The big problems for C++11 SFINAE

So I’ve finally started working on slides for my CppCon 2018 talk “Concepts As She Is Spoke.” (To see the original work to whose title this is an allusion, click through.) I intended to make it as bland and tutorial in nature as I can, but my natural tendency to find the hard problems keeps interfering. Let me take the 25 slides I just wrote and turn them into a blog post, so that I can qualmlessly delete these slides and start over.

Consider the following C++11-era function template.

template<class T>
std::string stringify(const T& t) {
    std::ostringstream oss; oss << t;
    return std::move(oss).str();
}

What kinds of arguments can we pass to this function?

“Anything with an operator<<,” I hear you say. That’s a reasonable first approximation.

If the caller tries to do stringify((const char *)nullptr), we’ll have trouble at runtime. That’s because the type (const char *) satisfies the syntactic requirements of this function (it has an operator<<), but some of its values violate the semantic requirements of this function (they cannot be printed). But we can write off this annoyance to “historical deficiencies of the C++ type system” and plow forward with our first approximation.

In C++11, if we wanted to sanity-check that the user had actually instantiated this template with an operator<<-able type (as opposed to getting some cryptic compiler error downstream), we’d write up a little type trait:

template<class T, class = void>
struct has_output_operator : std::false_type {};

template<class T> struct has_output_operator<T, decltype(void(
    std::declval<std::ostream&>() << std::declval<T>()
))> : std::true_type {};

template<class T>
inline constexpr bool has_output_operator_v = has_output_operator<T>::value;

template<class T>
std::string stringify(const T& t) {
    static_assert(has_output_operator_v<T>); // HERE
    std::ostringstream oss; oss << t;
    return std::move(oss).str();
}

The line marked HERE solves our problem. The actual trait definition takes seven lines of boilerplate, which is less cool, but it’s the best we could do in C++11. (And yeah I’m using several C++17isms in this code, such as inline and static_assert-without-message, but that’s not relevant.)


In C++11, to constrain stringify(x) to compile only when x has an operator<<, we’d have to write even a little more boilerplate:

template<class T, class = std::enable_if_t<has_output_operator_v<T>>>
std::string stringify(const T& t) {
    std::ostringstream oss; oss << t;
    return std::move(oss).str();
}

As mentioned in the previous post, we can apply a cute terse syntax at the cost of two more lines of boilerplate:

template<class T, class = std::enable_if_t<has_output_operator_v<T>>>
using HasOutputOperator = T;

template<class T>
std::string stringify(const HasOutputOperator<T>& t) {
    std::ostringstream oss; oss << t;
    return std::move(oss).str();
}

But pause. Anyone see a problem with this constrained template? (Either the terse one or the non-terse one; they have the same problem.)

Right! We’re constraining on properties of T, but we’re actually using properties of const T&, and the two sets of properties generally don’t have to match! A.k.a., value categories are why we can’t have nice things. (See my previous post on the subject.) What we probably meant to write was:

template<class T, class = std::enable_if_t<has_output_operator_v<const T&>>>
std::string stringify(const T& t)

or

template<class T>
std::string stringify(HasOutputOperator<const T&> t)

Sidebar: The terse syntax looks more philosophically correct in this instance, but notice that if we were taking T by value instead of by any kind of reference, we’d have to go back to putting the constraint at the end of the template clause, rather than inside the parameter declaration. Example:

template<class T, class = HasOutputOperator<T&>>
std::string stringify(T t)

Whether this nitpick about value categories actually matters to you probably depends on whether you’re planning to do your constraints on big OOPy concepts like SequenceContainer and MessageHolder, or tiny typesystemy concepts like CopyConstructible and Callable. If you’re sticking to the big OOPy concepts, then the only way you’ll really run into this issue is by people doing malicious things like

struct Evil {};
std::ostream& operator<< (std::ostream&, Evil&&);

static_assert(has_output_operator_v<Evil>);
static_assert(not has_output_operator_v<const Evil&>);

When working with generic code, you have to have a certain sense of laissez-faire; you can’t possibly design a generic algorithm that is foolproof against this kind of mischief. But certainly we want to avoid falling into traps unnecessarily or accidentally.

Step back: Why do we want a trait?

In this example, why is it important to us that certain specializations — let’s say, stringify(std::make_optional(1)) — should not participate in overload resolution?

I can think of two reasons.

  1. Metaprogramming higher up the call stack. One of our users wants to constrain and do one thing if stringify(x) compiles, and a different thing if stringify(x) does not compile. Therefore we need to make sure that stringify(x) will SFINAE.

  2. Implementation detail of crafting an overload set. We ourselves are planning to provide several overloads of stringify(x), and we want a different overload to be selected when operator<< is unavailable. Therefore we need to make sure that the other overload is a better match than this one, and one way to accomplish that is to make this one SFINAE away completely.

Notice that these two reasons are essentially one reason — someone wants to dispatch and do A or B depending on whether C is possible. The difference is just in who the someone is: in (1) it’s our caller, in (2) it’s ourselves.

So how do we craft an overload set in C++11, a la scenario (2)? Let’s say our “different overload” tries to call x.stringify() if it exists, and otherwise returns the string “unstringable”. Naïvely we’d try this:

template<class T, class = std::enable_if_t<has_output_operator_v<const T&>>>
std::string stringify(const T& t) {
    std::ostringstream oss; oss << t;
    return std::move(oss).str();
}

template<class T, class = std::enable_if_t<has_stringify_method_v<const T&>>>
std::string stringify(const T& t) {
    return t.stringify();
}

template<class T>
std::string stringify(const T& t) {
    return "unstringable";
}

which of course leads to ambiguities. So we need to force some kind of ordering onto the alternatives. We’d like to prioritize the x.stringify() approach over the oss << x approach, for performance. C++17 to the rescue!

template<class T>
std::string stringify(const T& t) {
    if constexpr (has_stringify_method_v<const T&>) {
        return t.stringify();
    } else if constexpr (has_output_operator_v<const T&>) {
        std::ostringstream oss; oss << t;
        return std::move(oss).str();
    } else {
        return "unstringable";
    }
}

Surprisingly, we didn’t run into any arcane pitfalls on this one!


But maybe we’d like to combine both (1) and (2), and say that a type with neither an operator<< nor a .stringify() method is not stringifiable by any means.

template<class T, class = std::enable_if_t<
    has_stringify_method_v<const T&> ||
    has_output_operator_v<const T&>
>>
std::string stringify(const T& t) {
    if constexpr (has_stringify_method_v<const T&>) {
        return t.stringify();
    } else if constexpr (has_output_operator_v<const T&>) {
        std::ostringstream oss; oss << t;
        return std::move(oss).str();
    } else {
        static_assert(false_v<T>, "unreachable");
    }
}

Now this is starting to get weird. That enable_if condition is quite bulky; should we provide a better way for our upstream consumers to ask whether stringify(x) compiles?

template<class T, class = void>
struct is_stringifiable : std::false_type {};

template<class T> struct is_stringifiable<T, decltype(void(
    stringify(std::declval<T>())
))> : std::true_type {};

template<class T>
inline constexpr bool is_stringifiable_v =
    is_stringifiable<T>::value;

This is kind of like the STL’s std::is_swappable_v: it’s named after a specific algorithm and simply tells you whether that algorithm is supported for this type or not.

Or should we save some boilerplate and just write it like this?

template<class T>
inline constexpr bool is_stringifiable_v =
    has_stringify_method_v<const T&> || has_output_operator_v<const T&>;

Then we could move it to the top of our header and rewrite stringify’s declaration to look a lot neater…

template<class T, class = std::enable_if_t<is_stringifiable_v<T>>>
std::string stringify(const T& t)

But this is crazy, right? “T can be stringified if and only if T is stringifiable”? The compiler is happy with it, but are we happy with it? (I’m not.)


Sidebar: We should take that same uncomfortable feeling and channel it any time we are shown an example like

template<class T> requires Sortable<T>
void sort(T t);

or

template<class T> requires Swappable<T>
void swap(T& a, T& b);

These are not appropriate uses of concepts.


Okay, so, to recap, here are the Big Problems, as I see them, with C++11 type traits and SFINAE. These are just the places where we ran into trouble with our simple stringify example, places where we could really use some help and guidance from the language to make our programming job easier.

  1. We want to assert-on, or constrain-on, properties of our parameters, but sometimes those properties are semantic — or even run-time dynamic — rather than syntactic. Recall stringify((const char*)nullptr).

  2. Sometimes the properties are purely syntactic, but we are nonetheless tripped up by value categories. Recall stringify(Evil{}).

  3. Sometimes we dodge the value-category bullet by defining a trait whose usefulness and generality are unnecessarily curtailed. Recall const HasOutputOperator<T>& versus HasOutputOperator<const T&>, and go look again at our last definition of is_stringifiable_v, which sneaks in a const& in the middle of the trait chain for no good reason except that it made for easier reasoning.

  4. We are often unsure if our traits belong “above” or “below” their associated functions, and/or how many public traits we actually need. Recall is_stringifiable_v<T> and is_swappable_v<T>.

  5. The syntax for defining a new type trait in C++11 is bulky.

Now for the good news:

C++2a Concepts fixes (or ignores) every single one of these problems!

Posted 2018-09-08