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

[Alan Li]

================
@@ -204,6 +204,15 @@ def ROCDL_ReadlaneOp : ROCDL_IntrOp<"readlane", [], [0], [AllTypesMatch<["res",
    }];
 }
 
+// The LLVM intrinsic function name is rather mouthful,
+// so here we opt to use a shorter rocdl name.
+def ROCDL_WaveIdOp :  LLVM_IntrOpBase<ROCDL_Dialect, "wave_id",
+                        "amdgcn_s_get_waveid_in_workgroup", [], [], [], 1>,
----------------
lialan wrote:

@kuhar so it is only conditionally speculatable in that it cannot be hoisted out of parallel loops but speculatable within parallel loops. Something like this might help:

```C++
    ::mlir::Speculation::Speculatability
    WaveIdOp::getSpeculatability() {
      Operation* enclosingParallelLoopOp = 
          getParentOfType<::mlir::scf::ForallOp>();
      return enclosingParallelLoopOp ?
          ::mlir::Speculation::Speculatable : ::mlir::Speculation::NotSpeculatable;
    }
```

But question: should this go into ROCDL op definitions? I feel it will be a bit out of place, as the restriction is downstream-specific.

Any thoughts on this?

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