[llvm-dev] RFC: Contributing Bazel BUILD files in the "peripheral" support tier

[Chris Tetreault via llvm-dev]

Just to be clear, I’m still not in love with this plan. However, it’s being done in accordance with the new policy on this sort of thing, and I’m going to have faith that the policy has teeth if it becomes any sort of issue.

Thanks,
   Christopher Tetreault

From: Geoffrey Martin-Noble <gcmn at google.com>
Sent: Thursday, December 3, 2020 4:27 PM
To: LLVM Dev <llvm-dev at lists.llvm.org>
Cc: tstellar at redhat.com; Renato Golin <rengolin at gmail.com>; Chris Tetreault <ctetreau at quicinc.com>
Subject: [EXT] Re: [llvm-dev] RFC: Contributing Bazel BUILD files in the "peripheral" support tier

Apologies for the delayed response here. I was out of the "office".

Thanks for taking another look :-)

I want to respond first to the process question of pitch vs RFC. My impression was that the pitch process should be used in the case that an RFC couldn't reach consensus. I asked a few times in the last thread (https://groups.google.com/g/llvm-dev/c/u07o3QREVUg/m/uVlV3pMTBAAJ and https://groups.google.com/g/llvm-dev/c/u07o3QREVUg/m/wF5mu-dpBAAJ) whether I should move this to a pitch, but feel like there wasn't a clear response in the context of Renato's support tiers RFC.

It seems like Tom and Renato still disagree about whether I should move this to a pitch. I would appreciate some consensus on that point at least :-D I do see the appeal of a living document for this sort of thing, so definitely see the appeal there, but also it seems like the pitch process is a heavier-weight and more unusual one, so I was hesitant. My inclination is to continue this as an RFC unless we are unable to reach consensus on the issue as outlined in the pitch process description. It does feel like this is really not quite as big a decision as you seem to be suggesting. It's also an easily reversible one since there are no build dependencies and everything is contained.

On Mon, Nov 16, 2020 at 9:41 PM Tom Stellard <tstellar at redhat.com<mailto:tstellar at redhat.com>> wrote:
> This should have approximately the same impact on the community as the
> current GN build in `llvm/utils/gn` does today. That is, it should not
> affect anyone who doesn't care.
>

I want to push back on this a little bit, because having the code in
tree does impact everyone, even people who don't care about it.  It
increases disk usage, commit traffic, checkout times, bugzilla / issue
traffic, and CI builds to name a few things.  There are costs to having
this in tree, the question (as always) is do the benefits outweigh the
costs?
Yes my apologies that this was poorly phrased. I was aiming for a pithy summary and a clear statement that our goal here is not to significantly impact contributors uninterested in Bazel. My impression is that the GN build has achieved that goal. I definitely agree that any addition to the monorepo should have a clear weighing of costs vs benefits and that the costs are never actually zero. I do think the costs here are really quite low however. I am happy to address your concerns and also think that it is important to note that if additional issues arise we are still agreeing to be on the hook for addressing them (e.g. if in practice this causes some unforseen issue with the release) and deleting this contribution if we cannot do so in a timely manner (`rm -rf utils/bazel` is all it requires).

Personally, I do not think we should have alternative build systems in
tree.  However, I still think you should try to propose this as a pitch.
I would much rather this go through a fair process and land than for it
to be rejected based on a contentious thread.

Here is why I'm not convinced this should be in tree:

To me it's not clear why having the build files in-tree is better than
having a separate repo with an llvm-project sub-module.  The in tree
bazel files will be broken from time to time, since most developers will
not be updating them, however, with the sub-module approach you can
ensure that the build will always work by pinning the llvm-bazel repo to
a known-working commit of llvm-project.  Can you expand on the pros/cons
of in-tree vs out-of-tree with sub-modules.
Out-of-tree with a submodule is the current approach we have with https://github.com/google/llvm-bazel. It's certainly doable, but involves quite a bit of bookkeeping to track which version corresponds to a given version of LLVM such that someone can fetch the correct configuration (you'll note that the repository has about 7k tags at the moment). To make things somewhat more complicated, the typical way to fetch something for use in Bazel is with an http_archive<https://docs.bazel.build/versions/master/repo/http.html#http_archive> which requires one to specify the archive digest to avoid refetching on each build. This doesn't work particularly well with tags that change which commit they point to. I'm not saying these issues aren't solvable, but they add quite a bit of complexity.

The other point is that I think this makes contributing to the Bazel configuration quite a bit more complex because you have to apply patches across multiple repositories to also be kept in sync. Given that LLVM has a monorepo, it still seems like the logical place for a build configuration of LLVM used by multiple projects.

Other concerns I have from reviewing the patch:
It seems like these are mostly concerns with the specific implementation. Would you be alright with saving the specific details for an eventual review on the patch if this moves forward? I've made brief responses below.
* It looks like there is a build configuration for at least one external
project (zlib) and possibly another (vulkan-headers?).  Do we really
want to have build configurations for non-LLVM projects in our tree?  Is
there any limit to the number of external projects that can and will be
added?
These are dependencies of the LLVM Project and LLVM keeps its dependencies pretty tightly managed AFAIU. These configurations are also pretty trivial, "here are the source files", type things, so I think it's even a bit generous to call them configurations: we're just informing Bazel where the files are located.

* There are 3 files (abi-breaking.h.cmake, config.h.cmake,
llvm-config.h.cmake) that have been copied from the llvm tree into
utils/bazel/, is there some way we can avoid carrying multiple copies of
the same file in tree?

* Similarly, there are some files that are normally generated at build
time clang/Config/config.h, llvm/Config/config.h,
llvm/Config/llvm-config.h that have been copied into utils/bazel.  Is it
really necessary
to have these in tree?  Especially since some of the templates, like
llvm-config.h.cmake, are also in utils/bazel?
The copy here is pretty much orthogonal to the actual build configuration. The intent is to have a literal change detector test for changes to these cmake configurations, since they would invalidate assumptions in the Bazel configuration. Chandler and I went back and forth on a few different ways to do this. We can certainly look at other options. The issue is that I don't think there's actually a useful way to interpret the .cmake template files since changes to them are also made as changes to the cmake configuration and without these being in sync the files just drift. Happy to discuss other options for how to handle this. We could, for instance, have some other process that just looks at the git diff/log for these files.

* I still worry about the bazel files causing merging conflicts when
backported to the stable branch.  If these are added to tree, could we
have a rule where commits to utils/bazel/ cannot include changes to
other files?
I'd certainly be open to discussing restrictions that would avoid additional burden on release managers. I think that one makes contributing to the Bazel configuration more difficult because you cannot do it as part of a patch that requires a change, but if it's something that would cause issues with the release then we can avoid it. My intuition is that this wouldn't actually come up often, however. For example, just looking at the gn directory I see several commits in the last week that touch this and other files. Have you actually run into issues? Since this is unsupported the conflicts could also be resolved pretty much however you wanted (e.g. delete the conflict markers, delete the file), so they seem pretty trivial to deal with if they only happen occasionally. My preference would therefore be to see if this is actually a problem in practice before putting rules in place.


On Tue, Nov 17, 2020 at 2:27 AM Renato Golin <rengolin at gmail.com<mailto:rengolin at gmail.com>> wrote:
Hi Geoffrey,

Thanks for the re-submission.

I have some comments below that may sound negative, but they're probably just a reflection of my own ignorance. I want to make sure the submission is clear, so it can be accepted on its own right.

On Tue, 17 Nov 2020 at 03:02, Geoffrey Martin-Noble via llvm-dev <llvm-dev at lists.llvm.org<mailto:llvm-dev at lists.llvm.org>> wrote:
 This should not affect development of core tier components. One reason we propose adding this to the root utils/ directory instead of under llvm/utils (where GN is located) is to avoid unnecessarily sending messages to llvm-commits. Others have raised the concern that the existence of an alternative build system might lead to lack of maintenance for the CMake build system. Given that supporting CMake will remain a requirement and maintenance of a Bazel build system will continue to happen regardless, we do not expect any significant impact in this way.

I was under the impression that "utils" was actually "llvm/utils", which would be in the same place as GN. I don't think we should treat GN and Bazel as different and I really wouldn't like to have a different quality control (for post commit reviews).

If the Bazel commits are too verbose (for example, committing auto-generated code), then we should really clean that up and commit the script that generates them and make that part of the build.

I understand the need to move the noise away, but move it too far away and it's no better than in a separate repo.
I am happy to put this in either location and agree it should be in the same place as GN. If we were to decide that it should go `utils/` then I would also propose we move GN to there as well. I believe the GN files were contributed prior to the existence of the monorepo, so a top-level `utils/` wouldn't have been an option. I think living under the root `utils/` directory makes more sense because these are not configurations for only the LLVM subproject (we also build MLIR and Clang with perhaps more to come). I believe it was Mehdi's suggestion that this would help mitigate some of the costs to having it in the monorepo because Tom mentioned commit list traffic as a concern. I don't think I agree that one directory up is akin to a separate repo though :-D

That said, this is a really minor point for me. I'm happy to put this wherever people prefer :-)


A number of people raised the question of "why not a separate repository". This is indeed possible: It's what we've done with https://github.com/google/llvm-bazel, which is currently used by https://github.com/google/iree. It is significantly more infrastructure, coordination, and complexity for something that is specifically a configuration for the LLVM project itself, not its own dependent or adjacent project.

I was also under the impression that one of the big reasons why we needed it to be in LLVM is that, like CMake, it needed files all over the place. This would indeed be a major infrastructure undertaking.

But given that it's all being hosted in a single directory, and outside of the LLVM tree, I really can't see what's so much harder about an extra checkout in the same tree.
Bazel *wants* the build files to be all over the place, but I've tricked it with some repository rule symlinking. That's also true of the LLVM GN configuration, I believe. My assumption is that having BUILD files actually throughout the repository would be something that would receive quite a bit of pushback and would be confusing for people who would naturally expect these BUILD files to be maintained as a supported build system. I would happily put a BUILD.bazel file at the root of each subproject and drop the symlinking madness, but I suspect this would not be embraced as a solution ;-P

I believe this contribution will significantly improve the situation for downstream users that use Bazel while having minimal impact on the community at large.

It's not clear to me yet if LLVM/Bazel is only used in Google projects or any other non-Google project. All that you listed so far seem to be exclusive to Google.

This is not a problem per se, but it does promote the idea that Google could common it up internally instead.

The main reasons why it would be upstream are that it's either a product by or requirement to the project itself, or it helps unite cross-industry collaboration that wouldn't be possible otherwise.

It's clearly not the former (and why it's in the periphery tier), but it's also not clear it's in the latter either.

I can really only speak for Google projects. I have also noticed several other Bazel build configurations in the wild, e.g. PlaidML<https://github.com/plaidml/plaidml/blob/master/vendor/llvm/llvm.BUILD> (Intel) or this bazel_llvm<https://github.com/ChrisCummins/bazel_llvm> project that I found after someone contributed a doc fix. I believe in the last thread someone from Facebook mentioned that Bazel build files would also be relatively easily translatable to their internal Bazel-derived build system, Buck. Someone from Lyft also expressed interest in using a Bazel build configuration if it was in-tree. But I can't really speak to the motivations, road maps, etc. for any of these people, companies, or projects (if you're reading, please chime in ;-P).
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.llvm.org/pipermail/llvm-dev/attachments/20201207/5d94a223/attachment-0001.html>