LI74482: INCORRECT OUTPUT WHEN USING -QNOSAVE/-QFIXED

April 2009 Update for XL Fortran Advanced Edition for Blue Gene/L, V11.1
September 2010 Update for XL Fortran Advanced Edition for Blue Gene/L, V11.1

APAR status

• Closed as program error.

Error description

• When the following test case is compiled with the following
  compile line, an incorrect output is obtained at execution time.
  
  Compile Line:
  mpxlf_r -I/usr/local/include -O4 -qnosave -qfixed mpitest.f
  
  ===== TESTCASE:
  <<<<<<<<<<<<< Start of mpitest.f >>>>>>>>>>>>>>>>>>>>>>>>
        PROGRAM testmain
  
        ! This program shows an MPI bug on jump,
        ! if compiled with "mpxlf_r -O4 -qnosave" and
        ! executed on one processor.
  
        IMPLICIT NONE
  
        INCLUDE 'mpif.h'
        INTEGER :: ierr(3),stt(MPI_STATUS_SIZE)
  
        CALL MPI_INIT(ierr)
  
        CALL testmpi(1)
  
        CALL MPI_FINALIZE(ierr)
  
        END PROGRAM
  
  !===============================================================
  
        SUBROUTINE testmpi(n)
  
        IMPLICIT NONE
  
        ! The bug does not appear, if the dimension of y
        ! is specified within this subroutine (i.e. "y(1)").
        INTEGER, INTENT (IN) :: n
        INTEGER :: j, x, y(n)
  !
        INCLUDE 'mpif.h'
        INTEGER :: ierr(3),irank,isize
  
        DO j= 1,n
          y(j)= 1
        ENDDO
  
        ! At first, x is set to 3. Then it is changed to 4.
        x = 3
        DO j= 1,n
          x = max( x, 4 )
        END DO
  
        CALL MPI_COMM_RANK (MPI_COMM_WORLD,irank,ierr)
        CALL MPI_COMM_SIZE (MPI_COMM_WORLD,isize,ierr)
  
        ! Here, x is 4.
        WRITE(*,fmt='(2(A,i2),A)')
       & 'irank =',irank,':  x =',x,'  before MPI_BCAST'
  
        ! The bug appears, if the following IF-block is not
        ! executed.
        ! It does not apper, if MPI_BCAST is uncommented
        IF ( isize > 1 ) THEN
          WRITE(*,fmt='(A,i2,A)') 'irank =',irank,':  isize > 1'
          CALL MPI_BCAST(x,1,MPI_INTEGER,0,MPI_COMM_WORLD,ierr)
        ENDIF
  
        ! Now, x has changed from 4 back to 3.
        WRITE(*,fmt='(2(A,i2),A)')
       & 'irank =',irank,':  x =',x,'  after MPI_BCAST'
  
        END SUBROUTINE testmpi
  
  
  <<<<<<<<<<<<<<<< end of mpitest.f >>>>>>>>>>>>>>>>>>>>>>
  
  ===== ACTUAL OUTPUT:
  $ ./a.out
  irank = 0:  x = 4  before MPI_BCAST
  irank = 0:  x = 3  after MPI_BCAST
  
  ===== EXPECTED OUTPUT:
  $ ./a.out
  irank = 0:  x = 4  before MPI_BCAST
  irank = 0:  x = 4  after MPI_BCAST
  

Local fix

• A workaround to this problem is statically allocate y in the
  subroutine.
  

Problem summary

• This specific problem was found in MPI code,
  but it has the potential to effect any code
  containing loops compiled with loop opts enabled.
  
  The compiler was performing an address-taken
  optimzation and placing statements between the
  latch directive and the latch branch of the loop. This has
  been corrected.
  

Problem conclusion

• When searching for the location to place the
  statement, the compiler checks whether it is in the
  latch block of a region. If it is, it walks up past
  the latch directive to ensure the statement is
  inserted in the loop body.
  

Temporary fix

Comments

APAR Information

• APAR is sysrouted FROM one or more of the following:

  IZ38370

• APAR is sysrouted TO one or more of the following:

Fix information

• Fixed component name

  XL FORTRAN AE B

• Fixed component ID

  5799HJG00

Applicable component levels

• RB10 PSN IZ38370

     UP06/09/13