#include "mlir/Dialect/Func/IR/FuncOps.h"  // from @llvm-project
#include "mlir/IR/BuiltinOps.h"  // from @llvm-project
#include "tensorflow/core/common_runtime/function_optimization_registry.h"
#include "tensorflow/core/common_runtime/optimization_registry.h"

namespace tensorflow {

// -------------------------------------------------------------------------- //

#include <string>

#include "absl/algorithm/container.h"
#include "tensorflow/compiler/jit/xla_cluster_util.h"
#include "xla/status_macros.h"
#include "tensorflow/core/framework/node_def.pb.h"
#include "tensorflow/core/framework/tensor.pb.h"
#include "tensorflow/core/util/dump_graph.h"

namespace tensorflow {

using tsl::StatusOr;

class CloneConstantsForBetterClusteringPassImpl {
 public:

// Permutation from the OIHW (== (output features, input features, height,
// width)) tensor format to HWIO. This is commonly used to transpose convolution
// weights represented as OIHW format to HWIO, which is more desirable for
// certain downstream optimization passes (e.g. XLA).
inline constexpr std::array<int64_t, 4> kOihwToHwioPermutation = {2, 3, 1, 0};

// Returns true if the value has static shape.

    raise ValueError(
        'Representative dataset is not a mapping (got: '
        f'{type(representative_dataset)}), but there is more than one '
        'signature key provided. Please provide a map of '
        '{signature_key -> dataset} with more than one signature key.'
    )
  else:
    representative_dataset_map = {
        list(signature_def_map.keys())[0]: representative_dataset,
    }

  auto posix_file = static_cast<PosixFile*>(file->plugin_file);
  char* dst = buffer;
  int64_t read = 0;

  while (n > 0) {
    // Some platforms, notably macs, throw `EINVAL` if `pread` is asked to read
    // more than fits in a 32-bit integer.
    size_t requested_read_length;
    if (n > INT32_MAX)
      requested_read_length = INT32_MAX;
    else
      requested_read_length = n;

        return failure();
      }

      int64_t channel_id = channel_id_++;
      // FIXME(b/226139061): Mode should be set to CrossReplicaAndPartition
      // in order to use XLA:GPU for more than one workers.
      // The mode is set to use CrossReplica to keep the
      // behavior on the primary user of this optimized path, because

      // trait, then we don't need to chain resource ops accessing this resource
      // between iterations: Every iteration will create a new independent
      // resource. This enables more parallelism across iterations.
      if (!side_effect_analysis.IsUniqueResourceAllocationId(
              resource_and_device.first)) {
        chain_resource_to_ops_map[resource_and_device].insert(&op);

auto *mlir_tf_quant_op_count = ::tensorflow::monitoring::Counter<1>::New(
    "/tensorflow/core/tf2xla/tf_quant_op_count" /*metric_name*/,
    "Counts the number of ops that has qint types" /*metric description*/,
    "op_name" /*metric label*/);

// Returns wether a type is illegal. Here we consider TF qint types illegal.
// See pass description in passes.td for more info about how illegal types are
// treated in this pass.
bool IsIllegalType(Type type) {

#include "absl/strings/strip.h"
#include "tensorflow/compiler/mlir/tensorflow/utils/dump_graph.h"
#include "xla/parse_flags_from_env.h"
#include "tensorflow/core/platform/macros.h"
#include "tensorflow/core/tpu/kernels/sparse_core_xla_flags_defaults.h"
#include "tensorflow/core/util/command_line_flags.h"

namespace tensorflow {
namespace {

BuildXlaOpsPassFlags* build_ops_flags;
MarkForCompilationPassFlags* mark_for_compilation_flags;

// recursively mutate the IR (resulting in quadratic behavior when moving
// operations). A rough sketch of the coarsening algorithm is shown below:
//
// // The algorithm iterates until a fixpoint is reached, i.e. when no more
// // islands can be merged.
// while (changed) {
//   // In the first phase we try to merge islands with their nearest consumer
//   // iff the consumer is another island.