Description
In a pair programming session today, we discovered numba-mlir https://github.com/numba/numba-mlir a proof of concept to replace numba's current LLVM backend with an MLIR-based backend. This is not necessarily usable as-is, but it's interesting because they have MLIR passes that we might one day be able to adapt to our purposes, some of which have been upstreamed.
For example, lift-cf-to-scf and the underlying CFGtoSCF could be something we use to avoid having to recognize and emit loops in the frontend. I.e., we could emit cf dialect ops to MLIR, and then call that pass to lift to scf, and then call a polygeist-inspired pass to lift scf to affine, or else support scf in our pipelines directly.
I think that would be a nice way to do it in the long term, though it would take some extra infrastructure to manage since the lift-cf-to-scf doesn't seem to recognize static loop bounds, which is critical for us in FHE.