FW: numerical results differ after irrelevant code change

[Michael D. Berger]

> -----Original Message-----
> From: Robert Dewar [mailto:dewar@adacore.com] 
> Sent: Sunday, May 08, 2011 13:02
> To: Michael D. Berger
> Cc: gcc@gcc.gnu.org
> Subject: Re: numerical results differ after irrelevant code change
> 
> On 5/8/2011 12:48 PM, Michael D. Berger wrote:
> 
> > I made the changes you suggest.  While I was previously getting
> > -1.16e-16 and -1.03e-16 depending presence of an "extra"
> > class member, I now get 1.11e-16 (sic not -).  But it is now 
> > independent of the extra class member, which is excellent.  
> Hopefully, 
> > things will stay that way.
> 
> This is a fairly big change to make in your program, since it 
> completely alters the fpt semantics, but probably in a 
> desirable way. Sounds like you don't have any real careful 
> analysis of the reliability of the results or needed 
> precision anyway, you just don't like the non-repeatability, 
> and indeed the change to restrict the precision to 64-bits 
> should work to solve that problem.
> 

As I should have said originally, I just did the
-march=pentium4 -mfpmath=sse, I didn't change
to a 64 bit system.  Will that get the repeatabiility?
It has on the brief test I did.

Thanks,
Mike.