test fails on HP

[David Relson]

At 01:55 PM 1/10/03, Timothy VanFosson wrote:

>At 06:35 PM 1/9/2003 -0500,  David Relson wrote:
>
>>This is the right place for your question.  Does your HP machine have a 
>>risc processor?  We've had some problems with HP risc processors, but I 
>>think they're all fixed in the current cvs repository.  Of course you may 
>>have a problem we haven't yet seen and fixed.
>>
>>Can you download the current cvs sources from SourceForge?  If so, be 
>>sure to get the "mime" branch.  It's the most up-to-date code as of the 
>>moment.  If you don't have the necessary access, let me know and I'll 
>>generate a current snapshot for you.
>>
>>Anyhow, once you have it, run "make check" and let us know whether your 
>>problem is fixed or not.
>
>Yes.  The HP has as PA-RISC 2.0 processor.  Using the latest code from the 
>repository did not make the problem go away, although it is 
>different.  Here are the diffs from test 2b.g.vvv using the latest code:
>
>renoir [535] % diff -u msg.2b.g.vvv ../../t.systest.d/outputs/msg.2b.g.vvv
>--- msg.2b.g.vvv        Fri Jan 10 12:17:26 2003
>+++ ../../t.systest.d/outputs/msg.2b.g.vvv      Sun Dec  8 12:36:13 2002
>@@ -1,7 +1,7 @@
>  X-Bogosity: Yes, tests=bogofilter, spamicity=1.000000
>-       0.010000  0.010000  his
>-       0.010000  0.000102  let
>-       0.010000  0.000001  without
>+       0.010000  0.010000  call
>+       0.010000  0.000102  effort
>+       0.010000  0.000001  having
>         0.990000  0.000102  abuse
>         0.990000  0.010000  bordercolor
>         0.990000  0.500001  car
>
>Any ideas on where to start looking?
>
>tv

Timothy,

Looking at name of the file in question is msg.2b.g.vv, the letter "g" 
tells me that it's the Graham algorithm being tested.  The Graham algorithm 
uses an array of 15 extrema, i.e. tokens with scores furthest from 
EVEN_ODDS, i.e. 0.500000.  Seeing that the words used on your machine have 
the same scores, i.e. 0.010000, as the reference results tells me that the 
scores are being calculated correctly.  Seeing that your machine has chosen 
different words indicates that the issue is with sorting and/or comparing 
probabilities.  In turn, that indicates floating point issues, specifically 
differences between values in registers and in ram.  I thought that those 
had all been fixed, but looking in graham.c (where the Graham code 
resides), I quickly spotted some floating ops that I'd swear had been 
fixed.  So, I edited the code and generated the attached patch.  Give it a 
try and let me know the result.

David
-------------- next part --------------
A non-text attachment was scrubbed...
Name: graham.c.patch
Type: application/octet-stream
Size: 584 bytes
Desc: not available
URL: <https://www.bogofilter.org/pipermail/bogofilter-dev/attachments/20030110/5c489f69/attachment.obj>