fix opt/8634

[Zdenek Dvorak]

Hello,

> As for Zdenek's proposed prepass patch... I think it would be
> better to not iterate over the entire instruction stream twice.
> We could, for instance, change the return value of purge_addressof_1
> to be three-state, with the new return location indicating that
> we should re-process this insn at the end.

why not? Is this time critical? Unless it is, I believe a bit slower
but easier to understand solution should be preferable.

Zdenek