[llvm-dev] RFC: Coroutine Optimization Passes

[Gor Nishanov via llvm-dev]

Hi all:

I've included below a brief description of coroutine related optimization
passes and some questions/thoughts related to them. Looking forward to your
feedback, comments and questions.

Thank you!

Roadmap:
========
1) Get agreement on coroutine representation and overall direction.
  .. repeat 1) until happy

  http://lists.llvm.org/pipermail/llvm-dev/2016-June/100838.html (Initial)
  http://lists.llvm.org/pipermail/llvm-dev/2016-July/102133.html (Round 2)

2) Get agreement on how coroutine transformation passes integrate into the
   optimizer pipeline. (this mail) <=== WE ARE HERE
  .. repeat 2) until happy

3) update IR/Intrinsics.td + doc/Coroutines.rst + doc/LangRef.rst

  https://github.com/GorNishanov/llvm/blob/coro-rfc/docs/Coroutines.rst

4) trickle in coroutine transformation passes + tests in digestible chunks.
5) get clang changes in (sometime after step 3).
6) fix bugs / remove limitations / add functionality to coroutine passes
 .. repeat 6) until happy

Overview:
=========
LLVM coroutines are functions that have one or more `suspend points`.
When a suspend point is reached, the execution of a coroutine is suspended and
control is returned back to its caller. A suspended coroutine can be resumed
to continue execution from the last suspend point or it can be destroyed.

CoroSplit Pass: (CGSCC Pass @ new extension point EP_CGSCCOptimizerLate)
---------------
Overall idea is that a coroutine is presented to an llvm as an ordinary function
with suspension points marked up with intrinsics.  We let the optimizer party
on the coroutine as a single function for as long as possible. Shortly before
the coroutine is eligible to be inlined into its callers, we outline parts of
the coroutine that correspond to the code that needs to get executed after the
coroutine is resumed or destroyed. We also create a struct that will keep the
objects that need to persist across suspend points. This transformation is
done by CoroSplit CGSCC pass run as a part of the IPO optimizations pipeline.

CoroElide Pass: (Function Pass @ EP_ScalarOptimizerLate extension point)
---------------
The coroutine resumption intrinsics get replaced with direct calls to those
outlined functions where possible. Then inliner can inline much leaner and nicer
coroutine or any of the outlined parts as desired.

If we discover that the lifetime of a coroutine is fully enclosed in the
lifetime of the caller, we remove dynamic allocation of coroutine state and
replace it with an `alloca` on the caller's frame.

These optimizations are done by a function pass CoroElide run by the main IPO
optimization at EP_ScalarOptimizerLate extension point.

CoroEarly Pass: (Function Pass @ EP_EarlyAsPossible)
---------------
Pretty boring pass that lowers coroutine intrinsics that are not needed for
later coroutine passes.

CoroCleanup Pass: (Function Pass @ EP_OptimizerLast)
--------------
Another boring pass that lowers all remaining coroutine intrinsics that were not
processed by earlier coroutine passes.

Questions / Concerns / Thoughts:
================================

Coroutine attribute or not.
---------------------------
I added a AttrKind::coroutine to flag the pre-split coroutine. The intention is
to lessen the impact of CoroSpit pass since it can simply check for an attribute
to learn if there is any work to be done on a function or not. Without the
attribute, it would need to examine every functions body to see if it has an
llvm.coro.begin. On the other hand, three other coroutine passes have to look at
every function body anyway to lower other coroutine related intrinsics, so we
only saving a little. Another negative aspect of having this attribute is that
there is a possibility of other passes checking if a function is a coroutine or
not a doing and doing something differently if it is. I'd like to avoid this
development and keep all coroutine logic in the coroutine passes and not affect
other optimizations passes. Currently I am leaning towards removing this
attribute.

Integration with CGSCC pipeline.
--------------------------------

As mentioned in the introduction, I would like to run entire IPO pipeline on
pre-split coroutine, then, split it, add the new functions to the SCC and
repeat the pipeline on the updated SCC before proceeding to the next SCC.

To facilitate that, CoroSplit pass ignores the coroutine the first time it sees
it, and requests restart of the pipeline and splits the coroutine when it sees
it the second time around. There are at least four different ways how I can
request a restart of the pipeline, and I would like to solicit your advice on
which one to pick.

Option 1: Force revisit of the SCC if there are any coroutine in it. This makes
  CGPassManager aware of a coroutine attribute and keep restarting the pipeline
  until no functions in SCC has coroutine attribute on them.

  https://reviews.llvm.org/D21569

Option 2: Add a ref parameter bool& Devirt parameter to runSCC virtual method.
  If a pass sets Devirt to true, it indicates to CGPassManager that the changes
  were made that require restart of the pipeline.

  https://reviews.llvm.org/D21570

Option 3: Similar to Option 2, but instead of an extra parameter to runSCC, I
  add another virtual function to CGSCC pass:

  virtual bool restartRequested() const { return false; }

  a pass can override it and return true, if restart is required.

Option 4: CoroSplit can add a fake indirect call to be replaced later with a
  direct call in a CoroElide function pass. CGPassManager::RefreshCallGraph will
  detect devirtualization and restart the pipeline. Hacky, but, no changes to
  CGPassManager required.

Out of these four options, I have a slight preference for 2 or 3, but, I would
like to hear your opinion.

Thoughts on ThinLTO and LTO
===========================

CoroEarly can run during regular compilation. CoroSplit, CoroEarly and
CoroCleanup would run during the link optimization step to take advantage of
more inlining opportunities.

This is all for now.
All the feedback and comments are greatly appreciated!!!

Thank you,
Gor