2 / Copyright (c) 2008 Eric Niebler
4 / Distributed under the Boost Software License, Version 1.0. (See accompanying
5 / file LICENSE_1_0.txt or copy at http://www.boost.org/LICENSE_1_0.txt)
8 [section:implementation Appendix D: Implementation Notes]
10 [section:sfinae Quick-n-Dirty Type Categorization]
12 Much has already been written about dispatching on type traits using
13 SFINAE (Substitution Failure Is Not An Error) techniques in C++. There
14 is a Boost library, Boost.Enable_if, to make the technique idiomatic.
15 Proto dispatches on type traits extensively, but it doesn't use
16 `enable_if<>` very often. Rather, it dispatches based on the presence
17 or absence of nested types, often typedefs for void.
19 Consider the implementation of `is_expr<>`. It could have been written
20 as something like this:
24 : is_base_and_derived<proto::some_expr_base, T>
27 Rather, it is implemented as this:
29 template<typename T, typename Void = void>
35 struct is_expr<T, typename T::proto_is_expr_>
39 This relies on the fact that the specialization will be preferred
40 if `T` has a nested `proto_is_expr_` that is a typedef for `void`.
41 All Proto expression types have such a nested typedef.
43 Why does Proto do it this way? The reason is because, after running
44 extensive benchmarks while trying to improve compile times, I have
45 found that this approach compiles faster. It requires exactly one
46 template instantiation. The other approach requires at least 2:
47 `is_expr<>` and `is_base_and_derived<>`, plus whatever templates
48 `is_base_and_derived<>` may instantiate.
52 [section:function_arity Detecting the Arity of Function Objects]
54 In several places, Proto needs to know whether or not a function
55 object `Fun` can be called with certain parameters and take a
56 fallback action if not. This happens in _callable_context_ and
57 in the _call_ transform. How does Proto know? It involves some
58 tricky metaprogramming. Here's how.
60 Another way of framing the question is by trying to implement
61 the following `can_be_called<>` Boolean metafunction, which
62 checks to see if a function object `Fun` can be called with
63 parameters of type `A` and `B`:
65 template<typename Fun, typename A, typename B>
68 First, we define the following `dont_care` struct, which has an
69 implicit conversion from anything. And not just any implicit
70 conversion; it has a ellipsis conversion, which is the worst possible
71 conversion for the purposes of overload resolution:
78 We also need some private type known only to us with an overloaded
79 comma operator (!), and some functions that detect the presence of
80 this type and return types with different sizes, as follows:
84 private_type const &operator,(int) const;
87 typedef char yes_type; // sizeof(yes_type) == 1
88 typedef char (&no_type)[2]; // sizeof(no_type) == 2
91 no_type is_private_type(T const &);
93 yes_type is_private_type(private_type const &);
95 Next, we implement a binary function object wrapper with a very
96 strange conversion operator, whose meaning will become clear later.
98 template<typename Fun>
102 typedef private_type const &(*pointer_to_function)(dont_care, dont_care);
103 operator pointer_to_function() const;
106 With all of these bits and pieces, we can implement `can_be_called<>` as
109 template<typename Fun, typename A, typename B>
112 static funwrap2<Fun> &fun;
116 static bool const value = (
117 sizeof(no_type) == sizeof(is_private_type( (fun(a,b), 0) ))
120 typedef mpl::bool_<value> type;
123 The idea is to make it so that `fun(a,b)` will always compile by adding
124 our own binary function overload, but doing it in such a way that we can
125 detect whether our overload was selected or not. And we rig it so that
126 our overload is selected if there is really no better option. What follows
127 is a description of how `can_be_called<>` works.
129 We wrap `Fun` in a type that has an implicit conversion to a pointer to
130 a binary function. An object `fun` of class type can be invoked as
131 `fun(a, b)` if it has such a conversion operator, but since it involves
132 a user-defined conversion operator, it is less preferred than an
133 overloaded `operator()`, which requires no such conversion.
135 The function pointer can accept any two arguments by virtue
136 of the `dont_care` type. The conversion sequence for each argument is
137 guaranteed to be the worst possible conversion sequence: an implicit
138 conversion through an ellipsis, and a user-defined conversion to
139 `dont_care`. In total, it means that `funwrap2<Fun>()(a, b)` will
140 always compile, but it will select our overload only if there really is
143 If there is a better option --- for example if `Fun` has an overloaded
144 function call operator such as `void operator()(A a, B b)` --- then
145 `fun(a, b)` will resolve to that one instead. The question now is how
146 to detect which function got picked by overload resolution.
148 Notice how `fun(a, b)` appears in `can_be_called<>`: `(fun(a, b), 0)`.
149 Why do we use the comma operator there? The reason is because we are
150 using this expression as the argument to a function. If the return type
151 of `fun(a, b)` is `void`, it cannot legally be used as an argument to
152 a function. The comma operator sidesteps the issue.
154 This should also make plain the purpose of the overloaded comma operator
155 in `private_type`. The return type of the pointer to function is
156 `private_type`. If overload resolution selects our overload, then the
157 type of `(fun(a, b), 0)` is `private_type`. Otherwise, it is `int`.
158 That fact is used to dispatch to either overload of `is_private_type()`,
159 which encodes its answer in the size of its return type.
161 That's how it works with binary functions. Now repeat the above process
162 for functions up to some predefined function arity, and you're done.
167 [section:ppmp_vs_tmp Avoiding Template Instiations With The Preprocessor]