empty function optimizations

[Joern Rennecke]

> This would break the gurantee of generating identical code with or without
> debugging information.  Debugging information should be completely
> removable with a simple run of `strip'.

The rule of thumb is that -O1 should work reasonably well with debugging,
while -O2 may be more aggressive.
SO we could make this optimization default to on for optimize >= 2
(or just make it dependent on -fexpensive-optimization).
Of course, gcc has to tell the linker about the uses of a label are
only jumps for this optimization to ork safely, we can make the creation
of this information dependent on the flag setting.