This is the mail archive of the
gcc-patches@gcc.gnu.org
mailing list for the GCC project.
Re: patch for estimate_probability instability
Geoff Keating <geoffk@geoffk.org> writes:
|> > Date: Thu, 15 Nov 2001 21:59:45 +0100
|> > From: Jan Hubicka <jh@suse.cz>
|> > Cc: Jan Hubicka <jh@suse.cz>, gcc-patches@gcc.gnu.org
|> > Content-Disposition: inline
|> > User-Agent: Mutt/1.3.20i
|> >
|> > > Jan Hubicka <jh@suse.cz> writes:
|> > >
|> > > > Hi,
|> > > > I've installed attached patch to cfg-brach to avoid estimate-branches
|> > > > instability and bootstrap misscomparisons.
|> > > >
|> > > > I am not quite sure if such a sollution can be acceptable for Mainline.
|> > >
|> > > I think this is just papering over some other problem. Floating-point
|> > > is not a random number generator, so given the same input it should
|> > > produce the same output. What is being compiled differently to cause
|> > > the comparison failure?
|> > i386 fp is ranom number generator. If you store value to memory it
|> > gets truncated from 80bits. When optimizing value is not stored, when
|> > not optimizing it is stored causing different roudoff error in two runs.
|>
|> Can we fix that?
|>
|> Perhaps GCC needs to limit FP precision to 'double' on x86 as we
|> recommend for other code.
Wouldn't that break long double?
Andreas.
--
Andreas Schwab "And now for something
Andreas.Schwab@suse.de completely different."
SuSE Labs, SuSE GmbH, Schanzäckerstr. 10, D-90443 Nürnberg
Key fingerprint = 58CA 54C7 6D53 942B 1756 01D3 44D5 214B 8276 4ED5