tr1::reference_wrapper, sfinae and result_type definition

[Emmanuel Deloget]

Hello,

I'm trying to find a (sfinae powered) way to verify if a particular
type declares a subtype (either using typedef or by declaring a
subclass). To be more concrete, let's say that I'm trying to see how to
implement a tiny part of tr1::reference_wrapper, more exactly the
definition of its weak result type.

Let's imagine that class A and B inherit both std::unary_function and
std::binary_function. As such, they also inherit
unary_function::result_type and binary_function::result_type.
reference_wrapper<A> and reference_wrapper<B> shall now define their
weak result type.

Let's use these definitions:

------------8<----------------------------------------

struct A :
unary_function<int, int>,
binary_function<float, float, float>
{
typedef bool result_type;
};

struct B :
unary_function<int, int>,
binary_function<float, float, float>
{ };

------------>8----------------------------------------

A has in fact 3 different result_type (A::unary_function::result_type,
A::binary_function::result_type and A::result_type). The TR1 tells me
that I should define reference_wrapper<A>::result_type as
A::result_type (quite logical).

The problem comes if I have to consider B, as I find the wording of the
TR1 quite obscure in this case. As I understand it, I have two
possibilities:

1) either I don't define reference_wrapper<B>::result_type, because B
doesn't define result_type (although it inherit a result_type member
from both unary_function and binary_function)

2) or I define reference_wrapper<B>::result_type as either
B::unary_function::result_type or B::binary_function::result_type
(which doesn't make much sense, since these types are different. So
maybe I should only define reference_wrapper<B>::result_type if
B::unary_function::result_type and B::binary_function::result_type are
the same).

But that's not only the sole problem in this area. Another one comes if
I want to verify if a class defines a result_type subtype. Consequently
to the C++ name lookup scheme, if result_type is not defined in a class
C but is defined (and accessible) in a super class, then C::result_type
is known and can hardly be differentiated from a subtype of C.

It means that I cannot differentiate between the cases of class A and
class B above - I have no way to find that result_type is not defined
in class B but only in its superclass. The problem is that since B has
two superclass that define their own result_type, B::result_type
results in a compile time ambiguousity - the compiler shoke (and he's
quite right).

Until now, I used this code (inherited from boost, I think) to try to
see if a type was defined in a class:

------------8<----------------------------------------

template <class t> class has_result_type
{
template <class u> struct wrapper { }
typedef char one[1];
typedef char two[2];
template <class u> static one& test(wrapper<u::result_type>*);
template <class u> static two& test(...);
public:
static const bool value = sizeof(test<t>(0)) == sizeof(one);
};

------------>8----------------------------------------

Gnu's g++ compiler is quite happy with this code -
has_result_type<A>::value is true, and has_result_type<B>::value is
false. MS VC++.NET 2005 compiler is a lot more problematic: both
has_result_type<A>::value and has_result_type<B>::value are true -
which is nearly logical, since B really has (at least) a definition of
result_type. Finally, the Comeau C++ compiler shoke - because
u::result_type is ambiguous when u = B (this is probably the most
logical behavior, as I understand how the standard defines name
lookup).

As a consequence, either this code is nonportable or it is just non
standard compliant.

Which means, in turn, that I still have no way to differentiate between
class A and class B using a generic way.

So far, I have two questions:
1) what is the correct, official intrepretation of the weak result type
definition in the TR1? (I personnally feel that the wording of this
definition should be changed to be clearer)
2) regardless of this interpretation, how can I differientiate between
the two cases I presented (class A and class B)?

This issue has bugged be for two weeks now, and I'm not able to devise
a solution to this problem.

Thanks for your time, and happy answering!

-- Emmanuel Deloget, Artware

 

[Pete Becker]

Let's use these definitions:

------------8<----------------------------------------

struct A :
unary_function<int, int>,
binary_function<float, float, float>
{
typedef bool result_type;
};

struct B :
unary_function<int, int>,
binary_function<float, float, float>
{ };

------------>8----------------------------------------

A has in fact 3 different result_type (A::unary_function::result_type,
A::binary_function::result_type and A::result_type). The TR1 tells me
that I should define reference_wrapper<A>::result_type as
A::result_type (quite logical).

Not really. A and B should each be derived from either unary_function or
from binary_function, but not both. Keep in mind that the requirements
for defining result_type are for backward compatibility with today's
callable types and with today's compilers. Callable types in the current
standard have one function call operator, so when you write that type
you know exactly how many arguments that function call operator takes,
so you can choose unary_function or binary_function as appropriate. (See
chapter 6 of my book, "The Standard C++ Library Extensions", for more
details).

struct A : unary_function<int, int>
{
int operator()(int);
};

or

struct A : binary_function<float, float, float>
{
float operator()(float, float);
};

If you're trying to do something more sophisticated with multiple
function call operators and different numbers of arguments you've got
problems: the support isn't quite there today. Eventually you'll be able
to do something like this:

struct A
{
int operator()(int);
float operator()(float);
};

A a;
reference_wrapper<A> awrap = a;
a(3);
a(4.5);

But at present, reference_wrapper isn't able to figure out the return
type of that function call operation. It relies on result_of, which, for
now, relies on finding result_type in the type that you're using. And,
as you've mentioned, it's hard to have more than one result_type in a class.

--

-- Pete
Roundhouse Consulting, Ltd. (www.versatilecoding.com)
Author of "The Standard C++ Library Extensions: a Tutorial and
Reference." (www.petebecker.com/tr1book)

 

[Emmanuel Deloget]

Pete Becker a écrit :

Not really. A and B should each be derived from either unary_function or
from binary_function, but not both. Keep in mind that the requirements
for defining result_type are for backward compatibility with today's
callable types and with today's compilers. Callable types in the current
standard have one function call operator, so when you write that type
you know exactly how many arguments that function call operator takes,
so you can choose unary_function or binary_function as appropriate. (See
chapter 6 of my book, "The Standard C++ Library Extensions", for more
details).

[snip-a-lot]

If you're trying to do something more sophisticated with multiple
function call operators and different numbers of arguments you've got
problems: the support isn't quite there today. Eventually you'll be able
to do something like this:

struct A
{
int operator()(int);
float operator()(float);
};

A a;
reference_wrapper<A> awrap = a;
a(3);
a(4.5);

But at present, reference_wrapper isn't able to figure out the return
type of that function call operation. It relies on result_of, which, for
now, relies on finding result_type in the type that you're using. And,
as you've mentioned, it's hard to have more than one result_type in a class.

--

-- Pete
Roundhouse Consulting, Ltd. (www.versatilecoding.com)
Author of "The Standard C++ Library Extensions: a Tutorial and
Reference." (www.petebecker.com/tr1book)

First, thanks for you answer.

Second, it appears that if the class could inherit either
unary_function or binary_function (but not both) to be
reference_wrapped, then the TR1 (or at least the TR1 draft that can be
downloaded from the WG21 website) is not adequately worded (no offense
intended), as its words implies that both should be supported: 2.1.2
§3 and 2.1.2 §4 are not mutually exclusive, resulting in your
servitor trying to implement the union of both clauses. It would also
simplify things a lot (at least, I'd be able to code this ^_^) ( the
new C++0x working draft also use a similar wording BTW, in which mutual
exclusivity is not implied).

I guess I can now move on some another annoying problem

Regards,

-- Emmanuel Deloget, Artware