Frozen Collections - Fast Partially Immutable Collections

• Summary
• Creation
  • Short Form
  • Long Form
• Traits
• Performance Considerations
• Analysis and Optimizations
• Cargo Features

Summary

Frozen collections are designed to deliver improved read performance relative to the standard HashMap and HashSet types. They are ideal for use with long-lasting collections which get initialized when an application starts and remain unchanged permanently, or at least extended periods of time. This is a common pattern in service applications.

As part of creating a frozen collection, analysis is performed over the data that the collection will hold to determine the best layout and algorithm to use to deliver optimal performance. Depending on the situation, sometimes the analysis is done at compile-time whereas in other cases it is done at runtime when the collection is initialized. This analysis can take some time, but the value in spending this time up front is that the collections provide faster read-time performance.

Frozen maps are only partially immutable. The keys associated with a frozen map are determined at creation time and cannot change, but the values can be updated at will if you have a mutable reference to the map. Frozen sets however are completely immutable and so never change after creation.

See BENCHMARKS.md for current benchmark numbers.

Creation

Frozen collections are created with one of eight macros: fz_hash_map!, fz_ordered_map!, fz_scalar_map!, fz_string_map!, fz_hash_set!, fz_ordered_set!, fz_scalar_set!, or fz_string_set!. These macros analyze the data you provide and return a custom implementation type that's optimized for the data. All the possible implementations types implement the Map or Set traits.

The macros exist in a short form and a long form, described below.

Short Form

With the short form, you supply the data that goes into the collection and get in return an initialized collection of an unnamed type. For example:

use frozen_collections::*;

let m = fz_string_map!({
    "Alice": 1,
    "Bob": 2,
    "Sandy": 3,
    "Tom": 4,
});

At build time, the macro analyzes the data supplied and determines the best map implementation type to use. As such, the type of m is not known to this code. m will always implement the [Map] trait however, so you can leverage type inference even though you don't know the actual type of m:

use frozen_collections::*;

fn main() {
    let m = fz_string_map!({
        "Alice": 1,
        "Bob": 2,
        "Sandy": 3,
        "Tom": 4,
    });

    more(m);
}

fn more<M>(m: M)
where
    M: Map<&'static str, i32>
{
    assert!(m.contains_key(&"Alice"));
}

Rather than specifying all the data inline, you can also create a frozen collection by passing a vector as input:

use frozen_collections::*;

let v = vec![
    ("Alice", 1),
    ("Bob", 2),
    ("Sandy", 3),
    ("Tom", 4),
];

let m = fz_string_map!(v);

The inline form is preferred however since it results in faster code. However, whereas the inline form requires all the data to be provided at compile time, the vector form enables the content of the frozen collection to be determined at runtime.

Long Form

The long form lets you provide a type alias name which will be created to correspond to the collection implementation type chosen by the macro invocation. Note that you must use the long form if you want to declare a static frozen collection.

use frozen_collections::*;

fz_string_map!(static MAP: MyMapType<&str, i32>, {
    "Alice": 1,
    "Bob": 2,
    "Sandy": 3,
    "Tom": 4,
});

The above creates a static variable called MAP with keys that are strings and values which are integers. As before, you don't know the specific implementation type selected by the macro, but this time you have a type alias (i.e. MyMapType) representing that type. You can then use this alias anywhere you'd like to in your code where you'd like to mention the type explicitly.

To use the long form for non-static uses, replace static with let:

use frozen_collections::*;

fz_string_map!(let m: MyMapType<&str, i32>, {
    "Alice": 1,
    "Bob": 2,
    "Sandy": 3,
    "Tom": 4,
});

more(m);

struct S {
    m: MyMapType,
}
 
fn more(m: MyMapType) {
    assert!(m.contains_key("Alice"));
}

And like in the short form, you can also supply the collection's data via a vector:

use frozen_collections::*;

let v = vec![
    ("Alice", 1),
    ("Bob", 2),
    ("Sandy", 3),
    ("Tom", 4),
];

fz_string_map!(let m: MyMapType<&str, i32>, v);

Traits

The maps created by the frozen collections macros implement the following traits:

• Map. The primary representation of a map. This trait has MapQuery and MapIteration as super-traits.
• MapQuery. A trait for querying maps. This is an object-safe trait.
• MapIteration. A trait for iterating over maps.

The sets created by the frozen collection macros implement the following traits:

• Set. The primary representation of a set. This trait has SetQuery, SetIteration and SetOps as super-traits.
• SetQuery. A trait for querying sets. This is an object-safe trait.
• SetIteration. A trait for iterating over sets.
• SetOps. A trait for set operations like union and intersections.

Performance Considerations

The analysis performed when creating maps tries to find the best concrete implementation type given the data at hand. If all the data is visible to the macro at compile time, then you get the best possible performance. If you supply a vector instead, then the analysis can only be done at runtime and the resulting collection types are slightly slower.

When creating static collections, the collections produced can often be embedded directly as constant data into the binary of the application, thus requiring no initialization time and no heap space. at This also happens to be the fastest form for these collections. If possible, this happens automatically, you don't need to do anything special to enable this behavior.

Analysis and Optimizations

Unlike normal collections, the frozen collections require you to provide all the data for the collection when you create the collection. The data you supply is analyzed which determines which specific underlying implementation strategy to use and how to lay out data internally.

The available implementation strategies are:

• Scalar as Hash. When the keys are of an integer or enum type, this uses the keys themselves as hash codes, avoiding the overhead of hashing.

• Length as Hash. When the keys are of a string type, the length of the keys are used as hash code, avoiding the overhead of hashing.

• Dense Scalar Lookup. When the keys represent a contiguous range of integer or enum values, lookups use a simple array access instead of hashing.

• Sparse Scalar Lookup. When the keys represent a sparse range of integer or enum values, lookups use a sparse array access instead of hashing.

• Left Hand Substring Hashing. When the keys are of a string type, this uses sub-slices of the keys for hashing, reducing the overhead of hashing.

• Right Hand Substring Hashing. Similar to the Left Hand Substring Hashing from above, but using right-aligned sub-slices instead.

• Linear Scan. For very small collections, this avoids hashing completely by scanning through the entries in linear order.

• Ordered Scan. For very small collections where the keys implement the [Ord] trait, this avoids hashing completely by scanning through the entries in linear order. Unlike the Linear Scan strategy, this one can early exit when keys are not found during the scan.

• Classic Hashing. This is the fallback when none of the previous strategies apply. The frozen implementations are generally faster than HashMap and HashSet.

• Binary Search. For relatively small collections where the keys implement the [Ord] trait, classic binary searching is used.

• Eytzinger Search. For larger collections where the keys implement the [Ord] trait, a cache-friendly Eytzinger search is used.

Cargo Features

You can specify the following features when you include the frozen_collections crate in your Cargo.toml file:

• std. Enables small features only available when building with the standard library.

The std feature is enabled by default.