%32 = "tf.GatherV2"(%9, %31, %5) {batch_dims = 0 : i64, device = ""} : (tensor<32x1xf32>, tensor<2x16xi32>, tensor<i32>) -> tensor<2x16x1xf32>
  %33 = "tf.AddV2"(%17, %27) {device = ""} : (tensor<2x16xi32>, tensor<2x16xi32>) -> tensor<2x16xi32>
  %34 = "tf.GatherV2"(%9, %33, %5) {batch_dims = 0 : i64, device = ""} : (tensor<32x1xf32>, tensor<2x16xi32>, tensor<i32>) -> tensor<2x16x1xf32>

    // CHECK:      %[[ONE:.*]] = "tf.Const"() <{value = dense<1> : tensor<i32>}>
    // CHECK:      %[[RES_READ_VAL:[0-9]*]] = "tf.ReadVariableOp"
    // CHECK-SAME: (tensor<*x!tf_type.resource<tensor<2x8xi32>>>) -> tensor<2x8xi32>
    // CHECK:      "tf.AddV2"(%[[RES_READ_VAL]], %[[ONE]])
    // CHECK-SAME: (tensor<2x8xi32>, tensor<i32>) -> tensor<2x8xi32>
    // CHECK:      "tf.AssignVariableOp"

    // expected-remark@+1 {{Cost: 1}}
    %2 = "tf.Const"() {value = dense<64> : tensor<i32>} : () -> tensor<i32>
    // expected-remark@+1 {{Cost: 1}}
    %3 = "tf.Const"() {value = dense<2> : tensor<1xi32>} : () -> tensor<1xi32>
    // expected-remark@+1 {{Cost: 1}}
    %4 = "tf.Const"() {value = dense<1> : tensor<1xi32>} : () -> tensor<1xi32>
    // expected-remark@+1 {{Cost: 1}}

  // CHECK:  %5 = "tfl.arg_max"(%arg0, %cst) : (tensor<1x32x1xf32>, tensor<1xi32>) -> tensor<1x1xi32>
  // CHECK:  return %4, %5 : tensor<1x1xf32>, tensor<1x1xi32>
}

// CHECK-LABEL:   func @convert_pytorch_argmax
func.func @convert_pytorch_argmax(%arg0: tensor<1x9xi32>) -> tensor<1xi32> {
  %0 = mhlo.constant dense<0> : tensor<i32>

  %20 = "mhlo.gather"(%arg0, %19) <{dimension_numbers = #mhlo.gather<offset_dims = [0, 2, 3], collapsed_slice_dims = [1], start_index_map = [1], index_vector_dim = 1>, slice_sizes = dense<[1, 1, 2, 10]> : tensor<4xi64>}> : (tensor<1x2x2x10xf32>, tensor<4x1xi32>) -> tensor<1x4x2x10xf32>
  %21 = "mhlo.iota"() <{iota_dimension = 0 : i64}> : () -> tensor<4xf32>

func.func @testSubOfZero(%arg0: tensor<?x1xf32>, %arg1: tensor<4x1xf32>) -> (tensor<?x1xf32>, tensor<4x1xf32>) {
  %0 = "tf.Const"() {value = dense<0.0> : tensor<f32>} : () -> tensor<f32>
  %1 = "tf.Sub"(%arg0, %0) : (tensor<?x1xf32>, tensor<f32>) -> tensor<?x1xf32>
  %2 = "tf.Sub"(%arg1, %0) : (tensor<4x1xf32>, tensor<f32>) -> tensor<4x1xf32>
  func.return %1, %2: tensor<?x1xf32>, tensor<4x1xf32>

// CHECK: return %arg0, %arg1
}

  %119 = "tf.StridedSlice"(%116, %16, %15, %16) {begin_mask = 0 : i64, device = "", ellipsis_mask = 0 : i64, end_mask = 1 : i64, new_axis_mask = 0 : i64, shrink_axis_mask = 0 : i64} : (tensor<1xi32>, tensor<1xi32>, tensor<1xi32>, tensor<1xi32>) -> tensor<0xi32>

// Note: Argument quantization sequence omitted.
// CHECK: stablehlo.dot_general %{{.+}}, %{{.+}}, contracting_dims = [1] x [0] : (tensor<1x1024xi8>, tensor<1024x3xi8>) -> tensor<1x3xi32>

// Note: Result dequantization sequence omitted.
// CHECK: return %{{.+}} : tensor<1x3xf32>

// -----

// Tests that a simple dot_general (lifted as a function) with CustomAggregators

    ```mlir
      %0 = "tf.Const"() {value = dense<[[42.0]]> : tensor<1x1xf32>} : () -> tensor<1x1xf32>
      %1 = "tf.Const"() {device = "", value = dense<[[42.0]]> : tensor<1x1xf32>} : () -> tensor<1x1xf32>
      %2 = "tf.Const"() {device = "baz", value = dense<[[42.0]]> : tensor<1x1xf32>} : () -> tensor<1x1xf32>
    ```

    then running this pass with 'default-device=foobar', we get:

    ```mlir

func.func @tensorlistConst(%arg0 : tensor<1xi32>) -> tensor<2x3xi32> {
  // CHECK-DAG: %[[ELEMENT0:.*]] = "tf.Const"() <{value = dense<[0, 1, 2]> : tensor<3xi32>}> : () -> tensor<3xi32>
  // CHECK-DAG: %[[ELEMENT1:.*]] = "tf.Const"() <{value = dense<[3, 4, 5]> : tensor<3xi32>}> : () -> tensor<3xi32>
  // CHECK: %[[LIST:.*]] = "tf.Pack"(%[[ELEMENT0]], %[[ELEMENT1]]) <{axis = 0 : i64}> : (tensor<3xi32>, tensor<3xi32>) -> tensor<2x3xi32>