Re: [boost] [result_of] fails with clan (C++11)

Subject: Re: [boost] [result_of] fails with clan (C++11)
From: Eric Niebler (eric_at_[hidden])
Date: 2012-11-28 21:20:40

On 11/28/2012 1:25 PM, Oliver Kowalke wrote:
> 2012/11/28 Oliver Kowalke <oliver.kowalke_at_[hidden]>
> 
>>
>>> You are asking result_of to compute the result of calling a unary
>>> function that takes an X& with zero arguments. That's nonsensical. If
>>> you intend to call it with an X&, then ask it what the result is when
>>> called with an X&:
>>>
>>>    boost::result_of< Fn(X&) >::type
>>>
>>> This should correctly report void regardless of whether result_of uses
>>> decltype or the TR1 result_of protocol.
>>
>> OK, I got it.
>> It's strange that clang 3.2 (c++11), gcc, ontel and msvc compile the code
>> without errors on Loinux and Windows.

I've already explained this. Decltype-based result_of is only turned on
by default for compilers that have strong-enough decltype support. That
is currently only very recent clang versions. That does NOT include gcc,
intel or msvc.

> I thought I could the problem can be solved like the following code, but
> it  does not compile for gcc (c++0x):
> 
> #ifndef BOOST_NO_CXX11_RVALUE_REFERENCES
> template< typename Fn >
> void g( BOOST_RV_REF( Fn) ) {
>     typedef typename remove_reference< Fn >::type Y;
>         BOOST_STATIC_ASSERT((
>             is_same< void, typename result_of< Y( X &) >::type >::value));
> }
> #else
> ...
> #endif

If Fn is void(&)(X&), then remove_reference<Fn>::type is void(X&). That
would mean that Y(X&) is the type of a function that returns a function.
That's totally bogus. There is no such thing. Apparently, the
decltype-based result_of is smart enough to see this as bogus, but the
TR1 result_of, which only does dumb type shunting with metaprogramming
hacks, doesn't notice the problem.

> g() should be callable with function pointers, copyable and moveable-only
> functors (signatue void( X&)).
> 
> Again - the original code works for gcc, intel, msvc and clang (Linux) and
> fails for clang on MacOS X.

And hopefully now that I've explained it a second time, you no longer
find it remarkable.

-- 
Eric Niebler
BoostPro Computing
http://www.boostpro.com