[LLVMdev] PM: High-level review of the new Pass Manager (so far)

[Chandler Carruth]

First off, sorry for not circling back to this sooner. I think it managed
to get caught in some part of the weird mail bounces from one address and I
had to find it in my other account.

Also, more improtantly, I got distracted. But I'm back to hacking on this
now and wanted to close the loop here.

On Tue Jun 17 2014 at 5:54:39 PM Duncan P. N. Exon Smith <
dexonsmith at apple.com> wrote:

> Hi Chandler,
>
> This is a high-level review of the new WIP `PassManager` infrastructure.
>
> For those that haven't dug into Chandler's commits, here's a very
> high-level overview (assuming IIUC):
>
>   - The driver supports simple declarative syntax for specifying passes
>     to run.  E.g., `module(a,b,function(c,d),e)` runs module passes `a`
>     and `b`, then function passes `c` and `d` for each function, and
>     then module pass `e`.
>
>   - There is a new concept of an AnalysisManager (AM), which runs
>     analyses on-demand and caches the results.
>
>   - Analysis passes are split conceptually from transformation passes.
>         - Analysis:       run: (IRUnit*, AM*) -> Result
>         - Transformation: run: (IRUnit*, AM*) -> PreservedAnalyses
>
>     Note that neither of these matches the old interface, which was
>     runOnIRUnit: IRUnit* -> bool.
>

Minor note for others: the PreservedAnalyses can be computed conservatively
from the bool by selecting between all (false in the old interface) and
none (true in the old interface)


>   - PassManagers interoperate via adaptors.  E.g.,
>     ModuleToFunctionPassAdaptor is a module transformation pass that
>     contains a FunctionPassManager (with some set of function passes).
>
>   - AnalysisManagers interoperate via proxies.  E.g.,
>     FunctionAnalysisManagerModuleProxy is a module analysis pass that
>     forwards to a FunctionAnalysisManager.
>
>   - LazyCallGraph and ModuleToPostOrderCGSCCPassAdaptor collude to visit
>     SCCs in post-order, including API for updating the SCC-graph
>     on-the-fly without invalidating the traversal.
>
> I think what's done generally looks great.  In particular, the
> declarative syntax is awesome and it's a huge step to have an
> AnalysisManager that can reason about analyses separately from the pass
> pipeline.  I haven't analysed the algorithms in LazyCallGraph carefully,
> but I think it's great that they keep the traversal valid amid a
> changing call graph.
>
> Questions and concerns:
>
>  1. PassConcept requires a `name()` for passes.  Why don't
>     AnalysisPassConcept and AnalysisResultConcept?
>

This was just me implementing things as I needed them. I hadn't done the
analysis debug logging, and so hadn't needed a name yet. I've added that at
this point for analysis passes (along with the name) but not for results.
I'm not sure results need a name, the pass having a name seems sufficient
to me.


>
>  2. AnalysisManagerBase defines two versions of `invalidate()` -- one of
>     which takes a `Module*` (instead of `IRUnitT`).  It's not clear to
>     me why the analysis managers (other than ModuleAnalysisManager) need
>     this.  What am I missing?
>

There are two ways invalidation happens, and they require different
interfaces.

a) A transformation pass returns some set of preserved analyses, and all
others are invalidated by the pass manager. The pass manager can't
enumerate the analysis passes and so just hands the preserved set and the
IR unit to the analysis manager.

b) A transformation pass directly invalidates a specific pass for its IR
unit, potentially to re-compute a fresh version while running. This can be
particularly important for cross-domain analyses. As a toy example, you
could imagine having the inliner invalidate and recompute the domtree for
the caller post inlining if it used the domtree over the caller to analyze
the call sites.

This may be a bit more clear now that the "invalidate" utility passes are
in place, but if not, I should add some comments to this effect.


>  3. The number of lines of code to enable transformation pass managers
>     (and analysis managers) to interoperate seems to scale quadratically
>     in the number of types of IRUnit.  Am I reading this wrong?  Is
>     there a plan to manage that, or is it an open problem?  (So far the
>     infrastructure only supports Module, SCC, and Function.  What
>     happens (e.g.) when we add BasicBlock passes?)
>

You're not wrong. I see several ways of managing it, but so far I've just
left it open because the overhead of the abstractions needed to manage it
seem larger than the problem. That may change, and we'd need to revisit
this if so.

However, I'm not sure the number of IR units is likely to grow that much.
Regarding basic block passes, I would like to just do away with them
completely. I think the complexity they bring is greater than the utility
they provide.

Notably, I'm not sure it will ever make sense to have analysis management
at a finer granularity than function, because I don't think there are any
more narrow domains which can compute analysis results in conjunction with
transformations independently on two units. Maybe I'm wrong?

I'm somewhat expecting the only other significant addition to be a loop
pass management layer, and I suspect that one will not provide the analyses
at all.

I'm curious what you and others think here. I do consider this area still
"open" in that I don't think we're painted into a corner (yet).


>
>  4. Previously, (e.g.) function analysis passes would typically hold
>     results only for one function -- when the next function was
>     analyzed, the previous function's results would be invalidated.
>     Now, the FunctionAnalysisManager will happily store results for
>     every function.  (Looking forward, BasicBlockAnalysisManager might be
>     able to store results for every BasicBlock.)
>
>     This is powerful, but costs memory.  What's the plan for keeping
>     memory usage in check (especially for large modules)?
>

The is both the benefit and challenge of the new system. The "plan" (such
as it is) is to teach the pass managers to GC analysis results. I'm not
focused on that until the system is, in essence, working because its hard
to know what the right balance is here without being able to run a big
pipeline of passes over a big pile of IR. I have to imagine that the module
-> SCC adaptor layer is going to do some amount of GC'ing of analysis
results for SCCs that are quiesced, the question is what the right signal
is for quiesced. We have a *lot* of options here while we're using the
caching approach, and I have no way to predict which will be the best.

Ultimately, this is allowing us to consciously make a time/space tradeoff.
While it means we have to think about that tradeoff at least we get to make
it now! =]


>  5. In the old infrastructure, `CGPassManager::runOnModule()` (in
>     `CallGraphSCCPass.cpp`) has a do-while loop that re-runs all passes
>     on a given SCC as long something new gets devirtualized, up to a
>     maximum number of iterations.  I don't see the equivalent loop in
>     CGSCCPassManager, and I don't see an obvious way to accomplish the
>     same optimization cycle.
>
>       - How will we re-implement this loop?  (Or did I miss it?)
>

I'm planning to have a iterate utility that can add a limited amount of
iteration based on specific conditions of any pass, including a CGSCC pass
manager.


>       - There's a potential infinite loop between devirtualization and
>         inlining on recursive virtual functions.  How will we avoid it?
>

I don't follow? We don't inline recursive functions. Eventually, we either
run out of call sites or see direct recursion and stop inlining. Am I
missing a case where this behaves differently during devirtualization?

(Arguably, we shouldn't just *stop*, we should do something more
intelligent, but that's orthogonal to the pass manager. That's just a
weakness of the inliner. And naturally, solving it includes setting limits.)


>  6. There aren't any machine-level pass managers yet.  Will they look
>     about the same?  (Are there open problems to solve there?)
>

I have NO IDEA! =[ This is a big open area.

My current plan is to start off the way the current pass manager works:
there are currently two independent module pass managers. At first, there
will still be, and only one will be the new pass manager.

My next planned step is to try to encapsulate codegen within a single
function pass, and the MC streaming within a single module pass. Then,
internally, codegen can continue using MachineFunctionPass and other
things. I don't know how well this will work with analyses though.

Failing that, there will be a new RFC thread about how to move CodeGen into
the new pass manager world because it'll be a substantial chunk of work,
and largely independent work. This is an area where I would love help of
course. =D I think the only thing it would depend on is having most of the
IR-level analysis up and running. From that point on, it should be
completely independent.


>
>  7. Andy asked in a separate thread [1] (my words): "Does this
>     infrastructure allow for invalidation of classes of analyses?"
>     (E.g., invalidate all analyses whose results depend on the
>     control-flow graph.)  Your response there [2] was (my words): "Not
>     directly, but `AnalysisManager` will be a good place for that API."
>
>     I thought a little about this, and it's not clear to me how to fit
>     this API into AnalysisManager cleanly.
>
>     One side seems straightforward.  Assuming you have a transformation
>     pass that modifies the CFG, tell the AnalysisManager that you've
>     modified the CFG, and return that all analyses have been preserved
>     (since everything else has).  This side is easy to make opt-in.
>
>     The other side is not opt-in, though.  *All* analyses that depend on
>     the CFG must respond, or the invalidation mechanism doesn't work.
>

My very hand-wavy idea has been to use something along the lines of the
preserved set for this. Essentially, analyses which rely on the CFG can
test for whether that has been preserved, and if not, become invalid. Then
we can just "try" to invalidate all the analysis results, and only those
that are safe will remain.

There are several related options with different API tradeoffs. I'd need to
sit down and implement it to see which pattern was really the best. Open to
suggestions here.

Another option I've toyed with, but not yet found a way that I like is to
have some shared IDs in addition to the unique IDs, and use those to form
category sets.


>
>  8. `CGPassManager` was one place where the old return value for a
>     transformation pass (`bool` representing "did-change") was useful.
>     Are there any others?


Not to my knowledge. =[


>     Is the API using `PreservedAnalyses::all()`
>     as a stand-in for "did-not-change"?  If so, does that conflict with
>     adding more nuanced API for on-the-fly invalidation of analyses?
>

Currently, yes, it is using preserving all analyses as a stand-in for "did
no work". This seems fine for invalidation purposes, but not so good for
the iteration thing I mentioned above. I think we'll need a better way to
indicate "no work done", but not sure what the best way is. Suggestions
very welcome here as well.
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.llvm.org/pipermail/llvm-dev/attachments/20150108/226a0932/attachment.html>