Fuzzing tools for MLIR

The project is highly experimental, and is not currently meant to be used. It aims at deriving fuzzers, reducers, or mutators for MLIR dialects, using IRDL and to avoid generating programs that do not verify.

Building

• Setup LLVM and IRDL projects
  • git submodule update --init --recursive
  • ./utils/build-llvm.sh
• Setup CMake
  • ./utils/setup.sh
• Build the different tools (only mlir-enumerate for now)
  • ./utils/build.sh

Running

First create the generation dir: mkdir generated.

./build/bin/mlir-enumerate dialects/arith.mlir -o generated will run the enumerator with the arith dialect, and generate the programs in the generated folder.

Testing MLIR code with xDSL-smt

The xdsl-smt project has a translation validation tool for arith that we can use to test random programs. The run_xdsl_tv.py script will run the translation validation tool on randomly generated programs.

# Install xdsl-smt in a virtual environment
cd PATH_TO_XDSL_SMT
python -m venv venv
source venv/bin/activate
pip install -e .

# Generate random programs and run the translation validation tool
cd PATH_TO_MLIR_FUZZ
python run_xdsl_tv.py

To test different passes, modify the is-miscompiling.sh script.