Failure at run-time with format F when using real KIND=3

The following program, compiled with FTN95 v. 9.03.0.0 Win32, raises an exception (Invalid floating point operation) at runtime for i=19. This error doesn't occur if KIND=2 is used for the real variables or F7.3 instead of F7.2 in the 9000 FORMAT. Am I missing an obvious pitfall in this case?

PROGRAM RUNGE
IMPLICIT NONE
INTEGER :: I, N 
REAL(KIND=3) :: A, B, T
A=0.0
B=5.0
N=1000
DO I=1,N
  T=(B-A)*FLOAT(I)/FLOAT(N)
  WRITE(*,9000) I,T
END DO
9000 FORMAT(1X,I4,3X,F7.2)
END PROGRAM RUNGE

There appears to be a problem with F7.2 for extended precision in this context but G7.2 seems to work OK.

This is a very strange error ! The following does not show the problem ?

  PROGRAM RUNGE
  IMPLICIT NONE
  INTEGER :: I, N
  REAL(KIND=3) :: A, B, T

  A=0.0
  B=5.0
  N=1001       !  35  !  1000
  DO I=1,N
    T=(B-A)*FLOAT(I)/FLOAT(N)
    WRITE(*,9000) I,T
  END DO
  9000 FORMAT(1X,I4,3X,F7.2)
  
  END PROGRAM RUNGE

N = 35 or N=1001 work ok, but why not N=1000 ?

This shows how unique this problem might be. The following variant shows there is something special about N = 1000 ; i = 19 ? Does anyone know the reason for the problem with this ?

  PROGRAM RUNGE
  IMPLICIT NONE
  INTEGER :: I, N
  REAL(KIND=3) :: A, B, T

  A=0.0
  B=5.0
  do n = 4,1004
!  N=1000       !  35  !  1000
  DO I=1,N
    T=(B-A)*FLOAT(I)/FLOAT(N)
    WRITE(*,9000) I,T
  END DO
  end do
  9000 FORMAT(1X,I4,3X,F7.2)
  
  END PROGRAM RUNGE

I am running using PLATO with Checkmate, Win32 [FTN95/x64 Ver. 9.03.0.0 Copyright (c) Silverfrost Ltd 1993-2024]

Even with Release, Win32 it crashes at i=19

The following change skips the error, do n = 994,1004 if ( n == 1000 ) cycle

??

Paul A similar error occurs with N=10000 or 100000 or 1000000 and G7.2 at I=199 John Thanks for confirming the bug. A similar error occurs with N=10000 and F7.2 at I=189 (not 199!). Neither the DO loop nor the reals A and B are needed, the reproducer can be reduced to the following:

PROGRAM RUNGE
IMPLICIT NONE
INTEGER :: I, N 
REAL(KIND=3) :: T
I=95
N=1000
T=FLOAT(I)/FLOAT(N)
WRITE(*,9000) I,T
9000 FORMAT(1X,I4,3X,F7.2)
END PROGRAM RUNGE

No error if you use dble(i)/dble(n).

However, the value displayed for the floating point value is still all ******'s.

The hexadecimal of this is interesting as a repeating pattern:

C28F5C28F5C28F5C

Bill I still get the same error with DBLE(I)/DBLE(N) using FTN95 version 9.03.0.0 Win32. Do you have an easy way to get the full hexadecimal 'physical' representation of a KIND=3 real (80-bit) number, since I can only use 64-bit integers (KIND=4) as targets with the TRANSFER method in FTN95?

This should work. This causes an error in both Checkmate and Release.

	PROGRAM RUNGE
	IMPLICIT NONE
	INTEGER :: I, N
        integer (KIND=4):: k=0
        integer(7):: my_loc,indx
	REAL (kind=3) :: T
	I=95
	N=1000
        t=0.0
	T=FLOAT(I)/FLOAT(N)
        my_loc = loc(t)
!        print *,my_loc ! decimal
        write(*,9002)my_loc,(core1(my_loc+indx-1),indx=1,10)
	WRITE(*,9000) I,T
9000	FORMAT(1X,I4,3X,F7.2)
9001	format(z20.20)
9002	format('Start Addr=',z8.8,': ',10z2.2)
	END PROGRAM RUNGE

FWIW: Fortran 9.02.00

Bill Thanks for the method of showing the physical representation of a 80-bit real number. I had all the knowledge to do it myself, as you kindly gave me a workaround for C_F_POINTER some time ago, but unfortunately I didn't recover the trick. Sorry for that and thanks again for your help.

No worries! That's what this forum is for, helping others!

This failure has now been fixed for the next release of salflibc.dll.

Paul Thank you for this.