(pronounced "you-dsl")

Embedded Universal DSL: a good DSL for us, by us

This repository contains the source code for EUDSL, a toolkit for the construction of embedded DSLs, in arbitrary languages, for targeting MLIR.

FYI: this project is currently "alpha" quality.

Currently, there are six components:

1. eudsl-tblgen: Python bindings to LLVM's TableGen library;
2. eudsl-nbgen: A source-to-source translator that translates MLIR headers¹ into direct nanobind bindings;
3. mlir-python-bindings: Scripts/CMake for nightly builds of the upstream MLIR Python bindings;
   • Available via pip install mlir-python-bindings -f https://llvm.github.io/eudsl.
4. eudsl-python-extras: Extra utilities for the upstream MLIR Python bindings;
   • Available via pip install eudsl-python-extras -f https://llvm.github.io/eudsl if used with the upstream bindings.
   • In reality, this package is meant to work in concert with any set of "host bindings" (some distribution of the actual MLIR Python bindings). So that means the pip install should be amended to

     $ EUDSL_PYTHON_EXTRAS_HOST_PACKAGE_PREFIX=<YOUR_HOST_MLIR_PYTHON_PACKAGE_PREFIX> \
           pip install eudsl-python-extras -f https://llvm.github.io/eudsl
     

     where EUDSL_PYTHON_EXTRAS_HOST_PACKAGE_PREFIX is (as it says) the package prefix for your chosen host bindings. When in doubt about this prefix, it is everything up until ir when you import your bindings, e.g., in import torch_mlir.ir, torch_mlir is the EUDSL_PYTHON_EXTRAS_HOST_PACKAGE_PREFIX for the torch-mlir bindings.
5. mlir-python-mcp: An MCP server exposing a persistent MLIR Python REPL, so that a coding agent can build and transform IR through the bindings rather than by editing MLIR text;
   • Available via pip install mlir-python-mcp -f https://llvm.github.io/eudsl.
   • See projects/mlir-python-mcp/README.md for the .mcp.json snippet and the list of tools.
6. eudsl-llvmpy: Python bindings for LLVM IR (and Machine IR), plus small DSLs for authoring both from Python;
   • The bindings are a hand-written nanobind wrapper over the LLVM C++ API, so the Python class hierarchy mirrors llvm::; the DSL lowers a decorated Python function (with ordinary arithmetic and if/while/for) to a compiled llvm::Function.
   • A third surface, llvm.mir, reaches one level lower — LLVM's post-instruction-selection Machine IR — with the same split: hand-written bindings over the MachineFunction/MachineInstr object model plus a @machine_function DSL for building generic (GlobalISel) MIR.
   • Available via pip install eudsl-llvmpy -f https://llvm.github.io/eudsl.
   • See projects/eudsl-llvmpy/README.md and projects/eudsl-llvmpy/tests.

Wasm Playground

We currently provide two online playgrounds where you can try out the WebAssembly version of mlir-python-bindings and eudsl-python-extras directly in your browser:

• https://llvm.github.io/eudsl/jupyter – A JupyterLite instance with a Pyodide kernel. You can install the MLIR Python bindings with: await piplite.install("mlir-python-bindings").

• https://llvm.github.io/eudsl/console – A Pyodide-based REPL with mlir-python-bindings and eudsl-python-extras preloaded. Just run: from mlir.ir import * to start coding.

Getting started

Python wheels of all the tools are available at the eudsl release page. They are also pip install-able with .e.g

$ pip install mlir-python-bindings -f https://llvm.github.io/eudsl
// or
$ pip install eudsl-python-extras -f https://llvm.github.io/eudsl

If you don't want to install locally, here is a colab notebook minimal working example.

Developing/building

Strong recommendation: check the CI scripts @ .github/workflows - they do a fresh checkout and build on every commit and are written to be read by a non-CI expert.

Firstly, you need a distribution of LLVM. You can build LLVM from source using our submodule by doing (on Mac/Linux or mingw):

$ cd <EUDSL_CHECKOUT_DIR>
$ ./scripts/build_llvm.sh

Alternatively you can download a distribution for your platform from the llvm release page.

Then each of the above tools can both be built as a conventional, standalone, CMake project and as a Python wheel. The wheel build looks something like:

$ cd <EUDSL_CHECKOUT_DIR>
$ export CMAKE_PREFIX_PATH=$PWD/llvm-install 
$ pip wheel projects/eudsl-nbgen -w wheelhouse -v

If you want to build an individual tool via CMake you can do something like:

$ cd <EUDSL_CHECKOUT_DIR>
$ pip install -r requirements.txt
$ export CMAKE_PREFIX_PATH=$PWD/llvm-install 
$ cmake -B $PWD/eudsl-nbgen-build -S $PWD/projects/eudsl-nbgen -DCMAKE_BUILD_TYPE=Release -DCMAKE_INSTALL_PREFIX=$PWD/eudsl-nbgen-install
$ cmake --build "$PWD/eudsl-build" --target install

If you want to build all the tools at once using CMake you can use the root CMakeLists.txt.

If you want to build all the tools at once using CMake directly against third_party/llvm-project then (roughly) the CMake invocation is:

cmake -GNinja -B build -S <EUDSL_CHECKOUT_DIR>/third_party/llvm-project/llvm \
  -DLLVM_EXTERNAL_PROJECTS=EUDSL \
  -DLLVM_EXTERNAL_EUDSL_SOURCE_DIR=<EUDSL_CHECKOUT_DIR> \
  -DLLVM_BUILD_LLVM_DYLIB=ON \
  -DLLVM_ENABLE_PROJECTS=clang;mlir

Footnotes

Footnotes

1. Yes C++ headers... ↩