1 ;;;; optimizers for list and sequence functions
3 ;;;; This software is part of the SBCL system. See the README file for
6 ;;;; This software is derived from the CMU CL system, which was
7 ;;;; written at Carnegie Mellon University and released into the
8 ;;;; public domain. The software is in the public domain and is
9 ;;;; provided with absolutely no warranty. See the COPYING and CREDITS
10 ;;;; files for more information.
14 ;;;; mapping onto lists: the MAPFOO functions
16 (defun mapfoo-transform (fn arglists accumulate take-car)
17 (collect ((do-clauses)
20 (let ((n-first (gensym)))
21 (dolist (a (if accumulate
23 `(,n-first ,@(rest arglists))))
25 (do-clauses `(,v ,a (cdr ,v)))
27 (args-to-fn (if take-car `(car ,v) v))))
29 (let* ((fn-sym (gensym)) ; for ONCE-ONLY-ish purposes
30 (call `(funcall ,fn-sym . ,(args-to-fn)))
31 (endtest `(or ,@(tests))))
35 (map-result (gensym)))
37 (,map-result (list nil)))
38 (do-anonymous ((,temp ,map-result) . ,(do-clauses))
39 (,endtest (cdr ,map-result))
40 (setq ,temp (last (nconc ,temp ,call)))))))
43 (map-result (gensym)))
45 (,map-result (list nil)))
46 (do-anonymous ((,temp ,map-result) . ,(do-clauses))
47 (,endtest (truly-the list (cdr ,map-result)))
48 (rplacd ,temp (setq ,temp (list ,call)))))))
51 (,n-first ,(first arglists)))
52 (do-anonymous ,(do-clauses)
53 (,endtest (truly-the list ,n-first))
56 (define-source-transform mapc (function list &rest more-lists)
57 (mapfoo-transform function (cons list more-lists) nil t))
59 (define-source-transform mapcar (function list &rest more-lists)
60 (mapfoo-transform function (cons list more-lists) :list t))
62 (define-source-transform mapcan (function list &rest more-lists)
63 (mapfoo-transform function (cons list more-lists) :nconc t))
65 (define-source-transform mapl (function list &rest more-lists)
66 (mapfoo-transform function (cons list more-lists) nil nil))
68 (define-source-transform maplist (function list &rest more-lists)
69 (mapfoo-transform function (cons list more-lists) :list nil))
71 (define-source-transform mapcon (function list &rest more-lists)
72 (mapfoo-transform function (cons list more-lists) :nconc nil))
74 ;;;; mapping onto sequences: the MAP function
76 ;;; MAP is %MAP plus a check to make sure that any length specified in
77 ;;; the result type matches the actual result. We also wrap it in a
78 ;;; TRULY-THE for the most specific type we can determine.
79 (deftransform map ((result-type-arg fun seq &rest seqs) * * :node node)
80 (let* ((seq-names (make-gensym-list (1+ (length seqs))))
81 (bare `(%map result-type-arg fun ,@seq-names))
82 (constant-result-type-arg-p (constant-lvar-p result-type-arg))
83 ;; what we know about the type of the result. (Note that the
84 ;; "result type" argument is not necessarily the type of the
85 ;; result, since NIL means the result has NULL type.)
86 (result-type (if (not constant-result-type-arg-p)
88 (let ((result-type-arg-value
89 (lvar-value result-type-arg)))
90 (if (null result-type-arg-value)
92 result-type-arg-value)))))
93 `(lambda (result-type-arg fun ,@seq-names)
94 (truly-the ,result-type
95 ,(cond ((policy node (< safety 3))
96 ;; ANSI requires the length-related type check only
97 ;; when the SAFETY quality is 3... in other cases, we
98 ;; skip it, because it could be expensive.
100 ((not constant-result-type-arg-p)
101 `(sequence-of-checked-length-given-type ,bare
104 (let ((result-ctype (ir1-transform-specifier-type
106 (if (array-type-p result-ctype)
107 (let ((dims (array-type-dimensions result-ctype)))
108 (unless (and (listp dims) (= (length dims) 1))
109 (give-up-ir1-transform "invalid sequence type"))
110 (let ((dim (first dims)))
113 `(vector-of-checked-length-given-length ,bare
115 ;; FIXME: this is wrong, as not all subtypes of
116 ;; VECTOR are ARRAY-TYPEs [consider, for
117 ;; example, (OR (VECTOR T 3) (VECTOR T
118 ;; 4))]. However, it's difficult to see what we
119 ;; should put here... maybe we should
120 ;; GIVE-UP-IR1-TRANSFORM if the type is a
121 ;; subtype of VECTOR but not an ARRAY-TYPE?
124 ;;; Return a DO loop, mapping a function FUN to elements of
125 ;;; sequences. SEQS is a list of lvars, SEQ-NAMES - list of variables,
126 ;;; bound to sequences, INTO - a variable, which is used in
127 ;;; MAP-INTO. RESULT and BODY are forms, which can use variables
128 ;;; FUNCALL-RESULT, containing the result of application of FUN, and
129 ;;; INDEX, containing the current position in sequences.
130 (defun build-sequence-iterator (seqs seq-names &key result into body)
131 (declare (type list seqs seq-names)
138 (let ((found-vector-p nil))
139 (flet ((process-vector (length)
140 (unless found-vector-p
141 (setq found-vector-p t)
142 (bindings `(index 0 (1+ index)))
143 (declarations `(type index index)))
144 (vector-lengths length)))
145 (loop for seq of-type lvar in seqs
146 for seq-name in seq-names
147 for type = (lvar-type seq)
148 do (cond ((csubtypep type (specifier-type 'list))
149 (with-unique-names (index)
150 (bindings `(,index ,seq-name (cdr ,index)))
151 (declarations `(type list ,index))
152 (places `(car ,index))
153 (tests `(endp ,index))))
154 ((csubtypep type (specifier-type 'vector))
155 (process-vector `(length ,seq-name))
156 (places `(locally (declare (optimize (insert-array-bounds-checks 0)))
157 (aref ,seq-name index))))
159 (give-up-ir1-transform
160 "can't determine sequence argument type"))))
162 (process-vector `(array-dimension ,into 0))))
164 (bindings `(length (min ,@(vector-lengths))))
165 (tests `(>= index length)))
167 ((or ,@(tests)) ,result)
168 (declare ,@(declarations))
169 (let ((funcall-result (funcall fun ,@(places))))
170 (declare (ignorable funcall-result))
173 ;;; Try to compile %MAP efficiently when we can determine sequence
174 ;;; argument types at compile time.
176 ;;; Note: This transform was written to allow open coding of
177 ;;; quantifiers by expressing them in terms of (MAP NIL ..). For
178 ;;; non-NIL values of RESULT-TYPE, it's still useful, but not
179 ;;; necessarily as efficient as possible. In particular, it will be
180 ;;; inefficient when RESULT-TYPE is a SIMPLE-ARRAY with specialized
181 ;;; numeric element types. It should be straightforward to make it
182 ;;; handle that case more efficiently, but it's left as an exercise to
183 ;;; the reader, because the code is complicated enough already and I
184 ;;; don't happen to need that functionality right now. -- WHN 20000410
185 (deftransform %map ((result-type fun seq &rest seqs) * *
186 :policy (>= speed space))
188 (unless (constant-lvar-p result-type)
189 (give-up-ir1-transform "RESULT-TYPE argument not constant"))
190 (labels ( ;; 1-valued SUBTYPEP, fails unless second value of SUBTYPEP is true
191 (fn-1subtypep (fn x y)
192 (multiple-value-bind (subtype-p valid-p) (funcall fn x y)
195 (give-up-ir1-transform
196 "can't analyze sequence type relationship"))))
197 (1subtypep (x y) (fn-1subtypep #'sb!xc:subtypep x y)))
198 (let* ((result-type-value (lvar-value result-type))
199 (result-supertype (cond ((null result-type-value) 'null)
200 ((1subtypep result-type-value 'vector)
202 ((1subtypep result-type-value 'list)
205 (give-up-ir1-transform
206 "result type unsuitable")))))
207 (cond ((and result-type-value (null seqs))
208 ;; The consing arity-1 cases can be implemented
209 ;; reasonably efficiently as function calls, and the cost
210 ;; of consing should be significantly larger than
211 ;; function call overhead, so we always compile these
212 ;; cases as full calls regardless of speed-versus-space
213 ;; optimization policy.
214 (cond ((subtypep result-type-value 'list)
215 '(%map-to-list-arity-1 fun seq))
216 ( ;; (This one can be inefficient due to COERCE, but
217 ;; the current open-coded implementation has the
219 (subtypep result-type-value 'vector)
220 `(coerce (%map-to-simple-vector-arity-1 fun seq)
221 ',result-type-value))
222 (t (bug "impossible (?) sequence type"))))
224 (let* ((seqs (cons seq seqs))
225 (seq-args (make-gensym-list (length seqs))))
226 (multiple-value-bind (push-dacc result)
227 (ecase result-supertype
228 (null (values nil nil))
229 (list (values `(push funcall-result acc)
231 (vector (values `(push funcall-result acc)
232 `(coerce (nreverse acc)
233 ',result-type-value))))
234 ;; (We use the same idiom, of returning a LAMBDA from
235 ;; DEFTRANSFORM, as is used in the DEFTRANSFORMs for
236 ;; FUNCALL and ALIEN-FUNCALL, and for the same
237 ;; reason: we need to get the runtime values of each
238 ;; of the &REST vars.)
239 `(lambda (result-type fun ,@seq-args)
240 (declare (ignore result-type))
241 (let ((fun (%coerce-callable-to-fun fun))
243 (declare (type list acc))
244 (declare (ignorable acc))
245 ,(build-sequence-iterator
248 :body push-dacc))))))))))
251 (deftransform map-into ((result fun &rest seqs)
255 (let ((seqs-names (mapcar (lambda (x)
259 `(lambda (result fun ,@seqs-names)
260 ,(build-sequence-iterator
262 :result '(when (array-has-fill-pointer-p result)
263 (setf (fill-pointer result) index))
265 :body '(locally (declare (optimize (insert-array-bounds-checks 0)))
266 (setf (aref result index) funcall-result)))
270 ;;; FIXME: once the confusion over doing transforms with known-complex
271 ;;; arrays is over, we should also transform the calls to (AND (ARRAY
272 ;;; * (*)) (NOT (SIMPLE-ARRAY * (*)))) objects.
273 (deftransform elt ((s i) ((simple-array * (*)) *) *)
276 (deftransform elt ((s i) (list *) * :policy (< safety 3))
279 (deftransform %setelt ((s i v) ((simple-array * (*)) * *) *)
282 (deftransform %setelt ((s i v) (list * *) * :policy (< safety 3))
283 '(setf (car (nthcdr i s)) v))
285 (deftransform %check-vector-sequence-bounds ((vector start end)
288 (if (policy node (< safety speed))
289 '(or end (length vector))
290 '(let ((length (length vector)))
291 (if (<= 0 start (or end length) length)
293 (sb!impl::signal-bounding-indices-bad-error vector start end)))))
296 (deftransform member ((item list &key key test test-not) * * :node node)
297 ;; Key can legally be NIL, but if it's NIL for sure we pretend it's
298 ;; not there at all. If it might be NIL, make up a form to that
299 ;; ensure it is a function.
300 (multiple-value-bind (key key-form)
302 (let ((key-type (lvar-type key))
303 (null-type (specifier-type 'null)))
304 (cond ((csubtypep key-type null-type)
306 ((csubtypep null-type key-type)
308 (%coerce-callable-to-fun key)
311 (values key '(%coerce-callable-to-fun key))))))
312 (multiple-value-bind (out-of-line funs test-expr)
313 (cond ((and (not key) (not test) (not test-not))
317 ((and key (not test) (not test-not))
320 '(eql item (%funcall key car))))
322 (values '%member-key-test
324 '(%funcall test item (%funcall key car))))
326 (values '%member-key-test-not
328 '(not (%funcall test-not item (%funcall key car)))))
330 (values '%member-test
332 '(%funcall test item car)))
334 (values '%member-test-not
336 '(not (%funcall test item car))))
339 (labels ((open-code (tail)
341 `(if (let ((car ',(car tail)))
344 ,(open-code (cdr tail)))))
348 `(%coerce-callable-to-fun ,fun))))
349 (if (and (constant-lvar-p list) (policy node (>= speed space)))
350 `(let ,(mapcar (lambda (fun) `(,fun ,(ensure-fun fun))) funs)
351 ,(open-code (lvar-value list)))
352 `(,out-of-line item list ,@(mapcar #'ensure-fun funs)))))))
354 (deftransform memq ((item list) (t (constant-arg list)))
357 `(if (eq item ',(car tail))
361 (rec (lvar-value list))))
363 ;;; FIXME: We have rewritten the original code that used DOLIST to this
364 ;;; more natural MACROLET. However, the original code suggested that when
365 ;;; this was done, a few bytes could be saved by a call to a shared
366 ;;; function. This remains to be done.
367 (macrolet ((def (fun eq-fun)
368 `(deftransform ,fun ((item list &key test) (t list &rest t) *)
370 ;; FIXME: The scope of this transformation could be
371 ;; widened somewhat, letting it work whenever the test is
372 ;; 'EQL and we know from the type of ITEM that it #'EQ
373 ;; works like #'EQL on it. (E.g. types FIXNUM, CHARACTER,
375 ;; If TEST is EQ, apply transform, else
376 ;; if test is not EQL, then give up on transform, else
377 ;; if ITEM is not a NUMBER or is a FIXNUM, apply
378 ;; transform, else give up on transform.
380 (unless (lvar-fun-is test '(eq))
381 (give-up-ir1-transform)))
382 ((types-equal-or-intersect (lvar-type item)
383 (specifier-type 'number))
384 (give-up-ir1-transform "Item might be a number.")))
385 `(,',eq-fun item list))))
390 (deftransform delete-if ((pred list) (t list))
392 '(do ((x list (cdr x))
395 (cond ((funcall pred (car x))
398 (rplacd splice (cdr x))))
399 (t (setq splice x)))))
401 (deftransform fill ((seq item &key (start 0) (end (length seq)))
402 (vector t &key (:start t) (:end index))
404 :policy (> speed space))
406 (let ((element-type (upgraded-element-type-specifier-or-give-up seq)))
408 `(with-array-data ((data seq)
411 (declare (type (simple-array ,element-type 1) data))
412 (declare (type fixnum start end))
413 (do ((i start (1+ i)))
415 (declare (type index i))
416 ;; WITH-ARRAY-DATA did our range checks once and for all, so
417 ;; it'd be wasteful to check again on every AREF...
418 (declare (optimize (safety 0)))
419 (setf (aref data i) item)))
420 ;; ... though we still need to check that the new element can fit
421 ;; into the vector in safe code. -- CSR, 2002-07-05
422 `((declare (type ,element-type item))))))
426 ;;; Return true if LVAR's only use is a non-NOTINLINE reference to a
427 ;;; global function with one of the specified NAMES.
428 (defun lvar-fun-is (lvar names)
429 (declare (type lvar lvar) (list names))
430 (let ((use (lvar-uses lvar)))
432 (let ((leaf (ref-leaf use)))
433 (and (global-var-p leaf)
434 (eq (global-var-kind leaf) :global-function)
435 (not (null (member (leaf-source-name leaf) names
436 :test #'equal))))))))
438 ;;; If LVAR is a constant lvar, the return the constant value. If it
439 ;;; is null, then return default, otherwise quietly give up the IR1
442 ;;; ### Probably should take an ARG and flame using the NAME.
443 (defun constant-value-or-lose (lvar &optional default)
444 (declare (type (or lvar null) lvar))
445 (cond ((not lvar) default)
446 ((constant-lvar-p lvar)
449 (give-up-ir1-transform))))
452 ;;;; hairy sequence transforms
454 ;;; FIXME: no hairy sequence transforms in SBCL?
456 ;;; There used to be a bunch of commented out code about here,
457 ;;; containing the (apparent) beginning of hairy sequence transform
458 ;;; infrastructure. People interested in implementing better sequence
459 ;;; transforms might want to look at it for inspiration, even though
460 ;;; the actual code is ancient CMUCL -- and hence bitrotted. The code
461 ;;; was deleted in 1.0.7.23.
463 ;;;; string operations
465 ;;; We transform the case-sensitive string predicates into a non-keyword
466 ;;; version. This is an IR1 transform so that we don't have to worry about
467 ;;; changing the order of evaluation.
468 (macrolet ((def (fun pred*)
469 `(deftransform ,fun ((string1 string2 &key (start1 0) end1
472 `(,',pred* string1 string2 start1 end1 start2 end2))))
473 (def string< string<*)
474 (def string> string>*)
475 (def string<= string<=*)
476 (def string>= string>=*)
477 (def string= string=*)
478 (def string/= string/=*))
480 ;;; Return a form that tests the free variables STRING1 and STRING2
481 ;;; for the ordering relationship specified by LESSP and EQUALP. The
482 ;;; start and end are also gotten from the environment. Both strings
483 ;;; must be SIMPLE-BASE-STRINGs.
484 (macrolet ((def (name lessp equalp)
485 `(deftransform ,name ((string1 string2 start1 end1 start2 end2)
486 (simple-base-string simple-base-string t t t t) *)
487 `(let* ((end1 (if (not end1) (length string1) end1))
488 (end2 (if (not end2) (length string2) end2))
489 (index (sb!impl::%sp-string-compare
490 string1 start1 end1 string2 start2 end2)))
492 (cond ((= index end1)
493 ,(if ',lessp 'index nil))
494 ((= (+ index (- start2 start1)) end2)
495 ,(if ',lessp nil 'index))
496 ((,(if ',lessp 'char< 'char>)
497 (schar string1 index)
506 ,(if ',equalp 'end1 nil))))))
509 (def string>* nil nil)
510 (def string>=* nil t))
512 (macrolet ((def (name result-fun)
513 `(deftransform ,name ((string1 string2 start1 end1 start2 end2)
514 (simple-base-string simple-base-string t t t t) *)
516 (sb!impl::%sp-string-compare
517 string1 start1 (or end1 (length string1))
518 string2 start2 (or end2 (length string2)))))))
520 (def string/=* identity))
523 ;;;; transforms for sequence functions
525 ;;; Moved here from generic/vm-tran.lisp to satisfy clisp. Only applies
526 ;;; to vectors based on simple arrays.
527 (def!constant vector-data-bit-offset
528 (* sb!vm:vector-data-offset sb!vm:n-word-bits))
530 (eval-when (:compile-toplevel)
531 (defun valid-bit-bash-saetp-p (saetp)
532 ;; BIT-BASHing isn't allowed on simple vectors that contain pointers
533 (and (not (eq t (sb!vm:saetp-specifier saetp)))
534 ;; Disallowing (VECTOR NIL) also means that we won't transform
535 ;; sequence functions into bit-bashing code and we let the
536 ;; generic sequence functions signal errors if necessary.
537 (not (zerop (sb!vm:saetp-n-bits saetp)))
538 ;; Due to limitations with the current BIT-BASHing code, we can't
539 ;; BIT-BASH reliably on arrays whose element types are larger
540 ;; than the word size.
541 (<= (sb!vm:saetp-n-bits saetp) sb!vm:n-word-bits)))
544 ;;; FIXME: In the copy loops below, we code the loops in a strange
547 ;;; (do ((i (+ src-offset length) (1- i)))
549 ;;; (... (aref foo (1- i)) ...))
551 ;;; rather than the more natural (and seemingly more efficient):
553 ;;; (do ((i (1- (+ src-offset length)) (1- i)))
555 ;;; (... (aref foo i) ...))
557 ;;; (more efficient because we don't have to do the index adjusting on
558 ;;; every iteration of the loop)
560 ;;; We do this to avoid a suboptimality in SBCL's backend. In the
561 ;;; latter case, the backend thinks I is a FIXNUM (which it is), but
562 ;;; when used as an array index, the backend thinks I is a
563 ;;; POSITIVE-FIXNUM (which it is). However, since the backend thinks of
564 ;;; these as distinct storage classes, it cannot coerce a move from a
565 ;;; FIXNUM TN to a POSITIVE-FIXNUM TN. The practical effect of this
566 ;;; deficiency is that we have two extra moves and increased register
567 ;;; pressure, which can lead to some spectacularly bad register
568 ;;; allocation. (sub-FIXME: the register allocation even with the
569 ;;; strangely written loops is not always excellent, either...). Doing
570 ;;; it the first way, above, means that I is always thought of as a
571 ;;; POSITIVE-FIXNUM and there are no issues.
573 ;;; Besides, the *-WITH-OFFSET machinery will fold those index
574 ;;; adjustments in the first version into the array addressing at no
575 ;;; performance penalty!
577 ;;; This transform is critical to the performance of string streams. If
578 ;;; you tweak it, make sure that you compare the disassembly, if not the
579 ;;; performance of, the functions implementing string streams
580 ;;; (e.g. SB!IMPL::STRING-OUCH).
582 ((define-replace-transforms ()
583 (loop for saetp across sb!vm:*specialized-array-element-type-properties*
584 for sequence-type = `(simple-array ,(sb!vm:saetp-specifier saetp) (*))
585 unless (= (sb!vm:saetp-typecode saetp) sb!vm::simple-array-nil-widetag)
587 `(deftransform replace ((seq1 seq2 &key (start1 0) (start2 0) end1 end2)
588 (,sequence-type ,sequence-type &rest t)
592 ((valid-bit-bash-saetp-p saetp) nil)
593 ;; If we're not bit-bashing, only allow cases where we
594 ;; can determine the order of copying up front. (There
595 ;; are actually more cases we can handle if we know the
596 ;; amount that we're copying, but this handles the
598 (t '(unless (= (constant-value-or-lose start1 0)
599 (constant-value-or-lose start2 0))
600 (give-up-ir1-transform))))
601 `(let* ((len1 (length seq1))
603 (end1 (or end1 len1))
604 (end2 (or end2 len2))
605 (replace-len1 (- end1 start1))
606 (replace-len2 (- end2 start2)))
607 ,(unless (policy node (= safety 0))
609 (unless (<= 0 start1 end1 len1)
610 (sb!impl::signal-bounding-indices-bad-error seq1 start1 end1))
611 (unless (<= 0 start2 end2 len2)
612 (sb!impl::signal-bounding-indices-bad-error seq2 start2 end2))))
614 ((valid-bit-bash-saetp-p saetp)
615 (let* ((n-element-bits (sb!vm:saetp-n-bits saetp))
616 (bash-function (intern (format nil "UB~D-BASH-COPY" n-element-bits)
617 (find-package "SB!KERNEL"))))
618 `(funcall (function ,bash-function) seq2 start2
619 seq1 start1 (min replace-len1 replace-len2))))
621 ;; We can expand the loop inline here because we
622 ;; would have given up the transform (see above)
623 ;; if we didn't have constant matching start
625 '(do ((i start1 (1+ i))
627 (min replace-len1 replace-len2))))
629 (declare (optimize (insert-array-bounds-checks 0)))
630 (setf (aref seq1 i) (aref seq2 i)))))
633 finally (return `(progn ,@forms)))))
634 (define-replace-transforms))
636 ;;; Expand simple cases of UB<SIZE>-BASH-COPY inline. "simple" is
637 ;;; defined as those cases where we are doing word-aligned copies from
638 ;;; both the source and the destination and we are copying from the same
639 ;;; offset from both the source and the destination. (The last
640 ;;; condition is there so we can determine the direction to copy at
641 ;;; compile time rather than runtime. Remember that UB<SIZE>-BASH-COPY
642 ;;; acts like memmove, not memcpy.) These conditions may seem rather
643 ;;; restrictive, but they do catch common cases, like allocating a (* 2
644 ;;; N)-size buffer and blitting in the old N-size buffer in.
646 (defun frob-bash-transform (src src-offset
648 length n-elems-per-word)
649 (declare (ignore src dst length))
650 (let ((n-bits-per-elem (truncate sb!vm:n-word-bits n-elems-per-word)))
651 (multiple-value-bind (src-word src-elt)
652 (truncate (lvar-value src-offset) n-elems-per-word)
653 (multiple-value-bind (dst-word dst-elt)
654 (truncate (lvar-value dst-offset) n-elems-per-word)
655 ;; Avoid non-word aligned copies.
656 (unless (and (zerop src-elt) (zerop dst-elt))
657 (give-up-ir1-transform))
658 ;; Avoid copies where we would have to insert code for
659 ;; determining the direction of copying.
660 (unless (= src-word dst-word)
661 (give-up-ir1-transform))
662 ;; FIXME: The cross-compiler doesn't optimize TRUNCATE properly,
663 ;; so we have to do its work here.
664 `(let ((end (+ ,src-word ,(if (= n-elems-per-word 1)
666 `(truncate (the index length) ,n-elems-per-word)))))
667 (declare (type index end))
668 ;; Handle any bits at the end.
669 (when (logtest length (1- ,n-elems-per-word))
670 (let* ((extra (mod length ,n-elems-per-word))
671 ;; FIXME: The shift amount on this ASH is
672 ;; *always* negative, but the backend doesn't
673 ;; have a NEGATIVE-FIXNUM primitive type, so we
674 ;; wind up with a pile of code that tests the
675 ;; sign of the shift count prior to shifting when
676 ;; all we need is a simple negate and shift
678 (mask (ash #.(1- (ash 1 sb!vm:n-word-bits))
679 (* (- extra ,n-elems-per-word)
681 (setf (sb!kernel:%vector-raw-bits dst end)
683 (logandc2 (sb!kernel:%vector-raw-bits dst end)
685 ,(ecase sb!c:*backend-byte-order*
687 (:big-endian `(* (- ,n-elems-per-word extra)
688 ,n-bits-per-elem)))))
689 (logand (sb!kernel:%vector-raw-bits src end)
691 ,(ecase sb!c:*backend-byte-order*
693 (:big-endian `(* (- ,n-elems-per-word extra)
694 ,n-bits-per-elem)))))))))
695 ;; Copy from the end to save a register.
698 (setf (sb!kernel:%vector-raw-bits dst (1- i))
699 (sb!kernel:%vector-raw-bits src (1- i)))))))))
701 #.(loop for i = 1 then (* i 2)
702 collect `(deftransform ,(intern (format nil "UB~D-BASH-COPY" i)
707 ((simple-unboxed-array (*))
709 (simple-unboxed-array (*))
713 (frob-bash-transform src src-offset
714 dst dst-offset length
715 ,(truncate sb!vm:n-word-bits i))) into forms
716 until (= i sb!vm:n-word-bits)
717 finally (return `(progn ,@forms)))
719 ;;; We expand copy loops inline in SUBSEQ and COPY-SEQ if we're copying
720 ;;; arrays with elements of size >= the word size. We do this because
721 ;;; we know the arrays cannot alias (one was just consed), therefore we
722 ;;; can determine at compile time the direction to copy, and for
723 ;;; word-sized elements, UB<WORD-SIZE>-BASH-COPY will do a bit of
724 ;;; needless checking to figure out what's going on. The same
725 ;;; considerations apply if we are copying elements larger than the word
726 ;;; size, with the additional twist that doing it inline is likely to
727 ;;; cons far less than calling REPLACE and letting generic code do the
730 ;;; However, we do not do this for elements whose size is < than the
731 ;;; word size because we don't want to deal with any alignment issues
732 ;;; inline. The UB*-BASH-COPY transforms might fix things up later
735 (defun maybe-expand-copy-loop-inline (src src-offset dst dst-offset length
737 (let ((saetp (find-saetp element-type)))
739 (if (>= (sb!vm:saetp-n-bits saetp) sb!vm:n-word-bits)
740 (expand-aref-copy-loop src src-offset dst dst-offset length)
741 `(locally (declare (optimize (safety 0)))
742 (replace ,dst ,src :start1 ,dst-offset :start2 ,src-offset :end1 ,length)))))
744 (defun expand-aref-copy-loop (src src-offset dst dst-offset length)
745 (if (eql src-offset dst-offset)
746 `(do ((i (+ ,src-offset ,length) (1- i)))
748 (declare (optimize (insert-array-bounds-checks 0)))
749 (setf (aref ,dst (1- i)) (aref ,src (1- i))))
750 ;; KLUDGE: The compiler is not able to derive that (+ offset
751 ;; length) must be a fixnum, but arrives at (unsigned-byte 29).
752 ;; We, however, know it must be so, as by this point the bounds
753 ;; have already been checked.
754 `(do ((i (truly-the fixnum (+ ,src-offset ,length)) (1- i))
755 (j (+ ,dst-offset ,length) (1- j)))
757 (declare (optimize (insert-array-bounds-checks 0))
758 (type (integer 0 #.sb!xc:array-dimension-limit) j i))
759 (setf (aref ,dst (1- j)) (aref ,src (1- i))))))
761 (deftransform subseq ((seq start &optional end)
762 ((or (simple-unboxed-array (*)) simple-vector) t &optional t)
764 (let ((array-type (lvar-type seq)))
765 (unless (array-type-p array-type)
766 (give-up-ir1-transform))
767 (let ((element-type (type-specifier (array-type-specialized-element-type array-type))))
768 `(let* ((length (length seq))
769 (end (or end length)))
770 ,(unless (policy node (= safety 0))
772 (unless (<= 0 start end length)
773 (sb!impl::signal-bounding-indices-bad-error seq start end))))
774 (let* ((size (- end start))
775 (result (make-array size :element-type ',element-type)))
776 ,(maybe-expand-copy-loop-inline 'seq (if (constant-lvar-p start)
779 'result 0 'size element-type)
782 (deftransform copy-seq ((seq) ((or (simple-unboxed-array (*)) simple-vector)) *)
783 (let ((array-type (lvar-type seq)))
784 (unless (array-type-p array-type)
785 (give-up-ir1-transform))
786 (let ((element-type (type-specifier (array-type-specialized-element-type array-type))))
787 `(let* ((length (length seq))
788 (result (make-array length :element-type ',element-type)))
789 ,(maybe-expand-copy-loop-inline 'seq 0 'result 0 'length element-type)
792 ;;; FIXME: it really should be possible to take advantage of the
793 ;;; macros used in code/seq.lisp here to avoid duplication of code,
794 ;;; and enable even funkier transformations.
795 (deftransform search ((pattern text &key (start1 0) (start2 0) end1 end2
799 (vector vector &rest t)
801 :policy (> speed (max space safety)))
803 (let ((from-end (when (lvar-p from-end)
804 (unless (constant-lvar-p from-end)
805 (give-up-ir1-transform ":FROM-END is not constant."))
806 (lvar-value from-end)))
808 (testp (lvar-p test)))
810 (let ((end1 (or end1 (length pattern)))
811 (end2 (or end2 (length text)))
813 '((key (coerce key 'function))))
815 '((test (coerce test 'function)))))
816 (declare (type index start1 start2 end1 end2))
818 '(index2 (- end2 (- end1 start1)) (1- index2))
819 '(index2 start2 (1+ index2))))
824 ;; INDEX2 is FIXNUM, not an INDEX, as right before the loop
825 ;; terminates is hits -1 when :FROM-END is true and :START2
827 (declare (type fixnum index2))
828 (when (do ((index1 start1 (1+ index1))
829 (index2 index2 (1+ index2)))
831 (declare (type index index1 index2))
833 '((when (= index2 end2)
834 (return-from search nil))))
839 '(funcall key (aref pattern index1))
840 '(aref pattern index1))
842 '(funcall key (aref text index2))
843 '(aref text index2)))
845 (return index2)))))))
847 ;;; FIXME: It seems as though it should be possible to make a DEFUN
848 ;;; %CONCATENATE (with a DEFTRANSFORM to translate constant RTYPE to
849 ;;; CTYPE before calling %CONCATENATE) which is comparably efficient,
850 ;;; at least once DYNAMIC-EXTENT works.
852 ;;; FIXME: currently KLUDGEed because of bug 188
854 ;;; FIXME: disabled for sb-unicode: probably want it back
856 (deftransform concatenate ((rtype &rest sequences)
857 (t &rest (or simple-base-string
858 (simple-array nil (*))))
860 :policy (< safety 3))
861 (loop for rest-seqs on sequences
862 for n-seq = (gensym "N-SEQ")
863 for n-length = (gensym "N-LENGTH")
864 for start = 0 then next-start
865 for next-start = (gensym "NEXT-START")
866 collect n-seq into args
867 collect `(,n-length (length ,n-seq)) into lets
868 collect n-length into all-lengths
869 collect next-start into starts
870 collect `(if (and (typep ,n-seq '(simple-array nil (*)))
872 (error 'nil-array-accessed-error)
873 (#.(let* ((i (position 'character sb!kernel::*specialized-array-element-types*))
874 (saetp (aref sb!vm:*specialized-array-element-type-properties* i))
875 (n-bits (sb!vm:saetp-n-bits saetp)))
876 (intern (format nil "UB~D-BASH-COPY" n-bits)
878 ,n-seq 0 res ,start ,n-length))
880 collect `(setq ,next-start (+ ,start ,n-length)) into forms
883 `(lambda (rtype ,@args)
884 (declare (ignore rtype))
886 (res (make-string (the index (+ ,@all-lengths))
887 :element-type 'base-char)))
888 (declare (type index ,@all-lengths))
889 (let (,@(mapcar (lambda (name) `(,name 0)) starts))
890 (declare (type index ,@starts))
894 ;;;; CONS accessor DERIVE-TYPE optimizers
896 (defoptimizer (car derive-type) ((cons))
897 (let ((type (lvar-type cons))
898 (null-type (specifier-type 'null)))
899 (cond ((eq type null-type)
902 (cons-type-car-type type)))))
904 (defoptimizer (cdr derive-type) ((cons))
905 (let ((type (lvar-type cons))
906 (null-type (specifier-type 'null)))
907 (cond ((eq type null-type)
910 (cons-type-cdr-type type)))))
912 ;;;; FIND, POSITION, and their -IF and -IF-NOT variants
914 ;;; We want to make sure that %FIND-POSITION is inline-expanded into
915 ;;; %FIND-POSITION-IF only when %FIND-POSITION-IF has an inline
916 ;;; expansion, so we factor out the condition into this function.
917 (defun check-inlineability-of-find-position-if (sequence from-end)
918 (let ((ctype (lvar-type sequence)))
919 (cond ((csubtypep ctype (specifier-type 'vector))
920 ;; It's not worth trying to inline vector code unless we
921 ;; know a fair amount about it at compile time.
922 (upgraded-element-type-specifier-or-give-up sequence)
923 (unless (constant-lvar-p from-end)
924 (give-up-ir1-transform
925 "FROM-END argument value not known at compile time")))
926 ((csubtypep ctype (specifier-type 'list))
927 ;; Inlining on lists is generally worthwhile.
930 (give-up-ir1-transform
931 "sequence type not known at compile time")))))
933 ;;; %FIND-POSITION-IF and %FIND-POSITION-IF-NOT for LIST data
934 (macrolet ((def (name condition)
935 `(deftransform ,name ((predicate sequence from-end start end key)
936 (function list t t t function)
938 :policy (> speed space))
943 (declare (type index index))
945 (if (and end (> end index))
946 (sb!impl::signal-bounding-indices-bad-error
948 (values find position)))
949 (let ((key-i (funcall key i)))
950 (when (and end (>= index end))
951 (return (values find position)))
952 (when (>= index start)
953 (,',condition (funcall predicate key-i)
954 ;; This hack of dealing with non-NIL
955 ;; FROM-END for list data by iterating
956 ;; forward through the list and keeping
957 ;; track of the last time we found a match
958 ;; might be more screwy than what the user
959 ;; expects, but it seems to be allowed by
960 ;; the ANSI standard. (And if the user is
961 ;; screwy enough to ask for FROM-END
962 ;; behavior on list data, turnabout is
965 ;; It's also not enormously efficient,
966 ;; calling PREDICATE and KEY more often
967 ;; than necessary; but all the
968 ;; alternatives seem to have their own
969 ;; efficiency problems.
973 (return (values i index))))))
975 (def %find-position-if when)
976 (def %find-position-if-not unless))
978 ;;; %FIND-POSITION for LIST data can be expanded into %FIND-POSITION-IF
979 ;;; without loss of efficiency. (I.e., the optimizer should be able
980 ;;; to straighten everything out.)
981 (deftransform %find-position ((item sequence from-end start end key test)
984 :policy (> speed space))
986 '(%find-position-if (let ((test-fun (%coerce-callable-to-fun test)))
987 ;; The order of arguments for asymmetric tests
988 ;; (e.g. #'<, as opposed to order-independent
989 ;; tests like #'=) is specified in the spec
990 ;; section 17.2.1 -- the O/Zi stuff there.
992 (funcall test-fun item i)))
997 (%coerce-callable-to-fun key)))
999 ;;; The inline expansions for the VECTOR case are saved as macros so
1000 ;;; that we can share them between the DEFTRANSFORMs and the default
1001 ;;; cases in the DEFUNs. (This isn't needed for the LIST case, because
1002 ;;; the DEFTRANSFORMs for LIST are less choosy about when to expand.)
1003 (defun %find-position-or-find-position-if-vector-expansion (sequence-arg
1009 (with-unique-names (offset block index n-sequence sequence n-end end)
1010 `(let ((,n-sequence ,sequence-arg)
1012 (with-array-data ((,sequence ,n-sequence :offset-var ,offset)
1014 (,end (%check-vector-sequence-bounds
1015 ,n-sequence ,start ,n-end)))
1017 (macrolet ((maybe-return ()
1018 ;; WITH-ARRAY-DATA has already performed bounds
1019 ;; checking, so we can safely elide the checks
1020 ;; in the inner loop.
1021 '(let ((,element (locally (declare (optimize (insert-array-bounds-checks 0)))
1022 (aref ,sequence ,index))))
1026 (- ,index ,offset)))))))
1029 ;; (If we aren't fastidious about declaring that
1030 ;; INDEX might be -1, then (FIND 1 #() :FROM-END T)
1031 ;; can send us off into never-never land, since
1032 ;; INDEX is initialized to -1.)
1033 of-type index-or-minus-1
1034 from (1- ,end) downto ,start do
1036 (loop for ,index of-type index from ,start below ,end do
1038 (values nil nil))))))
1040 (def!macro %find-position-vector-macro (item sequence
1041 from-end start end key test)
1042 (with-unique-names (element)
1043 (%find-position-or-find-position-if-vector-expansion
1049 ;; (See the LIST transform for a discussion of the correct
1050 ;; argument order, i.e. whether the searched-for ,ITEM goes before
1051 ;; or after the checked sequence element.)
1052 `(funcall ,test ,item (funcall ,key ,element)))))
1054 (def!macro %find-position-if-vector-macro (predicate sequence
1055 from-end start end key)
1056 (with-unique-names (element)
1057 (%find-position-or-find-position-if-vector-expansion
1063 `(funcall ,predicate (funcall ,key ,element)))))
1065 (def!macro %find-position-if-not-vector-macro (predicate sequence
1066 from-end start end key)
1067 (with-unique-names (element)
1068 (%find-position-or-find-position-if-vector-expansion
1074 `(not (funcall ,predicate (funcall ,key ,element))))))
1076 ;;; %FIND-POSITION, %FIND-POSITION-IF and %FIND-POSITION-IF-NOT for
1078 (deftransform %find-position-if ((predicate sequence from-end start end key)
1079 (function vector t t t function)
1081 :policy (> speed space))
1083 (check-inlineability-of-find-position-if sequence from-end)
1084 '(%find-position-if-vector-macro predicate sequence
1085 from-end start end key))
1087 (deftransform %find-position-if-not ((predicate sequence from-end start end key)
1088 (function vector t t t function)
1090 :policy (> speed space))
1092 (check-inlineability-of-find-position-if sequence from-end)
1093 '(%find-position-if-not-vector-macro predicate sequence
1094 from-end start end key))
1096 (deftransform %find-position ((item sequence from-end start end key test)
1097 (t vector t t t function function)
1099 :policy (> speed space))
1101 (check-inlineability-of-find-position-if sequence from-end)
1102 '(%find-position-vector-macro item sequence
1103 from-end start end key test))
1105 ;;; logic to unravel :TEST, :TEST-NOT, and :KEY options in FIND,
1106 ;;; POSITION-IF, etc.
1107 (define-source-transform effective-find-position-test (test test-not)
1108 (once-only ((test test)
1109 (test-not test-not))
1111 ((and ,test ,test-not)
1112 (error "can't specify both :TEST and :TEST-NOT"))
1113 (,test (%coerce-callable-to-fun ,test))
1115 ;; (Without DYNAMIC-EXTENT, this is potentially horribly
1116 ;; inefficient, but since the TEST-NOT option is deprecated
1117 ;; anyway, we don't care.)
1118 (complement (%coerce-callable-to-fun ,test-not)))
1120 (define-source-transform effective-find-position-key (key)
1121 (once-only ((key key))
1123 (%coerce-callable-to-fun ,key)
1126 (macrolet ((define-find-position (fun-name values-index)
1127 `(deftransform ,fun-name ((item sequence &key
1128 from-end (start 0) end
1130 (t (or list vector) &rest t))
1131 '(nth-value ,values-index
1132 (%find-position item sequence
1135 (effective-find-position-key key)
1136 (effective-find-position-test
1138 (define-find-position find 0)
1139 (define-find-position position 1))
1141 (macrolet ((define-find-position-if (fun-name values-index)
1142 `(deftransform ,fun-name ((predicate sequence &key
1145 (t (or list vector) &rest t))
1148 (%find-position-if (%coerce-callable-to-fun predicate)
1151 (effective-find-position-key key))))))
1152 (define-find-position-if find-if 0)
1153 (define-find-position-if position-if 1))
1155 ;;; the deprecated functions FIND-IF-NOT and POSITION-IF-NOT. We
1156 ;;; didn't bother to worry about optimizing them, except note that on
1157 ;;; Sat, Oct 06, 2001 at 04:22:38PM +0100, Christophe Rhodes wrote on
1160 ;;; My understanding is that while the :test-not argument is
1161 ;;; deprecated in favour of :test (complement #'foo) because of
1162 ;;; semantic difficulties (what happens if both :test and :test-not
1163 ;;; are supplied, etc) the -if-not variants, while officially
1164 ;;; deprecated, would be undeprecated were X3J13 actually to produce
1165 ;;; a revised standard, as there are perfectly legitimate idiomatic
1166 ;;; reasons for allowing the -if-not versions equal status,
1167 ;;; particularly remove-if-not (== filter).
1169 ;;; This is only an informal understanding, I grant you, but
1170 ;;; perhaps it's worth optimizing the -if-not versions in the same
1171 ;;; way as the others?
1173 ;;; FIXME: Maybe remove uses of these deprecated functions within the
1174 ;;; implementation of SBCL.
1175 (macrolet ((define-find-position-if-not (fun-name values-index)
1176 `(deftransform ,fun-name ((predicate sequence &key
1179 (t (or list vector) &rest t))
1182 (%find-position-if-not (%coerce-callable-to-fun predicate)
1185 (effective-find-position-key key))))))
1186 (define-find-position-if-not find-if-not 0)
1187 (define-find-position-if-not position-if-not 1))