This is the mail archive of the gcc-patches@gcc.gnu.org mailing list for the GCC project.


Index Nav: [Date Index] [Subject Index] [Author Index] [Thread Index]
Message Nav: [Date Prev] [Date Next] [Thread Prev] [Thread Next]
Other format: [Raw text]

Re: Remove NOTE_INSN_PREDICTION


> On Wednesday 21 July 2004 19:09, Jan Hubicka wrote:
> > > Hi,
> > >
> > > NOTE_INSN_PREDICTION was only still used for return predictions,
> > > which are broken since the patch to unify return statements for
> > > lower GIMPLE.  So we spend a lot of time (computing DOM and PDOM)
> > > in predict.c only to predict a single return with a very poor and
> > > broken prediction.
> > >
> > > With this patch, we no longer predict returns like this.  I will
> > > reimplement this in the tree branch prediction code instead.
> > >
> > > Bootstrapped on i686-suse-linux-gnu, testing ongoing.  OK when it
> > > passes?
> >
> > Yes, this is OK.
> 
> Thanks.
> 
> 
> >  I intended to propose this once we do have replacement
> > for this logic (via branch prediction on trees), but there is no need
> > to keep broken feature around anyway.  Tree profiling already replace
> > most of the logic, I just need to find time to do the merge.
> 
> You already merged the return predictors.  They're also broken
> because we only have one RETURN_EXPR on GIMPLE.  But at least
> we can fix those, whereas these RTL predictors are broken beyond
> repair.

Yep, Richard has changed this afterwards.  I will fix these ;)

Honza


Index Nav: [Date Index] [Subject Index] [Author Index] [Thread Index]
Message Nav: [Date Prev] [Date Next] [Thread Prev] [Thread Next]