[KURL] Avoid re-hashing input if is already canonicalized

third_party/WebKit/Source/platform/weborigin/KURL.cpp

Index: third_party/WebKit/Source/platform/weborigin/KURL.cpp
diff --git a/third_party/WebKit/Source/platform/weborigin/KURL.cpp b/third_party/WebKit/Source/platform/weborigin/KURL.cpp
index 666ec1043795a491dc0698c7ab5ff94c498730f4..168152b9adb9aaae34851e57e7b7c3da6662b9fb 100644
--- a/third_party/WebKit/Source/platform/weborigin/KURL.cpp
+++ b/third_party/WebKit/Source/platform/weborigin/KURL.cpp
@@ -718,10 +718,17 @@ void KURL::init(const KURL& base,
                 const String& relative,
                 const WTF::TextEncoding* queryEncoding) {
   if (!relative.isNull() && relative.is8Bit()) {
-    StringUTF8Adaptor relativeUTF8(relative);
-    init(base, relativeUTF8.data(), relativeUTF8.length(), queryEncoding);
-  } else
-    init(base, relative.characters16(), relative.length(), queryEncoding);

[esprehn, 2016/10/25 22:23:50]

We go down this path when relative.isNull() which is strange.

[Charlie Harrison, 2016/10/27 02:12:36]

The null check is just for the is8Bit check which reaches into m_impl. Other
than that the code is robust to nulls.

+    // Avoid using StringUTF8Adaptor because we want the possibly non-ascii data
+    // exposed to init<CHAR> as a String to avoid re-allocation.
+    String utf8 = relative.containsOnlyASCII()
+                      ? relative
+                      : String::fromUTF8(relative.utf8());
+    init(base, reinterpret_cast<const char*>(utf8.characters8()), utf8.length(),
+         queryEncoding, utf8);
+  } else {
+    init(base, relative.characters16(), relative.length(), queryEncoding,
+         relative);
+  }
   initProtocolIsInHTTPFamily();
   initInnerURL();
 }
@@ -730,7 +737,8 @@ template <typename CHAR>
 void KURL::init(const KURL& base,
                 const CHAR* relative,
                 int relativeLength,
-                const WTF::TextEncoding* queryEncoding) {
+                const WTF::TextEncoding* queryEncoding,
+                const String& originalRelative) {
   // As a performance optimization, we do not use the charset converter
   // if encoding is UTF-8 or other Unicode encodings. Note that this is
   // per HTML5 2.5.3 (resolving URL). The URL canonicalizer will be more
@@ -752,9 +760,15 @@ void KURL::init(const KURL& base,
                                    base.m_parsed, relative, relativeLength,
                                    charsetConverter, &output, &m_parsed);
 
-  // See FIXME in KURLPrivate in the header. If canonicalization has not
-  // changed the string, we can avoid an extra allocation by using assignment.
-  m_string = AtomicString::fromUTF8(output.data(), output.length());
+  // If the canonicalization has not changed the string, avoid re-allocating.
+  // Note: because we are canonicalizing into narrow characters, this check will
+  // always be false for 16 bit inputs.
+  if (output.length() == relativeLength &&
+      !memcmp(output.data(), relative, relativeLength)) {
+    m_string = AtomicString(originalRelative);
+  } else {
+    m_string = AtomicString::fromUTF8(output.data(), output.length());
+  }
 }
 
 void KURL::initInnerURL() {