Fast and Simple Rust Interner

This post describes a simple technique for writing interners in Rust which I havent seen documented before.

String interning is a classical optimization when you have to deal with many equal strings. The canonical example would be a compiler: most identifiers in a program are repeated several times.

Interning works by ensuring that theres only one canonical copy of each distinct string in memory. It can give the following benefits:

The simplest possible interner in Rust could look like this:

use std::collections::HashMap;

pub struct Interner {
    map: HashMap<String, u32>,
    vec: Vec<String>,

impl Interner {
    pub fn intern(&mut self, name: &str) -> u32 {
        if let Some(&idx) = {
            return idx;
        let idx = as u32;, idx);

        debug_assert!(self.lookup(idx) == name);
        debug_assert!(self.intern(name) == idx);


    pub fn lookup(&self, idx: u32) -> &str {
        self.vec[idx as usize].as_str()

To remove duplicates, we store strings in a HashMap. To map from an index back to the string, we also store strings in a Vec.

I didnt quite like this solution yesterday, for two reasons:

So Ive spent a part of the evening cobbling together a non-allocating trie-based interner. The result: trie does indeed asymptotically reduce the number of allocations from O(n) to O(log(n)). Unfortunately, it is slower, larger and way more complex than the above snippet. Minimizing allocations is important, but allocators are pretty fast, and that shouldnt be done at the expense of everything else. Also, Rust HashMap (implemented by @Amanieu based on Swiss Table) is fast.

For the curious, the Trie design I've used

The trie is build on per-byte basis (each node has at most 256 children). Each internal node is marked with a single byte. Leaf nodes are marked with substrings, so that only the common prefix requires node per byte.

To avoid allocating individual interned strings, we store them in a single long String. An interned string is represented by a Span (pair of indexes) inside the big buffer.

Trie itself is a tree structure, and we can use a standard trick of packing its nodes into array and using indexes to avoid allocating every node separately. However, nodes themselves can be of varying size, as each node can have different number of children. We can still array-allocate them, by rolling our own mini-allocator (using a segregated free list)!

Nodes children are represented as a sorted array of links. We use binary search for indexing and simple linear shift insertion. With at most 256 children per node, it shouldnt be that bad. Additionally, we pre-allocate 256 nodes and use array indexing for the first transition.

Links are organized in layers. The layer n stores a number of [Link] chunks of length 2n (in a single contiguous array). Each chunk represents the links for a single node (with possibly some extra capacity). Node can find its chunk because it knows the number of links (which gives the number of layers) and the first link in the layer. A new link for the node is added to the current chunk if theres space. If the chunk is full, it is copied to a chunk twice as big first. The old chunk is then added to the list of free chunks for reuse.

Heres the whole definition of the data structure:

pub struct Interner {
    trie: Vec<Node>,
    links: Vec<Layer>,
    strs: Vec<Span>,
    buf: String,

struct Span { start: u32, end: u32 }

struct Node {
    str: Option<u32>,
    n_links: u8,
    first_link: u32,
//  layer: u32 = first_link.next_power_of_two(),

struct Link { byte: u8, node: u32, }

struct Layer {
    links: Vec<Link>,
    free: Vec<u32>,

Isnt it incredibly cool that you can look only at the fields and understand how the thing works, without even seeing the rest 150 lines of relatively tricky implementation?

However, implementing a trie made me realize that theres a simple optimization we can apply to our naive interner to get rid of extra allocations. In the trie, I concatenate all interned strings into one giant String and use (u32, u32) index pairs as an internal representation of string slice.

If we translate this idea to our naive interner, we get:

struct Span { start: u32, end: u32 }

struct Interner {
    map: HashMap<Span, u32>,
    vec: Vec<Span>,
    buf: String,

impl Interner {
    pub fn intern(&mut self, name: &str) -> u32 { ... }

    pub fn lookup(&self, idx: u32) -> &str {
        let Span { start, end } = self.vec[idx as usize]
        &self.buf[start as usize..end as usize]

The problem here is that we cant actually write implementations of Eq and Hash for Span to make this work. In theory, this is possible: to compare two Spans, you resolve them to &str via buf, and then compare the strings. However, Rust API does not allow to express this idea. Moreover, even if HashMap allowed supplying a key closure at construction time, it wouldnt help!

impl HashMap<K, V, KeyFn, Key>
    KeyFn: Fn(&K) -> Key,
    Key: Hash + Eq,
    fn new_with_key_fn(key_fn: F) -> Self { ... }

Such API would run afoul of the borrow checker. The key_fn would have to borrow from the same struct. What would work is supplying a key_fn at call-site for every HashMap operation, but that would hurt ergonomics and ease of use a lot. This exact problem requires slightly unusual design of lazy values in Rust.

However, with a bit of unsafe, we can make something similar work. The trick is to add strings to buf in such a way that they are never moved, even if more strings are added on top. That way, we can just store &str in the HashMap. To achieve address stability, we use another trick from the typed_arena crate. If the buf is full (so that adding a new string would invalidate old pointers), we allocate a new buffer, twice as large, without coping the contents of the old one.

Heres the full implementation:

use std::{mem, collections::HashMap};

pub struct Interner {
    map: HashMap<&'static str, u32>,
    vec: Vec<&'static str>,
    buf: String,
    full: Vec<String>,

impl Interner {
    pub fn with_capacity(cap: usize) -> Interner {
        let cap = cap.next_power_of_two();
        Interner {
            map: HashMap::default(),
            vec: Vec::new(),
            buf: String::with_capacity(cap),
            full: Vec::new(),

    pub fn intern(&mut self, name: &str) -> u32 {
        if let Some(&id) = {
            return id;
        let name = unsafe { self.alloc(name) };
        let id = as u32;, id);

        debug_assert!(self.lookup(id) == name);
        debug_assert!(self.intern(name) == id);


    pub fn lookup(&self, id: u32) -> &str {
        self.vec[id as usize]

    unsafe fn alloc(&mut self, name: &str) -> &'static str {
        let cap = self.buf.capacity();
        if cap < self.buf.len() + name.len() {
            let new_cap = (cap.max(name.len()) + 1)
            let new_buf = String::with_capacity(new_cap);
            let old_buf = mem::replace(&mut self.buf, new_buf);

        let interned = {
            let start = self.buf.len();

        &*(interned as *const str)

The precise rule for increasing capacity is slightly more complicated:

let new_cap = (cap.max(name.len()) + 1).next_power_of_two();

Just doubling wont be enough, we also need to make sure that the new string actually fits.

We could have used a single bufs: Vec<String> in place of both buf and full. The benefit of splitting the last buffer into a dedicated field is that we statically guarantee that theres at least one buffer. That way, we void a bounds check and/or .unwrap when accessing the active buffer.

We also use &'static str to fake interior references. Miri (rust in-progress UB checker) is not entirely happy about this. I havent dug into this yet, it might be another instance of rust-lang/rust#61114. To be on the safe side, we can use *const str instead, with a bit of boilerplate to delegate PartialEq and Hash. Some kind of (hypothetical) 'unsafe lifetime could also be useful here! The critical detail that makes our use of fake 'static sound here is that the alloc function is private. The public lookup function shortens the lifetime to that of &self (via lifetime elision).

For the real implementation, I would change two things:

Thats all I have to say about fast and simple string interning in Rust! Discussion on /r/rust.