lastprivate behavior of SIMD clause

[ivan-pi]

Directive and clause it applies to

The omp simd directive should have lastprivate behavior by default.

Specification details

The 5.2 spec says under 5.1.1

The loop iteration variables in the associated loops of a simd construct with multiple associated loops are lastprivate.
...
The loop iteration variables in any associated loop of a loop-associates construct is otherwise private.

The OpenMP 6.0 spec says under 7.1.1

The loop-iteration variable in any affected loop of a loop or simd construct is lastprivate.

Pseudocode

The test could be an example similar to this one:

program foo
   implicit none
   integer(kind=4) :: i, j, me1, me2

   me1 = 1
   me2 = 1

   i = me1
   j = 42 !...dummy flag to test IV after loop exit

   !$omp simd
   do j=i+1, me2
     me1 = me1+1
   enddo

   print *, 'me1 ', me1
   print *, 'IV j should be 2, is: ', j

end program foo

Ideally the test should cover loops with both positive and non-positive trip count.

Any other comments

Related discussions:

• Fortran Discourse: https://fortran-lang.discourse.group/t/zero-iteration-do-loop-and-omp-simd-behavior/8392
• GCC Bugzilla: https://gcc.gnu.org/bugzilla/show_bug.cgi?id=110718
• LLVM Issues: [Flang] [OpenMP] Loop iteration variable in OpenMP simd construct incorrectly treated as private instead of lastprivate llvm/llvm-project#170784

[ivan-pi]

Besided the zero-trip-count case, it would be good to also capture multiple nested loops:

program test_omp_simd
implicit none

integer, parameter :: m = 32, n = 10
real :: a(m,n)
integer :: i, j

! Section 5.1.1, pg. 97, line 16-17
! The loop iteration variables in the associated loops of a
! simd construct with multiple associated loops are lastprivate.

!$omp simd collapse(2)
do j = 1, n
do i = 1, m
    a(i,j) = 10*i + j
end do
end do

if (i /= m+1) error stop 1
if (j /= n+1) error stop 2

associate(aseq => reshape([((10*i + j,i=1,m),j=1,n)],[m,n]) )
    if ( any(a /= aseq) ) error stop 3
end associate

print *, "PASS"

end program

A quick Compiler Explorer check show multiple compilers fail here: https://godbolt.org/z/4xj6zYKas