Users can create it through an explicit command-line option if desired.
Initializing the variable as an empty cache entry can wipe out a normal
variable of the same name that may have been set by a toolchain file.
Since commit v3.8.0-rc1~261^2~11 (CUDA: Use the host compiler for
linking CUDA executables and shared libs, 2016-09-19) we save the value
of `CMAKE_CUDA_HOST_COMPILER` persistently in the compiler information
file as a normal variable.
Fixes: #17935
The switch was not considering some languages, such as `ASM`.
Instead of memorizing the list of languages in the condition,
use a language specified by the includer.
Fixes: #17510
When no explicit `CMAKE_CUDA_HOST_COMPILER` is specified by the user
then we do not pass `-ccbin` to `nvcc`. In this case, nvcc's link line
we extract during the compiler identification step may not have the
absolute path to the host compiler it uses to drive linking. If it is
not absolute, use the `PATH=` from nvcc's output to search for it since
that is the one `nvcc` would use.
This fixes our internal `CMAKE_CUDA_HOST_LINK_LAUNCHER` value used to
construct link lines when using `CUDA` as the linker language. It needs
to match the host compiler `nvcc` uses internally during compilation.
Fixes: #17323
Do not assume that the implicit link line is the last line of the output
from `nvcc -v`. Instead first find the `LIBRARIES=` line, and then look
for that content on a later line. It appears twice. First on a call to
`nvlink`, which we ignore. Later it appears on the implicit link line.
Extract the latter line. On failure, abort with a `FATAL_ERROR` so that
the user does not try to build without proper link information.
Once we have the line, parse it with `separate_arguments` using the
`UNIX_COMMAND` option just like `CMakeParseImplicitLinkInfo` already
does. This robustly parses the command line and removes quoting.
Then extract the first argument as the host link launcher.
