C++: Internal compiler error

[Martin v. Loewis]

>   +	    /* VLAs never work as fields. */
>   +	    if (decl_context == FIELD && size != NULL_TREE
>   +		&& !processing_template_decl
>   +		&& !TREE_CONSTANT (cp_convert (sizetype, size)))
>   +	      {
>   +		cp_error ("size of member `%D' is not constant", dname);
>   +		/* Proceed with arbitrary constant size, so that offset
>   +		   computations don't get confused. */
>   +		size = integer_one_node;
>   +	      }
>   +
> 
> I'm not sure this is quite good enough.  Either it isn't, or
> compute_array_index_type is doing a lot more work than it needs to!
> For instance, does this handle `const int I = 7' as an array bound?
> Similarly `(int) (int) I'?  

Both work fine. cp_convert comes back with a constant in these cases,
as it uses decl_constant_value if it find that the declaration is a
constant expression. I only need the check for 'integral constant',
not the actual value.

> By the way, why does compute_array_index_type convert the argument to
> ssizetype? 

I don't know. It originates from the old grokdeclarator fragment

  /* The index is a signed object `sizetype' bits wide.  */
  tree index_type = signed_type (sizetype);

which predates egcs.

> And what if the original bound had type `long long'?  Aren't we then
> truncating the array?

No, it gives the error "overflow in array dimension".

> I wonder if you shouldn't check for this problem later in the game,
> after the array type is created a few lines below your patch.  Then,
> all you have to do is check that the TYPE_DOMAIN is TREE_CONSTANT, and
> fix up the type (for error-reporting purposes) if not.

I think the type cannot be changed afterwards, since it is already
hashed. Also, in this case, it would produce two diagnostic messages
for the same problem under -pedantic (first would be a pedwarn that
VLAs are not standard C++, and then it would cp_error that they won't
work at all as members).

Regards,
Martin