Do not construct a ScoreDoc for every non-zero hit, but only for those in the current top scoring set. This makes a substantial performance improvement for queries that match lots of documents.
lucene/java/trunk/src/java/org/apache/lucene/search/IndexSearcher.java