Currently, the "likelyToTakeSlowCaseMinimumCount" threshold is 100, which is larger than the number of times a typical bytecode instruction will execute before being DFG compiled (that threshold averages around 1000/15 ~ 66). This means that the rare case profiling doesn't actually help us avoid overspeculating.
Created attachment 246634 [details] the patch
Comment on attachment 246634 [details] the patch r=me
Landed in http://trac.webkit.org/changeset/180137