XMLReader only
Pros: fast, uses little memory
Cons: excessively hard to write and debug, requires lots of userland code to do anything useful. Userland code is slow and prone to error. Plus, it leaves you with more lines of code to maintain
XMLReader + SimpleXML
Pros: doesn’t use much memory (only the memory needed to process one node) and SimpleXML is, as the name implies, really easy to use.
Cons: creating a SimpleXMLElement object for each node is not very fast. You really have to benchmark it to understand whether it’s a problem for you. Even a modest machine would be able to process a thousand nodes per second, though.
XMLReader + DOM
Pros: uses about as much memory as SimpleXML, and XMLReader::expand() is faster than creating a new SimpleXMLElement. I wish it was possible to use simplexml_import_dom() but it doesn’t seem to work in that case
Cons: DOM is annoying to work with. It’s halfway between XMLReader and SimpleXML. Not as complicated and awkward as XMLReader, but light years away from working with SimpleXML.
My advice: write a prototype with SimpleXML, see if it works for you. If performance is paramount, try DOM. Stay as far away from XMLReader as possible. Remember that the more code you write, the higher the possibility of you introducing bugs or introducing performance regressions.