Update to LLVM 13.0.0

[chalcolith]

Updated our LLVM submodule to llvm-org-13.0.0 and updated code to compile with it.

[ponylang-main]

Hi @kulibali,

The changelog - changed label was added to this pull request; all PRs with a changelog label need to have release notes included as part of the PR. If you haven't added release notes already, please do.

Release notes are added by creating a uniquely named file in the .release-notes directory. We suggest you call the file 3837.md to match the number of this pull request.

The basic format of the release notes (using markdown) should be:

## Title

End user description of changes, why it's important,
problems it solves etc.

If a breaking change, make sure to include 1 or more
examples what code would look like prior to this change
and how to update it to work after this change.

Thanks.

[chalcolith]

This should wait for the official release.

No API changes at all, only the JIT needed a couple of things.

[SeanTAllen]

Awesome.

Do you know when the release is? Will the commit change then?

[chalcolith]

RC3 is scheduled for September 7 (https://lists.llvm.org/pipermail/llvm-dev/2021-August/152353.html); they usually mark it final a few days after RC3. When there is an official tag we'll need to update the submodule again in this branch.

[chalcolith]

Obligatory "It works fine on my Ubuntu 20.04 machine" post. I'll investigate further.

[SeanTAllen]

That's a cool error though.

[jemc]

Obligatory "It works fine on my Fedora 31 machine" post.

I wasn't able to reproduce the CI error.

I don't know much about the CI setup for LLVM building, but is it possible there is a CI cache issue?

[SeanTAllen]

The change to CMakeLists.txt would bust the cache.

You can see from the runs that the cache was rebuilt:

[SeanTAllen]

I'm going to try locally to see what I get for results.

[SeanTAllen]

Using an Ubuntu 20.04 system with WSL2, I am able to recreate the issue from CI:

Optimising
Verifying
Intrinsic name not mangled correctly for type arguments! Should be: llvm.masked.gather.v4p0s_s.v4p0p0s_s.0
<4 x %9929*> (<4 x %9929**>, i32, <4 x i1>, <4 x %9929*>)* @llvm.masked.gather.v4p0s_Strings.v4p0p0s_Strings
Intrinsic name not mangled correctly for type arguments! Should be: llvm.masked.gather.v4p0s_s.v4p0p0s_s.0
<4 x %9929*> (<4 x %9929**>, i32, <4 x i1>, <4 x %9929*>)* @llvm.masked.gather.v4p0s_Strings.v4p0p0s_Strings
Intrinsic name not mangled correctly for type arguments! Should be: llvm.masked.gather.v4p0s_s.v4p0p0s_s.0
<4 x %9929*> (<4 x %9929**>, i32, <4 x i1>, <4 x %9929*>)* @llvm.masked.gather.v4p0s_Strings.v4p0p0s_Strings
Intrinsic name not mangled correctly for type arguments! Should be: llvm.masked.gather.v4p0s_s.v4p0p0s_s.0
<4 x %9929*> (<4 x %9929**>, i32, <4 x i1>, <4 x %9929*>)* @llvm.masked.gather.v4p0s_Strings.v4p0p0s_Strings
Error:
Module verification failed: Intrinsic name not mangled correctly for type arguments! Should be: llvm.masked.gather.v4p0s_s.v4p0p0s_s.0
<4 x %9929*> (<4 x %9929**>, i32, <4 x i1>, <4 x %9929*>)* @llvm.masked.gather.v4p0s_Strings.v4p0p0s_Strings
Intrinsic name not mangled correctly for type arguments! Should be: llvm.masked.gather.v4p0s_s.v4p0p0s_s.0
<4 x %9929*> (<4 x %9929**>, i32, <4 x i1>, <4 x %9929*>)* @llvm.masked.gather.v4p0s_Strings.v4p0p0s_Strings
Intrinsic name not mangled correctly for type arguments! Should be: llvm.masked.gather.v4p0s_s.v4p0p0s_s.0
<4 x %9929*> (<4 x %9929**>, i32, <4 x i1>, <4 x %9929*>)* @llvm.masked.gather.v4p0s_Strings.v4p0p0s_Strings
Intrinsic name not mangled correctly for type arguments! Should be: llvm.masked.gather.v4p0s_s.v4p0p0s_s.0
<4 x %9929*> (<4 x %9929**>, i32, <4 x i1>, <4 x %9929*>)* @llvm.masked.gather.v4p0s_Strings.v4p0p0s_Strings

I used the following commands to build (same as CI):

make libs build_flags=-j8
make configure arch=x86-64
make build arch=x86-64 build_flags=-j8
make test-ci arch=x86-64

My clang version is clang version 10.0.0-4ubuntu1.

doing:

make libs build_flags=-j8
make configure
make build
make test-ci

also fails.

note in both cases, linking for debug works, it fails when doing optimized builds.

[SeanTAllen]

@kulibali are you sure you checked everything in and pushed it all?

the other question would be... did you try doing make test-ci locally that runs both debug and optimized versions of the test suite? or were you only doing debug versions?

[jemc]

make test-ci works for me locally

[SeanTAllen]

Turning off the inliner on line 1381 of genopt.cc fixes the issue for me, so it would appear to be related to name mangling of inlined functions.

    //pmb.Inliner = createFunctionInliningPass(275);

[SeanTAllen]

@jemc can you put the command you used to build here? is there a difference from either of mine?

[jemc]

The only difference is that I used 4 instead of 8 cores to build LLVM, so it wouldn't bring my system to a crawl:

make libs build_flags=-j4
make configure
make build
make test-ci

[jemc]

Looks like this is an issue with the LLVM vectorization optimizations - maybe my CPU is different enough from yours that LLVM is trying to make different vectorization choices?

[SeanTAllen]

maybe my CPU is different enough from yours that LLVM is trying to make different vectorization choices?

I have an I9 vPro.

[jemc]

@SeanTAllen - does it compile successfully for you if you disable Pony-specific optimizations?

That would hopefully indicate whether it may be a bug in our Pony-specific optimizations or a bug in LLVM itself.

[SeanTAllen]

@jemc

does it compile successfully for you if you disable Pony-specific optimizations?

Nope. That's the first thing I tried.

[jemc]

My CPU on this machine is:

Model name:                      Intel(R) Core(TM) i7-3740QM CPU @ 2.70GHz

[chalcolith]

Mine is Model name: AMD Ryzen 7 3700X 8-Core Processor

[jemc]

This may be a legitimate LLVM bug that we may want to consider reporting as part of their release candidate validation process.

[SeanTAllen]

@kulibali is there still time to report a bug on the 13.00 release?

[chalcolith]

Not sure. I only have superficial knowledge of the release process. I'm in a bit of a crunch at work today so if you have time to investigate that would be great.

[SeanTAllen]

@kulibali i have a deadline for the 15th so, I don't have time to look into it. I had a bit of downtime this morning while waiting on others that I used to do the triaging.

[SeanTAllen]

Turning off pmb.populateLTOPassManager(lpm) while leaving the inliner on also works.

[SeanTAllen]

Turning off lpm.add(createStripSymbolsPass()); also results in a build that works.

[SeanTAllen]

I dont understand how the pass manager works so this is probably wrong but... moving

pmb.populateLTOPassManager(lpm); to the last thing done in optimise also works. but i really dont understand what orderings are important and how the pmb relates to the other passes.

Would I be correct in assuming that effectively turns the pass off?

[SeanTAllen]

So I think this might be our best fix:

@kulibali @jemc

changing to:

createStripSymbolsPass(true)

which ONLY strips debug symbols. also works for me.

Here's the definition:

// These functions removes symbols from functions and modules.  If OnlyDebugInfo
// is true, only debugging information is removed from the module.
//
ModulePass *createStripSymbolsPass(bool OnlyDebugInfo = false);

[SeanTAllen]

@jemc how do you feel about the fix?

[jemc]

Sounds good to me - I gave my 👍 emoji earlier on your comment above.

[SeanTAllen]

I never ever notice emojis @jemc as I get no notification from GitHub about them.

[SeanTAllen]

I just noticed these emojis @jemc...

[chalcolith]

RC4 has now been tagged as llvmorg-13.0.0.

[SeanTAllen]

Awesome.

[SeanTAllen]

@kulibali so we can merge now, yes?

[chalcolith]

Yes.

[SeanTAllen]

Ugh I forgot to squash so this went in as 6 commits. Ooops.