Re: PATCH] tree-ssa-sink: Add heuristics for code sinking

[Richard Biener <richard.guenther@gmail.com>]

On Fri, Apr 14, 2023 at 10:42 AM Ajit Agarwal via Gcc-patches
<gcc-patches@gcc.gnu.org> wrote:
>
> Hello All:
>
> This patch add heuristics for code sinking opportunities.
> Bootstrapped and regtested for powerpc64-linux-gnu.
>
> Thanks & Regards
> Ajit
>
>         tree-ssa-sink: Add heuristics for code sinking.
>
>         Add following code sinking heuristics:
>
>         1. from code block dominates the call.
>         2. To Code block have uses inside the function call.
>         3. Loop headers.
>         4. Sinking from code block after call increases register
>         pressure.
>         5. Sinking calls.
>
>         2023-04-14  Ajit Kumar Agarwal  <aagarwa1@linux.ibm.com>
>
> gcc/ChangeLog:
>
>         * tree-ssa-sink.cc (statement_sink_location): Add heuristics
>         for code sinking.
> ---
>  gcc/tree-ssa-sink.cc | 33 +++++++++++++++++++++++++++++++++
>  1 file changed, 33 insertions(+)
>
> diff --git a/gcc/tree-ssa-sink.cc b/gcc/tree-ssa-sink.cc
> index 87b1d40c174..8de88b259a3 100644
> --- a/gcc/tree-ssa-sink.cc
> +++ b/gcc/tree-ssa-sink.cc
> @@ -465,6 +465,39 @@ statement_sink_location (gimple *stmt, basic_block frombb,
>           if (sinkbb == frombb)
>             return false;
>
> +         auto_vec<basic_block> h;
> +         h = get_all_dominated_blocks (CDI_DOMINATORS,
> +                                       frombb);
> +         bool is_call = false;
> +         while (h.length ())
> +           {
> +             basic_block bb = h.pop ();
> +
> +             if (bb == frombb)
> +               continue;
> +
> +             for (gimple_stmt_iterator gsi = gsi_last_bb (bb); !gsi_end_p (gsi);)
> +               {
> +                 gimple *stmt = gsi_stmt (gsi);
> +
> +                 if (is_gimple_call (stmt))
> +                   {
> +                     is_call = true;
> +                     break;
> +                   }
> +
> +                  if (!gsi_end_p (gsi))
> +                    gsi_prev (&gsi);
> +               }
> +            }
> +
> +           if (!is_gimple_call (stmt)
> +               && (gimple_bb (use) != frombb)
> +               && !is_gimple_call (use)
> +               && dominated_by_p (CDI_DOMINATORS, sinkbb, frombb)
> +               && is_call)
> +              return false;
> +

Sorry, but this lacks a comment, it doesn't explain why the existing heuristics
are not enough (select_best_block), it repeats dominance computing.

More so it lacks a testcase demonstrating the effect.

Richard.

>           if (sinkbb == gimple_bb (use))
>             *togsi = gsi_for_stmt (use);
>           else
> --
> 2.31.1
>