Skip to content
Bastiaan Veelo edited this page Feb 6, 2020 · 15 revisions

Generating Code

Once a grammar is defined and a parse tree was generated, how can it be used? We already saw how to use the parse tree in Using the Parse Tree, at runtime or compile-time. But since Pegged can parse at compile-time, the information can also be used to generate during compilation a string containing D code, which can in turn be mixed in.

So, what's the difference with what we saw in lesson Using the Parse Tree? None whatsoever, except be sure to do everything at compile-time, since you cannot mixin runtime strings.

This is exactly what Pegged itself does with the input you give it. So let's use Pegged as an example.

Part of the PEG grammar is :

Grammar      <- Spacing GrammarName Definition+ :eoi
Definition   <- LhsName Arrow Expression
Expression   <- Sequence (:OR Sequence)*
Sequence     <- Prefix+
Prefix       <- (POS / NEG / FUSE / DISCARD / KEEP / DROP)* Suffix
Suffix       <- Primary (OPTION / ZEROORMORE / ONEORMORE / Action)*
(...)

A Grammar is a bunch of Definitions up to the end of the input (to let nothing unparsed and provoke an error if something is left unparsed. A Definition is a name, an Arrow and an Expression and so on. Pretty standard stuff for you, now. Let's see how that can be used to create code.

As said in How Pegged Works, the basic PEG parsers are structs and functions like and (sequences), or (choices) or literal (literals). A PEG expression like:

A <- B C D+ E?

Is implemented in D code as:

ParseTree A (ParseTree p)
{
    return  and!(B, C, oneOrMore!(D), option!(E))(p);
}

So really, to create such a code from a parse tree we have to

  1. Get the rule name

  2. (Optionally, get the arrow type, not described here)

  3. Get the parsing expression parse tree

  4. Transform the expression into something like and!(B, C, oneOrMore!(D), option!(E))

  5. concatenate the parts : name + inheritance + parsing stuff inside the struct.

  6. (Optionally, collect the rule names to get better parse tree deforestation: cutting rules external to the grammar in the final parse tree)

Step 1) is standard identifier recognition. I'll focus on steps 3) and 4). If you look into pegged.grammar, you'll find the inner generateCode function, inside the grammar function. Its role is to transform a parse tree into D code. It's mainly a big switch on the root node name and recursion on the children. Pegged see parsing expression as choices (Expr1 / Expr2 / Epxr3 ...), composed of sequences (Expr1a Expr1b Expr1c ...), themselves composed of a prefix followed by a suffix and so on. Let's zoom on expressions and sequences:

case "Pegged.Expression":
	if (p.children.length > 1) // OR expression
	{
		result = "or!(";
		foreach(seq; p.children)
			result ~= generateCode(seq) ~ ", ";
		result = result[0..$-2] ~ ")";
	}
	else // One child -> just a sequence, no need for an or!( , )
	{
		result = generateCode(p.children[0]);
	}
	break;

When there is a choice node, its children are the different sub-rules. The final code must look something like or!(Expr1, Expr2, Expr3), which is exactly what the foreach loop build, invoking generateCode again on the children nodes. There is a slight simplification for 1-element choices: no need to create an or!(Expr), just return Expr. Note that the way the Pegged grammar is built, any lone top level expression is seen as 1-element choice (and then a 1-element sequence), so this simplification is used quite frequently.

The same goes for sequences:

case "Pegged.Sequence":
	if (p.children.length > 1) // real sequence
	{
		result = "and!(";
		foreach(seq; p.children)
		{
			string elementCode = generateCode(seq);
			if (elementCode[0..5] == "and!(") // flattening inner sequences
				elementCode = elementCode[5..$-1]; // cutting 'and!(' and ')'
			result ~= elementCode ~ ", ";
		}
		result = result[0..$-2] ~ ")";
	}
	else // One child -> just a Suffix, no need for a and!( , )
	{
		result = generateCode(p.children[0]);
	}
	break;

The difference here is that sequence of sequences are flattened into a single sequence: A (B C) D is rewritten into (the equivalent of) A B C D, that is and!(A, B, C, D).

You now understand how, given a rule such as:

A <- B C
   / D
   / E F (G / H)

Pegged generates or!(and!(B,C), D, and!(E, F, or!(G, H))).

The same process can be done in your code: define a DSL in Pegged, write a parse-tree-to-D-code function and mix it in:

mixin(sql(`

USE mydatabase;

SELECT customer, SUM(quantity) AS "Total Items"
FROM orders
GROUP BY customer;

`));

Next lesson: User-Defined Parsers


Pegged Tutorial

Clone this wiki locally