"tf_executor.yield"(%0) : (tensor<f32>) -> ()
      }) : (!tf_executor.control) -> (tensor<f32>, !tf_executor.control)
      %outputs_14, %control_15 = "tf_executor.island"() ({
        %0 = "tf.Const"() {device = "", value = dense<1.000000e+00> : tensor<f32>} : () -> tensor<f32>
        "tf_executor.yield"(%0) : (tensor<f32>) -> ()

      %cst: tensor<f32>) -> tensor<1x64x56x56xf32> {
  %tpos = stablehlo.transpose %input, dims = [0, 3, 1, 2]
      : (tensor<1x114x114x64xf32>) -> tensor<1x64x114x114xf32>
  %ret = "stablehlo.reduce_window"(%tpos, %cst)
      <{window_dimensions = array<i64: 1, 1, 3, 3>, window_strides = array<i64: 1, 1, 2, 2>}> ({
  ^bb0(%arg0: tensor<f32>, %arg1: tensor<f32>):
    %max = stablehlo.maximum %arg0, %arg1 : tensor<f32>

        "tf.Yield"(%4) : (tensor<f32>) -> ()
      }, {
        // CHECK: device = "/device:TPU_REPLICATED_CORE:0"
        %5 = "tf.Div" (%arg1, %2) {device = "/device:TPU_REPLICATED_CORE:0"} : (tensor<f32>, tensor<f32>) -> tensor<f32>
        "tf.Yield"(%5) : (tensor<f32>) -> ()
      }) {is_stateless = true, _xla_compile_device_type = "TPU", _replication_info = "x" } : (tensor<i1>) -> (tensor<f32>)

func.func @interleave(%arg0: f32) -> (f32, f32, f32) attributes {ignore_side_effects_for_testing} {
  %0 = "x.a"(%arg0) : (f32) -> f32
  %1 = "y.a"(%arg0) : (f32) -> f32
  %2 = "z.a"(%arg0) : (f32) -> f32
  %3 = "x.b"(%0) : (f32) -> f32
  %4 = "y.b"(%1) : (f32) -> f32
  %5 = "z.b"(%2) : (f32) -> f32
  %6 = "x.c"(%3) : (f32) -> f32
  %7 = "y.c"(%4) : (f32) -> f32
  %8 = "z.c"(%5) : (f32) -> f32
  func.return %6, %7, %8 : f32, f32, f32

// CHECK-NEXT: [[res:%.+]] = "quantfork.dcast"([[conv]]) : (tensor<*x!quant.uniform<i8:f32, 0.023529411764705882:-128>>) -> tensor<*xf32>

  %5 = "tfl.add"(%4, %arg3) {tac.device = "CPU", tac.inference_type = "QUANTIZED_INT8", fused_activation_function = "NONE"} : (tensor<1x384x128x!quant.uniform<i8:f32, 0.3:3>>, tensor<128x!quant.uniform<i8:f32, 0.2:-4>>) -> tensor<1x384x128x!quant.uniform<i8:f32, 0.3:-3>>
  func.return %5 : tensor<1x384x128x!quant.uniform<i8:f32, 0.3:-3>>

// CHECK:         %[[ADD:.*]] = f32[] add(f32[] %[[DIV]], f32[]
// CHECK:         %[[CLAMP:.*]] = f32[] clamp(f32[] %[[MIN:.*]], f32[] %[[ADD]],
// CHECK:         %[[ROUND:.*]] = f32[] round-nearest-even(f32[] %[[CLAMP]])
// CHECK:         %[[CONVERT_0:.*]] = u8[] convert(f32[] %[[ROUND]])
// CHECK:         %[[CONVERT_1:.*]] = s32[] convert(u8[] %[[CONVERT_0]])
// CHECK:         %[[SUB:.*]] = s32[] subtract(s32[] %[[CONVERT_1]],

func.func private @f(%arg0: f32, %arg1: f32) -> (f32, f32) {
  %0 = "some_op"(%arg0) : (f32) -> f32
  return %0, %arg1 : f32, f32
}

// CHECK-LABEL: @removes_second_passthrough_arg
func.func @removes_second_passthrough_arg(%arg0: f32, %arg1: f32) -> (f32, f32) {
  // CHECK: %0 = call @f(%arg0)
  %0, %1 = call @f(%arg0, %arg1) : (f32, f32) -> (f32, f32)
  // CHECK: return %0, %arg1
  return %0, %1 : f32, f32
}

// -----

    // CHECK-DAG: %[[CST_1:.*]] = "tf.Const"() <{value = dense<3> : tensor<i32>}> : () -> tensor<i32>
    // CHECK-NEXT: %[[UQ:.*]] = "tf.UniformQuantize"(%arg0, %cst, %cst_0) <{quantization_axis = -1 : i64, quantization_max_val = 127 : i64, quantization_min_val = -128 : i64}> : (tensor<1xf32>, tensor<f32>, tensor<i32>) -> tensor<1x!tf_type.qint8>

  %3 = tfrt_fallback_async.executeop key(1) cost(512) device("/job:localhost/replica:0/task:0/device:CPU:0") "tf.AddV2"(%0, %0) {T = f32} : 1
  %4 = tfrt_fallback_async.executeop key(1) cost(512) device("/job:localhost/replica:0/task:0/device:CPU:0") "tf.AddV2"(%0, %0) {T = f32} : 1
  %5 = tfrt_fallback_async.executeop key(1) cost(512) device("/job:localhost/replica:0/task:0/device:CPU:0") "tf.AddV2"(%0, %0) {T = f32} : 1