func.return %0 : tensor<2x2xi32>
}

// -----

func.func @unpack(%arg0: tensor<2x3xi32>) -> tensor<2xi32> {
  // CHECK: "tfl.unpack"(%arg0) <{axis = 1 : i32, num = 3 : i32}>
  %0:3 = "tfl.unpack"(%arg0) {axis = 1 : i32, num = 3 : i32} : (tensor<2x3xi32>) -> (tensor<2xi32>, tensor<2xi32>, tensor<2xi32>)
  func.return %0#0 : tensor<2xi32>
}

    %17 = "tf.Mul"(%16, %cst_0) : (tensor<2x2xf32>, tensor<f32>) -> tensor<2x2xf32>
    %18 = "tf.AddV2"(%17, %cst) : (tensor<2x2xf32>, tensor<f32>) -> tensor<2x2xf32>
    %19 = "tf.Floor"(%18) : (tensor<2x2xf32>) -> tensor<2x2xf32>
    %20 = "tf.ClipByValue"(%19, %cst_5, %cst_6) : (tensor<2x2xf32>, tensor<f32>, tensor<f32>) -> tensor<2x2xf32>

}
func.func @_func(%arg0: tensor<2x4xf32>, %arg1: tensor<4x2xf32>) -> tensor<2x2xf32> {
  %0 = "tf.MatMul"(%arg0, %arg1) {_XlaSharding = "\08\03\1A\02\02\01\22\02\00\01"} : (tensor<2x4xf32>, tensor<4x2xf32>) -> tensor<2x2xf32>
  %1 = "tf.Identity"(%0) : (tensor<2x2xf32>) -> tensor<2x2xf32>
  return %1 : tensor<2x2xf32>
}

// -----
// The following op sharding is used in the following test case:

  %sizes = "tf.Const"() {value = dense<[1, -1]> : tensor<2xi32>} : () -> (tensor<2xi32>)
  %0 = "tf.Slice"(%arg0, %arg1, %sizes) : (tensor<3x4xi32>, tensor<2xi32>, tensor<2xi32>) -> tensor<1x4xi32>
  func.return %0 : tensor<1x4xi32>
}

//===----------------------------------------------------------------------===//
// StridedSlice op legalizations.

    %8 = stablehlo.convert %7 : (tensor<1x2xi32>) -> tensor<1x2xf32>
    %9 = stablehlo.convert %2 : (tensor<2x3xi8>) -> tensor<2x3xf32>
    %10 = stablehlo.dot_general %8, %9, contracting_dims = [1] x [0] : (tensor<1x2xf32>, tensor<2x3xf32>) -> tensor<1x3xf32>
    %11 = stablehlo.convert %3 : (tensor<1x3xi32>) -> tensor<1x3xf32>

      %outputs_9, %control_10 = tf_executor.island(%control_4) wraps "tf.Identity"(%outputs_7) {device = ""} : (tensor<2x3xi32>) -> tensor<2x3xi32>
      tf_executor.fetch %outputs_9 : tensor<2x3xi32>
    }
    return %0 : tensor<2x3xi32>
  }
}

// CHECK: module attributes
// CHECK-SAME:  tfl.metadata = {{{.*}}keep_stablehlo_constant = "true"{{.*}}}

  %cst = arith.constant dense<[2, 2]> : tensor<2xi32>
  %cst_0 = arith.constant dense<4> : tensor<2x2xi32>
  %cst_1 = arith.constant dense<0> : tensor<2x2xi32>
  %cst_2 = arith.constant dense<0> : tensor<4x2xi32>
  %0 = "tf.SpaceToBatchND"(%arg0, %cst, %cst_0) : (tensor<1x128x128x3xf32>, tensor<2xi32>, tensor<2x2xi32>) -> tensor<4x68x68x3xf32>

  %group_assignment = "tf.Const"() { value = dense<[[0],[1]]> : tensor<2x1xi32> } : () -> tensor<2x1xi32>
  // CHECK: "mhlo.all_reduce"
  // CHECK{LITERAL}: replica_groups = dense<[[0], [1]]> : tensor<2x1xi64>
  // CHECK-NOT: channel_handle
  // CHECK: mhlo.add
  %0 = "tf.XlaAllReduce"(%input, %group_assignment) {reduce_op = "Add", mode = "CrossReplica"} : (tensor<f32>, tensor<2x1xi32>) -> tensor<f32>
  func.return %0 : tensor<f32>
}

// -----

  %ch0 = tfrt.new.chain

  %a1 = tfrt_fallback_async.const_dense_tensor dense<[[2, 2], [2, 2]]> : tensor<2x2xi32>
  %a2 = tfrt_fallback_async.const_dense_tensor dense<[[3, 3], [3, 3]]> : tensor<2x2xi32>
  %b = tfrt_fallback_async.const_dense_tensor dense<[[1, 1], [1, 1]]> : tensor<2x2xi32>

  // Two batch_size=2 batches get concatenated.

  func.func @dynamic_update_slice(%arg0: tensor<3x4xi32>, %arg1: tensor<2x2xi32>, %arg2: tensor<2xi32>) -> tensor<3x4xi32> {

    // CHECK: %[[SLICE0:.*]] = "mhlo.slice"(%[[ARG2]])
    // CHECK-DAG-SAME: start_indices = dense<0> : tensor<1xi64>
    // CHECK-DAG-SAME: limit_indices = dense<1> : tensor<1xi64>
    // CHECK-DAG-SAME: strides = dense<1> : tensor<1xi64>