Will fix it soon.
This is revealed by the fix to https://bugs.webkit.org/show_bug.cgi?id=124137
OK, I think this is fixed, as of http://trac.webkit.org/changeset/159394.
The problem was that we were constant-folding a variable but only along one control flow path, which was resulting in its live range appearing to be truncated.
This was ultimately solved by ensuring that liveness in the original bytecode, rather than liveness in the DFG IR after optimization, determines whether or not a variable is preserved by OSR.