| 1 | // Copyright 2020-2023 Daniel Lemire |
| 2 | // Copyright 2023 Matt Borland |
| 3 | // Distributed under the Boost Software License, Version 1.0. |
| 4 | // https://www.boost.org/LICENSE_1_0.txt |
| 5 | // |
| 6 | // Derivative of: https://github.com/fastfloat/fast_float |
| 7 | |
| 8 | #ifndef BOOST_CHARCONV_DETAIL_FASTFLOAT_DIGIT_COMPARISON_HPP |
| 9 | #define BOOST_CHARCONV_DETAIL_FASTFLOAT_DIGIT_COMPARISON_HPP |
| 10 | |
| 11 | #include <boost/charconv/detail/fast_float/float_common.hpp> |
| 12 | #include <boost/charconv/detail/fast_float/bigint.hpp> |
| 13 | #include <boost/charconv/detail/fast_float/ascii_number.hpp> |
| 14 | #include <algorithm> |
| 15 | #include <cstdint> |
| 16 | #include <cstring> |
| 17 | #include <iterator> |
| 18 | |
| 19 | namespace boost { namespace charconv { namespace detail { namespace fast_float { |
| 20 | |
| 21 | // 1e0 to 1e19 |
| 22 | constexpr static uint64_t powers_of_ten_uint64[] = { |
| 23 | 1UL, 10UL, 100UL, 1000UL, 10000UL, 100000UL, 1000000UL, 10000000UL, 100000000UL, |
| 24 | 1000000000UL, 10000000000UL, 100000000000UL, 1000000000000UL, 10000000000000UL, |
| 25 | 100000000000000UL, 1000000000000000UL, 10000000000000000UL, 100000000000000000UL, |
| 26 | 1000000000000000000UL, 10000000000000000000UL}; |
| 27 | |
| 28 | // calculate the exponent, in scientific notation, of the number. |
| 29 | // this algorithm is not even close to optimized, but it has no practical |
| 30 | // effect on performance: in order to have a faster algorithm, we'd need |
| 31 | // to slow down performance for faster algorithms, and this is still fast. |
| 32 | template <typename UC> |
| 33 | BOOST_FORCEINLINE BOOST_CHARCONV_FASTFLOAT_CONSTEXPR14 |
| 34 | int32_t scientific_exponent(parsed_number_string_t<UC> & num) noexcept { |
| 35 | uint64_t mantissa = num.mantissa; |
| 36 | int32_t exponent = int32_t(num.exponent); |
| 37 | while (mantissa >= 10000) { |
| 38 | mantissa /= 10000; |
| 39 | exponent += 4; |
| 40 | } |
| 41 | while (mantissa >= 100) { |
| 42 | mantissa /= 100; |
| 43 | exponent += 2; |
| 44 | } |
| 45 | while (mantissa >= 10) { |
| 46 | mantissa /= 10; |
| 47 | exponent += 1; |
| 48 | } |
| 49 | return exponent; |
| 50 | } |
| 51 | |
| 52 | // this converts a native floating-point number to an extended-precision float. |
| 53 | template <typename T> |
| 54 | BOOST_FORCEINLINE BOOST_CHARCONV_FASTFLOAT_CONSTEXPR20 |
| 55 | adjusted_mantissa to_extended(T value) noexcept { |
| 56 | using equiv_uint = typename binary_format<T>::equiv_uint; |
| 57 | constexpr equiv_uint exponent_mask = binary_format<T>::exponent_mask(); |
| 58 | constexpr equiv_uint mantissa_mask = binary_format<T>::mantissa_mask(); |
| 59 | constexpr equiv_uint hidden_bit_mask = binary_format<T>::hidden_bit_mask(); |
| 60 | |
| 61 | adjusted_mantissa am; |
| 62 | int32_t bias = binary_format<T>::mantissa_explicit_bits() - binary_format<T>::minimum_exponent(); |
| 63 | equiv_uint bits; |
| 64 | #if BOOST_CHARCONV_FASTFLOAT_HAS_BIT_CAST |
| 65 | bits = std::bit_cast<equiv_uint>(value); |
| 66 | #else |
| 67 | ::memcpy(dest: &bits, src: &value, n: sizeof(T)); |
| 68 | #endif |
| 69 | if ((bits & exponent_mask) == 0) { |
| 70 | // denormal |
| 71 | am.power2 = 1 - bias; |
| 72 | am.mantissa = bits & mantissa_mask; |
| 73 | } else { |
| 74 | // normal |
| 75 | am.power2 = int32_t((bits & exponent_mask) >> binary_format<T>::mantissa_explicit_bits()); |
| 76 | am.power2 -= bias; |
| 77 | am.mantissa = (bits & mantissa_mask) | hidden_bit_mask; |
| 78 | } |
| 79 | |
| 80 | return am; |
| 81 | } |
| 82 | |
| 83 | // get the extended precision value of the halfway point between b and b+u. |
| 84 | // we are given a native float that represents b, so we need to adjust it |
| 85 | // halfway between b and b+u. |
| 86 | template <typename T> |
| 87 | BOOST_FORCEINLINE BOOST_CHARCONV_FASTFLOAT_CONSTEXPR20 |
| 88 | adjusted_mantissa to_extended_halfway(T value) noexcept { |
| 89 | adjusted_mantissa am = to_extended(value); |
| 90 | am.mantissa <<= 1; |
| 91 | am.mantissa += 1; |
| 92 | am.power2 -= 1; |
| 93 | return am; |
| 94 | } |
| 95 | |
| 96 | // round an extended-precision float to the nearest machine float. |
| 97 | template <typename T, typename callback> |
| 98 | BOOST_FORCEINLINE BOOST_CHARCONV_FASTFLOAT_CONSTEXPR14 |
| 99 | void round(adjusted_mantissa& am, callback cb) noexcept { |
| 100 | int32_t mantissa_shift = 64 - binary_format<T>::mantissa_explicit_bits() - 1; |
| 101 | if (-am.power2 >= mantissa_shift) { |
| 102 | // have a denormal float |
| 103 | int32_t shift = -am.power2 + 1; |
| 104 | cb(am, std::min<int32_t>(a: shift, b: 64)); |
| 105 | // check for round-up: if rounding-nearest carried us to the hidden bit. |
| 106 | am.power2 = (am.mantissa < (uint64_t(1) << binary_format<T>::mantissa_explicit_bits())) ? 0 : 1; |
| 107 | return; |
| 108 | } |
| 109 | |
| 110 | // have a normal float, use the default shift. |
| 111 | cb(am, mantissa_shift); |
| 112 | |
| 113 | // check for carry |
| 114 | if (am.mantissa >= (uint64_t(2) << binary_format<T>::mantissa_explicit_bits())) { |
| 115 | am.mantissa = (uint64_t(1) << binary_format<T>::mantissa_explicit_bits()); |
| 116 | am.power2++; |
| 117 | } |
| 118 | |
| 119 | // check for infinite: we could have carried to an infinite power |
| 120 | am.mantissa &= ~(uint64_t(1) << binary_format<T>::mantissa_explicit_bits()); |
| 121 | if (am.power2 >= binary_format<T>::infinite_power()) { |
| 122 | am.power2 = binary_format<T>::infinite_power(); |
| 123 | am.mantissa = 0; |
| 124 | } |
| 125 | } |
| 126 | |
| 127 | template <typename callback> |
| 128 | BOOST_FORCEINLINE BOOST_CHARCONV_FASTFLOAT_CONSTEXPR14 |
| 129 | void round_nearest_tie_even(adjusted_mantissa& am, int32_t shift, callback cb) noexcept { |
| 130 | const uint64_t mask |
| 131 | = (shift == 64) |
| 132 | ? UINT64_MAX |
| 133 | : (uint64_t(1) << shift) - 1; |
| 134 | const uint64_t halfway |
| 135 | = (shift == 0) |
| 136 | ? 0 |
| 137 | : uint64_t(1) << (shift - 1); |
| 138 | uint64_t truncated_bits = am.mantissa & mask; |
| 139 | bool is_above = truncated_bits > halfway; |
| 140 | bool is_halfway = truncated_bits == halfway; |
| 141 | |
| 142 | // shift digits into position |
| 143 | if (shift == 64) { |
| 144 | am.mantissa = 0; |
| 145 | } else { |
| 146 | am.mantissa >>= shift; |
| 147 | } |
| 148 | am.power2 += shift; |
| 149 | |
| 150 | bool is_odd = (am.mantissa & 1) == 1; |
| 151 | am.mantissa += uint64_t(cb(is_odd, is_halfway, is_above)); |
| 152 | } |
| 153 | |
| 154 | BOOST_FORCEINLINE BOOST_CHARCONV_FASTFLOAT_CONSTEXPR14 |
| 155 | void round_down(adjusted_mantissa& am, int32_t shift) noexcept { |
| 156 | if (shift == 64) { |
| 157 | am.mantissa = 0; |
| 158 | } else { |
| 159 | am.mantissa >>= shift; |
| 160 | } |
| 161 | am.power2 += shift; |
| 162 | } |
| 163 | template <typename UC> |
| 164 | BOOST_FORCEINLINE BOOST_CHARCONV_FASTFLOAT_CONSTEXPR20 |
| 165 | void skip_zeros(UC const * & first, UC const * last) noexcept { |
| 166 | uint64_t val; |
| 167 | while (!cpp20_and_in_constexpr() && std::distance(first, last) >= int_cmp_len<UC>()) { |
| 168 | ::memcpy(dest: &val, src: first, n: sizeof(uint64_t)); |
| 169 | if (val != int_cmp_zeros<UC>()) { |
| 170 | break; |
| 171 | } |
| 172 | first += int_cmp_len<UC>(); |
| 173 | } |
| 174 | while (first != last) { |
| 175 | if (*first != UC('0')) { |
| 176 | break; |
| 177 | } |
| 178 | first++; |
| 179 | } |
| 180 | } |
| 181 | |
| 182 | // determine if any non-zero digits were truncated. |
| 183 | // all characters must be valid digits. |
| 184 | template <typename UC> |
| 185 | BOOST_FORCEINLINE BOOST_CHARCONV_FASTFLOAT_CONSTEXPR20 |
| 186 | bool is_truncated(UC const * first, UC const * last) noexcept { |
| 187 | // do 8-bit optimizations, can just compare to 8 literal 0s. |
| 188 | uint64_t val; |
| 189 | while (!cpp20_and_in_constexpr() && std::distance(first, last) >= int_cmp_len<UC>()) { |
| 190 | ::memcpy(dest: &val, src: first, n: sizeof(uint64_t)); |
| 191 | if (val != int_cmp_zeros<UC>()) { |
| 192 | return true; |
| 193 | } |
| 194 | first += int_cmp_len<UC>(); |
| 195 | } |
| 196 | while (first != last) { |
| 197 | if (*first != UC('0')) { |
| 198 | return true; |
| 199 | } |
| 200 | ++first; |
| 201 | } |
| 202 | return false; |
| 203 | } |
| 204 | template <typename UC> |
| 205 | BOOST_FORCEINLINE BOOST_CHARCONV_FASTFLOAT_CONSTEXPR20 |
| 206 | bool is_truncated(span<const UC> s) noexcept { |
| 207 | return is_truncated(s.ptr, s.ptr + s.len()); |
| 208 | } |
| 209 | |
| 210 | BOOST_FORCEINLINE BOOST_CHARCONV_FASTFLOAT_CONSTEXPR20 |
| 211 | void parse_eight_digits(const char16_t*& , limb& , size_t& , size_t& ) noexcept { |
| 212 | // currently unused |
| 213 | } |
| 214 | |
| 215 | BOOST_FORCEINLINE BOOST_CHARCONV_FASTFLOAT_CONSTEXPR20 |
| 216 | void parse_eight_digits(const char32_t*& , limb& , size_t& , size_t& ) noexcept { |
| 217 | // currently unused |
| 218 | } |
| 219 | |
| 220 | BOOST_FORCEINLINE BOOST_CHARCONV_FASTFLOAT_CONSTEXPR20 |
| 221 | void parse_eight_digits(const char*& p, limb& value, size_t& counter, size_t& count) noexcept { |
| 222 | value = value * 100000000 + parse_eight_digits_unrolled(chars: p); |
| 223 | p += 8; |
| 224 | counter += 8; |
| 225 | count += 8; |
| 226 | } |
| 227 | |
| 228 | template <typename UC> |
| 229 | BOOST_FORCEINLINE BOOST_CHARCONV_FASTFLOAT_CONSTEXPR14 |
| 230 | void parse_one_digit(UC const *& p, limb& value, size_t& counter, size_t& count) noexcept { |
| 231 | value = value * 10 + limb(*p - UC('0')); |
| 232 | p++; |
| 233 | counter++; |
| 234 | count++; |
| 235 | } |
| 236 | |
| 237 | BOOST_FORCEINLINE BOOST_CHARCONV_FASTFLOAT_CONSTEXPR20 |
| 238 | void add_native(bigint& big, limb power, limb value) noexcept { |
| 239 | big.mul(y: power); |
| 240 | big.add(y: value); |
| 241 | } |
| 242 | |
| 243 | BOOST_FORCEINLINE BOOST_CHARCONV_FASTFLOAT_CONSTEXPR20 |
| 244 | void round_up_bigint(bigint& big, size_t& count) noexcept { |
| 245 | // need to round-up the digits, but need to avoid rounding |
| 246 | // ....9999 to ...10000, which could cause a false halfway point. |
| 247 | add_native(big, power: 10, value: 1); |
| 248 | count++; |
| 249 | } |
| 250 | |
| 251 | // parse the significant digits into a big integer |
| 252 | template <typename UC> |
| 253 | inline BOOST_CHARCONV_FASTFLOAT_CONSTEXPR20 |
| 254 | void parse_mantissa(bigint& result, parsed_number_string_t<UC>& num, size_t max_digits, size_t& digits) noexcept { |
| 255 | // try to minimize the number of big integer and scalar multiplication. |
| 256 | // therefore, try to parse 8 digits at a time, and multiply by the largest |
| 257 | // scalar value (9 or 19 digits) for each step. |
| 258 | size_t counter = 0; |
| 259 | digits = 0; |
| 260 | limb value = 0; |
| 261 | #ifdef BOOST_CHARCONV_FASTFLOAT_64BIT_LIMB |
| 262 | constexpr size_t step = 19; |
| 263 | #else |
| 264 | constexpr size_t step = 9; |
| 265 | #endif |
| 266 | |
| 267 | // process all integer digits. |
| 268 | UC const * p = num.integer.ptr; |
| 269 | UC const * pend = p + num.integer.len(); |
| 270 | skip_zeros(p, pend); |
| 271 | // process all digits, in increments of step per loop |
| 272 | while (p != pend) { |
| 273 | if (std::is_same<UC,char>::value) { |
| 274 | while ((std::distance(p, pend) >= 8) && (step - counter >= 8) && (max_digits - digits >= 8)) { |
| 275 | parse_eight_digits(p, value, counter, digits); |
| 276 | } |
| 277 | } |
| 278 | while (counter < step && p != pend && digits < max_digits) { |
| 279 | parse_one_digit(p, value, counter, digits); |
| 280 | } |
| 281 | if (digits == max_digits) { |
| 282 | // add the temporary value, then check if we've truncated any digits |
| 283 | add_native(big&: result, power: limb(powers_of_ten_uint64[counter]), value); |
| 284 | bool truncated = is_truncated(p, pend); |
| 285 | if (num.fraction.ptr != nullptr) { |
| 286 | truncated |= is_truncated(num.fraction); |
| 287 | } |
| 288 | if (truncated) { |
| 289 | round_up_bigint(big&: result, count&: digits); |
| 290 | } |
| 291 | return; |
| 292 | } else { |
| 293 | add_native(big&: result, power: limb(powers_of_ten_uint64[counter]), value); |
| 294 | counter = 0; |
| 295 | value = 0; |
| 296 | } |
| 297 | } |
| 298 | |
| 299 | // add our fraction digits, if they're available. |
| 300 | if (num.fraction.ptr != nullptr) { |
| 301 | p = num.fraction.ptr; |
| 302 | pend = p + num.fraction.len(); |
| 303 | if (digits == 0) { |
| 304 | skip_zeros(p, pend); |
| 305 | } |
| 306 | // process all digits, in increments of step per loop |
| 307 | while (p != pend) { |
| 308 | if (std::is_same<UC,char>::value) { |
| 309 | while ((std::distance(p, pend) >= 8) && (step - counter >= 8) && (max_digits - digits >= 8)) { |
| 310 | parse_eight_digits(p, value, counter, digits); |
| 311 | } |
| 312 | } |
| 313 | while (counter < step && p != pend && digits < max_digits) { |
| 314 | parse_one_digit(p, value, counter, digits); |
| 315 | } |
| 316 | if (digits == max_digits) { |
| 317 | // add the temporary value, then check if we've truncated any digits |
| 318 | add_native(big&: result, power: limb(powers_of_ten_uint64[counter]), value); |
| 319 | bool truncated = is_truncated(p, pend); |
| 320 | if (truncated) { |
| 321 | round_up_bigint(big&: result, count&: digits); |
| 322 | } |
| 323 | return; |
| 324 | } else { |
| 325 | add_native(big&: result, power: limb(powers_of_ten_uint64[counter]), value); |
| 326 | counter = 0; |
| 327 | value = 0; |
| 328 | } |
| 329 | } |
| 330 | } |
| 331 | |
| 332 | if (counter != 0) { |
| 333 | add_native(big&: result, power: limb(powers_of_ten_uint64[counter]), value); |
| 334 | } |
| 335 | } |
| 336 | |
| 337 | template <typename T> |
| 338 | inline BOOST_CHARCONV_FASTFLOAT_CONSTEXPR20 |
| 339 | adjusted_mantissa positive_digit_comp(bigint& bigmant, int32_t exponent) noexcept { |
| 340 | BOOST_CHARCONV_FASTFLOAT_ASSERT(bigmant.pow10(uint32_t(exponent))); |
| 341 | adjusted_mantissa answer; |
| 342 | bool truncated; |
| 343 | answer.mantissa = bigmant.hi64(truncated); |
| 344 | int bias = binary_format<T>::mantissa_explicit_bits() - binary_format<T>::minimum_exponent(); |
| 345 | answer.power2 = bigmant.bit_length() - 64 + bias; |
| 346 | |
| 347 | round<T>(answer, [truncated](adjusted_mantissa& a, int32_t shift) { |
| 348 | round_nearest_tie_even(a, shift, [truncated](bool is_odd, bool is_halfway, bool is_above) -> bool { |
| 349 | return is_above || (is_halfway && truncated) || (is_odd && is_halfway); |
| 350 | }); |
| 351 | }); |
| 352 | |
| 353 | return answer; |
| 354 | } |
| 355 | |
| 356 | // the scaling here is quite simple: we have, for the real digits `m * 10^e`, |
| 357 | // and for the theoretical digits `n * 2^f`. Since `e` is always negative, |
| 358 | // to scale them identically, we do `n * 2^f * 5^-f`, so we now have `m * 2^e`. |
| 359 | // we then need to scale by `2^(f- e)`, and then the two significant digits |
| 360 | // are of the same magnitude. |
| 361 | template <typename T> |
| 362 | inline BOOST_CHARCONV_FASTFLOAT_CONSTEXPR20 |
| 363 | adjusted_mantissa negative_digit_comp(bigint& bigmant, adjusted_mantissa am, int32_t exponent) noexcept { |
| 364 | bigint& real_digits = bigmant; |
| 365 | int32_t real_exp = exponent; |
| 366 | |
| 367 | // get the value of `b`, rounded down, and get a bigint representation of b+h |
| 368 | adjusted_mantissa am_b = am; |
| 369 | // gcc7 buf: use a lambda to remove the noexcept qualifier bug with -Wnoexcept-type. |
| 370 | round<T>(am_b, [](adjusted_mantissa&a, int32_t shift) { round_down(am&: a, shift); }); |
| 371 | T b; |
| 372 | to_float(false, am_b, b); |
| 373 | adjusted_mantissa theor = to_extended_halfway(b); |
| 374 | bigint theor_digits(theor.mantissa); |
| 375 | int32_t theor_exp = theor.power2; |
| 376 | |
| 377 | // scale real digits and theor digits to be same power. |
| 378 | int32_t pow2_exp = theor_exp - real_exp; |
| 379 | uint32_t pow5_exp = uint32_t(-real_exp); |
| 380 | if (pow5_exp != 0) { |
| 381 | BOOST_CHARCONV_FASTFLOAT_ASSERT(theor_digits.pow5(pow5_exp)); |
| 382 | } |
| 383 | if (pow2_exp > 0) { |
| 384 | BOOST_CHARCONV_FASTFLOAT_ASSERT(theor_digits.pow2(uint32_t(pow2_exp))); |
| 385 | } else if (pow2_exp < 0) { |
| 386 | BOOST_CHARCONV_FASTFLOAT_ASSERT(real_digits.pow2(uint32_t(-pow2_exp))); |
| 387 | } |
| 388 | |
| 389 | // compare digits, and use it to director rounding |
| 390 | int ord = real_digits.compare(other: theor_digits); |
| 391 | adjusted_mantissa answer = am; |
| 392 | round<T>(answer, [ord](adjusted_mantissa& a, int32_t shift) { |
| 393 | round_nearest_tie_even(a, shift, [ord](bool is_odd, bool, bool) -> bool { |
| 394 | if (ord > 0) { |
| 395 | return true; |
| 396 | } else if (ord < 0) { |
| 397 | return false; |
| 398 | } else { |
| 399 | return is_odd; |
| 400 | } |
| 401 | }); |
| 402 | }); |
| 403 | |
| 404 | return answer; |
| 405 | } |
| 406 | |
| 407 | // parse the significant digits as a big integer to unambiguously round |
| 408 | // the significant digits. here, we are trying to determine how to round |
| 409 | // an extended float representation close to `b+h`, halfway between `b` |
| 410 | // (the float rounded-down) and `b+u`, the next positive float. this |
| 411 | // algorithm is always correct, and uses one of two approaches. when |
| 412 | // the exponent is positive relative to the significant digits (such as |
| 413 | // 1234), we create a big-integer representation, get the high 64-bits, |
| 414 | // determine if any lower bits are truncated, and use that to direct |
| 415 | // rounding. in case of a negative exponent relative to the significant |
| 416 | // digits (such as 1.2345), we create a theoretical representation of |
| 417 | // `b` as a big-integer type, scaled to the same binary exponent as |
| 418 | // the actual digits. we then compare the big integer representations |
| 419 | // of both, and use that to direct rounding. |
| 420 | template <typename T, typename UC> |
| 421 | inline BOOST_CHARCONV_FASTFLOAT_CONSTEXPR20 |
| 422 | adjusted_mantissa digit_comp(parsed_number_string_t<UC>& num, adjusted_mantissa am) noexcept { |
| 423 | // remove the invalid exponent bias |
| 424 | am.power2 -= invalid_am_bias; |
| 425 | |
| 426 | int32_t sci_exp = scientific_exponent(num); |
| 427 | size_t max_digits = binary_format<T>::max_digits(); |
| 428 | size_t digits = 0; |
| 429 | bigint bigmant; |
| 430 | parse_mantissa(bigmant, num, max_digits, digits); |
| 431 | // can't underflow, since digits is at most max_digits. |
| 432 | int32_t exponent = sci_exp + 1 - int32_t(digits); |
| 433 | if (exponent >= 0) { |
| 434 | return positive_digit_comp<T>(bigmant, exponent); |
| 435 | } else { |
| 436 | return negative_digit_comp<T>(bigmant, am, exponent); |
| 437 | } |
| 438 | } |
| 439 | |
| 440 | }}}} // namespace fast_float |
| 441 | |
| 442 | #endif |
| 443 | |