[Mlir-commits] [mlir] [mlir][amdgpu] implement amdgpu.global_load_async_to_lds for gfx1250 (PR #189279)

[Krzysztof Drewniak]

krzysz00 wrote:

Follow-up comments:

Since you can mask off a write by storing to `ptr addrspace(3) 0xff_ff_ff_ff` / null , and since the null memref isn't really a thing, we should add an optional `I1 mask`  that'll mask off each lane - and it'll just compile to selecting between the real index and the null.

https://github.com/llvm/llvm-project/pull/189279