// round 1
		{{range $i, $s := dup 4 .Shift1 -}}
			{{printf "arg0 = arg1 + bits.RotateLeft32((((arg2^arg3)&arg1)^arg3)+arg0+x%x+%#08x, %d)" (idx 1 $i) (index $.Table1 $i) $s | relabel}}
			{{rotate -}}
		{{end}}

		// round 2
		{{range $i, $s := dup 4 .Shift2 -}}
			{{printf "arg0 = arg1 + bits.RotateLeft32((((arg1^arg2)&arg3)^arg2)+arg0+x%x+%#08x, %d)" (idx 2 $i) (index $.Table2 $i) $s | relabel}}
			{{rotate -}}
		{{end}}

  %0:1 = "tfl.while"(%arg0, %arg1) ({
  ^bb0(%arg2: tensor<*xi32>, %arg3: tensor<*xf32>):
    %1 = func.call @WhileOp_cond(%arg2, %arg3) : (tensor<*xi32>, tensor<*xf32>) -> tensor<i1>
    "tfl.yield"(%1) : (tensor<i1>) -> ()
  },  {
  ^bb0(%arg2: tensor<*xi32>, %arg3: tensor<*xf32>):
    %1:2 = func.call @WhileOp_body(%arg2, %arg3) : (tensor<*xi32>, tensor<*xf32>) -> (tensor<*xi32>, tensor<*xf32>)

  // CHECK-SAME: %[[ARG3:.*]]: tensor<*x!tf_type.resource<tensor<3x3x1x32xf32>>> {tf.device = "/device:TPU:1"})
  func.func @main(%arg0: !tf_res_f32 {tf.device = "/device:TPU:0"},
             %arg1: !tf_res_f32 {tf.device = "/device:TPU:1"},
             %arg2: !tf_res_md_f32 {tf.device = "/device:TPU:0"},
             %arg3: !tf_res_md_f32 {tf.device = "/device:TPU:1"}) {

# CHECK-NEXT: %[[CUSTOM:.*]] = "tfl.custom_tf"(%arg0, %arg1) ({
# CHECK-NEXT: ^bb0(%arg2: tensor<4xi32>, %arg3: tensor<4xi32>):
# CHECK-NEXT:   %[[OUTPUTS:.*]] = "tf.BannaPotatoSaladWithColeslaw"(%arg2, %arg3) {T = i32, device = ""} : (tensor<4xi32>, tensor<4xi32>) -> tensor<*xi32>
# CHECK-NEXT:   "tfl.yield"(%[[OUTPUTS]]) : (tensor<*xi32>) -> ()

// CHECK-NEXT:   }
// CHECK-EMPTY:

^bb0(%arg0: tensor<4 x f32>, %arg1: tensor<4 x i8>, %arg2: tensor<4 x f32>, %arg3: tensor<4 x f32>):
  %0 = "tfl.pseudo_const" () {value = dense<0.0> : tensor<4xf32>} : () -> tensor<4xf32> loc("Const")
  %1 = "tfl.svdf"(%arg0, %arg1, %arg2, %arg3, %0) {fused_activation_function = "RELU", rank = 2 : i32} : (tensor<4xf32>, tensor<4xi8>, tensor<4xf32>, tensor<4xf32>, tensor<4xf32>) -> tensor<4xf32>

  %0 = "vhlo.dynamic_update_slice_v1"(%arg0, %arg1, %arg2, %arg3) : (tensor<4x4xi64>, tensor<2x3xi64>, tensor<i64>, tensor<i64>) -> tensor<4x4xi64>
  return %0 : tensor<4x4xi64>
}

//CHECK:func.func private @dynamic_update_slice(%arg0: tensor<4x4xi64>, %arg1: tensor<2x3xi64>, %arg2: tensor<i64>, %arg3: tensor<i64>) -> tensor<4x4xi64> {

}

// CHECK-LABEL: fused_batch_norm_v3_bf16
func.func @fused_batch_norm_v3_bf16(%arg0: tensor<8x8x8x8xbf16>, %arg1: tensor<8xf32>, %arg2: tensor<8xf32>, %arg3: tensor<8xf32>, %arg4: tensor<8xf32>) -> tensor<8x8x8x8xbf16> {

func.func @op_with_region(%arg0: tensor<1x16x16x320xf32>, %arg1: tensor<f32>) -> tensor<1x320xf32> {
  // CHECK:      %0 = "vhlo.reduce_v1"(%arg0, %arg1) <{{.*}}> ({
  // CHECK-NEXT:  ^bb0(%arg2: tensor<f32>, %arg3: tensor<f32>):
  // CHECK-NEXT:    %1 = "vhlo.add_v1"(%arg2, %arg3) : (tensor<f32>, tensor<f32>) -> tensor<f32>
  // CHECK-NEXT:    "vhlo.return_v1"(%1) : (tensor<f32>) -> ()
  // CHECK-NEXT: }) : (tensor<1x16x16x320xf32>, tensor<f32>) -> tensor<1x320xf32>

  %arg2: tensor<!tf_type.resource<tensor<f32>>>,
  %arg3: tensor<f32>) -> () {
  tf_executor.graph {
    // expected-warning@+2 {{unexpected control dependency path: path 0, node 0 (source)}}
    // expected-warning@+1 {{unexpected control dependency path: path 1, node 0 (source)}}
    %island1 = tf_executor.island {
      "tf.AssignVariableOp"(%arg0, %arg3) : (tensor<!tf_type.resource<tensor<f32>>>, tensor<f32>) -> ()

  // CHECK-NEXT: %1:10 = tf_mlrt.tf_async_while @"sort_map/while_body/TfMlrtAsyncWhileBody"(%0, %cst, %arg3, %arg3, %arg0, %arg1, %arg2, %arg1, %arg2, %arg4) {invariant_size = 6 : i32} : (tensor<i1>, tensor<i32>, tensor<3x3xf32>, tensor<3x3xf32>, tensor<i32>, tensor<?x!tf_type.resource>, tensor<*xf32>, tensor<?x!tf_type.resource>, tensor<*xf32>, tensor<i32>)...