[libcxx] r216909 - Fix PR#20834 - 'is_trivially_destructible yeilds wrong answer for arrays of unknown bound' Thanks to K-ballo for the bug report. Update a few of the other tests while we're here, and fix a typo in a test name.

[John McCall]

On Oct 1, 2014, at 5:09 PM, Argyrios Kyrtzidis <kyrtzidis at apple.com> wrote:
>> On Oct 1, 2014, at 4:52 PM, John McCall <rjmccall at apple.com> wrote:
>> 
>> On Oct 1, 2014, at 4:34 PM, Argyrios Kyrtzidis <kyrtzidis at apple.com> wrote:
>>> + John
>>> 
>>> John, could you advice here ?
>>> 
>>>> On Oct 1, 2014, at 4:20 PM, Marshall Clow <mclow.lists at gmail.com> wrote:
>>>> 
>>>> On Oct 1, 2014, at 3:36 PM, Marshall Clow <mclow.lists at gmail.com> wrote:
>>>>> 
>>>>> Richard Smith pointed out that destroying a pointer to an objective-C class might require an ARC call, and so the compiler needs to see the type to figure out if it is_destructible.
>>>>> 
>>>>> Note that is_destructible<ObjCForwardClass**>::value is true, so it’s just pointers to objective-C objects, not pointers in general.
>>>> 
>>>> Richard also pointed me to: http://clang.llvm.org/docs/AutomaticReferenceCounting.html#retainable-object-pointers
>>>> 
>>>> So, while I agree with Argyrios that this changed the behavior of is_trivially_destructible for forward-declared objective-C classes,
>>>> I am now convinced that the old behavior was incorrect, and the new behavior is “better”.
>>>> 
>>>> One could argue that a forward-declared objective-C class is not a complete type, since the compiler doesn’t know how to destroy it, and if so, then is_destructible is not required to give any answer at all ("T shall be a complete type”).
>> 
>> Pointers to forward-declared Objective-C classes are still complete types.  They should be both destructible and (outside of ARC) non-trivially destructible.
>> 
>> I assume the problem here is ObjC property syntax, which usually looks like “foo.prop” where foo has ObjC pointer type.  Type-checking a property access really does require the type to be complete.  However, pseudo-destructor calls are not property accesses, and should not require the pointee type to be complete.
>> 
>> Does that answer your question?
> 
> Are you suggesting that we make compiler changes to accept the pseudo-destructor call for ObjC pointers

Yes.  We already accept the pseudo-destructor call if the interface type is complete, presumably without actually looking anything up in the interface.  We just unnecessarily bail out early if the interface is incomplete, despite it being a pseudo-destructor call.

> , and if yes, what would the codegen behavior be for ARC ?

That doesn’t need to change.

John.
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.llvm.org/pipermail/cfe-commits/attachments/20141001/448a5624/attachment.html>