Re: Merging mem-ssa into mainline

["Richard Guenther" <richard dot guenther at gmail dot com>]

On 12/7/06, Diego Novillo <dnovillo@redhat.com> wrote:
> I will be merging mem-ssa into mainline over the next few days.  The
> experiments that I have been doing show marked improvements in compile
> times over the set of programs that I have in my sandbox.  The
> compiler should now be between 5% and 25% faster.
>
> Here's a summary of the speedups I obtained with respect to mainline
> @119145.  I am only including timings over the affected modules.  Many
> passes are now faster, but this is mostly due to speedups in the
> operand scanner and SSA updating.  Also, phases that show less than
> 0.5% change are not included:
>
> cc1-i-files
> -----------
> Phase                           Before  After   % change
> tree PTA                        16.39   15.35     -6.3%
> tree alias analysis             12.33   11.75     -4.7%
> tree SSA rewrite                 6.24    5.13    -17.8%
> tree SSA incremental            15.79   12.09    -23.4%
> tree operand scan               85.90   52.28    -39.1%
> TOTAL                          476.43  437.47     -8.2%
>
>
> DLV
> ---
> Phase                           Before  After   % change
> tree PTA                         4.22    3.69    -12.6%
> tree SSA incremental             4.09    2.89    -29.3%
> tree operand scan               20.85   14.47    -30.6%
> TOTAL                          101.39   91.69     -9.6%
>
>
> MICO
> ----
> Phase                           Before  After   % change
> tree SSA rewrite                 4.56    3.13    -31.4%
> tree SSA incremental            10.00    5.93    -40.7%
> tree operand scan               90.19   50.24    -44.3%
> TOTAL                          444.08  391.64    -11.8%
>
>
> SPEC2000
> --------
> Phase                           Before  After   % change
> tree alias analysis              6.01    5.31    -11.6%
> tree SSA rewrite                 4.54    2.66    -41.4%
> tree SSA incremental             9.47    5.08    -46.4%
> tree operand scan               49.93   43.83    -12.2%
> TOTAL                          287.97  271.64     -5.7%
>
>
> TRAMP3D
> -------
> Phase                           Before  After   % change
> tree SSA rewrite                 3.80    2.71    -28.7%
> tree SSA incremental             3.29    1.92    -41.6%
> tree operand scan               22.72   16.87    -25.7%
> expand                           5.88    5.06    -13.9%
> TOTAL                           88.78   77.68    -12.5%
>
>
> VARIOUS (PR12850)
> -------
> Phase                           Before  After   % change
> tree alias analysis              4.61    3.66    -20.6%
> tree SSA rewrite                 4.77    3.10    -35.0%
> tree SSA incremental             5.85    1.87    -68.0%
> tree operand scan               54.29   26.98    -50.3%
> TOTAL                          159.39  118.16    -25.9%

Do you have numbers for memory usage?  It would be nice to incrementally merge
the infrastructure changes with a one to two days delay inbetween each
change, so
regular performance / memory testers can see which change is causing what
improvement or regression.

Otherwise, nice numbers!

Richard.