[Mlir-commits] [mlir] [MLIR][NVVM] Add explicit aligned attribute to nvvm.barrier and nvvm.barrier.reduction (PR #200745)

[Guray Ozen]

================
@@ -1159,7 +1161,8 @@ def NVVM_BarrierOp : NVVM_VoidIntrinsicOp<"barrier",
 
   let arguments = (ins
       Optional<I32>:$barrierId,
-      Optional<I32>:$numberOfThreads);
+      Optional<I32>:$numberOfThreads,
+      DefaultValuedAttr<BoolAttr, "true">:$aligned);
----------------
grypp wrote:

I recommend keeping it `false` and let the downstream compiler decide alignment which has proper thread modeling. 

It's direct footgun to mark barrier aligned, then get UB 

```suggestion
      DefaultValuedAttr<BoolAttr, "false">:$aligned);
```

https://github.com/llvm/llvm-project/pull/200745