Crate extendr_api

source ·
Expand description

A safe and user friendly R extension interface.

  • Build rust extensions to R.
  • Convert R packages to Rust crates.

This library aims to provide an interface that will be familiar to first-time users of Rust or indeed any compiled language.

See Robj for much of the content of this crate. Robj provides a safe wrapper for the R object type.


Use attributes and macros to export to R.

use extendr_api::prelude::*;
// Export a function or impl to R.
fn fred(a: i32) -> i32 {
    a + 1

// define exports using extendr_module
extendr_module! {
   mod mymodule;
   fn fred;    

In R:

result <- fred(1)

Robj is a wrapper for R objects. The r!() and R!() macros let you build R objects using Rust and R syntax respectively.

use extendr_api::prelude::*;
test! {
    // An R object with a single string "hello"
    let character = r!("hello");
    let character = r!(["hello", "goodbye"]);
    // An R integer object with a single number 1L.
    // Note that in Rust, 1 is an integer and 1.0 is a real.
    let integer = r!(1);
    // An R real object with a single number 1.
    // Note that in R, 1 is a real and 1L is an integer.
    let real = r!(1.0);
    // An R real vector.
    let real_vector = r!([1.0, 2.0]);
    let real_vector = &[1.0, 2.0].iter().collect_robj();
    let real_vector = r!(vec![1.0, 2.0]);
    // An R function object.
    let function = R!("function(x, y) { x + y }")?;
    // A named list using the list! macro.
    let list = list!(a = 1, b = 2);
    // An unnamed list (of R objects) using the List wrapper.
    let list = r!(List::from_values(vec![1, 2, 3]));
    let list = r!(List::from_values(vec!["a", "b", "c"]));
    let list = r!(List::from_values(&[r!("a"), r!(1), r!(2.0)]));

    // A symbol
    let sym = sym!(wombat);

    // A R vector using collect_robj()
    let vector = (0..3).map(|x| x * 10).collect_robj();

In Rust, we prefer to use iterators rather than loops.

use extendr_api::prelude::*;
test! {
    // 1 ..= 100 is the same as 1:100
    let res = r!(1 ..= 100);
    assert_eq!(res, R!("1:100")?);
    // Rust arrays are zero-indexed so it is more common to use 0 .. 100.
    let res = r!(0 .. 100);
    assert_eq!(res.len(), 100);
    // Using map is a super fast way to generate vectors.
    let iter = (0..3).map(|i| format!("fred{}", i));
    let character = iter.collect_robj();
    assert_eq!(character, r!(["fred0", "fred1", "fred2"]));

To index a vector, first convert it to a slice and then remember to use 0-based indexing. In Rust, going out of bounds will cause and error (a panic) unlike C++ which may crash.

use extendr_api::prelude::*;
test! {
    let vals = r!([1.0, 2.0]);
    let slice = vals.as_real_slice().ok_or("expected slice")?;
    let one = slice[0];
    let two = slice[1];
    // let error = slice[2];
    assert_eq!(one, 1.0);
    assert_eq!(two, 2.0);

Much slower, but more general are these methods:

use extendr_api::prelude::*;
test! {
    let vals = r!([1.0, 2.0, 3.0]);
    // one-based indexing [[i]], returns an object.
    assert_eq!(vals.index(1)?, r!(1.0));
    // one-based slicing [x], returns an object.
    assert_eq!(vals.slice(1..=2)?, r!([1.0, 2.0]));
    // $ operator, returns an object
    let list = list!(a = 1.0, b = "xyz");
    assert_eq!(list.dollar("a")?, r!(1.0));

The R! macro lets you embed R code in Rust and takes Rust expressions in {{ }} pairs.

The Rraw! macro will not expand the {{ }} pairs.

use extendr_api::prelude::*;
test! {
    // The text "1 + 1" is parsed as R source code.
    // The result is 1.0 + 1.0 in Rust.
    assert_eq!(R!("1 + 1")?, r!(2.0));

    let a = 1.0;
    assert_eq!(R!("1 + {{a}}")?, r!(2.0));

        x <- {{ a }}
        x + 1
    ")?, r!(2.0));

        x <- "hello"
    "#)?, r!("hello"));

    // Use the R meaning of {{ }} and do not expand.
        x <- {{ 1 }}
        x + 1
    ")?, r!(2.0));

The r! macro converts a rust object to an R object and takes parameters.

use extendr_api::prelude::*;
test! {
    // The text "1.0+1.0" is parsed as Rust source code.
    let one = 1.0;
    assert_eq!(r!(one+1.0), r!(2.0));

You can call R functions and primitives using the call! macro.

use extendr_api::prelude::*;
test! {

    // As one R! macro call
    let confint1 = R!("confint(lm(weight ~ group - 1, PlantGrowth))")?;
    // As many parameterized calls.
    let mut formula = lang!("~", sym!(weight), lang!("-", sym!(group), 1.0));
    let plant_growth = global!(PlantGrowth)?;
    let model = call!("lm", formula, plant_growth)?;
    let confint2 = call!("confint", model)?;
    assert_eq!(confint1.as_real_vector(), confint2.as_real_vector());

Rust has a concept of “Owned” and “Borrowed” objects.

Owned objects, such as Vec and String allocate memory which is released when the object lifetime ends.

Borrowed objects such as &i32 and &str are just pointers to annother object’s memory and can’t live longer than the object they reference.

Borrowed objects are much faster than owned objects and use less memory but are used only for temporary access.

When we take a slice of an R vector, for example, we need the original R object to be alive or the data will be corrupted.

use extendr_api::prelude::*;
test! {
    // robj is an "Owned" object that controls the memory allocated.
    let robj = r!([1, 2, 3]);
    // Here slice is a "borrowed" reference to the bytes in robj.
    // and cannot live longer than robj.
    let slice = robj.as_integer_slice().ok_or("expected slice")?;
    assert_eq!(slice.len(), 3);

§Writing tests

To test the functions exposed to R, wrap your code in the test! macro. This macro starts up the necessary R machinery for tests to work.

use extendr_api::prelude::*;

fn things() ->  Strings {
    Strings::from_values(vec!["Test", "this"])

// define exports using extendr_module
extendr_module! {
   mod mymodule;
   fn things;    

mod test {
    use super::*;
    use extendr_api::prelude::*;

    fn test_simple_function() {
        assert_eq!(things().elt(0), "Test")

§Returning Result<T,E> to R

Currently, throw_r_error() leaks memory because it jumps to R without properly dropping some rust objects.

The memory-safe way to do error handling with extendr is to return a Result<T, E> to R. By default, any Err will trigger a panic! on the rust side which unwinds the stack. The rust error trace will be printed to stderr, not R terminal. Any Ok value is returned as is.

Alternatively, two experimental non-leaking features, result_list and result_condition, can be toggled to avoid panics on Err. Instead, an Err x is returned as either

  • list: list(ok=NULL, err=x) when result_list is enabled,
  • error condition: <error: extendr_error>, with x placed in condition$value, when resultd_condition is enabled.

It is currently solely up to the user to handle any result on R side.

The minimal overhead of calling an extendr function is in the ballpark of 2-4us. Returning a condition or list increases the overhead to 4-8us. Checking & handling the result on R side will likely increase overall overhead to 8-16us, depending on how efficiently the result is handled.

use extendr_api::prelude::*;
// simple function always returning an Err string
fn oups(a: i32) -> std::result::Result<i32, String> {
    Err("I did it again".to_string())

// define exports using extendr_module
extendr_module! {
   mod mymodule;
   fn oups;    

In R:

> ... long panic traceback from rust printed to stderr

lst <- oups(1)
> list(ok = NULL, err = "I did it again")

cnd <- oups(1)
> <error: extendr_error>
> "I did it again"

oups_handled <- function(a) {
  val_or_err <- oups(1)  
  if (inherits(val_or_err, "extendr_error")) stop(val_or_err)

§Feature gates

extendr-api has some optional features behind these feature gates:

  • ndarray: provides the conversion between R’s matrices and ndarray.
  • num-complex: provides the conversion between R’s complex numbers and num-complex.
  • serde: provides the Serde support.
  • graphics: provides the functionality to control or implement graphics devices.

extendr-api supports three ways of returning a Result<T,E> to R. Only one behavior feature can be enabled at a time.

  • result_panic: Default behavior, return Ok as is, panic! on any Err

Default behavior can be overridden by specifying extend_api features, i.e. extendr-api = {..., default-features = false, features= ["result_condition"]} These features are experimental and are subject to change.

  • result_list: return Ok as list(ok=?, err=NULL) or Err list(ok=NULL, err=?)
  • result_condition: return Ok as is or Err as $value in an R error condition.




  • Execute R code by parsing and evaluating tokens.
  • Execute R code by parsing and evaluating tokens but without expanding parameters.
  • Call a function or primitive defined by a text expression with arbitrary parameters. This currently works by parsing and evaluating the string in R, but will probably acquire some shortcuts for simple expressions, for example by caching symbols and constant values.
  • Create a dataframe.
  • Define a module and export symbols to R Example:
  • Create a factor.
  • Get a global variable.
  • A macro for constructing R langage objects.
  • Create a List R object from a list of name-value pairs.
  • Create a Pairlist R object from a list of name-value pairs.
  • Convert a rust expression to an R object.
  • Print via the R error stream.
  • Print with a newline via the R output stream.
  • Print via the R output stream.
  • Print with a newline via the R output stream.
  • The sym! macro install symbols. You should cache your symbols in variables as generating them is costly.
  • Macro for running tests.
  • Get a local variable from the calling function or a global variable if no such variable exists.


  • Enum use to unpack R objects into their specialist wrappers.
  • Type of R objects used by Robj::rtype.



  • Used for immutable dereferencing operations, like *v.
  • Used for mutable dereferencing operations, like in *v = 1;.
  • Simple and safe type conversions that may fail in a controlled way under some circumstances. It is the reciprocal of TryInto.
  • An attempted conversion that consumes self, which may or may not be expensive.


Attribute Macros§

  • The #[extendr]-macro may be placed on three items

Derive Macros§

  • Enable the construction of dataframes from arrays of structures.
  • Derives an implementation of From<Struct> for Robj and From<&Struct> for Robj on this struct.
  • Derives an implementation of TryFrom<Robj> for Struct and TryFrom<&Robj> for Struct on this struct.