%0 = "mhlo.reshape"(%arg0) : (tensor<1x1x512xf32>) -> tensor<1x512xf32>
  %1 = "mhlo.dot"(%0, %arg1) : (tensor<1x512xf32>, tensor<512x13x!quant.uniform<i8:f32, 0.00285>>) -> tensor<1x13xf32>
  %2 = "mhlo.reshape"(%1) : (tensor<1x13xf32>) -> tensor<1x1x13xf32>
  func.return %2 : tensor<1x1x13xf32>

// CHECK:      %[[RES:.*]] = "mhlo.dot_general"(%arg0, %arg1) <{

    %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>
    %12 = stablehlo.subtract %10, %11 : tensor<1x3xf32>  // q1 * q2 - z1 * q2
    %13 = stablehlo.multiply %12, %4 : tensor<1x3xf32>  // s1 * s2

      time_major = false} : (
        tensor<1x2x3xf32>,
        tensor<1x1xf32>, tensor<1x1xf32>, tensor<1x1xf32>, tensor<1x1xf32>,
        tensor<1x1xf32>, tensor<1x1xf32>, tensor<1x1xf32>, tensor<1x1xf32>,
        none, none, none,
        tensor<3xf32>, tensor<3xf32>, tensor<3xf32>, tensor<3xf32>,
        none, none,
        tensor<1x3xf32>, tensor<1x3xf32>,
        none, none, none, none) -> tensor<1x2x3xf32>

// CHECK:           }) : (tensor<1x?xf32>, tensor<1x0xf32>, tensor<1x0xf32>, tensor<1x0xf32>, tensor<1x0xf32>, tensor<1x3xf32>, tensor<1x3xf32>, tensor<1x3xf32>, tensor<1x3xf32>, none, none, none, tensor<1xf32>, tensor<1xf32>, tensor<1xf32>, tensor<1xf32>, tensor<3x1xf32>, tensor<3xf32>, tensor<1x3xf32>, tensor<1x1xf32>, none, none, none, none) -> tensor<1x3xf32>
// CHECK:           [[VAL_50:%.*]] = tensor.cast [[VAL_51:%.*]] : tensor<1x3xf32> to tensor<1x?xf32>

      time_major = false} : (
        tensor<1x2x3xf32>,
        tensor<1x1xf32>, tensor<1x1xf32>, tensor<1x1xf32>, tensor<1x1xf32>,
        tensor<1x1xf32>, tensor<1x1xf32>, tensor<1x1xf32>, tensor<1x1xf32>,
        none, none, none,
        tensor<3xf32>, tensor<3xf32>, tensor<3xf32>, tensor<3xf32>,
        none, none,
        tensor<1x3xf32>, tensor<1x3xf32>,
        none, none, none, none) -> tensor<1x2x3xf32>

  %outputs_2 = "tf.add"(%outputs_0, %outputs_1): (tensor<1x1xf32>, tensor<1x1xf32>) -> tensor<1x1xf32>
  return %outputs_2 : tensor<1x1xf32>
}

// CHECK-LABEL: @_func
func.func private @_func(%arg0: tensor<1x3xf32>, %arg1: tensor<3x1xf32>) -> (tensor<1x1xf32>) {

// CHECK-NEXT:    "tf.MatMul"(%arg0, [[TENSOR]]) : (tensor<1x3xf32>, tensor<3x1xf32>) -> tensor<1x1xf32>
// CHECK-NEXT:    "tf.IfrtCall"(%arg0, [[ARRAYKEY]]) <{program_id = 6515870160938153680 : i64, variable_arg_indices = [1 : i32]}> {__tpu_compile_metadata_text = "retvals { sharding { } }"} : (tensor<1x3xf32>, tensor<!tf_type.string>) -> tensor<1x1xf32>
// CHECK-NEXT:    return
//

    %result = "tf.IfrtCall"(%arg0, %2) <{program_id = 6515870160938153680 : i64, variable_arg_indices = []}> : (tensor<1x3xf32>, tensor<3x1xf32>) -> (tensor<1x1xf32>)
    return %result, %3 : tensor<1x1xf32>, tensor<1x1xf32>
  }
}

// -----
// Variable tensor is only for host
//
// CHECK-LABEL:  func.func @serving_default(%arg0: tensor<1x3xf32>) -> tensor<1x1xf32> {
// CHECK:  "tf.VarHandleOp"

      %1 = "quantfork.qcast"(%0) {volatile} : (tensor<1x3xf32>) -> tensor<1x3x!quant.uniform<i8:f32, 0.13170163023705575:-1>>
      %2 = "quantfork.dcast"(%1) : (tensor<1x3x!quant.uniform<i8:f32, 0.13170163023705575:-1>>) -> tensor<1x3xf32>
      %3 = stablehlo.reshape %2 : (tensor<1x3xf32>) -> tensor<3x1xf32>

  // Just to introduce extra usages for %cst.
  %3 = "tf.Mul"(%cst, %cst) {device = ""} : (tensor<1x3xf32>, tensor<1x3xf32>) -> tensor<1x3xf32>

  return %2: tensor<1x3xi32>
}
// The output of tf.Shape is considered a compile-time constant, so the
// constant leading to tf.Shape (which transitively becomes an input to the
// first arg of tf.Fill) is not duplicated.