Optimizer bug

mecej4 · ‎09-21-2010

The reproducer given below is a simplified adaptation (no reals, smaller arrays, bound checking not involved) of a recently reported bug.

[fortran]!  IFort 11.1.073 on Linux IA32 or X64
!  Different output with -save -O0
!                    and -save -O1
program optbug
   implicit none
   integer :: i, sout

   do i=1,2
      call sub(i,sout)
      write(*,'(1x,I3,2x,I4)')i,sout
   end do
end program optbug

! shift values in array dh every second call
 
subroutine sub(rinp,rout)
   implicit none
   integer, parameter :: dim = 2
   integer :: i, iframe, rinp, rout, dh(dim)
!  save iframe,dh

!  first call: set default values in array dh
   if(rinp.eq.1) then
      dh(1:dim)=(/ (i,i=1,dim) /)
      iframe=0
   endif
   if (iframe .eq. 0) then
      iframe = 1
      do i = 1,dim-1
         dh(i) = dh(i+1)
      end do
      dh(dim) = rinp
   else
      iframe = 0
   end if

   rout = dh(dim)
   return
end subroutine sub
[/fortran]

With the current issues (11.1.073) of the IA-32 and Intel64 compilers, with no options specified in ifort.cfg, here are the results: Note that the -save option is being used in lieu of the SAVE specification in the subroutine. The program assumes that the first argument to the subroutine will be equal to 1 at the first call.

[bash]$ ifort -save -O0 bb.f90 
$ ./a.out
   1     1
   2     1

$ ifort -save -O1 bb.f90 
$ ./a.out
 1 1
 2 2

[/bash]

Kevin_D_Intel · ‎09-21-2010

Thank you very much, mecej4.It is helpful to have check bounds eliminated.This variant is also fixed in the upcoming major release. I will include this in the report to Development and keep this thread updated along with the earlier one.