Re: [dataflow][PATCH] Compute dominance information in ce2 pass.

["Seongbae Park (박성배, 朴成培)" <seongbae dot park at gmail dot com>]

On 5/14/07, Seongbae Park <seongbae.park@gmail.com> wrote:
> Hi,
>
> The attached patch fixes the performance regression on SPEC CPU2000 wupwise.
> The mainline computes the dominance information in all ifconvert passes
> except ce1, but  df doesn't compute it at all.
>
> Bootstrap and regtest are still running.
>
> 2007-05-14  Seongbae Park <seongbae.park@gmail.com>
>
>         * ifcvt.c (if_convert): New parameter RECOMPUTE_DOMINANCE.
>         (rest_of_handle_if_conversion, rest_of_handle_if_after_combine,
>         rest_of_handle_if_after_reload): New parameter to if_convert.

The previous patch caused a bootstrap failure,
primarily because some of the if conversion routines were not executed
at all before my patch.
Enclosed is the new patch which passed the bootstrap and regression tests.

I think we probably want to add
df-based routines that do equivalent to
propagate_one_insn() so that people can compute
live sets at arbitrary points in the instruction stream
(without directly using df infrastructure).
I'll try to do that sometime.

The updated changelog is:
2007-05-16  Seongbae Park <seongbae.park@gmail.com>

        * ifcvt.c (dead_or_predicable): Compute the live set at EARLIEST.
        (if_convert): New parameter RECOMPUTE_DOMINANCE.
        (rest_of_handle_if_conversion, rest_of_handle_if_after_combine,
        rest_of_handle_if_after_reload): New parameter to if_convert.


OK for the branch ?
--
#pragma ident "Seongbae Park, compiler, http://seongbae.blogspot.com";

Attachment: df-ifcvt-dominance2.diff.txt
Description: Text document