swiftlang/swift

[SR-10667] Optimise chained SubSequence producing methods into fewer calls

Open

#53 066 ouverte le 11 mai 2019

Voir sur GitHub
 (1 commentaire) (0 réactions) (0 assignés)Swift (10 719 forks)batch import
SILGencompilergood first issueimprovementperformance

Métriques du dépôt

Stars
 (69 989 stars)
Métriques de merge PR
 (Merge moyen 8j 17h) (510 PRs mergées en 30 j)

Description

Previous ID SR-10667
Radar None
Original Reporter bnut (JIRA User)
Type Improvement
Status In Progress
Resolution
Votes 1
Component/s Compiler
Labels Improvement, Performance, SILGen, StarterBug
Assignee mkita (JIRA)
Priority Medium

md5: 3d2eff2a03f31963319181f67fe2b19a

Issue Description:

This code (on BidirectionalCollection) is nice and concise, but it's not efficient:

self.suffix(2).dropLast().last

It could be optimised by the compiler into something like this:

index(endIndex, offsetBy: -2, limitedBy: startIndex).map { self[$0] }

At @jckarter's suggestion this could be implemented with @_semantics similar to some existing optimisations on Array and String. An example of this is here: https://github.com/apple/swift/pull/5978.

There's a lot of permutations for suffix, prefix, dropFirst, dropLast, first, last, etc. so adding a special case for every one would be excessive. Ideally there is a more generic approach to pattern match and reduce these kinds of expressions. This might be done by having multiple small patterns, that can be combined to reduce the chain of calls iteratively.

Here are some ideas for further optimisations:

// Sequence:
suffix(m).suffix(n) -> suffix(Swift.min(m, n))
prefix(m).prefix(n) -> prefix(Swift.min(m, n))
dropFirst(m).dropFirst(n) -> dropFirst(m + n)
dropLast(m).dropLast(n) -> dropLast(m + n)

prefix(m).dropFirst(n) -> prefix(m - Swift.min(n,m))
suffix(m).dropFirst(n) -> suffix(m - Swift.min(n,m))

filter(where: predicate).first -> first(where: predicate)

// Collection
prefix(n).first -> first
suffix(n).last -> last

prefix(n).last  -> index(startIndex, offsetBy: n, limitedBy: endIndex).map { self[$0] }

// BidirectionalCollection
suffix(n).first -> index(endIndex, offsetBy: -n, limitedBy: startIndex).map { self[$0] }

Guide contributeur