[llvm] [mlir] [mlir][gpu] Change GPU modules to globals (PR #135478)

[Fabian Mora via llvm-commits]

fabianmcg wrote:

I'm not blocking or voicing opposition to any changes. I just have a few questions:

- What would we gain from moving the code to `gpu-to-llvm`?
- Is there something we can't do today that we would be able to do with that change?

> Would you consider this a step backwards from where you are heading? 

>From my initial idea, yes. Because, the idea was to reuse existing LLVM infra and not duplicate code. Moreover, the ops have clear low level meaning, thus they have direct translation, so I don't see any benefit in handling them during conversion.

I'll also say, that once project offload reaches maturity I'm going to propose migrating the existing GPU runtime wrappers to Offload's, which would fix issues like this.

https://github.com/llvm/llvm-project/pull/135478