Chapter Four: Pegular Expressions

Janet does not have native, built-in regular expressions.

You can use a third-party regular expression library if you really have to, I dunno, validate an email address or something. But most of the time, if you’re writing Janet, you’ll be writing PEGs instead.

PEG stands for “parsing expression grammar,” which is a mouthful, so I’m going to stick with the acronym, even though I just wrote a whole chapter about macros without abbreviating AST once.

As a first — extremely crude — approximation, you can think of PEGs as an alternative notation for writing regular expressions. That’s not actually correct — PEGs are quite a bit more powerful than regular expressions, for starters, and they behave differently in a few respects — but we have to start somewhere, and this will let us re-use a lot of our existing knowledge.

Here, let’s look at a few random regular expressions, and see how we’d write them in PEG format.

regex: .*
  peg: (any 1)

regex: (na)+
  peg: (some "na")

regex: \w{1,3}
  peg: (between 1 3 (choice :w "_"))
  peg: (between 1 3 (+ :w "_"))

regex: [^a-z-]
  peg: (not (choice "-" (range "az")))
  peg: (! (+ "-" (range "az")))

regex: [a-z][0-9]?
  peg: (sequence (range "az") (opt (range "09")))
  peg: (* (range "az") (? (range "09")))

Those are pretty random examples, and this is nowhere near an exhaustive list, but it’s enough for you to start forming a general idea. Let’s notice a few things from this:

1. PEGs are quite a bit more verbose than regular expressions.

2. PEGs use a lot of characters that usually mean something else.

3. PEGs are structured trees, rather than strings.

Alright, now let’s talk about some of the ways that these patterns differ from their regular expression equivalents.

First off, PEGs are always anchored to the beginning of the input, so there’s no equivalent “start of input” pattern. So (any 1) is actually equivalent to the regular expression ^.*.

Except, no, that’s not strictly true. Because PEGs do not backtrack. Which means that all repetition is implicitly ”possessive,” to use the regular expression term. So (any 1) is actually actually equivalent to ^.*+, which is not a construct that JavaScript’s regular expression engine supports.

PEGs do backtrack when using the choice combinator, as well as a few others. But backtracking is always obvious and explicit, as opposed to regular expressions’ implicit backtracking everywhere. This makes it less likely that you’ll accidentally write a PEG that executes in exponential time.

Alright. There’s one more thing we should talk about before we get to a concrete example: numbers.

We’ve seen 1 already, as a way to match any byte. You can write any other integer — 2, say, or even 3 — to match exactly that number of bytes.

But you can also write negative numbers. Negative numbers don’t advance the input at all, and they fail if you could advance that many characters. So -4 will fail unless there are fewer than four bytes left in the input. In practice I’ve only ever used this to write -1, which means “end of input.” I don’t think -1 is a particularly intuitive way to write “end of input,” so I wanted to call this out ahead of time.

Now that we’ve covered the basics, let’s look at a real example. Let’s write an HTML pretty printer.

(defn element-to-struct [tag attrs children]
  {:tag tag :attrs (struct ;attrs) :children children})

(def html-peg (peg/compile
  ~{:main (* :nodes -1)
    :nodes (any (+ :element :text))
    :element (unref
      {:main (/ (* :open-tag (group :nodes) :close-tag) ,element-to-struct)
       :open-tag (* "<" (<- :w+ :tag-name) (group (? (* :s+ :attributes))) ">")
       :attributes
         {:main (some (* :attribute (? :s+)))
          :attribute (* (<- :w+) "=" :quoted-string)
          :quoted-string (* `"` (<- (any (if-not `"` 1))) `"`)}
       :close-tag (* "</" (backmatch :tag-name) ">")})
    :text (<- (some (if-not "<" 1)))}))

(defn main [&]
  (def input (string/trim (file/read stdin :all)))
  (pp (peg/match html-peg input)))

Okay wow; we’re just diving right in huh.

First off, this isn’t really an HTML pretty printer; this is only an HTML parser. Well, strictly speaking, it’s a parser for a small subset of HTML — enough to make a point, without getting bogged down in minutiae.

So what are we looking at here?

First off, the outer pattern is a struct. The keys are names, and the values are patterns, and these patterns can reference other patterns by name — even recursively. Even mutually recursively, as you can see with :nodes and :element referring to one another.

We’ve seen named patterns like :w before, when I said it was an analog of regular expressions’ \w. But those are only the default pattern aliases, and by writing a struct like this we can create our own custom aliases, with scoping rules that make sense: patterns inside nested structs can refer to elements in the “outer struct,” but not the other way around.

Okay. Now let’s try to go through these individual patterns and make sure we understand them.

:main (* :nodes -1)
:nodes (any (+ :element :text))

The name :main is special, as that will be the pattern’s entry-point. This :main just calls :nodes, which matches zero or more :elements or :texts, and then asserts that there’s no input left. (* "x" -1) is like the regular expression ^x$.

:text (<- (some (if-not "<" 1)))

:text uses a combinator that we haven’t seen before: <-.

<- is an alias for (capture ...). We haven’t talked about captures yet, but they work similarly to regular expressions’ captures.

Just to quickly review, consider the regular expression <([^<]*)>. The parentheses around the innards there mean that there is a single “capture group,” and if we run this expression over a string, we can extract that match:

node
Welcome to Node.js v16.16.0.
Type ".help" for more information.
> /<([^<]*)>/.exec('<hello> there')
[
  '<hello>',
  'hello',
  index: 0,
  input: '<hello> there',
  groups: undefined
]

This returns an array of captured groups. The first element is the entire substring that matched the regular expression; the second is the text that matched the first (and in this case only) capture group.

> /<([^<]*)>/.exec('<hello> there')[1]
'hello'

PEGs work similarly: when you match a PEG over a string, you get a list of captures back.

repl:1:> (peg/match ~(* "<" (any (if-not ">" 1)) ">") "<hello>")
@[]

The list is empty here, because PEGs don’t implicitly capture anything. We have to explicitly ask for a capture, using <-:

repl:2:> (peg/match ~(* "<" (<- (any (if-not ">" 1))) ">") "<hello>")
@["hello"]

We could also capture the entire matching substring, if we wanted to:

repl:3:> (peg/match ~(<- (* "<" (<- (any (if-not ">" 1))) ">")) "<hello>")
@["hello" "<hello>"]

But note that captures show up “inside out.” (<- pat) first matches pat, which might push captures of its own, and then it pushes the text that pat matched.

So far this looks basically like regex country. But PEGs allow you to do so much more with captures. Here, let’s look at a slightly more interesting example:

repl:4:> (peg/match ~(* "<" (/ (<- (any (if-not ">" 1))) ,string/ascii-upper) ">") "<hello>")
@["HELLO"]

(/ ...) is an alias for (replace ...), which is a misleading name: if you pass it a function, it doesn’t replace the capture with the function, but actually maps the function over the captured value. And if you pass it a table or a struct, it looks up the capture as a key and replaces it with the value. If you pass any other values, then it actually replaces. (If actually you want to actually replace a capture with a function or a table, you have to wrap it in a function that ignores its argument.)

So we’re mapping the function string/ascii-upper over the value captured by (<- (any (if-not ">" 1))), which happens to produce a new string. But it doesn’t have to!

repl:5:> (peg/match ~(* "<" (/ (<- (any (if-not ">" 1))) ,length) ">") "<hello>")
@[5]

Our captures can be any Janet values — they don’t have to be strings. (<- pat) always captures the string that pat matches, but you can always map it, and there are other combinators that capture other things. Take $:

repl:6:> (peg/match ~(* "<" (* (<- (any (if-not ">" 1))) ($)) ">") "<hello>")
@["hello" 6]

($) is an alias for (position). It’s a pattern that always succeeds, consumes no input, and adds the current byte index to the capture stack. There’s also (line) and (column), which do what you expect.

But the most useful capture alternative is the constant operator. (constant x) always succeeds, consumes no input, and adds an arbitrary value to the capture stack. It’s useful for parsing text into something with a little more structure:

repl:7:> (peg/match ~(any (+
    (* "↑" (constant :up))
    (* "↓" (constant :down))
    (* "←" (constant :left))
    (* "→" (constant :right))
    (* "A" (constant :a))
    (* "B" (constant :b))
    (* "START" (constant :start))
    1))
    "↑↑↓↓←→←→ B A START")
@[:up :up :down :down :left :right :left :right :b :a :start]

Okay. This has been: PEG Captures 101. Now let’s get back to our HTML example.

:text (<- (some (if-not "<" 1)))

Right. So (<- (some (if-not "<" 1))) is equivalent to the regular expression ^([^<]++). It tries to match "<", and if that fails — if the next character is not < — then it advances by one character. And then it repeats, until it finds a < character or runs out of input, and finally it adds the entire string it consumed to the capture stack.

So if we give it the following input, it’s going to match the following substring:

hello yes this is <b>janet</b>

Easy. The next part is… not so easy.

:element (unref
  {:main (/ (* :open-tag (group :nodes) :close-tag) ,element-to-struct)
   :open-tag (* "<" (<- :w+ :tag-name) (group (? (* :s+ :attributes))) ">")
   :attributes
     {:main (some (* :attribute (? :s+)))
      :attribute (* (<- :w+) "=" :quoted-string)
      :quoted-string (* `"` (<- (any (if-not `"` 1))) `"`)}
   :close-tag (* "</" (backmatch :tag-name) ">")})

But we’ll take it one step at a time, and it’ll be fine.

The whole pattern is wrapped in unref, but I can’t actually explain that until the end, so we’ll skip over it for now and jump straight to :main. We’ll circle back to unref after we talk about backreferences.

:main (/ (* :open-tag (group :nodes) :close-tag) ,element-to-struct)

So an :element consists of an opening tag, some child nodes, and then a matching closing tag. Like <i>hello</i>.

But we don’t match :nodes; we match (group :nodes). Because recall that :nodes is going to push multiple nodes onto the capture stack:

:nodes (any (+ :element :text))

Specifically, anything captured in :element or :text. But (group :nodes) says “well, instead of pushing every capture individually, wrap all the captures into a tuple and push that tuple.” So we’ll match multiple nodes, but we’ll only have a single (possibly empty!) list of nodes on the capture stack when we’re done.

After we parse all of a tag’s individual components — tag name, attributes, and children — we’ll call element-to-struct to wrap it up into a nicer format. Note that element-to-struct actually takes three arguments: one for each of :element’s capture groups. (The tag name and attributes are captured by the :open-tag sub-pattern.)

But actually matching the tags is the interesting bit.

:open-tag (* "<" (<- :w+ :tag-name) (group (? (* :s+ :attributes))) ">")
:close-tag (* "</" (backmatch :tag-name) ">")

I want to draw your attention to (<- :w+ :tag-name). This is a tagged capture, and :tag-name is its “tag.” When you tag a capture, you can refer back to it later in the match — that’s exactly what (backmatch :tag-name) does.

But hark! There might be multiple tagged captures to contend with.

<p>If you have <em>nested</em> tags</p>

<p> will push a tagged capture to the stack, and so will <em>. So now there are two captures tagged :tag-name. But when we backmatch, we’re going to look for the most recent time we tagged a capture with :tag-name — which is going to be "em". This will match </em> successfully, but of course it will fail once we get to </p>.

And that’s bad! What we want to do is “scope” the tagged matches, so that parsing the <em> tag doesn’t leak out to our parsing of the <p> tag.

So that’s exactly what unref does. It says “after you’re done parsing this pattern, remove all of the tags that you associated with any captures.” By wrapping unref around our :element, we make these tagged captures local to each <tag>.

Okay, now you might be thinking: why is this a problem? Sure, we pushed "em" to the capture stack, but then we popped it off! We replaced it with an untagged struct when we called element-to-struct, right? Why can backmatch still see it?

Well, tagged captures are actually separate from the capture stack. backmatch doesn’t look for “the uppermost capture on the stack with this tag” — the tags don’t live on the capture stack at all. backmatch actually looks for “the last time we captured something with this tag.”

To help make this make sense, I’m going to describe a model of how you might implement a simple PEG matcher. We’ll keep track of two pieces of state: a stack of stacks, and a stack of tag scopes. We’ll start with a single stack on the stack-stack, and a single scope on the scope-stack, and different combinators will manipulate these.

The group combinator, for example, pushes a new stack onto the stack-stack, executes its pattern, and then pops that new stack and pushes it onto next highest stack (as a tuple). The replace combinator pushes a new stack, executes its pattern, then pops it off the stack-stack, passing its contents as positional arguments to its function. And then it pushes the return value to the new topmost stack on the stack-stack.

Meanwhile unref pushes a new tag scope, executes its pattern, and then pops the tag scope once it’s done. unref is the only combinator that affects the tag scope stack.

Alright. Now the only thing we haven’t talked about is the :attributes bit.

:attributes
  {:main (some (* :attribute (? :s+)))
   :attribute (* (<- :w+) "=" :quoted-string)
   :quoted-string (* `"` (<- (any (if-not `"` 1))) `"`)}

And I actually don’t think there’s much to say about this? You’ve seen it all already. This is easy. :s+ is “one or more whitespace characters,” and is one of many named patterns available by default.

Alright. That wasn’t so bad, was it?

(defn element-to-struct [tag attrs children]
  {:tag tag :attrs (struct ;attrs) :children children})

(def html-peg (peg/compile
  ~{:main (* :nodes -1)
    :nodes (any (+ :element :text))
    :element (unref
      {:main (/ (* :open-tag (group :nodes) :close-tag) ,element-to-struct)
       :open-tag (* "<" (<- :w+ :tag-name) (group (? (* :s+ :attributes))) ">")
       :attributes
         {:main (some (* :attribute (? :s+)))
          :attribute (* (<- :w+) "=" :quoted-string)
          :quoted-string (* `"` (<- (any (if-not `"` 1))) `"`)}
       :close-tag (* "</" (backmatch :tag-name) ">")})
    :text (<- (some (if-not "<" 1)))}))

(defn main [&]
  (def input (string/trim (file/read stdin :all)))
  (pp (peg/match html-peg input)))

When you look at it all at once, it is pretty intimidating. But just think what the equivalent regular expression would look like! Oh, wait. You can’t. Parsing HTML with regexes is famously impossible.

We’ve already seen a lot of useful PEG combinators, but we’re not limited to the built-in operations that Janet gives us. We can actually interleave arbitrary functions into a PEG, and use them to guide the matching process. This allows us to write custom predicates to express complicated matching logic that would be very difficult to implement natively in a PEG (“identifier with more vowels than consonants”), but it’s especially useful when we already have a regular function that knows how to parse strings.

For example, scan-number is a built-in function that parses numeric strings into numbers:

repl:1:> (scan-number "512")
512
repl:2:> (scan-number "512x")
nil

If we wanted to parse a number somewhere in a PEG, then… well, we’d use the built-in (number) operator that does exactly that. But let’s pretend that that doesn’t exist for a second, and try to implement it in terms of scan-number. Here’s a first attempt:

repl:1:> (peg/match ~(/ (<- (some (+ :d (set ".-+")))) ,scan-number) "123")
@[123]

That works, sometimes. But of course that number pattern is not very accurate, and we already saw that scan-number will return nil if we give it a bad input:

repl:2:> (peg/match ~(/ (<- (some (+ :d (set ".-+")))) ,scan-number) "1-12-3+3.-++")
@[nil]

But the match still succeeded, and captured nil, because that was what we told it to do.

So we could try to carefully write a valid number pattern here, such that we only ever pass valid input to scan-number. But we don’t want to do that. That sounds hard. We just want the pattern to fail if scan-number can’t actually parse a number.

Enter cmt:

repl:3:> (peg/match ~(cmt (<- (some (+ :d (set ".-+")))) ,scan-number) "1-12-3+3.-++")
nil

So cmt is very similar to replace, except that if your function returns something falsy (remember: just nil or false), then the cmt clause itself will fail to match. It’s sort of like a map vs filterMap situation.

cmt stands for “match-time capture,” apparently, even though the letters are not in that order. The name comes to us from LPEG, the Lua PEG library that inspired Janet’s PEG library, where all capture-related functions start with C. It’s a very useful function despite the confusing name, and there’s something else that makes it even more useful: the -> operator.

-> stands for backref, and it looks quite strange at first glance: all it does is re-capture a previously tagged capture. If you just used it by itself, it would duplicate previously tagged captures onto the capture stack and consume no input, which doesn’t sound very useful.

repl:1:> (peg/match ~(* (<- :d+ :num) (-> :num)) "123")
@["123" "123"]

But if you use it inside the pattern you pass to cmt, you can add previous captures as arguments to your custom mapping predicate.

Here’s a concrete, if extremely dumb, example: I’ve invented my own HTML dialect that is identical to regular HTML, except that <head> tags can optionally be closed with a </tail> tag, because that’s modestly whimsical.

Previously we were able to use backmatch to match closing tags, because they happened to be bytewise-identical to the values we captured in :open-tag:

:close-tag (* "</" (backmatch :tag-name) ">")

But now that’s no longer true, and backmatch isn’t sufficient to handle this very practical HTML dialect. We’ll have to write some logic:

(defn check-close-tag [open-tag close-tag]
  (or (= open-tag close-tag)
      (and (= open-tag "head")
           (= close-tag "tail"))))

:close-tag (* "</" (drop (cmt (* (-> :tag-name) (<- :w+)) ,check-close-tag)) ">")

Notice that we “re-capture” :tag-name, in addition to capturing the :w+. Because cmt needs a single pattern to execute, I stuck them together with *, but both of these captures will be passed as arguments to check-close-tag.

Neat.

We are now very close to knowing everything there is to know about PEGs, but I think we should talk about one more thing before we leave this chapter:

Regular expressions aren’t just useful for matching or extracting text. They’re also useful for changing text.

Regex replace is a common primitive operation; you use it all the time in your editor or with sed or whatever. And of course Janet has a native peg/replace function, and we’re going to talk about it soon.

But let’s just pretend, for a moment, that it doesn’t exist. Because it turns out that you don’t actually need a built-in PEG replace function: you can implement replacement as a special case of capturing.

It’s a pretty simple trick: we’re going to write a PEG that captures two things: the part of the string that matches the pattern we want to replace, and the entire rest of the string.

Just so we have something concrete to work with, let’s write a chaotic evil PEG: given a string, we’ll find all of the Oxford commas in that string, and replace them with Oxford semicolons.

So given input like this:

this is dumb, confusing, and upsetting

We’ll wind up with:

this is dumb, confusing; and upsetting

Naturally.

So the PEG itself is easy: we just want to match the literal string ", and", wherever it appears in the input:

repl:1:> (peg/match ~(any (+ ", and" 1)) "a, b, and c")
@[]

Okay. It did work; you just can’t tell. Let’s replace it, which will automatically capture the output, so we can at least see that it’s working:

repl:2:> (peg/match ~(any (+ (/ ", and" "; and") 1)) "a, b, and c")
@["; and"]

Okay. And now let’s also capture everything else:

repl:3:> (peg/match ~(any (+ (/ ", and" "; and") (<- 1))) "a, b, and c")
@["a" "," " " "b" "; and" " " "c"]

And we’re done! Sort of! We have the entire modified string, as a list of captures, and all we have to do now is stick them back together.

And I know: this looks unbelievably inefficient. And it would be, if we just called, like, string/concat on this result. But Janet has a way to efficiently join these matches together without even making these intermediate string allocations in the first place.

It’s called accumulate, although I’m going to use the short alias %:

repl:4:> (peg/match ~(% (any (+ (/ ", and" "; and") (<- 1)))) "a, b, and c")
@["a, b; and c"]

And accumulate is special-cased in the PEG engine: while Janet is executing the pattern inside an accumulate block, anything that would normally push captures onto the stack instead just copies it into a shared mutable buffer. And once it’s done with its pattern, that buffer becomes a string, and accumulate pushes it onto the capture stack.

So that’s a global replace. But what if you only want to replace the first occurrence?

Here’s one way:

repl:5:> (peg/match ~(% (any (+ (* (/ ", and" "; and") (<- (to -1))) (<- 1)))) "a, and b, and c")
@["a; and b, and c"]

After we match and replace the pattern, we immediately consume the rest of the string, so that the any repetition won’t fire again.

Hey look! We did it. accumulate was the last combinator on my list of combinators to tell you about, and I just told you about it. That means we’re almost done with the chapter now.

But we get to do something fun and easy first. There’s actually another way that we could have written that last pattern:

repl:6:> (peg/match ~(% (any (+ (* (/ ", and" "; and") '(to -1)) '1))) "a, and b, and c")
@["a; and b, and c"]

We replaced all of the (<- x) captures with just 'x, which does exactly the same thing. How does that work?

Well, 'x is actually just syntax sugar for (quote x). They both parse into exactly the same abstract syntax tree: if you’re writing a macro or a PEG engine or whatever else, you can’t actually tell whether it was originally written using the ' shorthand or not. So when the whole thing is quasiquoted:

repl:7:> ~(% (any (+ (* (/ ", and" "; and") '(to -1)) '1)))
(% (any (+ (* (/ ", and" "; and") (quote (to -1))) (quote 1))))

All of those single-quotes get expanded into (quote) forms, and quote is just another alias for capture in the PEG parser. But when you use the shorthand, you can save quite a few parentheses.

Now, it’s fun to work through these examples, and I think it’s valuable to understand how they work, just in case you ever find yourself needing to perform some weird text surgery deep inside some complicated PEG. But of course, in real life, you only have to write:

repl:1:> (peg/replace ", and" "; and" "a, b, and c")
@"a, b; and c"

There is also (peg/replace-all), (peg/find), which returns the index of the first match; and (peg/find-all), which returns all of the indices where the PEG would match.

Alright. That’s all of the important PEG stuff sorted, but I want to close with a few scattered, wandering observations:

1. PEGs operate on bytes, not characters.

2. PEGs are harder to debug than regular expressions.

3. You can compile PEGs ahead of time.

4. You can define your own combinators.

5. PEGs are the best.