get_level() always returns a result that is 0 or a positive
integer. So any negative argument to omp_get_ancestor_thread_num that is
negative is outside the range 0..omp_get_level(), so the result should be -1.
I suspect what is happening here is that the computation inside
omp_get_ancestor_thread_num (n) involves a subtract of n - num_threads which
wraps around because the value of n is at the limit of representable integers.
Note that the second trial input is different by 3 (and num_threads is 3), so
it just barely avoids the wrap.
