[Cfp-interest 1804] Re: Exact subnormal results

[Damian McGuckin]

Hi David,

On Tue, 20 Oct 2020, David Hough CFP wrote:

> Two problems in 754 - since 1977 - have been that the underflow 
> exception can arise without raising any flag, and it can arise without 
> producing a subnormal or zero result.

> This first case, associated with exact subnormal result, is unlike any 
> other 754 exception. However with alternate exception handling one can 
> detect the first case.

Is C2X talking about alternate exception handling?

> The second case means that underflow can't be detected just by looking 
> for a subnormal or zero result. Worse, in binary, a minimum normalized 
> result handled by default might be accompanied by an underflow flag in 
> some implementations and not others.

Can you point me somewhere I can read a bit more on the above?

> This has always been problematic to map into languages in a way that 
> applies to non-754 implementations.
>
> I consider the definition of underflow to be a bad design choice in 754,

Yes - you have several types of underflow, just as you have several types 
of invalid.  Mind you, there was a whole lot of politics going on at the 
time prior to IEEE754-1987 with underflow and what was accepted was better 
than at least one of the alternatives.

And we cannot fix the definition of underflow at the language level. We 
need to wait until 2029 for that definition and the chances of fixing it 
then are slim to none. And that is out of scope for this group and C2X.

> but it can't be fixed while maintaining upward compatibility of 
> implementations.

That is beyond my skill level.

The question remains of how do we address the issues that Fred raised? I 
thought I was on the right track but now I realise that is not the case.

But I figure by raising some of the underlying questions, somebody else 
might have a stroke of genius.

Regards - Damian

Pacific Engineering Systems International, 277-279 Broadway, Glebe NSW 2037
Ph:+61-2-8571-0847 .. Fx:+61-2-9692-9623 | unsolicited email not wanted here
Views & opinions here are mine and not those of any past or present employer