Hi, Honza, in addition to the code size problems, there are several runtime regression for the SPEC: (If I read the table correctly, if not, let me know)
SPEC/SPEC2006/INT/483.xalancbmk <https://lnt.opensuse.org/db_default/v4/SPEC/graph?plot.0=183.290.0> 146.131 4.89% SPEC/SPEC2006/FP/436.cactusADM <https://lnt.opensuse.org/db_default/v4/SPEC/graph?plot.0=171.100.0> 130.967 8.07% SPEC/SPEC2006/FP/435.gromacs <https://lnt.opensuse.org/db_default/v4/SPEC/graph?plot.0=177.90.0> 182.555 11.73% SPEC/SPEC2017/INT/541.leela_r <https://lnt.opensuse.org/db_default/v4/SPEC/graph?plot.0=174.397.0> 452.333 4.17% SPEC/SPEC2017/INT/520.omnetpp_r <https://lnt.opensuse.org/db_default/v4/SPEC/graph?plot.0=172.357.0> 395.582 4.98% do we have plan to study and fix these run-time regression? thanks. Qing > On Jan 12, 2019, at 12:32 PM, Jan Hubicka <hubi...@ucw.cz> wrote: > > Hello, > this patch sets inline-unit-growth to 40. The performance changes are > - Firefox, LTO > > https://treeherder.mozilla.org/perf.html#/compare?originalProject=try&originalRevision=f7bd026e1a931b9a284d1c85c2577a72dd592820&newProject=try&newRevision=74889968abcc688b8d161863566ed273c0401ee4&framework=1&filter=opt&showOnlyComparable=1&showOnlyImportant=1 > After fixes to inlining priorities this makes difference without > profile feedback only. > > Code size growth is about 9.15% with LTO and 3.95 with LTO and profile > feedback. > - Firefox noLTO > > https://treeherder.mozilla.org/perf.html#/compare?originalProject=try&originalRevision=c902b72340a3dca3114f58578c1c8f3e6a1cd89c&newProject=try&newRevision=4974da6f92c144a9c09765b56a564a640069ddb9&framework=1&showOnlyComparable=1&showOnlyImportant=1 > With about 7% code size growth > - SPEC > > https://lnt.opensuse.org/db_default/v4/CPP/latest_runs_report?num_runs=10&min_percentage_change=0.02&revisions=46e2bd1143b5c60af814916d7673879b34ceb3f6%2Cc0d79cfe9c4ec30823480f2f9b256600e8e3899f > - C++ benchmarks > > https://lnt.opensuse.org/db_default/v4/SPEC/latest_runs_report?num_runs=10&all_changes=on&min_percentage_change=0.02&revisions=46e2bd1143b5c60af814916d7673879b34ceb3f6%2Cc0d79cfe9c4ec30823480f2f9b256600e8e3899f > > I am not entirely happy about the code-size/performance tradeoffs but it > is concerned only for programs built with -O3 or having too many inline > keywords. I have looked into inlining decisions for Firefox, HHVM and > Clang and inliner gets out of growt bounds way too early and some of > more performance aware projects already sets the limit up. > > I will tune other metrics down to handle some of the code size problems. > > Honza > > Index: ChangeLog > =================================================================== > --- ChangeLog (revision 267882) > +++ ChangeLog (working copy) > @@ -1,3 +1,7 @@ > +2019-01-05 Jan Hubicka <hubi...@ucw.cz> > + > + * params.def (inline-unit-growth): Set to 40. > + > 2019-01-12 Jakub Jelinek <ja...@redhat.com> > > * tree-ssa-loop-ivopts.c (find_inv_vars): Fix a comment typo. > Index: params.def > =================================================================== > --- params.def (revision 267882) > +++ params.def (working copy) > @@ -227,7 +227,7 @@ DEFPARAM(PARAM_LARGE_UNIT_INSNS, > DEFPARAM(PARAM_INLINE_UNIT_GROWTH, > "inline-unit-growth", > "How much can given compilation unit grow because of the inlining (in > percent).", > - 20, 0, 0) > + 40, 0, 0) > DEFPARAM(PARAM_IPCP_UNIT_GROWTH, > "ipcp-unit-growth", > "How much can given compilation unit grow because of the > interprocedural constant propagation (in percent).",
