func.func @func_0_CPU(%arg0: tensor<10x10x10xf32>, %arg1: tensor<10xf32>) -> tensor<10x10x10xf32> attributes {tac.device = "CPU", tac.interface_name = "func_0"} {
  %0 = "tfl.add"(%arg0, %arg1) {fused_activation_function = "RELU", tac.device = "CPU"} : (tensor<10x10x10xf32>, tensor<10xf32>) -> tensor<10x10x10xf32>
  %1 = "tfl.mul"(%0, %arg1) {fused_activation_function = "RELU", tac.device = "CPU"} : (tensor<10x10x10xf32>, tensor<10xf32>) -> tensor<10x10x10xf32>

  // CHECK: %[[SCATTER1:.*]] = "tf.TensorScatterAdd"(%[[CST1]], %arg1, %arg2) : (tensor<14xf32>, tensor<1x1xi32>, tensor<1xf32>) -> tensor<14xf32>
  // CHECK: %[[ADD:.*]] = "tf.AddV2"(%[[MUL]], %[[SCATTER1]]) : (tensor<14xf32>, tensor<14xf32>) -> tensor<14xf32>
  // CHECK: return %[[ADD]] : tensor<14xf32>
}

  %1 = stablehlo.constant dense<2.000000e+00> : tensor<10xf32>
  %2 = stablehlo.dot_general %arg0, %0, contracting_dims = [1] x [0], precision = [DEFAULT, DEFAULT] : (tensor<?x12544xf32>, tensor<12544x10xf32>) -> tensor<?x10xf32>
  %3 = shape.shape_of %2 : tensor<?x10xf32> -> tensor<2xindex>
  %4 = stablehlo.dynamic_broadcast_in_dim %1, %3, dims = [1] : (tensor<10xf32>, tensor<2xindex>) -> tensor<?x10xf32>

// CHECK:           %[[VAL_3:.*]] = "tf.LeftShift"(%[[VAL_1]], %[[VAL_0]]) : (tensor<4xi32>, tensor<1xi32>) -> tensor<4xi32>
// CHECK:           return %[[VAL_2]], %[[VAL_3]] : tensor<4xi32>, tensor<4xi32>
// CHECK:         }
func.func @broadcast_shift_left(%arg0: tensor<1xi32>, %arg1: tensor<4xi32>) -> (tensor<4xi32>, tensor<4xi32>) {

        %3 = "tf.Mul"(%arg0, %0) : (tensor<16xf32>, tensor<f32>) -> tensor<16xf32>
        %4 = "tf.Mul"(%3, %0) : (tensor<16xf32>, tensor<f32>) -> tensor<16xf32>
        %5 = "tf.Mul"(%4, %1) : (tensor<16xf32>, tensor<f32>) -> tensor<16xf32>
        %6 = "tf.Mul"(%5, %2) : (tensor<16xf32>, tensor<f32>) -> tensor<16xf32>
        tf_device.return %6 : tensor<16xf32>
      }) {} : () -> tensor<16xf32>
      tf_executor.yield %3 : tensor<16xf32>
    }

  %min = arith.constant dense<0.0> : tensor<16xf32>
  %max = arith.constant dense<255.0> : tensor<16xf32>
  %mini = "tf.Identity"(%min) : (tensor<16xf32>) -> tensor<16xf32>
  %maxi = "tf.Identity"(%max) : (tensor<16xf32>) -> tensor<16xf32>
  %fq = "tf.FakeQuantWithMinMaxVarsPerChannel"(%in, %mini, %maxi) {num_bits = 5, narrow_range = false} : (tensor<3x3x3x16xf32>, tensor<16xf32>, tensor<16xf32>) -> tensor<3x3x3x16xf32>

  %8 = "tfl.mul"(%2, %3) {fused_activation_function = "NONE"} : (tensor<4xi32>, tensor<4xi32>) -> tensor<4xi32>

  func.return %5, %6, %7, %8 : tensor<i32>, tensor<4xi32>, tensor<4xi32>, tensor<4xi32>
}

// CHECK-LABEL: @add_dense_splat_int
func.func @add_dense_splat_int() -> tensor<4xi32> {
  %0 = arith.constant dense<[-10, -1, 42, 100]> : tensor<4xi32>

// RUN: tf-opt %s -tfl-legalize-tf='preserve-assert-op=true' | FileCheck %s --dump-input=fail

func.func @preserve_assert(%arg0: tensor<1xi32>, %arg1: tensor<1xi32>) -> tensor<1xi1> {
  %0 = "tf.LessEqual"(%arg0, %arg1) : (tensor<1xi32>, tensor<1xi32>) -> tensor<1xi1>
  "tf.Assert"(%0, %arg1) {summarize = 3} : (tensor<1xi1>, tensor<1xi32>) -> ()
  func.return %0 : tensor<1xi1>
  // CHECK-LABEL: preserve_assert
  // CHECK: tfl.less_equal
  // CHECK: Assert

      %outputs_45, %control_46 = tf_executor.island wraps "tf.Const"() {device = "/device:CPU:0", value = dense<3> : tensor<1xi32>} : () -> tensor<1xi32>
      %outputs_47, %control_48 = tf_executor.island wraps "tf.Const"() {device = "/device:CPU:0", value = dense<1> : tensor<1xi32>} : () -> tensor<1xi32>

// CHECK-NEXT:      %18 = "tfl.pseudo_const"() <{value = dense<2> : tensor<1xi32>}> : () -> tensor<1xi32>
// CHECK-NEXT:      %19 = "tfl.concatenation"(%18, %17, %16) <{axis = 0 : i32, fused_activation_function = "NONE"}> : (tensor<1xi32>, tensor<1xi32>, tensor<1xi32>) -> tensor<3xi32>
// CHECK-NEXT:      %20 = mhlo.dynamic_reshape %2, %19 : (tensor<2x3x4x?xf32>, tensor<3xi32>) -> tensor<2x3x?xf32>