[ROCm] Problem compiling for AMD GPUs

[rjfnobre]

The following error is produced when attempting to compile SYCL applications for AMD GPUs (tested on two different systems and with different SYCL applications, including with the simple_program example application from Getting Started with oneAPI DPC++):

llvm-foreach: No such file or directory
clang-14: error: amdgcn-link command failed with exit code 1 (use -v to see invocation)

Command used:

clang++ -fsycl -fsycl-targets=amdgcn-amd-amdhsa-sycldevice   -Xsycl-target-backend --offload-arch=gfx906  simple-sycl-app.cpp -o simple-sycl-app-amd.exe

The DPC++ compiler (git describe: 2020-12-23532-gd1ed8a1) has been compiled for the ROCm target, following the instructions in:
https://intel.github.io/llvm-docs/GetStartedGuide.html#build-dpc-toolchain-with-support-for-amd-rocm

This is a bug, right?

[alexbatashev]

@rjfnobre which target did you run to build the compiler? ninja sycl-toolchain? Can you try ninja llvm-foreach?

[rjfnobre]

The following commands have been used to build the compiler (as per the instructions in the Getting Started with oneAPI DPC++ guide) :

python $DPCPP_HOME/llvm/buildbot/configure.py --rocm
python $DPCPP_HOME/llvm/buildbot/compile.py

Executing ninja llvm-foreach results in the following output:

(base) bash-4.2$ ninja llvm-foreach
ninja: no work to do.

Executing ninja sycl-toolchain compiles the level-zero loader and copies the SYCL headers, which seems to be what should happen.

[npmiller]

The problem is that lld is missing, with AMD lld is used for linking, we should probably fix the buildbot script to add that and make the dependency clearer.

Adding lld to LLVM_ENABLE_PROJECTS, then building it with ninja lld should solve this issue.