so that the fmgr lookup only has to happen once per index scan and not once per tuple. Seems to save 5% or so of CPU time for an indexscan.