C++20’s parenthesized aggregate initialization has some downsides

Somehow the topic of P0960 parenthesized aggregate initialization has come up three times in the past week over on the cpplang Slack. The good news is that usually the asker is curious why some reasonable-looking C++20 code fails to compile in C++17 — indicating that C++20’s new rules are arguably more intuitive than C++17’s.

But let’s start at the beginning.

The basics of brace-initialization versus parens-initialization

Regular readers of this blog may remember my simple initialization guidelines from “The Knightmare of Initialization in C++” (2019-02-18):

  • Use = whenever you can.

  • Use initializer-list syntax {} only for element initializers (of containers and aggregates).

  • Use function-call syntax () to call a constructor, viewed as an object-factory.

And one more rule:

So for example we can write

int a[] = {1, 2};
std::array<int, 2> b = {1, 2};
std::pair<int, int> p = {1, 2};

struct Coord { int x, y; };
struct BadGrid { BadGrid(int width, int height); };
struct GoodGrid { explicit GoodGrid(int width, int height); };

Coord c = {1, 2};
auto g1 = BadGrid(10, 20);
auto g2 = GoodGrid(10, 20);

Notice that even when the type author of BadGrid has broken our rule that all constructors should be explicit, it’s still possible for us as the client-code author to ignore that “mistake” and use function-call syntax () anyway. It would have been physically possible for us to write

BadGrid g1 = {10, 20};  // a grid with dimensions 10x20

but that would have been a solecism according to the guidelines I just laid out: The resulting BadGrid is not conceptually a sequence of two elements {10, 20} in the same way that an array, pair, or Coord is conceptually a sequence of two elements. Therefore, even though we can implicitly convert the braced initializer list {10, 20} to a BadGrid, we should not.

This guideline is particularly relevant when dealing with STL containers like vector, for three reasons:

  • Every STL container does represent a sequence of elements.

  • Every STL container has an absolutely massive constructor overload set.

  • For historical reasons (with which I disagree), STL types deliberately make almost all of their constructors non-explicit.

Consider the following four initializations of vector:

std::vector<int> v1 = {10, 20};
std::vector<std::string> v2 = {"x", "y"};
auto v3 = std::vector<int>(10, 20);
auto v4 = std::vector<std::string>(10, "y");

These are all appropriate initializations according to my guidelines. The first two depict initializing a vector with a sequence of elements: {10, 20} or {"x", "y"}. In both cases, the resulting vector ends up with 2 elements. The second two initialize a vector using one of its many “object-factory” constructors: specifically, the one that takes a size and a fill value. v3 is initialized with 10 copies of 20; v4 is initialized with 10 copies of "y". In case v4, we physically could have written

std::vector<std::string> v4 = {10, "y"};  // Faux pas!

but according to my guidelines we should not write that, for the simple reason that we’re not intending to create a vector with the sequence of elements {10, "y"}. This situation is exactly analogous to the BadGrid g1 case above.

Aggregate versus non-aggregate initialization

The next thing you should know about initialization in C++ is that C++ treats aggregates differently from non-aggregates. In C++98, this distinction was extremely visible, because curly braces could be used only to initialize aggregates, by which I mean plain old C-style structs and arrays:

Coord ca = Coord(1, 2);  // syntax error in C++98
Coord cb = {1, 2};       // OK

BadGrid ga = BadGrid(1, 2);  // OK
BadGrid gb = {1, 2};         // syntax error in C++98

std::pair<int, int> pa(1, 2);     // OK
std::pair<int, int> pb = {1, 2};  // syntax error in C++98

In C++11 we got uniform initialization, which let us use {} to call constructors, like this:

Coord ca = Coord(1, 2);  // still a syntax error in C++11
Coord cb = {1, 2};       // OK

BadGrid ga = BadGrid(1, 2);  // OK
BadGrid gb = {1, 2};         // OK since C++11 (but solecism)

std::pair<int, int> pa(1, 2);     // OK (but solecism)
std::pair<int, int> pb = {1, 2};  // OK since C++11 (in fact, preferred)

So in C++11 we have two different meanings for {}-initialization: Sometimes it means we’re calling a constructor with a certain set of arguments (or maybe with a std::initializer_list), and sometimes it means we’re initializing the members of an aggregate. The second case, aggregate initialization, has several special powers:

First, aggregate initialization initializes each member directly. When you call a constructor, the arguments are matched up to the types of the constructor parameters; when you do aggregate initialization, the initializers are matched up directly to the types of the object’s data members. This matters mainly for immovable types like lock_guard:

struct Agg {
    std::lock_guard<std::mutex> lk_;
struct NonAgg {
    std::lock_guard<std::mutex> lk_;
    NonAgg(std::lock_guard<std::mutex> lk) : lk_(lk) {}

std::mutex m;
Agg a = { std::lock_guard(m) };      // OK
NonAgg na = { std::lock_guard(m) };  // oops, error

In the snippet above, the prvalue std::lock_guard(m) directly initializes a.lk_; but on the next line, the prvalue std::lock_guard(m) initializes only the constructor parameter lk — there’s no way for the author of that constructor to get lk’s value into the data member na.lk_, because lock_guard is immovable.

Even for movable types, the direct-initialization of aggregates can be a performance benefit. Recall from “The surprisingly high cost of static-lifetime constructors” (2018-06-26) that std::initializer_list is a view onto an immutable array. So:

std::string a[] = {"a", "b", "c"};

direct-initializes three std::string objects in an array a, whereas

std::vector<std::string> v = {"a", "b", "c"};

direct-initializes three std::string objects in an anonymous array, creates an initializer_list<string> referring to that array, and then calls vector’s constructor, which makes copies of those std::strings. Or again,

std::array<std::string, 3> b = {"a"s, "b"s, "c"s};

direct-initializes three std::string objects into the elements of b (which the Standard guarantees is an aggregate type), whereas

using S = std::string;
std::tuple<S, S, S> t = {"a"s, "b"s, "c"s};

direct-initializes three temporary std::string objects on the stack, and then calls tuple’s constructor with three std::string&&s. That constructor move-constructs into the elements of t and finally destroys the original temporaries.

Second, for backward-compatibility with C, aggregate initialization will value-initialize any trailing data members.

struct sockaddr_in {
    short          sin_family;
    unsigned short sin_port;
    struct in_addr sin_addr;
    char           sin_zero[8];

sockaddr_in s = {AF_INET};
assert(s.sin_port == 0);
assert(s.sin_addr.s_addr == 0);
assert(s.sin_zero[7] == 0);

This is useful for C compatibility, but it does lead to some surprising results: whether it’s “OK” to omit the initializers of trailing elements depends on whether the type being initialized is an aggregate or not, even though we’re uniformly using braced-initializer-list initialization syntax in both cases.

Coord c = {42};    // OK: Coord is an aggregate
BadGrid g = {42};  // ill-formed: BadGrid is not an aggregate

std::array<int,3> b = {1, 2};       // OK: array is an aggregate
std::tuple<int,int,int> t = {1, 2}; // ill-formed: tuple is not an aggregate

The coincidence of zero-argument initialization

Because C++11 introduced uniform initialization, T{} is permitted to call T’s default constructor if it has one; or, if T is an aggregate, then T{} will do aggregate initialization where every member is value-initialized.

using B = std::array<int,3>;
using T = std::tuple<int,int,int>;

B b = B{};  // value-initialize, i.e., {0,0,0}
T t = T{};  // call the zero-argument constructor, i.e., {0,0,0}

Also, ever since C++98, there’s been wording to make T() do value-initialization as a special case. (Sure, it looks like we’re calling T’s zero-argument constructor — and you can totally think of it that way in practice — but technically this is a special-case syntax for value-initializing T, which in turn calls T’s default constructor if it has one, but otherwise recursively value-initializes T’s members.)

B b = B();  // value-initialize, i.e., {0,0,0}
T t = T();  // value-initialize, i.e. call the default constructor, i.e., {0,0,0}

emplace_back presents a stumbling block

C++11 also introduced the idea of perfect-forwarding arguments through APIs like emplace_back. The idea of emplace_back is that you pass in some arguments Args&&... args, and then the vector will construct its new element using a placement-new expression like ::new (p) T(std::forward<Args>(args)...). Notice the parentheses there — not braces! This is important because we want to ensure that we get consistent behavior when emplacing an STL container. Compare the following snippet with our first section’s v3/v4 example:

std::vector<std::vector<int>> vvi;
vvi.emplace_back(10, 20);  // emplace vector<int>(10, 20)

std::vector<std::vector<std::string>> vvs;
vvs.emplace_back(10, "y"); // emplace vector<string>(10, "y")

But now, consider this C++17 code:

using T = std::tuple<int,int,int>; // non-aggregate
using B = std::array<int,3>;       // aggregate

std::vector<T> vt; // vector of non-aggregates
std::vector<B> vb; // vector of aggregates

vt.emplace_back(); // OK, emplaces T()
vb.emplace_back(); // 1: OK, emplaces B()

vt.emplace_back(1,2,3); // OK, emplaces T(1,2,3)
vb.emplace_back(1,2,3); // 2: Error in C++17!

Line 1 emplaces an array object constructed with B(), which, as we saw in the preceding section, means value-initialization: the same as {0,0,0}.

Line 2 attempts to emplace an array constructed with B(1,2,3) — and this fails, because B has no constructor taking three ints!

Whether a given emplace_back is legal depends on whether the type being initialized is an aggregate or not, even though we’re using the same emplace_back syntax in both cases.

C++20’s solution: Parens-init for aggregates

C++20 addressed the above emplace_back quirk by saying: well, if the problem is that B(1,2,3) isn’t legal syntax, let’s just make it legal! C++20 adopted P0960 “Allow initializing aggregates from a parenthesized list of values,” which extends the rules of initialization to cover the case where “the destination type is a (possibly cv-qualified) aggregate class A and the initializer is a parenthesized expression-list.” In that case, initialization proceeds just as in the curly-brace case, omitting a few minor quirks that are triggered (since C++11) by the curly-brace syntax specifically:

  • Curly-braced initializers are evaluated strictly left-to-right; parenthesized initializers can be evaluated in any order.

  • Curly-braced initializers forbid narrowing conversions (such as double-to-int); parenthesized initializers do not.

  • Curly-braced prvalues bound to reference data members can be lifetime-extended; parenthesized prvalues are never lifetime-extended.

  • Curly-braced initialization lists can involve brace elision (e.g. depicting {1, {2, 3}, 4} as {1, 2, 3, 4}); parenthesized initialization lists cannot.

  • Curly-braced initialization lists can include C++20 designated initializers; parenthesized initialization lists cannot.

The result is that the following is legal C++20 (but not legal C++17):

struct Coord { int x, y; }; // aggregate
std::vector<Coord> vc;
vc.emplace_back();        // OK since C++11, emplaces Coord() i.e. {0, 0}
vc.emplace_back(10, 20);  // OK since C++20, emplaces Coord(10, 20) i.e. {10, 20}

The unintended consequences

This being C++, naturally there are some rough edges and pitfalls. I count at least three.

First, remember that even though vc.emplace_back(10,20) emplaces the equivalent of {10,20}, it doesn’t actually use curly braces! emplace_back still uses the round-parens syntax Coord(10,20), and simply relies on aggregate parens-init to convert that into the equivalent of {10,20}. For non-aggregate types, where T(x,y) and T{x,y} do different things, you’re still going to get the T(x,y) behavior!

std::vector<std::vector<int>> vvi;
vvi.emplace_back(10, 20);  // OK since C++11, emplaces vector<int>(10, 20)
assert(vvi[0].size() == 10);  // not 2

Second, look back at that list of curly-brace quirks ignored by aggregate parens-init. Notice one quirk that’s not on that list: both kinds of aggregate initialization are allowed to omit trailing initializers!

Coord c1 = {10};       // OK since C++98, equivalent to {10, 0}
Coord c2 = Coord(10);  // OK since C++20, equivalent to {10, 0}
vc.emplace_back(10);   // OK since C++20, emplaces Coord(10) i.e. {10, 0}

In fact, since Coord(10) is legal in C++20, we can even write

auto c3 = static_cast<Coord>(10);  // OK since C++20, equivalent to {10, 0}

which strikes me as probably terrible.

Vice versa, even though C++20 made it legal to aggregate-initialize array types with parentheses in general, some of the syntactic space was deliberately reserved for future standardization. For example, parens-init works for array variables but not array prvalues:

using A = int[3];
A a1(1,2,3);         // OK since C++20
auto a2 = A(1,2,3);  // still an error!

Third, probably the most annoying pitfall (although it does satisfyingly confirm my prejudice against std::array): parens-init is basically useless for std::array!

using T = std::tuple<int, int, int>; // non-aggregate
using B = std::array<int, 3>; // aggregate
std::vector<T> vt;
std::vector<B> vb;
vt.push_back({1,2,3});  // OK since C++11
vb.push_back({1,2,3});  // OK since C++11

vt.emplace_back(1,2,3); // OK since C++11
vb.emplace_back(1,2,3); // still an error!

The reason is that B(1,2,3) remains ill-formed even in C++20; and the reason for that is the fourth quirk in the list above: parenthesized initializer lists don’t do brace elision the way braced initializer lists do! std::array<T,N> is guaranteed to be an aggregate initializable with N elements of type T, but nothing about the data’s underlying “shape” is guaranteed: does it have N data members of type T? or one data member of type T[N]? or N/2 data members of type T[2]? Brace elision allows us not to care about such details when we use curly-braced aggregate initialization, but not when we use parenthesized aggregate initialization.

B b1(1,2,3);              // still an error! (technically implementation-defined)
B b2{{1,2,3}};            // OK since C++11; direct-initializes (technically implementation-defined)
B b3({1,2,3});            // OK since C++11; move-constructs
vb.emplace_back({1,2,3}); // still an error!

The above b3 is already legal in C++11; it represents direct-initialization of a B from {1,2,3} (which overload resolution will decide to treat as B{1,2,3}), and C++20 doesn’t change this. And vb.emplace_back({1,2,3}) remains ill-formed, because emplace_back’s perfect forwarding wants its Args&&... to have nicely deducible types, and a brace-enclosed initializer list like {1,2,3} has no type.

The bottom line

C++20’s parenthesized aggregate initialization solves the tipmost link in a rather long chain of unintended consequences:

  • Aggregate types are quietly treated differently from non-aggregates; for example, they get more-direct initialization.

  • emplace_back uses parens-initialization rather than brace-initialization, and we can’t change that without breaking vvi.emplace_back(10, 20).

  • Therefore vc.emplace_back(10, 20) wants to parens-initialize Coord(10, 20); in C++17 that’s ill-formed unless we add a constructor Coord(int, int).

  • But giving a type a constructor makes it a non-aggregate, which means it loses the performance benefits of aggregate initialization. (Not that Coord cares about those benefits.)

  • C++20 makes vc.emplace_back(10, 20) legal, by permitting Coord(10, 20) to do aggregate initialization just like Coord{10, 20}.

But as a side effect, C++20 makes static_cast<Coord>(10) legal (surprise!); and even after all this rigmarole, C++20 fails to make either vb.emplace_back(1,2,3) or vb.emplace_back({1,2,3}) work as expected. That’s bad.

But I’ll leave you by reiterating the good news: When I’ve seen people bring up this topic on Slack, they’re usually not at all surprised that both of these lines

std::vector<Coord> vc;
vc.push_back({10, 20});  // A, OK since C++11
vc.emplace_back(10, 20); // B, OK since C++20

both work in C++20. Generally, they’re surprised only that line B failed to work before C++20! So, in at least that narrow sense, C++20’s parenthesized aggregate initialization has improved the story for newcomers to C++.

Posted 2022-06-03