Preload scan external css for @import

third_party/WebKit/Source/core/fetch/CSSStyleSheetResource.cpp

Index: third_party/WebKit/Source/core/fetch/CSSStyleSheetResource.cpp
diff --git a/third_party/WebKit/Source/core/fetch/CSSStyleSheetResource.cpp b/third_party/WebKit/Source/core/fetch/CSSStyleSheetResource.cpp
index 20bd6266d884287824c21202c8fbcb8805d2bda2..1167300feb0ae89577dbc2c370a3e7f99315787e 100644
--- a/third_party/WebKit/Source/core/fetch/CSSStyleSheetResource.cpp
+++ b/third_party/WebKit/Source/core/fetch/CSSStyleSheetResource.cpp
@@ -103,6 +103,14 @@ const String CSSStyleSheetResource::sheetText(MIMETypeCheck mimeTypeCheck) const
     return decodedText();
 }
 
+void CSSStyleSheetResource::appendData(const char* data, size_t length)
+{
+    Resource::appendData(data, length);
+    ResourceClientWalker<StyleSheetResourceClient> w(m_clients);
+    while (StyleSheetResourceClient* c = w.next())
+        c->didAppendData(this);

[kinuko, 2016/04/01 05:19:19]

So we call streamDecodedText in each appendData?  The code looks simple enough
but are we sure that this wouldn't add non-trivial computational/memory overhead
in worst cases?

(As has been commented in each experiment) in general it'd be good to make it
clearer what metrics we'll be watching to make sure this doesn't add regression.

[Charlie Harrison, 2016/04/01 13:07:18]

Yeah it's non ideal, but I couldn't see another way to avoid the copy without
just blindly trusting that the css is ascii. For metrics, Bryan is developing
parse time metrics which we'd look for for a regression.

+}
+
 void CSSStyleSheetResource::checkNotify()
 {
     // Decode the data to find out the encoding and keep the sheet text around during checkNotify()