X-Git-Url: http://repo.macrolet.net/gitweb/?a=blobdiff_plain;f=src%2Fcode%2Fsort.lisp;h=636c417594cb323b1b3cf977221374847fca961a;hb=bfe145acc01eb7a43790173db4f08610ae9cb07a;hp=443f83e7157c7ddaa0451d94372596aeeb4f46d2;hpb=1ff04b3ba4e6f3a0fc6ceea524e98720ecea7888;p=sbcl.git diff --git a/src/code/sort.lisp b/src/code/sort.lisp index 443f83e..636c417 100644 --- a/src/code/sort.lisp +++ b/src/code/sort.lisp @@ -11,155 +11,146 @@ (in-package "SB!IMPL") +(defun sort-vector (vector start end predicate-fun key-fun-or-nil) + (sort-vector vector start end predicate-fun key-fun-or-nil)) + +;;; This is MAYBE-INLINE because it's not too hard to have an +;;; application where sorting is a major bottleneck, and inlining it +;;; allows the compiler to make enough optimizations that it might be +;;; worth the (large) cost in space. +(declaim (maybe-inline sort)) (defun sort (sequence predicate &key key) #!+sb-doc "Destructively sort SEQUENCE. PREDICATE should return non-NIL if ARG1 is to precede ARG2." - (typecase sequence - (simple-vector - (if (> (the fixnum (length (the simple-vector sequence))) 0) - (sort-simple-vector sequence predicate key) - sequence)) - (list - (sort-list sequence predicate key)) - (vector - (if (> (the fixnum (length sequence)) 0) - (sort-vector sequence predicate key) - sequence)) - (t - (error 'simple-type-error - :datum sequence - :expected-type 'sequence - :format-control "~S is not a SEQUENCE." - :format-arguments (list sequence))))) - -;;;; sorting vectors - -;;; Make sorting functions for SIMPLE-VECTOR and miscellaneous other VECTORs. -(macrolet (;; BUILD-HEAP rearranges seq elements into a heap to start heap - ;; sorting. - (build-heap (seq type len-1 pred key) - (let ((i (gensym))) - `(do ((,i (floor ,len-1 2) (1- ,i))) - ((minusp ,i) ,seq) - (declare (fixnum ,i)) - (heapify ,seq ,type ,i ,len-1 ,pred ,key)))) - ;; HEAPIFY, assuming both sons of root are heaps, - ;; percolates the root element through the sons to form a - ;; heap at root. Root and max are zero based coordinates, - ;; but the heap algorithm only works on arrays indexed from - ;; 1 through N (not 0 through N-1); This is because a root - ;; at I has sons at 2*I and 2*I+1 which does not work for a - ;; root at 0. Because of this, boundaries, roots, and - ;; termination are computed using 1..N indexes. - (heapify (seq vector-ref root max pred key) - (let ((heap-root (gensym)) - (heap-max (gensym)) - (root-ele (gensym)) - (root-key (gensym)) - (heap-max/2 (gensym)) - (heap-l-son (gensym)) - (one-son (gensym)) - (one-son-ele (gensym)) - (one-son-key (gensym)) - (r-son-ele (gensym)) - (r-son-key (gensym)) - (var-root (gensym))) - `(let* ((,var-root ,root) ; (necessary to not clobber calling - ; root var) - (,heap-root (1+ ,root)) - (,heap-max (1+ ,max)) - (,root-ele (,vector-ref ,seq ,root)) - (,root-key (apply-key ,key ,root-ele)) - (,heap-max/2 (ash ,heap-max -1))) ; (floor heap-max 2) - (declare (fixnum ,var-root ,heap-root ,heap-max ,heap-max/2)) - (loop - (if (> ,heap-root ,heap-max/2) (return)) - (let* ((,heap-l-son (ash ,heap-root 1)) ; (* 2 heap-root) - ;; l-son index in seq (0..N-1) is one less than heap - ;; computation. - (,one-son (1- ,heap-l-son)) - (,one-son-ele (,vector-ref ,seq ,one-son)) - (,one-son-key (apply-key ,key ,one-son-ele))) - (declare (fixnum ,heap-l-son ,one-son)) - (if (< ,heap-l-son ,heap-max) - ;; There is a right son. - (let* ((,r-son-ele (,vector-ref ,seq ,heap-l-son)) - (,r-son-key (apply-key ,key ,r-son-ele))) - ;; Choose the greater of the two sons. - (when (funcall ,pred ,one-son-key ,r-son-key) - (setf ,one-son ,heap-l-son) - (setf ,one-son-ele ,r-son-ele) - (setf ,one-son-key ,r-son-key)))) - ;; If greater son is less than root, then we've - ;; formed a heap again.. - (if (funcall ,pred ,one-son-key ,root-key) (return)) - ;; ..else put greater son at root and make - ;; greater son node be the root. - (setf (,vector-ref ,seq ,var-root) ,one-son-ele) - (setf ,heap-root (1+ ,one-son)) ; (one plus to be in heap coordinates) - (setf ,var-root ,one-son))) ; actual index into vector for root ele - ;; Now really put percolated value into heap at the - ;; appropriate root node. - (setf (,vector-ref ,seq ,var-root) ,root-ele)))) - (def-vector-sort-fun (fun-name vector-ref) - `(defun ,fun-name (seq pred key) - (let ((len-1 (1- (length (the vector seq))))) - (declare (fixnum len-1)) - (build-heap seq ,vector-ref len-1 pred key) - (do* ((i len-1 i-1) - (i-1 (1- i) (1- i-1))) - ((zerop i) seq) - (declare (fixnum i i-1)) - (rotatef (,vector-ref seq 0) (,vector-ref seq i)) - (heapify seq ,vector-ref 0 i-1 pred key)))))) - (def-vector-sort-fun sort-vector aref) - (def-vector-sort-fun sort-simple-vector svref)) + (let ((predicate-fun (%coerce-callable-to-fun predicate))) + (typecase sequence + (list + (stable-sort-list sequence + predicate-fun + (if key (%coerce-callable-to-fun key) #'identity))) + (vector + (let ((key-fun-or-nil (and key (%coerce-callable-to-fun key)))) + (with-array-data ((vector (the vector sequence)) + (start 0) + (end (length sequence))) + (sort-vector vector start end predicate-fun key-fun-or-nil))) + sequence) + (t + (error 'simple-type-error + :datum sequence + :expected-type 'sequence + :format-control "~S is not a sequence." + :format-arguments (list sequence)))))) ;;;; stable sorting (defun stable-sort (sequence predicate &key key) #!+sb-doc - "Destructively sorts sequence. Predicate should return non-Nil if - Arg1 is to precede Arg2." - (typecase sequence - (simple-vector - (stable-sort-simple-vector sequence predicate key)) - (list - (sort-list sequence predicate key)) - (vector - (stable-sort-vector sequence predicate key)) - (t - (error 'simple-type-error - :datum sequence - :expected-type 'sequence - :format-control "~S is not a sequence." - :format-arguments (list sequence))))) + "Destructively sort SEQUENCE. PREDICATE should return non-NIL if + ARG1 is to precede ARG2." + (let ((predicate-fun (%coerce-callable-to-fun predicate))) + (typecase sequence + (simple-vector + (stable-sort-simple-vector sequence + predicate-fun + (and key (%coerce-callable-to-fun key)))) + (list + (stable-sort-list sequence + predicate-fun + (if key (%coerce-callable-to-fun key) #'identity))) + (vector + (stable-sort-vector sequence + predicate-fun + (and key (%coerce-callable-to-fun key)))) + (t + (error 'simple-type-error + :datum sequence + :expected-type 'sequence + :format-control "~S is not a sequence." + :format-arguments (list sequence)))))) + +;;; APPLY-KEYED-PRED saves us a function call sometimes. +(eval-when (:compile-toplevel :execute) + (sb!xc:defmacro apply-keyed-pred (one two pred key) + `(if ,key + (funcall ,pred (funcall ,key ,one) + (funcall ,key ,two)) + (funcall ,pred ,one ,two))) +) ; EVAL-WHEN + +;;;; stable sort of lists -;;; stable sort of lists +(defun last-cons-of (list) + (loop (let ((rest (rest list))) + (if rest + (setf list rest) + (return list))))) -;;; SORT-LIST uses a bottom up merge sort. First a pass is made over -;;; the list grabbing one element at a time and merging it with the -;;; next one form pairs of sorted elements. Then n is doubled, and -;;; elements are taken in runs of two, merging one run with the next -;;; to form quadruples of sorted elements. This continues until n is -;;; large enough that the inner loop only runs for one iteration; that -;;; is, there are only two runs that can be merged, the first run +;;; Destructively merge LIST-1 with LIST-2 (given that they're already +;;; sorted w.r.t. PRED-FUN on KEY-FUN, giving output sorted the same +;;; way). In the resulting list, elements of LIST-1 are guaranteed to +;;; come before equal elements of LIST-2. +;;; +;;; Return (VALUES HEAD TAILTAIL), where HEAD is the same value you'd +;;; expect from MERGE, and TAILTAIL is the last cons in the list (i.e. +;;; the last cons in the list which NRECONC calls TAIL). +(defun merge-lists* (list-1 list-2 pred-fun key-fun) + (declare (type list list-1 list-2)) + (declare (type function pred-fun key-fun)) + (cond ((null list-1) (values list-2 (last-cons-of list-2))) + ((null list-2) (values list-1 (last-cons-of list-1))) + (t (let* ((reversed-result-so-far nil) + (key-1 (funcall key-fun (car list-1))) + (key-2 (funcall key-fun (car list-2)))) + (loop + (macrolet ((frob (list-i key-i other-list) + `(progn + ;; basically + ;; (PUSH (POP ,LIST-I) REVERSED-RESULT-SO-FAR), + ;; except doing some fancy footwork to + ;; reuse the cons cell: + (psetf (cdr ,list-i) reversed-result-so-far + reversed-result-so-far ,list-i + ,list-i (cdr ,list-i)) + ;; Now maybe we're done. + (if (endp ,list-i) + (return (values (nreconc + reversed-result-so-far + ,other-list) + (last-cons-of + ,other-list))) + (setf ,key-i + (funcall key-fun (car ,list-i))))))) + ;; Note that by making KEY-2 the first arg to + ;; PRED-FUN, we arrange that if PRED-FUN is a function + ;; in the #'< style, the outcome is stably sorted. + (if (funcall pred-fun key-2 key-1) + (frob list-2 key-2 list-1) + (frob list-1 key-1 list-2)))))))) + +;;; STABLE-SORT-LIST uses a bottom-up merge sort. First a pass is made +;;; over the list grabbing one element at a time and merging it with +;;; the next one to form pairs of sorted elements. Then N is doubled, +;;; and elements are taken in runs of two, merging one run with the +;;; next to form quadruples of sorted elements. This continues until N +;;; is large enough that the inner loop only runs for one iteration; +;;; that is, there are only two runs that can be merged, the first run ;;; starting at the beginning of the list, and the second being the ;;; remaining elements. - -(defun sort-list (list pred key) +(defun stable-sort-list (list pred-fun key-fun) (let ((head (cons :header list)) ; head holds on to everything (n 1) ; bottom-up size of lists to be merged unsorted ; unsorted is the remaining list to be ; broken into n size lists and merged list-1 ; list-1 is one length n list to be merged last) ; last points to the last visited cell - (declare (fixnum n)) + (declare (type function pred-fun key-fun) + (type fixnum n)) (loop - ;; start collecting runs of n at the first element + ;; Start collecting runs of N at the first element. (setf unsorted (cdr head)) - ;; tack on the first merge of two n-runs to the head holder + ;; Tack on the first merge of two N-runs to the head holder. (setf last head) (let ((n-1 (1- n))) (declare (fixnum n-1)) @@ -168,21 +159,21 @@ (let ((temp (nthcdr n-1 list-1)) list-2) (cond (temp - ;; there are enough elements for a second run + ;; There are enough elements for a second run. (setf list-2 (cdr temp)) (setf (cdr temp) nil) (setf temp (nthcdr n-1 list-2)) (cond (temp (setf unsorted (cdr temp)) (setf (cdr temp) nil)) - ;; the second run goes off the end of the list + ;; The second run goes off the end of the list. (t (setf unsorted nil))) (multiple-value-bind (merged-head merged-last) - (merge-lists* list-1 list-2 pred key) - (setf (cdr last) merged-head) - (setf last merged-last)) + (merge-lists* list-1 list-2 pred-fun key-fun) + (setf (cdr last) merged-head + last merged-last)) (if (null unsorted) (return))) - ;; if there is only one run, then tack it on to the end + ;; If there is only one run, then tack it on to the end. (t (setf (cdr last) list-1) (return))))) (setf n (ash n 1)) ; (+ n n) @@ -192,46 +183,8 @@ ;; iteration to realize. (if (eq list-1 (cdr head)) (return list-1)))))) - -;;; APPLY-PRED saves us a function call sometimes. -(eval-when (:compile-toplevel :execute) - (sb!xc:defmacro apply-pred (one two pred key) - `(if ,key - (funcall ,pred (funcall ,key ,one) - (funcall ,key ,two)) - (funcall ,pred ,one ,two))) -) ; EVAL-WHEN - -(defvar *merge-lists-header* (list :header)) - -;;; MERGE-LISTS* originally written by Jim Large. -;;; modified to return a pointer to the end of the result -;;; and to not cons header each time its called. -;;; It destructively merges list-1 with list-2. In the resulting -;;; list, elements of list-2 are guaranteed to come after equal elements -;;; of list-1. -(defun merge-lists* (list-1 list-2 pred key) - (do* ((result *merge-lists-header*) - (P result)) ; points to last cell of result - ((or (null list-1) (null list-2)) ; done when either list used up - (if (null list-1) ; in which case, append the - (rplacd p list-2) ; other list - (rplacd p list-1)) - (do ((drag p lead) - (lead (cdr p) (cdr lead))) - ((null lead) - (values (prog1 (cdr result) ; Return the result sans header - (rplacd result nil)) ; (free memory, be careful) - drag)))) ; and return pointer to last element. - (cond ((apply-pred (car list-2) (car list-1) pred key) - (rplacd p list-2) ; Append the lesser list to last cell of - (setq p (cdr p)) ; result. Note: test must bo done for - (pop list-2)) ; LIST-2 < LIST-1 so merge will be - (T (rplacd p list-1) ; stable for LIST-1. - (setq p (cdr p)) - (pop list-1))))) - -;;; stable sort of vectors + +;;;; stable sort of vectors ;;; Stable sorting vectors is done with the same algorithm used for ;;; lists, using a temporary vector to merge back and forth between it @@ -269,9 +222,9 @@ (incf ,target-i) (incf ,i)) (return)) - ((apply-pred (,source-ref ,source ,j) - (,source-ref ,source ,i) - ,pred ,key) + ((apply-keyed-pred (,source-ref ,source ,j) + (,source-ref ,source ,i) + ,pred ,key) (setf (,target-ref ,target ,target-i) (,source-ref ,source ,j)) (incf ,j)) @@ -280,10 +233,10 @@ (incf ,i))) (incf ,target-i))))) -;;; VECTOR-MERGE-SORT is the same algorithm used to stable sort lists, but -;;; it uses a temporary vector. Direction determines whether we are merging -;;; into the temporary (T) or back into the given vector (NIL). - +;;; VECTOR-MERGE-SORT is the same algorithm used to stable sort lists, +;;; but it uses a temporary vector. DIRECTION determines whether we +;;; are merging into the temporary (T) or back into the given vector +;;; (NIL). (sb!xc:defmacro vector-merge-sort (vector pred key vector-ref) (let ((vector-len (gensym)) (n (gensym)) (direction (gensym)) (unsorted (gensym)) @@ -351,28 +304,31 @@ ) ; EVAL-when -;;; Temporary vector for stable sorting vectors. +;;; temporary vector for stable sorting vectors (defvar *merge-sort-temp-vector* (make-array 50)) (declaim (simple-vector *merge-sort-temp-vector*)) (defun stable-sort-simple-vector (vector pred key) - (declare (simple-vector vector)) + (declare (type simple-vector vector) + (type function pred) + (type (or null function) key)) (vector-merge-sort vector pred key svref)) (defun stable-sort-vector (vector pred key) + (declare (type function pred) + (type (or null function) key)) (vector-merge-sort vector pred key aref)) - + ;;;; merging (eval-when (:compile-toplevel :execute) ;;; MERGE-VECTORS returns a new vector which contains an interleaving -;;; of the elements of vector-1 and vector-2. Elements from vector-2 are -;;; chosen only if they are strictly less than elements of vector-1, -;;; (pred elt-2 elt-1), as specified in the manual. - +;;; of the elements of VECTOR-1 and VECTOR-2. Elements from VECTOR-2 +;;; are chosen only if they are strictly less than elements of +;;; VECTOR-1, (PRED ELT-2 ELT-1), as specified in the manual. (sb!xc:defmacro merge-vectors (vector-1 length-1 vector-2 length-2 result-vector pred key access) (let ((result-i (gensym)) @@ -397,8 +353,8 @@ (incf ,result-i) (incf ,i)) (return ,result-vector)) - ((apply-pred (,access ,vector-2 ,j) (,access ,vector-1 ,i) - ,pred ,key) + ((apply-keyed-pred (,access ,vector-2 ,j) (,access ,vector-1 ,i) + ,pred ,key) (setf (,access ,result-vector ,result-i) (,access ,vector-2 ,j)) (incf ,j)) @@ -411,27 +367,66 @@ (defun merge (result-type sequence1 sequence2 predicate &key key) #!+sb-doc - "The sequences Sequence1 and Sequence2 are destructively merged into - a sequence of type Result-Type using the Predicate to order the elements." - (if (eq result-type 'list) - (let ((result (merge-lists* (coerce sequence1 'list) - (coerce sequence2 'list) - predicate key))) - result) - (let* ((vector-1 (coerce sequence1 'vector)) - (vector-2 (coerce sequence2 'vector)) - (length-1 (length vector-1)) - (length-2 (length vector-2)) - (result (make-sequence-of-type result-type (+ length-1 length-2)))) - (declare (vector vector-1 vector-2) - (fixnum length-1 length-2)) - - #!+high-security - (check-type-var result result-type) - (if (and (simple-vector-p result) - (simple-vector-p vector-1) - (simple-vector-p vector-2)) - (merge-vectors vector-1 length-1 vector-2 length-2 - result predicate key svref) - (merge-vectors vector-1 length-1 vector-2 length-2 - result predicate key aref))))) + "Merge the sequences SEQUENCE1 and SEQUENCE2 destructively into a + sequence of type RESULT-TYPE using PREDICATE to order the elements." + ;; FIXME: This implementation is remarkably inefficient in various + ;; ways. In decreasing order of estimated user astonishment, I note: + ;; full calls to SPECIFIER-TYPE at runtime; copying input vectors + ;; to lists before doing MERGE-LISTS*; and walking input lists + ;; (because of the call to MERGE-LISTS*, which walks the list to + ;; find the last element for its second return value) even in cases + ;; like (MERGE 'LIST (LIST 1) (LIST 2 3 4 5 ... 1000)) where one list + ;; can be largely ignored. -- WHN 2003-01-05 + (let ((type (specifier-type result-type))) + (cond + ((csubtypep type (specifier-type 'list)) + ;; the VECTOR clause, below, goes through MAKE-SEQUENCE, so + ;; benefits from the error checking there. Short of + ;; reimplementing everything, we can't do the same for the LIST + ;; case, so do relevant length checking here: + (let ((s1 (coerce sequence1 'list)) + (s2 (coerce sequence2 'list)) + (pred-fun (%coerce-callable-to-fun predicate)) + (key-fun (if key + (%coerce-callable-to-fun key) + #'identity))) + (when (type= type (specifier-type 'list)) + (return-from merge (values (merge-lists* s1 s2 pred-fun key-fun)))) + (when (eq type *empty-type*) + (bad-sequence-type-error nil)) + (when (type= type (specifier-type 'null)) + (if (and (null s1) (null s2)) + (return-from merge 'nil) + ;; FIXME: This will break on circular lists (as, + ;; indeed, will the whole MERGE function). + (sequence-type-length-mismatch-error type + (+ (length s1) + (length s2))))) + (if (cons-type-p type) + (multiple-value-bind (min exactp) + (sb!kernel::cons-type-length-info type) + (let ((length (+ (length s1) (length s2)))) + (if exactp + (unless (= length min) + (sequence-type-length-mismatch-error type length)) + (unless (>= length min) + (sequence-type-length-mismatch-error type length))) + (values (merge-lists* s1 s2 pred-fun key-fun)))) + (sequence-type-too-hairy result-type)))) + ((csubtypep type (specifier-type 'vector)) + (let* ((vector-1 (coerce sequence1 'vector)) + (vector-2 (coerce sequence2 'vector)) + (length-1 (length vector-1)) + (length-2 (length vector-2)) + (result (make-sequence result-type + (+ length-1 length-2)))) + (declare (vector vector-1 vector-2) + (fixnum length-1 length-2)) + (if (and (simple-vector-p result) + (simple-vector-p vector-1) + (simple-vector-p vector-2)) + (merge-vectors vector-1 length-1 vector-2 length-2 + result predicate key svref) + (merge-vectors vector-1 length-1 vector-2 length-2 + result predicate key aref)))) + (t (bad-sequence-type-error result-type)))))