CVS: strange test failures on FreeBSD - ideas sought

[David Relson]

On Tue, 23 Sep 2003 04:01:36 +0200
Matthias Andree <matthias.andree at gmx.de> wrote:

> Hi,
> 
> on FreeBSD 4.9, running
> 
>         env BF_TEST_VERBOSE=1 gmake check TESTS=t.maildir\ t.MH
> 
> from a current CVS tree yields:
> 
> gmake[1]: Entering directory
> `/usr/home/emma/bogofilter-build/src/tests'# 4039 words, 48 messages
> # 1784 words, 21 messages
> 1c1
> < good/cur/002 H 3.12916e-13
> ---
> > good/cur/002 H 3.12972e-13
> FAIL: t.maildir
> 1c1
> < good/002 H 3.12916e-13
> ---
> > good/002 H 3.12972e-13
> FAIL: t.MH
> ===================
> 2 of 2 tests failed
> ===================
> 
> The difference between those figures is 5.6e-17, which is
> approximately 1/4 * DBL_EPSILON (2.22045e-16). Should we just quench
> values below DBL_EPSILON to 0?
> 
> On SuSE Linux 8.2, everything's fine. I cannot currently test SPARC
> for lack of formail(1). I don't see the difference between those
> machines at the moment. GSL is not it, neither is GCC. I tried without
> GSL and with GCC 2.95 on Linux.  FreeBSD runs on a K6-2/300 processor,
> Linux on a Duron/700 -- either is IA32.
> 
> Any ideas as to what causes these mismatches? I was under the
> impression K6-2/300 and Duron/700 were compatible, and the idea that
> Linux and FreeBSD might behave differently in bogofilter is
> intimidating. If I'd used some obscure, non-IEEE chip, I'd have had
> rounding policies to blame, but in this case?
> 
> I must be missing something not quite obvious.

Matthias,

My first thought was "need more info".  Scoring the messages with "-vvv"
and comparing the linux and FreeBSD outputs would tell whether parsing
is the same (which is probable because the word/message counts are
correct) and whether scoring calculations are the same.  

My second thought was gsl vs. dcdflib.  Are the same libraries present
in both environments?

David