It only makes sense to use the CMake package from the same ROCm
installation that the compiler uses. Ask the HIP compiler to report the
location of the ROCm installation. Verify up front that it contains the
expected CMake package file.
Since commit bd844387df (ROCMClang: Add the ROCm toolkit derived clang
compiler to CMake, 2020-08-28, v3.21.0-rc1~66^2~6) and commit ff0d2858e1
(HIP: Extract clang compiler details from hipcc, 2020-10-21,
v3.21.0-rc1~66^2~5), the separate `ROCMClang` compiler id for `hipcc`
has caused a few problems:
* The compiler id changed from behavior of CMake 3.20 and below,
breaking projects that already built with `hipcc` treated as `Clang`.
* The implementation of `target_compile_features` was incomplete for
the `ROCMClang` identity.
* Only `hipcc` was identified as `ROCMClang`, so after it is unwrapped
to the underlying `clang++`, future runs of new CMake versions on
an existing build tree would not repeat this.
* Clang should be usable as a HIP compiler without the `hipcc` wrapper.
Remove the `ROMClang` compiler identity, and revise HIP language support
to work directly with a Clang compiler.
Reject direct `hipcc` usage as a HIP compiler. For now it cannot be
supported because it interferes with flags CMake needs to pass to Clang.
Fixes: #22536, #22460, #22593
Our `find_program` calls specify `HINTS` to look in the toolchain's
directory first, and then in `PATH`. `CMAKE_PREFIX_PATH` may be
specified by the user to help find packages for project dependencies,
but this should not override the binutils.
Fixes: #22512
The search logic for the IAR linker, librarian and related tools is
moved from `Modules/Compiler/IAR-FindBinUtils.cmake` to
`Modules/CMakeFindBinUtils.cmake` by introducing an new elseif block
for the IAR toolchain.
The search logic was refactored to omit repeating itself.
Fixes: #22425
IAR is available for both Linux and Windows OSes. Moreover,
binaries `*.exe` could be called from a Linux box, or vice versa.
See also commit 01a4eec446 (IAR: Changes required for Linux,
2019-04-25, v3.15.0-rc1~185^2).
Fixes: #22312
Given the compiler to use, `CMakeFindBinUtils.cmake` automatically
determines a number of tools including linker (CMAKE_LINKER) and archiver
(CMAKE_AR) and stores them in a generated file `CMakeCCompiler.cmake` as
non-CACHE entries. The compiler-specific ARMClang.cmake then tries to
override CMAKE_LINKER and CMAKE_AR as CACHE entries.
Following the introduction of CMP0126, which is set to NEW in the test
for a working compiler, setting a CACHE entry does not replace a normal
entry of the same name anymore, resulting in a failed test due to wrong
linker and archiver.
To fix this, set CMAKE_LINKER and CMAKE_AR for ARMClang directly in
`CMakeFindBinUtils.cmake` as is done for other compilers. Check
for them in `ARMClang.cmake` to safeguard cases when a project explicitly
includes `ARMClang.cmake` prior to compiler determination (which some
projects do to work around other problems in older CMake versions).
Update the IAR linker and archiver rules to use the `<CMAKE_LINKER>`
and `<CMAKE_AR>` placeholders instead of hard-coding the tool names.
Fixes: #22395
In commit a90d2a9eed (IntelLLVM: Add support for Intel LLVM-based
compilers, 2020-11-02, v3.20.0-rc1~89^2~20) we accidentally left out
activation of the C/C++ standard level selection logic when IntelLLVM is
targeting the MSVC ABI.
Fixes: #22388
Microsoft intends to ship the final C++20 flag in VS 16.11 albeit with a few
parts missing due planned ABI breaking changes.
The current 16.11 Preview 1 toolchain version is 19.29.30129.3, so let's
restrict based on that.
The commit 98fea8205e (Compiler/TI: Avoid response file usage for
linker, 2020-07-11, v3.19.0-rc1~495^2) disabled linker file usage by
default. The previous settings were working, even if not for all cases.
Restore them and add an explanation in a comment.
Issue: #22233
Tell the Fortran compiler to write preprocessor output directly to a
file, as we do for the GNU compiler. The previous "redirect stdout"
approach could break checks using flags that add information to stdout
when called with `-###`.
Fixes: #22156
9c479c7c40 IntelLLVM: Add special case for ifx 2021.1 version extraction
b7193ab18f Intel: Update Classic compiler version detection for 2021
Acked-by: Kitware Robot <kwrobot@kitware.com>
Merge-request: !6052
c4941b7e66 ARMClang: Do not automatically add cpu/arch compile or link options
0078db3888 ARMClang: Separate cpu/arch flags from preceding flags
Acked-by: Kitware Robot <kwrobot@kitware.com>
Acked-by: Jaeden Amero <kitware@patater.com>
Merge-request: !6035
The compile options `--march=<arch>` and `--mcpu=<cpu>` and the
link option `--cpu=<cpu>` are automatically added by CMake based
on `CMAKE_SYSTEM_PROCESSOR` or `CMAKE_SYSTEM_ARCH`. But this is not
sufficient, because armclang also supports enabling or disabling
features using `+<feature>`:
-mcpu=<name>[+[no]<feature>+...]
For example:
-mcpu=cortex-a57+nocrypto+nofp+nosimd+crc
(Reference: https://developer.arm.com/documentation/dui0774/k/Compiler-Command-line-Options/-mcpu?lang=en)
The problem is, even if a project adds a flag with features it needs,
CMake still adds flags, resulting in code that is compiled with wrong
CPU features and unable to run.
Add policy `CMP0123` to not automatically add compile or link options,
and let projects set them instead.
Co-Author: Brad King <brad.king@kitware.com>
Fixes: #21173
Since commit 2c71d051fa (Makefiles Generators: use compiler for
dependencies generation, 2020-10-18, v3.20.0-rc1~392^2) we invoke `nvcc`
for CUDA < 10.2 a second time in order to generate a depfile. When
`CMAKE_CUDA_HOST_COMPILER` is set, the second invocation is missing its
`-ccbin=` option, even after refactoring in commit 8981e3e7cc
(NVIDIA-CUDA: rely on new capabilities for deps generation, 2020-12-02,
v3.20.0-rc1~362^2).
Ideally we should move the `-ccbin=` flag into `Compiler/NVIDIA-CUDA`,
but that will add `CMAKE_CUDA_HOST_COMPILER` support on Windows in
command-line generators but not the Visual Studio generators.
For now, add the flag to the depfile command specifically.
Fixes: #22037
This should be front end compatible with vanilla clang but giving it a
unique identifier allows a project to pass additional options unique to
Fujitsu and outside the scope of a CMake builtin.
Brad King