I've split a paper once.  I wasn't inclined to do it, but the paper was becoming huge so I was a bit worried -- 60+ pages, few people read papers that long.  The other criterion was there was a natural way to split the paper into two useful papers.  One turned out to be a "largely survey" 40+ page paper (sometimes it's tricky getting papers published when they have a lot of survey material).  This allowed for the main result to be a short 20+ page paper. 

In general it comes down to a cost-benefit analysis.  Size of the paper is a consideration.  How interesting the results are, that matters.  Some people go overboard on fragmentation of their papers.  For example, if you're publishing more than 4 papers per year all on essentially the same topic, it leads to confusion among readers as to what results appear where, which papers depend on which.