[Bug gcov-profile/112806] New: Profile-Guided Optimization (PGO) policy regarding explicit user optimization hint behavior

[Bug gcov-profile/112806] New: Profile-Guided Optimization (PGO) policy regarding explicit user optimization hint behavior

[zamazan4ik at tut dot by]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=112806

            Bug ID: 112806
           Summary: Profile-Guided Optimization (PGO) policy regarding
                    explicit user optimization hint behavior
           Product: gcc
           Version: unknown
            Status: UNCONFIRMED
          Severity: normal
          Priority: P3
         Component: gcov-profile
          Assignee: unassigned at gcc dot gnu.org
          Reporter: zamazan4ik at tut dot by
                CC: marxin at gcc dot gnu.org
  Target Milestone: ---

Hi!

Right now there is no documentation for the behavior of PGO in the presence of
user-provided optimization hints. It would be good to document how GCC's PGO
implementation handles conflicts between the hints the user explicitly wrote in
their code and the profile they're optimizing against. For example, how do the
following (non-exhaustive) hints behave: [[likely]] and other optimization
attributes, register or inline use, __builtin_expect, etc.

This will help to set user expectations - some users expect explicit
optimization hints to be ignored by PGO while other users assume the hints are
still honored (at least sometimes). If hints are honored, they should be
clearly documented as doing so if that behavior is intentional.

Similar issue in LLVM: https://github.com/llvm/llvm-project/issues/58189

[Bug gcov-profile/112806] Profile-Guided Optimization (PGO) policy regarding explicit user optimization hint behavior

[pinskia at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=112806

--- Comment #1 from Andrew Pinski <pinskia at gcc dot gnu.org> ---
[[likely]]/__builtin_expect is kinda of documented:

https://gcc.gnu.org/onlinedocs/gcc-13.2.0/gcc/Other-Builtins.html#index-fprofile-arcs-1

Everything else is just a hint.

[Bug gcov-profile/112806] Profile-Guided Optimization (PGO) policy regarding explicit user optimization hint behavior

[pinskia at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=112806

--- Comment #2 from Andrew Pinski <pinskia at gcc dot gnu.org> ---
(In reply to Andrew Pinski from comment #1)
> [[likely]]/__builtin_expect is kinda of documented:
> 
> https://gcc.gnu.org/onlinedocs/gcc-13.2.0/gcc/Other-Builtins.html#index-
> fprofile-arcs-1
> 
> Everything else is just a hint.

I should say a hint and mostly ignored.

[Bug gcov-profile/112806] Profile-Guided Optimization (PGO) policy regarding explicit user optimization hint behavior

[zamazan4ik at tut dot by]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=112806

--- Comment #3 from Alexander Zaitsev <zamazan4ik at tut dot by> ---
> https://gcc.gnu.org/onlinedocs/gcc-13.2.0/gcc/Other-Builtins.html#index-fprofile-arcs-1

I already read this and still do not understand the actual behavior. If PGO
profiles show that the branch is "cold" but a user write for this branch via
__builtin_expect/[[likely]] that the branch is "hot" - what decision will be
made by the optimizer?

On the link above there is only "In general, you should prefer to use actual
profile feedback for this (-fprofile-arcs), as programmers are notoriously bad
at predicting how their programs actually perform.". But it does not specify
the actual behavior - it's just a recommendation to use PGO instead of manual
[[likely]] hints.

[Bug gcov-profile/112806] Profile-Guided Optimization (PGO) policy regarding explicit user optimization hint behavior

[rguenth at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=112806

Richard Biener <rguenth at gcc dot gnu.org> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
                 CC|                            |hubicka at gcc dot gnu.org
            Version|unknown                     |14.0

--- Comment #4 from Richard Biener <rguenth at gcc dot gnu.org> ---
I think it depends on whether the code was executed at all during training
dependent on -fprofile-partial-training which does regular profile guessing
for untrained parts of the program.

In general the situation is complicated and it's difficult to document
behavior with a short description.