[PATCH 3/9] ipa-cp: Leave removal of unused parameters to IPA-SRA

[PATCH 3/9] ipa-cp: Leave removal of unused parameters to IPA-SRA

[Martin Jambor]

Hi,

I'm re-posting patches which I have posted at the end of stage 1 but
which have not passed review yet.

8<--------------------------------------------------------------------

Looking at some benchmarks I have noticed many cases when IPA-CP
cloned a function for all contexts just because it knew that some
parameters were not used at all.  Then IPA-SRA looked at the function
and cloned it again to split another parameter or two.  The latter
pass is better equipped to detect when parameters can be altogether
removed and so the IPA-CP cloning was for no good reason.

This patch simply alters the IPA-CP not to do that in the situations
where IPA-SRA can (for nodes which can be made local).

Bootstrapped and tested individually when I originally posted it and
now bootstrapped and LTO-bootstrapped and tested as part of the whole
series.  OK for master?


gcc/ChangeLog:

2022-11-11  Martin Jambor  <mjambor@suse.cz>

	* ipa-cp.cc (estimate_local_effects): Do not clone potentionally local
	nodes for all contexts just to remove unused parameters.
---
 gcc/ipa-cp.cc | 5 ++++-
 1 file changed, 4 insertions(+), 1 deletion(-)

diff --git a/gcc/ipa-cp.cc b/gcc/ipa-cp.cc
index cc031ebed0f..172ea353c49 100644
--- a/gcc/ipa-cp.cc
+++ b/gcc/ipa-cp.cc
@@ -3722,7 +3722,10 @@ estimate_local_effects (struct cgraph_node *node)
 						    &removable_params_cost);
   int devirt_bonus = devirtualization_time_bonus (node, &avals);
   if (always_const || devirt_bonus
-      || (removable_params_cost && node->can_change_signature))
+      || (removable_params_cost
+	  && node->can_change_signature
+	  /* If IPA-SRA can do it, it can do it better.  */
+	  && !node->can_be_local_p ()))
     {
       struct caller_statistics stats;
       ipa_call_estimates estimates;
-- 
2.38.1

Re: [PATCH 3/9] ipa-cp: Leave removal of unused parameters to IPA-SRA

[Jan Hubicka]

> Hi,
> 
> I'm re-posting patches which I have posted at the end of stage 1 but
> which have not passed review yet.
> 
> 8<--------------------------------------------------------------------
> 
> Looking at some benchmarks I have noticed many cases when IPA-CP
> cloned a function for all contexts just because it knew that some
> parameters were not used at all.  Then IPA-SRA looked at the function
> and cloned it again to split another parameter or two.  The latter
> pass is better equipped to detect when parameters can be altogether
> removed and so the IPA-CP cloning was for no good reason.
> 
> This patch simply alters the IPA-CP not to do that in the situations
> where IPA-SRA can (for nodes which can be made local).
> 
> Bootstrapped and tested individually when I originally posted it and
> now bootstrapped and LTO-bootstrapped and tested as part of the whole
> series.  OK for master?
> 
> 
> gcc/ChangeLog:
> 
> 2022-11-11  Martin Jambor  <mjambor@suse.cz>
> 
> 	* ipa-cp.cc (estimate_local_effects): Do not clone potentionally local
> 	nodes for all contexts just to remove unused parameters.
> ---
>  gcc/ipa-cp.cc | 5 ++++-
>  1 file changed, 4 insertions(+), 1 deletion(-)
> 
> diff --git a/gcc/ipa-cp.cc b/gcc/ipa-cp.cc
> index cc031ebed0f..172ea353c49 100644
> --- a/gcc/ipa-cp.cc
> +++ b/gcc/ipa-cp.cc
> @@ -3722,7 +3722,10 @@ estimate_local_effects (struct cgraph_node *node)
>  						    &removable_params_cost);
>    int devirt_bonus = devirtualization_time_bonus (node, &avals);
>    if (always_const || devirt_bonus
> -      || (removable_params_cost && node->can_change_signature))
> +      || (removable_params_cost
> +	  && node->can_change_signature
> +	  /* If IPA-SRA can do it, it can do it better.  */
> +	  && !node->can_be_local_p ()))
Perhaps we could dump that into dump file (i.e. not cloning because
ipa-sra will do it later).  That could save me from debugging session
in future :)
OK with that change.
Honza
>      {
>        struct caller_statistics stats;
>        ipa_call_estimates estimates;
> -- 
> 2.38.1
> 

Re: [PATCH 3/9] ipa-cp: Leave removal of unused parameters to IPA-SRA

[Martin Jambor]

Hi,

On Mon, Dec 12 2022, Jan Hubicka wrote:
>>
[...]
>> diff --git a/gcc/ipa-cp.cc b/gcc/ipa-cp.cc
>> index cc031ebed0f..172ea353c49 100644
>> --- a/gcc/ipa-cp.cc
>> +++ b/gcc/ipa-cp.cc
>> @@ -3722,7 +3722,10 @@ estimate_local_effects (struct cgraph_node *node)
>>  						    &removable_params_cost);
>>    int devirt_bonus = devirtualization_time_bonus (node, &avals);
>>    if (always_const || devirt_bonus
>> -      || (removable_params_cost && node->can_change_signature))
>> +      || (removable_params_cost
>> +	  && node->can_change_signature
>> +	  /* If IPA-SRA can do it, it can do it better.  */
>> +	  && !node->can_be_local_p ()))
> Perhaps we could dump that into dump file (i.e. not cloning because
> ipa-sra will do it later).  That could save me from debugging session
> in future :)
> OK with that change.
> Honza

OK, I plan to commit the following after a last round of testing.
Thanks!

Martin


Looking at some benchmarks I have noticed many cases when IPA-CP
cloned a function for all contexts just because it knew that some
parameters were not used at all.  Then IPA-SRA looked at the function
and cloned it again to split another parameter or two.  The latter
pass is better equipped to detect when parameters can be altogether
removed and so the IPA-CP cloning was for no good reason.

This patch simply alters the IPA-CP not to do that in the situations
where IPA-SRA can (for nodes which can be made local) with additional
dumping requested by Honza.

gcc/ChangeLog:

2022-12-13  Martin Jambor  <mjambor@suse.cz>

	* ipa-cp.cc (clone_for_param_removal_p): New function.
	(estimate_local_effects): Call it before considering cloning
	just to remove unused parameters.
---
 gcc/ipa-cp.cc | 25 ++++++++++++++++++++++++-
 1 file changed, 24 insertions(+), 1 deletion(-)

diff --git a/gcc/ipa-cp.cc b/gcc/ipa-cp.cc
index cc031ebed0f..300bec54bbd 100644
--- a/gcc/ipa-cp.cc
+++ b/gcc/ipa-cp.cc
@@ -3700,6 +3700,29 @@ get_max_overall_size (cgraph_node *node)
   return max_new_size;
 }
 
+/* Return true if NODE should be cloned just for a parameter removal, possibly
+   dumping a reason if not.  */
+
+static bool
+clone_for_param_removal_p (cgraph_node *node)
+{
+  if (!node->can_change_signature)
+    {
+      if (dump_file && (dump_flags & TDF_DETAILS))
+	fprintf (dump_file, "  Not considering cloning to remove parameters, "
+		 "function cannot change signature.\n");
+      return false;
+    }
+  if (node->can_be_local_p ())
+    {
+      if (dump_file && (dump_flags & TDF_DETAILS))
+	fprintf (dump_file, "  Not considering cloning to remove parameters, "
+		 "IPA-SRA can do it potentially better.\n");
+      return false;
+    }
+  return true;
+}
+
 /* Iterate over known values of parameters of NODE and estimate the local
    effects in terms of time and size they have.  */
 
@@ -3722,7 +3745,7 @@ estimate_local_effects (struct cgraph_node *node)
 						    &removable_params_cost);
   int devirt_bonus = devirtualization_time_bonus (node, &avals);
   if (always_const || devirt_bonus
-      || (removable_params_cost && node->can_change_signature))
+      || (removable_params_cost && clone_for_param_removal_p (node)))
     {
       struct caller_statistics stats;
       ipa_call_estimates estimates;
-- 
2.38.1