Re: 387 numerical instability

[Neil Ferguson <nferguso at eso dot org>]

I'd just like to say a belated "thanks" for your response - reading the PR, 
the attendant discussion and the links therein got me where I needed to be.

If anyone else finds themselves dealing with things like this, the following 
links may be of interest:

http://grouper.ieee.org/groups/754/
http://www.validlab.com/goldberg/paper.ps (found on the IEEE page)
http://www.cs.berkeley.edu/~wkahan/Mindless.pdf

Neil.

Ian Lance Taylor wrote:
> Neil Ferguson <nferguso@eso.org> writes:
>
> > Whilst reading the GCC man page documentation for the -mfpmath option,
> > I came across this sentence for -mfpmath=sse:
> >
> > > The resulting code should be considerably faster in the major-
> > > ity of cases and avoid the numerical instability problems of
> > > 387 code, but may break some existing code that expects tempo-
> > > raries to be 80bit.
> >
> > Can someone point me to a description of these "numerical instability
> > problems of 387 code", please?
> >
> > I've searched the GCC site & mailing lists, and I've searched with
> > Google, but I can't find anything that details the particular concerns
> > that led to this note in the manual. All I can find is a list message
> > containing the patch that put it into the man page.
>
> See, e.g.,
>     http://gcc.gnu.org/PR323
> and the many PRs which have been marked as a duplicate of that one.
>
> Ian