Re: State of AutoFDO in GCC

[Jan Hubicka <hubicka@ucw.cz>]

> On Fri, Apr 23, 2021 at 12:18 AM Martin Liška <mliska@suse.cz> wrote:
> 
> > On 4/23/21 9:00 AM, Richard Biener via Gcc wrote:
> > > On Fri, Apr 23, 2021 at 7:28 AM Xinliang David Li via Gcc
> > > <gcc@gcc.gnu.org> wrote:
> > >>
> > >> Hi, the create_gcov tool was probably removed with the assumption that
> > it
> > >> was only used with Google GCC branch, but it is actually used with GCC
> > >> trunk as well.
> > >>
> > >> Given that, the tool will be restored in the github repo. It seems to
> > build
> > >> and work fine with the regression test.
> > >>
> > >> The tool may ust work as it is right now, but there is no guarantee it
> > >> won't break in the future unless someone in the GCC community tries to
> > >> maintain it.
> >
> > Hi.
> >
> > The current situation is that AutoFDO doesn't work with pretty simple
> > test-cases
> > we have in testsuite:
> >
> > https://gcc.gnu.org/bugzilla/show_bug.cgi?id=71672
> > https://gcc.gnu.org/bugzilla/show_bug.cgi?id=81379
> >
> > These are ~5 years old and nothing has happened.

Basic functionality is to identify hot parts of programs and give basic
idea of branch probabilities.  Bugs above are about other optimizations
(indirect inlining, hot-cold partitioning and peeling). First one ought
to be working but still indirect call profiling accounts for relatively
minor portion of FDO benefits, so auto-FDO should be usefable even w/o
these working if tooling was fixed.  Expecting hot-cold partitioning and
peeling to work reliably for testcases designed for normalFDO is IMO bit
unrealistic.

Perf is not limited to x86, so we should get basic usability for all
major architectures.  David, what CPUs are supported with auto-FDO on
LLVM side?
> 
> This is a major feature in Clang. Deprecating it in GCC will be unfortunate.

I also like overall idea of auto-FDO just never had much time to get it
into shape.  Every stage1 I am trying to fix something that was broken
for a while (like ICF last stage1) and fix it, so option would be to
focus on autoFDO this stage1 even though I have quite few other things
in TODO list.  I never looked into the tools translating perf data to
gcc compatible format and how easy would be to keep this alive.

David, how does perf->LLVM path work these days?

Honza
> 
> David
> 
> >
> > Thoughts?
> > Martin
> >
> > >
> > > Having the tool third-party makes keeping the whole chain working more
> > > difficult.
> > >
> > > Richard.
> > >
> > >> Thanks,
> > >>
> > >> David
> > >>
> > >> On Thu, Apr 22, 2021 at 3:29 PM Jan Hubicka <hubicka@ucw.cz> wrote:
> > >>
> > >>>> On 4/22/21 9:58 PM, Eugene Rozenfeld via Gcc wrote:
> > >>>>> GCC documentation for AutoFDO points to create_gcov tool that
> > converts
> > >>> perf.data file into gcov format that can be consumed by gcc with
> > >>> -fauto-profile (
> > https://gcc.gnu.org/onlinedocs/gcc/Optimize-Options.html,
> > >>> https://gcc.gnu.org/wiki/AutoFDO/Tutorial).
> > >>>>>
> > >>>>> I noticed that the source code for create_gcov has been deleted from
> > >>> https://github.com/google/autofdo on April 7. I asked about that
> > change
> > >>> in that repo and got the following reply:
> > >>>>>
> > >>>>> https://github.com/google/autofdo/pull/107#issuecomment-819108738
> > >>>>>
> > >>>>> "Actually we didn't use create_gcov and havn't updated create_gcov
> > for
> > >>> years, and we also didn't have enough tests to guarantee it works (It
> > was
> > >>> gcc-4.8 when we used and verified create_gcov). If you need it, it is
> > >>> welcomed to update create_gcov and add it to the respository."
> > >>>>>
> > >>>>> Does this mean that AutoFDO is currently dead in gcc?
> > >>>>
> > >>>> Hello.
> > >>>>
> > >>>> Yes. I know that even basic test cases have been broken for years in
> > the
> > >>> GCC.
> > >>>> It's new to me that create_gcov was removed.
> > >>>>
> > >>>> I tend to send patch to GCC that will remove AutoFDO from GCC.
> > >>>> I known Bin spent some time working on AutoFDO, has he came up to
> > >>> something?
> > >>>
> > >>> The GCC side of auto-FDO is not that hard.  We have most of
> > >>> infrastructure in place, but stopping point for me was always
> > difficulty
> > >>> to get gcov-tool working.  If some maintainer steps up, I think I can
> > >>> fix GCC side.
> > >>>
> > >>> I am bit unsure how important feature it is - we have FDO that works
> > >>> quite well for most users but I know there are some users of the LLVM
> > >>> implementation and there is potential to tie this with other hardware
> > >>> events to asist i.e. if conversion (where one wants to know how well
> > CPU
> > >>> predicts the jump rather than just the jump probability) which I always
> > >>> found potentially interesting.
> > >>>
> > >>> Honza
> > >>>>
> > >>>> Martin
> > >>>>
> > >>>>>
> > >>>>> Thanks,
> > >>>>>
> > >>>>> Eugene
> > >>>>>
> > >>>>
> > >>>
> >
> >