mlir

MLIR

Purpose

Guide agents through MLIR (Multi-Level IR): ops, regions, blocks, and values; built-in dialects (arith, func, memref, affine, linalg); writing custom dialects with ODS; lowering passes with ConversionPattern; mlir-opt CLI; and ML compiler use cases (Torch-MLIR, IREE).

When to Use

• Building a domain-specific compiler IR (graphics, ML, hardware DSL)
• Lowering high-level ops to LLVM or GPU dialects
• Writing progressive lowering pipelines (linalg → loops → LLVM)
• Integrating with IREE or Torch-MLIR for ML deployment
• Creating reusable transformation passes across dialects
• Prototyping compiler optimizations at the right abstraction level

Workflow

1. MLIR structure

Module
└── func.func @main()
    └── region
        └── block ^bb0:
            └── operations (ops) producing SSA values

Key concepts:

• Operation — instruction-like node (arith.addi, memref.load)
• Region — container of blocks (functions, control flow)
• Block — CFG node with ordered ops
• Value — SSA result of an op or block argument

2. Built-in dialects

Dialect	Purpose
arith	Integer/float arithmetic
func	Function definitions and calls
memref	Buffer abstraction with shapes/strides
affine	Affine loop nests, map/set constraints
linalg	Structured linear algebra ops
scf	Structured control flow (for, if)
llvm	LLVM IR dialect for final lowering
gpu	GPU kernel launches

// example.mlir
func.func @add(%a: memref<4xf32>, %b: memref<4xf32>, %c: memref<4xf32>) {
  %c0 = arith.constant 0 : index
  %c4 = arith.constant 4 : index
  scf.for %i = %c0 to %c4 step %c1 {
    %av = memref.load %a[%i] : memref<4xf32>
    %bv = memref.load %b[%i] : memref<4xf32>
    %sum = arith.addf %av, %bv : f32
    memref.store %sum, %c[%i] : memref<4xf32>
  }
  return
}

3. mlir-opt CLI

# Parse and print
mlir-opt example.mlir

# Run canonicalization
mlir-opt example.mlir -canonicalize

# Lower affine to scf
mlir-opt affine.mlir -lower-affine

# Full pipeline toward LLVM
mlir-opt input.mlir \
  --linalg-bufferize \
  --convert-linalg-to-loops \
  --convert-scf-to-cf \
  --convert-arith-to-llvm \
  --convert-memref-to-llvm \
  --convert-func-to-llvm \
  -o llvm.mlir