Xref: news2.mv.net comp.os.msdos.djgpp:7289
From: jhvogel AT me DOT umn DOT edu (Jeff Vogel)
Newsgroups: comp.os.msdos.djgpp
Subject: Re: math addition error?
Date: 14 Aug 1996 15:23:19 -0500
Organization: me.umn.edu -- UMN Mechanical Engineering
Lines: 48
Message-ID: <jhvogel.840052956@me.umn.edu>
References: <Pine DOT SUN DOT 3 DOT 91 DOT 960813093235 DOT 28748U-100000 AT is>
NNTP-Posting-Host: ena.me.umn.edu
To: djgpp AT delorie DOT com
DJ-Gateway: from newsgroup comp.os.msdos.djgpp

Eli Zaretskii <eliz AT is DOT elta DOT co DOT il> writes:


>On 12 Aug 1996, Jeff Vogel wrote:

>> #include <stdio.h>
>> main()
>> {
>> double x=3.6053320875, y=3.6053320870;
>> printf( "%lf\n", x-y );
>           ^^^
>What happens if you put "%f" there?  You shouldn't have to use "%lf" in 
>printf calls; it is only required for scanf.

>If that doesn't help, seems like an emulator problem to me, since it 
>works on a Pentium (that has a built-in FPU).


I have now reproduced the error on the Pentium, by running in native
DOS (it doesn't happen on the Pentium in a DOS box), and with go32 set as:

set GO#@=emu c:/djgpp/bin/emu387.exe

(path to my emulator - this is ver 1.12 of DJGPP).

I have upgraded my machine do djgpp v2, and found no change in the behavior.
I have verified that I am using the new go32-v2, emu387.dxe, and a new cwsdpmi
(I think that is the name) is on the path, though not explicitly loaded.

I have also confirmed that the error is not too hard to duplicate if the two
numbers are between 2.0-4.0, but the error seems to never appear outside of tha
range. Let x be a number between 2 and 4, and y=x+5.e-10. Then y-x=5.e-10 about
half of the time, and y-x=-3.9999999995 the other (slightly less than) half of
the time. Further, it is interesting that if x=2.0, and you then add a number
to x that is between about 4.65e-10 and 9.33e-10, the answer appears to be still
exactly 2.0; but subtracting x from this new number gives: (x+dx)-x = dx-4.0.
However, if dx is outiside of that range (larger or smaller) no error occurs.

This is probably more detail than anyone wants. However, it is very important t
me that I find out how to fix it. I understand that there may be limited
interest in debugging stuff that depends largely on not having a co-processor,
but if I can't fix it, I will have to abandon djgpp.

If a solution requires working on emulator sources, I would be willing to try
to help.
-- 
	Jeff Vogel
	jhvogel AT me DOT umn DOT edu