Date: Thu, 14 Aug 2003 11:27:49 +0200
From: "Eli Zaretskii" <eliz AT elta DOT co DOT il>
Sender: halo1 AT zahav DOT net DOT il
To: rich AT phekda DOT freeserve DOT co DOT uk
Message-Id: <7458-Thu14Aug2003112749+0300-eliz@elta.co.il>
X-Mailer: emacs 21.3.50 (via feedmail 8 I) and Blat ver 1.8.9
CC: djgpp-workers AT delorie DOT com
In-reply-to: <3F3B2D8E.F90E84B5@phekda.freeserve.co.uk> (message from Richard
	Dawe on Thu, 14 Aug 2003 07:34:54 +0100)
Subject: Re: Anomaly in printf()
References: <46 DOT 3c4de930 DOT 2c6a2e90 AT aol DOT com> <Pine DOT OSF DOT 4 DOT 51 DOT 0308131046310 DOT 2694 AT sirppi DOT helsinki DOT fi> <3F3B2D8E DOT F90E84B5 AT phekda DOT freeserve DOT co DOT uk>
Reply-To: djgpp-workers AT delorie DOT com
Errors-To: nobody AT delorie DOT com
X-Mailing-List: djgpp-workers AT delorie DOT com
X-Unsubscribes-To: listserv AT delorie DOT com
Precedence: bulk

> Date: Thu, 14 Aug 2003 07:34:54 +0100
> From: Richard Dawe <rich AT phekda DOT freeserve DOT co DOT uk>
> 
> Esa A E Peuha wrote:
> > >
> > > A mistake is in multiplying by 0.1 instead of dividing by 10 and taking
> > > the remainder.
> > 
> > I agree.  This should definitely divide by ten to produce as accurate
> > result as possible.  Multiplying by 0.1 will cause errors when the
> > number has large absolute value.  Could someone test the following patch
> > before I commit it?
> [snip]
> 
> FWIW the patch makes no difference to the result of the Cygnus test suite for
> me. If anyone's interested, I've put the results here:
> 
>     http://www.phekda.freeserve.co.uk/richdawe/djgpp/2.04/
> 
> And K. B. Williams's test program (the one that displays "Begin with Rounding
> Mode at Startup") shows the same results, when compiled with a patched libc.

As I'd expect, FWIW: the problem seems to be that the result of either
multiplying by 0.1 or dividing by 10 does not have an exact FP
representation.  That is, we are seeing inaccuracies in the last
significant digit which are to be expected.

I don't know how to fix this without a radical change in the
algorithm.

Btw, the code we have in _doprnt was originally taken from the BSD
libc, IIRC.  So if someone has access to FreeBSD, it'd be interesting
to see if they exibit the same problem with K.B.'s test program.