Optimizing cuda to using LLVM

Question

I want to write an LLVM Pass for Cuda codes. I stuck on creating executable from bitcodes.

I used

clang-10 -O0 -S  -emit-llvm --cuda-gpu-arch=sm_61 --c