Optimizations in clangir?

[llvm]

Optimizations in clangir?

[lanza]

Hey Keyi! What exactly is your project? We certainly don't have as much of the Clang AST supported as Polygeist as of today. We hope to improve the story over the next year, but it'll be a relatively long process.

But the default cir output is unoptimized compared to Polygeist's. For instance, cgeist runs mem2reg and code motion optimizations, which are very helpful for downstream tools to analyze the IR.

Yes, we currently don't run any optimization passes. That's a bit of a "phase 2" topic. For "phase 1" we're prioritizing solving the problem of handling more of Clang's AST and codegen.

May I ask what your project and interest is? We're definitely interested more about it. We have a monthly meeting and the next one is this coming Monday if you'd like to join. @bcardosolopes will be sharing a reminder in https://discourse.llvm.org/tag/clang shortly.

[bcardosolopes]

For instance, cgeist runs mem2reg and code motion optimizations, which are very helpful for downstream tools to analyze the IR.

We currently have WIP support for CIR -> "in-tree MLIR dialects" lowering, so theoretically you could use that output as an input for poligesit already existing mem2reg and other code motion optimizations.

The problem here, as Nathan stated, is that we don't yet cover all AST nodes, which is what we are heavily working atm. Also, CIR -> MLIR in-tree dialects progress is a bit stalled since no one is actively working on it yet, but patches are welcome - and it's the probably best way to connect the two projects.

Is there any plan to add common optimizations to clangir before LLVM lowering pass?

We plan to work on that at some point in the future but it's not the current short-term priority.

[bcardosolopes]

@bcardosolopes will be sharing a reminder in https://discourse.llvm.org/tag/clang shortly

https://discourse.llvm.org/t/mlir-c-c-frontend-working-group-monday-march-27th/69511

[Kuree]

Thanks for the quick response!

May I ask what your project and interest is?

As much as I'd like to talk about it, unfortunately we are not ready to share the project details yet. Sorry about that.

Also, CIR -> MLIR in-tree dialects progress is a bit stalled since no one is actively working on it yet, but patches are welcome - and it's the probably best way to connect the two projects.

That sounds promising, but one potential issue is that Polygesit and clangir use slightly different version of MLIR. It's likely that we need to have two different LLVMs in order to connect these two projects. I will discuss it with my colleagues and see if there is any interest choosing this route.

Thank you again for the information. I will try to attend the Monday meeting.

[bcardosolopes]

That sounds promising, but one potential issue is that Polygesit and clangir use slightly different version of MLIR. It's likely that we need to have two different LLVMs in order to connect these two projects.

Indeed, we try to rebase on ToT often, but at times we might get few months behind, if it helps we can tag specific points in time for others to consume, or open to hear suggestions.

I will discuss it with my colleagues and see if there is any interest choosing this route.

Hit us up if you wanna discuss any piece further!