1 /* Quad-precision floating point e^x.
2 Copyright (C) 1999-2017 Free Software Foundation, Inc.
3 This file is part of the GNU C Library.
4 Contributed by Jakub Jelinek <jj@ultra.linux.cz>
5 Partly based on double-precision code
6 by Geoffrey Keating <geoffk@ozemail.com.au>
8 The GNU C Library is free software; you can redistribute it and/or
9 modify it under the terms of the GNU Lesser General Public
10 License as published by the Free Software Foundation; either
11 version 2.1 of the License, or (at your option) any later version.
13 The GNU C Library is distributed in the hope that it will be useful,
14 but WITHOUT ANY WARRANTY; without even the implied warranty of
15 MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
16 Lesser General Public License for more details.
18 You should have received a copy of the GNU Lesser General Public
19 License along with the GNU C Library; if not, see
20 <http://www.gnu.org/licenses/>. */
22 /* The basic design here is from
23 Abraham Ziv, "Fast Evaluation of Elementary Mathematical Functions with
24 Correctly Rounded Last Bit", ACM Trans. Math. Soft., 17 (3), September 1991,
27 We work with number pairs where the first number is the high part and
28 the second one is the low part. Arithmetic with the high part numbers must
29 be exact, without any roundoff errors.
31 The input value, X, is written as
32 X = n * ln(2)_0 + arg1[t1]_0 + arg2[t2]_0 + x
33 - n * ln(2)_1 + arg1[t1]_1 + arg2[t2]_1 + xl
36 - n is an integer, 16384 >= n >= -16495;
37 - ln(2)_0 is the first 93 bits of ln(2), and |ln(2)_0-ln(2)-ln(2)_1| < 2^-205
38 - t1 is an integer, 89 >= t1 >= -89
39 - t2 is an integer, 65 >= t2 >= -65
40 - |arg1[t1]-t1/256.0| < 2^-53
41 - |arg2[t2]-t2/32768.0| < 2^-53
42 - x + xl is whatever is left, |x + xl| < 2^-16 + 2^-53
44 Then e^x is approximated as
46 e^x = 2^n_1 ( 2^n_0 e^(arg1[t1]_0 + arg1[t1]_1) e^(arg2[t2]_0 + arg2[t2]_1)
47 + 2^n_0 e^(arg1[t1]_0 + arg1[t1]_1) e^(arg2[t2]_0 + arg2[t2]_1)
48 * p (x + xl + n * ln(2)_1))
50 - p(x) is a polynomial approximating e(x)-1
51 - e^(arg1[t1]_0 + arg1[t1]_1) is obtained from a table
52 - e^(arg2[t2]_0 + arg2[t2]_1) likewise
53 - n_1 + n_0 = n, so that |n_0| < -LDBL_MIN_EXP-1.
55 If it happens that n_1 == 0 (this is the usual case), that multiplication
67 #include <math_private.h>
69 #define _Float128 long double
74 static const long double C
[] = {
75 /* Smallest integer x for which e^x overflows. */
77 709.78271289338399678773454114191496482L,
79 /* Largest integer x for which e^x underflows. */
81 -744.44007192138126231410729844608163411L,
85 59421121885698253195157962752.0L,
88 #define THREEp103 C[3]
89 30423614405477505635920876929024.0L,
92 #define THREEp111 C[4]
93 7788445287802241442795744493830144.0L,
97 1.44269504088896340735992468100189204L,
99 /* first 93 bits of ln(2) */
101 0.693147180559945309417232121457981864L,
105 -1.94704509238074995158795957333327386E-31L,
107 /* very small number */
113 8.988465674311579538646525953945123668E+307L,
123 /* Chebyshev polynom coefficients for (exp(x)-1)/x */
131 1.66666666666666666666666666666666683E-01L,
132 4.16666666666666666666654902320001674E-02L,
133 8.33333333333333333333314659767198461E-03L,
134 1.38888888889899438565058018857254025E-03L,
135 1.98412698413981650382436541785404286E-04L,
139 __ieee754_expl (long double x
)
141 long double result
, x22
;
142 union ibm_extended_long_double ex2_u
, scale_u
;
145 /* Check for usual case. */
146 if (isless (x
, himark
) && isgreater (x
, lomark
))
148 int tval1
, tval2
, n_i
, exponent2
;
151 SET_RESTORE_ROUND (FE_TONEAREST
);
153 n
= __roundl (x
*M_1_LN2
);
157 tval1
= __roundl (x
*TWO8
);
158 x
-= __expl_table
[T_EXPL_ARG1
+2*tval1
];
159 xl
-= __expl_table
[T_EXPL_ARG1
+2*tval1
+1];
161 tval2
= __roundl (x
*TWO15
);
162 x
-= __expl_table
[T_EXPL_ARG2
+2*tval2
];
163 xl
-= __expl_table
[T_EXPL_ARG2
+2*tval2
+1];
167 /* Compute ex2 = 2^n_0 e^(argtable[tval1]) e^(argtable[tval2]). */
168 ex2_u
.ld
= (__expl_table
[T_EXPL_RES1
+ tval1
]
169 * __expl_table
[T_EXPL_RES2
+ tval2
]);
171 /* 'unsafe' is 1 iff n_1 != 0. */
172 unsafe
= fabsl(n_i
) >= -LDBL_MIN_EXP
- 1;
173 ex2_u
.d
[0].ieee
.exponent
+= n_i
>> unsafe
;
174 /* Fortunately, there are no subnormal lowpart doubles in
175 __expl_table, only normal values and zeros.
176 But after scaling it can be subnormal. */
177 exponent2
= ex2_u
.d
[1].ieee
.exponent
+ (n_i
>> unsafe
);
178 if (ex2_u
.d
[1].ieee
.exponent
== 0)
179 /* assert ((ex2_u.d[1].ieee.mantissa0|ex2_u.d[1].ieee.mantissa1) == 0) */;
180 else if (exponent2
> 0)
181 ex2_u
.d
[1].ieee
.exponent
= exponent2
;
182 else if (exponent2
<= -54)
184 ex2_u
.d
[1].ieee
.exponent
= 0;
185 ex2_u
.d
[1].ieee
.mantissa0
= 0;
186 ex2_u
.d
[1].ieee
.mantissa1
= 0;
191 two54
= 1.80143985094819840000e+16, /* 4350000000000000 */
192 twom54
= 5.55111512312578270212e-17; /* 3C90000000000000 */
193 ex2_u
.d
[1].d
*= two54
;
194 ex2_u
.d
[1].ieee
.exponent
+= n_i
>> unsafe
;
195 ex2_u
.d
[1].d
*= twom54
;
198 /* Compute scale = 2^n_1. */
200 scale_u
.d
[0].ieee
.exponent
+= n_i
- (n_i
>> unsafe
);
202 /* Approximate e^x2 - 1, using a seventh-degree polynomial,
203 with maximum error in [-2^-16-2^-53,2^-16+2^-53]
204 less than 4.8e-39. */
205 x22
= x
+ x
*x
*(P1
+x
*(P2
+x
*(P3
+x
*(P4
+x
*(P5
+x
*P6
)))));
207 /* Now we can test whether the result is ultimate or if we are unsure.
208 In the later case we should probably call a mpn based routine to give
210 Empirically, this routine is already ultimate in about 99.9986% of
211 cases, the test below for the round to nearest case will be false
212 in ~ 99.9963% of cases.
213 Without proc2 routine maximum error which has been seen is
216 union ieee854_long_double ex3_u;
219 fesetround (FE_TONEAREST);
221 ex3_u.d = (result - ex2_u.d) - x22 * ex2_u.d;
223 ex3_u.ieee.exponent += LDBL_MANT_DIG + 15 + IEEE854_LONG_DOUBLE_BIAS
224 - ex2_u.ieee.exponent;
229 if (fegetround () == FE_TONEAREST)
233 return __ieee754_expl_proc2 (origx);
237 /* Exceptional cases: */
238 else if (isless (x
, himark
))
241 /* e^-inf == 0, with no error. */
248 /* Return x, if x is a NaN or Inf; or overflow, otherwise. */
251 result
= x22
* ex2_u
.ld
+ ex2_u
.ld
;
254 return result
* scale_u
.ld
;
256 strong_alias (__ieee754_expl
, __expl_finite
)