From mboxrd@z Thu Jan  1 00:00:00 1970
Received: (from weis@localhost) by pauillac.inria.fr (8.7.6/8.7.3) id QAA23875 for caml-red; Fri, 13 Oct 2000 16:05:24 +0200 (MET DST)
Received: from concorde.inria.fr (concorde.inria.fr [192.93.2.39]) by pauillac.inria.fr (8.7.6/8.7.3) with ESMTP id PAA23164 for <weis@pauillac.inria.fr>; Fri, 13 Oct 2000 15:54:20 +0200 (MET DST)
Received: from mrwall.kal.com (mrwall.kal.com [194.193.14.236])
	by concorde.inria.fr (8.10.0/8.10.0) with SMTP id e9DDsJD15904
	for <Pierre.Weis@inria.fr>; Fri, 13 Oct 2000 15:54:20 +0200 (MET DST)
Received: from mrwall.kal.com [194.193.14.236]
	(HELO localhost)
	by mrwall.kal.com (AltaVista Mail V2.0J/2.0J BL25J listener)
	id 0000_0050_39e7_1438_6890;
	Fri, 13 Oct 2000 14:55:04 +0100
Received: from somewhere by smtpxd
Message-ID: <3145774E67D8D111BE6E00C0DF418B6631EE94@nt.kal.com>
From: Dave Berry <dave@kal.com>
To: Pierre Weis <Pierre.Weis@inria.fr>
Subject: RE: Undefined evaluation order
Date: Fri, 13 Oct 2000 14:56:50 +0100
MIME-Version: 1.0
X-Mailer: Internet Mail Service (5.0.1460.8)
Content-Type: text/plain;
	charset="iso-8859-1"
Sender: weis@pauillac.inria.fr

David,

Thanks for your reply.  Are you saying that if (a * b) would result in a NaN
then you always want (a * b * 0.0) to return a NaN, so that you are aware of
the potential problem? 

Dave.

> -----Original Message-----
> From: David McClain [mailto:dmcclain@azstarnet.com]
> Sent: Thursday, October 12, 2000 6:06 PM
> To: caml-list@inria.fr
> Subject: Re: Undefined evaluation order
> 
> 
> ... but the same would be true the other way too... (0.0 * a * b).  I am a
> numeric programmer and these things are unavoidable no matter how you
choose
> to order the evaluations.
> 
> If (a * b) raises a NaN then what would be the value of 0.0 times that?
The
> IEEE spec would say the result would have to continue to be a NaN.
> 
> I normally perform all arithmetic with exception processing supressed or
> deferred. The only time an exception is useful to me is if there is some
> remedial action that could be taken. I want my NaN's and INF's to appear
in
> my answers. In particular, if something should go awry at one point out of
> millions I don't want that one point to hose my entire calculation. (Note
> that I do not look kindly at the Fortran way of aborting an entire program
> for one bad point...) In signal and image processing, especially in a
> real-time environment, you just drop the bad points on the floor and
> continue running.
> 
> - DM