[][src]Crate dyn_cache

Caches for storing the results of repeated function calls. The caches use minimal dynamic dispatch to store arbitrarily many types of query results in a single store.

Cache storage is indexed by dynamic scopes:

let storage = dyn_cache::local::SharedLocalCache::default();

// scopes can be identified by ~anything Eq + Hash
let a_scope = 'a';
let b_scope = 'b';

// we use interior mutability here to demonstrate query side effects
let count = std::cell::Cell::new(0);
let increment = |&to_add: &i32| -> i32 {
    // let's pretend that there's some other interesting work happening here...
    let new = count.get() + to_add;
    count.set(new);
    new
};

// now we'll define some "queries" to the cache
let a_inc = |n| storage.cache(&a_scope, &n, &increment);
let b_inc = |n| storage.cache(&b_scope, &n, &increment);

assert_eq!(count.get(), 0, "haven't called any queries");

assert_eq!(a_inc(1), 1);
assert_eq!(count.get(), 1, "called 'a'(1) once");

assert_eq!(a_inc(1), 1);
assert_eq!(count.get(), 1, "called 'a'(1) twice, only ran once");

assert_eq!(b_inc(2), 3);
assert_eq!(count.get(), 3, "called 'a'(1) and 'b'(2)");

assert_eq!(a_inc(1), 1, "retains cached value");
assert_eq!(count.get(), 3, "queries only affect their own scope");

assert_eq!(a_inc(2), 5);
assert_eq!(count.get(), 5, "called 'a'(1), 'a'(2), 'b'(2)");

assert_eq!(a_inc(1), 6, "only the most recent revision is cached");
assert_eq!(count.get(), 6);

A single cache instance can hold multiple types of scope:

let storage = dyn_cache::local::SharedLocalCache::default();
let count = std::cell::Cell::new(0);
let increment = |&to_add: &i32| -> i32 {
    // let's pretend that there's some other interesting work happening here...
    let new = count.get() + to_add;
    count.set(new);
    new
};

let one_scope = 1u8;
let two_scope = 2i32;
let red_scope = b"red";
let blue_scope = "blue";

// each of these queries has a different type of scope
// and while the inputs/outputs are the same they could also
// vary without interfering with each other
let one_inc = |n| storage.cache(&one_scope, &n, increment);
let two_inc = |n| storage.cache(&two_scope, &n, increment);
let red_inc = |n| storage.cache(&red_scope, &n, increment);
let blue_inc = |n| storage.cache(&blue_scope, &n, increment);

assert_eq!(one_inc(1), 1);
assert_eq!(count.get(), 1);

assert_eq!(two_inc(1), 2);
assert_eq!(one_inc(1), 1, "still cached");
assert_eq!(count.get(), 2, "only one of the queries ran");

assert_eq!(red_inc(2), 4);
assert_eq!(two_inc(1), 2, "still cached");
assert_eq!(one_inc(1), 1, "still cached");
assert_eq!(count.get(), 4, "only one of the queries ran");

assert_eq!(blue_inc(3), 7);
assert_eq!(red_inc(2), 4, "still cached");
assert_eq!(two_inc(1), 2, "still cached");
assert_eq!(one_inc(1), 1, "still cached");
assert_eq!(count.get(), 7, "only one of the queries ran");

// invalidation still happens once per scope (type)
assert_eq!(blue_inc(5), 12, "blue has a different input");
assert_eq!(red_inc(2), 4, "still cached");
assert_eq!(two_inc(1), 2, "still cached");
assert_eq!(one_inc(1), 1, "still cached");
assert_eq!(count.get(), 12, "only one of the queries ran");

Cache types

There are two main flavors of cache available for use in this crate:

Shared typeSynchronized?
sync::SharedSendCacheMutex
local::SharedLocalCacheRefCell

These variants are used by calling sync::SharedSendCache::cache_with or local::SharedLocalCache::cache.

The shared cache types above are implemented by wrapping these "inner" types:

Mutable typeRequires Send?
sync::SendCacheyes
local::LocalCacheno

These "inner" caches require mutable access to call their functions like local::LocalCache::get which returns either a reference or a CacheMiss that can be passed back to the cache in local::LocalCache::store to initialize a value in the cache:

let mut cache = dyn_cache::local::LocalCache::default();
let scope = &'a';
let arg = &1;

let miss = cache.get(scope, arg).expect_err("first access will always be a miss");
let (entry, result): (_, Vec<usize>) = miss.init(|&n| {
    let v: Vec<usize> = vec![n; n];
    (v.clone(), v)
});
cache.store(entry);
assert_eq!(result, vec![1usize]);

let result: &Vec<usize> = cache.get(scope, arg).unwrap();
assert_eq!(result, &vec![1]);

See sync::SendCache::get and sync::SendCache::store for the thread-safe equivalents.

The shared variants are defined by wrapping these inner cache types in reference counting and synchronized mutability.

Query types

Each query type maps to a typed "namespace" within the unityped cache storage, each query having a distinct type each for its scope, input, and output.

Scopes

The scope of a query is its identifier within cache storage for the given input & output types. Scopes must implement Eq and Hash so that results can be efficiently and uniquely indexed.

Each scope identifies 0-1 (Input, Output) pairs in each namespace. The same type of scope can be used in multiple queries without collision if the types of inputs, outputs, or both differ.

Inputs

The input to a query determines when it is re-run. If a given query is present in the cache then the previous input is compared to the new input. If the input hasn't changed, the query can be skipped and its previously-stored output is returned.

Outputs

The only constraint on query outputs is that they are owned (Output: 'static). This imposes the inconvenient requirement that all access to stored values occurs during the scope of a closure (similar to thread-locals in the standard library).

The most common way to work around this requirement is to choose output types that cheaply implement std::clone::Clone.

Allocations

In order to store distinct query results in the same container, allocations and indirection are required.

Borrowed query parameters

All of the cache functions accept a reference to a type Key: ToOwned<Owned=Scope> so that the scope is only cloned on the first insertion to its storage and all subsequent lookups can be with a borrowed type.

Like the query scope, functions to get cache values accept a borrowed version of the input and only clone it when the input has changed.

Causes

There are three situations where these caches allocate:

  1. caching new types which haven't been seen by that cache instance yet
  2. storing the results of a new query
  3. updating the results of a stored query

There are several types of allocations performed by the caches in this crate:

AllocationCauses
box a new, empty namespace(1)
resize a cache's map of namespaces(1)
call .to_owned() on a scope/key(2)
resize a namespace's storage(2)
call .to_owned() on an input/arg(2), (3)
update an output's dependents(1), (2), (3)

Outside of these, only user-defined functions should perform any allocation.

Garbage Collection

All of the caches have a gc() method which retains only used values. A value is used if it or a value which depends on it has been used/rooted since the last call to gc().

let storage = dyn_cache::local::SharedLocalCache::default();
let a_scope = 'a';
let b_scope = 'b';

// we use interior mutability here to demonstrate query side effects
let count = std::cell::Cell::new(0);
let increment = |&to_add: &i32| -> i32 {
    // let's pretend that there's some other interesting work happening here...
    let new = count.get() + to_add;
    count.set(new);
    new
};

// we'll define the same "queries" to the cache as in the previous example
let a_inc = |n| storage.cache(&a_scope, &n, &increment);
let b_inc = |n| storage.cache(&b_scope, &n, &increment);

assert_eq!(a_inc(1), 1);
assert_eq!(count.get(), 1, "called 'a'(1) once");

assert_eq!(b_inc(2), 3);
assert_eq!(count.get(), 3, "called 'a'(1) and 'b'(2)");

// mark the end of this "revision" in the cache
// this won't drop anything yet, just marks all cached values as unused
storage.gc();

// run only one of the queries to mark it live
assert_eq!(a_inc(1), 1, "value is still cached");
assert_eq!(count.get(), 3, "nothing has touched our side effect tracker");

storage.gc(); // drops b_inc from storage

assert_eq!(b_inc(2), 5, "b_inc was dropped from the cache, ran again");
assert_eq!(count.get(), 5);

assert_eq!(a_inc(1), 1, "value is still cached");
assert_eq!(count.get(), 5);

Nesting

When a cache read fails, we expect that the value will be populated immediately after and a new node in the dependency graph is created. The new dependency node is marked as an incoming dependent on any cache values which are accessed during the initialization of the new value. The new node is then marked as a "root" for the garbage collector once it has been initialized and the cache populated. If in subsequent revisions the rooted value is accessed again it will be re-rooted and its dependents will be marked as live even if they were not directly accessed in that revision.

When a cache read succeeds, its dependency node is marked as being depended upon by the node (if any) which was being initialized during the read, linking the two dependencies together.

let storage = dyn_cache::local::SharedLocalCache::default();
let a_scope = 'a';
let b_scope = 'b';

let count = std::cell::Cell::new(0);
let increment = |&to_add: &i32| -> i32 {
    // let's pretend that there's some other interesting work happening here...
    let new = count.get() + to_add;
    count.set(new);
    new
};

let a_inc = |n| storage.cache(&a_scope, &n, &increment);

// this new query "depends on" a_inc by calling it in its own init closure
let b_inc = |n| storage.cache(&b_scope, &n, |&n| a_inc(n));

assert_eq!(b_inc(2), 2);
assert_eq!(count.get(), 2);

// until now, we haven't called a_inc directly
assert_eq!(a_inc(2), 2, "a_inc is indeed cached as a dep of b_inc");
assert_eq!(count.get(), 2);

storage.gc(); // mark both queries dead

// in this revision we'll only call b_inc directly
assert_eq!(b_inc(3), 5);
assert_eq!(count.get(), 5);

storage.gc(); // doesn't actually drop anything

// both queries should still have their outputs for input=3 cached
assert_eq!(b_inc(3), 5);
assert_eq!(a_inc(3), 5);
assert_eq!(count.get(), 5);

// we can also check to make sure that neither query is touching the cell
count.set(0);
assert_eq!(b_inc(3), 5);
assert_eq!(a_inc(3), 5);
assert_eq!(count.get(), 0);

Modules

local

A cache for types which are not thread-safe (?Send).

sync

A thread-safe cache which requires stored types implement Send.

Structs

CacheEntry

A fully-initialized input/output entry, ready to be written to the cache. Obtained from CacheMiss::init and passed to local::LocalCache::store or sync::SendCache::store.

CacheMiss

The result of a failed attempt to retrieve a value from the cache. Initialize a full CacheEntry for storage with CacheMiss::init.