mlir::CreateExecutorDialectToFunctionalConversionPass());
  // Guarantee all functions have one use, which enables more exact shape
  // inference.
  pm.addPass(mlir::TF::CreateGuaranteeAllFuncsOneUsePass());
  // Run shape inference so that tf_executor/tf_device ops created later will
  // likely to inherit more concrete types.
  pm.addPass(mlir::TF::CreateTFShapeInferencePass());

#include "tensorflow/compiler/mlir/tensorflow/utils/error_util.h"
#include "tensorflow/core/framework/function.h"
#include "tensorflow/core/framework/node_def_util.h"
#include "tensorflow/core/framework/tensor_shape.h"
#include "tensorflow/core/framework/types.pb.h"
#include "tensorflow/core/lib/llvm_rtti/llvm_rtti.h"
#include "tensorflow/core/platform/errors.h"
#include "tensorflow/core/platform/refcount.h"
#include "tsl/platform/errors.h"

  let description = [{
    This pass finds on-device resource variable reads and updates surrounding a
    `tf.TPUExecute` op and merges them into a `tf.TPUExecuteAndUpdateVariables`
    op. This allows the TPU execution to perform more efficient in-place
    variable updates.

    For example,

    ```mlir
      %0 = "tf.ReadVariableOp"(%arg0)
      %1 = "tf.ReadVariableOp"(%arg1)
      %2 = "tf.TPUExecute"(%0, %1, %compile)

// prior to importing the graph. The primary goal is for support of custom ops.
// This is not intended to be a general solution for custom ops for the future
// but mainly for supporting older models like mobilenet_ssd. More appropriate
// mechanisms, such as op hints or using functions to represent composable ops
// like https://github.com/tensorflow/community/pull/113 should be encouraged
// going forward.
// NOLINTNEXTLINE

  bool multiple_cpu_devices = false;
  bool multiple_gpu_devices = false;
  bool multiple_unknown_devices = false;

  // Returns 'true' if d0 and d1 are conflicting devices. If they are
  // compatible, update d1 with a more specific one.
  // TODO(sanjoy): Cache DeviceNameUtils::ParsedName inside device_info_cache.
  const auto is_multiple_devices =
      [&](const jit::DeviceId& d0, std::optional<jit::DeviceId>* d1) -> bool {

#include <memory>

#include "tensorflow/c/c_api.h"
#include "tensorflow/c/c_test_util.h"
#include "tensorflow/core/platform/logging.h"
#include "tensorflow/core/platform/strcat.h"
#include "tensorflow/core/platform/test.h"

using tensorflow::GraphDef;

namespace {

class CApiWhileLoopTest : public ::testing::Test {
 protected:

# LLVM/Clang: https://apt.llvm.org/
apt-key adv --fetch-keys https://apt.llvm.org/llvm-snapshot.gpg.key

# Set up custom sources
cat >/etc/apt/sources.list.d/custom.list <<SOURCES
# More Python versions: Deadsnakes
deb http://ppa.launchpad.net/deadsnakes/ppa/ubuntu focal main
deb-src http://ppa.launchpad.net/deadsnakes/ppa/ubuntu focal main

# LLVM/Clang 17 repository

# proto-file: devtools/build_cleaner/proto/actions.proto
# proto-message: ActionSpecs

# Python rules should not have more than one source file.
action_spec {
  action: CHECK_FILE_COUNT
  file_count_params {
    rule_selector {
      rule_kind_regex: "^.*py(type)?(_strict)?_(binary|library|test).*$"
      generator_function_regex: "^(?!boq_header)$"
    }
    max_source_count: 1
  }

// nested module which exhibit a more friendly "V2-like" structure.
// This is only intended for V1 compatibility mode where the bridge runs without
// feed/fetches on session create/extend.
struct TPUBridgeExecutorIslandOutlining
    : public impl::TPUBridgeExecutorIslandOutliningPassBase<
          TPUBridgeExecutorIslandOutlining> {
  void runOnOperation() override;
};

#include "tensorflow/core/framework/function.h"
#include "tensorflow/core/framework/op_kernel.h"
#include "tensorflow/core/framework/resource_mgr.h"
#include "tensorflow/core/framework/types.h"
#include "tensorflow/core/platform/errors.h"
#include "tensorflow/core/platform/status.h"
#include "tensorflow/core/tfrt/common/create_pjrt_client_util.h"
#include "tensorflow/core/tfrt/common/global_state.h"