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

[Mehdi Amini]

joker-eph wrote:

> When offload comes around, I imagine that pass will be replaced with something else entirely.

I think the issue is that offload will require to emit LLVM IR directly (to be able to reuse the offload infra, just like we do with the OpenMP flow), which means it has to be part of "translate" and not a separate pass. So at this point `gpu-to-mlir-gpu-runtime-calls ` would go back to "translate". 

`gpu-to-mlir-gpu-runtime-calls` is somehow "cleaner", that was my favorite option when we started all this infra, but @fabianmcg  convinced me otherwise because of the similarity with OpenMP where we already made this tradeoff.


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