Description Wladimir Safonov 2011-07-11 17:06:31 EDT

Created attachment 199451 [details]
Using INodes methods for obtaining next/prev/first/last siblings

Apparently migration to the new INodes API introduced a considerable performance leak in the class org.eclipse.xtext.parsetree.reconstr.impl.NodeIterator. In several places it makes use of the Lists.newArrayList(Iterable<INode>) to create a list of child nodes where eventually only one element will be picked up. This causes all children to be iterated through and collected in a new list. However dedicated methods on the INode can deliver the element of interest in one single get operation.

I attached a patch, which reduced serialization time for my model from 30 sec down to approx 2 sec. The model has a simple containment tree without any cross-references and has ~5 elements depth, but ~9K elements breadth, which is probably the reason of why iterations through the nodes take most of the serialization time.