[Mlir-commits] [mlir] [MLIR][NVVM] Adds an explicit aligned boolean attribute to nvvm.barrier (PR #192203)

[Guray Ozen]

================
@@ -1179,16 +1183,20 @@ def NVVM_BarrierOp : NVVM_SingleResultIntrinsicOp<"barrier",
   let results = (outs Optional<I32>:$res);
 
   let hasVerifier = 1;
+  let hasCanonicalizeMethod = 1;
 
   let assemblyFormat =
       "(`id` `=` $barrierId^)? (`number_of_threads` `=` $numberOfThreads^)? "
+      "custom<Aligned>($aligned) "
----------------
grypp wrote:

We have 3 OP for the [PTX's bar/barrier](https://docs.nvidia.com/cuda/parallel-thread-execution/index.html#parallel-synchronization-and-communication-instructions-bar)
1. nvvm.barrier0 # Will be removed
2. nvvm.barrier 
3. nvvm.barrier.arrive 


Given that the OP is already split, we can split it for reduction variant to simplify? So we can have the followings: 
1. nvvm.barrier
2. nvvm.barrier.arrive
3. nvvm.barrier.reduction

https://github.com/llvm/llvm-project/pull/192203