[hendzen]

Perhaps one could even use histograms to estimate selectivity of the query's join predicates to pick a good ordering of joins?

This is actually pretty much how a basic relational query optimizer works in practice.