none (none.delete@this.none.com) on August 14, 2012 6:25 am wrote:
> anon (anon.delete@this.anon.com) on August 14, 2012 6:12 am wrote:
> > none
> (none.delete@this.none.com) on August 14, 2012 5:40 am wrote:
> > > anon
>
> > (anon.delete@this.anon.com) on August 14, 2012 5:28 am wrote:
> > >
> [...]
> > >
> > >
> > > > IMHO predication is
> > >
> > not worth the trouble. With
> > advanced
> > > > >
> >
> > micro architectures, you'd better
> > > >
> > implement a good
> branch predictor
> > > rather than
> > > > > spending
> >
> time implementing
> > > > predication (which
> > > cost can be
> high).
> > >
> > >
> > > > What makes you say
> > >
> that?
> > >
> > >
> > Benchmarking.
> >
> > Please
> elaborate.
>
> NDA :)
>
> But you can guess that ARM didn't make such a radical
> change without doing benchmarking first.

Oh yes for full scale predication yes, because it's useless to execute down long paths of branches or potentially take expensive cache misses.

But that's not to say predication is useless in general. Something like select looks nice.

>
> > > > Intel and
> > >
> > AMD
> > both spend effort to make
> > > cmov for example run fast
> and low latency.
> > Intel
> > > > optimization manual
> > >
> recommends using it. Eliminating
> > fundamentally unpredictable
> > >
> > branches will
> > > reduce pressure on
> > branch predictor for
> predictable branches and
> > > > avoid
> > >
> >
> mispredicts.
> > >
> > > Aarch64 has the equivalent of cmov.
> >
>
> > It has some
> > predicated instructions, yes. Which would seem to add
> weight to the idea that
> > predication is useful.
>
> What kind of
> instruction would you like to see predicated?

Basically some conditional ALU operations should be enough.

> Even on Cortex-A9 predication
> usage is discouraged, in particular for ld/st which will block precious LSU
> resources until the flags are resolved.

I don't think load/store are quite so important.

I certainly agree that massive predication schemes as-envisaged for in-order non-powerwall-constrained devices that was dreamed up decades ago is no good for today. What ARM64 has looks nice though, and I'm sure it will be a win in some cases to eliminated unpredictable branches on implementations of it.