Original issue: https://github.com/RadeonOpenCompute/ROCm-CompilerSupport/issues/45
The source files are copied from https://github.com/RadeonOpenCompute/ROCm-CompilerSupport/
Simply install llvm/clang with clang and llvm-link in ${PATH} and run ./run.sh
Desired result: all commands run successfully. This can be achieved by using ROCm forked llvm
+ AMDGPU_ARCH=gfx803
+ echo 'Compile complete opencl code into object, step-by-step (source->bitcode->object)'
Compile complete opencl code into object, step-by-step (source->bitcode->object)
+ clang -cc1 -emit-llvm-bc -triple amdgcn-amd-amdhsa -target-cpu gfx803 -O3 -std=cl1.2 source.cl -o GPUBitcode.bc
+ clang -cc1 -emit-obj -triple amdgcn-amd-amdhsa -target-cpu gfx803 GPUBitcode.bc -o GPUkernel.o
+ echo 'Compile separated opencl code into object, step-by-step (source->bitcode->object)'
Compile separated opencl code into object, step-by-step (source->bitcode->object)
+ clang -cc1 -emit-llvm-bc -triple amdgcn-amd-amdhsa -target-cpu gfx803 -O3 -std=cl1.2 source1.cl -o GPUBitcode1.bc
+ clang -cc1 -emit-llvm-bc -triple amdgcn-amd-amdhsa -target-cpu gfx803 -O3 -std=cl1.2 source2.cl -o GPUBitcode2.bc
+ clang -cc1 -emit-obj -triple amdgcn-amd-amdhsa -target-cpu gfx803 GPUBitcode1.bc -o GPUkernel1.o
+ clang -cc1 -emit-obj -triple amdgcn-amd-amdhsa -target-cpu gfx803 GPUBitcode2.bc -o GPUkernel2.o
+ echo 'Compile separated opencl code into object, one-step (source->object)'
Compile separated opencl code into object, one-step (source->object)
+ clang -cc1 -emit-obj -triple amdgcn-amd-amdhsa -target-cpu gfx803 source1.cl -o GPUkernel1.o
+ clang -cc1 -emit-obj -triple amdgcn-amd-amdhsa -target-cpu gfx803 source2.cl -o GPUkernel2.o
+ echo 'linked opencl bitcode and then compile into object, (bitcode->object)'
linked opencl bitcode and then compile into object, (bitcode->object)
+ llvm-link GPUBitcode1.bc GPUBitcode2.bc -o GPUBitcode-linked.bc
+ clang -cc1 -emit-obj -triple amdgcn-amd-amdhsa -target-cpu gfx803 GPUBitcode-linked.bc -o GPUkernel-linked.o
Actual result, using vanilla llvm/clang :
+ AMDGPU_ARCH=gfx803
+ echo 'Compile complete opencl code into object, step-by-step (source->bitcode->object)'
Compile complete opencl code into object, step-by-step (source->bitcode->object)
+ clang -cc1 -emit-llvm-bc -triple amdgcn-amd-amdhsa -target-cpu gfx803 -O3 -std=cl1.2 source.cl -o GPUBitcode.bc
+ clang -cc1 -emit-obj -triple amdgcn-amd-amdhsa -target-cpu gfx803 GPUBitcode.bc -o GPUkernel.o
+ echo 'Compile separated opencl code into object, step-by-step (source->bitcode->object)'
Compile separated opencl code into object, step-by-step (source->bitcode->object)
+ clang -cc1 -emit-llvm-bc -triple amdgcn-amd-amdhsa -target-cpu gfx803 -O3 -std=cl1.2 source1.cl -o GPUBitcode1.bc
+ clang -cc1 -emit-llvm-bc -triple amdgcn-amd-amdhsa -target-cpu gfx803 -O3 -std=cl1.2 source2.cl -o GPUBitcode2.bc
+ clang -cc1 -emit-obj -triple amdgcn-amd-amdhsa -target-cpu gfx803 GPUBitcode1.bc -o GPUkernel1.o
fatal error: error in backend: Unsupported calling convention for call
+ clang -cc1 -emit-obj -triple amdgcn-amd-amdhsa -target-cpu gfx803 GPUBitcode2.bc -o GPUkernel2.o
+ echo 'Compile separated opencl code into object, one-step (source->object)'
Compile separated opencl code into object, one-step (source->object)
+ clang -cc1 -emit-obj -triple amdgcn-amd-amdhsa -target-cpu gfx803 source1.cl -o GPUkernel1.o
fatal error: error in backend: Unsupported calling convention for call
+ clang -cc1 -emit-obj -triple amdgcn-amd-amdhsa -target-cpu gfx803 source2.cl -o GPUkernel2.o
+ echo 'linked opencl bitcode and then compile into object, (bitcode->object)'
linked opencl bitcode and then compile into object, (bitcode->object)
+ llvm-link GPUBitcode1.bc GPUBitcode2.bc -o GPUBitcode-linked.bc
+ clang -cc1 -emit-obj -triple amdgcn-amd-amdhsa -target-cpu gfx803 GPUBitcode-linked.bc -o GPUkernel-linked.o
fatal error: error in backend: Unsupported calling convention for call