Understanding the ECMAScript spec, part 4

发布时间 · 标签: ECMAScript Understanding ECMAScript

All episodes

Meanwhile in other parts of the Web #

Jason Orendorff from Mozilla published a great in-depth analysis of JS syntactic quirks. Even though the implementation details differ, every JS engine faces the same problems with these quirks.

Cover grammars #

In this episode, we take a deeper look into cover grammars. They are a way to specify the grammar for syntactic constructs which look ambiguous at first.

Again, we'll skip the subscripts for [In, Yield, Await] for brevity, as they aren't important for this blog post. See part 3 for an explanation of their meaning and usage.

Finite lookaheads #

Typically, parsers decide which production to use based on a finite lookahead (a fixed amount of following tokens).

In some cases, the next token determines the production to use unambiguously. For example:

UpdateExpression :
LeftHandSideExpression
LeftHandSideExpression ++
LeftHandSideExpression --
++ UnaryExpression
-- UnaryExpression

If we're parsing an UpdateExpression and the next token is ++ or --, we know the production to use right away. If the next token is neither, it's still not too bad: we can parse a LeftHandSideExpression starting from the position we're at, and figure out what to do after we've parsed it.

If the token following the LeftHandSideExpression is ++, the production to use is UpdateExpression : LeftHandSideExpression ++. The case for -- is similar. And if the token following the LeftHandSideExpression is neither ++ nor --, we use the production UpdateExpression : LeftHandSideExpression.

Arrow function parameter list or a parenthesized expression? #

Distinguishing arrow function parameter lists from parenthesized expressions is more complicated.

For example:

let x = (a,

Is this the start of an arrow function, like this?

let x = (a, b) => { return a + b };

Or maybe it's a parenthesized expression, like this?

let x = (a, 3);

The parenthesized whatever-it-is can be arbitrarily long - we cannot know what it is based on a finite amount of tokens.

Let's imagine for a moment that we had the following straightforward productions:

AssignmentExpression :
...
ArrowFunction
ParenthesizedExpression

ArrowFunction :
ArrowParameterList => ConciseBody

Now we can't choose the production to use with a finite lookahead. If we had to parse a AssignmentExpression and the next token was (, how would we decide what to parse next? We could either parse an ArrowParameterList or a ParenthesizedExpression, but our guess could go wrong.

The very permissive new symbol: CPEAAPL #

The spec solves this problem by introducing the symbol CoverParenthesizedExpressionAndArrowParameterList (CPEAAPL for short). CPEAAPL is a symbol that is actually an ParenthesizedExpression or an ArrowParameterList behind the scenes, but we don't yet know which one.

The productions for CPEAAPL are very permissive, allowing all constructs that can occur in ParenthesizedExpressions and in ArrowParameterLists:

CPEAAPL :
( Expression )
( Expression , )
( )
( ... BindingIdentifier )
( ... BindingPattern )
( Expression , ... BindingIdentifier )
( Expression , ... BindingPattern )

For example, the following expressions are valid CPEAAPLs:

// Valid ParenthesizedExpression and ArrowParameterList:
(a, b)
(a, b = 1)

// Valid ParenthesizedExpression:
(1, 2, 3)
(function foo() { })

// Valid ArrowParameterList:
()
(a, b,)
(a, ...b)
(a = 1, ...b)

// Not valid either, but still a CPEAAPL:
(1, ...b)
(1, )

Trailing comma and the ... can occur only in ArrowParameterList. Some constructs, like b = 1 can occur in both, but they have different meanings: Inside ParenthesizedExpression it's an assignment, inside ArrowParameterList it's a parameter with a default value. Numbers and other PrimaryExpressions which are not valid parameter names (or parameter destructuring patterns) can only occur in ParenthesizedExpression. But they all can occur inside a CPEAAPL.

Using CPEAAPL in productions #

Now we can use the very permissive CPEAAPL in AssignmentExpression productions. (Note: ConditionalExpression leads to PrimaryExpression via a long production chain which is not shown here.)

AssignmentExpression :
ConditionalExpression
ArrowFunction
...

ArrowFunction :
ArrowParameters => ConciseBody

ArrowParameters :
BindingIdentifier
CPEAAPL

PrimaryExpression :
...
CPEAAPL

Imagine we're again in the situation that we need to parse an AssignmentExpression and the next token is (. Now we can parse a CPEAAPL and figure out later what production to use. It doesn't matter whether we're parsing an ArrowFunction or a ConditionalExpression, the next symbol to parse is CPEAAPL in any case!

After we've parsed the CPEAAPL, we can decide which production to use for the original AssignmentExpression (the one containing the CPEAAPL). This decision is made based on the token following the CPEAAPL.

If the token is =>, we use the production:

AssignmentExpression :
ArrowFunction

If the token is something else, we use the production:

AssignmentExpression :
ConditionalExpression

For example:

let x = (a, b) => { return a + b; };
// ^^^^^^
// CPEAAPL
// ^^
// The token following the CPEAAPL

let x = (a, 3);
// ^^^^^^
// CPEAAPL
// ^
// The token following the CPEAAPL

At that point we can keep the CPEAAPL as is and continue parsing the rest of the program. For example, if the CPEAAPL is inside an ArrowFunction, we don't yet need to look at whether it's a valid arrow function parameter list or not - that can be done later. (Real-world parsers might choose to do the validity check right away, but from the spec point of view, we don't need to.)

Restricting CPEAAPLs #

As we saw before, the grammar productions for CPEAAPL are very permissive and allow constructs (such as (1, ...a)) which are never valid. After we've done parsing the program according to the grammar, we need to disallow the corresponding illegal constructs.

The spec does this by adding the following restrictions:

Static Semantics: Early Errors

PrimaryExpression : CPEAAPL

It is a Syntax Error if CPEAAPL is not covering a ParenthesizedExpression.

Supplemental Syntax

When processing an instance of the production

PrimaryExpression : CPEAAPL

the interpretation of the CPEAAPL is refined using the following grammar:

ParenthesizedExpression : ( Expression )

This means: if a CPEAAPL occurs in the place of PrimaryExpression in the syntax tree, it is actually an ParenthesizedExpression and this is its only valid production.

Expression can never be empty, so ( ) is not a valid ParenthesizedExpression. Comma separated lists like (1, 2, 3) are created by the comma operator:

Expression :
AssignmentExpression
Expression , AssignmentExpression

Similarly, if a CPEAAPL occurs in the place of ArrowParameters, the following restrictions apply:

Static Semantics: Early Errors

ArrowParameters : CPEAAPL

It is a Syntax Error if CPEAAPL is not covering an ArrowFormalParameters.

Supplemental Syntax

When the production

ArrowParameters : CPEAAPL

is recognized the following grammar is used to refine the interpretation of CPEAAPL:

ArrowFormalParameters :
( UniqueFormalParameters )

Other cover grammars #

In addition to CPEAAPL, the spec uses cover grammars for other ambiguous-looking constructs.

ObjectLiteral is used as a cover grammar for ObjectAssignmentPattern which occurs inside arrow function parameter lists. This means that ObjectLiteral allows constructs which cannot occur inside actual object literals.

ObjectLiteral :
...
{ PropertyDefinitionList }

PropertyDefinition :
...
CoverInitializedName

CoverInitializedName :
IdentifierReference Initializer

Initializer :
= AssignmentExpression

For example:

let o = { a = 1 }; // syntax error

// Arrow function with a destructuring parameter with a default
// value:
let f = ({ a = 1 }) => { return a; };
f({}); // returns 1
f({a : 6}); // returns 6

Async arrow functions also look ambiguous with a finite lookahead:

let x = async(a,

Is this a call to a function called async or an async arrow function?

let x1 = async(a, b);
let x2 = async();
function async() { }

let x3 = async(a, b) => {};
let x4 = async();

To this end, the grammar defines a cover grammar symbol CoverCallExpressionAndAsyncArrowHead which works similarly to CPEAAPL.

Summary #

In this episode we looked into how the spec defines cover grammars and uses them in cases where we cannot identify the current syntactic construct based on a finite lookahead.

In particular, we looked into distinguishing arrow function parameter lists from parenthesized expressions and how the spec uses a cover grammar for first parsing ambiguous-looking constructs permissively and restricting them with static semantic rules later.