[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]

Floating Point Error

Some weeks ago I reported a bug which has annoyed me for some time
the source of which I believe I have now tracked down.  The bug is 

? (/ 0.0)
? (

Instead of 

? (/ 0.0)
> Error: Can't divide by zero.
> While executing: /
> Type Command-. to abort.
See the RestartsI menu item for further choices.
1 > 

Mainly this causes formatting mistakes, but it can lead to some
potentially serious numerical problems.   For example

In a standard regression test known as the Longley data set  (which
is a tricky test set because the data is highly correlated) I get the 
following differences in the Sum-of-Squares values depending on
whether the floating point bug that sometimes crops up has or not

1) when float bug not does not occur (/ 0.0) => > Error: Can't divide 
by zero.
(68445976650 834951.2065429687 6.844514169879346E+10  68260967824
               | here is a problem it the second Sum-of-Squares
2) when not working properly  (/ 0.0) => -1.7976931348623157E+308
(68445976650 832927.5610351562 6.8445143722438965E+10 68260967824

It appears that the floating-point bug occurs when I have Dove-Fax on 
at Init time, which causes its init to find the serial port  and test 
the modem out...  This occurs using Patch 2 on a clean copy of MCL. 
I have the modem and Imagewriter Printer hooked up to separate power  
switches so they are not always on at start up.  When the Dove Fax
is, the bug occurs as soon as I launch MCL, otherwise it does not. It 
does not occur when the Imagewriter alone is switched on. It also
does not happen if the fax is on, but the init is removed  from the
extension folder.  The dove fax version is 2.0.3.

Can anyone else reproduce this?