bnf bug

[Igor Schein <igor@txc.com>]

Hi, 

I've discovered a bizarre bug introduced in 2.0.12.alpha

Here's the session (2.0.13.alpha is used):

% gp -q -f -s 32000000
? f(t)=bnfclgp(x^3-t*x^2-(t+3)*x-1)
? for(k=511,521,f(k))
? #
   timer = 1 (on)
? a=f(522)
time = 1,510 ms.
[468, [78, 6], [[127, 59, 75; 0, 1, 0; 0, 0, 1], [23, 16, 20; 0, 1, 0;
0, 0, 1]]]
? b=f(522);
time = 3mn, 17,330 ms.
? a==b
time = 0 ms.
0

Note that f(522) produces 2 different outputs.  Which one is correct?
My guess is the second one is correct.

The timing was measured on Ultra-60,  so it should take at least twice as
long on a regular PC with Linux.

Thanks
Igor