3 releases (breaking)
0.3.0 | Mar 4, 2022 |
---|---|
0.2.0 | Dec 28, 2020 |
0.1.0 | Dec 3, 2020 |
#741 in Data structures
545KB
8K
SLoC
coca - Data Structures with Constant Capacity
Allocation-free data structures that make do with the memory they're given.
[dependencies]
coca = "0.3"
Overview
Rust's standard collection library provides efficient implementations of the most common general purpose programming data structures. By using the standard implementations, it should be possible for two libraries to communicate without significant data conversion.
However, these standard implementations manage their own memory using an
Allocator, defaulting to the global heap. This is generally
convenient, but may pose a problem in some cases, e.g. in (soft) real-time
applications, or in memory-constrained embedded systems, where the
alloc
crate may not even be available.
coca
aims to serve as a replacement in such environments by providing data
structures that operate on a given block of backing memory, without allocating
on their own. They are generic over the storage type, which may be any of the
following:
InlineStorage
: Store the contents inside thestruct
, without indirection. Requires capacities that are trulyconst
, i.e. statically known at compile time.AllocStorage
: Store the contents in globally allocated memory. Requires thealloc
feature flag.SliceStorage
: For array-like data structures only, store the contents in any given slice of uninitialized memory.ArenaStorage
:coca
includes an arena allocator, and allows ergonomic construction of data structures using arena-allocated memory as storage.
Within this paradigm, direct analogs to the following types are provided:
alloc::vec::Vec
alloc::string::String
alloc::collections::VecDeque
alloc::collections::BinaryHeap
slotmap::{SlotMap, DenseSlotMap}
Additionally, coca
also includes the following container types:
ListSet
, a set implemented as aVec
.ListMap
, an association list implemented as a pair of parallel arrays.CacheTable
, a forgetful hash map with a configurable eviction policy; ideal for caching, hence the name.OptionGroup
, a tuple or array of optional values with the occupancy flags packed into a single bitmap field.InlineObject
, a statically-sized container for dynamically-sized types, mainly trait objects; this requires unstable language features, and therefore needs theunstable
feature flag to be enabled.
Comparison with Other Libraries
First of all, unless you are trying to avoid hitting the global allocator, or
don't have one in your target environment, you are almost certainly better off
just using Rust's standard collections, or in the case of coca::collections::pool
,
the slotmap
crate. Even in such a scenario,
however, there are several crates filling a similar niche that are more mature
than coca
.
coca::arena
vs bumpalo
- Bumpalo is a
no_std
crate, but it does have a hard dependency on thealloc
crate, which it uses to automatically add chunks to its arenas whenever they run out of space. This helps in avoiding failed allocations, but makes it harder to bound memory usage. By contrast,coca
's dependency onalloc
is optional; acoca::arena::Arena
always returnsNone
(or panics, at your option) when it runs out of space, but can be constructed from any mutable byte slice. - Bumpalo has its own forks of Rust's standard
Vec
andString
that use itsBump
allocator, and optionally supports the nightly-onlyAllocator
API for compatibility with the other standard collections. On stable Rust,coca::arena
supports a wider variety of data structures, but it won't benefit from the stabilization offeature(allocator_api)
. - Uniquely,
coca
's arenas can be nested, allowing for stack-like de/allocation patterns.
coca::collections
vs heapless
heapless
provides a variety of data structures with statically known capacity, the equivalent ofcoca
'sInlineStorage
. It has no support for dynamic allocations.heapless
provides direct analogs tostd::collections::HashMap
andHashSet
, whichcoca
does not have (yet).- None of
coca
's data structures are thread-safe, whileheapless
provides multiple synchronization mechanisms: a lock-free memory pool with atomically reference-counting pointers, and both MPMC and SPSC lock-free queues. heapless
does not provide equivalents tostd::collections::VecDeque
,slotmap::SlotMap
orslotmap::DenseSlotMap
, whilecoca
does, on top of the more niche data structures (CacheTable
,OptionGroup
,InlineObject
).
coca::collections::vec
vs tinyvec
and arrayvec
- tinyvec uses no unsafe code, but requires the element type to implement the
Default
trait.coca
has no such restriction, and offers equivalents totinyvec::ArrayVec
andtinyvec::SliceVec
, but nottinyvec::TinyVec
, which is a small-size optimized vector with the ability to reallocate.coca
also requires a newer rust version (min. 1.59) than tinyvec (min. 1.34). - Both arrayvec and tinyvec have optional
serde
support, whilecoca
does not. coca::collections::Vec
supports more storage modes with just one implementation, meaning its instantiations inter-operate more easily, and you can write generic code to handle all of them. It is also generic over the index type, similar to what is offered by thetyped_index_collections
crate.
Feature Flags
alloc
: By default, coca isno_std
compatible; this feature flag enables some trait implementations for conveniently working with heap-allocated storage.profile
: Enables memory profiling in arenas; see the module-level documentation for details.unstable
: If you're working with the nightly rust toolchain, and don't mind depending on unstable features, you can enable this feature to get access toInlineObject
, allowing you to create trait objects without indirection.
License
Licensed under either Apache License, Version 2.0 or Zlib license at your option.
Unless you explicitly state otherwise, any contribution intentionally submitted for inclusion in this crate by you, as defined in the Apache-2.0 license, shall be dual licensed as above, without any additional terms or conditions.