MLIR-Tutorial

MLIR (Multi-Level Intermediate Representation) is a modular compiler infrastructure developed within the LLVM project.
It allows representing and transforming programs at multiple abstraction levels, using an extensible architecture made of dialects.

In many projects, it is necessary to define custom operations, types, and transformation rules. This is done by creating a custom MLIR dialect.

---

Why an out-of-tree dialect?

There are two main ways to develop an MLIR dialect:

• In-tree: integrated directly into the LLVM repository (like arith, math, or linalg dialects).
• Out-of-tree: developed in a separate repository, compiled against an existing LLVM/MLIR installation.

The out-of-tree approach is the most flexible for independent projects.
It allows:

• Maintaining a repository separate from LLVM.
• Evolving quickly without being tied to upstream LLVM commits.
• Testing against different LLVM versions (MLIR changes quickly).

---

Tutorial goals

This tutorial will guide you step by step through:

1. Setting up an MLIR build environment.
2. Creating a minimal out-of-tree dialect (with operations and types).
3. Writing a simple pass.
4. Lowering to standard dialects (arith, math).
5. Testing the dialect with lit and mlir-opt.

This tutorial is up to date for LLVM 22.

• Next section: CMake Environment

This site is open source. Improve this page.