Use ICU case conversion/transliterator for case conversion behind a flag

src/runtime/runtime-strings.cc

Index: src/runtime/runtime-strings.cc
diff --git a/src/runtime/runtime-strings.cc b/src/runtime/runtime-strings.cc
index 6786fa99fba0f92e1c41e5e228d9f6852349dd74..38be231f27e5ad0aaee57e0257d23b4835f6f0bd 100644
--- a/src/runtime/runtime-strings.cc
+++ b/src/runtime/runtime-strings.cc
@@ -12,6 +12,12 @@
 #include "src/string-builder.h"
 #include "src/string-search.h"
 
+#ifdef V8_I18N_SUPPORT
+#include "unicode/locid.h"
+#include "unicode/uchar.h"
+#include "unicode/unistr.h"
+#endif
+
 namespace v8 {
 namespace internal {
 
@@ -1077,20 +1083,210 @@ MUST_USE_RESULT static Object* ConvertCase(
   return ConvertCaseHelper(isolate, *s, *result, length, mapping);
 }
 
+#ifdef V8_I18N_SUPPORT
+namespace {
+
+MUST_USE_RESULT static Handle<String> ConvertCaseICU(Handle<String> s,
+                                                     Isolate* isolate,
+                                                     bool is_to_upper) {
+  DCHECK(s->IsFlat());
+  // Handle<String> flattened = String::Flatten(s);
+  String::FlatContent flat = s->GetFlatContent();

[adamk, 2016/04/07 20:50:55]

You'll want to add a DisallowHeapAllocation object on the stack here, like:

DisallowHeapAllocation no_gc;

What the DCHECK in GetFlatContent is doing is keeping you from holding onto a
pointer to the flat content across a GC, which might invalidate it.

Once you've added the above DisallowHeapAllocation, you'll need to make sure
that you do avoid allocations while it's on the stack (otherwise you'll get
other DCHECK failures). Specifically this would be the NewStringFromTwoByte()
call down at the bottom; the usual thing to do is put the rest of the function
inside an inner block, and allocate the DisallowHeapAllocation in there.

However...

+
+  const UChar* src;
+  if (flat.IsOneByte()) {
+    base::SmartArrayPointer<uc16> sap = s->ToWideCString();
+    src = reinterpret_cast<const UChar*>(sap.get());
+  } else {
+    src = reinterpret_cast<const UChar*>(flat.ToUC16Vector().start());
+  }
+
+  int32_t length = s->length();
+
+  // This UnicodeString ctor has copy-on-write semantics. It starts as a

[adamk, 2016/04/07 20:50:55]

...this comment worries me. Does this mean that converted.getBuffer() below
could be pointing at |src|? That doesn't sound valid to me, since
NewStringFromTwoByte needs to read from getBuffer _after_ doing an allocation,
which might cause GC. This function needs to make sure the FlatContent isn't
referenced when an allocation occurs.

+  // read-only alias but the buffer is copied when it's written to.
+  icu::UnicodeString converted(0, src, length);
+  const icu::Locale& root_locale = icu::Locale::getRoot();
+  if (is_to_upper)
+    converted.toUpper(root_locale);
+  else
+    converted.toLower(root_locale);
+
+  return isolate->factory()
+      ->NewStringFromTwoByte(Vector<const uint16_t>(
+          reinterpret_cast<const uint16_t*>(converted.getBuffer()),
+          converted.length()))
+      .ToHandleChecked();
+}
+
+inline bool IsASCIIUpper(uint16_t ch) { return ch >= 'A' && ch <= 'Z'; }
+
+inline uint16_t ToASCIILower(uint16_t ch) {
+  return ch | ((ch >= 'A' && ch <= 'Z') << 5);
+}
+
+inline uint16_t ToASCIIUpper(uint16_t ch) {
+  return ch & ~((ch >= 'a' && ch <= 'z') << 5);
+}
+
+MUST_USE_RESULT Handle<String> StringToLowerCase(Handle<String> s,
+                                                 Isolate* isolate) {
+  // Note: This is a hot function in the Dromaeo benchmark, specifically the
+  // no-op code path up through the first 'return' statement.
+
+  int length = s->length();
+  s = String::Flatten(s);
+  // First scan the string for uppercase and non-ASCII characters:
+  if (s->HasOnlyOneByteChars()) {
+    unsigned 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;
+      }
+    }
+
+    // Nothing to do if the string is all ASCII with no uppercase.
+    if (first_index_to_lower == 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 (untrue for ToUpperCase, and might
+    // be untrue in some locales, but this is the root locale)
+    Handle<SeqOneByteString> result =
+        isolate->factory()->NewRawOneByteString(length).ToHandleChecked();
+    if (s->IsSeqOneByteString()) {
+      SeqOneByteString* source = SeqOneByteString::cast(*s);
+      CopyChars(result->GetChars(), source->GetChars(), first_index_to_lower);
+    } else {
+      // Do we have to worry about External{One,Two}ByteString?
+      DCHECK(s->IsSeqTwoByteString());
+      SeqTwoByteString* source = SeqTwoByteString::cast(*s);
+      CopyChars(result->GetChars(), source->GetChars(), first_index_to_lower);
+    }
+
+    for (int index = first_index_to_lower; index < length; ++index) {
+      uint16_t ch = s->Get(index);
+      result->SeqOneByteStringSet(
+          index, V8_UNLIKELY(ch & ~0x7F) ? static_cast<uint16_t>(u_tolower(ch))
+                                         : ToASCIILower(ch));
+    }
+
+    return Handle<String>(*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 ConvertCaseICU(s, isolate, false);
+}
+
+const uint16_t sharp_s = 0x00DFu;
+
+MUST_USE_RESULT Handle<String> StringToUpperCase(Handle<String> s,
+                                                 Isolate* isolate) {
+  // This function could be optimized for no-op cases the way lower() 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 =
+        isolate->factory()->NewRawOneByteString(length).ToHandleChecked();
+
+    // Do a faster loop for the case where all the characters are ASCII.
+    uint16_t ored = 0;
+    for (int index = 0; index < length; ++index) {
+      uint16_t ch = s->Get(index);
+      ored |= ch;
+      result->SeqOneByteStringSet(index, ToASCIIUpper(ch));
+    }
+    if (!(ored & ~0x7F)) return Handle<String>(*result);
+
+    // Do a slower implementation for cases that include non-ASCII Latin-1
+    // characters.
+    int sharp_s_count = 0;
+
+    // There are two special cases.
+    //  1. latin-1 characters when converted to upper case are 16 bit
+    //  characters.
+    //  2. Lower case sharp-S converts to "SS" (two characters)
+    for (int32_t index = 0; index < length; ++index) {
+      uint16_t ch = s->Get(index);
+      if (V8_UNLIKELY(ch == sharp_s)) {
+        ++sharp_s_count;
+        continue;
+      }
+      uint16_t upper = static_cast<uint16_t>(u_toupper(static_cast<UChar>(ch)));
+      if (V8_UNLIKELY(upper > 0xff)) {
+        // Since this upper-cased character does not fit in an 8-bit string, we
+        // need to take the 16-bit path.
+        return ConvertCaseICU(s, isolate, true);
+      }
+      result->SeqOneByteStringSet(index, upper);
+    }
+
+    if (sharp_s_count == 0) return Handle<String>(*result);
+
+    // We have sharp_s_count sharp-s characters, but none of the other special
+    // characters.
+    result = isolate->factory()
+                 ->NewRawOneByteString(length + sharp_s_count)
+                 .ToHandleChecked();
+    for (int32_t index = 0, dest_index = 0; index < length; ++index) {
+      uint16_t ch = s->Get(index);
+      if (ch == sharp_s) {
+        result->SeqOneByteStringSet(dest_index++, 'S');
+        result->SeqOneByteStringSet(dest_index++, 'S');
+      } else {
+        uint16_t upper =
+            static_cast<uint16_t>(u_toupper(static_cast<UChar>(ch)));
+        result->SeqOneByteStringSet(dest_index++, upper);
+      }
+    }
+
+    return Handle<String>(*result);
+  }
+
+  return ConvertCaseICU(s, isolate, true);
+}
+
+}  // namespace
+#endif
 
 RUNTIME_FUNCTION(Runtime_StringToLowerCase) {
   HandleScope scope(isolate);
-  DCHECK(args.length() == 1);
+  DCHECK_EQ(args.length(), 1);
   CONVERT_ARG_HANDLE_CHECKED(String, s, 0);
-  return ConvertCase(s, isolate, isolate->runtime_state()->to_lower_mapping());
+#ifdef V8_I18N_SUPPORT
+  if (FLAG_icu_case_mapping)
+    return *StringToLowerCase(s, isolate);
+  else
+#endif
+    return ConvertCase(s, isolate,
+                       isolate->runtime_state()->to_lower_mapping());
 }
 
 
 RUNTIME_FUNCTION(Runtime_StringToUpperCase) {
   HandleScope scope(isolate);
-  DCHECK(args.length() == 1);
+  DCHECK_EQ(args.length(), 1);
   CONVERT_ARG_HANDLE_CHECKED(String, s, 0);
-  return ConvertCase(s, isolate, isolate->runtime_state()->to_upper_mapping());
+#ifdef V8_I18N_SUPPORT
+  if (FLAG_icu_case_mapping)
+    return *StringToUpperCase(s, isolate);
+  else
+#endif
+    return ConvertCase(s, isolate,
+                       isolate->runtime_state()->to_upper_mapping());
 }