// 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>

    %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

// 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"

// -----

func.func @foldDivF64() -> tensor<2xf64> {
  %0 = stablehlo.constant dense<[2.0, 3.0]> : tensor<2xf64>
  %1 = stablehlo.constant dense<[4.0, 6.0]> : tensor<2xf64>
  %2 = stablehlo.divide %0, %1 : tensor<2xf64>
  return %2 : tensor<2xf64>
}

// CHECK-LABEL: foldDivF64
// CHECK: stablehlo.constant dense<5.000000e-01> : tensor<2xf64>

// -----

func.func @foldDivI32() -> tensor<2xi32> {

  %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...

  %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...

  // CHECK-SAME: <{calibration_methods = [1 : i32, 1 : i32], ids = ["0", "1"], output_file_path = "main_0.pb"}> : (tensor<f32>, tensor<f32>, tensor<0xi64>, tensor<f32>, tensor<f32>, tensor<0xi64>) -> ()
  // CHECK: return