ALTREP and C++

10 min read

R C++ altrep

Traditionally, R vectors (numeric, …) are formed of one header ditrectly followed by the actual contiguous data. I’ll spare the details about the headers because even though I have a fair understanding, I don’t necessarily want to propagate my misconceptions 😬, and also because it is mostly irrelevant to this post.

Since R 3.5.0, the implementation of vectors benefits from ALTREP (alternative representation), which challenges this and decouples the header with the actual data. This has two main use cases:

  • when the memory comes from somewhere else (a memory mapped file, an arrow array or whatever), i.e. some contiguous memory that comes from somewhere. ALTREP lets you create an R vector that points to that memory.

  • when you only need part of the vector, e.g. a single value. Here ALTREP lets you define what is the ith value of the vector, without having to materialize it. The canonical example is for(i in 1:n){}. You never need the entire vector 1:n, you only need one value at a time.

As the projects I’m currently involved with (dplyr and arrow) might ultimately benefit from ALTREP, I’ve 🤹 with it in the altrepisode 📦.

In this post, I’m covering the first use case with numeric vectors, in particular creating an R altrep vector that borrows data from a C++ std::vector<double>.

The Altrep.h header

To manipulate ALTREP objects, create your own classes, you need to include the R_ext/Altrep.h header file. Unfortunately in R 3.5.0, the Altrep.h is not C++ friendly, so you need to be careful when you include it. The situation has been fixed recently so if you rely on R-devel things are easier.

In the meantime, here is my workaround:

// to manipulate R objects, aka SEXP
#include <R.h>
#include <Rinternals.h>
#include <Rversion.h>

// because we need to initialize the altrep class
#include <R_ext/Rdynload.h>

#if R_VERSION < R_Version(3, 6, 0)

// workaround because R's <R_ext/Altrep.h> not so conveniently uses `class`
// as a variable name, and C++ is not happy about that
// SEXP R_new_altrep(R_altrep_class_t class, SEXP data1, SEXP data2);
#define class klass

// Because functions declared in <R_ext/Altrep.h> have C linkage
extern "C" {
  #include <R_ext/Altrep.h>

// undo the workaround
#undef class

  #include <R_ext/Altrep.h>

The <200 lines of (fairly documented) C++ code stdvec_doubles.cpp contains the implementation that is discussed here. This uses Rcpp but only for code generation using the attributes feature. This is not using any Rcpp classes (e.g. NumericVector), which at this point would eliminate the benefits of ALTREP.


Let’s start by the end of the file to get some motivation. The doubles function below creates an R altrep object backed by a std::vector<double>.

//' an altrep object that wraps a std::vector<double>
//' @export
// [[Rcpp::export]]
SEXP doubles() {
  // create a new std::vector<double>
  // this uses `new` because we want the vector to survive
  // it is deleted when the altrep object is garbage collected
  auto v = new std::vector<double> {-2.0, -1.0, 0.0, 1.0, 2.0};

  // The altrep object owns the std::vector<double>
  return stdvec_double::Make(v, true);

The details will follow, but for now, let’s look at what the object look like when back on the R side:

x <- doubles()
## [1] -2 -1  0  1  2
## [1] "numeric"
##  num [1:5] -2 -1 0 1 2

It looks and feels like any other R numeric vector, that’s the point. As far as the R code is concerned, this is not different than an object that would have been created by c:

y <- c(-2.0, -1.0, 0.0, 1.0, 2.0)
identical(x, y)
## [1] TRUE

To see a difference, you have to look at the object with .Internal(inspect()), as ALTREP gives you a way to control how your ALTREP objects are inspected.

## @7fca66a1f990 14 REALSXP g0c0 [NAM(3)] std::vector<double> (len=5, ptr=0x7fca5f46da60)
## @7fca654f7dc8 14 REALSXP g0c4 [NAM(3)] (len=5, tl=0) -2,-1,0,1,2

Register the ALTREP class

The stdvec_double::Make function from the above code chunk creates an R object of ALTREP class (the ALTREP class is completely orthogonal to the R class, as again as far as R code is concerned, nothing has changed).

For this we need to register an R_altrep_class_t object with (in the case of an altrep numeric vector) the R_make_altreal_class function.

This is done at 📦 initialisation time, thanks to the new Rcpp::init attribute.

// static initialization of stdvec_double::class_t
R_altrep_class_t stdvec_double::class_t;

// Called the package is loaded (needs Rcpp
// [[Rcpp::init]]
void init_stdvec_double(DllInfo* dll){

ALTREP is a C api, relying on C functions, but because this is C++, I’ve squashed the functions together as static functions of the stdvec_double C++ class, hence the stdvec_double::Init call here. Init looks like this:

static void Init(DllInfo* dll){
  class_t = R_make_altreal_class("stdvec_double", "altrepisode", dll);

  // altrep
  R_set_altrep_Length_method(class_t, Length);
  R_set_altrep_Inspect_method(class_t, Inspect);

  // altvec
  R_set_altvec_Dataptr_method(class_t, Dataptr);
  R_set_altvec_Dataptr_or_null_method(class_t, Dataptr_or_null);

  // altreal
  R_set_altreal_Elt_method(class_t, real_Elt);
  R_set_altreal_Get_region_method(class_t, Get_region);

First, we register the class with R_make_altreal_class function, then we replace default methods with custom functions that know how to deal with out std::vector<double>.

  • Length : what is the length of the vector
  • Inspect : what happens when we .Internal(inspect()) the object
  • Dataptr : where is the data (more on that later) ?
  • Dataptr_or_null : where is the data (but don’t look too hard)
  • real_Elt : what is the ith element ?
  • Get_region : A contiguous region of the data

This is I believe 🤷 the bare minimum.

In addition to that, the stdvec_double class hosts:

  • Make : to construct one such objet from a pointer to a std::vector.
  • Finalize: to delete the object as the proper time if we own it
  • Ptr : to get the pointer
  • Get : to get a reference to the std::vector<double>


The stdvec_double::Make function creates the altrep R object backed by the `std::vector:

// Make an altrep object of class `stdvec_double::class_t`
static SEXP Make(std::vector<double>* data, bool owner){
  // The std::vector<double> pointer is wrapped into an R external pointer
  // `xp` needs protection because R_new_altrep allocates
  SEXP xp = PROTECT(R_MakeExternalPtr(data, R_NilValue, R_NilValue));

  // If we own the std::vector<double>*, we need to delete it
  // when the R object is being garbage collected
  if (owner) {
    R_RegisterCFinalizerEx(xp, stdvec_double::Finalize, TRUE);

  // make a new altrep object of class `stdvec_double::class_t`
  SEXP res = R_new_altrep(class_t, xp, R_NilValue);

  // xp no longer needs protection, as it has been adopted by `res`
  return res;

Eventually the R object (aka SEXP) is created with the R_new_altrep function, which takes the altrep class as first argument, and two other arbitrary R objects. These two R objects can be later accessed with R_altrep_data1 and R_altrep_data2 and can be just about anything you like.

Here we use an external pointer, created with R_MakeExternalPtr, as data1 and we don’t need anything for data2 so we use NULL. If we own the C++ vector, as indicated by the owner argument, we register a finalizer so that when the external pointer (the R object) is garbage collected, the destructor of the C++ object is invoked.

The Finalize, Ptr and Get functions are conveniences that allow us to go from the altrep R object to the C++ vector:

// finalizer for the external pointer
static void Finalize(SEXP xp){
  delete static_cast<std::vector<double>*>(R_ExternalPtrAddr(xp));

// get the std::vector<double>* from the altrep object `x`
static std::vector<double>* Ptr(SEXP x) {
  return static_cast<std::vector<double>*>(R_ExternalPtrAddr(R_altrep_data1(x)));

// same, but as a reference, for convenience
static std::vector<double>& Get(SEXP vec) {
  return *Ptr(vec) ;

Given the altrep object x we need to first get to its data1 with R_altrep_data1 and then cast that to the underlying C++ vector with R_ExternalPtrAddr and a static_cast<>. Once we have this, the rest follows naturally.

ALTREP methods

ALTREP is divided in several layers depending on the type of object we altrep (maybe it’s to soon for this to be a verb). The first layer is generic and apply to all ALTREP objects. There might be other methods, but for this I’ve implemented the Length and Inspect :

// The length of the object
static R_xlen_t Length(SEXP vec){
  return Get(vec).size();

// What gets printed when .Internal(inspect()) is used
static Rboolean Inspect(SEXP x, int pre, int deep, int pvec, void (*inspect_subtree)(SEXP, int, int, int)){
  Rprintf("std::vector<double> (len=%d, ptr=%p)\n", Length(x), Ptr(x));
  return TRUE;

In the Length method, we are given the altreped (still assuming this is a verb) object. From this object, we extract the std::vector<double>& with Get and then simply call size() on it.

The Inspect is a bit more involved, let’s just skip it 🙈.

ALTVEC methods

Then, we have methods that are only relevant to vector type R objects. In this implementation, I have defined the Dataptr and Dataptr_or_null methods:

// The start of the data, i.e. the underlying double* array from the std::vector<double>
// This is guaranteed to never allocate (in the R sense)
static const void* Dataptr_or_null(SEXP vec){
  return Get(vec).data();

// same in this case, writeable is ignored
static void* Dataptr(SEXP vec, Rboolean writeable){
  return Get(vec).data();

In this example, they are identical, but it’s not necessarily the case for all altrep class implementations.

The difference is that the Dataptr_or_null method is guaranteed to not allocate additional R memory. If you already have access to a contiguous chunk of memory, then return that, otherwise return a null pointer, but this should never allocate.

The Dataptr method is the big 🔨. Whatever your class has to do, it must return a pointer to contiguous chunk of memory where the data is. We’ll illustrate this better in another post, but in short if you already have that contiguous memory, then return it, if not do whatever it takes, allocate if you have to, but eventually get me that memory.

Dataptr is what most ALTREP unaware (e.g. the constructor of Rcpp::NumericVector or the mean function from base R) code will use.

ALTREAL methods

Eventually, methods specific to numeric vectors.

// the element at the index `i`
// this does not do bounds checking because that's expensive, so
// the caller must take care of that
static double real_Elt(SEXP vec, R_xlen_t i){
  return Get(vec)[i];

// Get a pointer to the region of the data starting at index `i`
// of at most `size` elements.
// The return values is the number of elements the region truly is so the caller
// must not go beyond
static R_xlen_t Get_region(SEXP vec, R_xlen_t start, R_xlen_t size, double* out){
  out = Get(vec).data() + start;
  R_xlen_t len = Get(vec).size() - start;
  return len > size ? len : size;

real_Elt gives you the element at offset i of the underlying object, here this calls the std::vector<double>::operator[] but you can imagine situations where this is implemented differently, e.g. for 1:n you can imagine this would just return i+1 without having to rely on anything else.

Get_region again is a bit more involved, so I’ll skip it too. You can do what I do and guess what it is supposed to do based on its name. There can be situations where you can have access to contiguous memory for part of the vector.

What else to read

I’ll write some more about this in another post. In the meantime, here are a few pointers:

  • Luke Tierney’s presentation about ALTREP.
  • The ALTREP Examples organization on github hosts packages that showcase ALTREP. Most of what I did on altrepisode is inspired from the simplemmap 📦. The mutable 📦 is newer and easier to grasp.
  • The altvecR 📦 from Simon Urbanek, which is a toy package that re-routes ALTREP/ALTVEC methods to R functions for experimentation.
  • The actual ALTREP code, mainly lives in the altrep.c and Altrep.h files.
  • My altrepisode. This contains the stdvec_double class described here and another class that I’ll discuss in a follow up post.
Support my work on patreon Blogging is one of the activities I have the freedom to do because of community sponsorship. If you like the content, would like to see more, or just generally like my work, please consider pledging.