Since commit v3.12.0-rc1~278^2 (CUDA: Pass more link libraries to device
linking, 2018-03-27) we consider every link item during device linking.
However, items that start in `-` may be host-specific link flags that
nvcc will not understand during device linking. Filter such items using
a white list.
In particular, this allows `-pthread` to be used for host linking while
not polluting the device link line.
Issue: #18008
The nvcc device linker is designed so that each static library
with device symbols only needs to be listed once as it doesn't
care about link order. If you provide the same static library
multiple times it will error out. To make sure this occurs
we find the unique set of link items.
Previously we dropped non-target items from the device link line because
nvcc rejects paths to shared library files, and only with target items
do we know the kind of library. However, this also prevents projects
from linking to system-provided libraries like `cublas_device` that
contain device code.
Fix this by passing more link items to device linking. Items that are
not file paths, such as `-lfoo`, can simply be passed unconditionally.
Items that are targets known to be shared libraries can still be
skipped. Items that are paths to library files can be passed directly
if they end in `.a`. Otherwise, pass them using `-Xnvlink` to bypass
nvcc's front-end. The nvlink tool knows to ignore shared library files.
Issue: #16317
If a static library has the property CUDA_RESOLVE_DEVICE_SYMBOLS enabled
it will now perform the device link step. The normal behavior is
to delay calling device link until the static library is consumed by
a shared library or an executable.
Previously we had a two issues when building cuda executables
that required separable compilation. The first was that we didn't
propagate FLAGS causing any -arch / -gencode flags to be dropped, and
secondly generators such as ninja would use the CXX language flags
instead of CUDA when the executable was mixed language.
