return success();
}

namespace {
// Canonicalize While op so that results and operands match and external values
// are via implicit capture rather than via block args.
struct WhileResultOperandsMatchAndImplicitCapture
    : public OpRewritePattern<WhileOp> {
  using OpRewritePattern<WhileOp>::OpRewritePattern;

  LogicalResult matchAndRewrite(WhileOp while_op,

          while_op.getInput().getTypes(), while_op.getOutput().getTypes(),
          while_op.getBody().getArgumentTypes());
      return PropagateShapeToRegions(compatible_types,
                                     {&while_op.getCond(), &while_op.getBody()},
                                     max_iterations);
    }
    return PropagateShapeToRegions(while_op.getInput().getTypes(),

      new_while_operands.push_back(while_op.getOperand(op_idx));
      new_result_types.push_back(while_op.getResult(op_idx).getType());
    }

    // Create the new while operation.
    auto new_while_op = rewriter.create<WhileRegionOp>(
        while_op.getLoc(), new_result_types, new_while_operands,
        while_op->getAttrs());

    // Move region bodies to the new while.

    for (auto type : while_op->getOperandTypes()) {
      if (mlir::isa<TupleType>(type)) return failure();
    }

    // Creates a TF::WhileRegionOp to replace the mhlo::WhileOp. HLO WhileOp
    // currently doesn't support stateless and shape invariant, so these
    // parameters are set to the default values.
    auto new_while = rewriter.create<TF::WhileRegionOp>(

  // Append lifted args' types to While node's T attribute.
  TF_ASSIGN_OR_RETURN(
      std::vector<DataType> data_types,
      UpdateTypesAttribute(lifted_arg_nodes_and_outside_compilation_nodes, "T",
                           n));

  // Add edges from outside compilation nodes to While node.
  std::vector<Node*> outside_compilation_nodes;
  outside_compilation_nodes.reserve(

    hand create a barrier at the end of while loop body thus blocking any
    parallelism across iterations.

    For example, the following while loop body has a `%barrier` at the end.
    Although there is no data/control dependency between `tf.AssignVariableOp`
    for `%arg0` to `tf.AssignVariableOp` for `%arg1` across any iteration, the

      }
      if (auto whileOp = dyn_cast<mlir::TFL::WhileOp>(inst)) {
        if (inst->getNumOperands() != inst->getNumResults()) {
          inst->emitOpError(
              "number of operands and results don't match, only canonical "
              "TFL While supported");
          return std::nullopt;
        }
        return BuildWhileOperator(whileOp, operands, results);
      }

      }
      if (map.usesValueReferences()) {
        clearValueReferenceQueue();
      }
    }

    void clearKeyReferenceQueue() {
      while (keyReferenceQueue.poll() != null) {}
    }

    void clearValueReferenceQueue() {
      while (valueReferenceQueue.poll() != null) {}
    }

    // recency queue, shared by expiration and eviction

    /**

      }
      if (map.usesValueReferences()) {
        clearValueReferenceQueue();
      }
    }

    void clearKeyReferenceQueue() {
      while (keyReferenceQueue.poll() != null) {}
    }

    void clearValueReferenceQueue() {
      while (valueReferenceQueue.poll() != null) {}
    }

    // recency queue, shared by expiration and eviction

    /**

          <description>
            Indicates the dependency is optional for use of this library. While the
            version of the dependency will be taken into account for dependency calculation if the
            library is used elsewhere, it will not be passed on transitively. Note: While the type
            of this field is {@code String} for technical reasons, the semantic type is actually