Re: Discussion paper: what are the level 2 datums?
Dan Zuras already expressed some of my concerns rather well,
so I can be brief here.
The notion that "untagged x (was BFP) and z (was DFP) cannot
be tested for equality really startled me. It is true that
mixed-radix expressions cause trouble in FP arithmetic,
and mixed-radix true comparisons, although easy to define
precisely, are devilishly difficult to implement well --
but one of the nice properties of Interval Arithmetic is
that mixed-format and even mixed-radix expressions can
generally be defined nicely, exploiting the hull concept,
so there is no need to restrict them. (Users would however
have to be aware of possible performance implications.)
Btw, the motion to support mixed-radix comparisons in an
early draft of P754-R was to specify the behaviour IF they
were provided, not to require or even recommend that they
be provided.
I was also taken aback by the notion that there was only ONE
NaI -- but I see that the analogy with 754 NaN works out fine.
My reaction was that there are (conceptually) so many NaIs
that they are ALL different -- even the same one used in two
places, which explains why they compare unequal to themselves.
Finally, I disagree with the intent to drop bare decorations.
Those are useful when there is no point in continuing to compute
with the interval portion. For ill-formed it probably makes
little difference because the same test is (almost always) needed
before every operation -- but if some of the other decorations
(unbounded, possibly undefined etc.) are deemed to pollute all
derived results anyway, it would be nice if the rest of the
computation could be skipped entirely, just propagating the
decoration.
Michel.
---Sent: 2010-10-06 01:35:51 UTC