xref: /netbsd-src/lib/libm/complex/catrig.c (revision 388550b026d49b7f7b7480b1113bf82bb8d6a480)
1 /*	$NetBSD: catrig.c,v 1.3 2022/04/19 20:32:16 rillig Exp $	*/
2 /*-
3  * Copyright (c) 2012 Stephen Montgomery-Smith <stephen@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 #include <sys/cdefs.h>
29 #if 0
30 __FBSDID("$FreeBSD: head/lib/msun/src/catrig.c 275819 2014-12-16 09:21:56Z ed $");
31 #endif
32 __RCSID("$NetBSD: catrig.c,v 1.3 2022/04/19 20:32:16 rillig Exp $");
33 
34 #include "namespace.h"
35 #ifdef __weak_alias
36 __weak_alias(casin, _casin)
37 #endif
38 #ifdef __weak_alias
39 __weak_alias(catan, _catan)
40 #endif
41 
42 #include <complex.h>
43 #include <float.h>
44 
45 #include "math.h"
46 #include "math_private.h"
47 
48 
49 
50 #undef isinf
51 #define isinf(x)	(fabs(x) == INFINITY)
52 #undef isnan
53 #define isnan(x)	((x) != (x))
54 #define	raise_inexact()	do { volatile float junk __unused = /*LINTED*/1 + tiny; } while (0)
55 #undef signbit
56 #define signbit(x)	(__builtin_signbit(x))
57 
58 /* We need that DBL_EPSILON^2/128 is larger than FOUR_SQRT_MIN. */
59 static const double
60 A_crossover =		10, /* Hull et al suggest 1.5, but 10 works better */
61 B_crossover =		0.6417,			/* suggested by Hull et al */
62 m_e =			2.7182818284590452e0,	/*  0x15bf0a8b145769.0p-51 */
63 m_ln2 =			6.9314718055994531e-1,	/*  0x162e42fefa39ef.0p-53 */
64 pio2_hi =		1.5707963267948966e0,	/*  0x1921fb54442d18.0p-52 */
65 RECIP_EPSILON =		1 / DBL_EPSILON,
66 SQRT_3_EPSILON =	2.5809568279517849e-8,	/*  0x1bb67ae8584caa.0p-78 */
67 SQRT_6_EPSILON =	3.6500241499888571e-8,	/*  0x13988e1409212e.0p-77 */
68 #if DBL_MAX_EXP == 1024	/* IEEE */
69 FOUR_SQRT_MIN =		0x1p-509,		/* >= 4 * sqrt(DBL_MIN) */
70 QUARTER_SQRT_MAX =	0x1p509,		/* <= sqrt(DBL_MAX) / 4 */
71 SQRT_MIN =		0x1p-511;		/* >= sqrt(DBL_MIN) */
72 #elif DBL_MAX_EXP == 127 /* VAX */
73 FOUR_SQRT_MIN =		0x1p-62,		/* >= 4 * sqrt(DBL_MIN) */
74 QUARTER_SQRT_MAX =	0x1p62,			/* <= sqrt(DBL_MAX) / 4 */
75 SQRT_MIN =		0x1p-64;		/* >= sqrt(DBL_MIN) */
76 #else
77 	#error "unsupported floating point format"
78 #endif
79 
80 
81 static const volatile double
82 pio2_lo =		6.1232339957367659e-17;	/*  0x11a62633145c07.0p-106 */
83 static const volatile float
84 tiny =			0x1p-100;
85 
86 static double complex clog_for_large_values(double complex z);
87 
88 /*
89  * Testing indicates that all these functions are accurate up to 4 ULP.
90  * The functions casin(h) and cacos(h) are about 2.5 times slower than asinh.
91  * The functions catan(h) are a little under 2 times slower than atanh.
92  *
93  * The code for casinh, casin, cacos, and cacosh comes first.  The code is
94  * rather complicated, and the four functions are highly interdependent.
95  *
96  * The code for catanh and catan comes at the end.  It is much simpler than
97  * the other functions, and the code for these can be disconnected from the
98  * rest of the code.
99  */
100 
101 /*
102  *			================================
103  *			| casinh, casin, cacos, cacosh |
104  *			================================
105  */
106 
107 /*
108  * The algorithm is very close to that in "Implementing the complex arcsine
109  * and arccosine functions using exception handling" by T. E. Hull, Thomas F.
110  * Fairgrieve, and Ping Tak Peter Tang, published in ACM Transactions on
111  * Mathematical Software, Volume 23 Issue 3, 1997, Pages 299-335,
112  * http://dl.acm.org/citation.cfm?id=275324.
113  *
114  * Throughout we use the convention z = x + I*y.
115  *
116  * casinh(z) = sign(x)*log(A+sqrt(A*A-1)) + I*asin(B)
117  * where
118  * A = (|z+I| + |z-I|) / 2
119  * B = (|z+I| - |z-I|) / 2 = y/A
120  *
121  * These formulas become numerically unstable:
122  *   (a) for Re(casinh(z)) when z is close to the line segment [-I, I] (that
123  *       is, Re(casinh(z)) is close to 0);
124  *   (b) for Im(casinh(z)) when z is close to either of the intervals
125  *       [I, I*infinity) or (-I*infinity, -I] (that is, |Im(casinh(z))| is
126  *       close to PI/2).
127  *
128  * These numerical problems are overcome by defining
129  * f(a, b) = (hypot(a, b) - b) / 2 = a*a / (hypot(a, b) + b) / 2
130  * Then if A < A_crossover, we use
131  *   log(A + sqrt(A*A-1)) = log1p((A-1) + sqrt((A-1)*(A+1)))
132  *   A-1 = f(x, 1+y) + f(x, 1-y)
133  * and if B > B_crossover, we use
134  *   asin(B) = atan2(y, sqrt(A*A - y*y)) = atan2(y, sqrt((A+y)*(A-y)))
135  *   A-y = f(x, y+1) + f(x, y-1)
136  * where without loss of generality we have assumed that x and y are
137  * non-negative.
138  *
139  * Much of the difficulty comes because the intermediate computations may
140  * produce overflows or underflows.  This is dealt with in the paper by Hull
141  * et al by using exception handling.  We do this by detecting when
142  * computations risk underflow or overflow.  The hardest part is handling the
143  * underflows when computing f(a, b).
144  *
145  * Note that the function f(a, b) does not appear explicitly in the paper by
146  * Hull et al, but the idea may be found on pages 308 and 309.  Introducing the
147  * function f(a, b) allows us to concentrate many of the clever tricks in this
148  * paper into one function.
149  */
150 
151 /*
152  * Function f(a, b, hypot_a_b) = (hypot(a, b) - b) / 2.
153  * Pass hypot(a, b) as the third argument.
154  */
155 static inline double
f(double a,double b,double hypot_a_b)156 f(double a, double b, double hypot_a_b)
157 {
158 	if (b < 0)
159 		return ((hypot_a_b - b) / 2);
160 	if (b == 0)
161 		return (a / 2);
162 	return (a * a / (hypot_a_b + b) / 2);
163 }
164 
165 /*
166  * All the hard work is contained in this function.
167  * x and y are assumed positive or zero, and less than RECIP_EPSILON.
168  * Upon return:
169  * rx = Re(casinh(z)) = -Im(cacos(y + I*x)).
170  * B_is_usable is set to 1 if the value of B is usable.
171  * If B_is_usable is set to 0, sqrt_A2my2 = sqrt(A*A - y*y), and new_y = y.
172  * If returning sqrt_A2my2 has potential to result in an underflow, it is
173  * rescaled, and new_y is similarly rescaled.
174  */
175 static inline void
do_hard_work(double x,double y,double * rx,int * B_is_usable,double * B,double * sqrt_A2my2,double * new_y)176 do_hard_work(double x, double y, double *rx, int *B_is_usable, double *B,
177     double *sqrt_A2my2, double *new_y)
178 {
179 	double R, S, A; /* A, B, R, and S are as in Hull et al. */
180 	double Am1, Amy; /* A-1, A-y. */
181 
182 	R = hypot(x, y + 1);		/* |z+I| */
183 	S = hypot(x, y - 1);		/* |z-I| */
184 
185 	/* A = (|z+I| + |z-I|) / 2 */
186 	A = (R + S) / 2;
187 	/*
188 	 * Mathematically A >= 1.  There is a small chance that this will not
189 	 * be so because of rounding errors.  So we will make certain it is
190 	 * so.
191 	 */
192 	if (A < 1)
193 		A = 1;
194 
195 	if (A < A_crossover) {
196 		/*
197 		 * Am1 = fp + fm, where fp = f(x, 1+y), and fm = f(x, 1-y).
198 		 * rx = log1p(Am1 + sqrt(Am1*(A+1)))
199 		 */
200 		if (y == 1 && x < DBL_EPSILON * DBL_EPSILON / 128) {
201 			/*
202 			 * fp is of order x^2, and fm = x/2.
203 			 * A = 1 (inexactly).
204 			 */
205 			*rx = sqrt(x);
206 		} else if (x >= DBL_EPSILON * fabs(y - 1)) {
207 			/*
208 			 * Underflow will not occur because
209 			 * x >= DBL_EPSILON^2/128 >= FOUR_SQRT_MIN
210 			 */
211 			Am1 = f(x, 1 + y, R) + f(x, 1 - y, S);
212 			*rx = log1p(Am1 + sqrt(Am1 * (A + 1)));
213 		} else if (y < 1) {
214 			/*
215 			 * fp = x*x/(1+y)/4, fm = x*x/(1-y)/4, and
216 			 * A = 1 (inexactly).
217 			 */
218 			*rx = x / sqrt((1 - y) * (1 + y));
219 		} else {		/* if (y > 1) */
220 			/*
221 			 * A-1 = y-1 (inexactly).
222 			 */
223 			*rx = log1p((y - 1) + sqrt((y - 1) * (y + 1)));
224 		}
225 	} else {
226 		*rx = log(A + sqrt(A * A - 1));
227 	}
228 
229 	*new_y = y;
230 
231 	if (y < FOUR_SQRT_MIN) {
232 		/*
233 		 * Avoid a possible underflow caused by y/A.  For casinh this
234 		 * would be legitimate, but will be picked up by invoking atan2
235 		 * later on.  For cacos this would not be legitimate.
236 		 */
237 		*B_is_usable = 0;
238 		*sqrt_A2my2 = A * (2 / DBL_EPSILON);
239 		*new_y = y * (2 / DBL_EPSILON);
240 		return;
241 	}
242 
243 	/* B = (|z+I| - |z-I|) / 2 = y/A */
244 	*B = y / A;
245 	*B_is_usable = 1;
246 
247 	if (*B > B_crossover) {
248 		*B_is_usable = 0;
249 		/*
250 		 * Amy = fp + fm, where fp = f(x, y+1), and fm = f(x, y-1).
251 		 * sqrt_A2my2 = sqrt(Amy*(A+y))
252 		 */
253 		if (y == 1 && x < DBL_EPSILON / 128) {
254 			/*
255 			 * fp is of order x^2, and fm = x/2.
256 			 * A = 1 (inexactly).
257 			 */
258 			*sqrt_A2my2 = sqrt(x) * sqrt((A + y) / 2);
259 		} else if (x >= DBL_EPSILON * fabs(y - 1)) {
260 			/*
261 			 * Underflow will not occur because
262 			 * x >= DBL_EPSILON/128 >= FOUR_SQRT_MIN
263 			 * and
264 			 * x >= DBL_EPSILON^2 >= FOUR_SQRT_MIN
265 			 */
266 			Amy = f(x, y + 1, R) + f(x, y - 1, S);
267 			*sqrt_A2my2 = sqrt(Amy * (A + y));
268 		} else if (y > 1) {
269 			/*
270 			 * fp = x*x/(y+1)/4, fm = x*x/(y-1)/4, and
271 			 * A = y (inexactly).
272 			 *
273 			 * y < RECIP_EPSILON.  So the following
274 			 * scaling should avoid any underflow problems.
275 			 */
276 			*sqrt_A2my2 = x * (4 / DBL_EPSILON / DBL_EPSILON) * y /
277 			    sqrt((y + 1) * (y - 1));
278 			*new_y = y * (4 / DBL_EPSILON / DBL_EPSILON);
279 		} else {		/* if (y < 1) */
280 			/*
281 			 * fm = 1-y >= DBL_EPSILON, fp is of order x^2, and
282 			 * A = 1 (inexactly).
283 			 */
284 			*sqrt_A2my2 = sqrt((1 - y) * (1 + y));
285 		}
286 	}
287 }
288 
289 /*
290  * casinh(z) = z + O(z^3)   as z -> 0
291  *
292  * casinh(z) = sign(x)*clog(sign(x)*z) + O(1/z^2)   as z -> infinity
293  * The above formula works for the imaginary part as well, because
294  * Im(casinh(z)) = sign(x)*atan2(sign(x)*y, fabs(x)) + O(y/z^3)
295  *    as z -> infinity, uniformly in y
296  */
297 double complex
casinh(double complex z)298 casinh(double complex z)
299 {
300 	double x, y, ax, ay, rx, ry, B, sqrt_A2my2, new_y;
301 	int B_is_usable;
302 	double complex w;
303 
304 	x = creal(z);
305 	y = cimag(z);
306 	ax = fabs(x);
307 	ay = fabs(y);
308 
309 	if (isnan(x) || isnan(y)) {
310 		/* casinh(+-Inf + I*NaN) = +-Inf + I*NaN */
311 		if (isinf(x))
312 			return (CMPLX(x, y + y));
313 		/* casinh(NaN + I*+-Inf) = opt(+-)Inf + I*NaN */
314 		if (isinf(y))
315 			return (CMPLX(y, x + x));
316 		/* casinh(NaN + I*0) = NaN + I*0 */
317 		if (y == 0)
318 			return (CMPLX(x + x, y));
319 		/*
320 		 * All other cases involving NaN return NaN + I*NaN.
321 		 * C99 leaves it optional whether to raise invalid if one of
322 		 * the arguments is not NaN, so we opt not to raise it.
323 		 */
324 		return (CMPLX(x + 0.0L + (y + 0), x + 0.0L + (y + 0)));
325 	}
326 
327 	if (ax > RECIP_EPSILON || ay > RECIP_EPSILON) {
328 		/* clog...() will raise inexact unless x or y is infinite. */
329 		if (signbit(x) == 0)
330 			w = clog_for_large_values(z) + m_ln2;
331 		else
332 			w = clog_for_large_values(-z) + m_ln2;
333 		return (CMPLX(copysign(creal(w), x), copysign(cimag(w), y)));
334 	}
335 
336 	/* Avoid spuriously raising inexact for z = 0. */
337 	if (x == 0 && y == 0)
338 		return (z);
339 
340 	/* All remaining cases are inexact. */
341 	raise_inexact();
342 
343 	if (ax < SQRT_6_EPSILON / 4 && ay < SQRT_6_EPSILON / 4)
344 		return (z);
345 
346 	do_hard_work(ax, ay, &rx, &B_is_usable, &B, &sqrt_A2my2, &new_y);
347 	if (B_is_usable)
348 		ry = asin(B);
349 	else
350 		ry = atan2(new_y, sqrt_A2my2);
351 	return (CMPLX(copysign(rx, x), copysign(ry, y)));
352 }
353 
354 /*
355  * casin(z) = reverse(casinh(reverse(z)))
356  * where reverse(x + I*y) = y + I*x = I*conj(z).
357  */
358 double complex
casin(double complex z)359 casin(double complex z)
360 {
361 	double complex w = casinh(CMPLX(cimag(z), creal(z)));
362 
363 	return (CMPLX(cimag(w), creal(w)));
364 }
365 
366 /*
367  * cacos(z) = PI/2 - casin(z)
368  * but do the computation carefully so cacos(z) is accurate when z is
369  * close to 1.
370  *
371  * cacos(z) = PI/2 - z + O(z^3)   as z -> 0
372  *
373  * cacos(z) = -sign(y)*I*clog(z) + O(1/z^2)   as z -> infinity
374  * The above formula works for the real part as well, because
375  * Re(cacos(z)) = atan2(fabs(y), x) + O(y/z^3)
376  *    as z -> infinity, uniformly in y
377  */
378 double complex
cacos(double complex z)379 cacos(double complex z)
380 {
381 	double x, y, ax, ay, rx, ry, B, sqrt_A2mx2, new_x;
382 	int sx, sy;
383 	int B_is_usable;
384 	double complex w;
385 
386 	x = creal(z);
387 	y = cimag(z);
388 	sx = signbit(x);
389 	sy = signbit(y);
390 	ax = fabs(x);
391 	ay = fabs(y);
392 
393 	if (isnan(x) || isnan(y)) {
394 		/* cacos(+-Inf + I*NaN) = NaN + I*opt(-)Inf */
395 		if (isinf(x))
396 			return (CMPLX(y + y, -INFINITY));
397 		/* cacos(NaN + I*+-Inf) = NaN + I*-+Inf */
398 		if (isinf(y))
399 			return (CMPLX(x + x, -y));
400 		/* cacos(0 + I*NaN) = PI/2 + I*NaN with inexact */
401 		if (x == 0)
402 			return (CMPLX(pio2_hi + pio2_lo, y + y));
403 		/*
404 		 * All other cases involving NaN return NaN + I*NaN.
405 		 * C99 leaves it optional whether to raise invalid if one of
406 		 * the arguments is not NaN, so we opt not to raise it.
407 		 */
408 		return (CMPLX(x + 0.0L + (y + 0), x + 0.0L + (y + 0)));
409 	}
410 
411 	if (ax > RECIP_EPSILON || ay > RECIP_EPSILON) {
412 		/* clog...() will raise inexact unless x or y is infinite. */
413 		w = clog_for_large_values(z);
414 		rx = fabs(cimag(w));
415 		ry = creal(w) + m_ln2;
416 		if (sy == 0)
417 			ry = -ry;
418 		return (CMPLX(rx, ry));
419 	}
420 
421 	/* Avoid spuriously raising inexact for z = 1. */
422 	if (x == 1 && y == 0)
423 		return (CMPLX(0, -y));
424 
425 	/* All remaining cases are inexact. */
426 	raise_inexact();
427 
428 	if (ax < SQRT_6_EPSILON / 4 && ay < SQRT_6_EPSILON / 4)
429 		return (CMPLX(pio2_hi - (x - pio2_lo), -y));
430 
431 	do_hard_work(ay, ax, &ry, &B_is_usable, &B, &sqrt_A2mx2, &new_x);
432 	if (B_is_usable) {
433 		if (sx == 0)
434 			rx = acos(B);
435 		else
436 			rx = acos(-B);
437 	} else {
438 		if (sx == 0)
439 			rx = atan2(sqrt_A2mx2, new_x);
440 		else
441 			rx = atan2(sqrt_A2mx2, -new_x);
442 	}
443 	if (sy == 0)
444 		ry = -ry;
445 	return (CMPLX(rx, ry));
446 }
447 
448 /*
449  * cacosh(z) = I*cacos(z) or -I*cacos(z)
450  * where the sign is chosen so Re(cacosh(z)) >= 0.
451  */
452 double complex
cacosh(double complex z)453 cacosh(double complex z)
454 {
455 	double complex w;
456 	double rx, ry;
457 
458 	w = cacos(z);
459 	rx = creal(w);
460 	ry = cimag(w);
461 	/* cacosh(NaN + I*NaN) = NaN + I*NaN */
462 	if (isnan(rx) && isnan(ry))
463 		return (CMPLX(ry, rx));
464 	/* cacosh(NaN + I*+-Inf) = +Inf + I*NaN */
465 	/* cacosh(+-Inf + I*NaN) = +Inf + I*NaN */
466 	if (isnan(rx))
467 		return (CMPLX(fabs(ry), rx));
468 	/* cacosh(0 + I*NaN) = NaN + I*NaN */
469 	if (isnan(ry))
470 		return (CMPLX(ry, ry));
471 	return (CMPLX(fabs(ry), copysign(rx, cimag(z))));
472 }
473 
474 /*
475  * Optimized version of clog() for |z| finite and larger than ~RECIP_EPSILON.
476  */
477 static double complex
clog_for_large_values(double complex z)478 clog_for_large_values(double complex z)
479 {
480 	double x, y;
481 	double ax, ay, t;
482 
483 	x = creal(z);
484 	y = cimag(z);
485 	ax = fabs(x);
486 	ay = fabs(y);
487 	if (ax < ay) {
488 		t = ax;
489 		ax = ay;
490 		ay = t;
491 	}
492 
493 	/*
494 	 * Avoid overflow in hypot() when x and y are both very large.
495 	 * Divide x and y by E, and then add 1 to the logarithm.  This depends
496 	 * on E being larger than sqrt(2).
497 	 * Dividing by E causes an insignificant loss of accuracy; however
498 	 * this method is still poor since it is uneccessarily slow.
499 	 */
500 	if (ax > DBL_MAX / 2)
501 		return (CMPLX(log(hypot(x / m_e, y / m_e)) + 1, atan2(y, x)));
502 
503 	/*
504 	 * Avoid overflow when x or y is large.  Avoid underflow when x or
505 	 * y is small.
506 	 */
507 	if (ax > QUARTER_SQRT_MAX || ay < SQRT_MIN)
508 		return (CMPLX(log(hypot(x, y)), atan2(y, x)));
509 
510 	return (CMPLX(log(ax * ax + ay * ay) / 2, atan2(y, x)));
511 }
512 
513 /*
514  *				=================
515  *				| catanh, catan |
516  *				=================
517  */
518 
519 /*
520  * sum_squares(x,y) = x*x + y*y (or just x*x if y*y would underflow).
521  * Assumes x*x and y*y will not overflow.
522  * Assumes x and y are finite.
523  * Assumes y is non-negative.
524  * Assumes fabs(x) >= DBL_EPSILON.
525  */
526 static inline double
sum_squares(double x,double y)527 sum_squares(double x, double y)
528 {
529 
530 	/* Avoid underflow when y is small. */
531 	if (y < SQRT_MIN)
532 		return (x * x);
533 
534 	return (x * x + y * y);
535 }
536 
537 /*
538  * real_part_reciprocal(x, y) = Re(1/(x+I*y)) = x/(x*x + y*y).
539  * Assumes x and y are not NaN, and one of x and y is larger than
540  * RECIP_EPSILON.  We avoid unwarranted underflow.  It is important to not use
541  * the code creal(1/z), because the imaginary part may produce an unwanted
542  * underflow.
543  * This is only called in a context where inexact is always raised before
544  * the call, so no effort is made to avoid or force inexact.
545  */
546 static inline double
real_part_reciprocal(double x,double y)547 real_part_reciprocal(double x, double y)
548 {
549 	double scale;
550 	uint32_t hx, hy;
551 	int32_t ix, iy;
552 
553 	/*
554 	 * This code is inspired by the C99 document n1124.pdf, Section G.5.1,
555 	 * example 2.
556 	 */
557 	GET_HIGH_WORD(hx, x);
558 	ix = hx & 0x7ff00000;
559 	GET_HIGH_WORD(hy, y);
560 	iy = hy & 0x7ff00000;
561 #define	BIAS	(DBL_MAX_EXP - 1)
562 /* XXX more guard digits are useful iff there is extra precision. */
563 #define	CUTOFF	(DBL_MANT_DIG / 2 + 1)	/* just half or 1 guard digit */
564 	if (ix - iy >= CUTOFF << 20 || isinf(x))
565 		return (1 / x);		/* +-Inf -> +-0 is special */
566 	if (iy - ix >= CUTOFF << 20)
567 		return (x / y / y);	/* should avoid double div, but hard */
568 	if (ix <= (BIAS + DBL_MAX_EXP / 2 - CUTOFF) << 20)
569 		return (x / (x * x + y * y));
570 	scale = 1;
571 	SET_HIGH_WORD(scale, 0x7ff00000 - ix);	/* 2**(1-ilogb(x)) */
572 	x *= scale;
573 	y *= scale;
574 	return (x / (x * x + y * y) * scale);
575 }
576 
577 /*
578  * catanh(z) = log((1+z)/(1-z)) / 2
579  *           = log1p(4*x / |z-1|^2) / 4
580  *             + I * atan2(2*y, (1-x)*(1+x)-y*y) / 2
581  *
582  * catanh(z) = z + O(z^3)   as z -> 0
583  *
584  * catanh(z) = 1/z + sign(y)*I*PI/2 + O(1/z^3)   as z -> infinity
585  * The above formula works for the real part as well, because
586  * Re(catanh(z)) = x/|z|^2 + O(x/z^4)
587  *    as z -> infinity, uniformly in x
588  */
589 double complex
catanh(double complex z)590 catanh(double complex z)
591 {
592 	double x, y, ax, ay, rx, ry;
593 
594 	x = creal(z);
595 	y = cimag(z);
596 	ax = fabs(x);
597 	ay = fabs(y);
598 
599 	/* This helps handle many cases. */
600 	if (y == 0 && ax <= 1)
601 		return (CMPLX(atanh(x), y));
602 
603 	/* To ensure the same accuracy as atan(), and to filter out z = 0. */
604 	if (x == 0)
605 		return (CMPLX(x, atan(y)));
606 
607 	if (isnan(x) || isnan(y)) {
608 		/* catanh(+-Inf + I*NaN) = +-0 + I*NaN */
609 		if (isinf(x))
610 			return (CMPLX(copysign(0, x), y + y));
611 		/* catanh(NaN + I*+-Inf) = sign(NaN)0 + I*+-PI/2 */
612 		if (isinf(y))
613 			return (CMPLX(copysign(0, x),
614 			    copysign(pio2_hi + pio2_lo, y)));
615 		/*
616 		 * All other cases involving NaN return NaN + I*NaN.
617 		 * C99 leaves it optional whether to raise invalid if one of
618 		 * the arguments is not NaN, so we opt not to raise it.
619 		 */
620 		return (CMPLX(x + 0.0L + (y + 0), x + 0.0L + (y + 0)));
621 	}
622 
623 	if (ax > RECIP_EPSILON || ay > RECIP_EPSILON)
624 		return (CMPLX(real_part_reciprocal(x, y),
625 		    copysign(pio2_hi + pio2_lo, y)));
626 
627 	if (ax < SQRT_3_EPSILON / 2 && ay < SQRT_3_EPSILON / 2) {
628 		/*
629 		 * z = 0 was filtered out above.  All other cases must raise
630 		 * inexact, but this is the only only that needs to do it
631 		 * explicitly.
632 		 */
633 		raise_inexact();
634 		return (z);
635 	}
636 
637 	if (ax == 1 && ay < DBL_EPSILON)
638 		rx = (m_ln2 - log(ay)) / 2;
639 	else
640 		rx = log1p(4 * ax / sum_squares(ax - 1, ay)) / 4;
641 
642 	if (ax == 1)
643 		ry = atan2(2, -ay) / 2;
644 	else if (ay < DBL_EPSILON)
645 		ry = atan2(2 * ay, (1 - ax) * (1 + ax)) / 2;
646 	else
647 		ry = atan2(2 * ay, (1 - ax) * (1 + ax) - ay * ay) / 2;
648 
649 	return (CMPLX(copysign(rx, x), copysign(ry, y)));
650 }
651 
652 /*
653  * catan(z) = reverse(catanh(reverse(z)))
654  * where reverse(x + I*y) = y + I*x = I*conj(z).
655  */
656 double complex
catan(double complex z)657 catan(double complex z)
658 {
659 	double complex w = catanh(CMPLX(cimag(z), creal(z)));
660 
661 	return (CMPLX(cimag(w), creal(w)));
662 }
663