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))))
33 `(let ((,fn-sym (%coerce-callable-to-fun ,fn)))
37 (map-result (gensym)))
38 `(let ((,map-result (list nil)))
39 (do-anonymous ((,temp ,map-result) . ,(do-clauses))
40 (,endtest (cdr ,map-result))
41 (setq ,temp (last (nconc ,temp ,call)))))))
44 (map-result (gensym)))
45 `(let ((,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)))))))
50 `(let ((,n-first ,(first arglists)))
51 (do-anonymous ,(do-clauses)
52 (,endtest (truly-the list ,n-first))
55 (define-source-transform mapc (function list &rest more-lists)
56 (mapfoo-transform function (cons list more-lists) nil t))
58 (define-source-transform mapcar (function list &rest more-lists)
59 (mapfoo-transform function (cons list more-lists) :list t))
61 (define-source-transform mapcan (function list &rest more-lists)
62 (mapfoo-transform function (cons list more-lists) :nconc t))
64 (define-source-transform mapl (function list &rest more-lists)
65 (mapfoo-transform function (cons list more-lists) nil nil))
67 (define-source-transform maplist (function list &rest more-lists)
68 (mapfoo-transform function (cons list more-lists) :list nil))
70 (define-source-transform mapcon (function list &rest more-lists)
71 (mapfoo-transform function (cons list more-lists) :nconc nil))
73 ;;;; mapping onto sequences: the MAP function
75 ;;; MAP is %MAP plus a check to make sure that any length specified in
76 ;;; the result type matches the actual result. We also wrap it in a
77 ;;; TRULY-THE for the most specific type we can determine.
78 (deftransform map ((result-type-arg fun seq &rest seqs) * * :node node)
79 (let* ((seq-names (make-gensym-list (1+ (length seqs))))
80 (bare `(%map result-type-arg fun ,@seq-names))
81 (constant-result-type-arg-p (constant-lvar-p result-type-arg))
82 ;; what we know about the type of the result. (Note that the
83 ;; "result type" argument is not necessarily the type of the
84 ;; result, since NIL means the result has NULL type.)
85 (result-type (if (not constant-result-type-arg-p)
87 (let ((result-type-arg-value
88 (lvar-value result-type-arg)))
89 (if (null result-type-arg-value)
91 result-type-arg-value)))))
92 `(lambda (result-type-arg fun ,@seq-names)
93 (truly-the ,result-type
94 ,(cond ((policy node (< safety 3))
95 ;; ANSI requires the length-related type check only
96 ;; when the SAFETY quality is 3... in other cases, we
97 ;; skip it, because it could be expensive.
99 ((not constant-result-type-arg-p)
100 `(sequence-of-checked-length-given-type ,bare
103 (let ((result-ctype (ir1-transform-specifier-type
105 (if (array-type-p result-ctype)
106 (let ((dims (array-type-dimensions result-ctype)))
107 (unless (and (listp dims) (= (length dims) 1))
108 (give-up-ir1-transform "invalid sequence type"))
109 (let ((dim (first dims)))
112 `(vector-of-checked-length-given-length ,bare
114 ;; FIXME: this is wrong, as not all subtypes of
115 ;; VECTOR are ARRAY-TYPEs [consider, for
116 ;; example, (OR (VECTOR T 3) (VECTOR T
117 ;; 4))]. However, it's difficult to see what we
118 ;; should put here... maybe we should
119 ;; GIVE-UP-IR1-TRANSFORM if the type is a
120 ;; subtype of VECTOR but not an ARRAY-TYPE?
123 ;;; Return a DO loop, mapping a function FUN to elements of
124 ;;; sequences. SEQS is a list of lvars, SEQ-NAMES - list of variables,
125 ;;; bound to sequences, INTO - a variable, which is used in
126 ;;; MAP-INTO. RESULT and BODY are forms, which can use variables
127 ;;; FUNCALL-RESULT, containing the result of application of FUN, and
128 ;;; INDEX, containing the current position in sequences.
129 (defun build-sequence-iterator (seqs seq-names &key result into body)
130 (declare (type list seqs seq-names)
137 (let ((found-vector-p nil))
138 (flet ((process-vector (length)
139 (unless found-vector-p
140 (setq found-vector-p t)
141 (bindings `(index 0 (1+ index)))
142 (declarations `(type index index)))
143 (vector-lengths length)))
144 (loop for seq of-type lvar in seqs
145 for seq-name in seq-names
146 for type = (lvar-type seq)
147 do (cond ((csubtypep type (specifier-type 'list))
148 (with-unique-names (index)
149 (bindings `(,index ,seq-name (cdr ,index)))
150 (declarations `(type list ,index))
151 (places `(car ,index))
152 (tests `(endp ,index))))
153 ((csubtypep type (specifier-type 'vector))
154 (process-vector `(length ,seq-name))
155 (places `(locally (declare (optimize (insert-array-bounds-checks 0)))
156 (aref ,seq-name index))))
158 (give-up-ir1-transform
159 "can't determine sequence argument type"))))
161 (process-vector `(array-dimension ,into 0))))
163 (bindings `(length (min ,@(vector-lengths))))
164 (tests `(>= index length)))
166 ((or ,@(tests)) ,result)
167 (declare ,@(declarations))
168 (let ((funcall-result (funcall fun ,@(places))))
169 (declare (ignorable funcall-result))
172 ;;; Try to compile %MAP efficiently when we can determine sequence
173 ;;; argument types at compile time.
175 ;;; Note: This transform was written to allow open coding of
176 ;;; quantifiers by expressing them in terms of (MAP NIL ..). For
177 ;;; non-NIL values of RESULT-TYPE, it's still useful, but not
178 ;;; necessarily as efficient as possible. In particular, it will be
179 ;;; inefficient when RESULT-TYPE is a SIMPLE-ARRAY with specialized
180 ;;; numeric element types. It should be straightforward to make it
181 ;;; handle that case more efficiently, but it's left as an exercise to
182 ;;; the reader, because the code is complicated enough already and I
183 ;;; don't happen to need that functionality right now. -- WHN 20000410
184 (deftransform %map ((result-type fun seq &rest seqs) * *
185 :policy (>= speed space))
187 (unless (constant-lvar-p result-type)
188 (give-up-ir1-transform "RESULT-TYPE argument not constant"))
189 (labels ( ;; 1-valued SUBTYPEP, fails unless second value of SUBTYPEP is true
190 (fn-1subtypep (fn x y)
191 (multiple-value-bind (subtype-p valid-p) (funcall fn x y)
194 (give-up-ir1-transform
195 "can't analyze sequence type relationship"))))
196 (1subtypep (x y) (fn-1subtypep #'sb!xc:subtypep x y)))
197 (let* ((result-type-value (lvar-value result-type))
198 (result-supertype (cond ((null result-type-value) 'null)
199 ((1subtypep result-type-value 'vector)
201 ((1subtypep result-type-value 'list)
204 (give-up-ir1-transform
205 "result type unsuitable")))))
206 (cond ((and result-type-value (null seqs))
207 ;; The consing arity-1 cases can be implemented
208 ;; reasonably efficiently as function calls, and the cost
209 ;; of consing should be significantly larger than
210 ;; function call overhead, so we always compile these
211 ;; cases as full calls regardless of speed-versus-space
212 ;; optimization policy.
213 (cond ((subtypep result-type-value 'list)
214 '(%map-to-list-arity-1 fun seq))
215 ( ;; (This one can be inefficient due to COERCE, but
216 ;; the current open-coded implementation has the
218 (subtypep result-type-value 'vector)
219 `(coerce (%map-to-simple-vector-arity-1 fun seq)
220 ',result-type-value))
221 (t (bug "impossible (?) sequence type"))))
223 (let* ((seqs (cons seq seqs))
224 (seq-args (make-gensym-list (length seqs))))
225 (multiple-value-bind (push-dacc result)
226 (ecase result-supertype
227 (null (values nil nil))
228 (list (values `(push funcall-result acc)
230 (vector (values `(push funcall-result acc)
231 `(coerce (nreverse acc)
232 ',result-type-value))))
233 ;; (We use the same idiom, of returning a LAMBDA from
234 ;; DEFTRANSFORM, as is used in the DEFTRANSFORMs for
235 ;; FUNCALL and ALIEN-FUNCALL, and for the same
236 ;; reason: we need to get the runtime values of each
237 ;; of the &REST vars.)
238 `(lambda (result-type fun ,@seq-args)
239 (declare (ignore result-type))
240 (let ((fun (%coerce-callable-to-fun fun))
242 (declare (type list acc))
243 (declare (ignorable acc))
244 ,(build-sequence-iterator
247 :body push-dacc))))))))))
250 (deftransform map-into ((result fun &rest seqs)
254 (let ((seqs-names (mapcar (lambda (x)
258 `(lambda (result fun ,@seqs-names)
259 ,(build-sequence-iterator
261 :result '(when (array-has-fill-pointer-p result)
262 (setf (fill-pointer result) index))
264 :body '(locally (declare (optimize (insert-array-bounds-checks 0)))
265 (setf (aref result index) funcall-result)))
269 ;;; FIXME: once the confusion over doing transforms with known-complex
270 ;;; arrays is over, we should also transform the calls to (AND (ARRAY
271 ;;; * (*)) (NOT (SIMPLE-ARRAY * (*)))) objects.
272 (deftransform elt ((s i) ((simple-array * (*)) *) *)
275 (deftransform elt ((s i) (list *) * :policy (< safety 3))
278 (deftransform %setelt ((s i v) ((simple-array * (*)) * *) *)
281 (deftransform %setelt ((s i v) (list * *) * :policy (< safety 3))
282 '(setf (car (nthcdr i s)) v))
284 (deftransform %check-vector-sequence-bounds ((vector start end)
287 (if (policy node (< safety speed))
288 '(or end (length vector))
289 '(let ((length (length vector)))
290 (if (<= 0 start (or end length) length)
292 (sb!impl::signal-bounding-indices-bad-error vector start end)))))
294 (defun transform-list-item-seek (name list key test test-not node)
295 ;; Key can legally be NIL, but if it's NIL for sure we pretend it's
296 ;; not there at all. If it might be NIL, make up a form to that
297 ;; ensure it is a function.
298 (multiple-value-bind (key key-form)
300 (let ((key-type (lvar-type key))
301 (null-type (specifier-type 'null)))
302 (cond ((csubtypep key-type null-type)
304 ((csubtypep null-type key-type)
306 (%coerce-callable-to-fun key)
309 (values key '(%coerce-callable-to-fun key))))))
310 (let* ((funs (remove nil (list (and key 'key) (cond (test 'test)
311 (test-not 'test-not)))))
312 (out-of-line (or (find-symbol (format nil "%~A~{-~A~}" name funs)
313 (load-time-value (find-package "SB!KERNEL")))
314 (bug "Unknown list item seek transform: name=~S, funs=~S"
316 (target-expr (if key '(%funcall key target) 'target))
317 (test-expr (cond (test `(%funcall test item ,target-expr))
318 (test-not `(not (%funcall test-not item ,target-expr)))
319 (t `(eql item ,target-expr)))))
320 (labels ((open-code (tail)
322 `(if (let ((this ',(car tail)))
325 `(and this (let ((target (car this)))
328 `(let ((target this))
333 ,(open-code (cdr tail)))))
337 `(%coerce-callable-to-fun ,fun))))
338 (let* ((cp (constant-lvar-p list))
339 (c-list (when cp (lvar-value list))))
340 (cond ((and cp c-list (policy node (>= speed space)))
341 `(let ,(mapcar (lambda (fun) `(,fun ,(ensure-fun fun))) funs)
342 ,(open-code c-list)))
343 ((and cp (not c-list))
344 ;; constant nil list -- nothing to find!
347 `(,out-of-line item list ,@(mapcar #'ensure-fun funs)))))))))
349 (deftransform member ((item list &key key test test-not) * * :node node)
350 (transform-list-item-seek 'member list key test test-not node))
352 (deftransform assoc ((item list &key key test test-not) * * :node node)
353 (transform-list-item-seek 'assoc list key test test-not node))
355 (deftransform memq ((item list) (t (constant-arg list)))
358 `(if (eq item ',(car tail))
362 (rec (lvar-value list))))
364 ;;; A similar transform used to apply to MEMBER and ASSOC, but since
365 ;;; TRANSFORM-LIST-ITEM-SEEK now takes care of them those transform
366 ;;; would never fire, and (%MEMBER-TEST ITEM LIST #'EQ) should be
367 ;;; almost as fast as MEMQ.
368 (deftransform delete ((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.")))
387 (deftransform delete-if ((pred list) (t list))
389 '(do ((x list (cdr x))
392 (cond ((funcall pred (car x))
395 (rplacd splice (cdr x))))
396 (t (setq splice x)))))
398 (deftransform fill ((seq item &key (start 0) (end (length seq)))
399 (vector t &key (:start t) (:end index))
401 :policy (> speed space))
403 (let ((element-type (upgraded-element-type-specifier-or-give-up seq)))
405 `(with-array-data ((data seq)
408 (declare (type (simple-array ,element-type 1) data))
409 (declare (type fixnum start end))
410 (do ((i start (1+ i)))
412 (declare (type index i))
413 ;; WITH-ARRAY-DATA did our range checks once and for all, so
414 ;; it'd be wasteful to check again on every AREF...
415 (declare (optimize (safety 0)))
416 (setf (aref data i) item)))
417 ;; ... though we still need to check that the new element can fit
418 ;; into the vector in safe code. -- CSR, 2002-07-05
419 `((declare (type ,element-type item))))))
423 ;;; Return true if LVAR's only use is a non-NOTINLINE reference to a
424 ;;; global function with one of the specified NAMES.
425 (defun lvar-fun-is (lvar names)
426 (declare (type lvar lvar) (list names))
427 (let ((use (lvar-uses lvar)))
429 (let ((leaf (ref-leaf use)))
430 (and (global-var-p leaf)
431 (eq (global-var-kind leaf) :global-function)
432 (not (null (member (leaf-source-name leaf) names
433 :test #'equal))))))))
435 ;;; If LVAR is a constant lvar, the return the constant value. If it
436 ;;; is null, then return default, otherwise quietly give up the IR1
439 ;;; ### Probably should take an ARG and flame using the NAME.
440 (defun constant-value-or-lose (lvar &optional default)
441 (declare (type (or lvar null) lvar))
442 (cond ((not lvar) default)
443 ((constant-lvar-p lvar)
446 (give-up-ir1-transform))))
449 ;;;; hairy sequence transforms
451 ;;; FIXME: no hairy sequence transforms in SBCL?
453 ;;; There used to be a bunch of commented out code about here,
454 ;;; containing the (apparent) beginning of hairy sequence transform
455 ;;; infrastructure. People interested in implementing better sequence
456 ;;; transforms might want to look at it for inspiration, even though
457 ;;; the actual code is ancient CMUCL -- and hence bitrotted. The code
458 ;;; was deleted in 1.0.7.23.
460 ;;;; string operations
462 ;;; We transform the case-sensitive string predicates into a non-keyword
463 ;;; version. This is an IR1 transform so that we don't have to worry about
464 ;;; changing the order of evaluation.
465 (macrolet ((def (fun pred*)
466 `(deftransform ,fun ((string1 string2 &key (start1 0) end1
469 `(,',pred* string1 string2 start1 end1 start2 end2))))
470 (def string< string<*)
471 (def string> string>*)
472 (def string<= string<=*)
473 (def string>= string>=*)
474 (def string= string=*)
475 (def string/= string/=*))
477 ;;; Return a form that tests the free variables STRING1 and STRING2
478 ;;; for the ordering relationship specified by LESSP and EQUALP. The
479 ;;; start and end are also gotten from the environment. Both strings
480 ;;; must be SIMPLE-BASE-STRINGs.
481 (macrolet ((def (name lessp equalp)
482 `(deftransform ,name ((string1 string2 start1 end1 start2 end2)
483 (simple-base-string simple-base-string t t t t) *)
484 `(let* ((end1 (if (not end1) (length string1) end1))
485 (end2 (if (not end2) (length string2) end2))
486 (index (sb!impl::%sp-string-compare
487 string1 start1 end1 string2 start2 end2)))
489 (cond ((= index end1)
490 ,(if ',lessp 'index nil))
491 ((= (+ index (- start2 start1)) end2)
492 ,(if ',lessp nil 'index))
493 ((,(if ',lessp 'char< 'char>)
494 (schar string1 index)
503 ,(if ',equalp 'end1 nil))))))
506 (def string>* nil nil)
507 (def string>=* nil t))
509 (macrolet ((def (name result-fun)
510 `(deftransform ,name ((string1 string2 start1 end1 start2 end2)
511 (simple-base-string simple-base-string t t t t) *)
513 (sb!impl::%sp-string-compare
514 string1 start1 (or end1 (length string1))
515 string2 start2 (or end2 (length string2)))))))
517 (def string/=* identity))
520 ;;;; transforms for sequence functions
522 ;;; Moved here from generic/vm-tran.lisp to satisfy clisp. Only applies
523 ;;; to vectors based on simple arrays.
524 (def!constant vector-data-bit-offset
525 (* sb!vm:vector-data-offset sb!vm:n-word-bits))
527 (eval-when (:compile-toplevel)
528 (defun valid-bit-bash-saetp-p (saetp)
529 ;; BIT-BASHing isn't allowed on simple vectors that contain pointers
530 (and (not (eq t (sb!vm:saetp-specifier saetp)))
531 ;; Disallowing (VECTOR NIL) also means that we won't transform
532 ;; sequence functions into bit-bashing code and we let the
533 ;; generic sequence functions signal errors if necessary.
534 (not (zerop (sb!vm:saetp-n-bits saetp)))
535 ;; Due to limitations with the current BIT-BASHing code, we can't
536 ;; BIT-BASH reliably on arrays whose element types are larger
537 ;; than the word size.
538 (<= (sb!vm:saetp-n-bits saetp) sb!vm:n-word-bits)))
541 ;;; FIXME: In the copy loops below, we code the loops in a strange
544 ;;; (do ((i (+ src-offset length) (1- i)))
546 ;;; (... (aref foo (1- i)) ...))
548 ;;; rather than the more natural (and seemingly more efficient):
550 ;;; (do ((i (1- (+ src-offset length)) (1- i)))
552 ;;; (... (aref foo i) ...))
554 ;;; (more efficient because we don't have to do the index adjusting on
555 ;;; every iteration of the loop)
557 ;;; We do this to avoid a suboptimality in SBCL's backend. In the
558 ;;; latter case, the backend thinks I is a FIXNUM (which it is), but
559 ;;; when used as an array index, the backend thinks I is a
560 ;;; POSITIVE-FIXNUM (which it is). However, since the backend thinks of
561 ;;; these as distinct storage classes, it cannot coerce a move from a
562 ;;; FIXNUM TN to a POSITIVE-FIXNUM TN. The practical effect of this
563 ;;; deficiency is that we have two extra moves and increased register
564 ;;; pressure, which can lead to some spectacularly bad register
565 ;;; allocation. (sub-FIXME: the register allocation even with the
566 ;;; strangely written loops is not always excellent, either...). Doing
567 ;;; it the first way, above, means that I is always thought of as a
568 ;;; POSITIVE-FIXNUM and there are no issues.
570 ;;; Besides, the *-WITH-OFFSET machinery will fold those index
571 ;;; adjustments in the first version into the array addressing at no
572 ;;; performance penalty!
574 ;;; This transform is critical to the performance of string streams. If
575 ;;; you tweak it, make sure that you compare the disassembly, if not the
576 ;;; performance of, the functions implementing string streams
577 ;;; (e.g. SB!IMPL::STRING-OUCH).
579 ((define-replace-transforms ()
580 (loop for saetp across sb!vm:*specialized-array-element-type-properties*
581 for sequence-type = `(simple-array ,(sb!vm:saetp-specifier saetp) (*))
582 unless (= (sb!vm:saetp-typecode saetp) sb!vm::simple-array-nil-widetag)
584 `(deftransform replace ((seq1 seq2 &key (start1 0) (start2 0) end1 end2)
585 (,sequence-type ,sequence-type &rest t)
589 ((valid-bit-bash-saetp-p saetp) nil)
590 ;; If we're not bit-bashing, only allow cases where we
591 ;; can determine the order of copying up front. (There
592 ;; are actually more cases we can handle if we know the
593 ;; amount that we're copying, but this handles the
595 (t '(unless (= (constant-value-or-lose start1 0)
596 (constant-value-or-lose start2 0))
597 (give-up-ir1-transform))))
598 `(let* ((len1 (length seq1))
600 (end1 (or end1 len1))
601 (end2 (or end2 len2))
602 (replace-len1 (- end1 start1))
603 (replace-len2 (- end2 start2)))
604 ,(unless (policy node (= safety 0))
606 (unless (<= 0 start1 end1 len1)
607 (sb!impl::signal-bounding-indices-bad-error seq1 start1 end1))
608 (unless (<= 0 start2 end2 len2)
609 (sb!impl::signal-bounding-indices-bad-error seq2 start2 end2))))
611 ((valid-bit-bash-saetp-p saetp)
612 (let* ((n-element-bits (sb!vm:saetp-n-bits saetp))
613 (bash-function (intern (format nil "UB~D-BASH-COPY" n-element-bits)
614 (find-package "SB!KERNEL"))))
615 `(funcall (function ,bash-function) seq2 start2
616 seq1 start1 (min replace-len1 replace-len2))))
618 ;; We can expand the loop inline here because we
619 ;; would have given up the transform (see above)
620 ;; if we didn't have constant matching start
622 '(do ((i start1 (1+ i))
624 (min replace-len1 replace-len2))))
626 (declare (optimize (insert-array-bounds-checks 0)))
627 (setf (aref seq1 i) (aref seq2 i)))))
630 finally (return `(progn ,@forms)))))
631 (define-replace-transforms))
633 ;;; Expand simple cases of UB<SIZE>-BASH-COPY inline. "simple" is
634 ;;; defined as those cases where we are doing word-aligned copies from
635 ;;; both the source and the destination and we are copying from the same
636 ;;; offset from both the source and the destination. (The last
637 ;;; condition is there so we can determine the direction to copy at
638 ;;; compile time rather than runtime. Remember that UB<SIZE>-BASH-COPY
639 ;;; acts like memmove, not memcpy.) These conditions may seem rather
640 ;;; restrictive, but they do catch common cases, like allocating a (* 2
641 ;;; N)-size buffer and blitting in the old N-size buffer in.
643 (defun frob-bash-transform (src src-offset
645 length n-elems-per-word)
646 (declare (ignore src dst length))
647 (let ((n-bits-per-elem (truncate sb!vm:n-word-bits n-elems-per-word)))
648 (multiple-value-bind (src-word src-elt)
649 (truncate (lvar-value src-offset) n-elems-per-word)
650 (multiple-value-bind (dst-word dst-elt)
651 (truncate (lvar-value dst-offset) n-elems-per-word)
652 ;; Avoid non-word aligned copies.
653 (unless (and (zerop src-elt) (zerop dst-elt))
654 (give-up-ir1-transform))
655 ;; Avoid copies where we would have to insert code for
656 ;; determining the direction of copying.
657 (unless (= src-word dst-word)
658 (give-up-ir1-transform))
659 ;; FIXME: The cross-compiler doesn't optimize TRUNCATE properly,
660 ;; so we have to do its work here.
661 `(let ((end (+ ,src-word ,(if (= n-elems-per-word 1)
663 `(truncate (the index length) ,n-elems-per-word)))))
664 (declare (type index end))
665 ;; Handle any bits at the end.
666 (when (logtest length (1- ,n-elems-per-word))
667 (let* ((extra (mod length ,n-elems-per-word))
668 ;; FIXME: The shift amount on this ASH is
669 ;; *always* negative, but the backend doesn't
670 ;; have a NEGATIVE-FIXNUM primitive type, so we
671 ;; wind up with a pile of code that tests the
672 ;; sign of the shift count prior to shifting when
673 ;; all we need is a simple negate and shift
675 (mask (ash #.(1- (ash 1 sb!vm:n-word-bits))
676 (* (- extra ,n-elems-per-word)
678 (setf (sb!kernel:%vector-raw-bits dst end)
680 (logandc2 (sb!kernel:%vector-raw-bits dst end)
682 ,(ecase sb!c:*backend-byte-order*
684 (:big-endian `(* (- ,n-elems-per-word extra)
685 ,n-bits-per-elem)))))
686 (logand (sb!kernel:%vector-raw-bits src end)
688 ,(ecase sb!c:*backend-byte-order*
690 (:big-endian `(* (- ,n-elems-per-word extra)
691 ,n-bits-per-elem)))))))))
692 ;; Copy from the end to save a register.
695 (setf (sb!kernel:%vector-raw-bits dst (1- i))
696 (sb!kernel:%vector-raw-bits src (1- i)))))))))
698 #.(loop for i = 1 then (* i 2)
699 collect `(deftransform ,(intern (format nil "UB~D-BASH-COPY" i)
704 ((simple-unboxed-array (*))
706 (simple-unboxed-array (*))
710 (frob-bash-transform src src-offset
711 dst dst-offset length
712 ,(truncate sb!vm:n-word-bits i))) into forms
713 until (= i sb!vm:n-word-bits)
714 finally (return `(progn ,@forms)))
716 ;;; We expand copy loops inline in SUBSEQ and COPY-SEQ if we're copying
717 ;;; arrays with elements of size >= the word size. We do this because
718 ;;; we know the arrays cannot alias (one was just consed), therefore we
719 ;;; can determine at compile time the direction to copy, and for
720 ;;; word-sized elements, UB<WORD-SIZE>-BASH-COPY will do a bit of
721 ;;; needless checking to figure out what's going on. The same
722 ;;; considerations apply if we are copying elements larger than the word
723 ;;; size, with the additional twist that doing it inline is likely to
724 ;;; cons far less than calling REPLACE and letting generic code do the
727 ;;; However, we do not do this for elements whose size is < than the
728 ;;; word size because we don't want to deal with any alignment issues
729 ;;; inline. The UB*-BASH-COPY transforms might fix things up later
732 (defun maybe-expand-copy-loop-inline (src src-offset dst dst-offset length
734 (let ((saetp (find-saetp element-type)))
736 (if (>= (sb!vm:saetp-n-bits saetp) sb!vm:n-word-bits)
737 (expand-aref-copy-loop src src-offset dst dst-offset length)
738 `(locally (declare (optimize (safety 0)))
739 (replace ,dst ,src :start1 ,dst-offset :start2 ,src-offset :end1 ,length)))))
741 (defun expand-aref-copy-loop (src src-offset dst dst-offset length)
742 (if (eql src-offset dst-offset)
743 `(do ((i (+ ,src-offset ,length) (1- i)))
745 (declare (optimize (insert-array-bounds-checks 0)))
746 (setf (aref ,dst (1- i)) (aref ,src (1- i))))
747 ;; KLUDGE: The compiler is not able to derive that (+ offset
748 ;; length) must be a fixnum, but arrives at (unsigned-byte 29).
749 ;; We, however, know it must be so, as by this point the bounds
750 ;; have already been checked.
751 `(do ((i (truly-the fixnum (+ ,src-offset ,length)) (1- i))
752 (j (+ ,dst-offset ,length) (1- j)))
754 (declare (optimize (insert-array-bounds-checks 0))
755 (type (integer 0 #.sb!xc:array-dimension-limit) j i))
756 (setf (aref ,dst (1- j)) (aref ,src (1- i))))))
758 (deftransform subseq ((seq start &optional end)
759 ((or (simple-unboxed-array (*)) simple-vector) t &optional t)
761 (let ((array-type (lvar-type seq)))
762 (unless (array-type-p array-type)
763 (give-up-ir1-transform))
764 (let ((element-type (type-specifier (array-type-specialized-element-type array-type))))
765 `(let* ((length (length seq))
766 (end (or end length)))
767 ,(unless (policy node (= safety 0))
769 (unless (<= 0 start end length)
770 (sb!impl::signal-bounding-indices-bad-error seq start end))))
771 (let* ((size (- end start))
772 (result (make-array size :element-type ',element-type)))
773 ,(maybe-expand-copy-loop-inline 'seq (if (constant-lvar-p start)
776 'result 0 'size element-type)
779 (deftransform copy-seq ((seq) ((or (simple-unboxed-array (*)) simple-vector)) *)
780 (let ((array-type (lvar-type seq)))
781 (unless (array-type-p array-type)
782 (give-up-ir1-transform))
783 (let ((element-type (type-specifier (array-type-specialized-element-type array-type))))
784 `(let* ((length (length seq))
785 (result (make-array length :element-type ',element-type)))
786 ,(maybe-expand-copy-loop-inline 'seq 0 'result 0 'length element-type)
789 ;;; FIXME: it really should be possible to take advantage of the
790 ;;; macros used in code/seq.lisp here to avoid duplication of code,
791 ;;; and enable even funkier transformations.
792 (deftransform search ((pattern text &key (start1 0) (start2 0) end1 end2
796 (vector vector &rest t)
798 :policy (> speed (max space safety)))
800 (let ((from-end (when (lvar-p from-end)
801 (unless (constant-lvar-p from-end)
802 (give-up-ir1-transform ":FROM-END is not constant."))
803 (lvar-value from-end)))
805 (testp (lvar-p test)))
807 (let ((end1 (or end1 (length pattern)))
808 (end2 (or end2 (length text)))
810 '((key (coerce key 'function))))
812 '((test (coerce test 'function)))))
813 (declare (type index start1 start2 end1 end2))
815 '(index2 (- end2 (- end1 start1)) (1- index2))
816 '(index2 start2 (1+ index2))))
821 ;; INDEX2 is FIXNUM, not an INDEX, as right before the loop
822 ;; terminates is hits -1 when :FROM-END is true and :START2
824 (declare (type fixnum index2))
825 (when (do ((index1 start1 (1+ index1))
826 (index2 index2 (1+ index2)))
828 (declare (type index index1 index2))
830 '((when (= index2 end2)
831 (return-from search nil))))
836 '(funcall key (aref pattern index1))
837 '(aref pattern index1))
839 '(funcall key (aref text index2))
840 '(aref text index2)))
842 (return index2)))))))
844 ;;; FIXME: It seems as though it should be possible to make a DEFUN
845 ;;; %CONCATENATE (with a DEFTRANSFORM to translate constant RTYPE to
846 ;;; CTYPE before calling %CONCATENATE) which is comparably efficient,
847 ;;; at least once DYNAMIC-EXTENT works.
849 ;;; FIXME: currently KLUDGEed because of bug 188
851 ;;; FIXME: disabled for sb-unicode: probably want it back
853 (deftransform concatenate ((rtype &rest sequences)
854 (t &rest (or simple-base-string
855 (simple-array nil (*))))
857 :policy (< safety 3))
858 (loop for rest-seqs on sequences
859 for n-seq = (gensym "N-SEQ")
860 for n-length = (gensym "N-LENGTH")
861 for start = 0 then next-start
862 for next-start = (gensym "NEXT-START")
863 collect n-seq into args
864 collect `(,n-length (length ,n-seq)) into lets
865 collect n-length into all-lengths
866 collect next-start into starts
867 collect `(if (and (typep ,n-seq '(simple-array nil (*)))
869 (error 'nil-array-accessed-error)
870 (#.(let* ((i (position 'character sb!kernel::*specialized-array-element-types*))
871 (saetp (aref sb!vm:*specialized-array-element-type-properties* i))
872 (n-bits (sb!vm:saetp-n-bits saetp)))
873 (intern (format nil "UB~D-BASH-COPY" n-bits)
875 ,n-seq 0 res ,start ,n-length))
877 collect `(setq ,next-start (+ ,start ,n-length)) into forms
880 `(lambda (rtype ,@args)
881 (declare (ignore rtype))
883 (res (make-string (the index (+ ,@all-lengths))
884 :element-type 'base-char)))
885 (declare (type index ,@all-lengths))
886 (let (,@(mapcar (lambda (name) `(,name 0)) starts))
887 (declare (type index ,@starts))
891 ;;;; CONS accessor DERIVE-TYPE optimizers
893 (defoptimizer (car derive-type) ((cons))
894 (let ((type (lvar-type cons))
895 (null-type (specifier-type 'null)))
896 (cond ((eq type null-type)
899 (cons-type-car-type type)))))
901 (defoptimizer (cdr derive-type) ((cons))
902 (let ((type (lvar-type cons))
903 (null-type (specifier-type 'null)))
904 (cond ((eq type null-type)
907 (cons-type-cdr-type type)))))
909 ;;;; FIND, POSITION, and their -IF and -IF-NOT variants
911 ;;; We want to make sure that %FIND-POSITION is inline-expanded into
912 ;;; %FIND-POSITION-IF only when %FIND-POSITION-IF has an inline
913 ;;; expansion, so we factor out the condition into this function.
914 (defun check-inlineability-of-find-position-if (sequence from-end)
915 (let ((ctype (lvar-type sequence)))
916 (cond ((csubtypep ctype (specifier-type 'vector))
917 ;; It's not worth trying to inline vector code unless we
918 ;; know a fair amount about it at compile time.
919 (upgraded-element-type-specifier-or-give-up sequence)
920 (unless (constant-lvar-p from-end)
921 (give-up-ir1-transform
922 "FROM-END argument value not known at compile time")))
923 ((csubtypep ctype (specifier-type 'list))
924 ;; Inlining on lists is generally worthwhile.
927 (give-up-ir1-transform
928 "sequence type not known at compile time")))))
930 ;;; %FIND-POSITION-IF and %FIND-POSITION-IF-NOT for LIST data
931 (macrolet ((def (name condition)
932 `(deftransform ,name ((predicate sequence from-end start end key)
933 (function list t t t function)
935 :policy (> speed space))
940 (declare (type index index))
942 (if (and end (> end index))
943 (sb!impl::signal-bounding-indices-bad-error
945 (values find position)))
946 (let ((key-i (funcall key i)))
947 (when (and end (>= index end))
948 (return (values find position)))
949 (when (>= index start)
950 (,',condition (funcall predicate key-i)
951 ;; This hack of dealing with non-NIL
952 ;; FROM-END for list data by iterating
953 ;; forward through the list and keeping
954 ;; track of the last time we found a match
955 ;; might be more screwy than what the user
956 ;; expects, but it seems to be allowed by
957 ;; the ANSI standard. (And if the user is
958 ;; screwy enough to ask for FROM-END
959 ;; behavior on list data, turnabout is
962 ;; It's also not enormously efficient,
963 ;; calling PREDICATE and KEY more often
964 ;; than necessary; but all the
965 ;; alternatives seem to have their own
966 ;; efficiency problems.
970 (return (values i index))))))
972 (def %find-position-if when)
973 (def %find-position-if-not unless))
975 ;;; %FIND-POSITION for LIST data can be expanded into %FIND-POSITION-IF
976 ;;; without loss of efficiency. (I.e., the optimizer should be able
977 ;;; to straighten everything out.)
978 (deftransform %find-position ((item sequence from-end start end key test)
981 :policy (> speed space))
983 '(%find-position-if (let ((test-fun (%coerce-callable-to-fun test)))
984 ;; The order of arguments for asymmetric tests
985 ;; (e.g. #'<, as opposed to order-independent
986 ;; tests like #'=) is specified in the spec
987 ;; section 17.2.1 -- the O/Zi stuff there.
989 (funcall test-fun item i)))
994 (%coerce-callable-to-fun key)))
996 ;;; The inline expansions for the VECTOR case are saved as macros so
997 ;;; that we can share them between the DEFTRANSFORMs and the default
998 ;;; cases in the DEFUNs. (This isn't needed for the LIST case, because
999 ;;; the DEFTRANSFORMs for LIST are less choosy about when to expand.)
1000 (defun %find-position-or-find-position-if-vector-expansion (sequence-arg
1006 (with-unique-names (offset block index n-sequence sequence n-end end)
1007 `(let ((,n-sequence ,sequence-arg)
1009 (with-array-data ((,sequence ,n-sequence :offset-var ,offset)
1011 (,end (%check-vector-sequence-bounds
1012 ,n-sequence ,start ,n-end)))
1014 (macrolet ((maybe-return ()
1015 ;; WITH-ARRAY-DATA has already performed bounds
1016 ;; checking, so we can safely elide the checks
1017 ;; in the inner loop.
1018 '(let ((,element (locally (declare (optimize (insert-array-bounds-checks 0)))
1019 (aref ,sequence ,index))))
1023 (- ,index ,offset)))))))
1026 ;; (If we aren't fastidious about declaring that
1027 ;; INDEX might be -1, then (FIND 1 #() :FROM-END T)
1028 ;; can send us off into never-never land, since
1029 ;; INDEX is initialized to -1.)
1030 of-type index-or-minus-1
1031 from (1- ,end) downto ,start do
1033 (loop for ,index of-type index from ,start below ,end do
1035 (values nil nil))))))
1037 (def!macro %find-position-vector-macro (item sequence
1038 from-end start end key test)
1039 (with-unique-names (element)
1040 (%find-position-or-find-position-if-vector-expansion
1046 ;; (See the LIST transform for a discussion of the correct
1047 ;; argument order, i.e. whether the searched-for ,ITEM goes before
1048 ;; or after the checked sequence element.)
1049 `(funcall ,test ,item (funcall ,key ,element)))))
1051 (def!macro %find-position-if-vector-macro (predicate sequence
1052 from-end start end key)
1053 (with-unique-names (element)
1054 (%find-position-or-find-position-if-vector-expansion
1060 `(funcall ,predicate (funcall ,key ,element)))))
1062 (def!macro %find-position-if-not-vector-macro (predicate sequence
1063 from-end start end key)
1064 (with-unique-names (element)
1065 (%find-position-or-find-position-if-vector-expansion
1071 `(not (funcall ,predicate (funcall ,key ,element))))))
1073 ;;; %FIND-POSITION, %FIND-POSITION-IF and %FIND-POSITION-IF-NOT for
1075 (deftransform %find-position-if ((predicate sequence from-end start end key)
1076 (function vector t t t function)
1078 :policy (> speed space))
1080 (check-inlineability-of-find-position-if sequence from-end)
1081 '(%find-position-if-vector-macro predicate sequence
1082 from-end start end key))
1084 (deftransform %find-position-if-not ((predicate sequence from-end start end key)
1085 (function vector t t t function)
1087 :policy (> speed space))
1089 (check-inlineability-of-find-position-if sequence from-end)
1090 '(%find-position-if-not-vector-macro predicate sequence
1091 from-end start end key))
1093 (deftransform %find-position ((item sequence from-end start end key test)
1094 (t vector t t t function function)
1096 :policy (> speed space))
1098 (check-inlineability-of-find-position-if sequence from-end)
1099 '(%find-position-vector-macro item sequence
1100 from-end start end key test))
1102 ;;; logic to unravel :TEST, :TEST-NOT, and :KEY options in FIND,
1103 ;;; POSITION-IF, etc.
1104 (define-source-transform effective-find-position-test (test test-not)
1105 (once-only ((test test)
1106 (test-not test-not))
1108 ((and ,test ,test-not)
1109 (error "can't specify both :TEST and :TEST-NOT"))
1110 (,test (%coerce-callable-to-fun ,test))
1112 ;; (Without DYNAMIC-EXTENT, this is potentially horribly
1113 ;; inefficient, but since the TEST-NOT option is deprecated
1114 ;; anyway, we don't care.)
1115 (complement (%coerce-callable-to-fun ,test-not)))
1117 (define-source-transform effective-find-position-key (key)
1118 (once-only ((key key))
1120 (%coerce-callable-to-fun ,key)
1123 (macrolet ((define-find-position (fun-name values-index)
1124 `(deftransform ,fun-name ((item sequence &key
1125 from-end (start 0) end
1127 (t (or list vector) &rest t))
1128 '(nth-value ,values-index
1129 (%find-position item sequence
1132 (effective-find-position-key key)
1133 (effective-find-position-test
1135 (define-find-position find 0)
1136 (define-find-position position 1))
1138 (macrolet ((define-find-position-if (fun-name values-index)
1139 `(deftransform ,fun-name ((predicate sequence &key
1142 (t (or list vector) &rest t))
1145 (%find-position-if (%coerce-callable-to-fun predicate)
1148 (effective-find-position-key key))))))
1149 (define-find-position-if find-if 0)
1150 (define-find-position-if position-if 1))
1152 ;;; the deprecated functions FIND-IF-NOT and POSITION-IF-NOT. We
1153 ;;; didn't bother to worry about optimizing them, except note that on
1154 ;;; Sat, Oct 06, 2001 at 04:22:38PM +0100, Christophe Rhodes wrote on
1157 ;;; My understanding is that while the :test-not argument is
1158 ;;; deprecated in favour of :test (complement #'foo) because of
1159 ;;; semantic difficulties (what happens if both :test and :test-not
1160 ;;; are supplied, etc) the -if-not variants, while officially
1161 ;;; deprecated, would be undeprecated were X3J13 actually to produce
1162 ;;; a revised standard, as there are perfectly legitimate idiomatic
1163 ;;; reasons for allowing the -if-not versions equal status,
1164 ;;; particularly remove-if-not (== filter).
1166 ;;; This is only an informal understanding, I grant you, but
1167 ;;; perhaps it's worth optimizing the -if-not versions in the same
1168 ;;; way as the others?
1170 ;;; FIXME: Maybe remove uses of these deprecated functions within the
1171 ;;; implementation of SBCL.
1172 (macrolet ((define-find-position-if-not (fun-name values-index)
1173 `(deftransform ,fun-name ((predicate sequence &key
1176 (t (or list vector) &rest t))
1179 (%find-position-if-not (%coerce-callable-to-fun predicate)
1182 (effective-find-position-key key))))))
1183 (define-find-position-if-not find-if-not 0)
1184 (define-find-position-if-not position-if-not 1))