// CHECK:           "mhlo.dot_general"({{.*}}, {{.*}}) <{
// CHECK-SAME:        lhs_batching_dimensions = [0]
// CHECK-SAME:        rhs_batching_dimensions = [0]
// CHECK-SAME:        lhs_contracting_dimensions = [2]
// CHECK-SAME:        rhs_contracting_dimensions = [1]

  // CHECK: %[[CONV2D_BACKPROP:[0-9]*]] = "tf.Conv2DBackpropFilter"
  // CHECK-SAME: (%[[IN_TRANSPOSE]], %[[FILTER_PERM]], %[[OUT_BP_TRANSPOSE]])
  // CHECK-SAME: data_format = "NCHW"
  // CHECK-SAME: dilations = [1, 4, 2, 3]
  // CHECK-SAME: explicit_paddings = [1, 2, 7, 8, 3, 4, 5, 6]
  // CHECK-SAME: padding = "EXPLICIT"
  // CHECK-SAME: strides = [5, 8, 6, 7]
  // CHECK-SAME: (tensor<1x3x32x32xf32>, tensor<4xi32>, tensor<1x8x32x32xf32>)

# it is being invoked.
# CHECK: module
# CHECK-SAME: tf.versions
# CHECK-SAME: bad_consumers
# CHECK-SAME: min_consumer
# CHECK-SAME: producer

# CHECK: tf_saved_model.session_initializer
# CHECK-SAME: initializers = [@[[restore:.*]]]

# CHECK: "tf_saved_model.asset"()
# CHECK-SAME: <{filename = [[filename:.*]], sym_name = "[[sym_name:.*]]"}> : () -> ()

# UNKNOWN-SAME:  outputs = "Add"
# UNKNOWN:           %[[add:.*]], %[[add_control:.*]] = tf_executor.island wraps "tf.Add"(%[[ARG_0]], %[[ARG_1]])
# UNKNOWN:           fetch %[[add]]

# DYNAMIC-LABEL: func @main
# DYNAMIC-SAME:  (%[[ARG_0:[a-z0-9]+]]: tensor<?x1x?xi32>, %[[ARG_1:[a-z0-9]+]]: tensor<1x?x1xi32>) -> tensor<*xi32>
# DYNAMIC-SAME:  control_outputs = ""
# DYNAMIC-SAME:  inputs = "input0,input1"

# itself is verified in the common graphdef converter, so here just assert
# it is being invoked.
# CHECK: module
# CHECK-SAME: tf.versions
# CHECK-SAME: bad_consumers
# CHECK-SAME: min_consumer
# CHECK-SAME: producer

# CHECK: "tf_saved_model.global_tensor"()
# CHECK: "tf_saved_model.session_initializer"() <{initializers = [@[[INIT_FUNC:[a-zA-Z_0-9]+]]]}> : () -> ()

// CHECK:      tf_device.replicate
// CHECK-SAME: [%[[OP_A]], %[[OP_B]], %[[OP_C]]] as %{{[a-z0-9]*}}: tensor<*xi1>
// CHECK-SAME: [%[[OP_D]], %[[OP_E]], %[[OP_F]]] as %{{[a-z0-9]*}}: tensor<*xi32>
// CHECK-SAME: [%[[OP_G]], %[[OP_H]], %[[OP_I]]] as %{{[a-z0-9]*}}: tensor<*xf32>
// CHECK-SAME: %[[OP_J]] as %{{[a-z0-9]*}}: tensor<*xi8>
// CHECK-SAME: %[[OP_K]] as %{{[a-z0-9]*}}: tensor<*xi16>
// CHECK-SAME: %[[OP_L]] as %{{[a-z0-9]*}}: tensor<*xi64>

  // CHECK-SAME: %arg0,
  // CHECK-SAME: %[[DQ_1]], %[[DQ_2]], %[[DQ_3]], %[[DQ_4]],
  // CHECK-SAME: %[[DQ_5]], %[[DQ_6]], %[[DQ_7]], %[[DQ_8]],
  // CHECK-SAME: %[[NONE]], %[[NONE]], %[[NONE]],
  // CHECK-SAME: %[[DQ_9]], %[[DQ_10]], %[[DQ_9]], %[[DQ_9]],
  // CHECK-SAME: %[[NONE]], %[[NONE]],
  // CHECK-SAME: %[[DQ_11]], %[[DQ_12]],
  // CHECK-SAME: %[[NONE]], %[[NONE]], %[[NONE]], %[[NONE]]) <{

  # multiple return statements.
  #
  # CHECK:      func {{@[a-zA-Z_0-9]+}}(
  # CHECK-SAME:   %arg0: tensor<f32> {tf._user_specified_name = "x", tf_saved_model.index_path = [0]}
  # CHECK-SAME: ) -> (
  # CHECK-SAME:   tensor<1xf32> {tf_saved_model.index_path = ["x"]})
  # CHECK-SAME: attributes {{.*}} tf_saved_model.exported_names = ["f0006_multiple_return_statements"]

// PerChannel-SAME:                                             %arg0: tensor<1x3x4x3xf32>,
// PerChannel-SAME:                                             %arg1: tensor<2x3x3x2xf32>)

// PerChannel-LABEL: func private @composite_depthwise_conv2d_fn_0(
// PerChannel-SAME:                                             %arg0: tensor<1x3x4x3xf32>,

      // CHECK-SAME: f = @_func
      // CHECK-SAME: device = "/device:test_device:0"
      %3 = "tf_device.launch_func"(%2) {device = "/device:test_device:0", func = @_func} : (tensor<?xf32>) -> tensor<?xf32>

      // CHECK: %[[CALL_OUTPUT_1:[0-9]*]] = "tf.PartitionedCall"(%[[CALL_OUTPUT_0]])
      // CHECK-SAME: f = @_func
      // CHECK-SAME: device = "/device:test_device:1"