[PATCH] SemaCXX: Support templates in availability attributes

[Duncan P. N. Exon Smith via cfe-commits]

> On 2016-Feb-23, at 11:18, Aaron Ballman <aaron at aaronballman.com> wrote:
> 
> On Tue, Feb 23, 2016 at 1:52 PM, Manman Ren <mren at apple.com> wrote:
>> This patch looks good to me. But I am not sure if Aaron has any comment.
>> 
>> On Feb 22, 2016, at 6:19 PM, Duncan P. N. Exon Smith <dexonsmith at apple.com>
>> wrote:
>> 
>> 
>> On 2016-Feb-22, at 17:24, Manman Ren <mren at apple.com> wrote:
>> 
>> 
>> 
>> On Feb 8, 2016, at 8:17 PM, Duncan P. N. Exon Smith <dexonsmith at apple.com>
>> wrote:
>> 
>> This patch adds support for templates in availability attributes.
>> - If the context for an availability diagnostic is a
>> `FunctionTemplateDecl`, look through it to the `FunctionDecl`.
>> 
>> 
>> AvailabilityResult Decl::getAvailability(std::string *Message) const {
>> +  if (auto *FTD = dyn_cast<FunctionTemplateDecl>(this))
>> +    return FTD->getTemplatedDecl()->getAvailability(Message);
>> +
>> AvailabilityResult Result = AR_Available;
>> 
>> 
>> This looks generally correct to me.
>> The UnavailableAttr is attached to the FunctionDecl, not the
>> FunctionTemplateDecl, so looking through sounds right.
>> 
>> - Add `__has_feature(attribute_availability_in_templates)`.
>> 
>> 
>> @Aaron, any comment on this?
>> This patch adds extra support for Availability attribute (similar to
>> attribute_availability_with_strict in r261548).
>> Not sure if has_attribute can be used for this purpose.
> 
> Given that we're already using __has_feature for the rest of the
> availability attribute stuff, I think it's better to keep it all
> grouped together instead of checking for some features with
> __has_feature and others with __has_attribute.

Besides that argument, this isn't adding any attributes, just checking
how they behave.

> If I understand
> properly, this is taking code that would have previously been
> ill-formed and making it well-formed, and that's why the feature
> testing macro is required?

Exactly.  Previously this was ill-formed:
```
class Unavail __attribute__((unavailable));

template <class T>
void foo(Unavail&) __attribute__((unavailable));
```

Same for `__attribute((availability(macosx,unavailable)))`, and other
triggers of "unavailable".

> 
> Thanks!
> 
> ~Aaron
> 
>> 
>> Manman
>> 
>> 
>> Is there anything else I should be testing to be sure availability
>> works correctly in templates?
>> 
>> 
>> Maybe
>> test<UnavailableClass>()
>> calling unavailable function from an unavailable template function
>> calling an unavailable template function
>> 
>> I think these all work with the current compiler. But I am not sure if we
>> have existing test coverage.
>> 
>> 
>> Thanks for the ideas; let me know if you have any others.
>> 
>> Can you have a look at the new patch?
>> 
>> Cheers,
>> Manman
>> 
>> I'm working on a patch to add
>> availability markup to the libc++ headers, and this is the only
>> problem I've hit so far.  Anyone have thoughts on other things I
>> should test?
>> 
>> 
>> <0001-SemaCXX-Support-templates-in-availability-attributes.patch>