C++11 attributes for C11 (was Re: [PATCH] Generalized attribute support)

[Aaron Ballman]

On Mon, Apr 21, 2014 at 5:17 PM, Richard Smith <richard at metafoo.co.uk> wrote:
> On Mon, Apr 21, 2014 at 2:04 PM, Aaron Ballman <aaron at aaronballman.com>
> wrote:
>>
>> On Sat, Apr 19, 2014 at 7:05 PM, Alp Toker <alp at nuanti.com> wrote:
>> > Small update for the list..
>> >
>> > We discussed this at LLVM Europe 2014 and came to the conclusion that we
>> > might as well offer the feature behind an -f flag.
>> >
>> > This should satisfy user requests for C++11-style attributes in C11
>> > without
>> > prejudicing any standards process.
>>
>> Why just C11? If we're using a feature flag, wouldn't it make sense to
>> expose this for C++98, or C99, or ObjC (presuming there are no
>> ambiguities)?
>
>
> Yes, if we have this, it should work in all language modes unless there's
> some fundamental incompatibility.
>
>>
>> > Thanks to the many detailed use-cases
>> > people sent in that have made me more comfortable picking this up again.
>> >
>> > So, I've got a slightly improved version of the patch and will aim to
>> > rebase
>> > and post it to the list for review in the coming weeks.
>> >
>> > All we need now is great name for the -f flag ;-)
>>
>> I like my bikesheds to be blue!
>>
>> -fcxx-style-attributes
>> -fcxx-attributes
>>
>> I don't have a strong opinion on the name, but I do think that the
>> feature flag should warn in nonsense situations (for instance
>> -fno-cxx-attributes -std=c++11), or maybe even -fcxx-attributes
>> -std=c++11 (since it's redundant, but I'm less tied to this).
>
>
> -fc++-attributes? We already have +s in -W and -std flags, why not in a -f
> flag? =)

That would be fine by me as well, though it does read a bit oddly.

> I don't see any particular reason to disallow or warn about -std=c++11
> -fno-c++-attributes, even though I don't immediately see a good reason to
> use that flag combination.

I was presuming it would mean we would have "C++11" where C++
attributes are not supported, which strikes me as being a bit weird
(and so we would simply ignore the flag and continue to support the
attributes). But if the switch works (the attributes aren't
supported), then I would agree we shouldn't warn.

> We should allow -std=c++11 -fc++-attributes, so
> that software that compiles with or without c++11 but needs c++ attributes
> can use 'clang $std -fc++-attributes'.

I agree with you (and Hal) on this.

>
> I think the only remaining argument against this is that we've historically
> tried to avoid creating language dialects by allowing some language features
> but not others. We've seen the damage done to C and C++ by GNU extensions,
> and we want to avoid doing the same thing ourselves. The counterpoint is
> that the alternative (__attribute__) in this case is already a language
> extension.

I agree that it's a concern we should be giving weight to.

~Aaron