GVN-PRE sucks -- well, for SPECCFP2000 mgrid anyway

[Jeffrey A Law]

On Sun, 2004-08-15 at 20:28, Richard Kenner wrote:
>     > 1) let pre be pre and have the register allocator "rematerialize" the 
>     > guys that have been pulled out of the loop.
> 
>     I've always tried to argue this approach, but it never seems to get 
>     places with people, because it means the register allocator doing *a lot*.
> 
> Not necessarly.  After all, reload has "rematerialized" constants almost
> since day one and has also rematerialized some addition operations more
> recently.
True, but its ability to rematerialize is poor at best.  And the
way it's structured, particularly for chains of insns producing an
ultimate value is, err, special.


> But I still agree with Kenny that it's the best approach: I don't think
> that complex interactions between passes to estimate register pressure
> will work well.
Likewise.

jeff