Newsgroups: comp.lang.c
Path: utzoo!henry
From: henry@utzoo.uucp (Henry Spencer)
Subject: Re: ambiguous ?
Message-ID: <1989Oct22.015131.25642@utzoo.uucp>
Organization: U of Toronto Zoology
References: <1989Oct20.175352.20598@utzoo.uucp> <14102@lanl.gov> <1989Oct21.072905.9039@utzoo.uucp> <11371@smoke.BRL.MIL>
Date: Sun, 22 Oct 89 01:51:31 GMT

In article <11371@smoke.BRL.MIL> gwyn@brl.arpa (Doug Gwyn) writes:
>>Most Fortran programmers, of course, either don't bother at all [to prove
>>their use of hardware floating-point is correct] ...
>>Avoiding this nasty compromise requires doing all math symbolically, using
>>complex and difficult exact representations, or at the very least using
>>a very carefully-designed interval-arithmetic package.  How one does any
>>of these things in Fortran is beyond me.
>
>To be fair, while I agree that most Fortran programmers don't do this
>properly, the main emphasis of Fortran is numerical programming, and
>there has been a lot of work put into resolving these problems.  The
>whole issue is a major branch of the field of numerical analysis.
>Some popular Fortran libraries are carefully designed in this regard.

I think you've very slightly missed my point, Doug.  What you're saying
is that great efforts have gone into engineering to *cope* with this
compromise, and make it a bit more manageable.  I agree.  But this sort
of thing -- coping intelligently with a compromise that cannot be avoided
in a cost-effective way -- is precisely what Jim was unwilling to accept.
If you insist on *eliminating* the non-ideal behavior, then the numerical
analysis work is irrelevant and more drastic measures are needed.  Then
Fortran falls down badly:  what you need is something like C++, where the
implementation of the arithmetic can be changed.
-- 
A bit of tolerance is worth a  |     Henry Spencer at U of Toronto Zoology
megabyte of flaming.           | uunet!attcan!utzoo!henry henry@zoo.toronto.edu