Description
Assuming node visitor interface is a good idea (is it? not 100% clear), there are several problems with existing xml_tree_walker:
- ::begin/::end are pretty pointless - could be called from external code
- No way to get callbacks before/after the subtree - this is what begin/end could have been. This was mentioned (if memory serves) in https://code.google.com/p/pugixml/issues/detail?id=219 but this link is now dead.
- No way to skip a subtree. This is mentioned in issue wish to add another xml_node::traverse function meber #77.
While manually implementing a recursive traversal is pretty trivial, all traversals that pugixml implements are stackless (which is not as simple/concise to write yourself), so there seems to be some value in a generic traversal. There are still several implementation options though:
- Interface with virtual functions. This will have to be a new interface to maintain binary compatibility - xml_tree_visitor or something like that.
- Pseudo-interface with templated member functions. This can be faster, although in a non-LTO build it's a weird tradeoff - templated implementations can't inline node structure accessors, while an implementation based on virtual dispatch can't inline user code. With LTO or header-only mode the approach based on compile-time polymorphism should be faster though. This can be seen as an extension of xml_node::find_node.
- Tree iterator with ability to descend into a subtree or skip it, so that the user code is in control of the flow.
These need to be evaluated for the performance and versatility in various scenarios. 1 and 2 are extensions of existring ::traverse() support while 3 is basically a separate idea.
Finally, there is - as it is - quite a few ways to traverse the nodes in pugixml. Ideally if we do need a new way it has to supplant the old one - that is, we should at the minimum deprecate existing ::traverse in favor of whatever the new interface is.