P4HIR: an experimental MLIR dialect that models P4C IR

[qobilidop]

I've been working on this project recently, and will do a lightning talk about it in the upcoming 2024 P4 Workshop. This issue is created for presenting the proposal, tracking the progress, and inviting further discussions with the community.

Proposal

P4HIR (P4 High-level IR) is an experimental MLIR dialect that aims to model the existing P4C IR by supporting bi-directional translations to and from the P4C IR. In more detail:

1. Support P4C IR → P4HIR translation, so we can further develop MLIR-based P4C backend in the future, while reusing the existing P4C frontend.
2. Support P4HIR → P4C IR translation, so we can further develop MLIR-based P4C frontend in the future, while reusing the existing P4C backend.
3. A P4C IR → P4HIR → P4C IR round-trip might not result in exactly the same code, but as long as we get semantically equivalent code, this shall be fine.

The current proposal is pretty high-level. There are a lot to figure out in the technical details. I'd like to discuss with the community on those as we make progress pushing this idea forward.

Plan and progress

• https://github.com/p4lang/p4mlir/pull/1
• Add a P4C backend that converts P4C IR to P4HIR
• Add an MLIR transformation that converts P4HIR to P4C IR
• Extend P4HIR to cover a core subset of P4C IR

Other notes

P4C - MLIR concepts mapping

To give a brief introduction to MLIR for P4C developers, we can map some P4C concepts to corresponding MLIR concepts.

P4C concept	MLIR concept
IR	dialect
IR node	operation (in regions with SSACFG semantics or graph-like semantics)
.def file	.td file (ODS, TableGen)
pass	pass

[fruffy]

I have also been thinking of starting something like this. Partially because I wanted to learn more about MLIR-style compilers. Is the work planned to be open-source? I would consider contributing it as an extension first (and with that a separate repository), since LLVM is such a large dependency.

It might be a good idea to have this back end produce code that is executable by a target, e.g., BMv2. Not sure if that is also planned.

[qobilidop]

Yes, open-source is the plan! I think having a separate repo is a good idea. I've been working in a local P4C fork in a subdirectory. But it shall be straightforward to move to a separate repo.

[qobilidop]

It might be a good idea to have this back end produce code that is executable by a target, e.g., BMv2. Not sure if that is also planned.

Agree with this! My thought is that the BMv2 JSON format could be defined as another MLIR dialect, and the BMv2 backend could be ported to be transformations from P4HIR dialect to the BMv2 dialect, with a relatively trivial final codegen from BMv2 dialect to BMv2 JSON.

This might be a great idea for a standalone follow-up work, once we have a relatively mature P4HIR that covers enough P4C IR to be useful.

[qobilidop]

After talking to @asl today, I learned that there is an existing work that basically does this already: https://github.com/barinsim/p4c-mlir

Here's my summary of what this project does:

• Defined a P4 dialect. See https://github.com/barinsim/p4c-mlir/blob/main/extensions/mlir/include/P4Ops.td for an overview of the operations defined. It seems to have captured most P4 primitives.
• Implemented the translation from P4C IR to the P4 dialect in https://github.com/barinsim/p4c-mlir/tree/main/extensions/mlir/p4c-mlir-translate. Also see https://github.com/barinsim/p4c-mlir/tree/main/extensions/mlir/test/translate for examples of this translation.
• Implemented the lowering from P4 dialect to LLVM dialect in https://github.com/barinsim/p4c-mlir/tree/main/extensions/mlir/p4dialect-llvm-translate. This part seems to be WIP.

I feel this could be a pretty good starting point for further development. But as @asl suggested, maybe we should hold one step back and present a detailed design for review before diving into the implementation.

@barinsim If you are still working on this and are interested in upstreaming these work, maybe we could have some discussions?