It's over 9 years old (as of 2024-02-18), there are many better guides! You might like https://rust-unofficial.github.io/too-many-lists/
% Let's build a binary tree!
Let's build a binary tree of strings in Rust. To recap, each node in a binary tree:
- must have a value
- may or may not have left and/or right children
So we can describe a single node like this:
struct Node {
val: &str,
l: Option<Node>,
r: Option<Node>,
}
Unfortunately the above won't compile! The first issue is explained in the Rust
book - if a struct will contain borrowed values (val
), we must tell the
compiler how long they're expected to last. The borrow checker can then enforce
that any use of the val
field obeys the restriction. The second issue is simply
that Rust cannot know how much memory to allocate to the data type if it
recursively contains itself. The fix is simply to fill this field with a
pointer instead of an instance of the type. Our fixed Node:
struct Node<'a> {
val: &'a str,
l: Option<Box<Node<'a>>>,
r: Option<Box<Node<'a>>>,
}
Now we know what our trees are made up of, but we have no way to build one. Let's create a method to allow insertion of items into a tree. My preference as a rust beginner is to write down the Rust-like pseudocode (i.e. minus mutability and lifetime specifiers) of the logic I'd like to use, then convince the borrow checker that my code is correct.
impl Node {
pub fn insert(&self, new_val: &str) {
if self.val == new_val {
return
}
let target_node = if new_val < self.val { &self.l } else { &self.r };
match target_node {
&Some(subnode) => subnode.insert(new_val),
&None => {
let new_node = Node { val: new_val, l: None, r: None };
let boxed_node = Some(Box::new(new_node));
*target_node = boxed_node;
}
}
}
}
There's nothing exciting going on here. We find which side of the current Node the new value should go on, then either add it as a subnode if there isn't one already on that side, or recurse down by calling insert on the subnode.
Needless to say, the borrow checker finds this piece of code very offensive! Since the first compiler complaints are about lifetimes, let's work those out first.
The first error we get is pretty clear (wrong number of lifetime parameters
)
and easily fixed, right? We're doing an impl
for Node
, but our type is
actually Node<'a>
, so tweak that and the problem is solved? Unfortunately not.
Let's consider at a bit of background here. Looking back at the section on
'Ownership' in the rust book, we can see that when creating functions we have
to declare the lifetimes as a parameter of the function to bring those
lifetime names in scope. This is the key - whenever a lifetime is referenced, it
has been declared somewhere beforehand. In fact, this is what we did for the
struct when saying that val
would last for the lifetime 'a
.
Coming back to the problem at hand, we know know we need to bring the lifetime into scope somehow. The correct way to do this in an impl is as a parameter to the impl itself, like so:
impl<'a> Node<'a> {
If we put this in, we now get an error about not being able to infer an appropriate lifetime for new_val when creating a new Node. There's a temptation when seeing an error like this to shove in a lifetime as a parameter to the function, add it to the function argument and hope for the best. In fact, a lot of the time this does work.
Unfortunately you need to take more care when implementing functions for an impl, as there's already a lifetime hanging around - the lifetime for the entire struct. Looking back at the definition of a Node, we've said that the value and sub-Nodes of a Node (and therefore all their values as well) must have the same lifetime as the Node itself. The error is about the value of a sub-Node, so let's make sure it has the same lifetime as the parent Node.
pub fn insert(&self, new_val: &'a str) {
We've used the lifetime brought in scope by the impl, used as the lifetime of the struct the impl is for, as the lifetime of the value.
Hooray! We've completed a first pass of convincing the borrow-checker that all the values involved live for the correct amount of time. The next stage is to make sure that we can actually modify values, as Rust requires us to explicitly keep track of mutability.
We get two different types of error if we try to compile now. The first talks
about moving values out of a borrow - when we assigned the value of
target_node
, the &
operator was used so we took a reference. The pattern
we match on attempts to take ownership (i.e. 'move') the value inside the
target_node option...which isn't allowed, because we're just borrowing it! The
Rust book chapter on patterns mentions the way forward here: ref
.
&Some(ref subnode) => subnode.insert(new_val),
The next error relates to target_node
. We want to replace the value being
pointed to (*target_node = ...
), so we need to make the
reference we take be mutable. This has the knock on effect of needing to
change our pattern matching, as mut
is technically part of the type.
let target_node = if new_val < self.val { &mut self.l } else { &mut self.r };
match target_node {
&mut Some(ref subnode) => subnode.insert(new_val),
&mut None => {
Attempting to compile now tells us that we can't take a mutable reference to
self.l
or self.r
as self
isn't actually mutable. We fix this in the
arguments to the function:
pub fn insert(&mut self, new_val: &'a str) {
As is typical when fixing an error by adding mutability or lifetime constraints,
changes are required elsewhere to comply with the new rules. In this case, our
ref subnode
is not mutable, so you can't call subnode.insert
(which now
requires self
to be mutable). Again referring back to the patterns section
of the Rust book, we find our answer:
&mut Some(ref mut subnode) => subnode.insert(new_val),
It's worth noting that this has worked because we're already taking a
mutable reference to the Option
- if that was immutable, so would the sub-Node
inside be.
You should now find that compilation is successful! The completed code is below along with a few test cases to see it working.
#[derive(PartialEq)]
struct Node<'a> {
val: &'a str,
l: Option<Box<Node<'a>>>,
r: Option<Box<Node<'a>>>,
}
impl<'a> Node<'a> {
pub fn insert(&mut self, new_val: &'a str) {
if self.val == new_val {
return
}
let target_node = if new_val < self.val { &mut self.l } else { &mut self.r };
match target_node {
&mut Some(ref mut subnode) => subnode.insert(new_val),
&mut None => {
let new_node = Node { val: new_val, l: None, r: None };
let boxed_node = Some(Box::new(new_node));
*target_node = boxed_node;
}
}
}
}
fn main () {
let mut x = Node { val: "m", l: None, r: None };
x.insert("z");
x.insert("b");
x.insert("c");
assert!(x == Node {
val: "m",
l: Some(Box::new(Node {
val: "b",
l: None,
r: Some(Box::new(Node { val: "c", l: None, r: None })),
})),
r: Some(Box::new(Node { val: "z", l: None, r: None })),
});
}
In the next instalment we'll look at implementing an iterator and why putting
'a
as a lifetime everywhere can be somewhat limiting.
nice article :) helped a lot!