[v4] tree-ssa-sink: Improve code sinking pass

  Hello All:

This patch improves code sinking pass to sink statements before call to reduce
register pressure.
Review comments are incorporated.

For example :

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;
    }
}

Code Sinking does the following:

void bar();
int j;
void foo(int a, int b, int c, int d, int e, int f)
{
  int l;
  
  if (a != 5)
    {
      l = a + b + c + d +e + f;     
      bar();
      j = l;
    }
}

Bootstrapped 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.

2023-05-24  Ajit Kumar Agarwal  <aagarwa1@linux.ibm.com>

gcc/ChangeLog:

	* tree-ssa-sink.cc (statement_sink_location): Move statements before
	calls.
	(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 | 15 +++++
 gcc/testsuite/gcc.dg/tree-ssa/ssa-sink-21.c | 19 ++++++
 gcc/tree-ssa-sink.cc                        | 74 +++++++++++++++++----
 3 files changed, 96 insertions(+), 12 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
  

Hi!

On Wed, 31 May 2023 12:31:35 +0530
Ajit Agarwal via Gcc-patches <gcc-patches@gcc.gnu.org> wrote:

> Hello All:
> 
> This patch improves code sinking pass to sink statements before call to reduce
> register pressure.
> Review comments are incorporated.
> 
> For example :
> 
> 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;
>     }
> }
> 
> Code Sinking does the following:
> 
> void bar();
> int j;
> void foo(int a, int b, int c, int d, int e, int f)
> {
>   int l;
>   
>   if (a != 5)
>     {
>       l = a + b + c + d +e + f;     
>       bar();
>       j = l;
>     }
> }
> 
> Bootstrapped 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.
> 
> 2023-05-24  Ajit Kumar Agarwal  <aagarwa1@linux.ibm.com>
> 
> gcc/ChangeLog:
> 
> 	* tree-ssa-sink.cc (statement_sink_location): Move statements before
> 	calls.
> 	(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 | 15 +++++
>  gcc/testsuite/gcc.dg/tree-ssa/ssa-sink-21.c | 19 ++++++
>  gcc/tree-ssa-sink.cc                        | 74 +++++++++++++++++----
>  3 files changed, 96 insertions(+), 12 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..49d5019ab93
> --- /dev/null
> +++ b/gcc/testsuite/gcc.dg/tree-ssa/ssa-sink-20.c
> @@ -0,0 +1,15 @@
> +/* { dg-do compile } */
> +/* { dg-options "-O2 -fdump-tree-optimized -fdump-tree-sink-stats" } */

You don't test the optimized dump, do you?

> +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 {l_12\s+=\s+_4\s+\+\s+f_11\(D\);\n\s+bar\s+\(\)} sink1 } } */
> 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..84e7938c54f
> --- /dev/null
> +++ b/gcc/testsuite/gcc.dg/tree-ssa/ssa-sink-21.c
> @@ -0,0 +1,19 @@
> +/* { dg-do compile } */
> +/* { dg-options "-O2 -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 {l_13\s+=\s+_4\s+\+\s+f_12\(D\);\n\s+bar\s+\(\)} sink1 } } */
> diff --git a/gcc/tree-ssa-sink.cc b/gcc/tree-ssa-sink.cc
> index b1ba7a2ad6c..ee8988bbb2c 100644
> --- a/gcc/tree-ssa-sink.cc
> +++ b/gcc/tree-ssa-sink.cc
> @@ -171,9 +171,28 @@ nearest_common_dominator_of_uses (def_operand_p def_p, bool *debug_stmts)
>    return commondom;
>  }
>  
> +/* Return TRUE if immediate uses of the defs in
> +   STMT occur in the same block as STMT, FALSE otherwise.  */
> +
> +bool
> +def_use_same_block (gimple *stmt)

Looks like this function should be static.

> +{
> +  def_operand_p def;
> +  ssa_op_iter iter;
> +
> +  FOR_EACH_SSA_DEF_OPERAND (def, stmt, iter, SSA_OP_DEF)
> +    {
> +      gimple *def_stmt = SSA_NAME_DEF_STMT (DEF_FROM_PTR (def));
> +      if ((gimple_bb (def_stmt) == gimple_bb (stmt)))
> +	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.
> +   statements. The best basic block should be in immediate dominator of
> +   best basic block if the use stmt is after the call.

s/in/an/ ?

>  
>     We want the most control dependent block in the shallowest loop nest.
>  
> @@ -190,7 +209,8 @@ 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)
>  {
>    basic_block best_bb = late_bb;
>    basic_block temp_bb = late_bb;
> @@ -230,14 +250,46 @@ select_best_block (basic_block early_bb,
>        if (threshold > 100)
>  	threshold = 100;
>      }
> -

superfluous whitespace change?

>    /* 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)

Excess braces in the two lines above and...

> +	  && !is_gimple_call (stmt)
> +	  && gsi_end_p (gsi_start_phis (new_best_bb))
> +	  && (gimple_bb (use) != early_bb)

...here

> +	  && !is_gimple_call (use)
> +	  && dominated_by_p (CDI_POST_DOMINATORS, new_best_bb, gimple_bb(use))

Missing space before the open parens: gimple_bb (use)

> +	  && dominated_by_p (CDI_DOMINATORS, new_best_bb, early_bb))
> +	{
> +	  if (def_use_same_block (use))
> +	    return best_bb;
> +
> +	  return new_best_bb;
> +	}
> +	return best_bb;
> +    }
>  

Many returns.
I'd have said
	  && !def_use_same_block (use)
	return new_best_bb;
      else
	return best_bb;

and rephrase the comment above list of Things to consider accordingly.

>    /* No better block found, so return EARLY_BB, which happens to be the
>       statement's original block.  */
> @@ -382,7 +434,7 @@ statement_sink_location (gimple *stmt, basic_block frombb,
>  	{
>  	  /* Do not sink loads from hard registers.  */
>  	  if (gimple_assign_single_p (stmt)
> -	      && VAR_P (gimple_assign_rhs1 (stmt))
> +	      && TREE_CODE (gimple_assign_rhs1 (stmt)) == VAR_DECL

Why? Please don't do this.
thanks,

>  	      && DECL_HARD_REGISTER (gimple_assign_rhs1 (stmt)))
>  	    return false;
>  
> @@ -439,7 +491,7 @@ 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 +508,17 @@ 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);
> +	  *togsi = gsi_after_labels (sinkbb);
>  
>  	  return true;
>  	}
> @@ -480,7 +530,7 @@ 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;
>
  

This is not a review of the patch itself, but...

On 5/31/23 2:01 AM, Ajit Agarwal wrote:
> 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.

I think the wording of your git log comment could be improved a little.
How about something like the following?

Currently, code sinking will sink code after function calls.  This increases
register pressure for callee-saved registers.  The following patch improves
code sinking by placing the sunk code before calls in the use block or in
the immediate dominator of the use blocks.



> gcc/ChangeLog:
> 
> 	* tree-ssa-sink.cc (statement_sink_location): Move statements before
> 	calls.
> 	(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.

Please don't forget to add "PR tree-optimization/81953" to both sections
of the ChangeLog entries.

Peter