Revert of Revert of HTMLConstructionSite: avoid n^2 running time for …

…large scripts. (patchset #1 id:1 of https://codereview.chromium.org/555223002/)

Reason for revert:
Not a culprit. Reverted to bring back the original patch.

Original issue's description:
> Revert of HTMLConstructionSite: avoid n^2 running time for large scripts. (patchset #5 id:80001 of https://codereview.chromium.org/494993002/)
> 
> Reason for revert:
> Speculative revert.
> 
> It may break the following blink sheriff bot:
> http://build.chromium.org/p/chromium.webkit/builders/Android%20Tests%20%28dbg%29/builds/21403
> 
> 
> 
> Original issue's description:
> > HTMLConstructionSite: avoid n^2 running time for large scripts.
> > 
> > Every time background parser sends chunk, tree is flushed.
> > 
> > If page contains very large script, then script node content is updated
> > many times. Every update is causes string concatenation.
> > 
> > Solution: do not flush pending text until it is mandatory.
> > 
> > Test: https://codereview.chromium.org/500363002
> > Test depends on: https://codereview.chromium.org/544453004/
> > 
> > BUG=410790
> > 
> > Committed: https://src.chromium.org/viewvc/blink?view=rev&revision=181635
> 
> [email protected],[email protected],[email protected]
> NOTREECHECKS=true
> NOTRY=true
> BUG=410790
> 
> Committed: https://src.chromium.org/viewvc/blink?view=rev&revision=181664

[email protected],[email protected],[email protected]
NOTREECHECKS=true
NOTRY=true
BUG=410790

Review URL: https://codereview.chromium.org/558743002

git-svn-id: svn://svn.chromium.org/blink/trunk@181668 bbb929c8-8fbe-4397-9dbb-9b2b20218538

Source/core/html/parser/HTMLConstructionSite.cpp

@@ -231,11 +231,15 @@ static String atomizeIfAllWhitespace(const String& string, WhitespaceMode whites
     return string;
 }
 
-void HTMLConstructionSite::flushPendingText()
+void HTMLConstructionSite::flushPendingText(FlushMode mode)
 {
     if (m_pendingText.isEmpty())
         return;
 
+    if (mode == FlushIfAtTextLimit
+        && !shouldUseLengthLimit(*m_pendingText.parent))
+        return;
+
     PendingText pendingText;
     // Hold onto the current pending text on the stack so that queueTask doesn't recurse infinitely.
     m_pendingText.swap(pendingText);
@@ -269,7 +273,7 @@ void HTMLConstructionSite::flushPendingText()
 
 void HTMLConstructionSite::queueTask(const HTMLConstructionSiteTask& task)
 {
-    flushPendingText();
+    flushPendingText(FlushAlways);
     ASSERT(m_pendingText.isEmpty());
     m_taskQueue.append(task);
 }
@@ -533,15 +537,15 @@ void HTMLConstructionSite::setCompatibilityModeFromDoctype(const String& name, c
 void HTMLConstructionSite::processEndOfFile()
 {
     ASSERT(currentNode());
-    flush();
+    flush(FlushAlways);
     openElements()->popAll();
 }
 
 void HTMLConstructionSite::finishedParsing()
 {
     // We shouldn't have any queued tasks but we might have pending text which we need to promote to tasks and execute.
     ASSERT(m_taskQueue.isEmpty());
-    flush();
+    flush(FlushAlways);
     m_document->finishedParsing();
 }
 
@@ -689,7 +693,7 @@ void HTMLConstructionSite::insertTextNode(const String& string, WhitespaceMode w
     // The nextChild != dummy.nextChild case occurs whenever foster parenting happened and we hit a new text node "<table>a</table>b"
     // In either case we have to flush the pending text into the task queue before making more.
     if (!m_pendingText.isEmpty() && (m_pendingText.parent != dummyTask.parent ||  m_pendingText.nextChild != dummyTask.nextChild))
-        flushPendingText();
+        flushPendingText(FlushAlways);
     m_pendingText.append(dummyTask.parent, dummyTask.nextChild, string, whitespaceMode);
 }
 

Source/core/html/parser/HTMLConstructionSite.h

@@ -92,6 +92,14 @@ enum WhitespaceMode {
     AllWhitespace,
 };
 
+enum FlushMode {
+    // Flush pending text. Flush queued tasks.
+    FlushAlways,
+
+    // Flush pending text if node has length limit. Flush queued tasks.
+    FlushIfAtTextLimit,
+};
+
 class AtomicHTMLToken;
 class Document;
 class Element;
@@ -113,16 +121,16 @@ class HTMLConstructionSite FINAL {
     void executeQueuedTasks();
 
     // flushPendingText turns pending text into queued Text insertions, but does not execute them.
-    void flushPendingText();
+    void flushPendingText(FlushMode);
 
     // Called before every token in HTMLTreeBuilder::processToken, thus inlined:
-    void flush()
+    void flush(FlushMode mode)
     {
         if (!hasPendingTasks())
             return;
-        flushPendingText();
+        flushPendingText(mode);
         executeQueuedTasks(); // NOTE: Possible reentrancy via JavaScript execution.
-        ASSERT(!hasPendingTasks());
+        ASSERT(mode == FlushIfAtTextLimit || !hasPendingTasks());
     }
 
     bool hasPendingTasks()

Source/core/html/parser/HTMLDocumentParser.cpp

@@ -498,9 +498,10 @@ void HTMLDocumentParser::processParsedChunkFromBackgroundParser(PassOwnPtr<Parse
         ASSERT(!m_token);
     }
 
-    // Make sure any pending text nodes are emitted before returning.
+    // Make sure all required pending text nodes are emitted before returning.
+    // This leaves "script", "style" and "svg" nodes text nodes intact.
     if (!isStopped())
-        m_treeBuilder->flush();
+        m_treeBuilder->flush(FlushIfAtTextLimit);
 }
 
 void HTMLDocumentParser::pumpPendingSpeculations()
@@ -635,7 +636,7 @@ void HTMLDocumentParser::pumpTokenizer(SynchronousMode mode)
     // There should only be PendingText left since the tree-builder always flushes
     // the task queue before returning. In case that ever changes, crash.
     if (mode == ForceSynchronous)
-        m_treeBuilder->flush();
+        m_treeBuilder->flush(FlushAlways);
     RELEASE_ASSERT(!isStopped());
 
     if (session.needsYield)

Source/core/html/parser/HTMLTreeBuilder.cpp

@@ -402,7 +402,7 @@ void HTMLTreeBuilder::processToken(AtomicHTMLToken* token)
 
     // Any non-character token needs to cause us to flush any pending text immediately.
     // NOTE: flush() can cause any queued tasks to execute, possibly re-entering the parser.
-    m_tree.flush();
+    m_tree.flush(FlushAlways);
     m_shouldSkipLeadingNewline = false;
 
     switch (token->type()) {
@@ -2689,7 +2689,7 @@ void HTMLTreeBuilder::processTokenInForeignContent(AtomicHTMLToken* token)
         return;
     }
 
-    m_tree.flush();
+    m_tree.flush(FlushAlways);
     HTMLStackItem* adjustedCurrentNode = adjustedCurrentStackItem();
 
     switch (token->type()) {

Source/core/html/parser/HTMLTreeBuilder.h

@@ -82,8 +82,9 @@ class HTMLTreeBuilder FINAL : public NoBaseWillBeGarbageCollectedFinalized<HTMLT
     // Done, close any open tags, etc.
     void finished();
 
-    // Synchronously empty any queues, possibly creating more DOM nodes.
-    void flush() { m_tree.flush(); }
+    // Synchronously flush pending text and queued tasks, possibly creating more DOM nodes.
+    // Flushing pending text depends on |mode|.
+    void flush(FlushMode mode) { m_tree.flush(mode); }
 
     void setShouldSkipLeadingNewline(bool shouldSkip) { m_shouldSkipLeadingNewline = shouldSkip; }