xref: /freebsd/lib/msun/src/catrigl.c (revision cbd30a72ca196976c1c700400ecd424baa1b9c16)
1 /*-
2  * Copyright (c) 2012 Stephen Montgomery-Smith <stephen@FreeBSD.ORG>
3  * Copyright (c) 2017 Mahdi Mokhtari <mmokhi@FreeBSD.org>
4  * All rights reserved.
5  *
6  * Redistribution and use in source and binary forms, with or without
7  * modification, are permitted provided that the following conditions
8  * are met:
9  * 1. Redistributions of source code must retain the above copyright
10  *    notice, this list of conditions and the following disclaimer.
11  * 2. Redistributions in binary form must reproduce the above copyright
12  *    notice, this list of conditions and the following disclaimer in the
13  *    documentation and/or other materials provided with the distribution.
14  *
15  * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND
16  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
17  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
18  * ARE DISCLAIMED.  IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE
19  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
20  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
21  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
22  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
23  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
24  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
25  * SUCH DAMAGE.
26  */
27 
28 /*
29  * The algorithm is very close to that in "Implementing the complex arcsine
30  * and arccosine functions using exception handling" by T. E. Hull, Thomas F.
31  * Fairgrieve, and Ping Tak Peter Tang, published in ACM Transactions on
32  * Mathematical Software, Volume 23 Issue 3, 1997, Pages 299-335,
33  * http://dl.acm.org/citation.cfm?id=275324.
34  *
35  * See catrig.c for complete comments.
36  *
37  * XXX comments were removed automatically, and even short ones on the right
38  * of statements were removed (all of them), contrary to normal style.  Only
39  * a few comments on the right of declarations remain.
40  */
41 
42 #include <sys/cdefs.h>
43 __FBSDID("$FreeBSD$");
44 
45 #include <complex.h>
46 #include <float.h>
47 
48 #include "invtrig.h"
49 #include "math.h"
50 #include "math_private.h"
51 
52 #undef isinf
53 #define isinf(x)	(fabsl(x) == INFINITY)
54 #undef isnan
55 #define isnan(x)	((x) != (x))
56 #define	raise_inexact()	do { volatile float junk = 1 + tiny; } while(0)
57 #undef signbit
58 #define signbit(x)	(__builtin_signbitl(x))
59 
60 static const long double
61 A_crossover =		10,
62 B_crossover =		0.6417,
63 FOUR_SQRT_MIN =		0x1p-8189L,
64 QUARTER_SQRT_MAX =	0x1p8189L,
65 RECIP_EPSILON =		1 / LDBL_EPSILON,
66 SQRT_MIN =		0x1p-8191L;
67 
68 #if LDBL_MANT_DIG == 64
69 static const union IEEEl2bits
70 um_e =		LD80C(0xadf85458a2bb4a9b,  1, 2.71828182845904523536e+0L),
71 um_ln2 =	LD80C(0xb17217f7d1cf79ac, -1, 6.93147180559945309417e-1L);
72 #define		m_e	um_e.e
73 #define		m_ln2	um_ln2.e
74 static const long double
75 /* The next 2 literals for non-i386.  Misrounding them on i386 is harmless. */
76 SQRT_3_EPSILON = 5.70316273435758915310e-10,	/*  0x9cc470a0490973e8.0p-94 */
77 SQRT_6_EPSILON = 8.06549008734932771664e-10;	/*  0xddb3d742c265539e.0p-94 */
78 #elif LDBL_MANT_DIG == 113
79 static const long double
80 m_e =		2.71828182845904523536028747135266250e0L,	/* 0x15bf0a8b1457695355fb8ac404e7a.0p-111 */
81 m_ln2 =		6.93147180559945309417232121458176568e-1L,	/* 0x162e42fefa39ef35793c7673007e6.0p-113 */
82 SQRT_3_EPSILON = 2.40370335797945490975336727199878124e-17,	/*  0x1bb67ae8584caa73b25742d7078b8.0p-168 */
83 SQRT_6_EPSILON = 3.39934988877629587239082586223300391e-17;	/*  0x13988e1409212e7d0321914321a55.0p-167 */
84 #else
85 #error "Unsupported long double format"
86 #endif
87 
88 static const volatile float
89 tiny =			0x1p-100;
90 
91 static long double complex clog_for_large_values(long double complex z);
92 
93 static inline long double
94 f(long double a, long double b, long double hypot_a_b)
95 {
96 	if (b < 0)
97 		return ((hypot_a_b - b) / 2);
98 	if (b == 0)
99 		return (a / 2);
100 	return (a * a / (hypot_a_b + b) / 2);
101 }
102 
103 static inline void
104 do_hard_work(long double x, long double y, long double *rx, int *B_is_usable,
105     long double *B, long double *sqrt_A2my2, long double *new_y)
106 {
107 	long double R, S, A;
108 	long double Am1, Amy;
109 
110 	R = hypotl(x, y + 1);
111 	S = hypotl(x, y - 1);
112 
113 	A = (R + S) / 2;
114 	if (A < 1)
115 		A = 1;
116 
117 	if (A < A_crossover) {
118 		if (y == 1 && x < LDBL_EPSILON * LDBL_EPSILON / 128) {
119 			*rx = sqrtl(x);
120 		} else if (x >= LDBL_EPSILON * fabsl(y - 1)) {
121 			Am1 = f(x, 1 + y, R) + f(x, 1 - y, S);
122 			*rx = log1pl(Am1 + sqrtl(Am1 * (A + 1)));
123 		} else if (y < 1) {
124 			*rx = x / sqrtl((1 - y) * (1 + y));
125 		} else {
126 			*rx = log1pl((y - 1) + sqrtl((y - 1) * (y + 1)));
127 		}
128 	} else {
129 		*rx = logl(A + sqrtl(A * A - 1));
130 	}
131 
132 	*new_y = y;
133 
134 	if (y < FOUR_SQRT_MIN) {
135 		*B_is_usable = 0;
136 		*sqrt_A2my2 = A * (2 / LDBL_EPSILON);
137 		*new_y = y * (2 / LDBL_EPSILON);
138 		return;
139 	}
140 
141 	*B = y / A;
142 	*B_is_usable = 1;
143 
144 	if (*B > B_crossover) {
145 		*B_is_usable = 0;
146 		if (y == 1 && x < LDBL_EPSILON / 128) {
147 			*sqrt_A2my2 = sqrtl(x) * sqrtl((A + y) / 2);
148 		} else if (x >= LDBL_EPSILON * fabsl(y - 1)) {
149 			Amy = f(x, y + 1, R) + f(x, y - 1, S);
150 			*sqrt_A2my2 = sqrtl(Amy * (A + y));
151 		} else if (y > 1) {
152 			*sqrt_A2my2 = x * (4 / LDBL_EPSILON / LDBL_EPSILON) * y /
153 			    sqrtl((y + 1) * (y - 1));
154 			*new_y = y * (4 / LDBL_EPSILON / LDBL_EPSILON);
155 		} else {
156 			*sqrt_A2my2 = sqrtl((1 - y) * (1 + y));
157 		}
158 	}
159 }
160 
161 long double complex
162 casinhl(long double complex z)
163 {
164 	long double x, y, ax, ay, rx, ry, B, sqrt_A2my2, new_y;
165 	int B_is_usable;
166 	long double complex w;
167 
168 	x = creall(z);
169 	y = cimagl(z);
170 	ax = fabsl(x);
171 	ay = fabsl(y);
172 
173 	if (isnan(x) || isnan(y)) {
174 		if (isinf(x))
175 			return (CMPLXL(x, y + y));
176 		if (isinf(y))
177 			return (CMPLXL(y, x + x));
178 		if (y == 0)
179 			return (CMPLXL(x + x, y));
180 		return (CMPLXL(x + 0.0L + (y + 0), x + 0.0L + (y + 0)));
181 	}
182 
183 	if (ax > RECIP_EPSILON || ay > RECIP_EPSILON) {
184 		if (signbit(x) == 0)
185 			w = clog_for_large_values(z) + m_ln2;
186 		else
187 			w = clog_for_large_values(-z) + m_ln2;
188 		return (CMPLXL(copysignl(creall(w), x),
189 		    copysignl(cimagl(w), y)));
190 	}
191 
192 	if (x == 0 && y == 0)
193 		return (z);
194 
195 	raise_inexact();
196 
197 	if (ax < SQRT_6_EPSILON / 4 && ay < SQRT_6_EPSILON / 4)
198 		return (z);
199 
200 	do_hard_work(ax, ay, &rx, &B_is_usable, &B, &sqrt_A2my2, &new_y);
201 	if (B_is_usable)
202 		ry = asinl(B);
203 	else
204 		ry = atan2l(new_y, sqrt_A2my2);
205 	return (CMPLXL(copysignl(rx, x), copysignl(ry, y)));
206 }
207 
208 long double complex
209 casinl(long double complex z)
210 {
211 	long double complex w;
212 
213 	w = casinhl(CMPLXL(cimagl(z), creall(z)));
214 	return (CMPLXL(cimagl(w), creall(w)));
215 }
216 
217 long double complex
218 cacosl(long double complex z)
219 {
220 	long double x, y, ax, ay, rx, ry, B, sqrt_A2mx2, new_x;
221 	int sx, sy;
222 	int B_is_usable;
223 	long double complex w;
224 
225 	x = creall(z);
226 	y = cimagl(z);
227 	sx = signbit(x);
228 	sy = signbit(y);
229 	ax = fabsl(x);
230 	ay = fabsl(y);
231 
232 	if (isnan(x) || isnan(y)) {
233 		if (isinf(x))
234 			return (CMPLXL(y + y, -INFINITY));
235 		if (isinf(y))
236 			return (CMPLXL(x + x, -y));
237 		if (x == 0)
238 			return (CMPLXL(pio2_hi + pio2_lo, y + y));
239 		return (CMPLXL(x + 0.0L + (y + 0), x + 0.0L + (y + 0)));
240 	}
241 
242 	if (ax > RECIP_EPSILON || ay > RECIP_EPSILON) {
243 		w = clog_for_large_values(z);
244 		rx = fabsl(cimagl(w));
245 		ry = creall(w) + m_ln2;
246 		if (sy == 0)
247 			ry = -ry;
248 		return (CMPLXL(rx, ry));
249 	}
250 
251 	if (x == 1 && y == 0)
252 		return (CMPLXL(0, -y));
253 
254 	raise_inexact();
255 
256 	if (ax < SQRT_6_EPSILON / 4 && ay < SQRT_6_EPSILON / 4)
257 		return (CMPLXL(pio2_hi - (x - pio2_lo), -y));
258 
259 	do_hard_work(ay, ax, &ry, &B_is_usable, &B, &sqrt_A2mx2, &new_x);
260 	if (B_is_usable) {
261 		if (sx == 0)
262 			rx = acosl(B);
263 		else
264 			rx = acosl(-B);
265 	} else {
266 		if (sx == 0)
267 			rx = atan2l(sqrt_A2mx2, new_x);
268 		else
269 			rx = atan2l(sqrt_A2mx2, -new_x);
270 	}
271 	if (sy == 0)
272 		ry = -ry;
273 	return (CMPLXL(rx, ry));
274 }
275 
276 long double complex
277 cacoshl(long double complex z)
278 {
279 	long double complex w;
280 	long double rx, ry;
281 
282 	w = cacosl(z);
283 	rx = creall(w);
284 	ry = cimagl(w);
285 	if (isnan(rx) && isnan(ry))
286 		return (CMPLXL(ry, rx));
287 	if (isnan(rx))
288 		return (CMPLXL(fabsl(ry), rx));
289 	if (isnan(ry))
290 		return (CMPLXL(ry, ry));
291 	return (CMPLXL(fabsl(ry), copysignl(rx, cimagl(z))));
292 }
293 
294 static long double complex
295 clog_for_large_values(long double complex z)
296 {
297 	long double x, y;
298 	long double ax, ay, t;
299 
300 	x = creall(z);
301 	y = cimagl(z);
302 	ax = fabsl(x);
303 	ay = fabsl(y);
304 	if (ax < ay) {
305 		t = ax;
306 		ax = ay;
307 		ay = t;
308 	}
309 
310 	if (ax > LDBL_MAX / 2)
311 		return (CMPLXL(logl(hypotl(x / m_e, y / m_e)) + 1,
312 		    atan2l(y, x)));
313 
314 	if (ax > QUARTER_SQRT_MAX || ay < SQRT_MIN)
315 		return (CMPLXL(logl(hypotl(x, y)), atan2l(y, x)));
316 
317 	return (CMPLXL(logl(ax * ax + ay * ay) / 2, atan2l(y, x)));
318 }
319 
320 static inline long double
321 sum_squares(long double x, long double y)
322 {
323 
324 	if (y < SQRT_MIN)
325 		return (x * x);
326 
327 	return (x * x + y * y);
328 }
329 
330 static inline long double
331 real_part_reciprocal(long double x, long double y)
332 {
333 	long double scale;
334 	uint16_t hx, hy;
335 	int16_t ix, iy;
336 
337 	GET_LDBL_EXPSIGN(hx, x);
338 	ix = hx & 0x7fff;
339 	GET_LDBL_EXPSIGN(hy, y);
340 	iy = hy & 0x7fff;
341 #define	BIAS	(LDBL_MAX_EXP - 1)
342 #define	CUTOFF	(LDBL_MANT_DIG / 2 + 1)
343 	if (ix - iy >= CUTOFF || isinf(x))
344 		return (1 / x);
345 	if (iy - ix >= CUTOFF)
346 		return (x / y / y);
347 	if (ix <= BIAS + LDBL_MAX_EXP / 2 - CUTOFF)
348 		return (x / (x * x + y * y));
349 	scale = 1;
350 	SET_LDBL_EXPSIGN(scale, 0x7fff - ix);
351 	x *= scale;
352 	y *= scale;
353 	return (x / (x * x + y * y) * scale);
354 }
355 
356 long double complex
357 catanhl(long double complex z)
358 {
359 	long double x, y, ax, ay, rx, ry;
360 
361 	x = creall(z);
362 	y = cimagl(z);
363 	ax = fabsl(x);
364 	ay = fabsl(y);
365 
366 	if (y == 0 && ax <= 1)
367 		return (CMPLXL(atanhl(x), y));
368 
369 	if (x == 0)
370 		return (CMPLXL(x, atanl(y)));
371 
372 	if (isnan(x) || isnan(y)) {
373 		if (isinf(x))
374 			return (CMPLXL(copysignl(0, x), y + y));
375 		if (isinf(y))
376 			return (CMPLXL(copysignl(0, x),
377 			    copysignl(pio2_hi + pio2_lo, y)));
378 		return (CMPLXL(x + 0.0L + (y + 0), x + 0.0L + (y + 0)));
379 	}
380 
381 	if (ax > RECIP_EPSILON || ay > RECIP_EPSILON)
382 		return (CMPLXL(real_part_reciprocal(x, y),
383 		    copysignl(pio2_hi + pio2_lo, y)));
384 
385 	if (ax < SQRT_3_EPSILON / 2 && ay < SQRT_3_EPSILON / 2) {
386 		raise_inexact();
387 		return (z);
388 	}
389 
390 	if (ax == 1 && ay < LDBL_EPSILON)
391 		rx = (m_ln2 - logl(ay)) / 2;
392 	else
393 		rx = log1pl(4 * ax / sum_squares(ax - 1, ay)) / 4;
394 
395 	if (ax == 1)
396 		ry = atan2l(2, -ay) / 2;
397 	else if (ay < LDBL_EPSILON)
398 		ry = atan2l(2 * ay, (1 - ax) * (1 + ax)) / 2;
399 	else
400 		ry = atan2l(2 * ay, (1 - ax) * (1 + ax) - ay * ay) / 2;
401 
402 	return (CMPLXL(copysignl(rx, x), copysignl(ry, y)));
403 }
404 
405 long double complex
406 catanl(long double complex z)
407 {
408 	long double complex w;
409 
410 	w = catanhl(CMPLXL(cimagl(z), creall(z)));
411 	return (CMPLXL(cimagl(w), creall(w)));
412 }
413