[Mlir-commits] [mlir] [MLIR][ROCDL] Added `SchedGroupBarrier` and `IglpOpt` ops (PR #112237)

[Giuseppe Rossini]

================
@@ -387,3 +388,26 @@ gpu.module @module_1 [#rocdl.target<O = 1, chip = "gfx900", abi = "500", link =
 
 gpu.module @module_2 [#rocdl.target<chip = "gfx900">, #rocdl.target<chip = "gfx90a">] {
 }
+
+// -----
+
+// ROCDL2LLVM: @rocdl_sched_barrier
+func.func @rocdl_sched_barrier() {
+  // ROCDL2LLVM-NEXT: call void @llvm.amdgcn.sched.barrier(i32 0)
+  rocdl.sched.barrier 0
+  llvm.return
+}
+
+// ROCDL2LLVM: @rocdl_sched_group_barrier
+func.func @rocdl_sched_group_barrier() {
+  // ROCDL2LLVM-NEXT: call void @llvm.amdgcn.sched.group.barrier(i32 8, i32 1, i32 0)
+  rocdl.sched.group.barrier 8, 1, 0
+  llvm.return
+}
+
+// ROCDL2LLVM: @rocdl_iglp_opt
+func.func @rocdl_iglp_opt() {
+  // ROCDL2LLVM-NEXT: call void @llvm.amdgcn.iglp.opt(i32 0)
+  rocdl.iglp.opt 0
+  llvm.return
+}
----------------
giuseros wrote:

Yep, looks perfect now!

https://github.com/llvm/llvm-project/pull/112237