Unable to compile intel/llvm on Linux Mint 20

[SylvainJoube]

Describe the bug

Unable to compile Intel/LLVM on branch sycl. Output for the build (at 97%) :

Scanning dependencies of target prepare-libspirv-nvptx64--nvidiacl.bc
[ 97%] Generating ../../lib/clc/libspirv-nvptx64--nvidiacl.bc
error: file too small to contain bitcode header
make[2]: *** [tools/libclc/CMakeFiles/prepare-libspirv-nvptx64--nvidiacl.bc.dir/build.make:62: lib/clc/libspirv-nvptx64--nvidiacl.bc] Error 1
make[1]: *** [CMakeFiles/Makefile2:77139: tools/libclc/CMakeFiles/prepare-libspirv-nvptx64--nvidiacl.bc.dir/all] Error 2
make[1]: *** Waiting for unfinished jobs....
make: *** [Makefile:152: all] Error 2

To Reproduce

Commands I used :

$ git clone https://github.com/intel/llvm
$ cd llvm
$ python ./buildbot/configure.py --cuda -t release --cmake-gen "Unix Makefiles"
$ cd build
$ ccmake .  # and disable the SYCL_ENABLE_WERROR flag
$ make install -j `nproc`

On branch sycl, at the following commit :

1811037bc8b3e9352b6806b24724eecb6d862895 (HEAD -> sycl, origin/sycl, origin/HEAD)
Date:   Sun Apr 25 14:11:44 2021 -0400).

Environment (please complete the following information):

OS : Linux Mint 20 (Ulyana) (based on Ubuntu focal)

Current compiler : gcc (Ubuntu 9.3.0-17ubuntu1~20.04) 9.3.0

Clang version :

clang version 10.0.0-4ubuntu1 
Target: x86_64-pc-linux-gnu
Thread model: posix
InstalledDir: /usr/bin

Target device and vendor : NVIDIA CUDA GPU

Installed CUDA version : cuda_11.3.r11.3/compiler.29745058_0

[bader]

Installed CUDA version : cuda_11.3.r11.3/compiler.29745058_0

FYI: CUDA 11 is not supported yet.

Output for the build (at 97%) :

Could you attach the full log, please? I suppose the problem happened before:

Scanning dependencies of target prepare-libspirv-nvptx64--nvidiacl.bc
[ 97%] Generating ../../lib/clc/libspirv-nvptx64--nvidiacl.bc
error: file too small to contain bitcode header

Tagging @steffenlarsen.

[SylvainJoube]

I thought CUDA 11 was fully supported yet, sorry, because @steffenlarsen just said on the IWOCL oneAPI Developer Summit (I'm currently following it) that his tests were working with CUDA 11 but I may have misunderstood what he was saying.

Here is the full log : cuda_11-3_mint-20_log.log

[bader]

The documentation still lists only CUDA 10.2: https://intel.github.io/llvm-docs/GetStartedGuide.html#build-dpc-toolchain-with-support-for-nvidia-cuda (although I'm not sure it's relevant to the issue). Our CI system uses CUDA 10.x to build DPC++ with CUDA support and there is no such issue there.

[steffenlarsen]

I thought CUDA 11 was fully supported yet, sorry, because @steffenlarsen just said on the IWOCL oneAPI Developer Summit (I'm currently following it) that his tests were working with CUDA 11 but I may have misunderstood what he was saying.

Here is the full log : cuda_11-3_mint-20_log.log

I am glad I made an impression! Yes, I am currently using CUDA 11.0, however we haven't yet gone as far as to mark it as "officially supported".

That said, I was able to reproduce your problem. Looks like using Unix Makefiles to build DPC++ for CUDA causes a problem in libclc. Exactly what the issue is still to be determined, but the quickest solution may be to use Ninja instead if possible.

[steffenlarsen]

If Ninja is not available, you can try and run the make command again. For me it succeeded after failing the first time. It smells a bit like a broken dependency in libclc.

[SylvainJoube]

Thank you Steffen, that helped me a lot. I installed Ninja and the build was successful (took my computer some time though).

So on my machine, following the Codeplay tutorial did not work. Instead of :

python ./buildbot/configure.py --cuda -t release --cmake-gen "Unix Makefiles"
cd build
make install -j `nproc`

I had to write :

python ./buildbot/configure.py --cuda -t release --cmake-gen "Ninja" --no-werror
cd build
ninja

thanks to your advice (and previously installed ninja-build since it was on the repos of my distrib). The --no-werror was necessary for the warnings not to cause errors. And that worked !

So thank you. I think the issue can be closed now. But it may be useful to inform the maintainers of the Codeplay page that there is this other way to compile llvm, for people not to have the same issue I had.

[steffenlarsen]

Great news, @SylvainJoube! I believe I found the issue and has a fix up (#3624) so if you ever feel like returning to using Unix Makefiles as the generator you should hopefully be able to.

Thank you for reporting this issue. I will discuss this with the tutorial maintainers so we avoid problems like this in the future. If you have any other issues, feel free to ping me directly on the issue. It saves @bader the trouble. 😄

[SylvainJoube]

Oh, great, thank you !

Got it, I'll remember it should I have another issue in the future :)