llvm::MapVector<Operation*, llvm::SmallVector<unsigned int, 4>>*
        arguments_to_erase) {
  auto arg = region->getArgument(argument_index);
  for (auto& usage : arg.getUses()) {
    auto* user_op = usage.getOwner();
    int operand_index = usage.getOperandNumber();
    PropagateUsage(user_op, operand_index, value, work_list,
                   arguments_to_erase);
  }
}

        workerClient.stop();
    }

    public void kill() {
        workerClient.stopNow();
    }

    DaemonForkOptions getForkOptions() {
        return forkOptions;
    }

    public int getUses() {
        return uses;
    }

    public KeepAliveMode getKeepAliveMode() {
        return forkOptions.getKeepAliveMode();
    }

    public LogLevel getLogLevel() {
        return logLevel;
    }

    const ShapedType type = mlir::dyn_cast<ShapedType>(op.getType());
    if (!type || !type.getElementType().isF32()) return quantizable_ops;

    const Value value = op.getResult();

    for (OpOperand& use : value.getUses()) {
      Operation* user = use.getOwner();
      const int operand_num = use.getOperandNumber();
      quantizable_ops.insert({user, operand_num});
    }
    return quantizable_ops;
  }

    if (!function_name.starts_with(kDequantizeFunctionName)) return failure();

    llvm::SmallVector<Operation*> users(op->getUsers().begin(),
                                        op->getUsers().end());

    bool changed = false;
    for (auto& use : op->getUses()) {
      Operation* user_op = use.getOwner();
      int user_idx = use.getOperandNumber();
      if (!IsOpWithInt8TypeOperand(user_op)) continue;

  for (auto& use : uses_optional.value()) {
    if (llvm::isa<TF::WhileOp>(use.getUser())) {
      if (*while_op != nullptr) {
        use.getUser()->emitOpError() << "multiple users of function.";
        return LogicalResult::failure();
      } else {
        *while_op = llvm::cast<TF::WhileOp>(use.getUser());
      }
    } else {
      use.getUser()->emitOpError() << "non while use of function.";

    Value res = read.getResource();
    Operation* scope = res.getParentBlock()->getParentOp();
    if (!scope->isProperAncestor(replicate)) continue;
    bool has_conflicting_write = false;
    for (OpOperand& use : res.getUses()) {
      Operation* using_op = use.getOwner();
      if (using_op == read) continue;
      if (!replicate->isProperAncestor(using_op)) continue;
      Operation* peer = GetAncestorBelow(using_op, replicate);

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

namespace {
constexpr char kDeviceAttr[] = "device";

// A type that abstracts over types that have uses accessible via `getUses`.
using Source = PointerUnion<Operation *, BlockArgument *>;

// We use union-find algorithm to build clusters of connected operations based
// on the user provided policy. If an operation can be clustered (one of the

// multiple such activations, one is returned (with no guarantee as to which
// one). If there are no activation functions that use the output, returns
// nullptr.
Operation *GetActivation(Value op) {
  for (auto &use : op.getUses()) {
    if (IsActivationFunction(use.getOwner())) return use.getOwner();
  }
  return nullptr;
}

// Finds and returns a BiasAdd that uses the result of `op` as the `value`

    ShapedType new_result_type = old_result_type.clone(quantized_type);

    // Insert CastOp if it does not exist yet. Otherwise, just rewire without
    // creating a CastOp.
    for (auto& connected_op : op.getResult().getUses()) {
      auto cast_op = llvm::dyn_cast_or_null<CastOp>(connected_op.getOwner());
      if (cast_op && cast_op.getType() == new_result_type) {
        quantize_op->setOperand(quantize_operand_num, cast_op);
        return;

           mlir::isa<IntegerType>(key_dtype) &&
           mlir::cast<IntegerType>(key_dtype).getWidth() == 64))) {
      return false;
    }

    for (auto& use : hashtable->getUses()) {
      Operation* user = use.getOwner();

      // Allow consuming hash table ops that can be covered by TensorFlow Lite
      // hash table kernels.