Use more efficient linker for flang debug builder.

[stevanradakovic]

Summary

Use lld linker for flang debug configuration.

[stevanradakovic]

@rovka thanks for the hint!

[awarzynski]

Which linker does this worker/builder use? It might be worth making sure it's not bfd.

[stevanradakovic]

@awarzynski I really can't tell from the logs. here's the local build that passed: http://ex40-01.tcwglab.linaro.org:8019/#/builders/18/builds/81

[awarzynski]

In D98846#2634578, @stevanradakovic wrote:

@awarzynski I really can't tell from the logs. here's the local build that passed: http://ex40-01.tcwglab.linaro.org:8019/#/builders/18/builds/81

Np, from here:

-- Linker detection: GNU ld

That's the bfd linker (yup, the log _does not_ make it easy to find!). For GNU gold you would get:

-- Linker detection: GNU Gold

LLD is even better :)

There are at least two ways to change the linker used by a worker/builder:

• change the system linker (check which ld and update the corresponding symbolic link)
• use LLVM_USE_LINKER when configuring CMake (more info: Building LLVM with CMake

More on linker-related issues when building LLVM: https://llvm.org/docs/GettingStarted.html#common-problems. More specifically, Flang has quite demanding memory requirements. It is really worth making sure that one uses the most efficient linker.

I hope that this helps :)

[stevanradakovic]

Thanks Andrzej!

I've installed lld on teh worker and used "-DLLVM_ENABLE_LLD=ON" flag here.
So what about parallel_link_jobs? Should we keep this setting or not?

[awarzynski]

In D98846#2636651, @stevanradakovic wrote:

Thanks Andrzej!

I've installed lld on teh worker and used "-DLLVM_ENABLE_LLD=ON" flag here.
So what about parallel_link_jobs? Should we keep this setting or not?

This will depend on the number of cores and amount of RAM available. In general, I've never had to use LLVM_PARALLEL_LINK_JOBS. Switching to a more efficient linker has always been sufficient. But this depends on the spec of the machine.

Note that your current config is not working:

CMake Error at cmake/modules/HandleLLVMOptions.cmake:286 (message):
  Host compiler does not support '-fuse-ld=lld'
Call Stack (most recent call first):
  CMakeLists.txt:697 (include)

I couldn't check which version of GCC introduced support for -fuse-ld=lld, but GCC 9.3.0 should be fine. Could you check gcc -fuse-ld=lld file.c for e.g.:

int main() { return 0;}

This worked for me on X86 and AArch64 (tested with GCC 10.2.0 and GCC 9.3.0).

Alternatively, you could use:

• GCC + ld.gold
• Clang + LLD

[stevanradakovic]

When I install lld package on the worker the current config works. Do you think that will be ok?

[stevanradakovic]

Right, and I have also linked /usr/bin/ld to /usr/bin/lld. Is it too much of a hack?

[awarzynski]

In D98846#2637021, @stevanradakovic wrote:

Right, and I have also linked /usr/bin/ld to /usr/bin/lld . Is it too much of a hack?

Personally I don't think that this a hack, but it's worth documenting somewhere what the default linker on that machine is. In particular, there is not much here ATM: http://ex40-01.tcwglab.linaro.org:8019/#/workers/100 (see Admin and Host - please, could you update?).

I'm not a Ubuntu user (or expert) myself, but AFAIK, /usr/bin/ld is just a symlink that points to another symlink (e.g. /usr/bin/ld.bfd). And GCC will probably use e.g. /usr/bin/ld.lld when going through -fuse-ld=lld. I think that the bottom line is that if you rely on /usr/bin/ld, then any system update/upgrade may restore the original config. So you may want to rely on /usr/bin/ld.lld instead.

[stevanradakovic]

Use lld linker instead of the bfd. Remove parallel link jobs setting.

[stevanradakovic]

Ok this seems more sensible then. I'll make sure that the lld is installed on the worker side in our setup. Let me know if there's anything else. Thanks!

[stevanradakovic]

@awarzynski ping

[awarzynski]

This makes sense. Please update the commit message before merging (this has changed a bit since uploading) ;-)

[stevanradakovic]

Update commit msg.

buildbot/osuosl/master/config/builders.py

'factory' : UnifiedTreeBuilder.getCmakeWithNinjaBuildFactory(
                     clean=True,
                     checks=['check-flang'],
                     depends_on_projects=['llvm','mlir','clang','flang'],
                     extra_configure_args=[
                         "-DLLVM_TARGETS_TO_BUILD=AArch64",
                         "-DFLANG_BUILD_NEW_DRIVER=ON",
                         "-DCMAKE_BUILD_TYPE=Debug",
                         "-DCMAKE_CXX_STANDARD=17",
+                        "-DLLVM_USE_LINKER=lld",
                     ])},
 
     {'name' : "flang-aarch64-ubuntu-latest-clang",
     'tags'  : ['flang', 'aarch64', 'release'],
     'workernames' : ["linaro-aarch64-flang-latest-clang"],
     'builddir': "flang-aarch64-ubuntu-latest-clang",
     'factory' : UnifiedTreeBuilder.getCmakeWithNinjaBuildFactory(
                     clean=True,