Enjoy template metaprogramming
This is not an official Boost library. It might be proposed as a successor for the current Boost.MPL in the future, but there is no guarantee. Also, I am currently working on Boost.Hana, a library which tries to merge MPL11 and Boost.Fusion into a single library. If that works out, it would probably replace the MPL11.
The library is unstable at the moment; do not use for production.
This was initially supposed to be a simple C++11 reimplementation of the Boost.MPL. However, for reasons documented in the rationales, the library was redesigned and the name does not fit so well anymore.
- Installation
- Introduction
- Metafunctions
- Datatypes and data constructors
- Methods
- Typeclasses
- Rewrite rules
- Acknowledgements
- Rationales
- Todo list
The MPL11 is a header only library. To use it in your own project, just add the include directory to your compiler's header search path and you are done.
The library has no dependencies - not even the standard library. However, it requires a C++14-able compiler. The test suite passes under the following compilers:
- clang version 3.4
- clang version 3.5.0
- GCC 4.9.0 20140302 (experimental)
- Apple LLVM version 5.1 (clang-503.0.38)
To compile the unit tests, you will also need CMake. Once you have it, you can go to the root of the project and do:
$ mkdir build
$ cd build
$ cmake ..
$ make tests # Compiles the unit tests.
A minified version of the library is also provided. To use it, simply include
the boost/mpl11.min.hpp
header, which contains the whole library. Note that
the minified header must not be used in conjunction with other headers from
the library.
The MPL11 is a C++11 library providing composable, high-level primitives for solving complex template metaprogramming problems. The library is built around a few core concepts; the aim of this tutorial is to present them, while the handful of tools provided by the library are left to the reference documentation.
This tutorial assumes a good understanding of template metaprogramming and basic functional programming concepts. Also, a good understanding of the Boost.MPL library will be helpful. However, the MPL11 diverges from the Boost.MPL in many ways, and one should be careful not to transfer knowledge between both libraries without checking the documentation.
Informally, a metafunction is a template representing a compile-time function taking types as arguments and returning a type as a result. Readers coming from the MPL should be careful here, since the formal definition differs from that of the MPL.
Formally, let f
be a C++ template with an arbitrary number of type
template parameters, and type template parameters only. Then, f
is a
metafunction if and only if there exists types x1, ..., xn
such
that f<x1, ..., xn>::type
is a valid type name. In this context,
x1, ..., xn
are the arguments off
.- Forming a specialization
f<x1, ..., xn>
is called suspendingf
withx1, ..., xn
. - A specialization
f<x1, ..., xn>
is called a thunk or a suspension. - The nested
::type
of a thunk is called the result of the thunk. If the thunk is of the formf<x1, ..., xn>
, we can also say it is the result off
withx1, ..., xn
. - Fetching the result of a thunk is called evaluating the thunk. If the
thunk is of the form
f<x1, ..., xn>
, we can also say invokingf
withx1, ..., xn
or applyingf
tox1, ..., xn
. - The arity of a metafunction is the number of arguments it can be invoked with. A metafunction with arity n is said to be a n-ary metafunction.
- A metafunction that can be invoked with any number of arguments is said to be variadic. By definition, a variadic metafunction is n-ary for any non-negative integer n.
It is important to note the difference between this definition and the one given by the Boost.MPL. With this definition, a metafunction can never be a normal C++ type; it must always be a template. Hence, Boost.MPL nullary metafunctions implemented as non-template classes are not considered as metafunctions. Here are some examples:
// A unary metafunction.
template <typename x>
struct unary { struct type; };
// A binary metafunction.
template <typename x, typename y>
struct binary { struct type; };
// A variadic metafunction.
template <typename ...>
struct variadic { struct type; };
// A nullary metafunction. It can only be invoked with
// 0 arguments, and it is therefore 0-ary (nullary).
template <typename ...> struct nullary;
template <> struct nullary<> { struct type; };
// Not a metafunction with the MPL11; it is not a template!
struct MPL_nullary { struct type; };
// Not a metafunction; it never has a result (a nested ::type)!
template <typename ...>
struct no_result { };
Informally, a boxed type is a type that has yet to be evaluated. Hence, before knowing the actual "value" of a boxed type, one must evaluate it, a process which is called unboxing.
Formally, for an arbitrary C++ type T
, a boxed T
is an arbitrary C++
type B
such that B::type
is T
. In this context, B
is called a box
(of T
) and fetching the nested ::type
inside of B
is called unboxing
T
.
struct T;
struct B { using type = T; }; // a boxed T (equivalently, a box of T)
B::type; // unboxing T
Conversely, wrapping an arbitrary type T
in a type B
such that B::type
is T
is called boxing T
(into B
or with B
). It is important to
note that B
may depend on T
, without which boxing would be of limited
interest.
struct T;
template <typename t>
struct B { using type = t; };
B<T>; // boxing T into B
Note that types may be boxed an arbitrary number of times. This is probably not useful, but the definition is general enough to allow it.
B<B<T>>; // this is a "box of B<T>", aka a "box of (box of T)"
There exists a special boxed type named undefined
(sometimes called
bottom) which has the characteristic of causing a compile-time error
when it is unboxed, even in SFINAE-able contexts. undefined
can be
seen as an invalid value, or the result of a computation that failed.
Here are some examples to illustrate the previous definition:
// This template takes an arbitrary type T and boxes it.
template <typename T>
struct box {
using type = T;
};
// These are not boxed.
class x;
struct y { char foo; };
char;
box<char>::type;
// These are boxed types.
box<char>; // a boxed `char`
box<box<char>>; // a boxed `box<char>`
box<box<char>>::type; // a boxed `char`
struct x { using type = char; }; // a boxed `char`
struct y { struct type; }; // a boxed `y::type`
struct z { using type = z; }; // self-referential? why not!
It is important to note that there are many different representations for a
boxed T
. This makes equivalence between boxed types a bit tricky. Consider
the following:
struct T;
struct B1 { using type = T; }; // a boxed T
struct B2 { using type = T; }; // a boxed T too
Certainly, B1
and B2
are equivalent w.r.t. to the type they box since they
both box the same type T
. However, B1
and B2
are not equivalent w.r.t.
the C++ type system because they are different types. Now, this is important
because it tells us that we can't use pattern matching to define a
metafunction taking a boxed type as an argument. Indeed, since the
representation of a boxed type is not unique, we can't know what form
will have our argument in advance, and therefore we can't pattern match.
Consider the following:
// B should be a boxed type.
template <typename B>
struct f;
// This should handle boxed chars, but we don't know
// what a boxed char may look like!
template <>
struct f<????> {
// ...
};
Now, we might be tempted to do the following:
// box is the template that we defined earlier. It takes an
// arbitrary type and boxes it.
template <>
struct f<box<char>> {
// ...
};
But then...
template <typename T>
struct bad {
using type = T;
};
// works, as expected
f<box<char>>::type;
// does not work, even though bad<char> is clearly a boxed char
f<bad<char>>::type;
Instead, we would have to do something more convoluted like:
template <typename T>
struct f_impl;
template <>
struct f_impl<char> {
using type = ...;
};
template <typename B>
struct f
: f_impl<typename B::type>
{ };
f<box<char>>::type; // works
f<bad<char>>::type; // works too
It is interesting to note that boxed types and thunks share a lot. In fact, a thunk is nothing but a boxed type that was formed by suspending a metafunction. Hence, whenever a boxed type is expected, a thunk may be used instead.
This section introduces the notion of laziness in the context of template metaprograms and explains how it relates to the previous notions. This is by no means a rigorous treatment of the broader subject of evaluation strategies, and knowledge of those concepts is expected.
Informally, laziness is a property of a metafunction meaning that the metafunction performs the least amount of work needed to give its result. It requires the metafunction to only evaluate the expressions that are actually needed in its body, and to evaluate them no more than once.
The second point is called memoization and it is handled behind the scenes by the compiler. While the C++ standard does not require compilers to memoize template instantiations, this is always the case in practice. Consider:
template <typename x>
struct f {
using type = x;
};
f<int>::type; // instantiate f<int>
f<int>::type; // f<int> is already instantiated, nothing is done.
The first point, however, must be handled manually when writing template metaprograms.
TODO Finish this section. Specifically, explain the following:
- What is a lazy metafunction (mf classes follow from that)
- The inverse of a lazy metafunction is a strict metafunction (broadly)
- Lazy metafunctions must take boxed types, otherwise they would always evaluate their arguments whether they are needed or not. This is equivalent to call-by-name.
- Strict metafunctions can take unboxed arguments, because they always evaluate their arguments anyways. However, strict metafunctions can still take boxed arguments and unbox them unconditionally; this is just a matter of convention.
- Metafunctions take boxed arguments by default in the MPL11.
- Metafunctions are lazy by default (i.e. whenever possible) in the MPL11.
- Strict metafunctions usually still take boxed arguments for consistency.
- Some metafunctions don't follow the convention, and in this case this behavior is documented.
- Metafunctions that don't follow the convention do it because it's necessary or because it's much easier to do such or such when breaking the convention.
- Why are lazy metafunctions useful? This could use the
drop
metafunction of the old introduction. Laziness often leads to increased expressiveness; for example it becomes easy to define new control structures and infinite data structures.- Consider keeping the optional section on non-strictness.
At this point, it is probably helpful to clarify that returning from a lazy
metafunction is no different than returning from a strict metafunction. For
example, consider the following lazy metafunction implementing an if
statement:
template <typename Condition, typename Then, typename Else>
struct if_ {
using Branch = std::conditional<Condition::type::value, Then, Else>::type;
using type = typename Branch::type;
};
Since if_
is lazy, its arguments are all boxed types. Here, we unbox
Branch
and return that instead of returning Branch
itself. This way,
if_<Condition, Then, Else>
is a thunk and we can pass it to other lazy
metafunctions as-is:
// A lazy metafunction.
template <typename x>
struct f;
using result = f< if_<Condition, Then, Else> >::type;
If we had defined if_
as follows
template <typename Condition, typename Then, typename Else>
struct if_ {
using Branch = std::conditional<Condition::type::value, Then, Else>::type;
using type = Branch; // Note that we don't unbox Branch here
};
then if_
would return a thunk and we would need to do the following instead:
using result = f< if_<Condition, Then, Else>::type >::type;
^^^^^^
Informally, a lifted metafunction is a representation of a metafunction that
allows it to be manipulated as a first class citizen in template metaprograms.
Formally, an arbitrary C++ type f
is a lifted metafunction if and only
if f::apply
is a metafunction. In general, lifted metafunctions inherit the
terminology of metafunctions, and the characteristics of a lifted metafunction
follow from that of its nested apply
metafunction. For example, the arity of
a lifted metafunction f
is that of f::apply
.
The definition of a lifted metafunction is almost the same as the definition of a metafunction class in the Boost.MPL. The only difference is the difference between metafunctions in both libraries.
Here are some examples of lifted metafunctions:
// A unary lifted metafunction.
struct unary {
template <typename x>
struct apply { struct type; };
};
// A binary lifted metafunction.
struct binary {
template <typename x, typename y>
struct apply { struct type; };
};
// A variadic lifted metafunction.
struct variadic {
template <typename ...>
struct variadic { struct type; };
};
// A nullary lifted metafunction.
struct nullary {
template <typename ...> struct apply;
};
template <> struct nullary::apply<> { struct type; };
// Not a lifted metafunction with the MPL11; its nested apply
// is not a metafunction!
struct MPL_nullary {
struct apply { struct type; };
};
// Not a lifted metafunction; its nested apply never has a result!
struct no_result {
template <typename ...>
struct apply { };
};
Any metafunction can be lifted, and the MPL11 defines a template to do just that.
template <template <typename ...> class f>
struct lift {
template <typename ...xs>
struct apply
: f<xs...>
{ };
};
lift
is essentially the same asquote
in the Boost.MPL. The namelift
was preferred because of its relation to a lift in category theory. For completeness,lift
is actually an equivalence of categories between the category where objects are C++ types and morphisms are templates to the category where objects are C++ types and morphisms are structs with a nestedapply
template.
At compile-time, a type becomes a value. A legitimate question would then be: what is the type of that value? In the MPL11, datatypes play that role. Closely related are data constructors, which are a way of creating values of a given datatype. For example, let's create a simple compile-time list:
// A "tag" representing the datatype.
struct List;
// The list constructor. It represents a value of type List that
// contains the provided elements.
template <typename ...Elements>
struct list { using mpl_datatype = List; };
// The cons constructor. It represents a value of type List
// with the provided head and tail.
template <typename Head, typename Tail>
struct cons { using mpl_datatype = List; };
// The nil constructor. It represents an empty List.
struct nil { using mpl_datatype = List; };
Data constructors must provide a nested ::mpl_datatype
alias representing
their datatype. One can then use the datatype
metafunction to retrieve it:
datatype<nil>::type; // == List
It is also possible to specialize datatype
instead of providing a
nested mpl_datatype
alias. So this definition of cons
is equally
valid (and the other constructors could be defined analogously):
template <typename Head, typename Tail>
struct cons; // no nested mpl_datatype
namespace boost { namespace mpl11 {
template <typename Head, typename Tail>
struct datatype<cons<Head, Tail>> {
using type = List;
};
}}
Being able to do this is paramount when adapting code over which we don't have
the control, but for simplicity we'll stick with the nested mpl_datatype
whenever possible. When unspecialized, datatype<T>
simply returns
T::mpl_datatype
if that expression is a valid type name, and Foreign
otherwise. Hence, Foreign
is the default datatype of everything.
Note that
datatype
is a strict metafunction and that it does not obey the convention of taking boxed arguments. Breaking the convention is necessary to allow user-defined specializations.
While it is not mandatory, it is often a good idea to box data constructors since it makes them usable as-is in lazy metafunctions. Let's rewrite the previous data constructors that way:
template <typename ...Elements>
struct list_impl {
using mpl_datatype = List;
};
template <typename Head, typename Tail>
struct cons_impl {
using mpl_datatype = List;
};
struct nil_impl { using mpl_datatype = List; };
template <typename ...Elements>
struct list {
using type = list_impl<Elements...>;
};
template <typename Head, typename Tail>
struct cons {
using type = cons_impl<Head, Tail>;
};
struct nil {
using type = nil_impl;
};
The downside is that we end up with twice the amount of code, half of which is complete boilerplate. Also, this approach causes twice as many templates to be instantiated each time we unbox a data constructor, which can hurt compile-time performance. Fortunately, we can use self-referential boxing to make this better.
template <typename ...Elements>
struct list {
using type = list;
using mpl_datatype = List;
};
template <typename Head, typename Tail>
struct cons {
using type = cons;
using mpl_datatype = List;
};
struct nil {
using type = nil;
using mpl_datatype = List;
};
That way, only one additional line of code is required per data constructor
and we only instantiate one template when we unbox it. Indeed,
cons<...>::type
is just cons<...>
, which is already instantiated.
You might wonder why I have even bothered with the inferior solution using
cons_impl
and friends in the first place, since the self-referential
solution is so obvious. This was to highlight the fact that boxed data
constructors do not need to be self-referential; it is merely a convenient
implementation trick. This is a subtle difference between the MPL11 and the
mpllibs library collection, which I wanted to point out.
There is something rather important that we have left undefined when we
created the list
and cons
data constructors: what do their arguments
look like?
template <typename ...Elements>
struct list;
template <typename Head, typename Tail>
struct cons;
There are two possible answers here, and both are valid. In the end, this is mainly a design choice. The first option is to require the arguments to be normal (non-boxed) types. We could then use the constructors like so:
using x = list<int, char, float>;
using y = cons<int, list<char, float>>;
using z = cons<int, cons<char, cons<float, nil>>>;
The second option is to require boxed arguments. We could then use the constructors like so:
using x = list<box<int>, box<char>, box<float>>;
using y = cons<box<int>, list<box<char>, box<float>>>;
using z = cons<box<int>, cons<box<char>, cons<box<float>, nil>>>;
Note that we do not need to box the second arguments to
cons
ourselves, because we have already madelist
,cons
andnil
boxed.
This is clearly less natural than the first solution. Still, for reasons
that will soon become clearer, the MPL11 List
constructors use the second
solution, and so will we for the rest of this tutorial. To reduce the
syntactic noise, we will define aliases that will make our life easier:
template <typename ...Elements>
using list_ = list<box<Elements>...>;
template <typename Head, typename Tail>
using cons_ = cons<box<Head>, box<Tail>>;
Note that we would not need to box the second argument of
cons_
because we expect it to be aList
, and allList
constructors are already boxed. Sobox<Tail>
is really redundant, but we still do it here for the sake of clarity.
We will say that a data constructor taking unboxed arguments is strict, and that a data constructor taking boxed arguments is lazy. An interesting observation is that some (but not all) constructors are also metafunctions. Specifically, boxed constructors taking type template parameters are metafunctions. Therefore, strict boxed constructors correspond to strict metafunctions, and lazy boxed constructors correspond to lazy metafunctions!
The MPL11 provides a way to convert values from one datatype to the other. The
usefulness of this is clearest when implementing numeric datatypes, but we'll
stick with List
because we already have it. Let's suppose we want to convert
values from a List
to an std::tuple
and vice-versa. First, we will need to
define a datatype of which std::tuple
can be a data constructor.
struct StdTuple;
namespace boost { namespace mpl11 {
template <typename ...T>
struct datatype<std::tuple<T...>> {
using type = StdTuple;
};
}}
We can now consider std::tuple
as a data constructor for the StdTuple
datatype. Note that unlike the arguments to list
, the arguments to
std::tuple
must be unboxed; this will be important for what follows.
The next step is to implement the conversion itself. This is done by
specializing the cast
lifted metafunction for the involved datatypes.
namespace boost { namespace mpl11 {
template <>
struct cast</*from*/ List, /*to*/ StdTuple> {
template <typename>
struct apply;
template <typename ...Elements>
struct apply<list<Elements...>> {
using type = std::tuple<typename Elements::type...>;
};
template <>
struct apply<nil> {
using type = std::tuple<>;
};
template <typename Head, typename Tail>
struct apply<cons<Head, Tail>> {
using type = /* omitted for simplicity */;
};
};
template <>
struct cast</*from*/ StdTuple, /*to*/ List> {
template <typename>
struct apply;
template <typename ...Elements>
struct apply<std::tuple<Elements...>> {
using type = list_<Elements...>;
};
};
}}
There are a few things to note here. First, it is very important to provide a
conversion for all the data constructors of a datatype. If, for instance, we
had forgotten to provide apply<nil>
, we could only convert from the list
and cons
constructors. Second, we had to unbox the elements of the list
when passing them to std::tuple
because std::tuple
expects unboxed types.
Similarly, we had to box the elements of the std::tuple
when passing them to
list
. Third, the cast
lifted metafunction is strict and it does not follow
the convention of taking boxed arguments, which makes it possible to pattern
match data constructors. Here is how we could now convert between the two
datatypes:
using my_list = list_<int, char, float>;
using my_tuple = std::tuple<int, char, float>;
cast<List, StdTuple>::apply<my_list>::type; // == my_tuple
cast<StdTuple, List>::apply<my_tuple>::type; // == my_list
Also note that casting from a datatype T
to itself is a noop, so you don't
have to worry about that trivial case. The library also defines a handy
cast_to
lifted metafunction that reduces the syntactic noise of cast
:
cast_to<StdTuple>::apply<my_list>::type; // == my_tuple
cast_to<List>::apply<my_tuple>::type; // == my_list
cast_to
simply deduces the datatype to cast from by using that of its
argument and then forwards to cast
. cast_to
is strict and does not
take boxed arguments.
TODO: Give more details on the
Foreign
datatype.
So far, we have created the List
datatype and a couple of constructors to
create "values" of that type, but we still don't have a way to manipulate
those values in a useful way. Let's define a head
metafunction that will
return the first element of a List
. We will stick to the convention of
taking boxed arguments.
template <typename List>
struct head_impl;
template <typename Head, typename ...Tail>
struct head_impl<list<Head, Tail...>>
: Head
{ };
template <typename Head, typename Tail>
struct head_impl<cons<Head, Tail>>
: Head
{ };
// head_impl<nil> is not defined since that's an error.
template <typename List>
struct head
: head_impl<typename List::type>
{ };
First, we need to add a level of indirection (head_impl
) because head
receives boxed arguments and we need to pattern match the constructors.
Second, note that head
is a strict metafunction because its argument
is always evaluated. Third, we inherit from Head
in head_impl
because
the List
constructors are lazy and hence Head
is boxed.
It is now possible to see why it was useful to make the List
constructors
lazy. Consider the following situation:
using refs = list<
std::add_lvalue_reference<int>,
std::add_lvalue_reference<void>
>;
head<refs>::type; // == int&
Since we can't form a reference to void
, we will trigger a compile-time
error if we evaluate the std::add_lvalue_reference<void>
thunk. However,
since we only ever use the value of the first element in the list, it would
be nice if we could only evaluate that element, right? Making list
a lazy
constructor makes that possible. If, instead, we had decided to make list
strict, we would need to write:
using refs = list<
std::add_lvalue_reference<int>::type,
std::add_lvalue_reference<void>::type
>;
which does not compile. The same reasoning is valid if the contents of the list were the results of complicated computations. By making the constructor lazy, we would only need to evaluate those computation whose result is actually used. In the end, the reasons for writing lazy data constructors are similar to the reasons for writing lazy metafunctions.
The head
metafunction we have so far is useful, but consider the following
datatype and lazy boxed constructor:
struct Vector;
template <typename ...Elements>
struct vector {
struct type {
using mpl_datatype = Vector;
};
};
template <typename Head, typename ...Tail>
struct vector<Head, Tail...> {
struct type {
using head = Head;
using mpl_datatype = Vector;
};
};
Since Vector
is really some kind of List
, it is only reasonable to expect
that we can invoke head
on a Vector
just like we do on a List
. But how
would we go about implementing head
for Vector
? Assuming we can't modify
the implementation of Vector
, the only way we have right now is to use
partial specialization of head_impl
on Vector
's constructor.
Unfortunately, that constructor is vector<...>::type
, not vector<...>
.
Since we can't partially specialize on a dependent type, we're out of luck.
To bypass this limitation, we will refine head
so it uses the datatype of
its argument.
template <typename Datatype>
struct head_impl;
template <typename List>
struct head
: head_impl<typename datatype<typename List::type>::type>::
template apply<typename List::type>
{ };
We now consider head_impl
as a lifted metafunction parameterized over the
datatype of its argument. Implementing head
for List
and Vector
is now
a breeze.
template <>
struct head_impl<List> {
template <typename> struct apply;
template <typename Head, typename ...Tail>
struct apply<list<Head, Tail...>>
: Head
{ };
template <typename Head, typename Tail>
struct apply<cons<Head, Tail>>
: Head
{ };
};
template <>
struct head_impl<Vector> {
template <typename v>
struct apply
: v::head
{ };
};
It is sometimes useful to exert a finer grained control over template specializations than what we currently have. A common idiom is for the primary template to provide an additional dummy parameter which can be SFINAE'd upon in partial specializations:
template <typename T, typename Enable = void>
struct trait;
template <typename T>
struct trait<T, std::enable_if_t<std::is_trivial<T>::value>> {
// ...
};
This enables the specialization to depend on an arbitrary compile-time boolean
expression (in fact on the syntactic validity of an arbitrary expression).
Note that all partial specializations using the enabler must have mutually
exclusive conditions or the specialization will be ambiguous; this can be
tricky at times. A variant of this trick is to use a default type of true_
instead of void
for the dummy template parameter. That makes it possible to
leave the std::enable_if_t
part for most use cases:
template <typename T, typename Enable = true_>
struct trait;
template <typename T>
struct trait<T, bool_<std::is_trivial<T>::value>> {
// ...
};
Note that we used
bool_<...>
instead ofstd::is_trivial<T>::type
because the latter isstd::true_type
, which is not necessarily the same astrue_
.Also, we don't use a straight
bool
for the enabler because partial specializations cannot have dependent non-type template arguments.
Since this functionality can really be useful, it might be a good idea to
support it in the head_impl
lifted metafunction. Fortunately, that only
requires changing the head_impl
forward declaration to:
template <typename Datatype, typename = true_>
struct head_impl;
TODO: Provide a use case where this is useful.
In this section, we went through the process of designing a simple yet powerful way of dispatching metafunctions. The subset of metafunctions using this dispatching technique are called methods in the MPL11.
TODO: Tackle binary operators and mixed-datatype dispatch.
Typeclasses come from the observation that some methods are naturally related
to each other. For example, take the head
, tail
and is_empty
methods.
When implementing any of these three methods, it is probable that the other
two should also be implemented. Hence, it would be logical to group them in
some way; that is the job of typeclasses. However, typeclasses are much more
than mere method bundles. They provide a clean way of specifying and extending
the interface of a datatype through a process called typeclass instantiation.
Let's make a typeclass out of the head
, tail
and is_empty
methods.
Datatypes supporting all three methods look somewhat like List
s; hence
we will call the typeclass Iterable
. We start by grouping the *_impl
lifted metafunctions of the methods together under the Iterable
banner.
In the following code snippets, the tail
and is_empty
methods will be
omitted when illustrating head
suffices.
struct Iterable {
template <typename Datatype, typename = true_>
struct head_impl;
// ...
};
template <typename Iter>
struct head
: Iterable::head_impl<typename datatype<typename Iter::type>::type>::
template apply<typename Iter::type>
{ };
// ...
To implement the methods for List
, we now have to write:
template <>
struct Iterable::head_impl<List> {
template <typename the_list>
struct apply {
// ...
};
};
Soon enough, we notice that we can regroup the datatype parametrization on
Iterable
instead of leaving it on each nested lifted metafunction.
template <typename Datatype, typename = true_>
struct Iterable {
struct head_impl;
// ...
};
template <typename Iter>
struct head
: Iterable<typename datatype<typename Iter::type>::type>::
head_impl::template apply<typename Iter::type>
{ };
// ...
The next logical step is to prune the superfluous indirection through the
*_impl
lifted metafunctions and to simply make them metafunctions.
template <typename Datatype, typename = true_>
struct Iterable {
template <typename Iter>
struct head_impl;
// ...
};
template <typename Iter>
struct head
: Iterable<typename datatype<typename Iter::type>::type>::
template head_impl<typename Iter::type>
{ };
// ...
Since it might be useful to query whether a datatype supports the operations
of Iterable
, we would like to have a boolean metafunction that does just
that. Fortunately, we can use the Iterable
for this task with a small
modification.
template <typename Datatype, typename = true_>
struct Iterable : false_ {
// ...
};
By default, Iterable
is therefore also a boolean metafunction returning
false
, meaning that arbitrary datatypes don't implement the head
, tail
and is_empty
metafunctions. In its current form, the Iterable
template is
called a typeclass, and metafunctions like head
following this
dispatching pattern through a typeclass are called typeclass methods.
In order to implement head
and friends, one would now write
template <>
struct Iterable<List> : true_ {
template <typename the_list>
struct head_impl {
// ...
};
// ...
};
static_assert(Iterable<List>{}, "List is an Iterable!");
The three methods Iterable
contains so far are very basic; for any given
datatype, it is not possible to provide a suitable default implementation.
However, there are other metafunctions that can be implemented in terms of
these three basic blocks. For example, consider the last
metafunction that
returns the last element of an Iterable
. A possible implementation would be:
template <typename iter>
struct last
: if_<is_empty<tail<iter>>,
head<iter>,
last<tail<iter>>
>
{ };
While we could provide this metafunction as-is, some datatypes might be able to provide a more efficient implementation. Therefore, we would like to make it a method, but one that can be defaulted to the above.
Note that method dispatching incurs some compile-time overhead; hence there is a tradeoff between using (typeclass) methods and regular metafunctions. The rule of thumb is that if a method is likely to never be specialized (i.e. the default implementation is almost always the best), then it should probably be a regular metafunction.
It turns out that providing a default implementation can be done easily using
typeclasses and a little convention. First, we make last
a normal typeclass
method.
template <typename Iter>
struct last
: Iterable<typename datatype<typename Iter::type>::type>::
template last_impl<typename Iter::type>
{ };
Then, we require specializations of Iterable
to inherit some special type
as follows:
template <>
struct Iterable<List> : instantiate<Iterable>::with<List> {
// ...
};
Here, instantiate<...>::with<...>
is true_
by default. Hence, it only
takes care of making Iterable
a true-valued boolean metafunction, which we
did by ourselves previously. However, instantiate
may be specialized by
typeclass designers in such a way that the member template with
also
contains default methods. In our case, we would provide a last_impl
metafunction corresponding to the default implementation of last
shown
above. This way, if a datatype does not implement the last
method, our
default implementation will be used.
namespace boost { namespace mpl11 {
template <>
struct instantiate<Iterable> {
template <typename Datatype>
struct with : true_ {
template <typename Iter>
struct last_impl {
// default implementation
};
};
};
}}
This technique provides a lot more flexibility. One notable improvement is the
ability to add new methods to Iterable
without breaking existing client code,
provided the new methods have a default implementation. Hence, in the MPL11,
all typeclass specializations are required to use this technique, regardless
of whether they actually need defaulted methods.
You might wonder why we use
instantiate<Typeclass>::with<Datatypes...>
instead of just usinginstantiate<Typeclass>
. This is because some typeclasses need to know the datatype(s) they operate on to provide meaningful defaults. Also, we don't useinstantiate<Typeclass, Datatypes...>
because that would make defaulted typeclass parameters hard to handle.
A typeclass specialization using the technique we just saw is called a
typeclass instantiation. When a typeclass instantiation exists for
a typeclass T
and a datatype D
, we say that D
is an instance of
T
. Equivalently, we say that D
instantiates T
, or sometimes that
D
is a T
. The set of definitions that must be provided for a
typeclass to be complete is called the minimal complete definition of
the typeclass. The minimal complete definition is typically the set of methods
without a default implementation, but it must be documented for each typeclass.
The term typeclass instantiation is borrowed from Haskell and should not be mistaken with template instantiation even though they share similarities, especially in the MPL11.
TODO
- Tackle mixed-datatype typeclass-method dispatch
TODO
The development of this library drew inspiration from a couple of projects with different levels of involvement in template metaprogramming. I would like to thank the people involved in these projects for their work, without which this library wouldn't be the same. The most notable sources of inspiration and enlightenment were:
This section contains rationales for some design decisions of the library. In its early development, the library was rewritten several times because fundamental aspects of it needed to be changed. Hence, only the rationales pertaining to the current design are kept here. If you have a question about a design decision that is not explained here, please contact the creator of the library (Louis Dionne).
The following points led to their removal:
- Lazy views were hard to implement because they required creating new iterators, which is a pain. Using a different abstraction for sequences made it much easier.
- Iterators being hard to implement and non-composable is a known problem, which is addressed e.g. by the Boost.Range library or in this paper by Andrei Alexandrescu.
- There is no performance gain to be achieved by using iterators. In fact, it often makes straightforward implementations more complicated, which can hide possible optimizations.
- Implementing infinite ranges using iterators is hacky at best.
There are two main reasons for this. First, if apply
were a method, one would
need to make every lifted metafunction an instance of the typeclass defining
apply
. Since lifted metafunctions are very common, that would be very
cumbersome. Second, making apply
a method requires using the usual method
dispatching mechanism, which adds some overhead.
In some previous design of the library, these were methods. However, allowing
and_
and or_
to be non-strict required special casing these methods. Since
I could not find any use case, I decided to make them normal metafunctions.
Switching the evaluation burden from the caller to the callee makes this useless in most cases.
It introduces a special case in the definition of metafunction that prevents us
from using f::apply<>
to invoke a nullary lifted metafunction. We have to use
apply<f>
, which will then use either f::apply<>
or f::apply
. This adds a
template instantiation and an overload resolution to each lifted metafunction
invocation, which significantly slows down compilation. Considering nullary
metafunctions are of limited use anyway (why would you want a function without
arguments in a purely functional setting?), this is not worth the trouble. Also,
note that MPL-style nullary non-template metafunctions fit in the definition of
a boxed type, so they're not completely forgotten.
-
What should we do for default typeclass methods? 1. We reuse the "official" method and we rebox the arguments.
cpp template <typename x, typename y> using equal_impl = not_<not_equal<box<x>, box<y>>>;
2. We create an "official" unboxed method and we use that. ```cpp template <typename x, typename y> using equal_impl = not_<not_equal_<x, y>>; ``` where `not_equal_` is a non-boxed version of `not_equal`. 3. We directly forward to the implementation of the method in the typeclass. ```cpp template <typename x, typename y> using equal_impl = not_<Comparable<Left, Right>::not_equal_impl<x, y>>; ```
-
Implement cross-type typeclasses.
-
Implement associative data structures.
-
Implement a small DSL to implement inline lifted metafunctions (like Boost.MPL's lambda). Consider let expressions. Using the Boost.MPL lingo, such a DSL should: - Allow leaving placeholders as-is inside a lambda, if this is desirable. - Allow performing placeholder substitution in a lambda without actually evaluating the expression, if this is desirable. - Allow "variadic placeholders", i.e. placeholders expanding to several types. One pitfall of this is using such a placeholder with a metafunction that is not variadic:
```cpp template <typename, typename> struct f; using F = lambda<f<_args>>; // error here, f is not unary using Result = F::apply<int, char>::type; ``` This fails because `f` requires 2 arguments.
-
Consider allowing types to be printed somehow. The idea is to have something like a
Show
typeclass that allows types to be pretty-printed for debugging purposes. -
Think about a convention or a system to customize some metafunction calls. Something neat would be to have a way of passing a custom predicate when comparing sequences; that would make
equal
as powerful as theequal
algorithm from the Boost.MPL. Maybe we can achieve the same effect in another way. -
Consider having a wrapper that allows treating template specializations as data. Something like sequence operations on template specializations and/or tree operations.
-
Consider adding
while_
anduntil
metafunctions. -
Consider ditching
Foreign
and making the default datatype the data constructor itself. -
Consider adding
Either
. -
Right now, we must include
boost/mpl11/core.hpp
to get theinstantiate<>
template in client code. Maybe typeclass headers should take care of it. Or maybeboost/mpl11/core.hpp
should never have to be included by clients altogether? -
Add interoperability with the Boost.MPL, Boost.Fusion and components in the
std
namespace. -
Use
constexpr
to perform numeric computations on homogeneous sequences of integral constants. -
Consider providing data constructors taking unboxed types for convenience.
-
Consider making
int_<>
a simple boxedint
without a value. -
Write a rationale for why we don't have parameterized datatypes. Or is this possible/useful?
-
Make bool.hpp lighter. In particular, it should probably not depend on integer.
-
Design a StaticConstant concept?
-
In the tutorial, when we specialize lifted metafunctions inside the
boost::mpl11
namespace, we don't make them boxed. This makes sense because they are lifted metafunctions, not boxed lifted metafunctions. What should we do? Should we 1. Make the specializations boxed 2. Do not take for granted that they are boxed when we use them in the library and box them redundantly, which is correct but suboptimal. 3. Explicitly state that nothing may be specialized inside theboost::mpl11
namespace unless specified otherwise. Then, specify what can be specialized on a per-component basis and then we apply (2) only to the components that might not be boxed. -
Consider having a public data constructor for
Foreign
, which would simply preserve type identity. Also; might consider changing the name ofForeign
. -
Consider regrouping the typeclass instantiations of a datatype in the datatype itself. This was done in a previous version, but it might have some advantages.
-
Consider providing automatic currying for metafunctions.
-
By making some std algorithms constexpr and providing a couple of containers with constexpr iterators, would we have constexpr for free almost everywhere?
-
Consider making
bool_
a lifted metafunction that behaves likeif_
. -
Provide a better syntax for casting. Consider
cast<Datatype(expr)>
. -
Seriously consider making datatypes lifted metafunctions.
-
Consider prototype-based objects?