Re: g77 problems

["Wes Bauske" <wsb paralleldata com>]

David Alan Gilbert wrote:
> 
> Wes Bauske wrote:
> 
> > Alpha's are VERY strict with respect to arguments of FP operations.
> > You cannot use un-normalized numbers! One normally gets those
> > by indexing in an array incorrectly or used an un-initialized
> > stack variable. If you run your code under gdb, and look at the
> > arguments to the FP operation, you most likely will find a
> > non-FP number. Now, if you REALLY intended to do this, you can
> > use the flag '-mieee', I believe, to have SW trap and handle these
> > sorts of problems. No guarantees that it will work but others
> > have reported success with it.
> 
> What happens with numbers which are so small/large that they can not
> be normalised (sorry I can't remember the term); as far as I understood
> they were valid numbers, but perhaps I'm confused.
> 

Underflows are set to zero according to the architecture
stuff I read. Don't know about overflows. I have an
algorithm that can underflow and is working correctly but
haven't tried it on an Alpha so I can't say it really works
from experience. You'd still have valid inputs to the FP
op though. Most problems I see in code are trying to feed
junk into a FP op.

If a number cannot be represented normalized, you're in
trouble because IEEE FP format has an implied 1 bit in the
mantissa. I use 8 byte floats if I need a larger range of
numbers. Most stuff I do works with 4 byte floats though.
(Digital signal processing)


Wes