V3 patch causes EH failures in ILP32 mode

[Mark Mitchell]

David Carlton wrote:

> Maybe one first step would be to only warn about a nonexistent copy
> constructor if the class has both a raw pointer member _and_ a
> destructor.  (Or (raw pointer && (destructor || assignment)),
> perhaps.)  And similarly for assignment.  

That sounds like a pretty good idea to me.  Reducing false positives is, 
in my mind, very important for these kinds of warnings.

 > It should at least reduce
> the false positive rate, and I would imagine (knowing nothing at all
> about the relevant bits of GCC) that it would be easier to implement
> than heuristics involving looking at the actual definitions of various
> member functions.

Yes, the full data flow stuff is definitely hard.  It could probably be 
done on tree-ssa, but I am opposed to issuing warnings from the back 
end.  The product I discussed above was a special-purpose analysis tool.

-- 
Mark Mitchell
CodeSourcery, LLC
mark@codesourcery.com