Re: reproducibility of tininess detection for binary formats

[David Hough 754R work <754r@xxxxxxxxxxx>]

The conclusion I came to is that there are two possible approaches to
reproducibility of underflow and FMA:

EITHER 

say that the cases that differ among implementations are not 
reproducible - an observation that will affect almost nobody except
test program writers 

OR

prescribe the behaviors of these cases for both reproducible and 
non-reproducible mode, accepting that certain existing implementations
won't be able to conform at full speed



The compromise in draft 1.5 means that each language standardization 
committee will have to decide:   is reproducibility worth enough on all
implementations to warrant the extra expense, required on some implementations,
to get underflow and fma right in reproducibility mode only - 
but not in the normal execution mode.       I have to think that most
language committees running by consensus
will decide that reproducibility can't be worth that much.


However I do agree with draft 1.5 that there is no particular reason why
the underflow rule has to be the same on binary and decimal formats.
I don't think that both radices would go through the exact same 
rounding and exception detection paths.