[Flang] [OpenMP] Do concurrent generate incorrect results when compared to the serial version

[Thirumalai-Shaktivel]

MRE:

program main
    integer, parameter :: V = 2048
    real(8) :: x(V), y(V)

    x = 100.0
    y = 200.0
    print *, cgdot(x, y, V)

contains
    function cgdot (x,y,N)
        integer :: N, i
        real(8) :: cgdot
        real(8), dimension(N) :: x, y
        cgdot=0.
        do concurrent(i=1:N)
            cgdot=cgdot+x(i)*y(i)
        enddo
        return
    end
end program

Problem:

$ flang -fopenmp -fdo-concurrent-to-openmp=host x.f90
$ ./a.out
 35780000.

Correct result

$ flang -fopenmp -fdo-concurrent-to-openmp=none x.f90
$ ./a.out
 40960000.

[m-saito-fj]

Hi.
In LLVM 21.1.2, when the code was written as follows, the results with fdo-concurrent-to-openmp=host and fdo-concurrent-to-
openmp=none were consistent.

program main
    integer, parameter :: V = 2048
    real(8) :: x(V), y(V)

    x = 100.0
    y = 200.0
    print *, cgdot(x, y, V)

contains
    function cgdot (x,y,N)
        integer :: N, i
        real(8) :: cgdot
        real(8), dimension(N) :: x, y
        cgdot=0.
        do concurrent(i=1:N) reduce(+:cgdot)
            cgdot=cgdot+x(i)*y(i)
        enddo
        return
    end
end program

I believe the reduce clause is necessary because there is a reduction operation.