On Jul 20, 6:28 pm, Paul Rubin <http://[EMAIL PROTECTED]> wrote:
> As the Intel FDIV bug > incident reminds us, even billions of test inputs are not enough to > show that the routine does the right thing for EVERY input. When I remember correctly the FDIV bug was due to a wrongly filled lookup table and occurred only for certain bitpattern in the divisor. I'm not sure how a formal proof on the structure of the algorithm could help here? Intel repaired the table i.e. the data not the algorithm that acted upon it. -- http://mail.python.org/mailman/listinfo/python-list