Re: [f-cpu] dynamic code reuse

[Nicolas Boulay <nico@seul.org>]

Imagine the instruction :

WATCH R1,R2,R3, [PC + #imm] W1, W2

R1,R2,R3, W1,W2 are register #imm is an immediat

Those instruction act as :

if (def cache{R1,R2,R3,PC})
{
 	fetch (PC + #imm)
	(W1,W2) = cache{R1,R2,R3,PC}
} 
else
{       fetch (PC+1)
	cache(R1,R2,R3,PC) = (W1,W2)@[PC + #imm] <- the result of registre W1,W2 
                                                      are cached at this point
}

It explicitly keep a result in a specific cache. 

nicO

Le Lundi 8 Décembre 2003 18:25, devik a écrit :
> You should post short example. I don't understand your idea yet.
>
> On Mon, 8 Dec 2003, Nicolas Boulay wrote:
> > Le Dimanche 7 Décembre 2003 16:36, Michael Riepe a écrit :
> > > On Sun, Dec 07, 2003 at 04:19:58PM +0000, Nicolas Boulay wrote:
> > > > There is a branch of studies concerning the reuse of soon calculated
> > > > value in code.
> > > >
> > > > http://citeseer.nj.nec.com/sodani97dynamic.html
> > > >
> > > > The idea from the paper is that there is always some result in the
> > > > code that have been soon calculated. The idea is to reuse the result
> > > > instead of recalulate it. Increase speed could reach 20%.
> > > >
> > > > This gave me an idea but with software support.
> > > >
> > > > This is a kind of jump instruction. The jump take some register and
> > > > one adresse. If the content of the r register + the PC match a
> > > > internal specific cache the result of 2w register is set and the jump
> > > > occur.
> > >
> > > Sounds a lot like `speculative execution' to me.
> > > Disclaimer: I didn't read the paper yet.
> >
> > nop this is a specultative reuse of a previous calculated result.
> >
> > > > To fill the cache, some thing "watch" the result calculated  where
> > > > the instruction is supposed to jump. In case of interrupt, this is
> > > > lost to avoid confusion and to stay stateless.
> > > >
> > > > This instruction needed to be handle carrefully by the compiler. So
> > > > what's up in the area ?
> > >
> > > Rien, I suppose.
> >
> > Devik ? what's your opinion about this feature to implemented in the
> > compiler ?
> >
> > nicO
> >

*************************************************************
To unsubscribe, send an e-mail to majordomo@seul.org with
unsubscribe f-cpu       in the body. http://f-cpu.seul.org/