jchlanda wrote: This patch serves the same purpose as: https://github.com/llvm/llvm-project/pull/149294, which is to try and simplify how we use clang as a tool when offloading to cuda/hip/sycl. https://github.com/llvm/llvm-project/pull/151114