[Mlir-commits] [mlir] [MLIR][ROCDL] Add conversion for gpu.subgroup_id to ROCDL (PR #136405)

[Jakub Kuderski]

================
@@ -204,6 +204,14 @@ def ROCDL_ReadlaneOp : ROCDL_IntrOp<"readlane", [], [0], [AllTypesMatch<["res",
    }];
 }
 
+// the intrinsic function name is too long so we use a shorter name for rocdl.
+def ROCDL_WaveIdOp :  LLVM_IntrOpBase<ROCDL_Dialect, "wave_id",
+                        "amdgcn_s_get_waveid_in_workgroup", [], [], [Pure], 1>,
----------------
kuhar wrote:

Is this really `Pure`? In my mental model, this would have a `read` memory effect on the hw register with the subgroup ID. `Pure` means speculatable + no memory effects

https://github.com/llvm/llvm-project/pull/136405