[SYCL] [CUDA] libclang_rt.profile-x86_64.a not found when compiling package in Debug

[cgleggett]

When building a package using clang++ from dpcpp/llvm with -DCMAKE_BUILD_TYPE=Debug, I'm seeing this error:

/usr/bin/ld: cannot find /opt/dpcpp/2020.09.21-cuda-9db44f8551a/lib/clang/12.0.0/lib/linux/libclang_rt.profile-x86_64.a

dpcpp was built with

buildbot/configure.py -s ../llvm -o . -t release --no-werror --shared-libs --cmake-opt=-DCMAKE_INSTALL_PREFIX=/opt/dpcpp/2020.09.21-cuda-9db44f8551a --cuda

I don't see libclang_rt.profile-x86_64.a anywhere in the llvm build directory, so it's not that it was built but not installed. In fact the whole /opt/dpcpp/2020.09.21-cuda-9db44f8551a/lib/clang/12.0.0/lib directory does not exist.

Is there another parameter that needs to be given to buildbot to build this?

The version of dpcpp is from git commit hash 9db44f8, but this seems the same in versions going back to March. I do see this library in the beta-0X packages however.

[vrpascuzzi]

Hi, @cgleggett.
To hopefully answer part of your question, I usually cp -r lib include bin ${LLVM_INSTALLDIR}/ as, indeed, the "installation" doesn't put all the required files. This, of course, also copies some build files -- e.g. .o, .spv -- as well (you can be smarter about the cp).

(P.S. It looks like /opt/dpcpp/2020.10.02-cuda-8149957a819a is better now.)

[cgleggett]

Actually, it wasn't being built at all, as some cmake option was missing. For reference, here's the magic sauce:

> cmake -DCMAKE_BUILD_TYPE=Release -DCMAKE_INSTALL_PREFIX=/opt/dpcpp/2020.10.02-cuda-8149957a819a \
-DLLVM_ENABLE_PROJECTS=clang;clang-tools-extra;compiler-rt;llvm-spirv;sycl;libclc \
-DLLVM_EXTERNAL_PROJECTS=sycl;llvm-spirv;opencl-aot;xpti;libdevice \
-DLLVM_EXTERNAL_SYCL_SOURCE_DIR=/opt/install/intel/llvm/sycl \
-DLLVM_EXTERNAL_LLVM_SPIRV_SOURCE_DIR=/opt/install/intel/llvm/llvm-spirv \
-DLLVM_EXTERNAL_LIBDEVICE_SOURCE_DIR=/opt/install/intel/llvm/libdevice \
-DLLVM_TARGETS_TO_BUILD=X86;NVPTX -DSYCL_BUILD_PI_CUDA=ON \
-DLIBCLC_TARGETS_TO_BUILD=nvptx64--;nvptx64--nvidiacl \
-DLLVM_ENABLE_EH=ON -DLLVM_ENABLE_PIC=ON -DLLVM_ENABLE_RTTI=ON \
-DOpenCL_INCLUDE_DIRS= -DOpenCL_LIBRARIES= -DOpenCL_INSTALL_KHRONOS_ICD_LOADER=TRUE \
-DGCC_INSTALL_PREFIX=/opt/gcc/8.4.0 \
-DCUDA_TOOLKIT_ROOT_DIR=/opt/cuda/10.1.243 -G Ninja \
/opt/install/intel/llvm/llvm

> ninja ocl-headers && ninja ocl-icd && ninja && ninja libsycldevice
> ninja install

[sebastian-freitag]

compiler-rt in LLVM_ENABLE_PROJECTS is the one that will provide libclang_rt.profile-x86_64.a
See here:
https://compiler-rt.llvm.org/