llvm-flang is failing to build about clang-builder

[banach-space]

Hi @mattgodbolt, hi @Sameeranjoshi 👋🏻

Thank you for your e-mail! I'm replying here so that more people can participate (in case I missed anything).

Sadly, LLVM Flang has rather extreme memory requirements. From my personal experience, 64GB should be sufficient. From the screenshot that you copied above, it looks like the build is crashing at compile rather than link time. This makes sense and (AFAIK) is due to a heavy use of templates in various parts of LLVM Flang (hence heavy use of memory at compile time).

I wasn't able to dig out the actual build commands that are used here. I've checked the action for building Clang and I saw ninja install. This might be too ambitious for LLVM Flang. From ninja docs

Builds are always run in parallel, based by default on the number of CPUs your system has.

AFAIK, ninja is equivalent to ninja -j64 (on a 64 core machine). I would start by experimenting with ninja -j32 (basically, nproc/2) and see what happens. You can then try gradually increasing the number of jobs to speed things up. Similar logic applies to make (though the default behaviour of make and ninja is different). I'm not aware of any CMake setting that would help here (e.g. tweaking the number of links jobs is unlikely to help).

thanks for working on this and for Compiler Explorer!
-Andrzej

[mattgodbolt]

Error is:

  CMake Error at CMakeLists.txt:97 (message):
    Clang is not enabled, but is required for the Flang driver
  
  
  -- Configuring incomplete, errors occurred!

So I guess maybe we just need to add clang to the list...

[mattgodbolt]

I added it to the list and we'll see...

[Sameeranjoshi]

Ya, please add clang to the list of builds as flang driver depends on clang driver. For recent developments I think Andrzej warzynski( ***@***.***) from arm may be the best person to ask if some new changes are in driver part. Other than that pinging the mailing list(flang-dev) would be the best option for letting people know about the godbolt flang compiler. I am not working on the project so I am not updated with current status. Thanks

…

On Mon 1 Nov, 2021, 10:39 PM Matt Godbolt, ***@***.***> wrote: I added it to the list and we'll see... — You are receiving this because you were mentioned. Reply to this email directly, view it on GitHub <#27 (comment)>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/AF6JNSS7RLOFSDH6BPGCMS3UJ3CUJANCNFSM5HEOISLQ> . Triage notifications on the go with GitHub Mobile for iOS <https://apps.apple.com/app/apple-store/id1477376905?ct=notification-email&mt=8&pt=524675> or Android <https://play.google.com/store/apps/details?id=com.github.android&referrer=utm_campaign%3Dnotification-email%26utm_medium%3Demail%26utm_source%3Dgithub>.

[mattgodbolt]

Thanks @Sameeranjoshi . The email address you mentioned got blanked for privacy reasons by github, but I can ask around.

In other news, llvm-flang now fails to build as it seems to kill the (new) runner:

The self-hosted runner: i-0752cedb3b50eb50c lost communication with the server. Verify the machine is running and has a healthy network connection. Anything in your workflow that terminates the runner process, starves it for CPU/Memory, or blocks its network access can cause this error.

It only seems to be the llvm-flang build that kills us, though that could be an issue on our side.

[Sameeranjoshi]

andrzej [dot] warzynski [at] gmail [dot] com

[mattgodbolt]

Thanks @Sameeranjoshi !

[mattgodbolt]

So it looks like it's running out of memory. I got a snapshot just before it died.

I'm guessing lots of linkers going on?

[mattgodbolt]

See 4f0ff86

[mattgodbolt]

Restricting to 4 links didn't help (assuming I did it right).

[mattgodbolt]

Thanks @banach-space -- I will try limiting the parallel using clang's own throttling - https://llvm.org/docs/CMake.html#frequently-used-llvm-related-variables suggests I can also reduce the number of compilations as well as link steps (when using ninja, specifically, which as you potin out we do)

Thanks!

[kiranchandramohan]

Can you also use gold or lld to link? -DLLVM_USE_LINKER=gold
Sometimes the compilation of files and link steps overlap and these linkers use less memory.

[mattgodbolt]

Thanks -- we can definitely give that a go. I am pretty sure we have gold available in our build env.

[mattgodbolt]

And we built!

https://github.com/compiler-explorer/compiler-workflows/actions/runs/1554449377

Thanks all!