Swift Collections 1.4.0

Related Projects Collections
1

The Swift Collections 1.4.0 package tag is now available.

This feature release supports Swift toolchain versions 6.0, 6.1 and 6.2. It includes a variety of bug fixes, and ships the following new features:

New ownership-aware ring buffer and hashed container implementations

In the DequeModule module, we have two new source-stable types that provide ownership-aware ring buffer implementations:

RigidDeque/UniqueDeque are to Deque like RigidArray/UniqueArray are to Array -- they provide noncopyable embodiments of the same basic data structure, with many of the same operations.

In the BasicContainers module, this release adds previews of four new types, implementing ownership-aware hashed containers:

These are direct analogues of the standard Set and Dictionary types. These types are built on top of the Equatable and Hashable protocol generalizations that were proposed in SE-0499; as that proposal is not yet implemented in any shipping toolchain, these new types are shipping as source-unstable previews, conditional on a new UnstableHashedContainers package trait. The final API of these types will also deeply depend on the struct Borrow and struct Inout proposals (and potentially other language/stdlib improvements) that are currently working their way through the Swift Evolution process. Accordingly, we may need to make source-breaking changes to the interfaces of these types -- they are not ready to be blessed as Public API. However, we encourage intrepid engineers to try them on for size, and report pain points. (Of which we expect there will be many in this first preview.)

We continue the pattern of Rigid- and Unique- naming prefixes with these new types:

  • The Unique types (UniqueArray, UniqueDeque, UniqueSet, UniqueDictionary etc.) are dynamically self-sizing containers that automatically reallocate their storage as needed to best accommodate their contents; the Unique prefix was chosen to highlight that these types are always uniquely held, avoiding the complications of mutating shared copies.
  • The Rigid types remove dynamic sizing, and they operate strictly within an explicitly configured capacity. Dynamic sizing is not always appropriate -- when targeting space- or time-constrained environments (think embedded use cases or real-time work), it is preferable to avoid implicit reallocations, and to instead choose to have explicit control over when (and if) storage is reallocated, and to what size. This is where the Rigid types come in: their instances are created with a specific capacity and it is a runtime error to exceed that. This makes them quite inflexible (hence the "rigid" qualifier), but in exchange, their operations provide far stricter complexity guarantees: they exhibit no random runtime latency spikes, and they can trivially fit in strict memory budgets.

Early drafts of borrowing sequence, generative iteration and container protocols

This release includes highly experimental but working implementations of new protocols supplying ownership-aware alternatives to the classic Sequence/Collection protocol hierarchy. These protocols and the generic operations built on top of them can be turned on by enabling the UnstableContainersPreview package trait.

In this version, the package has developed these protocols just enough to implement basic generic operations for moving data between containers like UniqueArray and RigidDeque. As we gain experience using these, future releases may start adding basic generic algorithms, more protocols (bidirectional, random-access, (per)mutable, range-replaceable containers etc.) convenience adapters, and other features -- or we may end up entirely overhauling or simply discarding some/all of them. Accordingly, the experimental interfaces enabled by UnstableContainersPreview are not source stable, and they are not intended for production use. We expect the eventual production version of these (or whatever designs they evolve into) to ship in the Swift Standard Library. We do highly recommend interested folks to try playing with these, to get a feel for the strange problems of Ownership.

Besides these protocols, the package also defines rudimentary substitutes of some basic primitives that belong in the Standard Library:

  • struct InputSpan<Element> the dual of OutputSpan -- while OutputSpan is primarily for moving items into somebody else's storage, InputSpan enables safely moving items out of storage.
  • struct Borrow<Target> represents a borrowing reference to an item. (This package models this with a pointer, which is an ill-fitting substitute for the real implementation in the stdlib.)
  • struct Inout<Target> represents a mutating reference to an item.

A formal way to access SortedSet and SortedDictionary types

The SortedCollections module contains (preexisting) early drafts of two sorted collection types SortedSet and SortedDictionary, built on top of an in-memory B-tree implementation. This release defines an UnstableSortedCollections package trait that can be used to enable building these types for experimentation without manually modifying the package. Like in previous releases, these implementations remain unfinished in this release, with known API issues; accordingly, these types remain unstable. (Issue #1 remains open.) Future package releases may change their interface in ways that break source compatibility, or they may remove these types altogether.

Minor interface-level changes

  • The Collections module no longer uses the unstable @_exported import feature. Instead, it publishes public typealiases of every type that it previously reexported from DequeModule, OrderedCollections, BitCollections, HeapModule and HashTreeCollections.

  • We renamed some RigidArray/UniqueArray operations to improve their clarity at the point of use. The old names are still available, but deprecated.

    Old name New name
    append(count:initializingWith:) append(addingCount:initializingWith:)
    insert(count:at:initializingWith:) insert(addingCount:at:initializingWith:)
    replaceSubrange(_:newCount:initializingWith:) replace(removing:addingCount:initializingWith:)
    replaceSubrange(_:moving:) replace(removing:moving:)
    replaceSubrange(_:copying:) replace(removing:copying:)
    copy() clone()
    copy(capacity:) clone(capacity:)

  • We have now defined a complete set of OutputSpan/InputSpan-based append/insert/replace/consume primitives, fully generalized to be implementable by piecewise contiguous containers. These operations pave the way for a Container-based analogue of the classic RangeReplaceableCollection protocol, with most of the user-facing operations becoming standard generic algorithms built on top of these primitives:

    mutating func append<E: Error>(
    addingCount newItemCount: Int,
    initializingWith initializer: (inout OutputSpan<Element>) throws(E) -> Void
)

mutating func insert<E: Error>(
   addingCount newItemCount: Int,
   at index: Int,
   initializingWith initializer: (inout OutputSpan<Element>) throws(E) -> Void
) throws(E)

mutating func replace<E: Error>(
    removing subrange: Range<Int>,
    consumingWith consumer: (inout InputSpan<Element>) -> Void,
    addingCount newItemCount: Int,
    initializingWith initializer: (inout OutputSpan<Element>) throws(E) -> Void
) throws(E)

mutating func consume(
    _ subrange: Range<Int>,
    consumingWith consumer: (inout InputSpan<Element>) -> Void
)

  • The package no longer uses the code generation tool gyb.

What's Changed

New Contributors

Full Changelog: Comparing 1.3.0...1.4.0 · apple/swift-collections · GitHub

18 Likes
2

Quick note on the hashed containers: the package uses compiler(>=6.4) conditionals to avoid build errors, but we don't have a toolchain snapshot that reports itself as 6.4 just yet.

To try the new set and dictionary types, we need to wait for a snapshot build that includes the relevant PR. Meanwhile, manually replacing compiler(>=6.4)/compiler(<6.4) with 6.3 will enable these types on the current main snapshots. Sorry for the speed bump!

3 Likes