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