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: [patch 2/2] PR27313 Transform conditional stores


On 5/15/07, Michael Matz <matz@suse.de> wrote:
Hi,

On Tue, 15 May 2007, Richard Guenther wrote:

> > > Yes this they are different targets.  This is why GCC have to
> > > becareful when implementing any kind of optimization, including this
> > > one.
> >
> > Okay, to perhaps move forward with this, do you have any suggestion on a
> > sensible heuristic when to not do the transformation I want?  Otherwise
> > I'm tempted to simply activate them by default on x86* (at least the SPEC
> > results support that), and leave other architectures alone.
>
> I suppose you can check BRANCH_COST - it's 20 for SPU, 3 for PPC and
> between 1 and 5 on x86{_64}.

Does PPC have a branch-predictor?  Because the costs for mispredicted
branches are the important ones (and I'm not sure they're correct for
x86*).

Well, BRANCH_COST is more like a cost for the average branch and thus a quality indicator for the branch prediction unit. SPU must suck there ;)

Richard.


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