FindCUDAToolkit: Missing NVCC executable considered fatal
We use a custom cross-compilation toolchain and compile CUDA using Clang. While working on !4442 (merged) I noticed that
FindCUDAToolkit considers a missing
nvcc executable to be a fatal error. This doesn't seem reasonable as it's not strictly required for using rest of the toolkit. In our case we don't use it and would rather not ship deadweight in the sysroot.
Would simply checking the existence of a
bin directory be an acceptable fallback?