The shim header approach would still have the benefit of being smaller and thus Clang being able to parse the version from it faster.
An upstream Clang change (<https://reviews.llvm.org/D110596>) has been merged to
make symlinking /usr/lib/cuda/include to /usr/include work.
- Bug#995122: Clang and /usr/lib/cuda Raul Tambre
- Bug#995122: cuda.h shim Raul Tambre
- Bug#995122: LLVM change Raul Tambre
- Bug#995122: LLVM change Andreas Beckmann
- Bug#995122: LLVM change Raul Tambre
