// device. Device attribute is used to infer the appropriate sharding
        // within TPUs for this op.
        // TODO(b/183598857): Use explicit sharding ops from the front-end.
        // For example, dequeue ops generated by
        // tensorflow/python/tpu/tpu_feed.py
        if (!tensorflow::IsTPUReplicatedCore(attr.getValue()) &&
            !isa<tf_device::LaunchOp>(op)) {

    std::vector<tpu::ShardingAndIndex>* arg_core_mapping,
    std::vector<std::vector<xla::Shape>>* per_core_arg_shapes);

// Compiles a serialized MLIR module into XLA HLO, generates all accompanying
// metadata and stores them in CompilationResult.
absl::StatusOr<XlaCompilationResult> LegalizeWithMlirBridge(
    const tpu::MlirToHloArgs& computation,
    const tpu::TPUCompileMetadataProto& metadata, bool use_tuple_args,

# Enable MLIR-Based new TPU Bridge

**MLIR-Based new TPU Bridge is an experimental feature, tread lightly.**

## For TF 1.x-Based Models

In tf.ConfigProto.Experimental, there is a knob controlling whether the new TPU
Bridge is enabled or not. You can set it by using the following example code:

```
session_config = tf.ConfigProto(
  ......
  experimental=tf.ConfigProto.Experimental(
    enable_mlir_bridge=True,
  ),
  ......
)
```

       "tf.TPUReplicateMetadata"() {_xla_compile_device_type = "TPU", _replication_info = "cluster", device = "device", num_replicas = 1, topology = "topology"} : () -> ()
        %index = "tf.opA"(%arg0) {_xla_compile_device_type = "TPU", _replication_info = "cluster"} : (tensor<i1>) -> tensor<i32>
        %input = "tf.opB"(%arg0) {_xla_compile_device_type = "TPU", _replication_info = "cluster"} : (tensor<i1>) -> tensor<i32>

  tf_executor.graph {
    %1 = tf_executor.island {
     "tf.SomeTpuOp"() {_xla_compile_device_type = "TPU"} : () -> ()
      "tf.SomeOtherTpuOp"() {_xla_compile_device_type = "TPU"} : () -> ()
      tf_executor.yield
    }
    %2 = tf_executor.island {
      "tf.TPUReplicateMetadata"() {_xla_compile_device_type = "TPU", _replication_info = "replicate", device = "device", num_replicas = 1, topology = "topology"} : () -> ()

#include "tensorflow/core/platform/status.h"
#include "tensorflow/core/protobuf/tpu/compile_metadata.pb.h"
#include "tensorflow/core/tpu/kernels/tpu_compile_op_support.h"
#include "tsl/platform/statusor.h"

namespace tensorflow {
namespace tf2xla {
namespace internal {

using metrics::IncrementTfMlirBridgeSecondPhaseCounter;
using metrics::MlirBridgeSecondPhaseMetric;
using tpu::MlirToHloArgs;

absl::StatusOr<XlaCompilationResult> LegalizeTfToHlo(

    %control_3 = tf_executor.island(%control_2) wraps "tf.NoOp"() {_replication_info = "cluster", _xla_compile_device_type = "TPU", device = ""} : () -> ()

  %1 = "tf.ReadVariableOp"(%arg1) : (!tf_res_f32) -> tensor<f32>
  %2 = "tf.ReadVariableOp"(%arg2) : (!tf_res_f32) -> tensor<f32>
  %device_output:2 = "tf_device.cluster_func"(%0, %1, %2) {_xla_compile_device_type = "TPU", _replication_info = "tpu", func = @device_func_0} : (tensor<i32>, tensor<f32>, tensor<f32>) -> (tensor<f32>, tensor<i32>)
  "tf.AssignVariableOp"(%arg1, %device_output#0) : (!tf_res_f32, tensor<f32>) -> ()

absl::Status CompileTensorflowGraphToHlo(
    const std::variant<tpu::MlirToHloArgs, tpu::FunctionToHloArgs>& computation,
    const tpu::TPUCompileMetadataProto& metadata, bool use_tuple_args,
    XlaShapeLayoutHelpers::ShapeDeterminationFns shape_determination_funcs,
    const std::vector<tensorflow::TensorShape>& arg_shapes,
    std::vector<tpu::ShardingAndIndex>* arg_core_mapping,

#include "tensorflow/core/tpu/kernels/tpu_compile_op_support.h"
#include "tsl/platform/statusor.h"

namespace tensorflow {
namespace tf2xla {
namespace internal {

// Legalize the given MLIR module to XLA HLO using a combination of the MLIR
// Bridge and XlaBuilder
absl::StatusOr<XlaCompilationResult> LegalizeTfToHlo(
    const tpu::MlirToHloArgs& computation,