Suppress some compiler notes
[sbcl.git] / src / compiler / float-tran.lisp
blobe367dfd5e6bf1d4d1d0620f076c634ea71a479c9
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
19 (movable foldable))
20 (defknown %double-float (real) double-float
21 (movable foldable))
23 (deftransform float ((n f) (* single-float) *)
24 '(%single-float n))
26 (deftransform float ((n f) (* double-float) *)
27 '(%double-float n))
29 (deftransform float ((n) *)
30 '(if (floatp n)
32 (%single-float n)))
34 (deftransform %single-float ((n) (single-float) *)
35 'n)
37 (deftransform %double-float ((n) (double-float) *)
38 'n)
40 ;;; RANDOM
41 (macrolet ((frob (fun type)
42 `(deftransform random ((num &optional state)
43 (,type &optional *) *)
44 "Use inline float operations."
45 '(,fun num (or state *random-state*)))))
46 (frob %random-single-float single-float)
47 (frob %random-double-float double-float))
49 ;;; Return an expression to generate an integer of N-BITS many random
50 ;;; bits, using the minimal number of random chunks possible.
51 (defun generate-random-expr-for-power-of-2 (n-bits state)
52 (declare (type (integer 1 #.sb!vm:n-word-bits) n-bits))
53 (multiple-value-bind (n-chunk-bits chunk-expr)
54 (cond ((<= n-bits n-random-chunk-bits)
55 (values n-random-chunk-bits `(random-chunk ,state)))
56 ((<= n-bits (* 2 n-random-chunk-bits))
57 (values (* 2 n-random-chunk-bits) `(big-random-chunk ,state)))
59 (error "Unexpectedly small N-RANDOM-CHUNK-BITS")))
60 (if (< n-bits n-chunk-bits)
61 `(logand ,(1- (ash 1 n-bits)) ,chunk-expr)
62 chunk-expr)))
64 ;;; This transform for compile-time constant word-sized integers
65 ;;; generates an accept-reject loop to achieve equidistribution of the
66 ;;; returned values. Several optimizations are done: If NUM is a power
67 ;;; of two no loop is needed. If the random chunk size is half the word
68 ;;; size only one chunk is used where sufficient. For values of NUM
69 ;;; where it is possible and results in faster code, the rejection
70 ;;; probability is reduced by accepting all values below the largest
71 ;;; multiple of the limit that fits into one or two chunks and and doing
72 ;;; a division to get the random value into the desired range.
73 (deftransform random ((num &optional state)
74 ((constant-arg (integer 1 #.(expt 2 sb!vm:n-word-bits)))
75 &optional *)
77 :policy (and (> speed compilation-speed)
78 (> speed space)))
79 "optimize to inlined RANDOM-CHUNK operations"
80 (let ((num (lvar-value num)))
81 (if (= num 1)
83 (flet ((chunk-n-bits-and-expr (n-bits)
84 (cond ((<= n-bits n-random-chunk-bits)
85 (values n-random-chunk-bits
86 '(random-chunk (or state *random-state*))))
87 ((<= n-bits (* 2 n-random-chunk-bits))
88 (values (* 2 n-random-chunk-bits)
89 '(big-random-chunk (or state *random-state*))))
91 (error "Unexpectedly small N-RANDOM-CHUNK-BITS")))))
92 (if (zerop (logand num (1- num)))
93 ;; NUM is a power of 2.
94 (let ((n-bits (integer-length (1- num))))
95 (multiple-value-bind (n-chunk-bits chunk-expr)
96 (chunk-n-bits-and-expr n-bits)
97 (if (< n-bits n-chunk-bits)
98 `(logand ,(1- (ash 1 n-bits)) ,chunk-expr)
99 chunk-expr)))
100 ;; Generate an accept-reject loop.
101 (let ((n-bits (integer-length num)))
102 (multiple-value-bind (n-chunk-bits chunk-expr)
103 (chunk-n-bits-and-expr n-bits)
104 (if (or (> (* num 3) (expt 2 n-chunk-bits))
105 (logbitp (- n-bits 2) num))
106 ;; Division can't help as the quotient is below 3,
107 ;; or is too costly as the rejection probability
108 ;; without it is already small (namely at most 1/4
109 ;; with the given test, which is experimentally a
110 ;; reasonable threshold and cheap to test for).
111 `(loop
112 (let ((bits ,(generate-random-expr-for-power-of-2
113 n-bits '(or state *random-state*))))
114 (when (< bits num)
115 (return bits))))
116 (let ((d (truncate (expt 2 n-chunk-bits) num)))
117 `(loop
118 (let ((bits ,chunk-expr))
119 (when (< bits ,(* num d))
120 (return (values (truncate bits ,d)))))))))))))))
123 ;;;; float accessors
125 (defknown make-single-float ((signed-byte 32)) single-float
126 (movable flushable))
128 (defknown make-double-float ((signed-byte 32) (unsigned-byte 32)) double-float
129 (movable flushable))
131 #-sb-xc-host
132 (deftransform make-single-float ((bits)
133 ((signed-byte 32)))
134 "Conditional constant folding"
135 (unless (constant-lvar-p bits)
136 (give-up-ir1-transform))
137 (let* ((bits (lvar-value bits))
138 (float (make-single-float bits)))
139 (when (float-nan-p float)
140 (give-up-ir1-transform))
141 float))
143 #-sb-xc-host
144 (deftransform make-double-float ((hi lo)
145 ((signed-byte 32) (unsigned-byte 32)))
146 "Conditional constant folding"
147 (unless (and (constant-lvar-p hi)
148 (constant-lvar-p lo))
149 (give-up-ir1-transform))
150 (let* ((hi (lvar-value hi))
151 (lo (lvar-value lo))
152 (float (make-double-float hi lo)))
153 (when (float-nan-p float)
154 (give-up-ir1-transform))
155 float))
157 (defknown single-float-bits (single-float) (signed-byte 32)
158 (movable foldable flushable))
160 (defknown double-float-high-bits (double-float) (signed-byte 32)
161 (movable foldable flushable))
163 (defknown double-float-low-bits (double-float) (unsigned-byte 32)
164 (movable foldable flushable))
166 (deftransform float-sign ((float &optional float2)
167 (single-float &optional single-float) *)
168 (if float2
169 (let ((temp (gensym)))
170 `(let ((,temp (abs float2)))
171 (if (minusp (single-float-bits float)) (- ,temp) ,temp)))
172 '(if (minusp (single-float-bits float)) -1f0 1f0)))
174 (deftransform float-sign ((float &optional float2)
175 (double-float &optional double-float) *)
176 (if float2
177 (let ((temp (gensym)))
178 `(let ((,temp (abs float2)))
179 (if (minusp (double-float-high-bits float)) (- ,temp) ,temp)))
180 '(if (minusp (double-float-high-bits float)) -1d0 1d0)))
182 ;;;; DECODE-FLOAT, INTEGER-DECODE-FLOAT, and SCALE-FLOAT
184 (defknown decode-single-float (single-float)
185 (values single-float single-float-exponent (single-float -1f0 1f0))
186 (movable foldable flushable))
188 (defknown decode-double-float (double-float)
189 (values double-float double-float-exponent (double-float -1d0 1d0))
190 (movable foldable flushable))
192 (defknown integer-decode-single-float (single-float)
193 (values single-float-significand single-float-int-exponent (integer -1 1))
194 (movable foldable flushable))
196 (defknown integer-decode-double-float (double-float)
197 (values double-float-significand double-float-int-exponent (integer -1 1))
198 (movable foldable flushable))
200 (defknown scale-single-float (single-float integer) single-float
201 (movable foldable flushable))
203 (defknown scale-double-float (double-float integer) double-float
204 (movable foldable flushable))
206 (deftransform decode-float ((x) (single-float) *)
207 '(decode-single-float x))
209 (deftransform decode-float ((x) (double-float) *)
210 '(decode-double-float x))
212 (deftransform integer-decode-float ((x) (single-float) *)
213 '(integer-decode-single-float x))
215 (deftransform integer-decode-float ((x) (double-float) *)
216 '(integer-decode-double-float x))
218 (deftransform scale-float ((f ex) (single-float *) *)
219 (cond #!+x86
220 ((csubtypep (lvar-type ex)
221 (specifier-type '(signed-byte 32)))
222 '(coerce (%scalbn (coerce f 'double-float) ex) 'single-float))
224 '(scale-single-float f ex))))
226 (deftransform scale-float ((f ex) (double-float *) *)
227 (cond #!+x86
228 ((csubtypep (lvar-type ex)
229 (specifier-type '(signed-byte 32)))
230 '(%scalbn f ex))
232 '(scale-double-float f ex))))
234 ;;; Given a number X, create a form suitable as a bound for an
235 ;;; interval. Make the bound open if OPEN-P is T. NIL remains NIL.
236 ;;; FIXME: as this is a constructor, shouldn't it be named MAKE-BOUND?
237 #!-sb-fluid (declaim (inline set-bound))
238 (defun set-bound (x open-p)
239 (if (and x open-p) (list x) x))
241 ;;; What is the CROSS-FLOAT-INFINITY-KLUDGE?
243 ;;; SBCL's own implementation of floating point supports floating
244 ;;; point infinities. Some of the old CMU CL :PROPAGATE-FLOAT-TYPE and
245 ;;; :PROPAGATE-FUN-TYPE code, like the DEFOPTIMIZERs below, uses this
246 ;;; floating point support. Thus, we have to avoid running it on the
247 ;;; cross-compilation host, since we're not guaranteed that the
248 ;;; cross-compilation host will support floating point infinities.
250 ;;; If we wanted to live dangerously, we could conditionalize the code
251 ;;; with #+(OR SBCL SB-XC) instead. That way, if the cross-compilation
252 ;;; host happened to be SBCL, we'd be able to run the infinity-using
253 ;;; code. Pro:
254 ;;; * SBCL itself gets built with more complete optimization.
255 ;;; Con:
256 ;;; * You get a different SBCL depending on what your cross-compilation
257 ;;; host is.
258 ;;; So far the pros and cons seem seem to be mostly academic, since
259 ;;; AFAIK (WHN 2001-08-28) the propagate-foo-type optimizations aren't
260 ;;; actually important in compiling SBCL itself. If this changes, then
261 ;;; we have to decide:
262 ;;; * Go for simplicity, leaving things as they are.
263 ;;; * Go for performance at the expense of conceptual clarity,
264 ;;; using #+(OR SBCL SB-XC) and otherwise leaving the build
265 ;;; process as is.
266 ;;; * Go for performance at the expense of build time, using
267 ;;; #+(OR SBCL SB-XC) and also making SBCL do not just
268 ;;; make-host-1.sh and make-host-2.sh, but a third step
269 ;;; make-host-3.sh where it builds itself under itself. (Such a
270 ;;; 3-step build process could also help with other things, e.g.
271 ;;; using specialized arrays to represent debug information.)
272 ;;; * Rewrite the code so that it doesn't depend on unportable
273 ;;; floating point infinities.
275 ;;; optimizers for SCALE-FLOAT. If the float has bounds, new bounds
276 ;;; are computed for the result, if possible.
277 #-sb-xc-host ; (See CROSS-FLOAT-INFINITY-KLUDGE.)
278 (progn
280 (defun scale-float-derive-type-aux (f ex same-arg)
281 (declare (ignore same-arg))
282 (flet ((scale-bound (x n)
283 ;; We need to be a bit careful here and catch any overflows
284 ;; that might occur. We can ignore underflows which become
285 ;; zeros.
286 (set-bound
287 (handler-case
288 (scale-float (type-bound-number x) n)
289 (floating-point-overflow ()
290 nil))
291 (consp x))))
292 (when (and (numeric-type-p f) (numeric-type-p ex))
293 (let ((f-lo (numeric-type-low f))
294 (f-hi (numeric-type-high f))
295 (ex-lo (numeric-type-low ex))
296 (ex-hi (numeric-type-high ex))
297 (new-lo nil)
298 (new-hi nil))
299 (when f-hi
300 (if (< (float-sign (type-bound-number f-hi)) 0.0)
301 (when ex-lo
302 (setf new-hi (scale-bound f-hi ex-lo)))
303 (when ex-hi
304 (setf new-hi (scale-bound f-hi ex-hi)))))
305 (when f-lo
306 (if (< (float-sign (type-bound-number f-lo)) 0.0)
307 (when ex-hi
308 (setf new-lo (scale-bound f-lo ex-hi)))
309 (when ex-lo
310 (setf new-lo (scale-bound f-lo ex-lo)))))
311 (make-numeric-type :class (numeric-type-class f)
312 :format (numeric-type-format f)
313 :complexp :real
314 :low new-lo
315 :high new-hi)))))
316 (defoptimizer (scale-single-float derive-type) ((f ex))
317 (two-arg-derive-type f ex #'scale-float-derive-type-aux
318 #'scale-single-float))
319 (defoptimizer (scale-double-float derive-type) ((f ex))
320 (two-arg-derive-type f ex #'scale-float-derive-type-aux
321 #'scale-double-float))
323 ;;; DEFOPTIMIZERs for %SINGLE-FLOAT and %DOUBLE-FLOAT. This makes the
324 ;;; FLOAT function return the correct ranges if the input has some
325 ;;; defined range. Quite useful if we want to convert some type of
326 ;;; bounded integer into a float.
327 (macrolet
328 ((frob (fun type most-negative most-positive)
329 (let ((aux-name (symbolicate fun "-DERIVE-TYPE-AUX")))
330 `(progn
331 (defun ,aux-name (num)
332 ;; When converting a number to a float, the limits are
333 ;; the same.
334 (let* ((lo (bound-func (lambda (x)
335 (if (< x ,most-negative)
336 ,most-negative
337 (coerce x ',type)))
338 (numeric-type-low num)
339 nil))
340 (hi (bound-func (lambda (x)
341 (if (< ,most-positive x )
342 ,most-positive
343 (coerce x ',type)))
344 (numeric-type-high num)
345 nil)))
346 (specifier-type `(,',type ,(or lo '*) ,(or hi '*)))))
348 (defoptimizer (,fun derive-type) ((num))
349 (handler-case
350 (one-arg-derive-type num #',aux-name #',fun)
351 (type-error ()
352 nil)))))))
353 (frob %single-float single-float
354 most-negative-single-float most-positive-single-float)
355 (frob %double-float double-float
356 most-negative-double-float most-positive-double-float))
357 ) ; PROGN
359 ;;;; float contagion
361 (defun safe-ctype-for-single-coercion-p (x)
362 ;; See comment in SAFE-SINGLE-COERCION-P -- this deals with the same
363 ;; problem, but in the context of evaluated and compiled (+ <int> <single>)
364 ;; giving different result if we fail to check for this.
365 (or (not (csubtypep x (specifier-type 'integer)))
366 #!+x86
367 (csubtypep x (specifier-type `(integer ,most-negative-exactly-single-float-fixnum
368 ,most-positive-exactly-single-float-fixnum)))
369 #!-x86
370 (csubtypep x (specifier-type 'fixnum))))
372 ;;; Do some stuff to recognize when the loser is doing mixed float and
373 ;;; rational arithmetic, or different float types, and fix it up. If
374 ;;; we don't, he won't even get so much as an efficiency note.
375 (deftransform float-contagion-arg1 ((x y) * * :defun-only t :node node)
376 (if (or (not (types-equal-or-intersect (lvar-type y) (specifier-type 'single-float)))
377 (safe-ctype-for-single-coercion-p (lvar-type x)))
378 `(,(lvar-fun-name (basic-combination-fun node))
379 (float x y) y)
380 (give-up-ir1-transform)))
381 (deftransform float-contagion-arg2 ((x y) * * :defun-only t :node node)
382 (if (or (not (types-equal-or-intersect (lvar-type x) (specifier-type 'single-float)))
383 (safe-ctype-for-single-coercion-p (lvar-type y)))
384 `(,(lvar-fun-name (basic-combination-fun node))
385 x (float y x))
386 (give-up-ir1-transform)))
388 (dolist (x '(+ * / -))
389 (%deftransform x '(function (rational float) *) #'float-contagion-arg1)
390 (%deftransform x '(function (float rational) *) #'float-contagion-arg2))
392 (dolist (x '(= < > + * / -))
393 (%deftransform x '(function (single-float double-float) *)
394 #'float-contagion-arg1)
395 (%deftransform x '(function (double-float single-float) *)
396 #'float-contagion-arg2))
398 (macrolet ((def (type &rest args)
399 `(deftransform * ((x y) (,type (constant-arg (member ,@args))) *
400 ;; Beware the SNaN!
401 :policy (zerop float-accuracy))
402 "optimize multiplication by one"
403 (let ((y (lvar-value y)))
404 (if (minusp y)
405 '(%negate x)
406 'x)))))
407 (def single-float 1.0 -1.0)
408 (def double-float 1.0d0 -1.0d0))
410 ;;; Return the reciprocal of X if it can be represented exactly, NIL otherwise.
411 (defun maybe-exact-reciprocal (x)
412 (unless (zerop x)
413 (handler-case
414 (multiple-value-bind (significand exponent sign)
415 (integer-decode-float x)
416 ;; only powers of 2 can be inverted exactly
417 (unless (zerop (logand significand (1- significand)))
418 (return-from maybe-exact-reciprocal nil))
419 (let ((expected (/ sign significand (expt 2 exponent)))
420 (reciprocal (/ x)))
421 (multiple-value-bind (significand exponent sign)
422 (integer-decode-float reciprocal)
423 ;; Denorms can't be inverted safely.
424 (and (eql expected (* sign significand (expt 2 exponent)))
425 reciprocal))))
426 (error () (return-from maybe-exact-reciprocal nil)))))
428 ;;; Replace constant division by multiplication with exact reciprocal,
429 ;;; if one exists.
430 (macrolet ((def (type)
431 `(deftransform / ((x y) (,type (constant-arg ,type)) *
432 :node node)
433 "convert to multiplication by reciprocal"
434 (let ((n (lvar-value y)))
435 (if (policy node (zerop float-accuracy))
436 `(* x ,(/ n))
437 (let ((r (maybe-exact-reciprocal n)))
438 (if r
439 `(* x ,r)
440 (give-up-ir1-transform
441 "~S does not have an exact reciprocal"
442 n))))))))
443 (def single-float)
444 (def double-float))
446 ;;; Optimize addition and subtraction of zero
447 (macrolet ((def (op type &rest args)
448 `(deftransform ,op ((x y) (,type (constant-arg (member ,@args))) *
449 ;; Beware the SNaN!
450 :policy (zerop float-accuracy))
451 'x)))
452 ;; No signed zeros, thanks.
453 (def + single-float 0 0.0)
454 (def - single-float 0 0.0)
455 (def + double-float 0 0.0 0.0d0)
456 (def - double-float 0 0.0 0.0d0))
458 ;;; On most platforms (+ x x) is faster than (* x 2)
459 (macrolet ((def (type &rest args)
460 `(deftransform * ((x y) (,type (constant-arg (member ,@args))))
461 '(+ x x))))
462 (def single-float 2 2.0)
463 (def double-float 2 2.0 2.0d0))
465 ;;; Prevent ZEROP, PLUSP, and MINUSP from losing horribly. We can't in
466 ;;; general float rational args to comparison, since Common Lisp
467 ;;; semantics says we are supposed to compare as rationals, but we can
468 ;;; do it for any rational that has a precise representation as a
469 ;;; float (such as 0).
470 (macrolet ((frob (op)
471 `(deftransform ,op ((x y) (float rational) *)
472 "open-code FLOAT to RATIONAL comparison"
473 (unless (constant-lvar-p y)
474 (give-up-ir1-transform
475 "The RATIONAL value isn't known at compile time."))
476 (let ((val (lvar-value y)))
477 (unless (eql (rational (float val)) val)
478 (give-up-ir1-transform
479 "~S doesn't have a precise float representation."
480 val)))
481 `(,',op x (float y x)))))
482 (frob <)
483 (frob >)
484 (frob =))
486 ;;;; irrational derive-type methods
488 ;;; Derive the result to be float for argument types in the
489 ;;; appropriate domain.
490 #+sb-xc-host ; (See CROSS-FLOAT-INFINITY-KLUDGE.)
491 (dolist (stuff '((asin (real -1.0 1.0))
492 (acos (real -1.0 1.0))
493 (acosh (real 1.0))
494 (atanh (real -1.0 1.0))
495 (sqrt (real 0.0))))
496 (destructuring-bind (name type) stuff
497 (let ((type (specifier-type type)))
498 (setf (fun-info-derive-type (fun-info-or-lose name))
499 (lambda (call)
500 (declare (type combination call))
501 (when (csubtypep (lvar-type
502 (first (combination-args call)))
503 type)
504 (specifier-type 'float)))))))
506 #+sb-xc-host ; (See CROSS-FLOAT-INFINITY-KLUDGE.)
507 (defoptimizer (log derive-type) ((x &optional y))
508 (when (and (csubtypep (lvar-type x)
509 (specifier-type '(real 0.0)))
510 (or (null y)
511 (csubtypep (lvar-type y)
512 (specifier-type '(real 0.0)))))
513 (specifier-type 'float)))
515 ;;;; irrational transforms
517 (defknown (%tan %sinh %asinh %atanh %log %logb %log10 %tan-quick)
518 (double-float) double-float
519 (movable foldable flushable))
521 (defknown (%sin %cos %tanh %sin-quick %cos-quick)
522 (double-float) (double-float -1.0d0 1.0d0)
523 (movable foldable flushable))
525 (defknown (%asin %atan)
526 (double-float)
527 (double-float #.(coerce (- (/ pi 2)) 'double-float)
528 #.(coerce (/ pi 2) 'double-float))
529 (movable foldable flushable))
531 (defknown (%acos)
532 (double-float) (double-float 0.0d0 #.(coerce pi 'double-float))
533 (movable foldable flushable))
535 (defknown (%cosh)
536 (double-float) (double-float 1.0d0)
537 (movable foldable flushable))
539 (defknown (%acosh %exp %sqrt)
540 (double-float) (double-float 0.0d0)
541 (movable foldable flushable))
543 (defknown %expm1
544 (double-float) (double-float -1d0)
545 (movable foldable flushable))
547 (defknown (%hypot)
548 (double-float double-float) (double-float 0d0)
549 (movable foldable flushable))
551 (defknown (%pow)
552 (double-float double-float) double-float
553 (movable foldable flushable))
555 (defknown (%atan2)
556 (double-float double-float)
557 (double-float #.(coerce (- pi) 'double-float)
558 #.(coerce pi 'double-float))
559 (movable foldable flushable))
561 (defknown (%scalb)
562 (double-float double-float) double-float
563 (movable foldable flushable))
565 (defknown (%scalbn)
566 (double-float (signed-byte 32)) double-float
567 (movable foldable flushable))
569 (defknown (%log1p)
570 (double-float) double-float
571 (movable foldable flushable))
573 (macrolet ((def (name prim rtype)
574 `(progn
575 (deftransform ,name ((x) (single-float) ,rtype)
576 `(coerce (,',prim (coerce x 'double-float)) 'single-float))
577 (deftransform ,name ((x) (double-float) ,rtype)
578 `(,',prim x)))))
579 (def exp %exp *)
580 (def log %log float)
581 (def sqrt %sqrt float)
582 (def asin %asin float)
583 (def acos %acos float)
584 (def atan %atan *)
585 (def sinh %sinh *)
586 (def cosh %cosh *)
587 (def tanh %tanh *)
588 (def asinh %asinh *)
589 (def acosh %acosh float)
590 (def atanh %atanh float))
592 ;;; The argument range is limited on the x86 FP trig. functions. A
593 ;;; post-test can detect a failure (and load a suitable result), but
594 ;;; this test is avoided if possible.
595 (macrolet ((def (name prim prim-quick)
596 (declare (ignorable prim-quick))
597 `(progn
598 (deftransform ,name ((x) (single-float) *)
599 #!+x86 (cond ((csubtypep (lvar-type x)
600 (specifier-type '(single-float
601 (#.(- (expt 2f0 63)))
602 (#.(expt 2f0 63)))))
603 `(coerce (,',prim-quick (coerce x 'double-float))
604 'single-float))
606 (compiler-notify
607 "unable to avoid inline argument range check~@
608 because the argument range (~S) was not within 2^63"
609 (type-specifier (lvar-type x)))
610 `(coerce (,',prim (coerce x 'double-float)) 'single-float)))
611 #!-x86 `(coerce (,',prim (coerce x 'double-float)) 'single-float))
612 (deftransform ,name ((x) (double-float) *)
613 #!+x86 (cond ((csubtypep (lvar-type x)
614 (specifier-type '(double-float
615 (#.(- (expt 2d0 63)))
616 (#.(expt 2d0 63)))))
617 `(,',prim-quick x))
619 (compiler-notify
620 "unable to avoid inline argument range check~@
621 because the argument range (~S) was not within 2^63"
622 (type-specifier (lvar-type x)))
623 `(,',prim x)))
624 #!-x86 `(,',prim x)))))
625 (def sin %sin %sin-quick)
626 (def cos %cos %cos-quick)
627 (def tan %tan %tan-quick))
629 (deftransform atan ((x y) (single-float single-float) *)
630 `(coerce (%atan2 (coerce x 'double-float) (coerce y 'double-float))
631 'single-float))
632 (deftransform atan ((x y) (double-float double-float) *)
633 `(%atan2 x y))
635 (deftransform expt ((x y) ((single-float 0f0) single-float) *)
636 `(coerce (%pow (coerce x 'double-float) (coerce y 'double-float))
637 'single-float))
638 (deftransform expt ((x y) ((double-float 0d0) double-float) *)
639 `(%pow x y))
640 (deftransform expt ((x y) ((single-float 0f0) (signed-byte 32)) *)
641 `(coerce (%pow (coerce x 'double-float) (coerce y 'double-float))
642 'single-float))
643 (deftransform expt ((x y) ((double-float 0d0) (signed-byte 32)) *)
644 `(%pow x (coerce y 'double-float)))
646 ;;; ANSI says log with base zero returns zero.
647 (deftransform log ((x y) (float float) float)
648 '(if (zerop y) y (/ (log x) (log y))))
650 ;;; Handle some simple transformations.
652 (deftransform abs ((x) ((complex double-float)) double-float)
653 '(%hypot (realpart x) (imagpart x)))
655 (deftransform abs ((x) ((complex single-float)) single-float)
656 '(coerce (%hypot (coerce (realpart x) 'double-float)
657 (coerce (imagpart x) 'double-float))
658 'single-float))
660 (deftransform phase ((x) ((complex double-float)) double-float)
661 '(%atan2 (imagpart x) (realpart x)))
663 (deftransform phase ((x) ((complex single-float)) single-float)
664 '(coerce (%atan2 (coerce (imagpart x) 'double-float)
665 (coerce (realpart x) 'double-float))
666 'single-float))
668 (deftransform phase ((x) ((float)) float)
669 '(if (minusp (float-sign x))
670 (float pi x)
671 (float 0 x)))
673 ;;; The number is of type REAL.
674 (defun numeric-type-real-p (type)
675 (and (numeric-type-p type)
676 (eq (numeric-type-complexp type) :real)))
678 ;;; Coerce a numeric type bound to the given type while handling
679 ;;; exclusive bounds.
680 (defun coerce-numeric-bound (bound type)
681 (when bound
682 (if (consp bound)
683 (list (coerce (car bound) type))
684 (coerce bound type))))
686 #-sb-xc-host ; (See CROSS-FLOAT-INFINITY-KLUDGE.)
687 (progn
689 ;;;; optimizers for elementary functions
690 ;;;;
691 ;;;; These optimizers compute the output range of the elementary
692 ;;;; function, based on the domain of the input.
694 ;;; Generate a specifier for a complex type specialized to the same
695 ;;; type as the argument.
696 (defun complex-float-type (arg)
697 (declare (type numeric-type arg))
698 (let* ((format (case (numeric-type-class arg)
699 ((integer rational) 'single-float)
700 (t (numeric-type-format arg))))
701 (float-type (or format 'float)))
702 (specifier-type `(complex ,float-type))))
704 ;;; Compute a specifier like '(OR FLOAT (COMPLEX FLOAT)), except float
705 ;;; should be the right kind of float. Allow bounds for the float
706 ;;; part too.
707 (defun float-or-complex-float-type (arg &optional lo hi)
708 (declare (type numeric-type arg))
709 (let* ((format (case (numeric-type-class arg)
710 ((integer rational) 'single-float)
711 (t (numeric-type-format arg))))
712 (float-type (or format 'float))
713 (lo (coerce-numeric-bound lo float-type))
714 (hi (coerce-numeric-bound hi float-type)))
715 (specifier-type `(or (,float-type ,(or lo '*) ,(or hi '*))
716 (complex ,float-type)))))
718 ) ; PROGN
720 (eval-when (:compile-toplevel :execute)
721 ;; So the problem with this hack is that it's actually broken. If
722 ;; the host does not have long floats, then setting *R-D-F-F* to
723 ;; LONG-FLOAT doesn't actually buy us anything. FIXME.
724 (setf *read-default-float-format*
725 #!+long-float 'long-float #!-long-float 'double-float))
726 ;;; Test whether the numeric-type ARG is within the domain specified by
727 ;;; DOMAIN-LOW and DOMAIN-HIGH, consider negative and positive zero to
728 ;;; be distinct.
729 #-sb-xc-host ; (See CROSS-FLOAT-INFINITY-KLUDGE.)
730 (defun domain-subtypep (arg domain-low domain-high)
731 (declare (type numeric-type arg)
732 (type (or real null) domain-low domain-high))
733 (let* ((arg-lo (numeric-type-low arg))
734 (arg-lo-val (type-bound-number arg-lo))
735 (arg-hi (numeric-type-high arg))
736 (arg-hi-val (type-bound-number arg-hi)))
737 ;; Check that the ARG bounds are correctly canonicalized.
738 (when (and arg-lo (floatp arg-lo-val) (zerop arg-lo-val) (consp arg-lo)
739 (minusp (float-sign arg-lo-val)))
740 (compiler-notify "float zero bound ~S not correctly canonicalized?" arg-lo)
741 (setq arg-lo 0e0 arg-lo-val arg-lo))
742 (when (and arg-hi (zerop arg-hi-val) (floatp arg-hi-val) (consp arg-hi)
743 (plusp (float-sign arg-hi-val)))
744 (compiler-notify "float zero bound ~S not correctly canonicalized?" arg-hi)
745 (setq arg-hi (ecase *read-default-float-format*
746 (double-float (load-time-value (make-unportable-float :double-float-negative-zero)))
747 #!+long-float
748 (long-float (load-time-value (make-unportable-float :long-float-negative-zero))))
749 arg-hi-val arg-hi))
750 (flet ((fp-neg-zero-p (f) ; Is F -0.0?
751 (and (floatp f) (zerop f) (minusp (float-sign f))))
752 (fp-pos-zero-p (f) ; Is F +0.0?
753 (and (floatp f) (zerop f) (plusp (float-sign f)))))
754 (and (or (null domain-low)
755 (and arg-lo (>= arg-lo-val domain-low)
756 (not (and (fp-pos-zero-p domain-low)
757 (fp-neg-zero-p arg-lo)))))
758 (or (null domain-high)
759 (and arg-hi (<= arg-hi-val domain-high)
760 (not (and (fp-neg-zero-p domain-high)
761 (fp-pos-zero-p arg-hi)))))))))
762 (eval-when (:compile-toplevel :execute)
763 (setf *read-default-float-format* 'single-float))
765 ;;; The basic interval type. It can handle open and closed intervals.
766 ;;; A bound is open if it is a list containing a number, just like
767 ;;; Lisp says. NIL means unbounded.
768 (defstruct (interval (:constructor %make-interval (low high))
769 (:copier nil))
770 low high)
772 #-sb-xc-host ; (See CROSS-FLOAT-INFINITY-KLUDGE.)
773 (progn
775 ;;; Handle monotonic functions of a single variable whose domain is
776 ;;; possibly part of the real line. ARG is the variable, FUN is the
777 ;;; function, and DOMAIN is a specifier that gives the (real) domain
778 ;;; of the function. If ARG is a subset of the DOMAIN, we compute the
779 ;;; bounds directly. Otherwise, we compute the bounds for the
780 ;;; intersection between ARG and DOMAIN, and then append a complex
781 ;;; result, which occurs for the parts of ARG not in the DOMAIN.
783 ;;; Negative and positive zero are considered distinct within
784 ;;; DOMAIN-LOW and DOMAIN-HIGH.
786 ;;; DEFAULT-LOW and DEFAULT-HIGH are the lower and upper bounds if we
787 ;;; can't compute the bounds using FUN.
788 (defun elfun-derive-type-simple (arg fun domain-low domain-high
789 default-low default-high
790 &optional (increasingp t))
791 (declare (type (or null real) domain-low domain-high))
792 (etypecase arg
793 (numeric-type
794 (cond ((eq (numeric-type-complexp arg) :complex)
795 (complex-float-type arg))
796 ((numeric-type-real-p arg)
797 ;; The argument is real, so let's find the intersection
798 ;; between the argument and the domain of the function.
799 ;; We compute the bounds on the intersection, and for
800 ;; everything else, we return a complex number of the
801 ;; appropriate type.
802 (multiple-value-bind (intersection difference)
803 (interval-intersection/difference (numeric-type->interval arg)
804 (make-interval
805 :low domain-low
806 :high domain-high))
807 (cond
808 (intersection
809 ;; Process the intersection.
810 (let* ((low (interval-low intersection))
811 (high (interval-high intersection))
812 (res-lo (or (bound-func fun (if increasingp low high) nil)
813 default-low))
814 (res-hi (or (bound-func fun (if increasingp high low) nil)
815 default-high))
816 (format (case (numeric-type-class arg)
817 ((integer rational) 'single-float)
818 (t (numeric-type-format arg))))
819 (bound-type (or format 'float))
820 (result-type
821 (make-numeric-type
822 :class 'float
823 :format format
824 :low (coerce-numeric-bound res-lo bound-type)
825 :high (coerce-numeric-bound res-hi bound-type))))
826 ;; If the ARG is a subset of the domain, we don't
827 ;; have to worry about the difference, because that
828 ;; can't occur.
829 (if (or (null difference)
830 ;; Check whether the arg is within the domain.
831 (domain-subtypep arg domain-low domain-high))
832 result-type
833 (list result-type
834 (specifier-type `(complex ,bound-type))))))
836 ;; No intersection so the result must be purely complex.
837 (complex-float-type arg)))))
839 (float-or-complex-float-type arg default-low default-high))))))
841 (macrolet
842 ((frob (name domain-low domain-high def-low-bnd def-high-bnd
843 &key (increasingp t))
844 (let ((num (gensym)))
845 `(defoptimizer (,name derive-type) ((,num))
846 (one-arg-derive-type
847 ,num
848 (lambda (arg)
849 (elfun-derive-type-simple arg #',name
850 ,domain-low ,domain-high
851 ,def-low-bnd ,def-high-bnd
852 ,increasingp))
853 #',name)))))
854 ;; These functions are easy because they are defined for the whole
855 ;; real line.
856 (frob exp nil nil 0 nil)
857 (frob sinh nil nil nil nil)
858 (frob tanh nil nil -1 1)
859 (frob asinh nil nil nil nil)
861 ;; These functions are only defined for part of the real line. The
862 ;; condition selects the desired part of the line.
863 (frob asin -1d0 1d0 (- (/ pi 2)) (/ pi 2))
864 ;; Acos is monotonic decreasing, so we need to swap the function
865 ;; values at the lower and upper bounds of the input domain.
866 (frob acos -1d0 1d0 0 pi :increasingp nil)
867 (frob acosh 1d0 nil nil nil)
868 (frob atanh -1d0 1d0 -1 1)
869 ;; Kahan says that (sqrt -0.0) is -0.0, so use a specifier that
870 ;; includes -0.0.
871 (frob sqrt (load-time-value (make-unportable-float :double-float-negative-zero)) nil 0 nil))
873 ;;; Compute bounds for (expt x y). This should be easy since (expt x
874 ;;; y) = (exp (* y (log x))). However, computations done this way
875 ;;; have too much roundoff. Thus we have to do it the hard way.
876 (defun safe-expt (x y)
877 (handler-case
878 (when (< (abs y) 10000)
879 (expt x y))
880 (error ()
881 nil)))
883 ;;; Handle the case when x >= 1.
884 (defun interval-expt-> (x y)
885 (case (interval-range-info y 0d0)
887 ;; Y is positive and log X >= 0. The range of exp(y * log(x)) is
888 ;; obviously non-negative. We just have to be careful for
889 ;; infinite bounds (given by nil).
890 (let ((lo (safe-expt (type-bound-number (interval-low x))
891 (type-bound-number (interval-low y))))
892 (hi (safe-expt (type-bound-number (interval-high x))
893 (type-bound-number (interval-high y)))))
894 (list (make-interval :low (or lo 1) :high hi))))
896 ;; Y is negative and log x >= 0. The range of exp(y * log(x)) is
897 ;; obviously [0, 1]. However, underflow (nil) means 0 is the
898 ;; result.
899 (let ((lo (safe-expt (type-bound-number (interval-high x))
900 (type-bound-number (interval-low y))))
901 (hi (safe-expt (type-bound-number (interval-low x))
902 (type-bound-number (interval-high y)))))
903 (list (make-interval :low (or lo 0) :high (or hi 1)))))
905 ;; Split the interval in half.
906 (destructuring-bind (y- y+)
907 (interval-split 0 y t)
908 (list (interval-expt-> x y-)
909 (interval-expt-> x y+))))))
911 ;;; Handle the case when x <= 1
912 (defun interval-expt-< (x y)
913 (case (interval-range-info x 0d0)
915 ;; The case of 0 <= x <= 1 is easy
916 (case (interval-range-info y)
918 ;; Y is positive and log X <= 0. The range of exp(y * log(x)) is
919 ;; obviously [0, 1]. We just have to be careful for infinite bounds
920 ;; (given by nil).
921 (let ((lo (safe-expt (type-bound-number (interval-low x))
922 (type-bound-number (interval-high y))))
923 (hi (safe-expt (type-bound-number (interval-high x))
924 (type-bound-number (interval-low y)))))
925 (list (make-interval :low (or lo 0) :high (or hi 1)))))
927 ;; Y is negative and log x <= 0. The range of exp(y * log(x)) is
928 ;; obviously [1, inf].
929 (let ((hi (safe-expt (type-bound-number (interval-low x))
930 (type-bound-number (interval-low y))))
931 (lo (safe-expt (type-bound-number (interval-high x))
932 (type-bound-number (interval-high y)))))
933 (list (make-interval :low (or lo 1) :high hi))))
935 ;; Split the interval in half
936 (destructuring-bind (y- y+)
937 (interval-split 0 y t)
938 (list (interval-expt-< x y-)
939 (interval-expt-< x y+))))))
941 ;; The case where x <= 0. Y MUST be an INTEGER for this to work!
942 ;; The calling function must insure this! For now we'll just
943 ;; return the appropriate unbounded float type.
944 (list (make-interval :low nil :high nil)))
946 (destructuring-bind (neg pos)
947 (interval-split 0 x t t)
948 (list (interval-expt-< neg y)
949 (interval-expt-< pos y))))))
951 ;;; Compute bounds for (expt x y).
952 (defun interval-expt (x y)
953 (case (interval-range-info x 1)
955 ;; X >= 1
956 (interval-expt-> x y))
958 ;; X <= 1
959 (interval-expt-< x y))
961 (destructuring-bind (left right)
962 (interval-split 1 x t t)
963 (list (interval-expt left y)
964 (interval-expt right y))))))
966 (defun fixup-interval-expt (bnd x-int y-int x-type y-type)
967 (declare (ignore x-int))
968 ;; Figure out what the return type should be, given the argument
969 ;; types and bounds and the result type and bounds.
970 (cond ((csubtypep x-type (specifier-type 'integer))
971 ;; an integer to some power
972 (case (numeric-type-class y-type)
973 (integer
974 ;; Positive integer to an integer power is either an
975 ;; integer or a rational.
976 (let ((lo (or (interval-low bnd) '*))
977 (hi (or (interval-high bnd) '*)))
978 (if (and (interval-low y-int)
979 (>= (type-bound-number (interval-low y-int)) 0))
980 (specifier-type `(integer ,lo ,hi))
981 (specifier-type `(rational ,lo ,hi)))))
982 (rational
983 ;; Positive integer to rational power is either a rational
984 ;; or a single-float.
985 (let* ((lo (interval-low bnd))
986 (hi (interval-high bnd))
987 (int-lo (if lo
988 (floor (type-bound-number lo))
989 '*))
990 (int-hi (if hi
991 (ceiling (type-bound-number hi))
992 '*))
993 (f-lo (or (bound-func #'float lo nil)
994 '*))
995 (f-hi (or (bound-func #'float hi nil)
996 '*)))
997 (specifier-type `(or (rational ,int-lo ,int-hi)
998 (single-float ,f-lo, f-hi)))))
999 (float
1000 ;; A positive integer to a float power is a float.
1001 (let ((format (numeric-type-format y-type)))
1002 (aver format)
1003 (modified-numeric-type
1004 y-type
1005 :low (coerce-numeric-bound (interval-low bnd) format)
1006 :high (coerce-numeric-bound (interval-high bnd) format))))
1008 ;; A positive integer to a number is a number (for now).
1009 (specifier-type 'number))))
1010 ((csubtypep x-type (specifier-type 'rational))
1011 ;; a rational to some power
1012 (case (numeric-type-class y-type)
1013 (integer
1014 ;; A positive rational to an integer power is always a rational.
1015 (specifier-type `(rational ,(or (interval-low bnd) '*)
1016 ,(or (interval-high bnd) '*))))
1017 (rational
1018 ;; A positive rational to rational power is either a rational
1019 ;; or a single-float.
1020 (let* ((lo (interval-low bnd))
1021 (hi (interval-high bnd))
1022 (int-lo (if lo
1023 (floor (type-bound-number lo))
1024 '*))
1025 (int-hi (if hi
1026 (ceiling (type-bound-number hi))
1027 '*))
1028 (f-lo (or (bound-func #'float lo nil)
1029 '*))
1030 (f-hi (or (bound-func #'float hi nil)
1031 '*)))
1032 (specifier-type `(or (rational ,int-lo ,int-hi)
1033 (single-float ,f-lo, f-hi)))))
1034 (float
1035 ;; A positive rational to a float power is a float.
1036 (let ((format (numeric-type-format y-type)))
1037 (aver format)
1038 (modified-numeric-type
1039 y-type
1040 :low (coerce-numeric-bound (interval-low bnd) format)
1041 :high (coerce-numeric-bound (interval-high bnd) format))))
1043 ;; A positive rational to a number is a number (for now).
1044 (specifier-type 'number))))
1045 ((csubtypep x-type (specifier-type 'float))
1046 ;; a float to some power
1047 (case (numeric-type-class y-type)
1048 ((or integer rational)
1049 ;; A positive float to an integer or rational power is
1050 ;; always a float.
1051 (let ((format (numeric-type-format x-type)))
1052 (aver format)
1053 (make-numeric-type
1054 :class 'float
1055 :format format
1056 :low (coerce-numeric-bound (interval-low bnd) format)
1057 :high (coerce-numeric-bound (interval-high bnd) format))))
1058 (float
1059 ;; A positive float to a float power is a float of the
1060 ;; higher type.
1061 (let ((format (float-format-max (numeric-type-format x-type)
1062 (numeric-type-format y-type))))
1063 (aver format)
1064 (make-numeric-type
1065 :class 'float
1066 :format format
1067 :low (coerce-numeric-bound (interval-low bnd) format)
1068 :high (coerce-numeric-bound (interval-high bnd) format))))
1070 ;; A positive float to a number is a number (for now)
1071 (specifier-type 'number))))
1073 ;; A number to some power is a number.
1074 (specifier-type 'number))))
1076 (defun merged-interval-expt (x y)
1077 (let* ((x-int (numeric-type->interval x))
1078 (y-int (numeric-type->interval y)))
1079 (mapcar (lambda (type)
1080 (fixup-interval-expt type x-int y-int x y))
1081 (flatten-list (interval-expt x-int y-int)))))
1083 (defun expt-derive-type-aux (x y same-arg)
1084 (declare (ignore same-arg))
1085 (cond ((or (not (numeric-type-real-p x))
1086 (not (numeric-type-real-p y)))
1087 ;; Use numeric contagion if either is not real.
1088 (numeric-contagion x y))
1089 ((csubtypep y (specifier-type 'integer))
1090 ;; A real raised to an integer power is well-defined.
1091 (merged-interval-expt x y))
1092 ;; A real raised to a non-integral power can be a float or a
1093 ;; complex number.
1094 ((or (csubtypep x (specifier-type '(rational 0)))
1095 (csubtypep x (specifier-type '(float (0d0)))))
1096 ;; But a positive real to any power is well-defined.
1097 (merged-interval-expt x y))
1098 ((and (csubtypep x (specifier-type 'rational))
1099 (csubtypep y (specifier-type 'rational)))
1100 ;; A rational to the power of a rational could be a rational
1101 ;; or a possibly-complex single float
1102 (specifier-type '(or rational single-float (complex single-float))))
1104 ;; a real to some power. The result could be a real or a
1105 ;; complex.
1106 (float-or-complex-float-type (numeric-contagion x y)))))
1108 (defoptimizer (expt derive-type) ((x y))
1109 (two-arg-derive-type x y #'expt-derive-type-aux #'expt))
1111 ;;; Note we must assume that a type including 0.0 may also include
1112 ;;; -0.0 and thus the result may be complex -infinity + i*pi.
1113 (defun log-derive-type-aux-1 (x)
1114 (elfun-derive-type-simple x #'log 0d0 nil nil nil))
1116 (defun log-derive-type-aux-2 (x y same-arg)
1117 (let ((log-x (log-derive-type-aux-1 x))
1118 (log-y (log-derive-type-aux-1 y))
1119 (accumulated-list nil))
1120 ;; LOG-X or LOG-Y might be union types. We need to run through
1121 ;; the union types ourselves because /-DERIVE-TYPE-AUX doesn't.
1122 (dolist (x-type (prepare-arg-for-derive-type log-x))
1123 (dolist (y-type (prepare-arg-for-derive-type log-y))
1124 (push (/-derive-type-aux x-type y-type same-arg) accumulated-list)))
1125 (apply #'type-union (flatten-list accumulated-list))))
1127 (defoptimizer (log derive-type) ((x &optional y))
1128 (if y
1129 (two-arg-derive-type x y #'log-derive-type-aux-2 #'log)
1130 (one-arg-derive-type x #'log-derive-type-aux-1 #'log)))
1132 (defun atan-derive-type-aux-1 (y)
1133 (elfun-derive-type-simple y #'atan nil nil (- (/ pi 2)) (/ pi 2)))
1135 (defun atan-derive-type-aux-2 (y x same-arg)
1136 (declare (ignore same-arg))
1137 ;; The hard case with two args. We just return the max bounds.
1138 (let ((result-type (numeric-contagion y x)))
1139 (cond ((and (numeric-type-real-p x)
1140 (numeric-type-real-p y))
1141 (let* (;; FIXME: This expression for FORMAT seems to
1142 ;; appear multiple times, and should be factored out.
1143 (format (case (numeric-type-class result-type)
1144 ((integer rational) 'single-float)
1145 (t (numeric-type-format result-type))))
1146 (bound-format (or format 'float)))
1147 (make-numeric-type :class 'float
1148 :format format
1149 :complexp :real
1150 :low (coerce (- pi) bound-format)
1151 :high (coerce pi bound-format))))
1153 ;; The result is a float or a complex number
1154 (float-or-complex-float-type result-type)))))
1156 (defoptimizer (atan derive-type) ((y &optional x))
1157 (if x
1158 (two-arg-derive-type y x #'atan-derive-type-aux-2 #'atan)
1159 (one-arg-derive-type y #'atan-derive-type-aux-1 #'atan)))
1161 (defun cosh-derive-type-aux (x)
1162 ;; We note that cosh x = cosh |x| for all real x.
1163 (elfun-derive-type-simple
1164 (if (numeric-type-real-p x)
1165 (abs-derive-type-aux x)
1167 #'cosh nil nil 0 nil))
1169 (defoptimizer (cosh derive-type) ((num))
1170 (one-arg-derive-type num #'cosh-derive-type-aux #'cosh))
1172 (defun phase-derive-type-aux (arg)
1173 (let* ((format (case (numeric-type-class arg)
1174 ((integer rational) 'single-float)
1175 (t (numeric-type-format arg))))
1176 (bound-type (or format 'float)))
1177 (cond ((numeric-type-real-p arg)
1178 (case (interval-range-info (numeric-type->interval arg) 0.0)
1180 ;; The number is positive, so the phase is 0.
1181 (make-numeric-type :class 'float
1182 :format format
1183 :complexp :real
1184 :low (coerce 0 bound-type)
1185 :high (coerce 0 bound-type)))
1187 ;; The number is always negative, so the phase is pi.
1188 (make-numeric-type :class 'float
1189 :format format
1190 :complexp :real
1191 :low (coerce pi bound-type)
1192 :high (coerce pi bound-type)))
1194 ;; We can't tell. The result is 0 or pi. Use a union
1195 ;; type for this.
1196 (list
1197 (make-numeric-type :class 'float
1198 :format format
1199 :complexp :real
1200 :low (coerce 0 bound-type)
1201 :high (coerce 0 bound-type))
1202 (make-numeric-type :class 'float
1203 :format format
1204 :complexp :real
1205 :low (coerce pi bound-type)
1206 :high (coerce pi bound-type))))))
1208 ;; We have a complex number. The answer is the range -pi
1209 ;; to pi. (-pi is included because we have -0.)
1210 (make-numeric-type :class 'float
1211 :format format
1212 :complexp :real
1213 :low (coerce (- pi) bound-type)
1214 :high (coerce pi bound-type))))))
1216 (defoptimizer (phase derive-type) ((num))
1217 (one-arg-derive-type num #'phase-derive-type-aux #'phase))
1219 ) ; PROGN
1221 (deftransform realpart ((x) ((complex rational)) *)
1222 '(%realpart x))
1223 (deftransform imagpart ((x) ((complex rational)) *)
1224 '(%imagpart x))
1226 ;;; Make REALPART and IMAGPART return the appropriate types. This
1227 ;;; should help a lot in optimized code.
1228 (defun realpart-derive-type-aux (type)
1229 (let ((class (numeric-type-class type))
1230 (format (numeric-type-format type)))
1231 (cond ((numeric-type-real-p type)
1232 ;; The realpart of a real has the same type and range as
1233 ;; the input.
1234 (make-numeric-type :class class
1235 :format format
1236 :complexp :real
1237 :low (numeric-type-low type)
1238 :high (numeric-type-high type)))
1240 ;; We have a complex number. The result has the same type
1241 ;; as the real part, except that it's real, not complex,
1242 ;; obviously.
1243 (make-numeric-type :class class
1244 :format format
1245 :complexp :real
1246 :low (numeric-type-low type)
1247 :high (numeric-type-high type))))))
1249 (defoptimizer (realpart derive-type) ((num))
1250 (one-arg-derive-type num #'realpart-derive-type-aux #'realpart))
1252 (defun imagpart-derive-type-aux (type)
1253 (let ((class (numeric-type-class type))
1254 (format (numeric-type-format type)))
1255 (cond ((numeric-type-real-p type)
1256 ;; The imagpart of a real has the same type as the input,
1257 ;; except that it's zero.
1258 (let ((bound-format (or format class 'real)))
1259 (make-numeric-type :class class
1260 :format format
1261 :complexp :real
1262 :low (coerce 0 bound-format)
1263 :high (coerce 0 bound-format))))
1265 ;; We have a complex number. The result has the same type as
1266 ;; the imaginary part, except that it's real, not complex,
1267 ;; obviously.
1268 (make-numeric-type :class class
1269 :format format
1270 :complexp :real
1271 :low (numeric-type-low type)
1272 :high (numeric-type-high type))))))
1274 (defoptimizer (imagpart derive-type) ((num))
1275 (one-arg-derive-type num #'imagpart-derive-type-aux #'imagpart))
1277 (defun complex-derive-type-aux-1 (re-type)
1278 (if (numeric-type-p re-type)
1279 (make-numeric-type :class (numeric-type-class re-type)
1280 :format (numeric-type-format re-type)
1281 :complexp (if (csubtypep re-type
1282 (specifier-type 'rational))
1283 :real
1284 :complex)
1285 :low (numeric-type-low re-type)
1286 :high (numeric-type-high re-type))
1287 (specifier-type 'complex)))
1289 (defun complex-derive-type-aux-2 (re-type im-type same-arg)
1290 (declare (ignore same-arg))
1291 (if (and (numeric-type-p re-type)
1292 (numeric-type-p im-type))
1293 ;; Need to check to make sure numeric-contagion returns the
1294 ;; right type for what we want here.
1296 ;; Also, what about rational canonicalization, like (complex 5 0)
1297 ;; is 5? So, if the result must be complex, we make it so.
1298 ;; If the result might be complex, which happens only if the
1299 ;; arguments are rational, we make it a union type of (or
1300 ;; rational (complex rational)).
1301 (let* ((element-type (numeric-contagion re-type im-type))
1302 (rat-result-p (csubtypep element-type
1303 (specifier-type 'rational))))
1304 (if rat-result-p
1305 (type-union element-type
1306 (specifier-type
1307 `(complex ,(numeric-type-class element-type))))
1308 (make-numeric-type :class (numeric-type-class element-type)
1309 :format (numeric-type-format element-type)
1310 :complexp (if rat-result-p
1311 :real
1312 :complex))))
1313 (specifier-type 'complex)))
1315 #-sb-xc-host ; (See CROSS-FLOAT-INFINITY-KLUDGE.)
1316 (defoptimizer (complex derive-type) ((re &optional im))
1317 (if im
1318 (two-arg-derive-type re im #'complex-derive-type-aux-2 #'complex)
1319 (one-arg-derive-type re #'complex-derive-type-aux-1 #'complex)))
1321 ;;; Define some transforms for complex operations. We do this in lieu
1322 ;;; of complex operation VOPs.
1323 (macrolet ((frob (type)
1324 `(progn
1325 (deftransform complex ((r) (,type))
1326 '(complex r ,(coerce 0 type)))
1327 (deftransform complex ((r i) (,type (and real (not ,type))))
1328 '(complex r (truly-the ,type (coerce i ',type))))
1329 (deftransform complex ((r i) ((and real (not ,type)) ,type))
1330 '(complex (truly-the ,type (coerce r ',type)) i))
1331 ;; negation
1332 #!-complex-float-vops
1333 (deftransform %negate ((z) ((complex ,type)) *)
1334 '(complex (%negate (realpart z)) (%negate (imagpart z))))
1335 ;; complex addition and subtraction
1336 #!-complex-float-vops
1337 (deftransform + ((w z) ((complex ,type) (complex ,type)) *)
1338 '(complex (+ (realpart w) (realpart z))
1339 (+ (imagpart w) (imagpart z))))
1340 #!-complex-float-vops
1341 (deftransform - ((w z) ((complex ,type) (complex ,type)) *)
1342 '(complex (- (realpart w) (realpart z))
1343 (- (imagpart w) (imagpart z))))
1344 ;; Add and subtract a complex and a real.
1345 #!-complex-float-vops
1346 (deftransform + ((w z) ((complex ,type) real) *)
1347 `(complex (+ (realpart w) z)
1348 (+ (imagpart w) ,(coerce 0 ',type))))
1349 #!-complex-float-vops
1350 (deftransform + ((z w) (real (complex ,type)) *)
1351 `(complex (+ (realpart w) z)
1352 (+ (imagpart w) ,(coerce 0 ',type))))
1353 ;; Add and subtract a real and a complex number.
1354 #!-complex-float-vops
1355 (deftransform - ((w z) ((complex ,type) real) *)
1356 `(complex (- (realpart w) z)
1357 (- (imagpart w) ,(coerce 0 ',type))))
1358 #!-complex-float-vops
1359 (deftransform - ((z w) (real (complex ,type)) *)
1360 `(complex (- z (realpart w))
1361 (- ,(coerce 0 ',type) (imagpart w))))
1362 ;; Multiply and divide two complex numbers.
1363 #!-complex-float-vops
1364 (deftransform * ((x y) ((complex ,type) (complex ,type)) *)
1365 '(let* ((rx (realpart x))
1366 (ix (imagpart x))
1367 (ry (realpart y))
1368 (iy (imagpart y)))
1369 (complex (- (* rx ry) (* ix iy))
1370 (+ (* rx iy) (* ix ry)))))
1371 (deftransform / ((x y) ((complex ,type) (complex ,type)) *)
1372 #!-complex-float-vops
1373 '(let* ((rx (realpart x))
1374 (ix (imagpart x))
1375 (ry (realpart y))
1376 (iy (imagpart y)))
1377 (if (> (abs ry) (abs iy))
1378 (let* ((r (/ iy ry))
1379 (dn (+ ry (* r iy))))
1380 (complex (/ (+ rx (* ix r)) dn)
1381 (/ (- ix (* rx r)) dn)))
1382 (let* ((r (/ ry iy))
1383 (dn (+ iy (* r ry))))
1384 (complex (/ (+ (* rx r) ix) dn)
1385 (/ (- (* ix r) rx) dn)))))
1386 #!+complex-float-vops
1387 `(let* ((cs (conjugate (sb!vm::swap-complex x)))
1388 (ry (realpart y))
1389 (iy (imagpart y)))
1390 (if (> (abs ry) (abs iy))
1391 (let* ((r (/ iy ry))
1392 (dn (+ ry (* r iy))))
1393 (/ (+ x (* cs r)) dn))
1394 (let* ((r (/ ry iy))
1395 (dn (+ iy (* r ry))))
1396 (/ (+ (* x r) cs) dn)))))
1397 ;; Multiply a complex by a real or vice versa.
1398 #!-complex-float-vops
1399 (deftransform * ((w z) ((complex ,type) real) *)
1400 '(complex (* (realpart w) z) (* (imagpart w) z)))
1401 #!-complex-float-vops
1402 (deftransform * ((z w) (real (complex ,type)) *)
1403 '(complex (* (realpart w) z) (* (imagpart w) z)))
1404 ;; Divide a complex by a real or vice versa.
1405 #!-complex-float-vops
1406 (deftransform / ((w z) ((complex ,type) real) *)
1407 '(complex (/ (realpart w) z) (/ (imagpart w) z)))
1408 (deftransform / ((x y) (,type (complex ,type)) *)
1409 #!-complex-float-vops
1410 '(let* ((ry (realpart y))
1411 (iy (imagpart y)))
1412 (if (> (abs ry) (abs iy))
1413 (let* ((r (/ iy ry))
1414 (dn (+ ry (* r iy))))
1415 (complex (/ x dn)
1416 (/ (- (* x r)) dn)))
1417 (let* ((r (/ ry iy))
1418 (dn (+ iy (* r ry))))
1419 (complex (/ (* x r) dn)
1420 (/ (- x) dn)))))
1421 #!+complex-float-vops
1422 '(let* ((ry (realpart y))
1423 (iy (imagpart y)))
1424 (if (> (abs ry) (abs iy))
1425 (let* ((r (/ iy ry))
1426 (dn (+ ry (* r iy))))
1427 (/ (complex x (- (* x r))) dn))
1428 (let* ((r (/ ry iy))
1429 (dn (+ iy (* r ry))))
1430 (/ (complex (* x r) (- x)) dn)))))
1431 ;; conjugate of complex number
1432 #!-complex-float-vops
1433 (deftransform conjugate ((z) ((complex ,type)) *)
1434 '(complex (realpart z) (- (imagpart z))))
1435 ;; CIS
1436 (deftransform cis ((z) ((,type)) *)
1437 '(complex (cos z) (sin z)))
1438 ;; comparison
1439 #!-complex-float-vops
1440 (deftransform = ((w z) ((complex ,type) (complex ,type)) *)
1441 '(and (= (realpart w) (realpart z))
1442 (= (imagpart w) (imagpart z))))
1443 #!-complex-float-vops
1444 (deftransform = ((w z) ((complex ,type) real) *)
1445 '(and (= (realpart w) z) (zerop (imagpart w))))
1446 #!-complex-float-vops
1447 (deftransform = ((w z) (real (complex ,type)) *)
1448 '(and (= (realpart z) w) (zerop (imagpart z)))))))
1450 (frob single-float)
1451 (frob double-float))
1453 ;;; Here are simple optimizers for SIN, COS, and TAN. They do not
1454 ;;; produce a minimal range for the result; the result is the widest
1455 ;;; possible answer. This gets around the problem of doing range
1456 ;;; reduction correctly but still provides useful results when the
1457 ;;; inputs are union types.
1458 #-sb-xc-host ; (See CROSS-FLOAT-INFINITY-KLUDGE.)
1459 (progn
1460 (defun trig-derive-type-aux (arg domain fun
1461 &optional def-lo def-hi (increasingp t))
1462 (etypecase arg
1463 (numeric-type
1464 (cond ((eq (numeric-type-complexp arg) :complex)
1465 (make-numeric-type :class (numeric-type-class arg)
1466 :format (numeric-type-format arg)
1467 :complexp :complex))
1468 ((numeric-type-real-p arg)
1469 (let* ((format (case (numeric-type-class arg)
1470 ((integer rational) 'single-float)
1471 (t (numeric-type-format arg))))
1472 (bound-type (or format 'float)))
1473 ;; If the argument is a subset of the "principal" domain
1474 ;; of the function, we can compute the bounds because
1475 ;; the function is monotonic. We can't do this in
1476 ;; general for these periodic functions because we can't
1477 ;; (and don't want to) do the argument reduction in
1478 ;; exactly the same way as the functions themselves do
1479 ;; it.
1480 (if (csubtypep arg domain)
1481 (let ((res-lo (bound-func fun (numeric-type-low arg) nil))
1482 (res-hi (bound-func fun (numeric-type-high arg) nil)))
1483 (unless increasingp
1484 (rotatef res-lo res-hi))
1485 (make-numeric-type
1486 :class 'float
1487 :format format
1488 :low (coerce-numeric-bound res-lo bound-type)
1489 :high (coerce-numeric-bound res-hi bound-type)))
1490 (make-numeric-type
1491 :class 'float
1492 :format format
1493 :low (and def-lo (coerce def-lo bound-type))
1494 :high (and def-hi (coerce def-hi bound-type))))))
1496 (float-or-complex-float-type arg def-lo def-hi))))))
1498 (defoptimizer (sin derive-type) ((num))
1499 (one-arg-derive-type
1501 (lambda (arg)
1502 ;; Derive the bounds if the arg is in [-pi/2, pi/2].
1503 (trig-derive-type-aux
1505 (specifier-type `(float ,(- (/ pi 2)) ,(/ pi 2)))
1506 #'sin
1507 -1 1))
1508 #'sin))
1510 (defoptimizer (cos derive-type) ((num))
1511 (one-arg-derive-type
1513 (lambda (arg)
1514 ;; Derive the bounds if the arg is in [0, pi].
1515 (trig-derive-type-aux arg
1516 (specifier-type `(float 0d0 ,pi))
1517 #'cos
1518 -1 1
1519 nil))
1520 #'cos))
1522 (defoptimizer (tan derive-type) ((num))
1523 (one-arg-derive-type
1525 (lambda (arg)
1526 ;; Derive the bounds if the arg is in [-pi/2, pi/2].
1527 (trig-derive-type-aux arg
1528 (specifier-type `(float ,(- (/ pi 2)) ,(/ pi 2)))
1529 #'tan
1530 nil nil))
1531 #'tan))
1533 (defoptimizer (conjugate derive-type) ((num))
1534 (one-arg-derive-type num
1535 (lambda (arg)
1536 (flet ((most-negative-bound (l h)
1537 (and l h
1538 (if (< (type-bound-number l) (- (type-bound-number h)))
1540 (set-bound (- (type-bound-number h)) (consp h)))))
1541 (most-positive-bound (l h)
1542 (and l h
1543 (if (> (type-bound-number h) (- (type-bound-number l)))
1545 (set-bound (- (type-bound-number l)) (consp l))))))
1546 (if (numeric-type-real-p arg)
1547 (lvar-type num)
1548 (let ((low (numeric-type-low arg))
1549 (high (numeric-type-high arg)))
1550 (let ((new-low (most-negative-bound low high))
1551 (new-high (most-positive-bound low high)))
1552 (modified-numeric-type arg :low new-low :high new-high))))))
1553 #'conjugate))
1555 (defoptimizer (cis derive-type) ((num))
1556 (one-arg-derive-type num
1557 (lambda (arg)
1558 (specifier-type
1559 `(complex ,(or (numeric-type-format arg) 'float))))
1560 #'cis))
1562 ) ; PROGN
1564 ;;;; TRUNCATE, FLOOR, CEILING, and ROUND
1566 (macrolet ((define-frobs (fun ufun)
1567 `(progn
1568 (defknown ,ufun (real) integer (movable foldable flushable))
1569 (deftransform ,fun ((x &optional by)
1570 (* &optional
1571 (constant-arg (member 1))))
1572 '(let ((res (,ufun x)))
1573 (values res (- x res)))))))
1574 (define-frobs truncate %unary-truncate)
1575 (define-frobs round %unary-round))
1577 (deftransform %unary-truncate ((x) (single-float))
1578 `(%unary-truncate/single-float x))
1579 (deftransform %unary-truncate ((x) (double-float))
1580 `(%unary-truncate/double-float x))
1582 ;;; Convert (TRUNCATE x y) to the obvious implementation.
1584 ;;; ...plus hair: Insert explicit coercions to appropriate float types: Python
1585 ;;; is reluctant it generate explicit integer->float coercions due to
1586 ;;; precision issues (see SAFE-SINGLE-COERCION-P &co), but this is not an
1587 ;;; issue here as there is no DERIVE-TYPE optimizer on specialized versions of
1588 ;;; %UNARY-TRUNCATE, so the derived type of TRUNCATE remains the best we can
1589 ;;; do here -- which is fine. Also take care not to add unnecassary division
1590 ;;; or multiplication by 1, since we are not able to always eliminate them,
1591 ;;; depending on FLOAT-ACCURACY. Finally, leave out the secondary value when
1592 ;;; we know it is unused: COERCE is not flushable.
1593 (macrolet ((def (type other-float-arg-types)
1594 (let ((unary (symbolicate "%UNARY-TRUNCATE/" type))
1595 (coerce (symbolicate "%" type)))
1596 `(deftransform truncate ((x &optional y)
1597 (,type
1598 &optional (or ,type ,@other-float-arg-types integer))
1599 * :result result)
1600 (let* ((result-type (and result
1601 (lvar-derived-type result)))
1602 (compute-all (and (values-type-p result-type)
1603 (not (type-single-value-p result-type)))))
1604 (if (or (not y)
1605 (and (constant-lvar-p y) (= 1 (lvar-value y))))
1606 (if compute-all
1607 `(let ((res (,',unary x)))
1608 (values res (- x (,',coerce res))))
1609 `(let ((res (,',unary x)))
1610 ;; Dummy secondary value!
1611 (values res x)))
1612 (if compute-all
1613 `(let* ((f (,',coerce y))
1614 (res (,',unary (/ x f))))
1615 (values res (- x (* f (,',coerce res)))))
1616 `(let* ((f (,',coerce y))
1617 (res (,',unary (/ x f))))
1618 ;; Dummy secondary value!
1619 (values res x)))))))))
1620 (def single-float ())
1621 (def double-float (single-float)))
1623 (defknown %unary-ftruncate (real) float (movable foldable flushable))
1624 (defknown %unary-ftruncate/single (single-float) single-float
1625 (movable foldable flushable))
1626 (defknown %unary-ftruncate/double (double-float) double-float
1627 (movable foldable flushable))
1629 #-sb-xc-host
1630 (defun %unary-ftruncate/single (x)
1631 (declare (muffle-conditions t))
1632 (declare (type single-float x))
1633 (declare (optimize speed (safety 0)))
1634 (let* ((bits (single-float-bits x))
1635 (exp (ldb sb!vm:single-float-exponent-byte bits))
1636 (biased (the single-float-exponent
1637 (- exp sb!vm:single-float-bias))))
1638 (declare (type (signed-byte 32) bits))
1639 (cond
1640 ((= exp sb!vm:single-float-normal-exponent-max) x)
1641 ((<= biased 0) (* x 0f0))
1642 ((>= biased (float-digits x)) x)
1644 (let ((frac-bits (- (float-digits x) biased)))
1645 (setf bits (logandc2 bits (- (ash 1 frac-bits) 1)))
1646 (make-single-float bits))))))
1648 #-sb-xc-host
1649 (defun %unary-ftruncate/double (x)
1650 (declare (muffle-conditions t))
1651 (declare (type double-float x))
1652 (declare (optimize speed (safety 0)))
1653 (let* ((high (double-float-high-bits x))
1654 (low (double-float-low-bits x))
1655 (exp (ldb sb!vm:double-float-exponent-byte high))
1656 (biased (the double-float-exponent
1657 (- exp sb!vm:double-float-bias))))
1658 (declare (type (signed-byte 32) high)
1659 (type (unsigned-byte 32) low))
1660 (cond
1661 ((= exp sb!vm:double-float-normal-exponent-max) x)
1662 ((<= biased 0) (* x 0d0))
1663 ((>= biased (float-digits x)) x)
1665 (let ((frac-bits (- (float-digits x) biased)))
1666 (cond ((< frac-bits 32)
1667 (setf low (logandc2 low (- (ash 1 frac-bits) 1))))
1669 (setf low 0)
1670 (setf high (logandc2 high (- (ash 1 (- frac-bits 32)) 1)))))
1671 (make-double-float high low))))))
1673 (macrolet
1674 ((def (float-type fun)
1675 `(deftransform %unary-ftruncate ((x) (,float-type))
1676 '(,fun x))))
1677 (def single-float %unary-ftruncate/single)
1678 (def double-float %unary-ftruncate/double))