Re: [RFC] Move ehcleanup pass to before early SRA

[Richard Biener <richard dot guenther at gmail dot com>]

On Wed, Jan 15, 2014 at 10:24 AM, Eric Botcazou <ebotcazou@adacore.com> wrote:
> Hi,
>
> we have run into optimization regressions in Ada caused by the interaction
> between the GIMPLE clobbers and SRA: on the one hand GIMPLE clobbers create
> artificial EH handlers for aggregate variables, on the other hand SRA refuses
> to scalarize objects that appear in a statement that can throw internally.
> The result is that GIMPLE clobbers block the scalarization of variables that
> used to be possible up to the 4.6 series.  Testcase attached, compile p.adb on
> x86-64 with -O2 -fdump-tree-ealias -fdump-tree-esra.
>
> This can be solved by moving the ehcleanup pass to before early SRA in
> the pipeline.  It has a small but measurable positive effect on some of our
> benchmarks (with a 4.7-based compiler).  Tested on x86-64/Linux.
>
> Thoughts?

What you want is tree-eh.c:optimize_clobbers, right?  Can't we
do this optimization during EH refactoring / lowering as well?

Also why does SRA refuse to scalarize here?  It will end up
scalarizing during regular opts as

<L1>:
  result = decls_support.get_private (100);
  result$i_12 = MEM[(struct decls_support__t_private *)&result];

where it can do the scalarization just fine if it inserts on all
non-EH successor edges (so not doing the scalarization only
if there is more than one non-EH successor edge would make
sense).

The current placement of EH cleanup is to make it catch
all opportunities before early inlining sees the function as
callee.  And I guess both FRE and DCE will expose quite
some EH cleanup opportunities especially with non-call-exceptions.

Richard.

>
> 2014-01-15  Eric Botcazou  <ebotcazou@adacore.com>
>
>         * passes.def (pass_early_local_passes): Run first EH cleanup pass early.
>
>
> --
> Eric Botcazou