1 | /* |
2 | * ==================================================== |
3 | * Copyright (C) 1993 by Sun Microsystems, Inc. All rights reserved. |
4 | * |
5 | * Developed at SunPro, a Sun Microsystems, Inc. business. |
6 | * Permission to use, copy, modify, and distribute this |
7 | * software is freely granted, provided that this notice |
8 | * is preserved. |
9 | * ==================================================== |
10 | */ |
11 | |
12 | /* Modifications for long double are |
13 | Copyright (C) 2001 Stephen L. Moshier <moshier@na-net.ornl.gov> |
14 | and are incorporated herein by permission of the author. The author |
15 | reserves the right to distribute this material elsewhere under different |
16 | copying permissions. These modifications are distributed here under |
17 | the following terms: |
18 | |
19 | This library is free software; you can redistribute it and/or |
20 | modify it under the terms of the GNU Lesser General Public |
21 | License as published by the Free Software Foundation; either |
22 | version 2.1 of the License, or (at your option) any later version. |
23 | |
24 | This library is distributed in the hope that it will be useful, |
25 | but WITHOUT ANY WARRANTY; without even the implied warranty of |
26 | MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU |
27 | Lesser General Public License for more details. |
28 | |
29 | You should have received a copy of the GNU Lesser General Public |
30 | License along with this library; if not, see |
31 | <http://www.gnu.org/licenses/>. */ |
32 | |
33 | /* |
34 | * __ieee754_jn(n, x), __ieee754_yn(n, x) |
35 | * floating point Bessel's function of the 1st and 2nd kind |
36 | * of order n |
37 | * |
38 | * Special cases: |
39 | * y0(0)=y1(0)=yn(n,0) = -inf with overflow signal; |
40 | * y0(-ve)=y1(-ve)=yn(n,-ve) are NaN with invalid signal. |
41 | * Note 2. About jn(n,x), yn(n,x) |
42 | * For n=0, j0(x) is called, |
43 | * for n=1, j1(x) is called, |
44 | * for n<x, forward recursion us used starting |
45 | * from values of j0(x) and j1(x). |
46 | * for n>x, a continued fraction approximation to |
47 | * j(n,x)/j(n-1,x) is evaluated and then backward |
48 | * recursion is used starting from a supposed value |
49 | * for j(n,x). The resulting value of j(0,x) is |
50 | * compared with the actual value to correct the |
51 | * supposed value of j(n,x). |
52 | * |
53 | * yn(n,x) is similar in all respects, except |
54 | * that forward recursion is used for all |
55 | * values of n>1. |
56 | * |
57 | */ |
58 | |
59 | #include <errno.h> |
60 | #include <float.h> |
61 | #include <math.h> |
62 | #include <math_private.h> |
63 | |
64 | static const long double |
65 | invsqrtpi = 5.64189583547756286948079e-1L, two = 2.0e0L, one = 1.0e0L; |
66 | |
67 | static const long double zero = 0.0L; |
68 | |
69 | long double |
70 | __ieee754_jnl (int n, long double x) |
71 | { |
72 | uint32_t se, i0, i1; |
73 | int32_t i, ix, sgn; |
74 | long double a, b, temp, di, ret; |
75 | long double z, w; |
76 | |
77 | /* J(-n,x) = (-1)^n * J(n, x), J(n, -x) = (-1)^n * J(n, x) |
78 | * Thus, J(-n,x) = J(n,-x) |
79 | */ |
80 | |
81 | GET_LDOUBLE_WORDS (se, i0, i1, x); |
82 | ix = se & 0x7fff; |
83 | |
84 | /* if J(n,NaN) is NaN */ |
85 | if (__glibc_unlikely ((ix == 0x7fff) && ((i0 & 0x7fffffff) != 0))) |
86 | return x + x; |
87 | if (n < 0) |
88 | { |
89 | n = -n; |
90 | x = -x; |
91 | se ^= 0x8000; |
92 | } |
93 | if (n == 0) |
94 | return (__ieee754_j0l (x)); |
95 | if (n == 1) |
96 | return (__ieee754_j1l (x)); |
97 | sgn = (n & 1) & (se >> 15); /* even n -- 0, odd n -- sign(x) */ |
98 | x = fabsl (x); |
99 | { |
100 | SET_RESTORE_ROUNDL (FE_TONEAREST); |
101 | if (__glibc_unlikely ((ix | i0 | i1) == 0 || ix >= 0x7fff)) |
102 | /* if x is 0 or inf */ |
103 | return sgn == 1 ? -zero : zero; |
104 | else if ((long double) n <= x) |
105 | { |
106 | /* Safe to use J(n+1,x)=2n/x *J(n,x)-J(n-1,x) */ |
107 | if (ix >= 0x412D) |
108 | { /* x > 2**302 */ |
109 | |
110 | /* ??? This might be a futile gesture. |
111 | If x exceeds X_TLOSS anyway, the wrapper function |
112 | will set the result to zero. */ |
113 | |
114 | /* (x >> n**2) |
115 | * Jn(x) = cos(x-(2n+1)*pi/4)*sqrt(2/x*pi) |
116 | * Yn(x) = sin(x-(2n+1)*pi/4)*sqrt(2/x*pi) |
117 | * Let s=sin(x), c=cos(x), |
118 | * xn=x-(2n+1)*pi/4, sqt2 = sqrt(2),then |
119 | * |
120 | * n sin(xn)*sqt2 cos(xn)*sqt2 |
121 | * ---------------------------------- |
122 | * 0 s-c c+s |
123 | * 1 -s-c -c+s |
124 | * 2 -s+c -c-s |
125 | * 3 s+c c-s |
126 | */ |
127 | long double s; |
128 | long double c; |
129 | __sincosl (x, &s, &c); |
130 | switch (n & 3) |
131 | { |
132 | case 0: |
133 | temp = c + s; |
134 | break; |
135 | case 1: |
136 | temp = -c + s; |
137 | break; |
138 | case 2: |
139 | temp = -c - s; |
140 | break; |
141 | case 3: |
142 | temp = c - s; |
143 | break; |
144 | } |
145 | b = invsqrtpi * temp / __ieee754_sqrtl (x); |
146 | } |
147 | else |
148 | { |
149 | a = __ieee754_j0l (x); |
150 | b = __ieee754_j1l (x); |
151 | for (i = 1; i < n; i++) |
152 | { |
153 | temp = b; |
154 | b = b * ((long double) (i + i) / x) - a; /* avoid underflow */ |
155 | a = temp; |
156 | } |
157 | } |
158 | } |
159 | else |
160 | { |
161 | if (ix < 0x3fde) |
162 | { /* x < 2**-33 */ |
163 | /* x is tiny, return the first Taylor expansion of J(n,x) |
164 | * J(n,x) = 1/n!*(x/2)^n - ... |
165 | */ |
166 | if (n >= 400) /* underflow, result < 10^-4952 */ |
167 | b = zero; |
168 | else |
169 | { |
170 | temp = x * 0.5; |
171 | b = temp; |
172 | for (a = one, i = 2; i <= n; i++) |
173 | { |
174 | a *= (long double) i; /* a = n! */ |
175 | b *= temp; /* b = (x/2)^n */ |
176 | } |
177 | b = b / a; |
178 | } |
179 | } |
180 | else |
181 | { |
182 | /* use backward recurrence */ |
183 | /* x x^2 x^2 |
184 | * J(n,x)/J(n-1,x) = ---- ------ ------ ..... |
185 | * 2n - 2(n+1) - 2(n+2) |
186 | * |
187 | * 1 1 1 |
188 | * (for large x) = ---- ------ ------ ..... |
189 | * 2n 2(n+1) 2(n+2) |
190 | * -- - ------ - ------ - |
191 | * x x x |
192 | * |
193 | * Let w = 2n/x and h=2/x, then the above quotient |
194 | * is equal to the continued fraction: |
195 | * 1 |
196 | * = ----------------------- |
197 | * 1 |
198 | * w - ----------------- |
199 | * 1 |
200 | * w+h - --------- |
201 | * w+2h - ... |
202 | * |
203 | * To determine how many terms needed, let |
204 | * Q(0) = w, Q(1) = w(w+h) - 1, |
205 | * Q(k) = (w+k*h)*Q(k-1) - Q(k-2), |
206 | * When Q(k) > 1e4 good for single |
207 | * When Q(k) > 1e9 good for double |
208 | * When Q(k) > 1e17 good for quadruple |
209 | */ |
210 | /* determine k */ |
211 | long double t, v; |
212 | long double q0, q1, h, tmp; |
213 | int32_t k, m; |
214 | w = (n + n) / (long double) x; |
215 | h = 2.0L / (long double) x; |
216 | q0 = w; |
217 | z = w + h; |
218 | q1 = w * z - 1.0L; |
219 | k = 1; |
220 | while (q1 < 1.0e11L) |
221 | { |
222 | k += 1; |
223 | z += h; |
224 | tmp = z * q1 - q0; |
225 | q0 = q1; |
226 | q1 = tmp; |
227 | } |
228 | m = n + n; |
229 | for (t = zero, i = 2 * (n + k); i >= m; i -= 2) |
230 | t = one / (i / x - t); |
231 | a = t; |
232 | b = one; |
233 | /* estimate log((2/x)^n*n!) = n*log(2/x)+n*ln(n) |
234 | * Hence, if n*(log(2n/x)) > ... |
235 | * single 8.8722839355e+01 |
236 | * double 7.09782712893383973096e+02 |
237 | * long double 1.1356523406294143949491931077970765006170e+04 |
238 | * then recurrent value may overflow and the result is |
239 | * likely underflow to zero |
240 | */ |
241 | tmp = n; |
242 | v = two / x; |
243 | tmp = tmp * __ieee754_logl (fabsl (v * tmp)); |
244 | |
245 | if (tmp < 1.1356523406294143949491931077970765006170e+04L) |
246 | { |
247 | for (i = n - 1, di = (long double) (i + i); i > 0; i--) |
248 | { |
249 | temp = b; |
250 | b *= di; |
251 | b = b / x - a; |
252 | a = temp; |
253 | di -= two; |
254 | } |
255 | } |
256 | else |
257 | { |
258 | for (i = n - 1, di = (long double) (i + i); i > 0; i--) |
259 | { |
260 | temp = b; |
261 | b *= di; |
262 | b = b / x - a; |
263 | a = temp; |
264 | di -= two; |
265 | /* scale b to avoid spurious overflow */ |
266 | if (b > 1e100L) |
267 | { |
268 | a /= b; |
269 | t /= b; |
270 | b = one; |
271 | } |
272 | } |
273 | } |
274 | /* j0() and j1() suffer enormous loss of precision at and |
275 | * near zero; however, we know that their zero points never |
276 | * coincide, so just choose the one further away from zero. |
277 | */ |
278 | z = __ieee754_j0l (x); |
279 | w = __ieee754_j1l (x); |
280 | if (fabsl (z) >= fabsl (w)) |
281 | b = (t * z / b); |
282 | else |
283 | b = (t * w / a); |
284 | } |
285 | } |
286 | if (sgn == 1) |
287 | ret = -b; |
288 | else |
289 | ret = b; |
290 | } |
291 | if (ret == 0) |
292 | { |
293 | ret = __copysignl (LDBL_MIN, ret) * LDBL_MIN; |
294 | __set_errno (ERANGE); |
295 | } |
296 | else |
297 | math_check_force_underflow (ret); |
298 | return ret; |
299 | } |
300 | strong_alias (__ieee754_jnl, __jnl_finite) |
301 | |
302 | long double |
303 | __ieee754_ynl (int n, long double x) |
304 | { |
305 | uint32_t se, i0, i1; |
306 | int32_t i, ix; |
307 | int32_t sign; |
308 | long double a, b, temp, ret; |
309 | |
310 | |
311 | GET_LDOUBLE_WORDS (se, i0, i1, x); |
312 | ix = se & 0x7fff; |
313 | /* if Y(n,NaN) is NaN */ |
314 | if (__builtin_expect ((ix == 0x7fff) && ((i0 & 0x7fffffff) != 0), 0)) |
315 | return x + x; |
316 | if (__builtin_expect ((ix | i0 | i1) == 0, 0)) |
317 | /* -inf or inf and divide-by-zero exception. */ |
318 | return ((n < 0 && (n & 1) != 0) ? 1.0L : -1.0L) / 0.0L; |
319 | if (__builtin_expect (se & 0x8000, 0)) |
320 | return zero / (zero * x); |
321 | sign = 1; |
322 | if (n < 0) |
323 | { |
324 | n = -n; |
325 | sign = 1 - ((n & 1) << 1); |
326 | } |
327 | if (n == 0) |
328 | return (__ieee754_y0l (x)); |
329 | { |
330 | SET_RESTORE_ROUNDL (FE_TONEAREST); |
331 | if (n == 1) |
332 | { |
333 | ret = sign * __ieee754_y1l (x); |
334 | goto out; |
335 | } |
336 | if (__glibc_unlikely (ix == 0x7fff)) |
337 | return zero; |
338 | if (ix >= 0x412D) |
339 | { /* x > 2**302 */ |
340 | |
341 | /* ??? See comment above on the possible futility of this. */ |
342 | |
343 | /* (x >> n**2) |
344 | * Jn(x) = cos(x-(2n+1)*pi/4)*sqrt(2/x*pi) |
345 | * Yn(x) = sin(x-(2n+1)*pi/4)*sqrt(2/x*pi) |
346 | * Let s=sin(x), c=cos(x), |
347 | * xn=x-(2n+1)*pi/4, sqt2 = sqrt(2),then |
348 | * |
349 | * n sin(xn)*sqt2 cos(xn)*sqt2 |
350 | * ---------------------------------- |
351 | * 0 s-c c+s |
352 | * 1 -s-c -c+s |
353 | * 2 -s+c -c-s |
354 | * 3 s+c c-s |
355 | */ |
356 | long double s; |
357 | long double c; |
358 | __sincosl (x, &s, &c); |
359 | switch (n & 3) |
360 | { |
361 | case 0: |
362 | temp = s - c; |
363 | break; |
364 | case 1: |
365 | temp = -s - c; |
366 | break; |
367 | case 2: |
368 | temp = -s + c; |
369 | break; |
370 | case 3: |
371 | temp = s + c; |
372 | break; |
373 | } |
374 | b = invsqrtpi * temp / __ieee754_sqrtl (x); |
375 | } |
376 | else |
377 | { |
378 | a = __ieee754_y0l (x); |
379 | b = __ieee754_y1l (x); |
380 | /* quit if b is -inf */ |
381 | GET_LDOUBLE_WORDS (se, i0, i1, b); |
382 | /* Use 0xffffffff since GET_LDOUBLE_WORDS sign-extends SE. */ |
383 | for (i = 1; i < n && se != 0xffffffff; i++) |
384 | { |
385 | temp = b; |
386 | b = ((long double) (i + i) / x) * b - a; |
387 | GET_LDOUBLE_WORDS (se, i0, i1, b); |
388 | a = temp; |
389 | } |
390 | } |
391 | /* If B is +-Inf, set up errno accordingly. */ |
392 | if (! isfinite (b)) |
393 | __set_errno (ERANGE); |
394 | if (sign > 0) |
395 | ret = b; |
396 | else |
397 | ret = -b; |
398 | } |
399 | out: |
400 | if (isinf (ret)) |
401 | ret = __copysignl (LDBL_MAX, ret) * LDBL_MAX; |
402 | return ret; |
403 | } |
404 | strong_alias (__ieee754_ynl, __ynl_finite) |
405 | |