[llvm-dev] Flang landing in the monorepo - next Monday!

Wed Jan 8 10:31:56 PST 2020

Hi Eric, David, all

I work with Peter at Arm in a team of 5 engineers working on F18, and will continue as it becomes llvm-project/flang. We depend on Fortran support in (PGI/Nvidia) flang for our commercial Fortran compiler and will move over to llvm-project/flang (I'll call it Flang from now on) when it is ready so we are committed long-term to the technology and the community. We have a further 12 or so engineers working on LLVM core codegen, mostly for Arm’s SVE architecture, and we have a long-term involvement and commitment to the LLVM project as a whole.

I totally agree with your points that Flang as it stands is not fully integrated into LLVM. Hal has explained the issue about Flang not using LLVM APIs for codegen so hopefully you will agree that will get better over time. The next large development in Flang will be to add the lowering code from the syntax tree to MLIR, which is now an LLVM API. This code is being prototyped in a separate branch of Flang [1], but Eric at Nvidia will start to land the changes soon.

I think your point on usage of LLVM data structures and support libraries is spot on and we need to take that on board. Over time we want Flang to use more of the common LLVM infrastructure and I think that being a part of the monorepo should help this to happen.

At Arm, we have worked on integrating Flang into Clang's driver [2] [3] [4] and we are trying to follow the existing structure of the driver and to take advantage of common functionality as much as possible in that exercise. I should also say that we are working on a patch to Flang that adds testing using llvm-lit/FileCheck in the style of llvm projects [5]. We'll be adding all tests for the new driver as lit tests. We are also working internally on porting over the existing Flang tests that use bespoke scripts to using lit (this code has not made it to github yet.)

As a community, we are also trying as much as possible to follow the LLVM coding style in Flang [6] and other design principles like modularity so each part of the compiler can be tested in isolation.

Our rationale for contributing Flang to LLVM at this point rather than waiting is so that we can work on integrating it more closely with LLVM as a community effort, in tree. The longer we develop Flang as a separate entity, the more its community ethos, its processes, its code style and structure, etc. will likely diverge from the LLVM norms. We think that the eventual inclusion Flang would be harder on that side even though the code itself may make more use of LLVM APIs and datastructures. I don’t think there is a right or wrong answer but this is the position we support at Arm and we look forward to working with the community to improve Flang in the ways you suggest.

Yours
Rich

[1] https://github.com/schweitzpgi/f18/tree/f18
[2] https://reviews.llvm.org/D63607
[3] https://github.com/flang-compiler/f18/pull/759
[4] https://github.com/flang-compiler/f18/pull/762 and https://github.com/flang-compiler/f18/pull/763
[5] https://github.com/flang-compiler/f18/pull/861
[6] https://github.com/flang-compiler/f18/blob/master/documentation/C%2B%2Bstyle.md

From: llvm-dev <llvm-dev-bounces at lists.llvm.org> On Behalf Of Eric Christopher via llvm-dev
Sent: 8 January, 2020 01:48
To: Finkel, Hal J. <hfinkel at anl.gov>
Cc: llvm-dev <llvm-dev at lists.llvm.org>; Mike Edwards <medwards at llvm.org>
Subject: Re: [llvm-dev] Flang landing in the monorepo - next Monday!

On Tue, Jan 7, 2020 at 5:29 PM Eric Christopher <echristo at gmail.com<mailto:echristo at gmail.com>> wrote:
Hi Hal,

On Tue, Jan 7, 2020 at 3:38 PM Finkel, Hal J. via llvm-dev <llvm-dev at lists.llvm.org<mailto:llvm-dev at lists.llvm.org>> wrote:

On 1/7/20 4:38 PM, Doerfert, Johannes via llvm-dev wrote:

On 01/07, David Blaikie via llvm-dev wrote:

Hey Peter - would you be able to link to/describe the history on this

process/decision. I can find one old thread where this was first proposed (

http://lists.llvm.org/pipermail/llvm-dev/2019-February/130497.html ) with

some general positive responses and a lot of questions.

I see there's a flang-dev list, though I'm not sure where code review is

happening (is there flang-commits or equivalent?) as there's not much

chatter on the mailing list that I can see.

There is a flang-commits. We have everything set up to use Phabricator

as MLIR did.

& I'm a bit concerned about flang, already a project with no LLVM API

usage, it seems, and a community that doesn't currently look like it has

much overlap with the LLVM developer community (a very rough glance at

flang-dev participants, but I could be wrong - the LLVM community is large,

to be sure) - I'm concerned that this might create a not very well

integrated community, as we saw in the past with lldb, for instance.

I think I mentioned this before but some people are waiting for the move

to be able to actually participate in the Flang development. So the

current github contributions do not necessarily reflect the LLVM

developers that will be involved.

I'll add to this that a number of us who have been involved with LLVM for a long time are involved in guiding the Flang project. While it's certainly true that a few of the people working hard on Flang are new to the LLVM community, there is a reasonable overlap with the existing LLVM community in the overall effort.

On flang-dev [0] we have long term LLVM developers and people that join

the LLVM developers community with the merge of Flang but even those

already actively participated at the LLVM-Dev!

The API usage is a valid concern, among others, which will require code

changes in the near future.

[0] http://lists.llvm.org/pipermail/flang-dev/2019-December/author.html

The other thing that I'll add here is that the existing Flang development has been broad and not deep - the development started with the lexical analysis (and preprocessor), then the semantic analysis, and so on. It is only very recently that any work on actual code generation has begun, and there, the initial focus has been on targeting LLVM via MLIR. I'm not saying that there aren't more places to use LLVM APIs (e.g., filesystem abstractions, ADT, etc.) -- there are -- but that's something we're planning to improve over time.

At a high level, however, the whole point of this Flang project is produce an LLVM-community-integrated Fortran frontend for LLVM. The current developers of the project, which is a growing community, are committed to realizing that outcome and are open to feedback on code structure, API usage, and other aspects of integrating with the rest of the LLVM project.
I added you to the original thread when I was replying because I knew you were around this effort, it might be nice to see if you can respond to the other thread too.

A lot of my concerns are echoed by David here and I have additional concerns as well. I don't think the project is ready for inclusion into the main llvm tree as I don't see the point right now. There's nothing llvm about it.

To elaborate a bit because this almost assuredly comes off poorly (and I apologize):

I am in favor of having a flang front end in tree. I have concerns about the design of flang versus other front ends, the lack of llvm based library use, and a number of other things that I tried to enumerate in previous emails. I don't know if anything has changed and the responses I got back originally were "we're going to do it anyway" so it didn't leave much room for engagement.

Thanks.

-eric

-eric

 -Hal

Thanks,

  Johannes

On Tue, Jan 7, 2020 at 6:04 AM Peter Waller via llvm-dev <

llvm-dev at lists.llvm.org<mailto:llvm-dev at lists.llvm.org>> wrote:

Hi All,

As discussed before Christmas, this is a reminder that we intend to

merge flang on the 13th January (next Monday) before the llvm-10 branch.

At the moment I'm proposing to do it at 10am GMT. I can be flexible on

this point if it requires close coordination with anyone in another

timezone, just let me know.

Previous discussion was in [llvm-dev] Flang landing in the monorepo

<http://lists.llvm.org/pipermail/llvm-dev/2019-December/137661.html><http://lists.llvm.org/pipermail/llvm-dev/2019-December/137661.html>.

The current plan of action is summarized at

<https://github.com/flang-compiler/f18/issues/876><https://github.com/flang-compiler/f18/issues/876>.

The result will look a lot like the recent MLIR merge from 24th Dec:

<https://github.com/llvm/llvm-project/commits/0f0d0ed1c78><https://github.com/llvm/llvm-project/commits/0f0d0ed1c78>, commit

0f0d0ed1c78 in the llvm-project monorepo.

Once it is done I'll mail the list. If you want to coordinate with me,

please get in touch.

Regards,

- Peter

_______________________________________________

LLVM Developers mailing list

llvm-dev at lists.llvm.org<mailto:llvm-dev at lists.llvm.org>

https://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-dev

_______________________________________________

LLVM Developers mailing list

llvm-dev at lists.llvm.org<mailto:llvm-dev at lists.llvm.org>

https://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-dev

_______________________________________________

LLVM Developers mailing list

llvm-dev at lists.llvm.org<mailto:llvm-dev at lists.llvm.org>

https://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-dev

--

Hal Finkel

Lead, Compiler Technology and Programming Languages

Leadership Computing Facility

Argonne National Laboratory
_______________________________________________
LLVM Developers mailing list
llvm-dev at lists.llvm.org<mailto:llvm-dev at lists.llvm.org>
https://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-dev
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.llvm.org/pipermail/llvm-dev/attachments/20200108/1dbb2d14/attachment.html>