1 /* Quad-precision floating point e^x.
2    Copyright (C) 1999-2018 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>
7 
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.
12 
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.
17 
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/>.  */
21 
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,
25    pp. 410-423.
26 
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.
30 
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
34 
35    where:
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
43 
44    Then e^x is approximated as
45 
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))
49    where:
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| < -FLT128_MIN_EXP-1.
54 
55    If it happens that n_1 == 0 (this is the usual case), that multiplication
56    is omitted.
57    */
58 
59 #ifndef _GNU_SOURCE
60 #define _GNU_SOURCE
61 #endif
62 
63 #include "quadmath-imp.h"
64 #include "expq_table.h"
65 
66 static const __float128 C[] = {
67 /* Smallest integer x for which e^x overflows.  */
68 #define himark C[0]
69  11356.523406294143949491931077970765Q,
70 
71 /* Largest integer x for which e^x underflows.  */
72 #define lomark C[1]
73 -11433.4627433362978788372438434526231Q,
74 
75 /* 3x2^96 */
76 #define THREEp96 C[2]
77  59421121885698253195157962752.0Q,
78 
79 /* 3x2^103 */
80 #define THREEp103 C[3]
81  30423614405477505635920876929024.0Q,
82 
83 /* 3x2^111 */
84 #define THREEp111 C[4]
85  7788445287802241442795744493830144.0Q,
86 
87 /* 1/ln(2) */
88 #define M_1_LN2 C[5]
89  1.44269504088896340735992468100189204Q,
90 
91 /* first 93 bits of ln(2) */
92 #define M_LN2_0 C[6]
93  0.693147180559945309417232121457981864Q,
94 
95 /* ln2_0 - ln(2) */
96 #define M_LN2_1 C[7]
97 -1.94704509238074995158795957333327386E-31Q,
98 
99 /* very small number */
100 #define TINY C[8]
101  1.0e-4900Q,
102 
103 /* 2^16383 */
104 #define TWO16383 C[9]
105  5.94865747678615882542879663314003565E+4931Q,
106 
107 /* 256 */
108 #define TWO8 C[10]
109  256,
110 
111 /* 32768 */
112 #define TWO15 C[11]
113  32768,
114 
115 /* Chebyshev polynom coefficients for (exp(x)-1)/x */
116 #define P1 C[12]
117 #define P2 C[13]
118 #define P3 C[14]
119 #define P4 C[15]
120 #define P5 C[16]
121 #define P6 C[17]
122  0.5Q,
123  1.66666666666666666666666666666666683E-01Q,
124  4.16666666666666666666654902320001674E-02Q,
125  8.33333333333333333333314659767198461E-03Q,
126  1.38888888889899438565058018857254025E-03Q,
127  1.98412698413981650382436541785404286E-04Q,
128 };
129 
130 __float128
expq(__float128 x)131 expq (__float128 x)
132 {
133   /* Check for usual case.  */
134   if (__builtin_isless (x, himark) && __builtin_isgreater (x, lomark))
135     {
136       int tval1, tval2, unsafe, n_i;
137       __float128 x22, n, t, result, xl;
138       ieee854_float128 ex2_u, scale_u;
139       fenv_t oldenv;
140 
141       feholdexcept (&oldenv);
142 #ifdef FE_TONEAREST
143       fesetround (FE_TONEAREST);
144 #endif
145 
146       /* Calculate n.  */
147       n = x * M_1_LN2 + THREEp111;
148       n -= THREEp111;
149       x = x - n * M_LN2_0;
150       xl = n * M_LN2_1;
151 
152       /* Calculate t/256.  */
153       t = x + THREEp103;
154       t -= THREEp103;
155 
156       /* Compute tval1 = t.  */
157       tval1 = (int) (t * TWO8);
158 
159       x -= __expq_table[T_EXPL_ARG1+2*tval1];
160       xl -= __expq_table[T_EXPL_ARG1+2*tval1+1];
161 
162       /* Calculate t/32768.  */
163       t = x + THREEp96;
164       t -= THREEp96;
165 
166       /* Compute tval2 = t.  */
167       tval2 = (int) (t * TWO15);
168 
169       x -= __expq_table[T_EXPL_ARG2+2*tval2];
170       xl -= __expq_table[T_EXPL_ARG2+2*tval2+1];
171 
172       x = x + xl;
173 
174       /* Compute ex2 = 2^n_0 e^(argtable[tval1]) e^(argtable[tval2]).  */
175       ex2_u.value = __expq_table[T_EXPL_RES1 + tval1]
176 		* __expq_table[T_EXPL_RES2 + tval2];
177       n_i = (int)n;
178       /* 'unsafe' is 1 iff n_1 != 0.  */
179       unsafe = abs(n_i) >= 15000;
180       ex2_u.ieee.exponent += n_i >> unsafe;
181 
182       /* Compute scale = 2^n_1.  */
183       scale_u.value = 1;
184       scale_u.ieee.exponent += n_i - (n_i >> unsafe);
185 
186       /* Approximate e^x2 - 1, using a seventh-degree polynomial,
187 	 with maximum error in [-2^-16-2^-53,2^-16+2^-53]
188 	 less than 4.8e-39.  */
189       x22 = x + x*x*(P1+x*(P2+x*(P3+x*(P4+x*(P5+x*P6)))));
190       math_force_eval (x22);
191 
192       /* Return result.  */
193       fesetenv (&oldenv);
194 
195       result = x22 * ex2_u.value + ex2_u.value;
196 
197       /* Now we can test whether the result is ultimate or if we are unsure.
198 	 In the later case we should probably call a mpn based routine to give
199 	 the ultimate result.
200 	 Empirically, this routine is already ultimate in about 99.9986% of
201 	 cases, the test below for the round to nearest case will be false
202 	 in ~ 99.9963% of cases.
203 	 Without proc2 routine maximum error which has been seen is
204 	 0.5000262 ulp.
205 
206 	  ieee854_float128 ex3_u;
207 
208 	  #ifdef FE_TONEAREST
209 	    fesetround (FE_TONEAREST);
210 	  #endif
211 	  ex3_u.value = (result - ex2_u.value) - x22 * ex2_u.value;
212 	  ex2_u.value = result;
213 	  ex3_u.ieee.exponent += FLT128_MANT_DIG + 15 + IEEE854_FLOAT128_BIAS
214 				 - ex2_u.ieee.exponent;
215 	  n_i = abs (ex3_u.value);
216 	  n_i = (n_i + 1) / 2;
217 	  fesetenv (&oldenv);
218 	  #ifdef FE_TONEAREST
219 	  if (fegetround () == FE_TONEAREST)
220 	    n_i -= 0x4000;
221 	  #endif
222 	  if (!n_i) {
223 	    return __ieee754_expl_proc2 (origx);
224 	  }
225        */
226       if (!unsafe)
227 	return result;
228       else
229 	{
230 	  result *= scale_u.value;
231 	  math_check_force_underflow_nonneg (result);
232 	  return result;
233 	}
234     }
235   /* Exceptional cases:  */
236   else if (__builtin_isless (x, himark))
237     {
238       if (isinfq (x))
239 	/* e^-inf == 0, with no error.  */
240 	return 0;
241       else
242 	/* Underflow */
243 	return TINY * TINY;
244     }
245   else
246     /* Return x, if x is a NaN or Inf; or overflow, otherwise.  */
247     return TWO16383*x;
248 }
249