[mlir] -DLLVM_DISTRIBUTION_COMPONENTS=mlir_c_runner_utils seems to unnecessarily require installing MLIRSparseTensorRuntime

[mgorny]

I'm finally working on packaging MLIR for Gentoo. The current main seems to work really great, for standalone builds and dylib installation both. However, there's one thing that seems wrong.

I'm using -DLLVM_DISTRIBUTION_COMPONENTS to skip installing the static libraries. However, if I include mlir_c_runner_utils shared library in them, CMake fails:

-- Configuring done (5.1s)
CMake Error: install(EXPORT "MLIRTargets" ...) includes target "mlir_c_runner_utils" which requires target "MLIRSparseTensorEnums" that is not in any export set.
CMake Error: install(EXPORT "MLIRTargets" ...) includes target "mlir_c_runner_utils" which requires target "MLIRSparseTensorRuntime" that is not in any export set.
-- Generating done (3.4s)
CMake Generate step failed.  Build files cannot be regenerated correctly.

Please correct me if I'm wrong, but if I understand correctly, MLIRSparseTensorRuntime is only used as an implementation detail of mlir_c_runner_utils and isn't used otherwise. So I don't think it should be necessary to install the static library at all, and something is going wrong within the build system dependencies.

My current (work-in-progress) CMake invocation is the following:

cmake -C /tmp/portage/llvm-core/mlir-20.0.0_pre20241221/work/mlir_build/gentoo_common_config.cmake -G Ninja -DCMAKE_INSTALL_PREFIX=/usr -DCMAKE_INSTALL_PREFIX=/usr/lib/llvm/20 -DLLVM_ROOT=/usr/lib/llvm/20 -DBUILD_SHARED_LIBS=OFF -DLLVM_BUILD_LLVM_DYLIB=ON -DMLIR_BUILD_MLIR_C_DYLIB=OFF -DMLIR_LINK_MLIR_DYLIB=ON -DMLIR_INCLUDE_TESTS=no -DLLVM_DISTRIBUTION_COMPONENTS=mlir-cmake-exports;mlir-headers;MLIR;mlir_arm_runner_utils;mlir_arm_sme_abi_stubs;mlir_async_runtime;mlir_c_runner_utils;mlir_float16_utils;mlir_runner_utils;mlir-cpu-runner;mlir-linalg-ods-yaml-gen;mlir-lsp-server;mlir-opt;mlir-pdll;mlir-pdll-lsp-server;mlir-query;mlir-reduce;mlir-rewrite;mlir-tblgen;mlir-translate;tblgen-lsp-server;tblgen-to-irdl; -DLLVM_BUILD_UTILS=ON -DPython3_EXECUTABLE=/usr/bin/python3.13 -DLLVM_BUILD_TOOLS=ON -DMLIR_ENABLE_CUDA_RUNNER=OFF -DMLIR_ENABLE_ROCM_RUNNER=OFF -DMLIR_ENABLE_SYCL_RUNNER=OFF -DMLIR_ENABLE_SPIRV_CPU_RUNNER=OFF -DMLIR_ENABLE_VULKAN_RUNNER=OFF -DMLIR_ENABLE_BINDINGS_PYTHON=OFF -DMLIR_INSTALL_AGGREGATE_OBJECTS=OFF -DCMAKE_BUILD_TYPE=RelWithDebInfo -DCMAKE_TOOLCHAIN_FILE=/tmp/portage/llvm-core/mlir-20.0.0_pre20241221/work/mlir_build/gentoo_toolchain.cmake /tmp/portage/llvm-core/mlir-20.0.0_pre20241221/work/mlir

[mgorny]

I think the problem is that these linked libraries are specified as PUBLIC, which implies that anything linking to them will also implicitly link to the static libraries. I think PUBLIC is rather an uncommon choice for shared libraries, and PRIVATE should be used instead, i.e.:

diff --git a/mlir/lib/ExecutionEngine/CMakeLists.txt b/mlir/lib/ExecutionEngine/CMakeLists.txt
index cf44a02cf5cb..bebe44b8a0d2 100644
--- a/mlir/lib/ExecutionEngine/CMakeLists.txt
+++ b/mlir/lib/ExecutionEngine/CMakeLists.txt
@@ -169,7 +169,7 @@ if(LLVM_ENABLE_PIC)
 
     EXCLUDE_FROM_LIBMLIR
 
-    LINK_LIBS PUBLIC
+    LINK_LIBS PRIVATE
     mlir_float16_utils
     MLIRSparseTensorEnums
     MLIRSparseTensorRuntime

And this indeed resolves the problem for me. This is probably the case for some of the other shared libraries there as well.