Re: Expression evaluation [was Re: I vote NO ...]
On May 18 2011, John Pryce wrote:
I am sorry, but that's not going to work. The reasons were raised
(mainly by outsiders) in the discussion of IEEE 754, but were swept
under the carpet.
What are you telling us? - That Motion 24 is useless? - That anything we
put in an interval standard to influence language standards is useless?
No, my point was NOT about motion 24, which is plausible, but about your
remark that 1788 is following 754's mistaken path to impose clause 10.
What I am saying is, roughly:
If you take the approach that established language standards will
have to throw away their existing semantic models and replace them
by (essentially) the one assumed by 754, you aren't going to get
anywhere.
Those models have often been proven to work - and allow fairly serious
optimisation - over 4-5 decades and dozens of radically different
architectures. Replacing them by one that is known to forbid many
(perhaps most) of the existing important optimisations just isn't going
to fly. 754 has been the de facto standard for a couple of decades,
and I don't know of ANY language that supports it in full, or even
supports most of it at all well. And the semantic problems are a major
part of the reason why not.
1788 has the opportunity to avoid many of the mistakes of 754 in this
area, but probably isn't going to be even as successful if it doesn't.
And one serious mistake that is well worth avoiding is to make claims
about languages that are false in fundamental ways. If that clause
really meant "shall specify" rather than "specifies" then, of course,
it is not false - merely unacceptable.
Or are you reminding us that, whatever good ideas we put in an interval
standard, there is no alternative to the hard graft of joining a language
standards group and working to influence it from the inside?
Not really. I am saying that there is a limit to how much you can demand
that they change, especially when they have very good reason not to do so.
And I am saying that demanding that they change to a 1-1 mapping between
their syntax and their evaluation is well beyond that limit, especially
for Fortran.
The real problem is that it will be much easier to get interval arithmetic
adopted as an optional extra, probably library-based, but it won't then
be adopted by the users who need it most. I don't have a good answer to
that, but following a path that has already been seen to not get anywhere
is not a good idea.
Regards,
Nick.