Update `_CUDAToolkit_find_and_add_import_lib` to create and add
dependencies to a target in one step that shared a single guard against
repeated definitions. Otherwise we were adding dependencies again on
every call.
Fixes: #20282
In the case that we use ExternalProject_Add to drive the subdirectory
build, replace the `forcebuild` step with the official `BUILD_ALWAYS`
option.
Issue: #20179
45d21dd5d4 CUDAToolkit: Use CMAKE_FIND_ROOT_PATH for all sdk lib searches
e357772f20 CUDAToolkit: Use HINTS as it has higher precedence for searches
c6ec51c625 CUDAToolkit: functions names now use CMake's reserved namespace
Acked-by: Kitware Robot <kwrobot@kitware.com>
Merge-request: !4278
Previously the search in the dynamic linker paths 'LIB', 'LD_LIBRARY_PATH'
and 'DYLD_LIBRARY_PATH' was dependent on the value of the environment
variable 'MKLROOT'. If MKLROOT was given, the dynamic linker paths where
not searched. This seems slightly counter-intuitive.
This PR changes the behavior so that MKLROOT is searched first, but if
unsuccesful, the dynamic linker paths are tried as well.
Fixes#17559
Replace our hard-coded default of cudart=static with a first-class abstraction to select the runtime library from an enumeration of logical names.
QCC is a wrapper around GCC, but it is not a fully transparent wrapper.
Some compile options need to be passed to GCC using a `-Wc` option.
QCC does not support --sysroot, so setting CMAKE_SYSROOT in a toolchain
file currently does not work. This means that it is likely that no one
is setting CMAKE_SYSROOT in existing QNC toolchain files. Override the
GCC option for sysroot in the QCC.cmake file with -Wc,-isysroot.
This exposes a further issue in that the QNX SDK does not follow the
same architectural folder structure as linux uses. That is, on linux
systems, architecture-specific libraries might be in
<sysroot>/usr/lib/<arch>
such as
/usr/lib/x86_64-linux-gnu/libcurl.so
CMake models this by suffixing the <arch> onto lib directories when
searching for libraries.
The QNX SDK is structured differently such that the <arch> should be
used as a prefix:
<sysroot>/<arch>/usr/lib
such as
<sysroot>/x86_64/usr/lib/libcurl.so
Add a variable for platform configuration to set whether to prefix or
suffix the <arch> and set that in the QCC.cmake.
Use the directory structure of the QNX SDK to compute the <arch> from
the implicit library directories. The assumption is that the arch will
be a single directory directly below the CMAKE_SYSROOT, below which the
usr/ prefix occurs.
It would not be appropriate to instruct users to make the <arch> part of
the sysroot when specified in the toolchain file because:
1. That would be non-DRY - The QCC wrapper already determines the <arch>
by the -V argument passed to the compiler, specified in the toolchain
file as the CMAKE_C_COMPILER_TARGET variable.
2. The includes in the QNX SDK are not below the <arch> directory.
So, the location of the <arch> in the full path is different on QNX
compared to, say an embedded linux platform, but the intent is the same.
Add documentation to recommend the use of CMAKE_SYSROOT in a QNX
toolchain file.
As the CMAKE_SYSROOT is always the same for QNX, it would be possible to
simply set it in QCC.cmake. However, that would change behavior for
existing users as when CMAKE_SYSROOT is set, files/paths outside of the
CMAKE_SYSROOT do not get found.
The <arch> prefixing is only enabled in cmSearchPath.cxx if
CMAKE_SYSROOT is set. This ensures that the user gets consistency in
the current state without CMAKE_SYSROOT, and gets better consistency
when using CMAKE_SYSROOT.
`clang-tidy` does not infer driver mode if it is not provided with a
JSON compilation database. This is exactly the way cmake launches it.
Hence clang-tidy will only use the default driver mode. Add an explicit
driver mode argument to avoid this.
This introduces 2 new INTERFACE IMPORTED libraries: GTest::gtest and
GTest::gtest_main. They link to GTest::GTest and GTest::Main targets
respectively, therefore working as aliases. These new names map the
names of the targets from upstream GTest's CMake package config.
Fixes: #20255
48bd2e0630 CUDAToolkit: Make sure to also search 'stubs' directory
69fcad9332 CUDAToolkit: Add support for cross-compiling
Acked-by: Kitware Robot <kwrobot@kitware.com>
Merge-request: !4253
Using MATCHES "^${_Name}_FOUND$" limits the name of the found variable,
since it can't contain any regular expression special-chars (period,
plus sign, etc.)
3ec82b713e cmMarkAsAdvancedCommand: ignore variables which don't exist in the cache
701a5c60e0 cmake: avoid marking local or unused variables as advanced
af158149e7 FindOpenSSL: do not mark a local variable as advanced
74f659f1f2 FindCurses: only mark CURSES_EXTRA_LIBRARY when it is used
7e2ae4e96d FindOpenGL: only mark declared cache variables as advanced
7cc02a0c29 FindLua: only mark LUA_MATH_LIBRARY as advanced if required
85cd26b8a6 FindBoost: only mark Boost_DIR as advanced if defined
338c7916ba CTest: avoid marking undeclared cache variables as advanced
...
Acked-by: Kitware Robot <kwrobot@kitware.com>
Merge-request: !4150
Bring whitespace and code style up to date in these scripts. Both
scripts share the same origin but have diverged over time, so
synchronize them again. This is relevant because BLAS and LAPACK
detection is often performed simultaneously, so both scripts should
evolve in sync. While at it, update a few comments.
This update is intended to have no functional changes.
In commit fb3370b6a1 (MSVC: Add abstraction for runtime library
selection, 2019-04-10, v3.15.0-rc1~229^2) we overlooked populating the
runtime library selection flags for the Microsoft assembler. It does
not actually have any such flags, but since its compiler id is `MSVC`
our generators expect the table to be populated. Use empty values.
Without this fix, enabling the `ASM_MASM` language with policy `CMP0091`
set to `NEW` causes an error due to the missing table entries.
Fixes: #20236, #19453
The change in commit 5861c6d450 (FindMPI: Preserve include order when
extracting component directories, 2019-12-19) converted a CMake-language
function to a macro. However, it contains a `return()` call that now
applies in the caller's scope, which is in correct. The `return()` is
meant to skip work when using a MPI compiler wrapper as the main
compiler. Move that condition to the call site instead.
Fixes: #20234
