is valid this use of forall

[lluvia_lists@lavabit.com]

>> forall (i=1:n) a(b(i)) = a(b(i)) + 1
> [..]
>> However, I'm debugging an erratic program based on that and I'm asking
>> myself if that could be reason. I'm using a cluster with different CPUs
>> of
>> different procesor families.
>
> I changed the forall by a do loop and I didn't see the strange behavior
> yet. I don't consider it a proof yet, but maybe it could be bug, maybe
> solved in since a newer version. I'm using GNU Fortran (SUSE Linux) 4.5.1
> 20101208 [gcc-4_5-branch revision 167585]

CONFIRMED: It is causing differences. I used the test of below and have an
infinite loop in a machine with one CPU and final execution at first
iteration in another machine. The exact result on the last one was:

 forall:            1 do:           5
 forall:            0 do:           0
 forall:            0 do:           0
 forall:            0 do:           0
 forall:            0 do:           0

Can anyone explain me if it is the desired behavior?

Thanks,

program test

  integer, parameter :: d=5
  integer :: a1(d), a2(d), b(d)
  integer :: i

  do
     ! generate a random array of indexes
     b  = 1
     a1 = 0
     a2 = 0
     forall (i=1:d) a1(b(i)) = a1(b(i)) + 1

     do i=1, d
        a2(b(i)) = a2(b(i)) + 1
     enddo

     if (any((a1-a2) .ne. 0)) then
        do i=1, d
           print *, "forall: ", a1(i), "do:", a2(i)
        enddo
        call system ("touch ERROR_TEST_ERROR")
        stop
     else
        print  *, "hello"
     endif
  enddo

end program test