Elliott Hughes | b8ee16f | 2014-11-06 11:16:55 -0800 | [diff] [blame] | 1 | /*- |
Elliott Hughes | c1e46b6 | 2023-07-19 14:06:31 -0700 | [diff] [blame] | 2 | * SPDX-License-Identifier: BSD-2-Clause |
Elliott Hughes | 8da8ca4 | 2018-05-08 13:35:33 -0700 | [diff] [blame] | 3 | * |
Elliott Hughes | b8ee16f | 2014-11-06 11:16:55 -0800 | [diff] [blame] | 4 | * Copyright (c) 2012 Stephen Montgomery-Smith <stephen@FreeBSD.ORG> |
| 5 | * All rights reserved. |
| 6 | * |
| 7 | * Redistribution and use in source and binary forms, with or without |
| 8 | * modification, are permitted provided that the following conditions |
| 9 | * are met: |
| 10 | * 1. Redistributions of source code must retain the above copyright |
| 11 | * notice, this list of conditions and the following disclaimer. |
| 12 | * 2. Redistributions in binary form must reproduce the above copyright |
| 13 | * notice, this list of conditions and the following disclaimer in the |
| 14 | * documentation and/or other materials provided with the distribution. |
| 15 | * |
| 16 | * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND |
| 17 | * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE |
| 18 | * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE |
| 19 | * ARE DISCLAIMED. IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE |
| 20 | * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL |
| 21 | * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS |
| 22 | * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) |
| 23 | * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT |
| 24 | * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY |
| 25 | * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF |
| 26 | * SUCH DAMAGE. |
| 27 | */ |
| 28 | |
| 29 | #include <sys/cdefs.h> |
Elliott Hughes | bac0ebb | 2021-01-26 14:17:20 -0800 | [diff] [blame] | 30 | __FBSDID("$FreeBSD$"); |
Elliott Hughes | b8ee16f | 2014-11-06 11:16:55 -0800 | [diff] [blame] | 31 | |
| 32 | #include <complex.h> |
| 33 | #include <float.h> |
| 34 | |
| 35 | #include "math.h" |
| 36 | #include "math_private.h" |
| 37 | |
| 38 | #undef isinf |
| 39 | #define isinf(x) (fabs(x) == INFINITY) |
| 40 | #undef isnan |
| 41 | #define isnan(x) ((x) != (x)) |
Elliott Hughes | 8da8ca4 | 2018-05-08 13:35:33 -0700 | [diff] [blame] | 42 | #define raise_inexact() do { volatile float junk __unused = 1 + tiny; } while(0) |
Elliott Hughes | b8ee16f | 2014-11-06 11:16:55 -0800 | [diff] [blame] | 43 | #undef signbit |
| 44 | #define signbit(x) (__builtin_signbit(x)) |
| 45 | |
| 46 | /* We need that DBL_EPSILON^2/128 is larger than FOUR_SQRT_MIN. */ |
| 47 | static const double |
| 48 | A_crossover = 10, /* Hull et al suggest 1.5, but 10 works better */ |
| 49 | B_crossover = 0.6417, /* suggested by Hull et al */ |
| 50 | FOUR_SQRT_MIN = 0x1p-509, /* >= 4 * sqrt(DBL_MIN) */ |
| 51 | QUARTER_SQRT_MAX = 0x1p509, /* <= sqrt(DBL_MAX) / 4 */ |
| 52 | m_e = 2.7182818284590452e0, /* 0x15bf0a8b145769.0p-51 */ |
| 53 | m_ln2 = 6.9314718055994531e-1, /* 0x162e42fefa39ef.0p-53 */ |
| 54 | pio2_hi = 1.5707963267948966e0, /* 0x1921fb54442d18.0p-52 */ |
| 55 | RECIP_EPSILON = 1 / DBL_EPSILON, |
| 56 | SQRT_3_EPSILON = 2.5809568279517849e-8, /* 0x1bb67ae8584caa.0p-78 */ |
| 57 | SQRT_6_EPSILON = 3.6500241499888571e-8, /* 0x13988e1409212e.0p-77 */ |
| 58 | SQRT_MIN = 0x1p-511; /* >= sqrt(DBL_MIN) */ |
| 59 | |
| 60 | static const volatile double |
| 61 | pio2_lo = 6.1232339957367659e-17; /* 0x11a62633145c07.0p-106 */ |
| 62 | static const volatile float |
| 63 | tiny = 0x1p-100; |
| 64 | |
| 65 | static double complex clog_for_large_values(double complex z); |
| 66 | |
| 67 | /* |
| 68 | * Testing indicates that all these functions are accurate up to 4 ULP. |
| 69 | * The functions casin(h) and cacos(h) are about 2.5 times slower than asinh. |
| 70 | * The functions catan(h) are a little under 2 times slower than atanh. |
| 71 | * |
| 72 | * The code for casinh, casin, cacos, and cacosh comes first. The code is |
| 73 | * rather complicated, and the four functions are highly interdependent. |
| 74 | * |
| 75 | * The code for catanh and catan comes at the end. It is much simpler than |
| 76 | * the other functions, and the code for these can be disconnected from the |
| 77 | * rest of the code. |
| 78 | */ |
| 79 | |
| 80 | /* |
| 81 | * ================================ |
| 82 | * | casinh, casin, cacos, cacosh | |
| 83 | * ================================ |
| 84 | */ |
| 85 | |
| 86 | /* |
| 87 | * The algorithm is very close to that in "Implementing the complex arcsine |
| 88 | * and arccosine functions using exception handling" by T. E. Hull, Thomas F. |
| 89 | * Fairgrieve, and Ping Tak Peter Tang, published in ACM Transactions on |
| 90 | * Mathematical Software, Volume 23 Issue 3, 1997, Pages 299-335, |
| 91 | * http://dl.acm.org/citation.cfm?id=275324. |
| 92 | * |
| 93 | * Throughout we use the convention z = x + I*y. |
| 94 | * |
| 95 | * casinh(z) = sign(x)*log(A+sqrt(A*A-1)) + I*asin(B) |
| 96 | * where |
| 97 | * A = (|z+I| + |z-I|) / 2 |
| 98 | * B = (|z+I| - |z-I|) / 2 = y/A |
| 99 | * |
| 100 | * These formulas become numerically unstable: |
| 101 | * (a) for Re(casinh(z)) when z is close to the line segment [-I, I] (that |
| 102 | * is, Re(casinh(z)) is close to 0); |
| 103 | * (b) for Im(casinh(z)) when z is close to either of the intervals |
| 104 | * [I, I*infinity) or (-I*infinity, -I] (that is, |Im(casinh(z))| is |
| 105 | * close to PI/2). |
| 106 | * |
| 107 | * These numerical problems are overcome by defining |
| 108 | * f(a, b) = (hypot(a, b) - b) / 2 = a*a / (hypot(a, b) + b) / 2 |
| 109 | * Then if A < A_crossover, we use |
| 110 | * log(A + sqrt(A*A-1)) = log1p((A-1) + sqrt((A-1)*(A+1))) |
| 111 | * A-1 = f(x, 1+y) + f(x, 1-y) |
| 112 | * and if B > B_crossover, we use |
| 113 | * asin(B) = atan2(y, sqrt(A*A - y*y)) = atan2(y, sqrt((A+y)*(A-y))) |
| 114 | * A-y = f(x, y+1) + f(x, y-1) |
| 115 | * where without loss of generality we have assumed that x and y are |
| 116 | * non-negative. |
| 117 | * |
| 118 | * Much of the difficulty comes because the intermediate computations may |
| 119 | * produce overflows or underflows. This is dealt with in the paper by Hull |
| 120 | * et al by using exception handling. We do this by detecting when |
| 121 | * computations risk underflow or overflow. The hardest part is handling the |
| 122 | * underflows when computing f(a, b). |
| 123 | * |
| 124 | * Note that the function f(a, b) does not appear explicitly in the paper by |
| 125 | * Hull et al, but the idea may be found on pages 308 and 309. Introducing the |
| 126 | * function f(a, b) allows us to concentrate many of the clever tricks in this |
| 127 | * paper into one function. |
| 128 | */ |
| 129 | |
| 130 | /* |
| 131 | * Function f(a, b, hypot_a_b) = (hypot(a, b) - b) / 2. |
| 132 | * Pass hypot(a, b) as the third argument. |
| 133 | */ |
| 134 | static inline double |
| 135 | f(double a, double b, double hypot_a_b) |
| 136 | { |
| 137 | if (b < 0) |
| 138 | return ((hypot_a_b - b) / 2); |
| 139 | if (b == 0) |
| 140 | return (a / 2); |
| 141 | return (a * a / (hypot_a_b + b) / 2); |
| 142 | } |
| 143 | |
| 144 | /* |
| 145 | * All the hard work is contained in this function. |
| 146 | * x and y are assumed positive or zero, and less than RECIP_EPSILON. |
| 147 | * Upon return: |
| 148 | * rx = Re(casinh(z)) = -Im(cacos(y + I*x)). |
| 149 | * B_is_usable is set to 1 if the value of B is usable. |
| 150 | * If B_is_usable is set to 0, sqrt_A2my2 = sqrt(A*A - y*y), and new_y = y. |
| 151 | * If returning sqrt_A2my2 has potential to result in an underflow, it is |
| 152 | * rescaled, and new_y is similarly rescaled. |
| 153 | */ |
| 154 | static inline void |
| 155 | do_hard_work(double x, double y, double *rx, int *B_is_usable, double *B, |
| 156 | double *sqrt_A2my2, double *new_y) |
| 157 | { |
| 158 | double R, S, A; /* A, B, R, and S are as in Hull et al. */ |
| 159 | double Am1, Amy; /* A-1, A-y. */ |
| 160 | |
| 161 | R = hypot(x, y + 1); /* |z+I| */ |
| 162 | S = hypot(x, y - 1); /* |z-I| */ |
| 163 | |
| 164 | /* A = (|z+I| + |z-I|) / 2 */ |
| 165 | A = (R + S) / 2; |
| 166 | /* |
| 167 | * Mathematically A >= 1. There is a small chance that this will not |
| 168 | * be so because of rounding errors. So we will make certain it is |
| 169 | * so. |
| 170 | */ |
| 171 | if (A < 1) |
| 172 | A = 1; |
| 173 | |
| 174 | if (A < A_crossover) { |
| 175 | /* |
| 176 | * Am1 = fp + fm, where fp = f(x, 1+y), and fm = f(x, 1-y). |
| 177 | * rx = log1p(Am1 + sqrt(Am1*(A+1))) |
| 178 | */ |
| 179 | if (y == 1 && x < DBL_EPSILON * DBL_EPSILON / 128) { |
| 180 | /* |
| 181 | * fp is of order x^2, and fm = x/2. |
| 182 | * A = 1 (inexactly). |
| 183 | */ |
| 184 | *rx = sqrt(x); |
| 185 | } else if (x >= DBL_EPSILON * fabs(y - 1)) { |
| 186 | /* |
| 187 | * Underflow will not occur because |
| 188 | * x >= DBL_EPSILON^2/128 >= FOUR_SQRT_MIN |
| 189 | */ |
| 190 | Am1 = f(x, 1 + y, R) + f(x, 1 - y, S); |
| 191 | *rx = log1p(Am1 + sqrt(Am1 * (A + 1))); |
| 192 | } else if (y < 1) { |
| 193 | /* |
| 194 | * fp = x*x/(1+y)/4, fm = x*x/(1-y)/4, and |
| 195 | * A = 1 (inexactly). |
| 196 | */ |
| 197 | *rx = x / sqrt((1 - y) * (1 + y)); |
| 198 | } else { /* if (y > 1) */ |
| 199 | /* |
| 200 | * A-1 = y-1 (inexactly). |
| 201 | */ |
| 202 | *rx = log1p((y - 1) + sqrt((y - 1) * (y + 1))); |
| 203 | } |
| 204 | } else { |
| 205 | *rx = log(A + sqrt(A * A - 1)); |
| 206 | } |
| 207 | |
| 208 | *new_y = y; |
| 209 | |
| 210 | if (y < FOUR_SQRT_MIN) { |
| 211 | /* |
| 212 | * Avoid a possible underflow caused by y/A. For casinh this |
| 213 | * would be legitimate, but will be picked up by invoking atan2 |
| 214 | * later on. For cacos this would not be legitimate. |
| 215 | */ |
| 216 | *B_is_usable = 0; |
| 217 | *sqrt_A2my2 = A * (2 / DBL_EPSILON); |
| 218 | *new_y = y * (2 / DBL_EPSILON); |
| 219 | return; |
| 220 | } |
| 221 | |
| 222 | /* B = (|z+I| - |z-I|) / 2 = y/A */ |
| 223 | *B = y / A; |
| 224 | *B_is_usable = 1; |
| 225 | |
| 226 | if (*B > B_crossover) { |
| 227 | *B_is_usable = 0; |
| 228 | /* |
| 229 | * Amy = fp + fm, where fp = f(x, y+1), and fm = f(x, y-1). |
| 230 | * sqrt_A2my2 = sqrt(Amy*(A+y)) |
| 231 | */ |
| 232 | if (y == 1 && x < DBL_EPSILON / 128) { |
| 233 | /* |
| 234 | * fp is of order x^2, and fm = x/2. |
| 235 | * A = 1 (inexactly). |
| 236 | */ |
| 237 | *sqrt_A2my2 = sqrt(x) * sqrt((A + y) / 2); |
| 238 | } else if (x >= DBL_EPSILON * fabs(y - 1)) { |
| 239 | /* |
| 240 | * Underflow will not occur because |
| 241 | * x >= DBL_EPSILON/128 >= FOUR_SQRT_MIN |
| 242 | * and |
| 243 | * x >= DBL_EPSILON^2 >= FOUR_SQRT_MIN |
| 244 | */ |
| 245 | Amy = f(x, y + 1, R) + f(x, y - 1, S); |
| 246 | *sqrt_A2my2 = sqrt(Amy * (A + y)); |
| 247 | } else if (y > 1) { |
| 248 | /* |
| 249 | * fp = x*x/(y+1)/4, fm = x*x/(y-1)/4, and |
| 250 | * A = y (inexactly). |
| 251 | * |
| 252 | * y < RECIP_EPSILON. So the following |
| 253 | * scaling should avoid any underflow problems. |
| 254 | */ |
| 255 | *sqrt_A2my2 = x * (4 / DBL_EPSILON / DBL_EPSILON) * y / |
| 256 | sqrt((y + 1) * (y - 1)); |
| 257 | *new_y = y * (4 / DBL_EPSILON / DBL_EPSILON); |
| 258 | } else { /* if (y < 1) */ |
| 259 | /* |
| 260 | * fm = 1-y >= DBL_EPSILON, fp is of order x^2, and |
| 261 | * A = 1 (inexactly). |
| 262 | */ |
| 263 | *sqrt_A2my2 = sqrt((1 - y) * (1 + y)); |
| 264 | } |
| 265 | } |
| 266 | } |
| 267 | |
| 268 | /* |
| 269 | * casinh(z) = z + O(z^3) as z -> 0 |
| 270 | * |
| 271 | * casinh(z) = sign(x)*clog(sign(x)*z) + O(1/z^2) as z -> infinity |
| 272 | * The above formula works for the imaginary part as well, because |
| 273 | * Im(casinh(z)) = sign(x)*atan2(sign(x)*y, fabs(x)) + O(y/z^3) |
| 274 | * as z -> infinity, uniformly in y |
| 275 | */ |
| 276 | double complex |
| 277 | casinh(double complex z) |
| 278 | { |
| 279 | double x, y, ax, ay, rx, ry, B, sqrt_A2my2, new_y; |
| 280 | int B_is_usable; |
| 281 | double complex w; |
| 282 | |
| 283 | x = creal(z); |
| 284 | y = cimag(z); |
| 285 | ax = fabs(x); |
| 286 | ay = fabs(y); |
| 287 | |
| 288 | if (isnan(x) || isnan(y)) { |
| 289 | /* casinh(+-Inf + I*NaN) = +-Inf + I*NaN */ |
| 290 | if (isinf(x)) |
Elliott Hughes | 8cff2f9 | 2015-08-28 20:21:43 -0700 | [diff] [blame] | 291 | return (CMPLX(x, y + y)); |
Elliott Hughes | b8ee16f | 2014-11-06 11:16:55 -0800 | [diff] [blame] | 292 | /* casinh(NaN + I*+-Inf) = opt(+-)Inf + I*NaN */ |
| 293 | if (isinf(y)) |
Elliott Hughes | 8cff2f9 | 2015-08-28 20:21:43 -0700 | [diff] [blame] | 294 | return (CMPLX(y, x + x)); |
Elliott Hughes | b8ee16f | 2014-11-06 11:16:55 -0800 | [diff] [blame] | 295 | /* casinh(NaN + I*0) = NaN + I*0 */ |
| 296 | if (y == 0) |
Elliott Hughes | 8cff2f9 | 2015-08-28 20:21:43 -0700 | [diff] [blame] | 297 | return (CMPLX(x + x, y)); |
Elliott Hughes | b8ee16f | 2014-11-06 11:16:55 -0800 | [diff] [blame] | 298 | /* |
| 299 | * All other cases involving NaN return NaN + I*NaN. |
| 300 | * C99 leaves it optional whether to raise invalid if one of |
| 301 | * the arguments is not NaN, so we opt not to raise it. |
| 302 | */ |
Elliott Hughes | ab52807 | 2018-07-24 00:01:52 +0000 | [diff] [blame] | 303 | return (CMPLX(nan_mix(x, y), nan_mix(x, y))); |
Elliott Hughes | b8ee16f | 2014-11-06 11:16:55 -0800 | [diff] [blame] | 304 | } |
| 305 | |
| 306 | if (ax > RECIP_EPSILON || ay > RECIP_EPSILON) { |
| 307 | /* clog...() will raise inexact unless x or y is infinite. */ |
| 308 | if (signbit(x) == 0) |
| 309 | w = clog_for_large_values(z) + m_ln2; |
| 310 | else |
| 311 | w = clog_for_large_values(-z) + m_ln2; |
Elliott Hughes | 8cff2f9 | 2015-08-28 20:21:43 -0700 | [diff] [blame] | 312 | return (CMPLX(copysign(creal(w), x), copysign(cimag(w), y))); |
Elliott Hughes | b8ee16f | 2014-11-06 11:16:55 -0800 | [diff] [blame] | 313 | } |
| 314 | |
| 315 | /* Avoid spuriously raising inexact for z = 0. */ |
| 316 | if (x == 0 && y == 0) |
| 317 | return (z); |
| 318 | |
| 319 | /* All remaining cases are inexact. */ |
| 320 | raise_inexact(); |
| 321 | |
| 322 | if (ax < SQRT_6_EPSILON / 4 && ay < SQRT_6_EPSILON / 4) |
| 323 | return (z); |
| 324 | |
| 325 | do_hard_work(ax, ay, &rx, &B_is_usable, &B, &sqrt_A2my2, &new_y); |
| 326 | if (B_is_usable) |
| 327 | ry = asin(B); |
| 328 | else |
| 329 | ry = atan2(new_y, sqrt_A2my2); |
Elliott Hughes | 8cff2f9 | 2015-08-28 20:21:43 -0700 | [diff] [blame] | 330 | return (CMPLX(copysign(rx, x), copysign(ry, y))); |
Elliott Hughes | b8ee16f | 2014-11-06 11:16:55 -0800 | [diff] [blame] | 331 | } |
| 332 | |
| 333 | /* |
| 334 | * casin(z) = reverse(casinh(reverse(z))) |
| 335 | * where reverse(x + I*y) = y + I*x = I*conj(z). |
| 336 | */ |
| 337 | double complex |
| 338 | casin(double complex z) |
| 339 | { |
Elliott Hughes | 8cff2f9 | 2015-08-28 20:21:43 -0700 | [diff] [blame] | 340 | double complex w = casinh(CMPLX(cimag(z), creal(z))); |
Elliott Hughes | b8ee16f | 2014-11-06 11:16:55 -0800 | [diff] [blame] | 341 | |
Elliott Hughes | 8cff2f9 | 2015-08-28 20:21:43 -0700 | [diff] [blame] | 342 | return (CMPLX(cimag(w), creal(w))); |
Elliott Hughes | b8ee16f | 2014-11-06 11:16:55 -0800 | [diff] [blame] | 343 | } |
| 344 | |
| 345 | /* |
| 346 | * cacos(z) = PI/2 - casin(z) |
| 347 | * but do the computation carefully so cacos(z) is accurate when z is |
| 348 | * close to 1. |
| 349 | * |
| 350 | * cacos(z) = PI/2 - z + O(z^3) as z -> 0 |
| 351 | * |
| 352 | * cacos(z) = -sign(y)*I*clog(z) + O(1/z^2) as z -> infinity |
| 353 | * The above formula works for the real part as well, because |
| 354 | * Re(cacos(z)) = atan2(fabs(y), x) + O(y/z^3) |
| 355 | * as z -> infinity, uniformly in y |
| 356 | */ |
| 357 | double complex |
| 358 | cacos(double complex z) |
| 359 | { |
| 360 | double x, y, ax, ay, rx, ry, B, sqrt_A2mx2, new_x; |
| 361 | int sx, sy; |
| 362 | int B_is_usable; |
| 363 | double complex w; |
| 364 | |
| 365 | x = creal(z); |
| 366 | y = cimag(z); |
| 367 | sx = signbit(x); |
| 368 | sy = signbit(y); |
| 369 | ax = fabs(x); |
| 370 | ay = fabs(y); |
| 371 | |
| 372 | if (isnan(x) || isnan(y)) { |
| 373 | /* cacos(+-Inf + I*NaN) = NaN + I*opt(-)Inf */ |
| 374 | if (isinf(x)) |
Elliott Hughes | 8cff2f9 | 2015-08-28 20:21:43 -0700 | [diff] [blame] | 375 | return (CMPLX(y + y, -INFINITY)); |
Elliott Hughes | b8ee16f | 2014-11-06 11:16:55 -0800 | [diff] [blame] | 376 | /* cacos(NaN + I*+-Inf) = NaN + I*-+Inf */ |
| 377 | if (isinf(y)) |
Elliott Hughes | 8cff2f9 | 2015-08-28 20:21:43 -0700 | [diff] [blame] | 378 | return (CMPLX(x + x, -y)); |
Elliott Hughes | b8ee16f | 2014-11-06 11:16:55 -0800 | [diff] [blame] | 379 | /* cacos(0 + I*NaN) = PI/2 + I*NaN with inexact */ |
| 380 | if (x == 0) |
Elliott Hughes | 8cff2f9 | 2015-08-28 20:21:43 -0700 | [diff] [blame] | 381 | return (CMPLX(pio2_hi + pio2_lo, y + y)); |
Elliott Hughes | b8ee16f | 2014-11-06 11:16:55 -0800 | [diff] [blame] | 382 | /* |
| 383 | * All other cases involving NaN return NaN + I*NaN. |
| 384 | * C99 leaves it optional whether to raise invalid if one of |
| 385 | * the arguments is not NaN, so we opt not to raise it. |
| 386 | */ |
Elliott Hughes | ab52807 | 2018-07-24 00:01:52 +0000 | [diff] [blame] | 387 | return (CMPLX(nan_mix(x, y), nan_mix(x, y))); |
Elliott Hughes | b8ee16f | 2014-11-06 11:16:55 -0800 | [diff] [blame] | 388 | } |
| 389 | |
| 390 | if (ax > RECIP_EPSILON || ay > RECIP_EPSILON) { |
| 391 | /* clog...() will raise inexact unless x or y is infinite. */ |
| 392 | w = clog_for_large_values(z); |
| 393 | rx = fabs(cimag(w)); |
| 394 | ry = creal(w) + m_ln2; |
| 395 | if (sy == 0) |
| 396 | ry = -ry; |
Elliott Hughes | 8cff2f9 | 2015-08-28 20:21:43 -0700 | [diff] [blame] | 397 | return (CMPLX(rx, ry)); |
Elliott Hughes | b8ee16f | 2014-11-06 11:16:55 -0800 | [diff] [blame] | 398 | } |
| 399 | |
| 400 | /* Avoid spuriously raising inexact for z = 1. */ |
| 401 | if (x == 1 && y == 0) |
Elliott Hughes | 8cff2f9 | 2015-08-28 20:21:43 -0700 | [diff] [blame] | 402 | return (CMPLX(0, -y)); |
Elliott Hughes | b8ee16f | 2014-11-06 11:16:55 -0800 | [diff] [blame] | 403 | |
| 404 | /* All remaining cases are inexact. */ |
| 405 | raise_inexact(); |
| 406 | |
| 407 | if (ax < SQRT_6_EPSILON / 4 && ay < SQRT_6_EPSILON / 4) |
Elliott Hughes | 8cff2f9 | 2015-08-28 20:21:43 -0700 | [diff] [blame] | 408 | return (CMPLX(pio2_hi - (x - pio2_lo), -y)); |
Elliott Hughes | b8ee16f | 2014-11-06 11:16:55 -0800 | [diff] [blame] | 409 | |
| 410 | do_hard_work(ay, ax, &ry, &B_is_usable, &B, &sqrt_A2mx2, &new_x); |
| 411 | if (B_is_usable) { |
| 412 | if (sx == 0) |
| 413 | rx = acos(B); |
| 414 | else |
| 415 | rx = acos(-B); |
| 416 | } else { |
| 417 | if (sx == 0) |
| 418 | rx = atan2(sqrt_A2mx2, new_x); |
| 419 | else |
| 420 | rx = atan2(sqrt_A2mx2, -new_x); |
| 421 | } |
| 422 | if (sy == 0) |
| 423 | ry = -ry; |
Elliott Hughes | 8cff2f9 | 2015-08-28 20:21:43 -0700 | [diff] [blame] | 424 | return (CMPLX(rx, ry)); |
Elliott Hughes | b8ee16f | 2014-11-06 11:16:55 -0800 | [diff] [blame] | 425 | } |
| 426 | |
| 427 | /* |
| 428 | * cacosh(z) = I*cacos(z) or -I*cacos(z) |
| 429 | * where the sign is chosen so Re(cacosh(z)) >= 0. |
| 430 | */ |
| 431 | double complex |
| 432 | cacosh(double complex z) |
| 433 | { |
| 434 | double complex w; |
| 435 | double rx, ry; |
| 436 | |
| 437 | w = cacos(z); |
| 438 | rx = creal(w); |
| 439 | ry = cimag(w); |
| 440 | /* cacosh(NaN + I*NaN) = NaN + I*NaN */ |
| 441 | if (isnan(rx) && isnan(ry)) |
Elliott Hughes | 8cff2f9 | 2015-08-28 20:21:43 -0700 | [diff] [blame] | 442 | return (CMPLX(ry, rx)); |
Elliott Hughes | b8ee16f | 2014-11-06 11:16:55 -0800 | [diff] [blame] | 443 | /* cacosh(NaN + I*+-Inf) = +Inf + I*NaN */ |
| 444 | /* cacosh(+-Inf + I*NaN) = +Inf + I*NaN */ |
| 445 | if (isnan(rx)) |
Elliott Hughes | 8cff2f9 | 2015-08-28 20:21:43 -0700 | [diff] [blame] | 446 | return (CMPLX(fabs(ry), rx)); |
Elliott Hughes | b8ee16f | 2014-11-06 11:16:55 -0800 | [diff] [blame] | 447 | /* cacosh(0 + I*NaN) = NaN + I*NaN */ |
| 448 | if (isnan(ry)) |
Elliott Hughes | 8cff2f9 | 2015-08-28 20:21:43 -0700 | [diff] [blame] | 449 | return (CMPLX(ry, ry)); |
| 450 | return (CMPLX(fabs(ry), copysign(rx, cimag(z)))); |
Elliott Hughes | b8ee16f | 2014-11-06 11:16:55 -0800 | [diff] [blame] | 451 | } |
| 452 | |
| 453 | /* |
| 454 | * Optimized version of clog() for |z| finite and larger than ~RECIP_EPSILON. |
| 455 | */ |
| 456 | static double complex |
| 457 | clog_for_large_values(double complex z) |
| 458 | { |
| 459 | double x, y; |
| 460 | double ax, ay, t; |
| 461 | |
| 462 | x = creal(z); |
| 463 | y = cimag(z); |
| 464 | ax = fabs(x); |
| 465 | ay = fabs(y); |
| 466 | if (ax < ay) { |
| 467 | t = ax; |
| 468 | ax = ay; |
| 469 | ay = t; |
| 470 | } |
| 471 | |
| 472 | /* |
| 473 | * Avoid overflow in hypot() when x and y are both very large. |
Elliott Hughes | 8da8ca4 | 2018-05-08 13:35:33 -0700 | [diff] [blame] | 474 | * Divide x and y by E, and then add 1 to the logarithm. This |
| 475 | * depends on E being larger than sqrt(2), since the return value of |
| 476 | * hypot cannot overflow if neither argument is greater in magnitude |
| 477 | * than 1/sqrt(2) of the maximum value of the return type. Likewise |
| 478 | * this determines the necessary threshold for using this method |
| 479 | * (however, actually use 1/2 instead as it is simpler). |
| 480 | * |
Elliott Hughes | b8ee16f | 2014-11-06 11:16:55 -0800 | [diff] [blame] | 481 | * Dividing by E causes an insignificant loss of accuracy; however |
| 482 | * this method is still poor since it is uneccessarily slow. |
| 483 | */ |
| 484 | if (ax > DBL_MAX / 2) |
Elliott Hughes | 8cff2f9 | 2015-08-28 20:21:43 -0700 | [diff] [blame] | 485 | return (CMPLX(log(hypot(x / m_e, y / m_e)) + 1, atan2(y, x))); |
Elliott Hughes | b8ee16f | 2014-11-06 11:16:55 -0800 | [diff] [blame] | 486 | |
| 487 | /* |
| 488 | * Avoid overflow when x or y is large. Avoid underflow when x or |
| 489 | * y is small. |
| 490 | */ |
| 491 | if (ax > QUARTER_SQRT_MAX || ay < SQRT_MIN) |
Elliott Hughes | 8cff2f9 | 2015-08-28 20:21:43 -0700 | [diff] [blame] | 492 | return (CMPLX(log(hypot(x, y)), atan2(y, x))); |
Elliott Hughes | b8ee16f | 2014-11-06 11:16:55 -0800 | [diff] [blame] | 493 | |
Elliott Hughes | 8cff2f9 | 2015-08-28 20:21:43 -0700 | [diff] [blame] | 494 | return (CMPLX(log(ax * ax + ay * ay) / 2, atan2(y, x))); |
Elliott Hughes | b8ee16f | 2014-11-06 11:16:55 -0800 | [diff] [blame] | 495 | } |
| 496 | |
| 497 | /* |
| 498 | * ================= |
| 499 | * | catanh, catan | |
| 500 | * ================= |
| 501 | */ |
| 502 | |
| 503 | /* |
| 504 | * sum_squares(x,y) = x*x + y*y (or just x*x if y*y would underflow). |
| 505 | * Assumes x*x and y*y will not overflow. |
| 506 | * Assumes x and y are finite. |
| 507 | * Assumes y is non-negative. |
| 508 | * Assumes fabs(x) >= DBL_EPSILON. |
| 509 | */ |
| 510 | static inline double |
| 511 | sum_squares(double x, double y) |
| 512 | { |
| 513 | |
| 514 | /* Avoid underflow when y is small. */ |
| 515 | if (y < SQRT_MIN) |
| 516 | return (x * x); |
| 517 | |
| 518 | return (x * x + y * y); |
| 519 | } |
| 520 | |
| 521 | /* |
| 522 | * real_part_reciprocal(x, y) = Re(1/(x+I*y)) = x/(x*x + y*y). |
| 523 | * Assumes x and y are not NaN, and one of x and y is larger than |
| 524 | * RECIP_EPSILON. We avoid unwarranted underflow. It is important to not use |
| 525 | * the code creal(1/z), because the imaginary part may produce an unwanted |
| 526 | * underflow. |
| 527 | * This is only called in a context where inexact is always raised before |
| 528 | * the call, so no effort is made to avoid or force inexact. |
| 529 | */ |
| 530 | static inline double |
| 531 | real_part_reciprocal(double x, double y) |
| 532 | { |
| 533 | double scale; |
| 534 | uint32_t hx, hy; |
| 535 | int32_t ix, iy; |
| 536 | |
| 537 | /* |
| 538 | * This code is inspired by the C99 document n1124.pdf, Section G.5.1, |
| 539 | * example 2. |
| 540 | */ |
| 541 | GET_HIGH_WORD(hx, x); |
| 542 | ix = hx & 0x7ff00000; |
| 543 | GET_HIGH_WORD(hy, y); |
| 544 | iy = hy & 0x7ff00000; |
| 545 | #define BIAS (DBL_MAX_EXP - 1) |
| 546 | /* XXX more guard digits are useful iff there is extra precision. */ |
| 547 | #define CUTOFF (DBL_MANT_DIG / 2 + 1) /* just half or 1 guard digit */ |
| 548 | if (ix - iy >= CUTOFF << 20 || isinf(x)) |
| 549 | return (1 / x); /* +-Inf -> +-0 is special */ |
| 550 | if (iy - ix >= CUTOFF << 20) |
| 551 | return (x / y / y); /* should avoid double div, but hard */ |
| 552 | if (ix <= (BIAS + DBL_MAX_EXP / 2 - CUTOFF) << 20) |
| 553 | return (x / (x * x + y * y)); |
| 554 | scale = 1; |
| 555 | SET_HIGH_WORD(scale, 0x7ff00000 - ix); /* 2**(1-ilogb(x)) */ |
| 556 | x *= scale; |
| 557 | y *= scale; |
| 558 | return (x / (x * x + y * y) * scale); |
| 559 | } |
| 560 | |
| 561 | /* |
| 562 | * catanh(z) = log((1+z)/(1-z)) / 2 |
| 563 | * = log1p(4*x / |z-1|^2) / 4 |
| 564 | * + I * atan2(2*y, (1-x)*(1+x)-y*y) / 2 |
| 565 | * |
| 566 | * catanh(z) = z + O(z^3) as z -> 0 |
| 567 | * |
| 568 | * catanh(z) = 1/z + sign(y)*I*PI/2 + O(1/z^3) as z -> infinity |
| 569 | * The above formula works for the real part as well, because |
| 570 | * Re(catanh(z)) = x/|z|^2 + O(x/z^4) |
| 571 | * as z -> infinity, uniformly in x |
| 572 | */ |
| 573 | double complex |
| 574 | catanh(double complex z) |
| 575 | { |
| 576 | double x, y, ax, ay, rx, ry; |
| 577 | |
| 578 | x = creal(z); |
| 579 | y = cimag(z); |
| 580 | ax = fabs(x); |
| 581 | ay = fabs(y); |
| 582 | |
| 583 | /* This helps handle many cases. */ |
| 584 | if (y == 0 && ax <= 1) |
Elliott Hughes | 8cff2f9 | 2015-08-28 20:21:43 -0700 | [diff] [blame] | 585 | return (CMPLX(atanh(x), y)); |
Elliott Hughes | b8ee16f | 2014-11-06 11:16:55 -0800 | [diff] [blame] | 586 | |
| 587 | /* To ensure the same accuracy as atan(), and to filter out z = 0. */ |
| 588 | if (x == 0) |
Elliott Hughes | 8cff2f9 | 2015-08-28 20:21:43 -0700 | [diff] [blame] | 589 | return (CMPLX(x, atan(y))); |
Elliott Hughes | b8ee16f | 2014-11-06 11:16:55 -0800 | [diff] [blame] | 590 | |
| 591 | if (isnan(x) || isnan(y)) { |
| 592 | /* catanh(+-Inf + I*NaN) = +-0 + I*NaN */ |
| 593 | if (isinf(x)) |
Elliott Hughes | 8cff2f9 | 2015-08-28 20:21:43 -0700 | [diff] [blame] | 594 | return (CMPLX(copysign(0, x), y + y)); |
Elliott Hughes | b8ee16f | 2014-11-06 11:16:55 -0800 | [diff] [blame] | 595 | /* catanh(NaN + I*+-Inf) = sign(NaN)0 + I*+-PI/2 */ |
| 596 | if (isinf(y)) |
Elliott Hughes | 8cff2f9 | 2015-08-28 20:21:43 -0700 | [diff] [blame] | 597 | return (CMPLX(copysign(0, x), |
Elliott Hughes | b8ee16f | 2014-11-06 11:16:55 -0800 | [diff] [blame] | 598 | copysign(pio2_hi + pio2_lo, y))); |
| 599 | /* |
| 600 | * All other cases involving NaN return NaN + I*NaN. |
| 601 | * C99 leaves it optional whether to raise invalid if one of |
| 602 | * the arguments is not NaN, so we opt not to raise it. |
| 603 | */ |
Elliott Hughes | ab52807 | 2018-07-24 00:01:52 +0000 | [diff] [blame] | 604 | return (CMPLX(nan_mix(x, y), nan_mix(x, y))); |
Elliott Hughes | b8ee16f | 2014-11-06 11:16:55 -0800 | [diff] [blame] | 605 | } |
| 606 | |
| 607 | if (ax > RECIP_EPSILON || ay > RECIP_EPSILON) |
Elliott Hughes | 8cff2f9 | 2015-08-28 20:21:43 -0700 | [diff] [blame] | 608 | return (CMPLX(real_part_reciprocal(x, y), |
Elliott Hughes | b8ee16f | 2014-11-06 11:16:55 -0800 | [diff] [blame] | 609 | copysign(pio2_hi + pio2_lo, y))); |
| 610 | |
| 611 | if (ax < SQRT_3_EPSILON / 2 && ay < SQRT_3_EPSILON / 2) { |
| 612 | /* |
| 613 | * z = 0 was filtered out above. All other cases must raise |
Elliott Hughes | 8da8ca4 | 2018-05-08 13:35:33 -0700 | [diff] [blame] | 614 | * inexact, but this is the only case that needs to do it |
Elliott Hughes | b8ee16f | 2014-11-06 11:16:55 -0800 | [diff] [blame] | 615 | * explicitly. |
| 616 | */ |
| 617 | raise_inexact(); |
| 618 | return (z); |
| 619 | } |
| 620 | |
| 621 | if (ax == 1 && ay < DBL_EPSILON) |
| 622 | rx = (m_ln2 - log(ay)) / 2; |
| 623 | else |
| 624 | rx = log1p(4 * ax / sum_squares(ax - 1, ay)) / 4; |
| 625 | |
| 626 | if (ax == 1) |
| 627 | ry = atan2(2, -ay) / 2; |
| 628 | else if (ay < DBL_EPSILON) |
| 629 | ry = atan2(2 * ay, (1 - ax) * (1 + ax)) / 2; |
| 630 | else |
| 631 | ry = atan2(2 * ay, (1 - ax) * (1 + ax) - ay * ay) / 2; |
| 632 | |
Elliott Hughes | 8cff2f9 | 2015-08-28 20:21:43 -0700 | [diff] [blame] | 633 | return (CMPLX(copysign(rx, x), copysign(ry, y))); |
Elliott Hughes | b8ee16f | 2014-11-06 11:16:55 -0800 | [diff] [blame] | 634 | } |
| 635 | |
| 636 | /* |
| 637 | * catan(z) = reverse(catanh(reverse(z))) |
| 638 | * where reverse(x + I*y) = y + I*x = I*conj(z). |
| 639 | */ |
| 640 | double complex |
| 641 | catan(double complex z) |
| 642 | { |
Elliott Hughes | 8cff2f9 | 2015-08-28 20:21:43 -0700 | [diff] [blame] | 643 | double complex w = catanh(CMPLX(cimag(z), creal(z))); |
Elliott Hughes | b8ee16f | 2014-11-06 11:16:55 -0800 | [diff] [blame] | 644 | |
Elliott Hughes | 8cff2f9 | 2015-08-28 20:21:43 -0700 | [diff] [blame] | 645 | return (CMPLX(cimag(w), creal(w))); |
Elliott Hughes | b8ee16f | 2014-11-06 11:16:55 -0800 | [diff] [blame] | 646 | } |
Elliott Hughes | 8da8ca4 | 2018-05-08 13:35:33 -0700 | [diff] [blame] | 647 | |
| 648 | #if LDBL_MANT_DIG == 53 |
| 649 | __weak_reference(cacosh, cacoshl); |
| 650 | __weak_reference(cacos, cacosl); |
| 651 | __weak_reference(casinh, casinhl); |
| 652 | __weak_reference(casin, casinl); |
| 653 | __weak_reference(catanh, catanhl); |
| 654 | __weak_reference(catan, catanl); |
| 655 | #endif |