Delete function with message

[Guillaume Racicot]

Hi there, I got an idea while writing a library. There are multiple cases where I got this:

template<typename T, std::enable_if_t<std::is_polymorphic<T>::value, int> = 0>
void doThings(T) {
    // ...
}

template<typename T, std::enable_if_t<!std::is_polymorphic<T>::value, int> = 0>
void doThings(T) {
    static_assert(!std::is_same<T, T>::value, "T must not be polymorphic.");
}

This code gives me clear error messages, nice. But there's a catch: You can't use SFINAE to check if the function is callable:

template<typename T>
auto callDoThingsMaybe(T t) -> decltype(doThings(t)) {
    doThings(t);
}

template<typename T>
auto callDoThingsMaybe(T t) -> decltype(doOtherThings(t)) {
    doOtherThings(t);
}

Where doOtherThings(T) only exists if T is polymorphic.

This code won't work, because  doThings(T) is not a SFINAE failure, and calling it will result in a hard error. If I would remove the overload that throws the static assert, SFINAE will work, but we will loose the nice error message.

I would like to make a proposal to enable sfinae with nice error messages.

Look at this alternative:

template<typename T, std::enable_if_t<!std::is_polymorphic<T>::value, int> = 0>
void doThings(T) = delete("T must not be polymorphic."); // proposed syntax

This will follow deleted function rules and give a nice error message when it is called. SFINAE will still work, and the compiler will still print nice error messages for users.

What do you think?

[Jim Porter]

On 7/5/2016 3:23 PM, Guillaume Racicot wrote:
> I would like to make a proposal to enable sfinae with nice error messages.

[snip]
> What do you think?

Isn't this what Concepts will do? You could constrain your function to
take types meeting a "Polymorphic" concept and, if you pass something
that's not polymorphic, the compiler would say something like
"constraint not satisfied because T is not a Polymorphic type".

- Jim

[Ville Voutilainen]

I think having such a facility would be immensely useful. One
food-for-thought point, though:
why not do it with an attribute?

template<typename T, std::enable_if_t<!std::is_polymorphic<T>::value, int> = 0>

[[diagnose("T must not be polymorphic.")]]
void doThings(T) = delete;

[Ville Voutilainen]

Correct. The reason why I envisioned the attribute approach is that I
envisioned a general
facility that would be "implementation, if you're going to issue a
diagnostic here, please
include the following text in the diagnostic".

One might also think of talking to the Contracts proposal authors
about means to do compile-time
contracts. It seems like the current proposal focuses on run-time contracts:
http://open-std.org/JTC1/SC22/WG21/docs/papers/2016/p0380r0.pdf

[Guillaume Racicot]

It would even better with concepts. It would actually look like this:

void doThing(auto&) = delete("For some reason X, you must send a Container")

void doThing(Container&) { /* ... */ }

It will provide more information for the users and give insights of what should be done to make things work.

[Jim Porter]

On 7/5/2016 4:14 PM, Ville Voutilainen wrote:
> Correct. The reason why I envisioned the attribute approach is that I
> envisioned a general
> facility that would be "implementation, if you're going to issue a
> diagnostic here, please
> include the following text in the diagnostic".

If this could be expanded to a recommendation to the compiler: "please
don't include a lot of extraneous error information", I'd be happy to
see it. Errors in nested template instantiations can get awfully
verbose, and I haven't come up with a good general-purpose way to keep
them brief.

- Jim

[Guillaume Racicot]

Indeed, it could be an attribute, but I think it feels nicer to get it directly in the delete syntax, as it is only an extension of the delete syntax.

I get the point tough, and one could easily write an implementation as a compiler extension with that.

I'm in the process of writing the proposal on github, I will include this alternative syntax. Thanks!

[Guillaume Racicot]

Here's the git repo of the proposal. Any contribution are welcome!

https://github.com/gracicot/delete-message-proposal/

[T. C.]

I'm fairly sure I've seen something like this before...

...ah yes, http://wg21.link/N4186

[Ville Voutilainen]

On 7 July 2016 at 02:36, T. C. <rs2...@gmail.com> wrote:
> I'm fairly sure I've seen something like this before...
>
> ...ah yes, http://wg21.link/N4186

Indeed, and it's also
https://cplusplus.github.io/EWG/ewg-active.html#152

[Guillaume Racicot]

No way!! Even the same syntax!