C******************************************************************************
C FILE: loop08.f
C DESCRIPTION:
C   Demonstrates loop defactorization
C******************************************************************************

	program loop08

	IMPLICIT NONE
	integer ARRAY_SIZE
	parameter(ARRAY_SIZE = 5000)
	integer	I, J
        real(8) A(ARRAY_SIZE), AA(ARRAY_SIZE)
        real(8) B(ARRAY_SIZE), BB(ARRAY_SIZE)
        real(8) C(ARRAY_SIZE), CC(ARRAY_SIZE)
        real(8) D(ARRAY_SIZE), DD(ARRAY_SIZE)
        integer(8) T1, T2, IRTC

C*****************************************************************************
C*				Factorized Loop				     *
C*****************************************************************************
        T1 = IRTC()

        DO I=1, ARRAY_SIZE
          A(I) = 0.0D0
          DO J = 1, ARRAY_SIZE
            A(I) = A(I) + B(J) * D(J) * C(I)
          ENDDO
        ENDDO

        T2 = IRTC()
        write(*,*)'Untuned loop took', (T2-T1)/1000000, 'msec'

C*****************************************************************************
C*				Unfactorized Loop			     *
C*****************************************************************************
        T1 = IRTC()

        DO I=1, ARRAY_SIZE
          AA(I) = 0.0D0
          DO J = 1, ARRAY_SIZE
            AA(I) = AA(I) + BB(J) * DD(J)
          ENDDO
          AA(I) = AA(I) * CC(I)
        ENDDO

        T2 = IRTC()
        write(*,*)'Tuned loop took', (T2-T1)/1000000, 'msec'

C       Need to actually use the results of the calculations or else the
C       optimizing compiler may just skip doing them...so we'll just
C       print them.  This will force the optimizer to actually do the work.
        print *, A(ARRAY_SIZE), AA(ARRAY_SIZE)

999   end