Hi,
the PR tree-opt/86020 track regression from patch I committed in
2017-05-22. I believe I did so accidentally becuase the patch does not
make much sense to me - the badness is documented as:
/* When profile is available. Compute badness as:
time_saved * caller_count
goodness = -------------------------------------------------
growth_of_caller * overall_growth * combined_size
badness = - goodness
Again use negative value to make calls with profile appear hotter
then calls without.
*/
And the code computes combined_size while I have chnaged it to be
inlined_time. I see no reason for penalizing functions running for long
time and I believed I meant it as a micro optimization since I read the
calcuation as time instead of size calculatoin.
I did some benchmarking on SPEC and Firefox and while there are changes
it seems neutral overall and thus I have decided to revert the patch.
Bootstrapped/regtested x86_64-linux, comitted.
PR tree-opt/86020
Revert:
2017-05-22 Jan Hubicka <hubi...@ucw.cz>
* ipa-inline.c (edge_badness): Use inlined_time instead of
inline_summaries->get.
Index: ipa-inline.c
===================================================================
--- ipa-inline.c (revision 267610)
+++ ipa-inline.c (working copy)
@@ -1173,7 +1173,7 @@
overall_growth += 256 * 256 - 256;
denominator *= overall_growth;
}
- denominator *= inlined_time;
+ denominator *= ipa_fn_summaries->get (caller)->self_size + growth;
badness = - numerator / denominator;
