// CHECK-DAG: %[[SCALE_BCAST:.+]] = "mhlo.broadcast_in_dim"(%[[SCALE]]) <{broadcast_dimensions = dense<1> : tensor<1xi64>}> : (tensor<256xf32>) -> tensor<4x256xf32>
  // CHECK-DAG: %[[OFFSET_BCAST:.+]] = "mhlo.broadcast_in_dim"(%[[OFFSET]]) <{broadcast_dimensions = dense<1> : tensor<1xi64>}> : (tensor<256xf32>) -> tensor<4x256xf32>

// CHECK: %[[XLA_CALL_MODULE:.+]] = "tf.XlaCallModule"(%[[CUSTOM_AGGREGATOR_0]], %[[CST]])
// CHECK-SAME: _entry_function = @composite_dot_general_fn_1, _original_entry_function = "composite_dot_general_fn_1"

    %4 = "tf.AddV2"(%3, %1) {device = ""} : (tensor<?xi64>, tensor<i64>) -> tensor<?xi64>
    %5 = "tf.LookupTableFindV2"(%0, %arg0, %cst) {device = ""} : (tensor<!tf_type.resource>, tensor<?x!tf_type.string>, tensor<i64>) -> tensor<*xi64>
    return %5 : tensor<*xi64>
  }

// Check that HashTable op is lifted.
// CHECK: func.func private @serving_default

func.func private @while_body(%arg0 : tensor<4xf32>, %arg1 : tensor<i32>) -> (tensor<4xi64>, tensor<i32>)
func.func @testWhileRegionIncompatibleCast(%arg0 : tensor<*xi64>, %arg1 : tensor<i32>) -> tensor<*xi64> {
  // CHECK: [[Result:%.*]]:2 = "tf.While"(%arg0, %arg1) <{body = @tf.WhileRegion_body, cond = @tf.WhileRegion_cond
  %0:2 = "tf.WhileRegion"(%arg0, %arg1) (
    {
      ^bb0(%carg0: tensor<*xi64>, %carg1: tensor<i32>):

// RUN: tf-opt %s -tfl-prepare-tf -tfl-legalize-tf='run-tfl-runtime-verification=false'  -tfl-optimize | FileCheck %s

func.func @broadcast_to_bf16(%arg0: tensor<3xbf16>, %arg1: tensor<2xi64>) -> tensor<3x3xbf16> {
  %0 = "tf.BroadcastTo"(%arg0, %arg1) : (tensor<3xbf16>, tensor<2xi64>) -> tensor<3x3xbf16>
  func.return %0: tensor<3x3xbf16>

// CHECK-LABEL: broadcast_to_bf16
// CHECK:  [[CST:%.*]] = arith.constant dense<1.000000e+00> : tensor<3x3xbf16>

      %0 = mhlo.constant dense<1.000000e+00> : tensor<f64>
      %1 = mhlo.convert %0 : (tensor<f64>) -> tensor<i64>
      %2 = mhlo.reshape %1 : (tensor<i64>) -> tensor<1xi64>
      %3 = "mhlo.dynamic_iota"(%2) {iota_dimension = 0 : i64} : (tensor<1xi64>) -> tensor<?xi32>
      %4 = mhlo.multiply %3, %3 : tensor<?xi32>
      return %4 : tensor<?xi32>
    }
  })";
  CellReader<int64_t> legal_error(kFailedLegalizationStreamz);

  %1:4 = "tf.CustomAggregator"(%arg0) {id = "1", calibration_method = 1 : i32, num_bins = 0 : i32, max_percentile = 0.000000e+00 : f32, min_percentile = 0.000000e+00 : f32} : (tensor<8x8x8x8xf32>) -> (tensor<8x8x8x8xf32>, tensor<f32>, tensor<f32>, tensor<*xi64>)
  func.return %0#0, %1#0 : tensor<8x8x8x8xf32>, tensor<8x8x8x8xf32>
}
// CHECK: func @customAggregator

    %1 = "tf.RngReadAndSkip"(%resource, %alg, %delta) : (tensor<!tf_type.resource<tensor<3xi64>>>, tensor<i32>, tensor<ui64>) -> tensor<3xi64>
    tf_device.return %1 : tensor<3xi64>
  }) : () -> tensor<3xi64>
  func.return %0 : tensor<3xi64>
}

// Test that ordering token resource of tf.CollectiveReduceV2 op is removed.
// CHECK-LABEL: func @decompose_collective_reduce_v2_op

  %cst = "tf.Const"() <{value = dense<0.000000e+00> : tensor<f32>}> : () -> tensor<f32>
  %cst_0 = "tf.Const"() <{value = dense<1> : tensor<1xi64>}> : () -> tensor<1xi64>
  %fill = "tf.Fill"(%cst_0, %cst) : (tensor<1xi64>, tensor<f32>) -> tensor<1xf32>
  tf_device.replicate([%0, %fill] as %arg_r0: tensor<1xf32>) {n = 2 : i32} {
    %1 = "tf_device.launch"() <{device = "TPU_REPLICATED_HOST_0"}> ({

    %3 = "tf.XlaCallModule"(%2#0, %0, %1) {Sout = [#tf_type.shape<1x3>], _entry_function = @composite_dot_general_fn_1, _original_entry_function = "composite_dot_general_fn_1", _tfl_quant_trait = "fully_quantizable", dim_args_spec = [], disabled_checks...