Re: C++ and Java CNI: Check Java references

[Andrew Haley <aph at cambridge dot redhat dot com>]

Fergus Henderson writes:
 > > +   
 > > +   /* Strip off all pointer conversions and if the thing at the root is
 > > +      an integer we won't bother to check it.  The main reason for this
 > > +      is that offsetof is often defined as
 > > + 
 > > +      ((size_t)&((type *)0)->memb)
 > > + 
 > > +      which implicitly dereferences a null pointer.  Let's not check that.  */
 > 
 > Hmm... that seems like a bit of a hack.

Fair enough.

 > It's also common to write expressions such as `&p[0]',

I think it's impossible in this case.  These are Java objects, and
C++-style arrays can't occur.

 > and I don't think it's a good idea to check for null
 > pointers in that case.
 > 
 > It would be nicer, I think, to skip the check iff the pointer dereference
 > is the operand of unary `&' (address-of).  Though it may not be as easy
 > to implement...

Hmm.  I think that this would require trees to be rewritten and the
pointer check to be reoved.  Actually, I could avoid all this with a
__builtin_offsetof.  Maybe that would fix the problem much more neatly.

 > > +     tree cond = build (NE_EXPR, boolean_type_node, exp, integer_zero_node);
 > 
 > I'm not sure if it makes any difference, but wouldn't it be more
 > appropriate to use null_pointer_node rather than integer_zero_node?

I'm pretty sure it makes no difference, but I agree that this would be
better.

Thanks!

Andrew.