Re: [PATCH] tree-ssa-loop-ivopts : Add live analysis in regs used in decision making

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

On Fri, Nov 10, 2023 at 7:42 AM Ajit Agarwal <aagarwa1@linux.ibm.com> wrote:
>
> Hello Richard:
>
>
> On 09/11/23 6:21 pm, Richard Biener wrote:
> > On Wed, Nov 8, 2023 at 4:00 PM Ajit Agarwal <aagarwa1@linux.ibm.com> wrote:
> >>
> >> tree-ssa-loop-ivopts : Add live analysis in regs used in decision making.
> >>
> >> Add live anaysis in regs used calculation in decision making of
> >> selecting ivopts candidates.
> >>
> >> 2023-11-08  Ajit Kumar Agarwal  <aagarwa1@linux.ibm.com>
> >>
> >> gcc/ChangeLog:
> >>
> >>         * tree-ssa-loop-ivopts.cc (get_regs_used): New function.
> >>         (determine_set_costs): Call to get_regs_used to use live
> >>         analysis.
> >> ---
> >>  gcc/tree-ssa-loop-ivopts.cc | 73 +++++++++++++++++++++++++++++++++++--
> >>  1 file changed, 70 insertions(+), 3 deletions(-)
> >>
> >> diff --git a/gcc/tree-ssa-loop-ivopts.cc b/gcc/tree-ssa-loop-ivopts.cc
> >> index c3336603778..e02fe7d434b 100644
> >> --- a/gcc/tree-ssa-loop-ivopts.cc
> >> +++ b/gcc/tree-ssa-loop-ivopts.cc
> >> @@ -6160,6 +6160,68 @@ ivopts_estimate_reg_pressure (struct ivopts_data *data, unsigned n_invs,
> >>    return cost + n_cands;
> >>  }
> >>
> >> +/* Return regs used based on live-in and liveout of given ssa variables.  */
> >
> > Please explain how the following code relates to anything like "live
> > analysis" and
> > where it uses live-in and live-out.  And what "live-in/out of a given
> > SSA variable"
> > should be.
> >
> > Also explain why you are doing this at all.  The patch doesn't come
> > with a testcase
> > or with any other hint that motivated you.
> >
> > Richard.
> >
>
> The function get_regs_used increments the regs_used based on live-in
> and live-out analysis of given ssa name. Instead of setting live-in and
> live-out bitmap I increment the regs_used.
>
> Below is how I identify live-in and live-out and increments the regs_used
> variable:
>
> a) For a given def_bb of gimple statement of ssa name there should be
> live-out and increments the regs_used.
>
> b) Visit each use of SSA_NAME and if it isn't in the same block as the def,
>  we identify live on entry blocks and increments regs_used.
>
> The below function is the modification of set_var_live_on_entry of tree-ssa-live.cc
> Where we set the bitmap of liveout and livein of basic block. Instead of setting bitmap, regs_used is incremented.

It clearly doesn't work that way, and the number doesn't in any way relate to
the number of registers used or register pressure.

> I identify regs_used as the number of live-in and liveout of given ssa name variable.
>
> For each iv candiate ssa variables I identify regs_used and take maximum of regs
> used for all the iv candidates that will be used in ivopts_estimate_register_pressure
> cost analysis.
>
> Motivation behind doing this optimization is I get good performance improvement
> for several spec cpu 2017 benchmarks for FP and INT around 2% to 7%.

An interesting GIGO effect.

> Also setting regs_used as number of iv candiates, which is not
> optimized and robust way of decision making for ivopts optimization I decide
> on live-in and live-out analysis which is more correct and appropriate way of
> identifying regs_used.
>
> And also there are no regressions in bootstrapped/regtested on powerpc64-linux-gnu.
>
> Thanks & Regards
> Ajit
>
> >> +static unsigned
> >> +get_regs_used (tree ssa_name)
> >> +{
> >> +  unsigned regs_used = 0;
> >> +  gimple *stmt;
> >> +  use_operand_p use;
> >> +  basic_block def_bb = NULL;
> >> +  imm_use_iterator imm_iter;
> >> +
> >> +  stmt = SSA_NAME_DEF_STMT (ssa_name);
> >> +  if (stmt)
> >> +    {
> >> +      def_bb = gimple_bb (stmt);
> >> +      /* Mark defs in liveout bitmap temporarily.  */
> >> +      if (def_bb)
> >> +       regs_used++;
> >> +    }
> >> +  else
> >> +    def_bb = ENTRY_BLOCK_PTR_FOR_FN (cfun);
> >> +
> >> +  /* An undefined local variable does not need to be very alive.  */
> >> +  if (virtual_operand_p (ssa_name)
> >> +      || ssa_undefined_value_p (ssa_name, false))
> >> +    return 0;
> >> +
> >> +  /* Visit each use of SSA_NAME and if it isn't in the same block as the def,
> >> +     add it to the list of live on entry blocks.  */
> >> +  FOR_EACH_IMM_USE_FAST (use, imm_iter, ssa_name)
> >> +    {
> >> +      gimple *use_stmt = USE_STMT (use);
> >> +      basic_block add_block = NULL;
> >> +
> >> +      if (gimple_code (use_stmt) == GIMPLE_PHI)
> >> +       {
> >> +         /* Uses in PHI's are considered to be live at exit of the SRC block
> >> +            as this is where a copy would be inserted.  Check to see if it is
> >> +            defined in that block, or whether its live on entry.  */
> >> +         int index = PHI_ARG_INDEX_FROM_USE (use);
> >> +         edge e = gimple_phi_arg_edge (as_a <gphi *> (use_stmt), index);
> >> +         if (e->src != def_bb)
> >> +           add_block = e->src;
> >> +       }
> >> +      else if (is_gimple_debug (use_stmt))
> >> +       continue;
> >> +      else
> >> +       {
> >> +         /* If its not defined in this block, its live on entry.  */
> >> +         basic_block use_bb = gimple_bb (use_stmt);
> >> +         if (use_bb != def_bb)
> >> +           add_block = use_bb;
> >> +       }
> >> +
> >> +      /* If there was a live on entry use, increment register used.  */
> >> +      if (add_block)
> >> +       {
> >> +         regs_used++;
> >> +       }
> >> +    }
> >> +  return regs_used;
> >> +}
> >> +
> >>  /* For each size of the induction variable set determine the penalty.  */
> >>
> >>  static void
> >> @@ -6200,15 +6262,20 @@ determine_set_costs (struct ivopts_data *data)
> >>        n++;
> >>      }
> >>
> >> +  unsigned max = 0;
> >>    EXECUTE_IF_SET_IN_BITMAP (data->relevant, 0, j, bi)
> >>      {
> >>        struct version_info *info = ver_info (data, j);
> >> -
> >>        if (info->inv_id && info->has_nonlin_use)
> >> -       n++;
> >> +       {
> >> +         tree ssa_name = ssa_name (j);
> >> +         n = get_regs_used (ssa_name);
> >> +         if (n >= max)
> >> +           max = n;
> >> +       }
> >>      }
> >>
> >> -  data->regs_used = n;
> >> +  data->regs_used = max;
> >>    if (dump_file && (dump_flags & TDF_DETAILS))
> >>      fprintf (dump_file, "  regs_used %d\n", n);
> >>
> >> --
> >> 2.39.3
> >>
> >>