Handle flattening in early inliner

Handle flattening in early inliner

[Jan Hubicka]

Hi,
this patch makes early inliner to handle flattening (that is flatten as much as
it can leaving rest of work for late inliner).  This helps to get more realistic
profiles on flattended nodes and also more effective profiling and profile feedback.

Bootstrapped/regtested x86_64-linux, comitted.

	* ipa-inline.c (cgraph_early_inlining): Handle flattening too.
Index: ipa-inline.c
===================================================================
--- ipa-inline.c	(revision 158462)
+++ ipa-inline.c	(working copy)
@@ -1667,6 +1667,17 @@ cgraph_early_inlining (void)
     }
   else
     {
+      if (lookup_attribute ("flatten",
+			    DECL_ATTRIBUTES (node->decl)) != NULL)
+	{
+	  if (dump_file)
+	    fprintf (dump_file,
+		     "Flattening %s\n", cgraph_node_name (node));
+	  cgraph_flatten (node);
+	  timevar_push (TV_INTEGRATION);
+	  todo |= optimize_inline_calls (current_function_decl);
+	  timevar_pop (TV_INTEGRATION);
+	}
       /* We iterate incremental inlining to get trivial cases of indirect
 	 inlining.  */
       while (iterations < PARAM_VALUE (PARAM_EARLY_INLINER_MAX_ITERATIONS)

Re: Handle flattening in early inliner

[Richard Guenther]

On Sun, Apr 18, 2010 at 1:44 AM, Jan Hubicka <hubicka@ucw.cz> wrote:
> Hi,
> this patch makes early inliner to handle flattening (that is flatten as much as
> it can leaving rest of work for late inliner).  This helps to get more realistic
> profiles on flattended nodes and also more effective profiling and profile feedback.
>
> Bootstrapped/regtested x86_64-linux, comitted.
>
>        * ipa-inline.c (cgraph_early_inlining): Handle flattening too.
> Index: ipa-inline.c
> ===================================================================
> --- ipa-inline.c        (revision 158462)
> +++ ipa-inline.c        (working copy)
> @@ -1667,6 +1667,17 @@ cgraph_early_inlining (void)
>     }
>   else
>     {
> +      if (lookup_attribute ("flatten",
> +                           DECL_ATTRIBUTES (node->decl)) != NULL)
> +       {
> +         if (dump_file)
> +           fprintf (dump_file,
> +                    "Flattening %s\n", cgraph_node_name (node));
> +         cgraph_flatten (node);
> +         timevar_push (TV_INTEGRATION);
> +         todo |= optimize_inline_calls (current_function_decl);
> +         timevar_pop (TV_INTEGRATION);
> +       }

you should be able to skip the following code for non-flatten
functions.  Thus place it in else { .. }.

Richard.

>       /* We iterate incremental inlining to get trivial cases of indirect
>         inlining.  */
>       while (iterations < PARAM_VALUE (PARAM_EARLY_INLINER_MAX_ITERATIONS)
>

Re: Handle flattening in early inliner

[Jan Hubicka]

> On Sun, Apr 18, 2010 at 1:44 AM, Jan Hubicka <hubicka@ucw.cz> wrote:
> > Hi,
> > this patch makes early inliner to handle flattening (that is flatten as much as
> > it can leaving rest of work for late inliner).  This helps to get more realistic
> > profiles on flattended nodes and also more effective profiling and profile feedback.
> >
> > Bootstrapped/regtested x86_64-linux, comitted.
> >
> >        * ipa-inline.c (cgraph_early_inlining): Handle flattening too.
> > Index: ipa-inline.c
> > ===================================================================
> > --- ipa-inline.c        (revision 158462)
> > +++ ipa-inline.c        (working copy)
> > @@ -1667,6 +1667,17 @@ cgraph_early_inlining (void)
> >     }
> >   else
> >     {
> > +      if (lookup_attribute ("flatten",
> > +                           DECL_ATTRIBUTES (node->decl)) != NULL)
> > +       {
> > +         if (dump_file)
> > +           fprintf (dump_file,
> > +                    "Flattening %s\n", cgraph_node_name (node));
> > +         cgraph_flatten (node);
> > +         timevar_push (TV_INTEGRATION);
> > +         todo |= optimize_inline_calls (current_function_decl);
> > +         timevar_pop (TV_INTEGRATION);
> > +       }
> 
> you should be able to skip the following code for non-flatten
> functions.  Thus place it in else { .. }.

The following code handle inlining of indirect calls via iteration, so I
decided to keep it so "flatten" decorated functions won't end up with less
inlining than non-flatten decorated ones (i.e. I am trying those attributes to
monotonously increase amount of inlining in general as that is what they are
supposed to do).  But it is not big deal.

Honza
> 
> Richard.
> 
> >       /* We iterate incremental inlining to get trivial cases of indirect
> >         inlining.  */
> >       while (iterations < PARAM_VALUE (PARAM_EARLY_INLINER_MAX_ITERATIONS)
> >

Re: Handle flattening in early inliner

[Richard Guenther]

2010/4/18 Jan Hubicka <hubicka@ucw.cz>:
>> On Sun, Apr 18, 2010 at 1:44 AM, Jan Hubicka <hubicka@ucw.cz> wrote:
>> > Hi,
>> > this patch makes early inliner to handle flattening (that is flatten as much as
>> > it can leaving rest of work for late inliner).  This helps to get more realistic
>> > profiles on flattended nodes and also more effective profiling and profile feedback.
>> >
>> > Bootstrapped/regtested x86_64-linux, comitted.
>> >
>> >        * ipa-inline.c (cgraph_early_inlining): Handle flattening too.
>> > Index: ipa-inline.c
>> > ===================================================================
>> > --- ipa-inline.c        (revision 158462)
>> > +++ ipa-inline.c        (working copy)
>> > @@ -1667,6 +1667,17 @@ cgraph_early_inlining (void)
>> >     }
>> >   else
>> >     {
>> > +      if (lookup_attribute ("flatten",
>> > +                           DECL_ATTRIBUTES (node->decl)) != NULL)
>> > +       {
>> > +         if (dump_file)
>> > +           fprintf (dump_file,
>> > +                    "Flattening %s\n", cgraph_node_name (node));
>> > +         cgraph_flatten (node);
>> > +         timevar_push (TV_INTEGRATION);
>> > +         todo |= optimize_inline_calls (current_function_decl);
>> > +         timevar_pop (TV_INTEGRATION);
>> > +       }
>>
>> you should be able to skip the following code for non-flatten
>> functions.  Thus place it in else { .. }.
>
> The following code handle inlining of indirect calls via iteration, so I
> decided to keep it so "flatten" decorated functions won't end up with less
> inlining than non-flatten decorated ones (i.e. I am trying those attributes to
> monotonously increase amount of inlining in general as that is what they are
> supposed to do).  But it is not big deal.

Ah, ok - but then you should technically increase iteration by one ;)

Well, it is not big deal as you say.

Richard.

> Honza
>>
>> Richard.
>>
>> >       /* We iterate incremental inlining to get trivial cases of indirect
>> >         inlining.  */
>> >       while (iterations < PARAM_VALUE (PARAM_EARLY_INLINER_MAX_ITERATIONS)
>> >
>

Re: Handle flattening in early inliner

[H.J. Lu]

On Sat, Apr 17, 2010 at 4:44 PM, Jan Hubicka <hubicka@ucw.cz> wrote:
> Hi,
> this patch makes early inliner to handle flattening (that is flatten as much as
> it can leaving rest of work for late inliner).  This helps to get more realistic
> profiles on flattended nodes and also more effective profiling and profile feedback.
>
> Bootstrapped/regtested x86_64-linux, comitted.
>
>        * ipa-inline.c (cgraph_early_inlining): Handle flattening too.

This caused:

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=44964

-- 
H.J.