%2 = mhlo.constant dense<3.40282347E+38> : tensor<f32>
  %3 = "mhlo.broadcast_in_dim"(%2) <{broadcast_dimensions = dense<> : tensor<0xi64>}> : (tensor<f32>) -> tensor<2xf32>
  %4 = mhlo.constant dense<3.000000e+00> : tensor<f32>
  %5 = "mhlo.broadcast_in_dim"(%4) <{broadcast_dimensions = dense<> : tensor<0xi64>}> : (tensor<f32>) -> tensor<2xf32>

    // expected-remark@+1 {{Cost: 2}}

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>)

func.func @slice_variable_start(%arg0: tensor<3x4xi32>, %arg1: tensor<2xi64>) -> tensor<1x4xi32> {
  // CHECK: %[[SLICED_START1:.*]] = "mhlo.slice"(%arg1)
  // CHECK-DAG-SAME: {limit_indices = dense<1> : tensor<1xi64>,
  // CHECK-DAG-SAME: start_indices = dense<0> : tensor<1xi64>,
  // CHECK-DAG-SAME: strides = dense<1> : tensor<1xi64>} : (tensor<2xi64>) -> tensor<1xi64>

// 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>

        padding = #vhlo.tensor_v1<dense<[[0, 0], [159, 0], [0, 0]]> : tensor<3x2xi64>>,
        window_dilations = #vhlo.tensor_v1<dense<1> : tensor<3xi64>>,
        window_dimensions = #vhlo.tensor_v1<dense<[1, 160, 1]> : tensor<3xi64>>,
        window_strides = #vhlo.tensor_v1<dense<1> : tensor<3xi64>>}> ({
    ^bb0(%arg23: tensor<f32>, %arg24: tensor<f32>):

  // 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>

    // 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>

  %1 = "tfl.shape"(%conv_s) : (tensor<?x26x26x26x16xf32>) -> tensor<5xi64>
  %2 = "tfl.broadcast_args"(%1, %cst) : (tensor<5xi64>, tensor<1xi64>) -> tensor<5xi64>
  %broad1 = "tfl.broadcast_to"(%conv_s, %2) : (tensor<?x26x26x26x16xf32>, tensor<5xi64>) -> tensor<?x26x26x26x16xf32>
  %broad2 = "tfl.broadcast_to"(%b, %2) : (tensor<16xf32>, tensor<5xi64>) -> tensor<?x26x26x26x16xf32>

  // CHECK-DAG: %[[X_NORMED:.+]] = mhlo.multiply %[[X]], %[[MULTIPLIER_BCAST]] : tensor<4x256xf32>
  // CHECK-DAG: %[[RHS_BCAST:.+]] = "mhlo.broadcast_in_dim"(%[[RHS]]) <{broadcast_dimensions = dense<1> : tensor<1xi64>}> : (tensor<256xf32>) -> tensor<4x256xf32>