Tuesday, February 4, 2014

Clang 3.4 and C++14

With each new release, gcc and clang add on more C++11 and C++14 features. While clang has been behind on some features, though ahead on others, they now claim to have C++1y all worked out.

This article is not comprehensive.
Clang's 3.4 C++ release notes:  http://llvm.org/releases/3.4/tools/clang/docs/ReleaseNotes.html#id1
libc++'s C++1y status: http://libcxx.llvm.org/cxx1y_status.html

Note: To compile these examples requires the flags, "-stdlib=libc++" and "-std=c++1y".



Variable templates.

This feature, from N3651, took me most be surprise, but it also seems quite obvious. In the simplest example, let def<T> be a variable that represents the default-constructed value of any type, T.

template<typename T>
constexpr T def = T();
auto x = def<int>; // x = int()
auto y = def<char>; // y = char() 

The proposal uses the example of pi, where it may be more useful to store it as a float or double, or even long double. By defining it as a template, one can have precision when needed and faster, but less precise, operations otherwise.

For another example, consider storing a few prime numbers, but not specifying the type of their container.

template<template<typename...> class Seq>
Seq<int> primes = { 1, 2, 3, 5, 7, 11, 13, 17, 19 };

auto vec = primes<std::vector>;
auto list = primes<std::list>;

Also, the standard library contains many template meta-functions, some with a static value member. Variable templates help there, too.

template<typename T, typename U>
constexpr bool is_same = std::is_same<T,U>::value;

bool t = is_same<int,int>;   // true
bool f = is_same<int,float>; // false

But since variable templates can be specialized just like template functions, it makes as much sense to define it this way:

template<typename T, typename U>
constexpr bool is_same = false;

template<typename T>
constexpr bool is_same<T,T> = true;

Except for when one requires that is_same refers to an integral_constant.

One thing worries me about this feature: How do we tell the difference between template meta-functions, template functions, template function objects, and variable templates? What naming conventions will be invented? Consider the above definition of is_same and the fallowing:

// A template lambda that looks like a template function.
template<typename T>
auto f = [](T t){ ... };

// A template meta-function that might be better as a variable template.
template<typename T>
struct Func : { using value = ...; };

They each has subtly different syntaxes. For example, N3545 adds an operator() overload to std::integral_constant which enables syntax like this: bool b = std::is_same<T,U>(), while N3655 adds std::is_same_t<T,U> as a synonym for std::is_same<T,U>::value. (Note: libc++ is missing std::is_same_t.) Even without variable templates, we have now three ways to refer to the same thing.

Finally, one problem I did have with it is I wrote a function like so:

template<typename T>
auto area( T r ) {
    return pi<T> * r * r;

and found that clang thought pi<T> was undefined at link time and clang's diagnostics did little to point that out.

/tmp/main-3487e1.o: In function `auto $_1::operator()<Circle<double> >(Circle<double>) const':
main.cpp:(.text+0x5e3d): undefined reference to `_ZL2piIdE'
clang: error: linker command failed with exit code 1 (use -v to see invocation

I solved this by explicitly instantiating pi for the types I needed by adding this to main:


Why in main and not in global scope? When I tried it right below the definition of pi, clang thought I wanted to specialize the type. Finally, attempting template<> pi<float>; left the value uninitialized. This is a bug in clang, and has been fixed. Until the next release, variable templates work as long as only non-template functions refer to them.



Generic lambdas and generalized capture.

Hey, didn't I already do an article about this? Well, that one covers Faisal Vali's fork of clang based off of the N3418, which has many more features than this iteration based off the more conservative N3559. Unfortunately it lacks the terseness and explicit template syntax (i.e. []<class T>(T t) f(t)), but it maintains automatic types for parameters ([](auto t){return f(t);}).

Defining lambdas as variable templates helps, but variable templates lack the abilities of functions, like implicit template parameters. For the situations where that may be helpful, it's there.

template<typename T>
auto convert = [](const auto& x) {
    return T(x);

Also, previously, clang couldn't capture values by move or forward into lambdas, which prohibited capturing move-only types by anything other than a reference. Transitively, that meant many perfect forwarding functions couldn't return lambdas.

Now, initialization is "general", to some degree.

std::unique_ptr<int> p = std::make_unique<int>(5);
auto add_p = [p=std::move(p)](int x){ return x + *p; };
std::cout << "5 + 5 = " << add_p(5) << std::endl;
(See also: std::make_unique)

Values can also be copied into a lambda using this syntax, but check out Scott Meyer's article for why [x] or [=] does not mean the same thing as [x=x] for mutable lambdas. (http://scottmeyers.blogspot.de/2014/02/capture-quirk-in-c14.html)

Values can also be defined and initialized in the capture expression.

std::vector<int> nums{ 5, 6, 7, 2, 9, 1 };
auto count = [i=0]( auto seq ) mutable {
    for( const auto& e : seq )
        i++; // Would error without "mutable".
    return i;

gcc has had at least partial support for this since 4.5, but should fully support it in 4.9.



Auto function return types.

This is also a feature gcc has had since 4.8 (and I wrote about, as well), but that was based off of N3386, whereas gcc 4.9 and clang 3.4 base off of N3638. I will not say much here because this is not an entirely new feature, not much has changed, and it's easy to groc.

Most notably, the syntax, decltype(auto), has been added to overcome some of the shortcomings of auto. For example, if we try to write a function that returns a reference with auto, a value is returned. But if we write it...

decltype(auto) ref(int& x) {
    return x;

decltype(auto) copy(int x) {
    return x;

Then a reference is returned when a is given, and a copy when a value is given. (Alternatively, the return type of ref could be auto&.)



More generalized constexprs.

The requirement that constexprs be single return statements worked well enough, but simple functions that required more than one line could not be constexpr. It sometimes forced inefficient implementations in order to have at least some of its results generated at compile-time, but not always all. The factorial function serves as a good example.

constexpr unsigned long long fact( unsigned long long x ) {
    return x <= 1 ? 1ull : x * fact(x-1);

but now we can write...

constexpr auto fact2( unsigned long long x ) {
    auto product = x;
    while( --x ) // Branching.
        product *= x; // Variable mutation.
    return product;

This version may be more efficient, both at compile time and run time.

The accompanying release of libc++ now labels many standard functions as constexpr thanks to N3469 (chrono), 3470 (containers), 3471 (utility), 3302 (std::complex), and 3789 (functional).

Note: gcc 4.9 does not yet implement branching and mutation in constexprs, but does include some of the library enhancements.



std::integer_sequence for working with tuples.

Although this library addition may not be of use to everyone, anyone who has attempted to unpack a tuple into a function (like this guy or that guy or this one or ...) will appreciate N3658 for "compile-time integer sequences". Thus far, no standard solution has existed. N3658 adds one template class, std::integer_sequence<T,t0,t1,...,tn>, and std::index_sequence<t0,...,tn>, which is an integer_sequence with T=size_t. This lets us write an apply_tuple function like so:

template<typename F, typename Tup, size_t ...I>
auto apply_tuple( F&& f, Tup&& t, std::index_sequence<I...> ) {
    return std::forward<F>(f) (
         std::get<I>( std::forward<Tup>(t) )... 
(See also: std::get)

For those who have not seen a function like this, the point of this function is just to capture the indexes from the index_sequence and call std::get variadically. It requires another function to create the index_sequence.

N3658 also supplies std::make_integer_sequence<T,N>, which expands to std::integer_sequence<T,0,1,...,N-1>, std::make_index_sequence<N>, and std::index_sequence_for<T...>, which expands to std::make_index_sequence<sizeof...(T)>.

// The auto return type especially helps here.
template<typename F, typename Tup >
auto apply_tuple( F&& f, Tup&& t ) {
    using T = std::decay_t<Tup>; // Thanks, N3655, for decay_t.

    constexpr auto size = std::tuple_size<T>(); // N3545 for the use of operator().
    using indicies = std::make_index_sequence<size>; 

    return apply_tuple( std::forward<F>(f), std::forward<Tup>(t), indicies() ); 
(See also: std::decay, std::tuple_size, gist

Unfortunately, even though the proposal uses a similar function as an example, there still exists no standard apply_tuple function, nor a standard way to extract an index_sequence from a tuple. Still, there may exist several conventions for applying tuples. For example, the function may be the first element or an outside component. The tuple may have an incomplete argument set and require additional arguments for apply_tuple to work.

Update: Two library proposals in the works address this issue: N3802 (apply), and N3416 (language extension: parameter packs).




While not accepted into C++14, libc++ has an implementation of N3672's optional hidden away in the experimental folder. Boost fans may think of it as the standard's answer to boost::optional as functional programers may think of it as like Haskell's Maybe.

Basically, some operations may not have a value to return. For example, a square root cannot be taken from a negative number, so one might want to write a "safe" square root function that returned a value only when x>0.

#include <experimental/optional>

template<typename T>
using optional = std::experimental::optional<T>;

optional<float> square_root( float x ) {
    return x > 0 ? std::sqrt(x) : optional<float>();

Using an optional is simple because they implicitly convert to bools and act like a pointer, but with value semantics (which is incidentally how libc++ implements it). Without optional, one might use a unique_ptr, but value semantics on initialization and assignment make optional more convenient.

auto qroot( float a, float b, float c ) 
    -> optional< std::tuple<float,float> >
    // Optionals implicitly convert to bools.
    if( auto root = square_root(b*b - 4*a*c) ) {
        float x1 = -b + *root / (2*a);
        float x2 = -b - *root / (2*a);
        return {{ x1, x2 }}; // Like optional{tuple{}}.
    return {}; // An empty optional.


Misc. improvements.

This version of libc++ allows one to retrieve a tuple's elements by type using std::get<T>.

std::tuple<int,char> t1{1,'a'};
std::tuple<int,int>  t2{1,2}; 
int x = std::get<int>(t1); // Fine.
int y = std::get<int>(t2); // Error, t2 contains two ints.

Clang now allows the use of single-quotes (') to separate digits. 1'000'000 becomes 1000000, and 1'0'0 becomes 100. (example) (It doesn't require that the separations make sense, but one cannot write 1''0 or '10.)

libc++ implements N3655, which adds several template aliases in the form of std::*_t = std::*::type to <type_traits>, such as std::result_of_t, std::is_integral_t, and many more. Unfortunately, while N3655 also adds std::is_same_t (see the top of the 7th page), libc++ does not define it. I do not know, but I believe this may be an oversight that will be fixed soon as it requires only one line.

N3421 adds specializations to the members of <functional>. If one wanted to send an addition function into another functions, one might write f(std::plus<int>(),args...), but we new longer need to specify the type and can instead write std::plus<>(). This instantiates a function object that can accept two values of any type to add them. Similarly, std::greater<>, std::less<>, std::equal_to<>, etc...




This may not be the most ground-breaking release, but C++14 expands on the concepts from C++11, improves the library, and adds a few missing features, and I find it impressive that the clang team has achieved this so preemptively. I selected to talk about the features I thought were most interesting, but I did not talk about, for example, sized deallocation, std::dynarray (<experimental/dynarry>), some additional overloads in <algorithm>, or Null Forward Iterators, to name a few. See the bottom for links to the full lists.

The GNU team still needs to do more work to catch up to clang. If one wanted to write code for both gcc 4.9 and clang 3.4, they could use generic lambdas, auto for return types, but not variable templates or generalized constexprs. For the library, gcc 4.9 includes std::make_unique (as did 4.8), the N3412 specializations in <functional>, integer sequences, constexpr library improvements, even experimental::optional (though I'm not sure where), and much more. It may be worth noting it does not seem to include the <type_traits> template aliases, like result_of_t.

See clang's full release notes related to C++14 here: http://llvm.org/releases/3.4/tools/clang/docs/ReleaseNotes.html#id1
For libc++'s improvements, see: http://libcxx.llvm.org/cxx1y_status.html
gcc 4.9's C++14 features: http://gcc.gnu.org/projects/cxx1y.html
And gcc's libstdc++ improvements:  http://gcc.gnu.org/onlinedocs/libstdc++/manual/status.html#status.iso.2014

The code I wrote to test these features: https://gist.github.com/splinterofchaos/8810949

Friday, December 28, 2012

Clang and Generic (Polymorphic) Lambdas.

Recently a Faisal Vali put forth an implementation of n3418, which he co-authored with Herb Stutter and Dave Abraham,  allowing generic lambdas using a fork of clang. It also includes auto type deduction, which I wrote about being implemented in gcc 4.8. There are a few caveats before continuing: This has not been merged into the mainline. It has a few bugs, but Vali is quick to fix them if you point one out. The implementation itself is a proof of concept (similar to automatic type deduction) and so it isn't unreasonable to assume some things might change. Section 2.4 of the proposal (named lambdas) has not yet been implemented. And while this doesn't allow us to do many things that were previously impossible, the possible used to be so verbose that no one would want to do it!

Generic lambdas are profound and may have a great impact on the style of code written in C++. Consider this a light (and lacking) overview of what is possible.

Before continuing, I want to note that I found evidence that some GCC developers had begun working on generic lambdas (from the mailing list: Latest experimental polymorphic lambda patches), however, I cannot find anything more recent than 2009 discussing this, and code using auto and lambdas does not compile.

Being terse.

This patch allows the writing of incredibly terse polymorphic functions, such as these:

auto add = [](auto x, auto y) x + y;
auto times = [](auto x, auto y) x * y; 
auto print_int = [](int x) void(std::cout << x);

No {}, no return, auto-deduced types, and void can even be used to throw away the value of state-full operations. x and y can be anything and the return type is entirely dependent on them. Why is this interesting? Say you want to find the product of a vector.

auto prod = std::accumulate ( 
    v.begin(), v.end(), 1,
    []( int x, int y ){ return x * y; }

Bit of a mouthful, right? v might store ints today, but tomorrow, maybe it will store long long ints to avoid overflow or just unsigned ints to avoid negatives. When the vector's declaration changes, the lambda's argument types need to change, which is a maintenance problem. I currently solve this by writing polymorphic function objects.

constexpr struct {
    template< class X, class Y >
    auto operator () ( X x, Y y )
        -> decltype(x*y)
        return x * y;
} timesObj{}; 

But the above and times are mostly equivalent. (A lambda can be thought of as an automatic function object. It even has a unique type. (see below: Overloading))

auto prod = std::accumulate ( 
    v.begin(), v.end(), 1,

This never needs to change unless the underlying operation (multiplication) changes.

Sometimes, an underlying operation is consistent across types. Using zip_tuple as an example from my article "Zipping and Mapping Tuples", one could write:

std::tuple<int,std::string> a{1,"hello "}, 

// r = {3,"hello world"}
auto r = zipTuple( ([](auto x, auto y) x + y), a, b );

Because of the comma operator, we must put the lambda in parentheses to make clear where it ends. 

Up until now, things like composition could not be lambdas.

template< class F, class G > struct Composition {
    F f;
    G g;

    template< class ...X >
    auto operator () ( X&& ...x ) 
        -> decltype( f(g(std::forward<X>(x)...)) )
        return f( g( std::forward<X>(x)... ) );

template< class F, class G, class C = Composition<F,G> >
C compose( const F& f, const G& g ) {
    return C{f,g};

int main () {
    auto add1 = []( auto x ) x + 1;
    auto Char = []( auto c ) char(c);
    // Prints "a + 1 = b"
    std::cout << "a + 1 = " << compose(Char,add1)('a') << std::endl;

compose is generic and it returns a generic function object. Generic lambdas make the same possible by returning another generic lambda that captures f and g.

auto compose = [](auto f, auto g) 
    [=](auto x) f(g(x));

However, this version of compose only allows one argument. Luckily, generic lambdas can be fully templated, variadic, and perfect forwarding.

auto compose = 
    []( auto f, auto g )
        [=]<class ...X>( X&& ...x ) 
            f( g(std::forward<X>(x)...) );

However, the syntax for these lambdas is so convenient, one might as well drop the functional programming theory and write

auto f = [](char c) char(c+1);

For an example of the power of nested lambdas, consider currying:

auto curry3 = 
    []( auto f )
        [=](auto x) [=](auto y) [=](auto z) 

auto sum3 = [](auto x, auto y, auto z) x + y + z;
auto ten = curry3(sum3)(2)(3)(5)

Nested lambdas especially aid in the use of monads, as I have written about previously ("Monads in C++").

auto justThree = Just(1) >>= [](auto x)
                 Just(2) >>= [](auto y)
                 mreturn<std::unique_ptr>( x + y ); // Or Just(x+y).

This also takes care of the return mreturn problem I discussed in that article.


Overloading functions is obviously useful, but impossible with lambdas alone. To fully take advantage of their brevity, we must have a way to overload them. In the proposal, Mathius Gaunard is attributed with the following:

template<class F1, class F2> struct overloaded : F1, F2
    overloaded(F1 x1, F2 x2) : F1(x1), F2(x2) {}
    using F1::operator();
    using F2::operator();

template<class F1, class F2>
overloaded<F1, F2> overload(F1 f1, F2 f2)
{ return overloaded<F1, F2>(f1, f2); } 

(See also: "Unifying Generic Functions and Function Objects")

This works because lambdas are function objects with a unique type, and can therefore act as the base class for overloaded. This is an unlikely solution because this fact is so rarely taken advantage of, however there is much advantage to take!

Unfortunately, one cannot inherit from function pointers, so, in order to overload lambdas and regular functions together requires a little more work. First, we must define a base type that can handle both function pointers and function objects. It's job is to just forward the arguments to its function.

template< class F > struct Forwarder : F {
    constexpr Forwarder( const F& f ) : F(f) { }

template< class R, class ...X > struct Forwarder<R(*)(X...)> {
    using type = R(*)(X...);
    type f;

    constexpr Forwarder( type f ) : f(f) { }

    constexpr R operator () ( X... x ) {
        return f(x...);

template< class R, class ...X > 
struct Forwarder<R(X...)> : Forwarder<R(*)(X...)>
    using type = R(*)(X...);
    constexpr Forwarder( type f ) 
        : Forwarder<R(*)(X...)>(f)

Function pointers get two specializations because decltype(f)=R(X) and decltype(&f)=R(*)(X). It makes the most sense to specialize for pointers, but only doing so would require we take the address of our functions when we call overload.

Next, Overloaded inherits from two Forwarders.

template< class F, class G >
struct Overloaded : Forwarder<F>, Forwarder<G> {
    constexpr Overloaded( const F& f, const G& g )
        : Forwarder<F>(f), Forwarder<G>(g)

template< class F > F overload( F&& f ) {
    return std::forward<F>(f);

template< class F, class G, class ...H,
          class O1 = Overloaded<F,G> > 
auto overload( const F& f, const G& g, const H& ...h ) 
    -> decltype( overload(O1(f,g),h...) )
    return overload( O1(f,g), h... );

Overloads can be of arity and domain (or argument type). The simplest example, for demonstration purposes, is a printing function.

auto prnt = overload (
    // Anything cout is already defined for.
    []( const auto& x ) 
        -> decltype( void(std::cout << x) )
    { std::cout << x; },

    // Any STL sequence.
    []<class Sequence>( const Sequence& s )
        -> decltype( void(std::begin(s)) )
        std::cout << "[ ";
        for( const auto& x : s ) 
            std::cout << x << ' ';
        std::cout << ']';

    // These are both sequences for which cout is defined.
    // Specializing disambiguates this.
    []( const char* const s ) { std::cout << s; },
    []( const std::string& s ) { std::cout << s; }

// ...
prnt("abc"); // Prints abc.
prnt(std::vector<int>{1,2,3}); // Prints [ 1 2 3 ]. 

Although defining all overloads in a single statement is an annoyance, they are grouped together, they don't require a template<...> line, and the visual clutter is overall less than if prnt were defined as the equivalent (explicit) function object.

Perhaps a function must be overloaded, but decltype or std::enable_if is too accepting and specializing for each type is redundant. For example, one might be annoyed by the last two string specializations of prnt. One solution is to define yet another overload type.

template< class X, class F >
struct UnaryOverload {
    F f;
    UnaryOverload( const F& f ) : f(f) { }

    using result_type = typename std::result_of< F(X) >::type;

    result_type operator () ( X x ) const {
        return f(x);

template< class ...X, class F >
auto unary_overload_set( const F& f ) 
    -> decltype( overload(UnaryOverload<X,F>(f)...) )
    return overload( UnaryOverload<X,F>(f)... );

auto prnt = overload (
    // ...

    unary_overload_set<const char* const,
                       const std::string&>(
        []( auto&& s ) { std::cout << s; }

One might write an overloading class to specialize for specific types, or a category of types, or more generally, a class might be written to do type conversion before calling the inner function, to prepare the output, or whatever one's needs may be. An overloading type might even select one of two functions based on an enable_if.

    // Where pred is a templated type defining pred<X>::value.
    auto h = enable_when<pred>( f, g );

The downsides of overloading function objects include that each overload must be defined all at once and none can be added. That isn't too bad since the point of lambdas is to be brief, but one should be mindful of extensibility when writing generic functions. (In other words, if an overload must be added, is it OK to modify the function object, or must the user be able to add overloads later?)


Without generic lambdas, recursion is possible.

std::function<int(int)> fact = [&]( int x ) x * fact(x-1);

Or, with function pointers, which a lambda may implicitly convert to.

// In global scope:
using IntToInt = int(*)(int);
IntToInt fact = []( auto x ) not x ? 1 : x * fact(x-1);

With generic lambdas, we could write it like this:

auto fact1 = []( auto f, int n ) -> int 
    not n ? 1 : f(f,n-1) * n;
auto fact = []( int x ) -> int 
    fact1( fact1, x );

One might notice that the Fibonacci sequence could be implemented in a similar fashion. In researching recursive lambdas, I came across the fixed-point combinator. Haskell has fix, which can be implemented like this:

auto fix = []( auto f )
    [=]( auto x ) -> decltype(x) f( f, x );

auto fact = fix (
    []( auto f, int n ) -> int
    not n ? 1 : f(f,n-1) * n

auto fib = fix (
    []( int f, int n ) -> int
    n == 0 ? 0 :
    n == 1 ? 1 :
    f(f,n-1) + f(f,n-2)

fix is a generalization of a certain type of recursion. For an idea of how one would implement fix without lambdas, see this Stack Overflow post.

Making prnt above variadic requires a different kind of recursion.

// Variadic void unary.
auto vvu_impl = overload (
    [] (auto,auto) {},
    []<class X, class ...Y>( const auto& self, const auto& u, 
                             X&& x, Y&& ...y ) 
        u( std::forward<X>(x) );
        self( self, u, std::forward<Y>(y)... );

auto vvu = []( const auto& u )
    [&]<class ...x>( const x& ...x )
        vvu_impl( vvu_impl, u, x... );

// Variadic print.
// vprint(x,y...) = prnt(x); prnt(y)...
auto vprint = vvu( prnt );

auto print_line = []<class ...X>( const X& ...x ) 
    vprint( x..., '\n' );
print_line( "abc ", 123, std::vector<int>{1} ); // prints "abc 123 [1]\n" 

We can generalize left-associativity as well.

auto chainl_impl = overload (
    []( auto self, auto b, auto r ) { return r; },
    []<class ...Z>( auto self, auto b, auto x, auto y, Z ...z ) 
        self( self, b, b(x,y), z... )

auto chainl = []( auto b )
    [=]<class ...X>( const X& ...x )
        chainl_impl( chainl_impl, b, x... );

auto sum = chainl(add);
auto three = sum(1,1,1);

// Variadic compose.
auto vcompose = chainl( compose );

auto inc = []( auto x ) ++x;
auto addThree = vcompose( inc, inc, inc );

A good exercise might be to (a) write a variadic version of fix and (b) use that version to reimplement chainl and vprint.

There are, of course, many types of recursion. Implementing recursive lambdas is more complicated than for regular functions, not by too much.


Polymorphic (generic) lambdas are very powerful indeed, but it may take a while before GCC, MSVC, and others catch up, much yet before Faisal Vali's branch is merged back into Clang. Still they may have a strong impact on code written in C++ in the future. Some thought that templates relieved a sort of functional language in C++, and others thought the same of constexpr. Generic lambdas reveal another, but more flexible one.

Lambdas cannot be marked constexpr. In terms of efficiency, I do not think this matters. They are implicitly inlined, so the compiler may still take advantage of any compile-time information it can gather. However, the result of a lambda expression could never be used in a template parameter, for example, which means they don't replace generic constexpr function objects.

Also, explicitly specifying a type is more verbose because the rules are the same as for template member functions, so lambdas can't replace template functions that require explicit parameters.

    auto f = []<class X>() { return x(); }; 
    f.operator()<int>(); // bad

The proposal to add polymorphic lambdas to C++ is not finalized and a few things are up in the air. For example, can we elide auto and just write [](x) f(x)? Should we be allowed to elid the enclosing braces and return? Are the implemented parts of this proposal useful? Remember that the standardization process is open to the public and that we can make our voices heard about the features that impact us.

Personally, I like the proposal as implemented currently. (Requiring auto, but eliding { return ... }.) I would go a step further and say that auto should be extended to allow variadic parameters. (i.e. [](auto ...x) f(x...)) And named lambdas (section 2.4) will be a very nice addition.

What are your thoughts?

Source for this article: https://gist.github.com/4347130 and https://gist.github.com/4381376
Another (google translated) article on generic lambdas: http://translate.google.com/translate?hl=en&sl=ja&u=http://d.hatena.ne.jp/osyo-manga/20121225/1356368196&prev=/search%3Fq%3Dgeneric%2Blambda%2Bc%252B%252B%2Bclang%26start%3D10%26hl%3Den%26safe%3Doff%26client%3Dubuntu%26sa%3DN%26tbo%3Dd%26channel%3Dfs%26biw%3D1535%26bih%3D870&sa=X&ei=v-LbUOG5BOmQ2gXw7ICABg&ved=0CFsQ7gEwBjgK
A long and thoughou article on the fixed-point combinator: http://mvanier.livejournal.com/2897.html

Friday, December 14, 2012

Quick and Easy -- Manipulating C++ Containers Functionally.

Update: Added examples for dup and foldMap.

Probably the most useful parts of the standard C++ library would be container and algorithms support. Who has worked in C++ for any non-trivial amount of time without using std::vector, list, map, or any of the others? <algorithm>, on the other hand, is more something everyone should know. It solves many of the problems that C++ developers encounter on a daily basis.

    "How do I test if there exists an element, x, where p(x) is true?" : std::any_of
    "How do I copy each element, x, where p(x)?" : std::copy_if
    "How do I removed each element, x, where p(x)?" : std::remove_if
    "How do I move elements from one container to another?" : std::move, <algorithm> version.
    "How do I find a subsequence?" : std::search
    "How do I sort an array?" std::sort
    "How do I find the sum of an array?" : std::accumulate

Any programmer worth half their salt could write any of these functions in their sleep--they're basic--and the thing is that these algorithms do get written, over and over and over again. Either because one does not realize a specific <algorithm> function exists, or because one is thinking on a low level and unable to see the higher level abstractions.

What I like most about the STL is that the only requirements for adapting any data type to a sequence are (1) define an iterator, and (2) define begin() and end(). After that, all (if not most) of <algorithm> becomes instantly usable with that type. (As well as the range-based for loop.) This makes it incredibly generic and useful.

What I dislike is its verbosity. For example:

    std::transform( xs.begin(), xs.end(), xs.begin(), f );

Wouldn't this be more clear if written...

    xs = std::transform(xs,f);

And this allows us to compose functions.

    std::transform( xs.begin(), xs.end(), xs.begin(), f );
    std::transform( xs.begin(), xs.end(), xs.begin(), g );

    // vs
    xs = std::transform( std::transform(xs,f), g );

    // Or, using actual composition:
    xs = std::transform( xs, compose(g,f) );

That's what this article will be about. An abstraction over the STL that lends itself to writing more terse, concise code without losing any clarity. This abstraction is less general, by design, because it works on entire containers, not iterators. I am not writing about a replacement for any <algorithm> functions, but an alternative inspired by functional programming.

However, I do go over many <algorithm> functions, so this can also be thought of as a review.

Filtering, Taking, and Dropping: Collecting data.

I've always found the erase-remove idiom an unintuitive solution to such a common problem. I certainly would not have figured it out on my own without the help of the C++ community to point it out. Requiring containers to define a predicated erase wouldn't be generic, and <algorithm> knows only of iterators, not containers, so the standard library can't offer anything simpler. filter fills this gap by combining its knowledge of containers and iterators.

template< class P, class S >
S filter( const P& p, S s ) {
    using F = std::function< bool(typename S::value_type) >;

    s.erase (
        std::remove_if( std::begin(s), std::end(s), std::not1(F(p)) ),

    return s;

// ...

std::vector<int> naturals = {1,2,3,4,5,6,7,8,9/*,...*/};
auto evens = filter( [](int x){ return x%2==0; }, naturals );

See also: std::not1.

It does two things: First, it inverses the predicate meaning we can use positive logic (defining what we want to keep, rather than throw away) and second, it abstracts the erase-remove idiom.

Using filter, we can write a rather quick-and-dirty qsort.

// For each x in s, returns pair( p(x), not p(x) ).
template< class P, class S >
std::pair<S,S> partition( const P& p, S s ) {
    using F = std::function< bool(typename S::value_type) >;

    // There does exist std::partition_copy, 
    // however this function demonstrates a use of filter.
    return std::make_pair ( 
        filter( p,    s ),
        filter( std::not1(F(p)), s )

// Fake Quick-Sort: A non-in-place, qsort-inspired function.
template< class S >
S fake_qsort( S s ) {
    using X = typename S::value_type;

    if( s.size() < 2 )
        return s;

    X pivot = s.back();

    S low, high; 
    std::tie(low,high) = partition (
        [&]( const X& x ) { return x <= pivot; },

    low = fake_qsort( std::move(low) );
    low.push_back( pivot );
    // Append the sorted high to the sorted low.
    high = fake_qsort( std::move(high) );
    std::move( std::begin(high), std::end(high), 
               std::back_inserter(low) );

    return low;
See also: std::partition, std::partition_copy, and std::sort.

take is a function that may seem entirely trivial, at least at first.

template< class S, class _X = decltype( *std::begin(std::declval<S>()) ),
          class X = typename std::decay<_X>::type >
std::vector<X> take( size_t n, const S& s ) {
    std::vector<X> r;
    std::copy_n( std::begin(s), n, std::back_inserter(r) );
    return r;

template< class P, class S, 
          class _X = decltype( *std::begin(std::declval<S>()) ), 
          class X  = typename std::decay<_X>::type >
std::vector<X> takeWhile( const P& p, const S& s ) {
    std::vector<X> r;
    std::copy( std::begin(s), 
               std::find_if( std::begin(s), std::end(s), p ),
               std::back_inserter(r) );
    return r;

It also breaks the convention of returning s's type. There's a reason for that. Infinite lists. Consider this Haskell code:

    take 10 [1..] == [1,2,3,4,5,6,7,8,9,10]

[1...] is an infinite list, starting at one. Obviously, it doesn't actually exist in memory. take returns a finite list that does.

The concept of iterators that represent infinite ranges in C++ isn't new, but neither is it common. std::insert_iterator could insert a theoretically infinite number of elements into a container. std::istream_ and ostream_iterator may read from or write to a file infinitely.

We can create pseudo-containers to represent infinite ranges and plug them into take.

template< class X > struct Reader {
    using iterator = std::istream_iterator<X>;

    iterator b;
    iterator e;

    Reader( iterator b = iterator( std::cin ), 
            iterator e = iterator() )
        : b(b), e(e)

    iterator begin() const { return b; }
    iterator end()   const { return e; }

// Probably ill-advised, 
// but this is /one/ way of doing IO before main().
std::vector<int> three = take( 3, Reader<int>() );

Sometimes we want to take the contents of an entire container, so dup may be helpful.

std::vector<X> dup( const S& s ) {
    std::vector<X> r;
    std::copy( std::begin(s), 
               std::back_inserter(r) );
    return r;

std::ifstream in( "in.txt" );
// Reader's constructor implicitly converts in to an iterator.
// Some may consider this bad style and require the constructor be "explicit".
std::vector<int> contents = dup( Reader<int>(in) );

The counterpart to take is drop, but it does not have take's quirks.

template< class S >
S drop( size_t n, const S& s ) {
    return S (
        std::next( std::begin(s), n ),

// Predicate version
template< class P, class S >
S dropWhile( const P& p, const S& s ) {
    return S (
        std::find_if_not( std::begin(s), std::end(s), p ),

Reader<int> r = drop( 2, Reader<int>() );

drop makes no promises about infinite lists, but unlike most container- or range-based algorithms, it can work on them. In the above example, two integers are read from std::cin, and their values lost.

For another example of the use of pseudo-containers, consider this solution the the first Euler Project problem using boost::irange.

#include <boost/range/irange.hpp>
void euler1() {
    // multiples of...
    auto three = boost::irange( 3, 1001, 3 );
    auto five  = boost::irange( 5, 1001, 5 );

    // Ensure that the final sum has no duplicates.
    std::vector<int> all;
    std::set_union( std::begin(three), std::end(three),
                    std::begin(five),  std::end(five),
                    std::back_inserter(all) );

    std::cout << "The sum of every multiple of 3 or 5 bellow 1000 :" 
        << std::accumulate( std::begin(all), std::end(all), 0 ) 
        << std::endl;

Folding: Reducing a list from many to one. (std::accumulate)

Accumulating is the "imperative" description of folding. Historically, you'd call the variable you update with the results of each calculation the accumulator. To accumulate, then, is to iterate through a sequence, updating the accumulator with each iteration.

Folding is another way to think of it. A fold is a transformation from a list of values to just one value. Haskell defines foldl and foldr, meaning "fold left" and "right".

template< class F, class X, class S >
constexpr X foldl( F&& f, X x, const S& s ) {
    return std::accumulate (
        std::begin(s), std::end(s),
        std::move(x), std::forward<F>(f) 

int main() {
    std::vector<int> v = { 5, 3, 2 };
    std::cout << "((10 - 5) - 3) - 2) = " << foldl( std::minus<int>(), 10, v ) << std::endl;

foldl is really just another name for accumulate. The accumulation function (here, std::minus) expects the accumulator as the left argument and value to accumulate as its right. foldr is reversed: Not only does it iterate in reverse, but expects the accumulator in the right-hand argument.

// A function wrapper that flips the argument order.
template< class F > struct Flip {
    F f = F();

    constexpr Flip( F f ) : f(std::move(f)) { }

    template< class X, class Y >
    constexpr auto operator () ( X&& x, Y&& y )
        -> typename std::result_of< F(Y,X) >::type
        return f( std::forward<Y>(y), std::forward<X>(x) );

template< class F, class X, class S >
constexpr X foldr( F&& f, X x, const S& s ) {
    using It = decltype(std::begin(s));
    using RIt = std::reverse_iterator<It>;
    return std::accumulate (
        // Just foldl in reverse.
        RIt(std::end(s)), RIt(std::begin(s)),

int main() {
    std::vector<int> v = { 5, 3, 2 };
    std::cout << "(2 - (3 - (5-10))) = " << foldr( std::minus<int>(), 10, v ) << std::endl;

Folding is great for monoids; types that have a binary operation with, often, the the signature "X(const X&, const X&)".

std::vector<std::string> strs = { "st", "ri", "ng" };
// std::string associated with (+) is a monoid.
std::cout << "'st' + 'ri' + 'ng' = " << 
    foldl( std::plus<std::string>(), std::string(), strs ) << std::endl;

using Func = std::function< int(int) >;

auto comp = []( Func f, Func g ) {
    return [f,g]( int x ){ return f(g(x)); };

auto inc = []( int x ) { return x+1; };
auto id  = []( int x ) { return x;   };

std::vector<Func> incs = { inc, inc, inc };
// Functions can be monoids under composition.
std::cout << "(inc . inc . inc)(1) = " << 
    foldl( comp, Func(id), incs )(1) << std::endl;

Functional programmers also like to build lists using fold. They build lists starting at the tail, so they typically prefer foldr to foldl. std::forward_list works like [] in Haskell and linked lists in other functional languages. This snippet simply copies the values from the std::vector, v.

using List = std::forward_list<int>;
auto cons = []( List l, int x ) {
    l.push_front( x );
    return std::move(l);

List l = foldr( cons, List(), v );

Note: This one is not an example of a monoid.

Zip and Map: many to many. (std::transform)

To zip two sequences together by some function is the same as calling std::transform. Transform implies modifying each member by some function. Zip implies the same, but with the visual metaphor of combining two lists into one, starting at one end and working up.

template< class F, template<class...>class S, class X, class Y,
          class Res = typename std::result_of< F(X,Y) >::type >
S<Res> zip( F&& f, const S<X>& v, const S<Y>& w ) {
    S<Res> r;
    std::transform( std::begin(v), std::end(v),
                    std::forward<F>(f) );
    return r;

int main() {
    std::vector<int> v = { 5, 3, 2 };
    auto doubleV = zip( std::plus<int>(), v, v );

Note: The only way I have discovered to write zip variadically is with tuples. Since this article is not on tuples, refer to the definition of transform in "Zipping and Mapping Tuples".

Note2: An in-place version of this function is possible, but showing both general and optimized versions of each function would be redundant, and the topic of optimization is worth discussing on its own.

Mapping is similar to zipping--in fact the two-argument forms of zip(f,xs) and map(f,xs) should be equivalent. The three argument form, like map(f,xs,ys), applies f to every combination of x and y.

    map(f,{x,y},{a,b}) == { f(x,a), f(x,b), f(y,a), f(y,b) }

If xs is size N and ys is of size M, then map(f,xs,ys) returns a sequence of size N x M.

template< class F, template<class...>class S, class X,
          class Res = typename std::result_of< F(X) >::type >
S<Res> map( const F& f, const S<X>& s ) {
    S<Res> r;
    std::transform( std::begin(s), std::end(s),
                    std::forward<F>(f) );
    return r;

template< class F, template<class...>class S, class X, class Y,
          class Res = typename std::result_of< F(X,Y) >::type >
S<Res> map( const F& f, const S<X>& xs, const S<Y>& ys ) {
    S<Res> r;

    for( const X& x : xs ) 
        for( const Y& y : ys )
            r.emplace_back( f(x,y) );

    return r;

int main() {
    std::vector<int> v = { 5, 3, 2 };
    std::vector<int> w = { 9, 8, 7 };
    auto sums = map( std::plus<int>(), v, w );

map is a bread and butter function in functional programming.

    // Convert a sequence from one type to another:
    auto ints = map( toInt, floats );

    // In a game loop:
    actors = map( update, actors ); 

    // A deck of cards (four suites with twelve values).
    auto deck = map( make_card, suites, value );

    // Making variants of the same thing from simpler data.
    auto inits = { 1, 2, 3, 4 };
    auto cs = map (
        []( int i ) { return std::complex<float>(i,0.1); },

    // Checking for collisions:
    ColisionObject collisions = map( make_collision, actors, actors );

    // AI:
    states = map( successor, actions, states );

One downfall of map is that it may create redundancies, which makes filter useful in conjunction.

    states = filter (
        map( successor, actions, states )

While this may turn an algorithm from one-pass (update and add if valid) to two-pass (update all states, then filter), it also makes simpler algorithms that can be optimized more easily by the compiler at times. For example,

    for( auto x : xs ) {
        for( auto y : ys ) {
            z = x * y;
            if( pred(z) ) r.push_back(z);

    // or:
    auto r = filter( pred, map(std::multiplies<int>(),xs,ys) );
While only profiling can tell in any given instance, the second example may be faster under some circumstances. The compiler may be able to vectorize the call to map, but have difficulties applying the same optimization to the first because it cannot evaluate both the multiplication and predicate in one vectorized step.

Sometimes, the goal is to calculate something given the data, rather than map it. Naively, one might write something like

    auto r = fold( f, map(g,xs) );

But isn't creating the new container inefficient? What if an in-place version of map were implemented, wouldn't transforming xs before folding still be inefficient? Thus, foldMap is useful.
template< class Fold, class Map, class X, class S >
X foldMap( const Fold& f, const Map& m, X x, const S& s ) {
    for( auto& y : s )
        x = f( std::move(x), m(y) );
    return x;

#include <cctype>
int main() {
    const char* names[] = { "jonh", "mary", "cary" };
    auto appendNames = []( std::string x, std::string y ) {
        return x + " " + y; 
    auto capitolizeName = []( std::string name ) {
        name[0] = std::toupper( name[0] );
        return name;
    std::cout << "Names : " 
        << foldMap (
        ) << std::endl;


Haskell's Data.List is actually a lot like <algorithm>, though on a higher level of abstraction. There are some things that can only be done with iterators, but many that can also only be done with whole containers. Data.List gives some good inspiration for helpful algorithms, even in C++.

But unlike in C++, Haskell uses simple linked lists by default and all of Data.List's function work only on linked lists. This gives both Haskell and functional programming a bad name when people compare Haskell code using linked lists to C++ code using std::vector. (See "C++ Benchmark -- std::vector vs. std::list vs. std::deque") When libraries are written to optimize inefficiencies in the linked list, like Data.Text, they re-implement Data.List's interface and often achieve equivalent efficiency to well-optimized C, but not without plenty of redundancy.

In C++, we can write one static interface that is both generic and efficient. Writing functional code does not mean writing slow code. The mathematical nature of these operations can even help the compiler optimize. The high-level interface of Data.List fits snugly atop of the low-level interface of iterators.

Source for this article: https://gist.github.com/4290166

Wednesday, December 12, 2012

Zipping and Mapping tuples.

Previously, I discussed some basic things that can be done with tuples. I showed how a tuple can be applied to a function, however I did not show how member-wise transformations could be done.

The code of this article builds on the code of the prior.


If we have several tuples, what if we want to apply a function to the nth element of each one?

template< template<size_t> class Fi = Get, size_t i,
          class F, class ...T >
constexpr auto zipRow( const F& f, T&& ...t )
    -> decltype( f(Fi<i>()(std::forward<T>(t))...) )
    return f( Fi<i>()( std::forward<T>(t) )... );

Not hard at all! It basically squishes that row (thinking of t as a column and ti... as a row), using f, into one value. Now, let's say we want to zip together t... into one tuple.

template< template<size_t> class Fi = Get, size_t ...i,
          class Ret, class F, class ...T >
constexpr auto zipIndexList( IndexList<i...>, 
                             const Ret& r, const F& f, T&& ...t )
    -> decltype( r(zipRow<Fi,i>(f,std::forward<T>(t)...)...) )
    return r( zipRow< Fi, i >( f, std::forward<T>(t)... )... );

template< template<size_t> class Fi = Get,
          class Ret, class F, class T, class ...U,
          class _T = typename std::decay<T>::type,
          class IL = typename IListFrom<_T>::type >
constexpr auto zipTupleTo( const Ret& r, const F& f, T&& t, U&& ...u )
    -> decltype( zipIndexList<Fi>(IL(),r,f,std::forward<T>(t),std::forward<U>(u)...) )
    return zipIndexList<Fi>( IL(), r, f, std::forward<T>(t), std::forward<U>(u)... );

int main() {
    auto zipped = zipTupleTo( tuple, std::plus<int>(), tuple(1,10), 
                                                       tuple(2,20) );
    std::cout << " 1 +  2 = " << std::get<0>(zipped) << std::endl;
    std::cout << "10 + 20 = " << std::get<1>(zipped) << std::endl;

In zipIndexList, r represents the function defining how the output is returned. tuple (gist), from the previous article, is just a function object form of std::make_tuple that can be passed to higher order functions. By supplying it as our r, we're saying "just make it a tuple again."

Since most often, we want to zip back into a tuple, it makes sense to define zipTuple like so:

template< template<size_t> class Fi = Get,
          class F, class ...T >
constexpr auto zipTuple( const F& f, T&& ...t )
    -> decltype( zipTupleTo<Fi>(tuple,f,std::forward<T>(t)...) )
    return zipTupleTo<Fi>( tuple, f, std::forward<T>(t)... );

zipTuple is to tuples what std::transform is to sequences. The drawback of std::transform is that it only allows for a unary transformation or binary. Let's write a version that accepts any number of arguments.

// We require these polymorphic function objects.
constexpr struct Inc {
    template< class X >
    constexpr X operator () ( X x ) { return ++x; }
} inc{};

constexpr struct Eq {
    template< class X >
    constexpr bool operator () ( const X& a, const X& b ) 
    { return a == b; }
} eq{};

struct Or {
    template< class X >
    constexpr bool operator () ( const X& a, const X& b ) 
    { return a || b; }

// Wrapper to dereference arguments before applying.
// indirect : (a -> b) -> (a* -> b)
template< class F > struct Indirect {
    F f = F();

    constexpr Indirect( F f ) : f(std::move(f)) { }

    template< class ...It >
    constexpr auto operator () ( It ...it )
        -> decltype( f(*it...) )
        return f( *it... );

template< class F, class I = Indirect<F> > 
constexpr I indirect( F f ) {
    return I( std::move(f) );

#include <vector>
#include <algorithm>
template< class F, class ...X,
          class Result = typename std::result_of<F(X...)>::type,
          class Ret = std::vector<Result> >
Ret transform( const F& f, const std::vector<X>& ...vs )
    Ret r;

    const auto ends = tuple( vs.end()... );

    // Iterate through each vector in parallel.
    for( auto its  = tuple( vs.begin()... ); 
         // This unrolls to: not (it0==end0 || it1==end1 || ...)
         not foldl( Or(), zipTuple(eq,its,ends) );
         // Increment each iterator.
         its = zipTuple( inc, its ) )
        r.emplace_back (
            applyTuple( indirect(f), its )

    return r;

int main() {
    std::vector<int> v = {1,10,100},
                     w = {2,20,200},
                     x = {3,30,300};

    auto vw = transform (
        [](int x, int y, int z){ return x+y+z; }, 
        v, w, x 
    std::cout << "  1 +   2 +   3 = " << vw[0] << std::endl;
    std::cout << " 10 +  20 +  30 = " << vw[1] << std::endl;
    std::cout << "100 + 200 + 300 = " << vw[2] << std::endl;

Note: foldl (gist).


Suppose we want to know the results of adding every combination of {1,2,3} with {9,8,7}. We could write a function that cross-applied every variable from each tuple, but slightly more generally, we can start by taking the Cartesian product.

// product : {x,y} x {a,b} -> {{x,a},{x,b},{y,a},{y,b}}
constexpr struct Product {
    // {...xi...} x {...aj...} -> {xi,aj}
    template< size_t i, size_t j, class T, class U >
    static constexpr auto zip( const T& t, const U& u )
        -> decltype( tuple(std::get<i>(t),std::get<j>(u)) )
        return tuple( std::get<i>(t), std::get<j>(u) );

    // {...xi...} x {a0,a1,a2...} -> { {xi,a0}, {xi,a1}, ... }
    template< size_t i, size_t ...j, class T, class U >
    static constexpr auto withI( IndexList<j...>, const T& t, const U& u )
        -> decltype( tuple(zip<i,j>(t,u)...) )
        return tuple( zip<i,j>(t,u)... );
    // {x...} x {a...} -> { {x,a}... }
    template< size_t ...i, size_t ...j, class T, class U >
    static constexpr auto withIndexes( IndexList<i...>, IndexList<j...> js,
                                       const T& t, const U& u )
        -> decltype( std::tuple_cat(withI<i>(js,t,u)...) )
        return std::tuple_cat( withI<i>(js,t,u)... );

    template< class T, class U,
              class IL  = typename IListFrom<T>::type,
              class IL2 = typename IListFrom<U>::type >
    constexpr auto operator () ( const T& t, const U& u )
        -> decltype( withIndexes(IL(),IL2(),t,u) )
        return withIndexes( IL(), IL2(), t, u );
} product{};

We can now define a map operation to apply the product.

template< class F > struct ApplyF {
    F f = F();

    constexpr ApplyF( F f ) : f(std::move(f)) { }

    template< class T >
    constexpr auto operator () ( T&& t ) 
        -> decltype( applyTuple(f,std::forward<T>(t)) )
        return applyTuple( f, std::forward<T>(t) );

template< class F > 
constexpr ApplyF<F> applyF( F f ) {
    return ApplyF<F>(std::move(f));

constexpr struct MapTuple {
    template< class F, class T, class U >
    constexpr auto operator () ( const F& f, const T& t, const U& u )
        -> decltype( zipTuple(applyF(f),product(t,u)) )
        return zipTuple( applyF(f), product(t,u) );
} mapTuple{};

int main() {
    auto sums = mapTuple( std::plus<int>(), tuple(1,2,3), tuple(7,8,9) );
    std::cout << "map (+) (1,2,3) (7,8,9) = ";
    forEach( printItem, sums );
    std::cout << std::endl;

This prints out:

map (+) (1,2,3) (7,8,9) = 8 9 10 9 10 11 10 11 12 

Zipping applies elements across from each other. Mapping applies everything to everything. (Note: a unary definition of map would be equivalent to a unary definition of zip.)

Tuples as function environments.

This might seem a little off topic, but Haskell has this neat function, id. It works like this:

    id x = x

Simple, right?

    (id f) x y = f x y = id f x y

 id has this neat property that if applied multiple arguments, it applies the tail arguments to the first. This is an artifact of Haskell's curried notation, but we can emulate this behaviour:

constexpr struct Id {
    template< class X >
    constexpr X operator () ( X&& x ) {
        return std::forward<X>(x);

    template< class F, class X, class ...Y >
    constexpr auto operator () ( const F& f, X&& x, Y&& ...y )
        -> typename std::result_of< F(X,Y...) >::type
        return f( std::forward<X>(x), std::forward<Y>(y)... );
} id{};
And now tuples take on a new role: Contained function environments. Consider:

    applyTuple( id, tuple(std::plus<int>(),1,2) );

What does this output? How about

    mapTuple( id, tuple(inc,dec), tuple(5,9) );

    auto pm = tuple(std::plus<int>(),std::minus<int>());
    zipTuple( id, pm, tuple(10,5), tuple(10,5) );


    mapTuple( id, pm, tuple(1,2), tuple(3,4);

I leave implementing the three-tuple version of mapTuple as an exercise, but here's a hint: cross( cross({f},{x}), {y}) = {{{f,x},{y}}}, but you need to take it from that to {{f,x,y}}. (Another good exercise might be to write zipTuple in terms of transposition (wiki).)


This flushes out some basic applications of tuples to functions. applyTuple unpacks a tuple and applies it to a function. foldl and foldr let one apply binary functions to nary tuples, or even singletons (maths concept, not design pattern). zipTuple transforms multiples tuples by a functions, member-wise. mapTuple performs a function for every combination of arguments.

Tuples have unusual mathematical properties compared to other data structures due to the profundity of what they generalize. They can help us shorthand functions to operate in parallel (zip), be passed around as partial or complete function environments, control variadic template parameters, and much, much more that I have either not discussed or yet realized.

One use I haven't discussed, for example, is as a relation, but for an example of this use of tuples, look no further than std::map.

I hope this post has been interesting. Happy coding!

Source for this article: https://gist.github.com/4268029

Monday, December 10, 2012

Fun with tuples.

std::tuple is an odd-but-fun part of the new standard. A lot can be done with them. The members of a tuple can be applied to functions, both in groups and individually. The tuple itself can be treated as a function environment. They can be reorganized, appended, and truncated. The list goes on.

The thing is, what can tuples be used for? Any POD struct or class can be a tuple instead, although this may or may not be desirable. Still, we can write generic algorithms with tuples, whereas we cannot with structs. If we wrote a function that printed any tuple, then any POD we turned into a tuple would suddenly become printable.


Tuples are indexed, which makes accessing them odd.

// A normal struct
struct Vec { int x, y; };
Vec a = { 1, 2 };
a.x = 2; // Normal access.

std::tuple<int,int> b(1,2);
std::get<0>(b) = 2; // Weird access. 

One might think "I'll just write an accessor function!", and that certainly would work. It becomes easier if std::get is made into a function object.

// std::tuple_element<i,T> does not perfect forward.
template< size_t i, class T >
using Elem = decltype( std::get<i>(std::declval<T>()) );

template< size_t i > struct Get {
    template< class T >
    constexpr auto operator () ( T&& t ) 
        -> Elem< i, T >
        return std::get<i>( std::forward<T>(t) );

constexpr auto getx = Get<0>();
constexpr auto gety = Get<1>();

getx(b) = 2; // Less weird.

I define Elem because using std::tuple_element returns what the tuple actually holds. std::get<0>(b) would return an int&, but tuple_element<0,decltype(b)>::type would be int.

One might find it useful to index the tuple backwards, so let's define a function, rget.

 template< size_t i, class T, 
          class _T = typename std::decay<T>::type,
          size_t N = std::tuple_size<_T>::value - 1, // Highest index
          size_t j = N - i >
constexpr auto rget( T&& t ) 
    -> Elem< j, T >
    return std::get<j>( std::forward<T>(t) );

Now we can also define functions to get the first and last elements of any tuple.

constexpr auto head = Get<0>();
constexpr auto last = RGet<0>(); // RGet is defined similarly to Get.

int main() {
    constexpr auto t = std::make_tuple( 1, 'a', "str" );
    std::cout << "head = " << head(t) << std::endl; // prints 1
    std::cout << "last = " << last(t) << std::endl; // prints str

Just for fun, let's also write a function that, if the index is too high, wraps around to the begining.

template< size_t i, class T, 
          class _T = typename std::decay<T>::type,
          size_t N = std::tuple_size<_T>::value,
          size_t j = i % N >
constexpr auto mod_get( T&& t ) 
    -> Elem< j, T >
    return std::get<j>( std::forward<T>(t) );

Now, let's say we want to call a function for every member of a tuple. We need a way of indexing it and applying some function for each index. It starts with the definition of a type to "hold" a list of indexes.

template< size_t ...i > struct IndexList {};

Now, if given a tuple of size 3, we want an IndexList<0,1,2> to represent each index. There are many solutions for how to do this, but they all have in common being obtuse or difficult to understand. The following solution is designed first and foremost to be obvious and intuitive.

template< size_t ... > struct EnumBuilder;

// Increment cur until cur == end.
template< size_t end, size_t cur, size_t ...i >
struct EnumBuilder< end, cur, i... > 
    // Recurse, adding cur to i...
    : EnumBuilder< end, cur+1, i..., cur >

// cur == end; the list has been built.
template< size_t end, size_t ...i >
struct EnumBuilder< end, end, i... > {
    using type = IndexList< i... >;

template< size_t b, size_t e >
struct Enumerate {
    using type = typename EnumBuilder< e, b >::type;

template< class > struct IListFrom;

template< class ...X > 
struct IListFrom< std::tuple<X...> > {
    static constexpr size_t N = sizeof ...(X);
    using type = typename Enumerate< 0, N >::type;

Now, a function that applies each index, and one that prints a tuple's elements for testing.

template< size_t i, size_t ...j, class F, class T >
void forEachIndex( IndexList<i,j...>, const F& f, const T& t ) {
    f( std::get<i>(t) );
    // Recurs, removing the first index.
    forEachIndex( IndexList<j...>(), f, t ); 

template< class F, class T >
void forEachIndex( IndexList<>, const F& f, const T& t ) {
    // No more indexes.

template< class F, class T >
void forEach( const F& f, const T& t ) {
    constexpr size_t N = std::tuple_size<T>::value;
    using IL = typename Enumerate<0,N>::type;
    forEachIndex( IL(), f, t );

constexpr struct PrintItem {
    template< class X >
    void operator () ( const X& x ) const {
        std::cout << x << ' ';
} printItem{};

int main() {
    constexpr auto t = std::make_tuple( 1, "and", 2 );
    std::cout << "t = "; 
    forEach( printItem, t ); // Prints "1 and 2"
    std::cout << std::endl;

Applying a tuple to a function.

This is probably one of the most pondered questions about tuples. "How do I apply one to a function?" This is easy since we already have IndexList defined.

template< size_t ...i, class F, class T >
constexpr auto applyIndexList( IndexList<i...>, F f, T&& t )
    -> typename std::result_of< F( Elem<i,T>... ) >::type
    return f( std::get<i>(std::forward<T>(t))... );

template< class F, class T,
          class _T = typename std::decay<T>::type;
          class IL = typename IListFrom<_T>::type >
constexpr auto applyTuple( const F& f, T&& t )
    -> decltype( applyIndexList(IL(),f,std::forward<T>(t)) )
    return applyIndexList( IL(), f, std::forward<T>(t) );

// Functional programmers may recognize this as cons.
constexpr struct PushFront {
    template< class ...X, class Y >
    constexpr auto operator () ( std::tuple<X...> t, Y y )
        -> std::tuple< Y, X... >
        return std::tuple_cat( tuple(std::move(y)), std::move(t) );
} pushFront{};

// Chain Left.
constexpr struct ChainL {
    template< class F, class X >
    constexpr X operator () ( const F&, X x ) {
        return x;

    template< class F, class X, class Y, class ...Z >
    constexpr auto operator () ( const F& b, const X& x, const Y& y, const Z& ...z) 
        -> decltype( (*this)(b, b(x,y), z... ) )
        return (*this)(b, b(x,y), z... );
} chainl{};

// Fold Left.
constexpr struct FoldL {
    // Given f and {x,y,z}, returns f( f(x,y), z ).
    template< class F, class T >
    constexpr auto operator () ( const F& f, const T& t ) 
        -> decltype( applyTuple(chainl,pushFront(t,f)) )
        return applyTuple( chainl, pushFront(t,f) );
} foldl{};

auto ten = foldl( std::plus<int>(), std::make_tuple(1,2,3,4) );

We can call applyIndexList with different index lists to get interesting results.

// Because std::make_tuple can't be passed 
// to higher order functions.
constexpr struct MakeTuple {
    template< class ...X >
    constexpr std::tuple<X...> operator () ( X ...x ) {
        return std::tuple<X...>( std::move(x)... );
} tuple{}; // function tuple that construct std::tuples.

// Returns the initial elements. (All but the last.)
// init( {1,2,3} ) = {1,2}
template< class T,
          size_t N = std::tuple_size<T>::value, 
          class IL = typename Enumerate< 0, N-1 >::type >
constexpr auto init( const T& t )
    -> decltype( applyIndexList(IL(),tuple,t) )
     // Construct a new tuple from the initial indexes.
     return applyIndexList( IL(), tuple, t );

// Returns a new tuple with every value from t except the first.
// tail( {1,2,3} ) = {2,3}
template< class T,
          size_t N = std::tuple_size<T>::value, 
          class IL = typename Enumerate< 1, N >::type >
constexpr auto tail( const T& t )
    -> decltype( applyIndexList(IL(),tuple,t) )
    return applyIndexList( IL(), tuple, t );

Remember Get and RGet from above? They're templated function objects based on an index. We can write a more generic applyIndexList that allows specifying such a function and without losing the default behaviour.

template< template<size_t> class Fi = Get, size_t ...i, class F, class T >
constexpr auto applyIndexList( IndexList<i...>, const F& f, T&& t )
    -> typename std::result_of< F( 
        typename std::result_of< Fi<i>(T) >::type...
    ) >::type
    return f( Fi<i>()(std::forward<T>(t))... );

template< template<size_t> class Fi = Get, class F, class T,
          class _T = typename std::decay<T>::type,
          class IL = typename IListFrom<_T>::type >
constexpr auto applyTuple( const F& f, T&& t )
    -> decltype( applyIndexList<Fi>(IL(),f,std::forward<T>(t)) )
    return applyIndexList<Fi>( IL(), f, std::forward<T>(t) );

// Reconstruct t in reverse.
template< class T >
constexpr auto reverse( const T& t ) 
    -> decltype( applyTuple<RGet>(tuple,t) )
    return applyTuple< RGet >( tuple, t );

// Fold Right.
constexpr struct FoldR {
    // Given f and {x,y,z}, returns f( f(z,y), x ).
    template< class F, class T >
    constexpr auto operator () ( const F& f, const T& t ) 
        -> decltype( foldl(f,reverse(t)) )
        return foldl( f, reverse(t) );
} foldr{};

This leaves us with two ways of transforming tuples: modifying the index list and defining an alternative get function. foldr and foldl give us a way to fold a tuple into one value.

Tuples and functions.

Perhaps we have a function that takes a tuple, but its arguments are not tuples. Haskell defined a function, curry, to transform the function from a pair-taking one to a two-argument function. Haskell does not have the same expressiveness with variadic types, so they can't write this more general C++ version.

// curry : ( (a,b...) -> c ) x a x b... -> c
template< class F, class ...X >
constexpr auto curry( const F& f, X&& ...x )
    -> decltype( f( std::forward_as_tuple(std::forward<X>(x)...) ) )
    return f( std::forward_as_tuple(std::forward<X>(x)...) );

// Pair Sum.
// psum : (int,int) -> int
unsigned int psum( std::tuple<int,int> p ) {
    return head(p) + last(p);

auto five = curry( psum, 3, 2 );
Haskell also defines uncurry, which is the same as applyTuple. The most important distinction between Haskell's curry and uncurry and this is that Haskell's curry is a unary higher order function, whereas this is binary. However, the two are the same if one considers the unary version a partial application of the binary one.

Tuples can be used as a sort of partial application on their own. One might store some values in a tuple, and add more arguments later to be applied to some function. For a somewhat contrived example, consider the following:

constexpr struct PushBack {
    template< class ...X, class Y >
    constexpr auto operator () ( std::tuple<X...> t, Y y )
        -> std::tuple< X..., Y >
        return std::tuple_cat( std::move(t), tuple(std::move(y)) );
} pushBack{};

#include <cmath>

// Quadratic root.
constexpr struct QRoot {
    using result = std::tuple<float,float>;

    result operator () ( float a, float b, float c ) const {
        float root = std::sqrt( b*b - 4*a*c );
        float den  = 2 * a;
        return result( (-b+root)/den, (-b-root)/den );
} qroot{};

std::ostream& operator << ( std::ostream& os, const QRoot::result r ) {
    return os << std::get<0>(r) << " or " << std::get<1>(r);

int main() {
    auto ab = std::make_tuple( 1, 3 );
    auto qroot_ab = [&] ( float c ) {
        return applyTuple( qroot, pushBack(ab,c) );
    std::cout << "qroot(1,3,-4) = " << qroot_ab(-4) << std::endl;
    std::cout << "qroot(1,3,-5) = " << qroot_ab(-5) << std::endl;

    auto bc = std::make_tuple( 3, -4 );
    auto qroot_bc = [&] ( float a ) {
        return applyTuple( qroot, pushFront(bc,a) );
    std::cout << "qroot(1,3,-4) = " << qroot_bc(1) << std::endl;
    std::cout << "qroot(1,3,-5) = " << qroot_bc(2) << std::endl;

One persuasive use-case of tuples is being able to freely manipulate variadic parameters.

template< class ...X >
constexpr auto third_arg( X&& ...x )
    -> Elem< 2, std::tuple<X...> >
    return std::get<2>( std::forward_as_tuple(std::forward<X>(x)...) );

Variadic parameters can be forwarded into a tuple, meaning anything we can do with tuples, we can do with variadic parameters. Arguments can be compacted into tuples, modified, reordered, and re-expanded into functions. One annoyance with the ... is that it eats up everything to the right. The following is ill-formed.

template< class ...X, class F >
constexpr auto backwards( X&& ...x, const F& f )
    -> typename std::result_of< F(X...) >::type
    return f( std::forward<X>(x)... )

I leave the solution as an exercise for the reader.


This article is introductory, at best. I must admit I have much less practice with them than other C++11 features, but this seems to be true of GCC, too! Attempting to compile the following will cause an internal error since at least 4.7.

#include <tuple>

template< class X > struct Hold {
 X x;
 constexpr Hold( X x ) : x(std::move(x)) { }

constexpr auto a = Hold<int>( 1 ); //ok
auto b = Hold<std::tuple<char>>( std::tuple<char>('c') ); // not
constexpr auto c = Hold<std::tuple<char>>( std::tuple<char>('c') ); // not

int main() {} 

I believe it's quite possible that whole programs could be written with tuples and basic data types, whether or not it should be preferred. We could look at them as a general utility class, but I think it would be appropriate to see them as lightweight, generalized, composable, structs. I plan on writing a follow-up to cover applying multiple tuples, transforming tuples by functions, and a few other things. If anyone has any interesting use-cases or neat tricks for tuples, I'd love to hear about it in the comments.  

The next article covers element-wise application, applying multiple tuples, and more: "Zipping and Mapping Tuples".

Source code: https://gist.github.com/4256092  
Tuple reference: http://en.cppreference.com/w/cpp/utility/tuple

Sunday, December 9, 2012

GCC 4.8 Has Automatic Return Type Deduction.

Note: GCC 4.8 is still in development; this article is based on Ubuntu's snapshot package of 4.8. I do not know about availability on other platforms. I say "has" because it does work and code can be written using it right now, even if it's in testing.  

Update: It turns out this feature has been implemented to test n3386. You can read the discussion and even see the patch on the mailing list: http://gcc.gnu.org/ml/gcc-patches/2012-03/msg01599.html

Are your two favourite C++11 features decltype and declval? I have mixed feelings. On one hand, it lets me write code like this

template< class X, class Y >
constexpr auto add( X&& x, Y&& y )
    -> decltype( std::declval<X>() + std::declval<Y>() )
    return std::forward<X>(x) + std::forward<Y>(y);

and know that it will work on any type and do the optimal thing if x or y should be moved or copied (like if X=std::string). On the other hand, it's tedious. "forward" and "declval" are both seven letter words that have to be typed out every time for every function, per variable. Then there's the std:: prefix and <X>(x) suffix. The only benefit of using declval over forward is a savings of one letter not typed.

But someone must have realized there's a better way. If the function is only one statement, and the return type is the declval of that statement, couldn't the compiler just figure out what I mean when I say this?

template< class X, class Y >
constexpr auto add( X&& x, Y&& y ) {
    return std::forward<X>(x) + std::forward<Y>(y);

March of this year, one Jason Merrill proposed just this (n3386) and GCC has already implemented it in 4.8 (change log), though it requires compiling with -std=c++1y. One can play with 4.8 on Ubuntu with gcc-snapshot. (Note that it doesn't modify your existing gcc install(s) and puts it in /usr/lib/gcc-snapshot/bin/g++. Also, I have been unable to install any but the third-to-most recent package.) I hope it is not too much trouble to install on other distros/platforms.

So if your favourite c++11 feature is decltype and declval, prepare to never use them again. The compiler can deduce the type for you implicitly, and better, and it works even if the function is longer than one line. Take for example, reasonably complex template functions like the liftM function I wrote for "Arrows and Keisli":

constexpr struct LiftM {
    template< class F, class M, class R = Return<typename std::decay<M>::type> >
    constexpr auto operator () ( F&& f, M&& m )
        -> decltype( std::declval<M>() >>= compose(R(),std::declval<F>()) )
        return std::forward<M>(m) >>= compose( R(), std::forward<F>(f) );

    template< class F, class A, class B >
    constexpr auto operator () ( F&& f, A&& a, B&& b )
        -> decltype( std::declval<A>() >>= compose (
                rcloset( LiftM(), std::declval<B>() ),
            ) )
        return std::forward<A>(a) >>= compose (
            rcloset( LiftM(), std::forward<B>(b) ),
} liftM{};

Could be written instead:

constexpr struct LiftM {
    template< class F, class M >
    constexpr auto operator () ( F&& f, M&& m ) {
        using R = Return< typename std::decay<M>::type >;
        return std::forward<M>(m) >>= compose( R(), std::forward<F>(f) );

    template< class F, class A, class B >
    constexpr auto operator () ( F&& f, A&& a, B&& b ) {
        return std::forward<A>(a) >>= compose (
            rcloset( LiftM(), std::forward<B>(b) ),
} liftM{};

Automatic type deduction didn't exactly make this function more obvious or simple, but it did remove the visual cruft and duplication of the definition. Now, if I improve this function to make it more clear, I won't have a decltype expression to have to also edit.

To be fair, this doesn't entirely replace decltype. auto doesn't perfect forward. But it seems to work as expected, most of the time.

For another example of the use-case of auto return type deduction, consider this program:

#include <tuple>

int main() {
    auto x = std::get<0>( std::tuple<>() );

This, small, simple, and obviously wrong program generates an error message 95 lines long. Why? GCC has to check make sure this isn't valid for the std::pair and std::array versions of get, and when it checks the tuple version, it has to instantiate std::tuple_element recursively to find the type of the element. It actually checks for the pair version first, so one has to search the message for the obviously correct version and figure out why it failed. A simple one-off bug in your program could cause a massive and difficult to parse error message. We can improve this simply.

#include <tuple>

template< unsigned int i, class T >
auto get( T&& t ) {
    using Tup = typename std::decay<T>::type;
    static_assert( i < std::tuple_size<Tup>::value,
                   "get: Index too high!" );
    return std::get<i>( std::forward<T>(t) );

int main() {
    int x = get<0>( std::tuple<>() );

How much does this shrink the error by? Actually, it grew to 112 lines, but right at the top is

auto.cpp: In instantiation of 'auto get(T&&) [with unsigned int i = 0u; T = std::tuple<>]':
auto.cpp:13:36:   required from here
auto.cpp:7:5: error: static assertion failed: get: Index too high!

The error message might be a little bigger, but it tells you right off the bat what the problem is, meaning one has less to parse.

Similar to this static_assert example, typedefs done as default template arguments can be moved to the function body in many cases.

template< class X, class Y = A<X>, class Z = B<Y> >
Z f( X x ) {
    Z z;
    return z;

can now be written

template< class X > // Simpler signature.
auto f( X x ) {
    using Y = A<X>; // Type instantiation logic
    using Z = B<Y>; // done in-function.
    Z z;
    return z;

Looking forward.

This release of GCC also implements inheriting constructors, alignas, and attribute syntax. It also may have introduced a few bugs; for example, my library, which compiles with 4.7, does not with 4.8, producing many undefined references.

The other features of this release might not be quite so compelling, but automatic type deduction alone is one powerful enough to change the way coding is done in C++--again. Heavily templated code will become a breeze to write and maintain as figuring out the return type is often the hardest part. I find it encouraging that this has been implemented so quickly. Of coarse, it's not standard, at least not yet.

On a final note, I wonder how this will interact with the static if. It would be nice, were the following well-formed:

template< class X >
auto f( X x ) {
    if( std::is_same<int,X>::value )
        return "Hi"; // Int? Return string.
        return x / 2; // Other? halve.