1.0.19.3: more careful PROGV and SET
[sbcl/tcr.git] / src / compiler / float-tran.lisp
blob262f86777cae2a22f469e98076068ceda5b4af71
1 ;;;; This file contains floating-point-specific transforms, and may be
2 ;;;; somewhat implementation-dependent in its assumptions of what the
3 ;;;; formats are.
5 ;;;; This software is part of the SBCL system. See the README file for
6 ;;;; more information.
7 ;;;;
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.
14 (in-package "SB!C")
16 ;;;; coercions
18 (defknown %single-float (real) single-float (movable foldable))
19 (defknown %double-float (real) double-float (movable foldable))
21 (deftransform float ((n f) (* single-float) *)
22 '(%single-float n))
24 (deftransform float ((n f) (* double-float) *)
25 '(%double-float n))
27 (deftransform float ((n) *)
28 '(if (floatp n)
30 (%single-float n)))
32 (deftransform %single-float ((n) (single-float) *)
33 'n)
35 (deftransform %double-float ((n) (double-float) *)
36 'n)
38 ;;; RANDOM
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
48 ;;;
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
57 ;; catch myself.
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
62 (32 'random-chunk)
63 (64 'sb!kernel::big-random-chunk))))
64 (if (numeric-type-p type)
65 (let ((num-high (numeric-type-high (lvar-type num))))
66 (aver num-high)
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))
72 num-high limit)
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*))
78 #!-(or x86 x86-64)
79 `(rem (,random-chunk (or state *random-state*)) num)
80 #!+(or x86 x86-64)
81 ;; Use multiplication, which is faster.
82 `(values (sb!bignum::%multiply
83 (,random-chunk (or state *random-state*))
84 num)))))
85 ((> num-high random-fixnum-max)
86 (give-up-ir1-transform
87 "The range is too large to ensure an accurate result."))
88 #!+(or x86 x86-64)
89 ((< num-high limit)
90 `(values (sb!bignum::%multiply
91 (,random-chunk (or state *random-state*))
92 num)))
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
97 ;; 2004-11.
98 '(rem (random-chunk (or state *random-state*)) num))))
100 ;;;; float accessors
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) *)
119 (if float2
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) *)
127 (if float2
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))))
180 '(%scalbn f ex)
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
195 ;;; code. Pro:
196 ;;; * SBCL itself gets built with more complete optimization.
197 ;;; Con:
198 ;;; * You get a different SBCL depending on what your cross-compilation
199 ;;; host is.
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
207 ;;; process as is.
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.)
220 (progn
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
227 ;; zeros.
228 (set-bound
229 (handler-case
230 (scale-float (type-bound-number x) n)
231 (floating-point-overflow ()
232 nil))
233 (consp x))))
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))
239 (new-lo nil)
240 (new-hi nil))
241 (when f-hi
242 (if (< (float-sign (type-bound-number f-hi)) 0.0)
243 (when ex-lo
244 (setf new-hi (scale-bound f-hi ex-lo)))
245 (when ex-hi
246 (setf new-hi (scale-bound f-hi ex-hi)))))
247 (when f-lo
248 (if (< (float-sign (type-bound-number f-lo)) 0.0)
249 (when ex-hi
250 (setf new-lo (scale-bound f-lo ex-hi)))
251 (when ex-lo
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)
255 :complexp :real
256 :low new-lo
257 :high new-hi)))))
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.
269 (macrolet
270 ((frob (fun type most-negative most-positive)
271 (let ((aux-name (symbolicate fun "-DERIVE-TYPE-AUX")))
272 `(progn
273 (defun ,aux-name (num)
274 ;; When converting a number to a float, the limits are
275 ;; the same.
276 (let* ((lo (bound-func (lambda (x)
277 (if (< x ,most-negative)
278 ,most-negative
279 (coerce x ',type)))
280 (numeric-type-low num)))
281 (hi (bound-func (lambda (x)
282 (if (< ,most-positive x )
283 ,most-positive
284 (coerce x ',type)))
285 (numeric-type-high num))))
286 (specifier-type `(,',type ,(or lo '*) ,(or hi '*)))))
288 (defoptimizer (,fun derive-type) ((num))
289 (handler-case
290 (one-arg-derive-type num #',aux-name #',fun)
291 (type-error ()
292 nil)))))))
293 (frob %single-float single-float
294 most-negative-single-float most-positive-single-float)
295 (frob %double-float double-float
296 most-negative-double-float most-positive-double-float))
297 ) ; PROGN
299 ;;;; float contagion
301 (defun safe-ctype-for-single-coercion-p (x)
302 ;; See comment in SAFE-SINGLE-COERCION-P -- this deals with the same
303 ;; problem, but in the context of evaluated and compiled (+ <int> <single>)
304 ;; giving different result if we fail to check for this.
305 (or (not (csubtypep x (specifier-type 'integer)))
306 (csubtypep x (specifier-type `(integer ,most-negative-exactly-single-float-fixnum
307 ,most-positive-exactly-single-float-fixnum)))))
309 ;;; Do some stuff to recognize when the loser is doing mixed float and
310 ;;; rational arithmetic, or different float types, and fix it up. If
311 ;;; we don't, he won't even get so much as an efficiency note.
312 (deftransform float-contagion-arg1 ((x y) * * :defun-only t :node node)
313 (if (or (not (types-equal-or-intersect (lvar-type y) (specifier-type 'single-float)))
314 (safe-ctype-for-single-coercion-p (lvar-type x)))
315 `(,(lvar-fun-name (basic-combination-fun node))
316 (float x y) y)
317 (give-up-ir1-transform)))
318 (deftransform float-contagion-arg2 ((x y) * * :defun-only t :node node)
319 (if (or (not (types-equal-or-intersect (lvar-type x) (specifier-type 'single-float)))
320 (safe-ctype-for-single-coercion-p (lvar-type y)))
321 `(,(lvar-fun-name (basic-combination-fun node))
322 x (float y x))
323 (give-up-ir1-transform)))
325 (dolist (x '(+ * / -))
326 (%deftransform x '(function (rational float) *) #'float-contagion-arg1)
327 (%deftransform x '(function (float rational) *) #'float-contagion-arg2))
329 (dolist (x '(= < > + * / -))
330 (%deftransform x '(function (single-float double-float) *)
331 #'float-contagion-arg1)
332 (%deftransform x '(function (double-float single-float) *)
333 #'float-contagion-arg2))
335 ;;; Prevent ZEROP, PLUSP, and MINUSP from losing horribly. We can't in
336 ;;; general float rational args to comparison, since Common Lisp
337 ;;; semantics says we are supposed to compare as rationals, but we can
338 ;;; do it for any rational that has a precise representation as a
339 ;;; float (such as 0).
340 (macrolet ((frob (op)
341 `(deftransform ,op ((x y) (float rational) *)
342 "open-code FLOAT to RATIONAL comparison"
343 (unless (constant-lvar-p y)
344 (give-up-ir1-transform
345 "The RATIONAL value isn't known at compile time."))
346 (let ((val (lvar-value y)))
347 (unless (eql (rational (float val)) val)
348 (give-up-ir1-transform
349 "~S doesn't have a precise float representation."
350 val)))
351 `(,',op x (float y x)))))
352 (frob <)
353 (frob >)
354 (frob =))
356 ;;;; irrational derive-type methods
358 ;;; Derive the result to be float for argument types in the
359 ;;; appropriate domain.
360 #+sb-xc-host ; (See CROSS-FLOAT-INFINITY-KLUDGE.)
361 (dolist (stuff '((asin (real -1.0 1.0))
362 (acos (real -1.0 1.0))
363 (acosh (real 1.0))
364 (atanh (real -1.0 1.0))
365 (sqrt (real 0.0))))
366 (destructuring-bind (name type) stuff
367 (let ((type (specifier-type type)))
368 (setf (fun-info-derive-type (fun-info-or-lose name))
369 (lambda (call)
370 (declare (type combination call))
371 (when (csubtypep (lvar-type
372 (first (combination-args call)))
373 type)
374 (specifier-type 'float)))))))
376 #+sb-xc-host ; (See CROSS-FLOAT-INFINITY-KLUDGE.)
377 (defoptimizer (log derive-type) ((x &optional y))
378 (when (and (csubtypep (lvar-type x)
379 (specifier-type '(real 0.0)))
380 (or (null y)
381 (csubtypep (lvar-type y)
382 (specifier-type '(real 0.0)))))
383 (specifier-type 'float)))
385 ;;;; irrational transforms
387 (defknown (%tan %sinh %asinh %atanh %log %logb %log10 %tan-quick)
388 (double-float) double-float
389 (movable foldable flushable))
391 (defknown (%sin %cos %tanh %sin-quick %cos-quick)
392 (double-float) (double-float -1.0d0 1.0d0)
393 (movable foldable flushable))
395 (defknown (%asin %atan)
396 (double-float)
397 (double-float #.(coerce (- (/ pi 2)) 'double-float)
398 #.(coerce (/ pi 2) 'double-float))
399 (movable foldable flushable))
401 (defknown (%acos)
402 (double-float) (double-float 0.0d0 #.(coerce pi 'double-float))
403 (movable foldable flushable))
405 (defknown (%cosh)
406 (double-float) (double-float 1.0d0)
407 (movable foldable flushable))
409 (defknown (%acosh %exp %sqrt)
410 (double-float) (double-float 0.0d0)
411 (movable foldable flushable))
413 (defknown %expm1
414 (double-float) (double-float -1d0)
415 (movable foldable flushable))
417 (defknown (%hypot)
418 (double-float double-float) (double-float 0d0)
419 (movable foldable flushable))
421 (defknown (%pow)
422 (double-float double-float) double-float
423 (movable foldable flushable))
425 (defknown (%atan2)
426 (double-float double-float)
427 (double-float #.(coerce (- pi) 'double-float)
428 #.(coerce pi 'double-float))
429 (movable foldable flushable))
431 (defknown (%scalb)
432 (double-float double-float) double-float
433 (movable foldable flushable))
435 (defknown (%scalbn)
436 (double-float (signed-byte 32)) double-float
437 (movable foldable flushable))
439 (defknown (%log1p)
440 (double-float) double-float
441 (movable foldable flushable))
443 (macrolet ((def (name prim rtype)
444 `(progn
445 (deftransform ,name ((x) (single-float) ,rtype)
446 `(coerce (,',prim (coerce x 'double-float)) 'single-float))
447 (deftransform ,name ((x) (double-float) ,rtype)
448 `(,',prim x)))))
449 (def exp %exp *)
450 (def log %log float)
451 (def sqrt %sqrt float)
452 (def asin %asin float)
453 (def acos %acos float)
454 (def atan %atan *)
455 (def sinh %sinh *)
456 (def cosh %cosh *)
457 (def tanh %tanh *)
458 (def asinh %asinh *)
459 (def acosh %acosh float)
460 (def atanh %atanh float))
462 ;;; The argument range is limited on the x86 FP trig. functions. A
463 ;;; post-test can detect a failure (and load a suitable result), but
464 ;;; this test is avoided if possible.
465 (macrolet ((def (name prim prim-quick)
466 (declare (ignorable prim-quick))
467 `(progn
468 (deftransform ,name ((x) (single-float) *)
469 #!+x86 (cond ((csubtypep (lvar-type x)
470 (specifier-type '(single-float
471 (#.(- (expt 2f0 64)))
472 (#.(expt 2f0 64)))))
473 `(coerce (,',prim-quick (coerce x 'double-float))
474 'single-float))
476 (compiler-notify
477 "unable to avoid inline argument range check~@
478 because the argument range (~S) was not within 2^64"
479 (type-specifier (lvar-type x)))
480 `(coerce (,',prim (coerce x 'double-float)) 'single-float)))
481 #!-x86 `(coerce (,',prim (coerce x 'double-float)) 'single-float))
482 (deftransform ,name ((x) (double-float) *)
483 #!+x86 (cond ((csubtypep (lvar-type x)
484 (specifier-type '(double-float
485 (#.(- (expt 2d0 64)))
486 (#.(expt 2d0 64)))))
487 `(,',prim-quick x))
489 (compiler-notify
490 "unable to avoid inline argument range check~@
491 because the argument range (~S) was not within 2^64"
492 (type-specifier (lvar-type x)))
493 `(,',prim x)))
494 #!-x86 `(,',prim x)))))
495 (def sin %sin %sin-quick)
496 (def cos %cos %cos-quick)
497 (def tan %tan %tan-quick))
499 (deftransform atan ((x y) (single-float single-float) *)
500 `(coerce (%atan2 (coerce x 'double-float) (coerce y 'double-float))
501 'single-float))
502 (deftransform atan ((x y) (double-float double-float) *)
503 `(%atan2 x y))
505 (deftransform expt ((x y) ((single-float 0f0) single-float) *)
506 `(coerce (%pow (coerce x 'double-float) (coerce y 'double-float))
507 'single-float))
508 (deftransform expt ((x y) ((double-float 0d0) double-float) *)
509 `(%pow x y))
510 (deftransform expt ((x y) ((single-float 0f0) (signed-byte 32)) *)
511 `(coerce (%pow (coerce x 'double-float) (coerce y 'double-float))
512 'single-float))
513 (deftransform expt ((x y) ((double-float 0d0) (signed-byte 32)) *)
514 `(%pow x (coerce y 'double-float)))
516 ;;; ANSI says log with base zero returns zero.
517 (deftransform log ((x y) (float float) float)
518 '(if (zerop y) y (/ (log x) (log y))))
520 ;;; Handle some simple transformations.
522 (deftransform abs ((x) ((complex double-float)) double-float)
523 '(%hypot (realpart x) (imagpart x)))
525 (deftransform abs ((x) ((complex single-float)) single-float)
526 '(coerce (%hypot (coerce (realpart x) 'double-float)
527 (coerce (imagpart x) 'double-float))
528 'single-float))
530 (deftransform phase ((x) ((complex double-float)) double-float)
531 '(%atan2 (imagpart x) (realpart x)))
533 (deftransform phase ((x) ((complex single-float)) single-float)
534 '(coerce (%atan2 (coerce (imagpart x) 'double-float)
535 (coerce (realpart x) 'double-float))
536 'single-float))
538 (deftransform phase ((x) ((float)) float)
539 '(if (minusp (float-sign x))
540 (float pi x)
541 (float 0 x)))
543 ;;; The number is of type REAL.
544 (defun numeric-type-real-p (type)
545 (and (numeric-type-p type)
546 (eq (numeric-type-complexp type) :real)))
548 ;;; Coerce a numeric type bound to the given type while handling
549 ;;; exclusive bounds.
550 (defun coerce-numeric-bound (bound type)
551 (when bound
552 (if (consp bound)
553 (list (coerce (car bound) type))
554 (coerce bound type))))
556 #-sb-xc-host ; (See CROSS-FLOAT-INFINITY-KLUDGE.)
557 (progn
559 ;;;; optimizers for elementary functions
560 ;;;;
561 ;;;; These optimizers compute the output range of the elementary
562 ;;;; function, based on the domain of the input.
564 ;;; Generate a specifier for a complex type specialized to the same
565 ;;; type as the argument.
566 (defun complex-float-type (arg)
567 (declare (type numeric-type arg))
568 (let* ((format (case (numeric-type-class arg)
569 ((integer rational) 'single-float)
570 (t (numeric-type-format arg))))
571 (float-type (or format 'float)))
572 (specifier-type `(complex ,float-type))))
574 ;;; Compute a specifier like '(OR FLOAT (COMPLEX FLOAT)), except float
575 ;;; should be the right kind of float. Allow bounds for the float
576 ;;; part too.
577 (defun float-or-complex-float-type (arg &optional lo hi)
578 (declare (type numeric-type arg))
579 (let* ((format (case (numeric-type-class arg)
580 ((integer rational) 'single-float)
581 (t (numeric-type-format arg))))
582 (float-type (or format 'float))
583 (lo (coerce-numeric-bound lo float-type))
584 (hi (coerce-numeric-bound hi float-type)))
585 (specifier-type `(or (,float-type ,(or lo '*) ,(or hi '*))
586 (complex ,float-type)))))
588 ) ; PROGN
590 (eval-when (:compile-toplevel :execute)
591 ;; So the problem with this hack is that it's actually broken. If
592 ;; the host does not have long floats, then setting *R-D-F-F* to
593 ;; LONG-FLOAT doesn't actually buy us anything. FIXME.
594 (setf *read-default-float-format*
595 #!+long-float 'long-float #!-long-float 'double-float))
596 ;;; Test whether the numeric-type ARG is within in domain specified by
597 ;;; DOMAIN-LOW and DOMAIN-HIGH, consider negative and positive zero to
598 ;;; be distinct.
599 #-sb-xc-host ; (See CROSS-FLOAT-INFINITY-KLUDGE.)
600 (defun domain-subtypep (arg domain-low domain-high)
601 (declare (type numeric-type arg)
602 (type (or real null) domain-low domain-high))
603 (let* ((arg-lo (numeric-type-low arg))
604 (arg-lo-val (type-bound-number arg-lo))
605 (arg-hi (numeric-type-high arg))
606 (arg-hi-val (type-bound-number arg-hi)))
607 ;; Check that the ARG bounds are correctly canonicalized.
608 (when (and arg-lo (floatp arg-lo-val) (zerop arg-lo-val) (consp arg-lo)
609 (minusp (float-sign arg-lo-val)))
610 (compiler-notify "float zero bound ~S not correctly canonicalized?" arg-lo)
611 (setq arg-lo 0e0 arg-lo-val arg-lo))
612 (when (and arg-hi (zerop arg-hi-val) (floatp arg-hi-val) (consp arg-hi)
613 (plusp (float-sign arg-hi-val)))
614 (compiler-notify "float zero bound ~S not correctly canonicalized?" arg-hi)
615 (setq arg-hi (ecase *read-default-float-format*
616 (double-float (load-time-value (make-unportable-float :double-float-negative-zero)))
617 #!+long-float
618 (long-float (load-time-value (make-unportable-float :long-float-negative-zero))))
619 arg-hi-val arg-hi))
620 (flet ((fp-neg-zero-p (f) ; Is F -0.0?
621 (and (floatp f) (zerop f) (minusp (float-sign f))))
622 (fp-pos-zero-p (f) ; Is F +0.0?
623 (and (floatp f) (zerop f) (plusp (float-sign f)))))
624 (and (or (null domain-low)
625 (and arg-lo (>= arg-lo-val domain-low)
626 (not (and (fp-pos-zero-p domain-low)
627 (fp-neg-zero-p arg-lo)))))
628 (or (null domain-high)
629 (and arg-hi (<= arg-hi-val domain-high)
630 (not (and (fp-neg-zero-p domain-high)
631 (fp-pos-zero-p arg-hi)))))))))
632 (eval-when (:compile-toplevel :execute)
633 (setf *read-default-float-format* 'single-float))
635 #-sb-xc-host ; (See CROSS-FLOAT-INFINITY-KLUDGE.)
636 (progn
638 ;;; Handle monotonic functions of a single variable whose domain is
639 ;;; possibly part of the real line. ARG is the variable, FCN is the
640 ;;; function, and DOMAIN is a specifier that gives the (real) domain
641 ;;; of the function. If ARG is a subset of the DOMAIN, we compute the
642 ;;; bounds directly. Otherwise, we compute the bounds for the
643 ;;; intersection between ARG and DOMAIN, and then append a complex
644 ;;; result, which occurs for the parts of ARG not in the DOMAIN.
646 ;;; Negative and positive zero are considered distinct within
647 ;;; DOMAIN-LOW and DOMAIN-HIGH.
649 ;;; DEFAULT-LOW and DEFAULT-HIGH are the lower and upper bounds if we
650 ;;; can't compute the bounds using FCN.
651 (defun elfun-derive-type-simple (arg fcn domain-low domain-high
652 default-low default-high
653 &optional (increasingp t))
654 (declare (type (or null real) domain-low domain-high))
655 (etypecase arg
656 (numeric-type
657 (cond ((eq (numeric-type-complexp arg) :complex)
658 (complex-float-type arg))
659 ((numeric-type-real-p arg)
660 ;; The argument is real, so let's find the intersection
661 ;; between the argument and the domain of the function.
662 ;; We compute the bounds on the intersection, and for
663 ;; everything else, we return a complex number of the
664 ;; appropriate type.
665 (multiple-value-bind (intersection difference)
666 (interval-intersection/difference (numeric-type->interval arg)
667 (make-interval
668 :low domain-low
669 :high domain-high))
670 (cond
671 (intersection
672 ;; Process the intersection.
673 (let* ((low (interval-low intersection))
674 (high (interval-high intersection))
675 (res-lo (or (bound-func fcn (if increasingp low high))
676 default-low))
677 (res-hi (or (bound-func fcn (if increasingp high low))
678 default-high))
679 (format (case (numeric-type-class arg)
680 ((integer rational) 'single-float)
681 (t (numeric-type-format arg))))
682 (bound-type (or format 'float))
683 (result-type
684 (make-numeric-type
685 :class 'float
686 :format format
687 :low (coerce-numeric-bound res-lo bound-type)
688 :high (coerce-numeric-bound res-hi bound-type))))
689 ;; If the ARG is a subset of the domain, we don't
690 ;; have to worry about the difference, because that
691 ;; can't occur.
692 (if (or (null difference)
693 ;; Check whether the arg is within the domain.
694 (domain-subtypep arg domain-low domain-high))
695 result-type
696 (list result-type
697 (specifier-type `(complex ,bound-type))))))
699 ;; No intersection so the result must be purely complex.
700 (complex-float-type arg)))))
702 (float-or-complex-float-type arg default-low default-high))))))
704 (macrolet
705 ((frob (name domain-low domain-high def-low-bnd def-high-bnd
706 &key (increasingp t))
707 (let ((num (gensym)))
708 `(defoptimizer (,name derive-type) ((,num))
709 (one-arg-derive-type
710 ,num
711 (lambda (arg)
712 (elfun-derive-type-simple arg #',name
713 ,domain-low ,domain-high
714 ,def-low-bnd ,def-high-bnd
715 ,increasingp))
716 #',name)))))
717 ;; These functions are easy because they are defined for the whole
718 ;; real line.
719 (frob exp nil nil 0 nil)
720 (frob sinh nil nil nil nil)
721 (frob tanh nil nil -1 1)
722 (frob asinh nil nil nil nil)
724 ;; These functions are only defined for part of the real line. The
725 ;; condition selects the desired part of the line.
726 (frob asin -1d0 1d0 (- (/ pi 2)) (/ pi 2))
727 ;; Acos is monotonic decreasing, so we need to swap the function
728 ;; values at the lower and upper bounds of the input domain.
729 (frob acos -1d0 1d0 0 pi :increasingp nil)
730 (frob acosh 1d0 nil nil nil)
731 (frob atanh -1d0 1d0 -1 1)
732 ;; Kahan says that (sqrt -0.0) is -0.0, so use a specifier that
733 ;; includes -0.0.
734 (frob sqrt (load-time-value (make-unportable-float :double-float-negative-zero)) nil 0 nil))
736 ;;; Compute bounds for (expt x y). This should be easy since (expt x
737 ;;; y) = (exp (* y (log x))). However, computations done this way
738 ;;; have too much roundoff. Thus we have to do it the hard way.
739 (defun safe-expt (x y)
740 (handler-case
741 (when (< (abs y) 10000)
742 (expt x y))
743 (error ()
744 nil)))
746 ;;; Handle the case when x >= 1.
747 (defun interval-expt-> (x y)
748 (case (sb!c::interval-range-info y 0d0)
750 ;; Y is positive and log X >= 0. The range of exp(y * log(x)) is
751 ;; obviously non-negative. We just have to be careful for
752 ;; infinite bounds (given by nil).
753 (let ((lo (safe-expt (type-bound-number (sb!c::interval-low x))
754 (type-bound-number (sb!c::interval-low y))))
755 (hi (safe-expt (type-bound-number (sb!c::interval-high x))
756 (type-bound-number (sb!c::interval-high y)))))
757 (list (sb!c::make-interval :low (or lo 1) :high hi))))
759 ;; Y is negative and log x >= 0. The range of exp(y * log(x)) is
760 ;; obviously [0, 1]. However, underflow (nil) means 0 is the
761 ;; result.
762 (let ((lo (safe-expt (type-bound-number (sb!c::interval-high x))
763 (type-bound-number (sb!c::interval-low y))))
764 (hi (safe-expt (type-bound-number (sb!c::interval-low x))
765 (type-bound-number (sb!c::interval-high y)))))
766 (list (sb!c::make-interval :low (or lo 0) :high (or hi 1)))))
768 ;; Split the interval in half.
769 (destructuring-bind (y- y+)
770 (sb!c::interval-split 0 y t)
771 (list (interval-expt-> x y-)
772 (interval-expt-> x y+))))))
774 ;;; Handle the case when x <= 1
775 (defun interval-expt-< (x y)
776 (case (sb!c::interval-range-info x 0d0)
778 ;; The case of 0 <= x <= 1 is easy
779 (case (sb!c::interval-range-info y)
781 ;; Y is positive and log X <= 0. The range of exp(y * log(x)) is
782 ;; obviously [0, 1]. We just have to be careful for infinite bounds
783 ;; (given by nil).
784 (let ((lo (safe-expt (type-bound-number (sb!c::interval-low x))
785 (type-bound-number (sb!c::interval-high y))))
786 (hi (safe-expt (type-bound-number (sb!c::interval-high x))
787 (type-bound-number (sb!c::interval-low y)))))
788 (list (sb!c::make-interval :low (or lo 0) :high (or hi 1)))))
790 ;; Y is negative and log x <= 0. The range of exp(y * log(x)) is
791 ;; obviously [1, inf].
792 (let ((hi (safe-expt (type-bound-number (sb!c::interval-low x))
793 (type-bound-number (sb!c::interval-low y))))
794 (lo (safe-expt (type-bound-number (sb!c::interval-high x))
795 (type-bound-number (sb!c::interval-high y)))))
796 (list (sb!c::make-interval :low (or lo 1) :high hi))))
798 ;; Split the interval in half
799 (destructuring-bind (y- y+)
800 (sb!c::interval-split 0 y t)
801 (list (interval-expt-< x y-)
802 (interval-expt-< x y+))))))
804 ;; The case where x <= 0. Y MUST be an INTEGER for this to work!
805 ;; The calling function must insure this! For now we'll just
806 ;; return the appropriate unbounded float type.
807 (list (sb!c::make-interval :low nil :high nil)))
809 (destructuring-bind (neg pos)
810 (interval-split 0 x t t)
811 (list (interval-expt-< neg y)
812 (interval-expt-< pos y))))))
814 ;;; Compute bounds for (expt x y).
815 (defun interval-expt (x y)
816 (case (interval-range-info x 1)
818 ;; X >= 1
819 (interval-expt-> x y))
821 ;; X <= 1
822 (interval-expt-< x y))
824 (destructuring-bind (left right)
825 (interval-split 1 x t t)
826 (list (interval-expt left y)
827 (interval-expt right y))))))
829 (defun fixup-interval-expt (bnd x-int y-int x-type y-type)
830 (declare (ignore x-int))
831 ;; Figure out what the return type should be, given the argument
832 ;; types and bounds and the result type and bounds.
833 (cond ((csubtypep x-type (specifier-type 'integer))
834 ;; an integer to some power
835 (case (numeric-type-class y-type)
836 (integer
837 ;; Positive integer to an integer power is either an
838 ;; integer or a rational.
839 (let ((lo (or (interval-low bnd) '*))
840 (hi (or (interval-high bnd) '*)))
841 (if (and (interval-low y-int)
842 (>= (type-bound-number (interval-low y-int)) 0))
843 (specifier-type `(integer ,lo ,hi))
844 (specifier-type `(rational ,lo ,hi)))))
845 (rational
846 ;; Positive integer to rational power is either a rational
847 ;; or a single-float.
848 (let* ((lo (interval-low bnd))
849 (hi (interval-high bnd))
850 (int-lo (if lo
851 (floor (type-bound-number lo))
852 '*))
853 (int-hi (if hi
854 (ceiling (type-bound-number hi))
855 '*))
856 (f-lo (if lo
857 (bound-func #'float lo)
858 '*))
859 (f-hi (if hi
860 (bound-func #'float hi)
861 '*)))
862 (specifier-type `(or (rational ,int-lo ,int-hi)
863 (single-float ,f-lo, f-hi)))))
864 (float
865 ;; A positive integer to a float power is a float.
866 (modified-numeric-type y-type
867 :low (interval-low bnd)
868 :high (interval-high bnd)))
870 ;; A positive integer to a number is a number (for now).
871 (specifier-type 'number))))
872 ((csubtypep x-type (specifier-type 'rational))
873 ;; a rational to some power
874 (case (numeric-type-class y-type)
875 (integer
876 ;; A positive rational to an integer power is always a rational.
877 (specifier-type `(rational ,(or (interval-low bnd) '*)
878 ,(or (interval-high bnd) '*))))
879 (rational
880 ;; A positive rational to rational power is either a rational
881 ;; or a single-float.
882 (let* ((lo (interval-low bnd))
883 (hi (interval-high bnd))
884 (int-lo (if lo
885 (floor (type-bound-number lo))
886 '*))
887 (int-hi (if hi
888 (ceiling (type-bound-number hi))
889 '*))
890 (f-lo (if lo
891 (bound-func #'float lo)
892 '*))
893 (f-hi (if hi
894 (bound-func #'float hi)
895 '*)))
896 (specifier-type `(or (rational ,int-lo ,int-hi)
897 (single-float ,f-lo, f-hi)))))
898 (float
899 ;; A positive rational to a float power is a float.
900 (modified-numeric-type y-type
901 :low (interval-low bnd)
902 :high (interval-high bnd)))
904 ;; A positive rational to a number is a number (for now).
905 (specifier-type 'number))))
906 ((csubtypep x-type (specifier-type 'float))
907 ;; a float to some power
908 (case (numeric-type-class y-type)
909 ((or integer rational)
910 ;; A positive float to an integer or rational power is
911 ;; always a float.
912 (make-numeric-type
913 :class 'float
914 :format (numeric-type-format x-type)
915 :low (interval-low bnd)
916 :high (interval-high bnd)))
917 (float
918 ;; A positive float to a float power is a float of the
919 ;; higher type.
920 (make-numeric-type
921 :class 'float
922 :format (float-format-max (numeric-type-format x-type)
923 (numeric-type-format y-type))
924 :low (interval-low bnd)
925 :high (interval-high bnd)))
927 ;; A positive float to a number is a number (for now)
928 (specifier-type 'number))))
930 ;; A number to some power is a number.
931 (specifier-type 'number))))
933 (defun merged-interval-expt (x y)
934 (let* ((x-int (numeric-type->interval x))
935 (y-int (numeric-type->interval y)))
936 (mapcar (lambda (type)
937 (fixup-interval-expt type x-int y-int x y))
938 (flatten-list (interval-expt x-int y-int)))))
940 (defun expt-derive-type-aux (x y same-arg)
941 (declare (ignore same-arg))
942 (cond ((or (not (numeric-type-real-p x))
943 (not (numeric-type-real-p y)))
944 ;; Use numeric contagion if either is not real.
945 (numeric-contagion x y))
946 ((csubtypep y (specifier-type 'integer))
947 ;; A real raised to an integer power is well-defined.
948 (merged-interval-expt x y))
949 ;; A real raised to a non-integral power can be a float or a
950 ;; complex number.
951 ((or (csubtypep x (specifier-type '(rational 0)))
952 (csubtypep x (specifier-type '(float (0d0)))))
953 ;; But a positive real to any power is well-defined.
954 (merged-interval-expt x y))
955 ((and (csubtypep x (specifier-type 'rational))
956 (csubtypep x (specifier-type 'rational)))
957 ;; A rational to the power of a rational could be a rational
958 ;; or a possibly-complex single float
959 (specifier-type '(or rational single-float (complex single-float))))
961 ;; a real to some power. The result could be a real or a
962 ;; complex.
963 (float-or-complex-float-type (numeric-contagion x y)))))
965 (defoptimizer (expt derive-type) ((x y))
966 (two-arg-derive-type x y #'expt-derive-type-aux #'expt))
968 ;;; Note we must assume that a type including 0.0 may also include
969 ;;; -0.0 and thus the result may be complex -infinity + i*pi.
970 (defun log-derive-type-aux-1 (x)
971 (elfun-derive-type-simple x #'log 0d0 nil nil nil))
973 (defun log-derive-type-aux-2 (x y same-arg)
974 (let ((log-x (log-derive-type-aux-1 x))
975 (log-y (log-derive-type-aux-1 y))
976 (accumulated-list nil))
977 ;; LOG-X or LOG-Y might be union types. We need to run through
978 ;; the union types ourselves because /-DERIVE-TYPE-AUX doesn't.
979 (dolist (x-type (prepare-arg-for-derive-type log-x))
980 (dolist (y-type (prepare-arg-for-derive-type log-y))
981 (push (/-derive-type-aux x-type y-type same-arg) accumulated-list)))
982 (apply #'type-union (flatten-list accumulated-list))))
984 (defoptimizer (log derive-type) ((x &optional y))
985 (if y
986 (two-arg-derive-type x y #'log-derive-type-aux-2 #'log)
987 (one-arg-derive-type x #'log-derive-type-aux-1 #'log)))
989 (defun atan-derive-type-aux-1 (y)
990 (elfun-derive-type-simple y #'atan nil nil (- (/ pi 2)) (/ pi 2)))
992 (defun atan-derive-type-aux-2 (y x same-arg)
993 (declare (ignore same-arg))
994 ;; The hard case with two args. We just return the max bounds.
995 (let ((result-type (numeric-contagion y x)))
996 (cond ((and (numeric-type-real-p x)
997 (numeric-type-real-p y))
998 (let* (;; FIXME: This expression for FORMAT seems to
999 ;; appear multiple times, and should be factored out.
1000 (format (case (numeric-type-class result-type)
1001 ((integer rational) 'single-float)
1002 (t (numeric-type-format result-type))))
1003 (bound-format (or format 'float)))
1004 (make-numeric-type :class 'float
1005 :format format
1006 :complexp :real
1007 :low (coerce (- pi) bound-format)
1008 :high (coerce pi bound-format))))
1010 ;; The result is a float or a complex number
1011 (float-or-complex-float-type result-type)))))
1013 (defoptimizer (atan derive-type) ((y &optional x))
1014 (if x
1015 (two-arg-derive-type y x #'atan-derive-type-aux-2 #'atan)
1016 (one-arg-derive-type y #'atan-derive-type-aux-1 #'atan)))
1018 (defun cosh-derive-type-aux (x)
1019 ;; We note that cosh x = cosh |x| for all real x.
1020 (elfun-derive-type-simple
1021 (if (numeric-type-real-p x)
1022 (abs-derive-type-aux x)
1024 #'cosh nil nil 0 nil))
1026 (defoptimizer (cosh derive-type) ((num))
1027 (one-arg-derive-type num #'cosh-derive-type-aux #'cosh))
1029 (defun phase-derive-type-aux (arg)
1030 (let* ((format (case (numeric-type-class arg)
1031 ((integer rational) 'single-float)
1032 (t (numeric-type-format arg))))
1033 (bound-type (or format 'float)))
1034 (cond ((numeric-type-real-p arg)
1035 (case (interval-range-info (numeric-type->interval arg) 0.0)
1037 ;; The number is positive, so the phase is 0.
1038 (make-numeric-type :class 'float
1039 :format format
1040 :complexp :real
1041 :low (coerce 0 bound-type)
1042 :high (coerce 0 bound-type)))
1044 ;; The number is always negative, so the phase is pi.
1045 (make-numeric-type :class 'float
1046 :format format
1047 :complexp :real
1048 :low (coerce pi bound-type)
1049 :high (coerce pi bound-type)))
1051 ;; We can't tell. The result is 0 or pi. Use a union
1052 ;; type for this.
1053 (list
1054 (make-numeric-type :class 'float
1055 :format format
1056 :complexp :real
1057 :low (coerce 0 bound-type)
1058 :high (coerce 0 bound-type))
1059 (make-numeric-type :class 'float
1060 :format format
1061 :complexp :real
1062 :low (coerce pi bound-type)
1063 :high (coerce pi bound-type))))))
1065 ;; We have a complex number. The answer is the range -pi
1066 ;; to pi. (-pi is included because we have -0.)
1067 (make-numeric-type :class 'float
1068 :format format
1069 :complexp :real
1070 :low (coerce (- pi) bound-type)
1071 :high (coerce pi bound-type))))))
1073 (defoptimizer (phase derive-type) ((num))
1074 (one-arg-derive-type num #'phase-derive-type-aux #'phase))
1076 ) ; PROGN
1078 (deftransform realpart ((x) ((complex rational)) *)
1079 '(sb!kernel:%realpart x))
1080 (deftransform imagpart ((x) ((complex rational)) *)
1081 '(sb!kernel:%imagpart x))
1083 ;;; Make REALPART and IMAGPART return the appropriate types. This
1084 ;;; should help a lot in optimized code.
1085 (defun realpart-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 realpart of a real has the same type and range as
1090 ;; the input.
1091 (make-numeric-type :class class
1092 :format format
1093 :complexp :real
1094 :low (numeric-type-low type)
1095 :high (numeric-type-high type)))
1097 ;; We have a complex number. The result has the same type
1098 ;; as the real part, except that it's real, not complex,
1099 ;; obviously.
1100 (make-numeric-type :class class
1101 :format format
1102 :complexp :real
1103 :low (numeric-type-low type)
1104 :high (numeric-type-high type))))))
1105 #-sb-xc-host ; (See CROSS-FLOAT-INFINITY-KLUDGE.)
1106 (defoptimizer (realpart derive-type) ((num))
1107 (one-arg-derive-type num #'realpart-derive-type-aux #'realpart))
1108 (defun imagpart-derive-type-aux (type)
1109 (let ((class (numeric-type-class type))
1110 (format (numeric-type-format type)))
1111 (cond ((numeric-type-real-p type)
1112 ;; The imagpart of a real has the same type as the input,
1113 ;; except that it's zero.
1114 (let ((bound-format (or format class 'real)))
1115 (make-numeric-type :class class
1116 :format format
1117 :complexp :real
1118 :low (coerce 0 bound-format)
1119 :high (coerce 0 bound-format))))
1121 ;; We have a complex number. The result has the same type as
1122 ;; the imaginary part, except that it's real, not complex,
1123 ;; obviously.
1124 (make-numeric-type :class class
1125 :format format
1126 :complexp :real
1127 :low (numeric-type-low type)
1128 :high (numeric-type-high type))))))
1129 #-sb-xc-host ; (See CROSS-FLOAT-INFINITY-KLUDGE.)
1130 (defoptimizer (imagpart derive-type) ((num))
1131 (one-arg-derive-type num #'imagpart-derive-type-aux #'imagpart))
1133 (defun complex-derive-type-aux-1 (re-type)
1134 (if (numeric-type-p re-type)
1135 (make-numeric-type :class (numeric-type-class re-type)
1136 :format (numeric-type-format re-type)
1137 :complexp (if (csubtypep re-type
1138 (specifier-type 'rational))
1139 :real
1140 :complex)
1141 :low (numeric-type-low re-type)
1142 :high (numeric-type-high re-type))
1143 (specifier-type 'complex)))
1145 (defun complex-derive-type-aux-2 (re-type im-type same-arg)
1146 (declare (ignore same-arg))
1147 (if (and (numeric-type-p re-type)
1148 (numeric-type-p im-type))
1149 ;; Need to check to make sure numeric-contagion returns the
1150 ;; right type for what we want here.
1152 ;; Also, what about rational canonicalization, like (complex 5 0)
1153 ;; is 5? So, if the result must be complex, we make it so.
1154 ;; If the result might be complex, which happens only if the
1155 ;; arguments are rational, we make it a union type of (or
1156 ;; rational (complex rational)).
1157 (let* ((element-type (numeric-contagion re-type im-type))
1158 (rat-result-p (csubtypep element-type
1159 (specifier-type 'rational))))
1160 (if rat-result-p
1161 (type-union element-type
1162 (specifier-type
1163 `(complex ,(numeric-type-class element-type))))
1164 (make-numeric-type :class (numeric-type-class element-type)
1165 :format (numeric-type-format element-type)
1166 :complexp (if rat-result-p
1167 :real
1168 :complex))))
1169 (specifier-type 'complex)))
1171 #-sb-xc-host ; (See CROSS-FLOAT-INFINITY-KLUDGE.)
1172 (defoptimizer (complex derive-type) ((re &optional im))
1173 (if im
1174 (two-arg-derive-type re im #'complex-derive-type-aux-2 #'complex)
1175 (one-arg-derive-type re #'complex-derive-type-aux-1 #'complex)))
1177 ;;; Define some transforms for complex operations. We do this in lieu
1178 ;;; of complex operation VOPs.
1179 (macrolet ((frob (type)
1180 `(progn
1181 ;; negation
1182 (deftransform %negate ((z) ((complex ,type)) *)
1183 '(complex (%negate (realpart z)) (%negate (imagpart z))))
1184 ;; complex addition and subtraction
1185 (deftransform + ((w z) ((complex ,type) (complex ,type)) *)
1186 '(complex (+ (realpart w) (realpart z))
1187 (+ (imagpart w) (imagpart z))))
1188 (deftransform - ((w z) ((complex ,type) (complex ,type)) *)
1189 '(complex (- (realpart w) (realpart z))
1190 (- (imagpart w) (imagpart z))))
1191 ;; Add and subtract a complex and a real.
1192 (deftransform + ((w z) ((complex ,type) real) *)
1193 '(complex (+ (realpart w) z) (imagpart w)))
1194 (deftransform + ((z w) (real (complex ,type)) *)
1195 '(complex (+ (realpart w) z) (imagpart w)))
1196 ;; Add and subtract a real and a complex number.
1197 (deftransform - ((w z) ((complex ,type) real) *)
1198 '(complex (- (realpart w) z) (imagpart w)))
1199 (deftransform - ((z w) (real (complex ,type)) *)
1200 '(complex (- z (realpart w)) (- (imagpart w))))
1201 ;; Multiply and divide two complex numbers.
1202 (deftransform * ((x y) ((complex ,type) (complex ,type)) *)
1203 '(let* ((rx (realpart x))
1204 (ix (imagpart x))
1205 (ry (realpart y))
1206 (iy (imagpart y)))
1207 (complex (- (* rx ry) (* ix iy))
1208 (+ (* rx iy) (* ix ry)))))
1209 (deftransform / ((x y) ((complex ,type) (complex ,type)) *)
1210 '(let* ((rx (realpart x))
1211 (ix (imagpart x))
1212 (ry (realpart y))
1213 (iy (imagpart y)))
1214 (if (> (abs ry) (abs iy))
1215 (let* ((r (/ iy ry))
1216 (dn (* ry (+ 1 (* r r)))))
1217 (complex (/ (+ rx (* ix r)) dn)
1218 (/ (- ix (* rx r)) dn)))
1219 (let* ((r (/ ry iy))
1220 (dn (* iy (+ 1 (* r r)))))
1221 (complex (/ (+ (* rx r) ix) dn)
1222 (/ (- (* ix r) rx) dn))))))
1223 ;; Multiply a complex by a real or vice versa.
1224 (deftransform * ((w z) ((complex ,type) real) *)
1225 '(complex (* (realpart w) z) (* (imagpart w) z)))
1226 (deftransform * ((z w) (real (complex ,type)) *)
1227 '(complex (* (realpart w) z) (* (imagpart w) z)))
1228 ;; Divide a complex by a real.
1229 (deftransform / ((w z) ((complex ,type) real) *)
1230 '(complex (/ (realpart w) z) (/ (imagpart w) z)))
1231 ;; conjugate of complex number
1232 (deftransform conjugate ((z) ((complex ,type)) *)
1233 '(complex (realpart z) (- (imagpart z))))
1234 ;; CIS
1235 (deftransform cis ((z) ((,type)) *)
1236 '(complex (cos z) (sin z)))
1237 ;; comparison
1238 (deftransform = ((w z) ((complex ,type) (complex ,type)) *)
1239 '(and (= (realpart w) (realpart z))
1240 (= (imagpart w) (imagpart z))))
1241 (deftransform = ((w z) ((complex ,type) real) *)
1242 '(and (= (realpart w) z) (zerop (imagpart w))))
1243 (deftransform = ((w z) (real (complex ,type)) *)
1244 '(and (= (realpart z) w) (zerop (imagpart z)))))))
1246 (frob single-float)
1247 (frob double-float))
1249 ;;; Here are simple optimizers for SIN, COS, and TAN. They do not
1250 ;;; produce a minimal range for the result; the result is the widest
1251 ;;; possible answer. This gets around the problem of doing range
1252 ;;; reduction correctly but still provides useful results when the
1253 ;;; inputs are union types.
1254 #-sb-xc-host ; (See CROSS-FLOAT-INFINITY-KLUDGE.)
1255 (progn
1256 (defun trig-derive-type-aux (arg domain fcn
1257 &optional def-lo def-hi (increasingp t))
1258 (etypecase arg
1259 (numeric-type
1260 (cond ((eq (numeric-type-complexp arg) :complex)
1261 (make-numeric-type :class (numeric-type-class arg)
1262 :format (numeric-type-format arg)
1263 :complexp :complex))
1264 ((numeric-type-real-p arg)
1265 (let* ((format (case (numeric-type-class arg)
1266 ((integer rational) 'single-float)
1267 (t (numeric-type-format arg))))
1268 (bound-type (or format 'float)))
1269 ;; If the argument is a subset of the "principal" domain
1270 ;; of the function, we can compute the bounds because
1271 ;; the function is monotonic. We can't do this in
1272 ;; general for these periodic functions because we can't
1273 ;; (and don't want to) do the argument reduction in
1274 ;; exactly the same way as the functions themselves do
1275 ;; it.
1276 (if (csubtypep arg domain)
1277 (let ((res-lo (bound-func fcn (numeric-type-low arg)))
1278 (res-hi (bound-func fcn (numeric-type-high arg))))
1279 (unless increasingp
1280 (rotatef res-lo res-hi))
1281 (make-numeric-type
1282 :class 'float
1283 :format format
1284 :low (coerce-numeric-bound res-lo bound-type)
1285 :high (coerce-numeric-bound res-hi bound-type)))
1286 (make-numeric-type
1287 :class 'float
1288 :format format
1289 :low (and def-lo (coerce def-lo bound-type))
1290 :high (and def-hi (coerce def-hi bound-type))))))
1292 (float-or-complex-float-type arg def-lo def-hi))))))
1294 (defoptimizer (sin derive-type) ((num))
1295 (one-arg-derive-type
1297 (lambda (arg)
1298 ;; Derive the bounds if the arg is in [-pi/2, pi/2].
1299 (trig-derive-type-aux
1301 (specifier-type `(float ,(- (/ pi 2)) ,(/ pi 2)))
1302 #'sin
1303 -1 1))
1304 #'sin))
1306 (defoptimizer (cos derive-type) ((num))
1307 (one-arg-derive-type
1309 (lambda (arg)
1310 ;; Derive the bounds if the arg is in [0, pi].
1311 (trig-derive-type-aux arg
1312 (specifier-type `(float 0d0 ,pi))
1313 #'cos
1314 -1 1
1315 nil))
1316 #'cos))
1318 (defoptimizer (tan derive-type) ((num))
1319 (one-arg-derive-type
1321 (lambda (arg)
1322 ;; Derive the bounds if the arg is in [-pi/2, pi/2].
1323 (trig-derive-type-aux arg
1324 (specifier-type `(float ,(- (/ pi 2)) ,(/ pi 2)))
1325 #'tan
1326 nil nil))
1327 #'tan))
1329 (defoptimizer (conjugate derive-type) ((num))
1330 (one-arg-derive-type num
1331 (lambda (arg)
1332 (flet ((most-negative-bound (l h)
1333 (and l h
1334 (if (< (type-bound-number l) (- (type-bound-number h)))
1336 (set-bound (- (type-bound-number h)) (consp h)))))
1337 (most-positive-bound (l h)
1338 (and l h
1339 (if (> (type-bound-number h) (- (type-bound-number l)))
1341 (set-bound (- (type-bound-number l)) (consp l))))))
1342 (if (numeric-type-real-p arg)
1343 (lvar-type num)
1344 (let ((low (numeric-type-low arg))
1345 (high (numeric-type-high arg)))
1346 (let ((new-low (most-negative-bound low high))
1347 (new-high (most-positive-bound low high)))
1348 (modified-numeric-type arg :low new-low :high new-high))))))
1349 #'conjugate))
1351 (defoptimizer (cis derive-type) ((num))
1352 (one-arg-derive-type num
1353 (lambda (arg)
1354 (sb!c::specifier-type
1355 `(complex ,(or (numeric-type-format arg) 'float))))
1356 #'cis))
1358 ) ; PROGN
1360 ;;;; TRUNCATE, FLOOR, CEILING, and ROUND
1362 (macrolet ((define-frobs (fun ufun)
1363 `(progn
1364 (defknown ,ufun (real) integer (movable foldable flushable))
1365 (deftransform ,fun ((x &optional by)
1366 (* &optional
1367 (constant-arg (member 1))))
1368 '(let ((res (,ufun x)))
1369 (values res (- x res)))))))
1370 (define-frobs truncate %unary-truncate)
1371 (define-frobs round %unary-round))
1373 ;;; Convert (TRUNCATE x y) to the obvious implementation. We only want
1374 ;;; this when under certain conditions and let the generic TRUNCATE
1375 ;;; handle the rest. (Note: if Y = 1, the divide and multiply by Y
1376 ;;; should be removed by other DEFTRANSFORMs.)
1377 (deftransform truncate ((x &optional y)
1378 (float &optional (or float integer)))
1379 (let ((defaulted-y (if y 'y 1)))
1380 `(let ((res (%unary-truncate (/ x ,defaulted-y))))
1381 (values res (- x (* ,defaulted-y res))))))
1383 (deftransform floor ((number &optional divisor)
1384 (float &optional (or integer float)))
1385 (let ((defaulted-divisor (if divisor 'divisor 1)))
1386 `(multiple-value-bind (tru rem) (truncate number ,defaulted-divisor)
1387 (if (and (not (zerop rem))
1388 (if (minusp ,defaulted-divisor)
1389 (plusp number)
1390 (minusp number)))
1391 (values (1- tru) (+ rem ,defaulted-divisor))
1392 (values tru rem)))))
1394 (deftransform ceiling ((number &optional divisor)
1395 (float &optional (or integer float)))
1396 (let ((defaulted-divisor (if divisor 'divisor 1)))
1397 `(multiple-value-bind (tru rem) (truncate number ,defaulted-divisor)
1398 (if (and (not (zerop rem))
1399 (if (minusp ,defaulted-divisor)
1400 (minusp number)
1401 (plusp number)))
1402 (values (1+ tru) (- rem ,defaulted-divisor))
1403 (values tru rem)))))
1405 (defknown %unary-ftruncate (real) float (movable foldable flushable))
1406 (defknown %unary-ftruncate/single (single-float) single-float
1407 (movable foldable flushable))
1408 (defknown %unary-ftruncate/double (double-float) double-float
1409 (movable foldable flushable))
1411 (defun %unary-ftruncate/single (x)
1412 (declare (type single-float x))
1413 (declare (optimize speed (safety 0)))
1414 (let* ((bits (single-float-bits x))
1415 (exp (ldb sb!vm:single-float-exponent-byte bits))
1416 (biased (the single-float-exponent
1417 (- exp sb!vm:single-float-bias))))
1418 (declare (type (signed-byte 32) bits))
1419 (cond
1420 ((= exp sb!vm:single-float-normal-exponent-max) x)
1421 ((<= biased 0) (* x 0f0))
1422 ((>= biased (float-digits x)) x)
1424 (let ((frac-bits (- (float-digits x) biased)))
1425 (setf bits (logandc2 bits (- (ash 1 frac-bits) 1)))
1426 (make-single-float bits))))))
1428 (defun %unary-ftruncate/double (x)
1429 (declare (type double-float x))
1430 (declare (optimize speed (safety 0)))
1431 (let* ((high (double-float-high-bits x))
1432 (low (double-float-low-bits x))
1433 (exp (ldb sb!vm:double-float-exponent-byte high))
1434 (biased (the double-float-exponent
1435 (- exp sb!vm:double-float-bias))))
1436 (declare (type (signed-byte 32) high)
1437 (type (unsigned-byte 32) low))
1438 (cond
1439 ((= exp sb!vm:double-float-normal-exponent-max) x)
1440 ((<= biased 0) (* x 0d0))
1441 ((>= biased (float-digits x)) x)
1443 (let ((frac-bits (- (float-digits x) biased)))
1444 (cond ((< frac-bits 32)
1445 (setf low (logandc2 low (- (ash 1 frac-bits) 1))))
1447 (setf low 0)
1448 (setf high (logandc2 high (- (ash 1 (- frac-bits 32)) 1)))))
1449 (make-double-float high low))))))
1451 (macrolet
1452 ((def (float-type fun)
1453 `(deftransform %unary-ftruncate ((x) (,float-type))
1454 '(,fun x))))
1455 (def single-float %unary-ftruncate/single)
1456 (def double-float %unary-ftruncate/double))