[Mlir-commits] [mlir] [mlir][amdgpu] Add gfx12+ split barrier wrappers (PR #175579)

[Nicolai Hähnle]

https://github.com/nhaehnle commented:

Overall, this seems quite reasonable to me.

Bikeshed: The name `lds_barrier_signal` sounds to me as if an object called an "LDS barrier" is being signaled. This is particularly problematic since we *do* have objects that I've heard many people refer to as "LDS barrier", namely the things you interact with using `llvm.amdgcn.ds.atomic.async.barrier.arrive`. Since you're describing a default workgroup barrier operation with additional memory ordering semantics, I'd put the "modifier" that describes the memory semantics at the end. In other words, something like `barrier.signal.lds`. (But of course I can also see that there is some unfortunate precedent already.)

https://github.com/llvm/llvm-project/pull/175579