Re: State of AutoFDO in GCC

[Xinliang David Li <davidxl@google.com>]

On Mon, Apr 26, 2021 at 8:12 AM Andi Kleen <ak@linux.intel.com> wrote:

> Jan Hubicka <hubicka@ucw.cz> writes:
> >
> > Is there a way to get this working w/o using older perf?
>
> It's usually rather simple to fix up autofdo for new perf.
> I did it before here
>
> https://github.com/andikleen/autofdo/commits/perf4-3
>
> I think it would work always if it just ignored unknown records
> (which is quite possible). perf adds new records now and then,
> but they can be usually ignored by old tools.
> Only the assert happy coding style in autofdo prevents it.
>
> BTW longer term my feeling is autofdo should be replaced with something
> like https://lists.llvm.org/pipermail/llvm-dev/2020-August/144101.html
> I think that would fix most of the weirdnesses in the current autofdo
> implementation.
>

There are multiple directional changes in this new tool:
1) it uses perf-script trace output (in text) as input profile data;
2) it uses pseudo probe like instrumentation FDO to do profile matching;
3) it supports full profile context sensitivity (not just from inline
stacks).

#1 is most relevant to this discussion -- using text format reduces the
pain introduced by perf binary format change.

David



>
> -Andi
>