Live range splitting in new allocator

[Joern Rennecke]

> <<So please at least consider breaking debugging.  It _will_ get fixed -
> too many people care about it.  But it should not get fixed by
> generating bad code.
> >>
> 
> It seems perfectly reasonable to me to take this attitude when optimization
> is turned on.

The rule of thumb so far was that -O1 should be reasonably debuggable, while
-O2 is free to do more aggressive stuff that might break debugging.