[llvm-dev] New pass manager for optimization pipeline status and questions

[Arthur Eubanks via llvm-dev]

On Wed, Jul 22, 2020 at 4:05 PM Chen, Yuanfang <Yuanfang.Chen at sony.com>
wrote:

>
>
> *From:* llvm-dev <llvm-dev-bounces at lists.llvm.org> *On Behalf Of *Arthur
> Eubanks via llvm-dev
> *Sent:* Wednesday, July 22, 2020 2:39 PM
> *To:* llvm-dev <llvm-dev at lists.llvm.org>
> *Subject:* [llvm-dev] New pass manager for optimization pipeline status
> and questions
>
>
>
> Hi all,
>
>
>
> I wanted to give a quick update on the status of NPM for the IR
> optimization pipeline and ask some questions.
>
>
>
> In the past I believe there were thoughts that NPM was basically ready
> because all of check-llvm and check-clang passed when
> -DENABLE_EXPERIMENTAL_NEW_PASS_MANAGER=ON was specified. But that CMake
> flag did not apply to opt and any tests running something like `opt
> -foo-pass -bar-pass` (which is the vast majority of check-llvm tests) were
> still using the legacy PM. The intended way to use NPM was to use the
> -passes flag, e.g. `opt -passes='foo,bar'`.
>
>
>
> I've added a -enable-new-pm flag to opt to force running NPM passes even
> when `opt -foo-pass` is used. This is because I didn't want to go through
> every single test and figure out which ones should be using both -foo-pass
> and -passes=foo. Switching on -enable-new-pm currently leads to ~1800
> check-llvm failures. I've documented the failed tests count per directory
> in https://bugs.llvm.org/show_bug.cgi?id=46651 (some have been fixed
> since that was posted).
>
>
>
> This has led to real bugs in NPM being discovered and fixed (e.g. some
> optnone issues).
>
>
>
> But a large portion of the remaining failures are because codegen-only
> passes haven't been ported to NPM yet. That's fine for the optimization
> pipeline NPM transition since it doesn't affect the optimization pipeline,
> but it does present an issue with the approach of the -enable-new-pm flag
> (which would by default become true alongside the NPM transition). Lots of
> tests are testing codegen-specific passes via opt (e.g. `opt
> -amdgpu-lower-intrinsics`) and they can't use NPM (yet).
>
>
>
> I think the ideal way is just to port these to NPM. The problem is if the
> opt pipeline NPM switch is blocked on this, we’re forcing the targets to
> start porting which I’m not sure if target owners want to do.
>
> On the other end, we’re (almost) actually ready to make these target IR
> passes use NPM for testing purpose with `opt` tool. Except there needs a
> way to expose these passes through llvm/lib/Target, llvm/lib/CodeGen rather
> than llvm/lib/Passes to `opt`. As part of the codegen using NPM work, this
> is almost done.
>
Grepping for INITIALIZE_PASS in llvm/lib/Target, there are 220 files with
that. That means there are ~200 target specific passes. There's no way
we'll be able to port all of those in a reasonable amount of time.
Plus it's not a blocker if we can detect passes are target specific and
simply revert to the legacy PM in opt. Then as codegen passes are ported to
NPM we can start running those on NPM instead reverting back to the legacy
PM. The exact mechanism for this TBD, although probably something similar
to below.

>
>
> I was thinking either we have a way of identifying codegen-only passes and
> revert back to the legacy PM in opt whenever we see one, or we go back to
> considering the originally intended approach of adding an equivalent
> `-passes=` RUN to all tests that should be also running under NPM.
>
>
>
> I would prefer the former since it sounds less pervasive.
>
>
>
> I'm not sure of a nice and clean solution to identify codegen-only passes.
> We could go and update every instance of INITIALIZE_PASS to take another
> parameter indicating if it's codegen-only. Or we could just have a central
> list somewhere where we check if the pass is in some hardcoded list or has
> some prefix (e.g. "x86-").
>
>
>
> The latter seems in line with the progress has been made on codegen using
> NPM work where each target would maintain their own pass registry like
> PassRegister.def. For x86, it is (tentatively) called X86PassRegistry.def. (
> https://reviews.llvm.org/D83613). If It is useful for the opt pipeline
> NPM switch, we could find a way to factor it out.
>
If we can add an equivalent of something like PassBuilder::isAAPassName()
to https://reviews.llvm.org/D83613
(X86CodeGenPassBuilder::isX86CodeGenPassName()?) that
would definitely work.

>
>
> The approach of adding equivalent `-passes=` RUN lines to all relevant
> tests seems daunting, but not exactly sure how daunting. Maybe it's
> possible to script something and see what fails? We'd still need some way
> to identify codegen-only passes to make sure we don't miss anything, and
> we'd need to distinguish between analyses and normal passes. Also, it would
> slow down test execution since we'd run a lot more tests twice, but maybe
> that's not such a big deal? Maybe it's good to have most tests running
> against the legacy PM even when NPM is on by default?
>
>
>
> Thoughts?
>
>
>
> This is split off from
> http://lists.llvm.org/pipermail/llvm-dev/2020-July/143395.html.
>
>


>
>
 I can't speak for other targets but I'd love to get AMDGPU passes
> converted to the NPM. Is there a howto somewhere?
>
I don't think so, but it's usually fairly straightforward, looking at other
passes that exist in both PMs is usually good. I should rewrite
https://llvm.org/docs/WritingAnLLVMPass.html to reference NPM.
But specifically for codegen passes we're currently blocked on ychen's work
for codegen passes infra in NPM:
http://lists.llvm.org/pipermail/llvm-dev/2020-July/143309.html.
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.llvm.org/pipermail/llvm-dev/attachments/20200723/261fc44e/attachment.html>