llvm dev - Jan 2021 - New pass manager for optimization pipeline status and questions

To echo the others' replies, the timeline is not fixed but we're looking
at
flipping the default in `opt` as soon as the AMDGPU issues are addressed,
and the CMake option shortly after that - order of within a week.
Variations on the timeline depend on the bugs being filed in the meantime,
hence the importance of testing and bug filing in the next couple of weeks.
The bugs that are not addressed before the switch would still be worked on
after.

As others have said, the cmake option to continue to use the legacy pass
manager will remain in place, and the legacy pass manager code will not be
removed from the tree for months after, while the transition stabilises.
There's also the option to revert the switch once made for major
regressions that may turn up, but the hope is to not get too many last
minute surprises.


On Tue, Jan 12, 2021 at 12:51 PM Reid Kleckner <rnk at google.com> wrote:
> Keep in mind that LLVM will continue to support the old pass manager via a
> cmake option, so if you are a vendor, you can fallback to the old pass
> manager and migrate on your own timeline. However, there are costs to
> diverging from upstream. As the community migrates to the NPM, the old pass
> manager will become less tested over time, and it may accumulate bugs or
> performance regressions.
>
> On Mon, Jan 11, 2021 at 3:45 PM Sjoerd Meijer via llvm-dev <
> llvm-dev at lists.llvm.org> wrote:
>
>> Hi Alina & Arthur,
>>
>> I've investigated the performance impact for us and can now say a
little
>> bit more now where we are. Switching now would lead to performance
>> regressions (*) in the initial set of 4 benchmarks we care about. One
>> benchmark is overall neutral but shows regressions where we don't
really
>> want them. Hopefully, your fix for PR48715
>> <https://bugs.llvm.org/show_bug.cgi?id=48715> that I raised today
will
>> solve that, many thanks for the amazingly speedy reply! A second
benchmark
>> is a bit of disaster, still need to look into that, and a third shows a
>> relatively small regression but it's significant for that benchmark
and am
>> looking into that now, and need to look into the fourth benchmark.
>>
>> Code-generation is *very* different for the cases I am look at and I am
>> profiling and analysing things, for which I need some time. This leads
to
>> my question: can you remind me about the timelines? I hope we can work
in
>> tandem to have at least the major issues resolved before we switch.
>>
>> Thanks for working on this, and for your help and speedy replies,
>> Sjoerd.
>>
>> (*) I am mostly looking at smaller cores, and very tight loops
>> (baremetal) and guess that most people look at bigger cores where some
of
>> these codegen differences have no or a different impact.
>>
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL:
<http://lists.llvm.org/pipermail/llvm-dev/attachments/20210112/c5cb6386/attachment.html>

Hello,

Thanks for the update and elaborating on this. I am generally okay with the
direction of switching when the blockers are resolved (and the timeline).
After the code-size problems, we have now increased our efforts to look at
performance and are filing our reports. I just wanted to reiterate that we are
looking at severe and generic problems/regressions that should affect all
targets, which I think need to be solved first. We have made quite some progress
and are looking at the last code base but that has the biggest regressions. I
hope that we will have filled all our blockers next week, so that we can make up
the balance of blockers in a week time.

It's understood we could fall back to the legacy pass manager, but I hope
that such a divergence is a last-resort that we don't need to use.

Cheers,
Sjoerd.
________________________________
From: Alina Sbirlea <alina.sbirlea at gmail.com>
Sent: 12 January 2021 22:36
To: Reid Kleckner <rnk at google.com>
Cc: Sjoerd Meijer <Sjoerd.Meijer at arm.com>; llvm-dev <llvm-dev at
lists.llvm.org>; Arthur Eubanks <aeubanks at google.com>
Subject: Re: [llvm-dev] New pass manager for optimization pipeline status and
questions

To echo the others' replies, the timeline is not fixed but we're looking
at flipping the default in `opt` as soon as the AMDGPU issues are addressed, and
the CMake option shortly after that - order of within a week. Variations on the
timeline depend on the bugs being filed in the meantime, hence the importance of
testing and bug filing in the next couple of weeks. The bugs that are not
addressed before the switch would still be worked on after.

As others have said, the cmake option to continue to use the legacy pass manager
will remain in place, and the legacy pass manager code will not be removed from
the tree for months after, while the transition stabilises. There's also the
option to revert the switch once made for major regressions that may turn up,
but the hope is to not get too many last minute surprises.


On Tue, Jan 12, 2021 at 12:51 PM Reid Kleckner <rnk at
google.com<mailto:rnk at google.com>> wrote:
Keep in mind that LLVM will continue to support the old pass manager via a cmake
option, so if you are a vendor, you can fallback to the old pass manager and
migrate on your own timeline. However, there are costs to diverging from
upstream. As the community migrates to the NPM, the old pass manager will become
less tested over time, and it may accumulate bugs or performance regressions.

On Mon, Jan 11, 2021 at 3:45 PM Sjoerd Meijer via llvm-dev <llvm-dev at
lists.llvm.org<mailto:llvm-dev at lists.llvm.org>> wrote:
Hi Alina & Arthur,

I've investigated the performance impact for us and can now say a little bit
more now where we are. Switching now would lead to performance regressions (*)
in the initial set of 4 benchmarks we care about. One benchmark is overall
neutral but shows regressions where we don't really want them. Hopefully,
your fix for PR48715<https://bugs.llvm.org/show_bug.cgi?id=48715> that I
raised today will solve that, many thanks for the amazingly speedy reply! A
second benchmark is a bit of disaster, still need to look into that, and a third
shows a relatively small regression but it's significant for that benchmark
and am looking into that now, and need to look into the fourth benchmark.

Code-generation is *very* different for the cases I am look at and I am
profiling and analysing things, for which I need some time. This leads to my
question: can you remind me about the timelines? I hope we can work in tandem to
have at least the major issues resolved before we switch.

Thanks for working on this, and for your help and speedy replies,
Sjoerd.

(*) I am mostly looking at smaller cores, and very tight loops (baremetal) and
guess that most people look at bigger cores where some of these codegen
differences have no or a different impact.
-------------- next part --------------
An HTML attachment was scrubbed...
URL:
<http://lists.llvm.org/pipermail/llvm-dev/attachments/20210113/864766ac/attachment.html>