Re: C++0x Constructor Delegation take 2

[Jason Merrill <jason at redhat dot com>]

> http://gcc.gnu.org/ml/gcc-patches/2007-04/msg00620.html

Thanks a lot for the submission, and I'm sorry it has taken this long to 
review the patch.  Please do CC me on C++ patches, and ping me if a week 
goes by without a review.

I agree with Doug's comments.

The tests for flag_cpp0x now need to be "cxx_dialect != cxx98".

> +static void
> +perform_target_ctor (tree init)
> +{
> +  tree decl = current_class_ref;
> +  tree type = current_class_type;
> +
> +  if (init == void_type_node)
> +    init = NULL_TREE;
> +
> +  finish_expr_stmt (build_aggr_init (decl, init, 0));
> +
> +  if (TYPE_HAS_NONTRIVIAL_DESTRUCTOR (type))
> +    {
> +      tree expr = build_delete (type, decl, sfk_complete_destructor,
> +			   LOOKUP_NONVIRTUAL|LOOKUP_DESTRUCTOR, 0);
> +      if (expr != error_mark_node)
> +	finish_eh_cleanup (expr);
> +    }
> +}

This assumes that we'll only do delegation for the complete constructor, 
which is not the case.  For instance, this testcase breaks:

int c;
struct A
{
  A() { ++c; }
};

struct B: virtual A
{
  B(): B (1) {}
  B(int) {}
};

struct C: public B
{
  C(): A(), B() { }
};

int main()
{
  C cee;
  if (c > 1)
    return c;
  else if (c == 0)
    return 1;
}

We end up constructing the A subobject twice, because the B() 
not-in-charge constructor calls the B(int) in-charge constructor, which 
wrongly calls the A constructor.

Jason