[J3] Question on overflow and run-time detection of infinity?

[Van Snyder]

On Fri, 2018-11-23 at 14:09 +0000, Clune, Thomas L. (GSFC-6101) via J3
wrote:

> My knowledge of the details of the IEEE specs is at best charitably
> described as marginal, esp. with regard to signaling.    And this
> question my well be in the quality of implementation realm rather than
> what the standard intends.   With those provisos …
> 
> 
> 
> I have encountered a situation where a denominator for an expression
> for a diagnostic can be 0 on semi-rare occasions.   In this case both
> numerator and denominator will be 0.  Both are otherwise non-negative,
> so +Infinity seemed a sensible value to return to the client code.
> So I inserted the following:
> 
> 
> 
>       USE, INTRINSIC :: IEEE_ARITHMETIC, only: IEEE_POSITIVE_INF,
> IEEE_VALUE
>       REAL(KIND=REAL64) :: relative_deviation
> 
> 
>       … 
>       REAL(KIND=REAL64) :: abs_mean
> 
> 
>       abs_mean = ABS(this%get_mean_cycle_time())
>       IF (abs_mean > 0) then ! non-zero denominator !
>          relative_deviation = 100*(this%
> get_standard_deviation()/abs_mean)
>       ELSE ! zero denominator
>          relative_deviation = IEEE_VALUE(1.0_REAL64,
> IEEE_POSITIVE_INF)
>       END IF

A similar question came up a very long time ago.  How does one assign a
signaling NaN using intrinsic assignment?

> No further calculations are ever done with relative_deviation except a
> subsequent formatted output.     Two compilers print “Infinity” as
> expected even when using aggressive runtime checking (e.g., with NAG I
> am using -C=all and -ieee=stop).   The 3rd compiler (gfortran 8.2.0)
> halts at the above assignment to IEEE_POSITIVE_INF with the useful,
> albeit undesired run-time message:
> 
> 
>         Program received signal SIGFPE: Floating-point exception -
>         erroneous arithmetic operation.
> 
> 
> 
> The flag that toggles is   -ffpe-trap=overflow.   I can turn that
> option off, but I would like the compile to helpfully identify other
> bad denominators (and similar things) elsewhere in the application.
> 
> 
> 
> 
> The question:   Is not the intent of the standard to allow using the
> IEEE functionality to return values such as Infinity without halting
> the run while still enabling run-time checking of calculations that
> incidentally compute erroneous operations?   
> 
> 
> Thanks,
> 
> 
> - Tom
> 
> 
> 
> 

-------------- next part --------------
An HTML attachment was scrubbed...
URL: <https://mailman.j3-fortran.org/pipermail/j3/attachments/20181125/9fd90908/attachment.html>