Re: [PATCH]: Speed up SSA-CCP

[Daniel Berlin <dberlin at dberlin dot org>]

On May 22, 2004, at 12:24 PM, Diego Novillo wrote:

> On Sat, 2004-05-22 at 09:06, Daniel Berlin wrote:
> > This speeds up SSA-CCP by moving more things to VARYING in one step
> > instead of two.
> > Kenneth Zadeck said it was worth about 30-40% performance in their
> > compiler.
> > SSA-CCP is already fast in our compiler, but it can't hurt to make it
> > faster (this patch can't ever slow it down, since it can only lessen
> > the number of iterations it takes to finish the algorithm, not 
> > increase
> > it).
> >
> >
> > 2004-05-22  Daniel Berlin  <dberlin@dberlin.org>
> >        Kenneth Zadeck <zadeck@naturalbridge.com>
> >
> >     * tree-ssa-ccp.c (varying_ssa_edges): New worklist.
> >     (add_var_to_ssa_edges_worklist): Add value argument.
> >     Update callers.
> >     Use new worklist.
> >     (tree_ssa_ccp): Use new worklist.
> It looks good.  Some questions before adding it:
>
>      1. Please factor out the loop that goes through the edge lists, so
>         that the main iteration loop in tree_ssa_ccp() calls
>         process_edge_list (ssa_edges) and process_edge_list
>         (varying_ssa_edges).

Okeydokey

>      2. Is Kenneth's copyright assignment already on file?

He doesn't need one (he didn't directly write any of this code, but he 
still deserves credit for it), but he would be covered by Apple's 
anyway.

>      3. Could you add a small example that illustrates how the new list
>         speeds up the demotion to VARYING?
Sure.

>      4. Any timings worth mentioning?
On three test cases where CCP took ~3 seconds, it now takes 1.8, 2.3, 
and 1.7 seconds
I didn't bother timing incredibly large pieces of code :)

> Thanks.  Diego.