Re: Updated IPCP cost model

["H.J. Lu" <hjl dot tools at gmail dot com>]

On Sun, Aug 24, 2008 at 10:55 AM, Jan Hubicka <jh@suse.cz> wrote:
> Hi,
> this is revised version of patch I would like to commit tonight.  This
> version is one the ipcp scores I posted earlier are based on.  Main
> change is that there is new ipa-cp-clone option that controls clonning.
> This way we can do classical ipcp at -O2/-Os and perform clonning at
> -O3.
>
> I've also added testcase testing the cost model.
>
> Bootstrapped/regtested i686-linux, I will test it on SPEC tonight too
> and will send separate patch to enable at -O2 if things works as
> expected.
>        * doc/invoke.texi (-fipa-cp-clone): New option.
>        (-fipa-cp): Update docs.
>        (--param ipcp-unit-growth):New.
>        * ipa-cp.c: Include fibheap.h, params.h
>        (ipcp_initialize_node_lattices): When not clonning, all externally
>        visible functions are bottom.
>        (ipcp_need_redirect_p): Accept clones.
>        (ipcp_insert_stage): Use cost driven heuristics.
>        (max_count, dead_nodes): New static vars.
>        (ipcp_need_original_clone_p, ipcp_estimate_clonning_cost,
>        ipcp_const_param_count): New functions.
>        * common.opt (ipa-cp-clone): New command line option.
>        * params.def (ipcp-unit-growth): New.
>
>        * gcc.dg/ipa/ipacost-1.c: New testcase.
>        * gcc.dg/ipa/ipacost-2.c: New testcase.
>        * gcc.dg/ipa/ipa-7.c: Update template.
>

Hi,

I think one of your checkins caused

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=37227


H.J.