In commit bb61c2d024 (Clang: use -imsvc for system include dirs when
running on Windows, 2020-09-16, v3.19.0-rc1~162^2) we added `-imsvc`
for all Clang compilers targeting the MSVC ABI. However, the option
only exists for the MSVC-like front-end. The GNU-like front-ends
use `-isystem`.
Fixes: #21789
Fixes#21750, #21763
Given that NVCC can be provided by multiple different sources (NVIDIA HPC SDK, CUDA Toolkit, distro)
each of which has a different layout, we need to extract the CUDA toolkit root from the compiler
itself, allowing us to support numerious different scattered toolkit layouts.
The NVIDIA HPC SDK specifically ships two copies of nvcc one in
`compilers/bin/` and one in `cuda/bin`. Thus when using
`compilers/bin/nvcc` the Toolkit root logic fails.
This module has been discouraged since commit 1e20be319f (STYLE: add
documentation for MACRO_ADD_FILE_DEPENDENCIES(), 2009-07-14,
v2.8.0~433). It has been listed in the `cmake-modules(7)` manual
as deprecated since commit df780bcc01 (Help: Move deprecated modules to
appropriate section., 2018-11-12, v3.14.0-rc1~374^2). Add the explicit
deprecation mark in its documentation.
Xcode 12 doesn't allow nested builds within the same build directory.
That means we can no longer do an install by building the install target
when IOS_INSTALL_COMBINED is true. We can, however, still do an install
by running the cmake_install.cmake script or executing cmake --install,
since there is no outer build and therefore the associated SDK can be
built as a sub-build.
The non-build methods previously didn't work when
IOS_INSTALL_COMBINED was true because the generated install script
and the CMakeIOSInstallCombined script both made certain assumptions
that relied on being part of a build. Those assumptions are now
removed. A side-effect of this work is that cpack now also works from the
command line when IOS_INSTALL_COMBINED is true.
Relates: #21282Fixes: #20023
In 315a200f0c (FindGit: Cache the GIT_EXECUTABLE version for the
current run, 2021-01-20), the GIT_VERSION_STRING was meant to be
cached after the first time it was computed for a given GIT_EXECUTABLE
location. That logic assumed GIT_VERSION_STRING would be visible in
the current scope, but it might not be. The global property alone is
enough to check whether the version has been determined previously,
so don't switch the logic based on whether GIT_VERSION_STRING is
defined or not.
Relates: #21703
The cost of setting up and executing a separate sub-build to do the
download, update and patch steps required for FetchContent population
can be significant with some platforms and CMake generators. Avoid the
sub-build altogether by invoking the step scripts directly.
Previously, if no generator was set (e.g. population was being done in
script mode), a generator needed to be available on the default PATH.
Since we no longer use a sub-build, this restriction is also now gone.
Fixes: #21703
The mkdir, download, update and patch steps are used by
FetchContent during the configure phase of the main build. Because
these steps need a target, this has so far required a sub-build to be
set up. The changes here factor out the preparation of the scripts
from the creation of the targets, allowing future work to leverage these
steps without a sub-build (see #21703).
As part of the refactoring, some rationalisation of the stamp files,
repository info files and script names was done to make things more
consistent between download methods and step implementations.
Every download method now records its own specific repository info
in a file and that file is a dependency of the download step. The source
directory is also written to that file, so if the SOURCE_DIR changes, the
download will be retriggered (the existing implementation fails in this
scenario). Each download method now also has just one driver script
that implements the whole step (it may pull in other scripts to do its
task though). The patch step gained support for USES_TERMINAL as
a result of generalising the implementation for custom commands.
Fixes: #21748
Teach CMake how to extract `CMAKE_<LANG>_LIBRARY_ARCHITECTURE` from
versioned paths such as `/usr/lib/gcc/x86_64-linux-gnu/9`. These kind of
paths are generated by NVHPC compilers.
Update the list of known versions.
Run the command
cmake -DBOOST_DIR=/path/to/boost_1_75_0 \
-P Utilities/Scripts/BoostScanDeps.cmake
to extract dependencies from the 1.75.0 source tree.
They differ from 1.74's dependencies by the addition of
dependencies of the json component, so add a new version
block to FindBoost.
Fixes: #21773
Refactoring in commit 7cc815a2a6 (CUDAToolkit: Detect CUDA SDK that
don't have nvcc, 2020-07-24, v3.19.0-rc1~366^2) accidentally broke
use of the `CUDA_PATH` environment variable.
Fixes: #21740
The previous implementation was scanning the documentation in
the file at runtime to determine the set of supported keywords for
each public function. This was fragile, made it difficult to restructure
the documentation and was sometimes observable in runtime
performance measurements. Change to a more conventional
approach where supported keywords are explicitly listed in the
code.
The internal _ExternalProject_SELF variable is no longer needed.
CMake now provides CMAKE_CURRENT_FUNCTION_LIST_DIR which
can be used for the same purpose and avoids having to set a
variable when the module is read. This also removes the
requirement that the module must be included by the current or a
parent scope. It is now enough that the module has been included
once somewhere before calling any of its functions.
The above changes combined mean that the module can now use
include_guard() and avoid having to re-parse the very long file every
time.
ac6a4d4884 ExternalProject: Improve robustness of update step
17c4c8b92b Tests: Prevent the noisy CMP0114 warnings in ExternalProjectUpdate test
1cb65e680d ExternalProject: Prevent the noisy detached head messages on checkout
Acked-by: Kitware Robot <kwrobot@kitware.com>
Merge-request: !5389
Refactor the update logic to make it easier to follow. The following
fixes/improvements are some consequences of this change:
* Absorb a confusing git checkout failure message when the failure
is allowed and we act on that failure appropriately.
* Fix an unnecessary fetch in some scenarios when checking out a
git hash we already have locally.
* Stash and restore any local changes even when not rebasing.
* Avoid unsafe rebasing where we are not on a branch that is
already tracking the requested branch.
* When fetching, use --tags --force to ensure we get all the tags
and commits leading up to them regardless of whether the tags
are on branches or not. Also update our local tags if they move
on the remote.
Fixes: #20677
When checking out a specific commit by its git hash, git will output
a very noisy message about checking out a detached HEAD. This
is not particularly helpful for us here and makes the output overall
quite verbose. Add a git config setting to prevent it.
Using a single ID 'IntelLLVM' for the suite of Intel compilers based on
the LLVM backend. The 'IntelLLVM' ID are used for C, C++, and Fortran.
Data Parallel C++ will be handled in a separate commit.
The C and C++ definitions are based on the Clang definitions. The Intel
LLVM-based C and C++ compilers are based on the Clang front end, so
existing Clang options are more likely to be a good match than options
for the older Intel compilers.
Fortran is based on the older Fortran front end with the LLVM backend.
It has a similar interface to the older versions, though many options
are shared with the C and C++ compilers.
Fixes: #21561
Signed-off-by: William R. Dieter <william.r.dieter@intel.com>
Identify the compilers as `NVHPC` to distinguish it from the older PGI
compilers from which they evolved, and from other `NVIDIA` compilers.
Fixes: #20887
