%0 = "tfl.pseudo_const" () {value = dense<1.0> : tensor<4xf32>} : () -> tensor<4xf32> loc("Const")
  %1 = "tfl.mul"(%arg0, %0) {fused_activation_function = "NONE"} : (tensor<4xf32>, tensor<4xf32>) -> tensor<4xf32> loc("mul0")
  %2 = "tfl.mul"(%1, %0) {fused_activation_function = "NONE"} : (tensor<4xf32>, tensor<4xf32>) -> tensor<4xf32> loc("mul1")
  %3 = "tfl.exp"(%2)  : (tensor<4xf32>) -> tensor<4xf32> loc("exp")
  func.return %3 : tensor<4xf32>

  %0 = "tf.FakeQuantWithMinMaxVars"(%arg0, %cst, %cst_0) {
    device = "", narrow_range = false, num_bits = 8 : i64
  } : (tensor<3xf32>, tensor<f32>, tensor<f32>) -> tensor<3xf32>
  %1 = "tf.FakeQuantWithMinMaxVarsPerChannel"(%arg1, %cst_1, %cst_2) {
    device = "", narrow_range = true, num_bits = 8 : i64
  } : (tensor<4x3xf32>, tensor<3xf32>, tensor<3xf32>) -> tensor<4x3xf32>
  func.return %0, %1 : tensor<3xf32>, tensor<4x3xf32>
}

        %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>
    }

func.func @sharding_custom_call_removed(%arg0: tensor<3xf32>) -> tensor<3xf32> {
  %1 = stablehlo.custom_call @Sharding(%arg0) {mhlo.sharding = ""} : (tensor<3xf32>) -> tensor<3xf32>
  return %1 : tensor<3xf32>
}
// CHECK-NOT: custom_call

// -----

// Tests that a custom_call that is not @Sharding is not removed.

// CHECK-LABEL: custom_call_not_removed
func.func @custom_call_not_removed(%arg0: tensor<3xf32>) -> tensor<3xf32> {

// CHECK-LABEL: func @main
func.func @main(%arg0 : tensor<10xf32>, %arg1 : tensor<10xf32>) -> tensor<10x10xf32> {
// CHECK-SAME: (%[[ARG0:.*]]: tensor<10xf32>, %[[ARG1:.*]]: tensor<10xf32>)
// CHECK-NEXT:    %[[ADD:.*]] = "tf.Add"(%[[ARG0]], %[[ARG1]]) : (tensor<10xf32>, tensor<10xf32>) -> tensor<10xf32>
// CHECK-NEXT:    %[[MAGIC:.*]] = "magic.op"(%[[ADD]], %[[ADD]]) : (tensor<10xf32>, tensor<10xf32>) -> tensor<10x10xf32>
// CHECK-NEXT:    return %[[MAGIC]]

  // Confirm that attributes that cannot be stored in the flatbuffer options
  // for a given operator are dropped silently.
  %1 = "tfl.squared_difference"(%arg0, %0) {fused_activation_function = "NONE"} : (tensor<4xf32>, tensor<4xf32>) -> tensor<4xf32> loc("squared_difference")
  %2 = "tfl.mul"(%arg0, %1) {fused_activation_function = "NONE"} : (tensor<4xf32>, tensor<4xf32>) -> tensor<4xf32> loc("mul")

    %cst_0 = "tf.Const"() {value = dense<1.000000e+00> : tensor<2xf32>} : () -> tensor<2xf32>

    %var_0 = "tf.VarHandleOp"() {shared_name = "var_0"} : () -> tensor<!tf_type.resource<tensor<2xf32>>>
    "tf.AssignVariableOp"(%var_0, %cst_0) : (tensor<!tf_type.resource<tensor<2xf32>>>, tensor<2xf32>) -> ()

    %var_1 = "tf.VarHandleOp"() {shared_name = "var_1"} : () -> tensor<!tf_type.resource<tensor<2xf32>>>

  // CHECK-DAG: %[[CST_0:.+]] = mhlo.constant dense<[1.000000e-01, 2.000000e-01]> : tensor<2xf32>
  // CHECK-DAG: %[[CST_1:.+]] = mhlo.constant dense<[3.000000e-01, 4.000000e-01]> : tensor<2xf32>
  // CHECK: %[[CST_BCAST:.+]] = "mhlo.broadcast_in_dim"(%[[CST_0]]) <{broadcast_dimensions = dense<3> : tensor<1xi64>}> : (tensor<2xf32>) -> tensor<1x1x3x2xf32>

module {
  // CHECK-LABEL: func @main
  // CHECK-SAME: %[[ARG_0:.*]]: tensor<10xf32>, %[[ARG_1:.*]]: tensor<10xf32>, %[[ARG_2:.*]]: tensor<6xi32>
  func.func @main(
      %arg0: tensor<10xf32>, %arg1: tensor<10xf32>, %arg2: tensor<6xi32>
    ) -> (tensor<10xf32>, tensor<6xi32>) {
    // CHECK: %[[CAST_0:.*]] = "tf.Cast"(%[[ARG_0]]) <{Truncate = false}> : (tensor<10xf32>) -> tensor<10xbf16>

    %add = "tf.AddV2"(%call, %call) : (tensor<4xf32>, tensor<4xf32>) -> tensor<4xf32>
    "tf.Yield"(%call, %add) : (tensor<4xf32>, tensor<4xf32>) -> ()
  },  {
    %call_0 = func.call @test_if_else1(%arg1) : (tensor<4xf32>) -> tensor<4xf32>
    "tf.Yield"(%call_0, %call_0) : (tensor<4xf32>, tensor<4xf32>) -> ()
  }) {is_stateless = false} : (tensor<i1>) -> (tensor<*xf32>, tensor<4xf32>)