Skip to content
Bastiaan Veelo edited this page Feb 15, 2016 · 24 revisions

Semantic Actions

Syntax

Semantic actions are introduced with the {ActionName} syntax, just behind the expression you want to give the action to. When the expression returns from parsing, ActionName is called on the expression parse tree. An action can be a named callable or a function/delegate literal (anonymous function), as long as it accepts a ParseTree as argument and returns a ParseTree.

Actions:
Rule1 <- Expr1 Expr2 {ActionA} Expr3 {ActionB)
Rule2 <- (Expr1 / Expr2) {ActionA} Expr3
Rule3 <{ActionB} Expr2 / Expr3 (!Expr1 Expr2)*
Rule4 <- (Expr1 {ActionA} Expr2 {ActionB}) {ActionC}

The previous example demonstrates some ways an action can be declared:

  • for Rule1, ActionA is called once Expr2 has finished parsing and then ActionB for Expr3.

  • for Rule2, ActionA is called once the 'or' expression returns, be it with an Expr1 or an Expr2. Expr3 has no associated action.

  • Rule3 is an example of rule-level action: once the right-hand side expression returns, ActionB is called on its output.

  • Rule4 is an example of nested actions: ActionA is called after Expr1, ActionB after Expr2 and than ActionC on the global sequence output.

Anonymous Functions as Actions

Actions can be declared as anonymous functions, examples of which include:

Actions:
Rule1 <- Expr1 { function ParseTree(ParseTree p) {return p;} } # Verbose form
Rule2 <- Expr2 { delegate ParseTree(ParseTree p) {return p;} } # Verbose form
Rule3 <- Expr3 { (p) {return p;} } # Parameter and return-type inferred 
Rule4 <- Expr4 { p => p.children[0] } # Lambda syntax

These actions can be freely mixed with normal callable actions, e.g.:

Actions:
Rule1 <- Expr1 { myAction1, p => p.children[0], myAction2 } # Mixed callable/lambda action list

When using anonymous functions as actions inside ordered choice rules, it may be necessary to explicitly declare the parameter type to avoid compilation errors involving delegates and nested templates, e.g.:

Actions:
Rule1 <- 
    / Expr1 { (ParseTree a) => doStuff(a) } # instead of { (a) => doStuff(a) }
    / Expr2 { (ParseTree a) => doOtherStuff(a) } # ditto

Note: avoid using delimited strings with identifier delimiters (http://dlang.org/lex.html#DelimitedString) inside anonymous function actions, as they may not be correctly parsed (in order to keep the Pegged grammar relatively simple). However delimited strings using nested delimiters ({} [] <> ()) are OK.

What Can Be Done With Semantic Actions

What are actions good for? They offer the user an opportunity to act on the parse tree during the parsing process. Since they are passed the complete output of an expression, they can modify the output, or construct some value based on the passed argument. Let's demonstrate the two uses:

PT cutChildren(PT)(PT p)
{
    p.children = null;
    return p;
}

cutChildren is a parse-tree-pruning action: it just nullifies the children array in the parse tree. Put it after expressions where you don't care for children expressions and just want to keep the captures.

Note that cutChildren is a function template. It's because it must be able to accept any kind of parse tree. (TODO: more explanations. For now, just make your semantic actions templates and everything will be alright).

Now, this action keeps only the first match of a rule:

PT first(PT)(PT p)
{
    if (p.matches.length > 1)
        p.matches.length = 1;
    return p;
}

Now, let's use actions to validate some XML nodes. First, a grammar:

mixin(grammar(`
XML:
    Node       <- OpeningTag (Node / Text)* ClosingTag
    OpeningTag <- "<"  identifier ">"
    Closingtag <- "</" identifier ">"
    Text       <~ (!OpeningTag !ClosingTag .)*  # Any char, as long as it's not a tag
`));

Suppose we want to validate the tags: any opening tag must be closed by an equivalent closing tag. We will use actions to push the tag identifier on a stack, which will be popped by closing tags.

import std.array;
string[] nameStack;

PT opening(PT)(PT p)
{
    nameStack ~= p.matches[0];
    return p;
}

PT closing(PT)(PT p)
{
    if (nameStack.back != p.matches[0])
        p.successful = false; // Make the rule fail
    else
        nameStack.popBack;
    return p;
}

And, adding actions to the XML grammar:

mixin(grammar(`
XML:
    Node       <- OpeningTag{opening} (Node / Text)* ClosingTag{closing}
    OpeningTag <- "<"  identifier ">"
    Closingtag <- "</" identifier ">"
    Text       <~ (!OpeningTag !ClosingTag .)*  # Any char, as long as it's not a tag
`));

Now, let's test it:

assert( XML("<a> Hello <b> World </b> ! </a>").successful);
assert(!XML("<a> Hello <b> World </c> ! </a>").successful); // <b> closed by a </c>
assert(!XML("<a> Hello <b> World </a> ! </b>").successful); // <a> and <b> incorrectly nested

As you can see, correctly nested nodes get parsed, but not incorrectly closed and nested nodes. This means actions do validation while parsing and, if the parsing is successful, you can be sure the input is a correctly nested collection of nodes and that the parse tree is also correct for any following function to act upon.

Expression-Level or Rule-Level Actions?

There is no real difference between

Rule1 <- Expr1 {Action}

and

Rule1 <{Action} Expr1

But there is a difference between

Rule1 <-  A {ActionA} B {ActionB}
Rule2 <-  B A
A <- ...
B <- ...

and

Rule1 <-  A B
Rule2 <-  B A
A <{ActionA} ...
B <{ActionB} ...

The latter means that ActionA gets called every time A is used, in Rule1 as well as in Rule2. Whereas for the former grammar, the actions are activated only for Rule1. Just decide what you want for your grammar.

Success and Failure

Actions are called when a rule finishes parsing, successfully or not. Note that sequences stop once a rule fail. Like this:

Rule1 <- A {foo} B {bar} C {baz}

If A succeeds, foo is called on its result. If B then fails, bar is still called on its resulting parse tree. Unless bar changes B.successful to true (a legal, but dangerous move), B {bar} will fail and C {baz} is not tested by Pegged.

A previous version of Pegged called actions only on successful parses. This forbade 'recovery rules', that'd accept a failed parse attempt and correct it to make it OK. So I'm testing letting rules be called after any parse result.

Side-Effects

A consequence of the previous section is that side-effects (like storing a value in a variable) can be activated inside a failed rule. In the previous example, foo was called. If foo acts like opening, then something somewhere was changed. Even though Rule1 failed, some partial modification was made. Maybe a cleaner solution would be to test each rule in a sandbox and undo any changes that were made by a failed rule. For actions acting on a grammar inner fields (something I plan to add), it's feasible: the current state can be duplicated, passed to a rule as a context variable and returned with the parse tree. If the rule fails, then the returned context is discarded. When the rule succeeds, the caller context is assigned the returned context.

Of course, all this copying and passing around'll probably slow the parsing process down and I cannot protect against any global variable modification.

Multiple Actions

You can have as many actions as you wish between the curly braces (min. 1, {} is not a valid action). Separate the actions names by a comma:

Rule1 <- A {foo, bar, baz} B {foo, baz}

Once A returns, foo is called on its result, then bar on foo's result and then baz. For B, it'll be foo and then baz.

Predefined Actions

For now, Pegged has no predefined actions. Here is the minimum list I want to add:

  • discardChildren (nullifies the children field of a parse tree)
  • discardMatches (nullifies the matches field of a parse tree)
  • fuseMatches (concatenates a parse tree's matches together)
  • failure (parsetree.successful = false)
  • success (parsetree.successful = true)
  • discard (discards anything the parse tree achieved: reset .end to be equal to .begin, discards the matches and the children)

Possible Extensions

I'm playing with the following ideas concerning actions:

  • Permitting other arguments, like this: {dropChild(1)}, which would call dropChild(ruleOutput, 1). Heck, in DMD with UFCS, it's just ruleOutput.dropChild(1) which is quite easy to code.

  • For now, actions are what I'd call internal actions: they act on the parse tree and any external action is a side-effect (assigning to external variables, for example). I could introduce 'external actions', for example with <ActionName> . These would return any D type and plug into one another during parsing. This would allow Pegged to have the same run-of-the-mill example of calculus on arithmetic expressions as other parser generators. We'll see...


Next lesson Generating Code


Pegged Tutorial