operator??

Document #: P2561R0
Date: 2022-07-09
Project: Programming Language C++
Audience: EWG
Reply-to: Barry Revzin
<>

1 Preface

It is important to clarify a few things up front. It is not the position of this paper that exceptions are bad. Or that exceptions are good. It is not the goal of this paper to convince you to start using exceptions, nor is it to convince you to stop using exceptions.

This paper simply recognizes that there are many code bases (or parts thereof) that do not use exceptions and probably will not in the future. That could be for performance or space reasons. It could be because exceptions are unsupported on a particular platform. It could be for code understandability reasons. Regardless, some code bases do not use exceptions. Moreover, some problems are not solved well by exceptions – even in code bases that otherwise use them to solve problems that they are more tailored to solve.

The problem is, C++ does not currently have a good story for error handling without exceptions. We’re moving away from returning bool or error codes in favor of solutions like std::expected ([P0323R12]), but the ergonomics of such types are not there yet. Bad ergonomics leads to code that is clunkier than it needs to be, harder to follow, and, significantly and ironically, error-prone.

We should try to improve such uses too.

2 Introduction

Let’s start with a fairly small example of a series of functions that can generate errors, but don’t themselves handle them - they just need to propagate them up. With exceptions, this might look like:

auto foo(int i) -> int; // might throw an E
auto bar(int i) -> int; // might throw an E

auto strcat(int i) -> std::string {
    int f = foo(i);
    int b = bar(i);
    return std::format("{}{}", f, b);
}

There’s a lot to like about exceptions. One nice advantage is the zero syntactic overhead necessary for propagating errors. Errors just propagate. You don’t even have to know which functions can fail.

We don’t even need to declare variables to hold the results of foo and bar, we can even use those expressions inline:

auto foo(int i) -> int; // might throw an E
auto bar(int i) -> int; // might throw an E

auto strcat(int i) -> std::string {
    return std::format("{}{}", foo(i), bar(i));
}

But with the newly adopted std::expected<T, E>, it’s not quite so nice:

auto foo(int i) -> std::expected<int, E>;
auto bar(int i) -> std::expected<int, E>;

auto strcat(int i) -> std::expected<std::string, E>
{
    auto f = foo(i);
    if (not f) {
        return std::unexpected(f.error());
    }

    auto b = bar(i);
    if (not b) {
        return std::unexpected(b.error());
    }

    return std::format("{}{}", *f, *b);
}

This is significantly longer and more tedious because we have to do manual error propagation. This manual error propagation is most of the code in this short example, and is bad not just because of the lengthy boilerplate, but also because:

In an effort to avoid… that… many libraries or code bases that use this sort approach to error handling provide a macro, which usually looks like this (Boost.LEAF, Boost.Outcome, mediapipe, SerenityOS, etc. Although not all do, neither folly’s fb::Expected nor tl::expected nor llvm::Expected provide such):

auto strcat(int i) -> std::expected<std::string, E>
{
    SOMETHING_TRY(int f, foo(i));
    SOMETHING_TRY(int b, bar(i));
    return std::format("{}{}", f, b);
}

Which avoids all those problems, though each such library type will have its own corresponding macro. Also these TRY macros (not all of them have TRY in the name) need to be written on their own line, since they are declarations - thus the one-line version of strcat in the exception version isn’t possible.

Some more adventurous macros take advantage of the statement-expression extension, which would allow you to do this:

auto strcat(int i) -> std::expected<std::string, E>
{
    int f = SOMETHING_TRY_EXPR(foo(i));
    int b = SOMETHING_TRY_EXPR(bar(i));
    return std::format("{}{}", f, b);
}

And thus also write both macros inline. But this relies on compiler extensions, and this particular extension isn’t quite as efficient as it could be - and in particular it doesn’t move when it should.

Both macros also suffer when the function in question returns expected<void, E>, since you cannot declare (or assign to) a variable to hold that value, so the macro needs to emit different code to handle this case (Boost.LEAF, Boost.Outcome, etc.)

To that end, in search for nice syntax, some people turn to coroutines:

auto strcat(int i) -> std::expected<std::string, E>
{
    int f = co_await foo(i);
    int b = co_await bar(i);
    co_return std::format("{}{}", f, b);

    // ... or
    co_return std::format("{}{}", co_await foo(i), co_await bar(i));
}

This can be made to work in a fully-conformant way (at the syntactic cost of having to now write co_return), and we can use the same syntax for both the void and non-void cases.

However, currently even the simple cases allocate which make this approach unusuable in many production contexts. The coroutine machinery also isn’t fully composable and runs into problems once you start doing something like optional<expected<T, E>> (or vice versa) or task<optional<T>>.

Which means the best-case today still involves being jealous of exceptions macros.

3 An automatic propagation operator

Let’s talk about Rust.

Rust’s primary form of error handling is a sum type named Result<T, E>. Taking our original example here and rewriting it in Rust (as one does) would look like this:

Rust
C++
fn strcat(i: i32) -> Result<String, E> {
    let f = match foo(i) {
        Ok(i) => i,
        Err(e) => return Err(e),
    };

    let b = match bar(i) {
        Ok(i) => i,
        Err(e) => return Err(e),
    }

    Ok(format!("{}{}", f, b))
}
auto strcat(int i) -> std::expected<std::string, E> {
    auto f = foo(i);
    if (not f) {
        return std::unexpected(f.error());
    }

    auto b = bar(i);
    if (not b) {
        return std::unexpected(b.error());
    }

    return std::format("{}{}", *f, *b);
}

This fully manual version is already better than the C++ version due to pattern matching’s ability to just give a name to the thing we care about (the value) and avoid giving a name to the thing we don’t care about (the Result object).

But this isn’t the way you do things in Rust.

Originally, there was a try! macro which was defined mostly as that match expression I have above. But then this got generalized into operator?, whose behavior is driven by the Try trait. That allows simply writing this:

Rust
C++ with exceptions
fn strcat(i: i32) -> Result<String, E> {
    let f = foo(i)?;
    let b = bar(i)?;
    Ok(format!("{}{}", f, b))

    // ... or simply ...
    Ok(format!("{}{}", foo(i)?, bar(i)?))
}
auto strcat(int i) -> std::string {
    int f = foo(i);
    int b = bar(i);
    return std::format("{}{}", f, b);

    // ... or simply ...
    return std::format("{}{}", foo(i), bar(i));
}

Now, Rust still has manual error propagation, but it’s the minimal possible syntactic overhead: one character per expression.

Importantly, one character per expression is still actually an enormous amount more overhead than zero characters per expression, since that implies that you cannot have error-neutral functions - they have to manually propagate errors too.

But to those people who write code using types like std::expected today, who may use the kinds of macros I showed earlier or foray into coroutines, this is kind of a dream?

3.1 Syntax for C++

Before diving too much into semantics, let’s just start by syntax. Unfortunately, C++ cannot simply grab the Rust syntax of a postfix ? here, because we also have the conditional operator ?:, with which it can be ambiguous:

auto res = a ? * b ? * c : d;

That could be parsed two ways:

auto res1 = a ? (*(b?) * c) : d;
auto res2 = ((a?) * b) ? (*c) : d;

What if you assume that a ? is a conditional operator and try to parse that until it fails, then back up and try again to parse a postfix ? operator? Is that really a viable strategy? If we assume both ?s are the beginning of a conditional, then that will eventually fail since we hit a ; before a second : - but it’s the outer ? that failed, not the inner - do we retry the inner first (which would lead to the res1 parse eventually) or the outer first (which would lead to the res2 one)?

Maybe this is doable with parsing heroics, but at some point I have to ask if it’s worth it. Especially since we can just pick something else: ??

This is only one character longer, and just as questioning. It’s easily unambiguous by virtue of not even being a valid token sequence today. But it’s worth commenting further on this choice of syntax.

3.1.1 Why not try?

For those libraries that provide this operation as a macro, the name is usually TRY and [P0779R0] previously suggested this sort of facility under the name operator try. As mentioned, Rust previously had an error propagation macro named try! and multiple other languages have such an error propagation operator (Zig, Swift, Midori, etc.).

The problem is, in C++, try is strongly associated with exceptions. That’s what a try block is for: to catch exceptions. In [P0709R4], there was a proposal for a try expression (in §4.5.1). That, too, was tied in with exceptions.

Having a facility for error propagation in C++ which has nothing to do with exceptions still use the keyword try would be, I think, quite misleading. And the goal here isn’t to interact with exceptions at all - it’s simply to provide automated error propagation for those error handling cases that don’t use exceptions.

3.1.2 ?? in other languages

?? is called a “null (or nil) coalescing operator” in some languages (like C# or JavaScript or Swift) where x ?? y is roughly equivalent to what C++ would spell as x ? *x : y except that x is only evaluated once. Kotlin spells this operator ?:, but it behaves differently from the gcc extension since x ?: y in gcc evaluates as x ? x : y rather than x ? *x : y.

For x being some kind of std::optional<T> or std::expected<T, E>, this can mostly already be spelled x.value_or(y). The difference is that here y is unconditionally evaluated, which is why [P2218R0] proposes a separate opt.value_or_else(f) which invokes f. Which would make a proper equivalence be spelled x.value_or_else([&]{ return y; }).

I’m not aware of any proposals to add this particular operator in C++, but because we already have two types that directly provide that functionality (as would many other non-std flavors thereof), and because it’s fairly straightforward to write such an algorithm generically, it wouldn’t seem especially valuable to have a dedicated operator for this functionality – so it’s probably safe to take for this use-case.

Of course, now we have to talk about semantics.

3.2 Semantics

This paper suggests that ?? evaluate as follows:

auto strcat(int i) -> std::expected<std::string, E>
{


    int f = foo(i)??;









    int b = bar(i)??;








    return std::format("{}{}", f, b);
}
auto strcat(int i) -> std::expected<std::string, E>
{
    using _Return = std::f_traits<
        std::expected<std::string, E>>;

    auto&& __f = foo(i);
    using _TraitsF = std::try_traits<
        std::remove_cvref_t<decltype(__f)>>;
    if (not _TraitsF::is_ok(__f)) {
        return _Return::from_error(
            _TraitsF::extract_error(FWD(__f)));
    }
    int f = _TraitsF::extract_value(FWD(__f));

    auto&& __b = bar(i);
    using _TraitsB = std::try_traits<
        std::remove_cvref_t<decltype(__b)>>;
    if (not _TraitsB::is_ok(__b)) {
        return _Return::from_error(
            _TraitsB::extract_error(FWD(__b)));
    }
    int b = _TraitsB::extract_value(FWD(__b));

    return std::format("{}{}", f, b);
}

The functionality here is driven by a new traits type called std::try_traits, such that a given specialization supports:

Note that this does not support deducing return type, since we need the return type in order to know how construct it - the above desugaring uses the return type of std::expected<std::string, E> to know how to re-wrap the potential error that foo(i) or bar(i) could return. This is important because it avoids the overhead that nicer syntax like std::unexpected or outcome::failure introduces (neither of which allow for deducing return type anyway, at least unless the function unconditionally fails), while still allowing nicer syntax.

These functions are all very easy to implement for the kinds of types that would want to support a facility like ??. Here are examples for optional and expected (with constexpr omitted to fit):

template <class T>
struct try_traits<optional<T>> {
  using value_type = T;
  using error_type = nullopt_t;

  auto is_ok(optional<T> const& o) -> bool {
    return o.has_value();
  }

  // extractors
  auto extract_value(auto&& o) -> auto&& {
    return *FWD(o);
  }
  auto extract_error(auto&&) -> error_type {
    return nullopt;
  }

  // factories
  auto from_value(auto&& v) -> optional<T> {
    return optional<T>(in_place, FWD(v));
  }
  auto from_error(nullopt_t) -> optional<T> {
    return {};
  }
};
template <class T, class E>
struct try_traits<expected<T, E>> {
  using value_type = T;
  using error_type = E;

  auto is_ok(expected<T, E> const& e) -> bool {
    return e.has_value();
  }

  // extractors
  auto extract_value(auto&& e) -> auto&& {
    return *FWD(e);
  }
  auto extract_error(auto&& e) -> auto&& {
    return FWD(e).error();
  }

  // factories
  auto from_value(auto&& v) -> expected<T, E> {
    return expected<T, E>(in_place, FWD(v));
  }
  auto from_error(auto&& e) -> expected<T, E> {
    return expected<T, E>(unexpect, FWD(e));
  }
};

This also helps demonstrate the requirements for what try_traits<O> have to return:

In the above case, try_traits<expected<T, E>>::extract_error will always give some kind of reference to E (either E&, E const&, E&&, or E const&&, depending on the value category of the argument), while try_traits<optional<T>>::extract_error will always be std::nullopt_t, by value. Both are fine, it simply depends on the type.

Since the extractors are only invoked on an O directly, you can safely assume that the object passed in is basically a forwarding reference to O, so auto&& is fine (at least pending something like [P2481R0]). The extractors have the implicit precondition that the object is in the state specified (e.g. extract_value(o) should only be called if is_ok(o), with the converse for extract_error(o)) The factories can accept anything though, and should probably be constrained.

The choice of desugaring based specifically on the return type (rather than relying on each object to produce some kind of construction disambiguator like nullopt_t or unexpected<E>) is not only that we can be more performant, but also we can allow conversions between different kinds of error types, which is useful when joining various libraries together:

auto foo(int i) -> tl::expected<int, E>;
auto bar(int i) -> std::expected<int, E>;

auto strcat(int i) -> Result<string, E>
{
    // this works
    return std::format("{}{}", foo(i)??, bar(i)??);
}

As long as each of these various error types opts into try_traits so that they can properly be constructed from an error, this will work just fine.

3.2.1 Short-circuiting fold

One of the algorithms considered in the ranges::fold paper ([P2322R5]) was a short-circuiting fold. That paper ultimately didn’t propose such an algorithm, since there isn’t really a good way to generically write such a thing. Probably the best option in the paper was to have a mutating accumulation function that returns bool on failure?

But with this facility, there is a clear direction for how to write a generic, short-circuiting fold:

template <typename T>
concept Try = requires (T t) {
    typename try_traits<T>::value_type;
    typename try_traits<T>::error_type;

    { try_traits<T>::is_ok(t) } -> boolean-testable;
    // etc. ...
};

template <input_iterator I,
          sentinel_for<I> S,
          class T,
          invocable<T, iter_refrence_t<R>> F,
          Try Return = invoke_result_t<F&, T, iter_reference_t<R>>
    requires same_as<
        typename try_traits<Return>::value_type,
        T>
constexpr auto try_fold(I first, S last, T init, F accum) -> Ret
{
    for (; first != last; ++first) {
        init = std::invoke(accum,
            std::move(init),
            *first)??;
    }

    return try_traits<Ret>::from_value(std::move(init));
}

This try_fold can be used with an accumulation function that returns optional<T> or expected<T, E> or boost::outcome::result<T> or … Any type that opts into being a Try will work.

Note that this may not be exactly the way we’d specify this algorithm, since we probably want to return something like a pair<I, Ret> instead, so the body wouldn’t be able to use ?? and would have to go through try_traits manually for the error propogation. But that’s still okay, since the important part was being able to have a generic algorithm to begin with.

3.2.2 Range of expected to expected of Range

There’s an algorithm in Haskell called sequence which takes a t (m a) and yields a m (t a). In C++ terms, that might be an algorithm that takes a range of expected<T, E> and yields a expected<vector<T>, E> - which contains either all the results or the first error.

With the same Try concept from a above, this can be generalized to also work for optional<T> or any number of other Result-like types:

template <ranges::input_range R,
          Try T = remove_cvref_t<ranges::range_reference_t<R>>,
          typename Traits = try_traits<T>,
          typename Result = Traits::rebind<vector<typename Traits::value_type>>>
auto sequence(R&& r) -> Result
{
    vector<typename Traits::value_type> results;
    for (auto it = ranges::begin(r); it != ranges::end(r); ++it) {
        results.push_back((*it)??);
    }
    return Result::from_value(std::move(results));
}

3.3 Potential directions to go from here

This paper is proposing just ?? and the machinery necessary to make that work (including a concept, opt-ins for optional and expected, but not the short-circuiting fold algorithm).

However, it’s worth it for completeness to point out a few other directions that such an operator can take us.

3.3.1 Error continuations

Several languages have a facility that allows for continuing to invoke member functions on optional values. This facility is called something different in every language (optional chaining in Swift, null-conditional operator in C#, safe call operator in Kotlin), but somehow it’s all spelled the same and does the same thing anyway.

Given a std::optional<std::string> named opt, what that operator – spelled ?. – means is approximately:

expression
C++ equivalent
opt?.size() opt.transform(&std::string::size) // technically UB
opt?.substr(from, to) opt.transform([&](auto& s){ return s.substr(from, to); })

Like the null coalescing meaning of ?? described above, the semantics of opt?.f() can be achieved using library facilities today. The expression E1?.E2, if E1 is an optional, basically means E1.transform([&](auto&& e){ return FWD(e).E2; })

Quite unlike ??, there is a significant drop in readability and just the general nice-ness of the syntax.

The try_traits facility very nearly gives us the tools necessary to support such a continuation operator. Since what we need to do is:

We mostly need one more customization point: to put the types back together. What I mean is, consider:

auto f(int) -> std::expected<std::string, E>;

auto x = f(42)?.size();

The type of x needs to be std::expected<size_t, E>, since that’s what the value case ends up being here. If we call that customization point rebind, as in:

template <typename T, typename E>
struct try_traits<expected<T, E>> {
    // ... rest as before ...

    template <class U>
    using rebind = expected<remove_cvref_t<U>, E>;
};

Then the above can be desugared into:

using _Traits = try_traits<remove_cvref_t<decltype(f(42))>>;
using _R = _Traits::rebind<decltype(_Traits::extract_value(f(42)).size())>;

auto&& e = f(42);
auto x = _Traits::is_ok(e)
       ? try_traits<_R>::from_value(_Traits::extract_value(FWD(e)).size())
       : try_traits<_R>::from_error(_Traits::extract_error(FWD(e)));

That may seem like a mouthful. Because it is a mouthful. But it’s a mouthful that the user doesn’t have to write any part of, they just put f(42)?.size() and this does do the right thing.

At least, this mostly does the right thing. We still have to talk about copy elision. Consider this version:

struct X {
    auto f() -> std::mutex;
};

auto g() -> Result<X, E>;
auto n = g()?.f();

Presumably, n is a Result<std::mutex, E>, but in order for this to work, we can’t just evaluate this as something like Result<std::mutex, E>(g().value().f()). std::mutex isn’t movable.

The only way for this to work today is be able to pass a callable all the way through into this Result’s constructor. Which is to say, we desugar like so:

auto&& e = g();
auto n = _Traits::is_ok(e)
       ? try_traits<_R>::from_value_func([&]() -> decltype(auto) {
            return _Traits::extract_value(FWD(e)).f()
         })
       : try_traits<_R>::from_error(_Traits::extract_error(FWD(e)));

By default, try_traits<R>::from_value_func(f) would just be try_traits<R>::from_value(f()).

This is weird, but it’s something to think about.

Note also error continuation would only help in the member function case. If we want to continue into a non-member function, you’d need the sort of .transform() member function anyway.

3.3.2 Not propagating errors

The ?? approach seems to work quite well at propagating errors: it’s syntactically cheap, performant, and allows for integrating multiple libraries.

But what if we didn’t want to propagate the error, but rather do something else with it? For std::optional and std::expected, we already have a UB-if-error accessor in the form of *x and a throw-if-error accessor in the form of x.value(). It seems like the corollary to an error-propagating x?? would be some sort of x!! that somehow forces the error differently.

While propagating the error only really has one way to go (you return it), there are quite a few different things you can do differently:

That’s a lot of different options, and the right one likely depends on context too.

An additional template parameter on the error type could drive what x!! does (as Boost.Outcome does, for instance), which would allow you to preserve the nice syntax if a particular error handling strategy is sufficiently common (maybe you always throw, so why would you want to write extra syntax for this case), but at a cost of suddenly having way more types. Although the try_traits approach does at least allow those “way more types” to interact well.

This behavior can be achieved by adding a new function to try_traits which desugars as follows:

auto val = expr!!;
auto&& __val = expr;
using _Traits = std::try_traits<
  std::remove_cvref_t<decltype(__val)>>;
if (not _Traits::is_ok(__val)) {
  _Traits::fail(FWD(__val));
}
auto val = _Traits::extract_value(FWD(__val));

But this doesn’t seem as valuable as ?? or even ?. since this case is easy to add as a member function. Indeed, that’s what x.value() and *x do for optional and expected (throw and undefined behavior, respectively).

Moreover, any of the kinds of behavior you want can be written as a free function:

template <class T, Try U = std::remove_cvref_t<T>>
auto narrow_value(T&& t) -> decltype(auto) {
    assert(std::try_traits<U>::is_ok(t));
    return std::try_traits<U>::extract_value(FWD(t));
}

template <class T, Try U = std::remove_cvref_t<T>>
auto wide_value(T&& t) -> decltype(auto) {
    if (not std::try_traits<U>::is_ok(t)) {
        [[unlikely]] throw std::try_traits<U>::extact_error(FWD(t));
    }
    return std::try_traits<U>::extract_value(FWD(t));
}

// etc.

Which further demonstrates the utility of the proposed facility.

4 References

[P0323R12] Vicente Botet, JF Bastien, Jonathan Wakely. 2022-01-07. std::expected.
https://wg21.link/p0323r12

[P0709R4] Herb Sutter. 2019-08-04. Zero-overhead deterministic exceptions: Throwing values.
https://wg21.link/p0709r4

[P0779R0] Niall Douglas. 2017-10-15. Proposing operator try() (with added native C++ macro functions!).
https://wg21.link/p0779r0

[P2218R0] Marc Mutz. 2020-09-15. More flexible optional::value_or().
https://wg21.link/p2218r0

[P2322R5] Barry Revzin. 2021-10-18. ranges::fold.
https://wg21.link/p2322r5

[P2481R0] Barry Revzin. 2021-10-15. Forwarding reference to specific type/template.
https://wg21.link/p2481r0