// expected-remark@above {{Predecessors: {6}}}
      // expected-remark@above {{Successors: {8}}}
      tf_executor.yield %read3 : tensor<32xf32>
      // expected-remark@above {{ID: 8}}
      // expected-remark@above {{Predecessors: {4,5,7}}}
    }
    tf_executor.fetch %island#0 : tensor<32xf32>
    // expected-remark@above {{ID: 10}}
    // expected-remark@above {{Predecessors: {9}}}
  }

      send_node_in_host_graph = n;
      break;
    }
  }
  if (send_node_in_host_graph) {
    // This is an "top-level" outside compilation. Clear the graph, and copy
    // SendFromHost and all its predecessors from `host_graph`.
    std::vector<Node*> nodes;
    nodes.reserve(g->num_op_nodes());
    for (Node* n : g->op_nodes()) {
      nodes.push_back(n);
    }
    for (Node* n : nodes) {

  let summary = "Colocates each Split op with its predecessor";
  let constructor = "TFTPU::CreateTPUColocateSplitsPass()";
  let description = [{
    It is beneficial for performance to assign a `Split` op to the same device
    as its predecessor. This is because the weight of cut edges is always
    minimized when the `Split` is with its predecessor. This colocation

      ^bb0(%barg0: tensor<i32>, %barg1: tensor<!tf_type.variant<tensor<?x1xf32>>>): // no predeceessors
        %cond = "tf.Less"(%barg0, %size) : (tensor<i32>, tensor<i32>) -> tensor<i1>
        "tf.Yield"(%cond) : (tensor<i1>) -> ()
    }, {
      ^bb0(%barg0: tensor<i32>, %barg1: tensor<!tf_type.variant<tensor<?x1xf32>>>): // no predeceessors
      %index = "tf.AddV2"(%barg0, %one) : (tensor<i32>, tensor<i32>) -> tensor<i32>