1 | /* |
2 | * IBM Accurate Mathematical Library |
3 | * written by International Business Machines Corp. |
4 | * Copyright (C) 2001-2018 Free Software Foundation, Inc. |
5 | * |
6 | * This program is free software; you can redistribute it and/or modify |
7 | * it under the terms of the GNU Lesser General Public License as published by |
8 | * the Free Software Foundation; either version 2.1 of the License, or |
9 | * (at your option) any later version. |
10 | * |
11 | * This program is distributed in the hope that it will be useful, |
12 | * but WITHOUT ANY WARRANTY; without even the implied warranty of |
13 | * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the |
14 | * GNU Lesser General Public License for more details. |
15 | * |
16 | * You should have received a copy of the GNU Lesser General Public License |
17 | * along with this program; if not, see <http://www.gnu.org/licenses/>. |
18 | */ |
19 | /****************************************************************************/ |
20 | /* */ |
21 | /* MODULE_NAME:usncs.c */ |
22 | /* */ |
23 | /* FUNCTIONS: usin */ |
24 | /* ucos */ |
25 | /* FILES NEEDED: dla.h endian.h mpa.h mydefs.h usncs.h */ |
26 | /* branred.c sincos.tbl */ |
27 | /* */ |
28 | /* An ultimate sin and cos routine. Given an IEEE double machine number x */ |
29 | /* it computes sin(x) or cos(x) with ~0.55 ULP. */ |
30 | /* Assumption: Machine arithmetic operations are performed in */ |
31 | /* round to nearest mode of IEEE 754 standard. */ |
32 | /* */ |
33 | /****************************************************************************/ |
34 | |
35 | |
36 | #include <errno.h> |
37 | #include <float.h> |
38 | #include "endian.h" |
39 | #include "mydefs.h" |
40 | #include "usncs.h" |
41 | #include "MathLib.h" |
42 | #include <math.h> |
43 | #include <math_private.h> |
44 | #include <math-underflow.h> |
45 | #include <libm-alias-double.h> |
46 | #include <fenv.h> |
47 | |
48 | /* Helper macros to compute sin of the input values. */ |
49 | #define POLYNOMIAL2(xx) ((((s5 * (xx) + s4) * (xx) + s3) * (xx) + s2) * (xx)) |
50 | |
51 | #define POLYNOMIAL(xx) (POLYNOMIAL2 (xx) + s1) |
52 | |
53 | /* The computed polynomial is a variation of the Taylor series expansion for |
54 | sin(a): |
55 | |
56 | a - a^3/3! + a^5/5! - a^7/7! + a^9/9! + (1 - a^2) * da / 2 |
57 | |
58 | The constants s1, s2, s3, etc. are pre-computed values of 1/3!, 1/5! and so |
59 | on. The result is returned to LHS. */ |
60 | #define TAYLOR_SIN(xx, a, da) \ |
61 | ({ \ |
62 | double t = ((POLYNOMIAL (xx) * (a) - 0.5 * (da)) * (xx) + (da)); \ |
63 | double res = (a) + t; \ |
64 | res; \ |
65 | }) |
66 | |
67 | #define SINCOS_TABLE_LOOKUP(u, sn, ssn, cs, ccs) \ |
68 | ({ \ |
69 | int4 k = u.i[LOW_HALF] << 2; \ |
70 | sn = __sincostab.x[k]; \ |
71 | ssn = __sincostab.x[k + 1]; \ |
72 | cs = __sincostab.x[k + 2]; \ |
73 | ccs = __sincostab.x[k + 3]; \ |
74 | }) |
75 | |
76 | #ifndef SECTION |
77 | # define SECTION |
78 | #endif |
79 | |
80 | extern const union |
81 | { |
82 | int4 i[880]; |
83 | double x[440]; |
84 | } __sincostab attribute_hidden; |
85 | |
86 | static const double |
87 | sn3 = -1.66666666666664880952546298448555E-01, |
88 | sn5 = 8.33333214285722277379541354343671E-03, |
89 | cs2 = 4.99999999999999999999950396842453E-01, |
90 | cs4 = -4.16666666666664434524222570944589E-02, |
91 | cs6 = 1.38888874007937613028114285595617E-03; |
92 | |
93 | int __branred (double x, double *a, double *aa); |
94 | |
95 | /* Given a number partitioned into X and DX, this function computes the cosine |
96 | of the number by combining the sin and cos of X (as computed by a variation |
97 | of the Taylor series) with the values looked up from the sin/cos table to |
98 | get the result. */ |
99 | static inline double |
100 | __always_inline |
101 | do_cos (double x, double dx) |
102 | { |
103 | mynumber u; |
104 | |
105 | if (x < 0) |
106 | dx = -dx; |
107 | |
108 | u.x = big + fabs (x); |
109 | x = fabs (x) - (u.x - big) + dx; |
110 | |
111 | double xx, s, sn, ssn, c, cs, ccs, cor; |
112 | xx = x * x; |
113 | s = x + x * xx * (sn3 + xx * sn5); |
114 | c = xx * (cs2 + xx * (cs4 + xx * cs6)); |
115 | SINCOS_TABLE_LOOKUP (u, sn, ssn, cs, ccs); |
116 | cor = (ccs - s * ssn - cs * c) - sn * s; |
117 | return cs + cor; |
118 | } |
119 | |
120 | /* Given a number partitioned into X and DX, this function computes the sine of |
121 | the number by combining the sin and cos of X (as computed by a variation of |
122 | the Taylor series) with the values looked up from the sin/cos table to get |
123 | the result. */ |
124 | static inline double |
125 | __always_inline |
126 | do_sin (double x, double dx) |
127 | { |
128 | double xold = x; |
129 | /* Max ULP is 0.501 if |x| < 0.126, otherwise ULP is 0.518. */ |
130 | if (fabs (x) < 0.126) |
131 | return TAYLOR_SIN (x * x, x, dx); |
132 | |
133 | mynumber u; |
134 | |
135 | if (x <= 0) |
136 | dx = -dx; |
137 | u.x = big + fabs (x); |
138 | x = fabs (x) - (u.x - big); |
139 | |
140 | double xx, s, sn, ssn, c, cs, ccs, cor; |
141 | xx = x * x; |
142 | s = x + (dx + x * xx * (sn3 + xx * sn5)); |
143 | c = x * dx + xx * (cs2 + xx * (cs4 + xx * cs6)); |
144 | SINCOS_TABLE_LOOKUP (u, sn, ssn, cs, ccs); |
145 | cor = (ssn + s * ccs - sn * c) + cs * s; |
146 | return __copysign (sn + cor, xold); |
147 | } |
148 | |
149 | /* Reduce range of x to within PI/2 with abs (x) < 105414350. The high part |
150 | is written to *a, the low part to *da. Range reduction is accurate to 136 |
151 | bits so that when x is large and *a very close to zero, all 53 bits of *a |
152 | are correct. */ |
153 | static inline int4 |
154 | __always_inline |
155 | reduce_sincos (double x, double *a, double *da) |
156 | { |
157 | mynumber v; |
158 | |
159 | double t = (x * hpinv + toint); |
160 | double xn = t - toint; |
161 | v.x = t; |
162 | double y = (x - xn * mp1) - xn * mp2; |
163 | int4 n = v.i[LOW_HALF] & 3; |
164 | |
165 | double b, db, t1, t2; |
166 | t1 = xn * pp3; |
167 | t2 = y - t1; |
168 | db = (y - t2) - t1; |
169 | |
170 | t1 = xn * pp4; |
171 | b = t2 - t1; |
172 | db += (t2 - b) - t1; |
173 | |
174 | *a = b; |
175 | *da = db; |
176 | return n; |
177 | } |
178 | |
179 | /* Compute sin or cos (A + DA) for the given quadrant N. */ |
180 | static double |
181 | __always_inline |
182 | do_sincos (double a, double da, int4 n) |
183 | { |
184 | double retval; |
185 | |
186 | if (n & 1) |
187 | /* Max ULP is 0.513. */ |
188 | retval = do_cos (a, da); |
189 | else |
190 | /* Max ULP is 0.501 if xx < 0.01588, otherwise ULP is 0.518. */ |
191 | retval = do_sin (a, da); |
192 | |
193 | return (n & 2) ? -retval : retval; |
194 | } |
195 | |
196 | |
197 | /*******************************************************************/ |
198 | /* An ultimate sin routine. Given an IEEE double machine number x */ |
199 | /* it computes the correctly rounded (to nearest) value of sin(x) */ |
200 | /*******************************************************************/ |
201 | #ifndef IN_SINCOS |
202 | double |
203 | SECTION |
204 | __sin (double x) |
205 | { |
206 | double t, a, da; |
207 | mynumber u; |
208 | int4 k, m, n; |
209 | double retval = 0; |
210 | |
211 | SET_RESTORE_ROUND_53BIT (FE_TONEAREST); |
212 | |
213 | u.x = x; |
214 | m = u.i[HIGH_HALF]; |
215 | k = 0x7fffffff & m; /* no sign */ |
216 | if (k < 0x3e500000) /* if x->0 =>sin(x)=x */ |
217 | { |
218 | math_check_force_underflow (x); |
219 | retval = x; |
220 | } |
221 | /*--------------------------- 2^-26<|x|< 0.855469---------------------- */ |
222 | else if (k < 0x3feb6000) |
223 | { |
224 | /* Max ULP is 0.548. */ |
225 | retval = do_sin (x, 0); |
226 | } /* else if (k < 0x3feb6000) */ |
227 | |
228 | /*----------------------- 0.855469 <|x|<2.426265 ----------------------*/ |
229 | else if (k < 0x400368fd) |
230 | { |
231 | t = hp0 - fabs (x); |
232 | /* Max ULP is 0.51. */ |
233 | retval = __copysign (do_cos (t, hp1), x); |
234 | } /* else if (k < 0x400368fd) */ |
235 | |
236 | /*-------------------------- 2.426265<|x|< 105414350 ----------------------*/ |
237 | else if (k < 0x419921FB) |
238 | { |
239 | n = reduce_sincos (x, &a, &da); |
240 | retval = do_sincos (a, da, n); |
241 | } /* else if (k < 0x419921FB ) */ |
242 | |
243 | /* --------------------105414350 <|x| <2^1024------------------------------*/ |
244 | else if (k < 0x7ff00000) |
245 | { |
246 | n = __branred (x, &a, &da); |
247 | retval = do_sincos (a, da, n); |
248 | } |
249 | /*--------------------- |x| > 2^1024 ----------------------------------*/ |
250 | else |
251 | { |
252 | if (k == 0x7ff00000 && u.i[LOW_HALF] == 0) |
253 | __set_errno (EDOM); |
254 | retval = x / x; |
255 | } |
256 | |
257 | return retval; |
258 | } |
259 | |
260 | |
261 | /*******************************************************************/ |
262 | /* An ultimate cos routine. Given an IEEE double machine number x */ |
263 | /* it computes the correctly rounded (to nearest) value of cos(x) */ |
264 | /*******************************************************************/ |
265 | |
266 | double |
267 | SECTION |
268 | __cos (double x) |
269 | { |
270 | double y, a, da; |
271 | mynumber u; |
272 | int4 k, m, n; |
273 | |
274 | double retval = 0; |
275 | |
276 | SET_RESTORE_ROUND_53BIT (FE_TONEAREST); |
277 | |
278 | u.x = x; |
279 | m = u.i[HIGH_HALF]; |
280 | k = 0x7fffffff & m; |
281 | |
282 | /* |x|<2^-27 => cos(x)=1 */ |
283 | if (k < 0x3e400000) |
284 | retval = 1.0; |
285 | |
286 | else if (k < 0x3feb6000) |
287 | { /* 2^-27 < |x| < 0.855469 */ |
288 | /* Max ULP is 0.51. */ |
289 | retval = do_cos (x, 0); |
290 | } /* else if (k < 0x3feb6000) */ |
291 | |
292 | else if (k < 0x400368fd) |
293 | { /* 0.855469 <|x|<2.426265 */ ; |
294 | y = hp0 - fabs (x); |
295 | a = y + hp1; |
296 | da = (y - a) + hp1; |
297 | /* Max ULP is 0.501 if xx < 0.01588 or 0.518 otherwise. |
298 | Range reduction uses 106 bits here which is sufficient. */ |
299 | retval = do_sin (a, da); |
300 | } /* else if (k < 0x400368fd) */ |
301 | |
302 | else if (k < 0x419921FB) |
303 | { /* 2.426265<|x|< 105414350 */ |
304 | n = reduce_sincos (x, &a, &da); |
305 | retval = do_sincos (a, da, n + 1); |
306 | } /* else if (k < 0x419921FB ) */ |
307 | |
308 | /* 105414350 <|x| <2^1024 */ |
309 | else if (k < 0x7ff00000) |
310 | { |
311 | n = __branred (x, &a, &da); |
312 | retval = do_sincos (a, da, n + 1); |
313 | } |
314 | |
315 | else |
316 | { |
317 | if (k == 0x7ff00000 && u.i[LOW_HALF] == 0) |
318 | __set_errno (EDOM); |
319 | retval = x / x; /* |x| > 2^1024 */ |
320 | } |
321 | |
322 | return retval; |
323 | } |
324 | |
325 | #ifndef __cos |
326 | libm_alias_double (__cos, cos) |
327 | #endif |
328 | #ifndef __sin |
329 | libm_alias_double (__sin, sin) |
330 | #endif |
331 | |
332 | #endif |
333 | |