When I've been banging on a hard problem for a long time with no results, I find it helpful to change the environment (as everyone else has been saying, go outside). Then I come back and start from scratch. (Even if your new version is broken, you can sometimes gain some insight into why the first version is broken.) Sometimes I draw pictures of the problem or data structures on paper.

One time I worked for three days on a search algorithm and it kept having the kind of bugs that, when you fix them, cause others to appear. I finally got so frustrated I went home an hour early. When I got home, I slammed a working version out from scratch in 30 minutes.

http://www.nodewarrior.org/chris/