// CHECK: return %[[identity]] : tensor<1x3x2x2xf32>

func.func @cast_bf16_matmul_to_fp32(%arg0: tensor<1x10xf32>) -> (tensor<1x2xf32>) {
  %cst = "tf.Const"() {device = "", value = dense<1.000000e+01> : tensor<10x2xbf16>} : () -> tensor<10x2xbf16>
  %0 = "tf.Cast"(%arg0) {Truncate = false, device = ""} : (tensor<1x10xf32>) -> tensor<1x10xbf16>

  // 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>

  // CHECK: "tfl.pack"(%arg0, %arg1) <{axis = -2 : i32, values_count = 2 : i32}>
  %0 = "tfl.pack"(%arg0, %arg1) {axis = -2 : i32, values_count = 2 : i32} : (tensor<1x4xi32>, tensor<1x4xi32>) -> tensor<1x2x4xi32>
  func.return %0 : tensor<1x2x4xi32>
}

func.func @packNegInputAxis3(%arg0: tensor<1x4xi32>, %arg1: tensor<1x4xi32>) -> tensor<2x1x4xi32> {

  // CHECK-SAME: (%arg0: tensor<1x1x9xi32>, %arg1: tensor<1xi32>) -> tensor<1x1x8xi32>
  func.func @xla_gather(%arg0: tensor<1x1x9xi32>, %arg1: tensor<1xi32>) -> tensor<*xi32> {
    %cst = "tf.Const"() {value = dense<[1, 1, 8]> : tensor<3xi32>} : () -> tensor<3xi32>
    %0 = "tf.XlaGather"(%arg0, %arg1, %cst) {dimension_numbers = "\0A\03\00\01\02\1A\01\02", indices_are_sorted = true} : (tensor<1x1x9xi32>, tensor<1xi32>, tensor<3xi32>) -> tensor<*xi32>

// CHECK:}

func.func @reshape(%arg0 : tensor<1x128xi32>) -> tensor<4x32x1xi32>{
  %0 = "vhlo.reshape_v1"(%arg0) : (tensor<1x128xi32>) -> tensor<4x32x1xi32>
  func.return %0 : tensor<4x32x1xi32>
}

//CHECK:func.func private @reshape(%arg0: tensor<1x128xi32>) -> tensor<4x32x1xi32> {
//CHECK-NEXT: %0 = "vhlo.reshape_v1"(%arg0) : (tensor<1x128xi32>) -> tensor<4x32x1xi32>
//CHECK-NEXT: return %0 : tensor<4x32x1xi32>

}

func.func @concat_v2_i64_axis(%arg0: tensor<2x1xi32>, %arg1: tensor<2x1xi32>, %arg2: tensor<2x1xi32>) -> tensor<2x3xi32> {
  %0 = "tf.Const"() { value = dense<-1> : tensor<i64> } : () -> tensor<i64>
  %1 = "tf.ConcatV2"(%arg0, %arg1, %arg2, %0) : (tensor<2x1xi32>, tensor<2x1xi32>, tensor<2x1xi32>, tensor<i64>) -> tensor<2x3xi32>

  %2 = "tf.Add"(%1, %1) {
    _xla_compile_device_type = "TPU", _replication_info = "cluster",
    device = "/task:0/device:TPU:0", dtype = f32
  } : (tensor<1x80xf32>, tensor<1x80xf32>) -> tensor<1x80xf32>
  %3 = "tf.ResourceGatherNd"(%arg0, %0) {
    Tindices = i32
  } : (tensor<*x!tf_type.resource<tensor<80xf32>>>, tensor<i32>) -> tensor<1x80xf32>

  %0 = "mhlo.reshape"(%arg0) : (tensor<1x1x512xf32>) -> tensor<1x512xf32>
  %1 = "mhlo.dot"(%0, %arg1) : (tensor<1x512xf32>, tensor<512x13x!quant.uniform<i8:f32, 0.00285>>) -> tensor<1x13xf32>
  %2 = "mhlo.reshape"(%1) : (tensor<1x13xf32>) -> tensor<1x1x13xf32>
  func.return %2 : tensor<1x1x13xf32>

// CHECK:      %[[RES:.*]] = "mhlo.dot_general"(%arg0, %arg1) <{
// CHECK-SAME:   dot_dimension_numbers = #mhlo.dot<

For example, if we have the code

```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

    %0 = "tf.Const"() {value = dense<1> : tensor<i64>} : () -> tensor<i64>
    %1 = "tf.Const"() {value = dense<0> : tensor<1x1xi32>} : () -> tensor<1x1xi32>
    %2 = "tf.Const"() {value = dense<[7, 7, 3, 64]> : tensor<4xi32>} : () -> tensor<4xi32>
    %3 = "tf.Const"() {value = dense<[[0, 0], [3, 3], [3, 3], [0, 0]]> : tensor<4x2xi32>} : () -> tensor<4x2xi32>