design question on inlining through statepoints and patchpoints

I've been looking at inlining invokes / calls done through statepoints
and I want to have a design discussion before I sink too much time
into something I'll have to throw away. I'm not actively working on
adding inlining support to patchpoints, but I suspect these issues are
applicable towards teaching LLVM to inline through patchpoints as
well.

There are two distinct problems to solve before LLVM can inline
through statepoints:

# Managing data flow for the extra metadata args.

LLVM needs some logic to "transfer" the extra live values attached to
a statepoint/patchpoint into the body of the inlinee somehow. How
this is handled depends on the semantics of the live values (something
the frontend knows). There needs to be a clean way for the frontend
to communicate this information to LLVM, or we need to devise a
convention that is sensible for the kinds of frontends we wish to
support. Initially I plan to sidestep this problem by only inlining
through statepoints have *no* extra live values / gc pointers.

# Managing the call graph

This is the problem we need to solve first. Currently LLVM views the
a statepoint or patchpoint call as

  1. A call to an intrisic. This does not add an edge to the call
     graph (not even to the dedicated external node).

  2. An escaping use of the callee.

IIUC, (2) is (conservatively) imprecise and (1) is incorrect. (1)
makes LLVM believe that a function that calls @f via a statepoint does
not call @f at all. (2) makes LLVM believe that @f is visible
externally, even if it has internal linkage.

Given this starting point, I can think of three ways to model
statepoint's (and patchpoint's) control flow semantics within a call
graph:

  1. Model calls to statepoint, patchpoint and stackmap intrinsics as
     calling the external node. Teach the inliner pass to
     "devirtualize" calls through statepoints when posssible, except
     that the "devirtualization" is only a facade (i.e. we don't
     mutate the IR to change the statepoint to a direct call). We add
     some abstraction to the inlining utility functions to inline
     through something more general than a CallSite.

  2. Introduce a new abstraction InlineSite (bikeshedding on the name
     is welcome). InlineSite sits on top of a CallSite and knows how
     to extract the semantic call out of a statepoint or a patchpoint
     (similar to the llvm::Statepoint class). The inliner and the
     call graph analysis works on top of this InlineSite abstraction
     instead of the CallSite abstraction.

  3. Change all the places that matter (CallGraph, CallGraphSCCPass
     etc.) from

       if (CallSite CS = ...)

     to

       if (Statepoint SP = ...)
          ...
       else if (CallSite CS = ...)

     or something equivalent to this.

Personally, I'd prefer going with (1) if it is viable, and (2) if not.

What do you think?

-- Sanjoy

Look at the LazyCallGraph pass? This is going to be the basis of the new pass manager, and I designed it specifically to help deal with these kinds of issues. I’d be interested if it just directly addresses the call graph issue without the need for any special handling. If not, I’d like to understand why not.

I’ll think some about the inliner side of this…

With respect to phase ordering, is the long term plan to run the statepoint placement/transformation phases late (after all optimizations)?
If so, will we need to support inlining post statepoint transformation?

Thanks,
Swaroop.

The long term plan is a) evolving, and b) dependent on the specific use case. :slight_smile:

It would definitely be nice if we could support both early and late safepoint insertion. I see no reason that LLVM as a project should pick one or the other since the infrastructure required is largely overlapping. (Obviously, I'm going to be mostly working on the parts that I need, but others are always welcome to extend in other directions.)

One of the challenges we've run into is that supporting deoptimization points (which in practice are safepoints) require a lot of the same infrastructure as early safepoint insertion. It's likely that we'll end with a scheme which inserts safepoint polls quite early (but with restricted semantics and optimization impact) and then converts them to explicit GC safepoints (with full invalidation semantics) quite late. We already have this distinction in tree in the form of PlaceSafepoints and RewriteStatepointsForGC. I suspect we'll move further in this direction.

I suspect that for languages without deoptimization, you'll want to insert safepoint polls quite late. Whether you do the same for safepoints-at-calls is debatable. I used to think that you should do that quite late, but I'm no longer sure that's always the right answer.

Philip

patches here (reverse chronological order):

http://reviews.llvm.org/D10633
http://reviews.llvm.org/D10632
http://reviews.llvm.org/D10631

I’m digging into the patches, and it has at least raised one high-level question for me that I wanted to ask here so that the response would be reasonably widely seen.

Essentially, what are the particular motivations for inlining through statepoints? How important are they for the actual users we’re building of the GC infrastructure? (Philip’s email starts to give some hints here, but I’m not really clear exactly how important this is… for example, I don’t understand what the deoptimization thing is all about. This is likely just my ignorance of common parlance and problems in GC land, but I think it would be useful to break it down so that we have a reference for what all is being discussed.)

I’m not really doubting the importance mind you, I’d just like to understand it better (and confirm my suspicion that this is really a “must have” feature rather than a “nice to have” feature).

Chandler,

Thanks for asking about the bigger picture. It’s definitely good to summarize that somewhere more visible than a patch review.

In my mind, there are two intersecting but mostly orthogonal reasons why we want to support the ability to inline through statepoints (and someday possibly patchpoints). They are: a) supporting early safepoint insertion and b) implementation flexibility. I’ll go into each below.

Today, we really only support the insertion of safepoints (both polls and call safepoints) after inlining has been done. This is the model that PlaceSafepoints and RewriteStatepointsForGC support in tree and that we’ve been using with some success. I believe that’s also the model that WebKit has used with patchpoints, though in a different form. Their high level optimization is done in an entirely different IR; whereas ours is done over LLVM IR.

There have been a couple of things that have come up which make me think it’s valuable to support the insertion of at least call safepoints earlier in the optimizer. For one, it makes thinking about escaped stack allocated objects (which must be updated in place) more obvious for the frontend. Second, talking to other frontend authors most people expect to use an early insertion model. Supporting both from a functional standpoint and having late insertion be an “optimization” seems to provide a more gentle introduction path. Third, it makes reasoning about deoptimization a lot cleaner.

I don’t want to get into the weeds of deoptimization right now - it would utterly derail the conversation and we’re not making any proposals for upstream in this area at the moment - but let me summarize the problem for context. Essentially, we need to be able to track a chain of abstract frame states through inlining and serialize them into the stackmap via a statepoint. Today, we handle tracking those abstract frames using an independent mechanism I’m going to skip over. This has worked, but has frankly been a major pain point. Having the ability to inline through statepoints opens up some possible designs in this space, but we’re not proposing anything concrete at this moment.

The more immediate benefit from adding support for inlining through statepoints is that it gives a lot more implementation flexibility. Today, we have a hard requirement that all inlining be done before statepoints are inserted. This influences design choices in sometimes odd ways.

As one in tree example, consider the insertion of safepoint polls done by PlaceSafepoints. We currently need to force inline newly inserted calls rather than just inserting them and worrying about the profitability later. In the future, we might want to support not inlining some of these poll sites. (Consider a really cold loop - you might prefer the cost of the extra call over the code size increase for the test and slow path.) Today, we’d have to make that choice within PlaceSafepoints. With support for inlining through statepoints, we could defer that to a distinct pass and thus factor the code more cleanly. In fact, we could even separate some of the existing optimizations for placement into a separate pass which removes provably redundant safepoints rather than combining that into the insertion logic itself. I’m not necessarily saying we will, but the ability to inline through statepoints gives us the option of considering choices like these which we can’t otherwise do.

Another (out of tree) example comes from our optimization pipeline. We use a set of known functions to represent high level concepts. Today, we have a hard requirement about inlining all of these known functions (well, all that can contain safepoints), before safepoint insertion. Part of our work (out of tree) at safepoint insertion is pruning deoptimization state which turn out not to be required. This often opens up optimization possibilities that would be visible if we could identify high level semantic constructs. Being able to swap the safepoint insertion and lowering phases (which does inlining internally) would be very helpful from an optimization perspective.

Philip

p.s. Just to be clear, the flexibility parts really only require the InlineFunction changes; not full support from the inliner. That’s only required for early call safepoint insertion.

p.p.s. Longer term, it really feels to me that statepoints/patchpoints are moving in the direction of being an arbitrary wrapped call. There’s some underlying call (or invoke) with additional layers of semantics wrapped around that. We seem to be settling on a relatively fixed set of composable semantics, so long term these may be IR extensions. No concrete proposals here yet though.