CUDA: Select CMAKE_CUDA_HOST_COMPILER default from CUDA bin directory

On my system, I have installed GCC 7 (at /usr/bin/{gcc,g++}) and GCC 6 (at /usr/bin/{gcc,g++}-6).

CUDA does not support GCC 7 so I've instructed it to use GCC 6 by setting up a symlink at $CUDA_ROOT/bin/{gcc,g++} which points to /usr/bin/{gcc,g++}-6.

This is enough to make nvcc work properly and this also worked with my pre-CMake3.8 find_package(CUDA) setup. The new first-class CUDA support however seems to ignore these symlinks and uses the wrong host compiler.

Now I can still compile the project by setting using the -DCMAKE_CUDA_HOST_COMPILER=/usr/bin/g++-6 flag, but isn't it possible to find this out automatically by looking at the symlink?

Arch Linux bug report