xref: /illumos-gate/usr/src/lib/libm/common/complex/cacos.c (revision 66582b606a8194f7f3ba5b3a3a6dca5b0d346361)
1 /*
2  * CDDL HEADER START
3  *
4  * The contents of this file are subject to the terms of the
5  * Common Development and Distribution License (the "License").
6  * You may not use this file except in compliance with the License.
7  *
8  * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE
9  * or http://www.opensolaris.org/os/licensing.
10  * See the License for the specific language governing permissions
11  * and limitations under the License.
12  *
13  * When distributing Covered Code, include this CDDL HEADER in each
14  * file and include the License file at usr/src/OPENSOLARIS.LICENSE.
15  * If applicable, add the following below this CDDL HEADER, with the
16  * fields enclosed by brackets "[]" replaced with your own identifying
17  * information: Portions Copyright [yyyy] [name of copyright owner]
18  *
19  * CDDL HEADER END
20  */
21 
22 /*
23  * Copyright 2011 Nexenta Systems, Inc.  All rights reserved.
24  */
25 /*
26  * Copyright 2006 Sun Microsystems, Inc.  All rights reserved.
27  * Use is subject to license terms.
28  */
29 
30 #pragma weak __cacos = cacos
31 
32 /* INDENT OFF */
33 /*
34  * dcomplex cacos(dcomplex z);
35  *
36  * Alogrithm
37  * (based on T.E.Hull, Thomas F. Fairgrieve and Ping Tak Peter Tang's
38  * paper "Implementing the Complex Arcsine and Arccosine Functins Using
39  * Exception Handling", ACM TOMS, Vol 23, pp 299-335)
40  *
41  * The principal value of complex inverse cosine function cacos(z),
42  * where z = x+iy, can be defined by
43  *
44  * 	cacos(z) = acos(B) - i sign(y) log (A + sqrt(A*A-1)),
45  *
46  * where the log function is the natural log, and
47  *             ____________           ____________
48  *       1    /     2    2      1    /     2    2
49  *  A = ---  / (x+1)  + y   +  ---  / (x-1)  + y
50  *       2 \/                   2 \/
51  *             ____________           ____________
52  *       1    /     2    2      1    /     2    2
53  *  B = ---  / (x+1)  + y   -  ---  / (x-1)  + y   .
54  *       2 \/                   2 \/
55  *
56  * The Branch cuts are on the real line from -inf to -1 and from 1 to inf.
57  * The real and imaginary parts are based on Abramowitz and Stegun
58  * [Handbook of Mathematic Functions, 1972].  The sign of the imaginary
59  * part is chosen to be the generally considered the principal value of
60  * this function.
61  *
62  * Notes:1. A is the average of the distances from z to the points (1,0)
63  *          and (-1,0) in the complex z-plane, and in particular A>=1.
64  *       2. B is in [-1,1], and A*B = x
65  *
66  * Basic relations
67  *    cacos(conj(z)) = conj(cacos(z))
68  *    cacos(-z)      = pi   - cacos(z)
69  *    cacos( z)      = pi/2 - casin(z)
70  *
71  * Special cases (conform to ISO/IEC 9899:1999(E)):
72  *    cacos(+-0  + i y  ) = pi/2 - i y for y is +-0, +-inf, NaN
73  *    cacos( x   + i inf) = pi/2 - i inf for all x
74  *    cacos( x   + i NaN) = NaN  + i NaN with invalid for non-zero finite x
75  *    cacos(-inf + i y  ) = pi   - i inf for finite +y
76  *    cacos( inf + i y  ) = 0    - i inf for finite +y
77  *    cacos(-inf + i inf) = 3pi/4- i inf
78  *    cacos( inf + i inf) = pi/4 - i inf
79  *    cacos(+-inf+ i NaN) = NaN  - i inf (sign of imaginary is unspecified)
80  *    cacos(NaN  + i y  ) = NaN  + i NaN with invalid for finite y
81  *    cacos(NaN  + i inf) = NaN  - i inf
82  *    cacos(NaN  + i NaN) = NaN  + i NaN
83  *
84  * Special Regions (better formula for accuracy and for avoiding spurious
85  * overflow or underflow) (all x and y are assumed nonnegative):
86  *  case 1: y = 0
87  *  case 2: tiny y relative to x-1: y <= ulp(0.5)*|x-1|
88  *  case 3: tiny y: y < 4 sqrt(u), where u = minimum normal number
89  *  case 4: huge y relative to x+1: y >= (1+x)/ulp(0.5)
90  *  case 5: huge x and y: x and y >= sqrt(M)/8, where M = maximum normal number
91  *  case 6: tiny x: x < 4 sqrt(u)
92  *  --------
93  *  case	1 & 2. y=0 or y/|x-1| is tiny. We have
94  *             ____________              _____________
95  *            /      2    2             /       y    2
96  *           / (x+-1)  + y   =  |x+-1| / 1 + (------)
97  *         \/                        \/       |x+-1|
98  *
99  *                                            1     y    2
100  *                           ~  |x+-1| ( 1 + --- (------)  )
101  *                                            2   |x+-1|
102  *
103  *                                          2
104  *                                         y
105  *                           = |x+-1| + --------.
106  *                                      2|x+-1|
107  *
108  *	Consequently, it is not difficult to see that
109  *                                 2
110  *                                y
111  *                    [ 1 + ------------ ,     if x < 1,
112  *                    [      2(1+x)(1-x)
113  *                    [
114  *                    [
115  *                    [ x,                     if x = 1 (y = 0),
116  *                    [
117  *		A ~=  [             2
118  *                    [        x * y
119  *                    [ x + ------------ ~ x,  if x > 1
120  *                    [      2(x+1)(x-1)
121  *
122  *	and hence
123  *                      ______                                 2
124  *                     / 2                    y               y
125  *               A + \/ A  - 1  ~  1 + ---------------- + -----------, if x < 1,
126  *                                     sqrt((x+1)(1-x))   2(x+1)(1-x)
127  *
128  *
129  *			        ~  x + sqrt((x-1)*(x+1)),             if x >= 1.
130  *
131  *                                         2
132  *                                        y
133  *                          [ x(1 - -----------) ~ x,  if x < 1,
134  *                          [       2(1+x)(1-x)
135  *		B = x/A  ~  [
136  *                          [ 1,                       if x = 1,
137  *			    [
138  *                          [           2
139  *                          [          y
140  *                          [ 1 - ------------ ,       if x > 1,
141  *                          [      2(x+1)(x-1)
142  *	Thus
143  *                            [ acos(x) - i y/sqrt((x-1)*(x+1)),      if x < 1,
144  *                            [
145  *		cacos(x+i*y)~ [ 0 - i 0,                              if x = 1,
146  *                            [
147  *                            [ y/sqrt(x*x-1) - i log(x+sqrt(x*x-1)), if x > 1.
148  *
149  *      Note: y/sqrt(x*x-1) ~ y/x when x >= 2**26.
150  *  case 3. y < 4 sqrt(u), where u = minimum normal x.
151  *	After case 1 and 2, this will only occurs when x=1. When x=1, we have
152  *	   A = (sqrt(4+y*y)+y)/2 ~ 1 + y/2 + y^2/8 + ...
153  *	and
154  *	   B = 1/A = 1 - y/2 + y^2/8 + ...
155  * 	Since
156  *         cos(sqrt(y)) ~ 1 - y/2 + ...
157  *      we have, for the real part,
158  *         acos(B) ~ acos(1 - y/2) ~ sqrt(y)
159  *	For the imaginary part,
160  *	   log(A+sqrt(A*A-1)) ~ log(1+y/2+sqrt(2*y/2))
161  *	                      = log(1+y/2+sqrt(y))
162  *	                      = (y/2+sqrt(y)) - (y/2+sqrt(y))^2/2 + ...
163  *	                      ~ sqrt(y) - y*(sqrt(y)+y/2)/2
164  *	                      ~ sqrt(y)
165  *
166  *  case 4. y >= (x+1)/ulp(0.5). In this case, A ~ y and B ~ x/y. Thus
167  *	   real part = acos(B) ~ pi/2
168  * 	and
169  *	   imag part = log(y+sqrt(y*y-one))
170  *
171  *  case 5. Both x and y are large: x and y > sqrt(M)/8, where M = maximum x
172  *	In this case,
173  *	   A ~ sqrt(x*x+y*y)
174  *	   B ~ x/sqrt(x*x+y*y).
175  *	Thus
176  *	   real part = acos(B) = atan(y/x),
177  *	   imag part = log(A+sqrt(A*A-1)) ~ log(2A)
178  *	             = log(2) + 0.5*log(x*x+y*y)
179  *	             = log(2) + log(y) + 0.5*log(1+(x/y)^2)
180  *
181  *  case 6. x < 4 sqrt(u). In this case, we have
182  *	    A ~ sqrt(1+y*y), B = x/sqrt(1+y*y).
183  *	Since B is tiny, we have
184  *	    real part = acos(B) ~ pi/2
185  *	    imag part = log(A+sqrt(A*A-1)) = log (A+sqrt(y*y))
186  *	              = log(y+sqrt(1+y*y))
187  *	              = 0.5*log(y^2+2ysqrt(1+y^2)+1+y^2)
188  *	              = 0.5*log(1+2y(y+sqrt(1+y^2)));
189  *	              = 0.5*log1p(2y(y+A));
190  *
191  * 	cacos(z) = acos(B) - i sign(y) log (A + sqrt(A*A-1)),
192  */
193 /* INDENT ON */
194 
195 #include "libm.h"
196 #include "complex_wrapper.h"
197 
198 /* INDENT OFF */
199 static const double
200 	zero = 0.0,
201 	one = 1.0,
202 	E = 1.11022302462515654042e-16,			/* 2**-53 */
203 	ln2 = 6.93147180559945286227e-01,
204 	pi = 3.1415926535897931159979634685,
205 	pi_l = 1.224646799147353177e-16,
206 	pi_2 = 1.570796326794896558e+00,
207 	pi_2_l = 6.123233995736765886e-17,
208 	pi_4 = 0.78539816339744827899949,
209 	pi_4_l = 3.061616997868382943e-17,
210 	pi3_4 = 2.356194490192344836998,
211 	pi3_4_l = 9.184850993605148829195e-17,
212 	Foursqrtu = 5.96667258496016539463e-154,	/* 2**(-509) */
213 	Acrossover = 1.5,
214 	Bcrossover = 0.6417,
215 	half = 0.5;
216 /* INDENT ON */
217 
218 dcomplex
219 cacos(dcomplex z) {
220 	double x, y, t, R, S, A, Am1, B, y2, xm1, xp1, Apx;
221 	int ix, iy, hx, hy;
222 	unsigned lx, ly;
223 	dcomplex ans;
224 
225 	x = D_RE(z);
226 	y = D_IM(z);
227 	hx = HI_WORD(x);
228 	lx = LO_WORD(x);
229 	hy = HI_WORD(y);
230 	ly = LO_WORD(y);
231 	ix = hx & 0x7fffffff;
232 	iy = hy & 0x7fffffff;
233 
234 	/* x is 0 */
235 	if ((ix | lx) == 0) {
236 		if (((iy | ly) == 0) || (iy >= 0x7ff00000)) {
237 			D_RE(ans) = pi_2;
238 			D_IM(ans) = -y;
239 			return (ans);
240 		}
241 	}
242 
243 	/* |y| is inf or NaN */
244 	if (iy >= 0x7ff00000) {
245 		if (ISINF(iy, ly)) {	/* cacos(x + i inf) = pi/2  - i inf */
246 			D_IM(ans) = -y;
247 			if (ix < 0x7ff00000) {
248 				D_RE(ans) = pi_2 + pi_2_l;
249 			} else if (ISINF(ix, lx)) {
250 				if (hx >= 0)
251 					D_RE(ans) = pi_4 + pi_4_l;
252 				else
253 					D_RE(ans) = pi3_4 + pi3_4_l;
254 			} else {
255 				D_RE(ans) = x;
256 			}
257 		} else {		/* cacos(x + i NaN) = NaN  + i NaN */
258 			D_RE(ans) = y + x;
259 			if (ISINF(ix, lx))
260 				D_IM(ans) = -fabs(x);
261 			else
262 				D_IM(ans) = y;
263 		}
264 		return (ans);
265 	}
266 
267 	x = fabs(x);
268 	y = fabs(y);
269 
270 	/* x is inf or NaN */
271 	if (ix >= 0x7ff00000) {	/* x is inf or NaN */
272 		if (ISINF(ix, lx)) {	/* x is INF */
273 			D_IM(ans) = -x;
274 			if (iy >= 0x7ff00000) {
275 				if (ISINF(iy, ly)) {
276 					/* INDENT OFF */
277 					/* cacos(inf + i inf) = pi/4 - i inf */
278 					/* cacos(-inf+ i inf) =3pi/4 - i inf */
279 					/* INDENT ON */
280 					if (hx >= 0)
281 						D_RE(ans) = pi_4 + pi_4_l;
282 					else
283 						D_RE(ans) = pi3_4 + pi3_4_l;
284 				} else
285 					/* INDENT OFF */
286 					/* cacos(inf + i NaN) = NaN  - i inf  */
287 					/* INDENT ON */
288 					D_RE(ans) = y + y;
289 			} else
290 				/* INDENT OFF */
291 				/* cacos(inf + iy ) = 0  - i inf */
292 				/* cacos(-inf+ iy  ) = pi - i inf */
293 				/* INDENT ON */
294 			if (hx >= 0)
295 				D_RE(ans) = zero;
296 			else
297 				D_RE(ans) = pi + pi_l;
298 		} else {		/* x is NaN */
299 			/* INDENT OFF */
300 			/*
301 			 * cacos(NaN + i inf) = NaN  - i inf
302 			 * cacos(NaN + i y  ) = NaN  + i NaN
303 			 * cacos(NaN + i NaN) = NaN  + i NaN
304 			 */
305 			/* INDENT ON */
306 			D_RE(ans) = x + y;
307 			if (iy >= 0x7ff00000) {
308 				D_IM(ans) = -y;
309 			} else {
310 				D_IM(ans) = x;
311 			}
312 		}
313 		if (hy < 0)
314 			D_IM(ans) = -D_IM(ans);
315 		return (ans);
316 	}
317 
318 	if ((iy | ly) == 0) {	/* region 1: y=0 */
319 		if (ix < 0x3ff00000) {	/* |x| < 1 */
320 			D_RE(ans) = acos(x);
321 			D_IM(ans) = zero;
322 		} else {
323 			D_RE(ans) = zero;
324 			if (ix >= 0x43500000)	/* |x| >= 2**54 */
325 				D_IM(ans) = ln2 + log(x);
326 			else if (ix >= 0x3ff80000)	/* x > Acrossover */
327 				D_IM(ans) = log(x + sqrt((x - one) * (x +
328 					one)));
329 			else {
330 				xm1 = x - one;
331 				D_IM(ans) = log1p(xm1 + sqrt(xm1 * (x + one)));
332 			}
333 		}
334 	} else if (y <= E * fabs(x - one)) {	/* region 2: y < tiny*|x-1| */
335 		if (ix < 0x3ff00000) {	/* x < 1 */
336 			D_RE(ans) = acos(x);
337 			D_IM(ans) = y / sqrt((one + x) * (one - x));
338 		} else if (ix >= 0x43500000) {	/* |x| >= 2**54 */
339 			D_RE(ans) = y / x;
340 			D_IM(ans) = ln2 + log(x);
341 		} else {
342 			t = sqrt((x - one) * (x + one));
343 			D_RE(ans) = y / t;
344 			if (ix >= 0x3ff80000)	/* x > Acrossover */
345 				D_IM(ans) = log(x + t);
346 			else
347 				D_IM(ans) = log1p((x - one) + t);
348 		}
349 	} else if (y < Foursqrtu) {	/* region 3 */
350 		t = sqrt(y);
351 		D_RE(ans) = t;
352 		D_IM(ans) = t;
353 	} else if (E * y - one >= x) {	/* region 4 */
354 		D_RE(ans) = pi_2;
355 		D_IM(ans) = ln2 + log(y);
356 	} else if (ix >= 0x5fc00000 || iy >= 0x5fc00000) {	/* x,y>2**509 */
357 		/* region 5: x+1 or y is very large (>= sqrt(max)/8) */
358 		t = x / y;
359 		D_RE(ans) = atan(y / x);
360 		D_IM(ans) = ln2 + log(y) + half * log1p(t * t);
361 	} else if (x < Foursqrtu) {
362 		/* region 6: x is very small, < 4sqrt(min) */
363 		D_RE(ans) = pi_2;
364 		A = sqrt(one + y * y);
365 		if (iy >= 0x3ff80000)	/* if y > Acrossover */
366 			D_IM(ans) = log(y + A);
367 		else
368 			D_IM(ans) = half * log1p((y + y) * (y + A));
369 	} else {	/* safe region */
370 		y2 = y * y;
371 		xp1 = x + one;
372 		xm1 = x - one;
373 		R = sqrt(xp1 * xp1 + y2);
374 		S = sqrt(xm1 * xm1 + y2);
375 		A = half * (R + S);
376 		B = x / A;
377 		if (B <= Bcrossover)
378 			D_RE(ans) = acos(B);
379 		else {		/* use atan and an accurate approx to a-x */
380 			Apx = A + x;
381 			if (x <= one)
382 				D_RE(ans) = atan(sqrt(half * Apx * (y2 / (R +
383 					xp1) + (S - xm1))) / x);
384 			else
385 				D_RE(ans) = atan((y * sqrt(half * (Apx / (R +
386 					xp1) + Apx / (S + xm1)))) / x);
387 		}
388 		if (A <= Acrossover) {
389 			/* use log1p and an accurate approx to A-1 */
390 			if (x < one)
391 				Am1 = half * (y2 / (R + xp1) + y2 / (S - xm1));
392 			else
393 				Am1 = half * (y2 / (R + xp1) + (S + xm1));
394 			D_IM(ans) = log1p(Am1 + sqrt(Am1 * (A + one)));
395 		} else {
396 			D_IM(ans) = log(A + sqrt(A * A - one));
397 		}
398 	}
399 	if (hx < 0)
400 		D_RE(ans) = pi - D_RE(ans);
401 	if (hy >= 0)
402 		D_IM(ans) = -D_IM(ans);
403 	return (ans);
404 }
405