This pass expands outside compilation attributes to Identity/Cast ops
at the head of TPU computation if it's only used by outside compiled ops.
### `-tf-tpu-identity-pruning`

_Removes Identity/IdentityN ops from the TPU computation_

### `-tf-tpu-merge-variables-with-execute`

_Merges device variable reads and updates into TPU execute ops_

  %3 = "tf.ReadVariableOp"(%0) {_xla_compile_device_type = "TPU", _replication_info = "a"} : (!tf_res) -> tensor<f32>
  %4 = "tf.ReadVariableOp"(%1) {_xla_compile_device_type = "TPU", _replication_info = "b"} : (!tf_res) -> tensor<f32>
  %5 = "tf.ReadVariableOp"(%2) {_xla_compile_device_type = "TPU", _replication_info = "c"} : (!tf_res) -> tensor<f32>
  %6 = "tf.Identity"(%3) {_xla_compile_device_type = "TPU", _replication_info = "a"} : (tensor<f32>) -> tensor<f32>

# ARM64 WHEEL
test:linux_arm64_wheel_test_filters --test_tag_filters=-no_oss,-no_aarch64,-oss_excluded,-oss_serial,-gpu,-tpu,-benchmark-test,-v1only,-no_oss_py38,-no_oss_py39,-no_oss_py310

  let summary = "Op that compiles a computation in MLIR into a TPU program, and loads and executes it on a TPU device.";

  let description = [{
For the internal use of the TPU compiler.

'static_shapes' are tensors specifying the maximum dimension sizes for the tensors specified in `dynamic_operands`.
'args' are inputs to the TPU computation.

  // For all other inputs that go from TPU op to TPU op, insert the
  // TPUOutput/Input pair.

  // Add TPUReplicatedInput/TPUReplicatedOutput pairs along each edge.
  // TODO(bfontain): Should be merged with the above loop.
  llvm::SetVector<Value> values_to_add_nodes;

  for (Operation* op : backward_pass_ops) {
    // TODO(bfontain): Should validate that all the TPU ops are in the same
    // replication region.

		return
	}
	tp0 := s.typePointersOfType(typ, addr)
	tp1 := s.typePointersOf(addr, size)
	failed := false
	for {
		var addr0, addr1 uintptr
		tp0, addr0 = tp0.next(addr + size)
		tp1, addr1 = tp1.next(addr + size)
		if addr0 != addr1 {
			failed = true
			break
		}
		if addr0 == 0 {
			break
		}
	}
	if failed {
		tp0 := s.typePointersOfType(typ, addr)

    ]),
    alwayslink = 1,
)

cc_library(
    name = "xla_tpu_jit",
    visibility = ["//visibility:public"],
    deps = if_libtpu([
        "//tensorflow/core/tpu/graph_rewrite:tpu_rewrite_pass_registration",
        "//tensorflow/core/tpu/graph_rewrite:configure_tpu_embedding_rewrite_registration",
        "@local_xla//xla/stream_executor/tpu:tpu_transfer_manager",
    ]),
    alwayslink = 1,
)

cc_library(

    ETCD_PID=
  fi
}

function print_color {
  message=$1
  prefix=${2:+$2: } # add colon only if defined
  color=${3:-1}     # default is red
  echo -n "$(tput bold)$(tput setaf "${color}")"
  echo "${prefix}${message}"
  echo -n "$(tput sgr0)"
}

function warning_log {
  print_color "$1" "W$(date "+%m%d %H:%M:%S")]" 1
}

function start_etcd {
    echo "Starting etcd"

// TODO(b/117085735): This form of clustering should be prevented.
TEST(XlaCompilationTest, NOT_DontClusterSpreadingNodes) {
  // MatMulSource below creates data for nodes on GPU0 and GPU1 and is placed
  // on GPU0. However, it should not be clustered with the next node on
  // GPU0, because that will prevent the node on GPU1 from beginning its work as
  // soon as the data has been produced.
  //
  // This graph is:

// _XlaSendFromHost but the _XlaCompileMlir has not yet been created for device
// cluster that contains the outside compiled ops. This placeholder should be
// replaced by the TPU cluster _XlaCompileMlir in a subsequent pass.
mlir::TF::_XlaCompileMlirPlaceholderProgramKeyOp
CreateCompilationKeyPlaceholder(Location loc, OpBuilder& builder) {
  auto result_type =