Opt Bisect layering

Andrew,

I would not make the caller pass the description of the IR unit. That is because it would result in the description generated every time even if OptBisect is disabled. Description generation is not very chip.

Thinking on the OptBisect extension, I believe passing the units are the right choice because OptPassGates may use them to make pass skipping decisions.

-Yevgeny Rouban

So… looking at OptBisect, I have a few thoughts:

  1. what’s the purpose of the virtual interface/OptPassGate? I’m guessing maybe that worked around the circular referencing in these APIs? hmm, no, I suppose that wouldn’t work/be relevant here.

  2. Why is OptBisector a ManagedStatic? That seems pretty antithetical to the role of LLVMContext. When/why would a user be bisecting over multiple LLVMContexts? & even then, maybe it’d be more suitable for that grouping (the scope for the bisection) to be API driven - passing the bisector into the LLVMContext ctor to define the set of contexts that share a bisector?

& now looking back at the patch-in-progress, I see it allows setting the OptBisector/OptPassGate as suggested in (2).

If that becomes the /only/ option (ie: LLVMContext has no default OptPassGate) then the virtual interface could be kept down in IR (though it’s still a bit questionable to have those Analysis types (Loop, Region, CallGraphSCC) even declared in IR). Then the implementation of OptBisector could be moved into Analysis - freely able to depend on the concrete Analysis types.

  • Dave

& now looking back at the patch-in-progress, I see it allows setting

the OptBisector/OptPassGate as suggested in (2).
Well, the patch currently discussed does not attempt to solve the passgate object management issue.
It is left for the discretion of passgate object provider.

>
> If that becomes the /only/ option (ie: LLVMContext has no default OptPassGate) then the virtual interface could be kept down in IR (though it's still a bit questionable to have those Analysis types (Loop, Region, CallGraphSCC) even declared in IR). Then the implementation of OptBisector could be moved into Analysis - freely able to depend on the concrete Analysis types.

To me this is a "Pass Manager catch" - entity that attempts to control all the passes needs to be part of (or tightly cooperate with) pass manager.
Pass manager is currently in IR, and perhaps rightfully so.
Yet passes that it controls work on "IR units" which are either IR or Analysis, thus Analysis leaks into the interfaces inevitably.
Kinda logical conflict it is...

regards,
Fedor.

>
> - Dave
>
> So... looking at OptBisect, I have a few thoughts:
>
> 1) what's the purpose of the virtual interface/OptPassGate? I'm guessing maybe that worked around the circular referencing in these APIs? hmm, no, I suppose that wouldn't work/be relevant here.
>
> 2) Why is OptBisector a ManagedStatic? That seems pretty antithetical to the role of LLVMContext. When/why would a user be bisecting over multiple LLVMContexts? & even then, maybe it'd be more suitable for that grouping (the scope for the bisection) to be API driven - passing the bisector into the LLVMContext ctor to define the set of contexts that share a bisector?
>
> Andrew,
>
> I would not make the caller pass the description of the IR unit. That is because it would result in the description generated every time even if OptBisect is disabled. Description generation is not very chip.
> Thinking on the OptBisect extension, I believe passing the units are the right choice because OptPassGates may use them to make pass skipping decisions.
>
> -Yevgeny Rouban
> -----------------------------------------------------------
>
> From: llvm-dev [mailto:llvm-dev-bounces@lists.llvm.org] On Behalf Of Kaylor, Andrew via llvm-dev
> Sent: Thursday, March 22, 2018 3:52 AM
> To: David Blaikie <dblaikie@gmail.com>; llvm-dev <llvm-dev@lists.llvm.org>; Friedman, Eli <efriedma@codeaurora.org>
> Subject: Re: [llvm-dev] Opt Bisect layering
>
> There is a patch under review right now from someone who wants to provide a mechanism to replace OptBisect as the pass gate keeping mechanism.
>
> https://reviews.llvm.org/D44464
>
> Possibly we could take this opportunity to move OptBisect to a different layer, though I don’t know where else it would belong.
>
> The other possibility is that we could have the caller pass in a description instead of a pointer to the pass and the IR unit. OptBisect isn’t doing anything with them other than building a string for output.

& now looking back at the patch-in-progress, I see it allows setting
the OptBisector/OptPassGate as suggested in (2).
Well, the patch currently discussed does not attempt to solve the
passgate object management issue.
It is left for the discretion of passgate object provider.

If that becomes the /only/ option (ie: LLVMContext has no default
OptPassGate) then the virtual interface could be kept down in IR (though
it’s still a bit questionable to have those Analysis types (Loop,
Region, CallGraphSCC) even declared in IR). Then the implementation of
OptBisector could be moved into Analysis - freely able to depend on the
concrete Analysis types.

To me this is a “Pass Manager catch” - entity that attempts to control
all the passes needs to be part of (or tightly cooperate with) pass manager.
Pass manager is currently in IR, and perhaps rightfully so.
Yet passes that it controls work on “IR units” which are either IR or
Analysis, thus Analysis leaks into the interfaces inevitably.
Kinda logical conflict it is…

This is in response to my “it’s still a bit questionable” comment? That’s not too important - I’m not pushing to change that if we can get the mechanical layering functional regardless, by only having forward declarations of those different Analysis entities in llvm/IR, but not need their definitions except in the implementation of this virtual interface which could live in llvm/Analysis.

But to discuss it anyway: It seems a bit different that the “Pass Manager catch” depends on the concrete types but the Pass Manager (PassManager.h) itself, does not - it’s only templates, none of it depends on Region, Loop, etc. If the catch could be implemented similarly to the manager itself, then it’d have the same layering requirements & no problem. But I haven’t looked closely enough at the APIs to figure out if/how that might be done - the current implementation/mechanisms are at odds because of the incompatibility of templates and virtual dispatch (can’t have a virtual function template - it’d have an unbounded/unknowable number of vtable entries, etc). Some sort of visitor-y thing might be needed/useful, I’m not sure. But again, not sure this is necessary to address/fix for the issues I’m seeing/pushing to deal with - but I’m happy to discuss/help design this area as well if you’d like :slight_smile:

  • Dave

Pass Manager (PassManager.h) itself, does not - it's only templates,

none of it depends on Region, Loop, etc.
Well, true but the problem happens when you try to instantiate the thing.
And for generic features like opt-bisection, ir-print-after-all etc that want:
- to have a say before/on/after every execution of every pass
- have a shared implementation of the main logic

you have to do most of the following:
1. instantiate your interfaces for all the IRUnits
2. have pass manager doing the job for you directly
3. extend pass interface with specific helpers for your job (skipFunction)

neither of those helps perfect layering...
And with new pass manager having no common Pass hierarchy this gets even more clumsy.

> I'm happy to discuss/help design this area as well if you'd like :slight_smile:
Yeah, I'm interested to continue this design discussion, although my interests
are primarily in the area of new pass manager currently.
I'm going to post a separate RFC on that topic.

regards,
Fedor.

>
> > & now looking back at the patch-in-progress, I see it allows setting
> the OptBisector/OptPassGate as suggested in (2).
> Well, the patch currently discussed does not attempt to solve the
> passgate object management issue.
> It is left for the discretion of passgate object provider.
>
> >
> > If that becomes the /only/ option (ie: LLVMContext has no default
> OptPassGate) then the virtual interface could be kept down in IR (though
> it's still a bit questionable to have those Analysis types (Loop,
> Region, CallGraphSCC) even declared in IR). Then the implementation of
> OptBisector could be moved into Analysis - freely able to depend on the
> concrete Analysis types.
>
> To me this is a "Pass Manager catch" - entity that attempts to control
> all the passes needs to be part of (or tightly cooperate with) pass manager.
> Pass manager is currently in IR, and perhaps rightfully so.
> Yet passes that it controls work on "IR units" which are either IR or
> Analysis, thus Analysis leaks into the interfaces inevitably.
> Kinda logical conflict it is...
>
> This is in response to my "it's still a bit questionable" comment? That's not too important - I'm not pushing to change that if we can get the mechanical layering functional regardless, by only having forward declarations of those different Analysis entities in llvm/IR, but not need their definitions except in the implementation of this virtual interface which could live in llvm/Analysis.
>
> But to discuss it anyway: It seems a bit different that the "Pass Manager catch" depends on the concrete types but the Pass Manager (PassManager.h) itself, does not - it's only templates, none of it depends on Region, Loop, etc. If the catch could be implemented similarly to the manager itself, then it'd have the same layering requirements & no problem. But I haven't looked closely enough at the APIs to figure out if/how that might be done - the current implementation/mechanisms are at odds because of the incompatibility of templates and virtual dispatch (can't have a virtual function template - it'd have an unbounded/unknowable number of vtable entries, etc). Some sort of visitor-y thing might be needed/useful, I'm not sure. But again, not sure this is necessary to address/fix for the issues I'm seeing/pushing to deal with - but I'm happy to discuss/help design this area as well if you'd like :slight_smile:
>
> - Dave
>
> regards,
> Fedor.
>
> >
> > - Dave
> >
> > So... looking at OptBisect, I have a few thoughts:
> >
> > 1) what's the purpose of the virtual interface/OptPassGate? I'm
> guessing maybe that worked around the circular referencing in these
> APIs? hmm, no, I suppose that wouldn't work/be relevant here.
> >
> > 2) Why is OptBisector a ManagedStatic? That seems pretty
> antithetical to the role of LLVMContext. When/why would a user be
> bisecting over multiple LLVMContexts? & even then, maybe it'd be more
> suitable for that grouping (the scope for the bisection) to be API
> driven - passing the bisector into the LLVMContext ctor to define the
> set of contexts that share a bisector?
> >
> > Andrew,
> >
> > I would not make the caller pass the description of the IR
> unit. That is because it would result in the description generated every
> time even if OptBisect is disabled. Description generation is not very chip.
> > Thinking on the OptBisect extension, I believe passing the
> units are the right choice because OptPassGates may use them to make
> pass skipping decisions.
> >
> > -Yevgeny Rouban
> > -----------------------------------------------------------
> >
> > From: llvm-dev [mailto:llvm-dev-bounces@lists.llvm.org] On
> Behalf Of Kaylor, Andrew via llvm-dev
> > Sent: Thursday, March 22, 2018 3:52 AM
> > To: David Blaikie <dblaikie@gmail.com>; llvm-dev
> <llvm-dev@lists.llvm.org>; Friedman, Eli <efriedma@codeaurora.org>
> > Subject: Re: [llvm-dev] Opt Bisect layering
> >
> > There is a patch under review right now from someone who
> wants to provide a mechanism to replace OptBisect as the pass gate
> keeping mechanism.
> >
> > ⚙ D44464 allow custom OptBisect classes set to LLVMContext
> >
> > Possibly we could take this opportunity to move OptBisect to
> a different layer, though I don’t know where else it would belong.
> >
> > The other possibility is that we could have the caller pass

Pass Manager (PassManager.h) itself, does not - it’s only templates,
none of it depends on Region, Loop, etc.
Well, true but the problem happens when you try to instantiate the thing.
And for generic features like opt-bisection, ir-print-after-all etc that
want:

  • to have a say before/on/after every execution of every pass
  • have a shared implementation of the main logic

you have to do most of the following:

  1. instantiate your interfaces for all the IRUnits
  2. have pass manager doing the job for you directly
  3. extend pass interface with specific helpers for your job
    (skipFunction)

neither of those helps perfect layering…
And with new pass manager having no common Pass hierarchy this gets even
more clumsy.

Not sure I follow, sorry - when you go to instantiate the pass manager & the catch system - at that point there’s a concrete set of passes (you have a dependence on Analysis and Transforms) and entities (regions, loops, etc) so the dependencies seem like they make sense.

I’m happy to discuss/help design this area as well if you’d like :slight_smile:
Yeah, I’m interested to continue this design discussion, although my
interests
are primarily in the area of new pass manager currently.
I’m going to post a separate RFC on that topic.

Sounds good.

Looping back for this thing - would it be reasonable to remove the default OptBisect from IR, move it into Transform or some other leafier dependency? Leaving only the basic interface (that can get away with forward declarations of Region, Loop, etc) in IR? Is that likely to be done in the patch under review, or shortly after it?

Ping on this - any chance we can look at fixing the OptBisect layering here/now?

Could we move the implementation into Analysis & require users to set it, rather than having it as a default value in IR?

David,

we definitely need to address this issue and I did not forget about it.

I’m still fleshing out a proposal on a generic solution for “pass execution control points”,
which was inspired by the new pass manager needs but then it appears
we can largerly reuse the implementation for both managers.

Since the implementation should be based on a special Analysis,
it definitely will be able to address this layering issue (as a side-effect :).

I very much hope to be able send something real to the list in less than a week.

regards,
Fedor.

Ping - did this end up progressing? (I might’ve missed or forgotten about anything coming out on the list)

In case anyone in this thread is interested, I have proposed a fix to the layering issue in https://reviews.llvm.org/D58406