MLIR Backend (Experimental)

Status: 1D scaffold, off by default behind the mlir cargo feature Crate: compiler/mlir-backend Library: melior 0.25.1 (Rust MLIR bindings, LLVM/MLIR 20)

Overview

The MLIR backend is the future tensor / autodiff / GPU lowering path (Phase 2+). It consumes the same typed High-Level IR (neuro-hir) the LLVM backend consumes. As of the 1D scaffold it emits a trivial, verifier-clean MLIR module, one func.func declaration per function and impl method, proving the HIR → melior → verified-MLIR pipeline end-to-end. Real body lowering (linalg / tensor dialects) is Phase 2+.

Feature Gate

The path is opt-in behind the off-by-default mlir feature (mlir = ["dep:melior", "dep:thiserror", "dep:neuro-hir"]):

• Disabled (default): the crate compiles to an empty placeholder and pulls in no MLIR toolchain (nor neuro-hir), so cargo build/test --workspace works on a stock LLVM 20 install with no MLIR on every CI OS.
• Enabled: pulls in melior + neuro-hir and exposes the entry points below. CI provisions MLIR only on Linux, where the --all-features lint job and a dedicated cargo test -p mlir-backend --features mlir smoke step exercise the gated code; the Windows/macOS legs build the placeholder.

See Installation → Optional: MLIR Backend for the MLIR 20 + libclang 20 toolchain setup.

Entry Points (feature mlir)

1pub fn lower_program(program: &HirProgram) -> Result<String, MlirError>;2pub fn emit_smoke_module() -> Result<String, MlirError>;

• lower_program, the HIR → MLIR scaffold: registers all dialects, walks the typed HIR, and returns the textual form of a verified module of func.func declarations.
• emit_smoke_module, the HIR-independent melior wiring check: builds + verifies func.func @neuro_smoke(index, index) -> index with an arith.addi body.

Lowering Rules (scaffold)

• Free functions and impl methods become func.func declarations (empty region, private visibility, external symbols, not definitions). A method receiver lowers to a pointer parameter. Structs and constants are skipped.
• HIR scalar types map to MLIR scalars: i8 to i64, i1 for bool, i32 for char, f16 / bf16 / f32 / f64.
• Every aggregate / reference / string type maps to an opaque !llvm.ptr until real tensor and struct lowering lands (Phase 2+).
• void is the empty result list in return position; anywhere else it is a MlirError::UnsupportedType.
• Function bodies are intentionally not lowered yet; that is the Phase 2 linalg/tensor work. The module is run through the MLIR verifier before its textual form is returned.

Coexistence with inkwell

mlir-sys carries no llvm-sys dependency and links its own MLIR key, so it coexists with inkwell's llvm-20 link without a Cargo links conflict. Pointing MLIR_SYS_200_PREFIX / TABLEGEN_200_PREFIX at the same LLVM 20 build as LLVM_SYS_201_PREFIX makes both bindings share one libLLVM-20 dylib. That prefix must include MLIR (mlir-c headers + libMLIR*); Arch's stock llvm20 omits MLIR, so build LLVM 20 with -DLLVM_ENABLE_PROJECTS=mlir.

melior 0.25.1 is the newest release targeting MLIR 20 (via mlir-sys 0.5.0); melior 0.26+ moved to MLIR 21/22.

Resources

• mlir-backend CONTEXT, slice contract
• melior, Rust MLIR bindings
• MLIR, Multi-Level Intermediate Representation