Inst not used as union

[Scott Tringali]

Bert Wesarg wrote:

> can someone confirm that in FillLoopAddrs() we still use Inst not as a union?
> See patch for a working solution.
> And to which branch should I commit this?

Thanks for fixing, I wonder how I missed that.  I don't think the cast 
to Inst* is needed anymore either, but my memory is fuzzy.  See 
AddBranchOffset.

Personally, I'd say commit anything that's not a user-visible regression 
from 5.5 to the mainline.