llvm dev - Dec 2021 - Preparing BOLT for LLVM monorepo

Hi Mehdi,

You raise a good point. At the moment we reluctantly went with the clang
dependency route for small tests and for more involved BOLT development we have
an external binary repo that stores large pre-built applications to be used as
tests. You are right with respect to the downsides, but the alternative, which
is to depend on the host compiler (we were previously using host_cc in LIT for
quite a while) can cause more headaches in terms of test stability because we
don’t have control over which compiler was used to build LLVM. BOLT operates at
level that changing even the linker can already introduce changes that may
modify the outcome of a test.

Regarding bugs in clang breaking BOLT tests, that is true. I wonder if this
affects lldb a lot, though. It also depends on the trunk clang to build its
tests. I imagine a debugger is also quite sensitive to the compiler used to
build its test inputs.

To summarize, I see 3 options here:
1 - depend on host compiler
2 - depend on trunk clang
3 - depend on pre-built binaries

#1 is too unstable at the moment. It’s hard to communicate with other devs
regarding why their test is failing because I can only repro it in my machine
with my host compiler.
#2 may be unstable because trunk clang can be broken, and new bugs may be
introduced while clang is not fixed. In my experience, BOLT is not easily broken
by LLVM code, though, so we can afford some time with the tests being offline.
#3 is a burden on the repo because the pre-built binaries may occupy
significantly more space than source code, and also are hard to read (even if
using yaml).

Best,
Rafael

From: Mehdi AMINI <joker.eph at gmail.com>
Date: Friday, December 17, 2021 at 8:26 PM
To: Rafael Auler <rafaelauler at fb.com>
Cc: Hans Wennborg <hans at chromium.org>, Maksim Panchenko <maks at
fb.com>, Xinliang David Li <xinliangli at gmail.com>, Fāng-ruì Sòng
<maskray at google.com>, Sriraman Tallam <tmsriram at google.com>,
Chris Lattner <clattner at llvm.org>, Andrey Bokhanko <andreybokhanko
at gmail.com>, vladislav.khmelevskyi at huawei.com <vladislav.khmelevskyi
at huawei.com>, tstellar at redhat.com <tstellar at redhat.com>,
echristo at gmail.com <echristo at gmail.com>, Nick Desaulniers
<ndesaulniers at google.com>, llvm-dev at lists.llvm.org <llvm-dev at
lists.llvm.org>
Subject: Re: Preparing BOLT for LLVM monorepo
Hi,

The instructions to build Bolt mention:
-DLLVM_ENABLE_PROJECTS="clang;lld;bolt"

It seems that clang is just a dependency here because you build C++ sources
during lit-tests execution, but it's not clear to me why we don't use
the host compiler or a prebuilt release here? Is there an intrinsic coupling
between clang and Bolt at HEAD?

Having to build clang at HEAD just to run the bolt test seems like a non-trivial
overhead for the casual developer, if the host compiler (or some clang pulled
with apt-get or your favorite package manager) would just work.
Also in CI, even if you have a beefy enough machine, you add risks for the Bolt
bot to be broken because Clang/LLVM itself would be broken: so instead of
tracking Bolt bugs you end up tracking clang issues. Also the fact that clang
being broken implicating you can't test Bolt means that in the meantime new
bugs can land in Bolt and complicate auto-bisection.
(The same question applies to lld I think?)

Thanks!

--
Mehdi







On Tue, Dec 7, 2021 at 5:36 PM Rafael Auler <rafaelauler at
fb.com<mailto:rafaelauler at fb.com>> wrote:
Hi folks,

https://github.com/facebookincubator/BOLT branch “main” contains a merge
proposal of BOLT into llvm-project. This is llvm from Nov 30th with 1016 commits
on top of it corresponding to the BOLT project.

These 1016 commits would ideally be committed in a merge commit, merging LLVM as
the first parent and BOLT as the second, and would be there only for the
purposes of preserving project history. In this way, they should be easily
skippable during a bisect of LLVM in the same way as the merge commit of flang.
These commits represent the linear history of BOLT on top of rebased LLVM, so
most commits are not buildable (since we can’t build a very old version of BOLT
on top of a recent LLVM base). That’s why this is for history/blame only.

We have addressed the issues in
https://github.com/facebookincubator/BOLT/issues/248 and we are happy to
continue working on any extra suggestions.

Would it be better if we put this branch as a PR into llvm-project as a way to
make it easier for people to review it? I don’t think we can put this into
phabricator. However, I guess github’s bot will probably auto-close the PR. Also
feel free to open new issues against our facebookincubator/BOLT project as a way
to review it.

Thanks

From: Hans Wennborg <hans at chromium.org<mailto:hans at
chromium.org>>
Date: Thursday, November 4, 2021 at 5:42 AM
To: Maksim Panchenko <maks at fb.com<mailto:maks at fb.com>>
Cc: Xinliang David Li <xinliangli at gmail.com<mailto:xinliangli at
gmail.com>>, Fāng-ruì Sòng <maskray at google.com<mailto:maskray at
google.com>>, Sriraman Tallam <tmsriram at
google.com<mailto:tmsriram at google.com>>, Chris Lattner <clattner
at llvm.org<mailto:clattner at llvm.org>>, Andrey Bokhanko
<andreybokhanko at gmail.com<mailto:andreybokhanko at gmail.com>>,
Rafael Auler <rafaelauler at fb.com<mailto:rafaelauler at fb.com>>,
vladislav.khmelevskyi at huawei.com<mailto:vladislav.khmelevskyi at
huawei.com> <vladislav.khmelevskyi at
huawei.com<mailto:vladislav.khmelevskyi at huawei.com>>, tstellar at
redhat.com<mailto:tstellar at redhat.com> <tstellar at
redhat.com<mailto:tstellar at redhat.com>>, joker.eph at
gmail.com<mailto:joker.eph at gmail.com> <joker.eph at
gmail.com<mailto:joker.eph at gmail.com>>, echristo at
gmail.com<mailto:echristo at gmail.com> <echristo at
gmail.com<mailto:echristo at gmail.com>>, Nick Desaulniers
<ndesaulniers at google.com<mailto:ndesaulniers at google.com>>,
llvm-dev at lists.llvm.org<mailto:llvm-dev at lists.llvm.org> <llvm-dev
at lists.llvm.org<mailto:llvm-dev at lists.llvm.org>>
Subject: Re: Preparing BOLT for LLVM monorepo
Hi Maksim,

On Tue, Nov 2, 2021 at 9:52 PM Maksim Panchenko <maks at
fb.com<mailto:maks at fb.com>> wrote:
> We are still working on finalizing the exact logistics of the merge.
>
> However, we expect to follow the Flang project's footsteps and run the
>
> "--no-ff" merge to preserve the history of ~1K commits. We would
like to
>
> ask for help and coordination from the release managers Tom Stellard and
>
> Hans Wennborg.
I'm no longer involved in release management, and don't really have
any opinion on how to merge this.

Thanks,
Hans
-------------- next part --------------
An HTML attachment was scrubbed...
URL:
<http://lists.llvm.org/pipermail/llvm-dev/attachments/20211220/27fc701e/attachment-0001.html>

On Mon, Dec 20, 2021 at 8:19 AM Rafael Auler <rafaelauler at fb.com>
wrote:
> Hi Mehdi,
>
>
>
> You raise a good point. At the moment we reluctantly went with the clang
> dependency route for small tests and for more involved BOLT development we
> have an external binary repo that stores large pre-built applications to be
> used as tests. You are right with respect to the downsides, but the
> alternative, which is to depend on the host compiler (we were previously
> using host_cc in LIT for quite a while) can cause more headaches in terms
> of test stability because we don’t have control over which compiler was
> used to build LLVM. BOLT operates at level that changing even the linker
> can already introduce changes that may modify the outcome of a test.
>
>
>
> Regarding bugs in clang breaking BOLT tests, that is true. I wonder if
> this affects lldb a lot, though. It also depends on the trunk clang to
> build its tests. I imagine a debugger is also quite sensitive to the
> compiler used to build its test inputs.
>
>
>
> To summarize, I see 3 options here:
>
> 1 - depend on host compiler
>
> 2 - depend on trunk clang
>
> 3 - depend on pre-built binaries
>
>
>
> #1 is too unstable at the moment. It’s hard to communicate with other devs
> regarding why their test is failing because I can only repro it in my
> machine with my host compiler.
>
> #2 may be unstable because trunk clang can be broken, and new bugs may be
> introduced while clang is not fixed. In my experience, BOLT is not easily
> broken by LLVM code, though, so we can afford some time with the tests
> being offline.
>
> #3 is a burden on the repo because the pre-built binaries may occupy
> significantly more space than source code, and also are hard to read (even
> if using yaml).
>
Thanks for the explanations, it makes sense.
You may want to look at some point in the future into a CMake option for
folks to provide a pre-built clang/lld (we even have such an option for
providing a prebuilt TableGen when building LLVM).
That way one could invoke `cmake -DLLVM_ENABLE_PROJECTS=bolt
-DBOLT_CLANG_PATH=/toolchain/bin/clang -DBOLD_LLD_PATH=/toolchain/bin/lld`
I would expect that for day-to-day development on a non-very-powerful
machine that could make quite a difference for the casual developer.

Best,

-- 
Mehdi




>
>
> Best,
>
> Rafael
>
>
>
> *From: *Mehdi AMINI <joker.eph at gmail.com>
> *Date: *Friday, December 17, 2021 at 8:26 PM
> *To: *Rafael Auler <rafaelauler at fb.com>
> *Cc: *Hans Wennborg <hans at chromium.org>, Maksim Panchenko <maks
at fb.com>,
> Xinliang David Li <xinliangli at gmail.com>, Fāng-ruì Sòng <
> maskray at google.com>, Sriraman Tallam <tmsriram at google.com>,
Chris Lattner
> <clattner at llvm.org>, Andrey Bokhanko <andreybokhanko at
gmail.com>,
> vladislav.khmelevskyi at huawei.com <vladislav.khmelevskyi at
huawei.com>,
> tstellar at redhat.com <tstellar at redhat.com>, echristo at
gmail.com <
> echristo at gmail.com>, Nick Desaulniers <ndesaulniers at
google.com>,
> llvm-dev at lists.llvm.org <llvm-dev at lists.llvm.org>
> *Subject: *Re: Preparing BOLT for LLVM monorepo
>
> Hi,
>
>
>
> The instructions to build Bolt mention:
> -DLLVM_ENABLE_PROJECTS="clang;lld;bolt"
>
>
>
> It seems that clang is just a dependency here because you build C++
> sources during lit-tests execution, but it's not clear to me why we
don't
> use the host compiler or a prebuilt release here? Is there an intrinsic
> coupling between clang and Bolt at HEAD?
>
>
>
> Having to build clang at HEAD just to run the bolt test seems like a
> non-trivial overhead for the casual developer, if the host compiler (or
> some clang pulled with apt-get or your favorite package manager) would just
> work.
>
> Also in CI, even if you have a beefy enough machine, you add risks for the
> Bolt bot to be broken because Clang/LLVM itself would be broken: so instead
> of tracking Bolt bugs you end up tracking clang issues. Also the fact that
> clang being broken implicating you can't test Bolt means that in the
> meantime new bugs can land in Bolt and complicate auto-bisection.
>
> (The same question applies to lld I think?)
>
>
>
> Thanks!
>
>
>
> --
>
> Mehdi
>
>
>
>
>
>
>
>
>
>
>
>
>
>
>
> On Tue, Dec 7, 2021 at 5:36 PM Rafael Auler <rafaelauler at fb.com>
wrote:
>
> Hi folks,
>
>
>
> https://github.com/facebookincubator/BOLT branch “main” contains a merge
> proposal of BOLT into llvm-project. This is llvm from Nov 30th with 1016
> commits on top of it corresponding to the BOLT project.
>
> These 1016 commits would ideally be committed in a merge commit, merging
> LLVM as the first parent and BOLT as the second, and would be there only
> for the purposes of preserving project history. In this way, they should be
> easily skippable during a bisect of LLVM in the same way as the merge
> commit of flang. These commits represent the linear history of BOLT on top
> of rebased LLVM, so most commits are not buildable (since we can’t build a
> very old version of BOLT on top of a recent LLVM base). That’s why this is
> for history/blame only.
>
>
>
> We have addressed the issues in
> https://github.com/facebookincubator/BOLT/issues/248 and we are happy to
> continue working on any extra suggestions.
>
> Would it be better if we put this branch as a PR into llvm-project as a
> way to make it easier for people to review it? I don’t think we can put
> this into phabricator. However, I guess github’s bot will probably
> auto-close the PR. Also feel free to open new issues against our
> facebookincubator/BOLT project as a way to review it.
>
> Thanks
>
>
>
> *From: *Hans Wennborg <hans at chromium.org>
> *Date: *Thursday, November 4, 2021 at 5:42 AM
> *To: *Maksim Panchenko <maks at fb.com>
> *Cc: *Xinliang David Li <xinliangli at gmail.com>, Fāng-ruì Sòng <
> maskray at google.com>, Sriraman Tallam <tmsriram at google.com>,
Chris Lattner
> <clattner at llvm.org>, Andrey Bokhanko <andreybokhanko at
gmail.com>, Rafael
> Auler <rafaelauler at fb.com>, vladislav.khmelevskyi at huawei.com
<
> vladislav.khmelevskyi at huawei.com>, tstellar at redhat.com <
> tstellar at redhat.com>, joker.eph at gmail.com <joker.eph at
gmail.com>,
> echristo at gmail.com <echristo at gmail.com>, Nick Desaulniers <
> ndesaulniers at google.com>, llvm-dev at lists.llvm.org <llvm-dev at
lists.llvm.org
> >
> *Subject: *Re: Preparing BOLT for LLVM monorepo
>
> Hi Maksim,
>
> On Tue, Nov 2, 2021 at 9:52 PM Maksim Panchenko <maks at fb.com>
wrote:
> > We are still working on finalizing the exact logistics of the merge.
> >
> > However, we expect to follow the Flang project's footsteps and run
the
> >
> > "--no-ff" merge to preserve the history of ~1K commits. We
would like to
> >
> > ask for help and coordination from the release managers Tom Stellard
and
> >
> > Hans Wennborg.
>
> I'm no longer involved in release management, and don't really have
> any opinion on how to merge this.
>
> Thanks,
> Hans
>
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL:
<http://lists.llvm.org/pipermail/llvm-dev/attachments/20211220/520492a5/attachment-0001.html>