Optimize case conversion with icu_case_mapping

Use FastAsciiConvert (as used by Unibrow) for i18n-aware
case conversion with --icu_case_mapping.

Move FastAsciiConvert to src/string-case.cc so that it can be used
by both runtime-{string,i18n}.

Add more tests.

BUG=v8:4477,v8:4476
TEST=intl/general/case*

Review-Url: https://codereview.chromium.org/2533983006
Cr-Commit-Position: refs/heads/master@{#41821}

BUILD.gn

@@ -1691,6 +1691,8 @@ v8_source_set("v8_base") {
     "src/startup-data-util.h",
     "src/string-builder.cc",
     "src/string-builder.h",
+    "src/string-case.cc",
+    "src/string-case.h",
     "src/string-search.h",
     "src/string-stream.cc",
     "src/string-stream.h",

src/js/i18n.js

@@ -2121,27 +2121,16 @@ OverrideFunction(GlobalString.prototype, 'normalize', function() {
 );
 
 function ToLowerCaseI18N() {
-  if (!IS_UNDEFINED(new.target)) {
-    throw %make_type_error(kOrdinaryFunctionCalledAsConstructor);
-  }
   CHECK_OBJECT_COERCIBLE(this, "String.prototype.toLowerCase");
-  var s = TO_STRING(this);
-  return %StringToLowerCaseI18N(s);
+  return %StringToLowerCaseI18N(TO_STRING(this));
 }
 
 function ToUpperCaseI18N() {
-  if (!IS_UNDEFINED(new.target)) {
-    throw %make_type_error(kOrdinaryFunctionCalledAsConstructor);
-  }
   CHECK_OBJECT_COERCIBLE(this, "String.prototype.toUpperCase");
-  var s = TO_STRING(this);
-  return %StringToUpperCaseI18N(s);
+  return %StringToUpperCaseI18N(TO_STRING(this));
 }
 
 function ToLocaleLowerCaseI18N(locales) {
-  if (!IS_UNDEFINED(new.target)) {
-    throw %make_type_error(kOrdinaryFunctionCalledAsConstructor);
-  }
   CHECK_OBJECT_COERCIBLE(this, "String.prototype.toLocaleLowerCase");
   return LocaleConvertCase(TO_STRING(this), locales, false);
 }
@@ -2149,9 +2138,6 @@ function ToLocaleLowerCaseI18N(locales) {
 %FunctionSetLength(ToLocaleLowerCaseI18N, 0);
 
 function ToLocaleUpperCaseI18N(locales) {
-  if (!IS_UNDEFINED(new.target)) {
-    throw %make_type_error(kOrdinaryFunctionCalledAsConstructor);
-  }
   CHECK_OBJECT_COERCIBLE(this, "String.prototype.toLocaleUpperCase");
   return LocaleConvertCase(TO_STRING(this), locales, true);
 }

src/runtime/runtime-i18n.cc

@@ -8,13 +8,15 @@
 
 #include <memory>
 
-#include "src/api.h"
 #include "src/api-natives.h"
+#include "src/api.h"
 #include "src/arguments.h"
 #include "src/factory.h"
 #include "src/i18n.h"
 #include "src/isolate-inl.h"
 #include "src/messages.h"
+#include "src/string-case.h"
+#include "src/utils.h"
 
 #include "unicode/brkiter.h"
 #include "unicode/calendar.h"
@@ -1041,15 +1043,14 @@ bool ToUpperFastASCII(const Vector<const Char>& src,
 const uint16_t sharp_s = 0xDF;
 
 template <typename Char>
-bool ToUpperOneByte(const Vector<const Char>& src,
-                    Handle<SeqOneByteString> result, int* sharp_s_count) {
+bool ToUpperOneByte(const Vector<const Char>& src, uint8_t* dest,
+                    int* sharp_s_count) {
   // Still pretty-fast path for the input with non-ASCII Latin-1 characters.
 
   // There are two special cases.
   //  1. U+00B5 and U+00FF are mapped to a character beyond U+00FF.
   //  2. Lower case sharp-S converts to "SS" (two characters)
   *sharp_s_count = 0;
-  int32_t index = 0;
   for (auto it = src.begin(); it != src.end(); ++it) {
     uint16_t ch = static_cast<uint16_t>(*it);
     if (V8_UNLIKELY(ch == sharp_s)) {
@@ -1061,7 +1062,7 @@ bool ToUpperOneByte(const Vector<const Char>& src,
       // need to take the 16-bit path.
       return false;
     }
-    result->SeqOneByteStringSet(index++, ToLatin1Upper(ch));
+    *dest++ = ToLatin1Upper(ch);
   }
 
   return true;
@@ -1082,6 +1083,16 @@ void ToUpperWithSharpS(const Vector<const Char>& src,
   }
 }
 
+inline int FindFirstUpperOrNonAscii(Handle<String> s, int length) {
+  for (int index = 0; index < length; ++index) {
+    uint16_t ch = s->Get(index);
+    if (V8_UNLIKELY(IsASCIIUpper(ch) || ch & ~0x7F)) {
+      return index;
+    }
+  }
+  return length;
+}
+
 }  // namespace
 
 RUNTIME_FUNCTION(Runtime_StringToLowerCaseI18N) {
@@ -1091,60 +1102,65 @@ RUNTIME_FUNCTION(Runtime_StringToLowerCaseI18N) {
 
   int length = s->length();
   s = String::Flatten(s);
-  // First scan the string for uppercase and non-ASCII characters:
-  if (s->HasOnlyOneByteChars()) {
-    int first_index_to_lower = length;
-    for (int index = 0; index < length; ++index) {
-      // Blink specializes this path for one-byte strings, so it
-      // does not need to do a generic get, but can do the equivalent
-      // of SeqOneByteStringGet.
-      uint16_t ch = s->Get(index);
-      if (V8_UNLIKELY(IsASCIIUpper(ch) || ch & ~0x7F)) {
-        first_index_to_lower = index;
-        break;
-      }
-    }
 
+  if (!s->HasOnlyOneByteChars()) {
+    // Use a slower implementation for strings with characters beyond U+00FF.
+    return LocaleConvertCase(s, isolate, false, "");
+  }
+
+  // We depend here on the invariant that the length of a Latin1
+  // string is invariant under ToLowerCase, and the result always
+  // fits in the Latin1 range in the *root locale*. It does not hold
+  // for ToUpperCase even in the root locale.
+
+  // Scan the string for uppercase and non-ASCII characters for strings
+  // shorter than a machine-word without any memory allocation overhead.
+  // TODO(jshin): Apply this to a longer input by breaking FastAsciiConvert()
+  // to two parts, one for scanning the prefix with no change and the other for
+  // handling ASCII-only characters.
+  int index_to_first_unprocessed = length;
+  const bool is_short = length < static_cast<int>(sizeof(uintptr_t));
+  if (is_short) {
+    index_to_first_unprocessed = FindFirstUpperOrNonAscii(s, length);
     // Nothing to do if the string is all ASCII with no uppercase.
-    if (first_index_to_lower == length) return *s;
+    if (index_to_first_unprocessed == length) return *s;
+  }
 
-    // We depend here on the invariant that the length of a Latin1
-    // string is invariant under ToLowerCase, and the result always
-    // fits in the Latin1 range in the *root locale*. It does not hold
-    // for ToUpperCase even in the root locale.
-    Handle<SeqOneByteString> result;
-    ASSIGN_RETURN_FAILURE_ON_EXCEPTION(
-        isolate, result, isolate->factory()->NewRawOneByteString(length));
+  Handle<SeqOneByteString> result =
+      isolate->factory()->NewRawOneByteString(length).ToHandleChecked();
 
-    DisallowHeapAllocation no_gc;
-    String::FlatContent flat = s->GetFlatContent();
-    if (flat.IsOneByte()) {
-      const uint8_t* src = flat.ToOneByteVector().start();
-      CopyChars(result->GetChars(), src,
-                static_cast<size_t>(first_index_to_lower));
-      for (int index = first_index_to_lower; index < length; ++index) {
-        uint16_t ch = static_cast<uint16_t>(src[index]);
-        result->SeqOneByteStringSet(index, ToLatin1Lower(ch));
-      }
-    } else {
-      const uint16_t* src = flat.ToUC16Vector().start();
-      CopyChars(result->GetChars(), src,
-                static_cast<size_t>(first_index_to_lower));
-      for (int index = first_index_to_lower; index < length; ++index) {
-        uint16_t ch = src[index];
-        result->SeqOneByteStringSet(index, ToLatin1Lower(ch));
-      }
+  DisallowHeapAllocation no_gc;
+  String::FlatContent flat = s->GetFlatContent();
+  uint8_t* dest = result->GetChars();
+  if (flat.IsOneByte()) {
+    const uint8_t* src = flat.ToOneByteVector().start();
+    bool has_changed_character = false;
+    index_to_first_unprocessed = FastAsciiConvert<true>(
+        reinterpret_cast<char*>(dest), reinterpret_cast<const char*>(src),
+        length, &has_changed_character);
+    // If not ASCII, we keep the result up to index_to_first_unprocessed and
+    // process the rest.
+    if (index_to_first_unprocessed == length)
+      return has_changed_character ? *result : *s;
+
+    for (int index = index_to_first_unprocessed; index < length; ++index) {
+      dest[index] = ToLatin1Lower(static_cast<uint16_t>(src[index]));
+    }
+  } else {
+    if (index_to_first_unprocessed == length) {
+      DCHECK(!is_short);
+      index_to_first_unprocessed = FindFirstUpperOrNonAscii(s, length);
+    }
+    // Nothing to do if the string is all ASCII with no uppercase.
+    if (index_to_first_unprocessed == length) return *s;
+    const uint16_t* src = flat.ToUC16Vector().start();
+    CopyChars(dest, src, index_to_first_unprocessed);
+    for (int index = index_to_first_unprocessed; index < length; ++index) {
+      dest[index] = ToLatin1Lower(static_cast<uint16_t>(src[index]));
     }
-
-    return *result;
   }
 
-  // Blink had an additional case here for ASCII 2-byte strings, but
-  // that is subsumed by the above code (assuming there isn't a false
-  // negative for HasOnlyOneByteChars).
-
-  // Do a slower implementation for cases that include non-ASCII characters.
-  return LocaleConvertCase(s, isolate, false, "");
+  return *result;
 }
 
 RUNTIME_FUNCTION(Runtime_StringToUpperCaseI18N) {
@@ -1152,35 +1168,38 @@ RUNTIME_FUNCTION(Runtime_StringToUpperCaseI18N) {
   DCHECK_EQ(args.length(), 1);
   CONVERT_ARG_HANDLE_CHECKED(String, s, 0);
 
-  // This function could be optimized for no-op cases the way lowercase
-  // counterpart is, but in empirical testing, few actual calls to upper()
-  // are no-ops. So, it wouldn't be worth the extra time for pre-scanning.
-
   int32_t length = s->length();
   s = String::Flatten(s);
 
   if (s->HasOnlyOneByteChars()) {
-    Handle<SeqOneByteString> result;
-    ASSIGN_RETURN_FAILURE_ON_EXCEPTION(
-        isolate, result, isolate->factory()->NewRawOneByteString(length));
+    Handle<SeqOneByteString> result =
+        isolate->factory()->NewRawOneByteString(length).ToHandleChecked();
 
     int sharp_s_count;
     bool is_result_single_byte;
     {
       DisallowHeapAllocation no_gc;
       String::FlatContent flat = s->GetFlatContent();
-      // If it was ok to slow down ASCII-only input slightly, ToUpperFastASCII
-      // could be removed  because ToUpperOneByte is pretty fast now (it
-      // does not call ICU API any more.).
+      uint8_t* dest = result->GetChars();
       if (flat.IsOneByte()) {
         Vector<const uint8_t> src = flat.ToOneByteVector();
-        if (ToUpperFastASCII(src, result)) return *result;
-        is_result_single_byte = ToUpperOneByte(src, result, &sharp_s_count);
+        bool has_changed_character = false;
+        int index_to_first_unprocessed =
+            FastAsciiConvert<false>(reinterpret_cast<char*>(result->GetChars()),
+                                    reinterpret_cast<const char*>(src.start()),
+                                    length, &has_changed_character);
+        if (index_to_first_unprocessed == length)
+          return has_changed_character ? *result : *s;
+        // If not ASCII, we keep the result up to index_to_first_unprocessed and
+        // process the rest.
+        is_result_single_byte =
+            ToUpperOneByte(src.SubVector(index_to_first_unprocessed, length),
+                           dest + index_to_first_unprocessed, &sharp_s_count);
       } else {
         DCHECK(flat.IsTwoByte());
         Vector<const uint16_t> src = flat.ToUC16Vector();
         if (ToUpperFastASCII(src, result)) return *result;
-        is_result_single_byte = ToUpperOneByte(src, result, &sharp_s_count);
+        is_result_single_byte = ToUpperOneByte(src, dest, &sharp_s_count);
       }
     }
 

src/runtime/runtime-strings.cc

@@ -7,6 +7,7 @@
 #include "src/arguments.h"
 #include "src/regexp/jsregexp-inl.h"
 #include "src/string-builder.h"
+#include "src/string-case.h"
 #include "src/string-search.h"
 
 namespace v8 {
@@ -694,122 +695,6 @@ MUST_USE_RESULT static Object* ConvertCaseHelper(
   }
 }
 
-
-static const uintptr_t kOneInEveryByte = kUintptrAllBitsSet / 0xFF;
-static const uintptr_t kAsciiMask = kOneInEveryByte << 7;
-
-// Given a word and two range boundaries returns a word with high bit
-// set in every byte iff the corresponding input byte was strictly in
-// the range (m, n). All the other bits in the result are cleared.
-// This function is only useful when it can be inlined and the
-// boundaries are statically known.
-// Requires: all bytes in the input word and the boundaries must be
-// ASCII (less than 0x7F).
-static inline uintptr_t AsciiRangeMask(uintptr_t w, char m, char n) {
-  // Use strict inequalities since in edge cases the function could be
-  // further simplified.
-  DCHECK(0 < m && m < n);
-  // Has high bit set in every w byte less than n.
-  uintptr_t tmp1 = kOneInEveryByte * (0x7F + n) - w;
-  // Has high bit set in every w byte greater than m.
-  uintptr_t tmp2 = w + kOneInEveryByte * (0x7F - m);
-  return (tmp1 & tmp2 & (kOneInEveryByte * 0x80));
-}
-
-
-#ifdef DEBUG
-static bool CheckFastAsciiConvert(char* dst, const char* src, int length,
-                                  bool changed, bool is_to_lower) {
-  bool expected_changed = false;
-  for (int i = 0; i < length; i++) {
-    if (dst[i] == src[i]) continue;
-    expected_changed = true;
-    if (is_to_lower) {
-      DCHECK('A' <= src[i] && src[i] <= 'Z');
-      DCHECK(dst[i] == src[i] + ('a' - 'A'));
-    } else {
-      DCHECK('a' <= src[i] && src[i] <= 'z');
-      DCHECK(dst[i] == src[i] - ('a' - 'A'));
-    }
-  }
-  return (expected_changed == changed);
-}
-#endif
-
-
-template <class Converter>
-static bool FastAsciiConvert(char* dst, const char* src, int length,
-                             bool* changed_out) {
-#ifdef DEBUG
-  char* saved_dst = dst;
-  const char* saved_src = src;
-#endif
-  DisallowHeapAllocation no_gc;
-  // We rely on the distance between upper and lower case letters
-  // being a known power of 2.
-  DCHECK('a' - 'A' == (1 << 5));
-  // Boundaries for the range of input characters than require conversion.
-  static const char lo = Converter::kIsToLower ? 'A' - 1 : 'a' - 1;
-  static const char hi = Converter::kIsToLower ? 'Z' + 1 : 'z' + 1;
-  bool changed = false;
-  uintptr_t or_acc = 0;
-  const char* const limit = src + length;
-
-  // dst is newly allocated and always aligned.
-  DCHECK(IsAligned(reinterpret_cast<intptr_t>(dst), sizeof(uintptr_t)));
-  // Only attempt processing one word at a time if src is also aligned.
-  if (IsAligned(reinterpret_cast<intptr_t>(src), sizeof(uintptr_t))) {
-    // Process the prefix of the input that requires no conversion one aligned
-    // (machine) word at a time.
-    while (src <= limit - sizeof(uintptr_t)) {
-      const uintptr_t w = *reinterpret_cast<const uintptr_t*>(src);
-      or_acc |= w;
-      if (AsciiRangeMask(w, lo, hi) != 0) {
-        changed = true;
-        break;
-      }
-      *reinterpret_cast<uintptr_t*>(dst) = w;
-      src += sizeof(uintptr_t);
-      dst += sizeof(uintptr_t);
-    }
-    // Process the remainder of the input performing conversion when
-    // required one word at a time.
-    while (src <= limit - sizeof(uintptr_t)) {
-      const uintptr_t w = *reinterpret_cast<const uintptr_t*>(src);
-      or_acc |= w;
-      uintptr_t m = AsciiRangeMask(w, lo, hi);
-      // The mask has high (7th) bit set in every byte that needs
-      // conversion and we know that the distance between cases is
-      // 1 << 5.
-      *reinterpret_cast<uintptr_t*>(dst) = w ^ (m >> 2);
-      src += sizeof(uintptr_t);
-      dst += sizeof(uintptr_t);
-    }
-  }
-  // Process the last few bytes of the input (or the whole input if
-  // unaligned access is not supported).
-  while (src < limit) {
-    char c = *src;
-    or_acc |= c;
-    if (lo < c && c < hi) {
-      c ^= (1 << 5);
-      changed = true;
-    }
-    *dst = c;
-    ++src;
-    ++dst;
-  }
-
-  if ((or_acc & kAsciiMask) != 0) return false;
-
-  DCHECK(CheckFastAsciiConvert(saved_dst, saved_src, length, changed,
-                               Converter::kIsToLower));
-
-  *changed_out = changed;
-  return true;
-}
-
-
 template <class Converter>
 MUST_USE_RESULT static Object* ConvertCase(
     Handle<String> s, Isolate* isolate,
@@ -833,12 +718,13 @@ MUST_USE_RESULT static Object* ConvertCase(
     String::FlatContent flat_content = s->GetFlatContent();
     DCHECK(flat_content.IsFlat());
     bool has_changed_character = false;
-    bool is_ascii = FastAsciiConvert<Converter>(
+    int index_to_first_unprocessed = FastAsciiConvert<Converter::kIsToLower>(
         reinterpret_cast<char*>(result->GetChars()),
         reinterpret_cast<const char*>(flat_content.ToOneByteVector().start()),
         length, &has_changed_character);
     // If not ASCII, we discard the result and take the 2 byte path.
-    if (is_ascii) return has_changed_character ? *result : *s;
+    if (index_to_first_unprocessed == length)
+      return has_changed_character ? *result : *s;
   }
 
   Handle<SeqString> result;  // Same length as input.
@@ -872,7 +758,6 @@ RUNTIME_FUNCTION(Runtime_StringToLowerCase) {
   return ConvertCase(s, isolate, isolate->runtime_state()->to_lower_mapping());
 }
 
-
 RUNTIME_FUNCTION(Runtime_StringToUpperCase) {
   HandleScope scope(isolate);
   DCHECK_EQ(args.length(), 1);

src/runtime/runtime-utils.h

@@ -6,6 +6,7 @@
 #define V8_RUNTIME_RUNTIME_UTILS_H_
 
 #include "src/base/logging.h"
+#include "src/globals.h"
 #include "src/runtime/runtime.h"
 
 namespace v8 {

src/string-case.cc

+// Copyright 2016 the V8 project authors. All rights reserved.
+// Use of this source code is governed by a BSD-style license that can be
+// found in the LICENSE file.
+
+#include "src/string-case.h"
+
+#include "src/assert-scope.h"
+#include "src/base/logging.h"
+#include "src/globals.h"
+#include "src/utils.h"
+
+namespace v8 {
+namespace internal {
+
+#ifdef DEBUG
+bool CheckFastAsciiConvert(char* dst, const char* src, int length, bool changed,
+                           bool is_to_lower) {
+  bool expected_changed = false;
+  for (int i = 0; i < length; i++) {
+    if (dst[i] == src[i]) continue;
+    expected_changed = true;
+    if (is_to_lower) {
+      DCHECK('A' <= src[i] && src[i] <= 'Z');
+      DCHECK(dst[i] == src[i] + ('a' - 'A'));
+    } else {
+      DCHECK('a' <= src[i] && src[i] <= 'z');
+      DCHECK(dst[i] == src[i] - ('a' - 'A'));
+    }
+  }
+  return (expected_changed == changed);
+}
+#endif
+
+const uintptr_t kOneInEveryByte = kUintptrAllBitsSet / 0xFF;
+const uintptr_t kAsciiMask = kOneInEveryByte << 7;
+
+// Given a word and two range boundaries returns a word with high bit
+// set in every byte iff the corresponding input byte was strictly in
+// the range (m, n). All the other bits in the result are cleared.
+// This function is only useful when it can be inlined and the
+// boundaries are statically known.
+// Requires: all bytes in the input word and the boundaries must be
+// ASCII (less than 0x7F).
+static inline uintptr_t AsciiRangeMask(uintptr_t w, char m, char n) {
+  // Use strict inequalities since in edge cases the function could be
+  // further simplified.
+  DCHECK(0 < m && m < n);
+  // Has high bit set in every w byte less than n.
+  uintptr_t tmp1 = kOneInEveryByte * (0x7F + n) - w;
+  // Has high bit set in every w byte greater than m.
+  uintptr_t tmp2 = w + kOneInEveryByte * (0x7F - m);
+  return (tmp1 & tmp2 & (kOneInEveryByte * 0x80));
+}
+
+template <bool is_lower>
+int FastAsciiConvert(char* dst, const char* src, int length,
+                     bool* changed_out) {
+#ifdef DEBUG
+  char* saved_dst = dst;
+#endif
+  const char* saved_src = src;
+  DisallowHeapAllocation no_gc;
+  // We rely on the distance between upper and lower case letters
+  // being a known power of 2.
+  DCHECK('a' - 'A' == (1 << 5));
+  // Boundaries for the range of input characters than require conversion.
+  static const char lo = is_lower ? 'A' - 1 : 'a' - 1;
+  static const char hi = is_lower ? 'Z' + 1 : 'z' + 1;
+  bool changed = false;
+  const char* const limit = src + length;
+
+  // dst is newly allocated and always aligned.
+  DCHECK(IsAligned(reinterpret_cast<intptr_t>(dst), sizeof(uintptr_t)));
+  // Only attempt processing one word at a time if src is also aligned.
+  if (IsAligned(reinterpret_cast<intptr_t>(src), sizeof(uintptr_t))) {
+    // Process the prefix of the input that requires no conversion one aligned
+    // (machine) word at a time.
+    while (src <= limit - sizeof(uintptr_t)) {
+      const uintptr_t w = *reinterpret_cast<const uintptr_t*>(src);
+      if ((w & kAsciiMask) != 0) return static_cast<int>(src - saved_src);
+      if (AsciiRangeMask(w, lo, hi) != 0) {
+        changed = true;
+        break;
+      }
+      *reinterpret_cast<uintptr_t*>(dst) = w;
+      src += sizeof(uintptr_t);
+      dst += sizeof(uintptr_t);
+    }
+    // Process the remainder of the input performing conversion when
+    // required one word at a time.
+    while (src <= limit - sizeof(uintptr_t)) {
+      const uintptr_t w = *reinterpret_cast<const uintptr_t*>(src);
+      if ((w & kAsciiMask) != 0) return static_cast<int>(src - saved_src);
+      uintptr_t m = AsciiRangeMask(w, lo, hi);
+      // The mask has high (7th) bit set in every byte that needs
+      // conversion and we know that the distance between cases is
+      // 1 << 5.
+      *reinterpret_cast<uintptr_t*>(dst) = w ^ (m >> 2);
+      src += sizeof(uintptr_t);
+      dst += sizeof(uintptr_t);
+    }
+  }
+  // Process the last few bytes of the input (or the whole input if
+  // unaligned access is not supported).
+  while (src < limit) {
+    char c = *src;
+    if ((c & kAsciiMask) != 0) return static_cast<int>(src - saved_src);
+    if (lo < c && c < hi) {
+      c ^= (1 << 5);
+      changed = true;
+    }
+    *dst = c;
+    ++src;
+    ++dst;
+  }
+
+  DCHECK(
+      CheckFastAsciiConvert(saved_dst, saved_src, length, changed, is_lower));
+
+  *changed_out = changed;
+  return length;
+}
+
+template int FastAsciiConvert<false>(char* dst, const char* src, int length,
+                                     bool* changed_out);
+template int FastAsciiConvert<true>(char* dst, const char* src, int length,
+                                    bool* changed_out);
+
+}  // namespace internal
+}  // namespace v8

src/string-case.h

+// Copyright 2016 the V8 project authors. All rights reserved.
+// Use of this source code is governed by a BSD-style license that can be
+// found in the LICENSE file.
+
+#ifndef V8_STRING_CASE_H_
+#define V8_STRING_CASE_H_
+
+namespace v8 {
+namespace internal {
+
+template <bool is_lower>
+int FastAsciiConvert(char* dst, const char* src, int length, bool* changed_out);
+
+}  // namespace internal
+}  // namespace v8
+
+#endif  // V8_STRING_CASE_H__

src/v8.gyp

@@ -1228,6 +1228,8 @@
         'startup-data-util.h',
         'string-builder.cc',
         'string-builder.h',
+        'string-case.cc',
+        'string-case.h',
         'string-search.h',
         'string-stream.cc',
         'string-stream.h',

test/intl/general/case-mapping.js

@@ -16,14 +16,33 @@ assertEquals("σς", "\u03A3\u03A3".toLowerCase());
 // Expand sharp s in latin1 fastpath
 assertEquals("ASSB", "A\u00DFB".toUpperCase());
 assertEquals("AB", "Ab".toUpperCase());
-// Find first upper case in fastpath
+// Find first uppercase in fastpath
+// Input length < a machine word size
+assertEquals("ab", "ab".toLowerCase());
 assertEquals("ab", "aB".toLowerCase());
 assertEquals("AÜ", "aü".toUpperCase());
 assertEquals("AÜ", "AÜ".toUpperCase());
 assertEquals("aü", "aü".toLowerCase());
+assertEquals("aü", "aÜ".toLowerCase());
 assertEquals("aü", "AÜ".toLowerCase());
 assertEquals("aü", "AÜ".toLowerCase());
 
+// Input length >= a machine word size
+assertEquals("abcdefghij", "abcdefghij".toLowerCase());
+assertEquals("abcdefghij", "abcdefghiJ".toLowerCase());
+assertEquals("abçdefghij", "abçdefghiJ".toLowerCase());
+assertEquals("abçdefghij", "abÇdefghiJ".toLowerCase());
+assertEquals("abcdefghiá", "abcdeFghiá".toLowerCase());
+assertEquals("abcdefghiá", "abcdeFghiÁ".toLowerCase());
+
+assertEquals("ABCDEFGHIJ", "ABCDEFGHIJ".toUpperCase());
+assertEquals("ABCDEFGHIJ", "ABCDEFGHIj".toUpperCase());
+assertEquals("ABÇDEFGHIJ", "ABÇDEFGHIj".toUpperCase());
+assertEquals("ABÇDEFGHIJ", "ABçDEFGHIj".toUpperCase());
+assertEquals("ABCDEFGHIÁ", "ABCDEfGHIÁ".toUpperCase());
+assertEquals("ABCDEFGHIÁ", "ABCDEfGHIá".toUpperCase());
+
+
 // Starts with fastpath, but switches to full Unicode path
 // U+00FF is uppercased to U+0178.
 assertEquals("AŸ", "aÿ".toUpperCase());
@@ -33,6 +52,10 @@ assertEquals("AΜ", "aµ".toUpperCase());
 // Buffer size increase
 assertEquals("CSSBẶ", "cßbặ".toUpperCase());
 assertEquals("FIFLFFIFFL", "\uFB01\uFB02\uFB03\uFB04".toUpperCase());
+assertEquals("ABCÀCSSA", "abcàcßa".toUpperCase());
+assertEquals("ABCDEFGHIÀCSSA", "ABCDEFGHIàcßa".toUpperCase());
+assertEquals("ABCDEFGHIÀCSSA", "abcdeFghiàcßa".toUpperCase());
+
 // OneByte input with buffer size increase: non-fast path
 assertEquals("ABCSS", "abCß".toLocaleUpperCase("tr"));