Incorrect bitfield aliasing with Tree SSA

[Adam Nemet]

Richard Guenther writes:
> Hm, you enlarge SFTs by 8 bytes this way.  Watch the memory stats in case
> this turns out to be bad.  [memory tags waste quite some padding and have
> unnecessarily large vars]

There does not seem to be any degradation on average.  I compared the
'Total' 'Allocated' entry under -fmem-report for cc1 .i files.

I will check in the fix per Dan's approval tomorrow if there are no
futher issues.

Adam