https://github.com/JuliaLang/julia
Raw File
Tip revision: 3b4e7b46ccf602e004905273ebebf85c47e220f7 authored by Keno Fischer on 10 October 2018, 00:26:12 UTC
HACK: Option to ignore all inlining heuristics
Tip revision: 3b4e7b4
match.scm
; tree regular expression pattern matching
; by Jeff Bezanson

; list of special pattern symbols that cannot be variable names
(define metasymbols '(_ ...))

; expression tree pattern matching
; matches expr against pattern p and returns an assoc list
;    ((var . expr) (var . expr) ...)
; mapping variables to captured subexpressions, or #f if no match.
; when a match succeeds, __ is always bound to the whole matched expression.
;
; p is an expression in the following pattern language:
;
; _       match anything, not captured
; <func>  any scheme function; matches if (func expr) returns #t
; <var>   match anything and capture as <var>.
;         future occurrences of <var> in the pattern must match the same thing.
; (head <p1> <p2> etc)   match an s-expr with 'head' matched literally,
;                        and the rest of the subpatterns matched recursively.
; (-/ <ex>)  match <ex> literally
; (-^ <p>)   complement of pattern <p>
; (-- <var> <p>)  match <p> and capture as <var> if match succeeds
; (-s)       match any symbol
; (<pat...> . <var>) match prefix and bind tail to <var>
;
; regular match constructs:
; ...                 match any number of anything
; (-$ <p1> <p2> etc)  match any of subpatterns <p1>, <p2>, etc
; (-* <p>)            match any number of <p>
; (-? <p>)            match 0 or 1 of <p>
; (-+ <p>)            match at least 1 of <p>
; all of these can be wrapped in (-- var   ) for capturing purposes
; This is NP-complete. Be careful.
;
(define (match- p expr state)
  (cond ((symbol? p)
         (cond ((eq? p '_) state)
               (else
                (let ((capt (assq p state)))
                  (if capt
                      (and (equal? expr (cdr capt)) state)
                      (cons (cons p expr) state))))))

        ((procedure? p)
         (and (p expr) state))

        ((pair? p)
         (cond ((eq? (car p) '-/)
                (and (equal? (cadr p) expr)             state))
               ((eq? (car p) '-^)
                (and (not (match- (cadr p) expr state)) state))
               ((eq? (car p) '--)
                (and (match- (caddr p) expr state)
                     (cons (cons (cadr p) expr) state)))
               ((eq? (car p) '-$)  ; greedy alternation for toplevel pattern
                (match-alt (cdr p) '() (list expr) state #f 1))
               ((eq? (car p) '-s)
                (and (symbol? expr) state))
               (else
                (and (pair? expr)
                     (equal? (car p) (car expr))
                     (match-seq (cdr p) (cdr expr) state (length
                                                          (cdr expr)))))))

        (else
         (and (equal? p expr) state))))

; match an alternation
(define (match-alt alt prest expr state var L)
  (if (null? alt) #f  ; no alternatives left
      (let ((subma (match- (car alt) (car expr) state)))
        (or (and subma
                 (match-seq prest (cdr expr)
                            (if var
                                (cons (cons var (car expr))
                                      subma)
                                subma)
                            (- L 1)))
            (match-alt (cdr alt) prest expr state var L)))))

; match generalized kleene star (try consuming min to max)
(define (match-star- p prest expr state var min max L sofar)
  (cond ; case 0: impossible to match
   ((> min max) #f)
   ; case 1: only allowed to match 0 subexpressions
   ((= max 0) (match-seq prest expr
                         (if var (cons (cons var (reverse sofar)) state)
                             state)
                         L))
   ; case 2: must match at least 1
   ((> min 0)
    (let ((subma (match- p (car expr) state)))
      (and subma
           (match-star- p prest (cdr expr) subma var (- min 1) (- max 1) (- L 1)
                        (cons (car expr) sofar)))))
   ; otherwise, must match either 0 or between 1 and max subexpressions
   (else
    (or (match-star- p prest expr state var 0 0   L sofar)
        (match-star- p prest expr state var 1 max L sofar)))))
(define (match-star p prest expr state var min max L)
  (match-star- p prest expr state var min max L '()))

; match sequences of expressions
(define (match-seq p expr state L)
  (cond ((not state) #f)
        ((symbol? p) (cons (cons p expr) state))
        ((null? p) (if (null? expr) state #f))
        (else
         (let ((capt? (and (pair? (car p))
                           (eq? (car (car p)) '--))))
           (let ((subp (if capt? (caddr (car p)) (car p)))
                 (var  (if capt? (cadr  (car p)) #f)))
             (let ((head (if (pair? subp) (car subp) '())))
               (cond ((eq? subp '...)
                      (match-star '_ (cdr p) expr state var 0 L L))
                     ((eq? head '-*)
                      (match-star (cadr subp) (cdr p) expr state var 0 L L))
                     ((eq? head '-+)
                      (match-star (cadr subp) (cdr p) expr state var 1 L L))
                     ((eq? head '-?)
                      (match-star (cadr subp) (cdr p) expr state var 0 1 L))
                     ((eq? head '-$)
                      (match-alt (cdr subp) (cdr p) expr state var L))
                     (else
                      (and (pair? expr)
                           (match-seq (cdr p) (cdr expr)
                                      (match- (car p) (car expr) state)
                                      (- L 1)))))))))))

(define (match p expr) (match- p expr (list (cons '__ expr))))

; try to transform expr using a pattern-lambda from plist
; returns the new expression, or expr if no matches
(define (apply-patterns plist expr)
  (cond ((vector? plist)
         (if (pair? expr)
             (let* ((relevant (table-ref (vector-ref plist 1) (car expr) '()))
                    (enew     (apply-patterns relevant expr)))
               (if (eq? enew expr)
                   (apply-patterns (vector-ref plist 2) expr)
                   enew))
             (apply-patterns (vector-ref plist 2) expr)))
        ((null? plist) expr)
        (else
         (let ((enew ((car plist) expr)))
           (if (not enew)
               (apply-patterns (cdr plist) expr)
               enew)))))

; top-down fixed-point macroexpansion. this is a typical algorithm,
; but it may leave some structure that matches a pattern unexpanded.
; the advantage is that non-terminating cases cannot arise as a result
; of expression composition. in other words, if the outer loop terminates
; on all inputs for a given set of patterns, then the whole algorithm
; terminates. pattern sets that violate this should be easier to detect,
; for example
; (pattern-lambda (/ 2 3) '(/ 3 2)), (pattern-lambda (/ 3 2) '(/ 2 3))
(define (pattern-expand plist expr)
  (if (or (not (pair? expr)) (memq (car expr) '(quote varlist inert)))
      expr
      (let ((enew (apply-patterns plist expr)))
        (if (eq? enew expr)
            ;; expr didn't change; move to subexpressions
            (let ((sub (lambda (subex)
                         (if (not (pair? subex))
                             subex
                             (pattern-expand plist subex)))))
              (if (eq? (car expr) 'lambda)
                  (list* 'lambda
                         (map sub (cadr expr))
                         (map sub (cddr expr)))
                  (map sub expr)))
            ;; expr changed; iterate
            (pattern-expand plist enew)))))

;; expand only outermost
(define (pattern-expand1 plist expr)
  (if (or (not (pair? expr)) (memq (car expr) '(quote inert)))
      expr
      (let ((enew (apply-patterns plist expr)))
        (if (eq? enew expr)
            expr
            ;; expr changed; iterate
            (pattern-expand plist enew)))))

;; finds and replaces pattern matches with their expansions
;; one pass, does not expand recursively
(define (pattern-replace plist expr)
  (if (or (not (pair? expr)) (memq (car expr) '(quote inert)))
      expr
      (let ((enew (apply-patterns plist expr)))
        (if (eq? enew expr)
            ;; expr didn't change; move to subexpressions
            (map (lambda (subex)
                   (if (not (pair? subex))
                       subex
                       (pattern-replace plist subex)))
                 expr)
            enew))))

(define-macro (pattern-set . pats)
  ; (pattern-lambda (x ...) ...) => x
  (define (pl-head p) (car (cadr p)))
  (receive
   (pls others) (separate (lambda (x)
                            (and (pair? x) (length= x 3)
                                 (eq? (car x) 'pattern-lambda)
                                 (pair? (cadr x))))
                          pats)
   (let ((heads (delete-duplicates (map pl-head pls)))
         (ht    (gensym)))
     `(let ((,ht (make-table)))
        ,@(map (lambda (h)
                 `(table-set! ,ht ',h (list
                                       ,@(filter (lambda (p)
                                                   (eq? (pl-head p) h))
                                                 pls))))
               heads)
        (vector 'pattern-set ,ht (list ,@others))))))

(define (plambda-expansion pat expr expander args)
  (let ((m (match pat expr)))
    (if m
        (apply expander (map (lambda (var) (cdr (or (assq var m) '(0 . #f))))
                             args))
        #f)))

(define-macro (pattern-lambda pat body)
  ; given a pattern p, return the list of capturing variables it uses
  (define (patargs- p)
    (cond ((and (symbol? p)
                (not (memq p '(_ ...))))
           (list p))

          ((pair? p)
           (if (eq? (car p) '-/)
               '()
               (delete-duplicates (apply append (map patargs- (to-proper (cdr p)))))))

          (else '())))
  (define (patargs p)
    (cons '__ (patargs- p)))
  (let* ((args (patargs pat))
         (expander `(lambda ,args ,body)))
    `(lambda (__ex__)
       (plambda-expansion ',pat __ex__ ,expander ',args))))
back to top