p0538.md

return with no argument

Pull request

Table of contents

• Problem
  • Use cases
  • What's wrong with the C++ rule?
• Background
• Proposal
• Details
  • Relation to upcoming proposals
• Rationale based on Carbon's goals
• Alternatives considered
  • Retain the C++ rule
  • Fully divorce functions and procedures

Problem

Use cases

We wish to support the following use cases:

• Functions that return values. This should be supported for any type for which a value can be formed, and such support should be uniform and independent of whether the return type is (). A return value must always be produced.
• Functions that do not return a value ("procedures"). In such functions, we do not need or want to be able to return a value. No return value is needed, so if control flow reaches the end of a procedure, it should return to its caller.
• Functions with parameterized return types, that may be either of the above, depending on the parameterization. For example, a call wrapper might be parameterized by the type of its callee, and might return the same type that its callee returns, or a type computed based on that type.

What's wrong with the C++ rule?

C++ treats void as a special case in a number of ways. We want to minimize the impact of this special-case treatment for the corresponding Carbon types. One way that this special treatment is visible in C++ is that functions with the return type void obey different rules: the operand of return becomes optional (but is still permitted), and reaching the end of the function becomes equivalent to return;. In a function template, this can even happen invisibly, with some instantiations having an implicit return; and others not.

This interferes with the ability to reason about programs. Consider:

template<typename T> auto f() {
  if (T::cond())
    return T::run();
}

Here, it is possible for control flow to reach the end of the function. However, a compiler can't warn on this without false positives, because it's possible that T::run() has type void, in which case this function has an implicit return; added before its }, and indeed, it might be the case that T::run() always has return type void for any T where T::cond() returns false.

Background

See the following issues:

• Proposal: function declaration syntax
• Proposal: return statements
• Leads question: what is the relationship between Void and ()?
• Leads question: should we allow return; in functions with a Void return type?

Proposal

Instead of applying special-case rules based on whether the return type of a function is (), we apply special-case rules based on whether a return type is provided.

Details

A function with no declared return type is a procedure. The return type of a procedure is implicitly (), and a procedure always returns the value () if and when it returns. Inside a procedure, return must have no argument, and if control flow reaches the end of a procedure, the behavior is as if return; is executed.

// F is a procedure.
fn F() {
  if (cond) {
    return;
  }
  if (cond2) {
    // Error: cannot return a value from a procedure.
    return F();
  }

  // Implicitly `return;` here.
}

A function with a declared return type is treated uniformly regardless of whether that return type happens to be (). Every return statement must return a value. There is no implicit return at the end of the function, and instead a compile error is produced if control flow can reach the end of the function, even if the return type is () -- or any other unit type.

fn G() -> () {
  if (cond) {
    // OK, F() returns ().
    return F();
  }
  if (cond2) {
    // Error: return without value in value-returning function.
    return;
  }

  // Error: control flow can reach end of value-returning function.
}

From the caller's perspective, there is no difference between a function declared as

fn DoTheThing() -> ();

and a function declared as

fn DoTheThing();

As a result, the choice to include or omit the -> () in the definition is an implementation detail, and the syntax used in a forward declaration is not required to match that used in a definition. The use of -> () in idiomatic Carbon code is expected to be rare, but it is permitted for uniformity, and in case there is a reason to desire the value-returning-function rules or to emphasize to the reader that () is the return type or similar.

Relation to upcoming proposals

Issue #510 asks whether we should support named return variables:

fn F() -> var ReturnType: x {
  // initialize x
  return;
}

If we do, functions using that syntax should follow the rules for procedures in this proposal, including the implicit return; if control reaches the end of the function. In particular,

fn F() { ... }

would be exactly equivalent to

fn F() -> var (): _ = () { ... }

Rationale based on Carbon's goals

• Software and language evolution
  • This proposal may decrease the number of places in which the return type of a procedure is used, by discouraging the use of return Procedure();. This in turn may make it easier to change a return type from () to something else, but this proposal by itself is insufficient to ensure that is always possible.
• Code that is easy to read, understand, and write
  • The conceptual integrity of the Carbon language is improved by making the same syntax result in the same semantics, regardless of whether a type happens to be () or not, and symmetrically by using different syntax for different semantics.
  • The readability of Carbon code is improved and a source of surprise is eliminated by removing the possibility of return F(); being mixed with return; in the same function.
• Practical safety guarantees and testing mechanisms
  • By making the presence or absence of an implicit return in a function be determined based on syntax alone, we permit checks for missing return statements to be provided in the definition of a template or generic, without needing to know the arguments. This is important for generics in particular, because we do not want monomorphization to be able to fail and because we do not in general guarantee that monomorphization will be performed.
• Fast and scalable development
  • The correct syntax for a return statement can be detected while parsing, using only syntactic information rather than contextual, semantic information. In practice, we will likely parse both kinds of return statement in all functions and check the return type from a context that has the semantic information, but the ability to do these checks syntactically may be useful for simple tools and editor integration.
• Interoperability with and migration from existing C++ code
  • This proposal rejects some constructs that would be valid in C++:

    return F();
    

    in a function with void return type would no longer be valid in a corresponding Carbon function with no specified return type, and would need to be translated into

    F();
    return;
    

    (possibly with braces added). However, the fact that this construct is valid in C++ is surprising to many, and the constructs that would be idiomatic in C++ are still valid under these rules.

Alternatives considered

Retain the C++ rule

The advantages of this approach compared to maintaining the C++ rule are discussed above. The advantage of maintaining the C++ rule would be that Carbon is more closely aligned with C++. However, the removed functionality -- specifically, the ability to return an expression of type void from a void returning function -- is still available, albeit with a more verbose syntax, and the existence of that functionality in C++ is a source of surprise to C++ programmers.

Fully divorce functions and procedures

We could treat the choice of function with () return type versus procedure as being part of the interface rather than being an implementation detail.

// F is a procedure.
fn F();
// F is a function returning ().
fn G() -> ();

// ...

// Error, procedure redeclared as a function.
fn F() -> () {
  return ();
}

// Error, function redeclared as a procedure.
fn G() {
}

Then, we could disallow any use of a procedure call in a context that depends on its return value, treating a procedure call as a statement rather than as an expression that can be used as a subexpression or an operand of an operator.

fn Func() -> ();
fn Proc();
// OK, x is of type ().
auto x = Func();
// Error, Proc is a procedure.
auto y = Proc();

Advantages:

• Removes all special treatment of () in this context. Procedures no longer need to say that their return type is implicitly () nor that they implicitly return ().
• Adding a return type to a procedure -- converting it to a function -- would be a non-breaking change.
  • But we don't have evidence that this is a common problem.
• Prevents a source of programming error where a returned () value is stored and used.
  • But it's not clear that this would be a frequent error, and it would likely be caught in other ways due to the limited API of ().

Disadvantages:

• Having distinct notions of function versus procedure would be a surprise for those coming from C++.
• Supporting auto x = F(); regardless of whether F is a function or procedure may be important for generic code.
• When migrating C++ code to Carbon, this makes the choice of function versus procedure load-bearing, as there may be uses that depend on a return value existing, for example in templates.