Re: [Bug-gnucap] insufficient limits checking in tanh model

[mcmahill]

On Sat, 5 Apr 2003, Al Davis wrote:

> On Saturday 05 April 2003 02:10 pm, address@hidden wrote:
> > I've been looking at gnucap-0.33 a little and it seems that
> > the bm_tanh.cc file has a potential problem on some systems. 
> > In this function:
> >
> > void EVAL_BM_TANH::tr_eval(ELEMENT* d)const
> > {
> >   double x = ioffset(d->_y0.x);
> >   double aa = x * _gain/_limit;
> >   double cosine = cosh(aa);
> >   double f1 = _gain / (cosine*cosine);
> >   double f0 = _limit * tanh(aa);
> >   d->_y0 = FPOLY1(x, f0, f1);
> >   tr_final_adjust(&(d->_y0), d->f_is_value());
> > }
> >
> > its possible for 'aa' to be large enough to overflow cosh and
> > product 'inf' as the result.  Then you're relying on _gain /
> > (inf * inf) to give 0.  In particular I've noticed that on an
> > alpha system (alpha processors do not implement full IEEE 754
> > math in hardware) I get floating point exceptions at that
> > line of the code.  If I compile with the -mieee flag to gcc,
> > traps are generated in the code to allow the software
> > floating point completion code to kick in and the program
> > runs without generating the SIGFPE's that I was getting
> > before.  However, not all systems implement this sort of
> > floating point math so you may find problems on some systems.
> 
> Can you post a circuit that demonstrates the problem?

actually its bm_tanh.1.ckt from the gnucap testsuite :)


> It doesn't happen on any system I have access to, but I 
> understand why it will on some others.
> 
> What does it do?
> 
> I expect it will print "floating point error" and continue, 
> eventually giving you the correct answer.  Does the command 
> "option noerror" suppress the message?  If that is what you 
> get, and you can suppress it, that means the sigfpe trap is 
> working correctly.

the sigfpe trap is working correctly.  I did see the signal handler
output.  It does seem to give the correct final answer.


> Aside from that ..
> 
> There should be a trap there.  Not all systems have sech, so I 
> can't eliminate the problem that way.  Besides, it might be 
> implemented as 1/cosh.  Also, tanh might be implemented as 
> sinh/cosh, making it also prone to the problem.  All of these 
> could be implemented using exp(x) and exp(-x), which is also 
> prone to the problem.
> 
> I think the only reliable way to handle it is to trap large 
> "aa". 
> 
> Try this:
> 

seems to work.

-Dan