Re: Infinite loop with GCC 4.3

[Trevor Daniels]

Mats

While what you say is in general good practice, it doesn't solve the problem 
with the x86 FPU: the problem is not due to rounding errors, 
but -inconsistent- rounding.  This means that any comparison of the results 
of two floating point calculations, with or without epsilon, may return a 
different result when carried out at different points in the computation, or 
with differing optimizations, so (for example) breaking assert statements or 
generating different results after optimization.

Trevor

----- Original Message ----- 
From: "Mats Bengtsson" <address@hidden>
To: "Joe Neeman" <address@hidden>
Cc: "lily-devel" <address@hidden>
Sent: Monday, May 26, 2008 9:09 AM
Subject: Re: Infinite loop with GCC 4.3


> As is taught in any course on numerical methods, you should always be very 
> careful with
> comparing numerical values. For example you should use tests of the form
> if (abs(x-y) < epsilon)
> instead of
> if (x==y)
> where epsilon is a fixed value somewhat larger than machine precisions.
>
> Following the same line of thought, you should use
> if (x >= y - epsilon)
> instead of
> if (x >= y )
> (though this is probably not mentioned as often in the course text books).
> and analogously it follows that you numerically cannot distinguish between
> if (x > y)
> and
> if (x >= y)
>
> Here you show an example where you cannot rely on
> if (x > y)
> but have to do something more or less equivalent to
> if ( x > y - epsilon)
> to guarantee numerical stability.
>
> I haven't read the corresponding thread in the GCC bug tracker, but I'm 
> sure that these
> aspects have been discussed in length there as well.
>
>    /Mats
>
> Joe Neeman wrote:
> > On Sun, 2008-05-25 at 13:41 +0200, John Mandereau wrote:
> >
> > > On 2008/05/23, Joe Neeman wrote:
> > >
> > > > Does re-building without optimisations help?
> > > Yes, it does!  Building with GCC 4.3.0 and without -O2 flag makes
> > > LilyPond produce the same output [*] as with GCC 4.1.2 with
> > > optimisations, but it is also (not very surpisingly) much slower, so
> > > I'll go on building Lily with GCC 4.1.2, unless Fedora 9 provides an
> > > upgrade to a newer and fixed version.
> >
> > Attached is the smallest example that I could find that reproduces the
> > problem. With g++ 4.3.0 on Fedora 9, I get an infinite loop with -O1 but
> > immediate termination with -O0. With g++ 4.2.3 on Debian, I get
> > termination regardless of the optimisation setting.
> >
> > I think it is also easy to argue that the program should terminate,
> > regardless of rounding (the value of 'i' should be the same every time
> > the loop executes. So on the second time through, 'x' should be at least
> > as large as 'i').
> >
> > Can you reproduce? If so, I'll go bug the gcc people.
> >
> > Joe
> >
> >   ------------------------------------------------------------------------
> >
> > _______________________________________________
> > lilypond-devel mailing list
> > address@hidden
> > http://lists.gnu.org/mailman/listinfo/lilypond-devel
>
> --
> =============================================
> Mats Bengtsson
> Signal Processing
> School of Electrical Engineering
> Royal Institute of Technology (KTH)
> SE-100 44  STOCKHOLM
> Sweden
> Phone: (+46) 8 790 8463 Fax:   (+46) 8 790 7260
> Email: address@hidden
> WWW: http://www.s3.kth.se/~mabe
> =============================================
>
>
>
> _______________________________________________
> lilypond-devel mailing list
> address@hidden
> http://lists.gnu.org/mailman/listinfo/lilypond-devel