1 /*- 2 * SPDX-License-Identifier: BSD-2-Clause 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 /* 33 * ld128 version of s_expl.c. See ../ld80/s_expl.c for most comments. 34 */ 35 36 #include <float.h> 37 38 #include "math.h" 39 #include "math_private.h" 40 #include "k_expl.h" 41 42 /* XXX Prevent compilers from erroneously constant folding these: */ 43 static const volatile long double 44 huge = 0x1p10000L, 45 tiny = 0x1p-10000L; 46 47 static const long double 48 twom10000 = 0x1p-10000L; 49 50 static const long double 51 /* log(2**16384 - 0.5) rounded towards zero: */ 52 /* log(2**16384 - 0.5 + 1) rounded towards zero for expm1l() is the same: */ 53 o_threshold = 11356.523406294143949491931077970763428L, 54 /* log(2**(-16381-64-1)) rounded towards zero: */ 55 u_threshold = -11433.462743336297878837243843452621503L; 56 57 long double 58 expl(long double x) 59 { 60 union ieee_ext_u u; 61 long double hi, lo, t, twopk; 62 int k; 63 uint16_t hx, ix; 64 65 /* Filter out exceptional cases. */ 66 u.extu_ld = x; 67 hx = GET_EXPSIGN(&u); 68 ix = hx & 0x7fff; 69 if (ix >= BIAS + 13) { /* |x| >= 8192 or x is NaN */ 70 if (ix == BIAS + LDBL_MAX_EXP) { 71 if (hx & 0x8000) /* x is -Inf or -NaN */ 72 RETURNF(-1 / x); 73 RETURNF(x + x); /* x is +Inf or +NaN */ 74 } 75 if (x > o_threshold) 76 RETURNF(huge * huge); 77 if (x < u_threshold) 78 RETURNF(tiny * tiny); 79 } else if (ix < BIAS - 114) { /* |x| < 0x1p-114 */ 80 RETURNF(1 + x); /* 1 with inexact iff x != 0 */ 81 } 82 83 ENTERI(); 84 85 twopk = 1; 86 __k_expl(x, &hi, &lo, &k); 87 t = SUM2P(hi, lo); 88 89 /* Scale by 2**k. */ 90 /* 91 * XXX sparc64 multiplication was so slow that scalbnl() is faster, 92 * but performance on aarch64 and riscv hasn't yet been quantified. 93 */ 94 if (k >= LDBL_MIN_EXP) { 95 if (k == LDBL_MAX_EXP) 96 RETURNI(t * 2 * 0x1p16383L); 97 SET_LDBL_EXPSIGN(twopk, BIAS + k); 98 RETURNI(t * twopk); 99 } else { 100 SET_LDBL_EXPSIGN(twopk, BIAS + k + 10000); 101 RETURNI(t * twopk * twom10000); 102 } 103 } 104 105 /* 106 * Our T1 and T2 are chosen to be approximately the points where method 107 * A and method B have the same accuracy. Tang's T1 and T2 are the 108 * points where method A's accuracy changes by a full bit. For Tang, 109 * this drop in accuracy makes method A immediately less accurate than 110 * method B, but our larger INTERVALS makes method A 2 bits more 111 * accurate so it remains the most accurate method significantly 112 * closer to the origin despite losing the full bit in our extended 113 * range for it. 114 * 115 * Split the interval [T1, T2] into two intervals [T1, T3] and [T3, T2]. 116 * Setting T3 to 0 would require the |x| < 0x1p-113 condition to appear 117 * in both subintervals, so set T3 = 2**-5, which places the condition 118 * into the [T1, T3] interval. 119 * 120 * XXX we now do this more to (partially) balance the number of terms 121 * in the C and D polys than to avoid checking the condition in both 122 * intervals. 123 * 124 * XXX these micro-optimizations are excessive. 125 */ 126 static const double 127 T1 = -0.1659, /* ~-30.625/128 * log(2) */ 128 T2 = 0.1659, /* ~30.625/128 * log(2) */ 129 T3 = 0.03125; 130 131 /* 132 * Domain [-0.1659, 0.03125], range ~[2.9134e-44, 1.8404e-37]: 133 * |(exp(x)-1-x-x**2/2)/x - p(x)| < 2**-122.03 134 * 135 * XXX none of the long double C or D coeffs except C10 is correctly printed. 136 * If you re-print their values in %.35Le format, the result is always 137 * different. For example, the last 2 digits in C3 should be 59, not 67. 138 * 67 is apparently from rounding an extra-precision value to 36 decimal 139 * places. 140 */ 141 static const long double 142 C3 = 1.66666666666666666666666666666666667e-1L, 143 C4 = 4.16666666666666666666666666666666645e-2L, 144 C5 = 8.33333333333333333333333333333371638e-3L, 145 C6 = 1.38888888888888888888888888891188658e-3L, 146 C7 = 1.98412698412698412698412697235950394e-4L, 147 C8 = 2.48015873015873015873015112487849040e-5L, 148 C9 = 2.75573192239858906525606685484412005e-6L, 149 C10 = 2.75573192239858906612966093057020362e-7L, 150 C11 = 2.50521083854417203619031960151253944e-8L, 151 C12 = 2.08767569878679576457272282566520649e-9L, 152 C13 = 1.60590438367252471783548748824255707e-10L; 153 154 /* 155 * XXX this has 1 more coeff than needed. 156 * XXX can start the double coeffs but not the double mults at C10. 157 * With my coeffs (C10-C17 double; s = best_s): 158 * Domain [-0.1659, 0.03125], range ~[-1.1976e-37, 1.1976e-37]: 159 * |(exp(x)-1-x-x**2/2)/x - p(x)| ~< 2**-122.65 160 */ 161 static const double 162 C14 = 1.1470745580491932e-11, /* 0x1.93974a81dae30p-37 */ 163 C15 = 7.6471620181090468e-13, /* 0x1.ae7f3820adab1p-41 */ 164 C16 = 4.7793721460260450e-14, /* 0x1.ae7cd18a18eacp-45 */ 165 C17 = 2.8074757356658877e-15, /* 0x1.949992a1937d9p-49 */ 166 C18 = 1.4760610323699476e-16; /* 0x1.545b43aabfbcdp-53 */ 167 168 /* 169 * Domain [0.03125, 0.1659], range ~[-2.7676e-37, -1.0367e-38]: 170 * |(exp(x)-1-x-x**2/2)/x - p(x)| < 2**-121.44 171 */ 172 static const long double 173 D3 = 1.66666666666666666666666666666682245e-1L, 174 D4 = 4.16666666666666666666666666634228324e-2L, 175 D5 = 8.33333333333333333333333364022244481e-3L, 176 D6 = 1.38888888888888888888887138722762072e-3L, 177 D7 = 1.98412698412698412699085805424661471e-4L, 178 D8 = 2.48015873015873015687993712101479612e-5L, 179 D9 = 2.75573192239858944101036288338208042e-6L, 180 D10 = 2.75573192239853161148064676533754048e-7L, 181 D11 = 2.50521083855084570046480450935267433e-8L, 182 D12 = 2.08767569819738524488686318024854942e-9L, 183 D13 = 1.60590442297008495301927448122499313e-10L; 184 185 /* 186 * XXX this has 1 more coeff than needed. 187 * XXX can start the double coeffs but not the double mults at D11. 188 * With my coeffs (D11-D16 double): 189 * Domain [0.03125, 0.1659], range ~[-1.1980e-37, 1.1980e-37]: 190 * |(exp(x)-1-x-x**2/2)/x - p(x)| ~< 2**-122.65 191 */ 192 static const double 193 D14 = 1.1470726176204336e-11, /* 0x1.93971dc395d9ep-37 */ 194 D15 = 7.6478532249581686e-13, /* 0x1.ae892e3D16fcep-41 */ 195 D16 = 4.7628892832607741e-14, /* 0x1.ad00Dfe41feccp-45 */ 196 D17 = 3.0524857220358650e-15; /* 0x1.D7e8d886Df921p-49 */ 197 198 long double 199 expm1l(long double x) 200 { 201 union ieee_ext_u u, v; 202 long double hx2_hi, hx2_lo, q, r, r1, t, twomk, twopk, x_hi; 203 long double x_lo, x2; 204 double dr, dx, fn, r2; 205 int k, n, n2; 206 uint16_t hx, ix; 207 208 /* Filter out exceptional cases. */ 209 u.extu_ld = x; 210 hx = GET_EXPSIGN(&u); 211 ix = hx & 0x7fff; 212 if (ix >= BIAS + 7) { /* |x| >= 128 or x is NaN */ 213 if (ix == BIAS + LDBL_MAX_EXP) { 214 if (hx & 0x8000) /* x is -Inf or -NaN */ 215 RETURNF(-1 / x - 1); 216 RETURNF(x + x); /* x is +Inf or +NaN */ 217 } 218 if (x > o_threshold) 219 RETURNF(huge * huge); 220 /* 221 * expm1l() never underflows, but it must avoid 222 * unrepresentable large negative exponents. We used a 223 * much smaller threshold for large |x| above than in 224 * expl() so as to handle not so large negative exponents 225 * in the same way as large ones here. 226 */ 227 if (hx & 0x8000) /* x <= -128 */ 228 RETURNF(tiny - 1); /* good for x < -114ln2 - eps */ 229 } 230 231 ENTERI(); 232 233 if (T1 < x && x < T2) { 234 x2 = x * x; 235 dx = x; 236 237 if (x < T3) { 238 if (ix < BIAS - 113) { /* |x| < 0x1p-113 */ 239 /* x (rounded) with inexact if x != 0: */ 240 RETURNI(x == 0 ? x : 241 (0x1p200 * x + fabsl(x)) * 0x1p-200); 242 } 243 q = x * x2 * C3 + x2 * x2 * (C4 + x * (C5 + x * (C6 + 244 x * (C7 + x * (C8 + x * (C9 + x * (C10 + 245 x * (C11 + x * (C12 + x * (C13 + 246 dx * (C14 + dx * (C15 + dx * (C16 + 247 dx * (C17 + dx * C18)))))))))))))); 248 } else { 249 q = x * x2 * D3 + x2 * x2 * (D4 + x * (D5 + x * (D6 + 250 x * (D7 + x * (D8 + x * (D9 + x * (D10 + 251 x * (D11 + x * (D12 + x * (D13 + 252 dx * (D14 + dx * (D15 + dx * (D16 + 253 dx * D17))))))))))))); 254 } 255 256 x_hi = (float)x; 257 x_lo = x - x_hi; 258 hx2_hi = x_hi * x_hi / 2; 259 hx2_lo = x_lo * (x + x_hi) / 2; 260 if (ix >= BIAS - 7) 261 RETURNI((hx2_hi + x_hi) + (hx2_lo + x_lo + q)); 262 else 263 RETURNI(x + (hx2_lo + q + hx2_hi)); 264 } 265 266 /* Reduce x to (k*ln2 + endpoint[n2] + r1 + r2). */ 267 fn = rnint((double)x * INV_L); 268 n = irint(fn); 269 n2 = (unsigned)n % INTERVALS; 270 k = n >> LOG2_INTERVALS; 271 r1 = x - fn * L1; 272 r2 = fn * -L2; 273 r = r1 + r2; 274 275 /* Prepare scale factor. */ 276 v.extu_ld = 1; 277 SET_EXPSIGN(&v, BIAS + k); 278 twopk = v.extu_ld; 279 280 /* 281 * Evaluate lower terms of 282 * expl(endpoint[n2] + r1 + r2) = tbl[n2] * expl(r1 + r2). 283 */ 284 dr = r; 285 q = r2 + r * r * (A2 + r * (A3 + r * (A4 + r * (A5 + r * (A6 + 286 dr * (A7 + dr * (A8 + dr * (A9 + dr * A10)))))))); 287 288 t = tbl[n2].lo + tbl[n2].hi; 289 290 if (k == 0) { 291 t = SUM2P(tbl[n2].hi - 1, tbl[n2].lo * (r1 + 1) + t * q + 292 tbl[n2].hi * r1); 293 RETURNI(t); 294 } 295 if (k == -1) { 296 t = SUM2P(tbl[n2].hi - 2, tbl[n2].lo * (r1 + 1) + t * q + 297 tbl[n2].hi * r1); 298 RETURNI(t / 2); 299 } 300 if (k < -7) { 301 t = SUM2P(tbl[n2].hi, tbl[n2].lo + t * (q + r1)); 302 RETURNI(t * twopk - 1); 303 } 304 if (k > 2 * LDBL_MANT_DIG - 1) { 305 t = SUM2P(tbl[n2].hi, tbl[n2].lo + t * (q + r1)); 306 if (k == LDBL_MAX_EXP) 307 RETURNI(t * 2 * 0x1p16383L - 1); 308 RETURNI(t * twopk - 1); 309 } 310 311 SET_EXPSIGN(&v, BIAS - k); 312 twomk = v.extu_ld; 313 314 if (k > LDBL_MANT_DIG - 1) 315 t = SUM2P(tbl[n2].hi, tbl[n2].lo - twomk + t * (q + r1)); 316 else 317 t = SUM2P(tbl[n2].hi - twomk, tbl[n2].lo + t * (q + r1)); 318 RETURNI(t * twopk); 319 } 320
Indexes created Wed Sep 24 05:09:52 GMT 2025