| 1 |
// fast_float by Daniel Lemire |
| 2 |
// fast_float by João Paulo Magalhaes |
| 3 |
// |
| 4 |
// |
| 5 |
// with contributions from Eugene Golushkov |
| 6 |
// with contributions from Maksim Kita |
| 7 |
// with contributions from Marcin Wojdyr |
| 8 |
// with contributions from Neal Richardson |
| 9 |
// with contributions from Tim Paine |
| 10 |
// with contributions from Fabio Pellacini |
| 11 |
// with contributions from Lénárd Szolnoki |
| 12 |
// with contributions from Jan Pharago |
| 13 |
// with contributions from Maya Warrier |
| 14 |
// with contributions from Taha Khokhar |
| 15 |
// with contributions from Anders Dalvander |
| 16 |
// |
| 17 |
// |
| 18 |
// Licensed under the Apache License, Version 2.0, or the |
| 19 |
// MIT License or the Boost License. This file may not be copied, |
| 20 |
// modified, or distributed except according to those terms. |
| 21 |
// |
| 22 |
// MIT License Notice |
| 23 |
// |
| 24 |
// MIT License |
| 25 |
// |
| 26 |
// Copyright (c) 2021 The fast_float authors |
| 27 |
// |
| 28 |
// Permission is hereby granted, free of charge, to any |
| 29 |
// person obtaining a copy of this software and associated |
| 30 |
// documentation files (the "Software"), to deal in the |
| 31 |
// Software without restriction, including without |
| 32 |
// limitation the rights to use, copy, modify, merge, |
| 33 |
// publish, distribute, sublicense, and/or sell copies of |
| 34 |
// the Software, and to permit persons to whom the Software |
| 35 |
// is furnished to do so, subject to the following |
| 36 |
// conditions: |
| 37 |
// |
| 38 |
// The above copyright notice and this permission notice |
| 39 |
// shall be included in all copies or substantial portions |
| 40 |
// of the Software. |
| 41 |
// |
| 42 |
// THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF |
| 43 |
// ANY KIND, EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED |
| 44 |
// TO THE WARRANTIES OF MERCHANTABILITY, FITNESS FOR A |
| 45 |
// PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT |
| 46 |
// SHALL THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY |
| 47 |
// CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION |
| 48 |
// OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR |
| 49 |
// IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER |
| 50 |
// DEALINGS IN THE SOFTWARE. |
| 51 |
// |
| 52 |
// Apache License (Version 2.0) Notice |
| 53 |
// |
| 54 |
// Copyright 2021 The fast_float authors |
| 55 |
// Licensed under the Apache License, Version 2.0 (the "License"); |
| 56 |
// you may not use this file except in compliance with the License. |
| 57 |
// You may obtain a copy of the License at |
| 58 |
// |
| 59 |
// http://www.apache.org/licenses/LICENSE-2.0 |
| 60 |
// |
| 61 |
// Unless required by applicable law or agreed to in writing, software |
| 62 |
// distributed under the License is distributed on an "AS IS" BASIS, |
| 63 |
// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. |
| 64 |
// See the License for the specific language governing permissions and |
| 65 |
// |
| 66 |
// BOOST License Notice |
| 67 |
// |
| 68 |
// Boost Software License - Version 1.0 - August 17th, 2003 |
| 69 |
// |
| 70 |
// Permission is hereby granted, free of charge, to any person or organization |
| 71 |
// obtaining a copy of the software and accompanying documentation covered by |
| 72 |
// this license (the "Software") to use, reproduce, display, distribute, |
| 73 |
// execute, and transmit the Software, and to prepare derivative works of the |
| 74 |
// Software, and to permit third-parties to whom the Software is furnished to |
| 75 |
// do so, all subject to the following: |
| 76 |
// |
| 77 |
// The copyright notices in the Software and this entire statement, including |
| 78 |
// the above license grant, this restriction and the following disclaimer, |
| 79 |
// must be included in all copies of the Software, in whole or in part, and |
| 80 |
// all derivative works of the Software, unless such copies or derivative |
| 81 |
// works are solely in the form of machine-executable object code generated by |
| 82 |
// a source language processor. |
| 83 |
// |
| 84 |
// THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR |
| 85 |
// IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, |
| 86 |
// FITNESS FOR A PARTICULAR PURPOSE, TITLE AND NON-INFRINGEMENT. IN NO EVENT |
| 87 |
// SHALL THE COPYRIGHT HOLDERS OR ANYONE DISTRIBUTING THE SOFTWARE BE LIABLE |
| 88 |
// FOR ANY DAMAGES OR OTHER LIABILITY, WHETHER IN CONTRACT, TORT OR OTHERWISE, |
| 89 |
// ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER |
| 90 |
// DEALINGS IN THE SOFTWARE. |
| 91 |
// |
| 92 |
|
| 93 |
#ifndef FASTFLOAT_CONSTEXPR_FEATURE_DETECT_H |
| 94 |
#define FASTFLOAT_CONSTEXPR_FEATURE_DETECT_H |
| 95 |
|
| 96 |
#ifdef __has_include |
| 97 |
#if __has_include(<version>) |
| 98 |
#include <version> |
| 99 |
#endif |
| 100 |
#endif |
| 101 |
|
| 102 |
// Testing for https://wg21.link/N3652, adopted in C++14 |
| 103 |
#if defined(__cpp_constexpr) && __cpp_constexpr >= 201304 |
| 104 |
#define FASTFLOAT_CONSTEXPR14 constexpr |
| 105 |
#else |
| 106 |
#define FASTFLOAT_CONSTEXPR14 |
| 107 |
#endif |
| 108 |
|
| 109 |
#if defined(__cpp_lib_bit_cast) && __cpp_lib_bit_cast >= 201806L |
| 110 |
#define FASTFLOAT_HAS_BIT_CAST 1 |
| 111 |
#else |
| 112 |
#define FASTFLOAT_HAS_BIT_CAST 0 |
| 113 |
#endif |
| 114 |
|
| 115 |
#if defined(__cpp_lib_is_constant_evaluated) && \ |
| 116 |
__cpp_lib_is_constant_evaluated >= 201811L |
| 117 |
#define FASTFLOAT_HAS_IS_CONSTANT_EVALUATED 1 |
| 118 |
#else |
| 119 |
#define FASTFLOAT_HAS_IS_CONSTANT_EVALUATED 0 |
| 120 |
#endif |
| 121 |
|
| 122 |
#if defined(__cpp_if_constexpr) && __cpp_if_constexpr >= 201606L |
| 123 |
#define FASTFLOAT_IF_CONSTEXPR17(x) if constexpr (x) |
| 124 |
#else |
| 125 |
#define FASTFLOAT_IF_CONSTEXPR17(x) if (x) |
| 126 |
#endif |
| 127 |
|
| 128 |
// Testing for relevant C++20 constexpr library features |
| 129 |
#if FASTFLOAT_HAS_IS_CONSTANT_EVALUATED && FASTFLOAT_HAS_BIT_CAST && \ |
| 130 |
defined(__cpp_lib_constexpr_algorithms) && \ |
| 131 |
__cpp_lib_constexpr_algorithms >= 201806L /*For std::copy and std::fill*/ |
| 132 |
#define FASTFLOAT_CONSTEXPR20 constexpr |
| 133 |
#define FASTFLOAT_IS_CONSTEXPR 1 |
| 134 |
#else |
| 135 |
#define FASTFLOAT_CONSTEXPR20 |
| 136 |
#define FASTFLOAT_IS_CONSTEXPR 0 |
| 137 |
#endif |
| 138 |
|
| 139 |
#if __cplusplus >= 201703L || (defined(_MSVC_LANG) && _MSVC_LANG >= 201703L) |
| 140 |
#define FASTFLOAT_DETAIL_MUST_DEFINE_CONSTEXPR_VARIABLE 0 |
| 141 |
#else |
| 142 |
#define FASTFLOAT_DETAIL_MUST_DEFINE_CONSTEXPR_VARIABLE 1 |
| 143 |
#endif |
| 144 |
|
| 145 |
#endif // FASTFLOAT_CONSTEXPR_FEATURE_DETECT_H |
| 146 |
|
| 147 |
#ifndef FASTFLOAT_FLOAT_COMMON_H |
| 148 |
#define FASTFLOAT_FLOAT_COMMON_H |
| 149 |
|
| 150 |
#include <cfloat> |
| 151 |
#include <cstddef> |
| 152 |
#include <cstdint> |
| 153 |
#include <cassert> |
| 154 |
#include <cstring> |
| 155 |
#include <limits> |
| 156 |
#include <type_traits> |
| 157 |
#include <system_error> |
| 158 |
#ifdef __has_include |
| 159 |
#if __has_include(<stdfloat>) && (__cplusplus > 202002L || (defined(_MSVC_LANG) && (_MSVC_LANG > 202002L))) |
| 160 |
#include <stdfloat> |
| 161 |
#endif |
| 162 |
#endif |
| 163 |
|
| 164 |
#define FASTFLOAT_VERSION_MAJOR 8 |
| 165 |
#define FASTFLOAT_VERSION_MINOR 2 |
| 166 |
#define FASTFLOAT_VERSION_PATCH 4 |
| 167 |
|
| 168 |
#define FASTFLOAT_STRINGIZE_IMPL(x) #x |
| 169 |
#define FASTFLOAT_STRINGIZE(x) FASTFLOAT_STRINGIZE_IMPL(x) |
| 170 |
|
| 171 |
#define FASTFLOAT_VERSION_STR \ |
| 172 |
FASTFLOAT_STRINGIZE(FASTFLOAT_VERSION_MAJOR) \ |
| 173 |
"." FASTFLOAT_STRINGIZE(FASTFLOAT_VERSION_MINOR) "." FASTFLOAT_STRINGIZE( \ |
| 174 |
FASTFLOAT_VERSION_PATCH) |
| 175 |
|
| 176 |
#define FASTFLOAT_VERSION \ |
| 177 |
(FASTFLOAT_VERSION_MAJOR * 10000 + FASTFLOAT_VERSION_MINOR * 100 + \ |
| 178 |
FASTFLOAT_VERSION_PATCH) |
| 179 |
|
| 180 |
namespace fast_float { |
| 181 |
|
| 182 |
enum class chars_format : uint64_t; |
| 183 |
|
| 184 |
namespace detail { |
| 185 |
constexpr chars_format basic_json_fmt = chars_format(1 << 5); |
| 186 |
constexpr chars_format basic_fortran_fmt = chars_format(1 << 6); |
| 187 |
} // namespace detail |
| 188 |
|
| 189 |
enum class chars_format : uint64_t { |
| 190 |
scientific = 1 << 0, |
| 191 |
fixed = 1 << 2, |
| 192 |
hex = 1 << 3, |
| 193 |
no_infnan = 1 << 4, |
| 194 |
// RFC 8259: https://datatracker.ietf.org/doc/html/rfc8259#section-6 |
| 195 |
json = uint64_t(detail::basic_json_fmt) | fixed | scientific | no_infnan, |
| 196 |
// Extension of RFC 8259 where, e.g., "inf" and "nan" are allowed. |
| 197 |
json_or_infnan = uint64_t(detail::basic_json_fmt) | fixed | scientific, |
| 198 |
fortran = uint64_t(detail::basic_fortran_fmt) | fixed | scientific, |
| 199 |
general = fixed | scientific, |
| 200 |
allow_leading_plus = 1 << 7, |
| 201 |
skip_white_space = 1 << 8, |
| 202 |
}; |
| 203 |
|
| 204 |
template <typename UC> struct from_chars_result_t { |
| 205 |
UC const *ptr; |
| 206 |
std::errc ec; |
| 207 |
|
| 208 |
// https://www.open-std.org/jtc1/sc22/wg21/docs/papers/2023/p2497r0.html |
| 209 |
constexpr explicit operator bool() const noexcept { |
| 210 |
return ec == std::errc(); |
| 211 |
} |
| 212 |
}; |
| 213 |
|
| 214 |
using from_chars_result = from_chars_result_t<char>; |
| 215 |
|
| 216 |
template <typename UC> struct parse_options_t { |
| 217 |
constexpr explicit parse_options_t(chars_format fmt = chars_format::general, |
| 218 |
UC dot = UC('.'), int b = 10) |
| 219 |
: format(fmt), decimal_point(dot), base(b) {} |
| 220 |
|
| 221 |
/** Which number formats are accepted */ |
| 222 |
chars_format format; |
| 223 |
/** The character used as decimal point */ |
| 224 |
UC decimal_point; |
| 225 |
/** The base used for integers */ |
| 226 |
int base; |
| 227 |
}; |
| 228 |
|
| 229 |
using parse_options = parse_options_t<char>; |
| 230 |
|
| 231 |
} // namespace fast_float |
| 232 |
|
| 233 |
#if FASTFLOAT_HAS_BIT_CAST |
| 234 |
#include <bit> |
| 235 |
#endif |
| 236 |
|
| 237 |
#if (defined(__x86_64) || defined(__x86_64__) || defined(_M_X64) || \ |
| 238 |
defined(__amd64) || defined(__aarch64__) || defined(_M_ARM64) || \ |
| 239 |
defined(__MINGW64__) || defined(__s390x__) || \ |
| 240 |
(defined(__ppc64__) || defined(__PPC64__) || defined(__ppc64le__) || \ |
| 241 |
defined(__PPC64LE__)) || \ |
| 242 |
defined(__loongarch64) || (defined(__riscv) && __riscv_xlen == 64)) |
| 243 |
#define FASTFLOAT_64BIT 1 |
| 244 |
#elif (defined(__i386) || defined(__i386__) || defined(_M_IX86) || \ |
| 245 |
defined(__arm__) || defined(_M_ARM) || defined(__ppc__) || \ |
| 246 |
defined(__MINGW32__) || defined(__EMSCRIPTEN__) || \ |
| 247 |
(defined(__riscv) && __riscv_xlen == 32)) |
| 248 |
#define FASTFLOAT_32BIT 1 |
| 249 |
#else |
| 250 |
// Need to check incrementally, since SIZE_MAX is a size_t, avoid overflow. |
| 251 |
// We can never tell the register width, but the SIZE_MAX is a good |
| 252 |
// approximation. UINTPTR_MAX and INTPTR_MAX are optional, so avoid them for max |
| 253 |
// portability. |
| 254 |
#if SIZE_MAX == 0xffff |
| 255 |
#error Unknown platform (16-bit, unsupported) |
| 256 |
#elif SIZE_MAX == 0xffffffff |
| 257 |
#define FASTFLOAT_32BIT 1 |
| 258 |
#elif SIZE_MAX == 0xffffffffffffffff |
| 259 |
#define FASTFLOAT_64BIT 1 |
| 260 |
#else |
| 261 |
#error Unknown platform (not 32-bit, not 64-bit?) |
| 262 |
#endif |
| 263 |
#endif |
| 264 |
|
| 265 |
#if ((defined(_WIN32) || defined(_WIN64)) && !defined(__clang__)) || \ |
| 266 |
(defined(_M_ARM64) && !defined(__MINGW32__)) |
| 267 |
#include <intrin.h> |
| 268 |
#endif |
| 269 |
|
| 270 |
#if defined(_MSC_VER) && !defined(__clang__) |
| 271 |
#define FASTFLOAT_VISUAL_STUDIO 1 |
| 272 |
#endif |
| 273 |
|
| 274 |
#if defined __BYTE_ORDER__ && defined __ORDER_BIG_ENDIAN__ |
| 275 |
#define FASTFLOAT_IS_BIG_ENDIAN (__BYTE_ORDER__ == __ORDER_BIG_ENDIAN__) |
| 276 |
#elif defined _WIN32 |
| 277 |
#define FASTFLOAT_IS_BIG_ENDIAN 0 |
| 278 |
#else |
| 279 |
#if defined(__APPLE__) || defined(__FreeBSD__) |
| 280 |
#include <machine/endian.h> |
| 281 |
#elif defined(sun) || defined(__sun) |
| 282 |
#include <sys/byteorder.h> |
| 283 |
#elif defined(__MVS__) |
| 284 |
#include <sys/endian.h> |
| 285 |
#else |
| 286 |
#ifdef __has_include |
| 287 |
#if __has_include(<endian.h>) |
| 288 |
#include <endian.h> |
| 289 |
#endif //__has_include(<endian.h>) |
| 290 |
#endif //__has_include |
| 291 |
#endif |
| 292 |
# |
| 293 |
#ifndef __BYTE_ORDER__ |
| 294 |
// safe choice |
| 295 |
#define FASTFLOAT_IS_BIG_ENDIAN 0 |
| 296 |
#endif |
| 297 |
# |
| 298 |
#ifndef __ORDER_LITTLE_ENDIAN__ |
| 299 |
// safe choice |
| 300 |
#define FASTFLOAT_IS_BIG_ENDIAN 0 |
| 301 |
#endif |
| 302 |
# |
| 303 |
#if __BYTE_ORDER__ == __ORDER_LITTLE_ENDIAN__ |
| 304 |
#define FASTFLOAT_IS_BIG_ENDIAN 0 |
| 305 |
#else |
| 306 |
#define FASTFLOAT_IS_BIG_ENDIAN 1 |
| 307 |
#endif |
| 308 |
#endif |
| 309 |
|
| 310 |
#if defined(__SSE2__) || (defined(FASTFLOAT_VISUAL_STUDIO) && \ |
| 311 |
(defined(_M_AMD64) || defined(_M_X64) || \ |
| 312 |
(defined(_M_IX86_FP) && _M_IX86_FP == 2))) |
| 313 |
#define FASTFLOAT_SSE2 1 |
| 314 |
#endif |
| 315 |
|
| 316 |
#if defined(__aarch64__) || defined(_M_ARM64) |
| 317 |
#define FASTFLOAT_NEON 1 |
| 318 |
#endif |
| 319 |
|
| 320 |
#if defined(FASTFLOAT_SSE2) || defined(FASTFLOAT_NEON) |
| 321 |
#define FASTFLOAT_HAS_SIMD 1 |
| 322 |
#endif |
| 323 |
|
| 324 |
#if defined(__GNUC__) |
| 325 |
// disable -Wcast-align=strict (GCC only) |
| 326 |
#define FASTFLOAT_SIMD_DISABLE_WARNINGS \ |
| 327 |
_Pragma("GCC diagnostic push") \ |
| 328 |
_Pragma("GCC diagnostic ignored \"-Wcast-align\"") |
| 329 |
#else |
| 330 |
#define FASTFLOAT_SIMD_DISABLE_WARNINGS |
| 331 |
#endif |
| 332 |
|
| 333 |
#if defined(__GNUC__) |
| 334 |
#define FASTFLOAT_SIMD_RESTORE_WARNINGS _Pragma("GCC diagnostic pop") |
| 335 |
#else |
| 336 |
#define FASTFLOAT_SIMD_RESTORE_WARNINGS |
| 337 |
#endif |
| 338 |
|
| 339 |
#ifdef FASTFLOAT_VISUAL_STUDIO |
| 340 |
#define fastfloat_really_inline __forceinline |
| 341 |
#else |
| 342 |
#define fastfloat_really_inline inline __attribute__((always_inline)) |
| 343 |
#endif |
| 344 |
|
| 345 |
#ifndef FASTFLOAT_ASSERT |
| 346 |
#define FASTFLOAT_ASSERT(x) \ |
| 347 |
{ ((void)(x)); } |
| 348 |
#endif |
| 349 |
|
| 350 |
#ifndef FASTFLOAT_DEBUG_ASSERT |
| 351 |
#define FASTFLOAT_DEBUG_ASSERT(x) \ |
| 352 |
{ ((void)(x)); } |
| 353 |
#endif |
| 354 |
|
| 355 |
// rust style `try!()` macro, or `?` operator |
| 356 |
#define FASTFLOAT_TRY(x) \ |
| 357 |
{ \ |
| 358 |
if (!(x)) \ |
| 359 |
return false; \ |
| 360 |
} |
| 361 |
|
| 362 |
#define FASTFLOAT_ENABLE_IF(...) \ |
| 363 |
typename std::enable_if<(__VA_ARGS__), int>::type |
| 364 |
|
| 365 |
namespace fast_float { |
| 366 |
|
| 367 |
fastfloat_really_inline constexpr bool cpp20_and_in_constexpr() { |
| 368 |
#if FASTFLOAT_HAS_IS_CONSTANT_EVALUATED |
| 369 |
return std::is_constant_evaluated(); |
| 370 |
#else |
| 371 |
return false; |
| 372 |
#endif |
| 373 |
} |
| 374 |
|
| 375 |
template <typename T> |
| 376 |
struct is_supported_float_type |
| 377 |
: std::integral_constant< |
| 378 |
bool, std::is_same<T, double>::value || std::is_same<T, float>::value |
| 379 |
#ifdef __STDCPP_FLOAT64_T__ |
| 380 |
|| std::is_same<T, std::float64_t>::value |
| 381 |
#endif |
| 382 |
#ifdef __STDCPP_FLOAT32_T__ |
| 383 |
|| std::is_same<T, std::float32_t>::value |
| 384 |
#endif |
| 385 |
#ifdef __STDCPP_FLOAT16_T__ |
| 386 |
|| std::is_same<T, std::float16_t>::value |
| 387 |
#endif |
| 388 |
#ifdef __STDCPP_BFLOAT16_T__ |
| 389 |
|| std::is_same<T, std::bfloat16_t>::value |
| 390 |
#endif |
| 391 |
> { |
| 392 |
}; |
| 393 |
|
| 394 |
template <typename T> |
| 395 |
using equiv_uint_t = typename std::conditional< |
| 396 |
sizeof(T) == 1, uint8_t, |
| 397 |
typename std::conditional< |
| 398 |
sizeof(T) == 2, uint16_t, |
| 399 |
typename std::conditional<sizeof(T) == 4, uint32_t, |
| 400 |
uint64_t>::type>::type>::type; |
| 401 |
|
| 402 |
template <typename T> struct is_supported_integer_type : std::is_integral<T> {}; |
| 403 |
|
| 404 |
template <typename UC> |
| 405 |
struct is_supported_char_type |
| 406 |
: std::integral_constant<bool, std::is_same<UC, char>::value || |
| 407 |
std::is_same<UC, wchar_t>::value || |
| 408 |
std::is_same<UC, char16_t>::value || |
| 409 |
std::is_same<UC, char32_t>::value |
| 410 |
#ifdef __cpp_char8_t |
| 411 |
|| std::is_same<UC, char8_t>::value |
| 412 |
#endif |
| 413 |
> { |
| 414 |
}; |
| 415 |
|
| 416 |
template <typename UC> |
| 417 |
inline FASTFLOAT_CONSTEXPR14 bool |
| 418 |
fastfloat_strncasecmp3(UC const *actual_mixedcase, |
| 419 |
UC const *expected_lowercase) { |
| 420 |
uint64_t mask{0}; |
| 421 |
FASTFLOAT_IF_CONSTEXPR17(sizeof(UC) == 1) { mask = 0x2020202020202020; } |
| 422 |
else FASTFLOAT_IF_CONSTEXPR17(sizeof(UC) == 2) { |
| 423 |
mask = 0x0020002000200020; |
| 424 |
} |
| 425 |
else FASTFLOAT_IF_CONSTEXPR17(sizeof(UC) == 4) { |
| 426 |
mask = 0x0000002000000020; |
| 427 |
} |
| 428 |
else { |
| 429 |
return false; |
| 430 |
} |
| 431 |
|
| 432 |
uint64_t val1{0}, val2{0}; |
| 433 |
if (cpp20_and_in_constexpr()) { |
| 434 |
for (size_t i = 0; i < 3; i++) { |
| 435 |
if ((actual_mixedcase[i] | 32) != expected_lowercase[i]) { |
| 436 |
return false; |
| 437 |
} |
| 438 |
} |
| 439 |
return true; |
| 440 |
} else { |
| 441 |
FASTFLOAT_IF_CONSTEXPR17(sizeof(UC) == 1 || sizeof(UC) == 2) { |
| 442 |
::memcpy(&val1, actual_mixedcase, 3 * sizeof(UC)); |
| 443 |
::memcpy(&val2, expected_lowercase, 3 * sizeof(UC)); |
| 444 |
val1 |= mask; |
| 445 |
val2 |= mask; |
| 446 |
return val1 == val2; |
| 447 |
} |
| 448 |
else FASTFLOAT_IF_CONSTEXPR17(sizeof(UC) == 4) { |
| 449 |
::memcpy(&val1, actual_mixedcase, 2 * sizeof(UC)); |
| 450 |
::memcpy(&val2, expected_lowercase, 2 * sizeof(UC)); |
| 451 |
val1 |= mask; |
| 452 |
if (val1 != val2) { |
| 453 |
return false; |
| 454 |
} |
| 455 |
return (actual_mixedcase[2] | 32) == (expected_lowercase[2]); |
| 456 |
} |
| 457 |
else { |
| 458 |
return false; |
| 459 |
} |
| 460 |
} |
| 461 |
} |
| 462 |
|
| 463 |
template <typename UC> |
| 464 |
inline FASTFLOAT_CONSTEXPR14 bool |
| 465 |
fastfloat_strncasecmp5(UC const *actual_mixedcase, |
| 466 |
UC const *expected_lowercase) { |
| 467 |
uint64_t mask{0}; |
| 468 |
uint64_t val1{0}, val2{0}; |
| 469 |
if (cpp20_and_in_constexpr()) { |
| 470 |
for (size_t i = 0; i < 5; i++) { |
| 471 |
if ((actual_mixedcase[i] | 32) != expected_lowercase[i]) { |
| 472 |
return false; |
| 473 |
} |
| 474 |
} |
| 475 |
return true; |
| 476 |
} else { |
| 477 |
FASTFLOAT_IF_CONSTEXPR17(sizeof(UC) == 1) { |
| 478 |
mask = 0x2020202020202020; |
| 479 |
::memcpy(&val1, actual_mixedcase, 5 * sizeof(UC)); |
| 480 |
::memcpy(&val2, expected_lowercase, 5 * sizeof(UC)); |
| 481 |
val1 |= mask; |
| 482 |
val2 |= mask; |
| 483 |
return val1 == val2; |
| 484 |
} |
| 485 |
else FASTFLOAT_IF_CONSTEXPR17(sizeof(UC) == 2) { |
| 486 |
mask = 0x0020002000200020; |
| 487 |
::memcpy(&val1, actual_mixedcase, 4 * sizeof(UC)); |
| 488 |
::memcpy(&val2, expected_lowercase, 4 * sizeof(UC)); |
| 489 |
val1 |= mask; |
| 490 |
if (val1 != val2) { |
| 491 |
return false; |
| 492 |
} |
| 493 |
return (actual_mixedcase[4] | 32) == (expected_lowercase[4]); |
| 494 |
} |
| 495 |
else FASTFLOAT_IF_CONSTEXPR17(sizeof(UC) == 4) { |
| 496 |
mask = 0x0000002000000020; |
| 497 |
::memcpy(&val1, actual_mixedcase, 2 * sizeof(UC)); |
| 498 |
::memcpy(&val2, expected_lowercase, 2 * sizeof(UC)); |
| 499 |
val1 |= mask; |
| 500 |
if (val1 != val2) { |
| 501 |
return false; |
| 502 |
} |
| 503 |
::memcpy(&val1, actual_mixedcase + 2, 2 * sizeof(UC)); |
| 504 |
::memcpy(&val2, expected_lowercase + 2, 2 * sizeof(UC)); |
| 505 |
val1 |= mask; |
| 506 |
if (val1 != val2) { |
| 507 |
return false; |
| 508 |
} |
| 509 |
return (actual_mixedcase[4] | 32) == (expected_lowercase[4]); |
| 510 |
} |
| 511 |
else { |
| 512 |
return false; |
| 513 |
} |
| 514 |
} |
| 515 |
} |
| 516 |
|
| 517 |
// Compares two ASCII strings in a case insensitive manner. |
| 518 |
template <typename UC> |
| 519 |
inline FASTFLOAT_CONSTEXPR14 bool |
| 520 |
fastfloat_strncasecmp(UC const *actual_mixedcase, UC const *expected_lowercase, |
| 521 |
size_t length) { |
| 522 |
uint64_t mask{0}; |
| 523 |
FASTFLOAT_IF_CONSTEXPR17(sizeof(UC) == 1) { mask = 0x2020202020202020; } |
| 524 |
else FASTFLOAT_IF_CONSTEXPR17(sizeof(UC) == 2) { |
| 525 |
mask = 0x0020002000200020; |
| 526 |
} |
| 527 |
else FASTFLOAT_IF_CONSTEXPR17(sizeof(UC) == 4) { |
| 528 |
mask = 0x0000002000000020; |
| 529 |
} |
| 530 |
else { |
| 531 |
return false; |
| 532 |
} |
| 533 |
|
| 534 |
if (cpp20_and_in_constexpr()) { |
| 535 |
for (size_t i = 0; i < length; i++) { |
| 536 |
if ((actual_mixedcase[i] | 32) != expected_lowercase[i]) { |
| 537 |
return false; |
| 538 |
} |
| 539 |
} |
| 540 |
return true; |
| 541 |
} else { |
| 542 |
uint64_t val1{0}, val2{0}; |
| 543 |
size_t sz{8 / (sizeof(UC))}; |
| 544 |
for (size_t i = 0; i < length; i += sz) { |
| 545 |
val1 = val2 = 0; |
| 546 |
sz = std::min(sz, length - i); |
| 547 |
::memcpy(&val1, actual_mixedcase + i, sz * sizeof(UC)); |
| 548 |
::memcpy(&val2, expected_lowercase + i, sz * sizeof(UC)); |
| 549 |
val1 |= mask; |
| 550 |
val2 |= mask; |
| 551 |
if (val1 != val2) { |
| 552 |
return false; |
| 553 |
} |
| 554 |
} |
| 555 |
return true; |
| 556 |
} |
| 557 |
} |
| 558 |
|
| 559 |
#ifndef FLT_EVAL_METHOD |
| 560 |
#error "FLT_EVAL_METHOD should be defined, please include cfloat." |
| 561 |
#endif |
| 562 |
|
| 563 |
// a pointer and a length to a contiguous block of memory |
| 564 |
template <typename T> struct span { |
| 565 |
T const *ptr; |
| 566 |
size_t length; |
| 567 |
|
| 568 |
constexpr span(T const *_ptr, size_t _length) : ptr(_ptr), length(_length) {} |
| 569 |
|
| 570 |
constexpr span() : ptr(nullptr), length(0) {} |
| 571 |
|
| 572 |
constexpr size_t len() const noexcept { return length; } |
| 573 |
|
| 574 |
FASTFLOAT_CONSTEXPR14 const T &operator[](size_t index) const noexcept { |
| 575 |
FASTFLOAT_DEBUG_ASSERT(index < length); |
| 576 |
return ptr[index]; |
| 577 |
} |
| 578 |
}; |
| 579 |
|
| 580 |
struct value128 { |
| 581 |
uint64_t low; |
| 582 |
uint64_t high; |
| 583 |
|
| 584 |
constexpr value128(uint64_t _low, uint64_t _high) : low(_low), high(_high) {} |
| 585 |
|
| 586 |
constexpr value128() : low(0), high(0) {} |
| 587 |
}; |
| 588 |
|
| 589 |
/* Helper C++14 constexpr generic implementation of leading_zeroes */ |
| 590 |
fastfloat_really_inline FASTFLOAT_CONSTEXPR14 int |
| 591 |
leading_zeroes_generic(uint64_t input_num, int last_bit = 0) { |
| 592 |
if (input_num & uint64_t(0xffffffff00000000)) { |
| 593 |
input_num >>= 32; |
| 594 |
last_bit |= 32; |
| 595 |
} |
| 596 |
if (input_num & uint64_t(0xffff0000)) { |
| 597 |
input_num >>= 16; |
| 598 |
last_bit |= 16; |
| 599 |
} |
| 600 |
if (input_num & uint64_t(0xff00)) { |
| 601 |
input_num >>= 8; |
| 602 |
last_bit |= 8; |
| 603 |
} |
| 604 |
if (input_num & uint64_t(0xf0)) { |
| 605 |
input_num >>= 4; |
| 606 |
last_bit |= 4; |
| 607 |
} |
| 608 |
if (input_num & uint64_t(0xc)) { |
| 609 |
input_num >>= 2; |
| 610 |
last_bit |= 2; |
| 611 |
} |
| 612 |
if (input_num & uint64_t(0x2)) { /* input_num >>= 1; */ |
| 613 |
last_bit |= 1; |
| 614 |
} |
| 615 |
return 63 - last_bit; |
| 616 |
} |
| 617 |
|
| 618 |
/* result might be undefined when input_num is zero */ |
| 619 |
fastfloat_really_inline FASTFLOAT_CONSTEXPR20 int |
| 620 |
leading_zeroes(uint64_t input_num) { |
| 621 |
assert(input_num > 0); |
| 622 |
if (cpp20_and_in_constexpr()) { |
| 623 |
return leading_zeroes_generic(input_num); |
| 624 |
} |
| 625 |
#ifdef FASTFLOAT_VISUAL_STUDIO |
| 626 |
#if defined(_M_X64) || defined(_M_ARM64) |
| 627 |
unsigned long leading_zero = 0; |
| 628 |
// Search the mask data from most significant bit (MSB) |
| 629 |
// to least significant bit (LSB) for a set bit (1). |
| 630 |
_BitScanReverse64(&leading_zero, input_num); |
| 631 |
return (int)(63 - leading_zero); |
| 632 |
#else |
| 633 |
return leading_zeroes_generic(input_num); |
| 634 |
#endif |
| 635 |
#else |
| 636 |
return __builtin_clzll(input_num); |
| 637 |
#endif |
| 638 |
} |
| 639 |
|
| 640 |
/* Helper C++14 constexpr generic implementation of countr_zero for 32-bit */ |
| 641 |
fastfloat_really_inline FASTFLOAT_CONSTEXPR14 int |
| 642 |
countr_zero_generic_32(uint32_t input_num) { |
| 643 |
if (input_num == 0) { |
| 644 |
return 32; |
| 645 |
} |
| 646 |
int last_bit = 0; |
| 647 |
if (!(input_num & 0x0000FFFF)) { |
| 648 |
input_num >>= 16; |
| 649 |
last_bit |= 16; |
| 650 |
} |
| 651 |
if (!(input_num & 0x00FF)) { |
| 652 |
input_num >>= 8; |
| 653 |
last_bit |= 8; |
| 654 |
} |
| 655 |
if (!(input_num & 0x0F)) { |
| 656 |
input_num >>= 4; |
| 657 |
last_bit |= 4; |
| 658 |
} |
| 659 |
if (!(input_num & 0x3)) { |
| 660 |
input_num >>= 2; |
| 661 |
last_bit |= 2; |
| 662 |
} |
| 663 |
if (!(input_num & 0x1)) { |
| 664 |
last_bit |= 1; |
| 665 |
} |
| 666 |
return last_bit; |
| 667 |
} |
| 668 |
|
| 669 |
/* count trailing zeroes for 32-bit integers */ |
| 670 |
fastfloat_really_inline FASTFLOAT_CONSTEXPR20 int |
| 671 |
countr_zero_32(uint32_t input_num) { |
| 672 |
if (cpp20_and_in_constexpr()) { |
| 673 |
return countr_zero_generic_32(input_num); |
| 674 |
} |
| 675 |
#ifdef FASTFLOAT_VISUAL_STUDIO |
| 676 |
unsigned long trailing_zero = 0; |
| 677 |
if (_BitScanForward(&trailing_zero, input_num)) { |
| 678 |
return (int)trailing_zero; |
| 679 |
} |
| 680 |
return 32; |
| 681 |
#else |
| 682 |
return input_num == 0 ? 32 : __builtin_ctz(input_num); |
| 683 |
#endif |
| 684 |
} |
| 685 |
|
| 686 |
// slow emulation routine for 32-bit |
| 687 |
fastfloat_really_inline constexpr uint64_t emulu(uint32_t x, uint32_t y) { |
| 688 |
return x * (uint64_t)y; |
| 689 |
} |
| 690 |
|
| 691 |
fastfloat_really_inline FASTFLOAT_CONSTEXPR14 uint64_t |
| 692 |
umul128_generic(uint64_t ab, uint64_t cd, uint64_t *hi) { |
| 693 |
uint64_t ad = emulu((uint32_t)(ab >> 32), (uint32_t)cd); |
| 694 |
uint64_t bd = emulu((uint32_t)ab, (uint32_t)cd); |
| 695 |
uint64_t adbc = ad + emulu((uint32_t)ab, (uint32_t)(cd >> 32)); |
| 696 |
uint64_t adbc_carry = (uint64_t)(adbc < ad); |
| 697 |
uint64_t lo = bd + (adbc << 32); |
| 698 |
*hi = emulu((uint32_t)(ab >> 32), (uint32_t)(cd >> 32)) + (adbc >> 32) + |
| 699 |
(adbc_carry << 32) + (uint64_t)(lo < bd); |
| 700 |
return lo; |
| 701 |
} |
| 702 |
|
| 703 |
#ifdef FASTFLOAT_32BIT |
| 704 |
|
| 705 |
// slow emulation routine for 32-bit |
| 706 |
#if !defined(__MINGW64__) |
| 707 |
fastfloat_really_inline FASTFLOAT_CONSTEXPR14 uint64_t _umul128(uint64_t ab, |
| 708 |
uint64_t cd, |
| 709 |
uint64_t *hi) { |
| 710 |
return umul128_generic(ab, cd, hi); |
| 711 |
} |
| 712 |
#endif // !__MINGW64__ |
| 713 |
|
| 714 |
#endif // FASTFLOAT_32BIT |
| 715 |
|
| 716 |
// compute 64-bit a*b |
| 717 |
fastfloat_really_inline FASTFLOAT_CONSTEXPR20 value128 |
| 718 |
full_multiplication(uint64_t a, uint64_t b) { |
| 719 |
if (cpp20_and_in_constexpr()) { |
| 720 |
value128 answer; |
| 721 |
answer.low = umul128_generic(a, b, &answer.high); |
| 722 |
return answer; |
| 723 |
} |
| 724 |
value128 answer; |
| 725 |
#if defined(_M_ARM64) && !defined(__MINGW32__) |
| 726 |
// ARM64 has native support for 64-bit multiplications, no need to emulate |
| 727 |
// But MinGW on ARM64 doesn't have native support for 64-bit multiplications |
| 728 |
answer.high = __umulh(a, b); |
| 729 |
answer.low = a * b; |
| 730 |
#elif defined(FASTFLOAT_32BIT) || (defined(_WIN64) && !defined(__clang__) && \ |
| 731 |
!defined(_M_ARM64) && !defined(__GNUC__)) |
| 732 |
answer.low = _umul128(a, b, &answer.high); // _umul128 not available on ARM64 |
| 733 |
#elif defined(FASTFLOAT_64BIT) && defined(__SIZEOF_INT128__) |
| 734 |
__uint128_t r = ((__uint128_t)a) * b; |
| 735 |
answer.low = uint64_t(r); |
| 736 |
answer.high = uint64_t(r >> 64); |
| 737 |
#else |
| 738 |
answer.low = umul128_generic(a, b, &answer.high); |
| 739 |
#endif |
| 740 |
return answer; |
| 741 |
} |
| 742 |
|
| 743 |
struct adjusted_mantissa { |
| 744 |
uint64_t mantissa{0}; |
| 745 |
int32_t power2{0}; // a negative value indicates an invalid result |
| 746 |
adjusted_mantissa() = default; |
| 747 |
|
| 748 |
constexpr bool operator==(adjusted_mantissa const &o) const { |
| 749 |
return mantissa == o.mantissa && power2 == o.power2; |
| 750 |
} |
| 751 |
|
| 752 |
constexpr bool operator!=(adjusted_mantissa const &o) const { |
| 753 |
return mantissa != o.mantissa || power2 != o.power2; |
| 754 |
} |
| 755 |
}; |
| 756 |
|
| 757 |
// Bias so we can get the real exponent with an invalid adjusted_mantissa. |
| 758 |
constexpr static int32_t invalid_am_bias = -0x8000; |
| 759 |
|
| 760 |
// used for binary_format_lookup_tables<T>::max_mantissa |
| 761 |
constexpr uint64_t constant_55555 = 5 * 5 * 5 * 5 * 5; |
| 762 |
|
| 763 |
template <typename T, typename U = void> struct binary_format_lookup_tables; |
| 764 |
|
| 765 |
template <typename T> struct binary_format : binary_format_lookup_tables<T> { |
| 766 |
using equiv_uint = equiv_uint_t<T>; |
| 767 |
|
| 768 |
static constexpr int mantissa_explicit_bits(); |
| 769 |
static constexpr int minimum_exponent(); |
| 770 |
static constexpr int infinite_power(); |
| 771 |
static constexpr int sign_index(); |
| 772 |
static constexpr int |
| 773 |
min_exponent_fast_path(); // used when fegetround() == FE_TONEAREST |
| 774 |
static constexpr int max_exponent_fast_path(); |
| 775 |
static constexpr int max_exponent_round_to_even(); |
| 776 |
static constexpr int min_exponent_round_to_even(); |
| 777 |
static constexpr uint64_t max_mantissa_fast_path(int64_t power); |
| 778 |
static constexpr uint64_t |
| 779 |
max_mantissa_fast_path(); // used when fegetround() == FE_TONEAREST |
| 780 |
static constexpr int largest_power_of_ten(); |
| 781 |
static constexpr int smallest_power_of_ten(); |
| 782 |
static constexpr T exact_power_of_ten(int64_t power); |
| 783 |
static constexpr size_t max_digits(); |
| 784 |
static constexpr equiv_uint exponent_mask(); |
| 785 |
static constexpr equiv_uint mantissa_mask(); |
| 786 |
static constexpr equiv_uint hidden_bit_mask(); |
| 787 |
}; |
| 788 |
|
| 789 |
template <typename U> struct binary_format_lookup_tables<double, U> { |
| 790 |
static constexpr double powers_of_ten[] = { |
| 791 |
1e0, 1e1, 1e2, 1e3, 1e4, 1e5, 1e6, 1e7, 1e8, 1e9, 1e10, 1e11, |
| 792 |
1e12, 1e13, 1e14, 1e15, 1e16, 1e17, 1e18, 1e19, 1e20, 1e21, 1e22}; |
| 793 |
|
| 794 |
// Largest integer value v so that (5**index * v) <= 1<<53. |
| 795 |
// 0x20000000000000 == 1 << 53 |
| 796 |
static constexpr uint64_t max_mantissa[] = { |
| 797 |
0x20000000000000, |
| 798 |
0x20000000000000 / 5, |
| 799 |
0x20000000000000 / (5 * 5), |
| 800 |
0x20000000000000 / (5 * 5 * 5), |
| 801 |
0x20000000000000 / (5 * 5 * 5 * 5), |
| 802 |
0x20000000000000 / (constant_55555), |
| 803 |
0x20000000000000 / (constant_55555 * 5), |
| 804 |
0x20000000000000 / (constant_55555 * 5 * 5), |
| 805 |
0x20000000000000 / (constant_55555 * 5 * 5 * 5), |
| 806 |
0x20000000000000 / (constant_55555 * 5 * 5 * 5 * 5), |
| 807 |
0x20000000000000 / (constant_55555 * constant_55555), |
| 808 |
0x20000000000000 / (constant_55555 * constant_55555 * 5), |
| 809 |
0x20000000000000 / (constant_55555 * constant_55555 * 5 * 5), |
| 810 |
0x20000000000000 / (constant_55555 * constant_55555 * 5 * 5 * 5), |
| 811 |
0x20000000000000 / (constant_55555 * constant_55555 * constant_55555), |
| 812 |
0x20000000000000 / (constant_55555 * constant_55555 * constant_55555 * 5), |
| 813 |
0x20000000000000 / |
| 814 |
(constant_55555 * constant_55555 * constant_55555 * 5 * 5), |
| 815 |
0x20000000000000 / |
| 816 |
(constant_55555 * constant_55555 * constant_55555 * 5 * 5 * 5), |
| 817 |
0x20000000000000 / |
| 818 |
(constant_55555 * constant_55555 * constant_55555 * 5 * 5 * 5 * 5), |
| 819 |
0x20000000000000 / |
| 820 |
(constant_55555 * constant_55555 * constant_55555 * constant_55555), |
| 821 |
0x20000000000000 / (constant_55555 * constant_55555 * constant_55555 * |
| 822 |
constant_55555 * 5), |
| 823 |
0x20000000000000 / (constant_55555 * constant_55555 * constant_55555 * |
| 824 |
constant_55555 * 5 * 5), |
| 825 |
0x20000000000000 / (constant_55555 * constant_55555 * constant_55555 * |
| 826 |
constant_55555 * 5 * 5 * 5), |
| 827 |
0x20000000000000 / (constant_55555 * constant_55555 * constant_55555 * |
| 828 |
constant_55555 * 5 * 5 * 5 * 5)}; |
| 829 |
}; |
| 830 |
|
| 831 |
#if FASTFLOAT_DETAIL_MUST_DEFINE_CONSTEXPR_VARIABLE |
| 832 |
|
| 833 |
template <typename U> |
| 834 |
constexpr double binary_format_lookup_tables<double, U>::powers_of_ten[]; |
| 835 |
|
| 836 |
template <typename U> |
| 837 |
constexpr uint64_t binary_format_lookup_tables<double, U>::max_mantissa[]; |
| 838 |
|
| 839 |
#endif |
| 840 |
|
| 841 |
template <typename U> struct binary_format_lookup_tables<float, U> { |
| 842 |
static constexpr float powers_of_ten[] = {1e0f, 1e1f, 1e2f, 1e3f, 1e4f, 1e5f, |
| 843 |
1e6f, 1e7f, 1e8f, 1e9f, 1e10f}; |
| 844 |
|
| 845 |
// Largest integer value v so that (5**index * v) <= 1<<24. |
| 846 |
// 0x1000000 == 1<<24 |
| 847 |
static constexpr uint64_t max_mantissa[] = { |
| 848 |
0x1000000, |
| 849 |
0x1000000 / 5, |
| 850 |
0x1000000 / (5 * 5), |
| 851 |
0x1000000 / (5 * 5 * 5), |
| 852 |
0x1000000 / (5 * 5 * 5 * 5), |
| 853 |
0x1000000 / (constant_55555), |
| 854 |
0x1000000 / (constant_55555 * 5), |
| 855 |
0x1000000 / (constant_55555 * 5 * 5), |
| 856 |
0x1000000 / (constant_55555 * 5 * 5 * 5), |
| 857 |
0x1000000 / (constant_55555 * 5 * 5 * 5 * 5), |
| 858 |
0x1000000 / (constant_55555 * constant_55555), |
| 859 |
0x1000000 / (constant_55555 * constant_55555 * 5)}; |
| 860 |
}; |
| 861 |
|
| 862 |
#if FASTFLOAT_DETAIL_MUST_DEFINE_CONSTEXPR_VARIABLE |
| 863 |
|
| 864 |
template <typename U> |
| 865 |
constexpr float binary_format_lookup_tables<float, U>::powers_of_ten[]; |
| 866 |
|
| 867 |
template <typename U> |
| 868 |
constexpr uint64_t binary_format_lookup_tables<float, U>::max_mantissa[]; |
| 869 |
|
| 870 |
#endif |
| 871 |
|
| 872 |
template <> |
| 873 |
inline constexpr int binary_format<double>::min_exponent_fast_path() { |
| 874 |
#if (FLT_EVAL_METHOD != 1) && (FLT_EVAL_METHOD != 0) |
| 875 |
return 0; |
| 876 |
#else |
| 877 |
return -22; |
| 878 |
#endif |
| 879 |
} |
| 880 |
|
| 881 |
template <> |
| 882 |
inline constexpr int binary_format<float>::min_exponent_fast_path() { |
| 883 |
#if (FLT_EVAL_METHOD != 1) && (FLT_EVAL_METHOD != 0) |
| 884 |
return 0; |
| 885 |
#else |
| 886 |
return -10; |
| 887 |
#endif |
| 888 |
} |
| 889 |
|
| 890 |
template <> |
| 891 |
inline constexpr int binary_format<double>::mantissa_explicit_bits() { |
| 892 |
return 52; |
| 893 |
} |
| 894 |
|
| 895 |
template <> |
| 896 |
inline constexpr int binary_format<float>::mantissa_explicit_bits() { |
| 897 |
return 23; |
| 898 |
} |
| 899 |
|
| 900 |
template <> |
| 901 |
inline constexpr int binary_format<double>::max_exponent_round_to_even() { |
| 902 |
return 23; |
| 903 |
} |
| 904 |
|
| 905 |
template <> |
| 906 |
inline constexpr int binary_format<float>::max_exponent_round_to_even() { |
| 907 |
return 10; |
| 908 |
} |
| 909 |
|
| 910 |
template <> |
| 911 |
inline constexpr int binary_format<double>::min_exponent_round_to_even() { |
| 912 |
return -4; |
| 913 |
} |
| 914 |
|
| 915 |
template <> |
| 916 |
inline constexpr int binary_format<float>::min_exponent_round_to_even() { |
| 917 |
return -17; |
| 918 |
} |
| 919 |
|
| 920 |
template <> inline constexpr int binary_format<double>::minimum_exponent() { |
| 921 |
return -1023; |
| 922 |
} |
| 923 |
|
| 924 |
template <> inline constexpr int binary_format<float>::minimum_exponent() { |
| 925 |
return -127; |
| 926 |
} |
| 927 |
|
| 928 |
template <> inline constexpr int binary_format<double>::infinite_power() { |
| 929 |
return 0x7FF; |
| 930 |
} |
| 931 |
|
| 932 |
template <> inline constexpr int binary_format<float>::infinite_power() { |
| 933 |
return 0xFF; |
| 934 |
} |
| 935 |
|
| 936 |
template <> inline constexpr int binary_format<double>::sign_index() { |
| 937 |
return 63; |
| 938 |
} |
| 939 |
|
| 940 |
template <> inline constexpr int binary_format<float>::sign_index() { |
| 941 |
return 31; |
| 942 |
} |
| 943 |
|
| 944 |
template <> |
| 945 |
inline constexpr int binary_format<double>::max_exponent_fast_path() { |
| 946 |
return 22; |
| 947 |
} |
| 948 |
|
| 949 |
template <> |
| 950 |
inline constexpr int binary_format<float>::max_exponent_fast_path() { |
| 951 |
return 10; |
| 952 |
} |
| 953 |
|
| 954 |
template <> |
| 955 |
inline constexpr uint64_t binary_format<double>::max_mantissa_fast_path() { |
| 956 |
return uint64_t(2) << mantissa_explicit_bits(); |
| 957 |
} |
| 958 |
|
| 959 |
template <> |
| 960 |
inline constexpr uint64_t binary_format<float>::max_mantissa_fast_path() { |
| 961 |
return uint64_t(2) << mantissa_explicit_bits(); |
| 962 |
} |
| 963 |
|
| 964 |
// credit: Jakub Jelínek |
| 965 |
#ifdef __STDCPP_FLOAT16_T__ |
| 966 |
template <typename U> struct binary_format_lookup_tables<std::float16_t, U> { |
| 967 |
static constexpr std::float16_t powers_of_ten[] = {1e0f16, 1e1f16, 1e2f16, |
| 968 |
1e3f16, 1e4f16}; |
| 969 |
|
| 970 |
// Largest integer value v so that (5**index * v) <= 1<<11. |
| 971 |
// 0x800 == 1<<11 |
| 972 |
static constexpr uint64_t max_mantissa[] = {0x800, |
| 973 |
0x800 / 5, |
| 974 |
0x800 / (5 * 5), |
| 975 |
0x800 / (5 * 5 * 5), |
| 976 |
0x800 / (5 * 5 * 5 * 5), |
| 977 |
0x800 / (constant_55555)}; |
| 978 |
}; |
| 979 |
|
| 980 |
#if FASTFLOAT_DETAIL_MUST_DEFINE_CONSTEXPR_VARIABLE |
| 981 |
|
| 982 |
template <typename U> |
| 983 |
constexpr std::float16_t |
| 984 |
binary_format_lookup_tables<std::float16_t, U>::powers_of_ten[]; |
| 985 |
|
| 986 |
template <typename U> |
| 987 |
constexpr uint64_t |
| 988 |
binary_format_lookup_tables<std::float16_t, U>::max_mantissa[]; |
| 989 |
|
| 990 |
#endif |
| 991 |
|
| 992 |
template <> |
| 993 |
inline constexpr std::float16_t |
| 994 |
binary_format<std::float16_t>::exact_power_of_ten(int64_t power) { |
| 995 |
// Work around clang bug https://godbolt.org/z/zedh7rrhc |
| 996 |
return (void)powers_of_ten[0], powers_of_ten[power]; |
| 997 |
} |
| 998 |
|
| 999 |
template <> |
| 1000 |
inline constexpr binary_format<std::float16_t>::equiv_uint |
| 1001 |
binary_format<std::float16_t>::exponent_mask() { |
| 1002 |
return 0x7C00; |
| 1003 |
} |
| 1004 |
|
| 1005 |
template <> |
| 1006 |
inline constexpr binary_format<std::float16_t>::equiv_uint |
| 1007 |
binary_format<std::float16_t>::mantissa_mask() { |
| 1008 |
return 0x03FF; |
| 1009 |
} |
| 1010 |
|
| 1011 |
template <> |
| 1012 |
inline constexpr binary_format<std::float16_t>::equiv_uint |
| 1013 |
binary_format<std::float16_t>::hidden_bit_mask() { |
| 1014 |
return 0x0400; |
| 1015 |
} |
| 1016 |
|
| 1017 |
template <> |
| 1018 |
inline constexpr int binary_format<std::float16_t>::max_exponent_fast_path() { |
| 1019 |
return 4; |
| 1020 |
} |
| 1021 |
|
| 1022 |
template <> |
| 1023 |
inline constexpr int binary_format<std::float16_t>::mantissa_explicit_bits() { |
| 1024 |
return 10; |
| 1025 |
} |
| 1026 |
|
| 1027 |
template <> |
| 1028 |
inline constexpr uint64_t |
| 1029 |
binary_format<std::float16_t>::max_mantissa_fast_path() { |
| 1030 |
return uint64_t(2) << mantissa_explicit_bits(); |
| 1031 |
} |
| 1032 |
|
| 1033 |
template <> |
| 1034 |
inline constexpr uint64_t |
| 1035 |
binary_format<std::float16_t>::max_mantissa_fast_path(int64_t power) { |
| 1036 |
// caller is responsible to ensure that |
| 1037 |
// power >= 0 && power <= 4 |
| 1038 |
// |
| 1039 |
// Work around clang bug https://godbolt.org/z/zedh7rrhc |
| 1040 |
return (void)max_mantissa[0], max_mantissa[power]; |
| 1041 |
} |
| 1042 |
|
| 1043 |
template <> |
| 1044 |
inline constexpr int binary_format<std::float16_t>::min_exponent_fast_path() { |
| 1045 |
return 0; |
| 1046 |
} |
| 1047 |
|
| 1048 |
template <> |
| 1049 |
inline constexpr int |
| 1050 |
binary_format<std::float16_t>::max_exponent_round_to_even() { |
| 1051 |
return 5; |
| 1052 |
} |
| 1053 |
|
| 1054 |
template <> |
| 1055 |
inline constexpr int |
| 1056 |
binary_format<std::float16_t>::min_exponent_round_to_even() { |
| 1057 |
return -22; |
| 1058 |
} |
| 1059 |
|
| 1060 |
template <> |
| 1061 |
inline constexpr int binary_format<std::float16_t>::minimum_exponent() { |
| 1062 |
return -15; |
| 1063 |
} |
| 1064 |
|
| 1065 |
template <> |
| 1066 |
inline constexpr int binary_format<std::float16_t>::infinite_power() { |
| 1067 |
return 0x1F; |
| 1068 |
} |
| 1069 |
|
| 1070 |
template <> inline constexpr int binary_format<std::float16_t>::sign_index() { |
| 1071 |
return 15; |
| 1072 |
} |
| 1073 |
|
| 1074 |
template <> |
| 1075 |
inline constexpr int binary_format<std::float16_t>::largest_power_of_ten() { |
| 1076 |
return 4; |
| 1077 |
} |
| 1078 |
|
| 1079 |
template <> |
| 1080 |
inline constexpr int binary_format<std::float16_t>::smallest_power_of_ten() { |
| 1081 |
return -27; |
| 1082 |
} |
| 1083 |
|
| 1084 |
template <> |
| 1085 |
inline constexpr size_t binary_format<std::float16_t>::max_digits() { |
| 1086 |
return 22; |
| 1087 |
} |
| 1088 |
#endif // __STDCPP_FLOAT16_T__ |
| 1089 |
|
| 1090 |
// credit: Jakub Jelínek |
| 1091 |
#ifdef __STDCPP_BFLOAT16_T__ |
| 1092 |
template <typename U> struct binary_format_lookup_tables<std::bfloat16_t, U> { |
| 1093 |
static constexpr std::bfloat16_t powers_of_ten[] = {1e0bf16, 1e1bf16, 1e2bf16, |
| 1094 |
1e3bf16}; |
| 1095 |
|
| 1096 |
// Largest integer value v so that (5**index * v) <= 1<<8. |
| 1097 |
// 0x100 == 1<<8 |
| 1098 |
static constexpr uint64_t max_mantissa[] = {0x100, 0x100 / 5, 0x100 / (5 * 5), |
| 1099 |
0x100 / (5 * 5 * 5), |
| 1100 |
0x100 / (5 * 5 * 5 * 5)}; |
| 1101 |
}; |
| 1102 |
|
| 1103 |
#if FASTFLOAT_DETAIL_MUST_DEFINE_CONSTEXPR_VARIABLE |
| 1104 |
|
| 1105 |
template <typename U> |
| 1106 |
constexpr std::bfloat16_t |
| 1107 |
binary_format_lookup_tables<std::bfloat16_t, U>::powers_of_ten[]; |
| 1108 |
|
| 1109 |
template <typename U> |
| 1110 |
constexpr uint64_t |
| 1111 |
binary_format_lookup_tables<std::bfloat16_t, U>::max_mantissa[]; |
| 1112 |
|
| 1113 |
#endif |
| 1114 |
|
| 1115 |
template <> |
| 1116 |
inline constexpr std::bfloat16_t |
| 1117 |
binary_format<std::bfloat16_t>::exact_power_of_ten(int64_t power) { |
| 1118 |
// Work around clang bug https://godbolt.org/z/zedh7rrhc |
| 1119 |
return (void)powers_of_ten[0], powers_of_ten[power]; |
| 1120 |
} |
| 1121 |
|
| 1122 |
template <> |
| 1123 |
inline constexpr int binary_format<std::bfloat16_t>::max_exponent_fast_path() { |
| 1124 |
return 3; |
| 1125 |
} |
| 1126 |
|
| 1127 |
template <> |
| 1128 |
inline constexpr binary_format<std::bfloat16_t>::equiv_uint |
| 1129 |
binary_format<std::bfloat16_t>::exponent_mask() { |
| 1130 |
return 0x7F80; |
| 1131 |
} |
| 1132 |
|
| 1133 |
template <> |
| 1134 |
inline constexpr binary_format<std::bfloat16_t>::equiv_uint |
| 1135 |
binary_format<std::bfloat16_t>::mantissa_mask() { |
| 1136 |
return 0x007F; |
| 1137 |
} |
| 1138 |
|
| 1139 |
template <> |
| 1140 |
inline constexpr binary_format<std::bfloat16_t>::equiv_uint |
| 1141 |
binary_format<std::bfloat16_t>::hidden_bit_mask() { |
| 1142 |
return 0x0080; |
| 1143 |
} |
| 1144 |
|
| 1145 |
template <> |
| 1146 |
inline constexpr int binary_format<std::bfloat16_t>::mantissa_explicit_bits() { |
| 1147 |
return 7; |
| 1148 |
} |
| 1149 |
|
| 1150 |
template <> |
| 1151 |
inline constexpr uint64_t |
| 1152 |
binary_format<std::bfloat16_t>::max_mantissa_fast_path() { |
| 1153 |
return uint64_t(2) << mantissa_explicit_bits(); |
| 1154 |
} |
| 1155 |
|
| 1156 |
template <> |
| 1157 |
inline constexpr uint64_t |
| 1158 |
binary_format<std::bfloat16_t>::max_mantissa_fast_path(int64_t power) { |
| 1159 |
// caller is responsible to ensure that |
| 1160 |
// power >= 0 && power <= 3 |
| 1161 |
// |
| 1162 |
// Work around clang bug https://godbolt.org/z/zedh7rrhc |
| 1163 |
return (void)max_mantissa[0], max_mantissa[power]; |
| 1164 |
} |
| 1165 |
|
| 1166 |
template <> |
| 1167 |
inline constexpr int binary_format<std::bfloat16_t>::min_exponent_fast_path() { |
| 1168 |
return 0; |
| 1169 |
} |
| 1170 |
|
| 1171 |
template <> |
| 1172 |
inline constexpr int |
| 1173 |
binary_format<std::bfloat16_t>::max_exponent_round_to_even() { |
| 1174 |
return 3; |
| 1175 |
} |
| 1176 |
|
| 1177 |
template <> |
| 1178 |
inline constexpr int |
| 1179 |
binary_format<std::bfloat16_t>::min_exponent_round_to_even() { |
| 1180 |
return -24; |
| 1181 |
} |
| 1182 |
|
| 1183 |
template <> |
| 1184 |
inline constexpr int binary_format<std::bfloat16_t>::minimum_exponent() { |
| 1185 |
return -127; |
| 1186 |
} |
| 1187 |
|
| 1188 |
template <> |
| 1189 |
inline constexpr int binary_format<std::bfloat16_t>::infinite_power() { |
| 1190 |
return 0xFF; |
| 1191 |
} |
| 1192 |
|
| 1193 |
template <> inline constexpr int binary_format<std::bfloat16_t>::sign_index() { |
| 1194 |
return 15; |
| 1195 |
} |
| 1196 |
|
| 1197 |
template <> |
| 1198 |
inline constexpr int binary_format<std::bfloat16_t>::largest_power_of_ten() { |
| 1199 |
return 38; |
| 1200 |
} |
| 1201 |
|
| 1202 |
template <> |
| 1203 |
inline constexpr int binary_format<std::bfloat16_t>::smallest_power_of_ten() { |
| 1204 |
return -60; |
| 1205 |
} |
| 1206 |
|
| 1207 |
template <> |
| 1208 |
inline constexpr size_t binary_format<std::bfloat16_t>::max_digits() { |
| 1209 |
return 98; |
| 1210 |
} |
| 1211 |
#endif // __STDCPP_BFLOAT16_T__ |
| 1212 |
|
| 1213 |
template <> |
| 1214 |
inline constexpr uint64_t |
| 1215 |
binary_format<double>::max_mantissa_fast_path(int64_t power) { |
| 1216 |
// caller is responsible to ensure that |
| 1217 |
// power >= 0 && power <= 22 |
| 1218 |
// |
| 1219 |
// Work around clang bug https://godbolt.org/z/zedh7rrhc |
| 1220 |
return (void)max_mantissa[0], max_mantissa[power]; |
| 1221 |
} |
| 1222 |
|
| 1223 |
template <> |
| 1224 |
inline constexpr uint64_t |
| 1225 |
binary_format<float>::max_mantissa_fast_path(int64_t power) { |
| 1226 |
// caller is responsible to ensure that |
| 1227 |
// power >= 0 && power <= 10 |
| 1228 |
// |
| 1229 |
// Work around clang bug https://godbolt.org/z/zedh7rrhc |
| 1230 |
return (void)max_mantissa[0], max_mantissa[power]; |
| 1231 |
} |
| 1232 |
|
| 1233 |
template <> |
| 1234 |
inline constexpr double |
| 1235 |
binary_format<double>::exact_power_of_ten(int64_t power) { |
| 1236 |
// Work around clang bug https://godbolt.org/z/zedh7rrhc |
| 1237 |
return (void)powers_of_ten[0], powers_of_ten[power]; |
| 1238 |
} |
| 1239 |
|
| 1240 |
template <> |
| 1241 |
inline constexpr float binary_format<float>::exact_power_of_ten(int64_t power) { |
| 1242 |
// Work around clang bug https://godbolt.org/z/zedh7rrhc |
| 1243 |
return (void)powers_of_ten[0], powers_of_ten[power]; |
| 1244 |
} |
| 1245 |
|
| 1246 |
template <> inline constexpr int binary_format<double>::largest_power_of_ten() { |
| 1247 |
return 308; |
| 1248 |
} |
| 1249 |
|
| 1250 |
template <> inline constexpr int binary_format<float>::largest_power_of_ten() { |
| 1251 |
return 38; |
| 1252 |
} |
| 1253 |
|
| 1254 |
template <> |
| 1255 |
inline constexpr int binary_format<double>::smallest_power_of_ten() { |
| 1256 |
return -342; |
| 1257 |
} |
| 1258 |
|
| 1259 |
template <> inline constexpr int binary_format<float>::smallest_power_of_ten() { |
| 1260 |
return -64; |
| 1261 |
} |
| 1262 |
|
| 1263 |
template <> inline constexpr size_t binary_format<double>::max_digits() { |
| 1264 |
return 769; |
| 1265 |
} |
| 1266 |
|
| 1267 |
template <> inline constexpr size_t binary_format<float>::max_digits() { |
| 1268 |
return 114; |
| 1269 |
} |
| 1270 |
|
| 1271 |
template <> |
| 1272 |
inline constexpr binary_format<float>::equiv_uint |
| 1273 |
binary_format<float>::exponent_mask() { |
| 1274 |
return 0x7F800000; |
| 1275 |
} |
| 1276 |
|
| 1277 |
template <> |
| 1278 |
inline constexpr binary_format<double>::equiv_uint |
| 1279 |
binary_format<double>::exponent_mask() { |
| 1280 |
return 0x7FF0000000000000; |
| 1281 |
} |
| 1282 |
|
| 1283 |
template <> |
| 1284 |
inline constexpr binary_format<float>::equiv_uint |
| 1285 |
binary_format<float>::mantissa_mask() { |
| 1286 |
return 0x007FFFFF; |
| 1287 |
} |
| 1288 |
|
| 1289 |
template <> |
| 1290 |
inline constexpr binary_format<double>::equiv_uint |
| 1291 |
binary_format<double>::mantissa_mask() { |
| 1292 |
return 0x000FFFFFFFFFFFFF; |
| 1293 |
} |
| 1294 |
|
| 1295 |
template <> |
| 1296 |
inline constexpr binary_format<float>::equiv_uint |
| 1297 |
binary_format<float>::hidden_bit_mask() { |
| 1298 |
return 0x00800000; |
| 1299 |
} |
| 1300 |
|
| 1301 |
template <> |
| 1302 |
inline constexpr binary_format<double>::equiv_uint |
| 1303 |
binary_format<double>::hidden_bit_mask() { |
| 1304 |
return 0x0010000000000000; |
| 1305 |
} |
| 1306 |
|
| 1307 |
template <typename T> |
| 1308 |
fastfloat_really_inline FASTFLOAT_CONSTEXPR20 void |
| 1309 |
to_float(bool negative, adjusted_mantissa am, T &value) { |
| 1310 |
using equiv_uint = equiv_uint_t<T>; |
| 1311 |
equiv_uint word = equiv_uint(am.mantissa); |
| 1312 |
word = equiv_uint(word | equiv_uint(am.power2) |
| 1313 |
<< binary_format<T>::mantissa_explicit_bits()); |
| 1314 |
word = |
| 1315 |
equiv_uint(word | equiv_uint(negative) << binary_format<T>::sign_index()); |
| 1316 |
#if FASTFLOAT_HAS_BIT_CAST |
| 1317 |
value = std::bit_cast<T>(word); |
| 1318 |
#else |
| 1319 |
::memcpy(&value, &word, sizeof(T)); |
| 1320 |
#endif |
| 1321 |
} |
| 1322 |
|
| 1323 |
template <typename = void> struct space_lut { |
| 1324 |
static constexpr bool value[] = { |
| 1325 |
0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 1, 1, 1, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, |
| 1326 |
0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, |
| 1327 |
0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, |
| 1328 |
0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, |
| 1329 |
0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, |
| 1330 |
0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, |
| 1331 |
0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, |
| 1332 |
0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, |
| 1333 |
0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, |
| 1334 |
0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, |
| 1335 |
0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0}; |
| 1336 |
}; |
| 1337 |
|
| 1338 |
#if FASTFLOAT_DETAIL_MUST_DEFINE_CONSTEXPR_VARIABLE |
| 1339 |
|
| 1340 |
template <typename T> constexpr bool space_lut<T>::value[]; |
| 1341 |
|
| 1342 |
#endif |
| 1343 |
|
| 1344 |
template <typename UC> constexpr bool is_space(UC c) { |
| 1345 |
return c < 256 && space_lut<>::value[uint8_t(c)]; |
| 1346 |
} |
| 1347 |
|
| 1348 |
template <typename UC> static constexpr uint64_t int_cmp_zeros() { |
| 1349 |
static_assert((sizeof(UC) == 1) || (sizeof(UC) == 2) || (sizeof(UC) == 4), |
| 1350 |
"Unsupported character size"); |
| 1351 |
return (sizeof(UC) == 1) ? 0x3030303030303030 |
| 1352 |
: (sizeof(UC) == 2) |
| 1353 |
? (uint64_t(UC('0')) << 48 | uint64_t(UC('0')) << 32 | |
| 1354 |
uint64_t(UC('0')) << 16 | UC('0')) |
| 1355 |
: (uint64_t(UC('0')) << 32 | UC('0')); |
| 1356 |
} |
| 1357 |
|
| 1358 |
template <typename UC> static constexpr int int_cmp_len() { |
| 1359 |
return sizeof(uint64_t) / sizeof(UC); |
| 1360 |
} |
| 1361 |
|
| 1362 |
template <typename UC> constexpr UC const *str_const_nan(); |
| 1363 |
|
| 1364 |
template <> constexpr char const *str_const_nan<char>() { return "nan"; } |
| 1365 |
|
| 1366 |
template <> constexpr wchar_t const *str_const_nan<wchar_t>() { return L"nan"; } |
| 1367 |
|
| 1368 |
template <> constexpr char16_t const *str_const_nan<char16_t>() { |
| 1369 |
return u"nan"; |
| 1370 |
} |
| 1371 |
|
| 1372 |
template <> constexpr char32_t const *str_const_nan<char32_t>() { |
| 1373 |
return U"nan"; |
| 1374 |
} |
| 1375 |
|
| 1376 |
#ifdef __cpp_char8_t |
| 1377 |
template <> constexpr char8_t const *str_const_nan<char8_t>() { |
| 1378 |
return u8"nan"; |
| 1379 |
} |
| 1380 |
#endif |
| 1381 |
|
| 1382 |
template <typename UC> constexpr UC const *str_const_inf(); |
| 1383 |
|
| 1384 |
template <> constexpr char const *str_const_inf<char>() { return "infinity"; } |
| 1385 |
|
| 1386 |
template <> constexpr wchar_t const *str_const_inf<wchar_t>() { |
| 1387 |
return L"infinity"; |
| 1388 |
} |
| 1389 |
|
| 1390 |
template <> constexpr char16_t const *str_const_inf<char16_t>() { |
| 1391 |
return u"infinity"; |
| 1392 |
} |
| 1393 |
|
| 1394 |
template <> constexpr char32_t const *str_const_inf<char32_t>() { |
| 1395 |
return U"infinity"; |
| 1396 |
} |
| 1397 |
|
| 1398 |
#ifdef __cpp_char8_t |
| 1399 |
template <> constexpr char8_t const *str_const_inf<char8_t>() { |
| 1400 |
return u8"infinity"; |
| 1401 |
} |
| 1402 |
#endif |
| 1403 |
|
| 1404 |
template <typename = void> struct int_luts { |
| 1405 |
static constexpr uint8_t chdigit[] = { |
| 1406 |
255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, |
| 1407 |
255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, |
| 1408 |
255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, |
| 1409 |
255, 255, 255, 0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 255, 255, |
| 1410 |
255, 255, 255, 255, 255, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, |
| 1411 |
20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31, 32, 33, 34, |
| 1412 |
35, 255, 255, 255, 255, 255, 255, 10, 11, 12, 13, 14, 15, 16, 17, |
| 1413 |
18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31, 32, |
| 1414 |
33, 34, 35, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, |
| 1415 |
255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, |
| 1416 |
255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, |
| 1417 |
255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, |
| 1418 |
255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, |
| 1419 |
255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, |
| 1420 |
255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, |
| 1421 |
255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, |
| 1422 |
255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, |
| 1423 |
255}; |
| 1424 |
|
| 1425 |
static constexpr size_t maxdigits_u64[] = { |
| 1426 |
64, 41, 32, 28, 25, 23, 22, 21, 20, 19, 18, 18, 17, 17, 16, 16, 16, 16, |
| 1427 |
15, 15, 15, 15, 14, 14, 14, 14, 14, 14, 14, 13, 13, 13, 13, 13, 13}; |
| 1428 |
|
| 1429 |
static constexpr uint64_t min_safe_u64[] = { |
| 1430 |
9223372036854775808ull, 12157665459056928801ull, 4611686018427387904, |
| 1431 |
7450580596923828125, 4738381338321616896, 3909821048582988049, |
| 1432 |
9223372036854775808ull, 12157665459056928801ull, 10000000000000000000ull, |
| 1433 |
5559917313492231481, 2218611106740436992, 8650415919381337933, |
| 1434 |
2177953337809371136, 6568408355712890625, 1152921504606846976, |
| 1435 |
2862423051509815793, 6746640616477458432, 15181127029874798299ull, |
| 1436 |
1638400000000000000, 3243919932521508681, 6221821273427820544, |
| 1437 |
11592836324538749809ull, 876488338465357824, 1490116119384765625, |
| 1438 |
2481152873203736576, 4052555153018976267, 6502111422497947648, |
| 1439 |
10260628712958602189ull, 15943230000000000000ull, 787662783788549761, |
| 1440 |
1152921504606846976, 1667889514952984961, 2386420683693101056, |
| 1441 |
3379220508056640625, 4738381338321616896}; |
| 1442 |
}; |
| 1443 |
|
| 1444 |
#if FASTFLOAT_DETAIL_MUST_DEFINE_CONSTEXPR_VARIABLE |
| 1445 |
|
| 1446 |
template <typename T> constexpr uint8_t int_luts<T>::chdigit[]; |
| 1447 |
|
| 1448 |
template <typename T> constexpr size_t int_luts<T>::maxdigits_u64[]; |
| 1449 |
|
| 1450 |
template <typename T> constexpr uint64_t int_luts<T>::min_safe_u64[]; |
| 1451 |
|
| 1452 |
#endif |
| 1453 |
|
| 1454 |
template <typename UC> |
| 1455 |
fastfloat_really_inline constexpr uint8_t ch_to_digit(UC c) { |
| 1456 |
// wchar_t and char can be signed, so we need to be careful. |
| 1457 |
using UnsignedUC = typename std::make_unsigned<UC>::type; |
| 1458 |
return int_luts<>::chdigit[static_cast<unsigned char>( |
| 1459 |
static_cast<UnsignedUC>(c) & |
| 1460 |
static_cast<UnsignedUC>( |
| 1461 |
-((static_cast<UnsignedUC>(c) & ~0xFFull) == 0)))]; |
| 1462 |
} |
| 1463 |
|
| 1464 |
fastfloat_really_inline constexpr size_t max_digits_u64(int base) { |
| 1465 |
return int_luts<>::maxdigits_u64[base - 2]; |
| 1466 |
} |
| 1467 |
|
| 1468 |
// If a u64 is exactly max_digits_u64() in length, this is |
| 1469 |
// the value below which it has definitely overflowed. |
| 1470 |
fastfloat_really_inline constexpr uint64_t min_safe_u64(int base) { |
| 1471 |
return int_luts<>::min_safe_u64[base - 2]; |
| 1472 |
} |
| 1473 |
|
| 1474 |
static_assert(std::is_same<equiv_uint_t<double>, uint64_t>::value, |
| 1475 |
"equiv_uint should be uint64_t for double"); |
| 1476 |
static_assert(std::numeric_limits<double>::is_iec559, |
| 1477 |
"double must fulfill the requirements of IEC 559 (IEEE 754)"); |
| 1478 |
|
| 1479 |
static_assert(std::is_same<equiv_uint_t<float>, uint32_t>::value, |
| 1480 |
"equiv_uint should be uint32_t for float"); |
| 1481 |
static_assert(std::numeric_limits<float>::is_iec559, |
| 1482 |
"float must fulfill the requirements of IEC 559 (IEEE 754)"); |
| 1483 |
|
| 1484 |
#ifdef __STDCPP_FLOAT64_T__ |
| 1485 |
static_assert(std::is_same<equiv_uint_t<std::float64_t>, uint64_t>::value, |
| 1486 |
"equiv_uint should be uint64_t for std::float64_t"); |
| 1487 |
static_assert( |
| 1488 |
std::numeric_limits<std::float64_t>::is_iec559, |
| 1489 |
"std::float64_t must fulfill the requirements of IEC 559 (IEEE 754)"); |
| 1490 |
|
| 1491 |
template <> |
| 1492 |
struct binary_format<std::float64_t> : public binary_format<double> {}; |
| 1493 |
#endif // __STDCPP_FLOAT64_T__ |
| 1494 |
|
| 1495 |
#ifdef __STDCPP_FLOAT32_T__ |
| 1496 |
static_assert(std::is_same<equiv_uint_t<std::float32_t>, uint32_t>::value, |
| 1497 |
"equiv_uint should be uint32_t for std::float32_t"); |
| 1498 |
static_assert( |
| 1499 |
std::numeric_limits<std::float32_t>::is_iec559, |
| 1500 |
"std::float32_t must fulfill the requirements of IEC 559 (IEEE 754)"); |
| 1501 |
|
| 1502 |
template <> |
| 1503 |
struct binary_format<std::float32_t> : public binary_format<float> {}; |
| 1504 |
#endif // __STDCPP_FLOAT32_T__ |
| 1505 |
|
| 1506 |
#ifdef __STDCPP_FLOAT16_T__ |
| 1507 |
static_assert( |
| 1508 |
std::is_same<binary_format<std::float16_t>::equiv_uint, uint16_t>::value, |
| 1509 |
"equiv_uint should be uint16_t for std::float16_t"); |
| 1510 |
static_assert( |
| 1511 |
std::numeric_limits<std::float16_t>::is_iec559, |
| 1512 |
"std::float16_t must fulfill the requirements of IEC 559 (IEEE 754)"); |
| 1513 |
#endif // __STDCPP_FLOAT16_T__ |
| 1514 |
|
| 1515 |
#ifdef __STDCPP_BFLOAT16_T__ |
| 1516 |
static_assert( |
| 1517 |
std::is_same<binary_format<std::bfloat16_t>::equiv_uint, uint16_t>::value, |
| 1518 |
"equiv_uint should be uint16_t for std::bfloat16_t"); |
| 1519 |
static_assert( |
| 1520 |
std::numeric_limits<std::bfloat16_t>::is_iec559, |
| 1521 |
"std::bfloat16_t must fulfill the requirements of IEC 559 (IEEE 754)"); |
| 1522 |
#endif // __STDCPP_BFLOAT16_T__ |
| 1523 |
|
| 1524 |
constexpr chars_format operator~(chars_format rhs) noexcept { |
| 1525 |
using int_type = std::underlying_type<chars_format>::type; |
| 1526 |
return static_cast<chars_format>(~static_cast<int_type>(rhs)); |
| 1527 |
} |
| 1528 |
|
| 1529 |
constexpr chars_format operator&(chars_format lhs, chars_format rhs) noexcept { |
| 1530 |
using int_type = std::underlying_type<chars_format>::type; |
| 1531 |
return static_cast<chars_format>(static_cast<int_type>(lhs) & |
| 1532 |
static_cast<int_type>(rhs)); |
| 1533 |
} |
| 1534 |
|
| 1535 |
constexpr chars_format operator|(chars_format lhs, chars_format rhs) noexcept { |
| 1536 |
using int_type = std::underlying_type<chars_format>::type; |
| 1537 |
return static_cast<chars_format>(static_cast<int_type>(lhs) | |
| 1538 |
static_cast<int_type>(rhs)); |
| 1539 |
} |
| 1540 |
|
| 1541 |
constexpr chars_format operator^(chars_format lhs, chars_format rhs) noexcept { |
| 1542 |
using int_type = std::underlying_type<chars_format>::type; |
| 1543 |
return static_cast<chars_format>(static_cast<int_type>(lhs) ^ |
| 1544 |
static_cast<int_type>(rhs)); |
| 1545 |
} |
| 1546 |
|
| 1547 |
fastfloat_really_inline FASTFLOAT_CONSTEXPR14 chars_format & |
| 1548 |
operator&=(chars_format &lhs, chars_format rhs) noexcept { |
| 1549 |
return lhs = (lhs & rhs); |
| 1550 |
} |
| 1551 |
|
| 1552 |
fastfloat_really_inline FASTFLOAT_CONSTEXPR14 chars_format & |
| 1553 |
operator|=(chars_format &lhs, chars_format rhs) noexcept { |
| 1554 |
return lhs = (lhs | rhs); |
| 1555 |
} |
| 1556 |
|
| 1557 |
fastfloat_really_inline FASTFLOAT_CONSTEXPR14 chars_format & |
| 1558 |
operator^=(chars_format &lhs, chars_format rhs) noexcept { |
| 1559 |
return lhs = (lhs ^ rhs); |
| 1560 |
} |
| 1561 |
|
| 1562 |
namespace detail { |
| 1563 |
// adjust for deprecated feature macros |
| 1564 |
constexpr chars_format adjust_for_feature_macros(chars_format fmt) { |
| 1565 |
return fmt |
| 1566 |
#ifdef FASTFLOAT_ALLOWS_LEADING_PLUS |
| 1567 |
| chars_format::allow_leading_plus |
| 1568 |
#endif |
| 1569 |
#ifdef FASTFLOAT_SKIP_WHITE_SPACE |
| 1570 |
| chars_format::skip_white_space |
| 1571 |
#endif |
| 1572 |
; |
| 1573 |
} |
| 1574 |
} // namespace detail |
| 1575 |
} // namespace fast_float |
| 1576 |
|
| 1577 |
#endif |
| 1578 |
|
| 1579 |
|
| 1580 |
#ifndef FASTFLOAT_FAST_FLOAT_H |
| 1581 |
#define FASTFLOAT_FAST_FLOAT_H |
| 1582 |
|
| 1583 |
|
| 1584 |
namespace fast_float { |
| 1585 |
/** |
| 1586 |
* This function parses the character sequence [first,last) for a number. It |
| 1587 |
* parses floating-point numbers expecting a locale-indepent format equivalent |
| 1588 |
* to what is used by std::strtod in the default ("C") locale. The resulting |
| 1589 |
* floating-point value is the closest floating-point values (using either float |
| 1590 |
* or double), using the "round to even" convention for values that would |
| 1591 |
* otherwise fall right in-between two values. That is, we provide exact parsing |
| 1592 |
* according to the IEEE standard. |
| 1593 |
* |
| 1594 |
* Given a successful parse, the pointer (`ptr`) in the returned value is set to |
| 1595 |
* point right after the parsed number, and the `value` referenced is set to the |
| 1596 |
* parsed value. In case of error, the returned `ec` contains a representative |
| 1597 |
* error, otherwise the default (`std::errc()`) value is stored. |
| 1598 |
* |
| 1599 |
* The implementation does not throw and does not allocate memory (e.g., with |
| 1600 |
* `new` or `malloc`). |
| 1601 |
* |
| 1602 |
* Like the C++17 standard, the `fast_float::from_chars` functions take an |
| 1603 |
* optional last argument of the type `fast_float::chars_format`. It is a bitset |
| 1604 |
* value: we check whether `fmt & fast_float::chars_format::fixed` and `fmt & |
| 1605 |
* fast_float::chars_format::scientific` are set to determine whether we allow |
| 1606 |
* the fixed point and scientific notation respectively. The default is |
| 1607 |
* `fast_float::chars_format::general` which allows both `fixed` and |
| 1608 |
* `scientific`. |
| 1609 |
*/ |
| 1610 |
template <typename T, typename UC = char, |
| 1611 |
typename = FASTFLOAT_ENABLE_IF(is_supported_float_type<T>::value)> |
| 1612 |
FASTFLOAT_CONSTEXPR20 from_chars_result_t<UC> |
| 1613 |
from_chars(UC const *first, UC const *last, T &value, |
| 1614 |
chars_format fmt = chars_format::general) noexcept; |
| 1615 |
|
| 1616 |
/** |
| 1617 |
* Like from_chars, but accepts an `options` argument to govern number parsing. |
| 1618 |
* Both for floating-point types and integer types. |
| 1619 |
*/ |
| 1620 |
template <typename T, typename UC = char> |
| 1621 |
FASTFLOAT_CONSTEXPR20 from_chars_result_t<UC> |
| 1622 |
from_chars_advanced(UC const *first, UC const *last, T &value, |
| 1623 |
parse_options_t<UC> options) noexcept; |
| 1624 |
|
| 1625 |
/** |
| 1626 |
* This function multiplies an integer number by a power of 10 and returns |
| 1627 |
* the result as a double precision floating-point value that is correctly |
| 1628 |
* rounded. The resulting floating-point value is the closest floating-point |
| 1629 |
* value, using the "round to nearest, tie to even" convention for values that |
| 1630 |
* would otherwise fall right in-between two values. That is, we provide exact |
| 1631 |
* conversion according to the IEEE standard. |
| 1632 |
* |
| 1633 |
* On overflow infinity is returned, on underflow 0 is returned. |
| 1634 |
* |
| 1635 |
* The implementation does not throw and does not allocate memory (e.g., with |
| 1636 |
* `new` or `malloc`). |
| 1637 |
*/ |
| 1638 |
FASTFLOAT_CONSTEXPR20 inline double |
| 1639 |
integer_times_pow10(uint64_t mantissa, int decimal_exponent) noexcept; |
| 1640 |
FASTFLOAT_CONSTEXPR20 inline double |
| 1641 |
integer_times_pow10(int64_t mantissa, int decimal_exponent) noexcept; |
| 1642 |
|
| 1643 |
/** |
| 1644 |
* This function is a template overload of `integer_times_pow10()` |
| 1645 |
* that returns a floating-point value of type `T` that is one of |
| 1646 |
* supported floating-point types (e.g. `double`, `float`). |
| 1647 |
*/ |
| 1648 |
template <typename T> |
| 1649 |
FASTFLOAT_CONSTEXPR20 |
| 1650 |
typename std::enable_if<is_supported_float_type<T>::value, T>::type |
| 1651 |
integer_times_pow10(uint64_t mantissa, int decimal_exponent) noexcept; |
| 1652 |
template <typename T> |
| 1653 |
FASTFLOAT_CONSTEXPR20 |
| 1654 |
typename std::enable_if<is_supported_float_type<T>::value, T>::type |
| 1655 |
integer_times_pow10(int64_t mantissa, int decimal_exponent) noexcept; |
| 1656 |
|
| 1657 |
/** |
| 1658 |
* from_chars for integer types. |
| 1659 |
*/ |
| 1660 |
template <typename T, typename UC = char, |
| 1661 |
typename = FASTFLOAT_ENABLE_IF(is_supported_integer_type<T>::value)> |
| 1662 |
FASTFLOAT_CONSTEXPR20 from_chars_result_t<UC> |
| 1663 |
from_chars(UC const *first, UC const *last, T &value, int base = 10) noexcept; |
| 1664 |
|
| 1665 |
} // namespace fast_float |
| 1666 |
|
| 1667 |
#endif // FASTFLOAT_FAST_FLOAT_H |
| 1668 |
|
| 1669 |
#ifndef FASTFLOAT_ASCII_NUMBER_H |
| 1670 |
#define FASTFLOAT_ASCII_NUMBER_H |
| 1671 |
|
| 1672 |
#include <cctype> |
| 1673 |
#include <cstdint> |
| 1674 |
#include <cstring> |
| 1675 |
#include <iterator> |
| 1676 |
#include <limits> |
| 1677 |
#include <type_traits> |
| 1678 |
|
| 1679 |
|
| 1680 |
#ifdef FASTFLOAT_SSE2 |
| 1681 |
#include <emmintrin.h> |
| 1682 |
#endif |
| 1683 |
|
| 1684 |
#ifdef FASTFLOAT_NEON |
| 1685 |
#include <arm_neon.h> |
| 1686 |
#endif |
| 1687 |
|
| 1688 |
namespace fast_float { |
| 1689 |
|
| 1690 |
template <typename UC> fastfloat_really_inline constexpr bool has_simd_opt() { |
| 1691 |
#ifdef FASTFLOAT_HAS_SIMD |
| 1692 |
return std::is_same<UC, char16_t>::value; |
| 1693 |
#else |
| 1694 |
return false; |
| 1695 |
#endif |
| 1696 |
} |
| 1697 |
|
| 1698 |
// Next function can be micro-optimized, but compilers are entirely |
| 1699 |
// able to optimize it well. |
| 1700 |
template <typename UC> |
| 1701 |
fastfloat_really_inline constexpr bool is_integer(UC c) noexcept { |
| 1702 |
return (unsigned)(c - UC('0')) <= 9u; |
| 1703 |
} |
| 1704 |
|
| 1705 |
fastfloat_really_inline constexpr uint64_t byteswap(uint64_t val) { |
| 1706 |
return (val & 0xFF00000000000000) >> 56 | (val & 0x00FF000000000000) >> 40 | |
| 1707 |
(val & 0x0000FF0000000000) >> 24 | (val & 0x000000FF00000000) >> 8 | |
| 1708 |
(val & 0x00000000FF000000) << 8 | (val & 0x0000000000FF0000) << 24 | |
| 1709 |
(val & 0x000000000000FF00) << 40 | (val & 0x00000000000000FF) << 56; |
| 1710 |
} |
| 1711 |
|
| 1712 |
fastfloat_really_inline constexpr uint32_t byteswap_32(uint32_t val) { |
| 1713 |
return (val >> 24) | ((val >> 8) & 0x0000FF00u) | ((val << 8) & 0x00FF0000u) | |
| 1714 |
(val << 24); |
| 1715 |
} |
| 1716 |
|
| 1717 |
// Read 8 UC into a u64. Truncates UC if not char. |
| 1718 |
template <typename UC> |
| 1719 |
fastfloat_really_inline FASTFLOAT_CONSTEXPR20 uint64_t |
| 1720 |
read8_to_u64(UC const *chars) { |
| 1721 |
if (cpp20_and_in_constexpr() || !std::is_same<UC, char>::value) { |
| 1722 |
uint64_t val = 0; |
| 1723 |
for (int i = 0; i < 8; ++i) { |
| 1724 |
val |= uint64_t(uint8_t(*chars)) << (i * 8); |
| 1725 |
++chars; |
| 1726 |
} |
| 1727 |
return val; |
| 1728 |
} |
| 1729 |
uint64_t val; |
| 1730 |
::memcpy(&val, chars, sizeof(uint64_t)); |
| 1731 |
#if FASTFLOAT_IS_BIG_ENDIAN == 1 |
| 1732 |
// Need to read as-if the number was in little-endian order. |
| 1733 |
val = byteswap(val); |
| 1734 |
#endif |
| 1735 |
return val; |
| 1736 |
} |
| 1737 |
|
| 1738 |
// Read 4 UC into a u32. Truncates UC if not char. |
| 1739 |
template <typename UC> |
| 1740 |
fastfloat_really_inline FASTFLOAT_CONSTEXPR20 uint32_t |
| 1741 |
read4_to_u32(UC const *chars) { |
| 1742 |
if (cpp20_and_in_constexpr() || !std::is_same<UC, char>::value) { |
| 1743 |
uint32_t val = 0; |
| 1744 |
for (int i = 0; i < 4; ++i) { |
| 1745 |
val |= uint32_t(uint8_t(*chars)) << (i * 8); |
| 1746 |
++chars; |
| 1747 |
} |
| 1748 |
return val; |
| 1749 |
} |
| 1750 |
uint32_t val; |
| 1751 |
::memcpy(&val, chars, sizeof(uint32_t)); |
| 1752 |
#if FASTFLOAT_IS_BIG_ENDIAN == 1 |
| 1753 |
val = byteswap_32(val); |
| 1754 |
#endif |
| 1755 |
return val; |
| 1756 |
} |
| 1757 |
#ifdef FASTFLOAT_SSE2 |
| 1758 |
|
| 1759 |
fastfloat_really_inline uint64_t simd_read8_to_u64(__m128i const data) { |
| 1760 |
FASTFLOAT_SIMD_DISABLE_WARNINGS |
| 1761 |
__m128i const packed = _mm_packus_epi16(data, data); |
| 1762 |
#ifdef FASTFLOAT_64BIT |
| 1763 |
return uint64_t(_mm_cvtsi128_si64(packed)); |
| 1764 |
#else |
| 1765 |
uint64_t value; |
| 1766 |
// Visual Studio + older versions of GCC don't support _mm_storeu_si64 |
| 1767 |
_mm_storel_epi64(reinterpret_cast<__m128i *>(&value), packed); |
| 1768 |
return value; |
| 1769 |
#endif |
| 1770 |
FASTFLOAT_SIMD_RESTORE_WARNINGS |
| 1771 |
} |
| 1772 |
|
| 1773 |
fastfloat_really_inline uint64_t simd_read8_to_u64(char16_t const *chars) { |
| 1774 |
FASTFLOAT_SIMD_DISABLE_WARNINGS |
| 1775 |
return simd_read8_to_u64( |
| 1776 |
_mm_loadu_si128(reinterpret_cast<__m128i const *>(chars))); |
| 1777 |
FASTFLOAT_SIMD_RESTORE_WARNINGS |
| 1778 |
} |
| 1779 |
|
| 1780 |
#elif defined(FASTFLOAT_NEON) |
| 1781 |
|
| 1782 |
fastfloat_really_inline uint64_t simd_read8_to_u64(uint16x8_t const data) { |
| 1783 |
FASTFLOAT_SIMD_DISABLE_WARNINGS |
| 1784 |
uint8x8_t utf8_packed = vmovn_u16(data); |
| 1785 |
return vget_lane_u64(vreinterpret_u64_u8(utf8_packed), 0); |
| 1786 |
FASTFLOAT_SIMD_RESTORE_WARNINGS |
| 1787 |
} |
| 1788 |
|
| 1789 |
fastfloat_really_inline uint64_t simd_read8_to_u64(char16_t const *chars) { |
| 1790 |
FASTFLOAT_SIMD_DISABLE_WARNINGS |
| 1791 |
return simd_read8_to_u64( |
| 1792 |
vld1q_u16(reinterpret_cast<uint16_t const *>(chars))); |
| 1793 |
FASTFLOAT_SIMD_RESTORE_WARNINGS |
| 1794 |
} |
| 1795 |
|
| 1796 |
#endif // FASTFLOAT_SSE2 |
| 1797 |
|
| 1798 |
// MSVC SFINAE is broken pre-VS2017 |
| 1799 |
#if defined(_MSC_VER) && _MSC_VER <= 1900 |
| 1800 |
template <typename UC> |
| 1801 |
#else |
| 1802 |
template <typename UC, FASTFLOAT_ENABLE_IF(!has_simd_opt<UC>()) = 0> |
| 1803 |
#endif |
| 1804 |
// dummy for compile |
| 1805 |
uint64_t simd_read8_to_u64(UC const *) { |
| 1806 |
return 0; |
| 1807 |
} |
| 1808 |
|
| 1809 |
// credit @aqrit |
| 1810 |
fastfloat_really_inline FASTFLOAT_CONSTEXPR14 uint32_t |
| 1811 |
parse_eight_digits_unrolled(uint64_t val) { |
| 1812 |
uint64_t const mask = 0x000000FF000000FF; |
| 1813 |
uint64_t const mul1 = 0x000F424000000064; // 100 + (1000000ULL << 32) |
| 1814 |
uint64_t const mul2 = 0x0000271000000001; // 1 + (10000ULL << 32) |
| 1815 |
val -= 0x3030303030303030; |
| 1816 |
val = (val * 10) + (val >> 8); // val = (val * 2561) >> 8; |
| 1817 |
val = (((val & mask) * mul1) + (((val >> 16) & mask) * mul2)) >> 32; |
| 1818 |
return uint32_t(val); |
| 1819 |
} |
| 1820 |
|
| 1821 |
// Call this if chars are definitely 8 digits. |
| 1822 |
template <typename UC> |
| 1823 |
fastfloat_really_inline FASTFLOAT_CONSTEXPR20 uint32_t |
| 1824 |
parse_eight_digits_unrolled(UC const *chars) noexcept { |
| 1825 |
if (cpp20_and_in_constexpr() || !has_simd_opt<UC>()) { |
| 1826 |
return parse_eight_digits_unrolled(read8_to_u64(chars)); // truncation okay |
| 1827 |
} |
| 1828 |
return parse_eight_digits_unrolled(simd_read8_to_u64(chars)); |
| 1829 |
} |
| 1830 |
|
| 1831 |
// credit @aqrit |
| 1832 |
fastfloat_really_inline constexpr bool |
| 1833 |
is_made_of_eight_digits_fast(uint64_t val) noexcept { |
| 1834 |
return !((((val + 0x4646464646464646) | (val - 0x3030303030303030)) & |
| 1835 |
0x8080808080808080)); |
| 1836 |
} |
| 1837 |
|
| 1838 |
fastfloat_really_inline constexpr bool |
| 1839 |
is_made_of_four_digits_fast(uint32_t val) noexcept { |
| 1840 |
return !((((val + 0x46464646) | (val - 0x30303030)) & 0x80808080)); |
| 1841 |
} |
| 1842 |
|
| 1843 |
fastfloat_really_inline FASTFLOAT_CONSTEXPR14 uint32_t |
| 1844 |
parse_four_digits_unrolled(uint32_t val) noexcept { |
| 1845 |
val -= 0x30303030; |
| 1846 |
val = (val * 10) + (val >> 8); |
| 1847 |
return (((val & 0x00FF00FF) * 0x00640001) >> 16) & 0xFFFF; |
| 1848 |
} |
| 1849 |
|
| 1850 |
#ifdef FASTFLOAT_HAS_SIMD |
| 1851 |
|
| 1852 |
// Call this if chars might not be 8 digits. |
| 1853 |
// Using this style (instead of is_made_of_eight_digits_fast() then |
| 1854 |
// parse_eight_digits_unrolled()) ensures we don't load SIMD registers twice. |
| 1855 |
fastfloat_really_inline FASTFLOAT_CONSTEXPR20 bool |
| 1856 |
simd_parse_if_eight_digits_unrolled(char16_t const *chars, |
| 1857 |
uint64_t &i) noexcept { |
| 1858 |
if (cpp20_and_in_constexpr()) { |
| 1859 |
return false; |
| 1860 |
} |
| 1861 |
#ifdef FASTFLOAT_SSE2 |
| 1862 |
FASTFLOAT_SIMD_DISABLE_WARNINGS |
| 1863 |
__m128i const data = |
| 1864 |
_mm_loadu_si128(reinterpret_cast<__m128i const *>(chars)); |
| 1865 |
|
| 1866 |
// (x - '0') <= 9 |
| 1867 |
// http://0x80.pl/articles/simd-parsing-int-sequences.html |
| 1868 |
__m128i const t0 = _mm_add_epi16(data, _mm_set1_epi16(32720)); |
| 1869 |
__m128i const t1 = _mm_cmpgt_epi16(t0, _mm_set1_epi16(-32759)); |
| 1870 |
|
| 1871 |
if (_mm_movemask_epi8(t1) == 0) { |
| 1872 |
i = i * 100000000 + parse_eight_digits_unrolled(simd_read8_to_u64(data)); |
| 1873 |
return true; |
| 1874 |
} else |
| 1875 |
return false; |
| 1876 |
FASTFLOAT_SIMD_RESTORE_WARNINGS |
| 1877 |
#elif defined(FASTFLOAT_NEON) |
| 1878 |
FASTFLOAT_SIMD_DISABLE_WARNINGS |
| 1879 |
uint16x8_t const data = vld1q_u16(reinterpret_cast<uint16_t const *>(chars)); |
| 1880 |
|
| 1881 |
// (x - '0') <= 9 |
| 1882 |
// http://0x80.pl/articles/simd-parsing-int-sequences.html |
| 1883 |
uint16x8_t const t0 = vsubq_u16(data, vmovq_n_u16('0')); |
| 1884 |
uint16x8_t const mask = vcltq_u16(t0, vmovq_n_u16('9' - '0' + 1)); |
| 1885 |
|
| 1886 |
if (vminvq_u16(mask) == 0xFFFF) { |
| 1887 |
i = i * 100000000 + parse_eight_digits_unrolled(simd_read8_to_u64(data)); |
| 1888 |
return true; |
| 1889 |
} else |
| 1890 |
return false; |
| 1891 |
FASTFLOAT_SIMD_RESTORE_WARNINGS |
| 1892 |
#else |
| 1893 |
(void)chars; |
| 1894 |
(void)i; |
| 1895 |
return false; |
| 1896 |
#endif // FASTFLOAT_SSE2 |
| 1897 |
} |
| 1898 |
|
| 1899 |
#endif // FASTFLOAT_HAS_SIMD |
| 1900 |
|
| 1901 |
// MSVC SFINAE is broken pre-VS2017 |
| 1902 |
#if defined(_MSC_VER) && _MSC_VER <= 1900 |
| 1903 |
template <typename UC> |
| 1904 |
#else |
| 1905 |
template <typename UC, FASTFLOAT_ENABLE_IF(!has_simd_opt<UC>()) = 0> |
| 1906 |
#endif |
| 1907 |
// dummy for compile |
| 1908 |
bool simd_parse_if_eight_digits_unrolled(UC const *, uint64_t &) { |
| 1909 |
return 0; |
| 1910 |
} |
| 1911 |
|
| 1912 |
template <typename UC, FASTFLOAT_ENABLE_IF(!std::is_same<UC, char>::value) = 0> |
| 1913 |
fastfloat_really_inline FASTFLOAT_CONSTEXPR20 void |
| 1914 |
loop_parse_if_eight_digits(UC const *&p, UC const *const pend, uint64_t &i) { |
| 1915 |
if (!has_simd_opt<UC>()) { |
| 1916 |
return; |
| 1917 |
} |
| 1918 |
while ((std::distance(p, pend) >= 8) && |
| 1919 |
simd_parse_if_eight_digits_unrolled( |
| 1920 |
p, i)) { // in rare cases, this will overflow, but that's ok |
| 1921 |
p += 8; |
| 1922 |
} |
| 1923 |
} |
| 1924 |
|
| 1925 |
fastfloat_really_inline FASTFLOAT_CONSTEXPR20 void |
| 1926 |
loop_parse_if_eight_digits(char const *&p, char const *const pend, |
| 1927 |
uint64_t &i) { |
| 1928 |
// optimizes better than parse_if_eight_digits_unrolled() for UC = char. |
| 1929 |
while ((std::distance(p, pend) >= 8) && |
| 1930 |
is_made_of_eight_digits_fast(read8_to_u64(p))) { |
| 1931 |
i = i * 100000000 + |
| 1932 |
parse_eight_digits_unrolled(read8_to_u64( |
| 1933 |
p)); // in rare cases, this will overflow, but that's ok |
| 1934 |
p += 8; |
| 1935 |
} |
| 1936 |
} |
| 1937 |
|
| 1938 |
enum class parse_error { |
| 1939 |
no_error, |
| 1940 |
// [JSON-only] The minus sign must be followed by an integer. |
| 1941 |
missing_integer_after_sign, |
| 1942 |
// A sign must be followed by an integer or dot. |
| 1943 |
missing_integer_or_dot_after_sign, |
| 1944 |
// [JSON-only] The integer part must not have leading zeros. |
| 1945 |
leading_zeros_in_integer_part, |
| 1946 |
// [JSON-only] The integer part must have at least one digit. |
| 1947 |
no_digits_in_integer_part, |
| 1948 |
// [JSON-only] If there is a decimal point, there must be digits in the |
| 1949 |
// fractional part. |
| 1950 |
no_digits_in_fractional_part, |
| 1951 |
// The mantissa must have at least one digit. |
| 1952 |
no_digits_in_mantissa, |
| 1953 |
// Scientific notation requires an exponential part. |
| 1954 |
missing_exponential_part, |
| 1955 |
}; |
| 1956 |
|
| 1957 |
template <typename UC> struct parsed_number_string_t { |
| 1958 |
int64_t exponent{0}; |
| 1959 |
uint64_t mantissa{0}; |
| 1960 |
UC const *lastmatch{nullptr}; |
| 1961 |
bool negative{false}; |
| 1962 |
bool valid{false}; |
| 1963 |
bool too_many_digits{false}; |
| 1964 |
// contains the range of the significant digits |
| 1965 |
span<UC const> integer{}; // non-nullable |
| 1966 |
span<UC const> fraction{}; // nullable |
| 1967 |
parse_error error{parse_error::no_error}; |
| 1968 |
}; |
| 1969 |
|
| 1970 |
using byte_span = span<char const>; |
| 1971 |
using parsed_number_string = parsed_number_string_t<char>; |
| 1972 |
|
| 1973 |
template <typename UC> |
| 1974 |
fastfloat_really_inline FASTFLOAT_CONSTEXPR20 parsed_number_string_t<UC> |
| 1975 |
report_parse_error(UC const *p, parse_error error) { |
| 1976 |
parsed_number_string_t<UC> answer; |
| 1977 |
answer.valid = false; |
| 1978 |
answer.lastmatch = p; |
| 1979 |
answer.error = error; |
| 1980 |
return answer; |
| 1981 |
} |
| 1982 |
|
| 1983 |
// Assuming that you use no more than 19 digits, this will |
| 1984 |
// parse an ASCII string. |
| 1985 |
template <bool basic_json_fmt, typename UC> |
| 1986 |
fastfloat_really_inline FASTFLOAT_CONSTEXPR20 parsed_number_string_t<UC> |
| 1987 |
parse_number_string(UC const *p, UC const *pend, |
| 1988 |
parse_options_t<UC> options) noexcept { |
| 1989 |
chars_format const fmt = detail::adjust_for_feature_macros(options.format); |
| 1990 |
UC const decimal_point = options.decimal_point; |
| 1991 |
|
| 1992 |
parsed_number_string_t<UC> answer; |
| 1993 |
answer.valid = false; |
| 1994 |
answer.too_many_digits = false; |
| 1995 |
// assume p < pend, so dereference without checks; |
| 1996 |
answer.negative = (*p == UC('-')); |
| 1997 |
// C++17 20.19.3.(7.1) explicitly forbids '+' sign here |
| 1998 |
if ((*p == UC('-')) || (uint64_t(fmt & chars_format::allow_leading_plus) && |
| 1999 |
!basic_json_fmt && *p == UC('+'))) { |
| 2000 |
++p; |
| 2001 |
if (p == pend) { |
| 2002 |
return report_parse_error<UC>( |
| 2003 |
p, parse_error::missing_integer_or_dot_after_sign); |
| 2004 |
} |
| 2005 |
FASTFLOAT_IF_CONSTEXPR17(basic_json_fmt) { |
| 2006 |
if (!is_integer(*p)) { // a sign must be followed by an integer |
| 2007 |
return report_parse_error<UC>(p, |
| 2008 |
parse_error::missing_integer_after_sign); |
| 2009 |
} |
| 2010 |
} |
| 2011 |
else { |
| 2012 |
if (!is_integer(*p) && |
| 2013 |
(*p != |
| 2014 |
decimal_point)) { // a sign must be followed by an integer or the dot |
| 2015 |
return report_parse_error<UC>( |
| 2016 |
p, parse_error::missing_integer_or_dot_after_sign); |
| 2017 |
} |
| 2018 |
} |
| 2019 |
} |
| 2020 |
UC const *const start_digits = p; |
| 2021 |
|
| 2022 |
uint64_t i = 0; // an unsigned int avoids signed overflows (which are bad) |
| 2023 |
|
| 2024 |
while ((p != pend) && is_integer(*p)) { |
| 2025 |
// a multiplication by 10 is cheaper than an arbitrary integer |
| 2026 |
// multiplication |
| 2027 |
i = 10 * i + |
| 2028 |
uint64_t(*p - |
| 2029 |
UC('0')); // might overflow, we will handle the overflow later |
| 2030 |
++p; |
| 2031 |
} |
| 2032 |
UC const *const end_of_integer_part = p; |
| 2033 |
int64_t digit_count = int64_t(end_of_integer_part - start_digits); |
| 2034 |
answer.integer = span<UC const>(start_digits, size_t(digit_count)); |
| 2035 |
FASTFLOAT_IF_CONSTEXPR17(basic_json_fmt) { |
| 2036 |
// at least 1 digit in integer part, without leading zeros |
| 2037 |
if (digit_count == 0) { |
| 2038 |
return report_parse_error<UC>(p, parse_error::no_digits_in_integer_part); |
| 2039 |
} |
| 2040 |
if ((start_digits[0] == UC('0') && digit_count > 1)) { |
| 2041 |
return report_parse_error<UC>(start_digits, |
| 2042 |
parse_error::leading_zeros_in_integer_part); |
| 2043 |
} |
| 2044 |
} |
| 2045 |
|
| 2046 |
int64_t exponent = 0; |
| 2047 |
bool const has_decimal_point = (p != pend) && (*p == decimal_point); |
| 2048 |
if (has_decimal_point) { |
| 2049 |
++p; |
| 2050 |
UC const *before = p; |
| 2051 |
// can occur at most twice without overflowing, but let it occur more, since |
| 2052 |
// for integers with many digits, digit parsing is the primary bottleneck. |
| 2053 |
loop_parse_if_eight_digits(p, pend, i); |
| 2054 |
|
| 2055 |
while ((p != pend) && is_integer(*p)) { |
| 2056 |
uint8_t digit = uint8_t(*p - UC('0')); |
| 2057 |
++p; |
| 2058 |
i = i * 10 + digit; // in rare cases, this will overflow, but that's ok |
| 2059 |
} |
| 2060 |
exponent = before - p; |
| 2061 |
answer.fraction = span<UC const>(before, size_t(p - before)); |
| 2062 |
digit_count -= exponent; |
| 2063 |
} |
| 2064 |
FASTFLOAT_IF_CONSTEXPR17(basic_json_fmt) { |
| 2065 |
// at least 1 digit in fractional part |
| 2066 |
if (has_decimal_point && exponent == 0) { |
| 2067 |
return report_parse_error<UC>(p, |
| 2068 |
parse_error::no_digits_in_fractional_part); |
| 2069 |
} |
| 2070 |
} |
| 2071 |
else if (digit_count == 0) { // we must have encountered at least one integer! |
| 2072 |
return report_parse_error<UC>(p, parse_error::no_digits_in_mantissa); |
| 2073 |
} |
| 2074 |
int64_t exp_number = 0; // explicit exponential part |
| 2075 |
if ((uint64_t(fmt & chars_format::scientific) && (p != pend) && |
| 2076 |
((UC('e') == *p) || (UC('E') == *p))) || |
| 2077 |
(uint64_t(fmt & detail::basic_fortran_fmt) && (p != pend) && |
| 2078 |
((UC('+') == *p) || (UC('-') == *p) || (UC('d') == *p) || |
| 2079 |
(UC('D') == *p)))) { |
| 2080 |
UC const *location_of_e = p; |
| 2081 |
if ((UC('e') == *p) || (UC('E') == *p) || (UC('d') == *p) || |
| 2082 |
(UC('D') == *p)) { |
| 2083 |
++p; |
| 2084 |
} |
| 2085 |
bool neg_exp = false; |
| 2086 |
if ((p != pend) && (UC('-') == *p)) { |
| 2087 |
neg_exp = true; |
| 2088 |
++p; |
| 2089 |
} else if ((p != pend) && |
| 2090 |
(UC('+') == |
| 2091 |
*p)) { // '+' on exponent is allowed by C++17 20.19.3.(7.1) |
| 2092 |
++p; |
| 2093 |
} |
| 2094 |
if ((p == pend) || !is_integer(*p)) { |
| 2095 |
if (!uint64_t(fmt & chars_format::fixed)) { |
| 2096 |
// The exponential part is invalid for scientific notation, so it must |
| 2097 |
// be a trailing token for fixed notation. However, fixed notation is |
| 2098 |
// disabled, so report a scientific notation error. |
| 2099 |
return report_parse_error<UC>(p, parse_error::missing_exponential_part); |
| 2100 |
} |
| 2101 |
// Otherwise, we will be ignoring the 'e'. |
| 2102 |
p = location_of_e; |
| 2103 |
} else { |
| 2104 |
while ((p != pend) && is_integer(*p)) { |
| 2105 |
uint8_t digit = uint8_t(*p - UC('0')); |
| 2106 |
if (exp_number < 0x10000000) { |
| 2107 |
exp_number = 10 * exp_number + digit; |
| 2108 |
} |
| 2109 |
++p; |
| 2110 |
} |
| 2111 |
if (neg_exp) { |
| 2112 |
exp_number = -exp_number; |
| 2113 |
} |
| 2114 |
exponent += exp_number; |
| 2115 |
} |
| 2116 |
} else { |
| 2117 |
// If it scientific and not fixed, we have to bail out. |
| 2118 |
if (uint64_t(fmt & chars_format::scientific) && |
| 2119 |
!uint64_t(fmt & chars_format::fixed)) { |
| 2120 |
return report_parse_error<UC>(p, parse_error::missing_exponential_part); |
| 2121 |
} |
| 2122 |
} |
| 2123 |
answer.lastmatch = p; |
| 2124 |
answer.valid = true; |
| 2125 |
|
| 2126 |
// If we frequently had to deal with long strings of digits, |
| 2127 |
// we could extend our code by using a 128-bit integer instead |
| 2128 |
// of a 64-bit integer. However, this is uncommon. |
| 2129 |
// |
| 2130 |
// We can deal with up to 19 digits. |
| 2131 |
if (digit_count > 19) { // this is uncommon |
| 2132 |
// It is possible that the integer had an overflow. |
| 2133 |
// We have to handle the case where we have 0.0000somenumber. |
| 2134 |
// We need to be mindful of the case where we only have zeroes... |
| 2135 |
// E.g., 0.000000000...000. |
| 2136 |
UC const *start = start_digits; |
| 2137 |
while ((start != pend) && (*start == UC('0') || *start == decimal_point)) { |
| 2138 |
if (*start == UC('0')) { |
| 2139 |
digit_count--; |
| 2140 |
} |
| 2141 |
start++; |
| 2142 |
} |
| 2143 |
|
| 2144 |
if (digit_count > 19) { |
| 2145 |
answer.too_many_digits = true; |
| 2146 |
// Let us start again, this time, avoiding overflows. |
| 2147 |
// We don't need to call if is_integer, since we use the |
| 2148 |
// pre-tokenized spans from above. |
| 2149 |
i = 0; |
| 2150 |
p = answer.integer.ptr; |
| 2151 |
UC const *int_end = p + answer.integer.len(); |
| 2152 |
uint64_t const minimal_nineteen_digit_integer{1000000000000000000}; |
| 2153 |
while ((i < minimal_nineteen_digit_integer) && (p != int_end)) { |
| 2154 |
i = i * 10 + uint64_t(*p - UC('0')); |
| 2155 |
++p; |
| 2156 |
} |
| 2157 |
if (i >= minimal_nineteen_digit_integer) { // We have a big integer |
| 2158 |
exponent = end_of_integer_part - p + exp_number; |
| 2159 |
} else { // We have a value with a fractional component. |
| 2160 |
p = answer.fraction.ptr; |
| 2161 |
UC const *frac_end = p + answer.fraction.len(); |
| 2162 |
while ((i < minimal_nineteen_digit_integer) && (p != frac_end)) { |
| 2163 |
i = i * 10 + uint64_t(*p - UC('0')); |
| 2164 |
++p; |
| 2165 |
} |
| 2166 |
exponent = answer.fraction.ptr - p + exp_number; |
| 2167 |
} |
| 2168 |
// We have now corrected both exponent and i, to a truncated value |
| 2169 |
} |
| 2170 |
} |
| 2171 |
answer.exponent = exponent; |
| 2172 |
answer.mantissa = i; |
| 2173 |
return answer; |
| 2174 |
} |
| 2175 |
|
| 2176 |
template <typename T, typename UC> |
| 2177 |
fastfloat_really_inline FASTFLOAT_CONSTEXPR20 from_chars_result_t<UC> |
| 2178 |
parse_int_string(UC const *p, UC const *pend, T &value, |
| 2179 |
parse_options_t<UC> options) { |
| 2180 |
chars_format const fmt = detail::adjust_for_feature_macros(options.format); |
| 2181 |
int const base = options.base; |
| 2182 |
|
| 2183 |
from_chars_result_t<UC> answer; |
| 2184 |
|
| 2185 |
UC const *const first = p; |
| 2186 |
|
| 2187 |
bool const negative = (*p == UC('-')); |
| 2188 |
#ifdef FASTFLOAT_VISUAL_STUDIO |
| 2189 |
#pragma warning(push) |
| 2190 |
#pragma warning(disable : 4127) |
| 2191 |
#endif |
| 2192 |
if (!std::is_signed<T>::value && negative) { |
| 2193 |
#ifdef FASTFLOAT_VISUAL_STUDIO |
| 2194 |
#pragma warning(pop) |
| 2195 |
#endif |
| 2196 |
answer.ec = std::errc::invalid_argument; |
| 2197 |
answer.ptr = first; |
| 2198 |
return answer; |
| 2199 |
} |
| 2200 |
if ((*p == UC('-')) || |
| 2201 |
(uint64_t(fmt & chars_format::allow_leading_plus) && (*p == UC('+')))) { |
| 2202 |
++p; |
| 2203 |
} |
| 2204 |
|
| 2205 |
UC const *const start_num = p; |
| 2206 |
|
| 2207 |
while (p != pend && *p == UC('0')) { |
| 2208 |
++p; |
| 2209 |
} |
| 2210 |
|
| 2211 |
bool const has_leading_zeros = p > start_num; |
| 2212 |
|
| 2213 |
UC const *const start_digits = p; |
| 2214 |
|
| 2215 |
FASTFLOAT_IF_CONSTEXPR17((std::is_same<T, std::uint8_t>::value)) { |
| 2216 |
if (base == 10) { |
| 2217 |
const size_t len = (size_t)(pend - p); |
| 2218 |
if (len == 0) { |
| 2219 |
if (has_leading_zeros) { |
| 2220 |
value = 0; |
| 2221 |
answer.ec = std::errc(); |
| 2222 |
answer.ptr = p; |
| 2223 |
} else { |
| 2224 |
answer.ec = std::errc::invalid_argument; |
| 2225 |
answer.ptr = first; |
| 2226 |
} |
| 2227 |
return answer; |
| 2228 |
} |
| 2229 |
|
| 2230 |
uint32_t digits; |
| 2231 |
|
| 2232 |
#if FASTFLOAT_HAS_IS_CONSTANT_EVALUATED && FASTFLOAT_HAS_BIT_CAST |
| 2233 |
if (std::is_constant_evaluated()) { |
| 2234 |
uint8_t str[4]{}; |
| 2235 |
for (size_t j = 0; j < 4 && j < len; ++j) { |
| 2236 |
str[j] = static_cast<uint8_t>(p[j]); |
| 2237 |
} |
| 2238 |
digits = std::bit_cast<uint32_t>(str); |
| 2239 |
#if FASTFLOAT_IS_BIG_ENDIAN |
| 2240 |
digits = byteswap_32(digits); |
| 2241 |
#endif |
| 2242 |
} |
| 2243 |
#else |
| 2244 |
if (false) { |
| 2245 |
} |
| 2246 |
#endif |
| 2247 |
else if (len >= 4) { |
| 2248 |
::memcpy(&digits, p, 4); |
| 2249 |
#if FASTFLOAT_IS_BIG_ENDIAN |
| 2250 |
digits = byteswap_32(digits); |
| 2251 |
#endif |
| 2252 |
} else { |
| 2253 |
uint32_t b0 = static_cast<uint8_t>(p[0]); |
| 2254 |
uint32_t b1 = (len > 1) ? static_cast<uint8_t>(p[1]) : 0xFFu; |
| 2255 |
uint32_t b2 = (len > 2) ? static_cast<uint8_t>(p[2]) : 0xFFu; |
| 2256 |
uint32_t b3 = 0xFFu; |
| 2257 |
digits = b0 | (b1 << 8) | (b2 << 16) | (b3 << 24); |
| 2258 |
} |
| 2259 |
|
| 2260 |
uint32_t magic = |
| 2261 |
((digits + 0x46464646u) | (digits - 0x30303030u)) & 0x80808080u; |
| 2262 |
uint32_t tz = (uint32_t)countr_zero_32(magic); // 7, 15, 23, 31, or 32 |
| 2263 |
uint32_t nd = (tz == 32) ? 4 : (tz >> 3); |
| 2264 |
nd = (uint32_t)std::min((size_t)nd, len); |
| 2265 |
if (nd == 0) { |
| 2266 |
if (has_leading_zeros) { |
| 2267 |
value = 0; |
| 2268 |
answer.ec = std::errc(); |
| 2269 |
answer.ptr = p; |
| 2270 |
return answer; |
| 2271 |
} |
| 2272 |
answer.ec = std::errc::invalid_argument; |
| 2273 |
answer.ptr = first; |
| 2274 |
return answer; |
| 2275 |
} |
| 2276 |
if (nd > 3) { |
| 2277 |
const UC *q = p + nd; |
| 2278 |
size_t rem = len - nd; |
| 2279 |
while (rem) { |
| 2280 |
if (*q < UC('0') || *q > UC('9')) |
| 2281 |
break; |
| 2282 |
++q; |
| 2283 |
--rem; |
| 2284 |
} |
| 2285 |
answer.ec = std::errc::result_out_of_range; |
| 2286 |
answer.ptr = q; |
| 2287 |
return answer; |
| 2288 |
} |
| 2289 |
|
| 2290 |
digits ^= 0x30303030u; |
| 2291 |
digits <<= ((4 - nd) * 8); |
| 2292 |
|
| 2293 |
uint32_t check = ((digits >> 24) & 0xff) | ((digits >> 8) & 0xff00) | |
| 2294 |
((digits << 8) & 0xff0000); |
| 2295 |
if (check > 0x00020505) { |
| 2296 |
answer.ec = std::errc::result_out_of_range; |
| 2297 |
answer.ptr = p + nd; |
| 2298 |
return answer; |
| 2299 |
} |
| 2300 |
value = (uint8_t)((0x640a01 * digits) >> 24); |
| 2301 |
answer.ec = std::errc(); |
| 2302 |
answer.ptr = p + nd; |
| 2303 |
return answer; |
| 2304 |
} |
| 2305 |
} |
| 2306 |
|
| 2307 |
FASTFLOAT_IF_CONSTEXPR17((std::is_same<T, std::uint16_t>::value)) { |
| 2308 |
if (base == 10) { |
| 2309 |
const size_t len = size_t(pend - p); |
| 2310 |
if (len == 0) { |
| 2311 |
if (has_leading_zeros) { |
| 2312 |
value = 0; |
| 2313 |
answer.ec = std::errc(); |
| 2314 |
answer.ptr = p; |
| 2315 |
} else { |
| 2316 |
answer.ec = std::errc::invalid_argument; |
| 2317 |
answer.ptr = first; |
| 2318 |
} |
| 2319 |
return answer; |
| 2320 |
} |
| 2321 |
|
| 2322 |
if (len >= 4) { |
| 2323 |
uint32_t digits = read4_to_u32(p); |
| 2324 |
if (is_made_of_four_digits_fast(digits)) { |
| 2325 |
uint32_t v = parse_four_digits_unrolled(digits); |
| 2326 |
if (len >= 5 && is_integer(p[4])) { |
| 2327 |
v = v * 10 + uint32_t(p[4] - '0'); |
| 2328 |
if (len >= 6 && is_integer(p[5])) { |
| 2329 |
answer.ec = std::errc::result_out_of_range; |
| 2330 |
const UC *q = p + 5; |
| 2331 |
while (q != pend && is_integer(*q)) { |
| 2332 |
q++; |
| 2333 |
} |
| 2334 |
answer.ptr = q; |
| 2335 |
return answer; |
| 2336 |
} |
| 2337 |
if (v > 65535) { |
| 2338 |
answer.ec = std::errc::result_out_of_range; |
| 2339 |
answer.ptr = p + 5; |
| 2340 |
return answer; |
| 2341 |
} |
| 2342 |
value = uint16_t(v); |
| 2343 |
answer.ec = std::errc(); |
| 2344 |
answer.ptr = p + 5; |
| 2345 |
return answer; |
| 2346 |
} |
| 2347 |
// 4 digits |
| 2348 |
value = uint16_t(v); |
| 2349 |
answer.ec = std::errc(); |
| 2350 |
answer.ptr = p + 4; |
| 2351 |
return answer; |
| 2352 |
} |
| 2353 |
} |
| 2354 |
} |
| 2355 |
} |
| 2356 |
|
| 2357 |
uint64_t i = 0; |
| 2358 |
if (base == 10) { |
| 2359 |
loop_parse_if_eight_digits(p, pend, i); // use SIMD if possible |
| 2360 |
} |
| 2361 |
while (p != pend) { |
| 2362 |
uint8_t digit = ch_to_digit(*p); |
| 2363 |
if (digit >= base) { |
| 2364 |
break; |
| 2365 |
} |
| 2366 |
i = uint64_t(base) * i + digit; // might overflow, check this later |
| 2367 |
p++; |
| 2368 |
} |
| 2369 |
|
| 2370 |
size_t digit_count = size_t(p - start_digits); |
| 2371 |
|
| 2372 |
if (digit_count == 0) { |
| 2373 |
if (has_leading_zeros) { |
| 2374 |
value = 0; |
| 2375 |
answer.ec = std::errc(); |
| 2376 |
answer.ptr = p; |
| 2377 |
} else { |
| 2378 |
answer.ec = std::errc::invalid_argument; |
| 2379 |
answer.ptr = first; |
| 2380 |
} |
| 2381 |
return answer; |
| 2382 |
} |
| 2383 |
|
| 2384 |
answer.ptr = p; |
| 2385 |
|
| 2386 |
// check u64 overflow |
| 2387 |
size_t max_digits = max_digits_u64(base); |
| 2388 |
if (digit_count > max_digits) { |
| 2389 |
answer.ec = std::errc::result_out_of_range; |
| 2390 |
return answer; |
| 2391 |
} |
| 2392 |
// this check can be eliminated for all other types, but they will all require |
| 2393 |
// a max_digits(base) equivalent |
| 2394 |
if (digit_count == max_digits && i < min_safe_u64(base)) { |
| 2395 |
answer.ec = std::errc::result_out_of_range; |
| 2396 |
return answer; |
| 2397 |
} |
| 2398 |
|
| 2399 |
// check other types overflow |
| 2400 |
if (!std::is_same<T, uint64_t>::value) { |
| 2401 |
if (i > uint64_t(std::numeric_limits<T>::max()) + uint64_t(negative)) { |
| 2402 |
answer.ec = std::errc::result_out_of_range; |
| 2403 |
return answer; |
| 2404 |
} |
| 2405 |
} |
| 2406 |
|
| 2407 |
if (negative) { |
| 2408 |
#ifdef FASTFLOAT_VISUAL_STUDIO |
| 2409 |
#pragma warning(push) |
| 2410 |
#pragma warning(disable : 4146) |
| 2411 |
#endif |
| 2412 |
// this weird workaround is required because: |
| 2413 |
// - converting unsigned to signed when its value is greater than signed max |
| 2414 |
// is UB pre-C++23. |
| 2415 |
// - reinterpret_casting (~i + 1) would work, but it is not constexpr |
| 2416 |
// this is always optimized into a neg instruction (note: T is an integer |
| 2417 |
// type) |
| 2418 |
value = T(-std::numeric_limits<T>::max() - |
| 2419 |
T(i - uint64_t(std::numeric_limits<T>::max()))); |
| 2420 |
#ifdef FASTFLOAT_VISUAL_STUDIO |
| 2421 |
#pragma warning(pop) |
| 2422 |
#endif |
| 2423 |
} else { |
| 2424 |
value = T(i); |
| 2425 |
} |
| 2426 |
|
| 2427 |
answer.ec = std::errc(); |
| 2428 |
return answer; |
| 2429 |
} |
| 2430 |
|
| 2431 |
} // namespace fast_float |
| 2432 |
|
| 2433 |
#endif |
| 2434 |
|
| 2435 |
#ifndef FASTFLOAT_FAST_TABLE_H |
| 2436 |
#define FASTFLOAT_FAST_TABLE_H |
| 2437 |
|
| 2438 |
#include <cstdint> |
| 2439 |
|
| 2440 |
namespace fast_float { |
| 2441 |
|
| 2442 |
/** |
| 2443 |
* When mapping numbers from decimal to binary, |
| 2444 |
* we go from w * 10^q to m * 2^p but we have |
| 2445 |
* 10^q = 5^q * 2^q, so effectively |
| 2446 |
* we are trying to match |
| 2447 |
* w * 2^q * 5^q to m * 2^p. Thus the powers of two |
| 2448 |
* are not a concern since they can be represented |
| 2449 |
* exactly using the binary notation, only the powers of five |
| 2450 |
* affect the binary significand. |
| 2451 |
*/ |
| 2452 |
|
| 2453 |
/** |
| 2454 |
* The smallest non-zero float (binary64) is 2^-1074. |
| 2455 |
* We take as input numbers of the form w x 10^q where w < 2^64. |
| 2456 |
* We have that w * 10^-343 < 2^(64-344) 5^-343 < 2^-1076. |
| 2457 |
* However, we have that |
| 2458 |
* (2^64-1) * 10^-342 = (2^64-1) * 2^-342 * 5^-342 > 2^-1074. |
| 2459 |
* Thus it is possible for a number of the form w * 10^-342 where |
| 2460 |
* w is a 64-bit value to be a non-zero floating-point number. |
| 2461 |
********* |
| 2462 |
* Any number of form w * 10^309 where w>= 1 is going to be |
| 2463 |
* infinite in binary64 so we never need to worry about powers |
| 2464 |
* of 5 greater than 308. |
| 2465 |
*/ |
| 2466 |
template <class unused = void> struct powers_template { |
| 2467 |
|
| 2468 |
constexpr static int smallest_power_of_five = |
| 2469 |
binary_format<double>::smallest_power_of_ten(); |
| 2470 |
constexpr static int largest_power_of_five = |
| 2471 |
binary_format<double>::largest_power_of_ten(); |
| 2472 |
constexpr static int number_of_entries = |
| 2473 |
2 * (largest_power_of_five - smallest_power_of_five + 1); |
| 2474 |
// Powers of five from 5^-342 all the way to 5^308 rounded toward one. |
| 2475 |
constexpr static uint64_t power_of_five_128[number_of_entries] = { |
| 2476 |
0xeef453d6923bd65a, 0x113faa2906a13b3f, |
| 2477 |
0x9558b4661b6565f8, 0x4ac7ca59a424c507, |
| 2478 |
0xbaaee17fa23ebf76, 0x5d79bcf00d2df649, |
| 2479 |
0xe95a99df8ace6f53, 0xf4d82c2c107973dc, |
| 2480 |
0x91d8a02bb6c10594, 0x79071b9b8a4be869, |
| 2481 |
0xb64ec836a47146f9, 0x9748e2826cdee284, |
| 2482 |
0xe3e27a444d8d98b7, 0xfd1b1b2308169b25, |
| 2483 |
0x8e6d8c6ab0787f72, 0xfe30f0f5e50e20f7, |
| 2484 |
0xb208ef855c969f4f, 0xbdbd2d335e51a935, |
| 2485 |
0xde8b2b66b3bc4723, 0xad2c788035e61382, |
| 2486 |
0x8b16fb203055ac76, 0x4c3bcb5021afcc31, |
| 2487 |
0xaddcb9e83c6b1793, 0xdf4abe242a1bbf3d, |
| 2488 |
0xd953e8624b85dd78, 0xd71d6dad34a2af0d, |
| 2489 |
0x87d4713d6f33aa6b, 0x8672648c40e5ad68, |
| 2490 |
0xa9c98d8ccb009506, 0x680efdaf511f18c2, |
| 2491 |
0xd43bf0effdc0ba48, 0x212bd1b2566def2, |
| 2492 |
0x84a57695fe98746d, 0x14bb630f7604b57, |
| 2493 |
0xa5ced43b7e3e9188, 0x419ea3bd35385e2d, |
| 2494 |
0xcf42894a5dce35ea, 0x52064cac828675b9, |
| 2495 |
0x818995ce7aa0e1b2, 0x7343efebd1940993, |
| 2496 |
0xa1ebfb4219491a1f, 0x1014ebe6c5f90bf8, |
| 2497 |
0xca66fa129f9b60a6, 0xd41a26e077774ef6, |
| 2498 |
0xfd00b897478238d0, 0x8920b098955522b4, |
| 2499 |
0x9e20735e8cb16382, 0x55b46e5f5d5535b0, |
| 2500 |
0xc5a890362fddbc62, 0xeb2189f734aa831d, |
| 2501 |
0xf712b443bbd52b7b, 0xa5e9ec7501d523e4, |
| 2502 |
0x9a6bb0aa55653b2d, 0x47b233c92125366e, |
| 2503 |
0xc1069cd4eabe89f8, 0x999ec0bb696e840a, |
| 2504 |
0xf148440a256e2c76, 0xc00670ea43ca250d, |
| 2505 |
0x96cd2a865764dbca, 0x380406926a5e5728, |
| 2506 |
0xbc807527ed3e12bc, 0xc605083704f5ecf2, |
| 2507 |
0xeba09271e88d976b, 0xf7864a44c633682e, |
| 2508 |
0x93445b8731587ea3, 0x7ab3ee6afbe0211d, |
| 2509 |
0xb8157268fdae9e4c, 0x5960ea05bad82964, |
| 2510 |
0xe61acf033d1a45df, 0x6fb92487298e33bd, |
| 2511 |
0x8fd0c16206306bab, 0xa5d3b6d479f8e056, |
| 2512 |
0xb3c4f1ba87bc8696, 0x8f48a4899877186c, |
| 2513 |
0xe0b62e2929aba83c, 0x331acdabfe94de87, |
| 2514 |
0x8c71dcd9ba0b4925, 0x9ff0c08b7f1d0b14, |
| 2515 |
0xaf8e5410288e1b6f, 0x7ecf0ae5ee44dd9, |
| 2516 |
0xdb71e91432b1a24a, 0xc9e82cd9f69d6150, |
| 2517 |
0x892731ac9faf056e, 0xbe311c083a225cd2, |
| 2518 |
0xab70fe17c79ac6ca, 0x6dbd630a48aaf406, |
| 2519 |
0xd64d3d9db981787d, 0x92cbbccdad5b108, |
| 2520 |
0x85f0468293f0eb4e, 0x25bbf56008c58ea5, |
| 2521 |
0xa76c582338ed2621, 0xaf2af2b80af6f24e, |
| 2522 |
0xd1476e2c07286faa, 0x1af5af660db4aee1, |
| 2523 |
0x82cca4db847945ca, 0x50d98d9fc890ed4d, |
| 2524 |
0xa37fce126597973c, 0xe50ff107bab528a0, |
| 2525 |
0xcc5fc196fefd7d0c, 0x1e53ed49a96272c8, |
| 2526 |
0xff77b1fcbebcdc4f, 0x25e8e89c13bb0f7a, |
| 2527 |
0x9faacf3df73609b1, 0x77b191618c54e9ac, |
| 2528 |
0xc795830d75038c1d, 0xd59df5b9ef6a2417, |
| 2529 |
0xf97ae3d0d2446f25, 0x4b0573286b44ad1d, |
| 2530 |
0x9becce62836ac577, 0x4ee367f9430aec32, |
| 2531 |
0xc2e801fb244576d5, 0x229c41f793cda73f, |
| 2532 |
0xf3a20279ed56d48a, 0x6b43527578c1110f, |
| 2533 |
0x9845418c345644d6, 0x830a13896b78aaa9, |
| 2534 |
0xbe5691ef416bd60c, 0x23cc986bc656d553, |
| 2535 |
0xedec366b11c6cb8f, 0x2cbfbe86b7ec8aa8, |
| 2536 |
0x94b3a202eb1c3f39, 0x7bf7d71432f3d6a9, |
| 2537 |
0xb9e08a83a5e34f07, 0xdaf5ccd93fb0cc53, |
| 2538 |
0xe858ad248f5c22c9, 0xd1b3400f8f9cff68, |
| 2539 |
0x91376c36d99995be, 0x23100809b9c21fa1, |
| 2540 |
0xb58547448ffffb2d, 0xabd40a0c2832a78a, |
| 2541 |
0xe2e69915b3fff9f9, 0x16c90c8f323f516c, |
| 2542 |
0x8dd01fad907ffc3b, 0xae3da7d97f6792e3, |
| 2543 |
0xb1442798f49ffb4a, 0x99cd11cfdf41779c, |
| 2544 |
0xdd95317f31c7fa1d, 0x40405643d711d583, |
| 2545 |
0x8a7d3eef7f1cfc52, 0x482835ea666b2572, |
| 2546 |
0xad1c8eab5ee43b66, 0xda3243650005eecf, |
| 2547 |
0xd863b256369d4a40, 0x90bed43e40076a82, |
| 2548 |
0x873e4f75e2224e68, 0x5a7744a6e804a291, |
| 2549 |
0xa90de3535aaae202, 0x711515d0a205cb36, |
| 2550 |
0xd3515c2831559a83, 0xd5a5b44ca873e03, |
| 2551 |
0x8412d9991ed58091, 0xe858790afe9486c2, |
| 2552 |
0xa5178fff668ae0b6, 0x626e974dbe39a872, |
| 2553 |
0xce5d73ff402d98e3, 0xfb0a3d212dc8128f, |
| 2554 |
0x80fa687f881c7f8e, 0x7ce66634bc9d0b99, |
| 2555 |
0xa139029f6a239f72, 0x1c1fffc1ebc44e80, |
| 2556 |
0xc987434744ac874e, 0xa327ffb266b56220, |
| 2557 |
0xfbe9141915d7a922, 0x4bf1ff9f0062baa8, |
| 2558 |
0x9d71ac8fada6c9b5, 0x6f773fc3603db4a9, |
| 2559 |
0xc4ce17b399107c22, 0xcb550fb4384d21d3, |
| 2560 |
0xf6019da07f549b2b, 0x7e2a53a146606a48, |
| 2561 |
0x99c102844f94e0fb, 0x2eda7444cbfc426d, |
| 2562 |
0xc0314325637a1939, 0xfa911155fefb5308, |
| 2563 |
0xf03d93eebc589f88, 0x793555ab7eba27ca, |
| 2564 |
0x96267c7535b763b5, 0x4bc1558b2f3458de, |
| 2565 |
0xbbb01b9283253ca2, 0x9eb1aaedfb016f16, |
| 2566 |
0xea9c227723ee8bcb, 0x465e15a979c1cadc, |
| 2567 |
0x92a1958a7675175f, 0xbfacd89ec191ec9, |
| 2568 |
0xb749faed14125d36, 0xcef980ec671f667b, |
| 2569 |
0xe51c79a85916f484, 0x82b7e12780e7401a, |
| 2570 |
0x8f31cc0937ae58d2, 0xd1b2ecb8b0908810, |
| 2571 |
0xb2fe3f0b8599ef07, 0x861fa7e6dcb4aa15, |
| 2572 |
0xdfbdcece67006ac9, 0x67a791e093e1d49a, |
| 2573 |
0x8bd6a141006042bd, 0xe0c8bb2c5c6d24e0, |
| 2574 |
0xaecc49914078536d, 0x58fae9f773886e18, |
| 2575 |
0xda7f5bf590966848, 0xaf39a475506a899e, |
| 2576 |
0x888f99797a5e012d, 0x6d8406c952429603, |
| 2577 |
0xaab37fd7d8f58178, 0xc8e5087ba6d33b83, |
| 2578 |
0xd5605fcdcf32e1d6, 0xfb1e4a9a90880a64, |
| 2579 |
0x855c3be0a17fcd26, 0x5cf2eea09a55067f, |
| 2580 |
0xa6b34ad8c9dfc06f, 0xf42faa48c0ea481e, |
| 2581 |
0xd0601d8efc57b08b, 0xf13b94daf124da26, |
| 2582 |
0x823c12795db6ce57, 0x76c53d08d6b70858, |
| 2583 |
0xa2cb1717b52481ed, 0x54768c4b0c64ca6e, |
| 2584 |
0xcb7ddcdda26da268, 0xa9942f5dcf7dfd09, |
| 2585 |
0xfe5d54150b090b02, 0xd3f93b35435d7c4c, |
| 2586 |
0x9efa548d26e5a6e1, 0xc47bc5014a1a6daf, |
| 2587 |
0xc6b8e9b0709f109a, 0x359ab6419ca1091b, |
| 2588 |
0xf867241c8cc6d4c0, 0xc30163d203c94b62, |
| 2589 |
0x9b407691d7fc44f8, 0x79e0de63425dcf1d, |
| 2590 |
0xc21094364dfb5636, 0x985915fc12f542e4, |
| 2591 |
0xf294b943e17a2bc4, 0x3e6f5b7b17b2939d, |
| 2592 |
0x979cf3ca6cec5b5a, 0xa705992ceecf9c42, |
| 2593 |
0xbd8430bd08277231, 0x50c6ff782a838353, |
| 2594 |
0xece53cec4a314ebd, 0xa4f8bf5635246428, |
| 2595 |
0x940f4613ae5ed136, 0x871b7795e136be99, |
| 2596 |
0xb913179899f68584, 0x28e2557b59846e3f, |
| 2597 |
0xe757dd7ec07426e5, 0x331aeada2fe589cf, |
| 2598 |
0x9096ea6f3848984f, 0x3ff0d2c85def7621, |
| 2599 |
0xb4bca50b065abe63, 0xfed077a756b53a9, |
| 2600 |
0xe1ebce4dc7f16dfb, 0xd3e8495912c62894, |
| 2601 |
0x8d3360f09cf6e4bd, 0x64712dd7abbbd95c, |
| 2602 |
0xb080392cc4349dec, 0xbd8d794d96aacfb3, |
| 2603 |
0xdca04777f541c567, 0xecf0d7a0fc5583a0, |
| 2604 |
0x89e42caaf9491b60, 0xf41686c49db57244, |
| 2605 |
0xac5d37d5b79b6239, 0x311c2875c522ced5, |
| 2606 |
0xd77485cb25823ac7, 0x7d633293366b828b, |
| 2607 |
0x86a8d39ef77164bc, 0xae5dff9c02033197, |
| 2608 |
0xa8530886b54dbdeb, 0xd9f57f830283fdfc, |
| 2609 |
0xd267caa862a12d66, 0xd072df63c324fd7b, |
| 2610 |
0x8380dea93da4bc60, 0x4247cb9e59f71e6d, |
| 2611 |
0xa46116538d0deb78, 0x52d9be85f074e608, |
| 2612 |
0xcd795be870516656, 0x67902e276c921f8b, |
| 2613 |
0x806bd9714632dff6, 0xba1cd8a3db53b6, |
| 2614 |
0xa086cfcd97bf97f3, 0x80e8a40eccd228a4, |
| 2615 |
0xc8a883c0fdaf7df0, 0x6122cd128006b2cd, |
| 2616 |
0xfad2a4b13d1b5d6c, 0x796b805720085f81, |
| 2617 |
0x9cc3a6eec6311a63, 0xcbe3303674053bb0, |
| 2618 |
0xc3f490aa77bd60fc, 0xbedbfc4411068a9c, |
| 2619 |
0xf4f1b4d515acb93b, 0xee92fb5515482d44, |
| 2620 |
0x991711052d8bf3c5, 0x751bdd152d4d1c4a, |
| 2621 |
0xbf5cd54678eef0b6, 0xd262d45a78a0635d, |
| 2622 |
0xef340a98172aace4, 0x86fb897116c87c34, |
| 2623 |
0x9580869f0e7aac0e, 0xd45d35e6ae3d4da0, |
| 2624 |
0xbae0a846d2195712, 0x8974836059cca109, |
| 2625 |
0xe998d258869facd7, 0x2bd1a438703fc94b, |
| 2626 |
0x91ff83775423cc06, 0x7b6306a34627ddcf, |
| 2627 |
0xb67f6455292cbf08, 0x1a3bc84c17b1d542, |
| 2628 |
0xe41f3d6a7377eeca, 0x20caba5f1d9e4a93, |
| 2629 |
0x8e938662882af53e, 0x547eb47b7282ee9c, |
| 2630 |
0xb23867fb2a35b28d, 0xe99e619a4f23aa43, |
| 2631 |
0xdec681f9f4c31f31, 0x6405fa00e2ec94d4, |
| 2632 |
0x8b3c113c38f9f37e, 0xde83bc408dd3dd04, |
| 2633 |
0xae0b158b4738705e, 0x9624ab50b148d445, |
| 2634 |
0xd98ddaee19068c76, 0x3badd624dd9b0957, |
| 2635 |
0x87f8a8d4cfa417c9, 0xe54ca5d70a80e5d6, |
| 2636 |
0xa9f6d30a038d1dbc, 0x5e9fcf4ccd211f4c, |
| 2637 |
0xd47487cc8470652b, 0x7647c3200069671f, |
| 2638 |
0x84c8d4dfd2c63f3b, 0x29ecd9f40041e073, |
| 2639 |
0xa5fb0a17c777cf09, 0xf468107100525890, |
| 2640 |
0xcf79cc9db955c2cc, 0x7182148d4066eeb4, |
| 2641 |
0x81ac1fe293d599bf, 0xc6f14cd848405530, |
| 2642 |
0xa21727db38cb002f, 0xb8ada00e5a506a7c, |
| 2643 |
0xca9cf1d206fdc03b, 0xa6d90811f0e4851c, |
| 2644 |
0xfd442e4688bd304a, 0x908f4a166d1da663, |
| 2645 |
0x9e4a9cec15763e2e, 0x9a598e4e043287fe, |
| 2646 |
0xc5dd44271ad3cdba, 0x40eff1e1853f29fd, |
| 2647 |
0xf7549530e188c128, 0xd12bee59e68ef47c, |
| 2648 |
0x9a94dd3e8cf578b9, 0x82bb74f8301958ce, |
| 2649 |
0xc13a148e3032d6e7, 0xe36a52363c1faf01, |
| 2650 |
0xf18899b1bc3f8ca1, 0xdc44e6c3cb279ac1, |
| 2651 |
0x96f5600f15a7b7e5, 0x29ab103a5ef8c0b9, |
| 2652 |
0xbcb2b812db11a5de, 0x7415d448f6b6f0e7, |
| 2653 |
0xebdf661791d60f56, 0x111b495b3464ad21, |
| 2654 |
0x936b9fcebb25c995, 0xcab10dd900beec34, |
| 2655 |
0xb84687c269ef3bfb, 0x3d5d514f40eea742, |
| 2656 |
0xe65829b3046b0afa, 0xcb4a5a3112a5112, |
| 2657 |
0x8ff71a0fe2c2e6dc, 0x47f0e785eaba72ab, |
| 2658 |
0xb3f4e093db73a093, 0x59ed216765690f56, |
| 2659 |
0xe0f218b8d25088b8, 0x306869c13ec3532c, |
| 2660 |
0x8c974f7383725573, 0x1e414218c73a13fb, |
| 2661 |
0xafbd2350644eeacf, 0xe5d1929ef90898fa, |
| 2662 |
0xdbac6c247d62a583, 0xdf45f746b74abf39, |
| 2663 |
0x894bc396ce5da772, 0x6b8bba8c328eb783, |
| 2664 |
0xab9eb47c81f5114f, 0x66ea92f3f326564, |
| 2665 |
0xd686619ba27255a2, 0xc80a537b0efefebd, |
| 2666 |
0x8613fd0145877585, 0xbd06742ce95f5f36, |
| 2667 |
0xa798fc4196e952e7, 0x2c48113823b73704, |
| 2668 |
0xd17f3b51fca3a7a0, 0xf75a15862ca504c5, |
| 2669 |
0x82ef85133de648c4, 0x9a984d73dbe722fb, |
| 2670 |
0xa3ab66580d5fdaf5, 0xc13e60d0d2e0ebba, |
| 2671 |
0xcc963fee10b7d1b3, 0x318df905079926a8, |
| 2672 |
0xffbbcfe994e5c61f, 0xfdf17746497f7052, |
| 2673 |
0x9fd561f1fd0f9bd3, 0xfeb6ea8bedefa633, |
| 2674 |
0xc7caba6e7c5382c8, 0xfe64a52ee96b8fc0, |
| 2675 |
0xf9bd690a1b68637b, 0x3dfdce7aa3c673b0, |
| 2676 |
0x9c1661a651213e2d, 0x6bea10ca65c084e, |
| 2677 |
0xc31bfa0fe5698db8, 0x486e494fcff30a62, |
| 2678 |
0xf3e2f893dec3f126, 0x5a89dba3c3efccfa, |
| 2679 |
0x986ddb5c6b3a76b7, 0xf89629465a75e01c, |
| 2680 |
0xbe89523386091465, 0xf6bbb397f1135823, |
| 2681 |
0xee2ba6c0678b597f, 0x746aa07ded582e2c, |
| 2682 |
0x94db483840b717ef, 0xa8c2a44eb4571cdc, |
| 2683 |
0xba121a4650e4ddeb, 0x92f34d62616ce413, |
| 2684 |
0xe896a0d7e51e1566, 0x77b020baf9c81d17, |
| 2685 |
0x915e2486ef32cd60, 0xace1474dc1d122e, |
| 2686 |
0xb5b5ada8aaff80b8, 0xd819992132456ba, |
| 2687 |
0xe3231912d5bf60e6, 0x10e1fff697ed6c69, |
| 2688 |
0x8df5efabc5979c8f, 0xca8d3ffa1ef463c1, |
| 2689 |
0xb1736b96b6fd83b3, 0xbd308ff8a6b17cb2, |
| 2690 |
0xddd0467c64bce4a0, 0xac7cb3f6d05ddbde, |
| 2691 |
0x8aa22c0dbef60ee4, 0x6bcdf07a423aa96b, |
| 2692 |
0xad4ab7112eb3929d, 0x86c16c98d2c953c6, |
| 2693 |
0xd89d64d57a607744, 0xe871c7bf077ba8b7, |
| 2694 |
0x87625f056c7c4a8b, 0x11471cd764ad4972, |
| 2695 |
0xa93af6c6c79b5d2d, 0xd598e40d3dd89bcf, |
| 2696 |
0xd389b47879823479, 0x4aff1d108d4ec2c3, |
| 2697 |
0x843610cb4bf160cb, 0xcedf722a585139ba, |
| 2698 |
0xa54394fe1eedb8fe, 0xc2974eb4ee658828, |
| 2699 |
0xce947a3da6a9273e, 0x733d226229feea32, |
| 2700 |
0x811ccc668829b887, 0x806357d5a3f525f, |
| 2701 |
0xa163ff802a3426a8, 0xca07c2dcb0cf26f7, |
| 2702 |
0xc9bcff6034c13052, 0xfc89b393dd02f0b5, |
| 2703 |
0xfc2c3f3841f17c67, 0xbbac2078d443ace2, |
| 2704 |
0x9d9ba7832936edc0, 0xd54b944b84aa4c0d, |
| 2705 |
0xc5029163f384a931, 0xa9e795e65d4df11, |
| 2706 |
0xf64335bcf065d37d, 0x4d4617b5ff4a16d5, |
| 2707 |
0x99ea0196163fa42e, 0x504bced1bf8e4e45, |
| 2708 |
0xc06481fb9bcf8d39, 0xe45ec2862f71e1d6, |
| 2709 |
0xf07da27a82c37088, 0x5d767327bb4e5a4c, |
| 2710 |
0x964e858c91ba2655, 0x3a6a07f8d510f86f, |
| 2711 |
0xbbe226efb628afea, 0x890489f70a55368b, |
| 2712 |
0xeadab0aba3b2dbe5, 0x2b45ac74ccea842e, |
| 2713 |
0x92c8ae6b464fc96f, 0x3b0b8bc90012929d, |
| 2714 |
0xb77ada0617e3bbcb, 0x9ce6ebb40173744, |
| 2715 |
0xe55990879ddcaabd, 0xcc420a6a101d0515, |
| 2716 |
0x8f57fa54c2a9eab6, 0x9fa946824a12232d, |
| 2717 |
0xb32df8e9f3546564, 0x47939822dc96abf9, |
| 2718 |
0xdff9772470297ebd, 0x59787e2b93bc56f7, |
| 2719 |
0x8bfbea76c619ef36, 0x57eb4edb3c55b65a, |
| 2720 |
0xaefae51477a06b03, 0xede622920b6b23f1, |
| 2721 |
0xdab99e59958885c4, 0xe95fab368e45eced, |
| 2722 |
0x88b402f7fd75539b, 0x11dbcb0218ebb414, |
| 2723 |
0xaae103b5fcd2a881, 0xd652bdc29f26a119, |
| 2724 |
0xd59944a37c0752a2, 0x4be76d3346f0495f, |
| 2725 |
0x857fcae62d8493a5, 0x6f70a4400c562ddb, |
| 2726 |
0xa6dfbd9fb8e5b88e, 0xcb4ccd500f6bb952, |
| 2727 |
0xd097ad07a71f26b2, 0x7e2000a41346a7a7, |
| 2728 |
0x825ecc24c873782f, 0x8ed400668c0c28c8, |
| 2729 |
0xa2f67f2dfa90563b, 0x728900802f0f32fa, |
| 2730 |
0xcbb41ef979346bca, 0x4f2b40a03ad2ffb9, |
| 2731 |
0xfea126b7d78186bc, 0xe2f610c84987bfa8, |
| 2732 |
0x9f24b832e6b0f436, 0xdd9ca7d2df4d7c9, |
| 2733 |
0xc6ede63fa05d3143, 0x91503d1c79720dbb, |
| 2734 |
0xf8a95fcf88747d94, 0x75a44c6397ce912a, |
| 2735 |
0x9b69dbe1b548ce7c, 0xc986afbe3ee11aba, |
| 2736 |
0xc24452da229b021b, 0xfbe85badce996168, |
| 2737 |
0xf2d56790ab41c2a2, 0xfae27299423fb9c3, |
| 2738 |
0x97c560ba6b0919a5, 0xdccd879fc967d41a, |
| 2739 |
0xbdb6b8e905cb600f, 0x5400e987bbc1c920, |
| 2740 |
0xed246723473e3813, 0x290123e9aab23b68, |
| 2741 |
0x9436c0760c86e30b, 0xf9a0b6720aaf6521, |
| 2742 |
0xb94470938fa89bce, 0xf808e40e8d5b3e69, |
| 2743 |
0xe7958cb87392c2c2, 0xb60b1d1230b20e04, |
| 2744 |
0x90bd77f3483bb9b9, 0xb1c6f22b5e6f48c2, |
| 2745 |
0xb4ecd5f01a4aa828, 0x1e38aeb6360b1af3, |
| 2746 |
0xe2280b6c20dd5232, 0x25c6da63c38de1b0, |
| 2747 |
0x8d590723948a535f, 0x579c487e5a38ad0e, |
| 2748 |
0xb0af48ec79ace837, 0x2d835a9df0c6d851, |
| 2749 |
0xdcdb1b2798182244, 0xf8e431456cf88e65, |
| 2750 |
0x8a08f0f8bf0f156b, 0x1b8e9ecb641b58ff, |
| 2751 |
0xac8b2d36eed2dac5, 0xe272467e3d222f3f, |
| 2752 |
0xd7adf884aa879177, 0x5b0ed81dcc6abb0f, |
| 2753 |
0x86ccbb52ea94baea, 0x98e947129fc2b4e9, |
| 2754 |
0xa87fea27a539e9a5, 0x3f2398d747b36224, |
| 2755 |
0xd29fe4b18e88640e, 0x8eec7f0d19a03aad, |
| 2756 |
0x83a3eeeef9153e89, 0x1953cf68300424ac, |
| 2757 |
0xa48ceaaab75a8e2b, 0x5fa8c3423c052dd7, |
| 2758 |
0xcdb02555653131b6, 0x3792f412cb06794d, |
| 2759 |
0x808e17555f3ebf11, 0xe2bbd88bbee40bd0, |
| 2760 |
0xa0b19d2ab70e6ed6, 0x5b6aceaeae9d0ec4, |
| 2761 |
0xc8de047564d20a8b, 0xf245825a5a445275, |
| 2762 |
0xfb158592be068d2e, 0xeed6e2f0f0d56712, |
| 2763 |
0x9ced737bb6c4183d, 0x55464dd69685606b, |
| 2764 |
0xc428d05aa4751e4c, 0xaa97e14c3c26b886, |
| 2765 |
0xf53304714d9265df, 0xd53dd99f4b3066a8, |
| 2766 |
0x993fe2c6d07b7fab, 0xe546a8038efe4029, |
| 2767 |
0xbf8fdb78849a5f96, 0xde98520472bdd033, |
| 2768 |
0xef73d256a5c0f77c, 0x963e66858f6d4440, |
| 2769 |
0x95a8637627989aad, 0xdde7001379a44aa8, |
| 2770 |
0xbb127c53b17ec159, 0x5560c018580d5d52, |
| 2771 |
0xe9d71b689dde71af, 0xaab8f01e6e10b4a6, |
| 2772 |
0x9226712162ab070d, 0xcab3961304ca70e8, |
| 2773 |
0xb6b00d69bb55c8d1, 0x3d607b97c5fd0d22, |
| 2774 |
0xe45c10c42a2b3b05, 0x8cb89a7db77c506a, |
| 2775 |
0x8eb98a7a9a5b04e3, 0x77f3608e92adb242, |
| 2776 |
0xb267ed1940f1c61c, 0x55f038b237591ed3, |
| 2777 |
0xdf01e85f912e37a3, 0x6b6c46dec52f6688, |
| 2778 |
0x8b61313bbabce2c6, 0x2323ac4b3b3da015, |
| 2779 |
0xae397d8aa96c1b77, 0xabec975e0a0d081a, |
| 2780 |
0xd9c7dced53c72255, 0x96e7bd358c904a21, |
| 2781 |
0x881cea14545c7575, 0x7e50d64177da2e54, |
| 2782 |
0xaa242499697392d2, 0xdde50bd1d5d0b9e9, |
| 2783 |
0xd4ad2dbfc3d07787, 0x955e4ec64b44e864, |
| 2784 |
0x84ec3c97da624ab4, 0xbd5af13bef0b113e, |
| 2785 |
0xa6274bbdd0fadd61, 0xecb1ad8aeacdd58e, |
| 2786 |
0xcfb11ead453994ba, 0x67de18eda5814af2, |
| 2787 |
0x81ceb32c4b43fcf4, 0x80eacf948770ced7, |
| 2788 |
0xa2425ff75e14fc31, 0xa1258379a94d028d, |
| 2789 |
0xcad2f7f5359a3b3e, 0x96ee45813a04330, |
| 2790 |
0xfd87b5f28300ca0d, 0x8bca9d6e188853fc, |
| 2791 |
0x9e74d1b791e07e48, 0x775ea264cf55347e, |
| 2792 |
0xc612062576589dda, 0x95364afe032a819e, |
| 2793 |
0xf79687aed3eec551, 0x3a83ddbd83f52205, |
| 2794 |
0x9abe14cd44753b52, 0xc4926a9672793543, |
| 2795 |
0xc16d9a0095928a27, 0x75b7053c0f178294, |
| 2796 |
0xf1c90080baf72cb1, 0x5324c68b12dd6339, |
| 2797 |
0x971da05074da7bee, 0xd3f6fc16ebca5e04, |
| 2798 |
0xbce5086492111aea, 0x88f4bb1ca6bcf585, |
| 2799 |
0xec1e4a7db69561a5, 0x2b31e9e3d06c32e6, |
| 2800 |
0x9392ee8e921d5d07, 0x3aff322e62439fd0, |
| 2801 |
0xb877aa3236a4b449, 0x9befeb9fad487c3, |
| 2802 |
0xe69594bec44de15b, 0x4c2ebe687989a9b4, |
| 2803 |
0x901d7cf73ab0acd9, 0xf9d37014bf60a11, |
| 2804 |
0xb424dc35095cd80f, 0x538484c19ef38c95, |
| 2805 |
0xe12e13424bb40e13, 0x2865a5f206b06fba, |
| 2806 |
0x8cbccc096f5088cb, 0xf93f87b7442e45d4, |
| 2807 |
0xafebff0bcb24aafe, 0xf78f69a51539d749, |
| 2808 |
0xdbe6fecebdedd5be, 0xb573440e5a884d1c, |
| 2809 |
0x89705f4136b4a597, 0x31680a88f8953031, |
| 2810 |
0xabcc77118461cefc, 0xfdc20d2b36ba7c3e, |
| 2811 |
0xd6bf94d5e57a42bc, 0x3d32907604691b4d, |
| 2812 |
0x8637bd05af6c69b5, 0xa63f9a49c2c1b110, |
| 2813 |
0xa7c5ac471b478423, 0xfcf80dc33721d54, |
| 2814 |
0xd1b71758e219652b, 0xd3c36113404ea4a9, |
| 2815 |
0x83126e978d4fdf3b, 0x645a1cac083126ea, |
| 2816 |
0xa3d70a3d70a3d70a, 0x3d70a3d70a3d70a4, |
| 2817 |
0xcccccccccccccccc, 0xcccccccccccccccd, |
| 2818 |
0x8000000000000000, 0x0, |
| 2819 |
0xa000000000000000, 0x0, |
| 2820 |
0xc800000000000000, 0x0, |
| 2821 |
0xfa00000000000000, 0x0, |
| 2822 |
0x9c40000000000000, 0x0, |
| 2823 |
0xc350000000000000, 0x0, |
| 2824 |
0xf424000000000000, 0x0, |
| 2825 |
0x9896800000000000, 0x0, |
| 2826 |
0xbebc200000000000, 0x0, |
| 2827 |
0xee6b280000000000, 0x0, |
| 2828 |
0x9502f90000000000, 0x0, |
| 2829 |
0xba43b74000000000, 0x0, |
| 2830 |
0xe8d4a51000000000, 0x0, |
| 2831 |
0x9184e72a00000000, 0x0, |
| 2832 |
0xb5e620f480000000, 0x0, |
| 2833 |
0xe35fa931a0000000, 0x0, |
| 2834 |
0x8e1bc9bf04000000, 0x0, |
| 2835 |
0xb1a2bc2ec5000000, 0x0, |
| 2836 |
0xde0b6b3a76400000, 0x0, |
| 2837 |
0x8ac7230489e80000, 0x0, |
| 2838 |
0xad78ebc5ac620000, 0x0, |
| 2839 |
0xd8d726b7177a8000, 0x0, |
| 2840 |
0x878678326eac9000, 0x0, |
| 2841 |
0xa968163f0a57b400, 0x0, |
| 2842 |
0xd3c21bcecceda100, 0x0, |
| 2843 |
0x84595161401484a0, 0x0, |
| 2844 |
0xa56fa5b99019a5c8, 0x0, |
| 2845 |
0xcecb8f27f4200f3a, 0x0, |
| 2846 |
0x813f3978f8940984, 0x4000000000000000, |
| 2847 |
0xa18f07d736b90be5, 0x5000000000000000, |
| 2848 |
0xc9f2c9cd04674ede, 0xa400000000000000, |
| 2849 |
0xfc6f7c4045812296, 0x4d00000000000000, |
| 2850 |
0x9dc5ada82b70b59d, 0xf020000000000000, |
| 2851 |
0xc5371912364ce305, 0x6c28000000000000, |
| 2852 |
0xf684df56c3e01bc6, 0xc732000000000000, |
| 2853 |
0x9a130b963a6c115c, 0x3c7f400000000000, |
| 2854 |
0xc097ce7bc90715b3, 0x4b9f100000000000, |
| 2855 |
0xf0bdc21abb48db20, 0x1e86d40000000000, |
| 2856 |
0x96769950b50d88f4, 0x1314448000000000, |
| 2857 |
0xbc143fa4e250eb31, 0x17d955a000000000, |
| 2858 |
0xeb194f8e1ae525fd, 0x5dcfab0800000000, |
| 2859 |
0x92efd1b8d0cf37be, 0x5aa1cae500000000, |
| 2860 |
0xb7abc627050305ad, 0xf14a3d9e40000000, |
| 2861 |
0xe596b7b0c643c719, 0x6d9ccd05d0000000, |
| 2862 |
0x8f7e32ce7bea5c6f, 0xe4820023a2000000, |
| 2863 |
0xb35dbf821ae4f38b, 0xdda2802c8a800000, |
| 2864 |
0xe0352f62a19e306e, 0xd50b2037ad200000, |
| 2865 |
0x8c213d9da502de45, 0x4526f422cc340000, |
| 2866 |
0xaf298d050e4395d6, 0x9670b12b7f410000, |
| 2867 |
0xdaf3f04651d47b4c, 0x3c0cdd765f114000, |
| 2868 |
0x88d8762bf324cd0f, 0xa5880a69fb6ac800, |
| 2869 |
0xab0e93b6efee0053, 0x8eea0d047a457a00, |
| 2870 |
0xd5d238a4abe98068, 0x72a4904598d6d880, |
| 2871 |
0x85a36366eb71f041, 0x47a6da2b7f864750, |
| 2872 |
0xa70c3c40a64e6c51, 0x999090b65f67d924, |
| 2873 |
0xd0cf4b50cfe20765, 0xfff4b4e3f741cf6d, |
| 2874 |
0x82818f1281ed449f, 0xbff8f10e7a8921a4, |
| 2875 |
0xa321f2d7226895c7, 0xaff72d52192b6a0d, |
| 2876 |
0xcbea6f8ceb02bb39, 0x9bf4f8a69f764490, |
| 2877 |
0xfee50b7025c36a08, 0x2f236d04753d5b4, |
| 2878 |
0x9f4f2726179a2245, 0x1d762422c946590, |
| 2879 |
0xc722f0ef9d80aad6, 0x424d3ad2b7b97ef5, |
| 2880 |
0xf8ebad2b84e0d58b, 0xd2e0898765a7deb2, |
| 2881 |
0x9b934c3b330c8577, 0x63cc55f49f88eb2f, |
| 2882 |
0xc2781f49ffcfa6d5, 0x3cbf6b71c76b25fb, |
| 2883 |
0xf316271c7fc3908a, 0x8bef464e3945ef7a, |
| 2884 |
0x97edd871cfda3a56, 0x97758bf0e3cbb5ac, |
| 2885 |
0xbde94e8e43d0c8ec, 0x3d52eeed1cbea317, |
| 2886 |
0xed63a231d4c4fb27, 0x4ca7aaa863ee4bdd, |
| 2887 |
0x945e455f24fb1cf8, 0x8fe8caa93e74ef6a, |
| 2888 |
0xb975d6b6ee39e436, 0xb3e2fd538e122b44, |
| 2889 |
0xe7d34c64a9c85d44, 0x60dbbca87196b616, |
| 2890 |
0x90e40fbeea1d3a4a, 0xbc8955e946fe31cd, |
| 2891 |
0xb51d13aea4a488dd, 0x6babab6398bdbe41, |
| 2892 |
0xe264589a4dcdab14, 0xc696963c7eed2dd1, |
| 2893 |
0x8d7eb76070a08aec, 0xfc1e1de5cf543ca2, |
| 2894 |
0xb0de65388cc8ada8, 0x3b25a55f43294bcb, |
| 2895 |
0xdd15fe86affad912, 0x49ef0eb713f39ebe, |
| 2896 |
0x8a2dbf142dfcc7ab, 0x6e3569326c784337, |
| 2897 |
0xacb92ed9397bf996, 0x49c2c37f07965404, |
| 2898 |
0xd7e77a8f87daf7fb, 0xdc33745ec97be906, |
| 2899 |
0x86f0ac99b4e8dafd, 0x69a028bb3ded71a3, |
| 2900 |
0xa8acd7c0222311bc, 0xc40832ea0d68ce0c, |
| 2901 |
0xd2d80db02aabd62b, 0xf50a3fa490c30190, |
| 2902 |
0x83c7088e1aab65db, 0x792667c6da79e0fa, |
| 2903 |
0xa4b8cab1a1563f52, 0x577001b891185938, |
| 2904 |
0xcde6fd5e09abcf26, 0xed4c0226b55e6f86, |
| 2905 |
0x80b05e5ac60b6178, 0x544f8158315b05b4, |
| 2906 |
0xa0dc75f1778e39d6, 0x696361ae3db1c721, |
| 2907 |
0xc913936dd571c84c, 0x3bc3a19cd1e38e9, |
| 2908 |
0xfb5878494ace3a5f, 0x4ab48a04065c723, |
| 2909 |
0x9d174b2dcec0e47b, 0x62eb0d64283f9c76, |
| 2910 |
0xc45d1df942711d9a, 0x3ba5d0bd324f8394, |
| 2911 |
0xf5746577930d6500, 0xca8f44ec7ee36479, |
| 2912 |
0x9968bf6abbe85f20, 0x7e998b13cf4e1ecb, |
| 2913 |
0xbfc2ef456ae276e8, 0x9e3fedd8c321a67e, |
| 2914 |
0xefb3ab16c59b14a2, 0xc5cfe94ef3ea101e, |
| 2915 |
0x95d04aee3b80ece5, 0xbba1f1d158724a12, |
| 2916 |
0xbb445da9ca61281f, 0x2a8a6e45ae8edc97, |
| 2917 |
0xea1575143cf97226, 0xf52d09d71a3293bd, |
| 2918 |
0x924d692ca61be758, 0x593c2626705f9c56, |
| 2919 |
0xb6e0c377cfa2e12e, 0x6f8b2fb00c77836c, |
| 2920 |
0xe498f455c38b997a, 0xb6dfb9c0f956447, |
| 2921 |
0x8edf98b59a373fec, 0x4724bd4189bd5eac, |
| 2922 |
0xb2977ee300c50fe7, 0x58edec91ec2cb657, |
| 2923 |
0xdf3d5e9bc0f653e1, 0x2f2967b66737e3ed, |
| 2924 |
0x8b865b215899f46c, 0xbd79e0d20082ee74, |
| 2925 |
0xae67f1e9aec07187, 0xecd8590680a3aa11, |
| 2926 |
0xda01ee641a708de9, 0xe80e6f4820cc9495, |
| 2927 |
0x884134fe908658b2, 0x3109058d147fdcdd, |
| 2928 |
0xaa51823e34a7eede, 0xbd4b46f0599fd415, |
| 2929 |
0xd4e5e2cdc1d1ea96, 0x6c9e18ac7007c91a, |
| 2930 |
0x850fadc09923329e, 0x3e2cf6bc604ddb0, |
| 2931 |
0xa6539930bf6bff45, 0x84db8346b786151c, |
| 2932 |
0xcfe87f7cef46ff16, 0xe612641865679a63, |
| 2933 |
0x81f14fae158c5f6e, 0x4fcb7e8f3f60c07e, |
| 2934 |
0xa26da3999aef7749, 0xe3be5e330f38f09d, |
| 2935 |
0xcb090c8001ab551c, 0x5cadf5bfd3072cc5, |
| 2936 |
0xfdcb4fa002162a63, 0x73d9732fc7c8f7f6, |
| 2937 |
0x9e9f11c4014dda7e, 0x2867e7fddcdd9afa, |
| 2938 |
0xc646d63501a1511d, 0xb281e1fd541501b8, |
| 2939 |
0xf7d88bc24209a565, 0x1f225a7ca91a4226, |
| 2940 |
0x9ae757596946075f, 0x3375788de9b06958, |
| 2941 |
0xc1a12d2fc3978937, 0x52d6b1641c83ae, |
| 2942 |
0xf209787bb47d6b84, 0xc0678c5dbd23a49a, |
| 2943 |
0x9745eb4d50ce6332, 0xf840b7ba963646e0, |
| 2944 |
0xbd176620a501fbff, 0xb650e5a93bc3d898, |
| 2945 |
0xec5d3fa8ce427aff, 0xa3e51f138ab4cebe, |
| 2946 |
0x93ba47c980e98cdf, 0xc66f336c36b10137, |
| 2947 |
0xb8a8d9bbe123f017, 0xb80b0047445d4184, |
| 2948 |
0xe6d3102ad96cec1d, 0xa60dc059157491e5, |
| 2949 |
0x9043ea1ac7e41392, 0x87c89837ad68db2f, |
| 2950 |
0xb454e4a179dd1877, 0x29babe4598c311fb, |
| 2951 |
0xe16a1dc9d8545e94, 0xf4296dd6fef3d67a, |
| 2952 |
0x8ce2529e2734bb1d, 0x1899e4a65f58660c, |
| 2953 |
0xb01ae745b101e9e4, 0x5ec05dcff72e7f8f, |
| 2954 |
0xdc21a1171d42645d, 0x76707543f4fa1f73, |
| 2955 |
0x899504ae72497eba, 0x6a06494a791c53a8, |
| 2956 |
0xabfa45da0edbde69, 0x487db9d17636892, |
| 2957 |
0xd6f8d7509292d603, 0x45a9d2845d3c42b6, |
| 2958 |
0x865b86925b9bc5c2, 0xb8a2392ba45a9b2, |
| 2959 |
0xa7f26836f282b732, 0x8e6cac7768d7141e, |
| 2960 |
0xd1ef0244af2364ff, 0x3207d795430cd926, |
| 2961 |
0x8335616aed761f1f, 0x7f44e6bd49e807b8, |
| 2962 |
0xa402b9c5a8d3a6e7, 0x5f16206c9c6209a6, |
| 2963 |
0xcd036837130890a1, 0x36dba887c37a8c0f, |
| 2964 |
0x802221226be55a64, 0xc2494954da2c9789, |
| 2965 |
0xa02aa96b06deb0fd, 0xf2db9baa10b7bd6c, |
| 2966 |
0xc83553c5c8965d3d, 0x6f92829494e5acc7, |
| 2967 |
0xfa42a8b73abbf48c, 0xcb772339ba1f17f9, |
| 2968 |
0x9c69a97284b578d7, 0xff2a760414536efb, |
| 2969 |
0xc38413cf25e2d70d, 0xfef5138519684aba, |
| 2970 |
0xf46518c2ef5b8cd1, 0x7eb258665fc25d69, |
| 2971 |
0x98bf2f79d5993802, 0xef2f773ffbd97a61, |
| 2972 |
0xbeeefb584aff8603, 0xaafb550ffacfd8fa, |
| 2973 |
0xeeaaba2e5dbf6784, 0x95ba2a53f983cf38, |
| 2974 |
0x952ab45cfa97a0b2, 0xdd945a747bf26183, |
| 2975 |
0xba756174393d88df, 0x94f971119aeef9e4, |
| 2976 |
0xe912b9d1478ceb17, 0x7a37cd5601aab85d, |
| 2977 |
0x91abb422ccb812ee, 0xac62e055c10ab33a, |
| 2978 |
0xb616a12b7fe617aa, 0x577b986b314d6009, |
| 2979 |
0xe39c49765fdf9d94, 0xed5a7e85fda0b80b, |
| 2980 |
0x8e41ade9fbebc27d, 0x14588f13be847307, |
| 2981 |
0xb1d219647ae6b31c, 0x596eb2d8ae258fc8, |
| 2982 |
0xde469fbd99a05fe3, 0x6fca5f8ed9aef3bb, |
| 2983 |
0x8aec23d680043bee, 0x25de7bb9480d5854, |
| 2984 |
0xada72ccc20054ae9, 0xaf561aa79a10ae6a, |
| 2985 |
0xd910f7ff28069da4, 0x1b2ba1518094da04, |
| 2986 |
0x87aa9aff79042286, 0x90fb44d2f05d0842, |
| 2987 |
0xa99541bf57452b28, 0x353a1607ac744a53, |
| 2988 |
0xd3fa922f2d1675f2, 0x42889b8997915ce8, |
| 2989 |
0x847c9b5d7c2e09b7, 0x69956135febada11, |
| 2990 |
0xa59bc234db398c25, 0x43fab9837e699095, |
| 2991 |
0xcf02b2c21207ef2e, 0x94f967e45e03f4bb, |
| 2992 |
0x8161afb94b44f57d, 0x1d1be0eebac278f5, |
| 2993 |
0xa1ba1ba79e1632dc, 0x6462d92a69731732, |
| 2994 |
0xca28a291859bbf93, 0x7d7b8f7503cfdcfe, |
| 2995 |
0xfcb2cb35e702af78, 0x5cda735244c3d43e, |
| 2996 |
0x9defbf01b061adab, 0x3a0888136afa64a7, |
| 2997 |
0xc56baec21c7a1916, 0x88aaa1845b8fdd0, |
| 2998 |
0xf6c69a72a3989f5b, 0x8aad549e57273d45, |
| 2999 |
0x9a3c2087a63f6399, 0x36ac54e2f678864b, |
| 3000 |
0xc0cb28a98fcf3c7f, 0x84576a1bb416a7dd, |
| 3001 |
0xf0fdf2d3f3c30b9f, 0x656d44a2a11c51d5, |
| 3002 |
0x969eb7c47859e743, 0x9f644ae5a4b1b325, |
| 3003 |
0xbc4665b596706114, 0x873d5d9f0dde1fee, |
| 3004 |
0xeb57ff22fc0c7959, 0xa90cb506d155a7ea, |
| 3005 |
0x9316ff75dd87cbd8, 0x9a7f12442d588f2, |
| 3006 |
0xb7dcbf5354e9bece, 0xc11ed6d538aeb2f, |
| 3007 |
0xe5d3ef282a242e81, 0x8f1668c8a86da5fa, |
| 3008 |
0x8fa475791a569d10, 0xf96e017d694487bc, |
| 3009 |
0xb38d92d760ec4455, 0x37c981dcc395a9ac, |
| 3010 |
0xe070f78d3927556a, 0x85bbe253f47b1417, |
| 3011 |
0x8c469ab843b89562, 0x93956d7478ccec8e, |
| 3012 |
0xaf58416654a6babb, 0x387ac8d1970027b2, |
| 3013 |
0xdb2e51bfe9d0696a, 0x6997b05fcc0319e, |
| 3014 |
0x88fcf317f22241e2, 0x441fece3bdf81f03, |
| 3015 |
0xab3c2fddeeaad25a, 0xd527e81cad7626c3, |
| 3016 |
0xd60b3bd56a5586f1, 0x8a71e223d8d3b074, |
| 3017 |
0x85c7056562757456, 0xf6872d5667844e49, |
| 3018 |
0xa738c6bebb12d16c, 0xb428f8ac016561db, |
| 3019 |
0xd106f86e69d785c7, 0xe13336d701beba52, |
| 3020 |
0x82a45b450226b39c, 0xecc0024661173473, |
| 3021 |
0xa34d721642b06084, 0x27f002d7f95d0190, |
| 3022 |
0xcc20ce9bd35c78a5, 0x31ec038df7b441f4, |
| 3023 |
0xff290242c83396ce, 0x7e67047175a15271, |
| 3024 |
0x9f79a169bd203e41, 0xf0062c6e984d386, |
| 3025 |
0xc75809c42c684dd1, 0x52c07b78a3e60868, |
| 3026 |
0xf92e0c3537826145, 0xa7709a56ccdf8a82, |
| 3027 |
0x9bbcc7a142b17ccb, 0x88a66076400bb691, |
| 3028 |
0xc2abf989935ddbfe, 0x6acff893d00ea435, |
| 3029 |
0xf356f7ebf83552fe, 0x583f6b8c4124d43, |
| 3030 |
0x98165af37b2153de, 0xc3727a337a8b704a, |
| 3031 |
0xbe1bf1b059e9a8d6, 0x744f18c0592e4c5c, |
| 3032 |
0xeda2ee1c7064130c, 0x1162def06f79df73, |
| 3033 |
0x9485d4d1c63e8be7, 0x8addcb5645ac2ba8, |
| 3034 |
0xb9a74a0637ce2ee1, 0x6d953e2bd7173692, |
| 3035 |
0xe8111c87c5c1ba99, 0xc8fa8db6ccdd0437, |
| 3036 |
0x910ab1d4db9914a0, 0x1d9c9892400a22a2, |
| 3037 |
0xb54d5e4a127f59c8, 0x2503beb6d00cab4b, |
| 3038 |
0xe2a0b5dc971f303a, 0x2e44ae64840fd61d, |
| 3039 |
0x8da471a9de737e24, 0x5ceaecfed289e5d2, |
| 3040 |
0xb10d8e1456105dad, 0x7425a83e872c5f47, |
| 3041 |
0xdd50f1996b947518, 0xd12f124e28f77719, |
| 3042 |
0x8a5296ffe33cc92f, 0x82bd6b70d99aaa6f, |
| 3043 |
0xace73cbfdc0bfb7b, 0x636cc64d1001550b, |
| 3044 |
0xd8210befd30efa5a, 0x3c47f7e05401aa4e, |
| 3045 |
0x8714a775e3e95c78, 0x65acfaec34810a71, |
| 3046 |
0xa8d9d1535ce3b396, 0x7f1839a741a14d0d, |
| 3047 |
0xd31045a8341ca07c, 0x1ede48111209a050, |
| 3048 |
0x83ea2b892091e44d, 0x934aed0aab460432, |
| 3049 |
0xa4e4b66b68b65d60, 0xf81da84d5617853f, |
| 3050 |
0xce1de40642e3f4b9, 0x36251260ab9d668e, |
| 3051 |
0x80d2ae83e9ce78f3, 0xc1d72b7c6b426019, |
| 3052 |
0xa1075a24e4421730, 0xb24cf65b8612f81f, |
| 3053 |
0xc94930ae1d529cfc, 0xdee033f26797b627, |
| 3054 |
0xfb9b7cd9a4a7443c, 0x169840ef017da3b1, |
| 3055 |
0x9d412e0806e88aa5, 0x8e1f289560ee864e, |
| 3056 |
0xc491798a08a2ad4e, 0xf1a6f2bab92a27e2, |
| 3057 |
0xf5b5d7ec8acb58a2, 0xae10af696774b1db, |
| 3058 |
0x9991a6f3d6bf1765, 0xacca6da1e0a8ef29, |
| 3059 |
0xbff610b0cc6edd3f, 0x17fd090a58d32af3, |
| 3060 |
0xeff394dcff8a948e, 0xddfc4b4cef07f5b0, |
| 3061 |
0x95f83d0a1fb69cd9, 0x4abdaf101564f98e, |
| 3062 |
0xbb764c4ca7a4440f, 0x9d6d1ad41abe37f1, |
| 3063 |
0xea53df5fd18d5513, 0x84c86189216dc5ed, |
| 3064 |
0x92746b9be2f8552c, 0x32fd3cf5b4e49bb4, |
| 3065 |
0xb7118682dbb66a77, 0x3fbc8c33221dc2a1, |
| 3066 |
0xe4d5e82392a40515, 0xfabaf3feaa5334a, |
| 3067 |
0x8f05b1163ba6832d, 0x29cb4d87f2a7400e, |
| 3068 |
0xb2c71d5bca9023f8, 0x743e20e9ef511012, |
| 3069 |
0xdf78e4b2bd342cf6, 0x914da9246b255416, |
| 3070 |
0x8bab8eefb6409c1a, 0x1ad089b6c2f7548e, |
| 3071 |
0xae9672aba3d0c320, 0xa184ac2473b529b1, |
| 3072 |
0xda3c0f568cc4f3e8, 0xc9e5d72d90a2741e, |
| 3073 |
0x8865899617fb1871, 0x7e2fa67c7a658892, |
| 3074 |
0xaa7eebfb9df9de8d, 0xddbb901b98feeab7, |
| 3075 |
0xd51ea6fa85785631, 0x552a74227f3ea565, |
| 3076 |
0x8533285c936b35de, 0xd53a88958f87275f, |
| 3077 |
0xa67ff273b8460356, 0x8a892abaf368f137, |
| 3078 |
0xd01fef10a657842c, 0x2d2b7569b0432d85, |
| 3079 |
0x8213f56a67f6b29b, 0x9c3b29620e29fc73, |
| 3080 |
0xa298f2c501f45f42, 0x8349f3ba91b47b8f, |
| 3081 |
0xcb3f2f7642717713, 0x241c70a936219a73, |
| 3082 |
0xfe0efb53d30dd4d7, 0xed238cd383aa0110, |
| 3083 |
0x9ec95d1463e8a506, 0xf4363804324a40aa, |
| 3084 |
0xc67bb4597ce2ce48, 0xb143c6053edcd0d5, |
| 3085 |
0xf81aa16fdc1b81da, 0xdd94b7868e94050a, |
| 3086 |
0x9b10a4e5e9913128, 0xca7cf2b4191c8326, |
| 3087 |
0xc1d4ce1f63f57d72, 0xfd1c2f611f63a3f0, |
| 3088 |
0xf24a01a73cf2dccf, 0xbc633b39673c8cec, |
| 3089 |
0x976e41088617ca01, 0xd5be0503e085d813, |
| 3090 |
0xbd49d14aa79dbc82, 0x4b2d8644d8a74e18, |
| 3091 |
0xec9c459d51852ba2, 0xddf8e7d60ed1219e, |
| 3092 |
0x93e1ab8252f33b45, 0xcabb90e5c942b503, |
| 3093 |
0xb8da1662e7b00a17, 0x3d6a751f3b936243, |
| 3094 |
0xe7109bfba19c0c9d, 0xcc512670a783ad4, |
| 3095 |
0x906a617d450187e2, 0x27fb2b80668b24c5, |
| 3096 |
0xb484f9dc9641e9da, 0xb1f9f660802dedf6, |
| 3097 |
0xe1a63853bbd26451, 0x5e7873f8a0396973, |
| 3098 |
0x8d07e33455637eb2, 0xdb0b487b6423e1e8, |
| 3099 |
0xb049dc016abc5e5f, 0x91ce1a9a3d2cda62, |
| 3100 |
0xdc5c5301c56b75f7, 0x7641a140cc7810fb, |
| 3101 |
0x89b9b3e11b6329ba, 0xa9e904c87fcb0a9d, |
| 3102 |
0xac2820d9623bf429, 0x546345fa9fbdcd44, |
| 3103 |
0xd732290fbacaf133, 0xa97c177947ad4095, |
| 3104 |
0x867f59a9d4bed6c0, 0x49ed8eabcccc485d, |
| 3105 |
0xa81f301449ee8c70, 0x5c68f256bfff5a74, |
| 3106 |
0xd226fc195c6a2f8c, 0x73832eec6fff3111, |
| 3107 |
0x83585d8fd9c25db7, 0xc831fd53c5ff7eab, |
| 3108 |
0xa42e74f3d032f525, 0xba3e7ca8b77f5e55, |
| 3109 |
0xcd3a1230c43fb26f, 0x28ce1bd2e55f35eb, |
| 3110 |
0x80444b5e7aa7cf85, 0x7980d163cf5b81b3, |
| 3111 |
0xa0555e361951c366, 0xd7e105bcc332621f, |
| 3112 |
0xc86ab5c39fa63440, 0x8dd9472bf3fefaa7, |
| 3113 |
0xfa856334878fc150, 0xb14f98f6f0feb951, |
| 3114 |
0x9c935e00d4b9d8d2, 0x6ed1bf9a569f33d3, |
| 3115 |
0xc3b8358109e84f07, 0xa862f80ec4700c8, |
| 3116 |
0xf4a642e14c6262c8, 0xcd27bb612758c0fa, |
| 3117 |
0x98e7e9cccfbd7dbd, 0x8038d51cb897789c, |
| 3118 |
0xbf21e44003acdd2c, 0xe0470a63e6bd56c3, |
| 3119 |
0xeeea5d5004981478, 0x1858ccfce06cac74, |
| 3120 |
0x95527a5202df0ccb, 0xf37801e0c43ebc8, |
| 3121 |
0xbaa718e68396cffd, 0xd30560258f54e6ba, |
| 3122 |
0xe950df20247c83fd, 0x47c6b82ef32a2069, |
| 3123 |
0x91d28b7416cdd27e, 0x4cdc331d57fa5441, |
| 3124 |
0xb6472e511c81471d, 0xe0133fe4adf8e952, |
| 3125 |
0xe3d8f9e563a198e5, 0x58180fddd97723a6, |
| 3126 |
0x8e679c2f5e44ff8f, 0x570f09eaa7ea7648, |
| 3127 |
}; |
| 3128 |
}; |
| 3129 |
|
| 3130 |
#if FASTFLOAT_DETAIL_MUST_DEFINE_CONSTEXPR_VARIABLE |
| 3131 |
|
| 3132 |
template <class unused> |
| 3133 |
constexpr uint64_t |
| 3134 |
powers_template<unused>::power_of_five_128[number_of_entries]; |
| 3135 |
|
| 3136 |
#endif |
| 3137 |
|
| 3138 |
using powers = powers_template<>; |
| 3139 |
|
| 3140 |
} // namespace fast_float |
| 3141 |
|
| 3142 |
#endif |
| 3143 |
|
| 3144 |
#ifndef FASTFLOAT_DECIMAL_TO_BINARY_H |
| 3145 |
#define FASTFLOAT_DECIMAL_TO_BINARY_H |
| 3146 |
|
| 3147 |
#include <cfloat> |
| 3148 |
#include <cinttypes> |
| 3149 |
#include <cmath> |
| 3150 |
#include <cstdint> |
| 3151 |
#include <cstdlib> |
| 3152 |
#include <cstring> |
| 3153 |
|
| 3154 |
namespace fast_float { |
| 3155 |
|
| 3156 |
// This will compute or rather approximate w * 5**q and return a pair of 64-bit |
| 3157 |
// words approximating the result, with the "high" part corresponding to the |
| 3158 |
// most significant bits and the low part corresponding to the least significant |
| 3159 |
// bits. |
| 3160 |
// |
| 3161 |
template <int bit_precision> |
| 3162 |
fastfloat_really_inline FASTFLOAT_CONSTEXPR20 value128 |
| 3163 |
compute_product_approximation(int64_t q, uint64_t w) { |
| 3164 |
int const index = 2 * int(q - powers::smallest_power_of_five); |
| 3165 |
// For small values of q, e.g., q in [0,27], the answer is always exact |
| 3166 |
// because The line value128 firstproduct = full_multiplication(w, |
| 3167 |
// power_of_five_128[index]); gives the exact answer. |
| 3168 |
value128 firstproduct = |
| 3169 |
full_multiplication(w, powers::power_of_five_128[index]); |
| 3170 |
static_assert((bit_precision >= 0) && (bit_precision <= 64), |
| 3171 |
" precision should be in (0,64]"); |
| 3172 |
constexpr uint64_t precision_mask = |
| 3173 |
(bit_precision < 64) ? (uint64_t(0xFFFFFFFFFFFFFFFF) >> bit_precision) |
| 3174 |
: uint64_t(0xFFFFFFFFFFFFFFFF); |
| 3175 |
if ((firstproduct.high & precision_mask) == |
| 3176 |
precision_mask) { // could further guard with (lower + w < lower) |
| 3177 |
// regarding the second product, we only need secondproduct.high, but our |
| 3178 |
// expectation is that the compiler will optimize this extra work away if |
| 3179 |
// needed. |
| 3180 |
value128 secondproduct = |
| 3181 |
full_multiplication(w, powers::power_of_five_128[index + 1]); |
| 3182 |
firstproduct.low += secondproduct.high; |
| 3183 |
if (secondproduct.high > firstproduct.low) { |
| 3184 |
firstproduct.high++; |
| 3185 |
} |
| 3186 |
} |
| 3187 |
return firstproduct; |
| 3188 |
} |
| 3189 |
|
| 3190 |
namespace detail { |
| 3191 |
/** |
| 3192 |
* For q in (0,350), we have that |
| 3193 |
* f = (((152170 + 65536) * q ) >> 16); |
| 3194 |
* is equal to |
| 3195 |
* floor(p) + q |
| 3196 |
* where |
| 3197 |
* p = log(5**q)/log(2) = q * log(5)/log(2) |
| 3198 |
* |
| 3199 |
* For negative values of q in (-400,0), we have that |
| 3200 |
* f = (((152170 + 65536) * q ) >> 16); |
| 3201 |
* is equal to |
| 3202 |
* -ceil(p) + q |
| 3203 |
* where |
| 3204 |
* p = log(5**-q)/log(2) = -q * log(5)/log(2) |
| 3205 |
*/ |
| 3206 |
constexpr fastfloat_really_inline int32_t power(int32_t q) noexcept { |
| 3207 |
return (((152170 + 65536) * q) >> 16) + 63; |
| 3208 |
} |
| 3209 |
} // namespace detail |
| 3210 |
|
| 3211 |
// create an adjusted mantissa, biased by the invalid power2 |
| 3212 |
// for significant digits already multiplied by 10 ** q. |
| 3213 |
template <typename binary> |
| 3214 |
fastfloat_really_inline FASTFLOAT_CONSTEXPR14 adjusted_mantissa |
| 3215 |
compute_error_scaled(int64_t q, uint64_t w, int lz) noexcept { |
| 3216 |
int hilz = int(w >> 63) ^ 1; |
| 3217 |
adjusted_mantissa answer; |
| 3218 |
answer.mantissa = w << hilz; |
| 3219 |
int bias = binary::mantissa_explicit_bits() - binary::minimum_exponent(); |
| 3220 |
answer.power2 = int32_t(detail::power(int32_t(q)) + bias - hilz - lz - 62 + |
| 3221 |
invalid_am_bias); |
| 3222 |
return answer; |
| 3223 |
} |
| 3224 |
|
| 3225 |
// w * 10 ** q, without rounding the representation up. |
| 3226 |
// the power2 in the exponent will be adjusted by invalid_am_bias. |
| 3227 |
template <typename binary> |
| 3228 |
fastfloat_really_inline FASTFLOAT_CONSTEXPR20 adjusted_mantissa |
| 3229 |
compute_error(int64_t q, uint64_t w) noexcept { |
| 3230 |
int lz = leading_zeroes(w); |
| 3231 |
w <<= lz; |
| 3232 |
value128 product = |
| 3233 |
compute_product_approximation<binary::mantissa_explicit_bits() + 3>(q, w); |
| 3234 |
return compute_error_scaled<binary>(q, product.high, lz); |
| 3235 |
} |
| 3236 |
|
| 3237 |
// Computers w * 10 ** q. |
| 3238 |
// The returned value should be a valid number that simply needs to be |
| 3239 |
// packed. However, in some very rare cases, the computation will fail. In such |
| 3240 |
// cases, we return an adjusted_mantissa with a negative power of 2: the caller |
| 3241 |
// should recompute in such cases. |
| 3242 |
template <typename binary> |
| 3243 |
fastfloat_really_inline FASTFLOAT_CONSTEXPR20 adjusted_mantissa |
| 3244 |
compute_float(int64_t q, uint64_t w) noexcept { |
| 3245 |
adjusted_mantissa answer; |
| 3246 |
if ((w == 0) || (q < binary::smallest_power_of_ten())) { |
| 3247 |
answer.power2 = 0; |
| 3248 |
answer.mantissa = 0; |
| 3249 |
// result should be zero |
| 3250 |
return answer; |
| 3251 |
} |
| 3252 |
if (q > binary::largest_power_of_ten()) { |
| 3253 |
// we want to get infinity: |
| 3254 |
answer.power2 = binary::infinite_power(); |
| 3255 |
answer.mantissa = 0; |
| 3256 |
return answer; |
| 3257 |
} |
| 3258 |
// At this point in time q is in [powers::smallest_power_of_five, |
| 3259 |
// powers::largest_power_of_five]. |
| 3260 |
|
| 3261 |
// We want the most significant bit of i to be 1. Shift if needed. |
| 3262 |
int lz = leading_zeroes(w); |
| 3263 |
w <<= lz; |
| 3264 |
|
| 3265 |
// The required precision is binary::mantissa_explicit_bits() + 3 because |
| 3266 |
// 1. We need the implicit bit |
| 3267 |
// 2. We need an extra bit for rounding purposes |
| 3268 |
// 3. We might lose a bit due to the "upperbit" routine (result too small, |
| 3269 |
// requiring a shift) |
| 3270 |
|
| 3271 |
value128 product = |
| 3272 |
compute_product_approximation<binary::mantissa_explicit_bits() + 3>(q, w); |
| 3273 |
// The computed 'product' is always sufficient. |
| 3274 |
// Mathematical proof: |
| 3275 |
// Noble Mushtak and Daniel Lemire, Fast Number Parsing Without Fallback (to |
| 3276 |
// appear) See script/mushtak_lemire.py |
| 3277 |
|
| 3278 |
// The "compute_product_approximation" function can be slightly slower than a |
| 3279 |
// branchless approach: value128 product = compute_product(q, w); but in |
| 3280 |
// practice, we can win big with the compute_product_approximation if its |
| 3281 |
// additional branch is easily predicted. Which is best is data specific. |
| 3282 |
int upperbit = int(product.high >> 63); |
| 3283 |
int shift = upperbit + 64 - binary::mantissa_explicit_bits() - 3; |
| 3284 |
|
| 3285 |
answer.mantissa = product.high >> shift; |
| 3286 |
|
| 3287 |
answer.power2 = int32_t(detail::power(int32_t(q)) + upperbit - lz - |
| 3288 |
binary::minimum_exponent()); |
| 3289 |
if (answer.power2 <= 0) { // we have a subnormal? |
| 3290 |
// Here have that answer.power2 <= 0 so -answer.power2 >= 0 |
| 3291 |
if (-answer.power2 + 1 >= |
| 3292 |
64) { // if we have more than 64 bits below the minimum exponent, you |
| 3293 |
// have a zero for sure. |
| 3294 |
answer.power2 = 0; |
| 3295 |
answer.mantissa = 0; |
| 3296 |
// result should be zero |
| 3297 |
return answer; |
| 3298 |
} |
| 3299 |
// next line is safe because -answer.power2 + 1 < 64 |
| 3300 |
answer.mantissa >>= -answer.power2 + 1; |
| 3301 |
// Thankfully, we can't have both "round-to-even" and subnormals because |
| 3302 |
// "round-to-even" only occurs for powers close to 0 in the 32-bit and |
| 3303 |
// and 64-bit case (with no more than 19 digits). |
| 3304 |
answer.mantissa += (answer.mantissa & 1); // round up |
| 3305 |
answer.mantissa >>= 1; |
| 3306 |
// There is a weird scenario where we don't have a subnormal but just. |
| 3307 |
// Suppose we start with 2.2250738585072013e-308, we end up |
| 3308 |
// with 0x3fffffffffffff x 2^-1023-53 which is technically subnormal |
| 3309 |
// whereas 0x40000000000000 x 2^-1023-53 is normal. Now, we need to round |
| 3310 |
// up 0x3fffffffffffff x 2^-1023-53 and once we do, we are no longer |
| 3311 |
// subnormal, but we can only know this after rounding. |
| 3312 |
// So we only declare a subnormal if we are smaller than the threshold. |
| 3313 |
answer.power2 = |
| 3314 |
(answer.mantissa < (uint64_t(1) << binary::mantissa_explicit_bits())) |
| 3315 |
? 0 |
| 3316 |
: 1; |
| 3317 |
return answer; |
| 3318 |
} |
| 3319 |
|
| 3320 |
// usually, we round *up*, but if we fall right in between and and we have an |
| 3321 |
// even basis, we need to round down |
| 3322 |
// We are only concerned with the cases where 5**q fits in single 64-bit word. |
| 3323 |
if ((product.low <= 1) && (q >= binary::min_exponent_round_to_even()) && |
| 3324 |
(q <= binary::max_exponent_round_to_even()) && |
| 3325 |
((answer.mantissa & 3) == 1)) { // we may fall between two floats! |
| 3326 |
// To be in-between two floats we need that in doing |
| 3327 |
// answer.mantissa = product.high >> (upperbit + 64 - |
| 3328 |
// binary::mantissa_explicit_bits() - 3); |
| 3329 |
// ... we dropped out only zeroes. But if this happened, then we can go |
| 3330 |
// back!!! |
| 3331 |
if ((answer.mantissa << shift) == product.high) { |
| 3332 |
answer.mantissa &= ~uint64_t(1); // flip it so that we do not round up |
| 3333 |
} |
| 3334 |
} |
| 3335 |
|
| 3336 |
answer.mantissa += (answer.mantissa & 1); // round up |
| 3337 |
answer.mantissa >>= 1; |
| 3338 |
if (answer.mantissa >= (uint64_t(2) << binary::mantissa_explicit_bits())) { |
| 3339 |
answer.mantissa = (uint64_t(1) << binary::mantissa_explicit_bits()); |
| 3340 |
answer.power2++; // undo previous addition |
| 3341 |
} |
| 3342 |
|
| 3343 |
answer.mantissa &= ~(uint64_t(1) << binary::mantissa_explicit_bits()); |
| 3344 |
if (answer.power2 >= binary::infinite_power()) { // infinity |
| 3345 |
answer.power2 = binary::infinite_power(); |
| 3346 |
answer.mantissa = 0; |
| 3347 |
} |
| 3348 |
return answer; |
| 3349 |
} |
| 3350 |
|
| 3351 |
} // namespace fast_float |
| 3352 |
|
| 3353 |
#endif |
| 3354 |
|
| 3355 |
#ifndef FASTFLOAT_BIGINT_H |
| 3356 |
#define FASTFLOAT_BIGINT_H |
| 3357 |
|
| 3358 |
#include <algorithm> |
| 3359 |
#include <cstdint> |
| 3360 |
#include <climits> |
| 3361 |
#include <cstring> |
| 3362 |
|
| 3363 |
|
| 3364 |
namespace fast_float { |
| 3365 |
|
| 3366 |
// the limb width: we want efficient multiplication of double the bits in |
| 3367 |
// limb, or for 64-bit limbs, at least 64-bit multiplication where we can |
| 3368 |
// extract the high and low parts efficiently. this is every 64-bit |
| 3369 |
// architecture except for sparc, which emulates 128-bit multiplication. |
| 3370 |
// we might have platforms where `CHAR_BIT` is not 8, so let's avoid |
| 3371 |
// doing `8 * sizeof(limb)`. |
| 3372 |
#if defined(FASTFLOAT_64BIT) && !defined(__sparc) |
| 3373 |
#define FASTFLOAT_64BIT_LIMB 1 |
| 3374 |
typedef uint64_t limb; |
| 3375 |
constexpr size_t limb_bits = 64; |
| 3376 |
#else |
| 3377 |
#define FASTFLOAT_32BIT_LIMB |
| 3378 |
typedef uint32_t limb; |
| 3379 |
constexpr size_t limb_bits = 32; |
| 3380 |
#endif |
| 3381 |
|
| 3382 |
typedef span<limb> limb_span; |
| 3383 |
|
| 3384 |
// number of bits in a bigint. this needs to be at least the number |
| 3385 |
// of bits required to store the largest bigint, which is |
| 3386 |
// `log2(10**(digits + max_exp))`, or `log2(10**(767 + 342))`, or |
| 3387 |
// ~3600 bits, so we round to 4000. |
| 3388 |
constexpr size_t bigint_bits = 4000; |
| 3389 |
constexpr size_t bigint_limbs = bigint_bits / limb_bits; |
| 3390 |
|
| 3391 |
// vector-like type that is allocated on the stack. the entire |
| 3392 |
// buffer is pre-allocated, and only the length changes. |
| 3393 |
template <uint16_t size> struct stackvec { |
| 3394 |
limb data[size]; |
| 3395 |
// we never need more than 150 limbs |
| 3396 |
uint16_t length{0}; |
| 3397 |
|
| 3398 |
stackvec() = default; |
| 3399 |
stackvec(stackvec const &) = delete; |
| 3400 |
stackvec &operator=(stackvec const &) = delete; |
| 3401 |
stackvec(stackvec &&) = delete; |
| 3402 |
stackvec &operator=(stackvec &&other) = delete; |
| 3403 |
|
| 3404 |
// create stack vector from existing limb span. |
| 3405 |
FASTFLOAT_CONSTEXPR20 stackvec(limb_span s) { |
| 3406 |
FASTFLOAT_ASSERT(try_extend(s)); |
| 3407 |
} |
| 3408 |
|
| 3409 |
FASTFLOAT_CONSTEXPR14 limb &operator[](size_t index) noexcept { |
| 3410 |
FASTFLOAT_DEBUG_ASSERT(index < length); |
| 3411 |
return data[index]; |
| 3412 |
} |
| 3413 |
|
| 3414 |
FASTFLOAT_CONSTEXPR14 const limb &operator[](size_t index) const noexcept { |
| 3415 |
FASTFLOAT_DEBUG_ASSERT(index < length); |
| 3416 |
return data[index]; |
| 3417 |
} |
| 3418 |
|
| 3419 |
// index from the end of the container |
| 3420 |
FASTFLOAT_CONSTEXPR14 const limb &rindex(size_t index) const noexcept { |
| 3421 |
FASTFLOAT_DEBUG_ASSERT(index < length); |
| 3422 |
size_t rindex = length - index - 1; |
| 3423 |
return data[rindex]; |
| 3424 |
} |
| 3425 |
|
| 3426 |
// set the length, without bounds checking. |
| 3427 |
FASTFLOAT_CONSTEXPR14 void set_len(size_t len) noexcept { |
| 3428 |
length = uint16_t(len); |
| 3429 |
} |
| 3430 |
|
| 3431 |
constexpr size_t len() const noexcept { return length; } |
| 3432 |
|
| 3433 |
constexpr bool is_empty() const noexcept { return length == 0; } |
| 3434 |
|
| 3435 |
constexpr size_t capacity() const noexcept { return size; } |
| 3436 |
|
| 3437 |
// append item to vector, without bounds checking |
| 3438 |
FASTFLOAT_CONSTEXPR14 void push_unchecked(limb value) noexcept { |
| 3439 |
data[length] = value; |
| 3440 |
length++; |
| 3441 |
} |
| 3442 |
|
| 3443 |
// append item to vector, returning if item was added |
| 3444 |
FASTFLOAT_CONSTEXPR14 bool try_push(limb value) noexcept { |
| 3445 |
if (len() < capacity()) { |
| 3446 |
push_unchecked(value); |
| 3447 |
return true; |
| 3448 |
} else { |
| 3449 |
return false; |
| 3450 |
} |
| 3451 |
} |
| 3452 |
|
| 3453 |
// add items to the vector, from a span, without bounds checking |
| 3454 |
FASTFLOAT_CONSTEXPR20 void extend_unchecked(limb_span s) noexcept { |
| 3455 |
limb *ptr = data + length; |
| 3456 |
std::copy_n(s.ptr, s.len(), ptr); |
| 3457 |
set_len(len() + s.len()); |
| 3458 |
} |
| 3459 |
|
| 3460 |
// try to add items to the vector, returning if items were added |
| 3461 |
FASTFLOAT_CONSTEXPR20 bool try_extend(limb_span s) noexcept { |
| 3462 |
if (len() + s.len() <= capacity()) { |
| 3463 |
extend_unchecked(s); |
| 3464 |
return true; |
| 3465 |
} else { |
| 3466 |
return false; |
| 3467 |
} |
| 3468 |
} |
| 3469 |
|
| 3470 |
// resize the vector, without bounds checking |
| 3471 |
// if the new size is longer than the vector, assign value to each |
| 3472 |
// appended item. |
| 3473 |
FASTFLOAT_CONSTEXPR20 |
| 3474 |
void resize_unchecked(size_t new_len, limb value) noexcept { |
| 3475 |
if (new_len > len()) { |
| 3476 |
size_t count = new_len - len(); |
| 3477 |
limb *first = data + len(); |
| 3478 |
limb *last = first + count; |
| 3479 |
::std::fill(first, last, value); |
| 3480 |
set_len(new_len); |
| 3481 |
} else { |
| 3482 |
set_len(new_len); |
| 3483 |
} |
| 3484 |
} |
| 3485 |
|
| 3486 |
// try to resize the vector, returning if the vector was resized. |
| 3487 |
FASTFLOAT_CONSTEXPR20 bool try_resize(size_t new_len, limb value) noexcept { |
| 3488 |
if (new_len > capacity()) { |
| 3489 |
return false; |
| 3490 |
} else { |
| 3491 |
resize_unchecked(new_len, value); |
| 3492 |
return true; |
| 3493 |
} |
| 3494 |
} |
| 3495 |
|
| 3496 |
// check if any limbs are non-zero after the given index. |
| 3497 |
// this needs to be done in reverse order, since the index |
| 3498 |
// is relative to the most significant limbs. |
| 3499 |
FASTFLOAT_CONSTEXPR14 bool nonzero(size_t index) const noexcept { |
| 3500 |
while (index < len()) { |
| 3501 |
if (rindex(index) != 0) { |
| 3502 |
return true; |
| 3503 |
} |
| 3504 |
index++; |
| 3505 |
} |
| 3506 |
return false; |
| 3507 |
} |
| 3508 |
|
| 3509 |
// normalize the big integer, so most-significant zero limbs are removed. |
| 3510 |
FASTFLOAT_CONSTEXPR14 void normalize() noexcept { |
| 3511 |
while (len() > 0 && rindex(0) == 0) { |
| 3512 |
length--; |
| 3513 |
} |
| 3514 |
} |
| 3515 |
}; |
| 3516 |
|
| 3517 |
fastfloat_really_inline FASTFLOAT_CONSTEXPR14 uint64_t |
| 3518 |
empty_hi64(bool &truncated) noexcept { |
| 3519 |
truncated = false; |
| 3520 |
return 0; |
| 3521 |
} |
| 3522 |
|
| 3523 |
fastfloat_really_inline FASTFLOAT_CONSTEXPR20 uint64_t |
| 3524 |
uint64_hi64(uint64_t r0, bool &truncated) noexcept { |
| 3525 |
truncated = false; |
| 3526 |
int shl = leading_zeroes(r0); |
| 3527 |
return r0 << shl; |
| 3528 |
} |
| 3529 |
|
| 3530 |
fastfloat_really_inline FASTFLOAT_CONSTEXPR20 uint64_t |
| 3531 |
uint64_hi64(uint64_t r0, uint64_t r1, bool &truncated) noexcept { |
| 3532 |
int shl = leading_zeroes(r0); |
| 3533 |
if (shl == 0) { |
| 3534 |
truncated = r1 != 0; |
| 3535 |
return r0; |
| 3536 |
} else { |
| 3537 |
int shr = 64 - shl; |
| 3538 |
truncated = (r1 << shl) != 0; |
| 3539 |
return (r0 << shl) | (r1 >> shr); |
| 3540 |
} |
| 3541 |
} |
| 3542 |
|
| 3543 |
fastfloat_really_inline FASTFLOAT_CONSTEXPR20 uint64_t |
| 3544 |
uint32_hi64(uint32_t r0, bool &truncated) noexcept { |
| 3545 |
return uint64_hi64(r0, truncated); |
| 3546 |
} |
| 3547 |
|
| 3548 |
fastfloat_really_inline FASTFLOAT_CONSTEXPR20 uint64_t |
| 3549 |
uint32_hi64(uint32_t r0, uint32_t r1, bool &truncated) noexcept { |
| 3550 |
uint64_t x0 = r0; |
| 3551 |
uint64_t x1 = r1; |
| 3552 |
return uint64_hi64((x0 << 32) | x1, truncated); |
| 3553 |
} |
| 3554 |
|
| 3555 |
fastfloat_really_inline FASTFLOAT_CONSTEXPR20 uint64_t |
| 3556 |
uint32_hi64(uint32_t r0, uint32_t r1, uint32_t r2, bool &truncated) noexcept { |
| 3557 |
uint64_t x0 = r0; |
| 3558 |
uint64_t x1 = r1; |
| 3559 |
uint64_t x2 = r2; |
| 3560 |
return uint64_hi64(x0, (x1 << 32) | x2, truncated); |
| 3561 |
} |
| 3562 |
|
| 3563 |
// add two small integers, checking for overflow. |
| 3564 |
// we want an efficient operation. for msvc, where |
| 3565 |
// we don't have built-in intrinsics, this is still |
| 3566 |
// pretty fast. |
| 3567 |
fastfloat_really_inline FASTFLOAT_CONSTEXPR20 limb |
| 3568 |
scalar_add(limb x, limb y, bool &overflow) noexcept { |
| 3569 |
limb z; |
| 3570 |
// gcc and clang |
| 3571 |
#if defined(__has_builtin) |
| 3572 |
#if __has_builtin(__builtin_add_overflow) |
| 3573 |
if (!cpp20_and_in_constexpr()) { |
| 3574 |
overflow = __builtin_add_overflow(x, y, &z); |
| 3575 |
return z; |
| 3576 |
} |
| 3577 |
#endif |
| 3578 |
#endif |
| 3579 |
|
| 3580 |
// generic, this still optimizes correctly on MSVC. |
| 3581 |
z = x + y; |
| 3582 |
overflow = z < x; |
| 3583 |
return z; |
| 3584 |
} |
| 3585 |
|
| 3586 |
// multiply two small integers, getting both the high and low bits. |
| 3587 |
fastfloat_really_inline FASTFLOAT_CONSTEXPR20 limb |
| 3588 |
scalar_mul(limb x, limb y, limb &carry) noexcept { |
| 3589 |
#ifdef FASTFLOAT_64BIT_LIMB |
| 3590 |
#if defined(__SIZEOF_INT128__) |
| 3591 |
// GCC and clang both define it as an extension. |
| 3592 |
__uint128_t z = __uint128_t(x) * __uint128_t(y) + __uint128_t(carry); |
| 3593 |
carry = limb(z >> limb_bits); |
| 3594 |
return limb(z); |
| 3595 |
#else |
| 3596 |
// fallback, no native 128-bit integer multiplication with carry. |
| 3597 |
// on msvc, this optimizes identically, somehow. |
| 3598 |
value128 z = full_multiplication(x, y); |
| 3599 |
bool overflow; |
| 3600 |
z.low = scalar_add(z.low, carry, overflow); |
| 3601 |
z.high += uint64_t(overflow); // cannot overflow |
| 3602 |
carry = z.high; |
| 3603 |
return z.low; |
| 3604 |
#endif |
| 3605 |
#else |
| 3606 |
uint64_t z = uint64_t(x) * uint64_t(y) + uint64_t(carry); |
| 3607 |
carry = limb(z >> limb_bits); |
| 3608 |
return limb(z); |
| 3609 |
#endif |
| 3610 |
} |
| 3611 |
|
| 3612 |
// add scalar value to bigint starting from offset. |
| 3613 |
// used in grade school multiplication |
| 3614 |
template <uint16_t size> |
| 3615 |
inline FASTFLOAT_CONSTEXPR20 bool small_add_from(stackvec<size> &vec, limb y, |
| 3616 |
size_t start) noexcept { |
| 3617 |
size_t index = start; |
| 3618 |
limb carry = y; |
| 3619 |
bool overflow; |
| 3620 |
while (carry != 0 && index < vec.len()) { |
| 3621 |
vec[index] = scalar_add(vec[index], carry, overflow); |
| 3622 |
carry = limb(overflow); |
| 3623 |
index += 1; |
| 3624 |
} |
| 3625 |
if (carry != 0) { |
| 3626 |
FASTFLOAT_TRY(vec.try_push(carry)); |
| 3627 |
} |
| 3628 |
return true; |
| 3629 |
} |
| 3630 |
|
| 3631 |
// add scalar value to bigint. |
| 3632 |
template <uint16_t size> |
| 3633 |
fastfloat_really_inline FASTFLOAT_CONSTEXPR20 bool |
| 3634 |
small_add(stackvec<size> &vec, limb y) noexcept { |
| 3635 |
return small_add_from(vec, y, 0); |
| 3636 |
} |
| 3637 |
|
| 3638 |
// multiply bigint by scalar value. |
| 3639 |
template <uint16_t size> |
| 3640 |
inline FASTFLOAT_CONSTEXPR20 bool small_mul(stackvec<size> &vec, |
| 3641 |
limb y) noexcept { |
| 3642 |
limb carry = 0; |
| 3643 |
for (size_t index = 0; index < vec.len(); index++) { |
| 3644 |
vec[index] = scalar_mul(vec[index], y, carry); |
| 3645 |
} |
| 3646 |
if (carry != 0) { |
| 3647 |
FASTFLOAT_TRY(vec.try_push(carry)); |
| 3648 |
} |
| 3649 |
return true; |
| 3650 |
} |
| 3651 |
|
| 3652 |
// add bigint to bigint starting from index. |
| 3653 |
// used in grade school multiplication |
| 3654 |
template <uint16_t size> |
| 3655 |
FASTFLOAT_CONSTEXPR20 bool large_add_from(stackvec<size> &x, limb_span y, |
| 3656 |
size_t start) noexcept { |
| 3657 |
// the effective x buffer is from `xstart..x.len()`, so exit early |
| 3658 |
// if we can't get that current range. |
| 3659 |
if (x.len() < start || y.len() > x.len() - start) { |
| 3660 |
FASTFLOAT_TRY(x.try_resize(y.len() + start, 0)); |
| 3661 |
} |
| 3662 |
|
| 3663 |
bool carry = false; |
| 3664 |
for (size_t index = 0; index < y.len(); index++) { |
| 3665 |
limb xi = x[index + start]; |
| 3666 |
limb yi = y[index]; |
| 3667 |
bool c1 = false; |
| 3668 |
bool c2 = false; |
| 3669 |
xi = scalar_add(xi, yi, c1); |
| 3670 |
if (carry) { |
| 3671 |
xi = scalar_add(xi, 1, c2); |
| 3672 |
} |
| 3673 |
x[index + start] = xi; |
| 3674 |
carry = c1 | c2; |
| 3675 |
} |
| 3676 |
|
| 3677 |
// handle overflow |
| 3678 |
if (carry) { |
| 3679 |
FASTFLOAT_TRY(small_add_from(x, 1, y.len() + start)); |
| 3680 |
} |
| 3681 |
return true; |
| 3682 |
} |
| 3683 |
|
| 3684 |
// add bigint to bigint. |
| 3685 |
template <uint16_t size> |
| 3686 |
fastfloat_really_inline FASTFLOAT_CONSTEXPR20 bool |
| 3687 |
large_add_from(stackvec<size> &x, limb_span y) noexcept { |
| 3688 |
return large_add_from(x, y, 0); |
| 3689 |
} |
| 3690 |
|
| 3691 |
// grade-school multiplication algorithm |
| 3692 |
template <uint16_t size> |
| 3693 |
FASTFLOAT_CONSTEXPR20 bool long_mul(stackvec<size> &x, limb_span y) noexcept { |
| 3694 |
limb_span xs = limb_span(x.data, x.len()); |
| 3695 |
stackvec<size> z(xs); |
| 3696 |
limb_span zs = limb_span(z.data, z.len()); |
| 3697 |
|
| 3698 |
if (y.len() != 0) { |
| 3699 |
limb y0 = y[0]; |
| 3700 |
FASTFLOAT_TRY(small_mul(x, y0)); |
| 3701 |
for (size_t index = 1; index < y.len(); index++) { |
| 3702 |
limb yi = y[index]; |
| 3703 |
stackvec<size> zi; |
| 3704 |
if (yi != 0) { |
| 3705 |
// re-use the same buffer throughout |
| 3706 |
zi.set_len(0); |
| 3707 |
FASTFLOAT_TRY(zi.try_extend(zs)); |
| 3708 |
FASTFLOAT_TRY(small_mul(zi, yi)); |
| 3709 |
limb_span zis = limb_span(zi.data, zi.len()); |
| 3710 |
FASTFLOAT_TRY(large_add_from(x, zis, index)); |
| 3711 |
} |
| 3712 |
} |
| 3713 |
} |
| 3714 |
|
| 3715 |
x.normalize(); |
| 3716 |
return true; |
| 3717 |
} |
| 3718 |
|
| 3719 |
// grade-school multiplication algorithm |
| 3720 |
template <uint16_t size> |
| 3721 |
FASTFLOAT_CONSTEXPR20 bool large_mul(stackvec<size> &x, limb_span y) noexcept { |
| 3722 |
if (y.len() == 1) { |
| 3723 |
FASTFLOAT_TRY(small_mul(x, y[0])); |
| 3724 |
} else { |
| 3725 |
FASTFLOAT_TRY(long_mul(x, y)); |
| 3726 |
} |
| 3727 |
return true; |
| 3728 |
} |
| 3729 |
|
| 3730 |
template <typename = void> struct pow5_tables { |
| 3731 |
static constexpr uint32_t large_step = 135; |
| 3732 |
static constexpr uint64_t small_power_of_5[] = { |
| 3733 |
1UL, |
| 3734 |
5UL, |
| 3735 |
25UL, |
| 3736 |
125UL, |
| 3737 |
625UL, |
| 3738 |
3125UL, |
| 3739 |
15625UL, |
| 3740 |
78125UL, |
| 3741 |
390625UL, |
| 3742 |
1953125UL, |
| 3743 |
9765625UL, |
| 3744 |
48828125UL, |
| 3745 |
244140625UL, |
| 3746 |
1220703125UL, |
| 3747 |
6103515625UL, |
| 3748 |
30517578125UL, |
| 3749 |
152587890625UL, |
| 3750 |
762939453125UL, |
| 3751 |
3814697265625UL, |
| 3752 |
19073486328125UL, |
| 3753 |
95367431640625UL, |
| 3754 |
476837158203125UL, |
| 3755 |
2384185791015625UL, |
| 3756 |
11920928955078125UL, |
| 3757 |
59604644775390625UL, |
| 3758 |
298023223876953125UL, |
| 3759 |
1490116119384765625UL, |
| 3760 |
7450580596923828125UL, |
| 3761 |
}; |
| 3762 |
#ifdef FASTFLOAT_64BIT_LIMB |
| 3763 |
constexpr static limb large_power_of_5[] = { |
| 3764 |
1414648277510068013UL, 9180637584431281687UL, 4539964771860779200UL, |
| 3765 |
10482974169319127550UL, 198276706040285095UL}; |
| 3766 |
#else |
| 3767 |
constexpr static limb large_power_of_5[] = { |
| 3768 |
4279965485U, 329373468U, 4020270615U, 2137533757U, 4287402176U, |
| 3769 |
1057042919U, 1071430142U, 2440757623U, 381945767U, 46164893U}; |
| 3770 |
#endif |
| 3771 |
}; |
| 3772 |
|
| 3773 |
#if FASTFLOAT_DETAIL_MUST_DEFINE_CONSTEXPR_VARIABLE |
| 3774 |
|
| 3775 |
template <typename T> constexpr uint32_t pow5_tables<T>::large_step; |
| 3776 |
|
| 3777 |
template <typename T> constexpr uint64_t pow5_tables<T>::small_power_of_5[]; |
| 3778 |
|
| 3779 |
template <typename T> constexpr limb pow5_tables<T>::large_power_of_5[]; |
| 3780 |
|
| 3781 |
#endif |
| 3782 |
|
| 3783 |
// big integer type. implements a small subset of big integer |
| 3784 |
// arithmetic, using simple algorithms since asymptotically |
| 3785 |
// faster algorithms are slower for a small number of limbs. |
| 3786 |
// all operations assume the big-integer is normalized. |
| 3787 |
struct bigint : pow5_tables<> { |
| 3788 |
// storage of the limbs, in little-endian order. |
| 3789 |
stackvec<bigint_limbs> vec; |
| 3790 |
|
| 3791 |
FASTFLOAT_CONSTEXPR20 bigint() : vec() {} |
| 3792 |
|
| 3793 |
bigint(bigint const &) = delete; |
| 3794 |
bigint &operator=(bigint const &) = delete; |
| 3795 |
bigint(bigint &&) = delete; |
| 3796 |
bigint &operator=(bigint &&other) = delete; |
| 3797 |
|
| 3798 |
FASTFLOAT_CONSTEXPR20 bigint(uint64_t value) : vec() { |
| 3799 |
#ifdef FASTFLOAT_64BIT_LIMB |
| 3800 |
vec.push_unchecked(value); |
| 3801 |
#else |
| 3802 |
vec.push_unchecked(uint32_t(value)); |
| 3803 |
vec.push_unchecked(uint32_t(value >> 32)); |
| 3804 |
#endif |
| 3805 |
vec.normalize(); |
| 3806 |
} |
| 3807 |
|
| 3808 |
// get the high 64 bits from the vector, and if bits were truncated. |
| 3809 |
// this is to get the significant digits for the float. |
| 3810 |
FASTFLOAT_CONSTEXPR20 uint64_t hi64(bool &truncated) const noexcept { |
| 3811 |
#ifdef FASTFLOAT_64BIT_LIMB |
| 3812 |
if (vec.len() == 0) { |
| 3813 |
return empty_hi64(truncated); |
| 3814 |
} else if (vec.len() == 1) { |
| 3815 |
return uint64_hi64(vec.rindex(0), truncated); |
| 3816 |
} else { |
| 3817 |
uint64_t result = uint64_hi64(vec.rindex(0), vec.rindex(1), truncated); |
| 3818 |
truncated |= vec.nonzero(2); |
| 3819 |
return result; |
| 3820 |
} |
| 3821 |
#else |
| 3822 |
if (vec.len() == 0) { |
| 3823 |
return empty_hi64(truncated); |
| 3824 |
} else if (vec.len() == 1) { |
| 3825 |
return uint32_hi64(vec.rindex(0), truncated); |
| 3826 |
} else if (vec.len() == 2) { |
| 3827 |
return uint32_hi64(vec.rindex(0), vec.rindex(1), truncated); |
| 3828 |
} else { |
| 3829 |
uint64_t result = |
| 3830 |
uint32_hi64(vec.rindex(0), vec.rindex(1), vec.rindex(2), truncated); |
| 3831 |
truncated |= vec.nonzero(3); |
| 3832 |
return result; |
| 3833 |
} |
| 3834 |
#endif |
| 3835 |
} |
| 3836 |
|
| 3837 |
// compare two big integers, returning the large value. |
| 3838 |
// assumes both are normalized. if the return value is |
| 3839 |
// negative, other is larger, if the return value is |
| 3840 |
// positive, this is larger, otherwise they are equal. |
| 3841 |
// the limbs are stored in little-endian order, so we |
| 3842 |
// must compare the limbs in ever order. |
| 3843 |
FASTFLOAT_CONSTEXPR20 int compare(bigint const &other) const noexcept { |
| 3844 |
if (vec.len() > other.vec.len()) { |
| 3845 |
return 1; |
| 3846 |
} else if (vec.len() < other.vec.len()) { |
| 3847 |
return -1; |
| 3848 |
} else { |
| 3849 |
for (size_t index = vec.len(); index > 0; index--) { |
| 3850 |
limb xi = vec[index - 1]; |
| 3851 |
limb yi = other.vec[index - 1]; |
| 3852 |
if (xi > yi) { |
| 3853 |
return 1; |
| 3854 |
} else if (xi < yi) { |
| 3855 |
return -1; |
| 3856 |
} |
| 3857 |
} |
| 3858 |
return 0; |
| 3859 |
} |
| 3860 |
} |
| 3861 |
|
| 3862 |
// shift left each limb n bits, carrying over to the new limb |
| 3863 |
// returns true if we were able to shift all the digits. |
| 3864 |
FASTFLOAT_CONSTEXPR20 bool shl_bits(size_t n) noexcept { |
| 3865 |
// Internally, for each item, we shift left by n, and add the previous |
| 3866 |
// right shifted limb-bits. |
| 3867 |
// For example, we transform (for u8) shifted left 2, to: |
| 3868 |
// b10100100 b01000010 |
| 3869 |
// b10 b10010001 b00001000 |
| 3870 |
FASTFLOAT_DEBUG_ASSERT(n != 0); |
| 3871 |
FASTFLOAT_DEBUG_ASSERT(n < sizeof(limb) * 8); |
| 3872 |
|
| 3873 |
size_t shl = n; |
| 3874 |
size_t shr = limb_bits - shl; |
| 3875 |
limb prev = 0; |
| 3876 |
for (size_t index = 0; index < vec.len(); index++) { |
| 3877 |
limb xi = vec[index]; |
| 3878 |
vec[index] = (xi << shl) | (prev >> shr); |
| 3879 |
prev = xi; |
| 3880 |
} |
| 3881 |
|
| 3882 |
limb carry = prev >> shr; |
| 3883 |
if (carry != 0) { |
| 3884 |
return vec.try_push(carry); |
| 3885 |
} |
| 3886 |
return true; |
| 3887 |
} |
| 3888 |
|
| 3889 |
// move the limbs left by `n` limbs. |
| 3890 |
FASTFLOAT_CONSTEXPR20 bool shl_limbs(size_t n) noexcept { |
| 3891 |
FASTFLOAT_DEBUG_ASSERT(n != 0); |
| 3892 |
if (n + vec.len() > vec.capacity()) { |
| 3893 |
return false; |
| 3894 |
} else if (!vec.is_empty()) { |
| 3895 |
// move limbs |
| 3896 |
limb *dst = vec.data + n; |
| 3897 |
limb const *src = vec.data; |
| 3898 |
std::copy_backward(src, src + vec.len(), dst + vec.len()); |
| 3899 |
// fill in empty limbs |
| 3900 |
limb *first = vec.data; |
| 3901 |
limb *last = first + n; |
| 3902 |
::std::fill(first, last, 0); |
| 3903 |
vec.set_len(n + vec.len()); |
| 3904 |
return true; |
| 3905 |
} else { |
| 3906 |
return true; |
| 3907 |
} |
| 3908 |
} |
| 3909 |
|
| 3910 |
// move the limbs left by `n` bits. |
| 3911 |
FASTFLOAT_CONSTEXPR20 bool shl(size_t n) noexcept { |
| 3912 |
size_t rem = n % limb_bits; |
| 3913 |
size_t div = n / limb_bits; |
| 3914 |
if (rem != 0) { |
| 3915 |
FASTFLOAT_TRY(shl_bits(rem)); |
| 3916 |
} |
| 3917 |
if (div != 0) { |
| 3918 |
FASTFLOAT_TRY(shl_limbs(div)); |
| 3919 |
} |
| 3920 |
return true; |
| 3921 |
} |
| 3922 |
|
| 3923 |
// get the number of leading zeros in the bigint. |
| 3924 |
FASTFLOAT_CONSTEXPR20 int ctlz() const noexcept { |
| 3925 |
if (vec.is_empty()) { |
| 3926 |
return 0; |
| 3927 |
} else { |
| 3928 |
#ifdef FASTFLOAT_64BIT_LIMB |
| 3929 |
return leading_zeroes(vec.rindex(0)); |
| 3930 |
#else |
| 3931 |
// no use defining a specialized leading_zeroes for a 32-bit type. |
| 3932 |
uint64_t r0 = vec.rindex(0); |
| 3933 |
return leading_zeroes(r0 << 32); |
| 3934 |
#endif |
| 3935 |
} |
| 3936 |
} |
| 3937 |
|
| 3938 |
// get the number of bits in the bigint. |
| 3939 |
FASTFLOAT_CONSTEXPR20 int bit_length() const noexcept { |
| 3940 |
int lz = ctlz(); |
| 3941 |
return int(limb_bits * vec.len()) - lz; |
| 3942 |
} |
| 3943 |
|
| 3944 |
FASTFLOAT_CONSTEXPR20 bool mul(limb y) noexcept { return small_mul(vec, y); } |
| 3945 |
|
| 3946 |
FASTFLOAT_CONSTEXPR20 bool add(limb y) noexcept { return small_add(vec, y); } |
| 3947 |
|
| 3948 |
// multiply as if by 2 raised to a power. |
| 3949 |
FASTFLOAT_CONSTEXPR20 bool pow2(uint32_t exp) noexcept { return shl(exp); } |
| 3950 |
|
| 3951 |
// multiply as if by 5 raised to a power. |
| 3952 |
FASTFLOAT_CONSTEXPR20 bool pow5(uint32_t exp) noexcept { |
| 3953 |
// multiply by a power of 5 |
| 3954 |
size_t large_length = sizeof(large_power_of_5) / sizeof(limb); |
| 3955 |
limb_span large = limb_span(large_power_of_5, large_length); |
| 3956 |
while (exp >= large_step) { |
| 3957 |
FASTFLOAT_TRY(large_mul(vec, large)); |
| 3958 |
exp -= large_step; |
| 3959 |
} |
| 3960 |
#ifdef FASTFLOAT_64BIT_LIMB |
| 3961 |
uint32_t small_step = 27; |
| 3962 |
limb max_native = 7450580596923828125UL; |
| 3963 |
#else |
| 3964 |
uint32_t small_step = 13; |
| 3965 |
limb max_native = 1220703125U; |
| 3966 |
#endif |
| 3967 |
while (exp >= small_step) { |
| 3968 |
FASTFLOAT_TRY(small_mul(vec, max_native)); |
| 3969 |
exp -= small_step; |
| 3970 |
} |
| 3971 |
if (exp != 0) { |
| 3972 |
// Work around clang bug https://godbolt.org/z/zedh7rrhc |
| 3973 |
// This is similar to https://github.com/llvm/llvm-project/issues/47746, |
| 3974 |
// except the workaround described there don't work here |
| 3975 |
FASTFLOAT_TRY(small_mul( |
| 3976 |
vec, limb(((void)small_power_of_5[0], small_power_of_5[exp])))); |
| 3977 |
} |
| 3978 |
|
| 3979 |
return true; |
| 3980 |
} |
| 3981 |
|
| 3982 |
// multiply as if by 10 raised to a power. |
| 3983 |
FASTFLOAT_CONSTEXPR20 bool pow10(uint32_t exp) noexcept { |
| 3984 |
FASTFLOAT_TRY(pow5(exp)); |
| 3985 |
return pow2(exp); |
| 3986 |
} |
| 3987 |
}; |
| 3988 |
|
| 3989 |
} // namespace fast_float |
| 3990 |
|
| 3991 |
#endif |
| 3992 |
|
| 3993 |
#ifndef FASTFLOAT_DIGIT_COMPARISON_H |
| 3994 |
#define FASTFLOAT_DIGIT_COMPARISON_H |
| 3995 |
|
| 3996 |
#include <algorithm> |
| 3997 |
#include <cstdint> |
| 3998 |
#include <cstring> |
| 3999 |
#include <iterator> |
| 4000 |
|
| 4001 |
|
| 4002 |
namespace fast_float { |
| 4003 |
|
| 4004 |
// 1e0 to 1e19 |
| 4005 |
constexpr static uint64_t powers_of_ten_uint64[] = {1UL, |
| 4006 |
10UL, |
| 4007 |
100UL, |
| 4008 |
1000UL, |
| 4009 |
10000UL, |
| 4010 |
100000UL, |
| 4011 |
1000000UL, |
| 4012 |
10000000UL, |
| 4013 |
100000000UL, |
| 4014 |
1000000000UL, |
| 4015 |
10000000000UL, |
| 4016 |
100000000000UL, |
| 4017 |
1000000000000UL, |
| 4018 |
10000000000000UL, |
| 4019 |
100000000000000UL, |
| 4020 |
1000000000000000UL, |
| 4021 |
10000000000000000UL, |
| 4022 |
100000000000000000UL, |
| 4023 |
1000000000000000000UL, |
| 4024 |
10000000000000000000UL}; |
| 4025 |
|
| 4026 |
// calculate the exponent, in scientific notation, of the number. |
| 4027 |
// this algorithm is not even close to optimized, but it has no practical |
| 4028 |
// effect on performance: in order to have a faster algorithm, we'd need |
| 4029 |
// to slow down performance for faster algorithms, and this is still fast. |
| 4030 |
fastfloat_really_inline FASTFLOAT_CONSTEXPR14 int32_t |
| 4031 |
scientific_exponent(uint64_t mantissa, int32_t exponent) noexcept { |
| 4032 |
while (mantissa >= 10000) { |
| 4033 |
mantissa /= 10000; |
| 4034 |
exponent += 4; |
| 4035 |
} |
| 4036 |
while (mantissa >= 100) { |
| 4037 |
mantissa /= 100; |
| 4038 |
exponent += 2; |
| 4039 |
} |
| 4040 |
while (mantissa >= 10) { |
| 4041 |
mantissa /= 10; |
| 4042 |
exponent += 1; |
| 4043 |
} |
| 4044 |
return exponent; |
| 4045 |
} |
| 4046 |
|
| 4047 |
// this converts a native floating-point number to an extended-precision float. |
| 4048 |
template <typename T> |
| 4049 |
fastfloat_really_inline FASTFLOAT_CONSTEXPR20 adjusted_mantissa |
| 4050 |
to_extended(T value) noexcept { |
| 4051 |
using equiv_uint = equiv_uint_t<T>; |
| 4052 |
constexpr equiv_uint exponent_mask = binary_format<T>::exponent_mask(); |
| 4053 |
constexpr equiv_uint mantissa_mask = binary_format<T>::mantissa_mask(); |
| 4054 |
constexpr equiv_uint hidden_bit_mask = binary_format<T>::hidden_bit_mask(); |
| 4055 |
|
| 4056 |
adjusted_mantissa am; |
| 4057 |
int32_t bias = binary_format<T>::mantissa_explicit_bits() - |
| 4058 |
binary_format<T>::minimum_exponent(); |
| 4059 |
equiv_uint bits; |
| 4060 |
#if FASTFLOAT_HAS_BIT_CAST |
| 4061 |
bits = std::bit_cast<equiv_uint>(value); |
| 4062 |
#else |
| 4063 |
::memcpy(&bits, &value, sizeof(T)); |
| 4064 |
#endif |
| 4065 |
if ((bits & exponent_mask) == 0) { |
| 4066 |
// denormal |
| 4067 |
am.power2 = 1 - bias; |
| 4068 |
am.mantissa = bits & mantissa_mask; |
| 4069 |
} else { |
| 4070 |
// normal |
| 4071 |
am.power2 = int32_t((bits & exponent_mask) >> |
| 4072 |
binary_format<T>::mantissa_explicit_bits()); |
| 4073 |
am.power2 -= bias; |
| 4074 |
am.mantissa = (bits & mantissa_mask) | hidden_bit_mask; |
| 4075 |
} |
| 4076 |
|
| 4077 |
return am; |
| 4078 |
} |
| 4079 |
|
| 4080 |
// get the extended precision value of the halfway point between b and b+u. |
| 4081 |
// we are given a native float that represents b, so we need to adjust it |
| 4082 |
// halfway between b and b+u. |
| 4083 |
template <typename T> |
| 4084 |
fastfloat_really_inline FASTFLOAT_CONSTEXPR20 adjusted_mantissa |
| 4085 |
to_extended_halfway(T value) noexcept { |
| 4086 |
adjusted_mantissa am = to_extended(value); |
| 4087 |
am.mantissa <<= 1; |
| 4088 |
am.mantissa += 1; |
| 4089 |
am.power2 -= 1; |
| 4090 |
return am; |
| 4091 |
} |
| 4092 |
|
| 4093 |
// round an extended-precision float to the nearest machine float. |
| 4094 |
template <typename T, typename callback> |
| 4095 |
fastfloat_really_inline FASTFLOAT_CONSTEXPR14 void round(adjusted_mantissa &am, |
| 4096 |
callback cb) noexcept { |
| 4097 |
int32_t mantissa_shift = 64 - binary_format<T>::mantissa_explicit_bits() - 1; |
| 4098 |
if (-am.power2 >= mantissa_shift) { |
| 4099 |
// have a denormal float |
| 4100 |
int32_t shift = -am.power2 + 1; |
| 4101 |
cb(am, std::min<int32_t>(shift, 64)); |
| 4102 |
// check for round-up: if rounding-nearest carried us to the hidden bit. |
| 4103 |
am.power2 = (am.mantissa < |
| 4104 |
(uint64_t(1) << binary_format<T>::mantissa_explicit_bits())) |
| 4105 |
? 0 |
| 4106 |
: 1; |
| 4107 |
return; |
| 4108 |
} |
| 4109 |
|
| 4110 |
// have a normal float, use the default shift. |
| 4111 |
cb(am, mantissa_shift); |
| 4112 |
|
| 4113 |
// check for carry |
| 4114 |
if (am.mantissa >= |
| 4115 |
(uint64_t(2) << binary_format<T>::mantissa_explicit_bits())) { |
| 4116 |
am.mantissa = (uint64_t(1) << binary_format<T>::mantissa_explicit_bits()); |
| 4117 |
am.power2++; |
| 4118 |
} |
| 4119 |
|
| 4120 |
// check for infinite: we could have carried to an infinite power |
| 4121 |
am.mantissa &= ~(uint64_t(1) << binary_format<T>::mantissa_explicit_bits()); |
| 4122 |
if (am.power2 >= binary_format<T>::infinite_power()) { |
| 4123 |
am.power2 = binary_format<T>::infinite_power(); |
| 4124 |
am.mantissa = 0; |
| 4125 |
} |
| 4126 |
} |
| 4127 |
|
| 4128 |
template <typename callback> |
| 4129 |
fastfloat_really_inline FASTFLOAT_CONSTEXPR14 void |
| 4130 |
round_nearest_tie_even(adjusted_mantissa &am, int32_t shift, |
| 4131 |
callback cb) noexcept { |
| 4132 |
uint64_t const mask = (shift == 64) ? UINT64_MAX : (uint64_t(1) << shift) - 1; |
| 4133 |
uint64_t const halfway = (shift == 0) ? 0 : uint64_t(1) << (shift - 1); |
| 4134 |
uint64_t truncated_bits = am.mantissa & mask; |
| 4135 |
bool is_above = truncated_bits > halfway; |
| 4136 |
bool is_halfway = truncated_bits == halfway; |
| 4137 |
|
| 4138 |
// shift digits into position |
| 4139 |
if (shift == 64) { |
| 4140 |
am.mantissa = 0; |
| 4141 |
} else { |
| 4142 |
am.mantissa >>= shift; |
| 4143 |
} |
| 4144 |
am.power2 += shift; |
| 4145 |
|
| 4146 |
bool is_odd = (am.mantissa & 1) == 1; |
| 4147 |
am.mantissa += uint64_t(cb(is_odd, is_halfway, is_above)); |
| 4148 |
} |
| 4149 |
|
| 4150 |
fastfloat_really_inline FASTFLOAT_CONSTEXPR14 void |
| 4151 |
round_down(adjusted_mantissa &am, int32_t shift) noexcept { |
| 4152 |
if (shift == 64) { |
| 4153 |
am.mantissa = 0; |
| 4154 |
} else { |
| 4155 |
am.mantissa >>= shift; |
| 4156 |
} |
| 4157 |
am.power2 += shift; |
| 4158 |
} |
| 4159 |
|
| 4160 |
template <typename UC> |
| 4161 |
fastfloat_really_inline FASTFLOAT_CONSTEXPR20 void |
| 4162 |
skip_zeros(UC const *&first, UC const *last) noexcept { |
| 4163 |
uint64_t val; |
| 4164 |
while (!cpp20_and_in_constexpr() && |
| 4165 |
std::distance(first, last) >= int_cmp_len<UC>()) { |
| 4166 |
::memcpy(&val, first, sizeof(uint64_t)); |
| 4167 |
if (val != int_cmp_zeros<UC>()) { |
| 4168 |
break; |
| 4169 |
} |
| 4170 |
first += int_cmp_len<UC>(); |
| 4171 |
} |
| 4172 |
while (first != last) { |
| 4173 |
if (*first != UC('0')) { |
| 4174 |
break; |
| 4175 |
} |
| 4176 |
first++; |
| 4177 |
} |
| 4178 |
} |
| 4179 |
|
| 4180 |
// determine if any non-zero digits were truncated. |
| 4181 |
// all characters must be valid digits. |
| 4182 |
template <typename UC> |
| 4183 |
fastfloat_really_inline FASTFLOAT_CONSTEXPR20 bool |
| 4184 |
is_truncated(UC const *first, UC const *last) noexcept { |
| 4185 |
// do 8-bit optimizations, can just compare to 8 literal 0s. |
| 4186 |
uint64_t val; |
| 4187 |
while (!cpp20_and_in_constexpr() && |
| 4188 |
std::distance(first, last) >= int_cmp_len<UC>()) { |
| 4189 |
::memcpy(&val, first, sizeof(uint64_t)); |
| 4190 |
if (val != int_cmp_zeros<UC>()) { |
| 4191 |
return true; |
| 4192 |
} |
| 4193 |
first += int_cmp_len<UC>(); |
| 4194 |
} |
| 4195 |
while (first != last) { |
| 4196 |
if (*first != UC('0')) { |
| 4197 |
return true; |
| 4198 |
} |
| 4199 |
++first; |
| 4200 |
} |
| 4201 |
return false; |
| 4202 |
} |
| 4203 |
|
| 4204 |
template <typename UC> |
| 4205 |
fastfloat_really_inline FASTFLOAT_CONSTEXPR20 bool |
| 4206 |
is_truncated(span<UC const> s) noexcept { |
| 4207 |
return is_truncated(s.ptr, s.ptr + s.len()); |
| 4208 |
} |
| 4209 |
|
| 4210 |
template <typename UC> |
| 4211 |
fastfloat_really_inline FASTFLOAT_CONSTEXPR20 void |
| 4212 |
parse_eight_digits(UC const *&p, limb &value, size_t &counter, |
| 4213 |
size_t &count) noexcept { |
| 4214 |
value = value * 100000000 + parse_eight_digits_unrolled(p); |
| 4215 |
p += 8; |
| 4216 |
counter += 8; |
| 4217 |
count += 8; |
| 4218 |
} |
| 4219 |
|
| 4220 |
template <typename UC> |
| 4221 |
fastfloat_really_inline FASTFLOAT_CONSTEXPR14 void |
| 4222 |
parse_one_digit(UC const *&p, limb &value, size_t &counter, |
| 4223 |
size_t &count) noexcept { |
| 4224 |
value = value * 10 + limb(*p - UC('0')); |
| 4225 |
p++; |
| 4226 |
counter++; |
| 4227 |
count++; |
| 4228 |
} |
| 4229 |
|
| 4230 |
fastfloat_really_inline FASTFLOAT_CONSTEXPR20 void |
| 4231 |
add_native(bigint &big, limb power, limb value) noexcept { |
| 4232 |
big.mul(power); |
| 4233 |
big.add(value); |
| 4234 |
} |
| 4235 |
|
| 4236 |
fastfloat_really_inline FASTFLOAT_CONSTEXPR20 void |
| 4237 |
round_up_bigint(bigint &big, size_t &count) noexcept { |
| 4238 |
// need to round-up the digits, but need to avoid rounding |
| 4239 |
// ....9999 to ...10000, which could cause a false halfway point. |
| 4240 |
add_native(big, 10, 1); |
| 4241 |
count++; |
| 4242 |
} |
| 4243 |
|
| 4244 |
// parse the significant digits into a big integer |
| 4245 |
template <typename UC> |
| 4246 |
inline FASTFLOAT_CONSTEXPR20 void |
| 4247 |
parse_mantissa(bigint &result, parsed_number_string_t<UC> &num, |
| 4248 |
size_t max_digits, size_t &digits) noexcept { |
| 4249 |
// try to minimize the number of big integer and scalar multiplication. |
| 4250 |
// therefore, try to parse 8 digits at a time, and multiply by the largest |
| 4251 |
// scalar value (9 or 19 digits) for each step. |
| 4252 |
size_t counter = 0; |
| 4253 |
digits = 0; |
| 4254 |
limb value = 0; |
| 4255 |
#ifdef FASTFLOAT_64BIT_LIMB |
| 4256 |
size_t step = 19; |
| 4257 |
#else |
| 4258 |
size_t step = 9; |
| 4259 |
#endif |
| 4260 |
|
| 4261 |
// process all integer digits. |
| 4262 |
UC const *p = num.integer.ptr; |
| 4263 |
UC const *pend = p + num.integer.len(); |
| 4264 |
skip_zeros(p, pend); |
| 4265 |
// process all digits, in increments of step per loop |
| 4266 |
while (p != pend) { |
| 4267 |
while ((std::distance(p, pend) >= 8) && (step - counter >= 8) && |
| 4268 |
(max_digits - digits >= 8)) { |
| 4269 |
parse_eight_digits(p, value, counter, digits); |
| 4270 |
} |
| 4271 |
while (counter < step && p != pend && digits < max_digits) { |
| 4272 |
parse_one_digit(p, value, counter, digits); |
| 4273 |
} |
| 4274 |
if (digits == max_digits) { |
| 4275 |
// add the temporary value, then check if we've truncated any digits |
| 4276 |
add_native(result, limb(powers_of_ten_uint64[counter]), value); |
| 4277 |
bool truncated = is_truncated(p, pend); |
| 4278 |
if (num.fraction.ptr != nullptr) { |
| 4279 |
truncated |= is_truncated(num.fraction); |
| 4280 |
} |
| 4281 |
if (truncated) { |
| 4282 |
round_up_bigint(result, digits); |
| 4283 |
} |
| 4284 |
return; |
| 4285 |
} else { |
| 4286 |
add_native(result, limb(powers_of_ten_uint64[counter]), value); |
| 4287 |
counter = 0; |
| 4288 |
value = 0; |
| 4289 |
} |
| 4290 |
} |
| 4291 |
|
| 4292 |
// add our fraction digits, if they're available. |
| 4293 |
if (num.fraction.ptr != nullptr) { |
| 4294 |
p = num.fraction.ptr; |
| 4295 |
pend = p + num.fraction.len(); |
| 4296 |
if (digits == 0) { |
| 4297 |
skip_zeros(p, pend); |
| 4298 |
} |
| 4299 |
// process all digits, in increments of step per loop |
| 4300 |
while (p != pend) { |
| 4301 |
while ((std::distance(p, pend) >= 8) && (step - counter >= 8) && |
| 4302 |
(max_digits - digits >= 8)) { |
| 4303 |
parse_eight_digits(p, value, counter, digits); |
| 4304 |
} |
| 4305 |
while (counter < step && p != pend && digits < max_digits) { |
| 4306 |
parse_one_digit(p, value, counter, digits); |
| 4307 |
} |
| 4308 |
if (digits == max_digits) { |
| 4309 |
// add the temporary value, then check if we've truncated any digits |
| 4310 |
add_native(result, limb(powers_of_ten_uint64[counter]), value); |
| 4311 |
bool truncated = is_truncated(p, pend); |
| 4312 |
if (truncated) { |
| 4313 |
round_up_bigint(result, digits); |
| 4314 |
} |
| 4315 |
return; |
| 4316 |
} else { |
| 4317 |
add_native(result, limb(powers_of_ten_uint64[counter]), value); |
| 4318 |
counter = 0; |
| 4319 |
value = 0; |
| 4320 |
} |
| 4321 |
} |
| 4322 |
} |
| 4323 |
|
| 4324 |
if (counter != 0) { |
| 4325 |
add_native(result, limb(powers_of_ten_uint64[counter]), value); |
| 4326 |
} |
| 4327 |
} |
| 4328 |
|
| 4329 |
template <typename T> |
| 4330 |
inline FASTFLOAT_CONSTEXPR20 adjusted_mantissa |
| 4331 |
positive_digit_comp(bigint &bigmant, int32_t exponent) noexcept { |
| 4332 |
FASTFLOAT_ASSERT(bigmant.pow10(uint32_t(exponent))); |
| 4333 |
adjusted_mantissa answer; |
| 4334 |
bool truncated; |
| 4335 |
answer.mantissa = bigmant.hi64(truncated); |
| 4336 |
int bias = binary_format<T>::mantissa_explicit_bits() - |
| 4337 |
binary_format<T>::minimum_exponent(); |
| 4338 |
answer.power2 = bigmant.bit_length() - 64 + bias; |
| 4339 |
|
| 4340 |
round<T>(answer, [truncated](adjusted_mantissa &a, int32_t shift) { |
| 4341 |
round_nearest_tie_even( |
| 4342 |
a, shift, |
| 4343 |
[truncated](bool is_odd, bool is_halfway, bool is_above) -> bool { |
| 4344 |
return is_above || (is_halfway && truncated) || |
| 4345 |
(is_odd && is_halfway); |
| 4346 |
}); |
| 4347 |
}); |
| 4348 |
|
| 4349 |
return answer; |
| 4350 |
} |
| 4351 |
|
| 4352 |
// the scaling here is quite simple: we have, for the real digits `m * 10^e`, |
| 4353 |
// and for the theoretical digits `n * 2^f`. Since `e` is always negative, |
| 4354 |
// to scale them identically, we do `n * 2^f * 5^-f`, so we now have `m * 2^e`. |
| 4355 |
// we then need to scale by `2^(f- e)`, and then the two significant digits |
| 4356 |
// are of the same magnitude. |
| 4357 |
template <typename T> |
| 4358 |
inline FASTFLOAT_CONSTEXPR20 adjusted_mantissa negative_digit_comp( |
| 4359 |
bigint &bigmant, adjusted_mantissa am, int32_t exponent) noexcept { |
| 4360 |
bigint &real_digits = bigmant; |
| 4361 |
int32_t real_exp = exponent; |
| 4362 |
|
| 4363 |
// get the value of `b`, rounded down, and get a bigint representation of b+h |
| 4364 |
adjusted_mantissa am_b = am; |
| 4365 |
// gcc7 buf: use a lambda to remove the noexcept qualifier bug with |
| 4366 |
// -Wnoexcept-type. |
| 4367 |
round<T>(am_b, |
| 4368 |
[](adjusted_mantissa &a, int32_t shift) { round_down(a, shift); }); |
| 4369 |
T b; |
| 4370 |
to_float(false, am_b, b); |
| 4371 |
adjusted_mantissa theor = to_extended_halfway(b); |
| 4372 |
bigint theor_digits(theor.mantissa); |
| 4373 |
int32_t theor_exp = theor.power2; |
| 4374 |
|
| 4375 |
// scale real digits and theor digits to be same power. |
| 4376 |
int32_t pow2_exp = theor_exp - real_exp; |
| 4377 |
uint32_t pow5_exp = uint32_t(-real_exp); |
| 4378 |
if (pow5_exp != 0) { |
| 4379 |
FASTFLOAT_ASSERT(theor_digits.pow5(pow5_exp)); |
| 4380 |
} |
| 4381 |
if (pow2_exp > 0) { |
| 4382 |
FASTFLOAT_ASSERT(theor_digits.pow2(uint32_t(pow2_exp))); |
| 4383 |
} else if (pow2_exp < 0) { |
| 4384 |
FASTFLOAT_ASSERT(real_digits.pow2(uint32_t(-pow2_exp))); |
| 4385 |
} |
| 4386 |
|
| 4387 |
// compare digits, and use it to direct rounding |
| 4388 |
int ord = real_digits.compare(theor_digits); |
| 4389 |
adjusted_mantissa answer = am; |
| 4390 |
round<T>(answer, [ord](adjusted_mantissa &a, int32_t shift) { |
| 4391 |
round_nearest_tie_even( |
| 4392 |
a, shift, [ord](bool is_odd, bool _, bool __) -> bool { |
| 4393 |
(void)_; // not needed, since we've done our comparison |
| 4394 |
(void)__; // not needed, since we've done our comparison |
| 4395 |
if (ord > 0) { |
| 4396 |
return true; |
| 4397 |
} else if (ord < 0) { |
| 4398 |
return false; |
| 4399 |
} else { |
| 4400 |
return is_odd; |
| 4401 |
} |
| 4402 |
}); |
| 4403 |
}); |
| 4404 |
|
| 4405 |
return answer; |
| 4406 |
} |
| 4407 |
|
| 4408 |
// parse the significant digits as a big integer to unambiguously round |
| 4409 |
// the significant digits. here, we are trying to determine how to round |
| 4410 |
// an extended float representation close to `b+h`, halfway between `b` |
| 4411 |
// (the float rounded-down) and `b+u`, the next positive float. this |
| 4412 |
// algorithm is always correct, and uses one of two approaches. when |
| 4413 |
// the exponent is positive relative to the significant digits (such as |
| 4414 |
// 1234), we create a big-integer representation, get the high 64-bits, |
| 4415 |
// determine if any lower bits are truncated, and use that to direct |
| 4416 |
// rounding. in case of a negative exponent relative to the significant |
| 4417 |
// digits (such as 1.2345), we create a theoretical representation of |
| 4418 |
// `b` as a big-integer type, scaled to the same binary exponent as |
| 4419 |
// the actual digits. we then compare the big integer representations |
| 4420 |
// of both, and use that to direct rounding. |
| 4421 |
template <typename T, typename UC> |
| 4422 |
inline FASTFLOAT_CONSTEXPR20 adjusted_mantissa |
| 4423 |
digit_comp(parsed_number_string_t<UC> &num, adjusted_mantissa am) noexcept { |
| 4424 |
// remove the invalid exponent bias |
| 4425 |
am.power2 -= invalid_am_bias; |
| 4426 |
|
| 4427 |
int32_t sci_exp = |
| 4428 |
scientific_exponent(num.mantissa, static_cast<int32_t>(num.exponent)); |
| 4429 |
size_t max_digits = binary_format<T>::max_digits(); |
| 4430 |
size_t digits = 0; |
| 4431 |
bigint bigmant; |
| 4432 |
parse_mantissa(bigmant, num, max_digits, digits); |
| 4433 |
// can't underflow, since digits is at most max_digits. |
| 4434 |
int32_t exponent = sci_exp + 1 - int32_t(digits); |
| 4435 |
if (exponent >= 0) { |
| 4436 |
return positive_digit_comp<T>(bigmant, exponent); |
| 4437 |
} else { |
| 4438 |
return negative_digit_comp<T>(bigmant, am, exponent); |
| 4439 |
} |
| 4440 |
} |
| 4441 |
|
| 4442 |
} // namespace fast_float |
| 4443 |
|
| 4444 |
#endif |
| 4445 |
|
| 4446 |
#ifndef FASTFLOAT_PARSE_NUMBER_H |
| 4447 |
#define FASTFLOAT_PARSE_NUMBER_H |
| 4448 |
|
| 4449 |
|
| 4450 |
#include <cmath> |
| 4451 |
#include <cstring> |
| 4452 |
#include <limits> |
| 4453 |
#include <system_error> |
| 4454 |
|
| 4455 |
namespace fast_float { |
| 4456 |
|
| 4457 |
namespace detail { |
| 4458 |
/** |
| 4459 |
* Special case +inf, -inf, nan, infinity, -infinity. |
| 4460 |
* The case comparisons could be made much faster given that we know that the |
| 4461 |
* strings a null-free and fixed. |
| 4462 |
**/ |
| 4463 |
template <typename T, typename UC> |
| 4464 |
from_chars_result_t<UC> |
| 4465 |
FASTFLOAT_CONSTEXPR14 parse_infnan(UC const *first, UC const *last, |
| 4466 |
T &value, chars_format fmt) noexcept { |
| 4467 |
from_chars_result_t<UC> answer{}; |
| 4468 |
answer.ptr = first; |
| 4469 |
answer.ec = std::errc(); // be optimistic |
| 4470 |
// assume first < last, so dereference without checks; |
| 4471 |
bool const minusSign = (*first == UC('-')); |
| 4472 |
// C++17 20.19.3.(7.1) explicitly forbids '+' sign here |
| 4473 |
if ((*first == UC('-')) || |
| 4474 |
(uint64_t(fmt & chars_format::allow_leading_plus) && |
| 4475 |
(*first == UC('+')))) { |
| 4476 |
++first; |
| 4477 |
} |
| 4478 |
if (last - first >= 3) { |
| 4479 |
if (fastfloat_strncasecmp3(first, str_const_nan<UC>())) { |
| 4480 |
answer.ptr = (first += 3); |
| 4481 |
value = minusSign ? -std::numeric_limits<T>::quiet_NaN() |
| 4482 |
: std::numeric_limits<T>::quiet_NaN(); |
| 4483 |
// Check for possible nan(n-char-seq-opt), C++17 20.19.3.7, |
| 4484 |
// C11 7.20.1.3.3. At least MSVC produces nan(ind) and nan(snan). |
| 4485 |
if (first != last && *first == UC('(')) { |
| 4486 |
for (UC const *ptr = first + 1; ptr != last; ++ptr) { |
| 4487 |
if (*ptr == UC(')')) { |
| 4488 |
answer.ptr = ptr + 1; // valid nan(n-char-seq-opt) |
| 4489 |
break; |
| 4490 |
} else if (!((UC('a') <= *ptr && *ptr <= UC('z')) || |
| 4491 |
(UC('A') <= *ptr && *ptr <= UC('Z')) || |
| 4492 |
(UC('0') <= *ptr && *ptr <= UC('9')) || *ptr == UC('_'))) |
| 4493 |
break; // forbidden char, not nan(n-char-seq-opt) |
| 4494 |
} |
| 4495 |
} |
| 4496 |
return answer; |
| 4497 |
} |
| 4498 |
if (fastfloat_strncasecmp3(first, str_const_inf<UC>())) { |
| 4499 |
if ((last - first >= 8) && |
| 4500 |
fastfloat_strncasecmp5(first + 3, str_const_inf<UC>() + 3)) { |
| 4501 |
answer.ptr = first + 8; |
| 4502 |
} else { |
| 4503 |
answer.ptr = first + 3; |
| 4504 |
} |
| 4505 |
value = minusSign ? -std::numeric_limits<T>::infinity() |
| 4506 |
: std::numeric_limits<T>::infinity(); |
| 4507 |
return answer; |
| 4508 |
} |
| 4509 |
} |
| 4510 |
answer.ec = std::errc::invalid_argument; |
| 4511 |
return answer; |
| 4512 |
} |
| 4513 |
|
| 4514 |
/** |
| 4515 |
* Returns true if the floating-pointing rounding mode is to 'nearest'. |
| 4516 |
* It is the default on most system. This function is meant to be inexpensive. |
| 4517 |
* Credit : @mwalcott3 |
| 4518 |
*/ |
| 4519 |
fastfloat_really_inline bool rounds_to_nearest() noexcept { |
| 4520 |
// https://lemire.me/blog/2020/06/26/gcc-not-nearest/ |
| 4521 |
#if (FLT_EVAL_METHOD != 1) && (FLT_EVAL_METHOD != 0) |
| 4522 |
return false; |
| 4523 |
#endif |
| 4524 |
// See |
| 4525 |
// A fast function to check your floating-point rounding mode |
| 4526 |
// https://lemire.me/blog/2022/11/16/a-fast-function-to-check-your-floating-point-rounding-mode/ |
| 4527 |
// |
| 4528 |
// This function is meant to be equivalent to : |
| 4529 |
// prior: #include <cfenv> |
| 4530 |
// return fegetround() == FE_TONEAREST; |
| 4531 |
// However, it is expected to be much faster than the fegetround() |
| 4532 |
// function call. |
| 4533 |
// |
| 4534 |
// The volatile keyword prevents the compiler from computing the function |
| 4535 |
// at compile-time. |
| 4536 |
// There might be other ways to prevent compile-time optimizations (e.g., |
| 4537 |
// asm). The value does not need to be std::numeric_limits<float>::min(), any |
| 4538 |
// small value so that 1 + x should round to 1 would do (after accounting for |
| 4539 |
// excess precision, as in 387 instructions). |
| 4540 |
static float volatile fmin = std::numeric_limits<float>::min(); |
| 4541 |
float fmini = fmin; // we copy it so that it gets loaded at most once. |
| 4542 |
// |
| 4543 |
// Explanation: |
| 4544 |
// Only when fegetround() == FE_TONEAREST do we have that |
| 4545 |
// fmin + 1.0f == 1.0f - fmin. |
| 4546 |
// |
| 4547 |
// FE_UPWARD: |
| 4548 |
// fmin + 1.0f > 1 |
| 4549 |
// 1.0f - fmin == 1 |
| 4550 |
// |
| 4551 |
// FE_DOWNWARD or FE_TOWARDZERO: |
| 4552 |
// fmin + 1.0f == 1 |
| 4553 |
// 1.0f - fmin < 1 |
| 4554 |
// |
| 4555 |
// Note: This may fail to be accurate if fast-math has been |
| 4556 |
// enabled, as rounding conventions may not apply. |
| 4557 |
#ifdef FASTFLOAT_VISUAL_STUDIO |
| 4558 |
#pragma warning(push) |
| 4559 |
// todo: is there a VS warning? |
| 4560 |
// see |
| 4561 |
// https://stackoverflow.com/questions/46079446/is-there-a-warning-for-floating-point-equality-checking-in-visual-studio-2013 |
| 4562 |
#elif defined(__clang__) |
| 4563 |
#pragma clang diagnostic push |
| 4564 |
#pragma clang diagnostic ignored "-Wfloat-equal" |
| 4565 |
#elif defined(__GNUC__) |
| 4566 |
#pragma GCC diagnostic push |
| 4567 |
#pragma GCC diagnostic ignored "-Wfloat-equal" |
| 4568 |
#endif |
| 4569 |
return (fmini + 1.0f == 1.0f - fmini); |
| 4570 |
#ifdef FASTFLOAT_VISUAL_STUDIO |
| 4571 |
#pragma warning(pop) |
| 4572 |
#elif defined(__clang__) |
| 4573 |
#pragma clang diagnostic pop |
| 4574 |
#elif defined(__GNUC__) |
| 4575 |
#pragma GCC diagnostic pop |
| 4576 |
#endif |
| 4577 |
} |
| 4578 |
|
| 4579 |
} // namespace detail |
| 4580 |
|
| 4581 |
template <typename T> struct from_chars_caller { |
| 4582 |
template <typename UC> |
| 4583 |
FASTFLOAT_CONSTEXPR20 static from_chars_result_t<UC> |
| 4584 |
call(UC const *first, UC const *last, T &value, |
| 4585 |
parse_options_t<UC> options) noexcept { |
| 4586 |
return from_chars_advanced(first, last, value, options); |
| 4587 |
} |
| 4588 |
}; |
| 4589 |
|
| 4590 |
#ifdef __STDCPP_FLOAT32_T__ |
| 4591 |
template <> struct from_chars_caller<std::float32_t> { |
| 4592 |
template <typename UC> |
| 4593 |
FASTFLOAT_CONSTEXPR20 static from_chars_result_t<UC> |
| 4594 |
call(UC const *first, UC const *last, std::float32_t &value, |
| 4595 |
parse_options_t<UC> options) noexcept { |
| 4596 |
// if std::float32_t is defined, and we are in C++23 mode; macro set for |
| 4597 |
// float32; set value to float due to equivalence between float and |
| 4598 |
// float32_t |
| 4599 |
float val; |
| 4600 |
auto ret = from_chars_advanced(first, last, val, options); |
| 4601 |
value = val; |
| 4602 |
return ret; |
| 4603 |
} |
| 4604 |
}; |
| 4605 |
#endif |
| 4606 |
|
| 4607 |
#ifdef __STDCPP_FLOAT64_T__ |
| 4608 |
template <> struct from_chars_caller<std::float64_t> { |
| 4609 |
template <typename UC> |
| 4610 |
FASTFLOAT_CONSTEXPR20 static from_chars_result_t<UC> |
| 4611 |
call(UC const *first, UC const *last, std::float64_t &value, |
| 4612 |
parse_options_t<UC> options) noexcept { |
| 4613 |
// if std::float64_t is defined, and we are in C++23 mode; macro set for |
| 4614 |
// float64; set value as double due to equivalence between double and |
| 4615 |
// float64_t |
| 4616 |
double val; |
| 4617 |
auto ret = from_chars_advanced(first, last, val, options); |
| 4618 |
value = val; |
| 4619 |
return ret; |
| 4620 |
} |
| 4621 |
}; |
| 4622 |
#endif |
| 4623 |
|
| 4624 |
template <typename T, typename UC, typename> |
| 4625 |
FASTFLOAT_CONSTEXPR20 from_chars_result_t<UC> |
| 4626 |
from_chars(UC const *first, UC const *last, T &value, |
| 4627 |
chars_format fmt /*= chars_format::general*/) noexcept { |
| 4628 |
return from_chars_caller<T>::call(first, last, value, |
| 4629 |
parse_options_t<UC>(fmt)); |
| 4630 |
} |
| 4631 |
|
| 4632 |
template <typename T> |
| 4633 |
fastfloat_really_inline FASTFLOAT_CONSTEXPR20 bool |
| 4634 |
clinger_fast_path_impl(uint64_t mantissa, int64_t exponent, bool is_negative, |
| 4635 |
T &value) noexcept { |
| 4636 |
// The implementation of the Clinger's fast path is convoluted because |
| 4637 |
// we want round-to-nearest in all cases, irrespective of the rounding mode |
| 4638 |
// selected on the thread. |
| 4639 |
// We proceed optimistically, assuming that detail::rounds_to_nearest() |
| 4640 |
// returns true. |
| 4641 |
if (binary_format<T>::min_exponent_fast_path() <= exponent && |
| 4642 |
exponent <= binary_format<T>::max_exponent_fast_path()) { |
| 4643 |
// Unfortunately, the conventional Clinger's fast path is only possible |
| 4644 |
// when the system rounds to the nearest float. |
| 4645 |
// |
| 4646 |
// We expect the next branch to almost always be selected. |
| 4647 |
// We could check it first (before the previous branch), but |
| 4648 |
// there might be performance advantages at having the check |
| 4649 |
// be last. |
| 4650 |
if (!cpp20_and_in_constexpr() && detail::rounds_to_nearest()) { |
| 4651 |
// We have that fegetround() == FE_TONEAREST. |
| 4652 |
// Next is Clinger's fast path. |
| 4653 |
if (mantissa <= binary_format<T>::max_mantissa_fast_path()) { |
| 4654 |
value = T(mantissa); |
| 4655 |
if (exponent < 0) { |
| 4656 |
value = value / binary_format<T>::exact_power_of_ten(-exponent); |
| 4657 |
} else { |
| 4658 |
value = value * binary_format<T>::exact_power_of_ten(exponent); |
| 4659 |
} |
| 4660 |
if (is_negative) { |
| 4661 |
value = -value; |
| 4662 |
} |
| 4663 |
return true; |
| 4664 |
} |
| 4665 |
} else { |
| 4666 |
// We do not have that fegetround() == FE_TONEAREST. |
| 4667 |
// Next is a modified Clinger's fast path, inspired by Jakub Jelínek's |
| 4668 |
// proposal |
| 4669 |
if (exponent >= 0 && |
| 4670 |
mantissa <= binary_format<T>::max_mantissa_fast_path(exponent)) { |
| 4671 |
#if defined(__clang__) || defined(FASTFLOAT_32BIT) |
| 4672 |
// Clang may map 0 to -0.0 when fegetround() == FE_DOWNWARD |
| 4673 |
if (mantissa == 0) { |
| 4674 |
value = is_negative ? T(-0.) : T(0.); |
| 4675 |
return true; |
| 4676 |
} |
| 4677 |
#endif |
| 4678 |
value = T(mantissa) * binary_format<T>::exact_power_of_ten(exponent); |
| 4679 |
if (is_negative) { |
| 4680 |
value = -value; |
| 4681 |
} |
| 4682 |
return true; |
| 4683 |
} |
| 4684 |
} |
| 4685 |
} |
| 4686 |
return false; |
| 4687 |
} |
| 4688 |
|
| 4689 |
/** |
| 4690 |
* This function overload takes parsed_number_string_t structure that is created |
| 4691 |
* and populated either by from_chars_advanced function taking chars range and |
| 4692 |
* parsing options or other parsing custom function implemented by user. |
| 4693 |
*/ |
| 4694 |
template <typename T, typename UC> |
| 4695 |
fastfloat_really_inline FASTFLOAT_CONSTEXPR20 from_chars_result_t<UC> |
| 4696 |
from_chars_advanced(parsed_number_string_t<UC> &pns, T &value) noexcept { |
| 4697 |
static_assert(is_supported_float_type<T>::value, |
| 4698 |
"only some floating-point types are supported"); |
| 4699 |
static_assert(is_supported_char_type<UC>::value, |
| 4700 |
"only char, wchar_t, char16_t and char32_t are supported"); |
| 4701 |
|
| 4702 |
from_chars_result_t<UC> answer; |
| 4703 |
|
| 4704 |
answer.ec = std::errc(); // be optimistic |
| 4705 |
answer.ptr = pns.lastmatch; |
| 4706 |
|
| 4707 |
if (!pns.too_many_digits && |
| 4708 |
clinger_fast_path_impl(pns.mantissa, pns.exponent, pns.negative, value)) |
| 4709 |
return answer; |
| 4710 |
|
| 4711 |
adjusted_mantissa am = |
| 4712 |
compute_float<binary_format<T>>(pns.exponent, pns.mantissa); |
| 4713 |
if (pns.too_many_digits && am.power2 >= 0) { |
| 4714 |
if (am != compute_float<binary_format<T>>(pns.exponent, pns.mantissa + 1)) { |
| 4715 |
am = compute_error<binary_format<T>>(pns.exponent, pns.mantissa); |
| 4716 |
} |
| 4717 |
} |
| 4718 |
// If we called compute_float<binary_format<T>>(pns.exponent, pns.mantissa) |
| 4719 |
// and we have an invalid power (am.power2 < 0), then we need to go the long |
| 4720 |
// way around again. This is very uncommon. |
| 4721 |
if (am.power2 < 0) { |
| 4722 |
am = digit_comp<T>(pns, am); |
| 4723 |
} |
| 4724 |
to_float(pns.negative, am, value); |
| 4725 |
// Test for over/underflow. |
| 4726 |
if ((pns.mantissa != 0 && am.mantissa == 0 && am.power2 == 0) || |
| 4727 |
am.power2 == binary_format<T>::infinite_power()) { |
| 4728 |
answer.ec = std::errc::result_out_of_range; |
| 4729 |
} |
| 4730 |
return answer; |
| 4731 |
} |
| 4732 |
|
| 4733 |
template <typename T, typename UC> |
| 4734 |
fastfloat_really_inline FASTFLOAT_CONSTEXPR20 from_chars_result_t<UC> |
| 4735 |
from_chars_float_advanced(UC const *first, UC const *last, T &value, |
| 4736 |
parse_options_t<UC> options) noexcept { |
| 4737 |
|
| 4738 |
static_assert(is_supported_float_type<T>::value, |
| 4739 |
"only some floating-point types are supported"); |
| 4740 |
static_assert(is_supported_char_type<UC>::value, |
| 4741 |
"only char, wchar_t, char16_t and char32_t are supported"); |
| 4742 |
|
| 4743 |
chars_format const fmt = detail::adjust_for_feature_macros(options.format); |
| 4744 |
|
| 4745 |
from_chars_result_t<UC> answer; |
| 4746 |
if (uint64_t(fmt & chars_format::skip_white_space)) { |
| 4747 |
while ((first != last) && fast_float::is_space(*first)) { |
| 4748 |
first++; |
| 4749 |
} |
| 4750 |
} |
| 4751 |
if (first == last) { |
| 4752 |
answer.ec = std::errc::invalid_argument; |
| 4753 |
answer.ptr = first; |
| 4754 |
return answer; |
| 4755 |
} |
| 4756 |
parsed_number_string_t<UC> pns = |
| 4757 |
uint64_t(fmt & detail::basic_json_fmt) |
| 4758 |
? parse_number_string<true, UC>(first, last, options) |
| 4759 |
: parse_number_string<false, UC>(first, last, options); |
| 4760 |
if (!pns.valid) { |
| 4761 |
if (uint64_t(fmt & chars_format::no_infnan)) { |
| 4762 |
answer.ec = std::errc::invalid_argument; |
| 4763 |
answer.ptr = first; |
| 4764 |
return answer; |
| 4765 |
} else { |
| 4766 |
return detail::parse_infnan(first, last, value, fmt); |
| 4767 |
} |
| 4768 |
} |
| 4769 |
|
| 4770 |
// call overload that takes parsed_number_string_t directly. |
| 4771 |
return from_chars_advanced(pns, value); |
| 4772 |
} |
| 4773 |
|
| 4774 |
template <typename T, typename UC, typename> |
| 4775 |
FASTFLOAT_CONSTEXPR20 from_chars_result_t<UC> |
| 4776 |
from_chars(UC const *first, UC const *last, T &value, int base) noexcept { |
| 4777 |
|
| 4778 |
static_assert(is_supported_integer_type<T>::value, |
| 4779 |
"only integer types are supported"); |
| 4780 |
static_assert(is_supported_char_type<UC>::value, |
| 4781 |
"only char, wchar_t, char16_t and char32_t are supported"); |
| 4782 |
|
| 4783 |
parse_options_t<UC> options; |
| 4784 |
options.base = base; |
| 4785 |
return from_chars_advanced(first, last, value, options); |
| 4786 |
} |
| 4787 |
|
| 4788 |
template <typename T> |
| 4789 |
FASTFLOAT_CONSTEXPR20 |
| 4790 |
typename std::enable_if<is_supported_float_type<T>::value, T>::type |
| 4791 |
integer_times_pow10(uint64_t mantissa, int decimal_exponent) noexcept { |
| 4792 |
T value; |
| 4793 |
if (clinger_fast_path_impl(mantissa, decimal_exponent, false, value)) |
| 4794 |
return value; |
| 4795 |
|
| 4796 |
adjusted_mantissa am = |
| 4797 |
compute_float<binary_format<T>>(decimal_exponent, mantissa); |
| 4798 |
to_float(false, am, value); |
| 4799 |
return value; |
| 4800 |
} |
| 4801 |
|
| 4802 |
template <typename T> |
| 4803 |
FASTFLOAT_CONSTEXPR20 |
| 4804 |
typename std::enable_if<is_supported_float_type<T>::value, T>::type |
| 4805 |
integer_times_pow10(int64_t mantissa, int decimal_exponent) noexcept { |
| 4806 |
const bool is_negative = mantissa < 0; |
| 4807 |
const uint64_t m = static_cast<uint64_t>(is_negative ? -mantissa : mantissa); |
| 4808 |
|
| 4809 |
T value; |
| 4810 |
if (clinger_fast_path_impl(m, decimal_exponent, is_negative, value)) |
| 4811 |
return value; |
| 4812 |
|
| 4813 |
adjusted_mantissa am = compute_float<binary_format<T>>(decimal_exponent, m); |
| 4814 |
to_float(is_negative, am, value); |
| 4815 |
return value; |
| 4816 |
} |
| 4817 |
|
| 4818 |
FASTFLOAT_CONSTEXPR20 inline double |
| 4819 |
integer_times_pow10(uint64_t mantissa, int decimal_exponent) noexcept { |
| 4820 |
return integer_times_pow10<double>(mantissa, decimal_exponent); |
| 4821 |
} |
| 4822 |
|
| 4823 |
FASTFLOAT_CONSTEXPR20 inline double |
| 4824 |
integer_times_pow10(int64_t mantissa, int decimal_exponent) noexcept { |
| 4825 |
return integer_times_pow10<double>(mantissa, decimal_exponent); |
| 4826 |
} |
| 4827 |
|
| 4828 |
// the following overloads are here to avoid surprising ambiguity for int, |
| 4829 |
// unsigned, etc. |
| 4830 |
template <typename T, typename Int> |
| 4831 |
FASTFLOAT_CONSTEXPR20 |
| 4832 |
typename std::enable_if<is_supported_float_type<T>::value && |
| 4833 |
std::is_integral<Int>::value && |
| 4834 |
!std::is_signed<Int>::value, |
| 4835 |
T>::type |
| 4836 |
integer_times_pow10(Int mantissa, int decimal_exponent) noexcept { |
| 4837 |
return integer_times_pow10<T>(static_cast<uint64_t>(mantissa), |
| 4838 |
decimal_exponent); |
| 4839 |
} |
| 4840 |
|
| 4841 |
template <typename T, typename Int> |
| 4842 |
FASTFLOAT_CONSTEXPR20 |
| 4843 |
typename std::enable_if<is_supported_float_type<T>::value && |
| 4844 |
std::is_integral<Int>::value && |
| 4845 |
std::is_signed<Int>::value, |
| 4846 |
T>::type |
| 4847 |
integer_times_pow10(Int mantissa, int decimal_exponent) noexcept { |
| 4848 |
return integer_times_pow10<T>(static_cast<int64_t>(mantissa), |
| 4849 |
decimal_exponent); |
| 4850 |
} |
| 4851 |
|
| 4852 |
template <typename Int> |
| 4853 |
FASTFLOAT_CONSTEXPR20 typename std::enable_if< |
| 4854 |
std::is_integral<Int>::value && !std::is_signed<Int>::value, double>::type |
| 4855 |
integer_times_pow10(Int mantissa, int decimal_exponent) noexcept { |
| 4856 |
return integer_times_pow10(static_cast<uint64_t>(mantissa), decimal_exponent); |
| 4857 |
} |
| 4858 |
|
| 4859 |
template <typename Int> |
| 4860 |
FASTFLOAT_CONSTEXPR20 typename std::enable_if< |
| 4861 |
std::is_integral<Int>::value && std::is_signed<Int>::value, double>::type |
| 4862 |
integer_times_pow10(Int mantissa, int decimal_exponent) noexcept { |
| 4863 |
return integer_times_pow10(static_cast<int64_t>(mantissa), decimal_exponent); |
| 4864 |
} |
| 4865 |
|
| 4866 |
template <typename T, typename UC> |
| 4867 |
FASTFLOAT_CONSTEXPR20 from_chars_result_t<UC> |
| 4868 |
from_chars_int_advanced(UC const *first, UC const *last, T &value, |
| 4869 |
parse_options_t<UC> options) noexcept { |
| 4870 |
|
| 4871 |
static_assert(is_supported_integer_type<T>::value, |
| 4872 |
"only integer types are supported"); |
| 4873 |
static_assert(is_supported_char_type<UC>::value, |
| 4874 |
"only char, wchar_t, char16_t and char32_t are supported"); |
| 4875 |
|
| 4876 |
chars_format const fmt = detail::adjust_for_feature_macros(options.format); |
| 4877 |
int const base = options.base; |
| 4878 |
|
| 4879 |
from_chars_result_t<UC> answer; |
| 4880 |
if (uint64_t(fmt & chars_format::skip_white_space)) { |
| 4881 |
while ((first != last) && fast_float::is_space(*first)) { |
| 4882 |
first++; |
| 4883 |
} |
| 4884 |
} |
| 4885 |
if (first == last || base < 2 || base > 36) { |
| 4886 |
answer.ec = std::errc::invalid_argument; |
| 4887 |
answer.ptr = first; |
| 4888 |
return answer; |
| 4889 |
} |
| 4890 |
|
| 4891 |
return parse_int_string(first, last, value, options); |
| 4892 |
} |
| 4893 |
|
| 4894 |
template <size_t TypeIx> struct from_chars_advanced_caller { |
| 4895 |
static_assert(TypeIx > 0, "unsupported type"); |
| 4896 |
}; |
| 4897 |
|
| 4898 |
template <> struct from_chars_advanced_caller<1> { |
| 4899 |
template <typename T, typename UC> |
| 4900 |
fastfloat_really_inline |
| 4901 |
FASTFLOAT_CONSTEXPR20 static from_chars_result_t<UC> |
| 4902 |
call(UC const *first, UC const *last, T &value, |
| 4903 |
parse_options_t<UC> options) noexcept { |
| 4904 |
return from_chars_float_advanced(first, last, value, options); |
| 4905 |
} |
| 4906 |
}; |
| 4907 |
|
| 4908 |
template <> struct from_chars_advanced_caller<2> { |
| 4909 |
template <typename T, typename UC> |
| 4910 |
fastfloat_really_inline FASTFLOAT_CONSTEXPR20 static from_chars_result_t<UC> |
| 4911 |
call(UC const *first, UC const *last, T &value, |
| 4912 |
parse_options_t<UC> options) noexcept { |
| 4913 |
return from_chars_int_advanced(first, last, value, options); |
| 4914 |
} |
| 4915 |
}; |
| 4916 |
|
| 4917 |
template <typename T, typename UC> |
| 4918 |
fastfloat_really_inline FASTFLOAT_CONSTEXPR20 from_chars_result_t<UC> |
| 4919 |
from_chars_advanced(UC const *first, UC const *last, T &value, |
| 4920 |
parse_options_t<UC> options) noexcept { |
| 4921 |
return from_chars_advanced_caller< |
| 4922 |
size_t(is_supported_float_type<T>::value) + |
| 4923 |
2 * size_t(is_supported_integer_type<T>::value)>::call(first, last, value, |
| 4924 |
options); |
| 4925 |
} |
| 4926 |
|
| 4927 |
} // namespace fast_float |
| 4928 |
|
| 4929 |
#endif |
| 4930 |
|