Re: [ tree-ssa ] Avoid useless reevaluations in CCP

[Daniel Berlin <dberlin at dberlin dot org>]

On Wednesday, February 19, 2003, at 01:06  PM, Diego Novillo wrote:

> On Wed, 2003-02-19 at 13:02, law at redhat dot com wrote:
>
> > Yea, but they never actually compare their scheme with the more
> > traditional schemes.  You've got to love papers which present no
> > actual data :-)  But yes, it looks promising.
> Shhh, no actual data is the secret recipe of many theses and 
> theoretical
> papers  :)

Well, that's because they know the only way you can completely get rid 
of *actual* reevaluations in CCP (you still might need to visit 
something more than once, but wouldn't need to actually evaluate it 
again, methinks) is to use something that captures value flow (which is 
what CCP is. A value flow problem), like a value flow graph.

Methinks you should be able to cut down on the number of actual visits 
by better statement and edge ordering (since it is, after all, only 
revisiting changed things to achieve a fixed point).
> Diego.