This comes from the profile in bug 229391. In that profile, we're spending a good bit of time (comparable to total time spent in painting, basically) in InitScanner() and ReleaseScanner(), because we have thousands of inline style rules being set in a row. The patch coming up makes CSS parsers not actually delete the scanner when they ReleaseScanner() and makes scanners deal with being reinitialized. It saves us about 1% of the total time on the testcase in the URL field.

Note to self: the line in nsCSSScanner.h that declares mURI is now: nsCOMPtr<nsIURI> mURI; // Cached so we know to not refetch mFileName per grammar nit on irc.

Comment on attachment 161824 [details] [diff] [review] Patch I'm not convinced ClearScanningState deserves to be its own function. It doesn't really clear all the state, so it doesn't seem useful to callers other than Init, and there is only one caller. Also, is there any reason CSSParserImpl::mScanner needs to be a pointer?

Fixed

So, maybe I'm missing something, but it looks like you're never setting mScannerInited to true.

Good catch. Checked in the fix.