I have a silly question for some possible future work. It looks to me like Apple uses LLVM for both OpenGL and OpenCL code generation. Does that mean that there are GPU back-ends available for use by front-ends? I looked in llvm/lib/Target and saw a Cell SPU target but nothing I recognized as targeting a GPU.
If there is a prospect for IR -> GPU compilation, that would allow for some...interesting possibilities in the future.