The Ninja generator has loosened the dependencies of object compilation. Object compilation now depends only on custom targets and custom commands associated with libraries on which the object’s target depends and no longer depends on the libraries themselves. Source files in dependent targets may now compile without waiting for their targets’ dependencies to link.
With BUILD_SHARED_LIBS, compiling units don’t wait for preceding shared libs.
See also; http://bb.pgr.jp/builders/i686-mingw32-RA-on-linux
Regardless of BUILD_SHARED_LIBS, compile units in add_executable() don’t wait for preceding libraries,
but targets by add_dependencies().
It doesn’t break anything in llvm tree. Assume;
add_executable(foo foo.cpp)
target_link_libraries(foo LLVMCore) # depends on intrinsics_gen
Compiling foo.cpp doesn’t wait for LLVMCore, but intrinsics_gen.
Linking foo waits for LLVMCore.
I have been working for cutting dependencies to increase parallelism.
For example, I introduced ENABLE_OBJLIB.
See also, https://reviews.llvm.org/rL305635
Ninja with CMake-3.9 doesn’t require parallelize with objlib.
This is great news! Do we know who contributed the changes to cut the extra library dependencies?
Do you think we should remove ENABLE_OBJLIB to simplify our CMake files in the near future? It seems to me that anyone who cares about highly parallel build throughput can upgrade CMake to get the good behavior. It’s probably easier and less error-prone than maintaining a special build configuration.
This is great news! Do we know who contributed the changes to cut the extra library dependencies?
Do you think we should remove ENABLE_OBJLIB to simplify our CMake files in the near future? It seems to me that anyone who cares about highly parallel build throughput can upgrade CMake to get the good behavior. It’s probably easier and less error-prone than maintaining a special build configuration.
At the moment, this facility (loose deps) is specific to Ninja generator.
There are a couple of users of ENABLE_OBJLIB.
Tablegen (except for ninja).
LIBCLANG_BUILD_STATIC. It doesn’t use LLVM_ENABLE_OBJLIB, but uses it internally.
Seen as this is related to 3.9. What are your thoughts on how to update the variables to solve this
Will we just add the policy to cmake or change RPATH on macos behave differently?
CMake Warning (dev):
Policy CMP0068 is not set: RPATH settings on macOS do not affect
install_name. Run "cmake --help-policy CMP0068" for policy details. Use
the cmake_policy command to set the policy and suppress this warning.
For compatibility with older versions of CMake, the install_name fields for
the following targets are still affected by RPATH settings:
LTO
libclang
This warning is for project developers. Use -Wno-dev to suppress it.
This is from an MacOS host building clang with cmake 3.9 and ninja
Looks like subsequent calls to cmake_minimum_required(VERSION 3.4.3) resets policies. You can test this by duplicating the call in llvm/CMakeLists.txt after setting a policy.
Perhaps clang, et al, including libcxx, etc., should do something like this:
Clang already does this check, so moving the call to cmake_minimum_required inside the check would help.
Also, it might be a good idea to remove all the obsolete cmake_policy calls that set the policy to NEW, since that’s the default once the minumum version has been set. Perhaps we could even put all of this in a macro that all projects could call without the need to duplicate the same logic.