X-Git-Url: http://repo.macrolet.net/gitweb/?a=blobdiff_plain;f=src%2Fcompiler%2Fseqtran.lisp;h=40040ab2576fb9d897a0a3f31b471ee59ead02a5;hb=8e4ec430504f0f563280be26034af590dff50d34;hp=a3126da6b8cdffe82a680d6c2b24e77195701e4e;hpb=8cbd7fc0f27222a778ce61bae7d943a5081362cc;p=sbcl.git diff --git a/src/compiler/seqtran.lisp b/src/compiler/seqtran.lisp index a3126da..40040ab 100644 --- a/src/compiler/seqtran.lisp +++ b/src/compiler/seqtran.lisp @@ -27,31 +27,30 @@ (args-to-fn (if take-car `(car ,v) v)))) (let* ((fn-sym (gensym)) ; for ONCE-ONLY-ish purposes - (call `(funcall ,fn-sym . ,(args-to-fn))) + (call `(%funcall ,fn-sym . ,(args-to-fn))) (endtest `(or ,@(tests)))) - (ecase accumulate - (:nconc - (let ((temp (gensym)) - (map-result (gensym))) - `(let ((,fn-sym ,fn) - (,map-result (list nil))) - (do-anonymous ((,temp ,map-result) . ,(do-clauses)) - (,endtest (cdr ,map-result)) - (setq ,temp (last (nconc ,temp ,call))))))) - (:list - (let ((temp (gensym)) - (map-result (gensym))) - `(let ((,fn-sym ,fn) - (,map-result (list nil))) - (do-anonymous ((,temp ,map-result) . ,(do-clauses)) - (,endtest (truly-the list (cdr ,map-result))) - (rplacd ,temp (setq ,temp (list ,call))))))) - ((nil) - `(let ((,fn-sym ,fn) - (,n-first ,(first arglists))) - (do-anonymous ,(do-clauses) - (,endtest (truly-the list ,n-first)) - ,call)))))))) + + `(let ((,fn-sym (%coerce-callable-to-fun ,fn))) + ,(ecase accumulate + (:nconc + (let ((temp (gensym)) + (map-result (gensym))) + `(let ((,map-result (list nil))) + (do-anonymous ((,temp ,map-result) . ,(do-clauses)) + (,endtest (cdr ,map-result)) + (setq ,temp (last (nconc ,temp ,call))))))) + (:list + (let ((temp (gensym)) + (map-result (gensym))) + `(let ((,map-result (list nil))) + (do-anonymous ((,temp ,map-result) . ,(do-clauses)) + (,endtest (truly-the list (cdr ,map-result))) + (rplacd ,temp (setq ,temp (list ,call))))))) + ((nil) + `(let ((,n-first ,(first arglists))) + (do-anonymous ,(do-clauses) + (,endtest (truly-the list ,n-first)) + ,call))))))))) (define-source-transform mapc (function list &rest more-lists) (mapfoo-transform function (cons list more-lists) nil t)) @@ -127,14 +126,15 @@ ;;; MAP-INTO. RESULT and BODY are forms, which can use variables ;;; FUNCALL-RESULT, containing the result of application of FUN, and ;;; INDEX, containing the current position in sequences. -(defun build-sequence-iterator (seqs seq-names &key result into body) +(defun build-sequence-iterator (seqs seq-names &key result into body fast) (declare (type list seqs seq-names) (type symbol into)) (collect ((bindings) (declarations) (vector-lengths) (tests) - (places)) + (places) + (around)) (let ((found-vector-p nil)) (flet ((process-vector (length) (unless found-vector-p @@ -151,10 +151,22 @@ (declarations `(type list ,index)) (places `(car ,index)) (tests `(endp ,index)))) - ((csubtypep type (specifier-type 'vector)) + ((or (csubtypep type (specifier-type '(simple-array * 1))) + (and (not fast) + (csubtypep type (specifier-type 'vector)))) (process-vector `(length ,seq-name)) (places `(locally (declare (optimize (insert-array-bounds-checks 0))) (aref ,seq-name index)))) + ((csubtypep type (specifier-type 'vector)) + (let ((data (gensym "DATA")) + (start (gensym "START")) + (end (gensym "END"))) + (around `(with-array-data ((,data ,seq-name) + (,start) + (,end (length ,seq-name))))) + (process-vector `(- ,end ,start)) + (places `(locally (declare (optimize (insert-array-bounds-checks 0))) + (aref ,data (truly-the index (+ index ,start))))))) (t (give-up-ir1-transform "can't determine sequence argument type")))) @@ -163,12 +175,18 @@ (when found-vector-p (bindings `(length (min ,@(vector-lengths)))) (tests `(>= index length))) - `(do (,@(bindings)) - ((or ,@(tests)) ,result) - (declare ,@(declarations)) - (let ((funcall-result (funcall fun ,@(places)))) - (declare (ignorable funcall-result)) - ,body))))) + (let ((body `(do (,@(bindings)) + ((or ,@(tests)) ,result) + (declare ,@(declarations)) + (let ((funcall-result (funcall fun ,@(places)))) + (declare (ignorable funcall-result)) + ,body)))) + (if (around) + (reduce (lambda (wrap body) (append wrap (list body))) + (around) + :from-end t + :initial-value body) + body))))) ;;; Try to compile %MAP efficiently when we can determine sequence ;;; argument types at compile time. @@ -183,7 +201,7 @@ ;;; the reader, because the code is complicated enough already and I ;;; don't happen to need that functionality right now. -- WHN 20000410 (deftransform %map ((result-type fun seq &rest seqs) * * - :policy (>= speed space)) + :node node :policy (>= speed space)) "open code" (unless (constant-lvar-p result-type) (give-up-ir1-transform "RESULT-TYPE argument not constant")) @@ -245,25 +263,45 @@ ,(build-sequence-iterator seqs seq-args :result result - :body push-dacc)))))))))) + :body push-dacc + :fast (policy node (> speed space)))))))))))) ;;; MAP-INTO (deftransform map-into ((result fun &rest seqs) (vector * &rest *) - *) + * :node node) "open code" (let ((seqs-names (mapcar (lambda (x) (declare (ignore x)) (gensym)) seqs))) `(lambda (result fun ,@seqs-names) - ,(build-sequence-iterator - seqs seqs-names - :result '(when (array-has-fill-pointer-p result) - (setf (fill-pointer result) index)) - :into 'result - :body '(locally (declare (optimize (insert-array-bounds-checks 0))) - (setf (aref result index) funcall-result))) + ,(if (and (policy node (> speed space)) + (not (csubtypep (lvar-type result) + (specifier-type '(simple-array * 1))))) + (let ((data (gensym "DATA")) + (start (gensym "START")) + (end (gensym "END"))) + `(with-array-data ((,data result) + (,start) + (,end)) + (declare (ignore ,end)) + ,(build-sequence-iterator + seqs seqs-names + :result '(when (array-has-fill-pointer-p result) + (setf (fill-pointer result) index)) + :into 'result + :body `(locally (declare (optimize (insert-array-bounds-checks 0))) + (setf (aref ,data (truly-the index (+ index ,start))) + funcall-result)) + :fast t))) + (build-sequence-iterator + seqs seqs-names + :result '(when (array-has-fill-pointer-p result) + (setf (fill-pointer result) index)) + :into 'result + :body '(locally (declare (optimize (insert-array-bounds-checks 0))) + (setf (aref result index) funcall-result)))) result))) @@ -277,7 +315,7 @@ '(nth i s)) (deftransform %setelt ((s i v) ((simple-array * (*)) * *) *) - '(%aset s i v)) + '(setf (aref s i) v)) (deftransform %setelt ((s i v) (list * *) * :policy (< safety 3)) '(setf (car (nthcdr i s)) v)) @@ -285,62 +323,214 @@ (deftransform %check-vector-sequence-bounds ((vector start end) (vector * *) * :node node) - (if (policy node (< safety speed)) + (if (policy node (= 0 insert-array-bounds-checks)) '(or end (length vector)) '(let ((length (length vector))) - (if (<= 0 start (or end length) length) - (or end length) - (sb!impl::signal-bounding-indices-bad-error vector start end))))) - -(macrolet ((def (name) - `(deftransform ,name ((e l &key (test #'eql)) * * - :node node) - (unless (constant-lvar-p l) - (give-up-ir1-transform)) - - (let ((val (lvar-value l))) - (unless (policy node - (or (= speed 3) - (and (>= speed space) - (<= (length val) 5)))) - (give-up-ir1-transform)) - - (labels ((frob (els) - (if els - `(if (funcall test e ',(car els)) - ',els - ,(frob (cdr els))) - nil))) - (frob val)))))) - (def member) - (def memq)) - -;;; FIXME: We have rewritten the original code that used DOLIST to this -;;; more natural MACROLET. However, the original code suggested that when -;;; this was done, a few bytes could be saved by a call to a shared -;;; function. This remains to be done. -(macrolet ((def (fun eq-fun) - `(deftransform ,fun ((item list &key test) (t list &rest t) *) - "convert to EQ test" - ;; FIXME: The scope of this transformation could be - ;; widened somewhat, letting it work whenever the test is - ;; 'EQL and we know from the type of ITEM that it #'EQ - ;; works like #'EQL on it. (E.g. types FIXNUM, CHARACTER, - ;; and SYMBOL.) - ;; If TEST is EQ, apply transform, else - ;; if test is not EQL, then give up on transform, else - ;; if ITEM is not a NUMBER or is a FIXNUM, apply - ;; transform, else give up on transform. - (cond (test - (unless (lvar-fun-is test '(eq)) - (give-up-ir1-transform))) - ((types-equal-or-intersect (lvar-type item) - (specifier-type 'number)) - (give-up-ir1-transform "Item might be a number."))) - `(,',eq-fun item list)))) - (def delete delq) - (def assoc assq) - (def member memq)) + (if (<= 0 start (or end length) length) + (or end length) + (sequence-bounding-indices-bad-error vector start end))))) + +(def!type eq-comparable-type () + '(or fixnum (not number))) + +;;; True if EQL comparisons involving type can be simplified to EQ. +(defun eq-comparable-type-p (type) + (csubtypep type (specifier-type 'eq-comparable-type))) + +(defun specialized-list-seek-function-name (function-name key-functions &optional variant) + (or (find-symbol (with-output-to-string (s) + ;; Write "%NAME-FUN1-FUN2-FUN3", etc. Not only is + ;; this ever so slightly faster then FORMAT, this + ;; way we are also proof against *PRINT-CASE* + ;; frobbing and such. + (write-char #\% s) + (write-string (symbol-name function-name) s) + (dolist (f key-functions) + (write-char #\- s) + (write-string (symbol-name f) s)) + (when variant + (write-char #\- s) + (write-string (symbol-name variant) s))) + (load-time-value (find-package "SB!KERNEL"))) + (bug "Unknown list item seek transform: name=~S, key-functions=~S variant=~S" + function-name key-functions variant))) + +(defparameter *list-open-code-limit* 128) + +(defun transform-list-item-seek (name item list key test test-not node) + (when (and test test-not) + (abort-ir1-transform "Both ~S and ~S supplied to ~S." :test :test-not name)) + ;; If TEST is EQL, drop it. + (when (and test (lvar-fun-is test '(eql))) + (setf test nil)) + ;; Ditto for KEY IDENTITY. + (when (and key (lvar-fun-is key '(identity))) + (setf key nil)) + ;; Key can legally be NIL, but if it's NIL for sure we pretend it's + ;; not there at all. If it might be NIL, make up a form to that + ;; ensures it is a function. + (multiple-value-bind (key key-form) + (when key + (let ((key-type (lvar-type key)) + (null-type (specifier-type 'null))) + (cond ((csubtypep key-type null-type) + (values nil nil)) + ((csubtypep null-type key-type) + (values key '(if key + (%coerce-callable-to-fun key) + #'identity))) + (t + (values key (ensure-lvar-fun-form key 'key)))))) + (let* ((c-test (cond ((and test (lvar-fun-is test '(eq))) + (setf test nil) + 'eq) + ((and (not test) (not test-not)) + (when (eq-comparable-type-p (lvar-type item)) + 'eq)))) + (funs (delete nil (list (when key (list key 'key)) + (when test (list test 'test)) + (when test-not (list test-not 'test-not))))) + (target-expr (if key '(%funcall key target) 'target)) + (test-expr (cond (test `(%funcall test item ,target-expr)) + (test-not `(not (%funcall test-not item ,target-expr))) + (c-test `(,c-test item ,target-expr)) + (t `(eql item ,target-expr))))) + (labels ((open-code (tail) + (when tail + `(if (let ((this ',(car tail))) + ,(ecase name + ((assoc rassoc) + (let ((cxx (if (eq name 'assoc) 'car 'cdr))) + `(and this (let ((target (,cxx this))) + ,test-expr)))) + (member + `(let ((target this)) + ,test-expr)))) + ',(ecase name + ((assoc rassoc) (car tail)) + (member tail)) + ,(open-code (cdr tail))))) + (ensure-fun (args) + (if (eq 'key (second args)) + key-form + (apply #'ensure-lvar-fun-form args)))) + (let* ((cp (constant-lvar-p list)) + (c-list (when cp (lvar-value list)))) + (cond ((and cp c-list (member name '(assoc rassoc member)) + (policy node (>= speed space)) + (not (nthcdr *list-open-code-limit* c-list))) + `(let ,(mapcar (lambda (fun) `(,(second fun) ,(ensure-fun fun))) funs) + ,(open-code c-list))) + ((and cp (not c-list)) + ;; constant nil list + (if (eq name 'adjoin) + '(list item) + nil)) + (t + ;; specialized out-of-line version + `(,(specialized-list-seek-function-name name (mapcar #'second funs) c-test) + item list ,@(mapcar #'ensure-fun funs))))))))) + +(defun transform-list-pred-seek (name pred list key node) + ;; If KEY is IDENTITY, drop it. + (when (and key (lvar-fun-is key '(identity))) + (setf key nil)) + ;; Key can legally be NIL, but if it's NIL for sure we pretend it's + ;; not there at all. If it might be NIL, make up a form to that + ;; ensures it is a function. + (multiple-value-bind (key key-form) + (when key + (let ((key-type (lvar-type key)) + (null-type (specifier-type 'null))) + (cond ((csubtypep key-type null-type) + (values nil nil)) + ((csubtypep null-type key-type) + (values key '(if key + (%coerce-callable-to-fun key) + #'identity))) + (t + (values key (ensure-lvar-fun-form key 'key)))))) + (let ((test-expr `(%funcall pred ,(if key '(%funcall key target) 'target))) + (pred-expr (ensure-lvar-fun-form pred 'pred))) + (when (member name '(member-if-not assoc-if-not rassoc-if-not)) + (setf test-expr `(not ,test-expr))) + (labels ((open-code (tail) + (when tail + `(if (let ((this ',(car tail))) + ,(ecase name + ((assoc-if assoc-if-not rassoc-if rassoc-if-not) + (let ((cxx (if (member name '(assoc-if assoc-if-not)) 'car 'cdr))) + `(and this (let ((target (,cxx this))) + ,test-expr)))) + ((member-if member-if-not) + `(let ((target this)) + ,test-expr)))) + ',(ecase name + ((assoc-if assoc-if-not rassoc-if rassoc-if-not) + (car tail)) + ((member-if member-if-not) + tail)) + ,(open-code (cdr tail)))))) + (let* ((cp (constant-lvar-p list)) + (c-list (when cp (lvar-value list)))) + (cond ((and cp c-list (policy node (>= speed space)) + (not (nthcdr *list-open-code-limit* c-list))) + `(let ((pred ,pred-expr) + ,@(when key `((key ,key-form)))) + ,(open-code c-list))) + ((and cp (not c-list)) + ;; constant nil list -- nothing to find! + nil) + (t + ;; specialized out-of-line version + `(,(specialized-list-seek-function-name name (when key '(key))) + ,pred-expr list ,@(when key (list key-form)))))))))) + +(macrolet ((def (name &optional if/if-not) + (let ((basic (symbolicate "%" name)) + (basic-eq (symbolicate "%" name "-EQ")) + (basic-key (symbolicate "%" name "-KEY")) + (basic-key-eq (symbolicate "%" name "-KEY-EQ"))) + `(progn + (deftransform ,name ((item list &key key test test-not) * * :node node) + (transform-list-item-seek ',name item list key test test-not node)) + (deftransform ,basic ((item list) (eq-comparable-type t)) + `(,',basic-eq item list)) + (deftransform ,basic-key ((item list) (eq-comparable-type t)) + `(,',basic-key-eq item list)) + ,@(when if/if-not + (let ((if-name (symbolicate name "-IF")) + (if-not-name (symbolicate name "-IF-NOT"))) + `((deftransform ,if-name ((pred list &key key) * * :node node) + (transform-list-pred-seek ',if-name pred list key node)) + (deftransform ,if-not-name ((pred list &key key) * * :node node) + (transform-list-pred-seek ',if-not-name pred list key node))))))))) + (def adjoin) + (def assoc t) + (def member t) + (def rassoc t)) + +(deftransform memq ((item list) (t (constant-arg list))) + (labels ((rec (tail) + (if tail + `(if (eq item ',(car tail)) + ',tail + ,(rec (cdr tail))) + nil))) + (rec (lvar-value list)))) + +;;; A similar transform used to apply to MEMBER and ASSOC, but since +;;; TRANSFORM-LIST-ITEM-SEEK now takes care of them those transform +;;; would never fire, and (%MEMBER-TEST ITEM LIST #'EQ) should be +;;; almost as fast as MEMQ. +(deftransform delete ((item list &key test) (t list &rest t) *) + "convert to EQ test" + (let ((type (lvar-type item))) + (unless (or (and test (lvar-fun-is test '(eq))) + (and (eq-comparable-type-p type) + (or (not test) (lvar-fun-is test '(eql))))) + (give-up-ir1-transform))) + `(delq item list)) (deftransform delete-if ((pred list) (t list)) "open code" @@ -353,255 +543,168 @@ (rplacd splice (cdr x)))) (t (setq splice x))))) -(deftransform fill ((seq item &key (start 0) (end (length seq))) - (vector t &key (:start t) (:end index)) +(deftransform fill ((seq item &key (start 0) (end nil)) + (list t &key (:start t) (:end t))) + '(list-fill* seq item start end)) + +(deftransform fill ((seq item &key (start 0) (end nil)) + (vector t &key (:start t) (:end t)) * - :policy (> speed space)) - "open code" - (let ((element-type (upgraded-element-type-specifier-or-give-up seq))) - (values - `(with-array-data ((data seq) - (start start) - (end end)) - (declare (type (simple-array ,element-type 1) data)) - (declare (type fixnum start end)) - (do ((i start (1+ i))) - ((= i end) seq) - (declare (type index i)) - ;; WITH-ARRAY-DATA did our range checks once and for all, so - ;; it'd be wasteful to check again on every AREF... - (declare (optimize (safety 0))) - (setf (aref data i) item))) - ;; ... though we still need to check that the new element can fit - ;; into the vector in safe code. -- CSR, 2002-07-05 - `((declare (type ,element-type item)))))) - -;;;; utilities - -;;; Return true if LVAR's only use is a non-NOTINLINE reference to a -;;; global function with one of the specified NAMES. -(defun lvar-fun-is (lvar names) - (declare (type lvar lvar) (list names)) - (let ((use (lvar-uses lvar))) - (and (ref-p use) - (let ((leaf (ref-leaf use))) - (and (global-var-p leaf) - (eq (global-var-kind leaf) :global-function) - (not (null (member (leaf-source-name leaf) names - :test #'equal)))))))) - -;;; If LVAR is a constant lvar, the return the constant value. If it -;;; is null, then return default, otherwise quietly give up the IR1 -;;; transform. -;;; -;;; ### Probably should take an ARG and flame using the NAME. -(defun constant-value-or-lose (lvar &optional default) - (declare (type (or lvar null) lvar)) - (cond ((not lvar) default) - ((constant-lvar-p lvar) - (lvar-value lvar)) - (t - (give-up-ir1-transform)))) - -;;; FIXME: Why is this code commented out? (Why *was* it commented -;;; out? We inherited this situation from cmucl-2.4.8, with no -;;; explanation.) Should we just delete this code? -#| -;;; This is a frob whose job it is to make it easier to pass around -;;; the arguments to IR1 transforms. It bundles together the name of -;;; the argument (which should be referenced in any expansion), and -;;; the continuation for that argument (or NIL if unsupplied.) -(defstruct (arg (:constructor %make-arg (name cont)) - (:copier nil)) - (name nil :type symbol) - (cont nil :type (or continuation null))) -(defmacro make-arg (name) - `(%make-arg ',name ,name)) - -;;; If Arg is null or its CONT is null, then return Default, otherwise -;;; return Arg's NAME. -(defun default-arg (arg default) - (declare (type (or arg null) arg)) - (if (and arg (arg-cont arg)) - (arg-name arg) - default)) - -;;; If Arg is null or has no CONT, return the default. Otherwise, Arg's -;;; CONT must be a constant continuation whose value we return. If not, we -;;; give up. -(defun arg-constant-value (arg default) - (declare (type (or arg null) arg)) - (if (and arg (arg-cont arg)) - (let ((cont (arg-cont arg))) - (unless (constant-continuation-p cont) - (give-up-ir1-transform "Argument is not constant: ~S." - (arg-name arg))) - (continuation-value from-end)) - default)) - -;;; If Arg is a constant and is EQL to X, then return T, otherwise NIL. If -;;; Arg is NIL or its CONT is NIL, then compare to the default. -(defun arg-eql (arg default x) - (declare (type (or arg null) x)) - (if (and arg (arg-cont arg)) - (let ((cont (arg-cont arg))) - (and (constant-continuation-p cont) - (eql (continuation-value cont) x))) - (eql default x))) - -(defstruct (iterator (:copier nil)) - ;; The kind of iterator. - (kind nil (member :normal :result)) - ;; A list of LET* bindings to create the initial state. - (binds nil :type list) - ;; A list of declarations for Binds. - (decls nil :type list) - ;; A form that returns the current value. This may be set with SETF to set - ;; the current value. - (current (error "Must specify CURRENT.")) - ;; In a :NORMAL iterator, a form that tests whether there is a current value. - (done nil) - ;; In a :RESULT iterator, a form that truncates the result at the current - ;; position and returns it. - (result nil) - ;; A form that returns the initial total number of values. The result is - ;; undefined after NEXT has been evaluated. - (length (error "Must specify LENGTH.")) - ;; A form that advances the state to the next value. It is an error to call - ;; this when the iterator is Done. - (next (error "Must specify NEXT."))) - -;;; Type of an index var that can go negative (in the from-end case.) -(deftype neg-index () - `(integer -1 ,most-positive-fixnum)) - -;;; Return an ITERATOR structure describing how to iterate over an arbitrary -;;; sequence. Sequence is a variable bound to the sequence, and Type is the -;;; type of the sequence. If true, INDEX is a variable that should be bound to -;;; the index of the current element in the sequence. -;;; -;;; If we can't tell whether the sequence is a list or a vector, or whether -;;; the iteration is forward or backward, then GIVE-UP. -(defun make-sequence-iterator (sequence type &key start end from-end index) - (declare (symbol sequence) (type ctype type) - (type (or arg null) start end from-end) - (type (or symbol null) index)) - (let ((from-end (arg-constant-value from-end nil))) - (cond ((csubtypep type (specifier-type 'vector)) - (let* ((n-stop (gensym)) - (n-idx (or index (gensym))) - (start (default-arg 0 start)) - (end (default-arg `(length ,sequence) end))) - (make-iterator - :kind :normal - :binds `((,n-idx ,(if from-end `(1- ,end) ,start)) - (,n-stop ,(if from-end `(1- ,start) ,end))) - :decls `((type neg-index ,n-idx ,n-stop)) - :current `(aref ,sequence ,n-idx) - :done `(,(if from-end '<= '>=) ,n-idx ,n-stop) - :next `(setq ,n-idx - ,(if from-end `(1- ,n-idx) `(1+ ,n-idx))) - :length (if from-end - `(- ,n-idx ,n-stop) - `(- ,n-stop ,n-idx))))) - ((csubtypep type (specifier-type 'list)) - (let* ((n-stop (if (and end (not from-end)) (gensym) nil)) - (n-current (gensym)) - (start-p (not (arg-eql start 0 0))) - (end-p (not (arg-eql end nil nil))) - (start (default-arg start 0)) - (end (default-arg end nil))) - (make-iterator - :binds `((,n-current - ,(if from-end - (if (or start-p end-p) - `(nreverse (subseq ,sequence ,start - ,@(when end `(,end)))) - `(reverse ,sequence)) - (if start-p - `(nthcdr ,start ,sequence) - sequence))) - ,@(when n-stop - `((,n-stop (nthcdr (the index - (- ,end ,start)) - ,n-current)))) - ,@(when index - `((,index ,(if from-end `(1- ,end) start))))) - :kind :normal - :decls `((list ,n-current ,n-end) - ,@(when index `((type neg-index ,index)))) - :current `(car ,n-current) - :done `(eq ,n-current ,n-stop) - :length `(- ,(or end `(length ,sequence)) ,start) - :next `(progn - (setq ,n-current (cdr ,n-current)) - ,@(when index - `((setq ,n-idx - ,(if from-end - `(1- ,index) - `(1+ ,index))))))))) + :node node) + (let* ((type (lvar-type seq)) + (element-ctype (array-type-upgraded-element-type type)) + (element-type (type-specifier element-ctype)) + (saetp (unless (eq *wild-type* element-ctype) + (find-saetp-by-ctype element-ctype)))) + (cond ((eq *wild-type* element-ctype) + (delay-ir1-transform node :constraint) + `(vector-fill* seq item start end)) + ((and saetp (sb!vm::valid-bit-bash-saetp-p saetp)) + (let* ((n-bits (sb!vm:saetp-n-bits saetp)) + (basher-name (format nil "UB~D-BASH-FILL" n-bits)) + (basher (or (find-symbol basher-name + (load-time-value (find-package :sb!kernel))) + (abort-ir1-transform + "Unknown fill basher, please report to sbcl-devel: ~A" + basher-name))) + (kind (cond ((sb!vm:saetp-fixnum-p saetp) :tagged) + ((member element-type '(character base-char)) :char) + ((eq element-type 'single-float) :single-float) + #!+#.(cl:if (cl:= 64 sb!vm:n-word-bits) '(and) '(or)) + ((eq element-type 'double-float) :double-float) + #!+#.(cl:if (cl:= 64 sb!vm:n-word-bits) '(and) '(or)) + ((equal element-type '(complex single-float)) + :complex-single-float) + (t + (aver (integer-type-p element-ctype)) + :bits))) + ;; BASH-VALUE is a word that we can repeatedly smash + ;; on the array: for less-than-word sized elements it + ;; contains multiple copies of the fill item. + (bash-value + (if (constant-lvar-p item) + (let ((tmp (lvar-value item))) + (unless (ctypep tmp element-ctype) + (abort-ir1-transform "~S is not ~S" tmp element-type)) + (let* ((bits + (ldb (byte n-bits 0) + (ecase kind + (:tagged + (ash tmp sb!vm:n-fixnum-tag-bits)) + (:char + (char-code tmp)) + (:bits + tmp) + (:single-float + (single-float-bits tmp)) + #!+#.(cl:if (cl:= 64 sb!vm:n-word-bits) '(and) '(or)) + (:double-float + (logior (ash (double-float-high-bits tmp) 32) + (double-float-low-bits tmp))) + #!+#.(cl:if (cl:= 64 sb!vm:n-word-bits) '(and) '(or)) + (:complex-single-float + (logior (ash (single-float-bits (imagpart tmp)) 32) + (ldb (byte 32 0) + (single-float-bits (realpart tmp)))))))) + (res bits)) + (loop for i of-type sb!vm:word from n-bits by n-bits + until (= i sb!vm:n-word-bits) + do (setf res (ldb (byte sb!vm:n-word-bits 0) + (logior res (ash bits i))))) + res)) + (progn + (delay-ir1-transform node :constraint) + `(let* ((bits (ldb (byte ,n-bits 0) + ,(ecase kind + (:tagged + `(ash item ,sb!vm:n-fixnum-tag-bits)) + (:char + `(char-code item)) + (:bits + `item) + (:single-float + `(single-float-bits item)) + #!+#.(cl:if (cl:= 64 sb!vm:n-word-bits) '(and) '(or)) + (:double-float + `(logior (ash (double-float-high-bits item) 32) + (double-float-low-bits item))) + #!+#.(cl:if (cl:= 64 sb!vm:n-word-bits) '(and) '(or)) + (:complex-single-float + `(logior (ash (single-float-bits (imagpart item)) 32) + (ldb (byte 32 0) + (single-float-bits (realpart item)))))))) + (res bits)) + (declare (type sb!vm:word res)) + ,@(unless (= sb!vm:n-word-bits n-bits) + `((loop for i of-type sb!vm:word from ,n-bits by ,n-bits + until (= i sb!vm:n-word-bits) + do (setf res + (ldb (byte ,sb!vm:n-word-bits 0) + (logior res (ash bits (truly-the (integer 0 ,(- sb!vm:n-word-bits n-bits)) i)))))))) + res))))) + (values + ;; KLUDGE: WITH-ARRAY data in its full glory is going to mess up + ;; dynamic-extent for MAKE-ARRAY :INITIAL-ELEMENT initialization. + (if (csubtypep (lvar-type seq) (specifier-type '(simple-array * (*)))) + `(let* ((len (length seq)) + (end (or end len)) + (bound (1+ end))) + ;; Minor abuse %CHECK-BOUND for bounds checking. + ;; (- END START) may still end up negative, but + ;; the basher handle that. + (,basher ,bash-value seq + (%check-bound seq bound start) + (- (if end (%check-bound seq bound end) len) + start))) + `(with-array-data ((data seq) + (start start) + (end end) + :check-fill-pointer t) + (declare (type (simple-array ,element-type 1) data)) + (declare (type index start end)) + (declare (optimize (safety 0) (speed 3))) + (,basher ,bash-value data start (- end start)) + seq)) + `((declare (type ,element-type item)))))) + ((policy node (> speed space)) + (values + `(with-array-data ((data seq) + (start start) + (end end) + :check-fill-pointer t) + (declare (type (simple-array ,element-type 1) data)) + (declare (type index start end)) + ;; WITH-ARRAY-DATA did our range checks once and for all, so + ;; it'd be wasteful to check again on every AREF... + (declare (optimize (safety 0) (speed 3))) + (do ((i start (1+ i))) + ((= i end) seq) + (declare (type index i)) + (setf (aref data i) item))) + ;; ... though we still need to check that the new element can fit + ;; into the vector in safe code. -- CSR, 2002-07-05 + `((declare (type ,element-type item))))) + ((csubtypep type (specifier-type 'string)) + '(string-fill* seq item start end)) (t - (give-up-ir1-transform - "can't tell whether sequence is a list or a vector"))))) - -;;; Make an iterator used for constructing result sequences. Name is a -;;; variable to be bound to the result sequence. Type is the type of result -;;; sequence to make. Length is an expression to be evaluated to get the -;;; maximum length of the result (not evaluated in list case.) -(defun make-result-sequence-iterator (name type length) - (declare (symbol name) (type ctype type)) - -;;; Define each NAME as a local macro that will call the value of the -;;; function arg with the given arguments. If the argument isn't known to be a -;;; function, give them an efficiency note and reference a coerced version. -(defmacro coerce-funs (specs &body body) - #!+sb-doc - "COERCE-FUNCTIONS ({(Name Fun-Arg Default)}*) Form*" - (collect ((binds) - (defs)) - (dolist (spec specs) - `(let ((body (progn ,@body)) - (n-fun (arg-name ,(second spec))) - (fun-cont (arg-cont ,(second spec)))) - (cond ((not fun-cont) - `(macrolet ((,',(first spec) (&rest args) - `(,',',(third spec) ,@args))) - ,body)) - ((not (csubtypep (continuation-type fun-cont) - (specifier-type 'function))) - (when (policy *compiler-error-context* - (> speed inhibit-warnings)) - (compiler-notify - "~S may not be a function, so must coerce at run-time." - n-fun)) - (once-only ((n-fun `(if (functionp ,n-fun) - ,n-fun - (symbol-function ,n-fun)))) - `(macrolet ((,',(first spec) (&rest args) - `(funcall ,',n-fun ,@args))) - ,body))) - (t - `(macrolet ((,',(first spec) (&rest args) - `(funcall ,',n-fun ,@args))) - ,body))))))) - -;;; Wrap code around the result of the body to define Name as a local macro -;;; that returns true when its arguments satisfy the test according to the Args -;;; Test and Test-Not. If both Test and Test-Not are supplied, abort the -;;; transform. -(defmacro with-sequence-test ((name test test-not) &body body) - `(let ((not-p (arg-cont ,test-not))) - (when (and (arg-cont ,test) not-p) - (abort-ir1-transform "Both ~S and ~S were supplied." - (arg-name ,test) - (arg-name ,test-not))) - (coerce-funs ((,name (if not-p ,test-not ,test) eql)) - ,@body))) -|# + '(vector-fill* seq item start end))))) + +(deftransform fill ((seq item &key (start 0) (end nil)) + ((and sequence (not vector) (not list)) t &key (:start t) (:end t))) + `(sb!sequence:fill seq item + :start start + :end (%check-generic-sequence-bounds seq start end))) ;;;; hairy sequence transforms ;;; FIXME: no hairy sequence transforms in SBCL? +;;; +;;; There used to be a bunch of commented out code about here, +;;; containing the (apparent) beginning of hairy sequence transform +;;; infrastructure. People interested in implementing better sequence +;;; transforms might want to look at it for inspiration, even though +;;; the actual code is ancient CMUCL -- and hence bitrotted. The code +;;; was deleted in 1.0.7.23. ;;;; string operations @@ -670,20 +773,6 @@ (def!constant vector-data-bit-offset (* sb!vm:vector-data-offset sb!vm:n-word-bits)) -(eval-when (:compile-toplevel) -(defun valid-bit-bash-saetp-p (saetp) - ;; BIT-BASHing isn't allowed on simple vectors that contain pointers - (and (not (eq t (sb!vm:saetp-specifier saetp))) - ;; Disallowing (VECTOR NIL) also means that we won't transform - ;; sequence functions into bit-bashing code and we let the - ;; generic sequence functions signal errors if necessary. - (not (zerop (sb!vm:saetp-n-bits saetp))) - ;; Due to limitations with the current BIT-BASHing code, we can't - ;; BIT-BASH reliably on arrays whose element types are larger - ;; than the word size. - (<= (sb!vm:saetp-n-bits saetp) sb!vm:n-word-bits))) -) ; EVAL-WHEN - ;;; FIXME: In the copy loops below, we code the loops in a strange ;;; fashion: ;;; @@ -721,60 +810,67 @@ ;;; you tweak it, make sure that you compare the disassembly, if not the ;;; performance of, the functions implementing string streams ;;; (e.g. SB!IMPL::STRING-OUCH). +(eval-when (#-sb-xc :compile-toplevel :load-toplevel :execute) + (defun make-replace-transform (saetp sequence-type1 sequence-type2) + `(deftransform replace ((seq1 seq2 &key (start1 0) (start2 0) end1 end2) + (,sequence-type1 ,sequence-type2 &rest t) + ,sequence-type1 + :node node) + `(let* ((len1 (length seq1)) + (len2 (length seq2)) + (end1 (or end1 len1)) + (end2 (or end2 len2)) + (replace-len (min (- end1 start1) (- end2 start2)))) + ,(unless (policy node (= insert-array-bounds-checks 0)) + `(progn + (unless (<= 0 start1 end1 len1) + (sequence-bounding-indices-bad-error seq1 start1 end1)) + (unless (<= 0 start2 end2 len2) + (sequence-bounding-indices-bad-error seq2 start2 end2)))) + ,',(cond + ((and saetp (sb!vm:valid-bit-bash-saetp-p saetp)) + (let* ((n-element-bits (sb!vm:saetp-n-bits saetp)) + (bash-function (intern (format nil "UB~D-BASH-COPY" + n-element-bits) + (find-package "SB!KERNEL")))) + `(funcall (function ,bash-function) seq2 start2 + seq1 start1 replace-len))) + (t + `(if (and + ;; If the sequence types are different, SEQ1 and + ;; SEQ2 must be distinct arrays. + ,(eql sequence-type1 sequence-type2) + (eq seq1 seq2) (> start1 start2)) + (do ((i (truly-the index (+ start1 replace-len -1)) + (1- i)) + (j (truly-the index (+ start2 replace-len -1)) + (1- j))) + ((< i start1)) + (declare (optimize (insert-array-bounds-checks 0))) + (setf (aref seq1 i) (aref seq2 j))) + (do ((i start1 (1+ i)) + (j start2 (1+ j)) + (end (+ start1 replace-len))) + ((>= i end)) + (declare (optimize (insert-array-bounds-checks 0))) + (setf (aref seq1 i) (aref seq2 j)))))) + seq1)))) + (macrolet ((define-replace-transforms () (loop for saetp across sb!vm:*specialized-array-element-type-properties* for sequence-type = `(simple-array ,(sb!vm:saetp-specifier saetp) (*)) unless (= (sb!vm:saetp-typecode saetp) sb!vm::simple-array-nil-widetag) - collect - `(deftransform replace ((seq1 seq2 &key (start1 0) (start2 0) end1 end2) - (,sequence-type ,sequence-type &rest t) - ,sequence-type - :node node) - ,(cond - ((valid-bit-bash-saetp-p saetp) nil) - ;; If we're not bit-bashing, only allow cases where we - ;; can determine the order of copying up front. (There - ;; are actually more cases we can handle if we know the - ;; amount that we're copying, but this handles the - ;; common cases.) - (t '(unless (= (constant-value-or-lose start1 0) - (constant-value-or-lose start2 0)) - (give-up-ir1-transform)))) - `(let* ((len1 (length seq1)) - (len2 (length seq2)) - (end1 (or end1 len1)) - (end2 (or end2 len2)) - (replace-len1 (- end1 start1)) - (replace-len2 (- end2 start2))) - ,(unless (policy node (= safety 0)) - `(progn - (unless (<= 0 start1 end1 len1) - (sb!impl::signal-bounding-indices-bad-error seq1 start1 end1)) - (unless (<= 0 start2 end2 len2) - (sb!impl::signal-bounding-indices-bad-error seq2 start2 end2)))) - ,',(cond - ((valid-bit-bash-saetp-p saetp) - (let* ((n-element-bits (sb!vm:saetp-n-bits saetp)) - (bash-function (intern (format nil "UB~D-BASH-COPY" n-element-bits) - (find-package "SB!KERNEL")))) - `(funcall (function ,bash-function) seq2 start2 - seq1 start1 (min replace-len1 replace-len2)))) - (t - ;; We can expand the loop inline here because we - ;; would have given up the transform (see above) - ;; if we didn't have constant matching start - ;; indices. - '(do ((i start1 (1+ i)) - (end (+ start1 - (min replace-len1 replace-len2)))) - ((>= i end)) - (declare (optimize (insert-array-bounds-checks 0))) - (setf (aref seq1 i) (aref seq2 i))))) - seq1)) + collect (make-replace-transform saetp sequence-type sequence-type) into forms - finally (return `(progn ,@forms))))) - (define-replace-transforms)) + finally (return `(progn ,@forms)))) + (define-one-transform (sequence-type1 sequence-type2) + (make-replace-transform nil sequence-type1 sequence-type2))) + (define-replace-transforms) + #!+sb-unicode + (progn + (define-one-transform (simple-array base-char (*)) (simple-array character (*))) + (define-one-transform (simple-array character (*)) (simple-array base-char (*))))) ;;; Expand simple cases of UB-BASH-COPY inline. "simple" is ;;; defined as those cases where we are doing word-aligned copies from @@ -839,7 +935,8 @@ (do ((i end (1- i))) ((<= i ,src-word)) (setf (sb!kernel:%vector-raw-bits dst (1- i)) - (sb!kernel:%vector-raw-bits src (1- i))))))))) + (sb!kernel:%vector-raw-bits src (1- i)))) + (values)))))) #.(loop for i = 1 then (* i 2) collect `(deftransform ,(intern (format nil "UB~D-BASH-COPY" i) @@ -890,39 +987,70 @@ ((<= i ,src-offset)) (declare (optimize (insert-array-bounds-checks 0))) (setf (aref ,dst (1- i)) (aref ,src (1- i)))) - `(do ((i (+ ,src-offset ,length) (1- i)) + ;; KLUDGE: The compiler is not able to derive that (+ offset + ;; length) must be a fixnum, but arrives at (unsigned-byte 29). + ;; We, however, know it must be so, as by this point the bounds + ;; have already been checked. + `(do ((i (truly-the fixnum (+ ,src-offset ,length)) (1- i)) (j (+ ,dst-offset ,length) (1- j))) ((<= i ,src-offset)) - (declare (optimize (insert-array-bounds-checks 0))) + (declare (optimize (insert-array-bounds-checks 0)) + (type (integer 0 #.sb!xc:array-dimension-limit) j i)) (setf (aref ,dst (1- j)) (aref ,src (1- i)))))) +;;; SUBSEQ, COPY-SEQ + +(deftransform subseq ((seq start &optional end) + (vector t &optional t) + * + :node node) + (let ((type (lvar-type seq))) + (cond + ((and (array-type-p type) + (csubtypep type (specifier-type '(or (simple-unboxed-array (*)) simple-vector))) + (policy node (> speed space))) + (let ((element-type (type-specifier (array-type-specialized-element-type type)))) + `(let* ((length (length seq)) + (end (or end length))) + ,(unless (policy node (zerop insert-array-bounds-checks)) + '(progn + (unless (<= 0 start end length) + (sequence-bounding-indices-bad-error seq start end)))) + (let* ((size (- end start)) + (result (make-array size :element-type ',element-type))) + ,(maybe-expand-copy-loop-inline 'seq (if (constant-lvar-p start) + (lvar-value start) + 'start) + 'result 0 'size element-type) + result)))) + (t + '(vector-subseq* seq start end))))) + +(deftransform subseq ((seq start &optional end) + (list t &optional t)) + `(list-subseq* seq start end)) + (deftransform subseq ((seq start &optional end) - ((or (simple-unboxed-array (*)) simple-vector) t &optional t) - * :node node) - (let ((array-type (lvar-type seq))) - (unless (array-type-p array-type) - (give-up-ir1-transform)) - (let ((element-type (type-specifier (array-type-specialized-element-type array-type)))) - `(let* ((length (length seq)) - (end (or end length))) - ,(unless (policy node (= safety 0)) - '(progn - (unless (<= 0 start end length) - (sb!impl::signal-bounding-indices-bad-error seq start end)))) - (let* ((size (- end start)) - (result (make-array size :element-type ',element-type))) - ,(maybe-expand-copy-loop-inline 'seq 'start 'result 0 'size element-type) - result))))) - -(deftransform copy-seq ((seq) ((or (simple-unboxed-array (*)) simple-vector)) *) - (let ((array-type (lvar-type seq))) - (unless (array-type-p array-type) - (give-up-ir1-transform)) - (let ((element-type (type-specifier (array-type-specialized-element-type array-type)))) - `(let* ((length (length seq)) - (result (make-array length :element-type ',element-type))) - ,(maybe-expand-copy-loop-inline 'seq 0 'result 0 'length element-type) - result)))) + ((and sequence (not vector) (not list)) t &optional t)) + '(sb!sequence:subseq seq start end)) + +(deftransform copy-seq ((seq) (vector)) + (let ((type (lvar-type seq))) + (cond ((and (array-type-p type) + (csubtypep type (specifier-type '(or (simple-unboxed-array (*)) simple-vector)))) + (let ((element-type (type-specifier (array-type-specialized-element-type type)))) + `(let* ((length (length seq)) + (result (make-array length :element-type ',element-type))) + ,(maybe-expand-copy-loop-inline 'seq 0 'result 0 'length element-type) + result))) + (t + '(vector-subseq* seq 0 nil))))) + +(deftransform copy-seq ((seq) (list)) + '(list-copy-seq* seq)) + +(deftransform copy-seq ((seq) ((and sequence (not vector) (not list)))) + '(sb!sequence:copy-seq seq)) ;;; FIXME: it really should be possible to take advantage of the ;;; macros used in code/seq.lisp here to avoid duplication of code, @@ -933,103 +1061,192 @@ from-end) (vector vector &rest t) * + :node node :policy (> speed (max space safety))) "open code" - (let ((from-end (when (lvar-p from-end) - (unless (constant-lvar-p from-end) - (give-up-ir1-transform ":FROM-END is not constant.")) - (lvar-value from-end))) - (keyp (lvar-p key)) - (testp (lvar-p test))) - `(block search - (let ((end1 (or end1 (length pattern))) - (end2 (or end2 (length text))) - ,@(when keyp - '((key (coerce key 'function)))) - ,@(when testp - '((test (coerce test 'function))))) - (declare (type index start1 start2 end1 end2)) - (do (,(if from-end - '(index2 (- end2 (- end1 start1)) (1- index2)) - '(index2 start2 (1+ index2)))) - (,(if from-end - '(< index2 start2) - '(>= index2 end2)) - nil) - ;; INDEX2 is FIXNUM, not an INDEX, as right before the loop - ;; terminates is hits -1 when :FROM-END is true and :START2 - ;; is 0. - (declare (type fixnum index2)) - (when (do ((index1 start1 (1+ index1)) - (index2 index2 (1+ index2))) - ((>= index1 end1) t) - (declare (type index index1 index2)) - ,@(unless from-end - '((when (= index2 end2) - (return-from search nil)))) - (unless (,@(if testp - '(funcall test) - '(eql)) - ,(if keyp - '(funcall key (aref pattern index1)) - '(aref pattern index1)) - ,(if keyp - '(funcall key (aref text index2)) - '(aref text index2))) - (return nil))) - (return index2))))))) - -;;; FIXME: It seems as though it should be possible to make a DEFUN -;;; %CONCATENATE (with a DEFTRANSFORM to translate constant RTYPE to -;;; CTYPE before calling %CONCATENATE) which is comparably efficient, -;;; at least once DYNAMIC-EXTENT works. + (flet ((maybe (x) + (when (lvar-p x) + (if (constant-lvar-p x) + (when (lvar-value x) + :yes) + :maybe)))) + (let ((from-end (when (lvar-p from-end) + (unless (constant-lvar-p from-end) + (give-up-ir1-transform ":FROM-END is not constant.")) + (lvar-value from-end))) + (key? (maybe key)) + (test? (maybe test)) + (check-bounds-p (policy node (plusp insert-array-bounds-checks)))) + `(block search + (flet ((oops (vector start end) + (sequence-bounding-indices-bad-error vector start end))) + (let* ((len1 (length pattern)) + (len2 (length text)) + (end1 (or end1 len1)) + (end2 (or end2 len2)) + ,@(case key? + (:yes `((key (%coerce-callable-to-fun key)))) + (:maybe `((key (when key + (%coerce-callable-to-fun key)))))) + ,@(when test? + `((test (%coerce-callable-to-fun test))))) + (declare (type index start1 start2 end1 end2)) + ,@(when check-bounds-p + `((unless (<= start1 end1 len1) + (oops pattern start1 end1)) + (unless (<= start2 end2 len2) + (oops pattern start2 end2)))) + (when (= end1 start1) + (return-from search (if from-end + end2 + start2))) + (do (,(if from-end + '(index2 (- end2 (- end1 start1)) (1- index2)) + '(index2 start2 (1+ index2)))) + (,(if from-end + '(< index2 start2) + '(>= index2 end2)) + nil) + ;; INDEX2 is FIXNUM, not an INDEX, as right before the loop + ;; terminates is hits -1 when :FROM-END is true and :START2 + ;; is 0. + (declare (type fixnum index2)) + (when (do ((index1 start1 (1+ index1)) + (index2 index2 (1+ index2))) + ((>= index1 end1) t) + (declare (type index index1 index2) + (optimize (insert-array-bounds-checks 0))) + ,@(unless from-end + '((when (= index2 end2) + (return-from search nil)))) + (unless (,@(if test? + `(funcall test) + `(eql)) + ,(case key? + (:yes `(funcall key (aref pattern index1))) + (:maybe `(let ((elt (aref pattern index1))) + (if key + (funcall key elt) + elt))) + (otherwise `(aref pattern index1))) + ,(case key? + (:yes `(funcall key (aref text index2))) + (:maybe `(let ((elt (aref text index2))) + (if key + (funcall key elt) + elt))) + (otherwise `(aref text index2)))) + (return nil))) + (return index2))))))))) + + +;;; Open-code CONCATENATE for strings. It would be possible to extend +;;; this transform to non-strings, but I chose to just do the case that +;;; should cover 95% of CONCATENATE performance complaints for now. +;;; -- JES, 2007-11-17 ;;; -;;; FIXME: currently KLUDGEed because of bug 188 +;;; Only handle the simple result type cases. If somebody does (CONCATENATE +;;; '(STRING 6) ...) their code won't be optimized, but nobody does that in +;;; practice. ;;; -;;; FIXME: disabled for sb-unicode: probably want it back -#!-sb-unicode -(deftransform concatenate ((rtype &rest sequences) - (t &rest (or simple-base-string - (simple-array nil (*)))) - simple-base-string - :policy (< safety 3)) - (loop for rest-seqs on sequences - for n-seq = (gensym "N-SEQ") - for n-length = (gensym "N-LENGTH") - for start = 0 then next-start - for next-start = (gensym "NEXT-START") - collect n-seq into args - collect `(,n-length (length ,n-seq)) into lets - collect n-length into all-lengths - collect next-start into starts - collect `(if (and (typep ,n-seq '(simple-array nil (*))) - (> ,n-length 0)) - (error 'nil-array-accessed-error) - (#.(let* ((i (position 'character sb!kernel::*specialized-array-element-types*)) - (saetp (aref sb!vm:*specialized-array-element-type-properties* i)) - (n-bits (sb!vm:saetp-n-bits saetp))) - (intern (format nil "UB~D-BASH-COPY" n-bits) - "SB!KERNEL")) - ,n-seq 0 res ,start ,n-length)) - into forms - collect `(setq ,next-start (+ ,start ,n-length)) into forms - finally - (return - `(lambda (rtype ,@args) - (declare (ignore rtype)) - (let* (,@lets - (res (make-string (the index (+ ,@all-lengths)) - :element-type 'base-char))) - (declare (type index ,@all-lengths)) - (let (,@(mapcar (lambda (name) `(,name 0)) starts)) - (declare (type index ,@starts)) - ,@forms) - res))))) +;;; Limit full open coding based on length of constant sequences. Default +;;; value is chosen so that other parts of the compiler (constraint propagation +;;; mainly) won't go nonlinear too badly. It's not an exact number -- but +;;; in the right ballpark. +(defvar *concatenate-open-code-limit* 129) + +(deftransform concatenate ((result-type &rest lvars) + ((constant-arg + (member string simple-string base-string simple-base-string)) + &rest sequence) + * :node node) + (let ((vars (loop for x in lvars collect (gensym))) + (type (lvar-value result-type))) + (if (policy node (<= speed space)) + ;; Out-of-line + `(lambda (.dummy. ,@vars) + (declare (ignore .dummy.)) + ,(ecase type + ((string simple-string) + `(%concatenate-to-string ,@vars)) + ((base-string simple-base-string) + `(%concatenate-to-base-string ,@vars)))) + ;; Inline + (let* ((element-type (ecase type + ((string simple-string) 'character) + ((base-string simple-base-string) 'base-char))) + (lvar-values (loop for lvar in lvars + collect (when (constant-lvar-p lvar) + (lvar-value lvar)))) + (lengths + (loop for value in lvar-values + for var in vars + collect (if value + (length value) + `(sb!impl::string-dispatch ((simple-array * (*)) + sequence) + ,var + (declare (muffle-conditions compiler-note)) + (length ,var))))) + (non-constant-start + (loop for value in lvar-values + while (and (stringp value) + (< (length value) *concatenate-open-code-limit*)) + sum (length value)))) + `(apply + (lambda ,vars + (declare (ignorable ,@vars)) + (declare (optimize (insert-array-bounds-checks 0))) + (let* ((.length. (+ ,@lengths)) + (.pos. ,non-constant-start) + (.string. (make-string .length. :element-type ',element-type))) + (declare (type index .length. .pos.) + (muffle-conditions compiler-note)) + ,@(loop with first-constants = t + for first = t then nil + for value in lvar-values + for var in vars + collect + (cond ((and (stringp value) + (< (length value) *concatenate-open-code-limit*)) + ;; Fold the array reads for constant arguments + `(progn + ,@(loop for c across value + for i from 0 + collect + ;; Without truly-the we get massive numbers + ;; of pointless error traps. + `(setf (aref .string. + (truly-the index ,(if first-constants + i + `(+ .pos. ,i)))) + ,c)) + ,(unless first-constants + `(incf (truly-the index .pos.) ,(length value))))) + (t + (prog1 + `(sb!impl::string-dispatch + (#!+sb-unicode + (simple-array character (*)) + (simple-array base-char (*)) + t) + ,var + (replace .string. ,var + ,@(cond ((not first-constants) + '(:start1 .pos.)) + ((plusp non-constant-start) + `(:start1 ,non-constant-start)))) + (incf (truly-the index .pos.) (length ,var))) + (setf first-constants nil))))) + .string.)) + lvars))))) ;;;; CONS accessor DERIVE-TYPE optimizers (defoptimizer (car derive-type) ((cons)) - (let ((type (lvar-type cons)) + ;; This and CDR needs to use LVAR-CONSERVATIVE-TYPE because type inference + ;; gets confused by things like (SETF CAR). + (let ((type (lvar-conservative-type cons)) (null-type (specifier-type 'null))) (cond ((eq type null-type) null-type) @@ -1037,7 +1254,7 @@ (cons-type-car-type type))))) (defoptimizer (cdr derive-type) ((cons)) - (let ((type (lvar-type cons)) + (let ((type (lvar-conservative-type cons)) (null-type (specifier-type 'null))) (cond ((eq type null-type) null-type) @@ -1072,41 +1289,49 @@ * :policy (> speed space)) "expand inline" - `(let ((index 0) - (find nil) + `(let ((find nil) (position nil)) - (declare (type index index)) - (dolist (i sequence - (if (and end (> end index)) - (sb!impl::signal-bounding-indices-bad-error - sequence start end) - (values find position))) - (let ((key-i (funcall key i))) - (when (and end (>= index end)) - (return (values find position))) - (when (>= index start) - (,',condition (funcall predicate key-i) - ;; This hack of dealing with non-NIL - ;; FROM-END for list data by iterating - ;; forward through the list and keeping - ;; track of the last time we found a match - ;; might be more screwy than what the user - ;; expects, but it seems to be allowed by - ;; the ANSI standard. (And if the user is - ;; screwy enough to ask for FROM-END - ;; behavior on list data, turnabout is - ;; fair play.) - ;; - ;; It's also not enormously efficient, - ;; calling PREDICATE and KEY more often - ;; than necessary; but all the - ;; alternatives seem to have their own - ;; efficiency problems. - (if from-end - (setf find i - position index) - (return (values i index)))))) - (incf index)))))) + (flet ((bounds-error () + (sequence-bounding-indices-bad-error sequence start end))) + (if (and end (> start end)) + (bounds-error) + (do ((slow sequence (cdr slow)) + (fast (cdr sequence) (cddr fast)) + (index 0 (+ index 1))) + ((cond ((null slow) + (if (and end (> end index)) + (bounds-error) + (return (values find position)))) + ((and end (>= index end)) + (return (values find position))) + ((eq slow fast) + (circular-list-error sequence))) + (bug "never")) + (declare (list slow fast)) + (when (>= index start) + (let* ((element (car slow)) + (key-i (funcall key element))) + (,',condition (funcall predicate key-i) + ;; This hack of dealing with non-NIL + ;; FROM-END for list data by iterating + ;; forward through the list and keeping + ;; track of the last time we found a + ;; match might be more screwy than what + ;; the user expects, but it seems to be + ;; allowed by the ANSI standard. (And + ;; if the user is screwy enough to ask + ;; for FROM-END behavior on list data, + ;; turnabout is fair play.) + ;; + ;; It's also not enormously efficient, + ;; calling PREDICATE and KEY more often + ;; than necessary; but all the + ;; alternatives seem to have their own + ;; efficiency problems. + (if from-end + (setf find element + position index) + (return (values element index))))))))))))) (def %find-position-if when) (def %find-position-if-not unless)) @@ -1141,13 +1366,12 @@ end-arg element done-p-expr) - (with-unique-names (offset block index n-sequence sequence n-end end) - `(let ((,n-sequence ,sequence-arg) - (,n-end ,end-arg)) + (with-unique-names (offset block index n-sequence sequence end) + `(let* ((,n-sequence ,sequence-arg)) (with-array-data ((,sequence ,n-sequence :offset-var ,offset) (,start ,start) - (,end (%check-vector-sequence-bounds - ,n-sequence ,start ,n-end))) + (,end ,end-arg) + :check-fill-pointer t) (block ,block (macrolet ((maybe-return () ;; WITH-ARRAY-DATA has already performed bounds @@ -1155,10 +1379,10 @@ ;; in the inner loop. '(let ((,element (locally (declare (optimize (insert-array-bounds-checks 0))) (aref ,sequence ,index)))) - (when ,done-p-expr - (return-from ,block - (values ,element - (- ,index ,offset))))))) + (when ,done-p-expr + (return-from ,block + (values ,element + (- ,index ,offset))))))) (if ,from-end (loop for ,index ;; (If we aren't fastidious about declaring that @@ -1169,7 +1393,7 @@ from (1- ,end) downto ,start do (maybe-return)) (loop for ,index of-type index from ,start below ,end do - (maybe-return)))) + (maybe-return)))) (values nil nil)))))) (def!macro %find-position-vector-macro (item sequence @@ -1235,7 +1459,56 @@ "expand inline" (check-inlineability-of-find-position-if sequence from-end) '(%find-position-vector-macro item sequence - from-end start end key test)) + from-end start end key test)) + +(deftransform %find-position ((item sequence from-end start end key test) + (t bit-vector t t t t t) + * :node node) + (when (and test (lvar-fun-is test '(eq eql equal))) + (setf test nil)) + (when (and key (lvar-fun-is key '(identity))) + (setf key nil)) + (when (or test key) + (delay-ir1-transform node :optimize) + (give-up-ir1-transform "non-trivial :KEY or :TEST")) + (catch 'not-a-bit + `(with-array-data ((bits sequence :offset-var offset) + (start start) + (end end) + :check-fill-pointer t) + (let ((p ,(if (constant-lvar-p item) + (case (lvar-value item) + (0 `(%bit-position/0 bits from-end start end)) + (1 `(%bit-position/1 bits from-end start end)) + (otherwise (throw 'not-a-bit `(values nil nil)))) + `(%bit-position item bits from-end start end)))) + (if p + (values item (the index (- (truly-the index p) offset))) + (values nil nil)))))) + +(deftransform %find-position ((item sequence from-end start end key test) + (character string t t t function function) + * + :policy (> speed space)) + (if (eq '* (upgraded-element-type-specifier sequence)) + (let ((form + `(sb!impl::string-dispatch ((simple-array character (*)) + (simple-array base-char (*)) + (simple-array nil (*))) + sequence + (%find-position item sequence from-end start end key test)))) + (if (csubtypep (lvar-type sequence) (specifier-type 'simple-string)) + form + ;; Otherwise we'd get three instances of WITH-ARRAY-DATA from + ;; %FIND-POSITION. + `(with-array-data ((sequence sequence :offset-var offset) + (start start) + (end end) + :check-fill-pointer t) + (multiple-value-bind (elt index) ,form + (values elt (when (fixnump index) (- index offset))))))) + ;; The type is known exactly, other transforms will take care of it. + (give-up-ir1-transform))) ;;; logic to unravel :TEST, :TEST-NOT, and :KEY options in FIND, ;;; POSITION-IF, etc. @@ -1320,3 +1593,114 @@ (effective-find-position-key key)))))) (define-find-position-if-not find-if-not 0) (define-find-position-if-not position-if-not 1)) + +(macrolet ((define-trimmer-transform (fun-name leftp rightp) + `(deftransform ,fun-name ((char-bag string) + (t simple-string)) + (let ((find-expr + (if (constant-lvar-p char-bag) + ;; If the bag is constant, use MEMBER + ;; instead of FIND, since we have a + ;; deftransform for MEMBER that can + ;; open-code all of the comparisons when + ;; the list is constant. -- JES, 2007-12-10 + `(not (member (schar string index) + ',(coerce (lvar-value char-bag) 'list) + :test #'char=)) + '(not (find (schar string index) char-bag :test #'char=))))) + `(flet ((char-not-in-bag (index) + ,find-expr)) + (let* ((end (length string)) + (left-end (if ,',leftp + (do ((index 0 (1+ index))) + ((or (= index (the fixnum end)) + (char-not-in-bag index)) + index) + (declare (fixnum index))) + 0)) + (right-end (if ,',rightp + (do ((index (1- end) (1- index))) + ((or (< index left-end) + (char-not-in-bag index)) + (1+ index)) + (declare (fixnum index))) + end))) + (if (and (eql left-end 0) + (eql right-end (length string))) + string + (subseq string left-end right-end)))))))) + (define-trimmer-transform string-left-trim t nil) + (define-trimmer-transform string-right-trim nil t) + (define-trimmer-transform string-trim t t)) + + +;;; (partially) constant-fold backq-* functions, or convert to their +;;; plain CL equivalent (now that they're not needed for pprinting). + +;; Pop constant values from the end, list/list* them if any, and link +;; the remainder with list* at runtime. +(defun transform-backq-list-or-list* (function values) + (let ((gensyms (make-gensym-list (length values))) + (reverse (reverse values)) + (constants '())) + (loop while (and reverse + (constant-lvar-p (car reverse))) + do (push (lvar-value (pop reverse)) + constants)) + (if (null constants) + `(lambda ,gensyms + (,function ,@gensyms)) + (let ((tail (apply function constants))) + (if (null reverse) + `',tail + (let* ((nvariants (length reverse)) + (variants (subseq gensyms 0 nvariants))) + `(lambda ,gensyms + (declare (ignore ,@(subseq gensyms nvariants))) + ,(if tail + `(list* ,@variants ',tail) + `(list ,@variants))))))))) + +(deftransform sb!impl::backq-list ((&rest elts)) + (transform-backq-list-or-list* 'list elts)) + +(deftransform sb!impl::backq-list* ((&rest elts)) + (transform-backq-list-or-list* 'list* elts)) + +;; Merge adjacent constant values +(deftransform sb!impl::backq-append ((&rest elts)) + (let ((gensyms (make-gensym-list (length elts))) + (acc nil) + (ignored '()) + (arguments '())) + (flet ((convert-accumulator () + (let ((constant (apply 'append (nreverse (shiftf acc nil))))) + (when constant + (push `',constant arguments))))) + (loop for gensym in gensyms + for (elt . next) on elts by #'cdr + do (cond ((constant-lvar-p elt) + (let ((elt (lvar-value elt))) + (when (and next (not (proper-list-p elt))) + (abort-ir1-transform + "Non-list or improper list spliced in ~ + the middle of a backquoted list.")) + (push gensym ignored) + (push elt acc))) + (t + (convert-accumulator) + (push gensym arguments))) + finally (convert-accumulator))) + (let ((arguments (nreverse arguments))) + `(lambda ,gensyms + (declare (ignore ,@ignored)) + (append ,@arguments))))) + +;; Nothing special for nconc +(define-source-transform sb!impl::backq-nconc (&rest elts) + `(nconc ,@elts)) + +;; cons and vector are handled with regular constant folding... +;; but we still want to convert backq-cons into cl:cons. +(deftransform sb!impl::backq-cons ((x y)) + `(cons x y))