%cst = arith.constant dense<[2, 100]> : tensor<2xi64>
  // CHECK: tac.cost = 4.000000e+01
  %0 = "tfl.concatenation"(%arg0, %arg1) {axis = 0 : i32, fused_activation_function = "NONE", tac.device = "GPU"} : (tensor<100xf32>, tensor<100xf32>) -> tensor<200xf32>
  // CHECK: tac.cost = 4.040000e+01
  %1 = "tfl.reshape"(%0, %cst) {tac.device = "GPU"} : (tensor<200xf32>, tensor<2xi64>) -> tensor<2x100xf32>
  func.return %1 : tensor<2x100xf32>
}

  // 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"

// 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

  %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"}> ({

  // CHECK-DAG: %cst = arith.constant dense<0.000000e+00> : tensor<10x10xf32>
  // CHECK-DAG: %cst_0 = arith.constant dense<1000> : tensor<2xi64>
  // CHECK-DAG: %cst_1 = arith.constant dense<1.000000e+00> : tensor<f32>
  // CHECK: %0 = "tfl.fill"(%cst_0, %cst_1) : (tensor<2xi64>, tensor<f32>) -> tensor<1000x1000xf32>
  // CHECK: return %cst, %0 : tensor<10x10xf32>, tensor<1000x1000xf32>

  %2 = "tf.LookupTableFindV2"(%0, %arg, %default) {device = "/CPU:0"} : (tensor<!tf_type.resource>, tensor<?x!tf_type.string>, tensor<i64>) -> tensor<*xi64>
  func.return %1, %2 : tensor<i64>, tensor<*xi64>
}

}

  %2 = "tf.XlaCallModule"(%1#0, %0) <{Sout = [#tf_type.shape<1x3>], dim_args_spec = [], disabled_checks = [], has_token_input_output = false, module = "", platforms = [], version = 5 : i64}> {_entry_function = @composite_dot_general_fn_1, _o...