Improve shaping segmentation for grapheme cluster based font fallback

third_party/WebKit/Source/platform/fonts/shaping/HarfBuzzShaper.cpp

Index: third_party/WebKit/Source/platform/fonts/shaping/HarfBuzzShaper.cpp
diff --git a/third_party/WebKit/Source/platform/fonts/shaping/HarfBuzzShaper.cpp b/third_party/WebKit/Source/platform/fonts/shaping/HarfBuzzShaper.cpp
index 6e47ccf7442173fb82c1ed7d5f1366239720b404..1f5b820f8b88233e0825355e3f9e0a9448a268af 100644
--- a/third_party/WebKit/Source/platform/fonts/shaping/HarfBuzzShaper.cpp
+++ b/third_party/WebKit/Source/platform/fonts/shaping/HarfBuzzShaper.cpp
@@ -34,12 +34,15 @@
 
 #include "hb.h"
 #include "platform/LayoutUnit.h"
+#include "platform/Logging.h"
 #include "platform/RuntimeEnabledFeatures.h"
 #include "platform/fonts/Character.h"
 #include "platform/fonts/Font.h"
+#include "platform/fonts/FontFallbackIterator.h"
 #include "platform/fonts/GlyphBuffer.h"
 #include "platform/fonts/UTF16TextIterator.h"
 #include "platform/fonts/shaping/HarfBuzzFace.h"
+#include "platform/fonts/shaping/RunSegmenter.h"
 #include "platform/text/TextBreakIterator.h"
 #include "wtf/Compiler.h"
 #include "wtf/MathExtras.h"
@@ -779,264 +782,261 @@ void HarfBuzzShaper::setFontFeatures()
     }
 }
 
-PassRefPtr<ShapeResult> HarfBuzzShaper::shapeResult()
+// A port of hb_icu_script_to_script because harfbuzz on CrOS is built
+// without hb-icu. See http://crbug.com/356929
+static inline hb_script_t ICUScriptToHBScript(UScriptCode script)
 {
-    if (!createHarfBuzzRuns())
-        return nullptr;
-    return shapeHarfBuzzRuns();
+    if (UNLIKELY(script == USCRIPT_INVALID_CODE))
+        return HB_SCRIPT_INVALID;
+
+    return hb_script_from_string(uscript_getShortName(script), -1);
 }
 
-struct CandidateRun {
-    UChar32 character;
-    unsigned start;
-    unsigned end;
-    const SimpleFontData* fontData;
-    UScriptCode script;
-};
+static inline hb_direction_t TextDirectionToHBDirection(TextDirection dir, FontOrientation orientation, const SimpleFontData* fontData)
+{
+    hb_direction_t harfBuzzDirection = isVerticalAnyUpright(orientation) && !fontData->isTextOrientationFallback() ? HB_DIRECTION_TTB : HB_DIRECTION_LTR;
+    return dir == RTL ? HB_DIRECTION_REVERSE(harfBuzzDirection) : harfBuzzDirection;
+}
 
-static inline bool collectCandidateRuns(const UChar* normalizedBuffer,
-    size_t bufferLength, const Font* font, Vector<CandidateRun>* runs, bool isSpaceNormalize)
+static const uint16_t* toUint16(const UChar* src)
 {
-    UTF16TextIterator iterator(normalizedBuffer, bufferLength);
-    UChar32 character;
-    unsigned startIndexOfCurrentRun = 0;
+    // FIXME: This relies on undefined behavior however it works on the
+    // current versions of all compilers we care about and avoids making
+    // a copy of the string.
+    static_assert(sizeof(UChar) == sizeof(uint16_t), "UChar should be the same size as uint16_t");
+    return reinterpret_cast<const uint16_t*>(src);
+}
 
-    if (!iterator.consume(character))
-        return false;
+static inline void addToHarfBuzzBufferInternal(hb_buffer_t* buffer,
+    const FontDescription& fontDescription, const UChar* normalizedBuffer,
+    unsigned startIndex, unsigned numCharacters)
+{
+    // TODO: Revisit whether we can always fill the hb_buffer_t with the
+    // full run text, but only specify startIndex and numCharacters for the part
+    // to be shaped. Then simplify/change the complicated index computations in
+    // extractShapeResults().
+    if (fontDescription.variant() == FontVariantSmallCaps) {
+        String upperText = String(normalizedBuffer + startIndex, numCharacters)
+            .upper();
+        // TextRun is 16 bit, therefore upperText is 16 bit, even after we call
+        // makeUpper().
+        ASSERT(!upperText.is8Bit());
+        hb_buffer_add_utf16(buffer, toUint16(upperText.characters16()),
+            numCharacters, 0, numCharacters);
+    } else {
+        hb_buffer_add_utf16(buffer, toUint16(normalizedBuffer + startIndex),
+            numCharacters, 0, numCharacters);
+    }
+}
 
-    const SimpleFontData* nextFontData = font->glyphDataForCharacter(character, false, isSpaceNormalize).fontData;
-    UErrorCode errorCode = U_ZERO_ERROR;
-    UScriptCode nextScript = uscript_getScript(character, &errorCode);
-    if (U_FAILURE(errorCode))
+inline bool HarfBuzzShaper::shapeRange(hb_buffer_t* harfBuzzBuffer,
+    unsigned startIndex,
+    unsigned numCharacters,
+    const SimpleFontData* currentFont,
+    UScriptCode currentRunScript,
+    hb_language_t language)
+{
+    const FontPlatformData* platformData = &(currentFont->platformData());
+    HarfBuzzFace* face = platformData->harfBuzzFace();
+    if (!face) {
+        WTF_LOG_ERROR("Could not create HarfBuzzFace from FontPlatformData.");
         return false;
+    }
 
-    do {
-        const UChar* currentCharacterPosition = iterator.characters();
-        const SimpleFontData* currentFontData = nextFontData;
-        UScriptCode currentScript = nextScript;
-
-        UChar32 lastCharacter = character;
-        for (iterator.advance(); iterator.consume(character); iterator.advance()) {
-            if (Character::treatAsZeroWidthSpace(character))
-                continue;
-            if ((U_GET_GC_MASK(character) & U_GC_M_MASK)
-                && (Character::isUnicodeVariationSelector(character)
-                    || currentFontData->canRenderCombiningCharacterSequence(
-                        currentCharacterPosition,
-                        iterator.glyphEnd() - currentCharacterPosition)))
-                continue;
+    hb_buffer_set_language(harfBuzzBuffer, language);
+    hb_buffer_set_script(harfBuzzBuffer, ICUScriptToHBScript(currentRunScript));
+    hb_buffer_set_direction(harfBuzzBuffer, TextDirectionToHBDirection(m_textRun.direction(),
+        m_font->fontDescription().orientation(), currentFont));
 
-            nextFontData = font->glyphDataForCharacter(character, false, isSpaceNormalize).fontData;
-            nextScript = uscript_getScript(character, &errorCode);
-            if (U_FAILURE(errorCode))
-                return false;
-            if (lastCharacter == zeroWidthJoinerCharacter)
-                currentFontData = nextFontData;
-            if ((nextFontData != currentFontData) || ((currentScript != nextScript) && (nextScript != USCRIPT_INHERITED) && (!uscript_hasScript(character, currentScript))))
-                break;
-            currentCharacterPosition = iterator.characters();
-            lastCharacter = character;
-        }
+    // Add a space as pre-context to the buffer. This prevents showing dotted-circle
+    // for combining marks at the beginning of runs.
+    static const uint16_t preContext = spaceCharacter;
+    hb_buffer_add_utf16(harfBuzzBuffer, &preContext, 1, 1, 0);
 
-        CandidateRun run = { lastCharacter, startIndexOfCurrentRun, static_cast<unsigned>(iterator.offset()), currentFontData, currentScript };
-        runs->append(run);
+    addToHarfBuzzBufferInternal(harfBuzzBuffer,
+        m_font->fontDescription(), m_normalizedBuffer.get(), startIndex,
+        numCharacters);
 
-        startIndexOfCurrentRun = iterator.offset();
-    } while (iterator.consume(character));
+    HarfBuzzScopedPtr<hb_font_t> harfBuzzFont(face->createFont(), hb_font_destroy);
+    hb_shape(harfBuzzFont.get(), harfBuzzBuffer, m_features.isEmpty() ? 0 : m_features.data(), m_features.size());
 
     return true;
 }
 
-static inline bool matchesAdjacentRun(UScriptCode* scriptExtensions, int length,
-    CandidateRun& adjacentRun)
+bool HarfBuzzShaper::extractShapeResults(hb_buffer_t* harfBuzzBuffer,
+    ShapeResult* shapeResult,
+    bool& fontCycleQueued, const HolesQueueItem& currentQueueItem,
+    const SimpleFontData* currentFont,
+    UScriptCode currentRunScript,
+    bool isLastResort)
 {
-    for (int i = 0; i < length; i++) {
-        if (scriptExtensions[i] == adjacentRun.script)
-            return true;
-    }
-    return false;
-}
+    enum ClusterResult {
+        Shaped,
+        NotDef,
+        Unknown
+    };
+    ClusterResult currentClusterResult = Unknown;
+    ClusterResult previousClusterResult = Unknown;
+    unsigned previousCluster = 0;
+    unsigned currentCluster = 0;
 
-static inline void resolveRunBasedOnScriptExtensions(Vector<CandidateRun>& runs,
-    CandidateRun& run, size_t i, size_t length, UScriptCode* scriptExtensions,
-    int extensionsLength, size_t& nextResolvedRun)
-{
-    // If uscript_getScriptExtensions returns 1 it only contains the script value,
-    // we only care about ScriptExtensions which is indicated by a value >= 2.
-    if (extensionsLength <= 1)
-        return;
+    // Find first notdef glyph in harfBuzzBuffer.
+    unsigned numGlyphs = hb_buffer_get_length(harfBuzzBuffer);
+    hb_glyph_info_t* glyphInfo = hb_buffer_get_glyph_infos(harfBuzzBuffer, 0);
 
-    if (i > 0 && matchesAdjacentRun(scriptExtensions, extensionsLength, runs[i - 1])) {
-        run.script = runs[i - 1].script;
-        return;
+    int lastChangePosition = 0;
+
+    if (!numGlyphs) {
+        WTF_LOG_ERROR("HarfBuzz returned empty glyph buffer after shaping.");
+        return false;
     }
 
-    for (size_t j = i + 1; j < length; j++) {
-        if (runs[j].script != USCRIPT_COMMON
-            && runs[j].script != USCRIPT_INHERITED
-            && matchesAdjacentRun(scriptExtensions, extensionsLength, runs[j])) {
-            nextResolvedRun = j;
-            break;
+    for (unsigned glyphIndex = 0; glyphIndex <= numGlyphs; ++glyphIndex) {
+        // Iterating by clusters, check for when the state switches from shaped
+        // to non-shaped and vice versa. Taking into account the edge cases of
+        // beginning of the run and end of the run.
+        previousCluster = currentCluster;
+        currentCluster = glyphInfo[glyphIndex].cluster;
+
+        if (glyphIndex < numGlyphs) {
+            // Still the same cluster, merge shaping status.
+            if (previousCluster == currentCluster && glyphIndex != 0) {
+                if (glyphInfo[glyphIndex].codepoint == 0) {
+                    currentClusterResult = NotDef;
+                } else {
+                    // We can only call the current cluster fully shapped, if
+                    // all characters that are part of it are shaped, so update
+                    // currentClusterResult to Shaped only if the previous
+                    // characters have been shaped, too.

[eae, 2015/10/14 13:00:40]

Got it, that makes sense. Thanks for adding the comment.

+                    currentClusterResult = currentClusterResult == Shaped ? Shaped : NotDef;
+                }
+                continue;
+            }
+            // We've moved to a new cluster.
+            previousClusterResult = currentClusterResult;
+            currentClusterResult = glyphInfo[glyphIndex].codepoint == 0 ? NotDef : Shaped;
+        } else {
+            // The code below operates on the "flanks"/changes between NotDef
+            // and Shaped. In order to keep the code below from explictly
+            // dealing with character indices and run end, we explicitly
+            // terminate the cluster/run here by setting the result value to the
+            // opposite of what it was, leading to atChange turning true.
+            previousClusterResult = currentClusterResult;
+            currentClusterResult = currentClusterResult == NotDef ? Shaped : NotDef;
         }
-    }
-}
 
-static inline void resolveRunBasedOnScriptValue(Vector<CandidateRun>& runs,
-    CandidateRun& run, size_t i, size_t length, size_t& nextResolvedRun)
-{
-    if (run.script != USCRIPT_COMMON)
-        return;
+        bool atChange = (previousClusterResult != currentClusterResult) && previousClusterResult != Unknown;
+        if (!atChange)
+            continue;
 
-    if (i > 0 && runs[i - 1].script != USCRIPT_COMMON) {
-        run.script = runs[i - 1].script;
-        return;
-    }
+        // Compute the range indices of consecutive shaped or .notdef glyphs.
+        // Cluster information for RTL runs becomes reversed, e.g. character 0
+        // has cluster index 5 in a run of 6 characters.
+        unsigned numCharacters = 0;
+        unsigned numGlyphsToInsert = 0;
+        unsigned startIndex = 0;
+        if (HB_DIRECTION_IS_FORWARD(hb_buffer_get_direction(harfBuzzBuffer))) {
+            startIndex = currentQueueItem.m_startIndex + glyphInfo[lastChangePosition].cluster;
+            if (glyphIndex == numGlyphs) {
+                numCharacters = currentQueueItem.m_numCharacters - glyphInfo[lastChangePosition].cluster;
+                numGlyphsToInsert = numGlyphs - lastChangePosition;
+            } else {
+                numCharacters = glyphInfo[glyphIndex].cluster - glyphInfo[lastChangePosition].cluster;
+                numGlyphsToInsert = glyphIndex - lastChangePosition;
+            }
+        } else {
+            // Direction Backwards
+            startIndex = currentQueueItem.m_startIndex + glyphInfo[glyphIndex - 1].cluster;
+            if (lastChangePosition == 0) {
+                numCharacters = currentQueueItem.m_numCharacters - glyphInfo[glyphIndex - 1].cluster;
+            } else {
+                numCharacters = glyphInfo[lastChangePosition - 1].cluster - glyphInfo[glyphIndex - 1].cluster;
+            }
+            numGlyphsToInsert = glyphIndex - lastChangePosition;
+        }
 
-    for (size_t j = i + 1; j < length; j++) {
-        if (runs[j].script != USCRIPT_COMMON
-            && runs[j].script != USCRIPT_INHERITED) {
-            nextResolvedRun = j;
-            break;
+        if (currentClusterResult == Shaped && !isLastResort) {
+            // Now it's clear that we need to continue processing.
+            if (!fontCycleQueued) {
+                appendToHolesQueue(HolesQueueNextFont, 0, 0);
+                fontCycleQueued = true;
+            }
+
+            // Here we need to put character positions.
+            ASSERT(numCharacters);
+            appendToHolesQueue(HolesQueueRange, startIndex, numCharacters);
         }
-    }
-}
 
-static inline bool resolveCandidateRuns(Vector<CandidateRun>& runs)
-{
-    UScriptCode scriptExtensions[USCRIPT_CODE_LIMIT];
-    UErrorCode errorCode = U_ZERO_ERROR;
-    size_t length = runs.size();
-    for (size_t i = 0; i < length; i++) {
-        CandidateRun& run = runs[i];
-        size_t nextResolvedRun = 0;
-
-        if (run.script == USCRIPT_INHERITED)
-            run.script = i > 0 ? runs[i - 1].script : USCRIPT_COMMON;
-
-        int extensionsLength = uscript_getScriptExtensions(run.character,
-            scriptExtensions, sizeof(scriptExtensions) / sizeof(scriptExtensions[0]),
-            &errorCode);
-        if (U_FAILURE(errorCode))
-            return false;
-
-        resolveRunBasedOnScriptExtensions(runs, run, i, length,
-            scriptExtensions, extensionsLength, nextResolvedRun);
-        resolveRunBasedOnScriptValue(runs, run, i, length,
-            nextResolvedRun);
-        for (size_t j = i; j < nextResolvedRun; j++)
-            runs[j].script = runs[nextResolvedRun].script;
-
-        i = std::max(i, nextResolvedRun);
+        // If numCharacters is 0, that means we hit a NotDef before shaping the
+        // whole grapheme. We do not append it here. For the next glyph we
+        // encounter, atChange will be true, and the characters corresponding to
+        // the grapheme will be added to the TODO queue again, attempting to
+        // shape the whole grapheme with the next font.
+        // When we're getting here with the last resort font, we have no other
+        // choice than adding boxes to the ShapeResult.
+        if ((currentClusterResult == NotDef && numCharacters) || isLastResort) {
+            // Here we need to specify glyph positions.
+            OwnPtr<ShapeResult::RunInfo> run = adoptPtr(new ShapeResult::RunInfo(currentFont,
+                TextDirectionToHBDirection(m_textRun.direction(),
+                m_font->fontDescription().orientation(), currentFont),
+                ICUScriptToHBScript(currentRunScript),
+                startIndex,
+                numGlyphsToInsert, numCharacters));
+            insertRunIntoShapeResult(shapeResult, run.release(), lastChangePosition, numGlyphsToInsert, currentQueueItem.m_startIndex, harfBuzzBuffer);
+        }
+        lastChangePosition = glyphIndex;
     }
     return true;
 }
 
-// For ideographic (CJK) documents, 90-95% of calls from width() are one character length
-// because most characters have break opportunities both before and after.
-bool HarfBuzzShaper::createHarfBuzzRunsForSingleCharacter()
+static inline const SimpleFontData* fontDataAdjustedForOrientation(const SimpleFontData* originalFont,
+    FontOrientation runOrientation,
+    OrientationIterator::RenderOrientation renderOrientation)
 {
-    ASSERT(m_normalizedBufferLength == 1);
-    UChar32 character = m_normalizedBuffer[0];
-    if (!U16_IS_SINGLE(character))
-        return false;
-    const SimpleFontData* fontData = m_font->glyphDataForCharacter(character, false, m_textRun.normalizeSpace()).fontData;
-    UErrorCode errorCode = U_ZERO_ERROR;
-    UScriptCode script = uscript_getScript(character, &errorCode);
-    if (U_FAILURE(errorCode))
-        return false;
-    addHarfBuzzRun(0, 1, fontData, script);
-    return true;
+    if (!isVerticalBaseline(runOrientation))
+        return originalFont;
+
+    if (runOrientation == FontOrientation::VerticalRotated
+        || (runOrientation == FontOrientation::VerticalMixed && renderOrientation == OrientationIterator::OrientationRotateSideways))
+        return originalFont->verticalRightOrientationFontData().get();
+
+    return originalFont;
 }
 
-bool HarfBuzzShaper::createHarfBuzzRuns()
+bool HarfBuzzShaper::collectFallbackHintChars(Vector<UChar32>& hint, bool needsList)
 {
-    if (m_normalizedBufferLength == 1)
-        return createHarfBuzzRunsForSingleCharacter();
-
-    Vector<CandidateRun> candidateRuns;
-    if (!collectCandidateRuns(m_normalizedBuffer.get(),
-        m_normalizedBufferLength, m_font, &candidateRuns, m_textRun.normalizeSpace()))
+    if (!m_holesQueue.size())
         return false;
 
-    if (!resolveCandidateRuns(candidateRuns))
-        return false;
+    hint.clear();
 
-    size_t length = candidateRuns.size();
-    for (size_t i = 0; i < length; ) {
-        CandidateRun& run = candidateRuns[i];
-        CandidateRun lastMatchingRun = run;
-        for (i++; i < length; i++) {
-            if (candidateRuns[i].script != run.script
-                || candidateRuns[i].fontData != run.fontData)
+    size_t numCharsAdded = 0;
+    for (auto it = m_holesQueue.begin(); it != m_holesQueue.end(); ++it) {
+        if (it->m_action == HolesQueueNextFont)
+            break;
+
+        UChar32 hintChar;
+        RELEASE_ASSERT(it->m_startIndex + it->m_numCharacters <= m_normalizedBufferLength);
+        UTF16TextIterator iterator(m_normalizedBuffer.get() + it->m_startIndex, it->m_numCharacters);
+        while (iterator.consume(hintChar)) {
+            hint.append(hintChar);
+            numCharsAdded++;
+            if (!needsList)
                 break;
-            lastMatchingRun = candidateRuns[i];
+            iterator.advance();
         }
-        addHarfBuzzRun(run.start, lastMatchingRun.end, run.fontData, run.script);
     }
-    return !m_harfBuzzRuns.isEmpty();
-}
-
-// A port of hb_icu_script_to_script because harfbuzz on CrOS is built
-// without hb-icu. See http://crbug.com/356929
-static inline hb_script_t ICUScriptToHBScript(UScriptCode script)
-{
-    if (UNLIKELY(script == USCRIPT_INVALID_CODE))
-        return HB_SCRIPT_INVALID;
-
-    return hb_script_from_string(uscript_getShortName(script), -1);
-}
-
-static inline hb_direction_t TextDirectionToHBDirection(TextDirection dir, FontOrientation orientation, const SimpleFontData* fontData)
-{
-    hb_direction_t harfBuzzDirection = isVerticalAnyUpright(orientation) && !fontData->isTextOrientationFallback() ? HB_DIRECTION_TTB : HB_DIRECTION_LTR;
-    return dir == RTL ? HB_DIRECTION_REVERSE(harfBuzzDirection) : harfBuzzDirection;
-}
-
-void HarfBuzzShaper::addHarfBuzzRun(unsigned startCharacter,
-    unsigned endCharacter, const SimpleFontData* fontData,
-    UScriptCode script)
-{
-    ASSERT(endCharacter > startCharacter);
-    ASSERT(script != USCRIPT_INVALID_CODE);
-
-    hb_direction_t direction = TextDirectionToHBDirection(m_textRun.direction(),
-        m_font->fontDescription().orientation(), fontData);
-    HarfBuzzRun harfBuzzRun = {
-        fontData, startCharacter, endCharacter - startCharacter,
-        direction, ICUScriptToHBScript(script)
-    };
-    m_harfBuzzRuns.append(harfBuzzRun);
-}
-
-static const uint16_t* toUint16(const UChar* src)
-{
-    // FIXME: This relies on undefined behavior however it works on the
-    // current versions of all compilers we care about and avoids making
-    // a copy of the string.
-    static_assert(sizeof(UChar) == sizeof(uint16_t), "UChar should be the same size as uint16_t");
-    return reinterpret_cast<const uint16_t*>(src);
+    return numCharsAdded > 0;
 }
 
-static inline void addToHarfBuzzBufferInternal(hb_buffer_t* buffer,
-    const FontDescription& fontDescription, const UChar* normalizedBuffer,
-    unsigned startIndex, unsigned numCharacters)
+void HarfBuzzShaper::appendToHolesQueue(HolesQueueItemAction action,
+    unsigned startIndex,
+    unsigned numCharacters)
 {
-    if (fontDescription.variant() == FontVariantSmallCaps
-        && u_islower(normalizedBuffer[startIndex])) {
-        String upperText = String(normalizedBuffer + startIndex, numCharacters)
-            .upper();
-        // TextRun is 16 bit, therefore upperText is 16 bit, even after we call
-        // makeUpper().
-        ASSERT(!upperText.is8Bit());
-        hb_buffer_add_utf16(buffer, toUint16(upperText.characters16()),
-            numCharacters, 0, numCharacters);
-    } else {
-        hb_buffer_add_utf16(buffer, toUint16(normalizedBuffer + startIndex),
-            numCharacters, 0, numCharacters);
-    }
+    m_holesQueue.append(HolesQueueItem(action, startIndex, numCharacters));
 }
 
-PassRefPtr<ShapeResult> HarfBuzzShaper::shapeHarfBuzzRuns()
+PassRefPtr<ShapeResult> HarfBuzzShaper::shapeResult()
 {
     RefPtr<ShapeResult> result = ShapeResult::create(m_font,
         m_normalizedBufferLength, m_textRun.direction());
@@ -1047,61 +1047,104 @@ PassRefPtr<ShapeResult> HarfBuzzShaper::shapeHarfBuzzRuns()
     CString locale = localeString.latin1();
     const hb_language_t language = hb_language_from_string(locale.data(), locale.length());
 
-    result->m_runs.resize(m_harfBuzzRuns.size());
-    for (unsigned i = 0; i < m_harfBuzzRuns.size(); ++i) {
-        unsigned runIndex = m_textRun.rtl() ? m_harfBuzzRuns.size() - i - 1 : i;
-        const HarfBuzzRun* currentRun = &m_harfBuzzRuns[runIndex];
-
-        const SimpleFontData* currentFontData = currentRun->m_fontData;
-        FontPlatformData* platformData = const_cast<FontPlatformData*>(&currentFontData->platformData());
-        HarfBuzzFace* face = platformData->harfBuzzFace();
-        if (!face)
-            return nullptr;
-
-        hb_buffer_set_language(harfBuzzBuffer.get(), language);
-        hb_buffer_set_script(harfBuzzBuffer.get(), currentRun->m_script);
-        hb_buffer_set_direction(harfBuzzBuffer.get(), currentRun->m_direction);
-
-        // Add a space as pre-context to the buffer. This prevents showing dotted-circle
-        // for combining marks at the beginning of runs.
-        static const uint16_t preContext = spaceCharacter;
-        hb_buffer_add_utf16(harfBuzzBuffer.get(), &preContext, 1, 1, 0);
-
-        addToHarfBuzzBufferInternal(harfBuzzBuffer.get(),
-            fontDescription, m_normalizedBuffer.get(), currentRun->m_startIndex,
-            currentRun->m_numCharacters);
-
-        HarfBuzzScopedPtr<hb_font_t> harfBuzzFont(face->createFont(), hb_font_destroy);
-        hb_shape(harfBuzzFont.get(), harfBuzzBuffer.get(), m_features.isEmpty() ? 0 : m_features.data(), m_features.size());
-        shapeResult(result.get(), i, currentRun, harfBuzzBuffer.get());
+    RunSegmenter::RunSegmenterRange segmentRange = {
+        0,
+        0,
+        USCRIPT_INVALID_CODE,
+        OrientationIterator::OrientationInvalid,
+        SmallCapsIterator::SmallCapsSameCase };
+    RunSegmenter runSegmenter(
+        m_normalizedBuffer.get(),
+        m_normalizedBufferLength,
+        m_font->fontDescription().orientation(),
+        fontDescription.variant());
+
+    Vector<UChar32> fallbackCharsHint;
+
+    // TODO: Check whether this treatAsZerowidthspace from the previous script
+    // segmentation plays a role here, does the new scriptRuniterator handle that correctly?
+    while (runSegmenter.consume(&segmentRange)) {
+        RefPtr<FontFallbackIterator> fallbackIterator = m_font->createFontFallbackIterator();
+
+        appendToHolesQueue(HolesQueueNextFont, 0, 0);
+        appendToHolesQueue(HolesQueueRange, segmentRange.start, segmentRange.end - segmentRange.start);
+
+        const SimpleFontData* currentFont = nullptr;
+
+        bool fontCycleQueued = false;
+        while (m_holesQueue.size()) {
+            HolesQueueItem currentQueueItem = m_holesQueue.takeFirst();
+
+            if (currentQueueItem.m_action == HolesQueueNextFont) {
+                // For now, we're building a character list with which we probe
+                // for needed fonts depending on the declared unicode-range of a
+                // segmented CSS font. Alternatively, we can build a fake font
+                // for the shaper and check whether any glyphs were found, or
+                // define a new API on the shaper which will give us coverage
+                // information?
+                if (!collectFallbackHintChars(fallbackCharsHint, fallbackIterator->needsHintList())) {
+                    // Give up shaping since we cannot retrieve a font fallback
+                    // font without a hintlist.
+                    m_holesQueue.clear();
+                    break;
+                }
+
+                currentFont = fallbackIterator->next(fallbackCharsHint);
+                if (!currentFont) {
+                    ASSERT(!m_holesQueue.size());
+                    break;
+                }
+                fontCycleQueued = false;
+                continue;
+            }
 
-        hb_buffer_reset(harfBuzzBuffer.get());
+            // TODO crbug.com/522964: Only use smallCapsFontData when the font does not support true smcp.  The spec
+            // says: "To match the surrounding text, a font may provide alternate glyphs for caseless characters when
+            // these features are enabled but when a user agent simulates small capitals, it must not attempt to
+            // simulate alternates for codepoints which are considered caseless."
+            const SimpleFontData* smallcapsAdjustedFont = segmentRange.smallCapsBehavior == SmallCapsIterator::SmallCapsUppercaseNeeded
+                ? currentFont->smallCapsFontData(fontDescription).get()
+                : currentFont;
+
+            // Compatibility with SimpleFontData approach of keeping a flag for overriding drawing direction.
+            // TODO: crbug.com/506224 This should go away in favor of storing that information elsewhere, for example in
+            // ShapeResult.
+            const SimpleFontData* directionAndSmallCapsAdjustedFont = fontDataAdjustedForOrientation(smallcapsAdjustedFont,
+                m_font->fontDescription().orientation(),
+                segmentRange.renderOrientation);
+
+            if (!shapeRange(harfBuzzBuffer.get(),
+                currentQueueItem.m_startIndex,
+                currentQueueItem.m_numCharacters,
+                directionAndSmallCapsAdjustedFont,
+                segmentRange.script,
+                language))
+                WTF_LOG_ERROR("Shaping range failed.");
+
+            if (!extractShapeResults(harfBuzzBuffer.get(),
+                result.get(),
+                fontCycleQueued,
+                currentQueueItem,
+                directionAndSmallCapsAdjustedFont,
+                segmentRange.script,
+                !fallbackIterator->hasNext()))
+                WTF_LOG_ERROR("Shape result extraction failed.");
+
+            hb_buffer_reset(harfBuzzBuffer.get());
+        }
     }
-
-    // We should have consumed all expansion opportunities.
-    // Failures here means that our logic does not match to the one in expansionOpportunityCount().
-    // FIXME: Ideally, we should ASSERT(!m_expansionOpportunityCount) here to ensure that,
-    // or unify the two logics (and the one in SimplePath too,) but there are some cases where our impl
-    // does not support justification very well yet such as U+3099, and it'll cause the ASSERT to fail.
-    // It's to be fixed because they're very rarely used, and a broken justification is still somewhat readable.
-
     return result.release();
 }
 
-void HarfBuzzShaper::shapeResult(ShapeResult* result, unsigned index,
-    const HarfBuzzRun* currentRun, hb_buffer_t* harfBuzzBuffer)
+// TODO crbug.com/542701: This should be a method on ShapeResult.
+void HarfBuzzShaper::insertRunIntoShapeResult(ShapeResult* result,
+    PassOwnPtr<ShapeResult::RunInfo> runToInsert, unsigned startGlyph, unsigned numGlyphs,
+    int bufferStartCharIndex, hb_buffer_t* harfBuzzBuffer)
 {
-    unsigned numGlyphs = hb_buffer_get_length(harfBuzzBuffer);
-    if (!numGlyphs) {
-        result->m_runs[index] = nullptr;
-        return;
-    }
+    ASSERT(numGlyphs > 0);
+    OwnPtr<ShapeResult::RunInfo> run(runToInsert);
 
-    OwnPtr<ShapeResult::RunInfo> run = adoptPtr(new ShapeResult::RunInfo(currentRun->m_fontData,
-        currentRun->m_direction, currentRun->m_script, currentRun->m_startIndex,
-        numGlyphs, currentRun->m_numCharacters));
-
-    const SimpleFontData* currentFontData = currentRun->m_fontData;
+    const SimpleFontData* currentFontData = run->m_fontData.get();
     hb_glyph_info_t* glyphInfos = hb_buffer_get_glyph_infos(harfBuzzBuffer, 0);
     hb_glyph_position_t* glyphPositions = hb_buffer_get_glyph_positions(harfBuzzBuffer, 0);
 
@@ -1111,21 +1154,29 @@ void HarfBuzzShaper::shapeResult(ShapeResult* result, unsigned index,
     float* directionOffset = m_font->fontDescription().isVerticalAnyUpright() ? &offsetY : &offsetX;
 
     // HarfBuzz returns result in visual order, no need to flip for RTL.
-    for (size_t i = 0; i < numGlyphs; ++i) {
+    for (unsigned i = 0; i < numGlyphs; ++i) {
         bool runEnd = i + 1 == numGlyphs;
-        uint16_t glyph = glyphInfos[i].codepoint;
-        offsetX = harfBuzzPositionToFloat(glyphPositions[i].x_offset);
-        offsetY = -harfBuzzPositionToFloat(glyphPositions[i].y_offset);
+        uint16_t glyph = glyphInfos[startGlyph + i].codepoint;
+        offsetX = harfBuzzPositionToFloat(glyphPositions[startGlyph + i].x_offset);
+        offsetY = -harfBuzzPositionToFloat(glyphPositions[startGlyph + i].y_offset);
 
         // One out of x_advance and y_advance is zero, depending on
         // whether the buffer direction is horizontal or vertical.
-        float advance = harfBuzzPositionToFloat(glyphPositions[i].x_advance - glyphPositions[i].y_advance);
-        unsigned currentCharacterIndex = currentRun->m_startIndex + glyphInfos[i].cluster;
+        float advance = harfBuzzPositionToFloat(glyphPositions[startGlyph + i].x_advance - glyphPositions[startGlyph + i].y_advance);
+        unsigned currentCharacterIndex = bufferStartCharIndex + glyphInfos[startGlyph + i].cluster;
         RELEASE_ASSERT(m_normalizedBufferLength > currentCharacterIndex);
-        bool isClusterEnd = runEnd || glyphInfos[i].cluster != glyphInfos[i + 1].cluster;
+        bool isClusterEnd = runEnd || glyphInfos[startGlyph + i].cluster != glyphInfos[startGlyph + i + 1].cluster;
         float spacing = 0;
 
-        run->m_glyphData[i].characterIndex = glyphInfos[i].cluster;
+        // The characterIndex of one ShapeResult run is normalized to the run's
+        // startIndex and length.  TODO crbug.com/542703: Consider changing that
+        // and instead pass the whole run to hb_buffer_t each time.
+        run->m_glyphData.resize(numGlyphs);
+        if (HB_DIRECTION_IS_FORWARD(hb_buffer_get_direction(harfBuzzBuffer))) {
+            run->m_glyphData[i].characterIndex = glyphInfos[startGlyph + i].cluster - glyphInfos[startGlyph].cluster;
+        } else {
+            run->m_glyphData[i].characterIndex = glyphInfos[startGlyph + i].cluster - glyphInfos[startGlyph + numGlyphs - 1].cluster;
+        }
 
         if (isClusterEnd)
             spacing += adjustSpacing(run.get(), i, currentCharacterIndex, *directionOffset, totalAdvance);
@@ -1151,11 +1202,33 @@ void HarfBuzzShaper::shapeResult(ShapeResult* result, unsigned index,
         result->m_glyphBoundingBox.unite(glyphBounds);
         glyphOrigin += FloatSize(advance + offsetX, offsetY);
     }
-
     run->m_width = std::max(0.0f, totalAdvance);
     result->m_width += run->m_width;
     result->m_numGlyphs += numGlyphs;
-    result->m_runs[index] = run.release();
+
+    // The runs are stored in result->m_runs in visual order. For LTR, we place
+    // the run to be inserted before the next run with a bigger character
+    // start index. For RTL, we place the run before the next run with a lower
+    // character index. Otherwise, for both directions, at the end.
+    if (HB_DIRECTION_IS_FORWARD(run->m_direction)) {
+        for (size_t pos = 0; pos < result->m_runs.size(); ++pos) {
+            if (result->m_runs.at(pos)->m_startIndex > run->m_startIndex) {
+                result->m_runs.insert(pos, run.release());
+                break;
+            }
+        }
+    } else {
+        for (size_t pos = 0; pos < result->m_runs.size(); ++pos) {
+            if (result->m_runs.at(pos)->m_startIndex < run->m_startIndex) {
+                result->m_runs.insert(pos, run.release());
+                break;
+            }
+        }
+    }
+    // If we didn't find an existing slot to place it, append.
+    if (run) {
+        result->m_runs.append(run.release());
+    }
 }
 
 PassRefPtr<ShapeResult> ShapeResult::createForTabulationCharacters(const Font* font,
@@ -1239,4 +1312,5 @@ float HarfBuzzShaper::adjustSpacing(ShapeResult::RunInfo* run, size_t glyphIndex
     return spacing;
 }
 
+
 } // namespace blink