LLVM is really powerful. 250 lines to create a basic but quite functional replacement for NVCC. Most work is replacing the CUDA stdlib, which also isn't that hard. NVPTX generates similar code to NVCC if you force it to, but LLVM will hapily reject optimizations such as loop unrolling & load store vectorization for dubious reasons such as "oh no the number of loop executions is uncomputable" where is NVCC is like "lol I will assume k is at least 4" out of thin air to yield code that hits fma optimization passes.