Destructor syntax

Pull request

Table of contents

Abstract

Fix destructor syntax ambiguity by switching to fn destroy mirroring standard function syntax. This is a purely syntactic change, maintaining destructor semantics.

Problem

The accepted destructor syntax includes out-of-line definitions such as:

class MyClass {
  destructor [addr self: Self*];
}
destructor MyClass [addr self: Self*] { ... }

The implicit parameter here could be interpreted as either an implicit parameter for MyClass or an implicit parameter for the destructor. How should ambiguities like this be resolved?

For comparison, note a generic might look like:

class GenericClass[T:! type](N:! T) { ... }
destructor GenericClass[T:! type](N:! T) [addr self: Self*] { ... }

The toolchain is able to parse this in constant time, but only because the lexer will pair brackets, so we can do lookahead at the bracket in GenericClass[ for the closing ], and look past that for the ( versus {. However, this is arbitrary lookahead and may be significantly less efficient in other parsers that people might want to use with Carbon, such as tree-sitter.

Background

In particular, we are discussing destruction as possibly similar to copy and move syntax, and trying to create a consistency between the functions.

Proposal

Destructor syntax will use standard function syntax, with destroy as a keyword for the function name.

For example, in contrast with problem examples:

class MyClass {
  fn destroy[addr self: Self*]();
}
fn MyClass.destroy[addr self: Self*]() { ... }

class GenericClass[T:! type](N:! T) { ... }
fn GenericClass[T:! type](N:! T).destroy[addr self: Self*]() { ... }

It is invalid to add other implicit or explicit parameters to the destroy function.

Not directly callable

Although the syntax of fn destroy looks similar to a regular function, the functions are not designed to be directly callable. This does not add support for my_var.destroy(). See Proposal #1154, alternative Allow functions to act as destructors for details.

Future work

Extend syntax to allow explicit marking of trivial destructors

Discussion has indicated potential utility in syntax to make the expectation of a trivial destructor explicit. This would allow a declarative way of ensuring no member accidentally caused a type to have non-trivial destruction.

Still, this requires a further extension of syntax that isn’t proposed at this time. Both determining syntax for such a feature and motivating it fully are left as future work.

Decide whether to desugar destructors to interfaces

Under this proposal, fn destroy remains a special function. We may want to make it desugar to an interface implementation, but even if we do so, the terse destructor syntax seems likely to remain. There are concerns about the ergonomics of requiring an impl in order to add a destructor to a type, and decisions would need to be made for how virtual destructors should be handled.

Copy and move functions

This proposal is set up for consistency with a possible fn copy and fn move, but those will be evaluated as part of copy and move semantics.

Rationale

  • Software and language evolution
    • Eliminates ambiguity in destructor syntax, by creating consistency with fn syntax.
    • Claiming destroy as a keyword is considered to be a good balance.
    • Syntax choices, particularly with the keyword as a function name, should not create a barrier for desugaring to an interface approach for destructions.
  • Code that is easy to read, understand, and write
    • Consistency with fn syntax should improve readability.
    • Features that impact data layout are consistently written like member declarations.

Alternatives considered

Destructor syntax options

The ambiguity between destructor MyClass [...] out-of-line destructor syntax and implicit parameters for generics is a sufficient barrier to change syntax. We do not want parsing Carbon to require arbitrary lookahead.

fn destroy was preferred because it builds on existing fn syntax.

Although adding a ., as in destructor MyClass.[...], was brought up, it didn’t present interesting advantages over fn destroy.

Destructor name options

We expect more name conflicts with C++ code using the destroy keyword than with the destructor keyword, for example with std::allocator::destroy, or visible searching LLVM code.

Still, the phrasing of destroy, particularly if we have copy and move to match, is preferred. Raw identifier syntax (r#destroy) is expected to be sufficient for name conflicts.

fn delete was mentioned as an option reusing current keywords, but declined due to the “heap allocated” implication of delete.

Non-keyword names were considered as part of proposal #1154: Destructors, and the trade-off considerations still apply.