1 | /* |
2 | * Double-precision e^x function. |
3 | * |
4 | * Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions. |
5 | * See https://llvm.org/LICENSE.txt for license information. |
6 | * SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception |
7 | */ |
8 | |
9 | #include <float.h> |
10 | #include <math.h> |
11 | #include <stdint.h> |
12 | #include "math_config.h" |
13 | |
14 | #define N (1 << EXP_TABLE_BITS) |
15 | #define InvLn2N __exp_data.invln2N |
16 | #define NegLn2hiN __exp_data.negln2hiN |
17 | #define NegLn2loN __exp_data.negln2loN |
18 | #define Shift __exp_data.shift |
19 | #define T __exp_data.tab |
20 | #define C2 __exp_data.poly[5 - EXP_POLY_ORDER] |
21 | #define C3 __exp_data.poly[6 - EXP_POLY_ORDER] |
22 | #define C4 __exp_data.poly[7 - EXP_POLY_ORDER] |
23 | #define C5 __exp_data.poly[8 - EXP_POLY_ORDER] |
24 | #define C6 __exp_data.poly[9 - EXP_POLY_ORDER] |
25 | |
26 | /* Handle cases that may overflow or underflow when computing the result that |
27 | is scale*(1+TMP) without intermediate rounding. The bit representation of |
28 | scale is in SBITS, however it has a computed exponent that may have |
29 | overflown into the sign bit so that needs to be adjusted before using it as |
30 | a double. (int32_t)KI is the k used in the argument reduction and exponent |
31 | adjustment of scale, positive k here means the result may overflow and |
32 | negative k means the result may underflow. */ |
33 | static inline double |
34 | specialcase (double_t tmp, uint64_t sbits, uint64_t ki) |
35 | { |
36 | double_t scale, y; |
37 | |
38 | if ((ki & 0x80000000) == 0) |
39 | { |
40 | /* k > 0, the exponent of scale might have overflowed by <= 460. */ |
41 | sbits -= 1009ull << 52; |
42 | scale = asdouble (i: sbits); |
43 | y = 0x1p1009 * (scale + scale * tmp); |
44 | return check_oflow (x: eval_as_double (x: y)); |
45 | } |
46 | /* k < 0, need special care in the subnormal range. */ |
47 | sbits += 1022ull << 52; |
48 | scale = asdouble (i: sbits); |
49 | y = scale + scale * tmp; |
50 | if (y < 1.0) |
51 | { |
52 | /* Round y to the right precision before scaling it into the subnormal |
53 | range to avoid double rounding that can cause 0.5+E/2 ulp error where |
54 | E is the worst-case ulp error outside the subnormal range. So this |
55 | is only useful if the goal is better than 1 ulp worst-case error. */ |
56 | double_t hi, lo; |
57 | lo = scale - y + scale * tmp; |
58 | hi = 1.0 + y; |
59 | lo = 1.0 - hi + y + lo; |
60 | y = eval_as_double (x: hi + lo) - 1.0; |
61 | /* Avoid -0.0 with downward rounding. */ |
62 | if (WANT_ROUNDING && y == 0.0) |
63 | y = 0.0; |
64 | /* The underflow exception needs to be signaled explicitly. */ |
65 | force_eval_double (x: opt_barrier_double (x: 0x1p-1022) * 0x1p-1022); |
66 | } |
67 | y = 0x1p-1022 * y; |
68 | return check_uflow (x: eval_as_double (x: y)); |
69 | } |
70 | |
71 | /* Top 12 bits of a double (sign and exponent bits). */ |
72 | static inline uint32_t |
73 | top12 (double x) |
74 | { |
75 | return asuint64 (f: x) >> 52; |
76 | } |
77 | |
78 | /* Computes exp(x+xtail) where |xtail| < 2^-8/N and |xtail| <= |x|. |
79 | If hastail is 0 then xtail is assumed to be 0 too. */ |
80 | static inline double |
81 | exp_inline (double x, double xtail, int hastail) |
82 | { |
83 | uint32_t abstop; |
84 | uint64_t ki, idx, top, sbits; |
85 | /* double_t for better performance on targets with FLT_EVAL_METHOD==2. */ |
86 | double_t kd, z, r, r2, scale, tail, tmp; |
87 | |
88 | abstop = top12 (x) & 0x7ff; |
89 | if (unlikely (abstop - top12 (0x1p-54) >= top12 (512.0) - top12 (0x1p-54))) |
90 | { |
91 | if (abstop - top12 (x: 0x1p-54) >= 0x80000000) |
92 | /* Avoid spurious underflow for tiny x. */ |
93 | /* Note: 0 is common input. */ |
94 | return WANT_ROUNDING ? 1.0 + x : 1.0; |
95 | if (abstop >= top12 (x: 1024.0)) |
96 | { |
97 | if (asuint64 (f: x) == asuint64 (f: -INFINITY)) |
98 | return 0.0; |
99 | if (abstop >= top12 (INFINITY)) |
100 | return 1.0 + x; |
101 | if (asuint64 (f: x) >> 63) |
102 | return __math_uflow (0); |
103 | else |
104 | return __math_oflow (0); |
105 | } |
106 | /* Large x is special cased below. */ |
107 | abstop = 0; |
108 | } |
109 | |
110 | /* exp(x) = 2^(k/N) * exp(r), with exp(r) in [2^(-1/2N),2^(1/2N)]. */ |
111 | /* x = ln2/N*k + r, with int k and r in [-ln2/2N, ln2/2N]. */ |
112 | z = InvLn2N * x; |
113 | #if TOINT_INTRINSICS |
114 | kd = roundtoint (z); |
115 | ki = converttoint (z); |
116 | #elif EXP_USE_TOINT_NARROW |
117 | /* z - kd is in [-0.5-2^-16, 0.5] in all rounding modes. */ |
118 | kd = eval_as_double (z + Shift); |
119 | ki = asuint64 (kd) >> 16; |
120 | kd = (double_t) (int32_t) ki; |
121 | #else |
122 | /* z - kd is in [-1, 1] in non-nearest rounding modes. */ |
123 | kd = eval_as_double (x: z + Shift); |
124 | ki = asuint64 (f: kd); |
125 | kd -= Shift; |
126 | #endif |
127 | r = x + kd * NegLn2hiN + kd * NegLn2loN; |
128 | /* The code assumes 2^-200 < |xtail| < 2^-8/N. */ |
129 | if (hastail) |
130 | r += xtail; |
131 | /* 2^(k/N) ~= scale * (1 + tail). */ |
132 | idx = 2 * (ki % N); |
133 | top = ki << (52 - EXP_TABLE_BITS); |
134 | tail = asdouble (T[idx]); |
135 | /* This is only a valid scale when -1023*N < k < 1024*N. */ |
136 | sbits = T[idx + 1] + top; |
137 | /* exp(x) = 2^(k/N) * exp(r) ~= scale + scale * (tail + exp(r) - 1). */ |
138 | /* Evaluation is optimized assuming superscalar pipelined execution. */ |
139 | r2 = r * r; |
140 | /* Without fma the worst case error is 0.25/N ulp larger. */ |
141 | /* Worst case error is less than 0.5+1.11/N+(abs poly error * 2^53) ulp. */ |
142 | #if EXP_POLY_ORDER == 4 |
143 | tmp = tail + r + r2 * C2 + r * r2 * (C3 + r * C4); |
144 | #elif EXP_POLY_ORDER == 5 |
145 | tmp = tail + r + r2 * (C2 + r * C3) + r2 * r2 * (C4 + r * C5); |
146 | #elif EXP_POLY_ORDER == 6 |
147 | tmp = tail + r + r2 * (0.5 + r * C3) + r2 * r2 * (C4 + r * C5 + r2 * C6); |
148 | #endif |
149 | if (unlikely (abstop == 0)) |
150 | return specialcase (tmp, sbits, ki); |
151 | scale = asdouble (i: sbits); |
152 | /* Note: tmp == 0 or |tmp| > 2^-200 and scale > 2^-739, so there |
153 | is no spurious underflow here even without fma. */ |
154 | return eval_as_double (x: scale + scale * tmp); |
155 | } |
156 | |
157 | double |
158 | exp (double x) |
159 | { |
160 | return exp_inline (x, xtail: 0, hastail: 0); |
161 | } |
162 | |
163 | /* May be useful for implementing pow where more than double |
164 | precision input is needed. */ |
165 | double |
166 | __exp_dd (double x, double xtail) |
167 | { |
168 | return exp_inline (x, xtail, hastail: 1); |
169 | } |
170 | #if USE_GLIBC_ABI |
171 | strong_alias (exp, __exp_finite) |
172 | hidden_alias (exp, __ieee754_exp) |
173 | hidden_alias (__exp_dd, __exp1) |
174 | # if LDBL_MANT_DIG == 53 |
175 | long double expl (long double x) { return exp (x); } |
176 | # endif |
177 | #endif |
178 | |