[Handshake] Canonicalize pack and unpack

[Dinistro]

Currently, there are no canonicalization patterns for pack and unpack.

[mortbopet]

All of this would of course be mitigated if there was just a standardized way, upstream, of handling tuple packing, unpacking, extracting, exploded (with accompanying canonicalization patterns).

One option could be to move tuple-type and pack/unpack lowering out of HandshakeToHW and instead as have a pre-pass that converts tuples to hw.structs, and pack/unpack ops to their corresponding hw counterparts. With this, the hw.struct_* canonicalizers should apply.

[Dinistro]

There is a rationale that explicitly states that they will never adopt tuple operations due to non generic semantics: https://mlir.llvm.org/docs/Rationale/Rationale/#tuple-types

One option could be to move tuple-type and pack/unpack lowering out of HandshakeToHW and instead as have a pre-pass that converts tuples to hw.structs, and pack/unpack ops to their corresponding hw counterparts. With this, the hw.struct_* canonicalizers should apply.

If that can handle the control signal accordingly, we could do so. Not sure how this would work if there is a fork in between though.

[mortbopet]

There is a rationale that explicitly states that they will never adopt tuple operations due to non generic semantics: https://mlir.llvm.org/docs/Rationale/Rationale/#tuple-types

I'm aware, and it makes sense. Personally, i'd be more in favor of a (yet another) upstream dialect containing a struct type, with optional field names, and struct create/extract operations.

If that can handle the control signal accordingly, we could do so. Not sure how this would work if there is a fork in between though.

Should be able to! Given that any value in a handshake.func carries handshake semantics, inserting hw operations is in theory therefore also supported, assuming that whatever lowering pass that follows supports the operation. hw struct operations aren't yet supported in HandshakeToHW but this is straight forward (mostly just adding a few template instantiations to the patterns) - and i will need this shortly, anyways :).

Not sure how this would work if there is a fork in between though.

In general, I'd expect canonicalization patterns to work on de-materialized IR.