topic floating-point exception in Intel® Fortran Compiler

floating-point exception

beam2000 — Sun, 06 Feb 2011 12:35:19 GMT

Created a new project on the basis of an example on usage of procedure CLEARSTATUSFPQQ from Intel Fortran Compiler User and Reference Guides:

! Program to demonstrate CLEARSTATUSFPQQ.

! This program uses polling to detect that a

! floating-point exception has occurred.

! So, build this console application with the default

! floating-point exception behavior, fpe3.

! You need to specify compiler option /debug or /Od (Windows)

! or -O0 (Linux) to get the correct results

! PROGRAM CLEARFP

USE IFPORT

REAL*4 A,B,C

INTEGER*2 STS

A = 2.0E0

B = 0.0E0

! Poll and display initial floating point status

CALL GETSTATUSFPQQ(STS)

WRITE(*,'(1X,A,Z4.4)') 'Initial fp status = ',STS

! Cause a divide-by-zero exception

! Poll and display the new floating point status

C = A/B

CALL GETSTATUSFPQQ(STS)

WRITE(*,'(1X,A,Z4.4)') 'After div-by-zero fp status = ',STS

! If a divide by zero error occurred, clear the floating point

! status register so future exceptions can be detected.

IF ((STS .AND. FPSW$ZERODIVIDE) > 0) THEN

CALL CLEARSTATUSFPQQ()

CALL GETSTATUSFPQQ(STS)

WRITE(*,'(1X,A,Z4.4)') 'After CLEARSTATUSFPQQ fp status = ',STS

ENDIF

END

Installed compiler directive: /debug and /fpe:3

But as a result received on the screen following output:

Initial fp status = 0000

After div-by-zero fp status = 000

That is the program does not find out divide by zero?

In the Compaq Visual Fortran 6 program finds divide by zero!

floating-point exception

psantos — Sun, 06 Feb 2011 14:56:20 GMT

Hello beam2000,

I have tried your program and I have the same behaviour, but only when compiled under IA32. If I compile in x64, the "fp status" prints 4 (divided by zero). There is something wrong here, but I don't know what it could be... I'm using Composer XE update 2. What compiler are you using?

Pedro

floating-point exception

beam2000 — Mon, 07 Feb 2011 07:19:41 GMT

Hello Pedro!

I have Intel Visual Fortran 11.1.060 [IA-32].

Sergey

floating-point exception

Steven_L_Intel1 — Mon, 07 Feb 2011 15:42:24 GMT

There are a couple of different things that may cause this difference. You might want to try adding the /arch:ia32 option (in Visual Studio that is project property Fortran > Code Generation > Enable Enhanced Instruction Set > No enhanced instruction sets. The SSE instructions used by Intel Fortran behave differently regarding the FP flags compared to the "X87" instructions in earlier versions.

A second possibility is a bug in the way the run-time library handles FP exceptions - use Intel Visual Fortran Composer XE 2011 to resolve that.

However, I'd suggest instead that you use the Fortran 2003 "IEEE Floating point" intrinsic modules to accomplish what you want. Here's an example:

[fortran]    program zerodivide
    
    use, intrinsic :: IEEE_EXCEPTIONS
    use, intrinsic :: IEEE_ARITHMETIC

    implicit none


    real :: a, b, c
    logical :: zerodivide_flag
        
    ! Disable halting on a zerodivide
    
    call IEEE_SET_HALTING_MODE (IEEE_DIVIDE_BY_ZERO, .FALSE.)
    
    a = 2.0
    
    ! In a loop, read B from the console and divide it into A
     do
      read (*,*,END=99) b
      ! Clear the exception flags
      call IEEE_SET_FLAG (IEEE_USUAL, .FALSE.)
      c = a / b
      ! Did we get an exception?
      call IEEE_GET_FLAG (IEEE_DIVIDE_BY_ZERO, zerodivide_flag)
      if (zerodivide_flag) then
        write (*,*) "Divide by zero!"
      else
        write (*,*) "Result is ", c
      end if
    end do
    
99  continue
    
    end program zerodivide
[/fortran]

floating-point exception

beam2000 — Tue, 08 Feb 2011 09:10:41 GMT

Thank you very much, Steve!!!

Sergey

floating-point exception

beam2000 — Wed, 09 Feb 2011 17:59:34 GMT

Hi Steve!

If to compile the program presented by you in configuration Win32 Release with Optimization directive: Speed the acceptable result turns out. If thus to change 29 line on

write (*, *) "Result is ok"

the result will be unacceptable. Why?

Sergey

floating-point exception

Steven_L_Intel1 — Wed, 09 Feb 2011 18:21:16 GMT

Sergey, if I understand correctly, you're saying that if you change the WRITE to one that does not use the result of the division then you don't get an error. This is because the compiler sees that the division can be removed and therefore no zerodivide occurs.

floating-point exception

beam2000 — Wed, 09 Feb 2011 18:49:55 GMT

Steve, but why in configuration Win32 Releasewith Optimization directive: Disable zerodivide occurs?

floating-point exception

Steven_L_Intel1 — Wed, 09 Feb 2011 19:05:28 GMT

Because you disabled the optimizer so it didn't remove the unused calculation.