[Mlir-commits] [mlir] [mlir][gpu] Add address space modifier to Barrier (PR #110527)

[Krzysztof Drewniak]

================
@@ -1371,11 +1372,25 @@ def GPU_BarrierOp : GPU_Op<"barrier"> {
     accessing the same memory can be avoided by synchronizing work items
     in-between these accesses.
 
+    The address space of visible memory accesses can be modified by adding a
+    list of address spaces required to be visible. By default all address spaces
+    are included.
----------------
krzysz00 wrote:

`gpu.barrier` is stronger than `amdgpu.lds_barrier`. The current semantics of `gpu.barrier` are not AMD's `s_barrier`, they're `(atomic fence) + s_barrier`.

(I think there are are even explicit calls to LLVM memory fencing in the AMD lowering path for `gpu.barrier`).

That is, `gpu.barrier` imposes a happens-before over all GPU memory by default, and this patch lets people loosen that.

https://github.com/llvm/llvm-project/pull/110527