1 | //===-- Implementation of exp10m1f 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/exp10m1f.h" |
10 | #include "src/__support/FPUtil/FEnvImpl.h" |
11 | #include "src/__support/FPUtil/FPBits.h" |
12 | #include "src/__support/FPUtil/PolyEval.h" |
13 | #include "src/__support/FPUtil/except_value_utils.h" |
14 | #include "src/__support/FPUtil/multiply_add.h" |
15 | #include "src/__support/FPUtil/rounding_mode.h" |
16 | #include "src/__support/common.h" |
17 | #include "src/__support/libc_errno.h" |
18 | #include "src/__support/macros/config.h" |
19 | #include "src/__support/macros/optimization.h" |
20 | |
21 | #include "explogxf.h" |
22 | |
23 | namespace LIBC_NAMESPACE_DECL { |
24 | |
25 | #ifndef LIBC_MATH_HAS_SKIP_ACCURATE_PASS |
26 | static constexpr size_t N_EXCEPTS_LO = 11; |
27 | |
28 | static constexpr fputil::ExceptValues<float, N_EXCEPTS_LO> EXP10M1F_EXCEPTS_LO = |
29 | {{ |
30 | // x = 0x1.0fe54ep-11, exp10m1f(x) = 0x1.3937eep-10 (RZ) |
31 | {0x3a07'f2a7U, 0x3a9c'9bf7U, 1U, 0U, 1U}, |
32 | // x = 0x1.80e6eap-11, exp10m1f(x) = 0x1.bb8272p-10 (RZ) |
33 | {0x3a40'7375U, 0x3add'c139U, 1U, 0U, 1U}, |
34 | // x = -0x1.2a33bcp-51, exp10m1f(x) = -0x1.57515ep-50 (RZ) |
35 | {0xa615'19deU, 0xa6ab'a8afU, 0U, 1U, 0U}, |
36 | // x = -0x0p+0, exp10m1f(x) = -0x0p+0 (RZ) |
37 | {0x8000'0000U, 0x8000'0000U, 0U, 0U, 0U}, |
38 | // x = -0x1.b59e08p-31, exp10m1f(x) = -0x1.f7d356p-30 (RZ) |
39 | {0xb05a'cf04U, 0xb0fb'e9abU, 0U, 1U, 1U}, |
40 | // x = -0x1.bf342p-12, exp10m1f(x) = -0x1.014e02p-10 (RZ) |
41 | {0xb9df'9a10U, 0xba80'a701U, 0U, 1U, 0U}, |
42 | // x = -0x1.6207fp-11, exp10m1f(x) = -0x1.9746cap-10 (RZ) |
43 | {0xba31'03f8U, 0xbacb'a365U, 0U, 1U, 1U}, |
44 | // x = -0x1.bd0c66p-11, exp10m1f(x) = -0x1.ffe168p-10 (RZ) |
45 | {0xba5e'8633U, 0xbaff'f0b4U, 0U, 1U, 1U}, |
46 | // x = -0x1.ffd84cp-10, exp10m1f(x) = -0x1.25faf2p-8 (RZ) |
47 | {0xbaff'ec26U, 0xbb92'fd79U, 0U, 1U, 0U}, |
48 | // x = -0x1.a74172p-9, exp10m1f(x) = -0x1.e57be2p-8 (RZ) |
49 | {0xbb53'a0b9U, 0xbbf2'bdf1U, 0U, 1U, 1U}, |
50 | // x = -0x1.cb694cp-9, exp10m1f(x) = -0x1.0764e4p-7 (RZ) |
51 | {0xbb65'b4a6U, 0xbc03'b272U, 0U, 1U, 0U}, |
52 | }}; |
53 | |
54 | static constexpr size_t N_EXCEPTS_HI = 19; |
55 | |
56 | static constexpr fputil::ExceptValues<float, N_EXCEPTS_HI> EXP10M1F_EXCEPTS_HI = |
57 | {{ |
58 | // (input, RZ output, RU offset, RD offset, RN offset) |
59 | // x = 0x1.8d31eep-8, exp10m1f(x) = 0x1.cc7e4cp-7 (RZ) |
60 | {0x3bc6'98f7U, 0x3c66'3f26U, 1U, 0U, 1U}, |
61 | // x = 0x1.915fcep-8, exp10m1f(x) = 0x1.d15f72p-7 (RZ) |
62 | {0x3bc8'afe7U, 0x3c68'afb9U, 1U, 0U, 0U}, |
63 | // x = 0x1.bcf982p-8, exp10m1f(x) = 0x1.022928p-6 (RZ) |
64 | {0x3bde'7cc1U, 0x3c81'1494U, 1U, 0U, 1U}, |
65 | // x = 0x1.99ff0ap-7, exp10m1f(x) = 0x1.dee416p-6 (RZ) |
66 | {0x3c4c'ff85U, 0x3cef'720bU, 1U, 0U, 0U}, |
67 | // x = 0x1.75ea14p-6, exp10m1f(x) = 0x1.b9ff16p-5 (RZ) |
68 | {0x3cba'f50aU, 0x3d5c'ff8bU, 1U, 0U, 0U}, |
69 | // x = 0x1.f81b64p-6, exp10m1f(x) = 0x1.2cb6bcp-4 (RZ) |
70 | {0x3cfc'0db2U, 0x3d96'5b5eU, 1U, 0U, 0U}, |
71 | // x = 0x1.fafecp+3, exp10m1f(x) = 0x1.8c880ap+52 (RZ) |
72 | {0x417d'7f60U, 0x59c6'4405U, 1U, 0U, 0U}, |
73 | // x = -0x1.3bf094p-8, exp10m1f(x) = -0x1.69ba4ap-7 (RZ) |
74 | {0xbb9d'f84aU, 0xbc34'dd25U, 0U, 1U, 0U}, |
75 | // x = -0x1.4558bcp-8, exp10m1f(x) = -0x1.746fb8p-7 (RZ) |
76 | {0xbba2'ac5eU, 0xbc3a'37dcU, 0U, 1U, 1U}, |
77 | // x = -0x1.4bb43p-8, exp10m1f(x) = -0x1.7babe4p-7 (RZ) |
78 | {0xbba5'da18U, 0xbc3d'd5f2U, 0U, 1U, 1U}, |
79 | // x = -0x1.776cc8p-8, exp10m1f(x) = -0x1.ad62c4p-7 (RZ) |
80 | {0xbbbb'b664U, 0xbc56'b162U, 0U, 1U, 0U}, |
81 | // x = -0x1.f024cp-8, exp10m1f(x) = -0x1.1b20d6p-6 (RZ) |
82 | {0xbbf8'1260U, 0xbc8d'906bU, 0U, 1U, 1U}, |
83 | // x = -0x1.f510eep-8, exp10m1f(x) = -0x1.1de9aap-6 (RZ) |
84 | {0xbbfa'8877U, 0xbc8e'f4d5U, 0U, 1U, 0U}, |
85 | // x = -0x1.0b43c4p-7, exp10m1f(x) = -0x1.30d418p-6 (RZ) |
86 | {0xbc05'a1e2U, 0xbc98'6a0cU, 0U, 1U, 0U}, |
87 | // x = -0x1.245ee4p-7, exp10m1f(x) = -0x1.4d2b86p-6 (RZ) |
88 | {0xbc12'2f72U, 0xbca6'95c3U, 0U, 1U, 0U}, |
89 | // x = -0x1.f9f2dap-7, exp10m1f(x) = -0x1.1e2186p-5 (RZ) |
90 | {0xbc7c'f96dU, 0xbd0f'10c3U, 0U, 1U, 0U}, |
91 | // x = -0x1.08e42p-6, exp10m1f(x) = -0x1.2b5c4p-5 (RZ) |
92 | {0xbc84'7210U, 0xbd15'ae20U, 0U, 1U, 1U}, |
93 | // x = -0x1.0cdc44p-5, exp10m1f(x) = -0x1.2a2152p-4 (RZ) |
94 | {0xbd06'6e22U, 0xbd95'10a9U, 0U, 1U, 1U}, |
95 | // x = -0x1.ca4322p-5, exp10m1f(x) = -0x1.ef073p-4 (RZ) |
96 | {0xbd65'2191U, 0xbdf7'8398U, 0U, 1U, 1U}, |
97 | }}; |
98 | #endif // !LIBC_MATH_HAS_SKIP_ACCURATE_PASS |
99 | |
100 | LLVM_LIBC_FUNCTION(float, exp10m1f, (float x)) { |
101 | using FPBits = fputil::FPBits<float>; |
102 | FPBits xbits(x); |
103 | |
104 | uint32_t x_u = xbits.uintval(); |
105 | uint32_t x_abs = x_u & 0x7fff'ffffU; |
106 | |
107 | // When x >= log10(2^128), or x is nan |
108 | if (LIBC_UNLIKELY(xbits.is_pos() && x_u >= 0x421a'209bU)) { |
109 | if (xbits.is_finite()) { |
110 | int rounding = fputil::quick_get_round(); |
111 | if (rounding == FE_DOWNWARD || rounding == FE_TOWARDZERO) |
112 | return FPBits::max_normal().get_val(); |
113 | |
114 | fputil::set_errno_if_required(ERANGE); |
115 | fputil::raise_except_if_required(FE_OVERFLOW); |
116 | } |
117 | |
118 | // x >= log10(2^128) and 10^x - 1 rounds to +inf, or x is +inf or nan |
119 | return x + FPBits::inf().get_val(); |
120 | } |
121 | |
122 | // When |x| <= log10(2) * 2^(-6) |
123 | if (LIBC_UNLIKELY(x_abs <= 0x3b9a'209bU)) { |
124 | #ifndef LIBC_MATH_HAS_SKIP_ACCURATE_PASS |
125 | if (auto r = EXP10M1F_EXCEPTS_LO.lookup(x_u); LIBC_UNLIKELY(r.has_value())) |
126 | return r.value(); |
127 | #endif // !LIBC_MATH_HAS_SKIP_ACCURATE_PASS |
128 | |
129 | double dx = x; |
130 | double dx_sq = dx * dx; |
131 | double c0 = dx * Exp10Base::COEFFS[0]; |
132 | double c1 = |
133 | fputil::multiply_add(dx, Exp10Base::COEFFS[2], Exp10Base::COEFFS[1]); |
134 | double c2 = |
135 | fputil::multiply_add(dx, Exp10Base::COEFFS[4], Exp10Base::COEFFS[3]); |
136 | // 10^dx - 1 ~ (1 + COEFFS[0] * dx + ... + COEFFS[4] * dx^5) - 1 |
137 | // = COEFFS[0] * dx + ... + COEFFS[4] * dx^5 |
138 | return static_cast<float>(fputil::polyeval(dx_sq, c0, c1, c2)); |
139 | } |
140 | |
141 | // When x <= log10(2^-25), or x is nan |
142 | if (LIBC_UNLIKELY(x_u >= 0xc0f0d2f1)) { |
143 | // exp10m1(-inf) = -1 |
144 | if (xbits.is_inf()) |
145 | return -1.0f; |
146 | // exp10m1(nan) = nan |
147 | if (xbits.is_nan()) |
148 | return x; |
149 | |
150 | int rounding = fputil::quick_get_round(); |
151 | if (rounding == FE_UPWARD || rounding == FE_TOWARDZERO || |
152 | (rounding == FE_TONEAREST && x_u == 0xc0f0d2f1)) |
153 | return -0x1.ffff'fep-1f; // -1.0f + 0x1.0p-24f |
154 | |
155 | fputil::set_errno_if_required(ERANGE); |
156 | fputil::raise_except_if_required(FE_UNDERFLOW); |
157 | return -1.0f; |
158 | } |
159 | |
160 | // Exact outputs when x = 1, 2, ..., 10. |
161 | // Quick check mask: 0x800f'ffffU = ~(bits of 1.0f | ... | bits of 10.0f) |
162 | if (LIBC_UNLIKELY((x_u & 0x800f'ffffU) == 0)) { |
163 | switch (x_u) { |
164 | case 0x3f800000U: // x = 1.0f |
165 | return 9.0f; |
166 | case 0x40000000U: // x = 2.0f |
167 | return 99.0f; |
168 | case 0x40400000U: // x = 3.0f |
169 | return 999.0f; |
170 | case 0x40800000U: // x = 4.0f |
171 | return 9'999.0f; |
172 | case 0x40a00000U: // x = 5.0f |
173 | return 99'999.0f; |
174 | case 0x40c00000U: // x = 6.0f |
175 | return 999'999.0f; |
176 | case 0x40e00000U: // x = 7.0f |
177 | return 9'999'999.0f; |
178 | case 0x41000000U: { // x = 8.0f |
179 | int rounding = fputil::quick_get_round(); |
180 | if (rounding == FE_UPWARD || rounding == FE_TONEAREST) |
181 | return 100'000'000.0f; |
182 | return 99'999'992.0f; |
183 | } |
184 | case 0x41100000U: { // x = 9.0f |
185 | int rounding = fputil::quick_get_round(); |
186 | if (rounding == FE_UPWARD || rounding == FE_TONEAREST) |
187 | return 1'000'000'000.0f; |
188 | return 999'999'936.0f; |
189 | } |
190 | case 0x41200000U: { // x = 10.0f |
191 | int rounding = fputil::quick_get_round(); |
192 | if (rounding == FE_UPWARD || rounding == FE_TONEAREST) |
193 | return 10'000'000'000.0f; |
194 | return 9'999'998'976.0f; |
195 | } |
196 | } |
197 | } |
198 | |
199 | #ifndef LIBC_MATH_HAS_SKIP_ACCURATE_PASS |
200 | if (auto r = EXP10M1F_EXCEPTS_HI.lookup(x_u); LIBC_UNLIKELY(r.has_value())) |
201 | return r.value(); |
202 | #endif // !LIBC_MATH_HAS_SKIP_ACCURATE_PASS |
203 | |
204 | // Range reduction: 10^x = 2^(mid + hi) * 10^lo |
205 | // rr = (2^(mid + hi), lo) |
206 | auto rr = exp_b_range_reduc<Exp10Base>(x); |
207 | |
208 | // The low part is approximated by a degree-5 minimax polynomial. |
209 | // 10^lo ~ 1 + COEFFS[0] * lo + ... + COEFFS[4] * lo^5 |
210 | double lo_sq = rr.lo * rr.lo; |
211 | double c0 = fputil::multiply_add(rr.lo, Exp10Base::COEFFS[0], 1.0); |
212 | double c1 = |
213 | fputil::multiply_add(rr.lo, Exp10Base::COEFFS[2], Exp10Base::COEFFS[1]); |
214 | double c2 = |
215 | fputil::multiply_add(rr.lo, Exp10Base::COEFFS[4], Exp10Base::COEFFS[3]); |
216 | double exp10_lo = fputil::polyeval(lo_sq, c0, c1, c2); |
217 | // 10^x - 1 = 2^(mid + hi) * 10^lo - 1 |
218 | // ~ mh * exp10_lo - 1 |
219 | return static_cast<float>(fputil::multiply_add(exp10_lo, rr.mh, -1.0)); |
220 | } |
221 | |
222 | } // namespace LIBC_NAMESPACE_DECL |
223 | |