Build error llvm 21.1.x (missing VectorTransformsEnums.h.inc)

[climbfuji]

Description

I am hitting a build error on an Alma Linux 9 system runnning in WSL 2 with LLVM 21.1.x that I did not encounter with 20.1.5 when using the exact same build options.

Clone, cmake (OK)

mkdir -p /home/dom/prod/llvm-21.1.1/src
cd /home/dom/prod/llvm-21.1.1/src

git clone -b llvmorg-21.1.1 --depth=1 https://github.com/llvm/llvm-project

cd llvm-project

/usr/bin/cmake \
--fresh \
-S llvm \
-B build \
-G "Unix Makefiles" \
-DCMAKE_INSTALL_PREFIX:STRING=/home/dom/prod/llvm-21.1.1 \
-DCMAKE_INSTALL_RPATH_USE_LINK_PATH:BOOL=ON \
-DCMAKE_INSTALL_RPATH:STRING="/home/dom/prod/llvm-21.1.1/lib;/home/dom/prod/llvm-21.1.1/lib64" \
-DCMAKE_BUILD_TYPE:STRING=Release \
-DCMAKE_INTERPROCEDURAL_OPTIMIZATION:BOOL=OFF \
-DCMAKE_POLICY_DEFAULT_CMP0090:STRING=NEW \
-DCMAKE_FIND_USE_PACKAGE_REGISTRY:BOOL=OFF \
-DCMAKE_EXPORT_COMPILE_COMMANDS:BOOL=ON \
-DLLVM_REQUIRES_RTTI:BOOL=ON \
-DLLVM_ENABLE_RTTI:BOOL=ON \
-DLLVM_ENABLE_LIBXML2:BOOL=OFF \
-DCLANG_DEFAULT_OPENMP_RUNTIME:STRING=libomp \
-DLIBOMP_USE_HWLOC:BOOL=OFF \
-DLLVM_ENABLE_ZSTD:BOOL=OFF \
-DCUDA_TOOLKIT_ROOT_DIR:STRING=IGNORE \
-DCUDA_SDK_ROOT_DIR:STRING=IGNORE \
-DCUDA_NVCC_EXECUTABLE:STRING=IGNORE \
-DLIBOMPTARGET_DEP_CUDA_DRIVER_LIBRARIES:STRING=IGNORE \
-DLIBOMPTARGET_ENABLE_DEBUG:BOOL=OFF \
-DLIBOMPTARGET_BUILD_AMDGPU_PLUGIN:BOOL=OFF \
-DLLDB_ENABLE_LIBEDIT:BOOL=OFF \
-DLLDB_ENABLE_LUA:BOOL=OFF \
-DLLDB_ENABLE_PYTHON:BOOL=OFF \
-DOPENMP_ENABLE_LIBOMPTARGET:BOOL=OFF \
-DLLVM_ENABLE_Z3_SOLVER:BOOL=OFF \
-DLINK_POLLY_INTO_TOOLS:BOOL=ON \
-DBUILD_SHARED_LIBS:BOOL=OFF \
-DLLVM_BUILD_LLVM_DYLIB:BOOL=ON \
-DLLVM_LINK_LLVM_DYLIB:BOOL=OFF \
-DLLVM_USE_SPLIT_DWARF:BOOL=ON \
-DLIBCXX_ENABLE_STATIC_ABI_LIBRARY:BOOL=ON \
-DCMAKE_FIND_PACKAGE_PREFER_CONFIG:BOOL=ON \
-DCMAKE_FIND_USE_PACKAGE_ROOT_PATH:BOOL=OFF \
-DCMAKE_FIND_USE_SYSTEM_PACKAGE_REGISTRY:BOOL=OFF \
-DCMAKE_FIND_USE_PACKAGE_REGISTRY:BOOL=OFF \
-DCMAKE_FIND_USE_SYSTEM_PATH:BOOL=OFF \
-DLLVM_TARGETS_TO_BUILD:STRING=X86 \
-DLLVM_ENABLE_PROJECTS:STRING="lldb;clang;clang-tools-extra;flang;lld;polly" \
-DLLVM_ENABLE_RUNTIMES:STRING="openmp;compiler-rt;libcxx;libcxxabi;libunwind" \
-DRUNTIMES_CMAKE_ARGS:STRING="-DCMAKE_INSTALL_RPATH_USE_LINK_PATH:BOOL=ON;-DCMAKE_C_FLAGS:STRING=--gcc-install-dir=/usr/lib/gcc/x86_64-redhat-linux/11;-DCMAKE_CXX_FLAGS:STRING=--gcc-install-dir=/usr/lib/gcc/x86_64-redhat-linux/11" \
2>&1 | tee log.cmake

Make

cd build
make VERBOSE=1 -j1 2>&1 | tee log.make.001

With llvm@20.1.5, this step completes successfully. With llvm@21.1.x (tried 21.1.1 and 21.1.2):

[ 32%] Building CXX object tools/mlir/lib/Conversion/ArithToAMDGPU/CMakeFiles/obj.MLIRArithToAMDGPU.dir/ArithToAMDGPU.cpp.o
cd /home/dom/prod/llvm-21.1.1/src/llvm-project/build/tools/mlir/lib/Conversion/ArithToAMDGPU && /usr/bin/c++ -DEXPERIMENTAL_KEY_INSTRUCTIONS -D_GNU_SOURCE -D__STDC_CONSTANT_MACROS -D__STDC_FORMAT_MACROS -D__STDC_LIMIT_MACROS -I/home/dom/prod/llvm-21.1.1/src/llvm-project/build/tools/mlir/lib/Conversion/ArithToAMDGPU -I/home/dom/prod/llvm-21.1.1/src/llvm-project/mlir/lib/Conversion/ArithToAMDGPU -I/home/dom/prod/llvm-21.1.1/src/llvm-project/build/tools/mlir/include -I/home/dom/prod/llvm-21.1.1/src/llvm-project/mlir/include -I/home/dom/prod/llvm-21.1.1/src/llvm-project/build/include -I/home/dom/prod/llvm-21.1.1/src/llvm-project/llvm/include -fPIC -fno-semantic-interposition -fvisibility-inlines-hidden -Werror=date-time -Wall -Wextra -Wno-unused-parameter -Wwrite-strings -Wcast-qual -Wno-missing-field-initializers -pedantic -Wno-long-long -Wimplicit-fallthrough -Wno-uninitialized -Wno-nonnull -Wno-class-memaccess -Wno-redundant-move -Wno-pessimizing-move -Wno-noexcept-type -Wdelete-non-virtual-dtor -Wsuggest-override -Wno-comment -Wno-misleading-indentation -Wctad-maybe-unsupported -ffunction-sections -fdata-sections -Wundef -Wno-unused-but-set-parameter -Wno-deprecated-copy -O3 -DNDEBUG -std=c++17  -fno-exceptions -funwind-tables -MD -MT tools/mlir/lib/Conversion/ArithToAMDGPU/CMakeFiles/obj.MLIRArithToAMDGPU.dir/ArithToAMDGPU.cpp.o -MF CMakeFiles/obj.MLIRArithToAMDGPU.dir/ArithToAMDGPU.cpp.o.d -o CMakeFiles/obj.MLIRArithToAMDGPU.dir/ArithToAMDGPU.cpp.o -c /home/dom/prod/llvm-21.1.1/src/llvm-project/mlir/lib/Conversion/ArithToAMDGPU/ArithToAMDGPU.cpp
In file included from /home/dom/prod/llvm-21.1.1/src/llvm-project/mlir/lib/Conversion/ArithToAMDGPU/ArithToAMDGPU.cpp:19:
/home/dom/prod/llvm-21.1.1/src/llvm-project/mlir/include/mlir/Dialect/Vector/Transforms/VectorRewritePatterns.h:19:10: fatal error: mlir/Dialect/Vector/Transforms/VectorTransformsEnums.h.inc: No such file or directory
   19 | #include "mlir/Dialect/Vector/Transforms/VectorTransformsEnums.h.inc"
      |          ^~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
compilation terminated.
make[2]: *** [tools/mlir/lib/Conversion/ArithToAMDGPU/CMakeFiles/obj.MLIRArithToAMDGPU.dir/build.make:76: tools/mlir/lib/Conversion/ArithToAMDGPU/CMakeFiles/obj.MLIRArithToAMDGPU.dir/ArithToAMDGPU.cpp.o] Error 1
make[2]: Leaving directory '/home/dom/prod/llvm-21.1.1/src/llvm-project/build'
make[1]: *** [CMakeFiles/Makefile2:60636: tools/mlir/lib/Conversion/ArithToAMDGPU/CMakeFiles/obj.MLIRArithToAMDGPU.dir/all] Error 2
make[1]: Leaving directory '/home/dom/prod/llvm-21.1.1/src/llvm-project/build'
make: *** [Makefile:156: all] Error 2

Solution (Workaround) for 21.1.x

cd tools/mlir/include/mlir/Dialect/Vector/Transforms
make VERBOSE=1 MLIRVectorTransformsIncGen 2>&1 | tee log.make.002
cd -
make VERBOSE=1 -j1 2>&1 | tee log.make.003

With this workaround, the build succeeds, and I can run make install afterwards.

[DavidSpickett]

Just looking at the errors and cmake config on the 21.x branch (I know nothing about MLIR itself):

• ArithToAMDGPU.cpp includes VectorRewritePatterns.h
• VectorRewritePatterns.h tries to include VectorTransformsEnums.h.inc

ArithToAMDGPU.cpp is included in MLIRArithToAMDGPU in mlir/lib/Conversion/ArithToAMDGPU/CMakeLists.txt. That library links MLIRVectorDialect, which is defined in mlir/lib/Dialect/Vector/IR/CMakeLists.txt.

In there it has:

  DEPENDS
  MLIRMaskableOpInterfaceIncGen
  MLIRMaskingOpInterfaceIncGen
  MLIRVectorOpsIncGen
  MLIRVectorAttributesIncGen

No MLIRVectorTransformsIncGen.

So if it's reproducible for you, could you try adding MLIRVectorTransformsIncGen in mlir/lib/Dialect/Vector/IR/CMakeLists.txt?

[climbfuji]

Thanks for the quick response. I will try that now.

[DavidSpickett]

Alternatively, some of the other DEPENDS in vector have:

add_public_tablegen_target(MLIRVectorOpsIncGen)
add_dependencies(mlir-generic-headers MLIRVectorOpsIncGen)

Not all of them though, so it's not necessarily a problem.

[climbfuji]

That didn't work (failed in cmake):

CMake Error at cmake/modules/AddLLVM.cmake:800 (target_link_libraries):
  Target "MLIRVectorTransformsIncGen" of type UTILITY may not be linked into
  another target.  One may link only to INTERFACE, OBJECT, STATIC or SHARED
  libraries, or to executables with the ENABLE_EXPORTS property set.
Call Stack (most recent call first):
  /home/dom/prod/llvm-21.1.1/src/llvm-project/mlir/cmake/modules/AddMLIR.cmake:387 (llvm_add_library)
  /home/dom/prod/llvm-21.1.1/src/llvm-project/mlir/cmake/modules/AddMLIR.cmake:674 (add_mlir_library)
  /home/dom/prod/llvm-21.1.1/src/llvm-project/mlir/lib/Dialect/Vector/IR/CMakeLists.txt:1 (add_mlir_dialect_library)

[DavidSpickett]

Thanks for trying. @banach-space might know the right thing to do.

Side note: the reason this is building mlir at all is because flang enables it. Took me a bit to realise that. Should not be any different to enabling it in LLVM_ENABLE_PROJECTS though.

[banach-space]

Sorry for the delay responding, I was away last week.

Unfortunately I am unable to reproduce this. I am using LLVM ToT (top of the tree) - could you try that instead of releases? I doubt that's going to change anything, but worth trying.

I suspect that these settings is what we are not testing in our CI (and hence not seeing any failures):

-DBUILD_SHARED_LIBS:BOOL=OFF \
-DLLVM_BUILD_LLVM_DYLIB:BOOL=ON \
-DLLVM_LINK_LLVM_DYLIB:BOOL=OFF \

Still, even with those settings the build works fine for me.

David's suggestion makes a lot of sense to me and doesn't trigger any CMake errors:

index 0248896e096a..60771004d127 100644
--- a/mlir/lib/Dialect/Vector/IR/CMakeLists.txt
+++ b/mlir/lib/Dialect/Vector/IR/CMakeLists.txt
@@ -11,6 +11,7 @@ add_mlir_dialect_library(MLIRVectorDialect
   MLIRMaskingOpInterfaceIncGen
   MLIRVectorOpsIncGen
   MLIRVectorAttributesIncGen
+  MLIRVectorTransformsIncGen

   LINK_LIBS PUBLIC
   MLIRAffineDialect

Could you double-check that? Thanks!

[climbfuji]

Thanks for your reply. I tried develop and I was not able to reproduce the issue there. Probably not worth trying to fix something on an older release tag.

[climbfuji]

@banach-space I suggest closing this, since I cannot reproduce it on develop.