1 ;;;; This file contains floating-point-specific transforms, and may be
2 ;;;; somewhat implementation-dependent in its assumptions of what the
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.
18 (defknown %single-float (real) single-float (movable foldable flushable))
19 (defknown %double-float (real) double-float (movable foldable flushable))
21 (deftransform float ((n f) (* single-float) *)
24 (deftransform float ((n f) (* double-float) *)
27 (deftransform float ((n) *)
32 (deftransform %single-float ((n) (single-float) *)
35 (deftransform %double-float ((n) (double-float) *)
39 (macrolet ((frob (fun type)
40 `(deftransform random ((num &optional state)
41 (,type &optional *) *)
42 "Use inline float operations."
43 '(,fun num (or state *random-state*)))))
44 (frob %random-single-float single-float)
45 (frob %random-double-float double-float))
47 ;;; Mersenne Twister RNG
49 ;;; FIXME: It's unpleasant to have RANDOM functionality scattered
50 ;;; through the code this way. It would be nice to move this into the
51 ;;; same file as the other RANDOM definitions.
52 (deftransform random ((num &optional state)
53 ((integer 1 #.(expt 2 sb!vm::n-word-bits)) &optional *))
54 ;; FIXME: I almost conditionalized this as #!+sb-doc. Find some way
55 ;; of automatically finding #!+sb-doc in proximity to DEFTRANSFORM
56 ;; to let me scan for places that I made this mistake and didn't
58 "use inline (UNSIGNED-BYTE 32) operations"
59 (let ((type (lvar-type num))
60 (limit (expt 2 sb!vm::n-word-bits))
61 (random-chunk (ecase sb!vm::n-word-bits
63 (64 'sb!kernel::big-random-chunk))))
64 (if (numeric-type-p type)
65 (let ((num-high (numeric-type-high (lvar-type num))))
67 (cond ((constant-lvar-p num)
68 ;; Check the worst case sum absolute error for the
69 ;; random number expectations.
70 (let ((rem (rem limit num-high)))
71 (unless (< (/ (* 2 rem (- num-high rem))
73 (expt 2 (- sb!kernel::random-integer-extra-bits)))
74 (give-up-ir1-transform
75 "The random number expectations are inaccurate."))
76 (if (= num-high limit)
77 `(,random-chunk (or state *random-state*))
79 `(rem (,random-chunk (or state *random-state*)) num)
81 ;; Use multiplication, which is faster.
82 `(values (sb!bignum::%multiply
83 (,random-chunk (or state *random-state*))
85 ((> num-high random-fixnum-max)
86 (give-up-ir1-transform
87 "The range is too large to ensure an accurate result."))
90 `(values (sb!bignum::%multiply
91 (,random-chunk (or state *random-state*))
94 `(rem (,random-chunk (or state *random-state*)) num))))
95 ;; KLUDGE: a relatively conservative treatment, but better
96 ;; than a bug (reported by PFD sbcl-devel towards the end of
98 '(rem (random-chunk (or state *random-state*)) num))))
102 (defknown make-single-float ((signed-byte 32)) single-float
103 (movable foldable flushable))
105 (defknown make-double-float ((signed-byte 32) (unsigned-byte 32)) double-float
106 (movable foldable flushable))
108 (defknown single-float-bits (single-float) (signed-byte 32)
109 (movable foldable flushable))
111 (defknown double-float-high-bits (double-float) (signed-byte 32)
112 (movable foldable flushable))
114 (defknown double-float-low-bits (double-float) (unsigned-byte 32)
115 (movable foldable flushable))
117 (deftransform float-sign ((float &optional float2)
118 (single-float &optional single-float) *)
120 (let ((temp (gensym)))
121 `(let ((,temp (abs float2)))
122 (if (minusp (single-float-bits float)) (- ,temp) ,temp)))
123 '(if (minusp (single-float-bits float)) -1f0 1f0)))
125 (deftransform float-sign ((float &optional float2)
126 (double-float &optional double-float) *)
128 (let ((temp (gensym)))
129 `(let ((,temp (abs float2)))
130 (if (minusp (double-float-high-bits float)) (- ,temp) ,temp)))
131 '(if (minusp (double-float-high-bits float)) -1d0 1d0)))
133 ;;;; DECODE-FLOAT, INTEGER-DECODE-FLOAT, and SCALE-FLOAT
135 (defknown decode-single-float (single-float)
136 (values single-float single-float-exponent (single-float -1f0 1f0))
137 (movable foldable flushable))
139 (defknown decode-double-float (double-float)
140 (values double-float double-float-exponent (double-float -1d0 1d0))
141 (movable foldable flushable))
143 (defknown integer-decode-single-float (single-float)
144 (values single-float-significand single-float-int-exponent (integer -1 1))
145 (movable foldable flushable))
147 (defknown integer-decode-double-float (double-float)
148 (values double-float-significand double-float-int-exponent (integer -1 1))
149 (movable foldable flushable))
151 (defknown scale-single-float (single-float integer) single-float
152 (movable foldable flushable))
154 (defknown scale-double-float (double-float integer) double-float
155 (movable foldable flushable))
157 (deftransform decode-float ((x) (single-float) *)
158 '(decode-single-float x))
160 (deftransform decode-float ((x) (double-float) *)
161 '(decode-double-float x))
163 (deftransform integer-decode-float ((x) (single-float) *)
164 '(integer-decode-single-float x))
166 (deftransform integer-decode-float ((x) (double-float) *)
167 '(integer-decode-double-float x))
169 (deftransform scale-float ((f ex) (single-float *) *)
170 (if (and #!+x86 t #!-x86 nil
171 (csubtypep (lvar-type ex)
172 (specifier-type '(signed-byte 32))))
173 '(coerce (%scalbn (coerce f 'double-float) ex) 'single-float)
174 '(scale-single-float f ex)))
176 (deftransform scale-float ((f ex) (double-float *) *)
177 (if (and #!+x86 t #!-x86 nil
178 (csubtypep (lvar-type ex)
179 (specifier-type '(signed-byte 32))))
181 '(scale-double-float f ex)))
183 ;;; What is the CROSS-FLOAT-INFINITY-KLUDGE?
185 ;;; SBCL's own implementation of floating point supports floating
186 ;;; point infinities. Some of the old CMU CL :PROPAGATE-FLOAT-TYPE and
187 ;;; :PROPAGATE-FUN-TYPE code, like the DEFOPTIMIZERs below, uses this
188 ;;; floating point support. Thus, we have to avoid running it on the
189 ;;; cross-compilation host, since we're not guaranteed that the
190 ;;; cross-compilation host will support floating point infinities.
192 ;;; If we wanted to live dangerously, we could conditionalize the code
193 ;;; with #+(OR SBCL SB-XC) instead. That way, if the cross-compilation
194 ;;; host happened to be SBCL, we'd be able to run the infinity-using
196 ;;; * SBCL itself gets built with more complete optimization.
198 ;;; * You get a different SBCL depending on what your cross-compilation
200 ;;; So far the pros and cons seem seem to be mostly academic, since
201 ;;; AFAIK (WHN 2001-08-28) the propagate-foo-type optimizations aren't
202 ;;; actually important in compiling SBCL itself. If this changes, then
203 ;;; we have to decide:
204 ;;; * Go for simplicity, leaving things as they are.
205 ;;; * Go for performance at the expense of conceptual clarity,
206 ;;; using #+(OR SBCL SB-XC) and otherwise leaving the build
208 ;;; * Go for performance at the expense of build time, using
209 ;;; #+(OR SBCL SB-XC) and also making SBCL do not just
210 ;;; make-host-1.sh and make-host-2.sh, but a third step
211 ;;; make-host-3.sh where it builds itself under itself. (Such a
212 ;;; 3-step build process could also help with other things, e.g.
213 ;;; using specialized arrays to represent debug information.)
214 ;;; * Rewrite the code so that it doesn't depend on unportable
215 ;;; floating point infinities.
217 ;;; optimizers for SCALE-FLOAT. If the float has bounds, new bounds
218 ;;; are computed for the result, if possible.
219 #-sb-xc-host ; (See CROSS-FLOAT-INFINITY-KLUDGE.)
222 (defun scale-float-derive-type-aux (f ex same-arg)
223 (declare (ignore same-arg))
224 (flet ((scale-bound (x n)
225 ;; We need to be a bit careful here and catch any overflows
226 ;; that might occur. We can ignore underflows which become
230 (scale-float (type-bound-number x) n)
231 (floating-point-overflow ()
234 (when (and (numeric-type-p f) (numeric-type-p ex))
235 (let ((f-lo (numeric-type-low f))
236 (f-hi (numeric-type-high f))
237 (ex-lo (numeric-type-low ex))
238 (ex-hi (numeric-type-high ex))
242 (if (< (float-sign (type-bound-number f-hi)) 0.0)
244 (setf new-hi (scale-bound f-hi ex-lo)))
246 (setf new-hi (scale-bound f-hi ex-hi)))))
248 (if (< (float-sign (type-bound-number f-lo)) 0.0)
250 (setf new-lo (scale-bound f-lo ex-hi)))
252 (setf new-lo (scale-bound f-lo ex-lo)))))
253 (make-numeric-type :class (numeric-type-class f)
254 :format (numeric-type-format f)
258 (defoptimizer (scale-single-float derive-type) ((f ex))
259 (two-arg-derive-type f ex #'scale-float-derive-type-aux
260 #'scale-single-float t))
261 (defoptimizer (scale-double-float derive-type) ((f ex))
262 (two-arg-derive-type f ex #'scale-float-derive-type-aux
263 #'scale-double-float t))
265 ;;; DEFOPTIMIZERs for %SINGLE-FLOAT and %DOUBLE-FLOAT. This makes the
266 ;;; FLOAT function return the correct ranges if the input has some
267 ;;; defined range. Quite useful if we want to convert some type of
268 ;;; bounded integer into a float.
271 (let ((aux-name (symbolicate fun "-DERIVE-TYPE-AUX")))
273 (defun ,aux-name (num)
274 ;; When converting a number to a float, the limits are
276 (let* ((lo (bound-func (lambda (x)
278 (numeric-type-low num)))
279 (hi (bound-func (lambda (x)
281 (numeric-type-high num))))
282 (specifier-type `(,',type ,(or lo '*) ,(or hi '*)))))
284 (defoptimizer (,fun derive-type) ((num))
285 (one-arg-derive-type num #',aux-name #',fun))))))
286 (frob %single-float single-float)
287 (frob %double-float double-float))
292 ;;; Do some stuff to recognize when the loser is doing mixed float and
293 ;;; rational arithmetic, or different float types, and fix it up. If
294 ;;; we don't, he won't even get so much as an efficiency note.
295 (deftransform float-contagion-arg1 ((x y) * * :defun-only t :node node)
296 `(,(lvar-fun-name (basic-combination-fun node))
298 (deftransform float-contagion-arg2 ((x y) * * :defun-only t :node node)
299 `(,(lvar-fun-name (basic-combination-fun node))
302 (dolist (x '(+ * / -))
303 (%deftransform x '(function (rational float) *) #'float-contagion-arg1)
304 (%deftransform x '(function (float rational) *) #'float-contagion-arg2))
306 (dolist (x '(= < > + * / -))
307 (%deftransform x '(function (single-float double-float) *)
308 #'float-contagion-arg1)
309 (%deftransform x '(function (double-float single-float) *)
310 #'float-contagion-arg2))
312 ;;; Prevent ZEROP, PLUSP, and MINUSP from losing horribly. We can't in
313 ;;; general float rational args to comparison, since Common Lisp
314 ;;; semantics says we are supposed to compare as rationals, but we can
315 ;;; do it for any rational that has a precise representation as a
316 ;;; float (such as 0).
317 (macrolet ((frob (op)
318 `(deftransform ,op ((x y) (float rational) *)
319 "open-code FLOAT to RATIONAL comparison"
320 (unless (constant-lvar-p y)
321 (give-up-ir1-transform
322 "The RATIONAL value isn't known at compile time."))
323 (let ((val (lvar-value y)))
324 (unless (eql (rational (float val)) val)
325 (give-up-ir1-transform
326 "~S doesn't have a precise float representation."
328 `(,',op x (float y x)))))
333 ;;;; irrational derive-type methods
335 ;;; Derive the result to be float for argument types in the
336 ;;; appropriate domain.
337 #+sb-xc-host ; (See CROSS-FLOAT-INFINITY-KLUDGE.)
338 (dolist (stuff '((asin (real -1.0 1.0))
339 (acos (real -1.0 1.0))
341 (atanh (real -1.0 1.0))
343 (destructuring-bind (name type) stuff
344 (let ((type (specifier-type type)))
345 (setf (fun-info-derive-type (fun-info-or-lose name))
347 (declare (type combination call))
348 (when (csubtypep (lvar-type
349 (first (combination-args call)))
351 (specifier-type 'float)))))))
353 #+sb-xc-host ; (See CROSS-FLOAT-INFINITY-KLUDGE.)
354 (defoptimizer (log derive-type) ((x &optional y))
355 (when (and (csubtypep (lvar-type x)
356 (specifier-type '(real 0.0)))
358 (csubtypep (lvar-type y)
359 (specifier-type '(real 0.0)))))
360 (specifier-type 'float)))
362 ;;;; irrational transforms
364 (defknown (%tan %sinh %asinh %atanh %log %logb %log10 %tan-quick)
365 (double-float) double-float
366 (movable foldable flushable))
368 (defknown (%sin %cos %tanh %sin-quick %cos-quick)
369 (double-float) (double-float -1.0d0 1.0d0)
370 (movable foldable flushable))
372 (defknown (%asin %atan)
374 (double-float #.(coerce (- (/ pi 2)) 'double-float)
375 #.(coerce (/ pi 2) 'double-float))
376 (movable foldable flushable))
379 (double-float) (double-float 0.0d0 #.(coerce pi 'double-float))
380 (movable foldable flushable))
383 (double-float) (double-float 1.0d0)
384 (movable foldable flushable))
386 (defknown (%acosh %exp %sqrt)
387 (double-float) (double-float 0.0d0)
388 (movable foldable flushable))
391 (double-float) (double-float -1d0)
392 (movable foldable flushable))
395 (double-float double-float) (double-float 0d0)
396 (movable foldable flushable))
399 (double-float double-float) double-float
400 (movable foldable flushable))
403 (double-float double-float)
404 (double-float #.(coerce (- pi) 'double-float)
405 #.(coerce pi 'double-float))
406 (movable foldable flushable))
409 (double-float double-float) double-float
410 (movable foldable flushable))
413 (double-float (signed-byte 32)) double-float
414 (movable foldable flushable))
417 (double-float) double-float
418 (movable foldable flushable))
420 (macrolet ((def (name prim rtype)
422 (deftransform ,name ((x) (single-float) ,rtype)
423 `(coerce (,',prim (coerce x 'double-float)) 'single-float))
424 (deftransform ,name ((x) (double-float) ,rtype)
428 (def sqrt %sqrt float)
429 (def asin %asin float)
430 (def acos %acos float)
436 (def acosh %acosh float)
437 (def atanh %atanh float))
439 ;;; The argument range is limited on the x86 FP trig. functions. A
440 ;;; post-test can detect a failure (and load a suitable result), but
441 ;;; this test is avoided if possible.
442 (macrolet ((def (name prim prim-quick)
443 (declare (ignorable prim-quick))
445 (deftransform ,name ((x) (single-float) *)
446 #!+x86 (cond ((csubtypep (lvar-type x)
447 (specifier-type '(single-float
448 (#.(- (expt 2f0 64)))
450 `(coerce (,',prim-quick (coerce x 'double-float))
454 "unable to avoid inline argument range check~@
455 because the argument range (~S) was not within 2^64"
456 (type-specifier (lvar-type x)))
457 `(coerce (,',prim (coerce x 'double-float)) 'single-float)))
458 #!-x86 `(coerce (,',prim (coerce x 'double-float)) 'single-float))
459 (deftransform ,name ((x) (double-float) *)
460 #!+x86 (cond ((csubtypep (lvar-type x)
461 (specifier-type '(double-float
462 (#.(- (expt 2d0 64)))
467 "unable to avoid inline argument range check~@
468 because the argument range (~S) was not within 2^64"
469 (type-specifier (lvar-type x)))
471 #!-x86 `(,',prim x)))))
472 (def sin %sin %sin-quick)
473 (def cos %cos %cos-quick)
474 (def tan %tan %tan-quick))
476 (deftransform atan ((x y) (single-float single-float) *)
477 `(coerce (%atan2 (coerce x 'double-float) (coerce y 'double-float))
479 (deftransform atan ((x y) (double-float double-float) *)
482 (deftransform expt ((x y) ((single-float 0f0) single-float) *)
483 `(coerce (%pow (coerce x 'double-float) (coerce y 'double-float))
485 (deftransform expt ((x y) ((double-float 0d0) double-float) *)
487 (deftransform expt ((x y) ((single-float 0f0) (signed-byte 32)) *)
488 `(coerce (%pow (coerce x 'double-float) (coerce y 'double-float))
490 (deftransform expt ((x y) ((double-float 0d0) (signed-byte 32)) *)
491 `(%pow x (coerce y 'double-float)))
493 ;;; ANSI says log with base zero returns zero.
494 (deftransform log ((x y) (float float) float)
495 '(if (zerop y) y (/ (log x) (log y))))
497 ;;; Handle some simple transformations.
499 (deftransform abs ((x) ((complex double-float)) double-float)
500 '(%hypot (realpart x) (imagpart x)))
502 (deftransform abs ((x) ((complex single-float)) single-float)
503 '(coerce (%hypot (coerce (realpart x) 'double-float)
504 (coerce (imagpart x) 'double-float))
507 (deftransform phase ((x) ((complex double-float)) double-float)
508 '(%atan2 (imagpart x) (realpart x)))
510 (deftransform phase ((x) ((complex single-float)) single-float)
511 '(coerce (%atan2 (coerce (imagpart x) 'double-float)
512 (coerce (realpart x) 'double-float))
515 (deftransform phase ((x) ((float)) float)
516 '(if (minusp (float-sign x))
520 ;;; The number is of type REAL.
521 (defun numeric-type-real-p (type)
522 (and (numeric-type-p type)
523 (eq (numeric-type-complexp type) :real)))
525 ;;; Coerce a numeric type bound to the given type while handling
526 ;;; exclusive bounds.
527 (defun coerce-numeric-bound (bound type)
530 (list (coerce (car bound) type))
531 (coerce bound type))))
533 #-sb-xc-host ; (See CROSS-FLOAT-INFINITY-KLUDGE.)
536 ;;;; optimizers for elementary functions
538 ;;;; These optimizers compute the output range of the elementary
539 ;;;; function, based on the domain of the input.
541 ;;; Generate a specifier for a complex type specialized to the same
542 ;;; type as the argument.
543 (defun complex-float-type (arg)
544 (declare (type numeric-type arg))
545 (let* ((format (case (numeric-type-class arg)
546 ((integer rational) 'single-float)
547 (t (numeric-type-format arg))))
548 (float-type (or format 'float)))
549 (specifier-type `(complex ,float-type))))
551 ;;; Compute a specifier like '(OR FLOAT (COMPLEX FLOAT)), except float
552 ;;; should be the right kind of float. Allow bounds for the float
554 (defun float-or-complex-float-type (arg &optional lo hi)
555 (declare (type numeric-type arg))
556 (let* ((format (case (numeric-type-class arg)
557 ((integer rational) 'single-float)
558 (t (numeric-type-format arg))))
559 (float-type (or format 'float))
560 (lo (coerce-numeric-bound lo float-type))
561 (hi (coerce-numeric-bound hi float-type)))
562 (specifier-type `(or (,float-type ,(or lo '*) ,(or hi '*))
563 (complex ,float-type)))))
567 (eval-when (:compile-toplevel :execute)
568 ;; So the problem with this hack is that it's actually broken. If
569 ;; the host does not have long floats, then setting *R-D-F-F* to
570 ;; LONG-FLOAT doesn't actually buy us anything. FIXME.
571 (setf *read-default-float-format*
572 #!+long-float 'long-float #!-long-float 'double-float))
573 ;;; Test whether the numeric-type ARG is within in domain specified by
574 ;;; DOMAIN-LOW and DOMAIN-HIGH, consider negative and positive zero to
576 #-sb-xc-host ; (See CROSS-FLOAT-INFINITY-KLUDGE.)
577 (defun domain-subtypep (arg domain-low domain-high)
578 (declare (type numeric-type arg)
579 (type (or real null) domain-low domain-high))
580 (let* ((arg-lo (numeric-type-low arg))
581 (arg-lo-val (type-bound-number arg-lo))
582 (arg-hi (numeric-type-high arg))
583 (arg-hi-val (type-bound-number arg-hi)))
584 ;; Check that the ARG bounds are correctly canonicalized.
585 (when (and arg-lo (floatp arg-lo-val) (zerop arg-lo-val) (consp arg-lo)
586 (minusp (float-sign arg-lo-val)))
587 (compiler-notify "float zero bound ~S not correctly canonicalized?" arg-lo)
588 (setq arg-lo 0e0 arg-lo-val arg-lo))
589 (when (and arg-hi (zerop arg-hi-val) (floatp arg-hi-val) (consp arg-hi)
590 (plusp (float-sign arg-hi-val)))
591 (compiler-notify "float zero bound ~S not correctly canonicalized?" arg-hi)
592 (setq arg-hi (ecase *read-default-float-format*
593 (double-float (load-time-value (make-unportable-float :double-float-negative-zero)))
595 (long-float (load-time-value (make-unportable-float :long-float-negative-zero))))
597 (flet ((fp-neg-zero-p (f) ; Is F -0.0?
598 (and (floatp f) (zerop f) (minusp (float-sign f))))
599 (fp-pos-zero-p (f) ; Is F +0.0?
600 (and (floatp f) (zerop f) (plusp (float-sign f)))))
601 (and (or (null domain-low)
602 (and arg-lo (>= arg-lo-val domain-low)
603 (not (and (fp-pos-zero-p domain-low)
604 (fp-neg-zero-p arg-lo)))))
605 (or (null domain-high)
606 (and arg-hi (<= arg-hi-val domain-high)
607 (not (and (fp-neg-zero-p domain-high)
608 (fp-pos-zero-p arg-hi)))))))))
609 (eval-when (:compile-toplevel :execute)
610 (setf *read-default-float-format* 'single-float))
612 #-sb-xc-host ; (See CROSS-FLOAT-INFINITY-KLUDGE.)
615 ;;; Handle monotonic functions of a single variable whose domain is
616 ;;; possibly part of the real line. ARG is the variable, FCN is the
617 ;;; function, and DOMAIN is a specifier that gives the (real) domain
618 ;;; of the function. If ARG is a subset of the DOMAIN, we compute the
619 ;;; bounds directly. Otherwise, we compute the bounds for the
620 ;;; intersection between ARG and DOMAIN, and then append a complex
621 ;;; result, which occurs for the parts of ARG not in the DOMAIN.
623 ;;; Negative and positive zero are considered distinct within
624 ;;; DOMAIN-LOW and DOMAIN-HIGH.
626 ;;; DEFAULT-LOW and DEFAULT-HIGH are the lower and upper bounds if we
627 ;;; can't compute the bounds using FCN.
628 (defun elfun-derive-type-simple (arg fcn domain-low domain-high
629 default-low default-high
630 &optional (increasingp t))
631 (declare (type (or null real) domain-low domain-high))
634 (cond ((eq (numeric-type-complexp arg) :complex)
635 (complex-float-type arg))
636 ((numeric-type-real-p arg)
637 ;; The argument is real, so let's find the intersection
638 ;; between the argument and the domain of the function.
639 ;; We compute the bounds on the intersection, and for
640 ;; everything else, we return a complex number of the
642 (multiple-value-bind (intersection difference)
643 (interval-intersection/difference (numeric-type->interval arg)
649 ;; Process the intersection.
650 (let* ((low (interval-low intersection))
651 (high (interval-high intersection))
652 (res-lo (or (bound-func fcn (if increasingp low high))
654 (res-hi (or (bound-func fcn (if increasingp high low))
656 (format (case (numeric-type-class arg)
657 ((integer rational) 'single-float)
658 (t (numeric-type-format arg))))
659 (bound-type (or format 'float))
664 :low (coerce-numeric-bound res-lo bound-type)
665 :high (coerce-numeric-bound res-hi bound-type))))
666 ;; If the ARG is a subset of the domain, we don't
667 ;; have to worry about the difference, because that
669 (if (or (null difference)
670 ;; Check whether the arg is within the domain.
671 (domain-subtypep arg domain-low domain-high))
674 (specifier-type `(complex ,bound-type))))))
676 ;; No intersection so the result must be purely complex.
677 (complex-float-type arg)))))
679 (float-or-complex-float-type arg default-low default-high))))))
682 ((frob (name domain-low domain-high def-low-bnd def-high-bnd
683 &key (increasingp t))
684 (let ((num (gensym)))
685 `(defoptimizer (,name derive-type) ((,num))
689 (elfun-derive-type-simple arg #',name
690 ,domain-low ,domain-high
691 ,def-low-bnd ,def-high-bnd
694 ;; These functions are easy because they are defined for the whole
696 (frob exp nil nil 0 nil)
697 (frob sinh nil nil nil nil)
698 (frob tanh nil nil -1 1)
699 (frob asinh nil nil nil nil)
701 ;; These functions are only defined for part of the real line. The
702 ;; condition selects the desired part of the line.
703 (frob asin -1d0 1d0 (- (/ pi 2)) (/ pi 2))
704 ;; Acos is monotonic decreasing, so we need to swap the function
705 ;; values at the lower and upper bounds of the input domain.
706 (frob acos -1d0 1d0 0 pi :increasingp nil)
707 (frob acosh 1d0 nil nil nil)
708 (frob atanh -1d0 1d0 -1 1)
709 ;; Kahan says that (sqrt -0.0) is -0.0, so use a specifier that
711 (frob sqrt (load-time-value (make-unportable-float :double-float-negative-zero)) nil 0 nil))
713 ;;; Compute bounds for (expt x y). This should be easy since (expt x
714 ;;; y) = (exp (* y (log x))). However, computations done this way
715 ;;; have too much roundoff. Thus we have to do it the hard way.
716 (defun safe-expt (x y)
718 (when (< (abs y) 10000)
723 ;;; Handle the case when x >= 1.
724 (defun interval-expt-> (x y)
725 (case (sb!c::interval-range-info y 0d0)
727 ;; Y is positive and log X >= 0. The range of exp(y * log(x)) is
728 ;; obviously non-negative. We just have to be careful for
729 ;; infinite bounds (given by nil).
730 (let ((lo (safe-expt (type-bound-number (sb!c::interval-low x))
731 (type-bound-number (sb!c::interval-low y))))
732 (hi (safe-expt (type-bound-number (sb!c::interval-high x))
733 (type-bound-number (sb!c::interval-high y)))))
734 (list (sb!c::make-interval :low (or lo 1) :high hi))))
736 ;; Y is negative and log x >= 0. The range of exp(y * log(x)) is
737 ;; obviously [0, 1]. However, underflow (nil) means 0 is the
739 (let ((lo (safe-expt (type-bound-number (sb!c::interval-high x))
740 (type-bound-number (sb!c::interval-low y))))
741 (hi (safe-expt (type-bound-number (sb!c::interval-low x))
742 (type-bound-number (sb!c::interval-high y)))))
743 (list (sb!c::make-interval :low (or lo 0) :high (or hi 1)))))
745 ;; Split the interval in half.
746 (destructuring-bind (y- y+)
747 (sb!c::interval-split 0 y t)
748 (list (interval-expt-> x y-)
749 (interval-expt-> x y+))))))
751 ;;; Handle the case when x <= 1
752 (defun interval-expt-< (x y)
753 (case (sb!c::interval-range-info x 0d0)
755 ;; The case of 0 <= x <= 1 is easy
756 (case (sb!c::interval-range-info y)
758 ;; Y is positive and log X <= 0. The range of exp(y * log(x)) is
759 ;; obviously [0, 1]. We just have to be careful for infinite bounds
761 (let ((lo (safe-expt (type-bound-number (sb!c::interval-low x))
762 (type-bound-number (sb!c::interval-high y))))
763 (hi (safe-expt (type-bound-number (sb!c::interval-high x))
764 (type-bound-number (sb!c::interval-low y)))))
765 (list (sb!c::make-interval :low (or lo 0) :high (or hi 1)))))
767 ;; Y is negative and log x <= 0. The range of exp(y * log(x)) is
768 ;; obviously [1, inf].
769 (let ((hi (safe-expt (type-bound-number (sb!c::interval-low x))
770 (type-bound-number (sb!c::interval-low y))))
771 (lo (safe-expt (type-bound-number (sb!c::interval-high x))
772 (type-bound-number (sb!c::interval-high y)))))
773 (list (sb!c::make-interval :low (or lo 1) :high hi))))
775 ;; Split the interval in half
776 (destructuring-bind (y- y+)
777 (sb!c::interval-split 0 y t)
778 (list (interval-expt-< x y-)
779 (interval-expt-< x y+))))))
781 ;; The case where x <= 0. Y MUST be an INTEGER for this to work!
782 ;; The calling function must insure this! For now we'll just
783 ;; return the appropriate unbounded float type.
784 (list (sb!c::make-interval :low nil :high nil)))
786 (destructuring-bind (neg pos)
787 (interval-split 0 x t t)
788 (list (interval-expt-< neg y)
789 (interval-expt-< pos y))))))
791 ;;; Compute bounds for (expt x y).
792 (defun interval-expt (x y)
793 (case (interval-range-info x 1)
796 (interval-expt-> x y))
799 (interval-expt-< x y))
801 (destructuring-bind (left right)
802 (interval-split 1 x t t)
803 (list (interval-expt left y)
804 (interval-expt right y))))))
806 (defun fixup-interval-expt (bnd x-int y-int x-type y-type)
807 (declare (ignore x-int))
808 ;; Figure out what the return type should be, given the argument
809 ;; types and bounds and the result type and bounds.
810 (cond ((csubtypep x-type (specifier-type 'integer))
811 ;; an integer to some power
812 (case (numeric-type-class y-type)
814 ;; Positive integer to an integer power is either an
815 ;; integer or a rational.
816 (let ((lo (or (interval-low bnd) '*))
817 (hi (or (interval-high bnd) '*)))
818 (if (and (interval-low y-int)
819 (>= (type-bound-number (interval-low y-int)) 0))
820 (specifier-type `(integer ,lo ,hi))
821 (specifier-type `(rational ,lo ,hi)))))
823 ;; Positive integer to rational power is either a rational
824 ;; or a single-float.
825 (let* ((lo (interval-low bnd))
826 (hi (interval-high bnd))
828 (floor (type-bound-number lo))
831 (ceiling (type-bound-number hi))
834 (bound-func #'float lo)
837 (bound-func #'float hi)
839 (specifier-type `(or (rational ,int-lo ,int-hi)
840 (single-float ,f-lo, f-hi)))))
842 ;; A positive integer to a float power is a float.
843 (modified-numeric-type y-type
844 :low (interval-low bnd)
845 :high (interval-high bnd)))
847 ;; A positive integer to a number is a number (for now).
848 (specifier-type 'number))))
849 ((csubtypep x-type (specifier-type 'rational))
850 ;; a rational to some power
851 (case (numeric-type-class y-type)
853 ;; A positive rational to an integer power is always a rational.
854 (specifier-type `(rational ,(or (interval-low bnd) '*)
855 ,(or (interval-high bnd) '*))))
857 ;; A positive rational to rational power is either a rational
858 ;; or a single-float.
859 (let* ((lo (interval-low bnd))
860 (hi (interval-high bnd))
862 (floor (type-bound-number lo))
865 (ceiling (type-bound-number hi))
868 (bound-func #'float lo)
871 (bound-func #'float hi)
873 (specifier-type `(or (rational ,int-lo ,int-hi)
874 (single-float ,f-lo, f-hi)))))
876 ;; A positive rational to a float power is a float.
877 (modified-numeric-type y-type
878 :low (interval-low bnd)
879 :high (interval-high bnd)))
881 ;; A positive rational to a number is a number (for now).
882 (specifier-type 'number))))
883 ((csubtypep x-type (specifier-type 'float))
884 ;; a float to some power
885 (case (numeric-type-class y-type)
886 ((or integer rational)
887 ;; A positive float to an integer or rational power is
891 :format (numeric-type-format x-type)
892 :low (interval-low bnd)
893 :high (interval-high bnd)))
895 ;; A positive float to a float power is a float of the
899 :format (float-format-max (numeric-type-format x-type)
900 (numeric-type-format y-type))
901 :low (interval-low bnd)
902 :high (interval-high bnd)))
904 ;; A positive float to a number is a number (for now)
905 (specifier-type 'number))))
907 ;; A number to some power is a number.
908 (specifier-type 'number))))
910 (defun merged-interval-expt (x y)
911 (let* ((x-int (numeric-type->interval x))
912 (y-int (numeric-type->interval y)))
913 (mapcar (lambda (type)
914 (fixup-interval-expt type x-int y-int x y))
915 (flatten-list (interval-expt x-int y-int)))))
917 (defun expt-derive-type-aux (x y same-arg)
918 (declare (ignore same-arg))
919 (cond ((or (not (numeric-type-real-p x))
920 (not (numeric-type-real-p y)))
921 ;; Use numeric contagion if either is not real.
922 (numeric-contagion x y))
923 ((csubtypep y (specifier-type 'integer))
924 ;; A real raised to an integer power is well-defined.
925 (merged-interval-expt x y))
926 ;; A real raised to a non-integral power can be a float or a
928 ((or (csubtypep x (specifier-type '(rational 0)))
929 (csubtypep x (specifier-type '(float (0d0)))))
930 ;; But a positive real to any power is well-defined.
931 (merged-interval-expt x y))
932 ((and (csubtypep x (specifier-type 'rational))
933 (csubtypep x (specifier-type 'rational)))
934 ;; A rational to the power of a rational could be a rational
935 ;; or a possibly-complex single float
936 (specifier-type '(or rational single-float (complex single-float))))
938 ;; a real to some power. The result could be a real or a
940 (float-or-complex-float-type (numeric-contagion x y)))))
942 (defoptimizer (expt derive-type) ((x y))
943 (two-arg-derive-type x y #'expt-derive-type-aux #'expt))
945 ;;; Note we must assume that a type including 0.0 may also include
946 ;;; -0.0 and thus the result may be complex -infinity + i*pi.
947 (defun log-derive-type-aux-1 (x)
948 (elfun-derive-type-simple x #'log 0d0 nil nil nil))
950 (defun log-derive-type-aux-2 (x y same-arg)
951 (let ((log-x (log-derive-type-aux-1 x))
952 (log-y (log-derive-type-aux-1 y))
953 (accumulated-list nil))
954 ;; LOG-X or LOG-Y might be union types. We need to run through
955 ;; the union types ourselves because /-DERIVE-TYPE-AUX doesn't.
956 (dolist (x-type (prepare-arg-for-derive-type log-x))
957 (dolist (y-type (prepare-arg-for-derive-type log-y))
958 (push (/-derive-type-aux x-type y-type same-arg) accumulated-list)))
959 (apply #'type-union (flatten-list accumulated-list))))
961 (defoptimizer (log derive-type) ((x &optional y))
963 (two-arg-derive-type x y #'log-derive-type-aux-2 #'log)
964 (one-arg-derive-type x #'log-derive-type-aux-1 #'log)))
966 (defun atan-derive-type-aux-1 (y)
967 (elfun-derive-type-simple y #'atan nil nil (- (/ pi 2)) (/ pi 2)))
969 (defun atan-derive-type-aux-2 (y x same-arg)
970 (declare (ignore same-arg))
971 ;; The hard case with two args. We just return the max bounds.
972 (let ((result-type (numeric-contagion y x)))
973 (cond ((and (numeric-type-real-p x)
974 (numeric-type-real-p y))
975 (let* (;; FIXME: This expression for FORMAT seems to
976 ;; appear multiple times, and should be factored out.
977 (format (case (numeric-type-class result-type)
978 ((integer rational) 'single-float)
979 (t (numeric-type-format result-type))))
980 (bound-format (or format 'float)))
981 (make-numeric-type :class 'float
984 :low (coerce (- pi) bound-format)
985 :high (coerce pi bound-format))))
987 ;; The result is a float or a complex number
988 (float-or-complex-float-type result-type)))))
990 (defoptimizer (atan derive-type) ((y &optional x))
992 (two-arg-derive-type y x #'atan-derive-type-aux-2 #'atan)
993 (one-arg-derive-type y #'atan-derive-type-aux-1 #'atan)))
995 (defun cosh-derive-type-aux (x)
996 ;; We note that cosh x = cosh |x| for all real x.
997 (elfun-derive-type-simple
998 (if (numeric-type-real-p x)
999 (abs-derive-type-aux x)
1001 #'cosh nil nil 0 nil))
1003 (defoptimizer (cosh derive-type) ((num))
1004 (one-arg-derive-type num #'cosh-derive-type-aux #'cosh))
1006 (defun phase-derive-type-aux (arg)
1007 (let* ((format (case (numeric-type-class arg)
1008 ((integer rational) 'single-float)
1009 (t (numeric-type-format arg))))
1010 (bound-type (or format 'float)))
1011 (cond ((numeric-type-real-p arg)
1012 (case (interval-range-info (numeric-type->interval arg) 0.0)
1014 ;; The number is positive, so the phase is 0.
1015 (make-numeric-type :class 'float
1018 :low (coerce 0 bound-type)
1019 :high (coerce 0 bound-type)))
1021 ;; The number is always negative, so the phase is pi.
1022 (make-numeric-type :class 'float
1025 :low (coerce pi bound-type)
1026 :high (coerce pi bound-type)))
1028 ;; We can't tell. The result is 0 or pi. Use a union
1031 (make-numeric-type :class 'float
1034 :low (coerce 0 bound-type)
1035 :high (coerce 0 bound-type))
1036 (make-numeric-type :class 'float
1039 :low (coerce pi bound-type)
1040 :high (coerce pi bound-type))))))
1042 ;; We have a complex number. The answer is the range -pi
1043 ;; to pi. (-pi is included because we have -0.)
1044 (make-numeric-type :class 'float
1047 :low (coerce (- pi) bound-type)
1048 :high (coerce pi bound-type))))))
1050 (defoptimizer (phase derive-type) ((num))
1051 (one-arg-derive-type num #'phase-derive-type-aux #'phase))
1055 (deftransform realpart ((x) ((complex rational)) *)
1056 '(sb!kernel:%realpart x))
1057 (deftransform imagpart ((x) ((complex rational)) *)
1058 '(sb!kernel:%imagpart x))
1060 ;;; Make REALPART and IMAGPART return the appropriate types. This
1061 ;;; should help a lot in optimized code.
1062 (defun realpart-derive-type-aux (type)
1063 (let ((class (numeric-type-class type))
1064 (format (numeric-type-format type)))
1065 (cond ((numeric-type-real-p type)
1066 ;; The realpart of a real has the same type and range as
1068 (make-numeric-type :class class
1071 :low (numeric-type-low type)
1072 :high (numeric-type-high type)))
1074 ;; We have a complex number. The result has the same type
1075 ;; as the real part, except that it's real, not complex,
1077 (make-numeric-type :class class
1080 :low (numeric-type-low type)
1081 :high (numeric-type-high type))))))
1082 #-sb-xc-host ; (See CROSS-FLOAT-INFINITY-KLUDGE.)
1083 (defoptimizer (realpart derive-type) ((num))
1084 (one-arg-derive-type num #'realpart-derive-type-aux #'realpart))
1085 (defun imagpart-derive-type-aux (type)
1086 (let ((class (numeric-type-class type))
1087 (format (numeric-type-format type)))
1088 (cond ((numeric-type-real-p type)
1089 ;; The imagpart of a real has the same type as the input,
1090 ;; except that it's zero.
1091 (let ((bound-format (or format class 'real)))
1092 (make-numeric-type :class class
1095 :low (coerce 0 bound-format)
1096 :high (coerce 0 bound-format))))
1098 ;; We have a complex number. The result has the same type as
1099 ;; the imaginary part, except that it's real, not complex,
1101 (make-numeric-type :class class
1104 :low (numeric-type-low type)
1105 :high (numeric-type-high type))))))
1106 #-sb-xc-host ; (See CROSS-FLOAT-INFINITY-KLUDGE.)
1107 (defoptimizer (imagpart derive-type) ((num))
1108 (one-arg-derive-type num #'imagpart-derive-type-aux #'imagpart))
1110 (defun complex-derive-type-aux-1 (re-type)
1111 (if (numeric-type-p re-type)
1112 (make-numeric-type :class (numeric-type-class re-type)
1113 :format (numeric-type-format re-type)
1114 :complexp (if (csubtypep re-type
1115 (specifier-type 'rational))
1118 :low (numeric-type-low re-type)
1119 :high (numeric-type-high re-type))
1120 (specifier-type 'complex)))
1122 (defun complex-derive-type-aux-2 (re-type im-type same-arg)
1123 (declare (ignore same-arg))
1124 (if (and (numeric-type-p re-type)
1125 (numeric-type-p im-type))
1126 ;; Need to check to make sure numeric-contagion returns the
1127 ;; right type for what we want here.
1129 ;; Also, what about rational canonicalization, like (complex 5 0)
1130 ;; is 5? So, if the result must be complex, we make it so.
1131 ;; If the result might be complex, which happens only if the
1132 ;; arguments are rational, we make it a union type of (or
1133 ;; rational (complex rational)).
1134 (let* ((element-type (numeric-contagion re-type im-type))
1135 (rat-result-p (csubtypep element-type
1136 (specifier-type 'rational))))
1138 (type-union element-type
1140 `(complex ,(numeric-type-class element-type))))
1141 (make-numeric-type :class (numeric-type-class element-type)
1142 :format (numeric-type-format element-type)
1143 :complexp (if rat-result-p
1146 (specifier-type 'complex)))
1148 #-sb-xc-host ; (See CROSS-FLOAT-INFINITY-KLUDGE.)
1149 (defoptimizer (complex derive-type) ((re &optional im))
1151 (two-arg-derive-type re im #'complex-derive-type-aux-2 #'complex)
1152 (one-arg-derive-type re #'complex-derive-type-aux-1 #'complex)))
1154 ;;; Define some transforms for complex operations. We do this in lieu
1155 ;;; of complex operation VOPs.
1156 (macrolet ((frob (type)
1159 (deftransform %negate ((z) ((complex ,type)) *)
1160 '(complex (%negate (realpart z)) (%negate (imagpart z))))
1161 ;; complex addition and subtraction
1162 (deftransform + ((w z) ((complex ,type) (complex ,type)) *)
1163 '(complex (+ (realpart w) (realpart z))
1164 (+ (imagpart w) (imagpart z))))
1165 (deftransform - ((w z) ((complex ,type) (complex ,type)) *)
1166 '(complex (- (realpart w) (realpart z))
1167 (- (imagpart w) (imagpart z))))
1168 ;; Add and subtract a complex and a real.
1169 (deftransform + ((w z) ((complex ,type) real) *)
1170 '(complex (+ (realpart w) z) (imagpart w)))
1171 (deftransform + ((z w) (real (complex ,type)) *)
1172 '(complex (+ (realpart w) z) (imagpart w)))
1173 ;; Add and subtract a real and a complex number.
1174 (deftransform - ((w z) ((complex ,type) real) *)
1175 '(complex (- (realpart w) z) (imagpart w)))
1176 (deftransform - ((z w) (real (complex ,type)) *)
1177 '(complex (- z (realpart w)) (- (imagpart w))))
1178 ;; Multiply and divide two complex numbers.
1179 (deftransform * ((x y) ((complex ,type) (complex ,type)) *)
1180 '(let* ((rx (realpart x))
1184 (complex (- (* rx ry) (* ix iy))
1185 (+ (* rx iy) (* ix ry)))))
1186 (deftransform / ((x y) ((complex ,type) (complex ,type)) *)
1187 '(let* ((rx (realpart x))
1191 (if (> (abs ry) (abs iy))
1192 (let* ((r (/ iy ry))
1193 (dn (* ry (+ 1 (* r r)))))
1194 (complex (/ (+ rx (* ix r)) dn)
1195 (/ (- ix (* rx r)) dn)))
1196 (let* ((r (/ ry iy))
1197 (dn (* iy (+ 1 (* r r)))))
1198 (complex (/ (+ (* rx r) ix) dn)
1199 (/ (- (* ix r) rx) dn))))))
1200 ;; Multiply a complex by a real or vice versa.
1201 (deftransform * ((w z) ((complex ,type) real) *)
1202 '(complex (* (realpart w) z) (* (imagpart w) z)))
1203 (deftransform * ((z w) (real (complex ,type)) *)
1204 '(complex (* (realpart w) z) (* (imagpart w) z)))
1205 ;; Divide a complex by a real.
1206 (deftransform / ((w z) ((complex ,type) real) *)
1207 '(complex (/ (realpart w) z) (/ (imagpart w) z)))
1208 ;; conjugate of complex number
1209 (deftransform conjugate ((z) ((complex ,type)) *)
1210 '(complex (realpart z) (- (imagpart z))))
1212 (deftransform cis ((z) ((,type)) *)
1213 '(complex (cos z) (sin z)))
1215 (deftransform = ((w z) ((complex ,type) (complex ,type)) *)
1216 '(and (= (realpart w) (realpart z))
1217 (= (imagpart w) (imagpart z))))
1218 (deftransform = ((w z) ((complex ,type) real) *)
1219 '(and (= (realpart w) z) (zerop (imagpart w))))
1220 (deftransform = ((w z) (real (complex ,type)) *)
1221 '(and (= (realpart z) w) (zerop (imagpart z)))))))
1224 (frob double-float))
1226 ;;; Here are simple optimizers for SIN, COS, and TAN. They do not
1227 ;;; produce a minimal range for the result; the result is the widest
1228 ;;; possible answer. This gets around the problem of doing range
1229 ;;; reduction correctly but still provides useful results when the
1230 ;;; inputs are union types.
1231 #-sb-xc-host ; (See CROSS-FLOAT-INFINITY-KLUDGE.)
1233 (defun trig-derive-type-aux (arg domain fcn
1234 &optional def-lo def-hi (increasingp t))
1237 (cond ((eq (numeric-type-complexp arg) :complex)
1238 (make-numeric-type :class (numeric-type-class arg)
1239 :format (numeric-type-format arg)
1240 :complexp :complex))
1241 ((numeric-type-real-p arg)
1242 (let* ((format (case (numeric-type-class arg)
1243 ((integer rational) 'single-float)
1244 (t (numeric-type-format arg))))
1245 (bound-type (or format 'float)))
1246 ;; If the argument is a subset of the "principal" domain
1247 ;; of the function, we can compute the bounds because
1248 ;; the function is monotonic. We can't do this in
1249 ;; general for these periodic functions because we can't
1250 ;; (and don't want to) do the argument reduction in
1251 ;; exactly the same way as the functions themselves do
1253 (if (csubtypep arg domain)
1254 (let ((res-lo (bound-func fcn (numeric-type-low arg)))
1255 (res-hi (bound-func fcn (numeric-type-high arg))))
1257 (rotatef res-lo res-hi))
1261 :low (coerce-numeric-bound res-lo bound-type)
1262 :high (coerce-numeric-bound res-hi bound-type)))
1266 :low (and def-lo (coerce def-lo bound-type))
1267 :high (and def-hi (coerce def-hi bound-type))))))
1269 (float-or-complex-float-type arg def-lo def-hi))))))
1271 (defoptimizer (sin derive-type) ((num))
1272 (one-arg-derive-type
1275 ;; Derive the bounds if the arg is in [-pi/2, pi/2].
1276 (trig-derive-type-aux
1278 (specifier-type `(float ,(- (/ pi 2)) ,(/ pi 2)))
1283 (defoptimizer (cos derive-type) ((num))
1284 (one-arg-derive-type
1287 ;; Derive the bounds if the arg is in [0, pi].
1288 (trig-derive-type-aux arg
1289 (specifier-type `(float 0d0 ,pi))
1295 (defoptimizer (tan derive-type) ((num))
1296 (one-arg-derive-type
1299 ;; Derive the bounds if the arg is in [-pi/2, pi/2].
1300 (trig-derive-type-aux arg
1301 (specifier-type `(float ,(- (/ pi 2)) ,(/ pi 2)))
1306 (defoptimizer (conjugate derive-type) ((num))
1307 (one-arg-derive-type num
1309 (flet ((most-negative-bound (l h)
1311 (if (< (type-bound-number l) (- (type-bound-number h)))
1313 (set-bound (- (type-bound-number h)) (consp h)))))
1314 (most-positive-bound (l h)
1316 (if (> (type-bound-number h) (- (type-bound-number l)))
1318 (set-bound (- (type-bound-number l)) (consp l))))))
1319 (if (numeric-type-real-p arg)
1321 (let ((low (numeric-type-low arg))
1322 (high (numeric-type-high arg)))
1323 (let ((new-low (most-negative-bound low high))
1324 (new-high (most-positive-bound low high)))
1325 (modified-numeric-type arg :low new-low :high new-high))))))
1328 (defoptimizer (cis derive-type) ((num))
1329 (one-arg-derive-type num
1331 (sb!c::specifier-type
1332 `(complex ,(or (numeric-type-format arg) 'float))))
1337 ;;;; TRUNCATE, FLOOR, CEILING, and ROUND
1339 (macrolet ((define-frobs (fun ufun)
1341 (defknown ,ufun (real) integer (movable foldable flushable))
1342 (deftransform ,fun ((x &optional by)
1344 (constant-arg (member 1))))
1345 '(let ((res (,ufun x)))
1346 (values res (- x res)))))))
1347 (define-frobs truncate %unary-truncate)
1348 (define-frobs round %unary-round))
1350 ;;; Convert (TRUNCATE x y) to the obvious implementation. We only want
1351 ;;; this when under certain conditions and let the generic TRUNCATE
1352 ;;; handle the rest. (Note: if Y = 1, the divide and multiply by Y
1353 ;;; should be removed by other DEFTRANSFORMs.)
1354 (deftransform truncate ((x &optional y)
1355 (float &optional (or float integer)))
1356 (let ((defaulted-y (if y 'y 1)))
1357 `(let ((res (%unary-truncate (/ x ,defaulted-y))))
1358 (values res (- x (* ,defaulted-y res))))))
1360 (deftransform floor ((number &optional divisor)
1361 (float &optional (or integer float)))
1362 (let ((defaulted-divisor (if divisor 'divisor 1)))
1363 `(multiple-value-bind (tru rem) (truncate number ,defaulted-divisor)
1364 (if (and (not (zerop rem))
1365 (if (minusp ,defaulted-divisor)
1368 (values (1- tru) (+ rem ,defaulted-divisor))
1369 (values tru rem)))))
1371 (deftransform ceiling ((number &optional divisor)
1372 (float &optional (or integer float)))
1373 (let ((defaulted-divisor (if divisor 'divisor 1)))
1374 `(multiple-value-bind (tru rem) (truncate number ,defaulted-divisor)
1375 (if (and (not (zerop rem))
1376 (if (minusp ,defaulted-divisor)
1379 (values (1+ tru) (- rem ,defaulted-divisor))
1380 (values tru rem)))))
1382 (defknown %unary-ftruncate (real) float (movable foldable flushable))
1383 (defknown %unary-ftruncate/single (single-float) single-float
1384 (movable foldable flushable))
1385 (defknown %unary-ftruncate/double (double-float) double-float
1386 (movable foldable flushable))
1388 (defun %unary-ftruncate/single (x)
1389 (declare (type single-float x))
1390 (declare (optimize speed (safety 0)))
1391 (let* ((bits (single-float-bits x))
1392 (exp (ldb sb!vm:single-float-exponent-byte bits))
1393 (biased (the single-float-exponent
1394 (- exp sb!vm:single-float-bias))))
1395 (declare (type (signed-byte 32) bits))
1397 ((= exp sb!vm:single-float-normal-exponent-max) x)
1398 ((<= biased 0) (* x 0f0))
1399 ((>= biased (float-digits x)) x)
1401 (let ((frac-bits (- (float-digits x) biased)))
1402 (setf bits (logandc2 bits (- (ash 1 frac-bits) 1)))
1403 (make-single-float bits))))))
1405 (defun %unary-ftruncate/double (x)
1406 (declare (type double-float x))
1407 (declare (optimize speed (safety 0)))
1408 (let* ((high (double-float-high-bits x))
1409 (low (double-float-low-bits x))
1410 (exp (ldb sb!vm:double-float-exponent-byte high))
1411 (biased (the double-float-exponent
1412 (- exp sb!vm:double-float-bias))))
1413 (declare (type (signed-byte 32) high)
1414 (type (unsigned-byte 32) low))
1416 ((= exp sb!vm:double-float-normal-exponent-max) x)
1417 ((<= biased 0) (* x 0d0))
1418 ((>= biased (float-digits x)) x)
1420 (let ((frac-bits (- (float-digits x) biased)))
1421 (cond ((< frac-bits 32)
1422 (setf low (logandc2 low (- (ash 1 frac-bits) 1))))
1425 (setf high (logandc2 high (- (ash 1 (- frac-bits 32)) 1)))))
1426 (make-double-float high low))))))
1429 ((def (float-type fun)
1430 `(deftransform %unary-ftruncate ((x) (,float-type))
1432 (def single-float %unary-ftruncate/single)
1433 (def double-float %unary-ftruncate/double))