In commit 4620cf77f2 (Clang: Remove unused CUDA implicit link variables,
2021-02-14) we removed some references. It turns out they are non-empty
and necessary if using a non-scattered installation.
Fixes: #21863
Each source compilation generates a dependencies file. These dependencies
files are consolidated in one file per target. This consolidation is done
as part of command 'cmake -E cmake_depends` launched before evaluation of
makefile dependency graph.
The consolidation uses the same approach as `CMake` dependencies management.
Fixes: #21321
These changes are in preparation of compiler generated dependencies support
for Makefiles generators
* compiler output and dependency target can be different for Makefiles generators
* resolve inconsistency naming for dependency file place-holder
For NVCC the compiler takes care of device linking when passed the "-dlink"
flag.
Clang doesn't support such magic and requires the buildsystem to do the work
that NVCC does behind the scenes.
The implementation is based on Bazel's device linking documentation:
7cabcdf073/third_party/nccl/build_defs.bzl.tpl (L259)Closes: #20726
__compiler_clang() doesn't call __compiler_gnu() if we're emulating MSVC. Thus
CMAKE_DEPFILE_FLAGS_CUDA remains unset and compiling doesn't work, due to NVCC
dependency injection workaround in CMakeCUDAInformation.cmake, which triggers
for Ninja if they're not set.
Always set the depfile flags to fix this. Most other compiler modules seem to
do the same.
When crosscompiling we pass the sysroot.
We need to try various architecture flags. Clang doesn't automatically
select one that works. First try the ones that are more likely to work
for modern installations:
* <=sm_50 is deprecated since CUDA 10.2, try sm_52 first for
future compatibility.
* <=sm_20 is removed since CUDA 9.0, try sm_30.
Otherwise fallback to Clang's current default. Currently that's `sm_20`,
the lowest it supports.
Separable compilation isn't supported yet.
Fixes: #16586
