The DFG graph has notions of Int32, cell, and array predictions, but no notion of double prediction. Furthermore, it has no notion of the least-upper-bound of predictions, or of distinguishing between a top prediction (where we know that the node can have every type because we've got confirmed sightings of assignments of all of the different types) and a bottom prediction (we have no information pointing in the direction of any particular type).
Created attachment 102097 [details] the patch This is performance-neutral, but shows sights of slight speed-ups in Kraken. It's meant as a stepping stone for other optimizations which would otherwise cause slow-downs due to overzealous int speculation.
Comment on attachment 102097 [details] the patch Clearing flags on attachment: 102097 Committed r91894: <http://trac.webkit.org/changeset/91894>
All reviewed patches have been landed. Closing bug.