Date: Mon, 4 Mar 1996 15:55:19 -0500 (EST)
From: Stephen L Moshier <moshier AT world DOT std DOT com>
Sender: Stephen L Moshier <moshier AT world DOT std DOT com>
Reply-To: Stephen L Moshier <moshier AT world DOT std DOT com>
Subject: Re: [moshier AT world DOT std DOT com: Re: math test suite] 
To: "J.T. Conklin" <jtc AT cygnus DOT com>
Cc: Roland McGrath <roland AT gnu DOT ai DOT mit DOT edu>, djgpp AT delorie DOT com
In-Reply-To: <199603042035.MAA26158@cygnus.com>
Message-Id: <Pine.3.89.9603041547.A11802-0100000@world.std.com>
Mime-Version: 1.0
Content-Type: TEXT/PLAIN; CHARSET=US-ASCII



On Mon, 4 Mar 1996, J.T. Conklin wrote:

> > Regarding a version modified by JT Conklin, all I know is that djgpp-msdos
> > has gone to something with fdlibm wrappers and some assembler code
> > bearing the name JT Conklin.  There is a hue and cry from the news groupies
> > about mysterious failures and crashes.  I have not investigated deeply
> > because it takes quite a long time to do a thorough job of testing.
> 
> Oh, I did that a long time ago.  At the time there didn't seem to be
> much interest in adopting it, so I stopped tracking it.  
> 
> There has been one bug found and fixed one of i386 assembly code
> routines since I made that first distribution.  I wish I had been
> contacted when they decided to go with it.
> 
> There is an unfortunate tendency within the free software community to
> integrate a piece of software without establishing correspondence with
> its maintainer(s).  In my opinion, this fragmentation diminishes the
> leverage free software has over proprietary alternatives. 
> 
> I was very happy to see Roland contact me instead of just integrating
> the NetBSD libm into glibc.  I hope that once it is integrated, the
> communication we've established will continue.
> 
> 	--jtc


There is one bug in the djgpp distribution that I have definitiely
identified:  the coprocessor is set to trap on divide by zero, instead
of quietly returning infinity.  Discovering this made me feel very uneasy, 
because the original fdlibm requires a strictly IEEE environment.
I hope you will be able to check out the code they are actually using
and help them ensure that it is working properly.