Dialect Conversion

[rahulmutt]

Where are we with supporting dialect conversions within pliron?

(1) I see an open PR by @greenhat here: #4 which has yet to be merged.
(2) I see a modified version of the PR by @zrr1999 here: zrr1999#1.

My questions are:

• what is left to update in the PR to finally merge it through?
• as a follow up what work is needed to setup a generalized pass / pass manager infrastructure (if we don't have it already)? (we can create separate issues for these)

[vaivaswatha]

The PR by @greenhat is really old and I haven't looked at it in a long time. The other one I'm not sure. I don't recall seeing it.

In general, regarding dialect conversions:

I'm not entirely convinced that we need a framework (or rather how that framework should look like) for conversions. The pattern applicator in MLIR is super complex and I'm not convinced of its necessity. So my thoughts were that we should have a couple of dialects first, and conversions implemented using plain interfaces (for example, how we convert pliron's LLVM dialect to actual LLVM IR) and see how that goes. With a couple of concrete examples, it's easier to decide how an infrastructure for dialect conversion should look like.

Regarding a pass manager, we don't have one at all, and we should have it. Some time ago, I wrote a pass manager for another compiler (https://github.com/FuelLabs/sway/blob/master/sway-ir/src/pass_manager.rs), but that's for a specific compiler, not an extensible infrastructure like pliron. So I didn't want to just blindly use that design and write one for pliron. So, I would really be happy to discuss how a pass manager for pliron should look like.

[rahulmutt]

I think that's a good approach. I do agree that the learning curve seems pretty steep for the MLIR pattern applicator. I was wondering if there was a macro approach that would make the entire thing cleaner on the Rust side.

Ok let's start a discussion on the pass manager then - gh issue or another discussion thread?

I'm not sure if this is part of the discussion, but do we intend to keep general purpose passes like canonicalization and cse in pliron?

[vaivaswatha]

Ok let's start a discussion on the pass manager then - gh issue or another discussion thread?

We could begin with a discussion thread, and when we have something concrete, create issues.

I'm not sure if this is part of the discussion, but do we intend to keep general purpose passes like canonicalization and cse in pliron?

Yes, as long as they are general, written using interfaces and can be generally applied to any IR (the specific ops may or may not implement the interface to participate). At least that's how I've been looking at it.