Interoperability between Rcpp and std::array

Rcpp is a mature and very widely used package providing seamless interoperability between C++ and the R language, which has a native C API. This package facilitates conversion between R data structures and the templated std::array class introduced in C++11, by using the interface provided in Rcpp. It originated with this Stack Overflow question.

std::array is a templated container type with a fixed number of elements, an object-orientated analogue of a C-style array type like int[3]. Client packages can interface this type with R if they add RcppArray to LinkingTo and include the header. (A simple example package is provided.)

// No need to include "Rcpp.h" as well
#include "RcppArray.h"

// [[Rcpp::export]]
RObject test() {
  Rcpp::NumericVector vec = Rcpp::NumericVector::create(1,2,3);
  std::array<double,3> arr = Rcpp::as<std::array<double,3>>(vec);
  // Do something with the array
  std::array<unsigned int,3> result = { 1, 2, 3 };
  return result;    // Implicitly Rcpp::wrap(result)

In either direction the element type of the array can be any atomic type that Rcpp knows how to convert, and isn’t limited to the types that R uses internally. In the example above, unsigned int is used even though it doesn’t directly correspond to an R vector mode: the wrap() function will convert it to an R numeric vector, i.e., double.


In addition to array, there is experimental support for std::span, introduced in C++20. This is a typed container of fixed or dynamic length which provides a “view” onto a contiguous block of data owned by another object. Once again, there is a simple example client package showing usage of this facility.

In this case there are some notable caveats. Client packages must obviously request C++20 support, and a configure script may well be needed to check for span availability. Moreover, since a span does not own the memory it points to, as<span<T,D>>() will only compile where the requested type T matches a type that R uses internally (viz. int, double, etc.). The latter limitation does not apply to wrap, however, because a new vector is created and the data converted.