OMP flush with component of UDT

[jimdempseyatthecove]

module foo
type someType
    logical :: flag = .false.
end type someType
end module foo
program omp_flush_issue
    use foo
    implicit none
    type(someType) :: myType
    !$omp flush(myType%flag)
end program omp_flush_issue

Compiling with Intel® Fortran Compiler 2025.0.0 [Intel(R) 64]...
omp_flush_issue.f90
C:\test\omp_flush_issue\omp_flush_issue\omp_flush_issue.f90(23): error #5082: Syntax error, found '%' when expecting one of: ) ,
C:\test\omp_flush_issue\omp_flush_issue\omp_flush_issue.f90(23): error #6404: This name does not have a type, and must have an explicit type. [FLAG]
compilation aborted for C:\test\omp_flush_issue\omp_flush_issue\omp_flush_issue.f90 (code 1)

As a hack fix, you can call a subroutine with (myType%flag) then issue the flush directive on the reference to the variable.

Jim Dempsey

[jimdempseyatthecove]

Example:

...
#if 0    
        !$omp flush(myContext%ACCvalid)
        do while(.not.myContext%ACCvalid)
        !$omp flush(myContext%ACCvalid)
        end do
#else
        call omp_flush_hack(myContext%ACCvalid)
#endif
...
subroutine omp_flush_hack(myACCvalid)
    logical :: myACCvalid
    !$omp flush(myACCvalid)
    do while(.not.myACCvalid)
    !$omp flush(myACCvalid)
    end do
end subroutine omp_flush_hack

Jim Dempsey

[Xiaoping_D_Intel]

At OpenMP 6.0 spec "5.2.1 OpenMP Argument Lists -> Restrictions -> Fortran" it is written that "Unless otherwise specified, a variable that is part of an aggregate variable must not be a variable list item or an extended list item" so the compiler issues the error on your first case.

Gfortran also provides similar compilation error:

$ gfortran -fopenmp -c test.f90

test.f90:10:16:

  10 |   !$omp flush(myType%flag)

   |        1

Error: Unexpected junk after $OMP FLUSH statement at (1)

[jimdempseyatthecove]

(unless I am missing something) this restriction is totally unnecessary. I suspect this is a committee disagreement as opposed to technical issue.

An alternate work around is

...
    associate(kludge=>myType%flag)
    !$omp flush(kludge)
    do while(.not.kludge)
    !$omp flush(kludge)
    end do
    end associate
...

FWIW the above can be used where some team members, but not all, need a synchronization point and where !$omp barrier/single cannot be used.

A potential issue with member variable within a UDT is:

If the UDT has (BIND(C) .OR. SEQUENCE) .AND. (the member variable is not naturally aligned .AND. falls across cache line boundary)

then, potentially, the entire variable might not get flushed.

This restriction means thread shared and modifiable variables cannot be within UDT types. Overly restrictive IMHO

Jim Dempsey