// the arguments into the order expected by XlaLaunch computations:
// 1) arguments
// 2) resource variable arguments
// See the documentation of EncapsulateSubgraphsInFunctions for the meaning
// of the arguments.
//
// TODO(b/113166435): Ordering constraints on XlaLaunch op can be relaxed.
Status RewriteSubgraph(const std::vector<OutputTensor>& arg_source_tensors,

    This pass rewrites `tf.PartitionedCall` and `tf.StatefulPartitionedCall`
    operations with `_xla_compile_device_type` attribute in a
    `tf_device.cluster` into `tf.XlaLaunch` operations. This makes the attached
    function execute with XLA. `tf.XlaLaunch` requires resource-type arguments
    come at the end, so this pass rewrites the called function if necessary.
    This pass assumes there are no nested `tf_device.cluster`s so we don't end

   public:
    // Allow using Device API (PjRt) for `device_type` in the XlaLaunch op.
    // Please note that `enabled_for_xla_launch_` needs to be true in addition
    // to the `device_type` being allowed in order to use the Device API for
    // single device compilation and execution in the XlaLaunch op.
    void AllowForDeviceInXlaLaunch(const DeviceType& device_type) {

  if (ctx->has_input(i) || ctx->has_input(++i)) {
    ctx->set_output(0, ctx->input(i));
  }
}

REGISTER_KERNEL_BUILDER(Name("XlaLaunch").Device(DEVICE_CPU), XlaLocalLaunchOp);

REGISTER_KERNEL_BUILDER(Name("XlaLaunchV2").Device(DEVICE_CPU), XlaLaunchV2Op);

REGISTER_KERNEL_BUILDER(Name("XlaLaunch")
                            .Device(DEVICE_GPU)
                            .HostMemory("constants")

  %unused = "tf.TestAsyncIdentity"(%x) {__op_key = 0: i32, T = i32} : (tensor<i32>) -> tensor<i32>
  // CHECK: mlrt.await_all_control [[unused]]
  return %x : tensor<i32>
}

// -----

// Test for XlaLaunch

func.func private @xla_func_0(%arg0: tensor<1x3xf32>, %arg1: tensor<1x3xf32>) -> tensor<1x3xf32> attributes {tf._XlaMustCompile = true, tf._noinline = true, tf._original_func_name = "should_not_be_used"} {

                              int device_ordinal, bool allocate_xla_tensors,
                              bool use_multiple_streams);

  // Builds a XlaCompiler::Argument vector from the arguments to an XlaLaunch
  // op.
  // Precondition: variables in `variable_args` are locked.
  static absl::StatusOr<std::vector<XlaCompiler::Argument>>
  BuildXlaCompilerArguments(absl::Span<int const> must_be_constant_idxs,

    with `_xla_compile_device_type` attribute into a `tf_device.cluster`.
    Notice this pass will only rewrite the outermost call if there are nested
    calls to avoid nested `tf.XlaLaunch` operations from being created later.

    For example, the following code

    ```mlir
    func.func @main() -> tensor<i32> {

  attributes {tf_saved_model.exported_names = ["main"]} {
  %0 = "tf.VarHandleOp"() {device = "/device:CPU:0", container = "", shared_name = "variable"} : () -> tensor<!tf_type.resource<tensor<1x3xf32>>>
  %1 = "tf.XlaLaunch"(%arg0, %0) {device = "/device:GPU:0", function = @xla_func, operandSegmentSizes = array<i32: 0, 2, 0>} : (tensor<1x3xf32>, tensor<!tf_type.resource<tensor<1x3xf32>>>) -> tensor<*xf32>
  func.return  %1 : tensor<*xf32>

}

// parent region.
std::unique_ptr<OperationPass<ModuleOp>> CreateXlaInlineDeviceOpsPass();

// Creates a pass that rewrites partitioned calls with `_xla_compile_device
// type` with `tf.XlaLaunch` ops.
std::unique_ptr<OperationPass<ModuleOp>> CreateXlaRewritePass();

// Create a pass that validates the input graph to the CPU/GPU bridge.
std::unique_ptr<OperationPass<ModuleOp>> CreateXlaValidateInputsPass();