From some profiles, it seems like a lot of time is spent on processing macro; this is caused by the fact that each macro is reparsed in isolating using the libclang API. It would be desirable to parse all macros in one shot, although then it could be difficult to understand which macros parsed successfully and which ones did not.