Inliner badness metric tweak

Inliner badness metric tweak

[Jan Hubicka]

Hi,
this patch makes inline metric to be
       estimate_speedup
  ---------------------------- (over)
  edge_growth^2*overall_growth

This is intended to make stronger push on inliner to preffer functions with
smaller growth and do not take overall growth (that is expected unit growth
after inlining all calls to a give function) so seriously.

This solve code size regression for tramp3d and also improve DLV. I tested it
on firefox with resonable results and on subset of SPEC that caused inliner
issues in past.

Lacking access to SUSE's periodic testers, I intend to commit it today and
watch them for possible regressions.

Bootstrapped/regtested x86_64-linux.

	PR ipa/65076
	* ipa-inline.c (edge_badness): Base denominator on callee's
	grwoth squared.

Index: ipa-inline.c
===================================================================
--- ipa-inline.c	(revision 221757)
+++ ipa-inline.c	(working copy)
@@ -1099,7 +1099,7 @@ edge_badness (struct cgraph_edge *edge,
 	numerator = numerator >> 11;
       denominator = growth;
       if (callee_info->growth > 0)
-	denominator *= callee_info->growth;
+	denominator *= callee_info->growth * callee_info->growth;
 
       badness = - numerator / denominator;
 

Re: Inliner badness metric tweak

[Richard Biener]

On Mon, Mar 30, 2015 at 3:45 AM, Jan Hubicka <hubicka@ucw.cz> wrote:
> Hi,
> this patch makes inline metric to be
>        estimate_speedup
>   ---------------------------- (over)
>   edge_growth^2*overall_growth
>
> This is intended to make stronger push on inliner to preffer functions with
> smaller growth and do not take overall growth (that is expected unit growth
> after inlining all calls to a give function) so seriously.
>
> This solve code size regression for tramp3d and also improve DLV. I tested it
> on firefox with resonable results and on subset of SPEC that caused inliner
> issues in past.
>
> Lacking access to SUSE's periodic testers, I intend to commit it today and
> watch them for possible regressions.
>
> Bootstrapped/regtested x86_64-linux.
>
>         PR ipa/65076
>         * ipa-inline.c (edge_badness): Base denominator on callee's
>         grwoth squared.
>
> Index: ipa-inline.c
> ===================================================================
> --- ipa-inline.c        (revision 221757)
> +++ ipa-inline.c        (working copy)
> @@ -1099,7 +1099,7 @@ edge_badness (struct cgraph_edge *edge,
>         numerator = numerator >> 11;
>        denominator = growth;
>        if (callee_info->growth > 0)
> -       denominator *= callee_info->growth;
> +       denominator *= callee_info->growth * callee_info->growth;

For sure a comment explaining this would be good ...

>        badness = - numerator / denominator;
>

Re: Inliner badness metric tweak

[H.J. Lu]

On Sun, Mar 29, 2015 at 6:45 PM, Jan Hubicka <hubicka@ucw.cz> wrote:
> Hi,
> this patch makes inline metric to be
>        estimate_speedup
>   ---------------------------- (over)
>   edge_growth^2*overall_growth
>
> This is intended to make stronger push on inliner to preffer functions with
> smaller growth and do not take overall growth (that is expected unit growth
> after inlining all calls to a give function) so seriously.
>
> This solve code size regression for tramp3d and also improve DLV. I tested it
> on firefox with resonable results and on subset of SPEC that caused inliner
> issues in past.
>
> Lacking access to SUSE's periodic testers, I intend to commit it today and
> watch them for possible regressions.
>
> Bootstrapped/regtested x86_64-linux.
>
>         PR ipa/65076
>         * ipa-inline.c (edge_badness): Base denominator on callee's
>         grwoth squared.

This caused:

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=65654

-- 
H.J.

Re: Inliner badness metric tweak

[Jan Hubicka]

> 
> This caused:
> 
> https://gcc.gnu.org/bugzilla/show_bug.cgi?id=65654

I will look into it; the assert checks that the expected function growth
match the real growth after inlining, so it is a semi-latent bug hit by this
patch.

Honza
> 
> -- 
> H.J.