X-Git-Url: http://repo.macrolet.net/gitweb/?a=blobdiff_plain;f=src%2Fcode%2Fsort.lisp;h=99494de18093b04510150e59a5aae7c4df16894c;hb=3a10f894e7867fa2c27a3af05380abc3247f728d;hp=de92015c3540b245fe52e62f1e3b2114ff6b1857;hpb=8ac4c19014a23665e5842d0a989cb9d22d1592ed;p=sbcl.git diff --git a/src/code/sort.lisp b/src/code/sort.lisp index de92015..99494de 100644 --- a/src/code/sort.lisp +++ b/src/code/sort.lisp @@ -16,89 +16,8 @@ ;;; to generalize the CMU CL code to allow START and END values, this ;;; code has been written from scratch following Chapter 7 of ;;; _Introduction to Algorithms_ by Corman, Rivest, and Shamir. -(macrolet ((%index (x) `(truly-the index ,x)) - (%parent (i) `(ash ,i -1)) - (%left (i) `(%index (ash ,i 1))) - (%right (i) `(%index (1+ (ash ,i 1)))) - (%heapify (i) - `(do* ((i ,i) - (left (%left i) (%left i))) - ((> left current-heap-size)) - (declare (type index i left)) - (let* ((i-elt (%elt i)) - (i-key (funcall keyfun i-elt)) - (left-elt (%elt left)) - (left-key (funcall keyfun left-elt))) - (multiple-value-bind (large large-elt large-key) - (if (funcall predicate i-key left-key) - (values left left-elt left-key) - (values i i-elt i-key)) - (let ((right (%right i))) - (multiple-value-bind (largest largest-elt) - (if (> right current-heap-size) - (values large large-elt) - (let* ((right-elt (%elt right)) - (right-key (funcall keyfun right-elt))) - (if (funcall predicate large-key right-key) - (values right right-elt) - (values large large-elt)))) - (cond ((= largest i) - (return)) - (t - (setf (%elt i) largest-elt - (%elt largest) i-elt - i largest))))))))) - (%sort-vector (keyfun &optional (vtype 'vector)) - `(macrolet (;; KLUDGE: In SBCL ca. 0.6.10, I had trouble getting - ;; type inference to propagate all the way - ;; through this tangled mess of inlining. The - ;; TRULY-THE here works around that. -- WHN - (%elt (i) - `(aref (truly-the ,',vtype vector) - (%index (+ (%index ,i) start-1))))) - (let ((start-1 (1- start)) ; Heaps prefer 1-based addressing. - (current-heap-size (- end start)) - (keyfun ,keyfun)) - (declare (type (integer -1 #.(1- most-positive-fixnum)) - start-1)) - (declare (type index current-heap-size)) - (declare (type function keyfun)) - (loop for i of-type index - from (ash current-heap-size -1) downto 1 do - (%heapify i)) - (loop - (when (< current-heap-size 2) - (return)) - (rotatef (%elt 1) (%elt current-heap-size)) - (decf current-heap-size) - (%heapify 1)))))) - - (declaim (inline sort-vector)) - (defun sort-vector (vector start end predicate key) - (declare (type vector vector)) - (declare (type index start end)) - (declare (type function predicate)) - (declare (type (or function null) key)) - ;; This used to be (OPTIMIZE (SPEED 3) (SAFETY 3)), but now - ;; (0.7.1.39) that (SAFETY 3) means "absolutely safe (including - ;; expensive things like %DETECT-STACK-EXHAUSTION)" we get closer - ;; to what we want by using (SPEED 2) (SAFETY 2): "pretty fast, - ;; pretty safe, and safety is no more important than speed". - (declare (optimize (speed 2) (safety 2) (debug 1) (space 1))) - (if (typep vector 'simple-vector) - ;; (VECTOR T) is worth optimizing for, and SIMPLE-VECTOR is - ;; what we get from (VECTOR T) inside WITH-ARRAY-DATA. - (if (null key) - ;; Special-casing the KEY=NIL case lets us avoid some - ;; function calls. - (%sort-vector #'identity simple-vector) - (%sort-vector key simple-vector)) - ;; It's hard to anticipate many speed-critical applications for - ;; sorting vector types other than (VECTOR T), so we just lump - ;; them all together in one slow dynamically typed mess. - (locally - (declare (optimize (speed 2) (space 2) (inhibit-warnings 3))) - (%sort-vector (or key #'identity)))))) +(defun sort-vector (vector start end predicate key) + (sort-vector vector start end predicate key)) ;;; This is MAYBE-INLINE because it's not too hard to have an ;;; application where sorting is a major bottleneck, and inlining it @@ -112,7 +31,7 @@ (let ((predicate-function (%coerce-callable-to-fun predicate)) (key-function (and key (%coerce-callable-to-fun key)))) (typecase sequence - (list (sort-list sequence predicate-function key-function)) + (list (stable-sort-list sequence predicate-function key-function)) (vector (with-array-data ((vector (the vector sequence)) (start 0) @@ -130,13 +49,13 @@ (defun stable-sort (sequence predicate &key key) #!+sb-doc - "Destructively sorts sequence. Predicate should return non-Nil if - Arg1 is to precede Arg2." + "Destructively sort 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)) + (stable-sort-list sequence predicate key)) (vector (stable-sort-vector sequence predicate key)) (t @@ -145,31 +64,91 @@ :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 key) (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 + last ; last points to the last visited cell + (pred-fun (%coerce-callable-to-fun pred)) + (key-fun (if key + (%coerce-callable-to-fun key) + #'identity))) (declare (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)) @@ -178,21 +157,21 @@ (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) @@ -202,46 +181,8 @@ ;; 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 @@ -279,9 +220,9 @@ (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)) @@ -373,7 +314,7 @@ (defun stable-sort-vector (vector pred key) (vector-merge-sort vector pred key aref)) - + ;;;; merging (eval-when (:compile-toplevel :execute) @@ -406,8 +347,8 @@ (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)) @@ -422,25 +363,58 @@ #!+sb-doc "Merge the sequences SEQUENCE1 and SEQUENCE2 destructively into a sequence of type RESULT-TYPE using 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 (aver (typep 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))))) + ;; 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 (csubtypep (specifier-type '(cons nil t)) type) + (if (and (null s1) (null s2)) + (sequence-type-length-mismatch-error type 0) + (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)))))