On Dec 6, 2010, at 10:33 AM, John McCall wrote:
>
> On Dec 6, 2010, at 10:25 AM, Argyrios Kyrtzidis wrote:
>
>>
>> On Dec 6, 2010, at 10:12 AM, John McCall wrote:
>>
>>>
>>> On Dec 6, 2010, at 10:06 AM, Argyrios Kyrtzidis wrote:
>>>
>>>> On Dec 5, 2010, at 8:33 PM, John McCall wrote:
>>>>
>>>>>
>>>>> On Dec 3, 2010, at 5:12 PM, Argyrios Kyrtzidis wrote:
>>>>>
>>>>>> Author: akirtzidis
>>>>>> Date: Fri Dec 3 19:12:11 2010
>>>>>> New Revision: 120872
>>>>>>
>>>>>> URL: http://llvm.org/viewvc/llvm-project?rev=120872&view=rev
>>>>>> Log:
>>>>>> Mark functions referenced by 'cleanup' attribute as used. Fixes
>>>>>> rdar://8728293
>>>>>
>>>>>>
>>>>>> Modified:
>>>>>> cfe/trunk/lib/Sema/SemaDeclAttr.cpp
>>>>>> cfe/trunk/test/Sema/warn-unused-function.c
>>>>>>
>>>>>> Modified: cfe/trunk/lib/Sema/SemaDeclAttr.cpp
>>>>>> URL:
>>>>>> http://llvm.org/viewvc/llvm-project/cfe/trunk/lib/Sema/SemaDeclAttr.cpp?rev=120872&r1=120871&r2=120872&view=diff
>>>>>> ==============================================================================
>>>>>> --- cfe/trunk/lib/Sema/SemaDeclAttr.cpp (original)
>>>>>> +++ cfe/trunk/lib/Sema/SemaDeclAttr.cpp Fri Dec 3 19:12:11 2010
>>>>>> @@ -1450,6 +1450,7 @@
>>>>>> }
>>>>>>
>>>>>> d->addAttr(::new (S.Context) CleanupAttr(Attr.getLoc(), S.Context, FD));
>>>>>> + FD->setUsed();
>>>>>> }
>>>>>
>>>>> Technically, you can use __attribute__((cleanup)) inside a template.
>>>>> This should use MarkDeclarationReferenced, and then there should be
>>>>> special instantiation logic for the attribute.
>>>>
>>>> Hmm, is this really the right approach ? If it's referenced from a
>>>> template which is not instantiated, it's not used from the codegen
>>>> perspective but how about from the user's perspective:
>>>>
>>>> -We emit a warning that the function 'foo' is unused.
>>>> -User removes function 'foo'
>>>> -Now we emit an error that 'foo' is not found.
>>>
>>> Oh, you mean if a function is referenced from a template definition that's
>>> never instantiated? Yeah, that's a valid user complaint. That's not
>>> specific to attribute((cleanup)), though; any reference without dependent
>>> ADL will have that issue.
>>
>> For this specific case, the lookup is always non-dependent, I think the
>> right user perspective is that the 'foo' is referenced if the attribute
>> references it even in uninstantiated templates.
>
> It could be dependent if we implemented overloading on it. :)
>
> I'm not sure why __attribute__((cleanup("foo"))) is different from, say, an
> explicit call to (foo)(x), or &foo. In all of these cases, from the user's
> perspective the function is used, because their code probably won't compile
> if you remove the declaration. So you should still use
> MarkDeclarationReferenced, and if we want to teach MarkDeclarationReferenced
> about this user notion of a dependent use, we've got one place to change.
I disagree with "from the user's perspective the function is used". The
function declaration is needed, sure, but that's very different from "used" in
the ODR sense, which implies that we need to have a definition. And I think
that user's don't expect to need definitions for something that's only used in
an uninstantiated template.
- Doug
_______________________________________________
cfe-commits mailing list
[email protected]
http://lists.cs.uiuc.edu/mailman/listinfo/cfe-commits
