Additions to go/ast and go/token to support parameterized functions and types

Authors: Rob Findley, Robert Griesemer

Last Updated: 2021-08-18

Abstract

This document proposes changes to go/ast to store the additional syntactic information necessary for the type parameters proposal (#43651), including the amendment for type sets (#45346). The changes to go/types related to type checking are discussed in a separate proposal.

Syntax Changes

See the type parameters proposal for a full discussion of the language changes to support parameterized functions and types, but to summarize the changes in syntax:

  • Type and function declarations get optional type parameters, as in type List[T any] ... or func f[T1, T2 any]() { ... }. Type parameters are a parameter list.
  • Parameterized types may be instantiated with one or more type arguments, to make them non-parameterized type expressions, as in l := &List[int]{} or type intList List[int]. Type arguments are an expression list.
  • Parameterized functions may be instantiated with one or more type arguments when they are called or used as function values, as in g := f[int] or x := f[int](). Function type arguments are an expression list.
  • Interface types can have new embedded elements that restrict the set of types that may implement them, for example interface { ~int64|~float64 }. Such elements are type expressions of the form T1 | T2 ... Tn where each term Ti stands for a type or a ~T where T is a type.

Proposal

The sections below describe new types and functions to be added, as well as their invariants. For a detailed discussion of these design choices, see the appendix.

For type parameters in type and function declarations

type TypeSpec struct {
	// ...existing fields

	TypeParams *FieldList
}

type FuncType struct {
	// ...existing fields

	TypeParams *FieldList
}

To represent type parameters in type and function declarations, both ast.TypeSpec and ast.FuncType gain a new TypeParams *FieldList field, which will be nil in the case of non-parameterized types and functions.

For type and function instantiation

To represent both type and function instantiation with type arguments, we introduce a new node type ast.IndexListExpr, which is an Expr node similar to ast.IndexExpr, but with a slice of indices rather than a single index:

type IndexListExpr struct {
	X Expr
	Lbrack token.Pos
	Indices []Expr
	Rbrack token.Pos
}

func (*IndexListExpr) End() token.Pos
func (*IndexListExpr) Pos() token.Pos

Type and function instance expressions will be parsed into a single IndexExpr if there is only one index, and an IndexListExpr if there is more than one index. Specifically, when encountering an expression f[expr1, ..., exprN] with N argument expressions, we parse as follows:

  1. If N == 1, as in normal index expressions f[expr], we parse an IndexExpr.
  2. If N > 1, parse an IndexListExpr with Indices set to the parsed expressions expr1, …, exprN
  3. If N == 0, as in the invalid expression f[], we parse an IndexExpr with BadExpr for its Index (this matches the current behavior for invalid index expressions).

There were several alternatives considered for representing this syntax. At least two of these alternatives were implemented. They are worth discussing:

  • Add a new ListExpr node type that holds an expression list, to serve as the Index field for an IndexExpr when N >= 2. This is an elegant solution, but results in inefficient storage and, more importantly, adds a new node type that exists only to alter the meaning of an existing node. This is inconsistent with the design of other nodes in go/ast, where additional nodes are preferred to overloading existing nodes. Compare with RangeStmt and TypeSwitchStmt, which are distinct nodes in go/ast. Having distinct nodes is generally easier to work with, as each node has a more uniform composition.
  • Overload ast.CallExpr to have a Brackets bool field, so f[T] would be analogous to f(T), but with Brackets set to true. This is roughly equivalent to the IndexListExpr node, and allows us to avoid adding a new type. However, it overloads the meaning of CallExpr and adds an additional field.
  • Add an Tail []Expr field to IndexExpr to hold additional type arguments. While this avoids a new node type, it adds an extra field to IndexExpr even when not needed.

For type restrictions

package token

const TILDE Token = 88

The new syntax for type restrictions in interfaces can be represented using existing node types.

We can represent the expression ~T1|T2 |~T3 in interface { ~T1|T2|~T3 } as a single embedded expression (i.e. an *ast.Field with empty Names), consisting of unary and binary expressions. Specifically, we can introduce a new token token.TILDE, and represent ~expr as an *ast.UnaryExpr where Op is token.TILDE. We can represent expr1|expr2 as an *ast.BinaryExpr where Op is token.OR, as would be done for a value expression involving bitwise-or.

Appendix: Considerations for API changes to go/ast

This section discusses what makes a change to go/ast break compatibility, what impact changes can have on users beyond pure compatibility, and what type of information is available to the parser at the time we choose a representation for syntax.

As described in the go1 compatibility promise, it is not enough for standard library packages to simply make no breaking API changes: valid programs must continue to both compile and run. Or put differently: the API of a library is both the structure and runtime behavior of its exported API.

This matters because the definition of a ‘valid program’ using go/ast is arguably a gray area. In go/ast, there is no separation between the interface to AST nodes and the data they contain: the node set consists entirely of pointers to structs where every field is exported. Is it a valid use of go/ast to assume that every field is exported (e.g. walk nodes using reflection)? Is it valid to assume that the set of nodes is complete (e.g. by panicking in the default clause of a type switch)? Which fields may be assumed to be non-nil?

For the purpose of this document, I propose the following heuristic:

A breaking change to go/ast (or go/parser) is any change that modifies (1) the parsed representation of existing, valid Go code, or (2) the per-node invariants that are preserved in the representation of invalid Go code. We consider all documented invariants plus any additional invariants that are assumed in significant amounts of code.

Of these two clauses, (1) is straightforward and hopefully uncontroversial: code that is valid in Go 1.17 must parse to an equivalent AST in Go 1.18. (2) is more subtle: there is no guarantee that the syntax tree of invalid code will not change. After all, use of type parameters is invalid in go1.17. Rather, the only guarantee is that if a property of existing fields holds for a node type N in all representations of code, valid or invalid, it should continue to hold. For example, ast.Walk assumes that ast.IndexExpr.Index is never nil. This must be preserved if we use IndexExpr to represent type instantiation, even for invalid instantiation expressions such as var l List[].

The rationale for this heuristic is pragmatic: there is too much code in the wild that makes assumptions about nodes in AST representations; that code should not break.

Notable edge cases:

  • It makes sense to preserve the property that all fields on Nodes are exported. cmd/gofmt makes this assumption, and it is reasonable to assume that other users will have made this assumption as well (and this was the original intent).
  • There is code in the wild that assumes the completeness of node sets, i.e. panicking if an unknown node is encountered. For example, see issue vscode-go#1551 for x/tools. If we were to consider this a valid use of go/ast, that would mean that we could never introduce a new node type. In order to avoid introducing new nodes, we'd have to pack new syntactic constructs into existing nodes, resulting in cumbersome APIs and increased memory usage. Also, from another perspective, assuming the completeness of node types is not so different from assuming the completeness of fields in struct literals, which is explicitly not guaranteed by the compatibility promise. We should therefore consider adding a new node type a valid change (and do our best to publicize this change to our users).

Finally, when selecting our representation, keep in mind that the parser has access to only local syntactic information. Therefore, it cannot differentiate between, for example, the representation of f[T] in var f []func(); T := 0; f[T]() and func f[S any](){} ... f[T]().