In my experience, it doesn't make error recovery difficult, not really, even without exception throwing. Then again, I mostly stick to the panic-mode recovery:

In my experience, it doesn't make error recovery difficult, not really, even without exception throwing. Then again, I mostly stick to the panic-mode recovery: on unexpected input, record an error in a global list of errors, skip tokens until end of phrase, return a bogus node. Depending on what exactly "bogus" means, you end with error-productions or just completely made up but otherwise perfectly valid nodes somewhere deep inside your parse/abstract tree which is why recording presence of errors somewhere off-side is a good idea. I imagine a slighlty more elaborate error recovery approach could return different kinds of bogus nodes and check for them and process them differently in different productions.