[cfe-dev] Access specifiers and anonymous unions

[John McCall]

On May 28, 2010, at 10:46 AM, Abramo Bagnara wrote:

> Il 28/05/2010 19:25, John McCall ha scritto:
>> On May 28, 2010, at 9:00 AM, Abramo Bagnara wrote:
>>> I might imagine this is done
>>> for access checking, but as I'd guess that access checking is done
>>> traversing and checking *all* the nesting records (also anonymous ones)
>>> the union fields would be unaccessible from outside class S also if
>>> correctly marked as public.
>> 
>> Your guess is wrong;  we assume that the decl's access specifier correctly
>> describes the access of the member when considered as a member of its
>> declaring class.  There is no value in remembering the access of the decl
>> within the anonymous union, first, because it is impossible to access
>> anything from the context of the anonymous union (as opposed to
>> from its enclosing context) after it is parsed, and second, because
>> (as you've rightly observed) that access is always public.
> 
> I'm not sure to understand the difference you see between
> 
> class s {
>  struct {
>    int a;
>    int b;
>  };
> };
> 
> and
> 
> class q {
>  struct {
>    int a;
>    int b;
>  } c;
> };
> 
> and why you treat the former as a special case, but that apart it's not
> a problem in either way once there is the possibility to know access
> specifier written in source.

Because it is a special case.  The first case is an anonymous union;  the
second case is just a field of anonymous type.  In the first case, 'a' and 'b'
are visible as members of 's';  in the second case they are not.

This is C++ [class.union]p4:
  A union for which objects or pointers are declared is not an anonymous union.

>> Is there a value to remembering the original access specifier, i.e. public,
>> that I'm not seeing?  I would much prefer not to change the current behavior,
>> but we could certainly provide a method that reports the "direct" access
>> specifier, i.e. public if the DC is an anonymous struct or union.
> 
> The value is for pretty printer like applications and for source
> checkers. However to add an AccessSpecifierAsWritten flag inside the
> decl is not the ideal solution, it can't represent something like this:
> 
> class s {
> private:
>  int a;
> };
> 
> (undistiguishable from class s { int a; })
> 
> or
> 
> class s {
> public:
> private:
>  int a;
> };
> 
> To represent correctly the above we might decide to have a node for
> access specifiers at the same level of other decls.

Yes, I've been wanting something like this for some time.

John.
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.llvm.org/pipermail/cfe-dev/attachments/20100528/1d3b9c5a/attachment.html>