1 /* Double-precision e^x function.
2    Copyright (C) 2018-2022 Free Software Foundation, Inc.
3    This file is part of the GNU C Library.
4 
5    The GNU C Library is free software; you can redistribute it and/or
6    modify it under the terms of the GNU Lesser General Public
7    License as published by the Free Software Foundation; either
8    version 2.1 of the License, or (at your option) any later version.
9 
10    The GNU C Library is distributed in the hope that it will be useful,
11    but WITHOUT ANY WARRANTY; without even the implied warranty of
12    MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
13    Lesser General Public License for more details.
14 
15    You should have received a copy of the GNU Lesser General Public
16    License along with the GNU C Library; if not, see
17    <https://www.gnu.org/licenses/>.  */
18 
19 #include <math.h>
20 #include <stdint.h>
21 #include <math-barriers.h>
22 #include <math-narrow-eval.h>
23 #include <math-svid-compat.h>
24 #include <libm-alias-finite.h>
25 #include <libm-alias-double.h>
26 #include "math_config.h"
27 
28 #define N (1 << EXP_TABLE_BITS)
29 #define InvLn2N __exp_data.invln2N
30 #define NegLn2hiN __exp_data.negln2hiN
31 #define NegLn2loN __exp_data.negln2loN
32 #define Shift __exp_data.shift
33 #define T __exp_data.tab
34 #define C2 __exp_data.poly[5 - EXP_POLY_ORDER]
35 #define C3 __exp_data.poly[6 - EXP_POLY_ORDER]
36 #define C4 __exp_data.poly[7 - EXP_POLY_ORDER]
37 #define C5 __exp_data.poly[8 - EXP_POLY_ORDER]
38 
39 /* Handle cases that may overflow or underflow when computing the result that
40    is scale*(1+TMP) without intermediate rounding.  The bit representation of
41    scale is in SBITS, however it has a computed exponent that may have
42    overflown into the sign bit so that needs to be adjusted before using it as
43    a double.  (int32_t)KI is the k used in the argument reduction and exponent
44    adjustment of scale, positive k here means the result may overflow and
45    negative k means the result may underflow.  */
46 static inline double
specialcase(double_t tmp,uint64_t sbits,uint64_t ki)47 specialcase (double_t tmp, uint64_t sbits, uint64_t ki)
48 {
49   double_t scale, y;
50 
51   if ((ki & 0x80000000) == 0)
52     {
53       /* k > 0, the exponent of scale might have overflowed by <= 460.  */
54       sbits -= 1009ull << 52;
55       scale = asdouble (sbits);
56       y = 0x1p1009 * (scale + scale * tmp);
57       return check_oflow (y);
58     }
59   /* k < 0, need special care in the subnormal range.  */
60   sbits += 1022ull << 52;
61   scale = asdouble (sbits);
62   y = scale + scale * tmp;
63   if (y < 1.0)
64     {
65       /* Round y to the right precision before scaling it into the subnormal
66 	 range to avoid double rounding that can cause 0.5+E/2 ulp error where
67 	 E is the worst-case ulp error outside the subnormal range.  So this
68 	 is only useful if the goal is better than 1 ulp worst-case error.  */
69       double_t hi, lo;
70       lo = scale - y + scale * tmp;
71       hi = 1.0 + y;
72       lo = 1.0 - hi + y + lo;
73       y = math_narrow_eval (hi + lo) - 1.0;
74       /* Avoid -0.0 with downward rounding.  */
75       if (WANT_ROUNDING && y == 0.0)
76 	y = 0.0;
77       /* The underflow exception needs to be signaled explicitly.  */
78       math_force_eval (math_opt_barrier (0x1p-1022) * 0x1p-1022);
79     }
80   y = 0x1p-1022 * y;
81   return check_uflow (y);
82 }
83 
84 /* Top 12 bits of a double (sign and exponent bits).  */
85 static inline uint32_t
top12(double x)86 top12 (double x)
87 {
88   return asuint64 (x) >> 52;
89 }
90 
91 #ifndef SECTION
92 # define SECTION
93 #endif
94 
95 double
96 SECTION
__exp(double x)97 __exp (double x)
98 {
99   uint32_t abstop;
100   uint64_t ki, idx, top, sbits;
101   /* double_t for better performance on targets with FLT_EVAL_METHOD==2.  */
102   double_t kd, z, r, r2, scale, tail, tmp;
103 
104   abstop = top12 (x) & 0x7ff;
105   if (__glibc_unlikely (abstop - top12 (0x1p-54)
106 			>= top12 (512.0) - top12 (0x1p-54)))
107     {
108       if (abstop - top12 (0x1p-54) >= 0x80000000)
109 	/* Avoid spurious underflow for tiny x.  */
110 	/* Note: 0 is common input.  */
111 	return WANT_ROUNDING ? 1.0 + x : 1.0;
112       if (abstop >= top12 (1024.0))
113 	{
114 	  if (asuint64 (x) == asuint64 (-INFINITY))
115 	    return 0.0;
116 	  if (abstop >= top12 (INFINITY))
117 	    return 1.0 + x;
118 	  if (asuint64 (x) >> 63)
119 	    return __math_uflow (0);
120 	  else
121 	    return __math_oflow (0);
122 	}
123       /* Large x is special cased below.  */
124       abstop = 0;
125     }
126 
127   /* exp(x) = 2^(k/N) * exp(r), with exp(r) in [2^(-1/2N),2^(1/2N)].  */
128   /* x = ln2/N*k + r, with int k and r in [-ln2/2N, ln2/2N].  */
129   z = InvLn2N * x;
130 #if TOINT_INTRINSICS
131   kd = roundtoint (z);
132   ki = converttoint (z);
133 #else
134   /* z - kd is in [-1, 1] in non-nearest rounding modes.  */
135   kd = math_narrow_eval (z + Shift);
136   ki = asuint64 (kd);
137   kd -= Shift;
138 #endif
139   r = x + kd * NegLn2hiN + kd * NegLn2loN;
140   /* 2^(k/N) ~= scale * (1 + tail).  */
141   idx = 2 * (ki % N);
142   top = ki << (52 - EXP_TABLE_BITS);
143   tail = asdouble (T[idx]);
144   /* This is only a valid scale when -1023*N < k < 1024*N.  */
145   sbits = T[idx + 1] + top;
146   /* exp(x) = 2^(k/N) * exp(r) ~= scale + scale * (tail + exp(r) - 1).  */
147   /* Evaluation is optimized assuming superscalar pipelined execution.  */
148   r2 = r * r;
149   /* Without fma the worst case error is 0.25/N ulp larger.  */
150   /* Worst case error is less than 0.5+1.11/N+(abs poly error * 2^53) ulp.  */
151   tmp = tail + r + r2 * (C2 + r * C3) + r2 * r2 * (C4 + r * C5);
152   if (__glibc_unlikely (abstop == 0))
153     return specialcase (tmp, sbits, ki);
154   scale = asdouble (sbits);
155   /* Note: tmp == 0 or |tmp| > 2^-65 and scale > 2^-739, so there
156      is no spurious underflow here even without fma.  */
157   return scale + scale * tmp;
158 }
159 #ifndef __exp
160 hidden_def (__exp)
161 strong_alias (__exp, __ieee754_exp)
162 libm_alias_finite (__ieee754_exp, __exp)
163 # if LIBM_SVID_COMPAT
164 versioned_symbol (libm, __exp, exp, GLIBC_2_29);
165 libm_alias_double_other (__exp, exp)
166 # else
167 libm_alias_double (__exp, exp)
168 # endif
169 #endif
170