Sources: background card; Lattner et al., MLIR: A Compiler Infrastructure for the End of Moore’s Law, arXiv:2002.11054; Lattner & Adve, LLVM: a compilation framework for lifelong program analysis & transformation, CGO 2004
LLVM IR is a single, fixed instruction set designed to sit close to machine code — it has one level of abstraction. MLIR is a framework for defining many IRs (dialects) at many levels of abstraction, plus the shared infrastructure (verifier, pass manager, rewrite engine, printer/parser) to manipulate all of them uniformly.
In practice they are complementary rather than competing: a compiler built on MLIR usually still ends by lowering into MLIR’s llvm dialect and handing off to real LLVM IR and the existing LLVM backend, to reuse LLVM’s mature code generation and target support. MLIR’s contribution is the earlier stages — representing a program at higher levels (source-language-ish, or domain-specific) before that final drop to LLVM.
For a project storing code and its IRs in a graph database, this suggests a natural layering: source-language ASTs at the top, one or more MLIR-dialect layers in the middle for language-agnostic and domain-specific transformations, and LLVM IR as the final common layer before machine code.