In C++11, attributes were added as a way to standardized features such as gnu
The language provisions for standard attributes as well as non-standard attributes through the use of namespaces, though the behavior of non-standard attributes was only settled for C++17. And sadly, as of 2018, neither GCC nor MSVC offer their vendor-specific attributes though the portable C++ standard syntax.
Most standard attributes were added in C++14 and 17. You can find a list on cppreference.
At some point before C++11 came out, the C++ Standard draft defined the
[[final]] attributes. These features were later converted to contextual keywords
final. The original proposal even suggested that default and deleted methods could be made possible with the use of attributes. We now know these features as
And ever since, when adding a language feature to C++, the question of whether that feature would be better served by a keyword or an attribute needs to be asked and answered.
It recently came up for the now accepted
[[no_unique_address]] and the currently discussed
[[move_relocates]] (which would, to the extent of my understanding, be akin to a destructive move, a nice performance optimization in some cases).
Systematically, a committee member will point out that “A compiler should be free to ignore attributes”. One other will add “Well, actually, an attribute should not change the semantic of a program”. Or more accurately,
Compiling a valid program with all instances of a particular attribute ignored must result in a correct implementation of the original program
This is sort of an unwritten rule, it’s not actually anywhere in the standard, though I found this quote in the initial attribute proposal:
What makes this a good candidate for attributes is that code that runs with these attributes also runs identically if the attributes are ignored, albeit with less type checking.
It makes a lot of sense in the case of non-standard attributes. We should let the dark past of not-portable code and vendor-extensions behind us. So having non-standard attributes being simply skipped over is very useful and desirable indeed.
But what about standard attributes ? The compilers are free to ignore those too.
So let’s say you use a
[[fallthrough]] attribute and have a 0-warning policy (say you compile with -WError), the compiler is free to emit a warning and to fail the build.
In the case of
sizeof may return a different result depending on whether the attribute is ignored or not, thereby letting you affecting the semantic of the program. Which both show that the committee doesn’t necessarily follow their own rules, but most importantly that attribute being ignorable do not match the developer’s intent.
Even as people learned to expect that vendors often implement the standard in partial and opinionated ways, they probably don’t sprinkle attributes just for kicks, or to give their code some artsy Christmas ornaments. If someone goes to the trouble of marking a function
[[nodiscard]], they probably really want the result of a function to be checked. Maybe not checking the return value could, for some reason, lead to a critical failure. Rockets blowing up, patients dying.
There is, for existing attributes, no implementability concern either. They all can be implemented on all hardware as they don’t impose hardware requirement. I suppose there are some hardware for which
[[caries_dependencies]] makes no sense, but on such hardware,
std::memory_order wouldn’t make sense either, making the point moot.
I’ve tried to ask several committee members the nagging question : Why ? Why attributes need to have no semantic meaning ?
The answer I got was : Because.
And it was hard for me to find more rationale behind that. And there is nothing quite irksome as a rule without rationale.
A reason that I was given is that the committee might, in absence of strong guidelines, use attributes to shovel more keywords into the language, since adding attributes is easier than adding keywords: introducing keywords might break someone’s code and understandably calls for a stronger motivation. And letting attributes be everything might make the language an ungodly soup of attributes.
That is certainly a valid concern. However, does the committee really needs to impose rules onto itself ? They still need to vote on each and every standard attribute that goes into the standard, and need to study the relevance of the attribute with or without the existence of this weird unwritten rule.
I don’t think there is a good general answer to whether attributes should be keyword, or whether they should be imbued with semantic meaning or not.
Instead, that question needs to be answered on a per-attribute basis. Should
alignas be a keyword ? Maybe not. Should
[[fallthrough]]be one ? Probably; that’s a good example of shoving a keyword as an attribute to avoid breaking user’s code.
Ultimately, that sort of consideration is highly subjective, but I doubt introducing seemingly arbitrary rules makes the design process any easier, probably quite the opposite.
Instead, each proposed standard attribute (or keyword) needs to be studied on its own merits, and maybe we will find attributes for which it makes sense that they could be ignored — Which I don’t think is the case for any of the existing attributes.
This might sound like bickering, and to some extent, it is. However it might matter in the context of reflection.
Reflection on attributes
I think reflection on attributes may be the most important aspect of reflection. Let say you want to serialize a class, and using reflection to visit all the members, you may need to filter out some members that you don’t want to serialize. One way to do that would be to (ab)use the type system but a better system would probably be to use attributes to tag the relevant members. That could open the door to amazing serialization, RCP and ORM libraries (…even though you probably should not use an ORM !)
Most people appear to see the value of such feature, however some argue that it would be better to add another syntax, that could be called decorator. It would essentially be the same thing as attributes, used in the same place as attributes, but with a new syntax that would be exempt of the “attribute should be ignorable” debate.
To me this really does not make sense. First, if a user chooses to impart semantic meaning to attributes by the mean of reflection, then it’s the user’s choice rather than the compiler’s, so portability concerns do not apply. And of course, this feature may and would be used to develop framework-specific behaviors and idioms, which some people seem to be rather strongly against. But that is a need that exists and is emulated today by the means of complicated macros or code generation, often both (Qt).
Adding a new syntax, giving it another name over some philosophical subtlety that will be loss on non-expert, would almost certainly add no value to the language. And would of course add delay while that new syntax is argued about. C++ is running out of tokens.
For example, which of the following do you think is the most readable ?
[[deprecated]] QSignal<void> f();
[[deprecated]] @@qt::signal@@ void f();
[[deprecated]] [[qt::signal]] void f();
There are other challenges with reflecting on attributes. Attributes can take parameters that are presently defined as being a token soup. Hem, I mean token-sequence sorry. A well balanced token-soup-sequence. Which I guess makes total sense for tooling and compilers. But is pretty pointless when it comes to reflection.
I think there are solutions to this problem, namely that reflectable-attributes be declared in the source before use and allow literals and
constexptras their parameters, rather than arbitrary tokens.
I’ve put more details on that on GitHub. Note that this is not a proposal and I don’t intend to make one, at least not while the “attribute should be ignorable” mantra exists.
In the end, attributes could have a variety of uses:
- Compiler directives for optimizations and diagnostics
- Instructions and meta data for external tooling
- Filtering and decoration by the use of both reflection of code injection (code injection, would let people use attributes as full blown decorators similar in expressiveness to the feature of the same name offered by Python)
- Contracts (
[[ensures:]]) ; though these syntax are different enough that i’m not sure they still qualify as attributes.
But as things stand, I feel like attributes are criminally underused and crippled.
What attributes would you like to see ?