You cannot select more than 25 topics
Topics must start with a letter or number, can include dashes ('-') and can be up to 35 characters long.
877 lines
49 KiB
Racket
877 lines
49 KiB
Racket
3 years ago
|
#lang racket/base
|
||
3 years ago
|
;; This module implements a parser form like the yaragg-parser-tools's
|
||
3 years ago
|
;; `parser', except that it works on an arbitrary CFG (returning
|
||
|
;; the first sucecssful parse).
|
||
|
|
||
|
;; I'm pretty sure that this is an implementation of Earley's
|
||
|
;; algorithm.
|
||
|
|
||
|
;; To a first approximation, it's a backtracking parser. Alternative
|
||
|
;; for a non-terminal are computed in parallel, and multiple attempts
|
||
|
;; to compute the same result block until the first one completes. If
|
||
|
;; you get into deadlock, such as when trying to match
|
||
|
;; <foo> := <foo>
|
||
|
;; then it means that there's no successful parse, so everything
|
||
|
;; that's blocked fails.
|
||
|
|
||
|
;; A cache holds the series of results for a particular non-terminal
|
||
|
;; at a particular starting location. (A series is used, instead of a
|
||
|
;; sinlge result, for backtracking.) Otherwise, the parser uses
|
||
|
;; backtracking search. Backtracking is implemented through explicit
|
||
|
;; success and failure continuations. Multiple results for a
|
||
|
;; particular nonterminal and location are kept only when they have
|
||
|
;; different lengths. (Otherwise, in the spirit of finding one
|
||
|
;; successful parse, only the first result is kept.)
|
||
|
|
||
3 years ago
|
;; The yaragg-parser-tools's `parse' is used to transform tokens in the
|
||
3 years ago
|
;; grammar to tokens specific to this parser. In other words, this
|
||
|
;; parser uses `parser' so that it doesn't have to know anything about
|
||
|
;; tokens.
|
||
|
;;
|
||
|
|
||
|
|
||
|
|
||
3 years ago
|
(require yaragg-parser-tools/yacc
|
||
|
yaragg-parser-tools/lex)
|
||
3 years ago
|
|
||
|
(require (for-syntax racket/base
|
||
|
syntax/boundmap
|
||
3 years ago
|
yaragg-parser-tools/private-lex/token-syntax))
|
||
3 years ago
|
|
||
|
(provide cfg-parser)
|
||
|
|
||
|
;; A raw token, wrapped so that we can recognize it:
|
||
|
(define-struct tok (name orig-name val start end))
|
||
|
|
||
|
;; Represents the thread scheduler:
|
||
|
(define-struct tasks (active active-back waits multi-waits cache progress?))
|
||
|
|
||
|
(define-for-syntax make-token-identifier-mapping make-hasheq)
|
||
|
(define-for-syntax (token-identifier-mapping-get t tok [fail #f])
|
||
|
(if fail
|
||
|
(hash-ref t (syntax-e tok) fail)
|
||
|
(hash-ref t (syntax-e tok))))
|
||
|
|
||
|
(define-for-syntax (token-identifier-mapping-put! t tok v)
|
||
|
(hash-set! t (syntax-e tok) v))
|
||
|
|
||
|
(define-for-syntax (token-identifier-mapping-map t f)
|
||
|
(hash-map t f))
|
||
|
|
||
|
;; Used to calculate information on the grammar, such as whether
|
||
|
;; a particular non-terminal is "simple" instead of recursively defined.
|
||
|
(define-for-syntax (nt-fixpoint nts proc nt-ids patss)
|
||
|
(define (ormap-all val f as bs)
|
||
|
(cond
|
||
|
[(null? as) val]
|
||
|
[else (ormap-all (or (f (car as) (car bs)) val)
|
||
|
f
|
||
|
(cdr as) (cdr bs))]))
|
||
|
(let loop ()
|
||
|
(when (ormap-all #f
|
||
|
(λ (nt pats)
|
||
|
(let ([old (bound-identifier-mapping-get nts nt)])
|
||
|
(let ([new (proc nt pats old)])
|
||
|
(if (equal? old new)
|
||
|
#f
|
||
|
(begin
|
||
|
(bound-identifier-mapping-put! nts nt new)
|
||
|
#t)))))
|
||
|
nt-ids patss)
|
||
|
(loop))))
|
||
|
|
||
|
;; Tries parse-a followed by parse-b. If parse-a is not simple,
|
||
|
;; then after parse-a succeeds once, we parallelize parse-b
|
||
|
;; and trying a second result for parse-a.
|
||
|
(define (parse-and simple-a? parse-a parse-b
|
||
|
stream last-consumed-token depth end success-k fail-k
|
||
|
max-depth tasks)
|
||
|
(define ((mk-got-k success-k fail-k) val stream last-consumed-token depth max-depth tasks next1-k)
|
||
|
(if simple-a?
|
||
|
(parse-b val stream last-consumed-token depth end
|
||
|
(mk-got2-k success-k fail-k next1-k)
|
||
|
(mk-fail2-k success-k fail-k next1-k)
|
||
|
max-depth tasks)
|
||
|
(parallel-or
|
||
|
(λ (success-k fail-k max-depth tasks)
|
||
|
(parse-b val stream last-consumed-token depth end
|
||
|
success-k fail-k
|
||
|
max-depth tasks))
|
||
|
(λ (success-k fail-k max-depth tasks)
|
||
|
(next1-k (mk-got-k success-k fail-k)
|
||
|
fail-k max-depth tasks))
|
||
|
success-k fail-k max-depth tasks)))
|
||
|
|
||
|
(define ((mk-got2-k success-k fail-k next1-k) val stream last-consumed-token depth max-depth tasks next-k)
|
||
|
(success-k val stream last-consumed-token depth max-depth tasks
|
||
|
(λ (success-k fail-k max-depth tasks)
|
||
|
(next-k (mk-got2-k success-k fail-k next1-k)
|
||
|
(mk-fail2-k success-k fail-k next1-k)
|
||
|
max-depth tasks))))
|
||
|
|
||
|
(define ((mk-fail2-k success-k fail-k next1-k) max-depth tasks)
|
||
|
(next1-k (mk-got-k success-k fail-k) fail-k max-depth tasks))
|
||
|
|
||
|
(parse-a stream last-consumed-token depth end
|
||
|
(mk-got-k success-k fail-k)
|
||
|
fail-k
|
||
|
max-depth tasks))
|
||
|
|
||
|
;; Parallel or for non-terminal alternatives
|
||
|
(define (parse-parallel-or parse-a parse-b stream last-consumed-token depth end success-k fail-k max-depth tasks)
|
||
|
(parallel-or (λ (success-k fail-k max-depth tasks)
|
||
|
(parse-a stream last-consumed-token depth end success-k fail-k max-depth tasks))
|
||
|
(λ (success-k fail-k max-depth tasks)
|
||
|
(parse-b stream last-consumed-token depth end success-k fail-k max-depth tasks))
|
||
|
success-k fail-k max-depth tasks))
|
||
|
|
||
|
;; Generic parallel-or
|
||
|
(define (parallel-or parse-a parse-b success-k fail-k max-depth tasks)
|
||
|
(define answer-key (gensym))
|
||
|
(define (gota-k val stream last-consumed-token depth max-depth tasks next-k)
|
||
|
(report-answer answer-key
|
||
|
max-depth
|
||
|
tasks
|
||
|
(list val stream last-consumed-token depth next-k)))
|
||
|
(define (faila-k max-depth tasks)
|
||
|
(report-answer answer-key
|
||
|
max-depth
|
||
|
tasks
|
||
|
null))
|
||
|
(let* ([tasks (queue-task tasks (λ (max-depth tasks)
|
||
|
(parse-a gota-k faila-k max-depth tasks)))]
|
||
|
[tasks (queue-task tasks (λ (max-depth tasks)
|
||
|
(parse-b gota-k faila-k max-depth tasks)))]
|
||
|
[queue-next (λ (next-k tasks)
|
||
|
(queue-task tasks (λ (max-depth tasks)
|
||
|
(next-k gota-k faila-k max-depth tasks))))])
|
||
|
(define ((mk-got-one immediate-next? get-nth success-k) val stream last-consumed-token depth max-depth tasks next-k)
|
||
|
(let ([tasks (if immediate-next?
|
||
|
(queue-next next-k tasks)
|
||
|
tasks)])
|
||
|
(success-k val stream last-consumed-token depth max-depth
|
||
|
tasks
|
||
|
(λ (success-k fail-k max-depth tasks)
|
||
|
(let ([tasks (if immediate-next?
|
||
|
tasks
|
||
|
(queue-next next-k tasks))])
|
||
|
(get-nth max-depth tasks success-k fail-k))))))
|
||
|
(define (get-first max-depth tasks success-k fail-k)
|
||
|
(wait-for-answer #f max-depth tasks answer-key
|
||
|
(mk-got-one #t get-first success-k)
|
||
|
(λ (max-depth tasks)
|
||
|
(get-second max-depth tasks success-k fail-k))
|
||
|
#f))
|
||
|
(define (get-second max-depth tasks success-k fail-k)
|
||
|
(wait-for-answer #f max-depth tasks answer-key
|
||
|
(mk-got-one #f get-second success-k)
|
||
|
fail-k #f))
|
||
|
(get-first max-depth tasks success-k fail-k)))
|
||
|
|
||
|
;; Non-terminal alternatives where the first is "simple" can be done
|
||
|
;; sequentially, which is simpler
|
||
|
(define (parse-or parse-a parse-b
|
||
|
stream last-consumed-token depth end success-k fail-k max-depth tasks)
|
||
|
(define ((mk-got-k success-k fail-k) val stream last-consumed-token depth max-depth tasks next-k)
|
||
|
(success-k val stream last-consumed-token depth
|
||
|
max-depth tasks
|
||
|
(λ (success-k fail-k max-depth tasks)
|
||
|
(next-k (mk-got-k success-k fail-k)
|
||
|
(mk-fail-k success-k fail-k)
|
||
|
max-depth tasks))))
|
||
|
(define ((mk-fail-k success-k fail-k) max-depth tasks)
|
||
|
(parse-b stream last-consumed-token depth end success-k fail-k max-depth tasks))
|
||
|
(parse-a stream last-consumed-token depth end
|
||
|
(mk-got-k success-k fail-k)
|
||
|
(mk-fail-k success-k fail-k)
|
||
|
max-depth tasks))
|
||
|
|
||
|
;; Starts a thread
|
||
|
(define (queue-task tasks t [progress? #t])
|
||
|
(make-tasks (tasks-active tasks)
|
||
|
(cons t (tasks-active-back tasks))
|
||
|
(tasks-waits tasks)
|
||
|
(tasks-multi-waits tasks)
|
||
|
(tasks-cache tasks)
|
||
|
(or progress? (tasks-progress? tasks))))
|
||
|
|
||
|
;; Reports an answer to a waiting thread:
|
||
|
(define (report-answer answer-key max-depth tasks val)
|
||
|
(define v (hash-ref (tasks-waits tasks) answer-key (λ () #f)))
|
||
|
(if v
|
||
|
(let ([tasks (make-tasks (cons (v val) (tasks-active tasks))
|
||
|
(tasks-active-back tasks)
|
||
|
(tasks-waits tasks)
|
||
|
(tasks-multi-waits tasks)
|
||
|
(tasks-cache tasks)
|
||
|
#t)])
|
||
|
(hash-remove! (tasks-waits tasks) answer-key)
|
||
|
(swap-task max-depth tasks))
|
||
|
;; We have an answer ready too fast; wait
|
||
|
(swap-task max-depth
|
||
|
(queue-task tasks
|
||
|
(λ (max-depth tasks)
|
||
|
(report-answer answer-key max-depth tasks val))
|
||
|
#f))))
|
||
|
|
||
|
;; Reports an answer to multiple waiting threads:
|
||
|
(define (report-answer-all answer-key max-depth tasks val k)
|
||
|
(define v (hash-ref (tasks-multi-waits tasks) answer-key (λ () null)))
|
||
|
(hash-remove! (tasks-multi-waits tasks) answer-key)
|
||
|
(let ([tasks (make-tasks (append (map (λ (a) (a val)) v)
|
||
|
(tasks-active tasks))
|
||
|
(tasks-active-back tasks)
|
||
|
(tasks-waits tasks)
|
||
|
(tasks-multi-waits tasks)
|
||
|
(tasks-cache tasks)
|
||
|
#t)])
|
||
|
(k max-depth tasks)))
|
||
|
|
||
|
;; Waits for an answer; if `multi?' is #f, this is sole waiter, otherwise
|
||
|
;; there might be many. Use wither #t or #f (and `report-answer' or
|
||
|
;; `report-answer-all', resptively) consistently for a particular answer key.
|
||
|
(define (wait-for-answer multi? max-depth tasks answer-key success-k fail-k deadlock-k)
|
||
|
(let ([wait (λ (val)
|
||
|
(λ (max-depth tasks)
|
||
|
(if val
|
||
|
(if (null? val)
|
||
|
(fail-k max-depth tasks)
|
||
|
(let-values ([(val stream last-consumed-token depth next-k) (apply values val)])
|
||
|
(success-k val stream last-consumed-token depth max-depth tasks next-k)))
|
||
|
(deadlock-k max-depth tasks))))])
|
||
|
(if multi?
|
||
|
(hash-set! (tasks-multi-waits tasks) answer-key
|
||
|
(cons wait (hash-ref (tasks-multi-waits tasks) answer-key
|
||
|
(λ () null))))
|
||
|
(hash-set! (tasks-waits tasks) answer-key wait))
|
||
|
(let ([tasks (make-tasks (tasks-active tasks)
|
||
|
(tasks-active-back tasks)
|
||
|
(tasks-waits tasks)
|
||
|
(tasks-multi-waits tasks)
|
||
|
(tasks-cache tasks)
|
||
|
#t)])
|
||
|
(swap-task max-depth tasks))))
|
||
|
|
||
|
;; Swap thread
|
||
|
(define (swap-task max-depth tasks)
|
||
|
;; Swap in first active:
|
||
|
(if (null? (tasks-active tasks))
|
||
|
(if (tasks-progress? tasks)
|
||
|
(swap-task max-depth
|
||
|
(make-tasks (reverse (tasks-active-back tasks))
|
||
|
null
|
||
|
(tasks-waits tasks)
|
||
|
(tasks-multi-waits tasks)
|
||
|
(tasks-cache tasks)
|
||
|
#f))
|
||
|
;; No progress, so issue failure for all multi-waits
|
||
|
(if (zero? (hash-count (tasks-multi-waits tasks)))
|
||
|
(error 'swap-task "Deadlock")
|
||
|
(swap-task max-depth
|
||
|
(make-tasks (apply
|
||
|
append
|
||
|
(hash-map (tasks-multi-waits tasks)
|
||
|
(λ (k l)
|
||
|
(map (λ (v) (v #f)) l))))
|
||
|
(tasks-active-back tasks)
|
||
|
(tasks-waits tasks)
|
||
|
(make-hasheq)
|
||
|
(tasks-cache tasks)
|
||
|
#t))))
|
||
|
(let ([t (car (tasks-active tasks))]
|
||
|
[tasks (make-tasks (cdr (tasks-active tasks))
|
||
|
(tasks-active-back tasks)
|
||
|
(tasks-waits tasks)
|
||
|
(tasks-multi-waits tasks)
|
||
|
(tasks-cache tasks)
|
||
|
(tasks-progress? tasks))])
|
||
|
(t max-depth tasks))))
|
||
|
|
||
|
;; Finds the symbolic representative of a token class
|
||
|
(define-for-syntax (map-token toks tok)
|
||
|
(car (token-identifier-mapping-get toks tok)))
|
||
|
|
||
|
(define no-pos-val (make-position #f #f #f))
|
||
|
(define-for-syntax no-pos
|
||
|
(let ([npv ((syntax-local-certifier) #'no-pos-val)])
|
||
|
(λ (stx) npv)))
|
||
|
(define-for-syntax ((at-tok-pos sel expr) stx)
|
||
|
#`(let ([v #,expr]) (if v (#,sel v) no-pos-val)))
|
||
|
|
||
|
;; Builds a matcher for a particular alternative
|
||
|
(define-for-syntax (build-match nts toks pat handle $ctx)
|
||
|
(let loop ([pat pat]
|
||
|
[pos 1])
|
||
|
(if (null? pat)
|
||
|
#`(success-k #,handle stream last-consumed-token depth max-depth tasks
|
||
|
(λ (success-k fail-k max-depth tasks)
|
||
|
(fail-k max-depth tasks)))
|
||
|
(let ([id (datum->syntax (car pat) (string->symbol (format "$~a" pos)))]
|
||
|
[id-start-pos (datum->syntax (car pat) (string->symbol (format "$~a-start-pos" pos)))]
|
||
|
[id-end-pos (datum->syntax (car pat) (string->symbol (format "$~a-end-pos" pos)))]
|
||
|
[n-end-pos (and (null? (cdr pat)) (datum->syntax (car pat) '$n-end-pos))])
|
||
|
(cond
|
||
|
[(bound-identifier-mapping-get nts (car pat) (λ () #f))
|
||
|
;; Match non-termimal
|
||
|
#`(parse-and
|
||
|
;; First part is simple? (If so, we don't have to parallelize the `and'.)
|
||
|
#,(let ([l (bound-identifier-mapping-get nts (car pat) (λ () #f))])
|
||
|
(or (not l)
|
||
|
(andmap values (caddr l))))
|
||
|
#,(car pat)
|
||
|
(let ([original-stream stream])
|
||
|
(λ (#,id stream last-consumed-token depth end success-k fail-k max-depth tasks)
|
||
|
(let-syntax ([#,id-start-pos (at-tok-pos #'(if (eq? original-stream stream)
|
||
|
tok-end
|
||
|
tok-start)
|
||
|
#'(if (eq? original-stream stream)
|
||
|
last-consumed-token
|
||
|
(and (pair? original-stream)
|
||
|
(car original-stream))))]
|
||
|
[#,id-end-pos (at-tok-pos #'tok-end #'last-consumed-token)]
|
||
|
#,@(if n-end-pos
|
||
|
#`([#,n-end-pos (at-tok-pos #'tok-end #'last-consumed-token)])
|
||
|
null))
|
||
|
#,(loop (cdr pat) (add1 pos)))))
|
||
|
stream last-consumed-token depth
|
||
|
#,(let ([cnt (apply +
|
||
|
(map (λ (item)
|
||
|
(cond
|
||
|
[(bound-identifier-mapping-get nts item (λ () #f))
|
||
|
=> (λ (l) (car l))]
|
||
|
[else 1]))
|
||
|
(cdr pat)))])
|
||
|
#`(- end #,cnt))
|
||
|
success-k fail-k max-depth tasks)]
|
||
|
[else
|
||
|
;; Match token
|
||
|
(let ([tok-id (map-token toks (car pat))])
|
||
|
#`(if (and (pair? stream)
|
||
|
(eq? '#,tok-id (tok-name (car stream))))
|
||
|
(let* ([stream-a (car stream)]
|
||
|
[#,id (tok-val stream-a)]
|
||
|
[last-consumed-token (car stream)]
|
||
|
[stream (cdr stream)]
|
||
|
[depth (add1 depth)])
|
||
|
(let ([max-depth (max max-depth depth)])
|
||
|
(let-syntax ([#,id-start-pos (at-tok-pos #'tok-start #'stream-a)]
|
||
|
[#,id-end-pos (at-tok-pos #'tok-end #'stream-a)]
|
||
|
#,@(if n-end-pos
|
||
|
#`([#,n-end-pos (at-tok-pos #'tok-end #'stream-a)])
|
||
|
null))
|
||
|
#,(loop (cdr pat) (add1 pos)))))
|
||
|
(fail-k max-depth tasks)))])))))
|
||
|
|
||
|
;; Starts parsing to match a non-terminal. There's a minor
|
||
|
;; optimization that checks for known starting tokens. Otherwise,
|
||
|
;; use the cache, block if someone else is already trying the match,
|
||
|
;; and cache the result if it's computed.
|
||
|
;; The cache maps nontermial+startingpos+iteration to a result, where
|
||
|
;; the iteration is 0 for the first match attempt, 1 for the second,
|
||
|
;; etc.
|
||
|
(define (parse-nt/share key min-cnt init-tokens stream last-consumed-token depth end max-depth tasks success-k fail-k k)
|
||
|
(if (and (positive? min-cnt)
|
||
|
(pair? stream)
|
||
|
(not (memq (tok-name (car stream)) init-tokens)))
|
||
|
;; No such leading token; give up
|
||
|
(fail-k max-depth tasks)
|
||
|
;; Run pattern
|
||
|
(let loop ([n 0]
|
||
|
[success-k success-k]
|
||
|
[fail-k fail-k]
|
||
|
[max-depth max-depth]
|
||
|
[tasks tasks]
|
||
|
[k k])
|
||
|
(define answer-key (gensym))
|
||
|
(define table-key (vector key depth n))
|
||
|
(define old-depth depth)
|
||
|
(define old-stream stream)
|
||
|
#;(printf "Loop ~a\n" table-key)
|
||
|
(cond
|
||
|
[(hash-ref (tasks-cache tasks) table-key (λ () #f))
|
||
|
=> (λ (result)
|
||
|
#;(printf "Reuse ~a\n" table-key)
|
||
|
(result success-k fail-k max-depth tasks))]
|
||
|
[else
|
||
|
#;(printf "Try ~a ~a\n" table-key (map tok-name stream))
|
||
|
(hash-set! (tasks-cache tasks) table-key
|
||
|
(λ (success-k fail-k max-depth tasks)
|
||
|
#;(printf "Wait ~a ~a\n" table-key answer-key)
|
||
|
(wait-for-answer #t max-depth tasks answer-key success-k fail-k
|
||
|
(λ (max-depth tasks)
|
||
|
#;(printf "Deadlock ~a ~a\n" table-key answer-key)
|
||
|
(fail-k max-depth tasks)))))
|
||
|
(let result-loop ([max-depth max-depth][tasks tasks][k k])
|
||
|
(define orig-stream stream)
|
||
|
(define (new-got-k val stream last-consumed-token depth max-depth tasks next-k)
|
||
|
;; Check whether we already have a result that consumed the same amount:
|
||
|
(define result-key (vector #f key old-depth depth))
|
||
|
(cond
|
||
|
[(hash-ref (tasks-cache tasks) result-key (λ () #f))
|
||
|
;; Go for the next-result
|
||
|
(result-loop max-depth
|
||
|
tasks
|
||
|
(λ (end max-depth tasks success-k fail-k)
|
||
|
(next-k success-k fail-k max-depth tasks)))]
|
||
|
[else
|
||
|
#;(printf "Success ~a ~a\n" table-key
|
||
|
(map tok-name (let loop ([d old-depth][s old-stream])
|
||
|
(if (= d depth)
|
||
|
null
|
||
|
(cons (car s) (loop (add1 d) (cdr s)))))))
|
||
|
(let ([next-k (λ (success-k fail-k max-depth tasks)
|
||
|
(loop (add1 n)
|
||
|
success-k
|
||
|
fail-k
|
||
|
max-depth
|
||
|
tasks
|
||
|
(λ (end max-depth tasks success-k fail-k)
|
||
|
(next-k success-k fail-k max-depth tasks))))])
|
||
|
(hash-set! (tasks-cache tasks) result-key #t)
|
||
|
(hash-set! (tasks-cache tasks) table-key
|
||
|
(λ (success-k fail-k max-depth tasks)
|
||
|
(success-k val stream last-consumed-token depth max-depth tasks next-k)))
|
||
|
(report-answer-all answer-key
|
||
|
max-depth
|
||
|
tasks
|
||
|
(list val stream last-consumed-token depth next-k)
|
||
|
(λ (max-depth tasks)
|
||
|
(success-k val stream last-consumed-token depth max-depth tasks next-k))))]))
|
||
|
(define (new-fail-k max-depth tasks)
|
||
|
#;(printf "Failure ~a\n" table-key)
|
||
|
(hash-set! (tasks-cache tasks) table-key
|
||
|
(λ (success-k fail-k max-depth tasks)
|
||
|
(fail-k max-depth tasks)))
|
||
|
(report-answer-all answer-key
|
||
|
max-depth
|
||
|
tasks
|
||
|
null
|
||
|
(λ (max-depth tasks)
|
||
|
(fail-k max-depth tasks))))
|
||
|
(k end max-depth tasks new-got-k new-fail-k))]))))
|
||
|
|
||
|
;; These temp identifiers can't be `gensym` or `generate-temporary`
|
||
|
;; because they have to be consistent between module loads
|
||
|
;; (IIUC, the parser is multi-threaded, and this approach is not thread-safe)
|
||
|
;; so I see no alternative to the old standby of making them ludicrously unlikely
|
||
|
(define-for-syntax start-id-temp 'start_jihqolbbafscgxvsufnepvmxqipnxgmlpxukmdoqxqzmzgaogaftbkbyqjttwwfimifowdxfyekjiixdmtprfkcvfciraehoeuaz)
|
||
|
(define-for-syntax atok-id-temp 'atok_wrutdjgecmybyfipiwsgjlvsveryodlgassuzcargiuznzgdghrykfqfbwcjgzdhdoeqxcucmtjkuyucskzethozhqkasphdwbht)
|
||
|
(define-syntax (cfg-parser stx)
|
||
|
(syntax-case stx ()
|
||
|
[(_ CLAUSE ...)
|
||
|
(let ([clauses (syntax->list #'(CLAUSE ...))])
|
||
|
(let-values ([(start grammar cfg-error parser-clauses src-pos?)
|
||
|
(let ([all-toks (apply
|
||
|
append
|
||
|
(for/list ([clause (in-list clauses)])
|
||
|
(syntax-case clause (tokens)
|
||
|
[(tokens T ...)
|
||
|
(apply
|
||
|
append
|
||
|
(for/list ([t (in-list (syntax->list #'(T ...)))])
|
||
|
(define v (syntax-local-value t (λ () #f)))
|
||
|
(cond
|
||
|
[(terminals-def? v)
|
||
|
(for/list ([v (in-list (syntax->list (terminals-def-t v)))])
|
||
|
(cons v #f))]
|
||
|
[(e-terminals-def? v)
|
||
|
(for/list ([v (in-list (syntax->list (e-terminals-def-t v)))])
|
||
|
(cons v #t))]
|
||
|
[else null])))]
|
||
|
[_else null])))]
|
||
|
[all-end-toks (apply
|
||
|
append
|
||
|
(for/list ([clause (in-list clauses)])
|
||
|
(syntax-case clause (end)
|
||
|
[(end T ...)
|
||
|
(syntax->list #'(T ...))]
|
||
|
[_else null])))])
|
||
|
(let loop ([clauses clauses]
|
||
|
[cfg-start #f]
|
||
|
[cfg-grammar #f]
|
||
|
[cfg-error #f]
|
||
|
[src-pos? #f]
|
||
|
[parser-clauses null])
|
||
|
(if (null? clauses)
|
||
|
(values cfg-start
|
||
|
cfg-grammar
|
||
|
cfg-error
|
||
|
(reverse parser-clauses)
|
||
|
src-pos?)
|
||
|
(syntax-case (car clauses) (start error grammar src-pos)
|
||
|
[(start TOK)
|
||
|
(loop (cdr clauses) #'TOK cfg-grammar cfg-error src-pos? parser-clauses)]
|
||
|
[(error EXPR)
|
||
|
(loop (cdr clauses) cfg-start cfg-grammar #'EXPR src-pos? parser-clauses)]
|
||
|
[(grammar [NT [PAT HANDLE0 HANDLE ...] ...] ...)
|
||
|
(let ([nts (make-bound-identifier-mapping)]
|
||
|
[toks (make-token-identifier-mapping)]
|
||
|
[end-toks (make-token-identifier-mapping)]
|
||
|
[nt-ids (syntax->list #'(NT ...))]
|
||
|
[patss (map (λ (stx)
|
||
|
(map syntax->list (syntax->list stx)))
|
||
|
(syntax->list #'((PAT ...) ...)))])
|
||
|
(for ([nt (in-list nt-ids)])
|
||
|
(bound-identifier-mapping-put! nts nt (list 0)))
|
||
|
(for ([t (in-list all-end-toks)])
|
||
|
(token-identifier-mapping-put! end-toks t #t))
|
||
|
(for ([t (in-list all-toks)]
|
||
|
#:unless (token-identifier-mapping-get end-toks (car t) (λ () #f)))
|
||
|
(define id (gensym (syntax-e (car t))))
|
||
|
(token-identifier-mapping-put! toks (car t) (cons id (cdr t))))
|
||
|
;; Compute min max size for each non-term:
|
||
|
(nt-fixpoint
|
||
|
nts
|
||
|
(λ (nt pats old-list)
|
||
|
(let ([new-cnt
|
||
|
(apply min (for/list ([pat (in-list pats)])
|
||
|
(for/sum ([elem (in-list pat)])
|
||
|
(car (bound-identifier-mapping-get
|
||
|
nts elem (λ () (list 1)))))))])
|
||
|
(if (new-cnt . > . (car old-list))
|
||
|
(cons new-cnt (cdr old-list))
|
||
|
old-list)))
|
||
|
nt-ids patss)
|
||
|
;; Compute set of toks that must appear at the beginning
|
||
|
;; for a non-terminal
|
||
|
(nt-fixpoint
|
||
|
nts
|
||
|
(λ (nt pats old-list)
|
||
|
(let ([new-list
|
||
|
(apply
|
||
|
append
|
||
|
(for/list ([pat (in-list pats)])
|
||
|
(let loop ([pat pat])
|
||
|
(if (pair? pat)
|
||
|
(let ([l (bound-identifier-mapping-get
|
||
|
nts
|
||
|
(car pat)
|
||
|
(λ ()
|
||
|
(list 1 (map-token toks (car pat)))))])
|
||
|
;; If the non-terminal can match 0 things,
|
||
|
;; then it might match something from the
|
||
|
;; next pattern element. Otherwise, it must
|
||
|
;; match the first element:
|
||
|
(if (zero? (car l))
|
||
|
(append (cdr l) (loop (cdr pat)))
|
||
|
(cdr l)))
|
||
|
null))))])
|
||
|
(let ([new (filter (λ (id)
|
||
|
(andmap (λ (id2)
|
||
|
(not (eq? id id2)))
|
||
|
(cdr old-list)))
|
||
|
new-list)])
|
||
|
(if (pair? new)
|
||
|
;; Drop dups in new list:
|
||
|
(let ([new (let loop ([new new])
|
||
|
(if (null? (cdr new))
|
||
|
new
|
||
|
(if (ormap (λ (id)
|
||
|
(eq? (car new) id))
|
||
|
(cdr new))
|
||
|
(loop (cdr new))
|
||
|
(cons (car new) (loop (cdr new))))))])
|
||
|
(cons (car old-list) (append new (cdr old-list))))
|
||
|
old-list))))
|
||
|
nt-ids patss)
|
||
|
;; Determine left-recursive clauses:
|
||
|
(for-each (λ (nt pats)
|
||
|
(let ([l (bound-identifier-mapping-get nts nt)])
|
||
|
(bound-identifier-mapping-put! nts nt (list (car l)
|
||
|
(cdr l)
|
||
|
(map (λ (x) #f) pats)))))
|
||
|
nt-ids patss)
|
||
|
(nt-fixpoint
|
||
|
nts
|
||
|
(λ (nt pats old-list)
|
||
|
(list (car old-list)
|
||
|
(cadr old-list)
|
||
|
(map (λ (pat simple?)
|
||
|
(or simple?
|
||
|
(let ([l (map (λ (elem)
|
||
|
(bound-identifier-mapping-get
|
||
|
nts
|
||
|
elem
|
||
|
(λ () #f)))
|
||
|
pat)])
|
||
|
(andmap (λ (i)
|
||
|
(or (not i)
|
||
|
(andmap values (caddr i))))
|
||
|
l))))
|
||
|
pats (caddr old-list))))
|
||
|
nt-ids patss)
|
||
|
;; Build a definition for each non-term:
|
||
|
(loop (cdr clauses)
|
||
|
cfg-start
|
||
|
(map (λ (nt pats handles $ctxs)
|
||
|
(define info (bound-identifier-mapping-get nts nt))
|
||
|
(list nt
|
||
|
#`(let ([key (gensym '#,nt)])
|
||
|
(λ (stream last-consumed-token depth end success-k fail-k max-depth tasks)
|
||
|
(parse-nt/share
|
||
|
key #,(car info) '#,(cadr info) stream last-consumed-token depth end
|
||
|
max-depth tasks
|
||
|
success-k fail-k
|
||
|
(λ (end max-depth tasks success-k fail-k)
|
||
|
#,(let loop ([pats pats]
|
||
|
[handles (syntax->list handles)]
|
||
|
[$ctxs (syntax->list $ctxs)]
|
||
|
[simple?s (caddr info)])
|
||
|
(if (null? pats)
|
||
|
#'(fail-k max-depth tasks)
|
||
|
#`(#,(if (or (null? (cdr pats))
|
||
|
(car simple?s))
|
||
|
#'parse-or
|
||
|
#'parse-parallel-or)
|
||
|
(λ (stream last-consumed-token depth end success-k fail-k max-depth tasks)
|
||
|
#,(build-match nts
|
||
|
toks
|
||
|
(car pats)
|
||
|
(car handles)
|
||
|
(car $ctxs)))
|
||
|
(λ (stream last-consumed-token depth end success-k fail-k max-depth tasks)
|
||
|
#,(loop (cdr pats)
|
||
|
(cdr handles)
|
||
|
(cdr $ctxs)
|
||
|
(cdr simple?s)))
|
||
|
stream last-consumed-token depth end success-k fail-k max-depth tasks)))))))))
|
||
|
nt-ids
|
||
|
patss
|
||
|
(syntax->list #'(((begin HANDLE0 HANDLE ...) ...) ...))
|
||
|
(syntax->list #'((HANDLE0 ...) ...)))
|
||
|
cfg-error
|
||
|
src-pos?
|
||
|
(list*
|
||
|
(with-syntax ([((tok tok-id . $e) ...)
|
||
|
(token-identifier-mapping-map toks
|
||
|
(λ (k v)
|
||
|
(list* k
|
||
|
(car v)
|
||
|
(if (cdr v)
|
||
|
#f
|
||
|
'$1))))]
|
||
|
[(pos ...)
|
||
|
(if src-pos?
|
||
|
#'($1-start-pos $1-end-pos)
|
||
|
#'(#f #f))]
|
||
|
;; rename `start` and `atok` to temp ids
|
||
|
;; so that "start" and "atok" can be used as literal string tokens in a grammar.
|
||
|
;; not sure why this works, but it passes all tests.
|
||
|
[%start start-id-temp]
|
||
|
[%atok atok-id-temp])
|
||
|
#`(grammar (%start [() null]
|
||
|
[(%atok %start) (cons $1 $2)])
|
||
|
(%atok [(tok) (make-tok 'tok-id 'tok $e pos ...)] ...)))
|
||
|
(with-syntax ([%start start-id-temp])
|
||
|
#`(start %start))
|
||
|
parser-clauses)))]
|
||
|
[(grammar . _)
|
||
|
(raise-syntax-error
|
||
|
#f
|
||
|
"bad grammar clause"
|
||
|
stx
|
||
|
(car clauses))]
|
||
|
[(src-pos)
|
||
|
(loop (cdr clauses)
|
||
|
cfg-start
|
||
|
cfg-grammar
|
||
|
cfg-error
|
||
|
#t
|
||
|
(cons (car clauses) parser-clauses))]
|
||
|
[_else
|
||
|
(loop (cdr clauses)
|
||
|
cfg-start
|
||
|
cfg-grammar
|
||
|
cfg-error
|
||
|
src-pos?
|
||
|
(cons (car clauses) parser-clauses))]))))])
|
||
|
#`(let ([orig-parse (parser
|
||
|
[error (λ (a b c)
|
||
|
(error 'cfg-parser "unexpected ~a token: ~a" b c))]
|
||
|
. #,parser-clauses)]
|
||
|
[error-proc #,cfg-error])
|
||
|
(letrec #,grammar
|
||
|
(λ (get-tok)
|
||
|
(let ([tok-list (orig-parse get-tok)])
|
||
|
(letrec ([success-k
|
||
|
(λ (val stream last-consumed-token depth max-depth tasks next)
|
||
|
(if (null? stream)
|
||
|
val
|
||
|
(next success-k fail-k max-depth tasks)))]
|
||
|
[fail-k (λ (max-depth tasks)
|
||
|
(cond
|
||
|
[(null? tok-list)
|
||
|
(if error-proc
|
||
|
(error-proc #t
|
||
|
'no-tokens
|
||
|
#f
|
||
|
(make-position #f #f #f)
|
||
|
(make-position #f #f #f))
|
||
|
(error
|
||
|
'cfg-parse
|
||
|
"no tokens"))]
|
||
|
[else
|
||
|
(let ([bad-tok (list-ref tok-list
|
||
|
(min (sub1 (length tok-list))
|
||
|
max-depth))])
|
||
|
(if error-proc
|
||
|
(error-proc #t
|
||
|
(tok-orig-name bad-tok)
|
||
|
(tok-val bad-tok)
|
||
|
(tok-start bad-tok)
|
||
|
(tok-end bad-tok))
|
||
|
(error
|
||
|
'cfg-parse
|
||
|
"failed at ~a"
|
||
|
(tok-val bad-tok))))]))])
|
||
|
(#,start tok-list
|
||
|
;; we simulate a token at the very beginning with zero width
|
||
|
;; for use with the position-generating code (*-start-pos, *-end-pos).
|
||
|
(if (null? tok-list)
|
||
|
(tok #f #f #f
|
||
|
(position 1
|
||
|
#,(if src-pos? #'1 #'#f)
|
||
|
#,(if src-pos? #'0 #'#f))
|
||
|
(position 1
|
||
|
#,(if src-pos? #'1 #'#f)
|
||
|
#,(if src-pos? #'0 #'#f)))
|
||
|
(tok (tok-name (car tok-list))
|
||
|
(tok-orig-name (car tok-list))
|
||
|
(tok-val (car tok-list))
|
||
|
(tok-start (car tok-list))
|
||
|
(tok-start (car tok-list))))
|
||
|
0
|
||
|
(length tok-list)
|
||
|
success-k
|
||
|
fail-k
|
||
|
0
|
||
|
(make-tasks null null
|
||
|
(make-hasheq) (make-hasheq)
|
||
|
(make-hash) #t)))))))))]))
|
||
|
|
||
|
|
||
|
(module* test racket/base
|
||
|
(require (submod "..")
|
||
3 years ago
|
yaragg-parser-tools/lex
|
||
3 years ago
|
racket/block
|
||
|
rackunit)
|
||
|
|
||
|
;; Test: parsing regular expressions.
|
||
|
;; Here is a test case on locations:
|
||
|
(block
|
||
|
(define-tokens regexp-tokens (ANCHOR STAR OR LIT LPAREN RPAREN EOF))
|
||
|
(define lex (lexer-src-pos ["|" (token-OR lexeme)]
|
||
|
["^" (token-ANCHOR lexeme)]
|
||
|
["*" (token-STAR lexeme)]
|
||
|
[(repetition 1 +inf.0 alphabetic) (token-LIT lexeme)]
|
||
|
["(" (token-LPAREN lexeme)]
|
||
|
[")" (token-RPAREN lexeme)]
|
||
|
[whitespace (return-without-pos (lex input-port))]
|
||
|
[(eof) (token-EOF 'eof)]))
|
||
|
(define -parse (cfg-parser
|
||
|
(tokens regexp-tokens)
|
||
|
(start top)
|
||
|
(end EOF)
|
||
|
(src-pos)
|
||
|
(grammar [top [(maybe-anchor regexp)
|
||
|
(cond [$1
|
||
|
`(anchored ,$2 ,(pos->sexp $1-start-pos) ,(pos->sexp $2-end-pos))]
|
||
|
[else
|
||
|
`(unanchored ,$2 ,(pos->sexp $1-start-pos) ,(pos->sexp $2-end-pos))])]]
|
||
|
[maybe-anchor [(ANCHOR) #t]
|
||
|
[() #f]]
|
||
|
[regexp [(regexp STAR) `(star ,$1 ,(pos->sexp $1-start-pos) ,(pos->sexp $2-end-pos))]
|
||
|
[(regexp OR regexp) `(or ,$1 ,$3 ,(pos->sexp $1-start-pos) ,(pos->sexp $3-end-pos))]
|
||
|
[(LPAREN regexp RPAREN) `(group ,$2 ,(pos->sexp $1-start-pos) ,(pos->sexp $3-end-pos))]
|
||
|
[(LIT) `(lit ,$1 ,(pos->sexp $1-start-pos) ,(pos->sexp $1-end-pos))]])))
|
||
|
(define (pos->sexp pos)
|
||
|
(position-offset pos))
|
||
|
|
||
|
(define (parse s)
|
||
|
(define ip (open-input-string s))
|
||
|
(port-count-lines! ip)
|
||
|
(-parse (λ () (lex ip))))
|
||
|
|
||
|
(check-equal? (parse "abc")
|
||
|
'(unanchored (lit "abc" 1 4) 1 4))
|
||
|
(check-equal? (parse "a | (b*) | c")
|
||
|
'(unanchored (or (or (lit "a" 1 2)
|
||
|
(group (star (lit "b" 6 7) 6 8) 5 9)
|
||
|
1 9)
|
||
|
(lit "c" 12 13)
|
||
|
1 13)
|
||
|
1 13)))
|
||
|
|
||
|
|
||
|
|
||
|
|
||
|
|
||
|
;; Tests used during development
|
||
|
(define-tokens non-terminals (PLUS MINUS STAR BAR COLON EOF))
|
||
|
|
||
|
(define lex
|
||
|
(lexer
|
||
|
["+" (token-PLUS '+)]
|
||
|
["-" (token-MINUS '-)]
|
||
|
["*" (token-STAR '*)]
|
||
|
["|" (token-BAR '||)]
|
||
|
[":" (token-COLON '|:|)]
|
||
|
[whitespace (lex input-port)]
|
||
|
[(eof) (token-EOF 'eof)]))
|
||
|
|
||
|
(define parse
|
||
|
(cfg-parser
|
||
|
(tokens non-terminals)
|
||
|
(start <program>)
|
||
|
(end EOF)
|
||
|
(error (λ (a b stx)
|
||
|
(error 'parse "failed at ~s" stx)))
|
||
|
(grammar [<program> [(PLUS) "plus"]
|
||
|
[(<minus-program> BAR <minus-program>) (list $1 $2 $3)]
|
||
|
[(<program> COLON) (list $1)]]
|
||
|
[<minus-program> [(MINUS) "minus"]
|
||
|
[(<program> STAR) (cons $1 $2)]]
|
||
|
[<simple> [(<alts> <alts> <alts> MINUS) "yes"]]
|
||
|
[<alts> [(PLUS) 'plus]
|
||
|
[(MINUS) 'minus]]
|
||
|
[<random> [() '0]
|
||
|
[(<random> PLUS) (add1 $1)]
|
||
|
[(<random> PLUS) (add1 $1)]])))
|
||
|
|
||
|
(let ([p (open-input-string #;"+*|-|-*|+**" #;"-|+*|+**"
|
||
|
#;"+*|+**|-" #;"-|-*|-|-*"
|
||
|
#;"-|-*|-|-**|-|-*|-|-**"
|
||
|
"-|-*|-|-**|-|-*|-|-***|-|-*|-|-**|-|-*|-|-****|-|-*|-|-**|-|-*|-|-***
|
||
|
|-|-*|-|-**|-|-*|-|-*****|-|-*|-|-**|-|-*|-|-***|-|-*|-|-**|-|-*|-|-****|
|
||
|
-|-*|-|-**|-|-*|-|-***|-|-*|-|-**|-|-*|-|-*****"
|
||
|
;; This one fails:
|
||
|
#;"+*")])
|
||
|
(check-equal? (parse (λ () (lex p)))
|
||
|
'((((((((((("minus" || "minus") . *) || (("minus" || "minus") . *)) . *) || (((("minus" || "minus") . *) || (("minus" || "minus") . *)) . *)) . *)
|
||
|
||
|
||
|
(((((("minus" || "minus") . *) || (("minus" || "minus") . *)) . *) || (((("minus" || "minus") . *) || (("minus" || "minus") . *)) . *)) . *))
|
||
|
.
|
||
|
*)
|
||
|
||
|
||
|
(((((((("minus" || "minus") . *) || (("minus" || "minus") . *)) . *) || (((("minus" || "minus") . *) || (("minus" || "minus") . *)) . *)) . *)
|
||
|
||
|
||
|
(((((("minus" || "minus") . *) || (("minus" || "minus") . *)) . *) || (((("minus" || "minus") . *) || (("minus" || "minus") . *)) . *)) . *))
|
||
|
.
|
||
|
*))
|
||
|
.
|
||
|
*)
|
||
|
||
|
||
|
(((((((((("minus" || "minus") . *) || (("minus" || "minus") . *)) . *) || (((("minus" || "minus") . *) || (("minus" || "minus") . *)) . *)) . *)
|
||
|
||
|
||
|
(((((("minus" || "minus") . *) || (("minus" || "minus") . *)) . *) || (((("minus" || "minus") . *) || (("minus" || "minus") . *)) . *)) . *))
|
||
|
.
|
||
|
*)
|
||
|
||
|
||
|
(((((((("minus" || "minus") . *) || (("minus" || "minus") . *)) . *) || (((("minus" || "minus") . *) || (("minus" || "minus") . *)) . *)) . *)
|
||
|
||
|
||
|
(((((("minus" || "minus") . *) || (("minus" || "minus") . *)) . *) || (((("minus" || "minus") . *) || (("minus" || "minus") . *)) . *)) . *))
|
||
|
.
|
||
|
*))
|
||
|
.
|
||
|
*)))))
|