numbacc: a Numba/MLIR backend for SPy

What it is¶

numba/numbacc is a next-generation AOT compiler project from the Numba team that uses SPy directly as its frontend — not a hypothetical integration, a working one. Its Makefile vendors spylang/spy as a pinned dependency, and nbcc/frontend/frontend.py reads SPy’s own typed AST objects (W_ASTFunc, SPyVM, FQN) straight out of the SPy VM, in-memory — no text serialization or re-parsing involved.

The pipeline¶

SPy (typed AST) → sealir (RVSDG representation) → egglog-based e-graph optimization → MLIR → LLVM → native code, plus a separate cutile_backend targeting NVIDIA’s real cuda-tile MLIR dialect for GPU.

A few notable pieces:

• sealir/RVSDG is a separate numba-org repo — a structured dataflow IR purpose-built for this project. It’s unrelated to the classic numba.core.ir used by Numba’s JIT.

• egglog-based equality saturation runs before MLIR lowering — a genuinely current compiler technique (e-graph rewriting), not something either SPy’s C backend or a hand-rolled Codon backend would get.

• MLIR’s linalg + transform dialects handle tensor/array tiling and fusion — mature, widely-adopted infrastructure across the ML compiler ecosystem (IREE, torch-mlir, and others), a stronger foundation than a bespoke fusion pass would be.

Not just a backend¶

numbacc also ships its own SPy-source library, mlir_tensor_lib.spy, defining a TensorType that splices literal MLIR ops (linalg.add, tensor.dim) directly into compiled output, using @blue-phase metaprogramming rather than compiler changes. This is separate from NumSPy’s existing array[DTYPE, NDIM] — a second, MLIR-native array abstraction, not (yet) unified with NumSPy’s.

Separately, examples/e2e/*.spy (loops, @struct classes, if/else) show a genuine general-purpose backend for SPy’s core language — no MLIR-specific syntax, ordinary SPy compiled straight through.

Caveats¶

• Early-stage: sparse README, few stars/watchers — a research prototype, not a supported product.

• Pinned to an old SPy commit rather than tracking main.

• A heavy toolchain: custom conda recipes for cross-compiled MLIR Python bindings, two separate MLIR version environments (LLVM 20/21), plus egglog, sealir, and cuda-tile as further dependencies.

• Whether NumSPy and mlir_tensor_lib’s TensorType are meant to converge eventually, or stay separate long-term, isn’t answered by the code.