// CHECK-LABEL: reduce_window_max_activation_transpose
// CHECK-SAME: (%[[ARG:.+]]: tensor<16x8xf32>) -> tensor<4x8xf32>
func.func @reduce_window_max_activation_transpose_rank2(%arg0: tensor<16x8xf32>) -> tensor<4x8xf32> {
  %0 = stablehlo.constant dense<0xFF800000> : tensor<f32>  // -inf
  %1 = stablehlo.transpose %arg0, dims = [1, 0] : (tensor<16x8xf32>) -> tensor<8x16xf32>
  %2 = "stablehlo.reduce_window"(%1, %0) ({

  %arg9: tensor<2048xf32>,
  %arg10: tensor<2048xf32>,
  %arg11: tensor<2048xf32>,
  %arg12: tensor<2048xf32>,
  %arg13: tensor<640x2048xf32>,
  %arg14: tensor<640xf32>,
  %arg15: tensor<2048xf32>,
  %arg16: tensor<2048xf32>,
  %arg17: tensor<2048xf32>,
  %arg18: tensor<2048xf32>,
  %arg19: tensor<1x640xf32>,
  %arg20: tensor<1x2048xf32>
) -> tensor<1x640xf32> {

    %add_out, %add_control = tf_executor.island wraps "tf.AddV2"(%const_out, %const_out) {_xla_compile_device_type = "TPU", _replication_info = "cluster"} : (tensor<4x4xf32>, tensor<4x4xf32>) -> tensor<4x4xf32>
    %replicated_out, %replicated_control = tf_executor.island wraps "tf.TPUReplicatedOutput"(%add_out) : (tensor<4x4xf32>) -> (tensor<4x4xf32>)

func.func @no_change_to_dag(%arg0: tensor<!tf_type.resource<tensor<10x3xf32>>>, %arg1: tensor<!tf_type.resource<tensor<10x3xf32>>>, %arg2: tensor<!tf_type.resource<tensor<10x3xf32>>>, %arg3: tensor<!tf_type.resource<tensor<10x3xf32>>>) -> (tensor<!tf_type.resource<tensor<10x3xf32>>>, tensor<!tf_type.resource<tensor<10x3xf32>>>, tensor<!tf_type.resource...

    %0 = "tf.Identity"(%cst) {device = ""} : (tensor<i1>) -> tensor<i1>
    %1 = "tf.PartitionedCall"(%arg0, %cst_1, %cst_0) <{config = "", config_proto = "", executor_type = "", f = @composite_matmul_with_bias_fn_1}> {_tfl_quant_trait = "fully_quantizable"} : (tensor<1x4xf32>, tensor<4x3xf32>, tensor<3xf32>) -> tensor<1x3xf32>

      tf_executor.yield %1, %1 : tensor<2048xf32>, tensor<2048xf32>
    }
    tf_executor.fetch %outputs#0, %outputs#1 : tensor<2048xf32>, tensor<2048xf32>
  }
  func.return %0#0, %0#1 : tensor<2048xf32>, tensor<2048xf32>
}
// CHECK-LABEL: func @single_op_island_duplicate_result
// CHECK: %[[CONST:.*]], %{{.*}} = tf_executor.island wraps "tf.Const"
// CHECK: tf_executor.fetch %[[CONST]], %[[CONST]]

      tf_executor.yield %1, %1 : tensor<2048xf32>, tensor<2048xf32>
    }
    tf_executor.fetch %outputs#0, %outputs#1 : tensor<2048xf32>, tensor<2048xf32>
  }
  func.return %0#0, %0#1 : tensor<2048xf32>, tensor<2048xf32>
}

// CHECK: func @tpu_load_embedding_ops_sink_controls
// CHECK: {{%.+}}, [[READ0:%.+]] = tf_executor.island wraps "tf.ReadVariableOp"

func.func @single_op_island_forward_block_arg(%arg0: tensor<?x?x?x?xbf16>) -> (tensor<2048xf32>, tensor<?x?x?x?xbf16>) {
  %0:2 = tf_executor.graph {
    %outputs, %control = tf_executor.island wraps "tf.Const"() {value = dense<0.000000e+00> : tensor<2048xf32>} : () -> tensor<2048xf32>
    tf_executor.fetch %outputs, %arg0 : tensor<2048xf32>, tensor<?x?x?x?xbf16>
  }
  return %0#0, %0#1 : tensor<2048xf32>, tensor<?x?x?x?xbf16>
}

  %0 = "tfl.unidirectional_sequence_lstm"(%arg0, %arg1, %arg2, %arg3, %arg4, %arg5, %arg6,...

// CHECK-LABEL: simple_strided_slice
func.func @simple_strided_slice(%input: tensor<4x8xf32>) -> tensor<3x2xf32> {
  // CHECK: %0 = mhlo.constant dense<[0, 1]> : tensor<2xi32>
  // CHECK-NEXT: %1 = mhlo.constant dense<[3, 7]> : tensor<2xi32>
  // CHECK-NEXT: %2 = mhlo.constant dense<[1, 3]> : tensor<2xi32>
  // CHECK-NEXT: %3 = "tf.StridedSlice"(%arg0, %0, %1, %2) : (tensor<4x8xf32>, tensor<2xi32>, tensor<2xi32>, tensor<2xi32>) -> tensor<3x2xf32>