Permalink
601 lines (492 sloc) 24.2 KB

A New Model for Collections and Indices

Summary

We propose a new model for Collections wherein responsibility for index traversal is moved from the index to the collection itself. For example, instead of writing i.successor(), one would write c.index(after: i). We also propose the following changes as a consequence of the new model:

  • A collection's Index can be any Comparable type.
  • The distinction between intervals and ranges disappears, leaving only ranges.
  • A closed range that includes the maximal value of its Bound type is now representable and does not trap.
  • Existing “private” in-place index traversal methods are now available publicly.

Motivation

In collections that don't support random access, (string views, sets, dictionaries, trees, etc.) it's very common that deriving one index value from another requires somehow inspecting the collection's data. For example, you could represent an index into a hash table as an offset into the underlying storage, except that one needs to actually look at structure of the hash table to reach the next bucket. In the current model, supporting i.successor() means that the index must additionally store not just an offset, but a reference to the collection's structure.

The consequences for performance aren't pretty:

  • Code that handles indices has to perform atomic reference counting, which has significant overhead and can prevent the optimizer from making other improvements.

  • Additional references to a collection's storage block the library-level copy-on-write optimization: in-place mutation of uniquely-referenced data. A live index makes underlying storage non-uniquely referenced, forcing unnecessary copies when the collection is mutated. In the standard library, Dictionary and Set use a double-indirection trick to work around this issue. Unfortunately, even this trick is not a solution, because (as we have recently realized) it isn't threadsafe. [^1]

By giving responsibility for traversal to the collection, we ensure that operations that need the collection's structure always have it, without the costs of holding references in indices.

Other Benefits

Although this change is primarily motivated by performance, it has other significant benefits:

  • Simplifies implementation of non-trivial indices.
  • Allows us to eliminate the Range/Interval distinction.
  • Making traversal a direct property of the Collection protocol, rather than its associated Index type, is closer to most peoples' mental model for collections, and simplifies the writing of many generic constraints.
  • Makes it feasible to fix existing concurrency issues in Set and Dictionary indices.
  • Allows String views to share a single index type, letting us eliminate the need for cumbersome index conversion functions (not part of this proposal, but planned).

Out of Scope

This proposal intentionally does not:

  • Expand the set of concrete collections provided by the standard library.
  • Expand the set of collection protocols to provide functionality beyond what is already provided (for example, protocols for sorted collections, queues etc.) Discussing how other concrete collections fit into the current protocol hierarchy is in scope, though.

Limitations of the Model

Ideally, our collection model would allow us to implement every interesting data structure with memory safety, optimal performance, value semantics, and a variety of other useful properties such as minimal invalidation of indexes upon mutation. In practice, these goals and the Swift language model interact in complicated ways, preventing some designs altogether, and suggesting a variety of implementation strategies for others that can be selected based on one's priorities. We've done some in-depth investigation of these implications, but presenting and explaining them is well beyond the scope of this proposal.

We can, however, be fairly sure that this change does not regress our ability to build any Collections that could have been built in Swift 2.2. After all, it is still possible to implement indices that store references and have the old traversal methods (the collection's traversal methods would simply forward to those of the index), so we haven't lost the ability to express anything.

Overview of Type And Protocol Changes

This section covers the proposed structural changes to the library at a high level. Details such as protocols introduced purely to work around compiler limitations (e.g. Indexable or IndexableBase) have been omitted. For a complete view of the code and documentation changes implementing this proposal, please see this pull request.

Collection Protocol Hierarchy

In the proposed model, indices don't have any requirements beyond Comparable, so the ForwardIndex, BidirectionalIndex, and RandomAccessIndex protocols are eliminated. Instead, we introduce BidirectionalCollection and RandomAccessCollection to provide the same traversal distinctions, as shown here:

                     +--------+
                     |Sequence|
                     +---+----+
                         |
                    +----+-----+
                    |Collection|
                    +----+-----+
                         |
          +--------------+-------------+
          |              |             |
          |     +--------+--------+    |
          |     |MutableCollection|    |
          |     +-----------------+    |
          |                            |
+---------+-------------+    +---------+----------------+
|BidirectionalCollection|    |RangeReplaceableCollection|
+---------+-------------+    +--------------------------+
          |
 +--------+-------------+
 |RandomAccessCollection|
 +----------------------+

These protocols compose naturally with the existing protocols MutableCollection and RangeReplaceableCollection to describe a collection's capabilities, e.g.

struct Array<Element>
  : RandomAccessCollection,
    MutableCollection,
    RangeReplaceableCollection { ... }

struct UnicodeScalarView : BidirectionalCollection { ... }

Range Types

The proposal adds several new types to support ranges:

  • The old Range<T>, ClosedInterval<T>, and OpenInterval<T> are replaced with four new generic range types:

    • Two for general ranges (whose bounds are Comparable): Range<T> and ClosedRange<T>. Having a separate ClosedRange type allows us to address the vexing inability of the old Range to represent a range containing the maximal value of its bound.

    • Two for ranges that additionally conform to RandomAccessCollection (requiring bounds that are Strideable with Stride conforming to Integer): CountableRange<T> and CountableClosedRange<T>. These types can be folded into Range and ClosedRange when Swift acquires conditional conformance capability.

The Associated Indices Type

The following code iterates over the indices of all elements in collection:

for index in collection.indices { ... }

In Swift 2, collection.indices returned a Range<Index>, but because a range is a simple pair of indices and indices can no longer be advanced on their own, Range<Index> is no longer iterable.

In order to keep code like the above working, Collection has acquired an associated Indices type that is always iterable, and three generic types were introduced to provide a default Indices for each Collection traversal category: DefaultIndices<C>, DefaultBidirectionalIndices<C>, and DefaultRandomAccessIndices<C>. These types store the underlying collection as a means of traversal. Collections like Array whose Indices don't need the collection simply use typealias Indices = CountableRange<Index>.

Expanded Default Slice Types

Because Swift doesn't support conditional protocol conformances and the three traversal distinctions have been moved into the Collection hierarchy, the four generic types Slice, MutableSlice, RangeReplaceableSlice, and MutableRangeReplaceableSlice have become twelve, with the addition of variations such as RangeReplaceableBidirectionalSlice.

The Comparable Requirement on Indices

In this model indices store the minimal amount of information required to describe an element's position. Usually an index can be represented with one or two Ints that efficiently encode the path to the element from the root of a data structure. Since one is free to choose the encoding of the “path”, we think it is possible to choose it in such a way that indices are cheaply comparable. That has been the case for all of the indices required to implement the standard library, and a few others we investigated while researching this change.

It's worth noting that this requirement isn't strictly necessary. Without it, though, indices would have no requirements beyond Equatable, and creation of a Range<T> would have to be allowed for any T conforming to Equatable. As a consequence, most interesting range operations, such as containment checks, would be unavailable unless T were also Comparable, and we'd be unable to provide bounds-checking in the general case.

That said, the requirement has real benefits. For example, it allows us to support distance measurement between arbitrary indices, even in collections without random access traversal. In the old model, x.distance(to: y) for these collections had the undetectable precondition that x precede y, with unpredictable consequences for violation in the general case.

Detailed API Changes

This section describes changes to methods, properties, and associated types at a high level. Details related to working around compiler limitations have been omitted. For a complete view of the code and documentation changes implementing this proposal, please see this pull request.

Collections

The following APIs were added:

protocol Collection {
  ...
  /// A type that can represent the number of steps between pairs of
  /// `Index` values where one value is reachable from the other.
  ///
  /// Reachability is defined by the ability to produce one value from
  /// the other via zero or more applications of `index(after: _)`.
  associatedtype IndexDistance : SignedInteger = Int

  /// A collection type whose elements are the indices of `self` that
  /// are valid for subscripting, in ascending order.
  associatedtype Indices : Collection = DefaultIndices<Self>

  /// The indices that are valid for subscripting `self`, in ascending order.
  ///
  /// - Note: `indices` can hold a strong reference to the collection itself,
  ///   causing the collection to be non-uniquely referenced.  If you need to
  ///   mutate the collection while iterating over its indices, use the
  ///   `index(after: _)` method starting with `startIndex` to produce indices
  ///   instead.
  /// 
  ///   ```
  ///   var c = [10, 20, 30, 40, 50]
  ///   var i = c.startIndex
  ///   while i != c.endIndex {
  ///       c[i] /= 5
  ///       i = c.index(after: i)
  ///   }
  ///   // c == [2, 4, 6, 8, 10]
  ///   ```
  var indices: Indices { get }

  /// Returns the position immediately after `i`.
  ///
  /// - Precondition: `(startIndex..<endIndex).contains(i)`
  @warn_unused_result
  func index(after i: Index) -> Index

  /// Replaces `i` with its successor.
  func formIndex(after i: inout Index)

  /// Returns the result of advancing `i` by `n` positions.
  ///
  /// - Returns:
  ///   - If `n > 0`, the `n`th index after `i`.
  ///   - If `n < 0`, the `n`th index before `i`.
  ///   - Otherwise, `i` unmodified.
  ///
  /// - Precondition: `n >= 0` unless `Self` conforms to
  ///   `BidirectionalCollection`.
  /// - Precondition:
  ///   - If `n > 0`, `n <= self.distance(from: i, to: self.endIndex)`
  ///   - If `n < 0`, `n >= self.distance(from: i, to: self.startIndex)`
  ///
  /// - Complexity:
  ///   - O(1) if `Self` conforms to `RandomAccessCollection`.
  ///   - O(`abs(n)`) otherwise.
  func index(_ i: Index, offsetBy n: IndexDistance) -> Index

  /// Returns the result of advancing `i` by `n` positions, or until it
  /// equals `limit`.
  ///
  /// - Returns:
  ///   - If `n > 0`, the `n`th index after `i` or `limit`, whichever
  ///     is reached first.
  ///   - If `n < 0`, the `n`th index before `i` or `limit`, whichever
  ///     is reached first.
  ///   - Otherwise, `i` unmodified.
  ///
  /// - Precondition: `n >= 0` unless `Self` conforms to
  ///   `BidirectionalCollection`.
  ///
  /// - Complexity:
  ///   - O(1) if `Self` conforms to `RandomAccessCollection`.
  ///   - O(`abs(n)`) otherwise.
  func index(
    _ i: Index, offsetBy n: IndexDistance, limitedBy limit: Index) -> Index

  /// Advances `i` by `n` positions.
  ///
  /// - Precondition: `n >= 0` unless `Self` conforms to
  ///   `BidirectionalCollection`.
  /// - Precondition:
  ///   - If `n > 0`, `n <= self.distance(from: i, to: self.endIndex)`
  ///   - If `n < 0`, `n >= self.distance(from: i, to: self.startIndex)`
  ///
  /// - Complexity:
  ///   - O(1) if `Self` conforms to `RandomAccessCollection`.
  ///   - O(`abs(n)`) otherwise.
  func formIndex(_ i: inout Index, offsetBy n: IndexDistance)

  /// Advances `i` by `n` positions, or until it equals `limit`.
  ///
  /// - Precondition: `n >= 0` unless `Self` conforms to
  ///   `BidirectionalCollection`.
  ///
  /// - Complexity:
  ///   - O(1) if `Self` conforms to `RandomAccessCollection`.
  ///   - O(`abs(n)`) otherwise.
  func formIndex(
    _ i: inout Index, offsetBy n: IndexDistance, limitedBy limit: Index)

  /// Returns the distance between `start` and `end`.
  ///
  /// - Precondition: `start <= end` unless `Self` conforms to
  ///   `BidirectionalCollection`.
  /// - Complexity:
  ///   - O(1) if `Self` conforms to `RandomAccessCollection`.
  ///   - O(`n`) otherwise, where `n` is the method's result.
  func distance(from start: Index, to end: Index) -> IndexDistance
}

protocol BidirectionalCollection {
  /// Returns the position immediately preceding `i`.
  ///
  /// - Precondition: `i > startIndex && i <= endIndex` 
  func index(before i: Index) -> Index

  /// Replaces `i` with its predecessor.
  ///
  /// - Precondition: `i > startIndex && i <= endIndex`
  func formIndex(before i: inout Index)
}

Note:

  • The formIndex overloads essentially enshrine the previously-hidden _successorInPlace et al., which can be important for performance when handling the rare heavyweight index type such as AnyIndex.

  • RandomAccessCollection does not add any syntactic requirements beyond those of BidirectionalCollection. Instead, it places tighter performance bounds on operations such as c.index(i, offsetBy: n) (O(1) instead of O(n)).

Ranges

The four range Range Types share the common interface shown below:

public struct Range<Bound: Comparable> : Equatable {
  
  /// Creates an instance with the given bounds.
  ///
  /// - Note: As this initializer does not check its precondition, it
  ///   should be used as an optimization only, when one is absolutely
  ///   certain that `lower <= upper`.  In general, the `..<` and `...`
  ///   operators are to be preferred for forming ranges.
  ///
  /// - Precondition: `lower <= upper`
  init(uncheckedBounds: (lower: Bound, upper: Bound))

  /// Returns `true` if the range contains the `value`.
  func contains(_ value: Bound) -> Bool
  
  /// Returns `true` iff `self` and `other` contain a value in common.
  func overlaps(_ other: Self) -> Bool

  /// Returns `true` iff `self.contains(x)` is `false` for all values of `x`.
  var isEmpty: Bool { get }
  
  /// The range's lower bound.
  var lowerBound: Bound { get }
  
  /// The range's upper bound.
  var upperBound: Bound { get }
  
  /// Returns `self` clamped to `limits`.
  ///
  /// The bounds of the result, even if it is empty, are always
  /// limited to the bounds of `limits`.
  func clamped(to limits: Self) -> Self
}

In addition, every implementable lossless conversion between range types is provided as a label-less init with one argument:

let a = 1..<10
let b = ClosedRange(a) // <=== Here

Note in particular:

  • In Range<T>, T is Comparable rather than an index type that can be advanced, so a generalized range is no longer a Collection, and startIndex/endIndex have become lowerBound/upperBound.
  • The semantic order of Interval's clamp method, which was unclear at its use-site, has been inverted and updated with a preposition for clarity.

Downsides

The proposed approach has several disadvantages, which we explore here in the interest of full disclosure:

  • In Swift 2, RandomAccessIndex has operations like + that provide easy access to arbitrary position offsets in some collections. That could also be seen as discouragement from trying to do random access operations with less-refined index protocols, because in those cases one has to resort to constructs like i.advancedBy(n). In this proposal, there is only c.index(i, offsetBy: n), which makes random access equally (in)convenient for all collections, and there is no particular syntactic penalty for doing things that might turn out to be inefficient.

  • Index movement is more complex in principle, since it now involves not only the index, but the collection as well. The impact of this complexity is limited somewhat because it's very common that code moving indices occurs in a method of the collection type, where “implicit self” kicks in. The net result is that index manipulations end up looking like free function calls:

    let j = index(after: i)           // self.index(after: i)
    let k = index(j, offsetBy: 5)     // self.index(j, offsetBy: 5)
  • The new index manipulation methods increase the API surface area of Collection, which is already quite large since algorithms are implemented as extensions.

  • Because Swift is unable to express conditional protocol conformances, implementing this change has required us to create a great deal of complexity in the standard library API. Aside from the two excess “Countable” range types, there are new overloads for slicing and twelve distinct slice types that capture all the combinations of traversal, mutability, and range-replaceability. While these costs are probably temporary, they are very real in the meantime.

  • The API complexity mentioned above stresses the type checker, requiring several changes just to get our test code to type-check in reasonable time. Again, an ostensibly temporary—but still real—cost.

Impact on existing code

Code that does not need to change:

  • Code that works with Array, ArraySlice, ContiguousArray, and their indices.

  • Code that operates on arbitrary collections and indices (on concrete instances or in generic context), but does no index traversal.

  • Iteration over collections' indices with c.indices does not change.

  • APIs of high-level collection algorithms don't change, even for algorithms that accept indices as parameters or return indices (e.g., index(of:), min(), sort(), prefix(), prefix(upTo:) etc.)

Code that needs to change:

  • Code that advances indices (i.successor(), i.predecessor(), i.advanced(by:) etc.) or calculates distances between indices (i.distance(to:)) now needs to call a method on the collection instead.

    // Before:
    var i = c.index { $0 % 2 == 0 }
    let j = i.successor()
    print(c[j])
    
    // After:
    var i = c.index { $0 % 2 == 0 }   // No change in algorithm API.
    let j = c.index(after: i)         // Advancing an index requires a collection instance.
    print(c[j])                       // No change in subscripting.

    The transformation from i.successor() to c.index(after: i) is non-trivial. Performing it correctly requires knowing how to get the corresponding collection. In general, it is not possible to perform this migration automatically. A very sophisticated migrator could handle some easy cases.

  • Custom collection implementations need to change. A simple fix would be to just move the methods from indices to collections to satisfy new protocol requirements. This is a more or less mechanical fix that does not require design work. This fix would allow the code to compile and run.

    In order to take advantage of performance improvements in the new model, and remove reference-counted stored properties from indices, the representation of the index might need to be redesigned.

    Implementing custom collections, as compared to using collections, is a niche case. We believe that for custom collection types it is sufficient to provide clear steps for manual migration that don't require a redesign. Implementing this in an automated migrator might be possible, but would be a heroic migration for a rare case.

Implementation Status

This pull request contains a complete implementation.

Alternatives considered

We considered index-less models, for example, D's std.range (see also On Iteration by Andrei Alexandrescu). Ranges work well for reference-typed collections, but it is not clear how to adjust the concept of D's range (similar to Slice in Swift) for mutable value-typed collections. In D, you process a collection by repeatedly slicing off elements. Once you have found an element that you would like to mutate, it is not clear how to actually change the original collection, if the collection and its slice are value types.


[^1]: Dictionary and Set use a double-indirection trick to avoid disturbing the reference count of the storage with indices.

    +--+    class                       struct
    |RC|---------+          +-----------------+
    +--+ Storage |<---------| DictionaryIndex |
      |          |          | value           |
      +----------+          +-----------------+
          ^
    +--+  |     class                struct
    |RC|-------------+        +------------+
    +--+ Indirection |<-------| Dictionary |
      |  ("owner")   |        | value      |
      +--------------+        +------------+

Instances of Dictionary point to an indirection, while instances of DictionaryIndex point to the storage itself. This allows us to have two separate reference counts. One of the refcounts tracks just the live Dictionary instances, which allows us to perform precise uniqueness checks.

The issue that we were previously unaware of is that this scheme is not thread-safe. When uniquely-referenced storage is being mutated in place, indices can be concurrently being incremented (on a different thread). This would be a read/write data race.

Fixing this data race (to provide memory safety) would require locking dictionary storage on every access, which would be an unacceptable performance penalty.