[cfe-dev] Creating New Types

[Bill Wendling]

On Aug 7, 2007, at 2:34 PM, Chris Lattner wrote:
> On Aug 7, 2007, at 1:10 PM, Bill Wendling wrote:
>> On 8/7/07, Chris Lattner <clattner at apple.com> wrote:
>>>>> (This is related to the description of "casting away constness" in
>>>>> the C++ standard [5.2.11p8].)
>>>>
>>>> How so?  const_cast should never change the underlying type (e.g.
>>>> from float to char).  In fact, my understanding of const cast is  
>>>> that
>>>> it isn't a "recursive" process, it just looks at the top-level
>>>> qualifiers and pointer/reference.
>>>
>>> Actually, const_cast has to *compare* the src and dest type
>>> recursively, but I still don't think it has to *reconstruct* a  
>>> mix of
>>> the two.
>>>
>> It seemed to say that if one pointer type with N pointers could be
>> implicitly cast to a "subtype" of another pointer type with M
>> pointers, where N < M for some type T, then we aren't casting away
>> constness. That's what I was doing with my example. I was going to  
>> use
>> the implicit cast checking stuff to make sure that this is valid. I
>> suppose it's possible to compare it recursively, if I can make sure
>> that the checking is valid for each step.
>
> I'm not certain I know what you mean, but I don't think that is  
> valid.  Can you give me an example C++ function that does this for  
> concreteness?
>
I think that there's a misunderstanding of why I'm doing this. I'm  
not doing this for the const_cast operator. But it's a check that  
needs to be performed for all of the other casting operators. For  
instance, take the reinterpret_cast which can change the underlying  
type; it's not allowed to cast away constness. So this is allowed:

int * ip;
float ** fpp = reinterpret_cast<float**>(ip);

But this isn't:

int const * ip;
float ** fpp = reinterpret_cast<float**>(ip);
float const ** cfpp = reinterpret_cast<float const **>(ip);

(T const * cannot be implicitly converted to T *.) But this is:

int const * ip;
float * const * fpp = reinterpret_cast<float* const *>(ip);

My understanding is that this is because there's an implicit cast  
between T const * to T const * (identity), which would be the check  
that needs to be performed due to [5.2.11p8].

-bw