[PATCH] Add Optimization for various IPA parameters.

Martin Jambor mjambor@suse.cz
Wed Jan 8 21:42:00 GMT 2020


Hi,

On Wed, Jan 08 2020, Jan Hubicka wrote:
>> Hi,
>> 
>> On Fri, Jan 03 2020, Martin Liška wrote:
>> > Hi.
>> >
>> > This is similar transformation for IPA passes. This time,
>> > one needs to use opt_for_fn in order to get the right
>> > parameter values.
>> >
>> > @Martin, Honza:
>> > There are last few remaining parameters which should use
>> > opt_for_fn:
>> >
>> > param_ipa_sra_max_replacements
>> 
>> IPA-CP: Always access param_ipa_sra_max_replacements through opt_for_fn
>> 
>> 2020-01-07  Martin Jambor  <mjambor@suse.cz>
>> 
>> 	* params.opt (param_ipa_sra_max_replacements): Mark as Optimization.
>> 	* ipa-sra.c (scanned_node): New variable.
>> 	(allocate_access): Use it to get param_ipa_sra_max_replacements.
>> 	(ipa_sra_summarize_function): Set up scanned_node.
>> 	(pull_accesses_from_callee): New parameter caller, use it to get
>> 	param_ipa_sra_max_replacements.
>> 	(param_splitting_across_edge): Pass the caller to
>> 	pull_accesses_from_callee.
>> ---
>>  gcc/ipa-sra.c  | 33 +++++++++++++++++++++------------
>>  gcc/params.opt |  2 +-
>>  2 files changed, 22 insertions(+), 13 deletions(-)
>> 
>> diff --git a/gcc/ipa-sra.c b/gcc/ipa-sra.c
>> index a051a9f2154..133ed687509 100644
>> --- a/gcc/ipa-sra.c
>> +++ b/gcc/ipa-sra.c
>> @@ -521,6 +521,10 @@ ipa_sra_call_summaries::duplicate (cgraph_edge *, cgraph_edge *,
>>  
>>  /* With all GTY stuff done, we can move to anonymous namespace.  */
>>  namespace {
>> +/* Functions which currently has its body analyzed.  */
>> +
>> +cgraph_node *scanned_node;
>> +
>>  /* Quick mapping from a decl to its param descriptor.  */
>>  
>>  hash_map<tree, gensum_param_desc *> *decl2desc;
>> @@ -1265,7 +1269,8 @@ allocate_access (gensum_param_desc *desc,
>>  		 HOST_WIDE_INT offset, HOST_WIDE_INT size)
>>  {
>>    if (desc->access_count
>> -      == (unsigned) param_ipa_sra_max_replacements)
>> +      == (unsigned) opt_for_fn (scanned_node->decl,
>> +				param_ipa_sra_max_replacements))
>>      {
>>        disqualify_split_candidate (desc, "Too many replacement candidates");
>>        return NULL;
>> @@ -2472,6 +2477,7 @@ ipa_sra_summarize_function (cgraph_node *node)
>>  	     node->order);
>>    if (!ipa_sra_preliminary_function_checks (node))
>>      return;
>> +  scanned_node = node;
>>    gcc_obstack_init (&gensum_obstack);
>>    isra_func_summary *ifs = func_sums->get_create (node);
>>    ifs->m_candidate = true;
>> @@ -2526,6 +2532,7 @@ ipa_sra_summarize_function (cgraph_node *node)
>>    delete decl2desc;
>>    decl2desc = NULL;
>>    obstack_free (&gensum_obstack, NULL);
>> +  scanned_node = NULL;
>
> It is your code.  having static var to track currently analyzed function
> is bit ugly, and I am not sure if you don't have current_function_decl
> set to that function in all cases.  But I will leave this to your
> decision.

It is my code but apparently I already forgot some changes to it.
Originally I only pushed/popped cfun when processing (some) scan results
(I originally hoped not to do it at all) but since May I actually push
it whenever the body of the function is scanned, so using cfun is indeed
OK.

So please use the following instead:

IPA-CP: Access param_ipa_sra_max_replacements through opt_for_fn

2020-01-08  Martin Jambor  <mjambor@suse.cz>

	* params.opt (param_ipa_sra_max_replacements): Mark as Optimization.
	* ipa-sra.c (pull_accesses_from_callee): New parameter caller, use it
	to get param_ipa_sra_max_replacements.
	(param_splitting_across_edge): Pass the caller to
	pull_accesses_from_callee.
---
 gcc/ipa-sra.c  | 24 +++++++++++++-----------
 gcc/params.opt |  2 +-
 2 files changed, 14 insertions(+), 12 deletions(-)

diff --git a/gcc/ipa-sra.c b/gcc/ipa-sra.c
index a051a9f2154..51d225ed772 100644
--- a/gcc/ipa-sra.c
+++ b/gcc/ipa-sra.c
@@ -3246,16 +3246,17 @@ all_callee_accesses_present_p (isra_param_desc *param_desc,
 enum acc_prop_kind {ACC_PROP_DONT, ACC_PROP_COPY, ACC_PROP_CERTAIN};
 
 
-/* Attempt to propagate all definite accesses from ARG_DESC to PARAM_DESC, if
-   they would not violate some constraint there.  If successful, return NULL,
-   otherwise return the string reason for failure (which can be written to the
-   dump file).  DELTA_OFFSET is the known offset of the actual argument withing
-   the formal parameter (so of ARG_DESCS within PARAM_DESCS), ARG_SIZE is the
-   size of the actual argument or zero, if not known. In case of success, set
-   *CHANGE_P to true if propagation actually changed anything.  */
+/* Attempt to propagate all definite accesses from ARG_DESC to PARAM_DESC,
+   (which belongs to CALLER) if they would not violate some constraint there.
+   If successful, return NULL, otherwise return the string reason for failure
+   (which can be written to the dump file).  DELTA_OFFSET is the known offset
+   of the actual argument withing the formal parameter (so of ARG_DESCS within
+   PARAM_DESCS), ARG_SIZE is the size of the actual argument or zero, if not
+   known. In case of success, set *CHANGE_P to true if propagation actually
+   changed anything.  */
 
 static const char *
-pull_accesses_from_callee (isra_param_desc *param_desc,
+pull_accesses_from_callee (cgraph_node *caller, isra_param_desc *param_desc,
 			   isra_param_desc *arg_desc,
 			   unsigned delta_offset, unsigned arg_size,
 			   bool *change_p)
@@ -3335,7 +3336,7 @@ pull_accesses_from_callee (isra_param_desc *param_desc,
       return NULL;
 
     if ((prop_count + pclen
-	 > (unsigned) param_ipa_sra_max_replacements)
+	 > (unsigned) opt_for_fn (caller->decl, param_ipa_sra_max_replacements))
 	|| size_would_violate_limit_p (param_desc,
 				       param_desc->size_reached + prop_size))
       return "propagating accesses would violate the count or size limit";
@@ -3455,7 +3456,8 @@ param_splitting_across_edge (cgraph_edge *cs)
 	  else
 	    {
 	      const char *pull_failure
-		= pull_accesses_from_callee (param_desc, arg_desc, 0, 0, &res);
+		= pull_accesses_from_callee (cs->caller, param_desc, arg_desc,
+					     0, 0, &res);
 	      if (pull_failure)
 		{
 		  if (dump_file && (dump_flags & TDF_DETAILS))
@@ -3516,7 +3518,7 @@ param_splitting_across_edge (cgraph_edge *cs)
 	  else
 	    {
 	      const char *pull_failure
-		= pull_accesses_from_callee (param_desc, arg_desc,
+		= pull_accesses_from_callee (cs->caller, param_desc, arg_desc,
 					     ipf->unit_offset,
 					     ipf->unit_size, &res);
 	      if (pull_failure)
diff --git a/gcc/params.opt b/gcc/params.opt
index 3cd6d7d481d..02584740640 100644
--- a/gcc/params.opt
+++ b/gcc/params.opt
@@ -235,7 +235,7 @@ Common Joined UInteger Var(param_ipa_max_switch_predicate_bounds) Init(5) Param
 Maximal number of boundary endpoints of case ranges of switch statement used during IPA functoin summary generation.
 
 -param=ipa-sra-max-replacements=
-Common Joined UInteger Var(param_ipa_sra_max_replacements) Init(8) IntegerRange(0, 16) Param
+Common Joined UInteger Var(param_ipa_sra_max_replacements) Optimization Init(8) IntegerRange(0, 16) Param
 Maximum pieces that IPA-SRA tracks per formal parameter, as a consequence, also the maximum number of replacements of a formal parameter.
 
 -param=ipa-sra-ptr-growth-factor=
-- 
2.24.1



More information about the Gcc-patches mailing list