It looks like the benefit of this optimization may have been lost since the BOL optimizations went in ('once through' flag).
Disabling this optimization seems to have no impact on SunSpider performance, and I measure a 2% win on v8 w/ YARR interpreter by disabling.
If disabling will simplify the code, fix bugs & improve performance, than that is what we should do.
Do we have evidence that this is still a win?
Created attachment 94517 [details]
Comment on attachment 94517 [details]
Removing this optimization for now in r87109, open to re-landing if we can demonstrate a significant performance win.