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

[Christian Sigg]

chsigg wrote:

Right, for offloading this needs to be part of 'translate', but for the gpu runtime calls it does not. It makes sense to me to have everything in the pass for lowering to gpu runtime calls while the offloading path is being built. Once that's complete, we remove the gpu runtime and that pass. At the current state, there would be close to zero code duplication.

But I'm not sure how much churn this would create for downstream users. I already have accidentally broken two projects, and I don't want to avoid that happening again if it's not really making things better. So I will leave things as is for now unless someone raises their hand.

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