Merge pull request #10002 from milseman/i_dream_of_grapheme

[stdlib] Grapheme break fast-paths for Cyrillic, Arabic, Hangul

Author: milseman

benchmark/single-source/StringWalk.swift

@@ -69,6 +69,7 @@ let utf16 = emoji + "the quick brown fox" + String(emoji.reversed() as Array<Cha
 let japanese = "今回のアップデートでSwiftに大幅な改良が施され、安定していてしかも直感的に使うことができるAppleプラットフォーム向けプログラミング言語になりました。"
 let chinese = "Swift 是面向 Apple 平台的编程语言，功能强大且直观易用，而本次更新对其进行了全面优化。"
 let korean = "이번 업데이트에서는 강력하면서도 직관적인 Apple 플랫폼용 프로그래밍 언어인 Swift를 완벽히 개선하였습니다."
+let russian = "в чащах юга жил-был цитрус? да, но фальшивый экземпляр"
 
 // A workload that's mostly Latin characters, with occasional emoji
 // interspersed. Common for tweets.
@@ -269,3 +270,33 @@ public func run_StringWalk_korean_characters_Backwards(_ N: Int) {
   }
 }
 
+
+@inline(never)
+public func run_StringWalk_russian_unicodeScalars(_ N: Int) {
+  for _ in 1...unicodeScalarsMultiplier*N {
+    count_unicodeScalars(russian.unicodeScalars)
+  }
+}
+
+@inline(never)
+public func run_StringWalk_russian_unicodeScalars_Backwards(_ N: Int) {
+  for _ in 1...unicodeScalarsMultiplier*N {
+    count_unicodeScalars_rev(russian.unicodeScalars.reversed())
+  }
+}
+
+
+@inline(never)
+public func run_StringWalk_russian_characters(_ N: Int) {
+  for _ in 1...charactersMultiplier*N {
+    count_characters(russian.characters)
+  }
+}
+
+@inline(never)
+public func run_StringWalk_russian_characters_Backwards(_ N: Int) {
+  for _ in 1...charactersMultiplier*N {
+    count_characters_rev(russian.characters.reversed())
+  }
+}
+

benchmark/single-source/StringWalk.swift.gyb

@@ -70,6 +70,7 @@ let utf16 = emoji + "the quick brown fox" + String(emoji.reversed() as Array<Cha
 let japanese = "今回のアップデートでSwiftに大幅な改良が施され、安定していてしかも直感的に使うことができるAppleプラットフォーム向けプログラミング言語になりました。"
 let chinese = "Swift 是面向 Apple 平台的编程语言，功能强大且直观易用，而本次更新对其进行了全面优化。"
 let korean = "이번 업데이트에서는 강력하면서도 직관적인 Apple 플랫폼용 프로그래밍 언어인 Swift를 완벽히 개선하였습니다."
+let russian = "в чащах юга жил-был цитрус? да, но фальшивый экземпляр"
 
 // A workload that's mostly Latin characters, with occasional emoji
 // interspersed. Common for tweets.
@@ -90,7 +91,7 @@ let baseMultiplier = 10_000
 let unicodeScalarsMultiplier = baseMultiplier
 let charactersMultiplier = baseMultiplier / 5
 
-% for Name in ["ascii", "utf16", "tweet", "japanese", "chinese", "korean"]:
+% for Name in ["ascii", "utf16", "tweet", "japanese", "chinese", "korean", "russian"]:
 %   for Kind in ["unicodeScalars", "characters"]:
 
 @inline(never)

benchmark/utils/main.swift

@@ -531,6 +531,10 @@ addTo(&stringTests, "StringWalk_korean_characters", run_StringWalk_korean_charac
 addTo(&stringTests, "StringWalk_korean_characters_Backwards", run_StringWalk_korean_characters_Backwards)
 addTo(&stringTests, "StringWalk_korean_unicodeScalars", run_StringWalk_korean_unicodeScalars)
 addTo(&stringTests, "StringWalk_korean_unicodeScalars_Backwards", run_StringWalk_korean_unicodeScalars_Backwards)
+addTo(&stringTests, "StringWalk_russian_characters", run_StringWalk_russian_characters)
+addTo(&stringTests, "StringWalk_russian_characters_Backwards", run_StringWalk_russian_characters_Backwards)
+addTo(&stringTests, "StringWalk_russian_unicodeScalars", run_StringWalk_russian_unicodeScalars)
+addTo(&stringTests, "StringWalk_russian_unicodeScalars_Backwards", run_StringWalk_russian_unicodeScalars_Backwards)
 addTo(&stringTests, "StringWalk_tweet_characters", run_StringWalk_tweet_characters)
 addTo(&stringTests, "StringWalk_tweet_characters_Backwards", run_StringWalk_tweet_characters_Backwards)
 addTo(&stringTests, "StringWalk_tweet_unicodeScalars", run_StringWalk_tweet_unicodeScalars)

stdlib/public/core/StringCharacterView.swift

@@ -297,33 +297,55 @@ extension String.CharacterView : BidirectionalCollection {
   internal static func _internalExtraCheckGraphemeBreakBetween(
     _ lhs: UInt16, _ rhs: UInt16
   ) -> Bool {
+    _sanityCheck(
+      lhs != _CR || rhs != _LF,
+      "CR-LF special case handled by _quickCheckGraphemeBreakBetween")
+
     // Whether the given scalar, when it appears paired with another scalar
     // satisfying this property, has a grapheme break between it and the other
     // scalar.
     func hasBreakWhenPaired(_ x: UInt16) -> Bool {
-      // TODO: This doesn't generate optimal code, tune/re-write at a lower level.
-
+      // TODO: This doesn't generate optimal code, tune/re-write at a lower
+      // level.
+      //
+      // NOTE: Order of case ranges affects codegen, and thus performance. All
+      // things being equal, keep existing order below.
+      switch x {
       // Unified CJK Han ideographs, common and some supplemental, amongst
       // others:
       //   0x3400-0xA4CF
-      if 0x3400 <= x && x <= 0xa4cf {
-        return true
-      }
+      case 0x3400...0xa4cf: return true
+      // TODO: CJK punctuation
 
+      // Repeat sub-300 check, this is beneficial for common cases of Latin
+      // characters embedded within non-Latin script (e.g. newlines, spaces,
+      // proper nouns and/or jargon, punctuation).
       //
+      // NOTE: CR-LF special case has already been checked.
+      case 0x0000...0x02ff: return true
+
+      // TODO: general punctuation
+
       // Non-combining kana:
       //   0x3041-0x3096
       //   0x30A1-0x30FA
-      //
-      // TODO: may be faster to verify whether only 3099 and 309A don't have
-      // this property, and compare not-equal rather than using two ranges.
-      if 0x3041 <= x && x <= 0x3096 || 0x30a1 <= x && x <= 0x30fa {
-        return true
-      }
+      case 0x3041...0x3096: return true
+      case 0x30a1...0x30fa: return true
+
+      // Non-combining modern (and some archaic) Cyrillic:
+      //   0x0400-0x0482 (first half of Cyrillic block)
+      case 0x0400...0x0482: return true
 
-      // TODO: sub-300 check would also be valuable, e.g. when breaking at the
-      // boundary between English embedded in Chinese.
-      return false
+      // Modern Arabic, excluding extenders and prependers:
+      //   0x061D-0x064A
+      case 0x061d...0x064a: return true
+
+      // Precomposed Hangul syllables:
+      //   0xAC00–0xD7AF
+      case 0xac00...0xd7af: return true
+
+      default: return false
+      }
     }
     return hasBreakWhenPaired(lhs) && hasBreakWhenPaired(rhs)
   }