// CHECK-NOT: tf.RestoreV2
  %0 = "tf.RestoreV2"(%cst, %cst_1, %cst_0): (tensor<!tf_type.string>, tensor<1x!tf_type.string>, tensor<1x!tf_type.string>) -> tensor<3x1xf32>
  %1 = "tf.VarHandleOp"() <{container = "", shared_name = "y"}> : () -> tensor<!tf_type.resource<tensor<3x1xf32>>>
  return
}


// CHECK-LABEL: func.func @used_restore_remains
func.func @used_restore_remains() {

module attributes {tf_saved_model.semantics} {
  func.func private @tf_quant__save(%arg: tensor<!tf_type.string>) -> () {
    tf_executor.graph {
      %out, %ctl = tf_executor.island wraps "tf.VarHandleOp"() {shared_name = "var_0"} : () -> tensor<!tf_type.resource<tensor<2xf32>>>
      %out_0, %ctl_0 = tf_executor.island wraps "tf.ReadVariableOp"(%out) : (tensor<!tf_type.resource<tensor<2xf32>>>) -> tensor<2xf32>

  int num_retvals = 0;
  std::vector<AbstractTensorHandle*> dummy_outputs;
  return op_ptr->Execute(absl::MakeSpan(dummy_outputs), &num_retvals);
}

// Op: VarHandleOp()
// Summary:
//
// Description:

    %0 = tf_executor.graph {
      %outputs_4, %control_5 = tf_executor.island wraps "tf.VarHandleOp"() {container = "", shared_name = "Variable"} : () -> tensor<!tf_type.resource<tensor<!tf_type.string>>>
      %outputs_10, %control_11 = tf_executor.island wraps "tf.VarHandleOp"() {container = "", shared_name = "Variable_1"} : () -> tensor<!tf_type.resource<tensor<!tf_type.string>>>

#include "tensorflow/compiler/tf2xla/xla_op_registry.h"
#include "tensorflow/core/framework/shape_inference.h"

namespace tensorflow {
namespace {

// Implementation of varhandle that binds a VarHandleOp to an XlaResource of the
// same name. It is not safe to use this op in a JIT context.
class XlaAotOnlyVarHandleOp : public XlaOpKernel {
 public:
  explicit XlaAotOnlyVarHandleOp(OpKernelConstruction* c);

# CHECK: [[v1:%.*]] = "tf.Identity"([[v0]])
# CHECK: [[handle:%.*]] = "tf.VarHandleOp"
# CHECK-SAME: shared_name = [[shared_name:".*"]]
# CHECK: "tf.AssignVariableOp"([[handle]], [[v1]])

# CHECK:      func {{@[a-zA-Z_0-9]+}}(
# CHECK-SAME: tf_saved_model.exported_names = ["key"]
# CHECK: tf.VarHandleOp
# CHECK-SAME: shared_name = [[shared_name]]


def Test():

// CHECK-NOT: "tf_saved_model.session_initializer"
"tf_saved_model.session_initializer"() { initializers = [@func_init] } : () -> ()

// CHECK-LABEL: _tfrt_resource_init
// CHECK: tf.VarHandleOp
// CHECK: tf.ReadVariableOp
// CHECK: tfrt_fallback_async.set_resource
// CHECK-SAME: {device = "/device:CPU:0", index = 0 : i64}


// CHECK-LABEL: func @init
// CHECK-SAME: {tfrt.cost_threshold = 1 : i64}

Clustered nodes: 2178
Unclustered nodes: 446
Number of clusters: 1

unclustered size 446
 AssignAddVariableOp 2
 Const 2
 DivNoNan 1
 Identity 1
 NoOp 1
 ReadVariableOp 2
 VarHandleOp 435
 _Retval 2
cluster 0 size 2178
 Add 17
 AddN 72
 ArgMax 1
 AssignAddVariableOp 1
 AssignSubVariableOp 106
 BiasAdd 1
 BiasAddGrad 1
 Cast 3
 Const 357
 Conv2D 53
 Conv2DBackpropFilter 53
 Conv2DBackpropInput 52

// CHECK-NEXT: [[o_chain:%.*]], [[o:%.*]] = tfrt_fallback_async.executeop.seq([[in_chain]]) key(0) cost({{.*}}) device("/job:localhost/replica:0/task:0/device:CPU:0") "tf.VarHandleOp"()
// CHECK-NEXT: [[o_chain_0:%.*]], [[o1:%.*]] = tfrt_fallback_async.executeop.seq([[in_chain]]) key(1) cost({{.*}}) device("/job:localhost/replica:0/task:0/device:CPU:0") "tf.ReadVariableOp"([[o]]) {dtype = f32} : 1

  func.return %0 : tensor<10xf32>
}

func.func @main(%arg0: tensor<7xf32>) -> tensor<10xf32> attributes {tf.entry_function = {control_outputs = "", inputs = "input:0", outputs = "output:0"}} {
  %0 = "tf.VarHandleOp"() {device = "/device:CPU:0", container = "", shared_name = "variable"} : () -> tensor<!tf_type.resource<tensor<f32>>>
  %1 = "tf.ReadVariableOp"(%0) {device = "/device:CPU:0"} : (tensor<!tf_type.resource<tensor<f32>>>) -> tensor<f32>