Re: cache misses in gcc 3.3

[Daniel Berlin <dberlin at dberlin dot org>]

On Monday, February 10, 2003, at 12:16  PM, Andi Kleen wrote:

> On Mon, Feb 10, 2003 at 05:30:25PM +0100, Diego Novillo wrote:
> > On Mon, 10 Feb 2003, Andi Kleen wrote:
> >
> > > Hopefully CSE can be fixed too - I assume tree-ssa will do that 
> > > eventually?
> > Well, by having PRE and other cleanups at the tree level, the
> > hope is to go into GCSE and rip out all the code that tries to
> > deal with cases that can be more easily caught at the tree level.
> GCSE doesn't seem to be that slow, just CSE is.
It's also not nearly as effective as it could be.  GCSE doesn't catch a 
lot of things, or perform very good load motion (not that it could, 
it's not got good enough information), and performs no store motion, or 
strength reduction.
It also increases register pressure in the current incarnation.

SSAPRE does all of these (store motion, load motion, strength 
reduction, and PRE).

The current GCSE implementation at the RTL level is edge-based LCM.
So it's not going to be slow unless you have a ton of expressions, or a 
ton of edges.


> CSE                   :   0.46 (11%) usr   0.00 ( 0%) sys   0.50 (10%) 
> wall
> global CSE            :   0.14 ( 3%) usr   0.00 ( 0%) sys   0.15 ( 3%) 
> wall
>
>
> -Andi