1 | //===-- Single-precision e^x function -------------------------------------===// |
2 | // |
3 | // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions. |
4 | // See https://llvm.org/LICENSE.txt for license information. |
5 | // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception |
6 | // |
7 | //===----------------------------------------------------------------------===// |
8 | |
9 | #include "src/math/expf.h" |
10 | #include "common_constants.h" // Lookup tables EXP_M1 and EXP_M2. |
11 | #include "src/__support/FPUtil/BasicOperations.h" |
12 | #include "src/__support/FPUtil/FEnvImpl.h" |
13 | #include "src/__support/FPUtil/FPBits.h" |
14 | #include "src/__support/FPUtil/PolyEval.h" |
15 | #include "src/__support/FPUtil/multiply_add.h" |
16 | #include "src/__support/FPUtil/nearest_integer.h" |
17 | #include "src/__support/FPUtil/rounding_mode.h" |
18 | #include "src/__support/common.h" |
19 | #include "src/__support/macros/optimization.h" // LIBC_UNLIKELY |
20 | |
21 | #include <errno.h> |
22 | |
23 | namespace LIBC_NAMESPACE { |
24 | |
25 | LLVM_LIBC_FUNCTION(float, expf, (float x)) { |
26 | using FPBits = typename fputil::FPBits<float>; |
27 | FPBits xbits(x); |
28 | |
29 | uint32_t x_u = xbits.uintval(); |
30 | uint32_t x_abs = x_u & 0x7fff'ffffU; |
31 | |
32 | // Exceptional values |
33 | if (LIBC_UNLIKELY(x_u == 0xc236'bd8cU)) { // x = -0x1.6d7b18p+5f |
34 | return 0x1.108a58p-66f - x * 0x1.0p-95f; |
35 | } |
36 | |
37 | // When |x| >= 89, |x| < 2^-25, or x is nan |
38 | if (LIBC_UNLIKELY(x_abs >= 0x42b2'0000U || x_abs <= 0x3280'0000U)) { |
39 | // |x| < 2^-25 |
40 | if (xbits.get_biased_exponent() <= 101) { |
41 | return 1.0f + x; |
42 | } |
43 | |
44 | // When x < log(2^-150) or nan |
45 | if (xbits.uintval() >= 0xc2cf'f1b5U) { |
46 | // exp(-Inf) = 0 |
47 | if (xbits.is_inf()) |
48 | return 0.0f; |
49 | // exp(nan) = nan |
50 | if (xbits.is_nan()) |
51 | return x; |
52 | if (fputil::fenv_is_round_up()) |
53 | return FPBits::min_subnormal().get_val(); |
54 | fputil::set_errno_if_required(ERANGE); |
55 | fputil::raise_except_if_required(FE_UNDERFLOW); |
56 | return 0.0f; |
57 | } |
58 | // x >= 89 or nan |
59 | if (xbits.is_pos() && (xbits.uintval() >= 0x42b2'0000)) { |
60 | // x is finite |
61 | if (xbits.uintval() < 0x7f80'0000U) { |
62 | int rounding = fputil::quick_get_round(); |
63 | if (rounding == FE_DOWNWARD || rounding == FE_TOWARDZERO) |
64 | return FPBits::max_normal().get_val(); |
65 | |
66 | fputil::set_errno_if_required(ERANGE); |
67 | fputil::raise_except_if_required(FE_OVERFLOW); |
68 | } |
69 | // x is +inf or nan |
70 | return x + FPBits::inf().get_val(); |
71 | } |
72 | } |
73 | // For -104 < x < 89, to compute exp(x), we perform the following range |
74 | // reduction: find hi, mid, lo such that: |
75 | // x = hi + mid + lo, in which |
76 | // hi is an integer, |
77 | // mid * 2^7 is an integer |
78 | // -2^(-8) <= lo < 2^-8. |
79 | // In particular, |
80 | // hi + mid = round(x * 2^7) * 2^(-7). |
81 | // Then, |
82 | // exp(x) = exp(hi + mid + lo) = exp(hi) * exp(mid) * exp(lo). |
83 | // We store exp(hi) and exp(mid) in the lookup tables EXP_M1 and EXP_M2 |
84 | // respectively. exp(lo) is computed using a degree-4 minimax polynomial |
85 | // generated by Sollya. |
86 | |
87 | // x_hi = (hi + mid) * 2^7 = round(x * 2^7). |
88 | float kf = fputil::nearest_integer(x: x * 0x1.0p7f); |
89 | // Subtract (hi + mid) from x to get lo. |
90 | double xd = static_cast<double>(fputil::multiply_add(x: kf, y: -0x1.0p-7f, z: x)); |
91 | int x_hi = static_cast<int>(kf); |
92 | x_hi += 104 << 7; |
93 | // hi = x_hi >> 7 |
94 | double exp_hi = EXP_M1[x_hi >> 7]; |
95 | // mid * 2^7 = x_hi & 0x0000'007fU; |
96 | double exp_mid = EXP_M2[x_hi & 0x7f]; |
97 | // Degree-4 minimax polynomial generated by Sollya with the following |
98 | // commands: |
99 | // > display = hexadecimal; |
100 | // > Q = fpminimax(expm1(x)/x, 3, [|D...|], [-2^-8, 2^-8]); |
101 | // > Q; |
102 | double exp_lo = |
103 | fputil::polyeval(x: xd, a0: 0x1p0, a: 0x1.ffffffffff777p-1, a: 0x1.000000000071cp-1, |
104 | a: 0x1.555566668e5e7p-3, a: 0x1.55555555ef243p-5); |
105 | return static_cast<float>(exp_hi * exp_mid * exp_lo); |
106 | } |
107 | |
108 | } // namespace LIBC_NAMESPACE |
109 | |