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 128-bit 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 division by zero 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 _Float128 |
65 | invsqrtpi = L(5.6418958354775628694807945156077258584405E-1), |
66 | two = 2, |
67 | one = 1, |
68 | zero = 0; |
69 | |
70 | |
71 | _Float128 |
72 | __ieee754_jnl (int n, _Float128 x) |
73 | { |
74 | u_int32_t se; |
75 | int32_t i, ix, sgn; |
76 | _Float128 a, b, temp, di, ret; |
77 | _Float128 z, w; |
78 | ieee854_long_double_shape_type u; |
79 | |
80 | |
81 | /* J(-n,x) = (-1)^n * J(n, x), J(n, -x) = (-1)^n * J(n, x) |
82 | * Thus, J(-n,x) = J(n,-x) |
83 | */ |
84 | |
85 | u.value = x; |
86 | se = u.parts32.w0; |
87 | ix = se & 0x7fffffff; |
88 | |
89 | /* if J(n,NaN) is NaN */ |
90 | if (ix >= 0x7fff0000) |
91 | { |
92 | if ((u.parts32.w0 & 0xffff) | u.parts32.w1 | u.parts32.w2 | u.parts32.w3) |
93 | return x + x; |
94 | } |
95 | |
96 | if (n < 0) |
97 | { |
98 | n = -n; |
99 | x = -x; |
100 | se ^= 0x80000000; |
101 | } |
102 | if (n == 0) |
103 | return (__ieee754_j0l (x)); |
104 | if (n == 1) |
105 | return (__ieee754_j1l (x)); |
106 | sgn = (n & 1) & (se >> 31); /* even n -- 0, odd n -- sign(x) */ |
107 | x = fabsl (x); |
108 | |
109 | { |
110 | SET_RESTORE_ROUNDL (FE_TONEAREST); |
111 | if (x == 0 || ix >= 0x7fff0000) /* if x is 0 or inf */ |
112 | return sgn == 1 ? -zero : zero; |
113 | else if ((_Float128) n <= x) |
114 | { |
115 | /* Safe to use J(n+1,x)=2n/x *J(n,x)-J(n-1,x) */ |
116 | if (ix >= 0x412D0000) |
117 | { /* x > 2**302 */ |
118 | |
119 | /* ??? Could use an expansion for large x here. */ |
120 | |
121 | /* (x >> n**2) |
122 | * Jn(x) = cos(x-(2n+1)*pi/4)*sqrt(2/x*pi) |
123 | * Yn(x) = sin(x-(2n+1)*pi/4)*sqrt(2/x*pi) |
124 | * Let s=sin(x), c=cos(x), |
125 | * xn=x-(2n+1)*pi/4, sqt2 = sqrt(2),then |
126 | * |
127 | * n sin(xn)*sqt2 cos(xn)*sqt2 |
128 | * ---------------------------------- |
129 | * 0 s-c c+s |
130 | * 1 -s-c -c+s |
131 | * 2 -s+c -c-s |
132 | * 3 s+c c-s |
133 | */ |
134 | _Float128 s; |
135 | _Float128 c; |
136 | __sincosl (x, &s, &c); |
137 | switch (n & 3) |
138 | { |
139 | case 0: |
140 | temp = c + s; |
141 | break; |
142 | case 1: |
143 | temp = -c + s; |
144 | break; |
145 | case 2: |
146 | temp = -c - s; |
147 | break; |
148 | case 3: |
149 | temp = c - s; |
150 | break; |
151 | } |
152 | b = invsqrtpi * temp / __ieee754_sqrtl (x); |
153 | } |
154 | else |
155 | { |
156 | a = __ieee754_j0l (x); |
157 | b = __ieee754_j1l (x); |
158 | for (i = 1; i < n; i++) |
159 | { |
160 | temp = b; |
161 | b = b * ((_Float128) (i + i) / x) - a; /* avoid underflow */ |
162 | a = temp; |
163 | } |
164 | } |
165 | } |
166 | else |
167 | { |
168 | if (ix < 0x3fc60000) |
169 | { /* x < 2**-57 */ |
170 | /* x is tiny, return the first Taylor expansion of J(n,x) |
171 | * J(n,x) = 1/n!*(x/2)^n - ... |
172 | */ |
173 | if (n >= 400) /* underflow, result < 10^-4952 */ |
174 | b = zero; |
175 | else |
176 | { |
177 | temp = x * 0.5; |
178 | b = temp; |
179 | for (a = one, i = 2; i <= n; i++) |
180 | { |
181 | a *= (_Float128) i; /* a = n! */ |
182 | b *= temp; /* b = (x/2)^n */ |
183 | } |
184 | b = b / a; |
185 | } |
186 | } |
187 | else |
188 | { |
189 | /* use backward recurrence */ |
190 | /* x x^2 x^2 |
191 | * J(n,x)/J(n-1,x) = ---- ------ ------ ..... |
192 | * 2n - 2(n+1) - 2(n+2) |
193 | * |
194 | * 1 1 1 |
195 | * (for large x) = ---- ------ ------ ..... |
196 | * 2n 2(n+1) 2(n+2) |
197 | * -- - ------ - ------ - |
198 | * x x x |
199 | * |
200 | * Let w = 2n/x and h=2/x, then the above quotient |
201 | * is equal to the continued fraction: |
202 | * 1 |
203 | * = ----------------------- |
204 | * 1 |
205 | * w - ----------------- |
206 | * 1 |
207 | * w+h - --------- |
208 | * w+2h - ... |
209 | * |
210 | * To determine how many terms needed, let |
211 | * Q(0) = w, Q(1) = w(w+h) - 1, |
212 | * Q(k) = (w+k*h)*Q(k-1) - Q(k-2), |
213 | * When Q(k) > 1e4 good for single |
214 | * When Q(k) > 1e9 good for double |
215 | * When Q(k) > 1e17 good for quadruple |
216 | */ |
217 | /* determine k */ |
218 | _Float128 t, v; |
219 | _Float128 q0, q1, h, tmp; |
220 | int32_t k, m; |
221 | w = (n + n) / (_Float128) x; |
222 | h = 2 / (_Float128) x; |
223 | q0 = w; |
224 | z = w + h; |
225 | q1 = w * z - 1; |
226 | k = 1; |
227 | while (q1 < L(1.0e17)) |
228 | { |
229 | k += 1; |
230 | z += h; |
231 | tmp = z * q1 - q0; |
232 | q0 = q1; |
233 | q1 = tmp; |
234 | } |
235 | m = n + n; |
236 | for (t = zero, i = 2 * (n + k); i >= m; i -= 2) |
237 | t = one / (i / x - t); |
238 | a = t; |
239 | b = one; |
240 | /* estimate log((2/x)^n*n!) = n*log(2/x)+n*ln(n) |
241 | * Hence, if n*(log(2n/x)) > ... |
242 | * single 8.8722839355e+01 |
243 | * double 7.09782712893383973096e+02 |
244 | * long double 1.1356523406294143949491931077970765006170e+04 |
245 | * then recurrent value may overflow and the result is |
246 | * likely underflow to zero |
247 | */ |
248 | tmp = n; |
249 | v = two / x; |
250 | tmp = tmp * __ieee754_logl (fabsl (v * tmp)); |
251 | |
252 | if (tmp < L(1.1356523406294143949491931077970765006170e+04)) |
253 | { |
254 | for (i = n - 1, di = (_Float128) (i + i); i > 0; i--) |
255 | { |
256 | temp = b; |
257 | b *= di; |
258 | b = b / x - a; |
259 | a = temp; |
260 | di -= two; |
261 | } |
262 | } |
263 | else |
264 | { |
265 | for (i = n - 1, di = (_Float128) (i + i); i > 0; i--) |
266 | { |
267 | temp = b; |
268 | b *= di; |
269 | b = b / x - a; |
270 | a = temp; |
271 | di -= two; |
272 | /* scale b to avoid spurious overflow */ |
273 | if (b > L(1e100)) |
274 | { |
275 | a /= b; |
276 | t /= b; |
277 | b = one; |
278 | } |
279 | } |
280 | } |
281 | /* j0() and j1() suffer enormous loss of precision at and |
282 | * near zero; however, we know that their zero points never |
283 | * coincide, so just choose the one further away from zero. |
284 | */ |
285 | z = __ieee754_j0l (x); |
286 | w = __ieee754_j1l (x); |
287 | if (fabsl (z) >= fabsl (w)) |
288 | b = (t * z / b); |
289 | else |
290 | b = (t * w / a); |
291 | } |
292 | } |
293 | if (sgn == 1) |
294 | ret = -b; |
295 | else |
296 | ret = b; |
297 | } |
298 | if (ret == 0) |
299 | { |
300 | ret = __copysignl (LDBL_MIN, ret) * LDBL_MIN; |
301 | __set_errno (ERANGE); |
302 | } |
303 | else |
304 | math_check_force_underflow (ret); |
305 | return ret; |
306 | } |
307 | strong_alias (__ieee754_jnl, __jnl_finite) |
308 | |
309 | _Float128 |
310 | __ieee754_ynl (int n, _Float128 x) |
311 | { |
312 | u_int32_t se; |
313 | int32_t i, ix; |
314 | int32_t sign; |
315 | _Float128 a, b, temp, ret; |
316 | ieee854_long_double_shape_type u; |
317 | |
318 | u.value = x; |
319 | se = u.parts32.w0; |
320 | ix = se & 0x7fffffff; |
321 | |
322 | /* if Y(n,NaN) is NaN */ |
323 | if (ix >= 0x7fff0000) |
324 | { |
325 | if ((u.parts32.w0 & 0xffff) | u.parts32.w1 | u.parts32.w2 | u.parts32.w3) |
326 | return x + x; |
327 | } |
328 | if (x <= 0) |
329 | { |
330 | if (x == 0) |
331 | return ((n < 0 && (n & 1) != 0) ? 1 : -1) / L(0.0); |
332 | if (se & 0x80000000) |
333 | return zero / (zero * x); |
334 | } |
335 | sign = 1; |
336 | if (n < 0) |
337 | { |
338 | n = -n; |
339 | sign = 1 - ((n & 1) << 1); |
340 | } |
341 | if (n == 0) |
342 | return (__ieee754_y0l (x)); |
343 | { |
344 | SET_RESTORE_ROUNDL (FE_TONEAREST); |
345 | if (n == 1) |
346 | { |
347 | ret = sign * __ieee754_y1l (x); |
348 | goto out; |
349 | } |
350 | if (ix >= 0x7fff0000) |
351 | return zero; |
352 | if (ix >= 0x412D0000) |
353 | { /* x > 2**302 */ |
354 | |
355 | /* ??? See comment above on the possible futility of this. */ |
356 | |
357 | /* (x >> n**2) |
358 | * Jn(x) = cos(x-(2n+1)*pi/4)*sqrt(2/x*pi) |
359 | * Yn(x) = sin(x-(2n+1)*pi/4)*sqrt(2/x*pi) |
360 | * Let s=sin(x), c=cos(x), |
361 | * xn=x-(2n+1)*pi/4, sqt2 = sqrt(2),then |
362 | * |
363 | * n sin(xn)*sqt2 cos(xn)*sqt2 |
364 | * ---------------------------------- |
365 | * 0 s-c c+s |
366 | * 1 -s-c -c+s |
367 | * 2 -s+c -c-s |
368 | * 3 s+c c-s |
369 | */ |
370 | _Float128 s; |
371 | _Float128 c; |
372 | __sincosl (x, &s, &c); |
373 | switch (n & 3) |
374 | { |
375 | case 0: |
376 | temp = s - c; |
377 | break; |
378 | case 1: |
379 | temp = -s - c; |
380 | break; |
381 | case 2: |
382 | temp = -s + c; |
383 | break; |
384 | case 3: |
385 | temp = s + c; |
386 | break; |
387 | } |
388 | b = invsqrtpi * temp / __ieee754_sqrtl (x); |
389 | } |
390 | else |
391 | { |
392 | a = __ieee754_y0l (x); |
393 | b = __ieee754_y1l (x); |
394 | /* quit if b is -inf */ |
395 | u.value = b; |
396 | se = u.parts32.w0 & 0xffff0000; |
397 | for (i = 1; i < n && se != 0xffff0000; i++) |
398 | { |
399 | temp = b; |
400 | b = ((_Float128) (i + i) / x) * b - a; |
401 | u.value = b; |
402 | se = u.parts32.w0 & 0xffff0000; |
403 | a = temp; |
404 | } |
405 | } |
406 | /* If B is +-Inf, set up errno accordingly. */ |
407 | if (! isfinite (b)) |
408 | __set_errno (ERANGE); |
409 | if (sign > 0) |
410 | ret = b; |
411 | else |
412 | ret = -b; |
413 | } |
414 | out: |
415 | if (isinf (ret)) |
416 | ret = __copysignl (LDBL_MAX, ret) * LDBL_MAX; |
417 | return ret; |
418 | } |
419 | strong_alias (__ieee754_ynl, __ynl_finite) |
420 | |