1 | //===-- Single-precision 10^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 | #ifndef LLVM_LIBC_SRC_MATH_GENERIC_EXP10F_IMPL_H |
10 | #define LLVM_LIBC_SRC_MATH_GENERIC_EXP10F_IMPL_H |
11 | |
12 | #include "explogxf.h" |
13 | #include "src/__support/FPUtil/BasicOperations.h" |
14 | #include "src/__support/FPUtil/FEnvImpl.h" |
15 | #include "src/__support/FPUtil/FPBits.h" |
16 | #include "src/__support/FPUtil/PolyEval.h" |
17 | #include "src/__support/FPUtil/multiply_add.h" |
18 | #include "src/__support/FPUtil/nearest_integer.h" |
19 | #include "src/__support/FPUtil/rounding_mode.h" |
20 | #include "src/__support/common.h" |
21 | #include "src/__support/macros/optimization.h" // LIBC_UNLIKELY |
22 | |
23 | #include <errno.h> |
24 | |
25 | namespace LIBC_NAMESPACE::generic { |
26 | |
27 | LIBC_INLINE float exp10f(float x) { |
28 | using FPBits = typename fputil::FPBits<float>; |
29 | FPBits xbits(x); |
30 | |
31 | uint32_t x_u = xbits.uintval(); |
32 | uint32_t x_abs = x_u & 0x7fff'ffffU; |
33 | |
34 | // When |x| >= log10(2^128), or x is nan |
35 | if (LIBC_UNLIKELY(x_abs >= 0x421a'209bU)) { |
36 | // When x < log10(2^-150) or nan |
37 | if (x_u > 0xc234'9e35U) { |
38 | // exp(-Inf) = 0 |
39 | if (xbits.is_inf()) |
40 | return 0.0f; |
41 | // exp(nan) = nan |
42 | if (xbits.is_nan()) |
43 | return x; |
44 | if (fputil::fenv_is_round_up()) |
45 | return FPBits::min_subnormal().get_val(); |
46 | fputil::set_errno_if_required(ERANGE); |
47 | fputil::raise_except_if_required(FE_UNDERFLOW); |
48 | return 0.0f; |
49 | } |
50 | // x >= log10(2^128) or nan |
51 | if (xbits.is_pos() && (x_u >= 0x421a'209bU)) { |
52 | // x is finite |
53 | if (x_u < 0x7f80'0000U) { |
54 | int rounding = fputil::quick_get_round(); |
55 | if (rounding == FE_DOWNWARD || rounding == FE_TOWARDZERO) |
56 | return FPBits::max_normal().get_val(); |
57 | |
58 | fputil::set_errno_if_required(ERANGE); |
59 | fputil::raise_except_if_required(FE_OVERFLOW); |
60 | } |
61 | // x is +inf or nan |
62 | return x + FPBits::inf().get_val(); |
63 | } |
64 | } |
65 | |
66 | // When |x| <= log10(2)*2^-6 |
67 | if (LIBC_UNLIKELY(x_abs <= 0x3b9a'209bU)) { |
68 | if (LIBC_UNLIKELY(x_u == 0xb25e'5bd9U)) { // x = -0x1.bcb7b2p-27f |
69 | if (fputil::fenv_is_round_to_nearest()) |
70 | return 0x1.fffffep-1f; |
71 | } |
72 | // |x| < 2^-25 |
73 | // 10^x ~ 1 + log(10) * x |
74 | if (LIBC_UNLIKELY(x_abs <= 0x3280'0000U)) { |
75 | return fputil::multiply_add(x, y: 0x1.26bb1cp+1f, z: 1.0f); |
76 | } |
77 | |
78 | return static_cast<float>(Exp10Base::powb_lo(dx: x)); |
79 | } |
80 | |
81 | // Exceptional value. |
82 | if (LIBC_UNLIKELY(x_u == 0x3d14'd956U)) { // x = 0x1.29b2acp-5f |
83 | if (fputil::fenv_is_round_up()) |
84 | return 0x1.1657c4p+0f; |
85 | } |
86 | |
87 | // Exact outputs when x = 1, 2, ..., 10. |
88 | // Quick check mask: 0x800f'ffffU = ~(bits of 1.0f | ... | bits of 10.0f) |
89 | if (LIBC_UNLIKELY((x_u & 0x800f'ffffU) == 0)) { |
90 | switch (x_u) { |
91 | case 0x3f800000U: // x = 1.0f |
92 | return 10.0f; |
93 | case 0x40000000U: // x = 2.0f |
94 | return 100.0f; |
95 | case 0x40400000U: // x = 3.0f |
96 | return 1'000.0f; |
97 | case 0x40800000U: // x = 4.0f |
98 | return 10'000.0f; |
99 | case 0x40a00000U: // x = 5.0f |
100 | return 100'000.0f; |
101 | case 0x40c00000U: // x = 6.0f |
102 | return 1'000'000.0f; |
103 | case 0x40e00000U: // x = 7.0f |
104 | return 10'000'000.0f; |
105 | case 0x41000000U: // x = 8.0f |
106 | return 100'000'000.0f; |
107 | case 0x41100000U: // x = 9.0f |
108 | return 1'000'000'000.0f; |
109 | case 0x41200000U: // x = 10.0f |
110 | return 10'000'000'000.0f; |
111 | } |
112 | } |
113 | |
114 | // Range reduction: 10^x = 2^(mid + hi) * 10^lo |
115 | // rr = (2^(mid + hi), lo) |
116 | auto rr = exp_b_range_reduc<Exp10Base>(x); |
117 | |
118 | // The low part is approximated by a degree-5 minimax polynomial. |
119 | // 10^lo ~ 1 + COEFFS[0] * lo + ... + COEFFS[4] * lo^5 |
120 | using fputil::multiply_add; |
121 | double lo2 = rr.lo * rr.lo; |
122 | // c0 = 1 + COEFFS[0] * lo |
123 | double c0 = multiply_add(x: rr.lo, y: Exp10Base::COEFFS[0], z: 1.0); |
124 | // c1 = COEFFS[1] + COEFFS[2] * lo |
125 | double c1 = multiply_add(x: rr.lo, y: Exp10Base::COEFFS[2], z: Exp10Base::COEFFS[1]); |
126 | // c2 = COEFFS[3] + COEFFS[4] * lo |
127 | double c2 = multiply_add(x: rr.lo, y: Exp10Base::COEFFS[4], z: Exp10Base::COEFFS[3]); |
128 | // p = c1 + c2 * lo^2 |
129 | // = COEFFS[1] + COEFFS[2] * lo + COEFFS[3] * lo^2 + COEFFS[4] * lo^3 |
130 | double p = multiply_add(x: lo2, y: c2, z: c1); |
131 | // 10^lo ~ c0 + p * lo^2 |
132 | // 10^x = 2^(mid + hi) * 10^lo |
133 | // ~ mh * (c0 + p * lo^2) |
134 | // = (mh * c0) + p * (mh * lo^2) |
135 | return static_cast<float>(multiply_add(x: p, y: lo2 * rr.mh, z: c0 * rr.mh)); |
136 | } |
137 | |
138 | } // namespace LIBC_NAMESPACE::generic |
139 | |
140 | #endif // LLVM_LIBC_SRC_MATH_GENERIC_EXP10F_IMPL_H |
141 | |