MLIR multi-level intermediate representation and dialect conversion

Document status: reviewed. Canonical source.

Summary

MLIR deliberately supports multiple abstraction levels and domain-specific dialects in one framework, with explicit legality targets and rewrite-based lowering. Its text, in-memory and compact serialized representations demonstrate that logical IR semantics need not be tied to a single storage encoding.

Evidence

Primary references: MLIR Language Reference, Dialect Conversion documentation, and MLIR rationale. Dialect conversion separates conversion targets, rewrite patterns and type conversion.

NUIF relevance

Borrow the ideas of dialect namespaces, explicit lowering passes, validation and partial legality. Do not copy SSA/control-flow machinery: NUIF is an authored document model, not compiler code IR.