You are viewing a plain text version of this content. The canonical link for it is here.
Posted to issues@mesos.apache.org by "Michael Park (JIRA)" <ji...@apache.org> on 2016/04/04 04:22:25 UTC

[jira] [Commented] (MESOS-4611) Passing a lambda to dispatch() always matches the template returning void

    [ https://issues.apache.org/jira/browse/MESOS-4611?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=15223603#comment-15223603 ] 

Michael Park commented on MESOS-4611:
-------------------------------------

I'm happy to shepherd this if someone would like to write a patch.

The issue is not that lambdas are not implicitly convertible to a {{std::function}}. They absolutely are.
However, for a function parameter with a template parameter which is deduced, the argument must be of exact type.

Recall that the type of a lambda is an anonymous type generated by the compiler, not {{std::function}}.

With that in mind, consider this small example:

{code}
template <typename T> struct S{ S(int); };
template <typename U> void F(S<U>);
{code}

Although {{int}} is implicitly convertible to {{S<T>}} for any {{T}},
Given {{F(42)}}, the compiler has no idea what to deduce {{U}} to be.

We can however, do:

{code}
S<double> x = 42;  // The `double` is irrelevant.
F(x);  // `T` can be deduced to be `double`.
{code}

If you were to replace {{S}} with {{std::function}}, {{42}} with the lambda, {{F}} with {{dispatch}},
you'll see that everything is consistent.

Now, for the fix we need to use a similar technique as the one we use in {{future.hpp}}.
That is, rather than using {{std::function}}, we simply use a template parameter {{F}} and use
{{result_of}} to constrain it correctly.

For example,

{code}
template <typename R>
Future<R> dispatch(const UPID& pid, const std::function<R()>& f);
{code}

should be something like:

{code}
template <typename F>
Future<typename result_of<F()>::type> dispatch(const UPID& pid, F&& f);
{code}

It's not that straight-forward since there are 3 overloads, but as I said, I can help out with this.

> Passing a lambda to dispatch() always matches the template returning void
> -------------------------------------------------------------------------
>
>                 Key: MESOS-4611
>                 URL: https://issues.apache.org/jira/browse/MESOS-4611
>             Project: Mesos
>          Issue Type: Bug
>          Components: libprocess
>            Reporter: Kevin Klues
>              Labels: dispatch, libprocess, mesosphere
>
> The following idiom does not currently compile:
> {code}
>   Future<Nothing> initialized = dispatch(pid, [] () -> Nothing {
>     return Nothing();
>   });
> {code}
> This seems non-intuitive because the following template exists for dispatch:
> {code}
> template <typename R>
> Future<R> dispatch(const UPID& pid, const std::function<R()>& f)
> {
>   std::shared_ptr<Promise<R>> promise(new Promise<R>()); 
>  
>   std::shared_ptr<std::function<void(ProcessBase*)>> f_(
>       new std::function<void(ProcessBase*)>(
>           [=](ProcessBase*) {
>             promise->set(f());
>           }));
>   internal::dispatch(pid, f_);
>   
>   return promise->future();
> }     
> {code}
> However, lambdas cannot be implicitly cast to a corresponding std::function<R()> type.
> To make this work, you have to explicitly type the lambda before passing it to dispatch.
> {code}
>   std::function<Nothing()> f = []() { return Nothing(); };
>   Future<Nothing> initialized = dispatch(pid, f);
> {code}
> We should add template support to allow lambdas to be passed to dispatch() without explicit typing. 



--
This message was sent by Atlassian JIRA
(v6.3.4#6332)