func.func @rng_bit_generator(%arg0: tensor<2xui64>) -> (tensor<2xui64>, tensor<10x12xui32>) {
  %output_state, %output = "vhlo.rng_bit_generator_v1"(%arg0) <{rng_algorithm = #vhlo<rng_algorithm_v1 DEFAULT>}> : (tensor<2xui64>) -> (tensor<2xui64>, tensor<10x12xui32>)
  func.return %output_state, %output : tensor<2xui64>, tensor<10x12xui32>
}

// CHECK:func.func private @rng_bit_generator(%arg0: tensor<2xui64>) -> (tensor<2xui64>, tensor<10x12xui32>) {

  %0 = mhlo.composite "aten.avg_pool2d.default" %arg0 {composite_attributes = {ceil_mode = false, count_include_pad = true, divisor_override = "py_None", kernel_size = dense<3> : tensor<2xi64>, padding = dense<0> : tensor<2xi64>, stride = dense<1> : tensor<2xi64>}, decomposition = @XlaCallModule_aten.avg_pool2d.default.impl_0} : (tensor<1x3x6x6xf32>) -> tensor<1x3x4x4xf32>
  %1 = "tf.Identity"(%0) {device = ""} : (tensor<1x3x4x4xf32>) -> tensor<*xf32>

  // CHECK-NEXT: %12 = "mhlo.pad"(%10, %11) <{edge_padding_high = dense<[3, 4]> : tensor<2xi64>, edge_padding_low = dense<0> : tensor<2xi64>, interior_padding = dense<0> : tensor<2xi64>}> : (tensor<3x4xi32>, tensor<i32>) -> tensor<6x8xi32>
  // CHECK-NEXT: %13 = "mhlo.slice"(%arg1) <{limit_indices = dense<1> : tensor<1xi64>, start_indices = dense<0> : tensor<1xi64>, strides = dense<1> : tensor<1xi64>}> : (tensor<2xi32>) -> tensor<1xi32>

// CHECK-NEXT:      %16 = "tfl.pseudo_const"() <{value = dense<[0, 1]> : tensor<2xi64>}> : () -> tensor<2xi64>
// CHECK-NEXT:      %17 = "tfl.gather"(%14, %16) <{axis = 0 : i32, batch_dims = 0 : i32}> : (tensor<3xi32>, tensor<2xi64>) -> tensor<2xi32>
// CHECK-NEXT:      %18 = "tfl.pseudo_const"() <{value = dense<1> : tensor<1xi64>}> : () -> tensor<1xi64>

  // CHECK-DAG: %[[cst:.*]] = arith.constant dense<[2, 5, 21]> : tensor<3xi64>
  // CHECK-DAG: %[[cst_1:.*]] = arith.constant dense<[21, 4]> : tensor<2xi64>
  // CHECK: %[[v0:.*]] = "tf.Reshape"(%arg0, %[[cst]]) : (tensor<2x5x7x3xf32>, tensor<3xi64>) -> tensor<2x5x21xf32>
  // CHECK: %[[v1:.*]] = "tf.Reshape"(%arg1, %[[cst_1]]) : (tensor<7x3x4xf32>, tensor<2xi64>) -> tensor<21x4xf32>

func.func @testBroadcastGradientArgI64() -> (tensor<2xi64>, tensor<0xi64>) {
  %s0 = "tf.Const"() {value = dense<> : tensor<0xi64>} : () -> tensor<0xi64>
  %s1 = "tf.Const"() {value = dense<[2, 4]> : tensor<2xi64>} : () -> tensor<2xi64>
  %r0, %r1 = "tf.BroadcastGradientArgs"(%s0, %s1) {} : (tensor<0xi64>, tensor<2xi64>) -> (tensor<2xi64>, tensor<0xi64>)

  %cst_0 = arith.constant dense<[2, -1]> : tensor<2xi64>
  %0 = "tf.FakeQuantWithMinMaxVars"(%arg0, %arg1, %arg2) {num_bits = 5, narrow_range = false} : (tensor<1x2xf32>, tensor<f32>, tensor<f32>) -> tensor<1x2xf32>
  %1 = "tf.Reshape"(%0, %cst_0) : (tensor<1x2xf32>, tensor<2xi64>) -> tensor<2x1xf32>
  %2 = "tf.Reshape"(%0, %cst_0) : (tensor<1x2xf32>, tensor<2xi64>) -> tensor<2x1xf32>
  func.return %1, %2 : tensor<2x1xf32>, tensor<2x1xf32>

  %cst_0 = arith.constant dense<[2, -1]> : tensor<2xi64>
  %0 = "tf.FakeQuantWithMinMaxVars"(%arg0, %arg1, %arg2) {num_bits = 3, narrow_range = false} : (tensor<1x2xf32>, tensor<f32>, tensor<f32>) -> tensor<1x2xf32>
  %1 = "tf.Reshape"(%0, %cst_0) : (tensor<1x2xf32>, tensor<2xi64>) -> tensor<2x1xf32>
  %2 = "tf.Reshape"(%0, %cst_0) : (tensor<1x2xf32>, tensor<2xi64>) -> tensor<2x1xf32>
  func.return %1, %2 : tensor<2x1xf32>, tensor<2x1xf32>

    %24 = "tf.Softmax"(%18) : (tensor<2x1001xf32>) -> tensor<2x1001xf32>
    %25 = "tf.ArgMax"(%24, %2) : (tensor<2x1001xf32>, tensor<i32>) -> tensor<2xi64>
    %26 = "tf.Cast"(%25) {Truncate = false} : (tensor<2xi64>) -> tensor<2xf32>
    %27 = "tf.Equal"(%14, %26) {incompatible_shape_error = true} : (tensor<2xf32>, tensor<2xf32>) -> tensor<2xi1>

    // CHECK: "mhlo.pad"(%[[ARG0]], %[[ARG1]])
    // CHECK-SAME-DAG: edge_padding_high = dense<[1, 2]> : tensor<2xi64>
    // CHECK-SAME-DAG: edge_padding_low = dense<[2, 1]> : tensor<2xi64>
    // CHECK-SAME-DAG: interior_padding = dense<[1, 0]> : tensor<2xi64>

    %0 = mhlo.constant dense<[2, 1]> : tensor<2xi32>
    %1 = mhlo.constant dense<[1, 2]> : tensor<2xi32>