[Mlir-commits] [mlir] [mlir][gpu] Allow gpu.dynamic_shared_memory return llvm.ptr (PR #96783)

[Guray Ozen]

grypp wrote:

> Just for my understanding, isn't the `gpu` dialect expected to provide a higher and independent level of abstraction on top of LLVM?

The OP is high level. I think a GPU OP returning an llvm.ptr feels unusual. 

I considered placing this OP in the LLVM dialect, but it isn't suitable since it's designed to map to LLVM intrinsics. 'gpu.dynamic_shared_memory' goes beyond intrinsics. So, I've improved the existing OP, and now it returns an 'llvm.ptr'. 

If it's too concerning, I can think about how/where else I can do this. 

https://github.com/llvm/llvm-project/pull/96783