See docs for details. More useful is the tutorial slides.

Basically, first we'll first need to install LLVM. E.g. as a submodule at ../llvm-project. Then to set it up, we'll need to build.

cd llvm-project
mkdir build && cd build
cmake -G Ninja ../llvm \
  -DLLVM_ENABLE_PROJECTS=mlir \
  -DLLVM_BUILD_EXAMPLES=ON \
  -DLLVM_ENABLE_ASSERTIONS=ON \
  -DCMAKE_BUILD_TYPE=Release \
  -DLLVM_ENABLE_RTTI=ON \
  -DLLVM_TARGETS_TO_BUILD="host"

cmake --build . --target check-mlir

Sometimes you might also want to change the targets in -DLLVM_TARGETS_TO_BUILD to include e.g. "NVPTX;AMDGPU".

Wherever you end up building LLVM, you'll need to change the appropriate variables in CMakeLists.txt, build.sh, and run.sh.

Setting up the first dialect is a huge pain. The documentation kinda sucks too, so this will be quite the learning experience.

• Build LLVM/MLIR as described above, e.g. in $LLVM_BUILD_DIR.

• Make sure to link them correctly to this project's cmake using both of $LLVM_DIR = $LLVM_BUILD_DIR/lib/cmake/llvm and $MLIR_DIR = $LLVM_BUILD_DIR/lib/cmake/mlir.

• Include the relevant LLVM/MLIR packages, which will be determined by their configuration files from $LLVM_DIR, $MLIR_DIR.

• Then add your own files

That is all of the necessary linking steps that I know of so far, but doesn't include e.g. testing.

Now we can actually work with MLIR, and hopefully it's not as totally tedious.

• include/[DialectName]/ all of the dialect headers, ODS files.

• lib/[DialectName]/ for the source files.

  • lib/[DialectName]/IR/ for the source files of IR (e.g. operations, types, etc.)

  • lib/[DialectName]/Transforms/ for the source files of the transformations

  • I don't know where the passes go according to the standard.

• test/[DialectName/ for the test files (?)

• [DialectName]-opt/ I don't know what this does yet.

• [DialectName]-translate/ for translating files to/from MLIR?

• [DialectName]-plugin/ plugins for IDEs (?)

• python/ if adding python bindings.

• Play with dynamic dialects? docs