// CHECK-NEXT: tf_mlrt.executeop([[y]]
  %v = "tf.AddV2"(%y, %z) {__op_key = 3: i32}: (tensor<1xi32>, tensor<1xi32>) -> tensor<1xi32>
  return %w, %u, %v : tensor<1xi32>, tensor<1xi32>, tensor<1xi32>
}

// -----

// Test node names are preserved.

// CHECK-LABEL: func @main
func.func @main(%x: tensor<i32>) -> tensor<i32> {
  // CHECK: tf_mlrt.executeop

module {
  func.func @simpleTest(%arg0: tensor<1xf32>, %arg1: tensor<1xf32>, %arg2: tensor<1xf32>, %arg3: tensor<1xf32>) -> tensor<2x1xf32> {
    %0 = func.call @func_0_GPU_FLOAT(%arg0, %arg1, %arg2) {tac.interface_name = "func_0"} : (tensor<1xf32>, tensor<1xf32>, tensor<1xf32>) -> tensor<1xf32>
    %1 = func.call @func_1_GPU_FLOAT(%arg0, %arg3) {tac.interface_name = "func_1"} : (tensor<1xf32>, tensor<1xf32>) -> tensor<1xf32>

    %4 = "tf.LayerNorm"(%a1, %a2, %a3, %a4) {_tfl_quant_trait = "fully_quantizable", device = ""} : (tensor<128x128xf32>, tensor<1xf32>, tensor<1xf32>, tensor<1xi32>) -> tensor<128x128xf32>
   "tfl.yield"(%4) : (tensor<128x128xf32>) -> ()
  }) {_tfl_quant_trait = "fully_quantizable", device = ""} : (tensor<128x128xf32>, tensor<1xf32>, tensor<1xf32>, tensor<1xi32>) -> tensor<128x128xf32>

  %b = arith.constant dense<0.0> : tensor<16xf32>
  %conv_3d = "tfl.conv_3d"(%arg0, %w, %b) {dilation_d_factor = 1 : i32, dilation_h_factor = 1 : i32, dilation_w_factor = 1 : i32, fused_activation_function = "NONE", padding = "SAME", stride_d = 1 : i32, stride_h = 1 : i32, stride_w = 1 : i32} : (tensor<1x32x32x32x8xf32>, tensor<1x1x1x8x16xf32>, tensor<16xf32>) -> tensor<1x32x32x32x16xf32>
  func.return %conv_3d : tensor<1x32x32x32x16xf32>

// PerTensor-LABEL: QuantizeGatherWeightOnly
func.func @QuantizeGatherWeightOnly(%arg0: tensor<3xi32>) -> tensor<3x3x3x3xf32> {
  %w = arith.constant dense<1.270000e+02> : tensor<64x3x3x3xf32>
  %emb = "tfl.gather"(%w, %arg0) {axis = 0 : i32, batch_dims = 0 : i32} : (tensor<64x3x3x3xf32>, tensor<3xi32>) -> tensor<3x3x3x3xf32>

  %cst = "tf.Const"() {value = dense<0.000000e+00> : tensor<10xf32>} : () -> tensor<10xf32>
  %cst_0 = "tf.Const"() {value = dense<[-1, 10]> : tensor<2xi32>} : () -> tensor<2xi32>
  %1 = "tf.MatMul"(%arg0, %arg1) {
    transpose_a = false, transpose_b = true
  } : (tensor<1x10xf32>, tensor<10x10xf32>) -> tensor<*xf32>
  %2 = "tf.Reshape"(%1, %cst_0) : (tensor<*xf32>, tensor<2xi32>) -> tensor<*xf32>

    %s0 = "tf.Const"() {value = dense<[501, 1, 32, 1280]> : tensor<4xi32>} : () -> tensor<4xi32>
    %s1 = "tf.Const"() {value = dense<[  1, 1,  1, 1280]> : tensor<4xi32>} : () -> tensor<4xi32>

    // CHECK: "tf.BroadcastGradientArgs"
    // CHECK-NOT: _xla_outside_compilation
    %r0, %r1 = "tf.BroadcastGradientArgs"(%s0, %s1) {} : (tensor<4xi32>, tensor<4xi32>) -> (tensor<1xi32>, tensor<3xi32>)
    tf_device.return

//          (tensor<?x3x4x2xf32>) -> tensor<i32>
// %5 = stablehlo.reshape %4 :
//          (tensor<i32>) -> tensor<1xi32>
// %6 = stablehlo.concatenate %5, %0, %1, %2, dim = 0 :
//          (tensor<1xi32>, tensor<1xi32>, tensor<1xi32>, tensor<1xi32>)
//            -> tensor<4xi32>
// %7 = stablehlo.dynamic_broadcast_in_dim %arg2, %6
// %8 = stablehlo.add %3, %7
// ```
//

// CHECK-DAG:       %[[VAL_3:.*]] = "tfl.pseudo_const"{{.*}}dense<2> : tensor<1xi32>
// CHECK-DAG:       %[[VAL_4:.*]] = "tfl.pseudo_const"{{.*}}dense<[2, 1]> : tensor<2xi32>
// CHECK:           %[[VAL_5:.*]] = "tfl.reshape"(%[[VAL_0]], %[[VAL_2]]) : (tensor<1xf32>, tensor<4xi32>) -> tensor<1x1x1x1xf32>
// CHECK:           %[[VAL_6:.*]] = "tfl.reshape"(%[[VAL_1]], %[[VAL_2]]) : (tensor<1xf32>, tensor<4xi32>) -> tensor<1x1x1x1xf32>

    %8 = stablehlo.convert %7 : (tensor<1x2xi32>) -> tensor<1x2xf32>
    %9 = stablehlo.convert %2 : (tensor<2x3xi8>) -> tensor<2x3xf32>
    %10 = stablehlo.dot_general %8, %9, contracting_dims = [1] x [0] : (tensor<1x2xf32>, tensor<2x3xf32>) -> tensor<1x3xf32>
    %11 = stablehlo.convert %3 : (tensor<1x3xi32>) -> tensor<1x3xf32>