1 ;;;; This file contains macro-like source transformations which
2 ;;;; convert uses of certain functions into the canonical form desired
3 ;;;; within the compiler. FIXME: and other IR1 transforms and stuff.
5 ;;;; This software is part of the SBCL system. See the README file for
8 ;;;; This software is derived from the CMU CL system, which was
9 ;;;; written at Carnegie Mellon University and released into the
10 ;;;; public domain. The software is in the public domain and is
11 ;;;; provided with absolutely no warranty. See the COPYING and CREDITS
12 ;;;; files for more information.
16 ;;; We turn IDENTITY into PROG1 so that it is obvious that it just
17 ;;; returns the first value of its argument. Ditto for VALUES with one
19 (define-source-transform identity (x) `(prog1 ,x))
20 (define-source-transform values (x) `(prog1 ,x))
22 ;;; CONSTANTLY is pretty much never worth transforming, but it's good to get the type.
23 (defoptimizer (constantly derive-type) ((value))
25 `(function (&rest t) (values ,(type-specifier (lvar-type value)) &optional))))
27 ;;; If the function has a known number of arguments, then return a
28 ;;; lambda with the appropriate fixed number of args. If the
29 ;;; destination is a FUNCALL, then do the &REST APPLY thing, and let
30 ;;; MV optimization figure things out.
31 (deftransform complement ((fun) * * :node node)
33 (multiple-value-bind (min max)
34 (fun-type-nargs (lvar-type fun))
36 ((and min (eql min max))
37 (let ((dums (make-gensym-list min)))
38 `#'(lambda ,dums (not (funcall fun ,@dums)))))
39 ((awhen (node-lvar node)
40 (let ((dest (lvar-dest it)))
41 (and (combination-p dest)
42 (eq (combination-fun dest) it))))
43 '#'(lambda (&rest args)
44 (not (apply fun args))))
46 (give-up-ir1-transform
47 "The function doesn't have a fixed argument count.")))))
50 (defun derive-symbol-value-type (lvar node)
51 (if (constant-lvar-p lvar)
52 (let* ((sym (lvar-value lvar))
53 (var (maybe-find-free-var sym))
55 (let ((*lexenv* (node-lexenv node)))
56 (lexenv-find var type-restrictions))))
57 (global-type (info :variable :type sym)))
59 (type-intersection local-type global-type)
63 (defoptimizer (symbol-value derive-type) ((symbol) node)
64 (derive-symbol-value-type symbol node))
66 (defoptimizer (symbol-global-value derive-type) ((symbol) node)
67 (derive-symbol-value-type symbol node))
71 ;;; Translate CxR into CAR/CDR combos.
72 (defun source-transform-cxr (form)
73 (if (/= (length form) 2)
75 (let* ((name (car form))
79 (leaf (leaf-source-name name))))))
80 (do ((i (- (length string) 2) (1- i))
82 `(,(ecase (char string i)
88 ;;; Make source transforms to turn CxR forms into combinations of CAR
89 ;;; and CDR. ANSI specifies that everything up to 4 A/D operations is
91 ;;; Don't transform CAD*R, they are treated specially for &more args
94 (/show0 "about to set CxR source transforms")
95 (loop for i of-type index from 2 upto 4 do
96 ;; Iterate over BUF = all names CxR where x = an I-element
97 ;; string of #\A or #\D characters.
98 (let ((buf (make-string (+ 2 i))))
99 (setf (aref buf 0) #\C
100 (aref buf (1+ i)) #\R)
101 (dotimes (j (ash 2 i))
102 (declare (type index j))
104 (declare (type index k))
105 (setf (aref buf (1+ k))
106 (if (logbitp k j) #\A #\D)))
107 (unless (member buf '("CADR" "CADDR" "CADDDR")
109 (setf (info :function :source-transform (intern buf))
110 #'source-transform-cxr)))))
111 (/show0 "done setting CxR source transforms")
113 ;;; Turn FIRST..FOURTH and REST into the obvious synonym, assuming
114 ;;; whatever is right for them is right for us. FIFTH..TENTH turn into
115 ;;; Nth, which can be expanded into a CAR/CDR later on if policy
117 (define-source-transform rest (x) `(cdr ,x))
118 (define-source-transform first (x) `(car ,x))
119 (define-source-transform second (x) `(cadr ,x))
120 (define-source-transform third (x) `(caddr ,x))
121 (define-source-transform fourth (x) `(cadddr ,x))
122 (define-source-transform fifth (x) `(nth 4 ,x))
123 (define-source-transform sixth (x) `(nth 5 ,x))
124 (define-source-transform seventh (x) `(nth 6 ,x))
125 (define-source-transform eighth (x) `(nth 7 ,x))
126 (define-source-transform ninth (x) `(nth 8 ,x))
127 (define-source-transform tenth (x) `(nth 9 ,x))
129 ;;; LIST with one arg is an extremely common operation (at least inside
130 ;;; SBCL itself); translate it to CONS to take advantage of common
131 ;;; allocation routines.
132 (define-source-transform list (&rest args)
134 (1 `(cons ,(first args) nil))
137 ;;; And similarly for LIST*.
138 (define-source-transform list* (arg &rest others)
139 (cond ((not others) arg)
140 ((not (cdr others)) `(cons ,arg ,(car others)))
143 (defoptimizer (list* derive-type) ((arg &rest args))
145 (specifier-type 'cons)
148 ;;; Translate RPLACx to LET and SETF.
149 (define-source-transform rplaca (x y)
154 (define-source-transform rplacd (x y)
160 (deftransform last ((list &optional n) (t &optional t))
161 (let ((c (constant-lvar-p n)))
163 (and c (eql 1 (lvar-value n))))
165 ((and c (eql 0 (lvar-value n)))
168 (let ((type (lvar-type n)))
169 (cond ((csubtypep type (specifier-type 'fixnum))
170 '(%lastn/fixnum list n))
171 ((csubtypep type (specifier-type 'bignum))
172 '(%lastn/bignum list n))
174 (give-up-ir1-transform "second argument type too vague"))))))))
176 (define-source-transform gethash (&rest args)
178 (2 `(sb!impl::gethash3 ,@args nil))
179 (3 `(sb!impl::gethash3 ,@args))
181 (define-source-transform get (&rest args)
183 (2 `(sb!impl::get2 ,@args))
184 (3 `(sb!impl::get3 ,@args))
187 (defvar *default-nthcdr-open-code-limit* 6)
188 (defvar *extreme-nthcdr-open-code-limit* 20)
190 (deftransform nthcdr ((n l) (unsigned-byte t) * :node node)
191 "convert NTHCDR to CAxxR"
192 (unless (constant-lvar-p n)
193 (give-up-ir1-transform))
194 (let ((n (lvar-value n)))
196 (if (policy node (and (= speed 3) (= space 0)))
197 *extreme-nthcdr-open-code-limit*
198 *default-nthcdr-open-code-limit*))
199 (give-up-ir1-transform))
204 `(cdr ,(frob (1- n))))))
207 ;;;; arithmetic and numerology
209 (define-source-transform plusp (x) `(> ,x 0))
210 (define-source-transform minusp (x) `(< ,x 0))
211 (define-source-transform zerop (x) `(= ,x 0))
213 (define-source-transform 1+ (x) `(+ ,x 1))
214 (define-source-transform 1- (x) `(- ,x 1))
216 (define-source-transform oddp (x) `(logtest ,x 1))
217 (define-source-transform evenp (x) `(not (logtest ,x 1)))
219 ;;; Note that all the integer division functions are available for
220 ;;; inline expansion.
222 (macrolet ((deffrob (fun)
223 `(define-source-transform ,fun (x &optional (y nil y-p))
230 #-sb-xc-host ; (See CROSS-FLOAT-INFINITY-KLUDGE.)
232 #-sb-xc-host ; (See CROSS-FLOAT-INFINITY-KLUDGE.)
235 ;;; This used to be a source transform (hence the lack of restrictions
236 ;;; on the argument types), but we make it a regular transform so that
237 ;;; the VM has a chance to see the bare LOGTEST and potentiall choose
238 ;;; to implement it differently. --njf, 06-02-2006
239 (deftransform logtest ((x y) * *)
240 `(not (zerop (logand x y))))
242 (deftransform logbitp
243 ((index integer) (unsigned-byte (or (signed-byte #.sb!vm:n-word-bits)
244 (unsigned-byte #.sb!vm:n-word-bits))))
245 `(if (>= index #.sb!vm:n-word-bits)
247 (not (zerop (logand integer (ash 1 index))))))
249 (define-source-transform byte (size position)
250 `(cons ,size ,position))
251 (define-source-transform byte-size (spec) `(car ,spec))
252 (define-source-transform byte-position (spec) `(cdr ,spec))
253 (define-source-transform ldb-test (bytespec integer)
254 `(not (zerop (mask-field ,bytespec ,integer))))
256 ;;; With the ratio and complex accessors, we pick off the "identity"
257 ;;; case, and use a primitive to handle the cell access case.
258 (define-source-transform numerator (num)
259 (once-only ((n-num `(the rational ,num)))
263 (define-source-transform denominator (num)
264 (once-only ((n-num `(the rational ,num)))
266 (%denominator ,n-num)
269 ;;;; interval arithmetic for computing bounds
271 ;;;; This is a set of routines for operating on intervals. It
272 ;;;; implements a simple interval arithmetic package. Although SBCL
273 ;;;; has an interval type in NUMERIC-TYPE, we choose to use our own
274 ;;;; for two reasons:
276 ;;;; 1. This package is simpler than NUMERIC-TYPE.
278 ;;;; 2. It makes debugging much easier because you can just strip
279 ;;;; out these routines and test them independently of SBCL. (This is a
282 ;;;; One disadvantage is a probable increase in consing because we
283 ;;;; have to create these new interval structures even though
284 ;;;; numeric-type has everything we want to know. Reason 2 wins for
287 ;;; Support operations that mimic real arithmetic comparison
288 ;;; operators, but imposing a total order on the floating points such
289 ;;; that negative zeros are strictly less than positive zeros.
290 (macrolet ((def (name op)
293 (if (and (floatp x) (floatp y) (zerop x) (zerop y))
294 (,op (float-sign x) (float-sign y))
296 (def signed-zero->= >=)
297 (def signed-zero-> >)
298 (def signed-zero-= =)
299 (def signed-zero-< <)
300 (def signed-zero-<= <=))
302 ;;; The basic interval type. It can handle open and closed intervals.
303 ;;; A bound is open if it is a list containing a number, just like
304 ;;; Lisp says. NIL means unbounded.
305 (defstruct (interval (:constructor %make-interval)
309 (defun make-interval (&key low high)
310 (labels ((normalize-bound (val)
313 (float-infinity-p val))
314 ;; Handle infinities.
318 ;; Handle any closed bounds.
321 ;; We have an open bound. Normalize the numeric
322 ;; bound. If the normalized bound is still a number
323 ;; (not nil), keep the bound open. Otherwise, the
324 ;; bound is really unbounded, so drop the openness.
325 (let ((new-val (normalize-bound (first val))))
327 ;; The bound exists, so keep it open still.
330 (error "unknown bound type in MAKE-INTERVAL")))))
331 (%make-interval :low (normalize-bound low)
332 :high (normalize-bound high))))
334 ;;; Given a number X, create a form suitable as a bound for an
335 ;;; interval. Make the bound open if OPEN-P is T. NIL remains NIL.
336 #!-sb-fluid (declaim (inline set-bound))
337 (defun set-bound (x open-p)
338 (if (and x open-p) (list x) x))
340 ;;; Apply the function F to a bound X. If X is an open bound and the
341 ;;; function is declared strictly monotonic, then the result will be
342 ;;; open. IF X is NIL, the result is NIL.
343 (defun bound-func (f x strict)
344 (declare (type function f))
347 (with-float-traps-masked (:underflow :overflow :inexact :divide-by-zero)
348 ;; With these traps masked, we might get things like infinity
349 ;; or negative infinity returned. Check for this and return
350 ;; NIL to indicate unbounded.
351 (let ((y (funcall f (type-bound-number x))))
353 (float-infinity-p y))
355 (set-bound y (and strict (consp x))))))
356 ;; Some numerical operations will signal SIMPLE-TYPE-ERROR, e.g.
357 ;; in the course of converting a bignum to a float. Default to
359 (simple-type-error ()))))
361 (defun safe-double-coercion-p (x)
362 (or (typep x 'double-float)
363 (<= most-negative-double-float x most-positive-double-float)))
365 (defun safe-single-coercion-p (x)
366 (or (typep x 'single-float)
368 ;; Fix for bug 420, and related issues: during type derivation we often
369 ;; end up deriving types for both
371 ;; (some-op <int> <single>)
373 ;; (some-op (coerce <int> 'single-float) <single>)
375 ;; or other equivalent transformed forms. The problem with this
376 ;; is that on x86 (+ <int> <single>) is on the machine level
379 ;; (coerce (+ (coerce <int> 'double-float)
380 ;; (coerce <single> 'double-float))
383 ;; so if the result of (coerce <int> 'single-float) is not exact, the
384 ;; derived types for the transformed forms will have an empty
385 ;; intersection -- which in turn means that the compiler will conclude
386 ;; that the call never returns, and all hell breaks lose when it *does*
387 ;; return at runtime. (This affects not just +, but other operators are
390 ;; See also: SAFE-CTYPE-FOR-SINGLE-COERCION-P
392 ;; FIXME: If we ever add SSE-support for x86, this conditional needs to
395 (not (typep x `(or (integer * (,most-negative-exactly-single-float-fixnum))
396 (integer (,most-positive-exactly-single-float-fixnum) *))))
397 (<= most-negative-single-float x most-positive-single-float))))
399 ;;; Apply a binary operator OP to two bounds X and Y. The result is
400 ;;; NIL if either is NIL. Otherwise bound is computed and the result
401 ;;; is open if either X or Y is open.
403 ;;; FIXME: only used in this file, not needed in target runtime
405 ;;; ANSI contaigon specifies coercion to floating point if one of the
406 ;;; arguments is floating point. Here we should check to be sure that
407 ;;; the other argument is within the bounds of that floating point
410 (defmacro safely-binop (op x y)
412 ((typep ,x 'double-float)
413 (when (safe-double-coercion-p ,y)
415 ((typep ,y 'double-float)
416 (when (safe-double-coercion-p ,x)
418 ((typep ,x 'single-float)
419 (when (safe-single-coercion-p ,y)
421 ((typep ,y 'single-float)
422 (when (safe-single-coercion-p ,x)
426 (defmacro bound-binop (op x y)
427 (with-unique-names (xb yb res)
429 (with-float-traps-masked (:underflow :overflow :inexact :divide-by-zero)
430 (let* ((,xb (type-bound-number ,x))
431 (,yb (type-bound-number ,y))
432 (,res (safely-binop ,op ,xb ,yb)))
434 (and (or (consp ,x) (consp ,y))
435 ;; Open bounds can very easily be messed up
436 ;; by FP rounding, so take care here.
439 ;; Multiplying a greater-than-zero with
440 ;; less than one can round to zero.
441 `(or (not (fp-zero-p ,res))
442 (cond ((and (consp ,x) (fp-zero-p ,xb))
444 ((and (consp ,y) (fp-zero-p ,yb))
447 ;; Dividing a greater-than-zero with
448 ;; greater than one can round to zero.
449 `(or (not (fp-zero-p ,res))
450 (cond ((and (consp ,x) (fp-zero-p ,xb))
452 ((and (consp ,y) (fp-zero-p ,yb))
455 ;; Adding or subtracting greater-than-zero
456 ;; can end up with identity.
457 `(and (not (fp-zero-p ,xb))
458 (not (fp-zero-p ,yb))))))))))))
460 (defun coercion-loses-precision-p (val type)
463 (double-float (subtypep type 'single-float))
464 (rational (subtypep type 'float))
465 (t (bug "Unexpected arguments to bounds coercion: ~S ~S" val type))))
467 (defun coerce-for-bound (val type)
469 (let ((xbound (coerce-for-bound (car val) type)))
470 (if (coercion-loses-precision-p (car val) type)
474 ((subtypep type 'double-float)
475 (if (<= most-negative-double-float val most-positive-double-float)
477 ((or (subtypep type 'single-float) (subtypep type 'float))
478 ;; coerce to float returns a single-float
479 (if (<= most-negative-single-float val most-positive-single-float)
481 (t (coerce val type)))))
483 (defun coerce-and-truncate-floats (val type)
486 (let ((xbound (coerce-for-bound (car val) type)))
487 (if (coercion-loses-precision-p (car val) type)
491 ((subtypep type 'double-float)
492 (if (<= most-negative-double-float val most-positive-double-float)
494 (if (< val most-negative-double-float)
495 most-negative-double-float most-positive-double-float)))
496 ((or (subtypep type 'single-float) (subtypep type 'float))
497 ;; coerce to float returns a single-float
498 (if (<= most-negative-single-float val most-positive-single-float)
500 (if (< val most-negative-single-float)
501 most-negative-single-float most-positive-single-float)))
502 (t (coerce val type))))))
504 ;;; Convert a numeric-type object to an interval object.
505 (defun numeric-type->interval (x)
506 (declare (type numeric-type x))
507 (make-interval :low (numeric-type-low x)
508 :high (numeric-type-high x)))
510 (defun type-approximate-interval (type)
511 (declare (type ctype type))
512 (let ((types (prepare-arg-for-derive-type type))
515 (let ((type (if (member-type-p type)
516 (convert-member-type type)
518 (unless (numeric-type-p type)
519 (return-from type-approximate-interval nil))
520 (let ((interval (numeric-type->interval type)))
523 (interval-approximate-union result interval)
527 (defun copy-interval-limit (limit)
532 (defun copy-interval (x)
533 (declare (type interval x))
534 (make-interval :low (copy-interval-limit (interval-low x))
535 :high (copy-interval-limit (interval-high x))))
537 ;;; Given a point P contained in the interval X, split X into two
538 ;;; intervals at the point P. If CLOSE-LOWER is T, then the left
539 ;;; interval contains P. If CLOSE-UPPER is T, the right interval
540 ;;; contains P. You can specify both to be T or NIL.
541 (defun interval-split (p x &optional close-lower close-upper)
542 (declare (type number p)
544 (list (make-interval :low (copy-interval-limit (interval-low x))
545 :high (if close-lower p (list p)))
546 (make-interval :low (if close-upper (list p) p)
547 :high (copy-interval-limit (interval-high x)))))
549 ;;; Return the closure of the interval. That is, convert open bounds
550 ;;; to closed bounds.
551 (defun interval-closure (x)
552 (declare (type interval x))
553 (make-interval :low (type-bound-number (interval-low x))
554 :high (type-bound-number (interval-high x))))
556 ;;; For an interval X, if X >= POINT, return '+. If X <= POINT, return
557 ;;; '-. Otherwise return NIL.
558 (defun interval-range-info (x &optional (point 0))
559 (declare (type interval x))
560 (let ((lo (interval-low x))
561 (hi (interval-high x)))
562 (cond ((and lo (signed-zero->= (type-bound-number lo) point))
564 ((and hi (signed-zero->= point (type-bound-number hi)))
569 ;;; Test to see whether the interval X is bounded. HOW determines the
570 ;;; test, and should be either ABOVE, BELOW, or BOTH.
571 (defun interval-bounded-p (x how)
572 (declare (type interval x))
579 (and (interval-low x) (interval-high x)))))
581 ;;; See whether the interval X contains the number P, taking into
582 ;;; account that the interval might not be closed.
583 (defun interval-contains-p (p x)
584 (declare (type number p)
586 ;; Does the interval X contain the number P? This would be a lot
587 ;; easier if all intervals were closed!
588 (let ((lo (interval-low x))
589 (hi (interval-high x)))
591 ;; The interval is bounded
592 (if (and (signed-zero-<= (type-bound-number lo) p)
593 (signed-zero-<= p (type-bound-number hi)))
594 ;; P is definitely in the closure of the interval.
595 ;; We just need to check the end points now.
596 (cond ((signed-zero-= p (type-bound-number lo))
598 ((signed-zero-= p (type-bound-number hi))
603 ;; Interval with upper bound
604 (if (signed-zero-< p (type-bound-number hi))
606 (and (numberp hi) (signed-zero-= p hi))))
608 ;; Interval with lower bound
609 (if (signed-zero-> p (type-bound-number lo))
611 (and (numberp lo) (signed-zero-= p lo))))
613 ;; Interval with no bounds
616 ;;; Determine whether two intervals X and Y intersect. Return T if so.
617 ;;; If CLOSED-INTERVALS-P is T, the treat the intervals as if they
618 ;;; were closed. Otherwise the intervals are treated as they are.
620 ;;; Thus if X = [0, 1) and Y = (1, 2), then they do not intersect
621 ;;; because no element in X is in Y. However, if CLOSED-INTERVALS-P
622 ;;; is T, then they do intersect because we use the closure of X = [0,
623 ;;; 1] and Y = [1, 2] to determine intersection.
624 (defun interval-intersect-p (x y &optional closed-intervals-p)
625 (declare (type interval x y))
626 (and (interval-intersection/difference (if closed-intervals-p
629 (if closed-intervals-p
634 ;;; Are the two intervals adjacent? That is, is there a number
635 ;;; between the two intervals that is not an element of either
636 ;;; interval? If so, they are not adjacent. For example [0, 1) and
637 ;;; [1, 2] are adjacent but [0, 1) and (1, 2] are not because 1 lies
638 ;;; between both intervals.
639 (defun interval-adjacent-p (x y)
640 (declare (type interval x y))
641 (flet ((adjacent (lo hi)
642 ;; Check to see whether lo and hi are adjacent. If either is
643 ;; nil, they can't be adjacent.
644 (when (and lo hi (= (type-bound-number lo) (type-bound-number hi)))
645 ;; The bounds are equal. They are adjacent if one of
646 ;; them is closed (a number). If both are open (consp),
647 ;; then there is a number that lies between them.
648 (or (numberp lo) (numberp hi)))))
649 (or (adjacent (interval-low y) (interval-high x))
650 (adjacent (interval-low x) (interval-high y)))))
652 ;;; Compute the intersection and difference between two intervals.
653 ;;; Two values are returned: the intersection and the difference.
655 ;;; Let the two intervals be X and Y, and let I and D be the two
656 ;;; values returned by this function. Then I = X intersect Y. If I
657 ;;; is NIL (the empty set), then D is X union Y, represented as the
658 ;;; list of X and Y. If I is not the empty set, then D is (X union Y)
659 ;;; - I, which is a list of two intervals.
661 ;;; For example, let X = [1,5] and Y = [-1,3). Then I = [1,3) and D =
662 ;;; [-1,1) union [3,5], which is returned as a list of two intervals.
663 (defun interval-intersection/difference (x y)
664 (declare (type interval x y))
665 (let ((x-lo (interval-low x))
666 (x-hi (interval-high x))
667 (y-lo (interval-low y))
668 (y-hi (interval-high y)))
671 ;; If p is an open bound, make it closed. If p is a closed
672 ;; bound, make it open.
676 (test-number (p int bound)
677 ;; Test whether P is in the interval.
678 (let ((pn (type-bound-number p)))
679 (when (interval-contains-p pn (interval-closure int))
680 ;; Check for endpoints.
681 (let* ((lo (interval-low int))
682 (hi (interval-high int))
683 (lon (type-bound-number lo))
684 (hin (type-bound-number hi)))
686 ;; Interval may be a point.
687 ((and lon hin (= lon hin pn))
688 (and (numberp p) (numberp lo) (numberp hi)))
689 ;; Point matches the low end.
690 ;; [P] [P,?} => TRUE [P] (P,?} => FALSE
691 ;; (P [P,?} => TRUE P) [P,?} => FALSE
692 ;; (P (P,?} => TRUE P) (P,?} => FALSE
693 ((and lon (= pn lon))
694 (or (and (numberp p) (numberp lo))
695 (and (consp p) (eq :low bound))))
696 ;; [P] {?,P] => TRUE [P] {?,P) => FALSE
697 ;; P) {?,P] => TRUE (P {?,P] => FALSE
698 ;; P) {?,P) => TRUE (P {?,P) => FALSE
699 ((and hin (= pn hin))
700 (or (and (numberp p) (numberp hi))
701 (and (consp p) (eq :high bound))))
702 ;; Not an endpoint, all is well.
705 (test-lower-bound (p int)
706 ;; P is a lower bound of an interval.
708 (test-number p int :low)
709 (not (interval-bounded-p int 'below))))
710 (test-upper-bound (p int)
711 ;; P is an upper bound of an interval.
713 (test-number p int :high)
714 (not (interval-bounded-p int 'above)))))
715 (let ((x-lo-in-y (test-lower-bound x-lo y))
716 (x-hi-in-y (test-upper-bound x-hi y))
717 (y-lo-in-x (test-lower-bound y-lo x))
718 (y-hi-in-x (test-upper-bound y-hi x)))
719 (cond ((or x-lo-in-y x-hi-in-y y-lo-in-x y-hi-in-x)
720 ;; Intervals intersect. Let's compute the intersection
721 ;; and the difference.
722 (multiple-value-bind (lo left-lo left-hi)
723 (cond (x-lo-in-y (values x-lo y-lo (opposite-bound x-lo)))
724 (y-lo-in-x (values y-lo x-lo (opposite-bound y-lo))))
725 (multiple-value-bind (hi right-lo right-hi)
727 (values x-hi (opposite-bound x-hi) y-hi))
729 (values y-hi (opposite-bound y-hi) x-hi)))
730 (values (make-interval :low lo :high hi)
731 (list (make-interval :low left-lo
733 (make-interval :low right-lo
736 (values nil (list x y))))))))
738 ;;; If intervals X and Y intersect, return a new interval that is the
739 ;;; union of the two. If they do not intersect, return NIL.
740 (defun interval-merge-pair (x y)
741 (declare (type interval x y))
742 ;; If x and y intersect or are adjacent, create the union.
743 ;; Otherwise return nil
744 (when (or (interval-intersect-p x y)
745 (interval-adjacent-p x y))
746 (flet ((select-bound (x1 x2 min-op max-op)
747 (let ((x1-val (type-bound-number x1))
748 (x2-val (type-bound-number x2)))
750 ;; Both bounds are finite. Select the right one.
751 (cond ((funcall min-op x1-val x2-val)
752 ;; x1 is definitely better.
754 ((funcall max-op x1-val x2-val)
755 ;; x2 is definitely better.
758 ;; Bounds are equal. Select either
759 ;; value and make it open only if
761 (set-bound x1-val (and (consp x1) (consp x2))))))
763 ;; At least one bound is not finite. The
764 ;; non-finite bound always wins.
766 (let* ((x-lo (copy-interval-limit (interval-low x)))
767 (x-hi (copy-interval-limit (interval-high x)))
768 (y-lo (copy-interval-limit (interval-low y)))
769 (y-hi (copy-interval-limit (interval-high y))))
770 (make-interval :low (select-bound x-lo y-lo #'< #'>)
771 :high (select-bound x-hi y-hi #'> #'<))))))
773 ;;; return the minimal interval, containing X and Y
774 (defun interval-approximate-union (x y)
775 (cond ((interval-merge-pair x y))
777 (make-interval :low (copy-interval-limit (interval-low x))
778 :high (copy-interval-limit (interval-high y))))
780 (make-interval :low (copy-interval-limit (interval-low y))
781 :high (copy-interval-limit (interval-high x))))))
783 ;;; basic arithmetic operations on intervals. We probably should do
784 ;;; true interval arithmetic here, but it's complicated because we
785 ;;; have float and integer types and bounds can be open or closed.
787 ;;; the negative of an interval
788 (defun interval-neg (x)
789 (declare (type interval x))
790 (make-interval :low (bound-func #'- (interval-high x) t)
791 :high (bound-func #'- (interval-low x) t)))
793 ;;; Add two intervals.
794 (defun interval-add (x y)
795 (declare (type interval x y))
796 (make-interval :low (bound-binop + (interval-low x) (interval-low y))
797 :high (bound-binop + (interval-high x) (interval-high y))))
799 ;;; Subtract two intervals.
800 (defun interval-sub (x y)
801 (declare (type interval x y))
802 (make-interval :low (bound-binop - (interval-low x) (interval-high y))
803 :high (bound-binop - (interval-high x) (interval-low y))))
805 ;;; Multiply two intervals.
806 (defun interval-mul (x y)
807 (declare (type interval x y))
808 (flet ((bound-mul (x y)
809 (cond ((or (null x) (null y))
810 ;; Multiply by infinity is infinity
812 ((or (and (numberp x) (zerop x))
813 (and (numberp y) (zerop y)))
814 ;; Multiply by closed zero is special. The result
815 ;; is always a closed bound. But don't replace this
816 ;; with zero; we want the multiplication to produce
817 ;; the correct signed zero, if needed. Use SIGNUM
818 ;; to avoid trying to multiply huge bignums with 0.0.
819 (* (signum (type-bound-number x)) (signum (type-bound-number y))))
820 ((or (and (floatp x) (float-infinity-p x))
821 (and (floatp y) (float-infinity-p y)))
822 ;; Infinity times anything is infinity
825 ;; General multiply. The result is open if either is open.
826 (bound-binop * x y)))))
827 (let ((x-range (interval-range-info x))
828 (y-range (interval-range-info y)))
829 (cond ((null x-range)
830 ;; Split x into two and multiply each separately
831 (destructuring-bind (x- x+) (interval-split 0 x t t)
832 (interval-merge-pair (interval-mul x- y)
833 (interval-mul x+ y))))
835 ;; Split y into two and multiply each separately
836 (destructuring-bind (y- y+) (interval-split 0 y t t)
837 (interval-merge-pair (interval-mul x y-)
838 (interval-mul x y+))))
840 (interval-neg (interval-mul (interval-neg x) y)))
842 (interval-neg (interval-mul x (interval-neg y))))
843 ((and (eq x-range '+) (eq y-range '+))
844 ;; If we are here, X and Y are both positive.
846 :low (bound-mul (interval-low x) (interval-low y))
847 :high (bound-mul (interval-high x) (interval-high y))))
849 (bug "excluded case in INTERVAL-MUL"))))))
851 ;;; Divide two intervals.
852 (defun interval-div (top bot)
853 (declare (type interval top bot))
854 (flet ((bound-div (x y y-low-p)
857 ;; Divide by infinity means result is 0. However,
858 ;; we need to watch out for the sign of the result,
859 ;; to correctly handle signed zeros. We also need
860 ;; to watch out for positive or negative infinity.
861 (if (floatp (type-bound-number x))
863 (- (float-sign (type-bound-number x) 0.0))
864 (float-sign (type-bound-number x) 0.0))
866 ((zerop (type-bound-number y))
867 ;; Divide by zero means result is infinity
870 (bound-binop / x y)))))
871 (let ((top-range (interval-range-info top))
872 (bot-range (interval-range-info bot)))
873 (cond ((null bot-range)
874 ;; The denominator contains zero, so anything goes!
875 (make-interval :low nil :high nil))
877 ;; Denominator is negative so flip the sign, compute the
878 ;; result, and flip it back.
879 (interval-neg (interval-div top (interval-neg bot))))
881 ;; Split top into two positive and negative parts, and
882 ;; divide each separately
883 (destructuring-bind (top- top+) (interval-split 0 top t t)
884 (interval-merge-pair (interval-div top- bot)
885 (interval-div top+ bot))))
887 ;; Top is negative so flip the sign, divide, and flip the
888 ;; sign of the result.
889 (interval-neg (interval-div (interval-neg top) bot)))
890 ((and (eq top-range '+) (eq bot-range '+))
893 :low (bound-div (interval-low top) (interval-high bot) t)
894 :high (bound-div (interval-high top) (interval-low bot) nil)))
896 (bug "excluded case in INTERVAL-DIV"))))))
898 ;;; Apply the function F to the interval X. If X = [a, b], then the
899 ;;; result is [f(a), f(b)]. It is up to the user to make sure the
900 ;;; result makes sense. It will if F is monotonic increasing (or, if
901 ;;; the interval is closed, non-decreasing).
903 ;;; (Actually most uses of INTERVAL-FUNC are coercions to float types,
904 ;;; which are not monotonic increasing, so default to calling
905 ;;; BOUND-FUNC with a non-strict argument).
906 (defun interval-func (f x &optional increasing)
907 (declare (type function f)
909 (let ((lo (bound-func f (interval-low x) increasing))
910 (hi (bound-func f (interval-high x) increasing)))
911 (make-interval :low lo :high hi)))
913 ;;; Return T if X < Y. That is every number in the interval X is
914 ;;; always less than any number in the interval Y.
915 (defun interval-< (x y)
916 (declare (type interval x y))
917 ;; X < Y only if X is bounded above, Y is bounded below, and they
919 (when (and (interval-bounded-p x 'above)
920 (interval-bounded-p y 'below))
921 ;; Intervals are bounded in the appropriate way. Make sure they
923 (let ((left (interval-high x))
924 (right (interval-low y)))
925 (cond ((> (type-bound-number left)
926 (type-bound-number right))
927 ;; The intervals definitely overlap, so result is NIL.
929 ((< (type-bound-number left)
930 (type-bound-number right))
931 ;; The intervals definitely don't touch, so result is T.
934 ;; Limits are equal. Check for open or closed bounds.
935 ;; Don't overlap if one or the other are open.
936 (or (consp left) (consp right)))))))
938 ;;; Return T if X >= Y. That is, every number in the interval X is
939 ;;; always greater than any number in the interval Y.
940 (defun interval->= (x y)
941 (declare (type interval x y))
942 ;; X >= Y if lower bound of X >= upper bound of Y
943 (when (and (interval-bounded-p x 'below)
944 (interval-bounded-p y 'above))
945 (>= (type-bound-number (interval-low x))
946 (type-bound-number (interval-high y)))))
948 ;;; Return T if X = Y.
949 (defun interval-= (x y)
950 (declare (type interval x y))
951 (and (interval-bounded-p x 'both)
952 (interval-bounded-p y 'both)
956 ;; Open intervals cannot be =
957 (return-from interval-= nil))))
958 ;; Both intervals refer to the same point
959 (= (bound (interval-high x)) (bound (interval-low x))
960 (bound (interval-high y)) (bound (interval-low y))))))
962 ;;; Return T if X /= Y
963 (defun interval-/= (x y)
964 (not (interval-intersect-p x y)))
966 ;;; Return an interval that is the absolute value of X. Thus, if
967 ;;; X = [-1 10], the result is [0, 10].
968 (defun interval-abs (x)
969 (declare (type interval x))
970 (case (interval-range-info x)
976 (destructuring-bind (x- x+) (interval-split 0 x t t)
977 (interval-merge-pair (interval-neg x-) x+)))))
979 ;;; Compute the square of an interval.
980 (defun interval-sqr (x)
981 (declare (type interval x))
982 (interval-func (lambda (x) (* x x)) (interval-abs x)))
984 ;;;; numeric DERIVE-TYPE methods
986 ;;; a utility for defining derive-type methods of integer operations. If
987 ;;; the types of both X and Y are integer types, then we compute a new
988 ;;; integer type with bounds determined by FUN when applied to X and Y.
989 ;;; Otherwise, we use NUMERIC-CONTAGION.
990 (defun derive-integer-type-aux (x y fun)
991 (declare (type function fun))
992 (if (and (numeric-type-p x) (numeric-type-p y)
993 (eq (numeric-type-class x) 'integer)
994 (eq (numeric-type-class y) 'integer)
995 (eq (numeric-type-complexp x) :real)
996 (eq (numeric-type-complexp y) :real))
997 (multiple-value-bind (low high) (funcall fun x y)
998 (make-numeric-type :class 'integer
1002 (numeric-contagion x y)))
1004 (defun derive-integer-type (x y fun)
1005 (declare (type lvar x y) (type function fun))
1006 (let ((x (lvar-type x))
1008 (derive-integer-type-aux x y fun)))
1010 ;;; simple utility to flatten a list
1011 (defun flatten-list (x)
1012 (labels ((flatten-and-append (tree list)
1013 (cond ((null tree) list)
1014 ((atom tree) (cons tree list))
1015 (t (flatten-and-append
1016 (car tree) (flatten-and-append (cdr tree) list))))))
1017 (flatten-and-append x nil)))
1019 ;;; Take some type of lvar and massage it so that we get a list of the
1020 ;;; constituent types. If ARG is *EMPTY-TYPE*, return NIL to indicate
1022 (defun prepare-arg-for-derive-type (arg)
1023 (flet ((listify (arg)
1028 (union-type-types arg))
1031 (unless (eq arg *empty-type*)
1032 ;; Make sure all args are some type of numeric-type. For member
1033 ;; types, convert the list of members into a union of equivalent
1034 ;; single-element member-type's.
1035 (let ((new-args nil))
1036 (dolist (arg (listify arg))
1037 (if (member-type-p arg)
1038 ;; Run down the list of members and convert to a list of
1040 (mapc-member-type-members
1042 (push (if (numberp member)
1043 (make-member-type :members (list member))
1047 (push arg new-args)))
1048 (unless (member *empty-type* new-args)
1051 ;;; Convert from the standard type convention for which -0.0 and 0.0
1052 ;;; are equal to an intermediate convention for which they are
1053 ;;; considered different which is more natural for some of the
1055 (defun convert-numeric-type (type)
1056 (declare (type numeric-type type))
1057 ;;; Only convert real float interval delimiters types.
1058 (if (eq (numeric-type-complexp type) :real)
1059 (let* ((lo (numeric-type-low type))
1060 (lo-val (type-bound-number lo))
1061 (lo-float-zero-p (and lo (floatp lo-val) (= lo-val 0.0)))
1062 (hi (numeric-type-high type))
1063 (hi-val (type-bound-number hi))
1064 (hi-float-zero-p (and hi (floatp hi-val) (= hi-val 0.0))))
1065 (if (or lo-float-zero-p hi-float-zero-p)
1067 :class (numeric-type-class type)
1068 :format (numeric-type-format type)
1070 :low (if lo-float-zero-p
1072 (list (float 0.0 lo-val))
1073 (float (load-time-value (make-unportable-float :single-float-negative-zero)) lo-val))
1075 :high (if hi-float-zero-p
1077 (list (float (load-time-value (make-unportable-float :single-float-negative-zero)) hi-val))
1084 ;;; Convert back from the intermediate convention for which -0.0 and
1085 ;;; 0.0 are considered different to the standard type convention for
1086 ;;; which and equal.
1087 (defun convert-back-numeric-type (type)
1088 (declare (type numeric-type type))
1089 ;;; Only convert real float interval delimiters types.
1090 (if (eq (numeric-type-complexp type) :real)
1091 (let* ((lo (numeric-type-low type))
1092 (lo-val (type-bound-number lo))
1094 (and lo (floatp lo-val) (= lo-val 0.0)
1095 (float-sign lo-val)))
1096 (hi (numeric-type-high type))
1097 (hi-val (type-bound-number hi))
1099 (and hi (floatp hi-val) (= hi-val 0.0)
1100 (float-sign hi-val))))
1102 ;; (float +0.0 +0.0) => (member 0.0)
1103 ;; (float -0.0 -0.0) => (member -0.0)
1104 ((and lo-float-zero-p hi-float-zero-p)
1105 ;; shouldn't have exclusive bounds here..
1106 (aver (and (not (consp lo)) (not (consp hi))))
1107 (if (= lo-float-zero-p hi-float-zero-p)
1108 ;; (float +0.0 +0.0) => (member 0.0)
1109 ;; (float -0.0 -0.0) => (member -0.0)
1110 (specifier-type `(member ,lo-val))
1111 ;; (float -0.0 +0.0) => (float 0.0 0.0)
1112 ;; (float +0.0 -0.0) => (float 0.0 0.0)
1113 (make-numeric-type :class (numeric-type-class type)
1114 :format (numeric-type-format type)
1120 ;; (float -0.0 x) => (float 0.0 x)
1121 ((and (not (consp lo)) (minusp lo-float-zero-p))
1122 (make-numeric-type :class (numeric-type-class type)
1123 :format (numeric-type-format type)
1125 :low (float 0.0 lo-val)
1127 ;; (float (+0.0) x) => (float (0.0) x)
1128 ((and (consp lo) (plusp lo-float-zero-p))
1129 (make-numeric-type :class (numeric-type-class type)
1130 :format (numeric-type-format type)
1132 :low (list (float 0.0 lo-val))
1135 ;; (float +0.0 x) => (or (member 0.0) (float (0.0) x))
1136 ;; (float (-0.0) x) => (or (member 0.0) (float (0.0) x))
1137 (list (make-member-type :members (list (float 0.0 lo-val)))
1138 (make-numeric-type :class (numeric-type-class type)
1139 :format (numeric-type-format type)
1141 :low (list (float 0.0 lo-val))
1145 ;; (float x +0.0) => (float x 0.0)
1146 ((and (not (consp hi)) (plusp hi-float-zero-p))
1147 (make-numeric-type :class (numeric-type-class type)
1148 :format (numeric-type-format type)
1151 :high (float 0.0 hi-val)))
1152 ;; (float x (-0.0)) => (float x (0.0))
1153 ((and (consp hi) (minusp hi-float-zero-p))
1154 (make-numeric-type :class (numeric-type-class type)
1155 :format (numeric-type-format type)
1158 :high (list (float 0.0 hi-val))))
1160 ;; (float x (+0.0)) => (or (member -0.0) (float x (0.0)))
1161 ;; (float x -0.0) => (or (member -0.0) (float x (0.0)))
1162 (list (make-member-type :members (list (float (load-time-value (make-unportable-float :single-float-negative-zero)) hi-val)))
1163 (make-numeric-type :class (numeric-type-class type)
1164 :format (numeric-type-format type)
1167 :high (list (float 0.0 hi-val)))))))
1173 ;;; Convert back a possible list of numeric types.
1174 (defun convert-back-numeric-type-list (type-list)
1177 (let ((results '()))
1178 (dolist (type type-list)
1179 (if (numeric-type-p type)
1180 (let ((result (convert-back-numeric-type type)))
1182 (setf results (append results result))
1183 (push result results)))
1184 (push type results)))
1187 (convert-back-numeric-type type-list))
1189 (convert-back-numeric-type-list (union-type-types type-list)))
1193 ;;; Take a list of types and return a canonical type specifier,
1194 ;;; combining any MEMBER types together. If both positive and negative
1195 ;;; MEMBER types are present they are converted to a float type.
1196 ;;; XXX This would be far simpler if the type-union methods could handle
1197 ;;; member/number unions.
1199 ;;; If we're about to generate an overly complex union of numeric types, start
1200 ;;; collapse the ranges together.
1202 ;;; FIXME: The MEMBER canonicalization parts of MAKE-DERIVED-UNION-TYPE and
1203 ;;; entire CONVERT-MEMBER-TYPE probably belong in the kernel's type logic,
1204 ;;; invoked always, instead of in the compiler, invoked only during some type
1206 (defvar *derived-numeric-union-complexity-limit* 6)
1208 (defun make-derived-union-type (type-list)
1209 (let ((xset (alloc-xset))
1212 (numeric-type *empty-type*))
1213 (dolist (type type-list)
1214 (cond ((member-type-p type)
1215 (mapc-member-type-members
1217 (if (fp-zero-p member)
1218 (unless (member member fp-zeroes)
1219 (pushnew member fp-zeroes))
1220 (add-to-xset member xset)))
1222 ((numeric-type-p type)
1223 (let ((*approximate-numeric-unions*
1224 (when (and (union-type-p numeric-type)
1225 (nthcdr *derived-numeric-union-complexity-limit*
1226 (union-type-types numeric-type)))
1228 (setf numeric-type (type-union type numeric-type))))
1230 (push type misc-types))))
1231 (if (and (xset-empty-p xset) (not fp-zeroes))
1232 (apply #'type-union numeric-type misc-types)
1233 (apply #'type-union (make-member-type :xset xset :fp-zeroes fp-zeroes)
1234 numeric-type misc-types))))
1236 ;;; Convert a member type with a single member to a numeric type.
1237 (defun convert-member-type (arg)
1238 (let* ((members (member-type-members arg))
1239 (member (first members))
1240 (member-type (type-of member)))
1241 (aver (not (rest members)))
1242 (specifier-type (cond ((typep member 'integer)
1243 `(integer ,member ,member))
1244 ((memq member-type '(short-float single-float
1245 double-float long-float))
1246 `(,member-type ,member ,member))
1250 ;;; This is used in defoptimizers for computing the resulting type of
1253 ;;; Given the lvar ARG, derive the resulting type using the
1254 ;;; DERIVE-FUN. DERIVE-FUN takes exactly one argument which is some
1255 ;;; "atomic" lvar type like numeric-type or member-type (containing
1256 ;;; just one element). It should return the resulting type, which can
1257 ;;; be a list of types.
1259 ;;; For the case of member types, if a MEMBER-FUN is given it is
1260 ;;; called to compute the result otherwise the member type is first
1261 ;;; converted to a numeric type and the DERIVE-FUN is called.
1262 (defun one-arg-derive-type (arg derive-fun member-fun
1263 &optional (convert-type t))
1264 (declare (type function derive-fun)
1265 (type (or null function) member-fun))
1266 (let ((arg-list (prepare-arg-for-derive-type (lvar-type arg))))
1272 (with-float-traps-masked
1273 (:underflow :overflow :divide-by-zero)
1275 `(eql ,(funcall member-fun
1276 (first (member-type-members x))))))
1277 ;; Otherwise convert to a numeric type.
1278 (let ((result-type-list
1279 (funcall derive-fun (convert-member-type x))))
1281 (convert-back-numeric-type-list result-type-list)
1282 result-type-list))))
1285 (convert-back-numeric-type-list
1286 (funcall derive-fun (convert-numeric-type x)))
1287 (funcall derive-fun x)))
1289 *universal-type*))))
1290 ;; Run down the list of args and derive the type of each one,
1291 ;; saving all of the results in a list.
1292 (let ((results nil))
1293 (dolist (arg arg-list)
1294 (let ((result (deriver arg)))
1296 (setf results (append results result))
1297 (push result results))))
1299 (make-derived-union-type results)
1300 (first results)))))))
1302 ;;; Same as ONE-ARG-DERIVE-TYPE, except we assume the function takes
1303 ;;; two arguments. DERIVE-FUN takes 3 args in this case: the two
1304 ;;; original args and a third which is T to indicate if the two args
1305 ;;; really represent the same lvar. This is useful for deriving the
1306 ;;; type of things like (* x x), which should always be positive. If
1307 ;;; we didn't do this, we wouldn't be able to tell.
1308 (defun two-arg-derive-type (arg1 arg2 derive-fun fun
1309 &optional (convert-type t))
1310 (declare (type function derive-fun fun))
1311 (flet ((deriver (x y same-arg)
1312 (cond ((and (member-type-p x) (member-type-p y))
1313 (let* ((x (first (member-type-members x)))
1314 (y (first (member-type-members y)))
1315 (result (ignore-errors
1316 (with-float-traps-masked
1317 (:underflow :overflow :divide-by-zero
1319 (funcall fun x y)))))
1320 (cond ((null result) *empty-type*)
1321 ((and (floatp result) (float-nan-p result))
1322 (make-numeric-type :class 'float
1323 :format (type-of result)
1326 (specifier-type `(eql ,result))))))
1327 ((and (member-type-p x) (numeric-type-p y))
1328 (let* ((x (convert-member-type x))
1329 (y (if convert-type (convert-numeric-type y) y))
1330 (result (funcall derive-fun x y same-arg)))
1332 (convert-back-numeric-type-list result)
1334 ((and (numeric-type-p x) (member-type-p y))
1335 (let* ((x (if convert-type (convert-numeric-type x) x))
1336 (y (convert-member-type y))
1337 (result (funcall derive-fun x y same-arg)))
1339 (convert-back-numeric-type-list result)
1341 ((and (numeric-type-p x) (numeric-type-p y))
1342 (let* ((x (if convert-type (convert-numeric-type x) x))
1343 (y (if convert-type (convert-numeric-type y) y))
1344 (result (funcall derive-fun x y same-arg)))
1346 (convert-back-numeric-type-list result)
1349 *universal-type*))))
1350 (let ((same-arg (same-leaf-ref-p arg1 arg2))
1351 (a1 (prepare-arg-for-derive-type (lvar-type arg1)))
1352 (a2 (prepare-arg-for-derive-type (lvar-type arg2))))
1354 (let ((results nil))
1356 ;; Since the args are the same LVARs, just run down the
1359 (let ((result (deriver x x same-arg)))
1361 (setf results (append results result))
1362 (push result results))))
1363 ;; Try all pairwise combinations.
1366 (let ((result (or (deriver x y same-arg)
1367 (numeric-contagion x y))))
1369 (setf results (append results result))
1370 (push result results))))))
1372 (make-derived-union-type results)
1373 (first results)))))))
1375 #+sb-xc-host ; (See CROSS-FLOAT-INFINITY-KLUDGE.)
1377 (defoptimizer (+ derive-type) ((x y))
1378 (derive-integer-type
1385 (values (frob (numeric-type-low x) (numeric-type-low y))
1386 (frob (numeric-type-high x) (numeric-type-high y)))))))
1388 (defoptimizer (- derive-type) ((x y))
1389 (derive-integer-type
1396 (values (frob (numeric-type-low x) (numeric-type-high y))
1397 (frob (numeric-type-high x) (numeric-type-low y)))))))
1399 (defoptimizer (* derive-type) ((x y))
1400 (derive-integer-type
1403 (let ((x-low (numeric-type-low x))
1404 (x-high (numeric-type-high x))
1405 (y-low (numeric-type-low y))
1406 (y-high (numeric-type-high y)))
1407 (cond ((not (and x-low y-low))
1409 ((or (minusp x-low) (minusp y-low))
1410 (if (and x-high y-high)
1411 (let ((max (* (max (abs x-low) (abs x-high))
1412 (max (abs y-low) (abs y-high)))))
1413 (values (- max) max))
1416 (values (* x-low y-low)
1417 (if (and x-high y-high)
1421 (defoptimizer (/ derive-type) ((x y))
1422 (numeric-contagion (lvar-type x) (lvar-type y)))
1426 #-sb-xc-host ; (See CROSS-FLOAT-INFINITY-KLUDGE.)
1428 (defun +-derive-type-aux (x y same-arg)
1429 (if (and (numeric-type-real-p x)
1430 (numeric-type-real-p y))
1433 (let ((x-int (numeric-type->interval x)))
1434 (interval-add x-int x-int))
1435 (interval-add (numeric-type->interval x)
1436 (numeric-type->interval y))))
1437 (result-type (numeric-contagion x y)))
1438 ;; If the result type is a float, we need to be sure to coerce
1439 ;; the bounds into the correct type.
1440 (when (eq (numeric-type-class result-type) 'float)
1441 (setf result (interval-func
1443 (coerce-for-bound x (or (numeric-type-format result-type)
1447 :class (if (and (eq (numeric-type-class x) 'integer)
1448 (eq (numeric-type-class y) 'integer))
1449 ;; The sum of integers is always an integer.
1451 (numeric-type-class result-type))
1452 :format (numeric-type-format result-type)
1453 :low (interval-low result)
1454 :high (interval-high result)))
1455 ;; general contagion
1456 (numeric-contagion x y)))
1458 (defoptimizer (+ derive-type) ((x y))
1459 (two-arg-derive-type x y #'+-derive-type-aux #'+))
1461 (defun --derive-type-aux (x y same-arg)
1462 (if (and (numeric-type-real-p x)
1463 (numeric-type-real-p y))
1465 ;; (- X X) is always 0.
1467 (make-interval :low 0 :high 0)
1468 (interval-sub (numeric-type->interval x)
1469 (numeric-type->interval y))))
1470 (result-type (numeric-contagion x y)))
1471 ;; If the result type is a float, we need to be sure to coerce
1472 ;; the bounds into the correct type.
1473 (when (eq (numeric-type-class result-type) 'float)
1474 (setf result (interval-func
1476 (coerce-for-bound x (or (numeric-type-format result-type)
1480 :class (if (and (eq (numeric-type-class x) 'integer)
1481 (eq (numeric-type-class y) 'integer))
1482 ;; The difference of integers is always an integer.
1484 (numeric-type-class result-type))
1485 :format (numeric-type-format result-type)
1486 :low (interval-low result)
1487 :high (interval-high result)))
1488 ;; general contagion
1489 (numeric-contagion x y)))
1491 (defoptimizer (- derive-type) ((x y))
1492 (two-arg-derive-type x y #'--derive-type-aux #'-))
1494 (defun *-derive-type-aux (x y same-arg)
1495 (if (and (numeric-type-real-p x)
1496 (numeric-type-real-p y))
1498 ;; (* X X) is always positive, so take care to do it right.
1500 (interval-sqr (numeric-type->interval x))
1501 (interval-mul (numeric-type->interval x)
1502 (numeric-type->interval y))))
1503 (result-type (numeric-contagion x y)))
1504 ;; If the result type is a float, we need to be sure to coerce
1505 ;; the bounds into the correct type.
1506 (when (eq (numeric-type-class result-type) 'float)
1507 (setf result (interval-func
1509 (coerce-for-bound x (or (numeric-type-format result-type)
1513 :class (if (and (eq (numeric-type-class x) 'integer)
1514 (eq (numeric-type-class y) 'integer))
1515 ;; The product of integers is always an integer.
1517 (numeric-type-class result-type))
1518 :format (numeric-type-format result-type)
1519 :low (interval-low result)
1520 :high (interval-high result)))
1521 (numeric-contagion x y)))
1523 (defoptimizer (* derive-type) ((x y))
1524 (two-arg-derive-type x y #'*-derive-type-aux #'*))
1526 (defun /-derive-type-aux (x y same-arg)
1527 (if (and (numeric-type-real-p x)
1528 (numeric-type-real-p y))
1530 ;; (/ X X) is always 1, except if X can contain 0. In
1531 ;; that case, we shouldn't optimize the division away
1532 ;; because we want 0/0 to signal an error.
1534 (not (interval-contains-p
1535 0 (interval-closure (numeric-type->interval y)))))
1536 (make-interval :low 1 :high 1)
1537 (interval-div (numeric-type->interval x)
1538 (numeric-type->interval y))))
1539 (result-type (numeric-contagion x y)))
1540 ;; If the result type is a float, we need to be sure to coerce
1541 ;; the bounds into the correct type.
1542 (when (eq (numeric-type-class result-type) 'float)
1543 (setf result (interval-func
1545 (coerce-for-bound x (or (numeric-type-format result-type)
1548 (make-numeric-type :class (numeric-type-class result-type)
1549 :format (numeric-type-format result-type)
1550 :low (interval-low result)
1551 :high (interval-high result)))
1552 (numeric-contagion x y)))
1554 (defoptimizer (/ derive-type) ((x y))
1555 (two-arg-derive-type x y #'/-derive-type-aux #'/))
1559 (defun ash-derive-type-aux (n-type shift same-arg)
1560 (declare (ignore same-arg))
1561 ;; KLUDGE: All this ASH optimization is suppressed under CMU CL for
1562 ;; some bignum cases because as of version 2.4.6 for Debian and 18d,
1563 ;; CMU CL blows up on (ASH 1000000000 -100000000000) (i.e. ASH of
1564 ;; two bignums yielding zero) and it's hard to avoid that
1565 ;; calculation in here.
1566 #+(and cmu sb-xc-host)
1567 (when (and (or (typep (numeric-type-low n-type) 'bignum)
1568 (typep (numeric-type-high n-type) 'bignum))
1569 (or (typep (numeric-type-low shift) 'bignum)
1570 (typep (numeric-type-high shift) 'bignum)))
1571 (return-from ash-derive-type-aux *universal-type*))
1572 (flet ((ash-outer (n s)
1573 (when (and (fixnump s)
1575 (> s sb!xc:most-negative-fixnum))
1577 ;; KLUDGE: The bare 64's here should be related to
1578 ;; symbolic machine word size values somehow.
1581 (if (and (fixnump s)
1582 (> s sb!xc:most-negative-fixnum))
1584 (if (minusp n) -1 0))))
1585 (or (and (csubtypep n-type (specifier-type 'integer))
1586 (csubtypep shift (specifier-type 'integer))
1587 (let ((n-low (numeric-type-low n-type))
1588 (n-high (numeric-type-high n-type))
1589 (s-low (numeric-type-low shift))
1590 (s-high (numeric-type-high shift)))
1591 (make-numeric-type :class 'integer :complexp :real
1594 (ash-outer n-low s-high)
1595 (ash-inner n-low s-low)))
1598 (ash-inner n-high s-low)
1599 (ash-outer n-high s-high))))))
1602 (defoptimizer (ash derive-type) ((n shift))
1603 (two-arg-derive-type n shift #'ash-derive-type-aux #'ash))
1605 #+sb-xc-host ; (See CROSS-FLOAT-INFINITY-KLUDGE.)
1606 (macrolet ((frob (fun)
1607 `#'(lambda (type type2)
1608 (declare (ignore type2))
1609 (let ((lo (numeric-type-low type))
1610 (hi (numeric-type-high type)))
1611 (values (if hi (,fun hi) nil) (if lo (,fun lo) nil))))))
1613 (defoptimizer (%negate derive-type) ((num))
1614 (derive-integer-type num num (frob -))))
1616 (defun lognot-derive-type-aux (int)
1617 (derive-integer-type-aux int int
1618 (lambda (type type2)
1619 (declare (ignore type2))
1620 (let ((lo (numeric-type-low type))
1621 (hi (numeric-type-high type)))
1622 (values (if hi (lognot hi) nil)
1623 (if lo (lognot lo) nil)
1624 (numeric-type-class type)
1625 (numeric-type-format type))))))
1627 (defoptimizer (lognot derive-type) ((int))
1628 (lognot-derive-type-aux (lvar-type int)))
1630 #-sb-xc-host ; (See CROSS-FLOAT-INFINITY-KLUDGE.)
1631 (defoptimizer (%negate derive-type) ((num))
1632 (flet ((negate-bound (b)
1634 (set-bound (- (type-bound-number b))
1636 (one-arg-derive-type num
1638 (modified-numeric-type
1640 :low (negate-bound (numeric-type-high type))
1641 :high (negate-bound (numeric-type-low type))))
1644 #+sb-xc-host ; (See CROSS-FLOAT-INFINITY-KLUDGE.)
1645 (defoptimizer (abs derive-type) ((num))
1646 (let ((type (lvar-type num)))
1647 (if (and (numeric-type-p type)
1648 (eq (numeric-type-class type) 'integer)
1649 (eq (numeric-type-complexp type) :real))
1650 (let ((lo (numeric-type-low type))
1651 (hi (numeric-type-high type)))
1652 (make-numeric-type :class 'integer :complexp :real
1653 :low (cond ((and hi (minusp hi))
1659 :high (if (and hi lo)
1660 (max (abs hi) (abs lo))
1662 (numeric-contagion type type))))
1664 #-sb-xc-host ; (See CROSS-FLOAT-INFINITY-KLUDGE.)
1665 (defun abs-derive-type-aux (type)
1666 (cond ((eq (numeric-type-complexp type) :complex)
1667 ;; The absolute value of a complex number is always a
1668 ;; non-negative float.
1669 (let* ((format (case (numeric-type-class type)
1670 ((integer rational) 'single-float)
1671 (t (numeric-type-format type))))
1672 (bound-format (or format 'float)))
1673 (make-numeric-type :class 'float
1676 :low (coerce 0 bound-format)
1679 ;; The absolute value of a real number is a non-negative real
1680 ;; of the same type.
1681 (let* ((abs-bnd (interval-abs (numeric-type->interval type)))
1682 (class (numeric-type-class type))
1683 (format (numeric-type-format type))
1684 (bound-type (or format class 'real)))
1689 :low (coerce-and-truncate-floats (interval-low abs-bnd) bound-type)
1690 :high (coerce-and-truncate-floats
1691 (interval-high abs-bnd) bound-type))))))
1693 #-sb-xc-host ; (See CROSS-FLOAT-INFINITY-KLUDGE.)
1694 (defoptimizer (abs derive-type) ((num))
1695 (one-arg-derive-type num #'abs-derive-type-aux #'abs))
1697 #+sb-xc-host ; (See CROSS-FLOAT-INFINITY-KLUDGE.)
1698 (defoptimizer (truncate derive-type) ((number divisor))
1699 (let ((number-type (lvar-type number))
1700 (divisor-type (lvar-type divisor))
1701 (integer-type (specifier-type 'integer)))
1702 (if (and (numeric-type-p number-type)
1703 (csubtypep number-type integer-type)
1704 (numeric-type-p divisor-type)
1705 (csubtypep divisor-type integer-type))
1706 (let ((number-low (numeric-type-low number-type))
1707 (number-high (numeric-type-high number-type))
1708 (divisor-low (numeric-type-low divisor-type))
1709 (divisor-high (numeric-type-high divisor-type)))
1710 (values-specifier-type
1711 `(values ,(integer-truncate-derive-type number-low number-high
1712 divisor-low divisor-high)
1713 ,(integer-rem-derive-type number-low number-high
1714 divisor-low divisor-high))))
1717 #-sb-xc-host ; (See CROSS-FLOAT-INFINITY-KLUDGE.)
1720 (defun rem-result-type (number-type divisor-type)
1721 ;; Figure out what the remainder type is. The remainder is an
1722 ;; integer if both args are integers; a rational if both args are
1723 ;; rational; and a float otherwise.
1724 (cond ((and (csubtypep number-type (specifier-type 'integer))
1725 (csubtypep divisor-type (specifier-type 'integer)))
1727 ((and (csubtypep number-type (specifier-type 'rational))
1728 (csubtypep divisor-type (specifier-type 'rational)))
1730 ((and (csubtypep number-type (specifier-type 'float))
1731 (csubtypep divisor-type (specifier-type 'float)))
1732 ;; Both are floats so the result is also a float, of
1733 ;; the largest type.
1734 (or (float-format-max (numeric-type-format number-type)
1735 (numeric-type-format divisor-type))
1737 ((and (csubtypep number-type (specifier-type 'float))
1738 (csubtypep divisor-type (specifier-type 'rational)))
1739 ;; One of the arguments is a float and the other is a
1740 ;; rational. The remainder is a float of the same
1742 (or (numeric-type-format number-type) 'float))
1743 ((and (csubtypep divisor-type (specifier-type 'float))
1744 (csubtypep number-type (specifier-type 'rational)))
1745 ;; One of the arguments is a float and the other is a
1746 ;; rational. The remainder is a float of the same
1748 (or (numeric-type-format divisor-type) 'float))
1750 ;; Some unhandled combination. This usually means both args
1751 ;; are REAL so the result is a REAL.
1754 (defun truncate-derive-type-quot (number-type divisor-type)
1755 (let* ((rem-type (rem-result-type number-type divisor-type))
1756 (number-interval (numeric-type->interval number-type))
1757 (divisor-interval (numeric-type->interval divisor-type)))
1758 ;;(declare (type (member '(integer rational float)) rem-type))
1759 ;; We have real numbers now.
1760 (cond ((eq rem-type 'integer)
1761 ;; Since the remainder type is INTEGER, both args are
1763 (let* ((res (integer-truncate-derive-type
1764 (interval-low number-interval)
1765 (interval-high number-interval)
1766 (interval-low divisor-interval)
1767 (interval-high divisor-interval))))
1768 (specifier-type (if (listp res) res 'integer))))
1770 (let ((quot (truncate-quotient-bound
1771 (interval-div number-interval
1772 divisor-interval))))
1773 (specifier-type `(integer ,(or (interval-low quot) '*)
1774 ,(or (interval-high quot) '*))))))))
1776 (defun truncate-derive-type-rem (number-type divisor-type)
1777 (let* ((rem-type (rem-result-type number-type divisor-type))
1778 (number-interval (numeric-type->interval number-type))
1779 (divisor-interval (numeric-type->interval divisor-type))
1780 (rem (truncate-rem-bound number-interval divisor-interval)))
1781 ;;(declare (type (member '(integer rational float)) rem-type))
1782 ;; We have real numbers now.
1783 (cond ((eq rem-type 'integer)
1784 ;; Since the remainder type is INTEGER, both args are
1786 (specifier-type `(,rem-type ,(or (interval-low rem) '*)
1787 ,(or (interval-high rem) '*))))
1789 (multiple-value-bind (class format)
1792 (values 'integer nil))
1794 (values 'rational nil))
1795 ((or single-float double-float #!+long-float long-float)
1796 (values 'float rem-type))
1798 (values 'float nil))
1801 (when (member rem-type '(float single-float double-float
1802 #!+long-float long-float))
1803 (setf rem (interval-func #'(lambda (x)
1804 (coerce-for-bound x rem-type))
1806 (make-numeric-type :class class
1808 :low (interval-low rem)
1809 :high (interval-high rem)))))))
1811 (defun truncate-derive-type-quot-aux (num div same-arg)
1812 (declare (ignore same-arg))
1813 (if (and (numeric-type-real-p num)
1814 (numeric-type-real-p div))
1815 (truncate-derive-type-quot num div)
1818 (defun truncate-derive-type-rem-aux (num div same-arg)
1819 (declare (ignore same-arg))
1820 (if (and (numeric-type-real-p num)
1821 (numeric-type-real-p div))
1822 (truncate-derive-type-rem num div)
1825 (defoptimizer (truncate derive-type) ((number divisor))
1826 (let ((quot (two-arg-derive-type number divisor
1827 #'truncate-derive-type-quot-aux #'truncate))
1828 (rem (two-arg-derive-type number divisor
1829 #'truncate-derive-type-rem-aux #'rem)))
1830 (when (and quot rem)
1831 (make-values-type :required (list quot rem)))))
1833 (defun ftruncate-derive-type-quot (number-type divisor-type)
1834 ;; The bounds are the same as for truncate. However, the first
1835 ;; result is a float of some type. We need to determine what that
1836 ;; type is. Basically it's the more contagious of the two types.
1837 (let ((q-type (truncate-derive-type-quot number-type divisor-type))
1838 (res-type (numeric-contagion number-type divisor-type)))
1839 (make-numeric-type :class 'float
1840 :format (numeric-type-format res-type)
1841 :low (numeric-type-low q-type)
1842 :high (numeric-type-high q-type))))
1844 (defun ftruncate-derive-type-quot-aux (n d same-arg)
1845 (declare (ignore same-arg))
1846 (if (and (numeric-type-real-p n)
1847 (numeric-type-real-p d))
1848 (ftruncate-derive-type-quot n d)
1851 (defoptimizer (ftruncate derive-type) ((number divisor))
1853 (two-arg-derive-type number divisor
1854 #'ftruncate-derive-type-quot-aux #'ftruncate))
1855 (rem (two-arg-derive-type number divisor
1856 #'truncate-derive-type-rem-aux #'rem)))
1857 (when (and quot rem)
1858 (make-values-type :required (list quot rem)))))
1860 (defun %unary-truncate-derive-type-aux (number)
1861 (truncate-derive-type-quot number (specifier-type '(integer 1 1))))
1863 (defoptimizer (%unary-truncate derive-type) ((number))
1864 (one-arg-derive-type number
1865 #'%unary-truncate-derive-type-aux
1868 (defoptimizer (%unary-truncate/single-float derive-type) ((number))
1869 (one-arg-derive-type number
1870 #'%unary-truncate-derive-type-aux
1873 (defoptimizer (%unary-truncate/double-float derive-type) ((number))
1874 (one-arg-derive-type number
1875 #'%unary-truncate-derive-type-aux
1878 (defoptimizer (%unary-ftruncate derive-type) ((number))
1879 (let ((divisor (specifier-type '(integer 1 1))))
1880 (one-arg-derive-type number
1882 (ftruncate-derive-type-quot-aux n divisor nil))
1883 #'%unary-ftruncate)))
1885 (defoptimizer (%unary-round derive-type) ((number))
1886 (one-arg-derive-type number
1889 (unless (numeric-type-real-p n)
1890 (return *empty-type*))
1891 (let* ((interval (numeric-type->interval n))
1892 (low (interval-low interval))
1893 (high (interval-high interval)))
1895 (setf low (car low)))
1897 (setf high (car high)))
1907 ;;; Define optimizers for FLOOR and CEILING.
1909 ((def (name q-name r-name)
1910 (let ((q-aux (symbolicate q-name "-AUX"))
1911 (r-aux (symbolicate r-name "-AUX")))
1913 ;; Compute type of quotient (first) result.
1914 (defun ,q-aux (number-type divisor-type)
1915 (let* ((number-interval
1916 (numeric-type->interval number-type))
1918 (numeric-type->interval divisor-type))
1919 (quot (,q-name (interval-div number-interval
1920 divisor-interval))))
1921 (specifier-type `(integer ,(or (interval-low quot) '*)
1922 ,(or (interval-high quot) '*)))))
1923 ;; Compute type of remainder.
1924 (defun ,r-aux (number-type divisor-type)
1925 (let* ((divisor-interval
1926 (numeric-type->interval divisor-type))
1927 (rem (,r-name divisor-interval))
1928 (result-type (rem-result-type number-type divisor-type)))
1929 (multiple-value-bind (class format)
1932 (values 'integer nil))
1934 (values 'rational nil))
1935 ((or single-float double-float #!+long-float long-float)
1936 (values 'float result-type))
1938 (values 'float nil))
1941 (when (member result-type '(float single-float double-float
1942 #!+long-float long-float))
1943 ;; Make sure that the limits on the interval have
1945 (setf rem (interval-func (lambda (x)
1946 (coerce-for-bound x result-type))
1948 (make-numeric-type :class class
1950 :low (interval-low rem)
1951 :high (interval-high rem)))))
1952 ;; the optimizer itself
1953 (defoptimizer (,name derive-type) ((number divisor))
1954 (flet ((derive-q (n d same-arg)
1955 (declare (ignore same-arg))
1956 (if (and (numeric-type-real-p n)
1957 (numeric-type-real-p d))
1960 (derive-r (n d same-arg)
1961 (declare (ignore same-arg))
1962 (if (and (numeric-type-real-p n)
1963 (numeric-type-real-p d))
1966 (let ((quot (two-arg-derive-type
1967 number divisor #'derive-q #',name))
1968 (rem (two-arg-derive-type
1969 number divisor #'derive-r #'mod)))
1970 (when (and quot rem)
1971 (make-values-type :required (list quot rem))))))))))
1973 (def floor floor-quotient-bound floor-rem-bound)
1974 (def ceiling ceiling-quotient-bound ceiling-rem-bound))
1976 ;;; Define optimizers for FFLOOR and FCEILING
1977 (macrolet ((def (name q-name r-name)
1978 (let ((q-aux (symbolicate "F" q-name "-AUX"))
1979 (r-aux (symbolicate r-name "-AUX")))
1981 ;; Compute type of quotient (first) result.
1982 (defun ,q-aux (number-type divisor-type)
1983 (let* ((number-interval
1984 (numeric-type->interval number-type))
1986 (numeric-type->interval divisor-type))
1987 (quot (,q-name (interval-div number-interval
1989 (res-type (numeric-contagion number-type
1992 :class (numeric-type-class res-type)
1993 :format (numeric-type-format res-type)
1994 :low (interval-low quot)
1995 :high (interval-high quot))))
1997 (defoptimizer (,name derive-type) ((number divisor))
1998 (flet ((derive-q (n d same-arg)
1999 (declare (ignore same-arg))
2000 (if (and (numeric-type-real-p n)
2001 (numeric-type-real-p d))
2004 (derive-r (n d same-arg)
2005 (declare (ignore same-arg))
2006 (if (and (numeric-type-real-p n)
2007 (numeric-type-real-p d))
2010 (let ((quot (two-arg-derive-type
2011 number divisor #'derive-q #',name))
2012 (rem (two-arg-derive-type
2013 number divisor #'derive-r #'mod)))
2014 (when (and quot rem)
2015 (make-values-type :required (list quot rem))))))))))
2017 (def ffloor floor-quotient-bound floor-rem-bound)
2018 (def fceiling ceiling-quotient-bound ceiling-rem-bound))
2020 ;;; functions to compute the bounds on the quotient and remainder for
2021 ;;; the FLOOR function
2022 (defun floor-quotient-bound (quot)
2023 ;; Take the floor of the quotient and then massage it into what we
2025 (let ((lo (interval-low quot))
2026 (hi (interval-high quot)))
2027 ;; Take the floor of the lower bound. The result is always a
2028 ;; closed lower bound.
2030 (floor (type-bound-number lo))
2032 ;; For the upper bound, we need to be careful.
2035 ;; An open bound. We need to be careful here because
2036 ;; the floor of '(10.0) is 9, but the floor of
2038 (multiple-value-bind (q r) (floor (first hi))
2043 ;; A closed bound, so the answer is obvious.
2047 (make-interval :low lo :high hi)))
2048 (defun floor-rem-bound (div)
2049 ;; The remainder depends only on the divisor. Try to get the
2050 ;; correct sign for the remainder if we can.
2051 (case (interval-range-info div)
2053 ;; The divisor is always positive.
2054 (let ((rem (interval-abs div)))
2055 (setf (interval-low rem) 0)
2056 (when (and (numberp (interval-high rem))
2057 (not (zerop (interval-high rem))))
2058 ;; The remainder never contains the upper bound. However,
2059 ;; watch out for the case where the high limit is zero!
2060 (setf (interval-high rem) (list (interval-high rem))))
2063 ;; The divisor is always negative.
2064 (let ((rem (interval-neg (interval-abs div))))
2065 (setf (interval-high rem) 0)
2066 (when (numberp (interval-low rem))
2067 ;; The remainder never contains the lower bound.
2068 (setf (interval-low rem) (list (interval-low rem))))
2071 ;; The divisor can be positive or negative. All bets off. The
2072 ;; magnitude of remainder is the maximum value of the divisor.
2073 (let ((limit (type-bound-number (interval-high (interval-abs div)))))
2074 ;; The bound never reaches the limit, so make the interval open.
2075 (make-interval :low (if limit
2078 :high (list limit))))))
2080 (floor-quotient-bound (make-interval :low 0.3 :high 10.3))
2081 => #S(INTERVAL :LOW 0 :HIGH 10)
2082 (floor-quotient-bound (make-interval :low 0.3 :high '(10.3)))
2083 => #S(INTERVAL :LOW 0 :HIGH 10)
2084 (floor-quotient-bound (make-interval :low 0.3 :high 10))
2085 => #S(INTERVAL :LOW 0 :HIGH 10)
2086 (floor-quotient-bound (make-interval :low 0.3 :high '(10)))
2087 => #S(INTERVAL :LOW 0 :HIGH 9)
2088 (floor-quotient-bound (make-interval :low '(0.3) :high 10.3))
2089 => #S(INTERVAL :LOW 0 :HIGH 10)
2090 (floor-quotient-bound (make-interval :low '(0.0) :high 10.3))
2091 => #S(INTERVAL :LOW 0 :HIGH 10)
2092 (floor-quotient-bound (make-interval :low '(-1.3) :high 10.3))
2093 => #S(INTERVAL :LOW -2 :HIGH 10)
2094 (floor-quotient-bound (make-interval :low '(-1.0) :high 10.3))
2095 => #S(INTERVAL :LOW -1 :HIGH 10)
2096 (floor-quotient-bound (make-interval :low -1.0 :high 10.3))
2097 => #S(INTERVAL :LOW -1 :HIGH 10)
2099 (floor-rem-bound (make-interval :low 0.3 :high 10.3))
2100 => #S(INTERVAL :LOW 0 :HIGH '(10.3))
2101 (floor-rem-bound (make-interval :low 0.3 :high '(10.3)))
2102 => #S(INTERVAL :LOW 0 :HIGH '(10.3))
2103 (floor-rem-bound (make-interval :low -10 :high -2.3))
2104 #S(INTERVAL :LOW (-10) :HIGH 0)
2105 (floor-rem-bound (make-interval :low 0.3 :high 10))
2106 => #S(INTERVAL :LOW 0 :HIGH '(10))
2107 (floor-rem-bound (make-interval :low '(-1.3) :high 10.3))
2108 => #S(INTERVAL :LOW '(-10.3) :HIGH '(10.3))
2109 (floor-rem-bound (make-interval :low '(-20.3) :high 10.3))
2110 => #S(INTERVAL :LOW (-20.3) :HIGH (20.3))
2113 ;;; same functions for CEILING
2114 (defun ceiling-quotient-bound (quot)
2115 ;; Take the ceiling of the quotient and then massage it into what we
2117 (let ((lo (interval-low quot))
2118 (hi (interval-high quot)))
2119 ;; Take the ceiling of the upper bound. The result is always a
2120 ;; closed upper bound.
2122 (ceiling (type-bound-number hi))
2124 ;; For the lower bound, we need to be careful.
2127 ;; An open bound. We need to be careful here because
2128 ;; the ceiling of '(10.0) is 11, but the ceiling of
2130 (multiple-value-bind (q r) (ceiling (first lo))
2135 ;; A closed bound, so the answer is obvious.
2139 (make-interval :low lo :high hi)))
2140 (defun ceiling-rem-bound (div)
2141 ;; The remainder depends only on the divisor. Try to get the
2142 ;; correct sign for the remainder if we can.
2143 (case (interval-range-info div)
2145 ;; Divisor is always positive. The remainder is negative.
2146 (let ((rem (interval-neg (interval-abs div))))
2147 (setf (interval-high rem) 0)
2148 (when (and (numberp (interval-low rem))
2149 (not (zerop (interval-low rem))))
2150 ;; The remainder never contains the upper bound. However,
2151 ;; watch out for the case when the upper bound is zero!
2152 (setf (interval-low rem) (list (interval-low rem))))
2155 ;; Divisor is always negative. The remainder is positive
2156 (let ((rem (interval-abs div)))
2157 (setf (interval-low rem) 0)
2158 (when (numberp (interval-high rem))
2159 ;; The remainder never contains the lower bound.
2160 (setf (interval-high rem) (list (interval-high rem))))
2163 ;; The divisor can be positive or negative. All bets off. The
2164 ;; magnitude of remainder is the maximum value of the divisor.
2165 (let ((limit (type-bound-number (interval-high (interval-abs div)))))
2166 ;; The bound never reaches the limit, so make the interval open.
2167 (make-interval :low (if limit
2170 :high (list limit))))))
2173 (ceiling-quotient-bound (make-interval :low 0.3 :high 10.3))
2174 => #S(INTERVAL :LOW 1 :HIGH 11)
2175 (ceiling-quotient-bound (make-interval :low 0.3 :high '(10.3)))
2176 => #S(INTERVAL :LOW 1 :HIGH 11)
2177 (ceiling-quotient-bound (make-interval :low 0.3 :high 10))
2178 => #S(INTERVAL :LOW 1 :HIGH 10)
2179 (ceiling-quotient-bound (make-interval :low 0.3 :high '(10)))
2180 => #S(INTERVAL :LOW 1 :HIGH 10)
2181 (ceiling-quotient-bound (make-interval :low '(0.3) :high 10.3))
2182 => #S(INTERVAL :LOW 1 :HIGH 11)
2183 (ceiling-quotient-bound (make-interval :low '(0.0) :high 10.3))
2184 => #S(INTERVAL :LOW 1 :HIGH 11)
2185 (ceiling-quotient-bound (make-interval :low '(-1.3) :high 10.3))
2186 => #S(INTERVAL :LOW -1 :HIGH 11)
2187 (ceiling-quotient-bound (make-interval :low '(-1.0) :high 10.3))
2188 => #S(INTERVAL :LOW 0 :HIGH 11)
2189 (ceiling-quotient-bound (make-interval :low -1.0 :high 10.3))
2190 => #S(INTERVAL :LOW -1 :HIGH 11)
2192 (ceiling-rem-bound (make-interval :low 0.3 :high 10.3))
2193 => #S(INTERVAL :LOW (-10.3) :HIGH 0)
2194 (ceiling-rem-bound (make-interval :low 0.3 :high '(10.3)))
2195 => #S(INTERVAL :LOW 0 :HIGH '(10.3))
2196 (ceiling-rem-bound (make-interval :low -10 :high -2.3))
2197 => #S(INTERVAL :LOW 0 :HIGH (10))
2198 (ceiling-rem-bound (make-interval :low 0.3 :high 10))
2199 => #S(INTERVAL :LOW (-10) :HIGH 0)
2200 (ceiling-rem-bound (make-interval :low '(-1.3) :high 10.3))
2201 => #S(INTERVAL :LOW (-10.3) :HIGH (10.3))
2202 (ceiling-rem-bound (make-interval :low '(-20.3) :high 10.3))
2203 => #S(INTERVAL :LOW (-20.3) :HIGH (20.3))
2206 (defun truncate-quotient-bound (quot)
2207 ;; For positive quotients, truncate is exactly like floor. For
2208 ;; negative quotients, truncate is exactly like ceiling. Otherwise,
2209 ;; it's the union of the two pieces.
2210 (case (interval-range-info quot)
2213 (floor-quotient-bound quot))
2215 ;; just like CEILING
2216 (ceiling-quotient-bound quot))
2218 ;; Split the interval into positive and negative pieces, compute
2219 ;; the result for each piece and put them back together.
2220 (destructuring-bind (neg pos) (interval-split 0 quot t t)
2221 (interval-merge-pair (ceiling-quotient-bound neg)
2222 (floor-quotient-bound pos))))))
2224 (defun truncate-rem-bound (num div)
2225 ;; This is significantly more complicated than FLOOR or CEILING. We
2226 ;; need both the number and the divisor to determine the range. The
2227 ;; basic idea is to split the ranges of NUM and DEN into positive
2228 ;; and negative pieces and deal with each of the four possibilities
2230 (case (interval-range-info num)
2232 (case (interval-range-info div)
2234 (floor-rem-bound div))
2236 (ceiling-rem-bound div))
2238 (destructuring-bind (neg pos) (interval-split 0 div t t)
2239 (interval-merge-pair (truncate-rem-bound num neg)
2240 (truncate-rem-bound num pos))))))
2242 (case (interval-range-info div)
2244 (ceiling-rem-bound div))
2246 (floor-rem-bound div))
2248 (destructuring-bind (neg pos) (interval-split 0 div t t)
2249 (interval-merge-pair (truncate-rem-bound num neg)
2250 (truncate-rem-bound num pos))))))
2252 (destructuring-bind (neg pos) (interval-split 0 num t t)
2253 (interval-merge-pair (truncate-rem-bound neg div)
2254 (truncate-rem-bound pos div))))))
2257 ;;; Derive useful information about the range. Returns three values:
2258 ;;; - '+ if its positive, '- negative, or nil if it overlaps 0.
2259 ;;; - The abs of the minimal value (i.e. closest to 0) in the range.
2260 ;;; - The abs of the maximal value if there is one, or nil if it is
2262 (defun numeric-range-info (low high)
2263 (cond ((and low (not (minusp low)))
2264 (values '+ low high))
2265 ((and high (not (plusp high)))
2266 (values '- (- high) (if low (- low) nil)))
2268 (values nil 0 (and low high (max (- low) high))))))
2270 (defun integer-truncate-derive-type
2271 (number-low number-high divisor-low divisor-high)
2272 ;; The result cannot be larger in magnitude than the number, but the
2273 ;; sign might change. If we can determine the sign of either the
2274 ;; number or the divisor, we can eliminate some of the cases.
2275 (multiple-value-bind (number-sign number-min number-max)
2276 (numeric-range-info number-low number-high)
2277 (multiple-value-bind (divisor-sign divisor-min divisor-max)
2278 (numeric-range-info divisor-low divisor-high)
2279 (when (and divisor-max (zerop divisor-max))
2280 ;; We've got a problem: guaranteed division by zero.
2281 (return-from integer-truncate-derive-type t))
2282 (when (zerop divisor-min)
2283 ;; We'll assume that they aren't going to divide by zero.
2285 (cond ((and number-sign divisor-sign)
2286 ;; We know the sign of both.
2287 (if (eq number-sign divisor-sign)
2288 ;; Same sign, so the result will be positive.
2289 `(integer ,(if divisor-max
2290 (truncate number-min divisor-max)
2293 (truncate number-max divisor-min)
2295 ;; Different signs, the result will be negative.
2296 `(integer ,(if number-max
2297 (- (truncate number-max divisor-min))
2300 (- (truncate number-min divisor-max))
2302 ((eq divisor-sign '+)
2303 ;; The divisor is positive. Therefore, the number will just
2304 ;; become closer to zero.
2305 `(integer ,(if number-low
2306 (truncate number-low divisor-min)
2309 (truncate number-high divisor-min)
2311 ((eq divisor-sign '-)
2312 ;; The divisor is negative. Therefore, the absolute value of
2313 ;; the number will become closer to zero, but the sign will also
2315 `(integer ,(if number-high
2316 (- (truncate number-high divisor-min))
2319 (- (truncate number-low divisor-min))
2321 ;; The divisor could be either positive or negative.
2323 ;; The number we are dividing has a bound. Divide that by the
2324 ;; smallest posible divisor.
2325 (let ((bound (truncate number-max divisor-min)))
2326 `(integer ,(- bound) ,bound)))
2328 ;; The number we are dividing is unbounded, so we can't tell
2329 ;; anything about the result.
2332 #+sb-xc-host ; (See CROSS-FLOAT-INFINITY-KLUDGE.)
2333 (defun integer-rem-derive-type
2334 (number-low number-high divisor-low divisor-high)
2335 (if (and divisor-low divisor-high)
2336 ;; We know the range of the divisor, and the remainder must be
2337 ;; smaller than the divisor. We can tell the sign of the
2338 ;; remainder if we know the sign of the number.
2339 (let ((divisor-max (1- (max (abs divisor-low) (abs divisor-high)))))
2340 `(integer ,(if (or (null number-low)
2341 (minusp number-low))
2344 ,(if (or (null number-high)
2345 (plusp number-high))
2348 ;; The divisor is potentially either very positive or very
2349 ;; negative. Therefore, the remainder is unbounded, but we might
2350 ;; be able to tell something about the sign from the number.
2351 `(integer ,(if (and number-low (not (minusp number-low)))
2352 ;; The number we are dividing is positive.
2353 ;; Therefore, the remainder must be positive.
2356 ,(if (and number-high (not (plusp number-high)))
2357 ;; The number we are dividing is negative.
2358 ;; Therefore, the remainder must be negative.
2362 #+sb-xc-host ; (See CROSS-FLOAT-INFINITY-KLUDGE.)
2363 (defoptimizer (random derive-type) ((bound &optional state))
2364 (let ((type (lvar-type bound)))
2365 (when (numeric-type-p type)
2366 (let ((class (numeric-type-class type))
2367 (high (numeric-type-high type))
2368 (format (numeric-type-format type)))
2372 :low (coerce 0 (or format class 'real))
2373 :high (cond ((not high) nil)
2374 ((eq class 'integer) (max (1- high) 0))
2375 ((or (consp high) (zerop high)) high)
2378 #-sb-xc-host ; (See CROSS-FLOAT-INFINITY-KLUDGE.)
2379 (defun random-derive-type-aux (type)
2380 (let ((class (numeric-type-class type))
2381 (high (numeric-type-high type))
2382 (format (numeric-type-format type)))
2386 :low (coerce 0 (or format class 'real))
2387 :high (cond ((not high) nil)
2388 ((eq class 'integer) (max (1- high) 0))
2389 ((or (consp high) (zerop high)) high)
2392 #-sb-xc-host ; (See CROSS-FLOAT-INFINITY-KLUDGE.)
2393 (defoptimizer (random derive-type) ((bound &optional state))
2394 (one-arg-derive-type bound #'random-derive-type-aux nil))
2396 ;;;; DERIVE-TYPE methods for LOGAND, LOGIOR, and friends
2398 ;;; Return the maximum number of bits an integer of the supplied type
2399 ;;; can take up, or NIL if it is unbounded. The second (third) value
2400 ;;; is T if the integer can be positive (negative) and NIL if not.
2401 ;;; Zero counts as positive.
2402 (defun integer-type-length (type)
2403 (if (numeric-type-p type)
2404 (let ((min (numeric-type-low type))
2405 (max (numeric-type-high type)))
2406 (values (and min max (max (integer-length min) (integer-length max)))
2407 (or (null max) (not (minusp max)))
2408 (or (null min) (minusp min))))
2411 ;;; See _Hacker's Delight_, Henry S. Warren, Jr. pp 58-63 for an
2412 ;;; explanation of LOG{AND,IOR,XOR}-DERIVE-UNSIGNED-{LOW,HIGH}-BOUND.
2413 ;;; Credit also goes to Raymond Toy for writing (and debugging!) similar
2414 ;;; versions in CMUCL, from which these functions copy liberally.
2416 (defun logand-derive-unsigned-low-bound (x y)
2417 (let ((a (numeric-type-low x))
2418 (b (numeric-type-high x))
2419 (c (numeric-type-low y))
2420 (d (numeric-type-high y)))
2421 (loop for m = (ash 1 (integer-length (lognor a c))) then (ash m -1)
2423 (unless (zerop (logand m (lognot a) (lognot c)))
2424 (let ((temp (logandc2 (logior a m) (1- m))))
2428 (setf temp (logandc2 (logior c m) (1- m)))
2432 finally (return (logand a c)))))
2434 (defun logand-derive-unsigned-high-bound (x y)
2435 (let ((a (numeric-type-low x))
2436 (b (numeric-type-high x))
2437 (c (numeric-type-low y))
2438 (d (numeric-type-high y)))
2439 (loop for m = (ash 1 (integer-length (logxor b d))) then (ash m -1)
2442 ((not (zerop (logand b (lognot d) m)))
2443 (let ((temp (logior (logandc2 b m) (1- m))))
2447 ((not (zerop (logand (lognot b) d m)))
2448 (let ((temp (logior (logandc2 d m) (1- m))))
2452 finally (return (logand b d)))))
2454 (defun logand-derive-type-aux (x y &optional same-leaf)
2456 (return-from logand-derive-type-aux x))
2457 (multiple-value-bind (x-len x-pos x-neg) (integer-type-length x)
2458 (declare (ignore x-pos))
2459 (multiple-value-bind (y-len y-pos y-neg) (integer-type-length y)
2460 (declare (ignore y-pos))
2462 ;; X must be positive.
2464 ;; They must both be positive.
2465 (cond ((and (null x-len) (null y-len))
2466 (specifier-type 'unsigned-byte))
2468 (specifier-type `(unsigned-byte* ,y-len)))
2470 (specifier-type `(unsigned-byte* ,x-len)))
2472 (let ((low (logand-derive-unsigned-low-bound x y))
2473 (high (logand-derive-unsigned-high-bound x y)))
2474 (specifier-type `(integer ,low ,high)))))
2475 ;; X is positive, but Y might be negative.
2477 (specifier-type 'unsigned-byte))
2479 (specifier-type `(unsigned-byte* ,x-len)))))
2480 ;; X might be negative.
2482 ;; Y must be positive.
2484 (specifier-type 'unsigned-byte))
2485 (t (specifier-type `(unsigned-byte* ,y-len))))
2486 ;; Either might be negative.
2487 (if (and x-len y-len)
2488 ;; The result is bounded.
2489 (specifier-type `(signed-byte ,(1+ (max x-len y-len))))
2490 ;; We can't tell squat about the result.
2491 (specifier-type 'integer)))))))
2493 (defun logior-derive-unsigned-low-bound (x y)
2494 (let ((a (numeric-type-low x))
2495 (b (numeric-type-high x))
2496 (c (numeric-type-low y))
2497 (d (numeric-type-high y)))
2498 (loop for m = (ash 1 (integer-length (logxor a c))) then (ash m -1)
2501 ((not (zerop (logandc2 (logand c m) a)))
2502 (let ((temp (logand (logior a m) (1+ (lognot m)))))
2506 ((not (zerop (logandc2 (logand a m) c)))
2507 (let ((temp (logand (logior c m) (1+ (lognot m)))))
2511 finally (return (logior a c)))))
2513 (defun logior-derive-unsigned-high-bound (x y)
2514 (let ((a (numeric-type-low x))
2515 (b (numeric-type-high x))
2516 (c (numeric-type-low y))
2517 (d (numeric-type-high y)))
2518 (loop for m = (ash 1 (integer-length (logand b d))) then (ash m -1)
2520 (unless (zerop (logand b d m))
2521 (let ((temp (logior (- b m) (1- m))))
2525 (setf temp (logior (- d m) (1- m)))
2529 finally (return (logior b d)))))
2531 (defun logior-derive-type-aux (x y &optional same-leaf)
2533 (return-from logior-derive-type-aux x))
2534 (multiple-value-bind (x-len x-pos x-neg) (integer-type-length x)
2535 (multiple-value-bind (y-len y-pos y-neg) (integer-type-length y)
2537 ((and (not x-neg) (not y-neg))
2538 ;; Both are positive.
2539 (if (and x-len y-len)
2540 (let ((low (logior-derive-unsigned-low-bound x y))
2541 (high (logior-derive-unsigned-high-bound x y)))
2542 (specifier-type `(integer ,low ,high)))
2543 (specifier-type `(unsigned-byte* *))))
2545 ;; X must be negative.
2547 ;; Both are negative. The result is going to be negative
2548 ;; and be the same length or shorter than the smaller.
2549 (if (and x-len y-len)
2551 (specifier-type `(integer ,(ash -1 (min x-len y-len)) -1))
2553 (specifier-type '(integer * -1)))
2554 ;; X is negative, but we don't know about Y. The result
2555 ;; will be negative, but no more negative than X.
2557 `(integer ,(or (numeric-type-low x) '*)
2560 ;; X might be either positive or negative.
2562 ;; But Y is negative. The result will be negative.
2564 `(integer ,(or (numeric-type-low y) '*)
2566 ;; We don't know squat about either. It won't get any bigger.
2567 (if (and x-len y-len)
2569 (specifier-type `(signed-byte ,(1+ (max x-len y-len))))
2571 (specifier-type 'integer))))))))
2573 (defun logxor-derive-unsigned-low-bound (x y)
2574 (let ((a (numeric-type-low x))
2575 (b (numeric-type-high x))
2576 (c (numeric-type-low y))
2577 (d (numeric-type-high y)))
2578 (loop for m = (ash 1 (integer-length (logxor a c))) then (ash m -1)
2581 ((not (zerop (logandc2 (logand c m) a)))
2582 (let ((temp (logand (logior a m)
2586 ((not (zerop (logandc2 (logand a m) c)))
2587 (let ((temp (logand (logior c m)
2591 finally (return (logxor a c)))))
2593 (defun logxor-derive-unsigned-high-bound (x y)
2594 (let ((a (numeric-type-low x))
2595 (b (numeric-type-high x))
2596 (c (numeric-type-low y))
2597 (d (numeric-type-high y)))
2598 (loop for m = (ash 1 (integer-length (logand b d))) then (ash m -1)
2600 (unless (zerop (logand b d m))
2601 (let ((temp (logior (- b m) (1- m))))
2603 ((>= temp a) (setf b temp))
2604 (t (let ((temp (logior (- d m) (1- m))))
2607 finally (return (logxor b d)))))
2609 (defun logxor-derive-type-aux (x y &optional same-leaf)
2611 (return-from logxor-derive-type-aux (specifier-type '(eql 0))))
2612 (multiple-value-bind (x-len x-pos x-neg) (integer-type-length x)
2613 (multiple-value-bind (y-len y-pos y-neg) (integer-type-length y)
2615 ((and (not x-neg) (not y-neg))
2616 ;; Both are positive
2617 (if (and x-len y-len)
2618 (let ((low (logxor-derive-unsigned-low-bound x y))
2619 (high (logxor-derive-unsigned-high-bound x y)))
2620 (specifier-type `(integer ,low ,high)))
2621 (specifier-type '(unsigned-byte* *))))
2622 ((and (not x-pos) (not y-pos))
2623 ;; Both are negative. The result will be positive, and as long
2625 (specifier-type `(unsigned-byte* ,(if (and x-len y-len)
2628 ((or (and (not x-pos) (not y-neg))
2629 (and (not y-pos) (not x-neg)))
2630 ;; Either X is negative and Y is positive or vice-versa. The
2631 ;; result will be negative.
2632 (specifier-type `(integer ,(if (and x-len y-len)
2633 (ash -1 (max x-len y-len))
2636 ;; We can't tell what the sign of the result is going to be.
2637 ;; All we know is that we don't create new bits.
2639 (specifier-type `(signed-byte ,(1+ (max x-len y-len)))))
2641 (specifier-type 'integer))))))
2643 (macrolet ((deffrob (logfun)
2644 (let ((fun-aux (symbolicate logfun "-DERIVE-TYPE-AUX")))
2645 `(defoptimizer (,logfun derive-type) ((x y))
2646 (two-arg-derive-type x y #',fun-aux #',logfun)))))
2651 (defoptimizer (logeqv derive-type) ((x y))
2652 (two-arg-derive-type x y (lambda (x y same-leaf)
2653 (lognot-derive-type-aux
2654 (logxor-derive-type-aux x y same-leaf)))
2656 (defoptimizer (lognand derive-type) ((x y))
2657 (two-arg-derive-type x y (lambda (x y same-leaf)
2658 (lognot-derive-type-aux
2659 (logand-derive-type-aux x y same-leaf)))
2661 (defoptimizer (lognor derive-type) ((x y))
2662 (two-arg-derive-type x y (lambda (x y same-leaf)
2663 (lognot-derive-type-aux
2664 (logior-derive-type-aux x y same-leaf)))
2666 (defoptimizer (logandc1 derive-type) ((x y))
2667 (two-arg-derive-type x y (lambda (x y same-leaf)
2669 (specifier-type '(eql 0))
2670 (logand-derive-type-aux
2671 (lognot-derive-type-aux x) y nil)))
2673 (defoptimizer (logandc2 derive-type) ((x y))
2674 (two-arg-derive-type x y (lambda (x y same-leaf)
2676 (specifier-type '(eql 0))
2677 (logand-derive-type-aux
2678 x (lognot-derive-type-aux y) nil)))
2680 (defoptimizer (logorc1 derive-type) ((x y))
2681 (two-arg-derive-type x y (lambda (x y same-leaf)
2683 (specifier-type '(eql -1))
2684 (logior-derive-type-aux
2685 (lognot-derive-type-aux x) y nil)))
2687 (defoptimizer (logorc2 derive-type) ((x y))
2688 (two-arg-derive-type x y (lambda (x y same-leaf)
2690 (specifier-type '(eql -1))
2691 (logior-derive-type-aux
2692 x (lognot-derive-type-aux y) nil)))
2695 ;;;; miscellaneous derive-type methods
2697 (defoptimizer (integer-length derive-type) ((x))
2698 (let ((x-type (lvar-type x)))
2699 (when (numeric-type-p x-type)
2700 ;; If the X is of type (INTEGER LO HI), then the INTEGER-LENGTH
2701 ;; of X is (INTEGER (MIN lo hi) (MAX lo hi), basically. Be
2702 ;; careful about LO or HI being NIL, though. Also, if 0 is
2703 ;; contained in X, the lower bound is obviously 0.
2704 (flet ((null-or-min (a b)
2705 (and a b (min (integer-length a)
2706 (integer-length b))))
2708 (and a b (max (integer-length a)
2709 (integer-length b)))))
2710 (let* ((min (numeric-type-low x-type))
2711 (max (numeric-type-high x-type))
2712 (min-len (null-or-min min max))
2713 (max-len (null-or-max min max)))
2714 (when (ctypep 0 x-type)
2716 (specifier-type `(integer ,(or min-len '*) ,(or max-len '*))))))))
2718 (defoptimizer (isqrt derive-type) ((x))
2719 (let ((x-type (lvar-type x)))
2720 (when (numeric-type-p x-type)
2721 (let* ((lo (numeric-type-low x-type))
2722 (hi (numeric-type-high x-type))
2723 (lo-res (if lo (isqrt lo) '*))
2724 (hi-res (if hi (isqrt hi) '*)))
2725 (specifier-type `(integer ,lo-res ,hi-res))))))
2727 (defoptimizer (char-code derive-type) ((char))
2728 (let ((type (type-intersection (lvar-type char) (specifier-type 'character))))
2729 (cond ((member-type-p type)
2732 ,@(loop for member in (member-type-members type)
2733 when (characterp member)
2734 collect (char-code member)))))
2735 ((sb!kernel::character-set-type-p type)
2738 ,@(loop for (low . high)
2739 in (character-set-type-pairs type)
2740 collect `(integer ,low ,high)))))
2741 ((csubtypep type (specifier-type 'base-char))
2743 `(mod ,base-char-code-limit)))
2746 `(mod ,char-code-limit))))))
2748 (defoptimizer (code-char derive-type) ((code))
2749 (let ((type (lvar-type code)))
2750 ;; FIXME: unions of integral ranges? It ought to be easier to do
2751 ;; this, given that CHARACTER-SET is basically an integral range
2752 ;; type. -- CSR, 2004-10-04
2753 (when (numeric-type-p type)
2754 (let* ((lo (numeric-type-low type))
2755 (hi (numeric-type-high type))
2756 (type (specifier-type `(character-set ((,lo . ,hi))))))
2758 ;; KLUDGE: when running on the host, we lose a slight amount
2759 ;; of precision so that we don't have to "unparse" types
2760 ;; that formally we can't, such as (CHARACTER-SET ((0
2761 ;; . 0))). -- CSR, 2004-10-06
2763 ((csubtypep type (specifier-type 'standard-char)) type)
2765 ((csubtypep type (specifier-type 'base-char))
2766 (specifier-type 'base-char))
2768 ((csubtypep type (specifier-type 'extended-char))
2769 (specifier-type 'extended-char))
2770 (t #+sb-xc-host (specifier-type 'character)
2771 #-sb-xc-host type))))))
2773 (defoptimizer (values derive-type) ((&rest values))
2774 (make-values-type :required (mapcar #'lvar-type values)))
2776 (defun signum-derive-type-aux (type)
2777 (if (eq (numeric-type-complexp type) :complex)
2778 (let* ((format (case (numeric-type-class type)
2779 ((integer rational) 'single-float)
2780 (t (numeric-type-format type))))
2781 (bound-format (or format 'float)))
2782 (make-numeric-type :class 'float
2785 :low (coerce -1 bound-format)
2786 :high (coerce 1 bound-format)))
2787 (let* ((interval (numeric-type->interval type))
2788 (range-info (interval-range-info interval))
2789 (contains-0-p (interval-contains-p 0 interval))
2790 (class (numeric-type-class type))
2791 (format (numeric-type-format type))
2792 (one (coerce 1 (or format class 'real)))
2793 (zero (coerce 0 (or format class 'real)))
2794 (minus-one (coerce -1 (or format class 'real)))
2795 (plus (make-numeric-type :class class :format format
2796 :low one :high one))
2797 (minus (make-numeric-type :class class :format format
2798 :low minus-one :high minus-one))
2799 ;; KLUDGE: here we have a fairly horrible hack to deal
2800 ;; with the schizophrenia in the type derivation engine.
2801 ;; The problem is that the type derivers reinterpret
2802 ;; numeric types as being exact; so (DOUBLE-FLOAT 0d0
2803 ;; 0d0) within the derivation mechanism doesn't include
2804 ;; -0d0. Ugh. So force it in here, instead.
2805 (zero (make-numeric-type :class class :format format
2806 :low (- zero) :high zero)))
2808 (+ (if contains-0-p (type-union plus zero) plus))
2809 (- (if contains-0-p (type-union minus zero) minus))
2810 (t (type-union minus zero plus))))))
2812 (defoptimizer (signum derive-type) ((num))
2813 (one-arg-derive-type num #'signum-derive-type-aux nil))
2815 ;;;; byte operations
2817 ;;;; We try to turn byte operations into simple logical operations.
2818 ;;;; First, we convert byte specifiers into separate size and position
2819 ;;;; arguments passed to internal %FOO functions. We then attempt to
2820 ;;;; transform the %FOO functions into boolean operations when the
2821 ;;;; size and position are constant and the operands are fixnums.
2823 (macrolet (;; Evaluate body with SIZE-VAR and POS-VAR bound to
2824 ;; expressions that evaluate to the SIZE and POSITION of
2825 ;; the byte-specifier form SPEC. We may wrap a let around
2826 ;; the result of the body to bind some variables.
2828 ;; If the spec is a BYTE form, then bind the vars to the
2829 ;; subforms. otherwise, evaluate SPEC and use the BYTE-SIZE
2830 ;; and BYTE-POSITION. The goal of this transformation is to
2831 ;; avoid consing up byte specifiers and then immediately
2832 ;; throwing them away.
2833 (with-byte-specifier ((size-var pos-var spec) &body body)
2834 (once-only ((spec `(macroexpand ,spec))
2836 `(if (and (consp ,spec)
2837 (eq (car ,spec) 'byte)
2838 (= (length ,spec) 3))
2839 (let ((,size-var (second ,spec))
2840 (,pos-var (third ,spec)))
2842 (let ((,size-var `(byte-size ,,temp))
2843 (,pos-var `(byte-position ,,temp)))
2844 `(let ((,,temp ,,spec))
2847 (define-source-transform ldb (spec int)
2848 (with-byte-specifier (size pos spec)
2849 `(%ldb ,size ,pos ,int)))
2851 (define-source-transform dpb (newbyte spec int)
2852 (with-byte-specifier (size pos spec)
2853 `(%dpb ,newbyte ,size ,pos ,int)))
2855 (define-source-transform mask-field (spec int)
2856 (with-byte-specifier (size pos spec)
2857 `(%mask-field ,size ,pos ,int)))
2859 (define-source-transform deposit-field (newbyte spec int)
2860 (with-byte-specifier (size pos spec)
2861 `(%deposit-field ,newbyte ,size ,pos ,int))))
2863 (defoptimizer (%ldb derive-type) ((size posn num))
2864 (let ((size (lvar-type size)))
2865 (if (and (numeric-type-p size)
2866 (csubtypep size (specifier-type 'integer)))
2867 (let ((size-high (numeric-type-high size)))
2868 (if (and size-high (<= size-high sb!vm:n-word-bits))
2869 (specifier-type `(unsigned-byte* ,size-high))
2870 (specifier-type 'unsigned-byte)))
2873 (defoptimizer (%mask-field derive-type) ((size posn num))
2874 (let ((size (lvar-type size))
2875 (posn (lvar-type posn)))
2876 (if (and (numeric-type-p size)
2877 (csubtypep size (specifier-type 'integer))
2878 (numeric-type-p posn)
2879 (csubtypep posn (specifier-type 'integer)))
2880 (let ((size-high (numeric-type-high size))
2881 (posn-high (numeric-type-high posn)))
2882 (if (and size-high posn-high
2883 (<= (+ size-high posn-high) sb!vm:n-word-bits))
2884 (specifier-type `(unsigned-byte* ,(+ size-high posn-high)))
2885 (specifier-type 'unsigned-byte)))
2888 (defun %deposit-field-derive-type-aux (size posn int)
2889 (let ((size (lvar-type size))
2890 (posn (lvar-type posn))
2891 (int (lvar-type int)))
2892 (when (and (numeric-type-p size)
2893 (numeric-type-p posn)
2894 (numeric-type-p int))
2895 (let ((size-high (numeric-type-high size))
2896 (posn-high (numeric-type-high posn))
2897 (high (numeric-type-high int))
2898 (low (numeric-type-low int)))
2899 (when (and size-high posn-high high low
2900 ;; KLUDGE: we need this cutoff here, otherwise we
2901 ;; will merrily derive the type of %DPB as
2902 ;; (UNSIGNED-BYTE 1073741822), and then attempt to
2903 ;; canonicalize this type to (INTEGER 0 (1- (ASH 1
2904 ;; 1073741822))), with hilarious consequences. We
2905 ;; cutoff at 4*SB!VM:N-WORD-BITS to allow inference
2906 ;; over a reasonable amount of shifting, even on
2907 ;; the alpha/32 port, where N-WORD-BITS is 32 but
2908 ;; machine integers are 64-bits. -- CSR,
2910 (<= (+ size-high posn-high) (* 4 sb!vm:n-word-bits)))
2911 (let ((raw-bit-count (max (integer-length high)
2912 (integer-length low)
2913 (+ size-high posn-high))))
2916 `(signed-byte ,(1+ raw-bit-count))
2917 `(unsigned-byte* ,raw-bit-count)))))))))
2919 (defoptimizer (%dpb derive-type) ((newbyte size posn int))
2920 (%deposit-field-derive-type-aux size posn int))
2922 (defoptimizer (%deposit-field derive-type) ((newbyte size posn int))
2923 (%deposit-field-derive-type-aux size posn int))
2925 (deftransform %ldb ((size posn int)
2926 (fixnum fixnum integer)
2927 (unsigned-byte #.sb!vm:n-word-bits))
2928 "convert to inline logical operations"
2929 `(logand (ash int (- posn))
2930 (ash ,(1- (ash 1 sb!vm:n-word-bits))
2931 (- size ,sb!vm:n-word-bits))))
2933 (deftransform %mask-field ((size posn int)
2934 (fixnum fixnum integer)
2935 (unsigned-byte #.sb!vm:n-word-bits))
2936 "convert to inline logical operations"
2938 (ash (ash ,(1- (ash 1 sb!vm:n-word-bits))
2939 (- size ,sb!vm:n-word-bits))
2942 ;;; Note: for %DPB and %DEPOSIT-FIELD, we can't use
2943 ;;; (OR (SIGNED-BYTE N) (UNSIGNED-BYTE N))
2944 ;;; as the result type, as that would allow result types that cover
2945 ;;; the range -2^(n-1) .. 1-2^n, instead of allowing result types of
2946 ;;; (UNSIGNED-BYTE N) and result types of (SIGNED-BYTE N).
2948 (deftransform %dpb ((new size posn int)
2950 (unsigned-byte #.sb!vm:n-word-bits))
2951 "convert to inline logical operations"
2952 `(let ((mask (ldb (byte size 0) -1)))
2953 (logior (ash (logand new mask) posn)
2954 (logand int (lognot (ash mask posn))))))
2956 (deftransform %dpb ((new size posn int)
2958 (signed-byte #.sb!vm:n-word-bits))
2959 "convert to inline logical operations"
2960 `(let ((mask (ldb (byte size 0) -1)))
2961 (logior (ash (logand new mask) posn)
2962 (logand int (lognot (ash mask posn))))))
2964 (deftransform %deposit-field ((new size posn int)
2966 (unsigned-byte #.sb!vm:n-word-bits))
2967 "convert to inline logical operations"
2968 `(let ((mask (ash (ldb (byte size 0) -1) posn)))
2969 (logior (logand new mask)
2970 (logand int (lognot mask)))))
2972 (deftransform %deposit-field ((new size posn int)
2974 (signed-byte #.sb!vm:n-word-bits))
2975 "convert to inline logical operations"
2976 `(let ((mask (ash (ldb (byte size 0) -1) posn)))
2977 (logior (logand new mask)
2978 (logand int (lognot mask)))))
2980 (defoptimizer (mask-signed-field derive-type) ((size x))
2981 (let ((size (lvar-type size)))
2982 (if (numeric-type-p size)
2983 (let ((size-high (numeric-type-high size)))
2984 (if (and size-high (<= 1 size-high sb!vm:n-word-bits))
2985 (specifier-type `(signed-byte ,size-high))
2990 ;;; Modular functions
2992 ;;; (ldb (byte s 0) (foo x y ...)) =
2993 ;;; (ldb (byte s 0) (foo (ldb (byte s 0) x) y ...))
2995 ;;; and similar for other arguments.
2997 (defun make-modular-fun-type-deriver (prototype kind width signedp)
2998 (declare (ignore kind))
3000 (binding* ((info (info :function :info prototype) :exit-if-null)
3001 (fun (fun-info-derive-type info) :exit-if-null)
3002 (mask-type (specifier-type
3004 ((nil) (let ((mask (1- (ash 1 width))))
3005 `(integer ,mask ,mask)))
3006 ((t) `(signed-byte ,width))))))
3008 (let ((res (funcall fun call)))
3010 (if (eq signedp nil)
3011 (logand-derive-type-aux res mask-type))))))
3014 (binding* ((info (info :function :info prototype) :exit-if-null)
3015 (fun (fun-info-derive-type info) :exit-if-null)
3016 (res (funcall fun call) :exit-if-null)
3017 (mask-type (specifier-type
3019 ((nil) (let ((mask (1- (ash 1 width))))
3020 `(integer ,mask ,mask)))
3021 ((t) `(signed-byte ,width))))))
3022 (if (eq signedp nil)
3023 (logand-derive-type-aux res mask-type)))))
3025 ;;; Try to recursively cut all uses of LVAR to WIDTH bits.
3027 ;;; For good functions, we just recursively cut arguments; their
3028 ;;; "goodness" means that the result will not increase (in the
3029 ;;; (unsigned-byte +infinity) sense). An ordinary modular function is
3030 ;;; replaced with the version, cutting its result to WIDTH or more
3031 ;;; bits. For most functions (e.g. for +) we cut all arguments; for
3032 ;;; others (e.g. for ASH) we have "optimizers", cutting only necessary
3033 ;;; arguments (maybe to a different width) and returning the name of a
3034 ;;; modular version, if it exists, or NIL. If we have changed
3035 ;;; anything, we need to flush old derived types, because they have
3036 ;;; nothing in common with the new code.
3037 (defun cut-to-width (lvar kind width signedp)
3038 (declare (type lvar lvar) (type (integer 0) width))
3039 (let ((type (specifier-type (if (zerop width)
3042 ((nil) 'unsigned-byte)
3045 (labels ((reoptimize-node (node name)
3046 (setf (node-derived-type node)
3048 (info :function :type name)))
3049 (setf (lvar-%derived-type (node-lvar node)) nil)
3050 (setf (node-reoptimize node) t)
3051 (setf (block-reoptimize (node-block node)) t)
3052 (reoptimize-component (node-component node) :maybe))
3053 (cut-node (node &aux did-something)
3054 (when (and (not (block-delete-p (node-block node)))
3056 (constant-p (ref-leaf node)))
3057 (let* ((constant-value (constant-value (ref-leaf node)))
3058 (new-value (if signedp
3059 (mask-signed-field width constant-value)
3060 (ldb (byte width 0) constant-value))))
3061 (unless (= constant-value new-value)
3062 (change-ref-leaf node (make-constant new-value))
3063 (let ((lvar (node-lvar node)))
3064 (setf (lvar-%derived-type lvar)
3065 (and (lvar-has-single-use-p lvar)
3066 (make-values-type :required (list (ctype-of new-value))))))
3067 (setf (block-reoptimize (node-block node)) t)
3068 (reoptimize-component (node-component node) :maybe)
3069 (return-from cut-node t))))
3070 (when (and (not (block-delete-p (node-block node)))
3071 (combination-p node)
3072 (eq (basic-combination-kind node) :known))
3073 (let* ((fun-ref (lvar-use (combination-fun node)))
3074 (fun-name (leaf-source-name (ref-leaf fun-ref)))
3075 (modular-fun (find-modular-version fun-name kind signedp width)))
3076 (when (and modular-fun
3077 (not (and (eq fun-name 'logand)
3079 (single-value-type (node-derived-type node))
3081 (binding* ((name (etypecase modular-fun
3082 ((eql :good) fun-name)
3084 (modular-fun-info-name modular-fun))
3086 (funcall modular-fun node width)))
3088 (unless (eql modular-fun :good)
3089 (setq did-something t)
3092 (find-free-fun name "in a strange place"))
3093 (setf (combination-kind node) :full))
3094 (unless (functionp modular-fun)
3095 (dolist (arg (basic-combination-args node))
3096 (when (cut-lvar arg)
3097 (setq did-something t))))
3099 (reoptimize-node node name))
3101 (cut-lvar (lvar &aux did-something)
3102 (do-uses (node lvar)
3103 (when (cut-node node)
3104 (setq did-something t)))
3108 (defun best-modular-version (width signedp)
3109 ;; 1. exact width-matched :untagged
3110 ;; 2. >/>= width-matched :tagged
3111 ;; 3. >/>= width-matched :untagged
3112 (let* ((uuwidths (modular-class-widths *untagged-unsigned-modular-class*))
3113 (uswidths (modular-class-widths *untagged-signed-modular-class*))
3114 (uwidths (merge 'list uuwidths uswidths #'< :key #'car))
3115 (twidths (modular-class-widths *tagged-modular-class*)))
3116 (let ((exact (find (cons width signedp) uwidths :test #'equal)))
3118 (return-from best-modular-version (values width :untagged signedp))))
3119 (flet ((inexact-match (w)
3121 ((eq signedp (cdr w)) (<= width (car w)))
3122 ((eq signedp nil) (< width (car w))))))
3123 (let ((tgt (find-if #'inexact-match twidths)))
3125 (return-from best-modular-version
3126 (values (car tgt) :tagged (cdr tgt)))))
3127 (let ((ugt (find-if #'inexact-match uwidths)))
3129 (return-from best-modular-version
3130 (values (car ugt) :untagged (cdr ugt))))))))
3132 (defoptimizer (logand optimizer) ((x y) node)
3133 (let ((result-type (single-value-type (node-derived-type node))))
3134 (when (numeric-type-p result-type)
3135 (let ((low (numeric-type-low result-type))
3136 (high (numeric-type-high result-type)))
3137 (when (and (numberp low)
3140 (let ((width (integer-length high)))
3141 (multiple-value-bind (w kind signedp)
3142 (best-modular-version width nil)
3144 ;; FIXME: This should be (CUT-TO-WIDTH NODE KIND WIDTH SIGNEDP).
3146 ;; FIXME: I think the FIXME (which is from APD) above
3147 ;; implies that CUT-TO-WIDTH should do /everything/
3148 ;; that's required, including reoptimizing things
3149 ;; itself that it knows are necessary. At the moment,
3150 ;; CUT-TO-WIDTH sets up some new calls with
3151 ;; combination-type :FULL, which later get noticed as
3152 ;; known functions and properly converted.
3154 ;; We cut to W not WIDTH if SIGNEDP is true, because
3155 ;; signed constant replacement needs to know which bit
3156 ;; in the field is the signed bit.
3157 (let ((xact (cut-to-width x kind (if signedp w width) signedp))
3158 (yact (cut-to-width y kind (if signedp w width) signedp)))
3159 (declare (ignore xact yact))
3160 nil) ; After fixing above, replace with T, meaning
3161 ; "don't reoptimize this (LOGAND) node any more".
3164 (defoptimizer (mask-signed-field optimizer) ((width x) node)
3165 (let ((result-type (single-value-type (node-derived-type node))))
3166 (when (numeric-type-p result-type)
3167 (let ((low (numeric-type-low result-type))
3168 (high (numeric-type-high result-type)))
3169 (when (and (numberp low) (numberp high))
3170 (let ((width (max (integer-length high) (integer-length low))))
3171 (multiple-value-bind (w kind)
3172 (best-modular-version (1+ width) t)
3174 ;; FIXME: This should be (CUT-TO-WIDTH NODE KIND W T).
3175 ;; [ see comment above in LOGAND optimizer ]
3176 (cut-to-width x kind w t)
3177 nil ; After fixing above, replace with T.
3180 ;;; miscellanous numeric transforms
3182 ;;; If a constant appears as the first arg, swap the args.
3183 (deftransform commutative-arg-swap ((x y) * * :defun-only t :node node)
3184 (if (and (constant-lvar-p x)
3185 (not (constant-lvar-p y)))
3186 `(,(lvar-fun-name (basic-combination-fun node))
3189 (give-up-ir1-transform)))
3191 (dolist (x '(= char= + * logior logand logxor))
3192 (%deftransform x '(function * *) #'commutative-arg-swap
3193 "place constant arg last"))
3195 ;;; Handle the case of a constant BOOLE-CODE.
3196 (deftransform boole ((op x y) * *)
3197 "convert to inline logical operations"
3198 (unless (constant-lvar-p op)
3199 (give-up-ir1-transform "BOOLE code is not a constant."))
3200 (let ((control (lvar-value op)))
3202 (#.sb!xc:boole-clr 0)
3203 (#.sb!xc:boole-set -1)
3204 (#.sb!xc:boole-1 'x)
3205 (#.sb!xc:boole-2 'y)
3206 (#.sb!xc:boole-c1 '(lognot x))
3207 (#.sb!xc:boole-c2 '(lognot y))
3208 (#.sb!xc:boole-and '(logand x y))
3209 (#.sb!xc:boole-ior '(logior x y))
3210 (#.sb!xc:boole-xor '(logxor x y))
3211 (#.sb!xc:boole-eqv '(logeqv x y))
3212 (#.sb!xc:boole-nand '(lognand x y))
3213 (#.sb!xc:boole-nor '(lognor x y))
3214 (#.sb!xc:boole-andc1 '(logandc1 x y))
3215 (#.sb!xc:boole-andc2 '(logandc2 x y))
3216 (#.sb!xc:boole-orc1 '(logorc1 x y))
3217 (#.sb!xc:boole-orc2 '(logorc2 x y))
3219 (abort-ir1-transform "~S is an illegal control arg to BOOLE."
3222 ;;;; converting special case multiply/divide to shifts
3224 ;;; If arg is a constant power of two, turn * into a shift.
3225 (deftransform * ((x y) (integer integer) *)
3226 "convert x*2^k to shift"
3227 (unless (constant-lvar-p y)
3228 (give-up-ir1-transform))
3229 (let* ((y (lvar-value y))
3231 (len (1- (integer-length y-abs))))
3232 (unless (and (> y-abs 0) (= y-abs (ash 1 len)))
3233 (give-up-ir1-transform))
3238 ;;; These must come before the ones below, so that they are tried
3239 ;;; first. Since %FLOOR and %CEILING are inlined, this allows
3240 ;;; the general case to be handled by TRUNCATE transforms.
3241 (deftransform floor ((x y))
3244 (deftransform ceiling ((x y))
3247 ;;; If arg is a constant power of two, turn FLOOR into a shift and
3248 ;;; mask. If CEILING, add in (1- (ABS Y)), do FLOOR and correct a
3250 (flet ((frob (y ceil-p)
3251 (unless (constant-lvar-p y)
3252 (give-up-ir1-transform))
3253 (let* ((y (lvar-value y))
3255 (len (1- (integer-length y-abs))))
3256 (unless (and (> y-abs 0) (= y-abs (ash 1 len)))
3257 (give-up-ir1-transform))
3258 (let ((shift (- len))
3260 (delta (if ceil-p (* (signum y) (1- y-abs)) 0)))
3261 `(let ((x (+ x ,delta)))
3263 `(values (ash (- x) ,shift)
3264 (- (- (logand (- x) ,mask)) ,delta))
3265 `(values (ash x ,shift)
3266 (- (logand x ,mask) ,delta))))))))
3267 (deftransform floor ((x y) (integer integer) *)
3268 "convert division by 2^k to shift"
3270 (deftransform ceiling ((x y) (integer integer) *)
3271 "convert division by 2^k to shift"
3274 ;;; Do the same for MOD.
3275 (deftransform mod ((x y) (integer integer) *)
3276 "convert remainder mod 2^k to LOGAND"
3277 (unless (constant-lvar-p y)
3278 (give-up-ir1-transform))
3279 (let* ((y (lvar-value y))
3281 (len (1- (integer-length y-abs))))
3282 (unless (and (> y-abs 0) (= y-abs (ash 1 len)))
3283 (give-up-ir1-transform))
3284 (let ((mask (1- y-abs)))
3286 `(- (logand (- x) ,mask))
3287 `(logand x ,mask)))))
3289 ;;; If arg is a constant power of two, turn TRUNCATE into a shift and mask.
3290 (deftransform truncate ((x y) (integer integer))
3291 "convert division by 2^k to shift"
3292 (unless (constant-lvar-p y)
3293 (give-up-ir1-transform))
3294 (let* ((y (lvar-value y))
3296 (len (1- (integer-length y-abs))))
3297 (unless (and (> y-abs 0) (= y-abs (ash 1 len)))
3298 (give-up-ir1-transform))
3299 (let* ((shift (- len))
3302 (values ,(if (minusp y)
3304 `(- (ash (- x) ,shift)))
3305 (- (logand (- x) ,mask)))
3306 (values ,(if (minusp y)
3307 `(ash (- ,mask x) ,shift)
3309 (logand x ,mask))))))
3311 ;;; And the same for REM.
3312 (deftransform rem ((x y) (integer integer) *)
3313 "convert remainder mod 2^k to LOGAND"
3314 (unless (constant-lvar-p y)
3315 (give-up-ir1-transform))
3316 (let* ((y (lvar-value y))
3318 (len (1- (integer-length y-abs))))
3319 (unless (and (> y-abs 0) (= y-abs (ash 1 len)))
3320 (give-up-ir1-transform))
3321 (let ((mask (1- y-abs)))
3323 (- (logand (- x) ,mask))
3324 (logand x ,mask)))))
3326 ;;; Return an expression to calculate the integer quotient of X and
3327 ;;; constant Y, using multiplication, shift and add/sub instead of
3328 ;;; division. Both arguments must be unsigned, fit in a machine word and
3329 ;;; Y must neither be zero nor a power of two. The quotient is rounded
3331 ;;; The algorithm is taken from the paper "Division by Invariant
3332 ;;; Integers using Multiplication", 1994 by Torbj\"{o}rn Granlund and
3333 ;;; Peter L. Montgomery, Figures 4.2 and 6.2, modified to exclude the
3334 ;;; case of division by powers of two.
3335 ;;; The algorithm includes an adaptive precision argument. Use it, since
3336 ;;; we often have sub-word value ranges. Careful, in this case, we need
3337 ;;; p s.t 2^p > n, not the ceiling of the binary log.
3338 ;;; Also, for some reason, the paper prefers shifting to masking. Mask
3339 ;;; instead. Masking is equivalent to shifting right, then left again;
3340 ;;; all the intermediate values are still words, so we just have to shift
3341 ;;; right a bit more to compensate, at the end.
3343 ;;; The following two examples show an average case and the worst case
3344 ;;; with respect to the complexity of the generated expression, under
3345 ;;; a word size of 64 bits:
3347 ;;; (UNSIGNED-DIV-TRANSFORMER 10 MOST-POSITIVE-WORD) ->
3348 ;;; (ASH (%MULTIPLY (LOGANDC2 X 0) 14757395258967641293) -3)
3350 ;;; (UNSIGNED-DIV-TRANSFORMER 7 MOST-POSITIVE-WORD) ->
3352 ;;; (T1 (%MULTIPLY NUM 2635249153387078803)))
3353 ;;; (ASH (LDB (BYTE 64 0)
3354 ;;; (+ T1 (ASH (LDB (BYTE 64 0)
3359 (defun gen-unsigned-div-by-constant-expr (y max-x)
3360 (declare (type (integer 3 #.most-positive-word) y)
3362 (aver (not (zerop (logand y (1- y)))))
3364 ;; the floor of the binary logarithm of (positive) X
3365 (integer-length (1- x)))
3366 (choose-multiplier (y precision)
3368 (shift l (1- shift))
3369 (expt-2-n+l (expt 2 (+ sb!vm:n-word-bits l)))
3370 (m-low (truncate expt-2-n+l y) (ash m-low -1))
3371 (m-high (truncate (+ expt-2-n+l
3372 (ash expt-2-n+l (- precision)))
3375 ((not (and (< (ash m-low -1) (ash m-high -1))
3377 (values m-high shift)))))
3378 (let ((n (expt 2 sb!vm:n-word-bits))
3379 (precision (integer-length max-x))
3381 (multiple-value-bind (m shift2)
3382 (choose-multiplier y precision)
3383 (when (and (>= m n) (evenp y))
3384 (setq shift1 (ld (logand y (- y))))
3385 (multiple-value-setq (m shift2)
3386 (choose-multiplier (/ y (ash 1 shift1))
3387 (- precision shift1))))
3390 `(truly-the word ,x)))
3392 (t1 (%multiply-high num ,(- m n))))
3393 (ash ,(word `(+ t1 (ash ,(word `(- num t1))
3396 ((and (zerop shift1) (zerop shift2))
3397 (let ((max (truncate max-x y)))
3398 ;; Explicit TRULY-THE needed to get the FIXNUM=>FIXNUM
3400 `(truly-the (integer 0 ,max)
3401 (%multiply-high x ,m))))
3403 `(ash (%multiply-high (logandc2 x ,(1- (ash 1 shift1))) ,m)
3404 ,(- (+ shift1 shift2)))))))))
3406 ;;; If the divisor is constant and both args are positive and fit in a
3407 ;;; machine word, replace the division by a multiplication and possibly
3408 ;;; some shifts and an addition. Calculate the remainder by a second
3409 ;;; multiplication and a subtraction. Dead code elimination will
3410 ;;; suppress the latter part if only the quotient is needed. If the type
3411 ;;; of the dividend allows to derive that the quotient will always have
3412 ;;; the same value, emit much simpler code to handle that. (This case
3413 ;;; may be rare but it's easy to detect and the compiler doesn't find
3414 ;;; this optimization on its own.)
3415 (deftransform truncate ((x y) (word (constant-arg word))
3417 :policy (and (> speed compilation-speed)
3419 "convert integer division to multiplication"
3420 (let* ((y (lvar-value y))
3421 (x-type (lvar-type x))
3422 (max-x (or (and (numeric-type-p x-type)
3423 (numeric-type-high x-type))
3424 most-positive-word)))
3425 ;; Division by zero, one or powers of two is handled elsewhere.
3426 (when (zerop (logand y (1- y)))
3427 (give-up-ir1-transform))
3428 `(let* ((quot ,(gen-unsigned-div-by-constant-expr y max-x))
3429 (rem (ldb (byte #.sb!vm:n-word-bits 0)
3430 (- x (* quot ,y)))))
3431 (values quot rem))))
3433 ;;;; arithmetic and logical identity operation elimination
3435 ;;; Flush calls to various arith functions that convert to the
3436 ;;; identity function or a constant.
3437 (macrolet ((def (name identity result)
3438 `(deftransform ,name ((x y) (* (constant-arg (member ,identity))) *)
3439 "fold identity operations"
3446 (def logxor -1 (lognot x))
3449 (deftransform logand ((x y) (* (constant-arg t)) *)
3450 "fold identity operation"
3451 (let ((y (lvar-value y)))
3452 (unless (and (plusp y)
3453 (= y (1- (ash 1 (integer-length y)))))
3454 (give-up-ir1-transform))
3455 (unless (csubtypep (lvar-type x)
3456 (specifier-type `(integer 0 ,y)))
3457 (give-up-ir1-transform))
3460 (deftransform mask-signed-field ((size x) ((constant-arg t) *) *)
3461 "fold identity operation"
3462 (let ((size (lvar-value size)))
3463 (unless (csubtypep (lvar-type x) (specifier-type `(signed-byte ,size)))
3464 (give-up-ir1-transform))
3467 ;;; These are restricted to rationals, because (- 0 0.0) is 0.0, not -0.0, and
3468 ;;; (* 0 -4.0) is -0.0.
3469 (deftransform - ((x y) ((constant-arg (member 0)) rational) *)
3470 "convert (- 0 x) to negate"
3472 (deftransform * ((x y) (rational (constant-arg (member 0))) *)
3473 "convert (* x 0) to 0"
3476 ;;; Return T if in an arithmetic op including lvars X and Y, the
3477 ;;; result type is not affected by the type of X. That is, Y is at
3478 ;;; least as contagious as X.
3480 (defun not-more-contagious (x y)
3481 (declare (type continuation x y))
3482 (let ((x (lvar-type x))
3484 (values (type= (numeric-contagion x y)
3485 (numeric-contagion y y)))))
3486 ;;; Patched version by Raymond Toy. dtc: Should be safer although it
3487 ;;; XXX needs more work as valid transforms are missed; some cases are
3488 ;;; specific to particular transform functions so the use of this
3489 ;;; function may need a re-think.
3490 (defun not-more-contagious (x y)
3491 (declare (type lvar x y))
3492 (flet ((simple-numeric-type (num)
3493 (and (numeric-type-p num)
3494 ;; Return non-NIL if NUM is integer, rational, or a float
3495 ;; of some type (but not FLOAT)
3496 (case (numeric-type-class num)
3500 (numeric-type-format num))
3503 (let ((x (lvar-type x))
3505 (if (and (simple-numeric-type x)
3506 (simple-numeric-type y))
3507 (values (type= (numeric-contagion x y)
3508 (numeric-contagion y y)))))))
3510 (def!type exact-number ()
3511 '(or rational (complex rational)))
3515 ;;; Only safely applicable for exact numbers. For floating-point
3516 ;;; x, one would have to first show that neither x or y are signed
3517 ;;; 0s, and that x isn't an SNaN.
3518 (deftransform + ((x y) (exact-number (constant-arg (eql 0))) *)
3523 (deftransform - ((x y) (exact-number (constant-arg (eql 0))) *)
3527 ;;; Fold (OP x +/-1)
3529 ;;; %NEGATE might not always signal correctly.
3531 ((def (name result minus-result)
3532 `(deftransform ,name ((x y)
3533 (exact-number (constant-arg (member 1 -1))))
3534 "fold identity operations"
3535 (if (minusp (lvar-value y)) ',minus-result ',result))))
3536 (def * x (%negate x))
3537 (def / x (%negate x))
3538 (def expt x (/ 1 x)))
3540 ;;; Fold (expt x n) into multiplications for small integral values of
3541 ;;; N; convert (expt x 1/2) to sqrt.
3542 (deftransform expt ((x y) (t (constant-arg real)) *)
3543 "recode as multiplication or sqrt"
3544 (let ((val (lvar-value y)))
3545 ;; If Y would cause the result to be promoted to the same type as
3546 ;; Y, we give up. If not, then the result will be the same type
3547 ;; as X, so we can replace the exponentiation with simple
3548 ;; multiplication and division for small integral powers.
3549 (unless (not-more-contagious y x)
3550 (give-up-ir1-transform))
3552 (let ((x-type (lvar-type x)))
3553 (cond ((csubtypep x-type (specifier-type '(or rational
3554 (complex rational))))
3556 ((csubtypep x-type (specifier-type 'real))
3560 ((csubtypep x-type (specifier-type 'complex))
3561 ;; both parts are float
3563 (t (give-up-ir1-transform)))))
3564 ((= val 2) '(* x x))
3565 ((= val -2) '(/ (* x x)))
3566 ((= val 3) '(* x x x))
3567 ((= val -3) '(/ (* x x x)))
3568 ((= val 1/2) '(sqrt x))
3569 ((= val -1/2) '(/ (sqrt x)))
3570 (t (give-up-ir1-transform)))))
3572 (deftransform expt ((x y) ((constant-arg (member -1 -1.0 -1.0d0)) integer) *)
3573 "recode as an ODDP check"
3574 (let ((val (lvar-value x)))
3576 '(- 1 (* 2 (logand 1 y)))
3581 ;;; KLUDGE: Shouldn't (/ 0.0 0.0), etc. cause exceptions in these
3582 ;;; transformations?
3583 ;;; Perhaps we should have to prove that the denominator is nonzero before
3584 ;;; doing them? -- WHN 19990917
3585 (macrolet ((def (name)
3586 `(deftransform ,name ((x y) ((constant-arg (integer 0 0)) integer)
3593 (macrolet ((def (name)
3594 `(deftransform ,name ((x y) ((constant-arg (integer 0 0)) integer)
3603 (macrolet ((def (name &optional float)
3604 (let ((x (if float '(float x) 'x)))
3605 `(deftransform ,name ((x y) (integer (constant-arg (member 1 -1)))
3607 "fold division by 1"
3608 `(values ,(if (minusp (lvar-value y))
3621 ;;;; character operations
3623 (deftransform char-equal ((a b) (base-char base-char))
3625 '(let* ((ac (char-code a))
3627 (sum (logxor ac bc)))
3629 (when (eql sum #x20)
3630 (let ((sum (+ ac bc)))
3631 (or (and (> sum 161) (< sum 213))
3632 (and (> sum 415) (< sum 461))
3633 (and (> sum 463) (< sum 477))))))))
3635 (deftransform char-upcase ((x) (base-char))
3637 '(let ((n-code (char-code x)))
3638 (if (or (and (> n-code #o140) ; Octal 141 is #\a.
3639 (< n-code #o173)) ; Octal 172 is #\z.
3640 (and (> n-code #o337)
3642 (and (> n-code #o367)
3644 (code-char (logxor #x20 n-code))
3647 (deftransform char-downcase ((x) (base-char))
3649 '(let ((n-code (char-code x)))
3650 (if (or (and (> n-code 64) ; 65 is #\A.
3651 (< n-code 91)) ; 90 is #\Z.
3656 (code-char (logxor #x20 n-code))
3659 ;;;; equality predicate transforms
3661 ;;; Return true if X and Y are lvars whose only use is a
3662 ;;; reference to the same leaf, and the value of the leaf cannot
3664 (defun same-leaf-ref-p (x y)
3665 (declare (type lvar x y))
3666 (let ((x-use (principal-lvar-use x))
3667 (y-use (principal-lvar-use y)))
3670 (eq (ref-leaf x-use) (ref-leaf y-use))
3671 (constant-reference-p x-use))))
3673 ;;; If X and Y are the same leaf, then the result is true. Otherwise,
3674 ;;; if there is no intersection between the types of the arguments,
3675 ;;; then the result is definitely false.
3676 (deftransform simple-equality-transform ((x y) * *
3679 ((same-leaf-ref-p x y) t)
3680 ((not (types-equal-or-intersect (lvar-type x) (lvar-type y)))
3682 (t (give-up-ir1-transform))))
3685 `(%deftransform ',x '(function * *) #'simple-equality-transform)))
3689 ;;; This is similar to SIMPLE-EQUALITY-TRANSFORM, except that we also
3690 ;;; try to convert to a type-specific predicate or EQ:
3691 ;;; -- If both args are characters, convert to CHAR=. This is better than
3692 ;;; just converting to EQ, since CHAR= may have special compilation
3693 ;;; strategies for non-standard representations, etc.
3694 ;;; -- If either arg is definitely a fixnum, we check to see if X is
3695 ;;; constant and if so, put X second. Doing this results in better
3696 ;;; code from the backend, since the backend assumes that any constant
3697 ;;; argument comes second.
3698 ;;; -- If either arg is definitely not a number or a fixnum, then we
3699 ;;; can compare with EQ.
3700 ;;; -- Otherwise, we try to put the arg we know more about second. If X
3701 ;;; is constant then we put it second. If X is a subtype of Y, we put
3702 ;;; it second. These rules make it easier for the back end to match
3703 ;;; these interesting cases.
3704 (deftransform eql ((x y) * * :node node)
3705 "convert to simpler equality predicate"
3706 (let ((x-type (lvar-type x))
3707 (y-type (lvar-type y))
3708 (char-type (specifier-type 'character)))
3709 (flet ((fixnum-type-p (type)
3710 (csubtypep type (specifier-type 'fixnum))))
3712 ((same-leaf-ref-p x y) t)
3713 ((not (types-equal-or-intersect x-type y-type))
3715 ((and (csubtypep x-type char-type)
3716 (csubtypep y-type char-type))
3718 ((or (fixnum-type-p x-type) (fixnum-type-p y-type))
3719 (commutative-arg-swap node))
3720 ((or (eq-comparable-type-p x-type) (eq-comparable-type-p y-type))
3722 ((and (not (constant-lvar-p y))
3723 (or (constant-lvar-p x)
3724 (and (csubtypep x-type y-type)
3725 (not (csubtypep y-type x-type)))))
3728 (give-up-ir1-transform))))))
3730 ;;; similarly to the EQL transform above, we attempt to constant-fold
3731 ;;; or convert to a simpler predicate: mostly we have to be careful
3732 ;;; with strings and bit-vectors.
3733 (deftransform equal ((x y) * *)
3734 "convert to simpler equality predicate"
3735 (let ((x-type (lvar-type x))
3736 (y-type (lvar-type y))
3737 (string-type (specifier-type 'string))
3738 (bit-vector-type (specifier-type 'bit-vector)))
3740 ((same-leaf-ref-p x y) t)
3741 ((and (csubtypep x-type string-type)
3742 (csubtypep y-type string-type))
3744 ((and (csubtypep x-type bit-vector-type)
3745 (csubtypep y-type bit-vector-type))
3746 '(bit-vector-= x y))
3747 ;; if at least one is not a string, and at least one is not a
3748 ;; bit-vector, then we can reason from types.
3749 ((and (not (and (types-equal-or-intersect x-type string-type)
3750 (types-equal-or-intersect y-type string-type)))
3751 (not (and (types-equal-or-intersect x-type bit-vector-type)
3752 (types-equal-or-intersect y-type bit-vector-type)))
3753 (not (types-equal-or-intersect x-type y-type)))
3755 (t (give-up-ir1-transform)))))
3757 ;;; Convert to EQL if both args are rational and complexp is specified
3758 ;;; and the same for both.
3759 (deftransform = ((x y) (number number) *)
3761 (let ((x-type (lvar-type x))
3762 (y-type (lvar-type y)))
3763 (cond ((or (and (csubtypep x-type (specifier-type 'float))
3764 (csubtypep y-type (specifier-type 'float)))
3765 (and (csubtypep x-type (specifier-type '(complex float)))
3766 (csubtypep y-type (specifier-type '(complex float))))
3767 #!+complex-float-vops
3768 (and (csubtypep x-type (specifier-type '(or single-float (complex single-float))))
3769 (csubtypep y-type (specifier-type '(or single-float (complex single-float)))))
3770 #!+complex-float-vops
3771 (and (csubtypep x-type (specifier-type '(or double-float (complex double-float))))
3772 (csubtypep y-type (specifier-type '(or double-float (complex double-float))))))
3773 ;; They are both floats. Leave as = so that -0.0 is
3774 ;; handled correctly.
3775 (give-up-ir1-transform))
3776 ((or (and (csubtypep x-type (specifier-type 'rational))
3777 (csubtypep y-type (specifier-type 'rational)))
3778 (and (csubtypep x-type
3779 (specifier-type '(complex rational)))
3781 (specifier-type '(complex rational)))))
3782 ;; They are both rationals and complexp is the same.
3786 (give-up-ir1-transform
3787 "The operands might not be the same type.")))))
3789 (defun maybe-float-lvar-p (lvar)
3790 (neq *empty-type* (type-intersection (specifier-type 'float)
3793 (flet ((maybe-invert (node op inverted x y)
3794 ;; Don't invert if either argument can be a float (NaNs)
3796 ((or (maybe-float-lvar-p x) (maybe-float-lvar-p y))
3797 (delay-ir1-transform node :constraint)
3798 `(or (,op x y) (= x y)))
3800 `(if (,inverted x y) nil t)))))
3801 (deftransform >= ((x y) (number number) * :node node)
3802 "invert or open code"
3803 (maybe-invert node '> '< x y))
3804 (deftransform <= ((x y) (number number) * :node node)
3805 "invert or open code"
3806 (maybe-invert node '< '> x y)))
3808 ;;; See whether we can statically determine (< X Y) using type
3809 ;;; information. If X's high bound is < Y's low, then X < Y.
3810 ;;; Similarly, if X's low is >= to Y's high, the X >= Y (so return
3811 ;;; NIL). If not, at least make sure any constant arg is second.
3812 (macrolet ((def (name inverse reflexive-p surely-true surely-false)
3813 `(deftransform ,name ((x y))
3814 "optimize using intervals"
3815 (if (and (same-leaf-ref-p x y)
3816 ;; For non-reflexive functions we don't need
3817 ;; to worry about NaNs: (non-ref-op NaN NaN) => false,
3818 ;; but with reflexive ones we don't know...
3820 '((and (not (maybe-float-lvar-p x))
3821 (not (maybe-float-lvar-p y))))))
3823 (let ((ix (or (type-approximate-interval (lvar-type x))
3824 (give-up-ir1-transform)))
3825 (iy (or (type-approximate-interval (lvar-type y))
3826 (give-up-ir1-transform))))
3831 ((and (constant-lvar-p x)
3832 (not (constant-lvar-p y)))
3835 (give-up-ir1-transform))))))))
3836 (def = = t (interval-= ix iy) (interval-/= ix iy))
3837 (def /= /= nil (interval-/= ix iy) (interval-= ix iy))
3838 (def < > nil (interval-< ix iy) (interval->= ix iy))
3839 (def > < nil (interval-< iy ix) (interval->= iy ix))
3840 (def <= >= t (interval->= iy ix) (interval-< iy ix))
3841 (def >= <= t (interval->= ix iy) (interval-< ix iy)))
3843 (defun ir1-transform-char< (x y first second inverse)
3845 ((same-leaf-ref-p x y) nil)
3846 ;; If we had interval representation of character types, as we
3847 ;; might eventually have to to support 2^21 characters, then here
3848 ;; we could do some compile-time computation as in transforms for
3849 ;; < above. -- CSR, 2003-07-01
3850 ((and (constant-lvar-p first)
3851 (not (constant-lvar-p second)))
3853 (t (give-up-ir1-transform))))
3855 (deftransform char< ((x y) (character character) *)
3856 (ir1-transform-char< x y x y 'char>))
3858 (deftransform char> ((x y) (character character) *)
3859 (ir1-transform-char< y x x y 'char<))
3861 ;;;; converting N-arg comparisons
3863 ;;;; We convert calls to N-arg comparison functions such as < into
3864 ;;;; two-arg calls. This transformation is enabled for all such
3865 ;;;; comparisons in this file. If any of these predicates are not
3866 ;;;; open-coded, then the transformation should be removed at some
3867 ;;;; point to avoid pessimization.
3869 ;;; This function is used for source transformation of N-arg
3870 ;;; comparison functions other than inequality. We deal both with
3871 ;;; converting to two-arg calls and inverting the sense of the test,
3872 ;;; if necessary. If the call has two args, then we pass or return a
3873 ;;; negated test as appropriate. If it is a degenerate one-arg call,
3874 ;;; then we transform to code that returns true. Otherwise, we bind
3875 ;;; all the arguments and expand into a bunch of IFs.
3876 (defun multi-compare (predicate args not-p type &optional force-two-arg-p)
3877 (let ((nargs (length args)))
3878 (cond ((< nargs 1) (values nil t))
3879 ((= nargs 1) `(progn (the ,type ,@args) t))
3882 `(if (,predicate ,(first args) ,(second args)) nil t)
3884 `(,predicate ,(first args) ,(second args))
3887 (do* ((i (1- nargs) (1- i))
3889 (current (gensym) (gensym))
3890 (vars (list current) (cons current vars))
3892 `(if (,predicate ,current ,last)
3894 `(if (,predicate ,current ,last)
3897 `((lambda ,vars (declare (type ,type ,@vars)) ,result)
3900 (define-source-transform = (&rest args) (multi-compare '= args nil 'number))
3901 (define-source-transform < (&rest args) (multi-compare '< args nil 'real))
3902 (define-source-transform > (&rest args) (multi-compare '> args nil 'real))
3903 ;;; We cannot do the inversion for >= and <= here, since both
3904 ;;; (< NaN X) and (> NaN X)
3905 ;;; are false, and we don't have type-information available yet. The
3906 ;;; deftransforms for two-argument versions of >= and <= takes care of
3907 ;;; the inversion to > and < when possible.
3908 (define-source-transform <= (&rest args) (multi-compare '<= args nil 'real))
3909 (define-source-transform >= (&rest args) (multi-compare '>= args nil 'real))
3911 (define-source-transform char= (&rest args) (multi-compare 'char= args nil
3913 (define-source-transform char< (&rest args) (multi-compare 'char< args nil
3915 (define-source-transform char> (&rest args) (multi-compare 'char> args nil
3917 (define-source-transform char<= (&rest args) (multi-compare 'char> args t
3919 (define-source-transform char>= (&rest args) (multi-compare 'char< args t
3922 (define-source-transform char-equal (&rest args)
3923 (multi-compare 'sb!impl::two-arg-char-equal args nil 'character t))
3924 (define-source-transform char-lessp (&rest args)
3925 (multi-compare 'sb!impl::two-arg-char-lessp args nil 'character t))
3926 (define-source-transform char-greaterp (&rest args)
3927 (multi-compare 'sb!impl::two-arg-char-greaterp args nil 'character t))
3928 (define-source-transform char-not-greaterp (&rest args)
3929 (multi-compare 'sb!impl::two-arg-char-greaterp args t 'character t))
3930 (define-source-transform char-not-lessp (&rest args)
3931 (multi-compare 'sb!impl::two-arg-char-lessp args t 'character t))
3933 ;;; This function does source transformation of N-arg inequality
3934 ;;; functions such as /=. This is similar to MULTI-COMPARE in the <3
3935 ;;; arg cases. If there are more than two args, then we expand into
3936 ;;; the appropriate n^2 comparisons only when speed is important.
3937 (declaim (ftype (function (symbol list t) *) multi-not-equal))
3938 (defun multi-not-equal (predicate args type)
3939 (let ((nargs (length args)))
3940 (cond ((< nargs 1) (values nil t))
3941 ((= nargs 1) `(progn (the ,type ,@args) t))
3943 `(if (,predicate ,(first args) ,(second args)) nil t))
3944 ((not (policy *lexenv*
3945 (and (>= speed space)
3946 (>= speed compilation-speed))))
3949 (let ((vars (make-gensym-list nargs)))
3950 (do ((var vars next)
3951 (next (cdr vars) (cdr next))
3954 `((lambda ,vars (declare (type ,type ,@vars)) ,result)
3956 (let ((v1 (first var)))
3958 (setq result `(if (,predicate ,v1 ,v2) nil ,result))))))))))
3960 (define-source-transform /= (&rest args)
3961 (multi-not-equal '= args 'number))
3962 (define-source-transform char/= (&rest args)
3963 (multi-not-equal 'char= args 'character))
3964 (define-source-transform char-not-equal (&rest args)
3965 (multi-not-equal 'char-equal args 'character))
3967 ;;; Expand MAX and MIN into the obvious comparisons.
3968 (define-source-transform max (arg0 &rest rest)
3969 (once-only ((arg0 arg0))
3971 `(values (the real ,arg0))
3972 `(let ((maxrest (max ,@rest)))
3973 (if (>= ,arg0 maxrest) ,arg0 maxrest)))))
3974 (define-source-transform min (arg0 &rest rest)
3975 (once-only ((arg0 arg0))
3977 `(values (the real ,arg0))
3978 `(let ((minrest (min ,@rest)))
3979 (if (<= ,arg0 minrest) ,arg0 minrest)))))
3981 ;;;; converting N-arg arithmetic functions
3983 ;;;; N-arg arithmetic and logic functions are associated into two-arg
3984 ;;;; versions, and degenerate cases are flushed.
3986 ;;; Left-associate FIRST-ARG and MORE-ARGS using FUNCTION.
3987 (declaim (ftype (sfunction (symbol t list t) list) associate-args))
3988 (defun associate-args (fun first-arg more-args identity)
3989 (let ((next (rest more-args))
3990 (arg (first more-args)))
3992 `(,fun ,first-arg ,(if arg arg identity))
3993 (associate-args fun `(,fun ,first-arg ,arg) next identity))))
3995 ;;; Reduce constants in ARGS list.
3996 (declaim (ftype (sfunction (symbol list t symbol) list) reduce-constants))
3997 (defun reduce-constants (fun args identity one-arg-result-type)
3998 (let ((one-arg-constant-p (ecase one-arg-result-type
4000 (integer #'integerp)))
4001 (reduced-value identity)
4003 (collect ((not-constants))
4005 (if (funcall one-arg-constant-p arg)
4006 (setf reduced-value (funcall fun reduced-value arg)
4008 (not-constants arg)))
4009 ;; It is tempting to drop constants reduced to identity here,
4010 ;; but if X is SNaN in (* X 1), we cannot drop the 1.
4013 `(,reduced-value ,@(not-constants))
4015 `(,reduced-value)))))
4017 ;;; Do source transformations for transitive functions such as +.
4018 ;;; One-arg cases are replaced with the arg and zero arg cases with
4019 ;;; the identity. ONE-ARG-RESULT-TYPE is the type to ensure (with THE)
4020 ;;; that the argument in one-argument calls is.
4021 (declaim (ftype (function (symbol list t &optional symbol list)
4022 (values t &optional (member nil t)))
4023 source-transform-transitive))
4024 (defun source-transform-transitive (fun args identity
4025 &optional (one-arg-result-type 'number)
4026 (one-arg-prefixes '(values)))
4029 (1 `(,@one-arg-prefixes (the ,one-arg-result-type ,(first args))))
4031 (t (let ((reduced-args (reduce-constants fun args identity one-arg-result-type)))
4032 (associate-args fun (first reduced-args) (rest reduced-args) identity)))))
4034 (define-source-transform + (&rest args)
4035 (source-transform-transitive '+ args 0))
4036 (define-source-transform * (&rest args)
4037 (source-transform-transitive '* args 1))
4038 (define-source-transform logior (&rest args)
4039 (source-transform-transitive 'logior args 0 'integer))
4040 (define-source-transform logxor (&rest args)
4041 (source-transform-transitive 'logxor args 0 'integer))
4042 (define-source-transform logand (&rest args)
4043 (source-transform-transitive 'logand args -1 'integer))
4044 (define-source-transform logeqv (&rest args)
4045 (source-transform-transitive 'logeqv args -1 'integer))
4046 (define-source-transform gcd (&rest args)
4047 (source-transform-transitive 'gcd args 0 'integer '(abs)))
4048 (define-source-transform lcm (&rest args)
4049 (source-transform-transitive 'lcm args 1 'integer '(abs)))
4051 ;;; Do source transformations for intransitive n-arg functions such as
4052 ;;; /. With one arg, we form the inverse. With two args we pass.
4053 ;;; Otherwise we associate into two-arg calls.
4054 (declaim (ftype (function (symbol symbol list t list &optional symbol)
4055 (values list &optional (member nil t)))
4056 source-transform-intransitive))
4057 (defun source-transform-intransitive (fun fun* args identity one-arg-prefixes
4058 &optional (one-arg-result-type 'number))
4060 ((0 2) (values nil t))
4061 (1 `(,@one-arg-prefixes (the ,one-arg-result-type ,(first args))))
4062 (t (let ((reduced-args
4063 (reduce-constants fun* (rest args) identity one-arg-result-type)))
4064 (associate-args fun (first args) reduced-args identity)))))
4066 (define-source-transform - (&rest args)
4067 (source-transform-intransitive '- '+ args 0 '(%negate)))
4068 (define-source-transform / (&rest args)
4069 (source-transform-intransitive '/ '* args 1 '(/ 1)))
4071 ;;;; transforming APPLY
4073 ;;; We convert APPLY into MULTIPLE-VALUE-CALL so that the compiler
4074 ;;; only needs to understand one kind of variable-argument call. It is
4075 ;;; more efficient to convert APPLY to MV-CALL than MV-CALL to APPLY.
4076 (define-source-transform apply (fun arg &rest more-args)
4077 (let ((args (cons arg more-args)))
4078 `(multiple-value-call ,fun
4079 ,@(mapcar (lambda (x) `(values ,x)) (butlast args))
4080 (values-list ,(car (last args))))))
4082 ;;;; transforming references to &REST argument
4084 ;;; We add magical &MORE arguments to all functions with &REST. If ARG names
4085 ;;; the &REST argument, this returns the lambda-vars for the context and
4087 (defun possible-rest-arg-context (arg)
4089 (let* ((var (lexenv-find arg vars))
4090 (info (when (lambda-var-p var)
4091 (lambda-var-arg-info var))))
4093 (eq :rest (arg-info-kind info))
4094 (consp (arg-info-default info)))
4095 (values-list (arg-info-default info))))))
4097 (defun mark-more-context-used (rest-var)
4098 (let ((info (lambda-var-arg-info rest-var)))
4099 (aver (eq :rest (arg-info-kind info)))
4100 (destructuring-bind (context count &optional used) (arg-info-default info)
4102 (setf (arg-info-default info) (list context count t))))))
4104 (defun mark-more-context-invalid (rest-var)
4105 (let ((info (lambda-var-arg-info rest-var)))
4106 (aver (eq :rest (arg-info-kind info)))
4107 (setf (arg-info-default info) t)))
4109 ;;; This determines of we the REF to a &REST variable is headed towards
4110 ;;; parts unknown, or if we can really use the context.
4111 (defun rest-var-more-context-ok (lvar)
4112 (let* ((use (lvar-use lvar))
4113 (var (when (ref-p use) (ref-leaf use)))
4114 (home (when (lambda-var-p var) (lambda-var-home var)))
4115 (info (when (lambda-var-p var) (lambda-var-arg-info var)))
4116 (restp (when info (eq :rest (arg-info-kind info)))))
4117 (flet ((ref-good-for-more-context-p (ref)
4118 (let ((dest (principal-lvar-end (node-lvar ref))))
4119 (and (combination-p dest)
4120 ;; If the destination is to anything but these, we're going to
4121 ;; actually need the rest list -- and since other operations
4122 ;; might modify the list destructively, the using the context
4123 ;; isn't good anywhere else either.
4124 (lvar-fun-is (combination-fun dest)
4125 '(%rest-values %rest-ref %rest-length
4126 %rest-null %rest-true))
4127 ;; If the home lambda is different and isn't DX, it might
4128 ;; escape -- in which case using the more context isn't safe.
4129 (let ((clambda (node-home-lambda dest)))
4130 (or (eq home clambda)
4131 (leaf-dynamic-extent clambda)))))))
4132 (let ((ok (and restp
4133 (consp (arg-info-default info))
4134 (not (lambda-var-specvar var))
4135 (not (lambda-var-sets var))
4136 (every #'ref-good-for-more-context-p (lambda-var-refs var)))))
4138 (mark-more-context-used var)
4140 (mark-more-context-invalid var)))
4143 ;;; VALUES-LIST -> %REST-VALUES
4144 (define-source-transform values-list (list)
4145 (multiple-value-bind (context count) (possible-rest-arg-context list)
4147 `(%rest-values ,list ,context ,count)
4150 ;;; NTH -> %REST-REF
4151 (define-source-transform nth (n list)
4152 (multiple-value-bind (context count) (possible-rest-arg-context list)
4154 `(%rest-ref ,n ,list ,context ,count)
4155 `(car (nthcdr ,n ,list)))))
4157 (define-source-transform elt (seq n)
4158 (if (policy *lexenv* (= safety 3))
4160 (multiple-value-bind (context count) (possible-rest-arg-context seq)
4162 `(%rest-ref ,n ,seq ,context ,count)
4165 ;;; CAxR -> %REST-REF
4166 (defun source-transform-car (list nth)
4167 (multiple-value-bind (context count) (possible-rest-arg-context list)
4169 `(%rest-ref ,nth ,list ,context ,count)
4172 (define-source-transform car (list)
4173 (source-transform-car list 0))
4175 (define-source-transform cadr (list)
4176 (or (source-transform-car list 1)
4177 `(car (cdr ,list))))
4179 (define-source-transform caddr (list)
4180 (or (source-transform-car list 2)
4181 `(car (cdr (cdr ,list)))))
4183 (define-source-transform cadddr (list)
4184 (or (source-transform-car list 3)
4185 `(car (cdr (cdr (cdr ,list))))))
4187 ;;; LENGTH -> %REST-LENGTH
4188 (defun source-transform-length (list)
4189 (multiple-value-bind (context count) (possible-rest-arg-context list)
4191 `(%rest-length ,list ,context ,count)
4193 (define-source-transform length (list) (source-transform-length list))
4194 (define-source-transform list-length (list) (source-transform-length list))
4196 ;;; ENDP, NULL and NOT -> %REST-NULL
4198 ;;; Outside &REST convert into an IF so that IF optimizations will eliminate
4199 ;;; redundant negations.
4200 (defun source-transform-null (x op)
4201 (multiple-value-bind (context count) (possible-rest-arg-context x)
4203 `(%rest-null ',op ,x ,context ,count))
4205 `(if (the list ,x) nil t))
4208 (define-source-transform not (x) (source-transform-null x 'not))
4209 (define-source-transform null (x) (source-transform-null x 'null))
4210 (define-source-transform endp (x) (source-transform-null x 'endp))
4212 (deftransform %rest-values ((list context count))
4213 (if (rest-var-more-context-ok list)
4214 `(%more-arg-values context 0 count)
4215 `(values-list list)))
4217 (deftransform %rest-ref ((n list context count))
4218 (cond ((rest-var-more-context-ok list)
4219 `(and (< (the index n) count)
4220 (%more-arg context n)))
4221 ((and (constant-lvar-p n) (zerop (lvar-value n)))
4226 (deftransform %rest-length ((list context count))
4227 (if (rest-var-more-context-ok list)
4231 (deftransform %rest-null ((op list context count))
4232 (aver (constant-lvar-p op))
4233 (if (rest-var-more-context-ok list)
4235 `(,(lvar-value op) list)))
4237 (deftransform %rest-true ((list context count))
4238 (if (rest-var-more-context-ok list)
4239 `(not (eql 0 count))
4242 ;;;; transforming FORMAT
4244 ;;;; If the control string is a compile-time constant, then replace it
4245 ;;;; with a use of the FORMATTER macro so that the control string is
4246 ;;;; ``compiled.'' Furthermore, if the destination is either a stream
4247 ;;;; or T and the control string is a function (i.e. FORMATTER), then
4248 ;;;; convert the call to FORMAT to just a FUNCALL of that function.
4250 ;;; for compile-time argument count checking.
4252 ;;; FIXME II: In some cases, type information could be correlated; for
4253 ;;; instance, ~{ ... ~} requires a list argument, so if the lvar-type
4254 ;;; of a corresponding argument is known and does not intersect the
4255 ;;; list type, a warning could be signalled.
4256 (defun check-format-args (string args fun)
4257 (declare (type string string))
4258 (unless (typep string 'simple-string)
4259 (setq string (coerce string 'simple-string)))
4260 (multiple-value-bind (min max)
4261 (handler-case (sb!format:%compiler-walk-format-string string args)
4262 (sb!format:format-error (c)
4263 (compiler-warn "~A" c)))
4265 (let ((nargs (length args)))
4268 (warn 'format-too-few-args-warning
4270 "Too few arguments (~D) to ~S ~S: requires at least ~D."
4271 :format-arguments (list nargs fun string min)))
4273 (warn 'format-too-many-args-warning
4275 "Too many arguments (~D) to ~S ~S: uses at most ~D."
4276 :format-arguments (list nargs fun string max))))))))
4278 (defoptimizer (format optimizer) ((dest control &rest args))
4279 (when (constant-lvar-p control)
4280 (let ((x (lvar-value control)))
4282 (check-format-args x args 'format)))))
4284 ;;; We disable this transform in the cross-compiler to save memory in
4285 ;;; the target image; most of the uses of FORMAT in the compiler are for
4286 ;;; error messages, and those don't need to be particularly fast.
4288 (deftransform format ((dest control &rest args) (t simple-string &rest t) *
4289 :policy (>= speed space))
4290 (unless (constant-lvar-p control)
4291 (give-up-ir1-transform "The control string is not a constant."))
4292 (let ((arg-names (make-gensym-list (length args))))
4293 `(lambda (dest control ,@arg-names)
4294 (declare (ignore control))
4295 (format dest (formatter ,(lvar-value control)) ,@arg-names))))
4297 (deftransform format ((stream control &rest args) (stream function &rest t))
4298 (let ((arg-names (make-gensym-list (length args))))
4299 `(lambda (stream control ,@arg-names)
4300 (funcall control stream ,@arg-names)
4303 (deftransform format ((tee control &rest args) ((member t) function &rest t))
4304 (let ((arg-names (make-gensym-list (length args))))
4305 `(lambda (tee control ,@arg-names)
4306 (declare (ignore tee))
4307 (funcall control *standard-output* ,@arg-names)
4310 (deftransform pathname ((pathspec) (pathname) *)
4313 (deftransform pathname ((pathspec) (string) *)
4314 '(values (parse-namestring pathspec)))
4318 `(defoptimizer (,name optimizer) ((control &rest args))
4319 (when (constant-lvar-p control)
4320 (let ((x (lvar-value control)))
4322 (check-format-args x args ',name)))))))
4325 #+sb-xc-host ; Only we should be using these
4328 (def compiler-error)
4330 (def compiler-style-warn)
4331 (def compiler-notify)
4332 (def maybe-compiler-notify)
4335 (defoptimizer (cerror optimizer) ((report control &rest args))
4336 (when (and (constant-lvar-p control)
4337 (constant-lvar-p report))
4338 (let ((x (lvar-value control))
4339 (y (lvar-value report)))
4340 (when (and (stringp x) (stringp y))
4341 (multiple-value-bind (min1 max1)
4343 (sb!format:%compiler-walk-format-string x args)
4344 (sb!format:format-error (c)
4345 (compiler-warn "~A" c)))
4347 (multiple-value-bind (min2 max2)
4349 (sb!format:%compiler-walk-format-string y args)
4350 (sb!format:format-error (c)
4351 (compiler-warn "~A" c)))
4353 (let ((nargs (length args)))
4355 ((< nargs (min min1 min2))
4356 (warn 'format-too-few-args-warning
4358 "Too few arguments (~D) to ~S ~S ~S: ~
4359 requires at least ~D."
4361 (list nargs 'cerror y x (min min1 min2))))
4362 ((> nargs (max max1 max2))
4363 (warn 'format-too-many-args-warning
4365 "Too many arguments (~D) to ~S ~S ~S: ~
4368 (list nargs 'cerror y x (max max1 max2))))))))))))))
4370 (defoptimizer (coerce derive-type) ((value type) node)
4372 ((constant-lvar-p type)
4373 ;; This branch is essentially (RESULT-TYPE-SPECIFIER-NTH-ARG 2),
4374 ;; but dealing with the niggle that complex canonicalization gets
4375 ;; in the way: (COERCE 1 'COMPLEX) returns 1, which is not of
4377 (let* ((specifier (lvar-value type))
4378 (result-typeoid (careful-specifier-type specifier)))
4380 ((null result-typeoid) nil)
4381 ((csubtypep result-typeoid (specifier-type 'number))
4382 ;; the difficult case: we have to cope with ANSI 12.1.5.3
4383 ;; Rule of Canonical Representation for Complex Rationals,
4384 ;; which is a truly nasty delivery to field.
4386 ((csubtypep result-typeoid (specifier-type 'real))
4387 ;; cleverness required here: it would be nice to deduce
4388 ;; that something of type (INTEGER 2 3) coerced to type
4389 ;; DOUBLE-FLOAT should return (DOUBLE-FLOAT 2.0d0 3.0d0).
4390 ;; FLOAT gets its own clause because it's implemented as
4391 ;; a UNION-TYPE, so we don't catch it in the NUMERIC-TYPE
4394 ((and (numeric-type-p result-typeoid)
4395 (eq (numeric-type-complexp result-typeoid) :real))
4396 ;; FIXME: is this clause (a) necessary or (b) useful?
4398 ((or (csubtypep result-typeoid
4399 (specifier-type '(complex single-float)))
4400 (csubtypep result-typeoid
4401 (specifier-type '(complex double-float)))
4403 (csubtypep result-typeoid
4404 (specifier-type '(complex long-float))))
4405 ;; float complex types are never canonicalized.
4408 ;; if it's not a REAL, or a COMPLEX FLOAToid, it's
4409 ;; probably just a COMPLEX or equivalent. So, in that
4410 ;; case, we will return a complex or an object of the
4411 ;; provided type if it's rational:
4412 (type-union result-typeoid
4413 (type-intersection (lvar-type value)
4414 (specifier-type 'rational))))))
4415 ((and (policy node (zerop safety))
4416 (csubtypep result-typeoid (specifier-type '(array * (*)))))
4417 ;; At zero safety the deftransform for COERCE can elide dimension
4418 ;; checks for the things like (COERCE X '(SIMPLE-VECTOR 5)) -- so we
4419 ;; need to simplify the type to drop the dimension information.
4420 (let ((vtype (simplify-vector-type result-typeoid)))
4422 (specifier-type vtype)
4427 ;; OK, the result-type argument isn't constant. However, there
4428 ;; are common uses where we can still do better than just
4429 ;; *UNIVERSAL-TYPE*: e.g. (COERCE X (ARRAY-ELEMENT-TYPE Y)),
4430 ;; where Y is of a known type. See messages on cmucl-imp
4431 ;; 2001-02-14 and sbcl-devel 2002-12-12. We only worry here
4432 ;; about types that can be returned by (ARRAY-ELEMENT-TYPE Y), on
4433 ;; the basis that it's unlikely that other uses are both
4434 ;; time-critical and get to this branch of the COND (non-constant
4435 ;; second argument to COERCE). -- CSR, 2002-12-16
4436 (let ((value-type (lvar-type value))
4437 (type-type (lvar-type type)))
4439 ((good-cons-type-p (cons-type)
4440 ;; Make sure the cons-type we're looking at is something
4441 ;; we're prepared to handle which is basically something
4442 ;; that array-element-type can return.
4443 (or (and (member-type-p cons-type)
4444 (eql 1 (member-type-size cons-type))
4445 (null (first (member-type-members cons-type))))
4446 (let ((car-type (cons-type-car-type cons-type)))
4447 (and (member-type-p car-type)
4448 (eql 1 (member-type-members car-type))
4449 (let ((elt (first (member-type-members car-type))))
4453 (numberp (first elt)))))
4454 (good-cons-type-p (cons-type-cdr-type cons-type))))))
4455 (unconsify-type (good-cons-type)
4456 ;; Convert the "printed" respresentation of a cons
4457 ;; specifier into a type specifier. That is, the
4458 ;; specifier (CONS (EQL SIGNED-BYTE) (CONS (EQL 16)
4459 ;; NULL)) is converted to (SIGNED-BYTE 16).
4460 (cond ((or (null good-cons-type)
4461 (eq good-cons-type 'null))
4463 ((and (eq (first good-cons-type) 'cons)
4464 (eq (first (second good-cons-type)) 'member))
4465 `(,(second (second good-cons-type))
4466 ,@(unconsify-type (caddr good-cons-type))))))
4467 (coerceable-p (part)
4468 ;; Can the value be coerced to the given type? Coerce is
4469 ;; complicated, so we don't handle every possible case
4470 ;; here---just the most common and easiest cases:
4472 ;; * Any REAL can be coerced to a FLOAT type.
4473 ;; * Any NUMBER can be coerced to a (COMPLEX
4474 ;; SINGLE/DOUBLE-FLOAT).
4476 ;; FIXME I: we should also be able to deal with characters
4479 ;; FIXME II: I'm not sure that anything is necessary
4480 ;; here, at least while COMPLEX is not a specialized
4481 ;; array element type in the system. Reasoning: if
4482 ;; something cannot be coerced to the requested type, an
4483 ;; error will be raised (and so any downstream compiled
4484 ;; code on the assumption of the returned type is
4485 ;; unreachable). If something can, then it will be of
4486 ;; the requested type, because (by assumption) COMPLEX
4487 ;; (and other difficult types like (COMPLEX INTEGER)
4488 ;; aren't specialized types.
4489 (let ((coerced-type (careful-specifier-type part)))
4491 (or (and (csubtypep coerced-type (specifier-type 'float))
4492 (csubtypep value-type (specifier-type 'real)))
4493 (and (csubtypep coerced-type
4494 (specifier-type `(or (complex single-float)
4495 (complex double-float))))
4496 (csubtypep value-type (specifier-type 'number)))))))
4497 (process-types (type)
4498 ;; FIXME: This needs some work because we should be able
4499 ;; to derive the resulting type better than just the
4500 ;; type arg of coerce. That is, if X is (INTEGER 10
4501 ;; 20), then (COERCE X 'DOUBLE-FLOAT) should say
4502 ;; (DOUBLE-FLOAT 10d0 20d0) instead of just
4504 (cond ((member-type-p type)
4507 (mapc-member-type-members
4509 (if (coerceable-p member)
4510 (push member members)
4511 (return-from punt *universal-type*)))
4513 (specifier-type `(or ,@members)))))
4514 ((and (cons-type-p type)
4515 (good-cons-type-p type))
4516 (let ((c-type (unconsify-type (type-specifier type))))
4517 (if (coerceable-p c-type)
4518 (specifier-type c-type)
4521 *universal-type*))))
4522 (cond ((union-type-p type-type)
4523 (apply #'type-union (mapcar #'process-types
4524 (union-type-types type-type))))
4525 ((or (member-type-p type-type)
4526 (cons-type-p type-type))
4527 (process-types type-type))
4529 *universal-type*)))))))
4531 (defoptimizer (compile derive-type) ((nameoid function))
4532 (when (csubtypep (lvar-type nameoid)
4533 (specifier-type 'null))
4534 (values-specifier-type '(values function boolean boolean))))
4536 ;;; FIXME: Maybe also STREAM-ELEMENT-TYPE should be given some loving
4537 ;;; treatment along these lines? (See discussion in COERCE DERIVE-TYPE
4538 ;;; optimizer, above).
4539 (defoptimizer (array-element-type derive-type) ((array))
4540 (let ((array-type (lvar-type array)))
4541 (labels ((consify (list)
4544 `(cons (eql ,(car list)) ,(consify (rest list)))))
4545 (get-element-type (a)
4547 (type-specifier (array-type-specialized-element-type a))))
4548 (cond ((eq element-type '*)
4549 (specifier-type 'type-specifier))
4550 ((symbolp element-type)
4551 (make-member-type :members (list element-type)))
4552 ((consp element-type)
4553 (specifier-type (consify element-type)))
4555 (error "can't understand type ~S~%" element-type))))))
4556 (labels ((recurse (type)
4557 (cond ((array-type-p type)
4558 (get-element-type type))
4559 ((union-type-p type)
4561 (mapcar #'recurse (union-type-types type))))
4563 *universal-type*))))
4564 (recurse array-type)))))
4566 (define-source-transform sb!impl::sort-vector (vector start end predicate key)
4567 ;; Like CMU CL, we use HEAPSORT. However, other than that, this code
4568 ;; isn't really related to the CMU CL code, since instead of trying
4569 ;; to generalize the CMU CL code to allow START and END values, this
4570 ;; code has been written from scratch following Chapter 7 of
4571 ;; _Introduction to Algorithms_ by Corman, Rivest, and Shamir.
4572 `(macrolet ((%index (x) `(truly-the index ,x))
4573 (%parent (i) `(ash ,i -1))
4574 (%left (i) `(%index (ash ,i 1)))
4575 (%right (i) `(%index (1+ (ash ,i 1))))
4578 (left (%left i) (%left i)))
4579 ((> left current-heap-size))
4580 (declare (type index i left))
4581 (let* ((i-elt (%elt i))
4582 (i-key (funcall keyfun i-elt))
4583 (left-elt (%elt left))
4584 (left-key (funcall keyfun left-elt)))
4585 (multiple-value-bind (large large-elt large-key)
4586 (if (funcall ,',predicate i-key left-key)
4587 (values left left-elt left-key)
4588 (values i i-elt i-key))
4589 (let ((right (%right i)))
4590 (multiple-value-bind (largest largest-elt)
4591 (if (> right current-heap-size)
4592 (values large large-elt)
4593 (let* ((right-elt (%elt right))
4594 (right-key (funcall keyfun right-elt)))
4595 (if (funcall ,',predicate large-key right-key)
4596 (values right right-elt)
4597 (values large large-elt))))
4598 (cond ((= largest i)
4601 (setf (%elt i) largest-elt
4602 (%elt largest) i-elt
4604 (%sort-vector (keyfun &optional (vtype 'vector))
4605 `(macrolet (;; KLUDGE: In SBCL ca. 0.6.10, I had
4606 ;; trouble getting type inference to
4607 ;; propagate all the way through this
4608 ;; tangled mess of inlining. The TRULY-THE
4609 ;; here works around that. -- WHN
4611 `(aref (truly-the ,',vtype ,',',vector)
4612 (%index (+ (%index ,i) start-1)))))
4613 (let (;; Heaps prefer 1-based addressing.
4614 (start-1 (1- ,',start))
4615 (current-heap-size (- ,',end ,',start))
4617 (declare (type (integer -1 #.(1- sb!xc:most-positive-fixnum))
4619 (declare (type index current-heap-size))
4620 (declare (type function keyfun))
4621 (loop for i of-type index
4622 from (ash current-heap-size -1) downto 1 do
4625 (when (< current-heap-size 2)
4627 (rotatef (%elt 1) (%elt current-heap-size))
4628 (decf current-heap-size)
4630 (if (typep ,vector 'simple-vector)
4631 ;; (VECTOR T) is worth optimizing for, and SIMPLE-VECTOR is
4632 ;; what we get from (VECTOR T) inside WITH-ARRAY-DATA.
4634 ;; Special-casing the KEY=NIL case lets us avoid some
4636 (%sort-vector #'identity simple-vector)
4637 (%sort-vector ,key simple-vector))
4638 ;; It's hard to anticipate many speed-critical applications for
4639 ;; sorting vector types other than (VECTOR T), so we just lump
4640 ;; them all together in one slow dynamically typed mess.
4642 (declare (optimize (speed 2) (space 2) (inhibit-warnings 3)))
4643 (%sort-vector (or ,key #'identity))))))
4645 ;;;; debuggers' little helpers
4647 ;;; for debugging when transforms are behaving mysteriously,
4648 ;;; e.g. when debugging a problem with an ASH transform
4649 ;;; (defun foo (&optional s)
4650 ;;; (sb-c::/report-lvar s "S outside WHEN")
4651 ;;; (when (and (integerp s) (> s 3))
4652 ;;; (sb-c::/report-lvar s "S inside WHEN")
4653 ;;; (let ((bound (ash 1 (1- s))))
4654 ;;; (sb-c::/report-lvar bound "BOUND")
4655 ;;; (let ((x (- bound))
4657 ;;; (sb-c::/report-lvar x "X")
4658 ;;; (sb-c::/report-lvar x "Y"))
4659 ;;; `(integer ,(- bound) ,(1- bound)))))
4660 ;;; (The DEFTRANSFORM doesn't do anything but report at compile time,
4661 ;;; and the function doesn't do anything at all.)
4664 (defknown /report-lvar (t t) null)
4665 (deftransform /report-lvar ((x message) (t t))
4666 (format t "~%/in /REPORT-LVAR~%")
4667 (format t "/(LVAR-TYPE X)=~S~%" (lvar-type x))
4668 (when (constant-lvar-p x)
4669 (format t "/(LVAR-VALUE X)=~S~%" (lvar-value x)))
4670 (format t "/MESSAGE=~S~%" (lvar-value message))
4671 (give-up-ir1-transform "not a real transform"))
4672 (defun /report-lvar (x message)
4673 (declare (ignore x message))))
4676 ;;;; Transforms for internal compiler utilities
4678 ;;; If QUALITY-NAME is constant and a valid name, don't bother
4679 ;;; checking that it's still valid at run-time.
4680 (deftransform policy-quality ((policy quality-name)
4682 (unless (and (constant-lvar-p quality-name)
4683 (policy-quality-name-p (lvar-value quality-name)))
4684 (give-up-ir1-transform))
4685 '(%policy-quality policy quality-name))