Let’s talk about C++ constructors #

Lately, I had an interesting talk with colleagues of mine about C++. I told them I needed a safe std::unique_ptr abstraction. I needed such a type because I was handling scarce resources that needed to be uniquely owned in our code base, but at the same time, checking every time whether a value is present (nullptr being the problem here) yields large numbers of bugs in production.

They asked “What’s a safest std::unique_ptr?” My main problem was (still is) that std::unique_ptr can be initialized with nullptr. The problem with this is that you cannot assume that you always own a T if you are given a std::unique_ptr<T>. References and wrapped references cannot be used either because we really want to own the data here.

The rest of the discussion is driven by that initial motivation.

As a (mainly) Haskell and Rust developer, I’m also pretty good at writing C and C++ — after all, I’ve been doing C and C++ since I’m 11 and today, I’m 28; I’ll let you do the math. I especially developed 3D stuff, ranging from toy experiments to demoscene productions, and lots of other low-levels things I won’t detail here.

See, I’ve failed a lot while trying things when I was younger — especially around the age of 15 / 16. That’s basically the age where I started to be very frustrated at C++ and, without even knowing it, was asking to find a better way to design code. Yes, I failed a lot, I have no shame stating it. But to me, failing has been a building block to my experience. The more I fail, the less likely I’ll fail again the same way. By definition, if you fail a lot on different things… you now have very good hints and ideas about how not to fail again. This is like learn the hard way, but I truly think it’s important, especially to get why something doesn’t work. Reading books and being told is important, but failing is even more.

Yes, today, with all the hindsight I’ve been through, all the languages I’ve been using – ranging from C/C++, D, Python, Haskell, Idris, OCaml, Rust, Java, JavaScript, Perl, C#, GLSL and several others — I think I have a pretty good idea about what’s out there in terms of language ideas.

The goal of this article is to explain exactly why I don’t like that much C++, why I think it has a lot of bad ideas, and to provide people with material to think and meditate. I truly think that a good developer needs to call themselves into question regularly and question their process and established knowledge. I think that, decades ago, C and C++ were the right way to do things. Today, I think those languages and especially what has been accepted as ground and established truth about how to do things has become obsolete.

I’m going to explain why in a series of blog articles, starting with this one. Let’s go.

• C++ and the wrong idea of constructors
  • Hidden costs
  • The C++ broken initialization design
  • Construction ≠ constructors
• C++ and its bad standard library
• How would you do that in Rust?

Next article in the series.

C++ and the wrong idea of constructors#

I want to start with a very good example of something that has been accepted as pretty nice when it was introduced and that is completely obsolete to me. Let’s talk about C++’s constructors. But not constructors by themselves. I want to focus on how most people use them, especially around the concept of fallible construction — we’re going to talk about exceptions, too!

For those who are not familiar with constructors in C++, let me explain a bit. In C, when you have a data type (like a struct), you can create an object with such a type by allocating memory for it (either on the stack or the heap) and then initialize its fields. Since C99, you also have the designated initializer syntax to initialize a struct directly while you’re declaring it. Some examples:

struct Foo {
  int x;
  char const* y;
};

int main() {
  Foo foo;
  foo.x = 3;
  foo.y = "Hello, world!";

  // designated initializer syntax
  Foo foo2 = {
    .x = 3,
    .y = "Hello, world!"
  };
}

Even though that code seems pretty harmless, to me, it’s very easy to misuse C’s construction syntax. I will not explain exactly why here because we’re interested in C++, but you’ll understand what’s the problem by reading on. It relates to being possible to leave a state or part of a state undefined without having your compiler prevent you from doing that.

Hidden costs#

In C++, you cannot go with any of those way to create a value without a mandatory C++ concept: constructors. Constructors are C++’s way to initialize objects by using the same kind of syntax. C++ has several constructors:

• Default constructor. This one is invoked when you want to create an object without constructing it with any parameter. It’s an important constructor in C++, as several containers and standard algorithms rely on its presence.
• Parameterized constructors: you can create as many parameterized constructors as you want and they will take parameters to initialize your objects based on those parameters.
• Copy constructors: special constructors called when you want to copy objects. In C++, a copy constructor is automatically called when you pass an object foo which type is Foo to a function that expects a Foo, for instance.
• Move constructors: special constructors called when you move objects around. In C++, a move constructor is automatically called when you pass a rvalue reference to something (Foo &&). You can obtain such references by using std::move or std::forward.

There is a slightly exception: aggregate initialization, which is a bit similar to the C’s designated initialization, but requires that you have not declared any constructor – among other restrictions — and is then always available, either for public or private code. I will then not talk about it. I also have to admit I never completely recall all the rules about POD designated initialization.

Let’s demonstrate some of the constructors from above in the following program:

#include <iostream>

struct Foo {
  // default constuctor
  Foo() {
    std::cout << "default ctor called" << std::endl;
  }

  // copy constructor
  Foo(Foo const &) {
    std::cout << "copy ctor called" << std::endl;
  }

  // move constructor
  Foo(Foo &&) {
    std::cout << "move ctor called" << std::endl;
  }

  // assignment operator by copy
  Foo & operator=(Foo const &) {
    std::cout << "operator= called by copy" << std::endl;
    return *this;
  }

  // assignment operator by move
  Foo & operator=(Foo &&) {
    std::cout << "operator= called by move" << std::endl;
    return *this;
  }
};

Foo get_foo() {
  return Foo();
}

int main() {
  Foo foo;
  auto foo2 = foo;
  auto foo3 = get_foo();

  return 0;
}

In the main, what constructors do you think each line is going to call? :) The answer is… it depends on your compiler. Yes, you heard me right. With my compiler, this is what I get:

• Foo foo;: default constructor, which should be the case for every compiler here.
• auto foo2 = foo;: copy constructor.
• auto foo3 = get_foo();: default constructor. This is due to something called (N)RVO.

This is already pretty bad, for several reasons:

• It’s hard to tell when we are going to copy something. Think of a more complex data structure, such as a std::vector<Something>, which might contain thousands of millions of Something. Because C++ copies by default with the value semantics, you might end up with several copies while your code doesn’t explicitly copy anything.
• Because copy constructors and move constructors are subject to compiler optimizations, we don’t really know when the (N)RVO optimization is going to kick in — we can have good confidence but we might be wrong. I know some people never really know what auto foo = something; does — and I feel them, I often spend some time trying to figure out as well.

The existence of copy constructors and the value semantics by default makes C++ hard to reason about when we are making copies. Other languages, such as Rust, mitigate that risk by using a move semantics by default. If you want to copy (i.e. clone) something, you have to explicitly ask for it. For instance, consider:

// C++
void foo(std::vector<Foo> foos) {
  // …
}

int main() {
  std::vector<Foo> foos;

  // …

  foo(foos);

  return 0;

vs.

fn foo(foos: Vec<Foo>) {
  // …
}

fn main {
  let foos = Vec::new();

  // …

  foo(foos);
}

The C++ program will copy the content of foos when calling foo. The copy here is due to the copy constructor of std::vector being called. The effect of is that the whole content of the heap-allocated region owns by the vector will be copied. I typically call that a deep copy, or simply a clone.

The Rust program will not make a copy of foos. Instead, it will move ownership from the caller to the calle. foos won’t be available anymore in main after the call to foo. So no heap allocation will occur here. If we wanted the same behavior as the C++ one, we would do this:

  foo(foos.clone());

That requires that Foo implement the Clone trait, but it’s off topic here. However, there a few places where Rust can still do hidden copies that can hurt your performance. By default, everything is moved. However, if your type implements the Copy trait, it means that it can be safely copied. Such a copy is always performed bit-wise. What’s interesting is that there is, in theory, nothing different at runtime between copying and moving in Rust. Moving might allow more optimizations to prevent actually copying data, but imagine you don’t have optimizations. Both copy and move semantics, in Rust, could be implemented with a memcpy. The main difference is that copying doesn’t lose ownership. So, because the original value, is not moved, can still be used after copy. Rust has implementations of Copy for arrays, which, thus, can yield bad performances if passed directly to a function, for instance:

fn bar(_: [u32; 1000000]) {}

Calling bar will make a copy of the array you pass as argument, which is bad news here. However, I think it’s pretty unlikely that you allocate such an array on your stack. Most non-primitive types from the standard library don’t have an implementation for Copy, which makes move follow the move semantics.

Let’s get back to C++, shall we. We’ve established that C++ makes it very easy to write code which has hidden copies, because of the existence of copy constructors and that value semantics depends on it. Let’s go on with a problem I have with constructors that is as bad to me.

The C++ broken initialization design#

Imagine that you want to be able to manipulate a kind of integer. That integer cannot be negative and it cannot be equal to 0. Also, we want it on 32-bit precision. There are several ways to do that. Let’s enumerate a few of them (non-exhaustive list):

1. Just use the int32_t type from the standard library and every time we try to create or modify it, forbid to put values which are less or equal to 0.
2. Use the uint32_t type from the standard library and prevent using 0. We don’t need to check for negative values because uint32_t cannot encode them.
3. Create a wrapper type that prevents from constructing values that violate the invariant (the invariant being that it cannot be less than 1.

Obviously, both (1.) and (2.) have a big drawback: you’re going to spend a lot of runtime to check that the invariant is not violated. Worse, the invariant leaks in the API. They have an API contract flaw: they don’t say to the programmers they cannot be equal to 0. For instance:

void foo(uint32_t non_zero) {
  // …
}

Even though the name of the variable is non_zero, what prevents a user from calling foo with 0? Nothing. They can just call foo(0) and the code will still compile. They will then get a really bad runtime undefined behavior, or maybe nothing for several runs and a sudden bug… or you will have to handle the invariant at every call, which doesn’t seem like a lot of fun — and it’s dangerous because you can forget to check it.

Instead, we want this:

void foo(NonZero non_zero) {
}

Here, the function will clearly not be callable with a uint32_t. So the provided value must be statically verified to be non-zero. The NonZero wrapper type must then ensure the invariant is non-violated. I now have a question: do you think NonZero has to check the invariant every time it’s used, or we can do better?

C++ states that NonZero should be constructed with a constructor, and my point is that constructors are broken, because not all values can be constructed via C++ constructors. But they are not broken the way you think. In theory, and actually, there is a way to use C++ constructors in sound ways. Try to spend a few minutes and think about it. Can you implement a struct NonZero with an API exposing only constructors to create and initialize non-zero values? There are two possibilities:

• Either you, indeed, use a constructor, but because you want to be able to prevent people from building non-zero values, you will have to fail in your constructor. And failure in constructors is possible by throwing an exception. I don’t accept that answer because exceptions should only be used in exceptional failure situations, and this is not one of them. I’ll explain why.
• By still allowing construction of the object but putting it in an invalid state. I consider this a non-answer as well, as you will have to keep track of the invariant every time you want to use the NonZero, and because it’s very easy to build a NonZero with a violated invariant with your code compiled correctly: auto foo = NonZero(0);. So, it’s a no.

I want to be very very crystal clear here. Exceptions in C++ are heavy. They require unwinding most of the time when some code throws them. They are needed to handle exceptional errors, such as an out of memory error, or a dead thread. People have been abusing exceptions for too long. Really. Stop using exceptions. Calling NonZero(x) with x being 0 is an error that originate from several places, because NonZero is a pretty abstract concept. You shouldn’t assume it’s exceptional that someone tries to build one with 0. It’s part of its API: it must fail to construct if built with 0. The API must convey the fact it can handle that kind of error. So it’s an error that is covered by the scope of the type and its carried invariant.

An example of exceptional error here would be that, when you construct your NonZero, you’re out of space on the stack. Throwing an error here seems okay to me. The problem, here, is not that we have a constuctor that fails. The problem is that we used a constructor to encode fallible construction. That’s the problem. My point here is that constructors must not be used for fallible constructions.

So… what are we left with? If you really want to only use constructors without exceptions, you have no other choices but to accept to have a dangling invariant. The invariant here is that the wrapped uint32_t must not be 0.

A small note on invariants: they must be held before and after you call a public function on your type. We can violate them inside our private code for performance and optimization purposes, but don’t forget that you must rebalance everything so that the invariant is held when you give control back to the user. This is especially hard to do in C++ because of exception safety, so a rule of thumb: try to never violate invariants, even in private code.

struct NonZero {
  // prevent people from creating default non-zero values; it makes no sense
  NonZero() = delete;

  NonZero(uint32_t value): _wrapped(value) {
    // that would break your application at runtime on debug and do nothing on release… :(
    assert(_wrapped != 0);
  }

  // let C++ automatically implement some stuff for us, which won’t break the invariant
  NonZero(NonZero const &) = default;
  NonZero(NonZero &&) = default;

  NonZero & operator=(NonZero const &) = default;
  NonZero & operator=(NonZero &&) = default;

private:
  // the invariant must be held on this value
  uint32_t _wrapped;
};

As you can see, having a NonZero with this definition doesn’t mean the value cannot be 0. Example:

int main() {
  auto nz = NonZero(0); // meh
  return 0;
}

This code compiles fine and will either abort at runtime on an assert or just silently do something very wrong on release code.

So what can we do to this situation?

Construction ≠ constructors#

The thing is that, if you’ve only done C++ in your life, you’re likely to think that the best thing to do is to give up and use exceptions. After all, they’re pretty nice.

#include <stdexcept> // you’ll need that for the exception type

struct NonZero {
  // prevent people from creating default non-zero values; it makes no sense
  NonZero() = delete;

  NonZero(uint32_t value): _wrapped(value) {
    if (_wrapped == 0) {
      throw std::invalid_argument("cannot create a NonZery with 0");
    }
  }

  // let C++ automatically implement some stuff for us, which won’t break the invariant
  NonZero(NonZero const &) = default;
  NonZero(NonZero &&) = default;

  NonZero & operator=(NonZero const &) = default;
  NonZero & operator=(NonZero &&) = default;

private:
  // the invariant must be held on this value
  uint32_t _wrapped;
};

Now, when you call NonZero(0), an exception is thrown. You can catch it with a try / catch block. But please wait a minute. Several points:

• Again, exceptions should not be used for such errors. I get it you’ve given up on that rule, but you shouldn’t. Save exceptions for exceptional cases. This is not exceptional. This is naaaaaaaaaaht.
• When you look at the constructor definition, nothing tells you it can fail. You have to read the body of the constructor to check for throw — and again, it could quickly get hard because you might call functions that throw too…

That last point is what makes C++ a very hard language what it comes to error handling to me. Error handling, most of the time, is based on exceptions, which are invisible at the type-level. I got some remarks stating that, nowadays, modern C++ code bases use the noexcept keyword when something cannot fail, and the rest of the time, we must assume something can fail. Okay, but then, fail how? What kind of error? How do I know that? Rely on whether it’s written in a Doxygen documentation, that, most of the time, doesn’t even exist? Also, just have a look at how people handle failures in C++. Do you see those try / catch everywhere noexcept is not annotated? No, because people don’t use noexcept and when they use library, they don’t even apply that advice to themselves either. The problem with exceptions not being visible in types and function signatures is that you have to read the code (or the missing Doxygen documentation :)) to find out what exceptions you need to catch. The other problem is that they can be ignored and implicitely and automatically propagated upwards in the call stack. There is no way to force a user to either handle the error or explicitly pass it to their caller.

All that to say: don’t do this. Don’t think constructors are a good idea to construct types in all possible situations. Failible constructions shouldn’t be done with public constructors and exceptions. However, it doesn’t mean constructors cannot be used to implement fallible constructions. That seems tricky, and you’re right. You might wonder:

“Okay okay, you’ve advanced points but… still, how would you do it?”

My implementation — and today, I would only accept that one — to this problem is this: since we cannot create a value without using a constructor, and since constructors fail to encode fallible constructions without using exceptions — which I don’t accept either, I’m going to use a private constructor to build my NonZero and not check the invariant. Then, using a static method, I can return a more typed object to encode a possible failure, by checking the invariant before creating the object.

Consider:

#include <optional>

struct NonZero {
  // prevent people from creating default non-zero values; it makes no sense
  NonZero() = delete;

  // let C++ automatically implement some stuff for us, which won’t break the invariant
  NonZero(NonZero const &) = default;
  NonZero(NonZero &&) = default;

  NonZero & operator=(NonZero const &) = default;
  NonZero & operator=(NonZero &&) = default;

  // static method used to create a NonZero; the only way to create one
  // using the public interface
  std::optional<NonZero> from_u32(uint32_t value) {
    std::optional<NonZero> r;

    if (value != 0) {
      // call to the private ctor; the invariant is already checked in this branch
      r = NonZero(value);
    }

    return r;
  }

private:
  // private constructor that doesn’t check the invariant
  NonZero(uint32_t value): _wrapped(value) {}

  // the invariant must be held on this value
  uint32_t _wrapped;
};

So let me explain a bit all this code. First, the NonZero(uint32_t) constructor is private. It has to be private because it doesn’t enforce the invariant. It allows to construct any NonZero. We need it because C++ requires you to use a constructor to create a value of type NonZero. Because I don’t want exceptions, I just don’t check the invariant here so that constructor cannot fail.

The static method called from_u32 is the only entry-point to create a value of type NonZero. As you can see, it returns a std::optional<NonZero>, which means that it can fail. If you call that function with 0, you will get an empty optional value. That function is implemented by creating an optional value, allocated on the stack with the default constructor — which makes it empty. Then, we check the invariant and if it’s not violated, we allocate and initialize the NonZero, and return the whole thing.

C++ and its bad standard library#

For the rest of the article, consider we add the following method to NonZero:

  uint32_t value() const {
    return _wrapped;
  }

If you’ve followed carefully what we’ve been doing here, a call to NonZero::value() will never return 0, because construction statically disallows building such non-zero values. You then don’t have to check for it!

However, we’re not completely done. We’ve used the std::optional standard type. How are we supposed to use that type? Looking at the official documentation, we get use the operator* or operator-> to access the underlying object. We have the has_value() method to check whether a value is present.

auto nz = NonZero::from_u32(0);

if (nz.has_value()) {
  std::cout << "We have a value! " << nz->value() << std::endl;
}

That seems nice, right? Well, not really. C++ doesn’t have exhaustive pattern matching, which makes it impossible to statically ensure you exhaustively check a std::optional. What it means is that, given the current contract based on has_value() and the dereference operators, we can still break our program and bring undefined behaviors without having our compiler complain.

auto nz = NonZero::from_u32(0);

std::cout << "We have a value… right? " << nz->value() << std::endl;

That program is perfectly fine and valid. However it’s not. It’s completely wrong, because std::optional’s API allows developers to access a value that might not be there.

This is the same situation as doing this:

int * ptr = nullptr;
int x = *ptr;

Statically fine. Dynamically pretty bad, right? C++’s standard library’s std::optional could have been done in a much safer way, by, for instance, providing you with some combinators to work with the underlying type. C++ doesn’t have exhaustive pattern matching but it still has higher-order functions. We could then have something like:

auto nz = NonZero::from_u32(0);

nz.maybe(
  // called if no value is present
  []() {
    std::cout << "No value there." << std::endl;
  },
  // called if a value is present
  [](NonZero const &v) {
    std::cout << "We have a value: " << v.value() << std::endl;
  }
);

For most C++ developers, that code will look like very verbose and wrong, and I agree. Without pattern matching, it’s very hard to safely and statically use a std::optional.

How would you do that in Rust?#

In Rust, we have exhaustive static dispatch. std::optional is called Option and Rust doesn’t have constructors at all.

Values can be constructed by providing all fields, if they are all available (i.e. that’s always the case in the module the type is defined in; for the rest, you need access to the fields by marking them all pub).

This is my NonZero wrapper in Rust:

struct NonZero {
  // we don’t make that value pub so no one can directly construct a NonZero
  // outside of our module
  wrapped: u32,
}

impl NonZero {
  // yes, new is nothing special in Rust, so we can use it as a method!
  // Self references the current type we’re adding an impl for (here, NonZero)
  pub new(wrapped: u32) -> Option<Self> {
    if wrapped == 0 {
      None
    } else {
      Some(NonZery { wrapped })
    }
  }
}

fn main() {
  match NonZero::new(0) {
    Some(value) => println!("we have a value! {}", value),
    None => println!("we don’t have a value… :("),
  }
}

That’s all. The program is statically safe, as the compiler can make sure we’re correctly using the Option<NonZero>. We could use combinators that do the pattern matching for us. For instance, imagine that you want the value 8080 if it’s not correctly constructed, or what it contains otherwise:

let port = nz.unwrap_or(8080);

Here, the type of port is NonZero. Option::unwrap_or might be implemented like this:

impl<T> Option<T> {
  pub fn unwrap_or(self, default_value: T) -> T {
    match self {
      Some(v) => v,
      None => default_value,
    }
  }
}

There are a lot of similar combinators, like Option::map, Option::or, Option::unwrap_or_else, etc. I’ll let you dig by yourselves. However, there is an important point to make here. In C++, there is the std::optional::value() method that returns the value if it’s present or just throws an error (which I dislike). In Rust, there is the std::Option::unwrap() method, that returns the value if it’s present or panics otherwise… which I also dislike. unwrap() can be dangerous in Rust and I think it should be marked unsafe. But a more detailed article is needed about why I want it to be marked unsafe and it’s off topic anyways.

So, we’re hitting the end of the article. It might be a lot of information; sorry. To sum up, in C++:

• Construction must not be assumed to be done exclusively via constructors.
• Pure construction (as in infallible) can use constructors if you want.
• Fallible construction can use:
  • A static function that returns a typed object via, for instance, std::optional or similar, and enforce invariants / pre-conditions.
  • A private constructor to build the result value after that invariant and pre-conditions get checked.
• We see that we can make C++ much easier to reason about by following those rules. It will not be as handy as in Rust where no constructor are needed, but it’s still a good improvement to me.

Next time, I’ll bring you on a tour with me around the concept of inclusion polymorphism, which is most of the time called inheritance in the world of OOP languages. But we’ll still stick to C++. ;)

Keep the vibes!

Discussion on the article here.

Keep on reading.