This example will show how to parse /etc/passwd using PEGs.
First we define an example /etc/passwd file:
(define *etc-passwd* "root:x:0:0:root:/root:/bin/bash daemon:x:1:1:daemon:/usr/sbin:/bin/sh bin:x:2:2:bin:/bin:/bin/sh sys:x:3:3:sys:/dev:/bin/sh nobody:x:65534:65534:nobody:/nonexistent:/bin/sh messagebus:x:103:107::/var/run/dbus:/bin/false ")
As a first pass at this, we might want to have all the entries in /etc/passwd in a list.
Doing this with string-based PEG syntax would look like this:
(define-peg-string-patterns "passwd <- entry* !. entry <-- (! NL .)* NL* NL < '\n'")
A passwd
file is 0 or more entries (entry*
) until the end
of the file (!.
(.
is any character, so !.
means
“not anything”)). We want to capture the data in the nonterminal
passwd
, but not tag it with the name, so we use <-
.
An entry is a series of 0 or more characters that aren’t newlines
((! NL .)*
) followed by 0 or more newlines (NL*
). We want
to tag all the entries with entry
, so we use <--
.
A newline is just a literal newline ('\n'
). We don’t want a
bunch of newlines cluttering up the output, so we use <
to throw
away the captured data.
Here is the same PEG defined using S-expressions:
(define-peg-pattern passwd body (and (* entry) (not-followed-by peg-any))) (define-peg-pattern entry all (and (* (and (not-followed-by NL) peg-any)) (* NL))) (define-peg-pattern NL none "\n")
Obviously this is much more verbose. On the other hand, it’s more
explicit, and thus easier to build automatically. However, there are
some tricks that make S-expressions easier to use in some cases. One is
the ignore
keyword; the string syntax has no way to say “throw
away this text” except breaking it out into a separate nonterminal.
For instance, to throw away the newlines we had to define NL
. In
the S-expression syntax, we could have simply written (ignore
"\n")
. Also, for the cases where string syntax is really much cleaner,
the peg
keyword can be used to embed string syntax in
S-expression syntax. For instance, we could have written:
(define-peg-pattern passwd body (peg "entry* !."))
However we define it, parsing *etc-passwd*
with the passwd
nonterminal yields the same results:
(peg:tree (match-pattern passwd *etc-passwd*)) ⇒ ((entry "root:x:0:0:root:/root:/bin/bash") (entry "daemon:x:1:1:daemon:/usr/sbin:/bin/sh") (entry "bin:x:2:2:bin:/bin:/bin/sh") (entry "sys:x:3:3:sys:/dev:/bin/sh") (entry "nobody:x:65534:65534:nobody:/nonexistent:/bin/sh") (entry "messagebus:x:103:107::/var/run/dbus:/bin/false"))
However, here is something to be wary of:
(peg:tree (match-pattern passwd "one entry")) ⇒ (entry "one entry")
By default, the parse trees generated by PEGs are compressed as much as
possible without losing information. It may not look like this is what
you want at first, but uncompressed parse trees are an enormous headache
(there’s no easy way to predict how deep particular lists will nest,
there are empty lists littered everywhere, etc. etc.). One side-effect
of this, however, is that sometimes the compressor is too aggressive.
No information is discarded when ((entry "one entry"))
is
compressed to (entry "one entry")
, but in this particular case it
probably isn’t what we want.
There are two functions for easily dealing with this:
keyword-flatten
and context-flatten
. The
keyword-flatten
function takes a list of keywords and a list to
flatten, then tries to coerce the list such that the first element of
all sublists is one of the keywords. The context-flatten
function is similar, but instead of a list of keywords it takes a
predicate that should indicate whether a given sublist is good enough
(refer to the API reference for more details).
What we want here is keyword-flatten
.
(keyword-flatten '(entry) (peg:tree (match-pattern passwd *etc-passwd*))) ⇒ ((entry "root:x:0:0:root:/root:/bin/bash") (entry "daemon:x:1:1:daemon:/usr/sbin:/bin/sh") (entry "bin:x:2:2:bin:/bin:/bin/sh") (entry "sys:x:3:3:sys:/dev:/bin/sh") (entry "nobody:x:65534:65534:nobody:/nonexistent:/bin/sh") (entry "messagebus:x:103:107::/var/run/dbus:/bin/false")) (keyword-flatten '(entry) (peg:tree (match-pattern passwd "one entry"))) ⇒ ((entry "one entry"))
Of course, this is a somewhat contrived example. In practice we would
probably just tag the passwd
nonterminal to remove the ambiguity
(using either the all
keyword for S-expressions or the <--
symbol for strings)..
(define-peg-pattern tag-passwd all (peg "entry* !.")) (peg:tree (match-pattern tag-passwd *etc-passwd*)) ⇒ (tag-passwd (entry "root:x:0:0:root:/root:/bin/bash") (entry "daemon:x:1:1:daemon:/usr/sbin:/bin/sh") (entry "bin:x:2:2:bin:/bin:/bin/sh") (entry "sys:x:3:3:sys:/dev:/bin/sh") (entry "nobody:x:65534:65534:nobody:/nonexistent:/bin/sh") (entry "messagebus:x:103:107::/var/run/dbus:/bin/false")) (peg:tree (match-pattern tag-passwd "one entry")) (tag-passwd (entry "one entry"))
If you’re ever uncertain about the potential results of parsing something, remember the two absolute rules:
For the purposes of (1), "parsing information" means things tagged with
the any
keyword or the <--
symbol. Plain strings will be
concatenated.
Let’s extend this example a bit more and actually pull some useful information out of the passwd file:
(define-peg-string-patterns "passwd <-- entry* !. entry <-- login C pass C uid C gid C nameORcomment C homedir C shell NL* login <-- text pass <-- text uid <-- [0-9]* gid <-- [0-9]* nameORcomment <-- text homedir <-- path shell <-- path path <-- (SLASH pathELEMENT)* pathELEMENT <-- (!NL !C !'/' .)* text <- (!NL !C .)* C < ':' NL < '\n' SLASH < '/'")
This produces rather pretty parse trees:
(passwd (entry (login "root") (pass "x") (uid "0") (gid "0") (nameORcomment "root") (homedir (path (pathELEMENT "root"))) (shell (path (pathELEMENT "bin") (pathELEMENT "bash")))) (entry (login "daemon") (pass "x") (uid "1") (gid "1") (nameORcomment "daemon") (homedir (path (pathELEMENT "usr") (pathELEMENT "sbin"))) (shell (path (pathELEMENT "bin") (pathELEMENT "sh")))) (entry (login "bin") (pass "x") (uid "2") (gid "2") (nameORcomment "bin") (homedir (path (pathELEMENT "bin"))) (shell (path (pathELEMENT "bin") (pathELEMENT "sh")))) (entry (login "sys") (pass "x") (uid "3") (gid "3") (nameORcomment "sys") (homedir (path (pathELEMENT "dev"))) (shell (path (pathELEMENT "bin") (pathELEMENT "sh")))) (entry (login "nobody") (pass "x") (uid "65534") (gid "65534") (nameORcomment "nobody") (homedir (path (pathELEMENT "nonexistent"))) (shell (path (pathELEMENT "bin") (pathELEMENT "sh")))) (entry (login "messagebus") (pass "x") (uid "103") (gid "107") nameORcomment (homedir (path (pathELEMENT "var") (pathELEMENT "run") (pathELEMENT "dbus"))) (shell (path (pathELEMENT "bin") (pathELEMENT "false")))))
Notice that when there’s no entry in a field (e.g. nameORcomment
for messagebus) the symbol is inserted. This is the “don’t throw away
any information” rule—we successfully matched a nameORcomment
of 0 characters (since we used *
when defining it). This is
usually what you want, because it allows you to e.g. use list-ref
to pull out elements (since they all have known offsets).
If you’d prefer not to have symbols for empty matches, you can replace
the *
with a +
and add a ?
after the
nameORcomment
in entry
. Then it will try to parse 1 or
more characters, fail (inserting nothing into the parse tree), but
continue because it didn’t have to match the nameORcomment to continue.
We can parse simple mathematical expressions with the following PEG:
(define-peg-string-patterns "expr <- sum sum <-- (product ('+' / '-') sum) / product product <-- (value ('*' / '/') product) / value value <-- number / '(' expr ')' number <-- [0-9]+")
Then:
(peg:tree (match-pattern expr "1+1/2*3+(1+1)/2")) ⇒ (sum (product (value (number "1"))) "+" (sum (product (value (number "1")) "/" (product (value (number "2")) "*" (product (value (number "3"))))) "+" (sum (product (value "(" (sum (product (value (number "1"))) "+" (sum (product (value (number "1"))))) ")") "/" (product (value (number "2")))))))
There is very little wasted effort in this PEG. The number
nonterminal has to be tagged because otherwise the numbers might run
together with the arithmetic expressions during the string concatenation
stage of parse-tree compression (the parser will see “1” followed by
“/” and decide to call it “1/”). When in doubt, tag.
It is very easy to turn these parse trees into lisp expressions:
(define (parse-sum sum left . rest) (if (null? rest) (apply parse-product left) (list (string->symbol (car rest)) (apply parse-product left) (apply parse-sum (cadr rest))))) (define (parse-product product left . rest) (if (null? rest) (apply parse-value left) (list (string->symbol (car rest)) (apply parse-value left) (apply parse-product (cadr rest))))) (define (parse-value value first . rest) (if (null? rest) (string->number (cadr first)) (apply parse-sum (car rest)))) (define parse-expr parse-sum)
(Notice all these functions look very similar; for a more complicated PEG, it would be worth abstracting.)
Then:
(apply parse-expr (peg:tree (match-pattern expr "1+1/2*3+(1+1)/2"))) ⇒ (+ 1 (+ (/ 1 (* 2 3)) (/ (+ 1 1) 2)))
But wait! The associativity is wrong! Where it says (/ 1 (* 2
3))
, it should say (* (/ 1 2) 3)
.
It’s tempting to try replacing e.g. "sum <-- (product ('+' / '-')
sum) / product"
with "sum <-- (sum ('+' / '-') product) /
product"
, but this is a Bad Idea. PEGs don’t support left recursion.
To see why, imagine what the parser will do here. When it tries to
parse sum
, it first has to try and parse sum
. But to do
that, it first has to try and parse sum
. This will continue
until the stack gets blown off.
So how does one parse left-associative binary operators with PEGs? Honestly, this is one of their major shortcomings. There’s no general-purpose way of doing this, but here the repetition operators are a good choice:
(use-modules (srfi srfi-1)) (define-peg-string-patterns "expr <- sum sum <-- (product ('+' / '-'))* product product <-- (value ('*' / '/'))* value value <-- number / '(' expr ')' number <-- [0-9]+") ;; take a deep breath... (define (make-left-parser next-func) (lambda (sum first . rest) ;; general form, comments below assume ;; that we're dealing with a sum expression (if (null? rest) ;; form (sum (product ...)) (apply next-func first) (if (string? (cadr first));; form (sum ((product ...) "+") (product ...)) (list (string->symbol (cadr first)) (apply next-func (car first)) (apply next-func (car rest))) ;; form (sum (((product ...) "+") ((product ...) "+")) (product ...)) (car (reduce ;; walk through the list and build a left-associative tree (lambda (l r) (list (list (cadr r) (car r) (apply next-func (car l))) (string->symbol (cadr l)))) 'ignore (append ;; make a list of all the products ;; the first one should be pre-parsed (list (list (apply next-func (caar first)) (string->symbol (cadar first)))) (cdr first) ;; the last one has to be added in (list (append rest '("done")))))))))) (define (parse-value value first . rest) (if (null? rest) (string->number (cadr first)) (apply parse-sum (car rest)))) (define parse-product (make-left-parser parse-value)) (define parse-sum (make-left-parser parse-product)) (define parse-expr parse-sum)
Then:
(apply parse-expr (peg:tree (match-pattern expr "1+1/2*3+(1+1)/2"))) ⇒ (+ (+ 1 (* (/ 1 2) 3)) (/ (+ 1 1) 2))
As you can see, this is much uglier (it could be made prettier by using
context-flatten
, but the way it’s written above makes it clear
how we deal with the three ways the zero-or-more *
expression can
parse). Fortunately, most of the time we can get away with only using
right-associativity.
For a more tantalizing example, consider the following grammar that parses (highly) simplified C functions:
(define-peg-string-patterns "cfunc <-- cSP ctype cSP cname cSP cargs cLB cSP cbody cRB ctype <-- cidentifier cname <-- cidentifier cargs <-- cLP (! (cSP cRP) carg cSP (cCOMMA / cRP) cSP)* cSP carg <-- cSP ctype cSP cname cbody <-- cstatement * cidentifier <- [a-zA-z][a-zA-Z0-9_]* cstatement <-- (!';'.)*cSC cSP cSC < ';' cCOMMA < ',' cLP < '(' cRP < ')' cLB < '{' cRB < '}' cSP < [ \t\n]*")
Then:
(match-pattern cfunc "int square(int a) { return a*a;}") ⇒ (32 (cfunc (ctype "int") (cname "square") (cargs (carg (ctype "int") (cname "a"))) (cbody (cstatement "return a*a"))))
And:
(match-pattern cfunc "int mod(int a, int b) { int c = a/b;return a-b*c; }") ⇒ (52 (cfunc (ctype "int") (cname "mod") (cargs (carg (ctype "int") (cname "a")) (carg (ctype "int") (cname "b"))) (cbody (cstatement "int c = a/b") (cstatement "return a- b*c"))))
By wrapping all the carg
nonterminals in a cargs
nonterminal, we were able to remove any ambiguity in the parsing
structure and avoid having to call context-flatten
on the output
of match-pattern
. We used the same trick with the cstatement
nonterminals, wrapping them in a cbody
nonterminal.
The whitespace nonterminal cSP
used here is a (very) useful
instantiation of a common pattern for matching syntactically irrelevant
information. Since it’s tagged with <
and ends with *
it
won’t clutter up the parse trees (all the empty lists will be discarded
during the compression step) and it will never cause parsing to fail.