MLIR paper

MLIR is infrastructure where dialects define coexisting levels of abstraction, not one IR that descends through fixed stages.

What the paper proposes

MLIR is argued as infrastructure, not as a compiler. An operation carries a name from a dialect namespace, operands and results, attributes, and regions; a dialect is the unit that owns a set of operations, types and invariants. Because operations from different dialects coexist in one module, a pass can rewrite a subset and leave the rest legal, which is what makes lowering incremental rather than a sequence of whole-program rewrites.

What it does not claim

The paper does not ship a dialect set, a frontend, or a performance result for a model. It argues a design point and then surveys the users of the day. Read the structural sections; treat the use-case survey as evidence that the design was usable in 2020, not as a description of what exists now.

What the project became

The core survived: an operation-based IR, per-operation verification, a pass manager with nested pipelines, dialect conversion. The surface did not. Dialects the paper treats as examples now have versioned specifications and compatibility promises, and whole compilation stacks have grown on top of the infrastructure. Tensor and buffer semantics were split into explicit conversions, and “progressive lowering” is now a set of named mechanisms that the Toy tutorial introduces better than the paper does.

Read it like this

Read the design principles and the IR structure, skip the evaluation table, then work through the first two chapters of the Toy tutorial and map each construct back to a paragraph in the paper. That mapping is the point. Afterwards you should be able to say what a dialect owns, why a pass restricted to one level can be correct, and where the paper’s vocabulary stops matching the documentation.

Things that break

Reading “multi-level IR” as one IR that holds every level at once and lowers in a single pass. The symptom is a pass that rewrites linalg straight to the LLVM dialect and then cannot verify anything in between, or an expectation that MLIR supplies a frontend. MLIR supplies the container; the levels are yours to define and to keep legal.

Sources