[SR-9184] UTF-8 String Regressions

[milseman]

Previous ID	SR-9184
Radar	None
Original Reporter	@milseman
Type	Task

Attachment: Download

Additional Detail from JIRA

Votes	0
Component/s	Standard Library
Labels	Task, Performance
Assignee	@milseman
Priority	Medium

md5: 5e76fa6a2d0620628262e8b5de6feea2

Sub-Tasks:

• SR-9192 Flatten/rework normalized code unit iterator
• SR-9212 String comparison binary-divergence fast-path not triggering
• SR-9270 Tweak String comparison inlinability
• SR-9272 Potential invalid benchmarks (used to skip allocations)
• SR-9287 Poor codegen for String iteration
• SR-9427 CharacterLiteralsSmall: Failure to constant-fold release

Issue Description:

UTF-8 String comes a lot of wins, but some current regressions:

regressions.md

This is an ☂️ bug, see Sub-Tasks below.

Current status of this set of regressions. Note this is only a listing of what were regressions when the PR was merged; omitted benchmarks exhibited improvement:

• To fix in 5.0:

  • Flattened normalized iterator (https://bugs.swift.org/browse/SR-9192):

    • StringComparison_abnormal

    • StringHashing_abnormal

    • StringHashing_zalgo

  • MapReduceString (https://bugs.swift.org/browse/SR-9460)

• 5.0 currently has parity with 4.2 performance:

  • StringFromLongWholeSubstringGeneric

  • StringComparison_fastPrenormal

  • WordCountUniqueASCII

  • RemoveWhereQuadraticString

  • StringWordBuilder

  • StringHashing_ascii

• 5.0 is faster than 4.2:

  • SubstringComparble

  • StringFromLongWholeSubstring

  • StringComparison_latin1

  • StringWalk

  • StringEqualPointerComparison

  • Dictionary3

  • StringWordBuilderReservingCapacity

  • StringHasPrefixUnicode

  • Dictionary

  • RemoveWhereFilterString

  • StringInterpolationManySmallSegments

  • StringMatch

  • Join

  • StringBuilderLong

  • RomanNumbers

• Resolved for Swift 5.0:

  • Chars2: We're happy with the size/speed tradeoff

    • While still slower than 4.2, this is now is significantly smaller in code size (~7x smaller)

• Post Swift 5.0:

  • Native Grapheme Breaking (https://bugs.swift.org/browse/SR-9423)

    • StringHasSuffixUnicode

    • CSVParsing2 (improved)

    • CharacterPropertiesStashedMemo

  • StrComplexWalk (https://bugs.swift.org/browse/SR-9424)

  • OpenClose (improved, further is https://bugs.swift.org/browse/SR-9425)

  • CharacterLiteralsSmall (https://bugs.swift.org/browse/SR-9427)

  • SortStringsUnicode (improved, further is https://bugs.swift.org/browse/SR-9432)

  • Dictionary2 (https://bugs.swift.org/browse/SR-9438)

  • Calculator (improved, further is https://bugs.swift.org/browse/SR-9440)

Every single one of these extracted and merged into a single script file normalized to around 1 second each: https://gist.github.com/milseman/737d74fd4a6817432b3b5f18c34d416d

[milseman]

Ok, so JIRA seems to just drop all my carefully set hyperlinks whenever I edit the description...

[palimondo]

@milseman Re: "Every single one of these …"
I see these contain stuff from StringWalk.swift (CharIndexing*, CharIteration* and StringWalk* families). I was about to enable all of these benchmarks (analogous to Existential Redux) of these after adjusting for correct fast runtimes and renaming to match proposed naming convention. Can I do that, or are you already working on these?

[milseman]

Go ahead! I may merge several of them together in the future; as we add more fast-paths they all start to measure the same fast-paths.

[milseman]

CC @airspeedswift @eeckstein, as I'm using this to track the remaining String regressions.