1 | /* |
2 | * Double-precision 2^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 Shift __exp_data.exp2_shift |
16 | #define T __exp_data.tab |
17 | #define C1 __exp_data.exp2_poly[0] |
18 | #define C2 __exp_data.exp2_poly[1] |
19 | #define C3 __exp_data.exp2_poly[2] |
20 | #define C4 __exp_data.exp2_poly[3] |
21 | #define C5 __exp_data.exp2_poly[4] |
22 | #define C6 __exp_data.exp2_poly[5] |
23 | |
24 | /* Handle cases that may overflow or underflow when computing the result that |
25 | is scale*(1+TMP) without intermediate rounding. The bit representation of |
26 | scale is in SBITS, however it has a computed exponent that may have |
27 | overflown into the sign bit so that needs to be adjusted before using it as |
28 | a double. (int32_t)KI is the k used in the argument reduction and exponent |
29 | adjustment of scale, positive k here means the result may overflow and |
30 | negative k means the result may underflow. */ |
31 | static inline double |
32 | specialcase (double_t tmp, uint64_t sbits, uint64_t ki) |
33 | { |
34 | double_t scale, y; |
35 | |
36 | if ((ki & 0x80000000) == 0) |
37 | { |
38 | /* k > 0, the exponent of scale might have overflowed by 1. */ |
39 | sbits -= 1ull << 52; |
40 | scale = asdouble (i: sbits); |
41 | y = 2 * (scale + scale * tmp); |
42 | return check_oflow (x: eval_as_double (x: y)); |
43 | } |
44 | /* k < 0, need special care in the subnormal range. */ |
45 | sbits += 1022ull << 52; |
46 | scale = asdouble (i: sbits); |
47 | y = scale + scale * tmp; |
48 | if (y < 1.0) |
49 | { |
50 | /* Round y to the right precision before scaling it into the subnormal |
51 | range to avoid double rounding that can cause 0.5+E/2 ulp error where |
52 | E is the worst-case ulp error outside the subnormal range. So this |
53 | is only useful if the goal is better than 1 ulp worst-case error. */ |
54 | double_t hi, lo; |
55 | lo = scale - y + scale * tmp; |
56 | hi = 1.0 + y; |
57 | lo = 1.0 - hi + y + lo; |
58 | y = eval_as_double (x: hi + lo) - 1.0; |
59 | /* Avoid -0.0 with downward rounding. */ |
60 | if (WANT_ROUNDING && y == 0.0) |
61 | y = 0.0; |
62 | /* The underflow exception needs to be signaled explicitly. */ |
63 | force_eval_double (x: opt_barrier_double (x: 0x1p-1022) * 0x1p-1022); |
64 | } |
65 | y = 0x1p-1022 * y; |
66 | return check_uflow (x: eval_as_double (x: y)); |
67 | } |
68 | |
69 | /* Top 12 bits of a double (sign and exponent bits). */ |
70 | static inline uint32_t |
71 | top12 (double x) |
72 | { |
73 | return asuint64 (f: x) >> 52; |
74 | } |
75 | |
76 | double |
77 | exp2 (double x) |
78 | { |
79 | uint32_t abstop; |
80 | uint64_t ki, idx, top, sbits; |
81 | /* double_t for better performance on targets with FLT_EVAL_METHOD==2. */ |
82 | double_t kd, r, r2, scale, tail, tmp; |
83 | |
84 | abstop = top12 (x) & 0x7ff; |
85 | if (unlikely (abstop - top12 (0x1p-54) >= top12 (512.0) - top12 (0x1p-54))) |
86 | { |
87 | if (abstop - top12 (x: 0x1p-54) >= 0x80000000) |
88 | /* Avoid spurious underflow for tiny x. */ |
89 | /* Note: 0 is common input. */ |
90 | return WANT_ROUNDING ? 1.0 + x : 1.0; |
91 | if (abstop >= top12 (x: 1024.0)) |
92 | { |
93 | if (asuint64 (f: x) == asuint64 (f: -INFINITY)) |
94 | return 0.0; |
95 | if (abstop >= top12 (INFINITY)) |
96 | return 1.0 + x; |
97 | if (!(asuint64 (f: x) >> 63)) |
98 | return __math_oflow (0); |
99 | else if (asuint64 (f: x) >= asuint64 (f: -1075.0)) |
100 | return __math_uflow (0); |
101 | } |
102 | if (2 * asuint64 (f: x) > 2 * asuint64 (f: 928.0)) |
103 | /* Large x is special cased below. */ |
104 | abstop = 0; |
105 | } |
106 | |
107 | /* exp2(x) = 2^(k/N) * 2^r, with 2^r in [2^(-1/2N),2^(1/2N)]. */ |
108 | /* x = k/N + r, with int k and r in [-1/2N, 1/2N]. */ |
109 | kd = eval_as_double (x: x + Shift); |
110 | ki = asuint64 (f: kd); /* k. */ |
111 | kd -= Shift; /* k/N for int k. */ |
112 | r = x - kd; |
113 | /* 2^(k/N) ~= scale * (1 + tail). */ |
114 | idx = 2 * (ki % N); |
115 | top = ki << (52 - EXP_TABLE_BITS); |
116 | tail = asdouble (T[idx]); |
117 | /* This is only a valid scale when -1023*N < k < 1024*N. */ |
118 | sbits = T[idx + 1] + top; |
119 | /* exp2(x) = 2^(k/N) * 2^r ~= scale + scale * (tail + 2^r - 1). */ |
120 | /* Evaluation is optimized assuming superscalar pipelined execution. */ |
121 | r2 = r * r; |
122 | /* Without fma the worst case error is 0.5/N ulp larger. */ |
123 | /* Worst case error is less than 0.5+0.86/N+(abs poly error * 2^53) ulp. */ |
124 | #if EXP2_POLY_ORDER == 4 |
125 | tmp = tail + r * C1 + r2 * C2 + r * r2 * (C3 + r * C4); |
126 | #elif EXP2_POLY_ORDER == 5 |
127 | tmp = tail + r * C1 + r2 * (C2 + r * C3) + r2 * r2 * (C4 + r * C5); |
128 | #elif EXP2_POLY_ORDER == 6 |
129 | tmp = tail + r * C1 + r2 * (0.5 + r * C3) + r2 * r2 * (C4 + r * C5 + r2 * C6); |
130 | #endif |
131 | if (unlikely (abstop == 0)) |
132 | return specialcase (tmp, sbits, ki); |
133 | scale = asdouble (i: sbits); |
134 | /* Note: tmp == 0 or |tmp| > 2^-65 and scale > 2^-928, so there |
135 | is no spurious underflow here even without fma. */ |
136 | return eval_as_double (x: scale + scale * tmp); |
137 | } |
138 | #if USE_GLIBC_ABI |
139 | strong_alias (exp2, __exp2_finite) |
140 | hidden_alias (exp2, __ieee754_exp2) |
141 | # if LDBL_MANT_DIG == 53 |
142 | long double exp2l (long double x) { return exp2 (x); } |
143 | # endif |
144 | #endif |
145 | |