1 /*-
2 * SPDX-License-Identifier: BSD-2-Clause-FreeBSD
3 *
4 * Copyright (c) 2009-2013 Steven G. Kargl
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 unmodified, this list of conditions, and the following
12 * disclaimer.
13 * 2. Redistributions in binary form must reproduce the above copyright
14 * notice, this list of conditions and the following disclaimer in the
15 * documentation and/or other materials provided with the distribution.
16 *
17 * THIS SOFTWARE IS PROVIDED BY THE AUTHOR ``AS IS'' AND ANY EXPRESS OR
18 * IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES
19 * OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED.
20 * IN NO EVENT SHALL THE AUTHOR BE LIABLE FOR ANY DIRECT, INDIRECT,
21 * INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT
22 * NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
23 * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
24 * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
25 * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF
26 * THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
27 *
28 * Optimized by Bruce D. Evans.
29 */
30
31 #include <sys/cdefs.h>
32 __FBSDID("$FreeBSD: head/lib/msun/ld128/s_expl.c 336545 2018-07-20 12:42:24Z bde $");
33
34 /*
35 * ld128 version of s_expl.c. See ../ld80/s_expl.c for most comments.
36 */
37
38 #include <float.h>
39
40 #include "fpmath.h"
41 #include "math.h"
42 #include "math_private.h"
43 #include "k_expl.h"
44
45 /* XXX Prevent compilers from erroneously constant folding these: */
46 static const volatile long double
47 huge = 0x1p10000L,
48 tiny = 0x1p-10000L;
49
50 static const long double
51 twom10000 = 0x1p-10000L;
52
53 static const long double
54 /* log(2**16384 - 0.5) rounded towards zero: */
55 /* log(2**16384 - 0.5 + 1) rounded towards zero for expm1l() is the same: */
56 o_threshold = 11356.523406294143949491931077970763428L,
57 /* log(2**(-16381-64-1)) rounded towards zero: */
58 u_threshold = -11433.462743336297878837243843452621503L;
59
60 long double
expl(long double x)61 expl(long double x)
62 {
63 union IEEEl2bits u;
64 long double hi, lo, t, twopk;
65 int k;
66 uint16_t hx, ix;
67
68 DOPRINT_START(&x);
69
70 /* Filter out exceptional cases. */
71 u.e = x;
72 hx = u.xbits.expsign;
73 ix = hx & 0x7fff;
74 if (ix >= BIAS + 13) { /* |x| >= 8192 or x is NaN */
75 if (ix == BIAS + LDBL_MAX_EXP) {
76 if (hx & 0x8000) /* x is -Inf or -NaN */
77 RETURNP(-1 / x);
78 RETURNP(x + x); /* x is +Inf or +NaN */
79 }
80 if (x > o_threshold)
81 RETURNP(huge * huge);
82 if (x < u_threshold)
83 RETURNP(tiny * tiny);
84 } else if (ix < BIAS - 114) { /* |x| < 0x1p-114 */
85 RETURN2P(1, x); /* 1 with inexact iff x != 0 */
86 }
87
88 ENTERI();
89
90 twopk = 1;
91 __k_expl(x, &hi, &lo, &k);
92 t = SUM2P(hi, lo);
93
94 /* Scale by 2**k. */
95 /* XXX sparc64 multiplication is so slow that scalbnl() is faster. */
96 if (k >= LDBL_MIN_EXP) {
97 if (k == LDBL_MAX_EXP)
98 RETURNI(t * 2 * 0x1p16383L);
99 SET_LDBL_EXPSIGN(twopk, BIAS + k);
100 RETURNI(t * twopk);
101 } else {
102 SET_LDBL_EXPSIGN(twopk, BIAS + k + 10000);
103 RETURNI(t * twopk * twom10000);
104 }
105 }
106
107 /*
108 * Our T1 and T2 are chosen to be approximately the points where method
109 * A and method B have the same accuracy. Tang's T1 and T2 are the
110 * points where method A's accuracy changes by a full bit. For Tang,
111 * this drop in accuracy makes method A immediately less accurate than
112 * method B, but our larger INTERVALS makes method A 2 bits more
113 * accurate so it remains the most accurate method significantly
114 * closer to the origin despite losing the full bit in our extended
115 * range for it.
116 *
117 * Split the interval [T1, T2] into two intervals [T1, T3] and [T3, T2].
118 * Setting T3 to 0 would require the |x| < 0x1p-113 condition to appear
119 * in both subintervals, so set T3 = 2**-5, which places the condition
120 * into the [T1, T3] interval.
121 *
122 * XXX we now do this more to (partially) balance the number of terms
123 * in the C and D polys than to avoid checking the condition in both
124 * intervals.
125 *
126 * XXX these micro-optimizations are excessive.
127 */
128 static const double
129 T1 = -0.1659, /* ~-30.625/128 * log(2) */
130 T2 = 0.1659, /* ~30.625/128 * log(2) */
131 T3 = 0.03125;
132
133 /*
134 * Domain [-0.1659, 0.03125], range ~[2.9134e-44, 1.8404e-37]:
135 * |(exp(x)-1-x-x**2/2)/x - p(x)| < 2**-122.03
136 *
137 * XXX none of the long double C or D coeffs except C10 is correctly printed.
138 * If you re-print their values in %.35Le format, the result is always
139 * different. For example, the last 2 digits in C3 should be 59, not 67.
140 * 67 is apparently from rounding an extra-precision value to 36 decimal
141 * places.
142 */
143 static const long double
144 C3 = 1.66666666666666666666666666666666667e-1L,
145 C4 = 4.16666666666666666666666666666666645e-2L,
146 C5 = 8.33333333333333333333333333333371638e-3L,
147 C6 = 1.38888888888888888888888888891188658e-3L,
148 C7 = 1.98412698412698412698412697235950394e-4L,
149 C8 = 2.48015873015873015873015112487849040e-5L,
150 C9 = 2.75573192239858906525606685484412005e-6L,
151 C10 = 2.75573192239858906612966093057020362e-7L,
152 C11 = 2.50521083854417203619031960151253944e-8L,
153 C12 = 2.08767569878679576457272282566520649e-9L,
154 C13 = 1.60590438367252471783548748824255707e-10L;
155
156 /*
157 * XXX this has 1 more coeff than needed.
158 * XXX can start the double coeffs but not the double mults at C10.
159 * With my coeffs (C10-C17 double; s = best_s):
160 * Domain [-0.1659, 0.03125], range ~[-1.1976e-37, 1.1976e-37]:
161 * |(exp(x)-1-x-x**2/2)/x - p(x)| ~< 2**-122.65
162 */
163 static const double
164 C14 = 1.1470745580491932e-11, /* 0x1.93974a81dae30p-37 */
165 C15 = 7.6471620181090468e-13, /* 0x1.ae7f3820adab1p-41 */
166 C16 = 4.7793721460260450e-14, /* 0x1.ae7cd18a18eacp-45 */
167 C17 = 2.8074757356658877e-15, /* 0x1.949992a1937d9p-49 */
168 C18 = 1.4760610323699476e-16; /* 0x1.545b43aabfbcdp-53 */
169
170 /*
171 * Domain [0.03125, 0.1659], range ~[-2.7676e-37, -1.0367e-38]:
172 * |(exp(x)-1-x-x**2/2)/x - p(x)| < 2**-121.44
173 */
174 static const long double
175 D3 = 1.66666666666666666666666666666682245e-1L,
176 D4 = 4.16666666666666666666666666634228324e-2L,
177 D5 = 8.33333333333333333333333364022244481e-3L,
178 D6 = 1.38888888888888888888887138722762072e-3L,
179 D7 = 1.98412698412698412699085805424661471e-4L,
180 D8 = 2.48015873015873015687993712101479612e-5L,
181 D9 = 2.75573192239858944101036288338208042e-6L,
182 D10 = 2.75573192239853161148064676533754048e-7L,
183 D11 = 2.50521083855084570046480450935267433e-8L,
184 D12 = 2.08767569819738524488686318024854942e-9L,
185 D13 = 1.60590442297008495301927448122499313e-10L;
186
187 /*
188 * XXX this has 1 more coeff than needed.
189 * XXX can start the double coeffs but not the double mults at D11.
190 * With my coeffs (D11-D16 double):
191 * Domain [0.03125, 0.1659], range ~[-1.1980e-37, 1.1980e-37]:
192 * |(exp(x)-1-x-x**2/2)/x - p(x)| ~< 2**-122.65
193 */
194 static const double
195 D14 = 1.1470726176204336e-11, /* 0x1.93971dc395d9ep-37 */
196 D15 = 7.6478532249581686e-13, /* 0x1.ae892e3D16fcep-41 */
197 D16 = 4.7628892832607741e-14, /* 0x1.ad00Dfe41feccp-45 */
198 D17 = 3.0524857220358650e-15; /* 0x1.D7e8d886Df921p-49 */
199
200 long double
expm1l(long double x)201 expm1l(long double x)
202 {
203 union IEEEl2bits u, v;
204 long double hx2_hi, hx2_lo, q, r, r1, t, twomk, twopk, x_hi;
205 long double x_lo, x2;
206 double dr, dx, fn, r2;
207 int k, n, n2;
208 uint16_t hx, ix;
209
210 DOPRINT_START(&x);
211
212 /* Filter out exceptional cases. */
213 u.e = x;
214 hx = u.xbits.expsign;
215 ix = hx & 0x7fff;
216 if (ix >= BIAS + 7) { /* |x| >= 128 or x is NaN */
217 if (ix == BIAS + LDBL_MAX_EXP) {
218 if (hx & 0x8000) /* x is -Inf or -NaN */
219 RETURNP(-1 / x - 1);
220 RETURNP(x + x); /* x is +Inf or +NaN */
221 }
222 if (x > o_threshold)
223 RETURNP(huge * huge);
224 /*
225 * expm1l() never underflows, but it must avoid
226 * unrepresentable large negative exponents. We used a
227 * much smaller threshold for large |x| above than in
228 * expl() so as to handle not so large negative exponents
229 * in the same way as large ones here.
230 */
231 if (hx & 0x8000) /* x <= -128 */
232 RETURN2P(tiny, -1); /* good for x < -114ln2 - eps */
233 }
234
235 ENTERI();
236
237 if (T1 < x && x < T2) {
238 x2 = x * x;
239 dx = x;
240
241 if (x < T3) {
242 if (ix < BIAS - 113) { /* |x| < 0x1p-113 */
243 /* x (rounded) with inexact if x != 0: */
244 RETURNPI(x == 0 ? x :
245 (0x1p200 * x + fabsl(x)) * 0x1p-200);
246 }
247 q = x * x2 * C3 + x2 * x2 * (C4 + x * (C5 + x * (C6 +
248 x * (C7 + x * (C8 + x * (C9 + x * (C10 +
249 x * (C11 + x * (C12 + x * (C13 +
250 dx * (C14 + dx * (C15 + dx * (C16 +
251 dx * (C17 + dx * C18))))))))))))));
252 } else {
253 q = x * x2 * D3 + x2 * x2 * (D4 + x * (D5 + x * (D6 +
254 x * (D7 + x * (D8 + x * (D9 + x * (D10 +
255 x * (D11 + x * (D12 + x * (D13 +
256 dx * (D14 + dx * (D15 + dx * (D16 +
257 dx * D17)))))))))))));
258 }
259
260 x_hi = (float)x;
261 x_lo = x - x_hi;
262 hx2_hi = x_hi * x_hi / 2;
263 hx2_lo = x_lo * (x + x_hi) / 2;
264 if (ix >= BIAS - 7)
265 RETURN2PI(hx2_hi + x_hi, hx2_lo + x_lo + q);
266 else
267 RETURN2PI(x, hx2_lo + q + hx2_hi);
268 }
269
270 /* Reduce x to (k*ln2 + endpoint[n2] + r1 + r2). */
271 fn = rnint((double)x * INV_L);
272 n = irint(fn);
273 n2 = (unsigned)n % INTERVALS;
274 k = n >> LOG2_INTERVALS;
275 r1 = x - fn * L1;
276 r2 = fn * -L2;
277 r = r1 + r2;
278
279 /* Prepare scale factor. */
280 v.e = 1;
281 v.xbits.expsign = BIAS + k;
282 twopk = v.e;
283
284 /*
285 * Evaluate lower terms of
286 * expl(endpoint[n2] + r1 + r2) = tbl[n2] * expl(r1 + r2).
287 */
288 dr = r;
289 q = r2 + r * r * (A2 + r * (A3 + r * (A4 + r * (A5 + r * (A6 +
290 dr * (A7 + dr * (A8 + dr * (A9 + dr * A10))))))));
291
292 t = tbl[n2].lo + tbl[n2].hi;
293
294 if (k == 0) {
295 t = SUM2P(tbl[n2].hi - 1, tbl[n2].lo * (r1 + 1) + t * q +
296 tbl[n2].hi * r1);
297 RETURNI(t);
298 }
299 if (k == -1) {
300 t = SUM2P(tbl[n2].hi - 2, tbl[n2].lo * (r1 + 1) + t * q +
301 tbl[n2].hi * r1);
302 RETURNI(t / 2);
303 }
304 if (k < -7) {
305 t = SUM2P(tbl[n2].hi, tbl[n2].lo + t * (q + r1));
306 RETURNI(t * twopk - 1);
307 }
308 if (k > 2 * LDBL_MANT_DIG - 1) {
309 t = SUM2P(tbl[n2].hi, tbl[n2].lo + t * (q + r1));
310 if (k == LDBL_MAX_EXP)
311 RETURNI(t * 2 * 0x1p16383L - 1);
312 RETURNI(t * twopk - 1);
313 }
314
315 v.xbits.expsign = BIAS - k;
316 twomk = v.e;
317
318 if (k > LDBL_MANT_DIG - 1)
319 t = SUM2P(tbl[n2].hi, tbl[n2].lo - twomk + t * (q + r1));
320 else
321 t = SUM2P(tbl[n2].hi - twomk, tbl[n2].lo + t * (q + r1));
322 RETURNI(t * twopk);
323 }
324