Many years ago, Peter Norvig wrote a beautiful article about creating a lisp interpreter in Python. It’s the most fun tutorial I’ve seen, not just because it teaches you about my favorite language family (Lisp), but because it cuts through to the essence of interpreters, is fun to follow and quick to finish.
Recently, I had some time and wanted to learn Rust. It’s a beautiful systems language, and I’ve seen some great work come out from those who adopt it. I thought, what better way to learn Rust, than to create a lisp interpreter in it?
Hence, Risp — a lisp in rust — was born. In this essay you and I will follow along with Norvig’s Lispy, but instead of Python, we’ll do it in Rust 🙂.
If you haven’t heard of lisp, some Paul Graham’s essays (one, two, three), alongside some Rich Hickey talks will get you fired up. In short, everything is a list, everything is an expression, and that makes for a very powerful language.
Our structure will be similar to Norvig’s tutorial, though I depart slightly in two ways:
Finally, this is the first program I wrote in Rust. I may have misused some things, so if you’re a Rust hacker, I’d love to hear your feedback 🙂.
With the notes out of the way, let’s get into it.
As Norvig suggests, our first goal is to create a subset of lisp, that can do what a basic calculator can do.
To make it as simple as possible to follow, for language 1, we’ll only support addition and subtraction. No variable definitions, no if statements, nada.
This departs a bit from Lispy, but I found this stopping point a lot more convenient when writing it in Rust. So, our goal:
(+ 10 5 2)//=> 17
(- 10 5 2) //=> 3The important process we need to remember is the flow of an interpreter:
our program ⟶ parse ⟶ abstract syntax tree ⟶ eval ⟶ result
We will need to parse our program and convert it into an abstract syntax tree. After that, we can eval the abstract syntax tree and get our result. (Refer to Norvig’s article for more detailed definitions and explanations).
Risp can have three kinds of values for now:
#[derive(Clone)]
enum RispExp {
Symbol(String),
Number(f64),
List(Vec<RispExp>),
} We’ll also need an error type. We’ll keep this simple, but if you’re curious there is a more robust approach.
#[derive(Debug)]
enum RispErr {
Reason(String),
}Finally, we’ll need an environment type. This is where we will store defined variables, built-in functions, and so forth:
#[derive(Clone)]
struct RispEnv {
data: HashMap<String, RispExp>,
}Our goal is to take our program, and build an abstract syntax tree from it. For us, that is going to be a RispExp. To do this, first we will take our program, and cut it up into a bunch of tokens:
tokenize("(+ 10 5)") //=> ["(", "+", "10", "5", ")"]Here’s how we can do that in Rust:
fn tokenize(expr: String) -> Vec<String> {
expr
.replace("(", " ( ")
.replace(")", " ) ")
.split_whitespace()
.map(|x| x.to_string())
.collect()
}Then, we can parse these tokens, into a RispExp:
fn parse<'a>(tokens: &'a [String]) -> Result<(RispExp, &'a [String]), RispErr> {
let (token, rest) = tokens.split_first()
.ok_or(
RispErr::Reason("could not get token".to_string())
)?;
match &token[..] {
"(" => read_seq(rest),
")" => Err(RispErr::Reason("unexpected `)`".to_string())),
_ => Ok((parse_atom(token), rest)),
}
}Note: I depart slightly from Norvig’s implementation, by returning the “next” slice. This lets us recurse and parse nested lists, without mutating the original list.
We get the token for the current position. If it’s the beginning of a list “(“, we start reading and parsing the tokens that follow, until we hit a closing parenthesis:
fn read_seq<'a>(tokens: &'a [String]) -> Result<(RispExp, &'a [String]), RispErr> {
let mut res: Vec<RispExp> = vec![];
let mut xs = tokens;
loop {
let (next_token, rest) = xs
.split_first()
.ok_or(RispErr::Reason("could not find closing `)`".to_string()))
?;
if next_token == ")" {
return Ok((RispExp::List(res), rest)) // skip `)`, head to the token after
}
let (exp, new_xs) = parse(&xs)?;
res.push(exp);
xs = new_xs;
}
}If it’s a closing tag of a list “)”, we return an error, as read_seq should have skipped past it.
Otherwise, it can only be an atom, so we parse that:
fn parse_atom(token: &str) -> RispExp {
let potential_float: Result<f64, ParseFloatError> = token.parse();
match potential_float {
Ok(v) => RispExp::Number(v),
Err(_) => RispExp::Symbol(token.to_string().clone())
}
}Let’s go ahead and create the default, global environment. As Norvig explains, environments are where we will store variable definitions and built-in functions.
To implement built-in operations (+, -), we need a way to save rust function references. Let’s update RispExp, so that we can store rust function references:
#[derive(Clone)]
enum RispExp {
Symbol(String),
Number(f64),
List(Vec<RispExp>),
Func(fn(&[RispExp]) -> Result<RispExp, RispErr>), // bam
}Then, we can create a default_env function, that returns a RispEnv, which implements +, and -
fn default_env() -> RispEnv {
let mut data: HashMap<String, RispExp> = HashMap::new();
data.insert(
"+".to_string(),
RispExp::Func(
|args: &[RispExp]| -> Result<RispExp, RispErr> {
let sum = parse_list_of_floats(args)?.iter().fold(0.0, |sum, a| sum + a);
Ok(RispExp::Number(sum))
}
)
);
data.insert(
"-".to_string(),
RispExp::Func(
|args: &[RispExp]| -> Result<RispExp, RispErr> {
let floats = parse_list_of_floats(args)?;
let first = *floats.first().ok_or(RispErr::Reason("expected at least one number".to_string()))?;
let sum_of_rest = floats[1..].iter().fold(0.0, |sum, a| sum + a);
Ok(RispExp::Number(first - sum_of_rest))
}
)
);
RispEnv {data}
}Note: I am following Clojure’s spec for + and -.
To make this simpler, I made a quick helper, which enforces that all RispExp that we receive are floats:
fn parse_list_of_floats(args: &[RispExp]) -> Result<Vec<f64>, RispErr> {
args
.iter()
.map(|x| parse_single_float(x))
.collect()
}
fn parse_single_float(exp: &RispExp) -> Result<f64, RispErr> {
match exp {
RispExp::Number(num) => Ok(*num),
_ => Err(RispErr::Reason("expected a number".to_string())),
}
}Now, time to implement eval.
If it’s a symbol, we’ll query for that symbol in the environment and return it (for now, it should be a RispExp::Func)
If it’s a number, we’ll simply return it.
If it’s a list, we’ll evaluate the first form. It should be a RispExp::Func. Then, we’ll call that function with all the other evaluated forms as the arguments.
fn eval(exp: &RispExp, env: &mut RispEnv) -> Result<RispExp, RispErr> {
match exp {
RispExp::Symbol(k) =>
env.data.get(k)
.ok_or(
RispErr::Reason(
format!("unexpected symbol k='{}'", k)
)
)
.map(|x| x.clone())
,
RispExp::Number(_a) => Ok(exp.clone()),
RispExp::List(list) => {
let first_form = list
.first()
.ok_or(RispErr::Reason("expected a non-empty list".to_string()))?;
let arg_forms = &list[1..];
let first_eval = eval(first_form, env)?;
match first_eval {
RispExp::Func(f) => {
let args_eval = arg_forms
.iter()
.map(|x| eval(x, env))
.collect::<Result<Vec<RispExp>, RispErr>>();
f(&args_eval?)
},
_ => Err(
RispErr::Reason("first form must be a function".to_string())
),
}
},
RispExp::Func(_) => Err(
RispErr::Reason("unexpected form".to_string())
),
}
}Aand, bam, we have eval.
Now, to make this fun and interactive, let’s make a repl.
We first need a way to convert our RispExp to a string. Let’s implement the Display trait
impl fmt::Display for RispExp {
fn fmt(&self, f: &mut fmt::Formatter) -> fmt::Result {
let str = match self {
RispExp::Symbol(s) => s.clone(),
RispExp::Number(n) => n.to_string(),
RispExp::List(list) => {
let xs: Vec<String> = list
.iter()
.map(|x| x.to_string())
.collect();
format!("({})", xs.join(","))
},
RispExp::Func(_) => "Function {}".to_string(),
};
write!(f, "{}", str)
}
}Then, let’s tie the interpreter process into a loop
fn parse_eval(expr: String, env: &mut RispEnv) -> Result<RispExp, RispErr> {
let (parsed_exp, _) = parse(&tokenize(expr))?;
let evaled_exp = eval(&parsed_exp, env)?;
Ok(evaled_exp)
}
fn slurp_expr() -> String {
let mut expr = String::new();
io::stdin().read_line(&mut expr)
.expect("Failed to read line");
expr
}
fn main() {
let env = &mut default_env();
loop {
println!("risp >");
let expr = slurp_expr();
match parse_eval(expr, env) {
Ok(res) => println!("// 🔥 => {}", res),
Err(e) => match e {
RispErr::Reason(msg) => println!("// 🙀 => {}", msg),
},
}
}
}We can now add and subtract!
risp >
(+ 10 5 (- 10 3 3))
// 🔥 => 19Okay, we have a basic calculator. Now, let’s add support for booleans, and introduce some equality comparators.
To implement bools, let’s include it in our RispExp
#[derive(Clone)]
enum RispExp {
Bool(bool), // bam
Symbol(String),
Number(f64),
List(Vec<RispExp>),
Func(fn(&[RispExp]) -> Result<RispExp, RispErr>),
}Rust will tell us to update Display
impl fmt::Display for RispExp {
fn fmt(&self, f: &mut fmt::Formatter) -> fmt::Result {
let str = match self {
RispExp::Bool(a) => a.to_string(),Then Rust will tell us we should change eval, to consider bools:
fn eval(exp: &RispExp, env: &mut RispEnv) -> Result<RispExp, RispErr> {
match exp {
...
RispExp::Bool(_a) => Ok(exp.clone()),Let’s also update our parse_atom function, to consider bools:
fn parse_atom(token: &str) -> RispExp {
match token.as_ref() {
"true" => RispExp::Bool(true),
"false" => RispExp::Bool(false),
_ => {
let potential_float: Result<f64, ParseFloatError> = token.parse();
match potential_float {
Ok(v) => RispExp::Number(v),
Err(_) => RispExp::Symbol(token.to_string().clone())
}
}
}
}Now, we should be good to go. To really see these in action though, let’s implement =, >, <, >=, <=
In clojure, these comparison operators are a bit special. They can take more than 2 args, and return true if they are in a monotonic order that satisfies the operator.
For example (> 6 5 3 2) is true, because 6 > 5 > 3 > 2. Let’s do this for Risp:
fn default_env() -> RispEnv {
let mut data: HashMap<String, RispExp> = HashMap::new();
...
data.insert(
"=".to_string(),
RispExp::Func(ensure_tonicity!(|a, b| a == b))
);
data.insert(
">".to_string(),
RispExp::Func(ensure_tonicity!(|a, b| a > b))
);
data.insert(
">=".to_string(),
RispExp::Func(ensure_tonicity!(|a, b| a >= b))
);
data.insert(
"<".to_string(),
RispExp::Func(ensure_tonicity!(|a, b| a < b))
);
data.insert(
"<=".to_string(),
RispExp::Func(ensure_tonicity!(|a, b| a <= b))
);
RispEnv {data}
}The key here is our helper macro ensure_tonicty. This takes a checker function, and ensures that the conditional passes in a monotonic way:
macro_rules! ensure_tonicity {
($check_fn:expr) => {{
|args: &[RispExp]| -> Result<RispExp, RispErr> {
let floats = parse_list_of_floats(args)?;
let first = floats.first().ok_or(RispErr::Reason("expected at least one number".to_string()))?;
let rest = &floats[1..];
fn f (prev: &f64, xs: &[f64]) -> bool {
match xs.first() {
Some(x) => $check_fn(prev, x) && f(x, &xs[1..]),
None => true,
}
};
Ok(RispExp::Bool(f(first, rest)))
}
}};
}We can now use comparators, and see booleans!
risp >
(> 6 4 3 1)
// 🔥 => trueOkay, now, let’s make this a language. Let’s introduce def and if.
To do this, let’s update eval to deal with built-in operators:
fn eval(exp: &RispExp, env: &mut RispEnv) -> Result<RispExp, RispErr> {
match exp {
...
RispExp::List(list) => {
let first_form = list
.first()
.ok_or(RispErr::Reason("expected a non-empty list".to_string()))?;
let arg_forms = &list[1..];
match eval_built_in_form(first_form, arg_forms, env) {
Some(res) => res,
None => {
let first_eval = eval(first_form, env)?;
match first_eval {
RispExp::Func(f) => {
let args_eval = arg_forms
.iter()
.map(|x| eval(x, env))
.collect::<Result<Vec<RispExp>, RispErr>>();
return f(&args_eval?);
},
_ => Err(
RispErr::Reason("first form must be a function".to_string())
),
}
}
}
},We take the first form, and try to eval it as a built-in. If we can, voila, otherwise we evaluate as normal.
Here’s how eval_built_in_form looks:
fn eval_built_in_form(
exp: &RispExp, arg_forms: &[RispExp], env: &mut RispEnv
) -> Option<Result<RispExp, RispErr>> {
match exp {
RispExp::Symbol(s) =>
match s.as_ref() {
"if" => Some(eval_if_args(arg_forms, env)),
"def" => Some(eval_def_args(arg_forms, env)),
_ => None,
}
,
_ => None,
}
}Here’s how we can implement if:
fn eval_if_args(arg_forms: &[RispExp], env: &mut RispEnv) -> Result<RispExp, RispErr> {
let test_form = arg_forms.first().ok_or(
RispErr::Reason(
"expected test form".to_string(),
)
)?;
let test_eval = eval(test_form, env)?;
match test_eval {
RispExp::Bool(b) => {
let form_idx = if b { 1 } else { 2 };
let res_form = arg_forms.get(form_idx)
.ok_or(RispErr::Reason(
format!("expected form idx={}", form_idx)
))?;
let res_eval = eval(res_form, env);
res_eval
},
_ => Err(
RispErr::Reason(format!("unexpected test form='{}'", test_form.to_string()))
)
}
}And here’s def:
fn eval_def_args(arg_forms: &[RispExp], env: &mut RispEnv) -> Result<RispExp, RispErr> {
let first_form = arg_forms.first().ok_or(
RispErr::Reason(
"expected first form".to_string(),
)
)?;
let first_str = match first_form {
RispExp::Symbol(s) => Ok(s.clone()),
_ => Err(RispErr::Reason(
"expected first form to be a symbol".to_string(),
))
}?;
let second_form = arg_forms.get(1).ok_or(
RispErr::Reason(
"expected second form".to_string(),
)
)?;
if arg_forms.len() > 2 {
return Err(
RispErr::Reason(
"def can only have two forms ".to_string(),
)
)
}
let second_eval = eval(second_form, env)?;
env.data.insert(first_str, second_eval);
Ok(first_form.clone())
}We now have some coool built-in functions.
risp >
(def a 1)
// 🔥 => a
risp >
(+ a 1)
// 🔥 => 2
risp >
(if (> 2 4 6) 1 2)
// 🔥 => 2
risp >
(if (< 2 4 6) 1 2)
// 🔥 => 1Now, let’s make this a full-on language. Let’s implement _lambdas_! Our syntax can look like this:
(def add-one (fn (a) (+ 1 a)))
(add-one 1) // => 2First things first, let’s introduce a Lambda type for our RispExp
#[derive(Clone)]
enum RispExp {
Bool(bool),
Symbol(String),
Number(f64),
List(Vec<RispExp>),
Func(fn(&[RispExp]) -> Result<RispExp, RispErr>),
Lambda(RispLambda) // bam
}
#[derive(Clone)]
struct RispLambda {
params_exp: Rc<RispExp>,
body_exp: Rc<RispExp>,
}Rust will tell us to update Display:
impl fmt::Display for RispExp {
fn fmt(&self, f: &mut fmt::Formatter) -> fmt::Result {
let str = match self {
...
RispExp::Lambda(_) => "Lambda {}".to_string(),Then Rust will tell us to update eval:
fn eval(exp: &RispExp, env: &mut RispEnv) -> Result<RispExp, RispErr> {
match exp {
...
RispExp::Lambda(_) => Err(RispErr::Reason("unexpected form".to_string())),Now, let’s update eval, to handle fn — this will be the built-in call that creates a Lambda expression:
fn eval_built_in_form(
exp: &RispExp, arg_forms: &[RispExp], env: &mut RispEnv
...
"fn" => Some(eval_lambda_args(arg_forms)),eval_lambda_args can look like this:
fn eval_lambda_args(arg_forms: &[RispExp]) -> Result<RispExp, RispErr> {
let params_exp = arg_forms.first().ok_or(
RispErr::Reason(
"expected args form".to_string(),
)
)?;
let body_exp = arg_forms.get(1).ok_or(
RispErr::Reason(
"expected second form".to_string(),
)
)?;
if arg_forms.len() > 2 {
return Err(
RispErr::Reason(
"fn definition can only have two forms ".to_string(),
)
)
}
Ok(
RispExp::Lambda(
RispLambda {
body_exp: Rc::new(body_exp.clone()),
params_exp: Rc::new(params_exp.clone()),
}
)
)
}For now we only have a global environment. To support lambdas, we need to introduce the concept of scoped environments. Whenever we call a lambda, we’ll need to instantiate a new environment.
To do this, let’s first update our RispEnv struct, to keep an outer reference:
#[derive(Clone)]
struct RispEnv<'a> {
data: HashMap<String, RispExp>,
outer: Option<&'a RispEnv<'a>>,
}Let’s update default_env, to specify the lifetime and return None as the outer environment:
fn default_env<'a>() -> RispEnv<'a> {
...
RispEnv {data, outer: None}
}Then, let’s update eval, to recursively search for symbols in our environment:
fn env_get(k: &str, env: &RispEnv) -> Option<RispExp> {
match env.data.get(k) {
Some(exp) => Some(exp.clone()),
None => {
match &env.outer {
Some(outer_env) => env_get(k, &outer_env),
None => None
}
}
}
}
fn eval(exp: &RispExp, env: &mut RispEnv) -> Result<RispExp, RispErr> {
match exp {
RispExp::Symbol(k) =>
env_get(k, env)
.ok_or(
RispErr::Reason(
format!("unexpected symbol k='{}'", k)
)
)
,Let’s update eval, so that we know what to do when the first form in a list is a lambda:
fn eval(exp: &RispExp, env: &mut RispEnv) -> Result<RispExp, RispErr> {
...
let first_eval = eval(first_form, env)?;
match first_eval {
RispExp::Func(f) => {
f(&eval_forms(arg_forms, env)?)
},
RispExp::Lambda(lambda) => {
let new_env = &mut env_for_lambda(lambda.params_exp, arg_forms, env)?;
eval(&lambda.body_exp, new_env)
},
_ => Err(
RispErr::Reason("first form must be a function".to_string())
),
}We first have a quick helper function to eval a list of expressions, as we’ll be doing that both for RispExp::Func and RispExp::Lambda
fn eval_forms(arg_forms: &[RispExp], env: &mut RispEnv) -> Result<Vec<RispExp>, RispErr> {
arg_forms
.iter()
.map(|x| eval(x, env))
.collect()
}Then, we create a function call env_for_lambda. This will get the params_exp, and create an environment, where each param corresponds to the argument at that index:
fn env_for_lambda<'a>(
params: Rc<RispExp>,
arg_forms: &[RispExp],
outer_env: &'a mut RispEnv,
) -> Result<RispEnv<'a>, RispErr> {
let ks = parse_list_of_symbol_strings(params)?;
if ks.len() != arg_forms.len() {
return Err(
RispErr::Reason(
format!("expected {} arguments, got {}", ks.len(), arg_forms.len())
)
);
}
let vs = eval_forms(arg_forms, outer_env)?;
let mut data: HashMap<String, RispExp> = HashMap::new();
for (k, v) in ks.iter().zip(vs.iter()) {
data.insert(k.clone(), v.clone());
}
Ok(
RispEnv {
data,
outer: Some(outer_env),
}
)
}To do this, we need the helper parse_list_of_symbol_strings, to make sure all of our param definitions are in fact symbols:
fn parse_list_of_symbol_strings(form: Rc<RispExp>) -> Result<Vec<String>, RispErr> {
let list = match form.as_ref() {
RispExp::List(s) => Ok(s.clone()),
_ => Err(RispErr::Reason(
"expected args form to be a list".to_string(),
))
}?;
list
.iter()
.map(
|x| {
match x {
RispExp::Symbol(s) => Ok(s.clone()),
_ => Err(RispErr::Reason(
"expected symbols in the argument list".to_string(),
))
}
}
).collect()
}With that, we can eval(lambda.body_exp, new_env), and…
We now support lambdas!
risp >
(def add-one (fn (a) (+ 1 a)))
// 🔥 => add-one
risp >
(add-one 1)
// 🔥 => 2And with that, we’ve reached the end of this adventure. I hope it’s been fun!
There’s still a bunch more to implement, and ways we can make this even more elegant. If you get to it, send me your thoughts 🙂.
Finally, I have to say, I loved using Rust. It’s the least mental overhead I’ve had to maintain with a systems language, and it was a blast to use. The community is alive and well, plus — their guides are phenomenal! Give it a shot if you haven’t already.
If you liked this post, please share it. For more posts and thoughts, follow me on twitter 🙂.
Special thanks to Mark Shlick, Taryn Hill, Kaczor Donald, for reviewing this essay.
Thanks to eridius for suggesting a cleaner implementation of parse Thanks to thenewwazoo for suggesting a better way to do error handling Thanks to phil_gk for suggesting the use the Display trait