index_vec

This crate helps with defining "newtype"-style wrappers around usize (or other integers), and Vec<T> so that some additional type safety can be gained at zero cost.

Example / Overview

use index_vec::{IndexVec, IndexSlice, index_vec};

index_vec::define_index_type! {
    // Define StrIdx to use only 32 bits internally (you can use usize, u16,
    // and even u8).
    pub struct StrIdx = u32;

    // The defaults are very reasonable, but this macro can let
    // you customize things quite a bit:

    // By default, creating a StrIdx would check an incoming `usize against
    // `u32::max_value()`, as u32 is the wrapped index type. Lets imagine that
    // StrIdx has to interface with an external system that uses signed ints.
    // We can change the checking behavior to complain on i32::max_value()
    // instead:
    MAX_INDEX = i32::max_value() as usize;

    // We can also disable checking all-together if we are more concerned with perf
    // than any overflow problems, or even do so, but only for debug builds: Quite
    // pointless here, but an okay example
    DISABLE_MAX_INDEX_CHECK = cfg!(not(debug_assertions));

    // And more too, see this macro's docs for more info.
}

// Create a vector which can be accessed using `StrIdx`s.
let mut strs: IndexVec<StrIdx, &'static str> = index_vec!["strs", "bar", "baz"];

// l is a `StrIdx`
let l = strs.last_idx();
assert_eq!(strs[l], "baz");

let new_i = strs.push("quux");
assert_eq!(strs[new_i], "quux");

// The slice APIs are wrapped as well.
let s: &IndexSlice<StrIdx, [&'static str]> = &strs[StrIdx::new(1)..];
assert_eq!(s[0], "bar");

// Indices are mostly interoperable with `usize`, and support
// a lot of what you might want to do to an index.

// Comparison
assert_eq!(StrIdx::new(0), 0usize);

// Addition
assert_eq!(StrIdx::new(0) + 1, StrIdx::new(1));

// Subtraction
assert_eq!(StrIdx::new(1) - 1, StrIdx::new(0));

// Wrapping
assert_eq!(StrIdx::new(5) % strs.len(), StrIdx::new(1));
// ...

Background

The goal is to help with the pattern of using a type FooIdx = usize to access a Vec<Foo> with something that can statically prevent using a FooIdx in a Vec<Bar>. It's most useful if you have a bunch of indices referring to different sorts of vectors.

The code was originally based on rustc's IndexVec code, however that has been almost entirely rewritten (except for the cases where it's trivial, e.g. the Vec wrapper).

Other crates

The indexed_vec crate predates this, and is a much closer copy of the code from rustc. Unfortunately, this means it does not compile on stable.

If you're looking for something further from a vec and closer to a map, you might find handy, slotmap, or slab to be closer what you want.

FAQ

Wouldn't define_index_type be better as a proc macro?

Probably. It's not a proc macro because I tend to avoid them where possible due to wanting to minimize compile times. If the issues around proc-macro compile times are fixed, then I'll revisit this.

I also may eventually add a proc-macro feature which is not required, but avoids some of the grossness.

Does define_index_type do too much?

Possibly. It defines a type, implements a bunch of functions on it, and quite a few traits. That said, it's intended to be a very painless journey from Vec<T> + usize to IndexVec<I, T>. If it left it up to the developer to do those things, it would be too annoying to be worth using.

The syntax for the options in define_index_type is terrible.

I'm open to suggestions.

Does it support no_std?

Yes, although it uses extern crate alloc;, of course.

Does it support serde?

Yes, but only if you turn on the serde feature.

What features are planned?

Planned is a bit strong but here are the things I would find useful.

• Support any remaining parts of the slice/vec api.
• Add typesafe wrappers for SmallVec/ArrayVec (behind a cargo feature, of course).
• Better syntax for the define_index_type macro (no concrete ideas).
• Allow the generated type to be a tuple struct, or use a specific field name.
• Allow use of indices for string types (the primary benefit here would probably be the ability to e.g. use u32 without too much pain rather than mixing up indices from different strings -- but you never know!)
• Allow index types such as NonZeroU32 and such, if it can be done sanely.
• ...

License

This is based on code from rustc's source, and retains it's status as dual-licensed under MIT (LICENSE-MIT) / Apache 2.0 (LICENSE-APACHE).