objects_must_conflict_p

[Eric Botcazou]

> Yuck!  It looks like we have to say they only are guaranteed to conflict
> if they have the *same* alias set.

This elegantly (and efficiently) solves the problem for aggregates vs 
aggregates, as well as for other combinations. I think this is probably good 
enough for the 3.3 branch (PR opt/11198 is a regression). Thanks!

But we pessimize a bit: for example int vs array of ints, or even different 
types of arrays of ints. I don't know whether any refinement is really worth 
implementing though.

-- 
Eric Botcazou