[Bug tree-optimization/38785] huge performance regression on EEMBC bitmnp01

["rguenther at suse dot de" <gcc-bugzilla at gcc dot gnu dot org>]

------- Comment #10 from rguenther at suse dot de  2009-01-14 20:51 -------
Subject: Re:  huge performance regression on
 EEMBC bitmnp01

On Wed, 14 Jan 2009, amylaar at gcc dot gnu dot org wrote:

> I think the disregard for conditional execution opportunities and the
> assumption that phi nodes have no execution cost are two separate issues.
> I'd like to address the latter first, because it causes exponential code and
> execution time growth.
> 
> A phi node joining two constants has at least the cost of a constant load.
> A phi node joining two different variables which are initialized by a graph
> with constant leafs costs at least a reg-reg copy on one arm, plus the cost
> of its parents if these are needed solely for this phi node.
> 
> Therefore, if an expression is only partially anticipatable, we should compare
> the cost of any phi node needed to compute it early with the estimated
> likelyhod that such a computatatio, once done, is actually needed, multiplied
> with the cost of the replaced operation.       
> 
> Can we use edge probabilities inside tree-pre to calculate execution
> probabilities?
> 
> Can we calculate the cost of replaced expressions?

You would completely underestimate the optimization opportunities PRE
unleashes.

Richard.


-- 


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=38785