So far our API and codebase passes shared pointers by value. This should be made const& to optimize the superflous refcounting in such cases. This was also mentioned in one of the talks here: http://channel9.msdn.com/Events/GoingNative/GoingNative-2012 I (milian) will try to find out which one it was if anyone doubts the above statement.
Instead of using IndexedString for string interning, use a thread local class which achieves the same without the QMutex overhead. Only convert to IndexedString when using the strings in the DUChain.
Instead of the monolithic QString/QByteArray in ParseJob::Contents we should have some API that yields the contents on a line-by-line basis. This would be especially good for files that are read from disk. For files that are open in the editor though we'll have to see what to do. Probably copying a QVector<QString> would be good there, since Kate stores the strings also on a line-by-line basis - this could potentially share some memory then.
Redo what I did for the IndexedString item repository for the other repository. This includes:
- use fast hash methods (no RunningHash, investigate hashlittle, ...) - no recursive mutex locking - no superfluous type conversions (like const char* <-> QString or similar)
Many of the AST nodes are quite large and/or very often instantiated. Especially UnqualifiedNameAST and NameAST are memory hogs. Either use more specialized AST nodes or use unions to reduce the memory size.