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

[Nicolai Hähnle]

================
@@ -977,6 +977,69 @@ def AMDGPU_LDSBarrierOp : AMDGPU_Op<"lds_barrier"> {
   let hasCanonicalizer = 1;
 }
 
+def AMDGPU_LDSBarrierSignalOp : AMDGPU_Op<"lds_barrier_signal"> {
+  let summary = "Signal phase of split LDS barrier.";
+  let description = [{
+    `amdgpu.lds_barrier_signal` is the signal phase of a split barrier that
+    initiates a wait for all operations that affect the Local Data Store (LDS)
+    issued from the workgroup. This operation allows computation to be overlapped
+    with LDS memory operations by separating the barrier signal from the wait.
----------------
nhaehnle wrote:

I don't feel entirely comfortable about the wording here. This barrier, which will be lowered to a fence+barrier, doesn't just *initiate* the wait for LDS operations, it will actually wait for them. If you expect that
```
   store (to LDS)
   lds_barrier_signal
   math
```
will allow the math to run under the store to LDS, you may be in for a bad time.

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