Re: [PATCH v1] tree-ssa-sink: Improve code sinking pass.

[Ajit Agarwal <aagarwa1@linux.ibm.com>]

Hello Richard:

On 22/05/23 6:26 pm, Richard Biener wrote:
> On Thu, May 18, 2023 at 9:14 AM Ajit Agarwal <aagarwa1@linux.ibm.com> wrote:
>>
>> Hello All:
>>
>> This patch improves code sinking pass to sink statements before call to reduce
>> register pressure.
>> Review comments are incorporated.
>>
>> Bootstrapped and regtested on powerpc64-linux-gnu.
>>
>> Thanks & Regards
>> Ajit
>>
>>
>> tree-ssa-sink: Improve code sinking pass.
>>
>> Code Sinking sinks the blocks after call. This increases
>> register pressure for callee-saved registers. Improves
>> code sinking before call in the use blocks or immediate
>> dominator of use blocks.
>>
>> 2023-05-18  Ajit Kumar Agarwal  <aagarwa1@linux.ibm.com>
>>
>> gcc/ChangeLog:
>>
>>         * tree-ssa-sink.cc (statement_sink_location): Modifed to
>>         move statements before calls.
>>         (block_call_p): New function.
>>         (def_use_same_block): New function.
>>         (select_best_block): Add heuristics to select the best
>>         blocks in the immediate post dominator.
>>
>> gcc/testsuite/ChangeLog:
>>
>>         * gcc.dg/tree-ssa/ssa-sink-20.c: New testcase.
>>         * gcc.dg/tree-ssa/ssa-sink-21.c: New testcase.
>> ---
>>  gcc/testsuite/gcc.dg/tree-ssa/ssa-sink-20.c |  16 ++
>>  gcc/testsuite/gcc.dg/tree-ssa/ssa-sink-21.c |  20 +++
>>  gcc/tree-ssa-sink.cc                        | 159 ++++++++++++++++++--
>>  3 files changed, 185 insertions(+), 10 deletions(-)
>>  create mode 100644 gcc/testsuite/gcc.dg/tree-ssa/ssa-sink-20.c
>>  create mode 100644 gcc/testsuite/gcc.dg/tree-ssa/ssa-sink-21.c
>>
>> diff --git a/gcc/testsuite/gcc.dg/tree-ssa/ssa-sink-20.c b/gcc/testsuite/gcc.dg/tree-ssa/ssa-sink-20.c
>> new file mode 100644
>> index 00000000000..716bc1f9257
>> --- /dev/null
>> +++ b/gcc/testsuite/gcc.dg/tree-ssa/ssa-sink-20.c
>> @@ -0,0 +1,16 @@
>> +/* { dg-do compile } */
>> +/* { dg-options "-O2 -fdump-tree-sink -fdump-tree-optimized -fdump-tree-sink-stats" } */
>> +
>> +void bar();
>> +int j;
>> +void foo(int a, int b, int c, int d, int e, int f)
>> +{
>> +  int l;
>> +  l = a + b + c + d +e + f;
>> +  if (a != 5)
>> +    {
>> +      bar();
>> +      j = l;
>> +    }
>> +}
>> +/* { dg-final { scan-tree-dump-times "Sunk statements: 5" 1 "sink" } } */
> 
> this doesn't verify the place we sink to?
>

I am not sure how to verify the place we sink to with dg-final.
 
>> diff --git a/gcc/testsuite/gcc.dg/tree-ssa/ssa-sink-21.c b/gcc/testsuite/gcc.dg/tree-ssa/ssa-sink-21.c
>> new file mode 100644
>> index 00000000000..ff41e2ea8ae
>> --- /dev/null
>> +++ b/gcc/testsuite/gcc.dg/tree-ssa/ssa-sink-21.c
>> @@ -0,0 +1,20 @@
>> +/* { dg-do compile } */
>> +/* { dg-options "-O2 -fdump-tree-sink-stats -fdump-tree-sink-stats" } */
>> +
>> +void bar();
>> +int j, x;
>> +void foo(int a, int b, int c, int d, int e, int f)
>> +{
>> +  int l;
>> +  l = a + b + c + d +e + f;
>> +  if (a != 5)
>> +    {
>> +      bar();
>> +      if (b != 3)
>> +        x = 3;
>> +      else
>> +        x = 5;
>> +      j = l;
>> +    }
>> +}
>> +/* { dg-final { scan-tree-dump-times "Sunk statements: 5" 1 "sink" } } */
> 
> likewise.  So both tests already pass before the patch?
> 
>> diff --git a/gcc/tree-ssa-sink.cc b/gcc/tree-ssa-sink.cc
>> index 87b1d40c174..76556e7795b 100644
>> --- a/gcc/tree-ssa-sink.cc
>> +++ b/gcc/tree-ssa-sink.cc
>> @@ -171,6 +171,72 @@ nearest_common_dominator_of_uses (def_operand_p def_p, bool *debug_stmts)
>>    return commondom;
>>  }
>>
>> +/* Return TRUE if immediate uses of the defs in
>> +   USE occur in the same block as USE, FALSE otherwise.  */
>> +
>> +bool
>> +def_use_same_block (gimple *stmt)
>> +{
>> +  use_operand_p use_p;
>> +  def_operand_p def_p;
>> +  imm_use_iterator imm_iter;
>> +  ssa_op_iter iter;
>> +
>> +  FOR_EACH_SSA_DEF_OPERAND (def_p, stmt, iter, SSA_OP_DEF)
>> +    {
>> +      FOR_EACH_IMM_USE_FAST (use_p, imm_iter, DEF_FROM_PTR (def_p))
>> +       {
>> +         if (is_gimple_debug (USE_STMT (use_p)))
>> +           continue;
>> +
>> +         if (use_p
> 
> use_p is never null
> 
>> +             && (gimple_bb (USE_STMT (use_p)) == gimple_bb (stmt)))
>> +           return true;
> 
> the function behavior is obviously odd ...
> 
>> +       }
>> +     }
>> +  return false;
>> +}
>> +
>> +/* Return TRUE if the block has only calls, FALSE otherwise. */
>> +
>> +bool
>> +block_call_p (basic_block bb)
>> +{
>> +  int i = 0;
>> +  bool is_call = false;
>> +  gimple_stmt_iterator gsi = gsi_last_bb (bb);
>> +  gimple *last_stmt = gsi_stmt (gsi);
>> +
>> +  if (last_stmt && gimple_code (last_stmt) == GIMPLE_COND)
>> +    {
>> +      if (!gsi_end_p (gsi))
>> +       gsi_prev (&gsi);
>> +
>> +       for (; !gsi_end_p (gsi);)
>> +        {
>> +          gimple *stmt = gsi_stmt (gsi);
>> +
>> +          /* We have already seen a call.  */
>> +          if (is_call)
>> +            return false;
> 
> Likewise.  Do you want to check whether a block has
> a single stmt and that is a call and that is followed by
> a condition?  It looks like a very convoluted way to write this.
> 
>> +
>> +          if (is_gimple_call (stmt))
>> +            is_call = true;
>> +          else
>> +            return false;
>> +
>> +          if (!gsi_end_p (gsi))
>> +            gsi_prev (&gsi);
>> +
>> +           ++i;
>> +       }
>> +     }
>> +  if (is_call && i == 1)
>> +    return true;
>> +
>> +  return false;
>> +}
>> +
>>  /* Given EARLY_BB and LATE_BB, two blocks in a path through the dominator
>>     tree, return the best basic block between them (inclusive) to place
>>     statements.
>> @@ -190,7 +256,8 @@ nearest_common_dominator_of_uses (def_operand_p def_p, bool *debug_stmts)
>>  static basic_block
>>  select_best_block (basic_block early_bb,
>>                    basic_block late_bb,
>> -                  gimple *stmt)
>> +                  gimple *stmt,
>> +                  gimple *use)
> 
> please update the function comment
> 
>>  {
>>    basic_block best_bb = late_bb;
>>    basic_block temp_bb = late_bb;
>> @@ -230,14 +297,47 @@ select_best_block (basic_block early_bb,
>>        if (threshold > 100)
>>         threshold = 100;
>>      }
>> -
>>    /* If BEST_BB is at the same nesting level, then require it to have
>>       significantly lower execution frequency to avoid gratuitous movement.  */
>>    if (bb_loop_depth (best_bb) == bb_loop_depth (early_bb)
>>        /* If result of comparsion is unknown, prefer EARLY_BB.
>>          Thus use !(...>=..) rather than (...<...)  */
>>        && !(best_bb->count * 100 >= early_bb->count * threshold))
>> -    return best_bb;
>> +    {
>> +      basic_block new_best_bb = get_immediate_dominator (CDI_DOMINATORS, best_bb);
>> +      /* Return best_bb if def and use are in same block otherwise new_best_bb.
>> +
>> +        Things to consider:
>> +
>> +          new_best_bb is not equal to best_bb and early_bb.
>> +
>> +          stmt is not call.
>> +
>> +          new_best_bb doesnt have any phis.
>> +
>> +          use basic block is not equal to early_bb.
>> +
>> +          use basic block post dominates to new_best_bb.
>> +
>> +          new_best_bb dominates early_bb. */
>> +      if (new_best_bb && use
>> +         && (new_best_bb != best_bb)
>> +         && (new_best_bb != early_bb)
>> +         && !is_gimple_call (stmt)
>> +         && gsi_end_p (gsi_start_phis (new_best_bb))
>> +         && (gimple_bb (use) != early_bb)
>> +         && !is_gimple_call (use)
>> +         && dominated_by_p (CDI_POST_DOMINATORS, new_best_bb, gimple_bb(use))
>> +         && dominated_by_p (CDI_DOMINATORS, new_best_bb, early_bb)
>> +         && block_call_p (new_best_bb))
>> +       {
>> +         if (def_use_same_block (use))
>> +           return best_bb;
> 
> given the odd implementation of the predicates this matches very very
> specific cases.
> 
> Consider
> 
>  if (..)
>   {
>     foo();
>     bar();
>     ... = l;
>   }
> 
> and C++ where foo and bar might throw.  You then likely want to sink
> before foo ().
> 
> What's the reason to only consider blocks with exactly 'call; cond;' ?
> 
>> +
>> +         return new_best_bb;
>> +       }
>> +       return best_bb;
>> +    }
>>
>>    /* No better block found, so return EARLY_BB, which happens to be the
>>       statement's original block.  */
>> @@ -439,7 +539,7 @@ statement_sink_location (gimple *stmt, basic_block frombb,
>>        if (!dominated_by_p (CDI_DOMINATORS, commondom, frombb))
>>         return false;
>>
>> -      commondom = select_best_block (frombb, commondom, stmt);
>> +      commondom = select_best_block (frombb, commondom, stmt, NULL);
>>
>>        if (commondom == frombb)
>>         return false;
>> @@ -456,19 +556,58 @@ statement_sink_location (gimple *stmt, basic_block frombb,
>>             continue;
>>           break;
>>         }
>> +
>>        use = USE_STMT (one_use);
>>
>>        if (gimple_code (use) != GIMPLE_PHI)
>>         {
>> -         sinkbb = select_best_block (frombb, gimple_bb (use), stmt);
>> +         sinkbb = select_best_block (frombb, gimple_bb (use), stmt, use);
>>
>>           if (sinkbb == frombb)
>>             return false;
>>
>> -         if (sinkbb == gimple_bb (use))
>> -           *togsi = gsi_for_stmt (use);
>> -         else
>> -           *togsi = gsi_after_labels (sinkbb);
>> +          gimple *def_stmt = SSA_NAME_DEF_STMT (DEF_FROM_PTR (def_p));
>> +
>> +          if ((gimple_bb (def_stmt) == gimple_bb (use))
>> +               && (gimple_bb (use) != sinkbb))
>> +            sinkbb = gimple_bb (use);
>> +
>> +           if (sinkbb == gimple_bb (use))
>> +             {
>> +               gimple_stmt_iterator gsi = gsi_last_bb (sinkbb);
>> +               gimple *def_stmt = SSA_NAME_DEF_STMT (DEF_FROM_PTR (def_p));
>> +               gimple *last_stmt = gsi_stmt (gsi);
>> +
>> +               /* Update sinking point as stmt before call if the sinking block
>> +                  has only calls. Otherwise update sinking point as the use
>> +                  stmt. */
>> +               if (gsi_stmt (gsi) == use
>> +                   && !is_gimple_call (last_stmt)
>> +                   && (gimple_code (last_stmt) != GIMPLE_SWITCH)
>> +                   && (gimple_code (last_stmt) != GIMPLE_COND)
>> +                   && (gimple_code (last_stmt) != GIMPLE_GOTO)
>> +                   && (!gimple_vdef (use) || !def_use_same_block (def_stmt)))
>> +                 {
>> +                   if (!gsi_end_p (gsi))
>> +                     gsi_prev (&gsi);
>> +
>> +                   gimple *stmt = gsi_stmt (gsi);
>> +
>> +                   if (!gsi_end_p (gsi))
>> +                     gsi_prev (&gsi);
>> +
>> +                   if (gsi_end_p (gsi) && stmt && is_gimple_call (stmt)
>> +                       && gsi_end_p (gsi_start_phis (sinkbb))
>> +                       && !is_gimple_call (def_stmt))
>> +                     *togsi = gsi_for_stmt (stmt);
>> +                   else
>> +                     *togsi = gsi_for_stmt (use);
>> +                  }
>> +               else
>> +                 *togsi = gsi_for_stmt(use);
>> +              }
>> +            else
>> +               *togsi = gsi_after_labels (sinkbb);
> 
> This is very convoluted.  I think that in the end you want to compute (once) the
> position of the first call in each block.  Since we're waking the CFG backwards
> in post-dominator order this information can be gathered during this walk.
> This would determine the location to sink to iff the use stmt is dominated by
> this location (you can for example use gimple_uid to mark stmts before it).
> 
> The alternative is to simply always sink to the start of blocks even for the
> use stmt block in case that has a call before the use (but you still need to
> efficiently compute that).
>

Incorporated the above comments and sent a separate patch.

Thanks & Regards
Ajit
 
> Richard.
> 
>>
>>           return true;
>>         }
>> @@ -480,7 +619,7 @@ statement_sink_location (gimple *stmt, basic_block frombb,
>>    if (!sinkbb)
>>      return false;
>>
>> -  sinkbb = select_best_block (frombb, sinkbb, stmt);
>> +  sinkbb = select_best_block (frombb, sinkbb, stmt, NULL);
>>    if (!sinkbb || sinkbb == frombb)
>>      return false;
>>
>> --
>> 2.31.1
>>