[llvm-dev] OptBisect implementation for new pass manager

[Chandler Carruth via llvm-dev]

Sorry I'm late to the thread (conference + vacation delayed me). I've tried
to skim the thread, but haven't found too much real conclusions to a few
points I'd like to make. If any of the below re-hashes stuff that was
already covered, my apologies and feel free to just mention by whom or what
date and I'll read more carefully.


I feel like the design of this is made unnecessarily complex and could be
simplified in a few ways. These all stem from a key aspect of bisection:
this is a *development* activity. It doesn't have to hit some specific
quality bar the way that `optnone` and -O0 (which are both exposed to
users) need to....

Some immediate simplifications:

1) I don't think we need to go out of our way to connect the IR pass
bisection (in the new PM) with codegen's IR pass bisection. We already have
two tools (`opt` and `llc` and can drive them separately IMO).

2) I also don't think we need the subtle *correctness* guarantees of
`optnone` which really and truly IMO require *passes* to make the decision
rather than some abstract pass pipeline system.

3) I think we really do want high *resolution* of bisection even if it
isn't 100% functional. Let's imagine that this skips a "necessary" pass for
some behavior. The IR will still be valid, and this step of bisection can
still be very useful for reducing crashes and assert failures.

4) I believe we can re-use the debug counter infrastructure that did not
exist when OptBisect was first introduced rather than rolling a custom
version. It's possible there are use cases this cannot handle, but it might
be worth trying to avoid inventing another thing here.

Given the above, I'd really be interested in seeing how far we can get with
a simple debug counter wired up to the new instrumentation framework, and
nothing more. Could we get that working? Can we see where that would be
genuinely insufficient for developers (as opposed to simply producing a
slightly different workflow or command sequence)?


Regarding #2 which is I think the most surprising thing... Keep in mind
that *after* we finish bisection, we can still run some minimal second set
of passes in order to generate "correct" code. Also, I believe debug
counters has the ability to disable only for a range of counts and then
re-enable. Well designed bisection test scripts should be able to preserve
and/or synthesize the necessary bits to keep IR "working" for whatever
constraints a particular reduction has.


Hope all of this makes some sense.
-Chandler

On Wed, Sep 26, 2018 at 9:54 AM Fedor Sergeev via llvm-dev <
llvm-dev at lists.llvm.org> wrote:

> Greetings!
>
> As the generic Pass Instrumentation framework for new pass manager is
> finally *in*,
> I'm glad to start the discussion on implementation of -opt-bisect
> through that framework.
>
> As it has already been discovered while porting other features (namely,
> -time-passes)
> blindly copying the currently existing legacy implementation is most
> likely not a perfect
> way forward. Now is a chance to take a fresh look at the overall
> approach and perhaps
> do better, without the restrictions that legacy pass manager framework
> imposed on
> the implementation.
>
> Kind of a summary of what we have now:
>    - There is a single OptBisect object, requested through LLVMContext
>      (managed as ManagedStatic).
>
>    - OptBisect is defined in lib/IR, but does use analyses,
>      which is a known layering issue
>
>    - Pass hierarchy provides skipModule etc helper functions
>
>    - Individual passes opt-in to OptBisect activities by manually
> calling skip* helper functions
>      whenever appropriate
>
> With current state of new-pm PassInstrumentation potential OptBisect
> implementation
> will have the following properties/issues:
>    - OptBisect object that exists per compilation pipeline, managed
> similar to PassBuilder/PassManagers
>      (which makes it more suitable for use in parallel compilations)
>
>    - no more layering issues imposed by implementation since
> instrumentations by design
>      can live anywhere - lib/Analysis, lib/Passes etc
>
>    - since Codegen is still legacy-only we will have to make a joint
> implementation that
>      provides a sequential passes numbering through both new-PM IR and
> legacy Codegen pipelines
>
>    - as of right now there is no mechanism for opt-in/opt-out, so it
> needs to be designed/implemented
>      Here I would like to ask:
>          - what would be preferable - opt-in or opt-out?
>
>          - with legacy implementation passes opt-in both for bisect and
> attribute-optnone support at once.
>            Do we need to follow that in new-pm implementation?
>
> Also, I would like to ask whether people see current user interface for
> opt-bisect limiting?
> Do we need better controls for more sophisticated bisection?
> Basically I'm looking for any ideas on improving opt-bisect user
> experience that might
> affect design approaches we take on the initial implementation.
>
> regards,
>    Fedor.
>
> _______________________________________________
> LLVM Developers mailing list
> llvm-dev at lists.llvm.org
> http://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-dev
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.llvm.org/pipermail/llvm-dev/attachments/20181004/fd85fbd2/attachment.html>