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 | #include <fenv_private.h> |
64 | #include <math-underflow.h> |
65 | |
66 | static const long double |
67 | invsqrtpi = 5.64189583547756286948079e-1L, two = 2.0e0L, one = 1.0e0L; |
68 | |
69 | static const long double zero = 0.0L; |
70 | |
71 | long double |
72 | __ieee754_jnl (int n, long double x) |
73 | { |
74 | uint32_t se, i0, i1; |
75 | int32_t i, ix, sgn; |
76 | long double a, b, temp, di, ret; |
77 | long double z, w; |
78 | |
79 | /* J(-n,x) = (-1)^n * J(n, x), J(n, -x) = (-1)^n * J(n, x) |
80 | * Thus, J(-n,x) = J(n,-x) |
81 | */ |
82 | |
83 | GET_LDOUBLE_WORDS (se, i0, i1, x); |
84 | ix = se & 0x7fff; |
85 | |
86 | /* if J(n,NaN) is NaN */ |
87 | if (__glibc_unlikely ((ix == 0x7fff) && ((i0 & 0x7fffffff) != 0))) |
88 | return x + x; |
89 | if (n < 0) |
90 | { |
91 | n = -n; |
92 | x = -x; |
93 | se ^= 0x8000; |
94 | } |
95 | if (n == 0) |
96 | return (__ieee754_j0l (x)); |
97 | if (n == 1) |
98 | return (__ieee754_j1l (x)); |
99 | sgn = (n & 1) & (se >> 15); /* even n -- 0, odd n -- sign(x) */ |
100 | x = fabsl (x); |
101 | { |
102 | SET_RESTORE_ROUNDL (FE_TONEAREST); |
103 | if (__glibc_unlikely ((ix | i0 | i1) == 0 || ix >= 0x7fff)) |
104 | /* if x is 0 or inf */ |
105 | return sgn == 1 ? -zero : zero; |
106 | else if ((long double) n <= x) |
107 | { |
108 | /* Safe to use J(n+1,x)=2n/x *J(n,x)-J(n-1,x) */ |
109 | if (ix >= 0x412D) |
110 | { /* x > 2**302 */ |
111 | |
112 | /* ??? This might be a futile gesture. |
113 | If x exceeds X_TLOSS anyway, the wrapper function |
114 | will set the result to zero. */ |
115 | |
116 | /* (x >> n**2) |
117 | * Jn(x) = cos(x-(2n+1)*pi/4)*sqrt(2/x*pi) |
118 | * Yn(x) = sin(x-(2n+1)*pi/4)*sqrt(2/x*pi) |
119 | * Let s=sin(x), c=cos(x), |
120 | * xn=x-(2n+1)*pi/4, sqt2 = sqrt(2),then |
121 | * |
122 | * n sin(xn)*sqt2 cos(xn)*sqt2 |
123 | * ---------------------------------- |
124 | * 0 s-c c+s |
125 | * 1 -s-c -c+s |
126 | * 2 -s+c -c-s |
127 | * 3 s+c c-s |
128 | */ |
129 | long double s; |
130 | long double c; |
131 | __sincosl (x, &s, &c); |
132 | switch (n & 3) |
133 | { |
134 | case 0: |
135 | temp = c + s; |
136 | break; |
137 | case 1: |
138 | temp = -c + s; |
139 | break; |
140 | case 2: |
141 | temp = -c - s; |
142 | break; |
143 | case 3: |
144 | temp = c - s; |
145 | break; |
146 | default: |
147 | __builtin_unreachable (); |
148 | } |
149 | b = invsqrtpi * temp / sqrtl (x); |
150 | } |
151 | else |
152 | { |
153 | a = __ieee754_j0l (x); |
154 | b = __ieee754_j1l (x); |
155 | for (i = 1; i < n; i++) |
156 | { |
157 | temp = b; |
158 | b = b * ((long double) (i + i) / x) - a; /* avoid underflow */ |
159 | a = temp; |
160 | } |
161 | } |
162 | } |
163 | else |
164 | { |
165 | if (ix < 0x3fde) |
166 | { /* x < 2**-33 */ |
167 | /* x is tiny, return the first Taylor expansion of J(n,x) |
168 | * J(n,x) = 1/n!*(x/2)^n - ... |
169 | */ |
170 | if (n >= 400) /* underflow, result < 10^-4952 */ |
171 | b = zero; |
172 | else |
173 | { |
174 | temp = x * 0.5; |
175 | b = temp; |
176 | for (a = one, i = 2; i <= n; i++) |
177 | { |
178 | a *= (long double) i; /* a = n! */ |
179 | b *= temp; /* b = (x/2)^n */ |
180 | } |
181 | b = b / a; |
182 | } |
183 | } |
184 | else |
185 | { |
186 | /* use backward recurrence */ |
187 | /* x x^2 x^2 |
188 | * J(n,x)/J(n-1,x) = ---- ------ ------ ..... |
189 | * 2n - 2(n+1) - 2(n+2) |
190 | * |
191 | * 1 1 1 |
192 | * (for large x) = ---- ------ ------ ..... |
193 | * 2n 2(n+1) 2(n+2) |
194 | * -- - ------ - ------ - |
195 | * x x x |
196 | * |
197 | * Let w = 2n/x and h=2/x, then the above quotient |
198 | * is equal to the continued fraction: |
199 | * 1 |
200 | * = ----------------------- |
201 | * 1 |
202 | * w - ----------------- |
203 | * 1 |
204 | * w+h - --------- |
205 | * w+2h - ... |
206 | * |
207 | * To determine how many terms needed, let |
208 | * Q(0) = w, Q(1) = w(w+h) - 1, |
209 | * Q(k) = (w+k*h)*Q(k-1) - Q(k-2), |
210 | * When Q(k) > 1e4 good for single |
211 | * When Q(k) > 1e9 good for double |
212 | * When Q(k) > 1e17 good for quadruple |
213 | */ |
214 | /* determine k */ |
215 | long double t, v; |
216 | long double q0, q1, h, tmp; |
217 | int32_t k, m; |
218 | w = (n + n) / (long double) x; |
219 | h = 2.0L / (long double) x; |
220 | q0 = w; |
221 | z = w + h; |
222 | q1 = w * z - 1.0L; |
223 | k = 1; |
224 | while (q1 < 1.0e11L) |
225 | { |
226 | k += 1; |
227 | z += h; |
228 | tmp = z * q1 - q0; |
229 | q0 = q1; |
230 | q1 = tmp; |
231 | } |
232 | m = n + n; |
233 | for (t = zero, i = 2 * (n + k); i >= m; i -= 2) |
234 | t = one / (i / x - t); |
235 | a = t; |
236 | b = one; |
237 | /* estimate log((2/x)^n*n!) = n*log(2/x)+n*ln(n) |
238 | * Hence, if n*(log(2n/x)) > ... |
239 | * single 8.8722839355e+01 |
240 | * double 7.09782712893383973096e+02 |
241 | * long double 1.1356523406294143949491931077970765006170e+04 |
242 | * then recurrent value may overflow and the result is |
243 | * likely underflow to zero |
244 | */ |
245 | tmp = n; |
246 | v = two / x; |
247 | tmp = tmp * __ieee754_logl (fabsl (v * tmp)); |
248 | |
249 | if (tmp < 1.1356523406294143949491931077970765006170e+04L) |
250 | { |
251 | for (i = n - 1, di = (long double) (i + i); i > 0; i--) |
252 | { |
253 | temp = b; |
254 | b *= di; |
255 | b = b / x - a; |
256 | a = temp; |
257 | di -= two; |
258 | } |
259 | } |
260 | else |
261 | { |
262 | for (i = n - 1, di = (long double) (i + i); i > 0; i--) |
263 | { |
264 | temp = b; |
265 | b *= di; |
266 | b = b / x - a; |
267 | a = temp; |
268 | di -= two; |
269 | /* scale b to avoid spurious overflow */ |
270 | if (b > 1e100L) |
271 | { |
272 | a /= b; |
273 | t /= b; |
274 | b = one; |
275 | } |
276 | } |
277 | } |
278 | /* j0() and j1() suffer enormous loss of precision at and |
279 | * near zero; however, we know that their zero points never |
280 | * coincide, so just choose the one further away from zero. |
281 | */ |
282 | z = __ieee754_j0l (x); |
283 | w = __ieee754_j1l (x); |
284 | if (fabsl (z) >= fabsl (w)) |
285 | b = (t * z / b); |
286 | else |
287 | b = (t * w / a); |
288 | } |
289 | } |
290 | if (sgn == 1) |
291 | ret = -b; |
292 | else |
293 | ret = b; |
294 | } |
295 | if (ret == 0) |
296 | { |
297 | ret = copysignl (LDBL_MIN, ret) * LDBL_MIN; |
298 | __set_errno (ERANGE); |
299 | } |
300 | else |
301 | math_check_force_underflow (ret); |
302 | return ret; |
303 | } |
304 | strong_alias (__ieee754_jnl, __jnl_finite) |
305 | |
306 | long double |
307 | __ieee754_ynl (int n, long double x) |
308 | { |
309 | uint32_t se, i0, i1; |
310 | int32_t i, ix; |
311 | int32_t sign; |
312 | long double a, b, temp, ret; |
313 | |
314 | |
315 | GET_LDOUBLE_WORDS (se, i0, i1, x); |
316 | ix = se & 0x7fff; |
317 | /* if Y(n,NaN) is NaN */ |
318 | if (__builtin_expect ((ix == 0x7fff) && ((i0 & 0x7fffffff) != 0), 0)) |
319 | return x + x; |
320 | if (__builtin_expect ((ix | i0 | i1) == 0, 0)) |
321 | /* -inf or inf and divide-by-zero exception. */ |
322 | return ((n < 0 && (n & 1) != 0) ? 1.0L : -1.0L) / 0.0L; |
323 | if (__builtin_expect (se & 0x8000, 0)) |
324 | return zero / (zero * x); |
325 | sign = 1; |
326 | if (n < 0) |
327 | { |
328 | n = -n; |
329 | sign = 1 - ((n & 1) << 1); |
330 | } |
331 | if (n == 0) |
332 | return (__ieee754_y0l (x)); |
333 | { |
334 | SET_RESTORE_ROUNDL (FE_TONEAREST); |
335 | if (n == 1) |
336 | { |
337 | ret = sign * __ieee754_y1l (x); |
338 | goto out; |
339 | } |
340 | if (__glibc_unlikely (ix == 0x7fff)) |
341 | return zero; |
342 | if (ix >= 0x412D) |
343 | { /* x > 2**302 */ |
344 | |
345 | /* ??? See comment above on the possible futility of this. */ |
346 | |
347 | /* (x >> n**2) |
348 | * Jn(x) = cos(x-(2n+1)*pi/4)*sqrt(2/x*pi) |
349 | * Yn(x) = sin(x-(2n+1)*pi/4)*sqrt(2/x*pi) |
350 | * Let s=sin(x), c=cos(x), |
351 | * xn=x-(2n+1)*pi/4, sqt2 = sqrt(2),then |
352 | * |
353 | * n sin(xn)*sqt2 cos(xn)*sqt2 |
354 | * ---------------------------------- |
355 | * 0 s-c c+s |
356 | * 1 -s-c -c+s |
357 | * 2 -s+c -c-s |
358 | * 3 s+c c-s |
359 | */ |
360 | long double s; |
361 | long double c; |
362 | __sincosl (x, &s, &c); |
363 | switch (n & 3) |
364 | { |
365 | case 0: |
366 | temp = s - c; |
367 | break; |
368 | case 1: |
369 | temp = -s - c; |
370 | break; |
371 | case 2: |
372 | temp = -s + c; |
373 | break; |
374 | case 3: |
375 | temp = s + c; |
376 | break; |
377 | default: |
378 | __builtin_unreachable (); |
379 | } |
380 | b = invsqrtpi * temp / sqrtl (x); |
381 | } |
382 | else |
383 | { |
384 | a = __ieee754_y0l (x); |
385 | b = __ieee754_y1l (x); |
386 | /* quit if b is -inf */ |
387 | GET_LDOUBLE_WORDS (se, i0, i1, b); |
388 | /* Use 0xffffffff since GET_LDOUBLE_WORDS sign-extends SE. */ |
389 | for (i = 1; i < n && se != 0xffffffff; i++) |
390 | { |
391 | temp = b; |
392 | b = ((long double) (i + i) / x) * b - a; |
393 | GET_LDOUBLE_WORDS (se, i0, i1, b); |
394 | a = temp; |
395 | } |
396 | } |
397 | /* If B is +-Inf, set up errno accordingly. */ |
398 | if (! isfinite (b)) |
399 | __set_errno (ERANGE); |
400 | if (sign > 0) |
401 | ret = b; |
402 | else |
403 | ret = -b; |
404 | } |
405 | out: |
406 | if (isinf (ret)) |
407 | ret = copysignl (LDBL_MAX, ret) * LDBL_MAX; |
408 | return ret; |
409 | } |
410 | strong_alias (__ieee754_ynl, __ynl_finite) |
411 | |