result.replaceAllUsesWith(if_op.getOperand(then_aliased_input + 1));
    }
  }
  if_op.replaceAllUsesWith(new_if);
  if_op.erase();
  return success();
}

// Handles stack usage by a tf.StatefulPartitionedCall or a tf.PartitionedCall.
// It will first check if the callee was previously handled, and try to reuse
// that result if so. Otherwise, it will clone and convert the callee function,

  // CHECK: [[results_0:%.*]] = call @callee(
  // CHECK-SAME: (!tf_mlrt.tensor) -> !tf_mlrt.tensor
  %1 = "tf.StatefulPartitionedCall"(%0) {config = "", config_proto = "", executor_type = "", f = @callee} : (tensor<i32>) -> (tensor<i32>)
  // CHECK-NEXT: [[results_1:%.*]] = call @callee(
  // CHECK-SAME: (!tf_mlrt.tensor) -> !tf_mlrt.tensor

  // expected-remark@above {{ID: 0}}
  %r1 = "tf.ReadVariableOp"(%arg0) : (tensor<!tf_type.resource>) -> tensor<1xf32>
  // expected-remark@above {{ID: 1}}
  %two = "tf.StatefulPartitionedCall"(%one, %one) {config="", config_proto="", executor_type="", f=@add} : (tensor<1xf32>, tensor<1xf32>) -> tensor<1xf32>
  // expected-remark@above {{ID: 2}}

                                                arg_buffer_size_is_fixed(i)};
    }
  }
  UpdateFuncType(func);
}

// Holds information about a decomposed callee function for
// PartitionedCall/StatefulPartitionedCall.
struct PartitionedCallDecompositionInfo {
  bool signature_change;
  func::FuncOp decomposed_callee;
  llvm::SmallDenseMap<int64_t, int64_t> buffer_arg_to_size_arg;

    // (checked later).
    if (auto device_type_attr =
            op.getAttrOfType<StringAttr>(mlir::TF::kCompileDeviceTypeAttr)) {
      // tf.StatefulPartitionedCall ops with and without
      // _tpu_replicate attributes may exist in the same graph. Ops without
      // the attribute but with _XlaMustCompile=true would have
      // _xla_compile_device_type="" after

    %0 = "tf.PartitionedCall"(%arg0) {config = "", config_proto = "", executor_type = "", f = @partitioned_called_func} : (tensor<20xi32>) -> tensor<*xi32>
    // CHECK: tf.StatefulPartitionedCall
    // CHECK-SAME: (tensor<20xi32>) -> tensor<20xi32>
    %1 = "tf.StatefulPartitionedCall"(%arg0) {config = "", config_proto = "", executor_type = "", f = @stateful_partitioned_call_func} : (tensor<20xi32>) -> tensor<*xi32>
    // CHECK: tf.TPUPartitionedCall

func.func @func_with_device_training_loop(%arg0: tensor<*xi32>, %arg1: tensor<*xi1>) -> (tensor<*xi32>, tensor<*xi1>) {
  %1:2 = "tf.StatefulPartitionedCall"(%arg0){f= @func_body, config="", config_proto="", executor_type=""}
         : (tensor<*xi32>) -> (tensor<*xi32>, tensor<*xi1>)

  func.return %arg0 : tensor<i32>
}

// -----

// Test valid tf.StatefulPartitionedCall
// CHECK-LABEL: func @testValidStatefulPartitionedCall
func.func @testValidStatefulPartitionedCall(%arg0: tensor<i32>) -> tensor<i32> {
  %0 = "tf.StatefulPartitionedCall"(%arg0) {config = "", config_proto = "", executor_type = "", f = @pcall_func} : (tensor<i32>) -> (tensor<i32>)

  // CHECK: call @callee() : () -> ()
  // CHECK: call @callee() : () -> ()
  "tf.PartitionedCall"() {config = "", config_proto = "", executor_type = "", f = @callee} : () -> ()
  "tf.StatefulPartitionedCall"() {config = "", config_proto = "", executor_type = "", f = @callee} : () -> ()
  "tf.LegacyCall"() {config = "", config_proto = "", executor_type = "", f = @callee} : () -> ()
  func.return
}

      "SpaceToBatchND",
      "SpaceToDepth",
      "SparseMatMul",
      "SparseToDense",
      "StackCloseV2",
      "StackPopV2",
      "StackPushV2",
      "StackV2",
      "StatefulPartitionedCall",
      "StatefulStandardNormalV2",
      "StatefulTruncatedNormal",
      "StatefulUniform",
      "StatefulUniformFullInt",
      "StatefulUniformInt",
      "StatelessCase",