[Cfp-interest 3205] Re: Various Updates to Annex F+G

[Damian McGuckin]

On Thu, 15 Aug 2024, Damian McGuckin wrote:

> Nowhere is the concept of normative text defined or discussed.
>
> With regard to F.10.1#14
>
> 	 Functions with a NaN argument return a NaN result and raise no
> 	 floating-point exception, except where explicitly stated otherwise.
>
> I would dearly love to add the word 'single' before that NaN.

I personal would normally expect that the normal behaviour if given a
NaN as one of several argument, one may or may not get a NaN returned.

Indeed, hypot() does not always do that.

In some of my own programs which model physics, depending on the other 
parameters, I may take branches which never need to access that argument 
which is a NaN.

So, in my own limited cases, I would only assume normative behaviour for
a case like

 	x = f(NaN)

would yield a result which is x.

The standard assumes normative behaviour of F.10.1#14 for

 	frexp(NaN, p)

where 'p' is a pointer to an integer.  I find that very abnormal 
personally but that is a highly subjective opinion.  I have used
'hypot' probably 100 times more than I have ever used 'frexp'.

Thanks - Damian