mirror of
https://github.com/llvm/llvm-project.git
synced 2025-04-17 07:26:41 +00:00
505e049aa0
LLD has a bug regarding ordering of static link libraries in the ELF backend, which has been reported as #116669. CMake 3.31.0 started properly deduplicating static libraries for LLD causing the following linking failure for `libclang-cpp.so` with `-DLLVM_LINK_LLVM_DYLIB=ON`: ``` ld.lld: error: undefined symbol: llvm::omp::getOpenMPClauseName(llvm::omp::Clause) >>> referenced by OpenMPKinds.cpp >>> tools/clang/lib/Basic/CMakeFiles/obj.clangBasic.dir/OpenMPKinds.cpp.o:(clang::getOpenMPSimpleClauseTypeName(llvm::omp::Clause, unsigned int)) >>> referenced by SemaOpenMP.cpp >>> tools/clang/lib/Sema/CMakeFiles/obj.clangSema.dir/SemaOpenMP.cpp.o:(clang::SemaOpenMP::CheckOMPRequiresDecl(clang::SourceLocation, llvm::ArrayRef<clang::OMPClause*>)) >>> referenced by SemaOpenMP.cpp >>> tools/clang/lib/Sema/CMakeFiles/obj.clangSema.dir/SemaOpenMP.cpp.o:(clang::SemaOpenMP::CheckOMPRequiresDecl(clang::SourceLocation, llvm::ArrayRef<clang::OMPClause*>)) >>> referenced 166 more times [tons more] ``` CMake 3.31 also introduced CMP0179, which builds on CMP0156 and makes the deduplication consistent across platforms. By coincidence this works around the above LLD deficiency and is the fix that CMake 3.31.1 will implement. However, the fix is to ignore CMP0156 unless CMP0179 is also enabled, i.e. no more deduplication. So enable CMP0179 to keep the benefits of deduplication from CMP0156 on LLD and fix the build for CMake 3.31.0. See: #116669 See: https://gitlab.kitware.com/cmake/cmake/-/issues/26447 Fixes: cb90d5b
=======================
LLVM Common CMake Utils
=======================
What goes here
--------------
These are CMake modules to be shared between LLVM projects strictly at build
time. In other words, they must not be included from an installed CMake module,
such as the ``Add*.cmake`` ones. Modules that are reachable from installed
modules should instead go in ``${project}/cmake/modules`` of the most upstream
project that uses them.
The advantage of not putting these modules in an existing location like
``llvm/cmake/modules`` is two-fold:
- Since they are not installed, we don't have to worry about any out-of-tree
downstream usage, and thus there is no need for stability.
- Since they are available as part of the source at build-time, we don't have
to do the usual stand-alone vs combined-build dances, avoiding much
complexity.
How to use
----------
For tools, please do:
.. code-block:: cmake
if(NOT DEFINED LLVM_COMMON_CMAKE_UTILS)
set(LLVM_COMMON_CMAKE_UTILS ${CMAKE_CURRENT_SOURCE_DIR}/../cmake)
endif()
# Add path for custom modules.
list(INSERT CMAKE_MODULE_PATH 0
# project-specific module dirs first
"${LLVM_COMMON_CMAKE_UTILS}/Modules"
)
Notes:
- The ``if(NOT DEFINED ...)`` guard is there because in combined builds, LLVM
will set this variable. This is useful for legacy builds where projects are
found in ``llvm/tools`` instead.
- ``INSERT ... 0`` ensures these new entries are prepended to the front of the
module path, so nothing might shadow them by mistake.
For runtime libs, we skip the ``if(NOT DEFINED`` part:
.. code-block:: cmake
set(LLVM_COMMON_CMAKE_UTILS ${CMAKE_CURRENT_SOURCE_DIR}/../cmake)
... # same as before
If ``llvm/tools`` legacy-style combined builds are deprecated, we should then
skip it everywhere, bringing the tools and runtimes boilerplate back in line.