VeniVidiVici
/
GDAL2
1
0
Fork 0
Code Issues Pull Requests Packages Projects Releases Wiki Activity

third_party/fast_float: resync with upstream

This commit is contained in:
Even Rouault 2024-11-05 23:01:19 +01:00
parent c578e9d910
commit 71078a1d16
No known key found for this signature in database
GPG Key ID: 33EBBFC47B3DD87D
10 changed files with 1856 additions and 1403 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

2
third_party/fast_float/PROVENANCE.TXT vendored
View File

@ -1,4 +1,4 @@
https://github.com/fastfloat/fast_float
Retrieved at commit https://github.com/fastfloat/fast_float/commit/a5ea2059295260922aa300d676a43a76b5e19a35
Retrieved at commit https://github.com/fastfloat/fast_float/commit/9058831e6884e95358bcad29139a8b9d6cf0b534
Using the MIT license choice.

403
third_party/fast_float/ascii_number.h vendored
View File

@ -5,6 +5,7 @@
#include <cstdint>
#include <cstring>
#include <iterator>
#include <limits>
#include <type_traits>
#include "float_common.h"
@ -19,8 +20,7 @@
namespace fast_float {
template <typename UC>
fastfloat_really_inline constexpr bool has_simd_opt() {
template <typename UC> fastfloat_really_inline constexpr bool has_simd_opt() {
#ifdef FASTFLOAT_HAS_SIMD
return std::is_same<UC, char16_t>::value;
#else
@ -36,24 +36,20 @@ fastfloat_really_inline constexpr bool is_integer(UC c) noexcept {
}
fastfloat_really_inline constexpr uint64_t byteswap(uint64_t val) {
return (val & 0xFF00000000000000) >> 56
| (val & 0x00FF000000000000) >> 40
| (val & 0x0000FF0000000000) >> 24
| (val & 0x000000FF00000000) >> 8
| (val & 0x00000000FF000000) << 8
| (val & 0x0000000000FF0000) << 24
| (val & 0x000000000000FF00) << 40
| (val & 0x00000000000000FF) << 56;
return (val & 0xFF00000000000000) >> 56 | (val & 0x00FF000000000000) >> 40 |
(val & 0x0000FF0000000000) >> 24 | (val & 0x000000FF00000000) >> 8 |
(val & 0x00000000FF000000) << 8 | (val & 0x0000000000FF0000) << 24 |
(val & 0x000000000000FF00) << 40 | (val & 0x00000000000000FF) << 56;
}
// Read 8 UC into a u64. Truncates UC if not char.
template <typename UC>
fastfloat_really_inline FASTFLOAT_CONSTEXPR20
uint64_t read8_to_u64(const UC *chars) {
fastfloat_really_inline FASTFLOAT_CONSTEXPR20 uint64_t
read8_to_u64(const UC *chars) {
if (cpp20_and_in_constexpr() || !std::is_same<UC, char>::value) {
uint64_t val = 0;
for(int i = 0; i < 8; ++i) {
val |= uint64_t(uint8_t(*chars)) << (i*8);
for (int i = 0; i < 8; ++i) {
val |= uint64_t(uint8_t(*chars)) << (i * 8);
++chars;
}
return val;
@ -69,44 +65,41 @@ uint64_t read8_to_u64(const UC *chars) {
#ifdef FASTFLOAT_SSE2
fastfloat_really_inline
uint64_t simd_read8_to_u64(const __m128i data) {
FASTFLOAT_SIMD_DISABLE_WARNINGS
fastfloat_really_inline uint64_t simd_read8_to_u64(const __m128i data) {
FASTFLOAT_SIMD_DISABLE_WARNINGS
const __m128i packed = _mm_packus_epi16(data, data);
#ifdef FASTFLOAT_64BIT
return uint64_t(_mm_cvtsi128_si64(packed));
#else
uint64_t value;
// Visual Studio + older versions of GCC don't support _mm_storeu_si64
_mm_storel_epi64(reinterpret_cast<__m128i*>(&value), packed);
_mm_storel_epi64(reinterpret_cast<__m128i *>(&value), packed);
return value;
#endif
FASTFLOAT_SIMD_RESTORE_WARNINGS
FASTFLOAT_SIMD_RESTORE_WARNINGS
}
fastfloat_really_inline
uint64_t simd_read8_to_u64(const char16_t* chars) {
FASTFLOAT_SIMD_DISABLE_WARNINGS
return simd_read8_to_u64(_mm_loadu_si128(reinterpret_cast<const __m128i*>(chars)));
FASTFLOAT_SIMD_RESTORE_WARNINGS
fastfloat_really_inline uint64_t simd_read8_to_u64(const char16_t *chars) {
FASTFLOAT_SIMD_DISABLE_WARNINGS
return simd_read8_to_u64(
_mm_loadu_si128(reinterpret_cast<const __m128i *>(chars)));
FASTFLOAT_SIMD_RESTORE_WARNINGS
}
#elif defined(FASTFLOAT_NEON)
fastfloat_really_inline
uint64_t simd_read8_to_u64(const uint16x8_t data) {
FASTFLOAT_SIMD_DISABLE_WARNINGS
fastfloat_really_inline uint64_t simd_read8_to_u64(const uint16x8_t data) {
FASTFLOAT_SIMD_DISABLE_WARNINGS
uint8x8_t utf8_packed = vmovn_u16(data);
return vget_lane_u64(vreinterpret_u64_u8(utf8_packed), 0);
FASTFLOAT_SIMD_RESTORE_WARNINGS
FASTFLOAT_SIMD_RESTORE_WARNINGS
}
fastfloat_really_inline
uint64_t simd_read8_to_u64(const char16_t* chars) {
FASTFLOAT_SIMD_DISABLE_WARNINGS
return simd_read8_to_u64(vld1q_u16(reinterpret_cast<const uint16_t*>(chars)));
FASTFLOAT_SIMD_RESTORE_WARNINGS
fastfloat_really_inline uint64_t simd_read8_to_u64(const char16_t *chars) {
FASTFLOAT_SIMD_DISABLE_WARNINGS
return simd_read8_to_u64(
vld1q_u16(reinterpret_cast<const uint16_t *>(chars)));
FASTFLOAT_SIMD_RESTORE_WARNINGS
}
#endif // FASTFLOAT_SSE2
@ -115,34 +108,16 @@ FASTFLOAT_SIMD_RESTORE_WARNINGS
#if defined(_MSC_VER) && _MSC_VER <= 1900
template <typename UC>
#else
template <typename UC, FASTFLOAT_ENABLE_IF(!has_simd_opt<UC>())>
template <typename UC, FASTFLOAT_ENABLE_IF(!has_simd_opt<UC>()) = 0>
#endif
// dummy for compile
uint64_t simd_read8_to_u64(UC const*) {
uint64_t simd_read8_to_u64(UC const *) {
return 0;
}
fastfloat_really_inline FASTFLOAT_CONSTEXPR20
void write_u64(uint8_t *chars, uint64_t val) {
if (cpp20_and_in_constexpr()) {
for(int i = 0; i < 8; ++i) {
*chars = uint8_t(val);
val >>= 8;
++chars;
}
return;
}
#if FASTFLOAT_IS_BIG_ENDIAN == 1
// Need to read as-if the number was in little-endian order.
val = byteswap(val);
#endif
::memcpy(chars, &val, sizeof(uint64_t));
}
// credit @aqrit
fastfloat_really_inline FASTFLOAT_CONSTEXPR14
uint32_t parse_eight_digits_unrolled(uint64_t val) {
fastfloat_really_inline FASTFLOAT_CONSTEXPR14 uint32_t
parse_eight_digits_unrolled(uint64_t val) {
const uint64_t mask = 0x000000FF000000FF;
const uint64_t mul1 = 0x000F424000000064; // 100 + (1000000ULL << 32)
const uint64_t mul2 = 0x0000271000000001; // 1 + (10000ULL << 32)
@ -152,38 +127,38 @@ uint32_t parse_eight_digits_unrolled(uint64_t val) {
return uint32_t(val);
}
// Call this if chars are definitely 8 digits.
template <typename UC>
fastfloat_really_inline FASTFLOAT_CONSTEXPR20
uint32_t parse_eight_digits_unrolled(UC const * chars) noexcept {
fastfloat_really_inline FASTFLOAT_CONSTEXPR20 uint32_t
parse_eight_digits_unrolled(UC const *chars) noexcept {
if (cpp20_and_in_constexpr() || !has_simd_opt<UC>()) {
return parse_eight_digits_unrolled(read8_to_u64(chars)); // truncation okay
}
return parse_eight_digits_unrolled(simd_read8_to_u64(chars));
}
// credit @aqrit
fastfloat_really_inline constexpr bool is_made_of_eight_digits_fast(uint64_t val) noexcept {
fastfloat_really_inline constexpr bool
is_made_of_eight_digits_fast(uint64_t val) noexcept {
return !((((val + 0x4646464646464646) | (val - 0x3030303030303030)) &
0x8080808080808080));
0x8080808080808080));
}
#ifdef FASTFLOAT_HAS_SIMD
// Call this if chars might not be 8 digits.
// Using this style (instead of is_made_of_eight_digits_fast() then parse_eight_digits_unrolled())
// ensures we don't load SIMD registers twice.
fastfloat_really_inline FASTFLOAT_CONSTEXPR20
bool simd_parse_if_eight_digits_unrolled(const char16_t* chars, uint64_t& i) noexcept {
// Using this style (instead of is_made_of_eight_digits_fast() then
// parse_eight_digits_unrolled()) ensures we don't load SIMD registers twice.
fastfloat_really_inline FASTFLOAT_CONSTEXPR20 bool
simd_parse_if_eight_digits_unrolled(const char16_t *chars,
uint64_t &i) noexcept {
if (cpp20_and_in_constexpr()) {
return false;
}
}
#ifdef FASTFLOAT_SSE2
FASTFLOAT_SIMD_DISABLE_WARNINGS
const __m128i data = _mm_loadu_si128(reinterpret_cast<const __m128i*>(chars));
FASTFLOAT_SIMD_DISABLE_WARNINGS
const __m128i data =
_mm_loadu_si128(reinterpret_cast<const __m128i *>(chars));
// (x - '0') <= 9
// http://0x80.pl/articles/simd-parsing-int-sequences.html
@ -193,13 +168,13 @@ FASTFLOAT_SIMD_DISABLE_WARNINGS
if (_mm_movemask_epi8(t1) == 0) {
i = i * 100000000 + parse_eight_digits_unrolled(simd_read8_to_u64(data));
return true;
}
else return false;
FASTFLOAT_SIMD_RESTORE_WARNINGS
} else
return false;
FASTFLOAT_SIMD_RESTORE_WARNINGS
#elif defined(FASTFLOAT_NEON)
FASTFLOAT_SIMD_DISABLE_WARNINGS
const uint16x8_t data = vld1q_u16(reinterpret_cast<const uint16_t*>(chars));
FASTFLOAT_SIMD_DISABLE_WARNINGS
const uint16x8_t data = vld1q_u16(reinterpret_cast<const uint16_t *>(chars));
// (x - '0') <= 9
// http://0x80.pl/articles/simd-parsing-int-sequences.html
const uint16x8_t t0 = vsubq_u16(data, vmovq_n_u16('0'));
@ -208,11 +183,12 @@ FASTFLOAT_SIMD_DISABLE_WARNINGS
if (vminvq_u16(mask) == 0xFFFF) {
i = i * 100000000 + parse_eight_digits_unrolled(simd_read8_to_u64(data));
return true;
}
else return false;
FASTFLOAT_SIMD_RESTORE_WARNINGS
} else
return false;
FASTFLOAT_SIMD_RESTORE_WARNINGS
#else
(void)chars; (void)i;
(void)chars;
(void)i;
return false;
#endif // FASTFLOAT_SSE2
}
@ -223,55 +199,90 @@ FASTFLOAT_SIMD_RESTORE_WARNINGS
#if defined(_MSC_VER) && _MSC_VER <= 1900
template <typename UC>
#else
template <typename UC, FASTFLOAT_ENABLE_IF(!has_simd_opt<UC>())>
template <typename UC, FASTFLOAT_ENABLE_IF(!has_simd_opt<UC>()) = 0>
#endif
// dummy for compile
bool simd_parse_if_eight_digits_unrolled(UC const*, uint64_t&) {
bool simd_parse_if_eight_digits_unrolled(UC const *, uint64_t &) {
return 0;
}
template <typename UC, FASTFLOAT_ENABLE_IF(!std::is_same<UC, char>::value)>
fastfloat_really_inline FASTFLOAT_CONSTEXPR20
void loop_parse_if_eight_digits(const UC*& p, const UC* const pend, uint64_t& i) {
template <typename UC, FASTFLOAT_ENABLE_IF(!std::is_same<UC, char>::value) = 0>
fastfloat_really_inline FASTFLOAT_CONSTEXPR20 void
loop_parse_if_eight_digits(const UC *&p, const UC *const pend, uint64_t &i) {
if (!has_simd_opt<UC>()) {
return;
}
while ((std::distance(p, pend) >= 8) && simd_parse_if_eight_digits_unrolled(p, i)) { // in rare cases, this will overflow, but that's ok
while ((std::distance(p, pend) >= 8) &&
simd_parse_if_eight_digits_unrolled(
p, i)) { // in rare cases, this will overflow, but that's ok
p += 8;
}
}
fastfloat_really_inline FASTFLOAT_CONSTEXPR20
void loop_parse_if_eight_digits(const char*& p, const char* const pend, uint64_t& i) {
fastfloat_really_inline FASTFLOAT_CONSTEXPR20 void
loop_parse_if_eight_digits(const char *&p, const char *const pend,
uint64_t &i) {
// optimizes better than parse_if_eight_digits_unrolled() for UC = char.
while ((std::distance(p, pend) >= 8) && is_made_of_eight_digits_fast(read8_to_u64(p))) {
i = i * 100000000 + parse_eight_digits_unrolled(read8_to_u64(p)); // in rare cases, this will overflow, but that's ok
while ((std::distance(p, pend) >= 8) &&
is_made_of_eight_digits_fast(read8_to_u64(p))) {
i = i * 100000000 +
parse_eight_digits_unrolled(read8_to_u64(
p)); // in rare cases, this will overflow, but that's ok
p += 8;
}
}
template <typename UC>
struct parsed_number_string_t {
enum class parse_error {
no_error,
// [JSON-only] The minus sign must be followed by an integer.
missing_integer_after_sign,
// A sign must be followed by an integer or dot.
missing_integer_or_dot_after_sign,
// [JSON-only] The integer part must not have leading zeros.
leading_zeros_in_integer_part,
// [JSON-only] The integer part must have at least one digit.
no_digits_in_integer_part,
// [JSON-only] If there is a decimal point, there must be digits in the
// fractional part.
no_digits_in_fractional_part,
// The mantissa must have at least one digit.
no_digits_in_mantissa,
// Scientific notation requires an exponential part.
missing_exponential_part,
};
template <typename UC> struct parsed_number_string_t {
int64_t exponent{0};
uint64_t mantissa{0};
UC const * lastmatch{nullptr};
UC const *lastmatch{nullptr};
bool negative{false};
bool valid{false};
bool too_many_digits{false};
// contains the range of the significant digits
span<const UC> integer{}; // non-nullable
span<const UC> fraction{}; // nullable
parse_error error{parse_error::no_error};
};
using byte_span = span<const char>;
using parsed_number_string = parsed_number_string_t<char>;
template <typename UC>
fastfloat_really_inline FASTFLOAT_CONSTEXPR20 parsed_number_string_t<UC>
report_parse_error(UC const *p, parse_error error) {
parsed_number_string_t<UC> answer;
answer.valid = false;
answer.lastmatch = p;
answer.error = error;
return answer;
}
// Assuming that you use no more than 19 digits, this will
// parse an ASCII string.
template <typename UC>
fastfloat_really_inline FASTFLOAT_CONSTEXPR20
parsed_number_string_t<UC> parse_number_string(UC const *p, UC const * pend, parse_options_t<UC> options) noexcept {
fastfloat_really_inline FASTFLOAT_CONSTEXPR20 parsed_number_string_t<UC>
parse_number_string(UC const *p, UC const *pend,
parse_options_t<UC> options) noexcept {
chars_format const fmt = options.format;
UC const decimal_point = options.decimal_point;
@ -286,19 +297,24 @@ parsed_number_string_t<UC> parse_number_string(UC const *p, UC const * pend, par
#endif
++p;
if (p == pend) {
return answer;
return report_parse_error<UC>(
p, parse_error::missing_integer_or_dot_after_sign);
}
if (fmt & FASTFLOAT_JSONFMT) {
if (!is_integer(*p)) { // a sign must be followed by an integer
return answer;
}
return report_parse_error<UC>(p,
parse_error::missing_integer_after_sign);
}
} else {
if (!is_integer(*p) && (*p != decimal_point)) { // a sign must be followed by an integer or the dot
return answer;
if (!is_integer(*p) &&
(*p !=
decimal_point)) { // a sign must be followed by an integer or the dot
return report_parse_error<UC>(
p, parse_error::missing_integer_or_dot_after_sign);
}
}
}
UC const * const start_digits = p;
UC const *const start_digits = p;
uint64_t i = 0; // an unsigned int avoids signed overflows (which are bad)
@ -306,16 +322,21 @@ parsed_number_string_t<UC> parse_number_string(UC const *p, UC const * pend, par
// a multiplication by 10 is cheaper than an arbitrary integer
// multiplication
i = 10 * i +
uint64_t(*p - UC('0')); // might overflow, we will handle the overflow later
uint64_t(*p -
UC('0')); // might overflow, we will handle the overflow later
++p;
}
UC const * const end_of_integer_part = p;
UC const *const end_of_integer_part = p;
int64_t digit_count = int64_t(end_of_integer_part - start_digits);
answer.integer = span<const UC>(start_digits, size_t(digit_count));
if (fmt & FASTFLOAT_JSONFMT) {
// at least 1 digit in integer part, without leading zeros
if (digit_count == 0 || (start_digits[0] == UC('0') && digit_count > 1)) {
return answer;
if (digit_count == 0) {
return report_parse_error<UC>(p, parse_error::no_digits_in_integer_part);
}
if ((start_digits[0] == UC('0') && digit_count > 1)) {
return report_parse_error<UC>(start_digits,
parse_error::leading_zeros_in_integer_part);
}
}
@ -323,7 +344,7 @@ parsed_number_string_t<UC> parse_number_string(UC const *p, UC const * pend, par
const bool has_decimal_point = (p != pend) && (*p == decimal_point);
if (has_decimal_point) {
++p;
UC const * before = p;
UC const *before = p;
// can occur at most twice without overflowing, but let it occur more, since
// for integers with many digits, digit parsing is the primary bottleneck.
loop_parse_if_eight_digits(p, pend, i);
@ -340,35 +361,39 @@ parsed_number_string_t<UC> parse_number_string(UC const *p, UC const * pend, par
if (fmt & FASTFLOAT_JSONFMT) {
// at least 1 digit in fractional part
if (has_decimal_point && exponent == 0) {
return answer;
return report_parse_error<UC>(p,
parse_error::no_digits_in_fractional_part);
}
}
else if (digit_count == 0) { // we must have encountered at least one integer!
return answer;
} else if (digit_count ==
0) { // we must have encountered at least one integer!
return report_parse_error<UC>(p, parse_error::no_digits_in_mantissa);
}
int64_t exp_number = 0; // explicit exponential part
if ( ((fmt & chars_format::scientific) &&
(p != pend) &&
((UC('e') == *p) || (UC('E') == *p)))
||
((fmt & FASTFLOAT_FORTRANFMT) &&
(p != pend) &&
((UC('+') == *p) || (UC('-') == *p) || (UC('d') == *p) || (UC('D') == *p)))) {
UC const * location_of_e = p;
if ((UC('e') == *p) || (UC('E') == *p) || (UC('d') == *p) || (UC('D') == *p)) {
int64_t exp_number = 0; // explicit exponential part
if (((fmt & chars_format::scientific) && (p != pend) &&
((UC('e') == *p) || (UC('E') == *p))) ||
((fmt & FASTFLOAT_FORTRANFMT) && (p != pend) &&
((UC('+') == *p) || (UC('-') == *p) || (UC('d') == *p) ||
(UC('D') == *p)))) {
UC const *location_of_e = p;
if ((UC('e') == *p) || (UC('E') == *p) || (UC('d') == *p) ||
(UC('D') == *p)) {
++p;
}
bool neg_exp = false;
if ((p != pend) && (UC('-') == *p)) {
neg_exp = true;
++p;
} else if ((p != pend) && (UC('+') == *p)) { // '+' on exponent is allowed by C++17 20.19.3.(7.1)
} else if ((p != pend) &&
(UC('+') ==
*p)) { // '+' on exponent is allowed by C++17 20.19.3.(7.1)
++p;
}
if ((p == pend) || !is_integer(*p)) {
if(!(fmt & chars_format::fixed)) {
// We are in error.
return answer;
if (!(fmt & chars_format::fixed)) {
// The exponential part is invalid for scientific notation, so it must
// be a trailing token for fixed notation. However, fixed notation is
// disabled, so report a scientific notation error.
return report_parse_error<UC>(p, parse_error::missing_exponential_part);
}
// Otherwise, we will be ignoring the 'e'.
p = location_of_e;
@ -380,12 +405,16 @@ parsed_number_string_t<UC> parse_number_string(UC const *p, UC const * pend, par
}
++p;
}
if(neg_exp) { exp_number = - exp_number; }
if (neg_exp) {
exp_number = -exp_number;
}
exponent += exp_number;
}
} else {
// If it scientific and not fixed, we have to bail out.
if((fmt & chars_format::scientific) && !(fmt & chars_format::fixed)) { return answer; }
if ((fmt & chars_format::scientific) && !(fmt & chars_format::fixed)) {
return report_parse_error<UC>(p, parse_error::missing_exponential_part);
}
}
answer.lastmatch = p;
answer.valid = true;
@ -400,9 +429,11 @@ parsed_number_string_t<UC> parse_number_string(UC const *p, UC const * pend, par
// We have to handle the case where we have 0.0000somenumber.
// We need to be mindful of the case where we only have zeroes...
// E.g., 0.000000000...000.
UC const * start = start_digits;
UC const *start = start_digits;
while ((start != pend) && (*start == UC('0') || *start == decimal_point)) {
if(*start == UC('0')) { digit_count --; }
if (*start == UC('0')) {
digit_count--;
}
start++;
}
@ -413,18 +444,17 @@ parsed_number_string_t<UC> parse_number_string(UC const *p, UC const * pend, par
// pre-tokenized spans from above.
i = 0;
p = answer.integer.ptr;
UC const* int_end = p + answer.integer.len();
const uint64_t minimal_nineteen_digit_integer{ 1000000000000000000 };
UC const *int_end = p + answer.integer.len();
const uint64_t minimal_nineteen_digit_integer{1000000000000000000};
while ((i < minimal_nineteen_digit_integer) && (p != int_end)) {
i = i * 10 + uint64_t(*p - UC('0'));
++p;
}
if (i >= minimal_nineteen_digit_integer) { // We have a big integers
exponent = end_of_integer_part - p + exp_number;
}
else { // We have a value with a fractional component.
} else { // We have a value with a fractional component.
p = answer.fraction.ptr;
UC const* frac_end = p + answer.fraction.len();
UC const *frac_end = p + answer.fraction.len();
while ((i < minimal_nineteen_digit_integer) && (p != frac_end)) {
i = i * 10 + uint64_t(*p - UC('0'));
++p;
@ -439,6 +469,111 @@ parsed_number_string_t<UC> parse_number_string(UC const *p, UC const * pend, par
return answer;
}
template <typename T, typename UC>
fastfloat_really_inline FASTFLOAT_CONSTEXPR20 from_chars_result_t<UC>
parse_int_string(UC const *p, UC const *pend, T &value, int base) {
from_chars_result_t<UC> answer;
UC const *const first = p;
bool negative = (*p == UC('-'));
if (!std::is_signed<T>::value && negative) {
answer.ec = std::errc::invalid_argument;
answer.ptr = first;
return answer;
}
#ifdef FASTFLOAT_ALLOWS_LEADING_PLUS // disabled by default
if ((*p == UC('-')) || (*p == UC('+'))) {
#else
if (*p == UC('-')) {
#endif
++p;
}
UC const *const start_num = p;
while (p != pend && *p == UC('0')) {
++p;
}
const bool has_leading_zeros = p > start_num;
UC const *const start_digits = p;
uint64_t i = 0;
if (base == 10) {
loop_parse_if_eight_digits(p, pend, i); // use SIMD if possible
}
while (p != pend) {
uint8_t digit = ch_to_digit(*p);
if (digit >= base) {
break;
}
i = uint64_t(base) * i + digit; // might overflow, check this later
p++;
}
size_t digit_count = size_t(p - start_digits);
if (digit_count == 0) {
if (has_leading_zeros) {
value = 0;
answer.ec = std::errc();
answer.ptr = p;
} else {
answer.ec = std::errc::invalid_argument;
answer.ptr = first;
}
return answer;
}
answer.ptr = p;
// check u64 overflow
size_t max_digits = max_digits_u64(base);
if (digit_count > max_digits) {
answer.ec = std::errc::result_out_of_range;
return answer;
}
// this check can be eliminated for all other types, but they will all require
// a max_digits(base) equivalent
if (digit_count == max_digits && i < min_safe_u64(base)) {
answer.ec = std::errc::result_out_of_range;
return answer;
}
// check other types overflow
if (!std::is_same<T, uint64_t>::value) {
if (i > uint64_t(std::numeric_limits<T>::max()) + uint64_t(negative)) {
answer.ec = std::errc::result_out_of_range;
return answer;
}
}
if (negative) {
#ifdef FASTFLOAT_VISUAL_STUDIO
#pragma warning(push)
#pragma warning(disable : 4146)
#endif
// this weird workaround is required because:
// - converting unsigned to signed when its value is greater than signed max
// is UB pre-C++23.
// - reinterpret_casting (~i + 1) would work, but it is not constexpr
// this is always optimized into a neg instruction (note: T is an integer
// type)
value = T(-std::numeric_limits<T>::max() -
T(i - uint64_t(std::numeric_limits<T>::max())));
#ifdef FASTFLOAT_VISUAL_STUDIO
#pragma warning(pop)
#endif
} else {
value = T(i);
}
answer.ec = std::errc();
return answer;
}
} // namespace fast_float
#endif

201
third_party/fast_float/bigint.h vendored
View File

@ -37,8 +37,7 @@ constexpr size_t bigint_limbs = bigint_bits / limb_bits;
// vector-like type that is allocated on the stack. the entire
// buffer is pre-allocated, and only the length changes.
template <uint16_t size>
struct stackvec {
template <uint16_t size> struct stackvec {
limb data[size];
// we never need more than 150 limbs
uint16_t length{0};
@ -54,16 +53,16 @@ struct stackvec {
FASTFLOAT_ASSERT(try_extend(s));
}
FASTFLOAT_CONSTEXPR14 limb& operator[](size_t index) noexcept {
FASTFLOAT_CONSTEXPR14 limb &operator[](size_t index) noexcept {
FASTFLOAT_DEBUG_ASSERT(index < length);
return data[index];
}
FASTFLOAT_CONSTEXPR14 const limb& operator[](size_t index) const noexcept {
FASTFLOAT_CONSTEXPR14 const limb &operator[](size_t index) const noexcept {
FASTFLOAT_DEBUG_ASSERT(index < length);
return data[index];
}
// index from the end of the container
FASTFLOAT_CONSTEXPR14 const limb& rindex(size_t index) const noexcept {
FASTFLOAT_CONSTEXPR14 const limb &rindex(size_t index) const noexcept {
FASTFLOAT_DEBUG_ASSERT(index < length);
size_t rindex = length - index - 1;
return data[rindex];
@ -73,15 +72,9 @@ struct stackvec {
FASTFLOAT_CONSTEXPR14 void set_len(size_t len) noexcept {
length = uint16_t(len);
}
constexpr size_t len() const noexcept {
return length;
}
constexpr bool is_empty() const noexcept {
return length == 0;
}
constexpr size_t capacity() const noexcept {
return size;
}
constexpr size_t len() const noexcept { return length; }
constexpr bool is_empty() const noexcept { return length == 0; }
constexpr size_t capacity() const noexcept { return size; }
// append item to vector, without bounds checking
FASTFLOAT_CONSTEXPR14 void push_unchecked(limb value) noexcept {
data[length] = value;
@ -98,7 +91,7 @@ struct stackvec {
}
// add items to the vector, from a span, without bounds checking
FASTFLOAT_CONSTEXPR20 void extend_unchecked(limb_span s) noexcept {
limb* ptr = data + length;
limb *ptr = data + length;
std::copy_n(s.ptr, s.len(), ptr);
set_len(len() + s.len());
}
@ -118,8 +111,8 @@ struct stackvec {
void resize_unchecked(size_t new_len, limb value) noexcept {
if (new_len > len()) {
size_t count = new_len - len();
limb* first = data + len();
limb* last = first + count;
limb *first = data + len();
limb *last = first + count;
::std::fill(first, last, value);
set_len(new_len);
} else {
@ -155,21 +148,21 @@ struct stackvec {
}
};
fastfloat_really_inline FASTFLOAT_CONSTEXPR14
uint64_t empty_hi64(bool& truncated) noexcept {
fastfloat_really_inline FASTFLOAT_CONSTEXPR14 uint64_t
empty_hi64(bool &truncated) noexcept {
truncated = false;
return 0;
}
fastfloat_really_inline FASTFLOAT_CONSTEXPR20
uint64_t uint64_hi64(uint64_t r0, bool& truncated) noexcept {
fastfloat_really_inline FASTFLOAT_CONSTEXPR20 uint64_t
uint64_hi64(uint64_t r0, bool &truncated) noexcept {
truncated = false;
int shl = leading_zeroes(r0);
return r0 << shl;
}
fastfloat_really_inline FASTFLOAT_CONSTEXPR20
uint64_t uint64_hi64(uint64_t r0, uint64_t r1, bool& truncated) noexcept {
fastfloat_really_inline FASTFLOAT_CONSTEXPR20 uint64_t
uint64_hi64(uint64_t r0, uint64_t r1, bool &truncated) noexcept {
int shl = leading_zeroes(r0);
if (shl == 0) {
truncated = r1 != 0;
@ -181,20 +174,20 @@ uint64_t uint64_hi64(uint64_t r0, uint64_t r1, bool& truncated) noexcept {
}
}
fastfloat_really_inline FASTFLOAT_CONSTEXPR20
uint64_t uint32_hi64(uint32_t r0, bool& truncated) noexcept {
fastfloat_really_inline FASTFLOAT_CONSTEXPR20 uint64_t
uint32_hi64(uint32_t r0, bool &truncated) noexcept {
return uint64_hi64(r0, truncated);
}
fastfloat_really_inline FASTFLOAT_CONSTEXPR20
uint64_t uint32_hi64(uint32_t r0, uint32_t r1, bool& truncated) noexcept {
fastfloat_really_inline FASTFLOAT_CONSTEXPR20 uint64_t
uint32_hi64(uint32_t r0, uint32_t r1, bool &truncated) noexcept {
uint64_t x0 = r0;
uint64_t x1 = r1;
return uint64_hi64((x0 << 32) | x1, truncated);
}
fastfloat_really_inline FASTFLOAT_CONSTEXPR20
uint64_t uint32_hi64(uint32_t r0, uint32_t r1, uint32_t r2, bool& truncated) noexcept {
fastfloat_really_inline FASTFLOAT_CONSTEXPR20 uint64_t
uint32_hi64(uint32_t r0, uint32_t r1, uint32_t r2, bool &truncated) noexcept {
uint64_t x0 = r0;
uint64_t x1 = r1;
uint64_t x2 = r2;
@ -205,17 +198,17 @@ uint64_t uint32_hi64(uint32_t r0, uint32_t r1, uint32_t r2, bool& truncated) noe
// we want an efficient operation. for msvc, where
// we don't have built-in intrinsics, this is still
// pretty fast.
fastfloat_really_inline FASTFLOAT_CONSTEXPR20
limb scalar_add(limb x, limb y, bool& overflow) noexcept {
fastfloat_really_inline FASTFLOAT_CONSTEXPR20 limb
scalar_add(limb x, limb y, bool &overflow) noexcept {
limb z;
// gcc and clang
#if defined(__has_builtin)
#if __has_builtin(__builtin_add_overflow)
if (!cpp20_and_in_constexpr()) {
overflow = __builtin_add_overflow(x, y, &z);
return z;
}
#endif
#if __has_builtin(__builtin_add_overflow)
if (!cpp20_and_in_constexpr()) {
overflow = __builtin_add_overflow(x, y, &z);
return z;
}
#endif
#endif
// generic, this still optimizes correctly on MSVC.
@ -225,24 +218,24 @@ limb scalar_add(limb x, limb y, bool& overflow) noexcept {
}
// multiply two small integers, getting both the high and low bits.
fastfloat_really_inline FASTFLOAT_CONSTEXPR20
limb scalar_mul(limb x, limb y, limb& carry) noexcept {
fastfloat_really_inline FASTFLOAT_CONSTEXPR20 limb
scalar_mul(limb x, limb y, limb &carry) noexcept {
#ifdef FASTFLOAT_64BIT_LIMB
#if defined(__SIZEOF_INT128__)
#if defined(__SIZEOF_INT128__)
// GCC and clang both define it as an extension.
__uint128_t z = __uint128_t(x) * __uint128_t(y) + __uint128_t(carry);
carry = limb(z >> limb_bits);
return limb(z);
#else
#else
// fallback, no native 128-bit integer multiplication with carry.
// on msvc, this optimizes identically, somehow.
value128 z = full_multiplication(x, y);
bool overflow;
z.low = scalar_add(z.low, carry, overflow);
z.high += uint64_t(overflow); // cannot overflow
z.high += uint64_t(overflow); // cannot overflow
carry = z.high;
return z.low;
#endif
#endif
#else
uint64_t z = uint64_t(x) * uint64_t(y) + uint64_t(carry);
carry = limb(z >> limb_bits);
@ -253,8 +246,8 @@ limb scalar_mul(limb x, limb y, limb& carry) noexcept {
// add scalar value to bigint starting from offset.
// used in grade school multiplication
template <uint16_t size>
inline FASTFLOAT_CONSTEXPR20
bool small_add_from(stackvec<size>& vec, limb y, size_t start) noexcept {
inline FASTFLOAT_CONSTEXPR20 bool small_add_from(stackvec<size> &vec, limb y,
size_t start) noexcept {
size_t index = start;
limb carry = y;
bool overflow;
@ -271,15 +264,15 @@ bool small_add_from(stackvec<size>& vec, limb y, size_t start) noexcept {
// add scalar value to bigint.
template <uint16_t size>
fastfloat_really_inline FASTFLOAT_CONSTEXPR20
bool small_add(stackvec<size>& vec, limb y) noexcept {
fastfloat_really_inline FASTFLOAT_CONSTEXPR20 bool
small_add(stackvec<size> &vec, limb y) noexcept {
return small_add_from(vec, y, 0);
}
// multiply bigint by scalar value.
template <uint16_t size>
inline FASTFLOAT_CONSTEXPR20
bool small_mul(stackvec<size>& vec, limb y) noexcept {
inline FASTFLOAT_CONSTEXPR20 bool small_mul(stackvec<size> &vec,
limb y) noexcept {
limb carry = 0;
for (size_t index = 0; index < vec.len(); index++) {
vec[index] = scalar_mul(vec[index], y, carry);
@ -293,12 +286,12 @@ bool small_mul(stackvec<size>& vec, limb y) noexcept {
// add bigint to bigint starting from index.
// used in grade school multiplication
template <uint16_t size>
FASTFLOAT_CONSTEXPR20
bool large_add_from(stackvec<size>& x, limb_span y, size_t start) noexcept {
FASTFLOAT_CONSTEXPR20 bool large_add_from(stackvec<size> &x, limb_span y,
size_t start) noexcept {
// the effective x buffer is from `xstart..x.len()`, so exit early
// if we can't get that current range.
if (x.len() < start || y.len() > x.len() - start) {
FASTFLOAT_TRY(x.try_resize(y.len() + start, 0));
FASTFLOAT_TRY(x.try_resize(y.len() + start, 0));
}
bool carry = false;
@ -324,15 +317,14 @@ bool large_add_from(stackvec<size>& x, limb_span y, size_t start) noexcept {
// add bigint to bigint.
template <uint16_t size>
fastfloat_really_inline FASTFLOAT_CONSTEXPR20
bool large_add_from(stackvec<size>& x, limb_span y) noexcept {
fastfloat_really_inline FASTFLOAT_CONSTEXPR20 bool
large_add_from(stackvec<size> &x, limb_span y) noexcept {
return large_add_from(x, y, 0);
}
// grade-school multiplication algorithm
template <uint16_t size>
FASTFLOAT_CONSTEXPR20
bool long_mul(stackvec<size>& x, limb_span y) noexcept {
FASTFLOAT_CONSTEXPR20 bool long_mul(stackvec<size> &x, limb_span y) noexcept {
limb_span xs = limb_span(x.data, x.len());
stackvec<size> z(xs);
limb_span zs = limb_span(z.data, z.len());
@ -360,8 +352,7 @@ bool long_mul(stackvec<size>& x, limb_span y) noexcept {
// grade-school multiplication algorithm
template <uint16_t size>
FASTFLOAT_CONSTEXPR20
bool large_mul(stackvec<size>& x, limb_span y) noexcept {
FASTFLOAT_CONSTEXPR20 bool large_mul(stackvec<size> &x, limb_span y) noexcept {
if (y.len() == 1) {
FASTFLOAT_TRY(small_mul(x, y[0]));
} else {
@ -370,36 +361,58 @@ bool large_mul(stackvec<size>& x, limb_span y) noexcept {
return true;
}
template <typename = void>
struct pow5_tables {
template <typename = void> struct pow5_tables {
static constexpr uint32_t large_step = 135;
static constexpr uint64_t small_power_of_5[] = {
1UL, 5UL, 25UL, 125UL, 625UL, 3125UL, 15625UL, 78125UL, 390625UL,
1953125UL, 9765625UL, 48828125UL, 244140625UL, 1220703125UL,
6103515625UL, 30517578125UL, 152587890625UL, 762939453125UL,
3814697265625UL, 19073486328125UL, 95367431640625UL, 476837158203125UL,
2384185791015625UL, 11920928955078125UL, 59604644775390625UL,
298023223876953125UL, 1490116119384765625UL, 7450580596923828125UL,
1UL,
5UL,
25UL,
125UL,
625UL,
3125UL,
15625UL,
78125UL,
390625UL,
1953125UL,
9765625UL,
48828125UL,
244140625UL,
1220703125UL,
6103515625UL,
30517578125UL,
152587890625UL,
762939453125UL,
3814697265625UL,
19073486328125UL,
95367431640625UL,
476837158203125UL,
2384185791015625UL,
11920928955078125UL,
59604644775390625UL,
298023223876953125UL,
1490116119384765625UL,
7450580596923828125UL,
};
#ifdef FASTFLOAT_64BIT_LIMB
constexpr static limb large_power_of_5[] = {
1414648277510068013UL, 9180637584431281687UL, 4539964771860779200UL,
10482974169319127550UL, 198276706040285095UL};
1414648277510068013UL, 9180637584431281687UL, 4539964771860779200UL,
10482974169319127550UL, 198276706040285095UL};
#else
constexpr static limb large_power_of_5[] = {
4279965485U, 329373468U, 4020270615U, 2137533757U, 4287402176U,
1057042919U, 1071430142U, 2440757623U, 381945767U, 46164893U};
4279965485U, 329373468U, 4020270615U, 2137533757U, 4287402176U,
1057042919U, 1071430142U, 2440757623U, 381945767U, 46164893U};
#endif
};
template <typename T>
constexpr uint32_t pow5_tables<T>::large_step;
#if FASTFLOAT_DETAIL_MUST_DEFINE_CONSTEXPR_VARIABLE
template <typename T>
constexpr uint64_t pow5_tables<T>::small_power_of_5[];
template <typename T> constexpr uint32_t pow5_tables<T>::large_step;
template <typename T>
constexpr limb pow5_tables<T>::large_power_of_5[];
template <typename T> constexpr uint64_t pow5_tables<T>::small_power_of_5[];
template <typename T> constexpr limb pow5_tables<T>::large_power_of_5[];
#endif
// big integer type. implements a small subset of big integer
// arithmetic, using simple algorithms since asymptotically
@ -409,13 +422,13 @@ struct bigint : pow5_tables<> {
// storage of the limbs, in little-endian order.
stackvec<bigint_limbs> vec;
FASTFLOAT_CONSTEXPR20 bigint(): vec() {}
FASTFLOAT_CONSTEXPR20 bigint() : vec() {}
bigint(const bigint &) = delete;
bigint &operator=(const bigint &) = delete;
bigint(bigint &&) = delete;
bigint &operator=(bigint &&other) = delete;
FASTFLOAT_CONSTEXPR20 bigint(uint64_t value): vec() {
FASTFLOAT_CONSTEXPR20 bigint(uint64_t value) : vec() {
#ifdef FASTFLOAT_64BIT_LIMB
vec.push_unchecked(value);
#else
@ -427,7 +440,7 @@ struct bigint : pow5_tables<> {
// get the high 64 bits from the vector, and if bits were truncated.
// this is to get the significant digits for the float.
FASTFLOAT_CONSTEXPR20 uint64_t hi64(bool& truncated) const noexcept {
FASTFLOAT_CONSTEXPR20 uint64_t hi64(bool &truncated) const noexcept {
#ifdef FASTFLOAT_64BIT_LIMB
if (vec.len() == 0) {
return empty_hi64(truncated);
@ -446,7 +459,8 @@ struct bigint : pow5_tables<> {
} else if (vec.len() == 2) {
return uint32_hi64(vec.rindex(0), vec.rindex(1), truncated);
} else {
uint64_t result = uint32_hi64(vec.rindex(0), vec.rindex(1), vec.rindex(2), truncated);
uint64_t result =
uint32_hi64(vec.rindex(0), vec.rindex(1), vec.rindex(2), truncated);
truncated |= vec.nonzero(3);
return result;
}
@ -459,7 +473,7 @@ struct bigint : pow5_tables<> {
// positive, this is larger, otherwise they are equal.
// the limbs are stored in little-endian order, so we
// must compare the limbs in ever order.
FASTFLOAT_CONSTEXPR20 int compare(const bigint& other) const noexcept {
FASTFLOAT_CONSTEXPR20 int compare(const bigint &other) const noexcept {
if (vec.len() > other.vec.len()) {
return 1;
} else if (vec.len() < other.vec.len()) {
@ -512,12 +526,12 @@ struct bigint : pow5_tables<> {
return false;
} else if (!vec.is_empty()) {
// move limbs
limb* dst = vec.data + n;
const limb* src = vec.data;
limb *dst = vec.data + n;
const limb *src = vec.data;
std::copy_backward(src, src + vec.len(), dst + vec.len());
// fill in empty limbs
limb* first = vec.data;
limb* last = first + n;
limb *first = vec.data;
limb *last = first + n;
::std::fill(first, last, 0);
vec.set_len(n + vec.len());
return true;
@ -560,18 +574,12 @@ struct bigint : pow5_tables<> {
return int(limb_bits * vec.len()) - lz;
}
FASTFLOAT_CONSTEXPR20 bool mul(limb y) noexcept {
return small_mul(vec, y);
}
FASTFLOAT_CONSTEXPR20 bool mul(limb y) noexcept { return small_mul(vec, y); }
FASTFLOAT_CONSTEXPR20 bool add(limb y) noexcept {
return small_add(vec, y);
}
FASTFLOAT_CONSTEXPR20 bool add(limb y) noexcept { return small_add(vec, y); }
// multiply as if by 2 raised to a power.
FASTFLOAT_CONSTEXPR20 bool pow2(uint32_t exp) noexcept {
return shl(exp);
}
FASTFLOAT_CONSTEXPR20 bool pow2(uint32_t exp) noexcept { return shl(exp); }
// multiply as if by 5 raised to a power.
FASTFLOAT_CONSTEXPR20 bool pow5(uint32_t exp) noexcept {
@ -597,9 +605,8 @@ struct bigint : pow5_tables<> {
// Work around clang bug https://godbolt.org/z/zedh7rrhc
// This is similar to https://github.com/llvm/llvm-project/issues/47746,
// except the workaround described there don't work here
FASTFLOAT_TRY(
small_mul(vec, limb(((void)small_power_of_5[0], small_power_of_5[exp])))
);
FASTFLOAT_TRY(small_mul(
vec, limb(((void)small_power_of_5[0], small_power_of_5[exp]))));
}
return true;

14
third_party/fast_float/constexpr_feature_detect.h vendored
View File

@ -20,16 +20,16 @@
#define FASTFLOAT_HAS_BIT_CAST 0
#endif
#if defined(__cpp_lib_is_constant_evaluated) && __cpp_lib_is_constant_evaluated >= 201811L
#if defined(__cpp_lib_is_constant_evaluated) && \
__cpp_lib_is_constant_evaluated >= 201811L
#define FASTFLOAT_HAS_IS_CONSTANT_EVALUATED 1
#else
#define FASTFLOAT_HAS_IS_CONSTANT_EVALUATED 0
#endif
// Testing for relevant C++20 constexpr library features
#if FASTFLOAT_HAS_IS_CONSTANT_EVALUATED \
&& FASTFLOAT_HAS_BIT_CAST \
&& __cpp_lib_constexpr_algorithms >= 201806L /*For std::copy and std::fill*/
#if FASTFLOAT_HAS_IS_CONSTANT_EVALUATED && FASTFLOAT_HAS_BIT_CAST && \
__cpp_lib_constexpr_algorithms >= 201806L /*For std::copy and std::fill*/
#define FASTFLOAT_CONSTEXPR20 constexpr
#define FASTFLOAT_IS_CONSTEXPR 1
#else
@ -37,4 +37,10 @@
#define FASTFLOAT_IS_CONSTEXPR 0
#endif
#if __cplusplus >= 201703L || (defined(_MSVC_LANG) && _MSVC_LANG >= 201703L)
#define FASTFLOAT_DETAIL_MUST_DEFINE_CONSTEXPR_VARIABLE 0
#else
#define FASTFLOAT_DETAIL_MUST_DEFINE_CONSTEXPR_VARIABLE 1
#endif
#endif // FASTFLOAT_CONSTEXPR_FEATURE_DETECT_H

124
third_party/fast_float/decimal_to_binary.h vendored
View File

@ -12,27 +12,34 @@
namespace fast_float {
// This will compute or rather approximate w * 5**q and return a pair of 64-bit words approximating
// the result, with the "high" part corresponding to the most significant bits and the
// low part corresponding to the least significant bits.
// This will compute or rather approximate w * 5**q and return a pair of 64-bit
// words approximating the result, with the "high" part corresponding to the
// most significant bits and the low part corresponding to the least significant
// bits.
//
template <int bit_precision>
fastfloat_really_inline FASTFLOAT_CONSTEXPR20
value128 compute_product_approximation(int64_t q, uint64_t w) {
fastfloat_really_inline FASTFLOAT_CONSTEXPR20 value128
compute_product_approximation(int64_t q, uint64_t w) {
const int index = 2 * int(q - powers::smallest_power_of_five);
// For small values of q, e.g., q in [0,27], the answer is always exact because
// The line value128 firstproduct = full_multiplication(w, power_of_five_128[index]);
// gives the exact answer.
value128 firstproduct = full_multiplication(w, powers::power_of_five_128[index]);
static_assert((bit_precision >= 0) && (bit_precision <= 64), " precision should be in (0,64]");
constexpr uint64_t precision_mask = (bit_precision < 64) ?
(uint64_t(0xFFFFFFFFFFFFFFFF) >> bit_precision)
: uint64_t(0xFFFFFFFFFFFFFFFF);
if((firstproduct.high & precision_mask) == precision_mask) { // could further guard with (lower + w < lower)
// regarding the second product, we only need secondproduct.high, but our expectation is that the compiler will optimize this extra work away if needed.
value128 secondproduct = full_multiplication(w, powers::power_of_five_128[index + 1]);
// For small values of q, e.g., q in [0,27], the answer is always exact
// because The line value128 firstproduct = full_multiplication(w,
// power_of_five_128[index]); gives the exact answer.
value128 firstproduct =
full_multiplication(w, powers::power_of_five_128[index]);
static_assert((bit_precision >= 0) && (bit_precision <= 64),
" precision should be in (0,64]");
constexpr uint64_t precision_mask =
(bit_precision < 64) ? (uint64_t(0xFFFFFFFFFFFFFFFF) >> bit_precision)
: uint64_t(0xFFFFFFFFFFFFFFFF);
if ((firstproduct.high & precision_mask) ==
precision_mask) { // could further guard with (lower + w < lower)
// regarding the second product, we only need secondproduct.high, but our
// expectation is that the compiler will optimize this extra work away if
// needed.
value128 secondproduct =
full_multiplication(w, powers::power_of_five_128[index + 1]);
firstproduct.low += secondproduct.high;
if(secondproduct.high > firstproduct.low) {
if (secondproduct.high > firstproduct.low) {
firstproduct.high++;
}
}
@ -55,43 +62,45 @@ namespace detail {
* where
* p = log(5**-q)/log(2) = -q * log(5)/log(2)
*/
constexpr fastfloat_really_inline int32_t power(int32_t q) noexcept {
return (((152170 + 65536) * q) >> 16) + 63;
}
constexpr fastfloat_really_inline int32_t power(int32_t q) noexcept {
return (((152170 + 65536) * q) >> 16) + 63;
}
} // namespace detail
// create an adjusted mantissa, biased by the invalid power2
// for significant digits already multiplied by 10 ** q.
template <typename binary>
fastfloat_really_inline FASTFLOAT_CONSTEXPR14
adjusted_mantissa compute_error_scaled(int64_t q, uint64_t w, int lz) noexcept {
fastfloat_really_inline FASTFLOAT_CONSTEXPR14 adjusted_mantissa
compute_error_scaled(int64_t q, uint64_t w, int lz) noexcept {
int hilz = int(w >> 63) ^ 1;
adjusted_mantissa answer;
answer.mantissa = w << hilz;
int bias = binary::mantissa_explicit_bits() - binary::minimum_exponent();
answer.power2 = int32_t(detail::power(int32_t(q)) + bias - hilz - lz - 62 + invalid_am_bias);
answer.power2 = int32_t(detail::power(int32_t(q)) + bias - hilz - lz - 62 +
invalid_am_bias);
return answer;
}
// w * 10 ** q, without rounding the representation up.
// the power2 in the exponent will be adjusted by invalid_am_bias.
template <typename binary>
fastfloat_really_inline FASTFLOAT_CONSTEXPR20
adjusted_mantissa compute_error(int64_t q, uint64_t w) noexcept {
fastfloat_really_inline FASTFLOAT_CONSTEXPR20 adjusted_mantissa
compute_error(int64_t q, uint64_t w) noexcept {
int lz = leading_zeroes(w);
w <<= lz;
value128 product = compute_product_approximation<binary::mantissa_explicit_bits() + 3>(q, w);
value128 product =
compute_product_approximation<binary::mantissa_explicit_bits() + 3>(q, w);
return compute_error_scaled<binary>(q, product.high, lz);
}
// w * 10 ** q
// The returned value should be a valid ieee64 number that simply need to be packed.
// However, in some very rare cases, the computation will fail. In such cases, we
// return an adjusted_mantissa with a negative power of 2: the caller should recompute
// in such cases.
// The returned value should be a valid ieee64 number that simply need to be
// packed. However, in some very rare cases, the computation will fail. In such
// cases, we return an adjusted_mantissa with a negative power of 2: the caller
// should recompute in such cases.
template <typename binary>
fastfloat_really_inline FASTFLOAT_CONSTEXPR20
adjusted_mantissa compute_float(int64_t q, uint64_t w) noexcept {
fastfloat_really_inline FASTFLOAT_CONSTEXPR20 adjusted_mantissa
compute_float(int64_t q, uint64_t w) noexcept {
adjusted_mantissa answer;
if ((w == 0) || (q < binary::smallest_power_of_ten())) {
answer.power2 = 0;
@ -105,7 +114,8 @@ adjusted_mantissa compute_float(int64_t q, uint64_t w) noexcept {
answer.mantissa = 0;
return answer;
}
// At this point in time q is in [powers::smallest_power_of_five, powers::largest_power_of_five].
// At this point in time q is in [powers::smallest_power_of_five,
// powers::largest_power_of_five].
// We want the most significant bit of i to be 1. Shift if needed.
int lz = leading_zeroes(w);
@ -114,26 +124,32 @@ adjusted_mantissa compute_float(int64_t q, uint64_t w) noexcept {
// The required precision is binary::mantissa_explicit_bits() + 3 because
// 1. We need the implicit bit
// 2. We need an extra bit for rounding purposes
// 3. We might lose a bit due to the "upperbit" routine (result too small, requiring a shift)
// 3. We might lose a bit due to the "upperbit" routine (result too small,
// requiring a shift)
value128 product = compute_product_approximation<binary::mantissa_explicit_bits() + 3>(q, w);
value128 product =
compute_product_approximation<binary::mantissa_explicit_bits() + 3>(q, w);
// The computed 'product' is always sufficient.
// Mathematical proof:
// Noble Mushtak and Daniel Lemire, Fast Number Parsing Without Fallback (to appear)
// See script/mushtak_lemire.py
// Noble Mushtak and Daniel Lemire, Fast Number Parsing Without Fallback (to
// appear) See script/mushtak_lemire.py
// The "compute_product_approximation" function can be slightly slower than a branchless approach:
// value128 product = compute_product(q, w);
// but in practice, we can win big with the compute_product_approximation if its additional branch
// is easily predicted. Which is best is data specific.
// The "compute_product_approximation" function can be slightly slower than a
// branchless approach: value128 product = compute_product(q, w); but in
// practice, we can win big with the compute_product_approximation if its
// additional branch is easily predicted. Which is best is data specific.
int upperbit = int(product.high >> 63);
int shift = upperbit + 64 - binary::mantissa_explicit_bits() - 3;
answer.mantissa = product.high >> (upperbit + 64 - binary::mantissa_explicit_bits() - 3);
answer.mantissa = product.high >> shift;
answer.power2 = int32_t(detail::power(int32_t(q)) + upperbit - lz - binary::minimum_exponent());
answer.power2 = int32_t(detail::power(int32_t(q)) + upperbit - lz -
binary::minimum_exponent());
if (answer.power2 <= 0) { // we have a subnormal?
// Here have that answer.power2 <= 0 so -answer.power2 >= 0
if(-answer.power2 + 1 >= 64) { // if we have more than 64 bits below the minimum exponent, you have a zero for sure.
if (-answer.power2 + 1 >=
64) { // if we have more than 64 bits below the minimum exponent, you
// have a zero for sure.
answer.power2 = 0;
answer.mantissa = 0;
// result should be zero
@ -152,20 +168,26 @@ adjusted_mantissa compute_float(int64_t q, uint64_t w) noexcept {
// up 0x3fffffffffffff x 2^-1023-53 and once we do, we are no longer
// subnormal, but we can only know this after rounding.
// So we only declare a subnormal if we are smaller than the threshold.
answer.power2 = (answer.mantissa < (uint64_t(1) << binary::mantissa_explicit_bits())) ? 0 : 1;
answer.power2 =
(answer.mantissa < (uint64_t(1) << binary::mantissa_explicit_bits()))
? 0
: 1;
return answer;
}
// usually, we round *up*, but if we fall right in between and and we have an
// even basis, we need to round down
// We are only concerned with the cases where 5**q fits in single 64-bit word.
if ((product.low <= 1) && (q >= binary::min_exponent_round_to_even()) && (q <= binary::max_exponent_round_to_even()) &&
((answer.mantissa & 3) == 1) ) { // we may fall between two floats!
if ((product.low <= 1) && (q >= binary::min_exponent_round_to_even()) &&
(q <= binary::max_exponent_round_to_even()) &&
((answer.mantissa & 3) == 1)) { // we may fall between two floats!
// To be in-between two floats we need that in doing
// answer.mantissa = product.high >> (upperbit + 64 - binary::mantissa_explicit_bits() - 3);
// ... we dropped out only zeroes. But if this happened, then we can go back!!!
if((answer.mantissa << (upperbit + 64 - binary::mantissa_explicit_bits() - 3)) == product.high) {
answer.mantissa &= ~uint64_t(1); // flip it so that we do not round up
// answer.mantissa = product.high >> (upperbit + 64 -
// binary::mantissa_explicit_bits() - 3);
// ... we dropped out only zeroes. But if this happened, then we can go
// back!!!
if ((answer.mantissa << shift) == product.high) {
answer.mantissa &= ~uint64_t(1); // flip it so that we do not round up
}
}

185
third_party/fast_float/digit_comparison.h vendored
View File

@ -13,19 +13,34 @@
namespace fast_float {
// 1e0 to 1e19
constexpr static uint64_t powers_of_ten_uint64[] = {
1UL, 10UL, 100UL, 1000UL, 10000UL, 100000UL, 1000000UL, 10000000UL, 100000000UL,
1000000000UL, 10000000000UL, 100000000000UL, 1000000000000UL, 10000000000000UL,
100000000000000UL, 1000000000000000UL, 10000000000000000UL, 100000000000000000UL,
1000000000000000000UL, 10000000000000000000UL};
constexpr static uint64_t powers_of_ten_uint64[] = {1UL,
10UL,
100UL,
1000UL,
10000UL,
100000UL,
1000000UL,
10000000UL,
100000000UL,
1000000000UL,
10000000000UL,
100000000000UL,
1000000000000UL,
10000000000000UL,
100000000000000UL,
1000000000000000UL,
10000000000000000UL,
100000000000000000UL,
1000000000000000000UL,
10000000000000000000UL};
// calculate the exponent, in scientific notation, of the number.
// this algorithm is not even close to optimized, but it has no practical
// effect on performance: in order to have a faster algorithm, we'd need
// to slow down performance for faster algorithms, and this is still fast.
template <typename UC>
fastfloat_really_inline FASTFLOAT_CONSTEXPR14
int32_t scientific_exponent(parsed_number_string_t<UC> & num) noexcept {
fastfloat_really_inline FASTFLOAT_CONSTEXPR14 int32_t
scientific_exponent(parsed_number_string_t<UC> &num) noexcept {
uint64_t mantissa = num.mantissa;
int32_t exponent = int32_t(num.exponent);
while (mantissa >= 10000) {
@ -45,15 +60,16 @@ int32_t scientific_exponent(parsed_number_string_t<UC> & num) noexcept {
// this converts a native floating-point number to an extended-precision float.
template <typename T>
fastfloat_really_inline FASTFLOAT_CONSTEXPR20
adjusted_mantissa to_extended(T value) noexcept {
fastfloat_really_inline FASTFLOAT_CONSTEXPR20 adjusted_mantissa
to_extended(T value) noexcept {
using equiv_uint = typename binary_format<T>::equiv_uint;
constexpr equiv_uint exponent_mask = binary_format<T>::exponent_mask();
constexpr equiv_uint mantissa_mask = binary_format<T>::mantissa_mask();
constexpr equiv_uint hidden_bit_mask = binary_format<T>::hidden_bit_mask();
adjusted_mantissa am;
int32_t bias = binary_format<T>::mantissa_explicit_bits() - binary_format<T>::minimum_exponent();
int32_t bias = binary_format<T>::mantissa_explicit_bits() -
binary_format<T>::minimum_exponent();
equiv_uint bits;
#if FASTFLOAT_HAS_BIT_CAST
bits = std::bit_cast<equiv_uint>(value);
@ -66,7 +82,8 @@ adjusted_mantissa to_extended(T value) noexcept {
am.mantissa = bits & mantissa_mask;
} else {
// normal
am.power2 = int32_t((bits & exponent_mask) >> binary_format<T>::mantissa_explicit_bits());
am.power2 = int32_t((bits & exponent_mask) >>
binary_format<T>::mantissa_explicit_bits());
am.power2 -= bias;
am.mantissa = (bits & mantissa_mask) | hidden_bit_mask;
}
@ -78,8 +95,8 @@ adjusted_mantissa to_extended(T value) noexcept {
// we are given a native float that represents b, so we need to adjust it
// halfway between b and b+u.
template <typename T>
fastfloat_really_inline FASTFLOAT_CONSTEXPR20
adjusted_mantissa to_extended_halfway(T value) noexcept {
fastfloat_really_inline FASTFLOAT_CONSTEXPR20 adjusted_mantissa
to_extended_halfway(T value) noexcept {
adjusted_mantissa am = to_extended(value);
am.mantissa <<= 1;
am.mantissa += 1;
@ -89,15 +106,18 @@ adjusted_mantissa to_extended_halfway(T value) noexcept {
// round an extended-precision float to the nearest machine float.
template <typename T, typename callback>
fastfloat_really_inline FASTFLOAT_CONSTEXPR14
void round(adjusted_mantissa& am, callback cb) noexcept {
fastfloat_really_inline FASTFLOAT_CONSTEXPR14 void round(adjusted_mantissa &am,
callback cb) noexcept {
int32_t mantissa_shift = 64 - binary_format<T>::mantissa_explicit_bits() - 1;
if (-am.power2 >= mantissa_shift) {
// have a denormal float
int32_t shift = -am.power2 + 1;
cb(am, std::min<int32_t>(shift, 64));
// check for round-up: if rounding-nearest carried us to the hidden bit.
am.power2 = (am.mantissa < (uint64_t(1) << binary_format<T>::mantissa_explicit_bits())) ? 0 : 1;
am.power2 = (am.mantissa <
(uint64_t(1) << binary_format<T>::mantissa_explicit_bits()))
? 0
: 1;
return;
}
@ -105,7 +125,8 @@ void round(adjusted_mantissa& am, callback cb) noexcept {
cb(am, mantissa_shift);
// check for carry
if (am.mantissa >= (uint64_t(2) << binary_format<T>::mantissa_explicit_bits())) {
if (am.mantissa >=
(uint64_t(2) << binary_format<T>::mantissa_explicit_bits())) {
am.mantissa = (uint64_t(1) << binary_format<T>::mantissa_explicit_bits());
am.power2++;
}
@ -119,16 +140,11 @@ void round(adjusted_mantissa& am, callback cb) noexcept {
}
template <typename callback>
fastfloat_really_inline FASTFLOAT_CONSTEXPR14
void round_nearest_tie_even(adjusted_mantissa& am, int32_t shift, callback cb) noexcept {
const uint64_t mask
= (shift == 64)
? UINT64_MAX
: (uint64_t(1) << shift) - 1;
const uint64_t halfway
= (shift == 0)
? 0
: uint64_t(1) << (shift - 1);
fastfloat_really_inline FASTFLOAT_CONSTEXPR14 void
round_nearest_tie_even(adjusted_mantissa &am, int32_t shift,
callback cb) noexcept {
const uint64_t mask = (shift == 64) ? UINT64_MAX : (uint64_t(1) << shift) - 1;
const uint64_t halfway = (shift == 0) ? 0 : uint64_t(1) << (shift - 1);
uint64_t truncated_bits = am.mantissa & mask;
bool is_above = truncated_bits > halfway;
bool is_halfway = truncated_bits == halfway;
@ -145,8 +161,8 @@ void round_nearest_tie_even(adjusted_mantissa& am, int32_t shift, callback cb) n
am.mantissa += uint64_t(cb(is_odd, is_halfway, is_above));
}
fastfloat_really_inline FASTFLOAT_CONSTEXPR14
void round_down(adjusted_mantissa& am, int32_t shift) noexcept {
fastfloat_really_inline FASTFLOAT_CONSTEXPR14 void
round_down(adjusted_mantissa &am, int32_t shift) noexcept {
if (shift == 64) {
am.mantissa = 0;
} else {
@ -155,10 +171,11 @@ void round_down(adjusted_mantissa& am, int32_t shift) noexcept {
am.power2 += shift;
}
template <typename UC>
fastfloat_really_inline FASTFLOAT_CONSTEXPR20
void skip_zeros(UC const * & first, UC const * last) noexcept {
fastfloat_really_inline FASTFLOAT_CONSTEXPR20 void
skip_zeros(UC const *&first, UC const *last) noexcept {
uint64_t val;
while (!cpp20_and_in_constexpr() && std::distance(first, last) >= int_cmp_len<UC>()) {
while (!cpp20_and_in_constexpr() &&
std::distance(first, last) >= int_cmp_len<UC>()) {
::memcpy(&val, first, sizeof(uint64_t));
if (val != int_cmp_zeros<UC>()) {
break;
@ -176,11 +193,12 @@ void skip_zeros(UC const * & first, UC const * last) noexcept {
// determine if any non-zero digits were truncated.
// all characters must be valid digits.
template <typename UC>
fastfloat_really_inline FASTFLOAT_CONSTEXPR20
bool is_truncated(UC const * first, UC const * last) noexcept {
fastfloat_really_inline FASTFLOAT_CONSTEXPR20 bool
is_truncated(UC const *first, UC const *last) noexcept {
// do 8-bit optimizations, can just compare to 8 literal 0s.
uint64_t val;
while (!cpp20_and_in_constexpr() && std::distance(first, last) >= int_cmp_len<UC>()) {
while (!cpp20_and_in_constexpr() &&
std::distance(first, last) >= int_cmp_len<UC>()) {
::memcpy(&val, first, sizeof(uint64_t));
if (val != int_cmp_zeros<UC>()) {
return true;
@ -196,15 +214,15 @@ bool is_truncated(UC const * first, UC const * last) noexcept {
return false;
}
template <typename UC>
fastfloat_really_inline FASTFLOAT_CONSTEXPR20
bool is_truncated(span<const UC> s) noexcept {
fastfloat_really_inline FASTFLOAT_CONSTEXPR20 bool
is_truncated(span<const UC> s) noexcept {
return is_truncated(s.ptr, s.ptr + s.len());
}
template <typename UC>
fastfloat_really_inline FASTFLOAT_CONSTEXPR20
void parse_eight_digits(const UC*& p, limb& value, size_t& counter, size_t& count) noexcept {
fastfloat_really_inline FASTFLOAT_CONSTEXPR20 void
parse_eight_digits(const UC *&p, limb &value, size_t &counter,
size_t &count) noexcept {
value = value * 100000000 + parse_eight_digits_unrolled(p);
p += 8;
counter += 8;
@ -212,22 +230,23 @@ void parse_eight_digits(const UC*& p, limb& value, size_t& counter, size_t& coun
}
template <typename UC>
fastfloat_really_inline FASTFLOAT_CONSTEXPR14
void parse_one_digit(UC const *& p, limb& value, size_t& counter, size_t& count) noexcept {
fastfloat_really_inline FASTFLOAT_CONSTEXPR14 void
parse_one_digit(UC const *&p, limb &value, size_t &counter,
size_t &count) noexcept {
value = value * 10 + limb(*p - UC('0'));
p++;
counter++;
count++;
}
fastfloat_really_inline FASTFLOAT_CONSTEXPR20
void add_native(bigint& big, limb power, limb value) noexcept {
fastfloat_really_inline FASTFLOAT_CONSTEXPR20 void
add_native(bigint &big, limb power, limb value) noexcept {
big.mul(power);
big.add(value);
}
fastfloat_really_inline FASTFLOAT_CONSTEXPR20
void round_up_bigint(bigint& big, size_t& count) noexcept {
fastfloat_really_inline FASTFLOAT_CONSTEXPR20 void
round_up_bigint(bigint &big, size_t &count) noexcept {
// need to round-up the digits, but need to avoid rounding
// ....9999 to ...10000, which could cause a false halfway point.
add_native(big, 10, 1);
@ -236,8 +255,9 @@ void round_up_bigint(bigint& big, size_t& count) noexcept {
// parse the significant digits into a big integer
template <typename UC>
inline FASTFLOAT_CONSTEXPR20
void parse_mantissa(bigint& result, parsed_number_string_t<UC>& num, size_t max_digits, size_t& digits) noexcept {
inline FASTFLOAT_CONSTEXPR20 void
parse_mantissa(bigint &result, parsed_number_string_t<UC> &num,
size_t max_digits, size_t &digits) noexcept {
// try to minimize the number of big integer and scalar multiplication.
// therefore, try to parse 8 digits at a time, and multiply by the largest
// scalar value (9 or 19 digits) for each step.
@ -251,12 +271,13 @@ void parse_mantissa(bigint& result, parsed_number_string_t<UC>& num, size_t max_
#endif
// process all integer digits.
UC const * p = num.integer.ptr;
UC const * pend = p + num.integer.len();
UC const *p = num.integer.ptr;
UC const *pend = p + num.integer.len();
skip_zeros(p, pend);
// process all digits, in increments of step per loop
while (p != pend) {
while ((std::distance(p, pend) >= 8) && (step - counter >= 8) && (max_digits - digits >= 8)) {
while ((std::distance(p, pend) >= 8) && (step - counter >= 8) &&
(max_digits - digits >= 8)) {
parse_eight_digits(p, value, counter, digits);
}
while (counter < step && p != pend && digits < max_digits) {
@ -289,7 +310,8 @@ void parse_mantissa(bigint& result, parsed_number_string_t<UC>& num, size_t max_
}
// process all digits, in increments of step per loop
while (p != pend) {
while ((std::distance(p, pend) >= 8) && (step - counter >= 8) && (max_digits - digits >= 8)) {
while ((std::distance(p, pend) >= 8) && (step - counter >= 8) &&
(max_digits - digits >= 8)) {
parse_eight_digits(p, value, counter, digits);
}
while (counter < step && p != pend && digits < max_digits) {
@ -317,19 +339,23 @@ void parse_mantissa(bigint& result, parsed_number_string_t<UC>& num, size_t max_
}
template <typename T>
inline FASTFLOAT_CONSTEXPR20
adjusted_mantissa positive_digit_comp(bigint& bigmant, int32_t exponent) noexcept {
inline FASTFLOAT_CONSTEXPR20 adjusted_mantissa
positive_digit_comp(bigint &bigmant, int32_t exponent) noexcept {
FASTFLOAT_ASSERT(bigmant.pow10(uint32_t(exponent)));
adjusted_mantissa answer;
bool truncated;
answer.mantissa = bigmant.hi64(truncated);
int bias = binary_format<T>::mantissa_explicit_bits() - binary_format<T>::minimum_exponent();
int bias = binary_format<T>::mantissa_explicit_bits() -
binary_format<T>::minimum_exponent();
answer.power2 = bigmant.bit_length() - 64 + bias;
round<T>(answer, [truncated](adjusted_mantissa& a, int32_t shift) {
round_nearest_tie_even(a, shift, [truncated](bool is_odd, bool is_halfway, bool is_above) -> bool {
return is_above || (is_halfway && truncated) || (is_odd && is_halfway);
});
round<T>(answer, [truncated](adjusted_mantissa &a, int32_t shift) {
round_nearest_tie_even(
a, shift,
[truncated](bool is_odd, bool is_halfway, bool is_above) -> bool {
return is_above || (is_halfway && truncated) ||
(is_odd && is_halfway);
});
});
return answer;
@ -341,15 +367,17 @@ adjusted_mantissa positive_digit_comp(bigint& bigmant, int32_t exponent) noexcep
// we then need to scale by `2^(f- e)`, and then the two significant digits
// are of the same magnitude.
template <typename T>
inline FASTFLOAT_CONSTEXPR20
adjusted_mantissa negative_digit_comp(bigint& bigmant, adjusted_mantissa am, int32_t exponent) noexcept {
bigint& real_digits = bigmant;
inline FASTFLOAT_CONSTEXPR20 adjusted_mantissa negative_digit_comp(
bigint &bigmant, adjusted_mantissa am, int32_t exponent) noexcept {
bigint &real_digits = bigmant;
int32_t real_exp = exponent;
// get the value of `b`, rounded down, and get a bigint representation of b+h
adjusted_mantissa am_b = am;
// gcc7 buf: use a lambda to remove the noexcept qualifier bug with -Wnoexcept-type.
round<T>(am_b, [](adjusted_mantissa&a, int32_t shift) { round_down(a, shift); });
// gcc7 buf: use a lambda to remove the noexcept qualifier bug with
// -Wnoexcept-type.
round<T>(am_b,
[](adjusted_mantissa &a, int32_t shift) { round_down(a, shift); });
T b;
to_float(false, am_b, b);
adjusted_mantissa theor = to_extended_halfway(b);
@ -371,18 +399,19 @@ adjusted_mantissa negative_digit_comp(bigint& bigmant, adjusted_mantissa am, int
// compare digits, and use it to director rounding
int ord = real_digits.compare(theor_digits);
adjusted_mantissa answer = am;
round<T>(answer, [ord](adjusted_mantissa& a, int32_t shift) {
round_nearest_tie_even(a, shift, [ord](bool is_odd, bool _, bool __) -> bool {
(void)_; // not needed, since we've done our comparison
(void)__; // not needed, since we've done our comparison
if (ord > 0) {
return true;
} else if (ord < 0) {
return false;
} else {
return is_odd;
}
});
round<T>(answer, [ord](adjusted_mantissa &a, int32_t shift) {
round_nearest_tie_even(
a, shift, [ord](bool is_odd, bool _, bool __) -> bool {
(void)_; // not needed, since we've done our comparison
(void)__; // not needed, since we've done our comparison
if (ord > 0) {
return true;
} else if (ord < 0) {
return false;
} else {
return is_odd;
}
});
});
return answer;
@ -402,8 +431,8 @@ adjusted_mantissa negative_digit_comp(bigint& bigmant, adjusted_mantissa am, int
// the actual digits. we then compare the big integer representations
// of both, and use that to direct rounding.
template <typename T, typename UC>
inline FASTFLOAT_CONSTEXPR20
adjusted_mantissa digit_comp(parsed_number_string_t<UC>& num, adjusted_mantissa am) noexcept {
inline FASTFLOAT_CONSTEXPR20 adjusted_mantissa
digit_comp(parsed_number_string_t<UC> &num, adjusted_mantissa am) noexcept {
// remove the invalid exponent bias
am.power2 -= invalid_am_bias;

58
third_party/fast_float/fast_float.h vendored
View File

@ -6,36 +6,50 @@
namespace fast_float {
/**
* This function parses the character sequence [first,last) for a number. It parses floating-point numbers expecting
* a locale-indepent format equivalent to what is used by std::strtod in the default ("C") locale.
* The resulting floating-point value is the closest floating-point values (using either float or double),
* using the "round to even" convention for values that would otherwise fall right in-between two values.
* That is, we provide exact parsing according to the IEEE standard.
* This function parses the character sequence [first,last) for a number. It
* parses floating-point numbers expecting a locale-indepent format equivalent
* to what is used by std::strtod in the default ("C") locale. The resulting
* floating-point value is the closest floating-point values (using either float
* or double), using the "round to even" convention for values that would
* otherwise fall right in-between two values. That is, we provide exact parsing
* according to the IEEE standard.
*
* Given a successful parse, the pointer (`ptr`) in the returned value is set to point right after the
* parsed number, and the `value` referenced is set to the parsed value. In case of error, the returned
* `ec` contains a representative error, otherwise the default (`std::errc()`) value is stored.
* Given a successful parse, the pointer (`ptr`) in the returned value is set to
* point right after the parsed number, and the `value` referenced is set to the
* parsed value. In case of error, the returned `ec` contains a representative
* error, otherwise the default (`std::errc()`) value is stored.
*
* The implementation does not throw and does not allocate memory (e.g., with `new` or `malloc`).
* The implementation does not throw and does not allocate memory (e.g., with
* `new` or `malloc`).
*
* Like the C++17 standard, the `fast_float::from_chars` functions take an optional last argument of
* the type `fast_float::chars_format`. It is a bitset value: we check whether
* `fmt & fast_float::chars_format::fixed` and `fmt & fast_float::chars_format::scientific` are set
* to determine whether we allow the fixed point and scientific notation respectively.
* The default is `fast_float::chars_format::general` which allows both `fixed` and `scientific`.
* Like the C++17 standard, the `fast_float::from_chars` functions take an
* optional last argument of the type `fast_float::chars_format`. It is a bitset
* value: we check whether `fmt & fast_float::chars_format::fixed` and `fmt &
* fast_float::chars_format::scientific` are set to determine whether we allow
* the fixed point and scientific notation respectively. The default is
* `fast_float::chars_format::general` which allows both `fixed` and
* `scientific`.
*/
template<typename T, typename UC = char>
FASTFLOAT_CONSTEXPR20
from_chars_result_t<UC> from_chars(UC const * first, UC const * last,
T &value, chars_format fmt = chars_format::general) noexcept;
template <typename T, typename UC = char,
typename = FASTFLOAT_ENABLE_IF(is_supported_float_type<T>())>
FASTFLOAT_CONSTEXPR20 from_chars_result_t<UC>
from_chars(UC const *first, UC const *last, T &value,
chars_format fmt = chars_format::general) noexcept;
/**
* Like from_chars, but accepts an `options` argument to govern number parsing.
*/
template<typename T, typename UC = char>
FASTFLOAT_CONSTEXPR20
from_chars_result_t<UC> from_chars_advanced(UC const * first, UC const * last,
T &value, parse_options_t<UC> options) noexcept;
template <typename T, typename UC = char>
FASTFLOAT_CONSTEXPR20 from_chars_result_t<UC>
from_chars_advanced(UC const *first, UC const *last, T &value,
parse_options_t<UC> options) noexcept;
/**
* from_chars for integer types.
*/
template <typename T, typename UC = char,
typename = FASTFLOAT_ENABLE_IF(!is_supported_float_type<T>())>
FASTFLOAT_CONSTEXPR20 from_chars_result_t<UC>
from_chars(UC const *first, UC const *last, T &value, int base = 10) noexcept;
} // namespace fast_float
#include "parse_number.h"

1326
third_party/fast_float/fast_table.h vendored
View File

File diff suppressed because it is too large Load Diff

589
third_party/fast_float/float_common.h vendored
View File

@ -7,7 +7,11 @@
#include <cstring>
#include <type_traits>
#include <system_error>
#ifdef __has_include
#if __has_include(<stdfloat>) && (__cplusplus > 202002L || _MSVC_LANG > 202002L)
#include <stdfloat>
#endif
#endif
#include "constexpr_feature_detect.h"
namespace fast_float {
@ -28,18 +32,16 @@ enum chars_format {
general = fixed | scientific
};
template <typename UC>
struct from_chars_result_t {
UC const* ptr;
template <typename UC> struct from_chars_result_t {
UC const *ptr;
std::errc ec;
};
using from_chars_result = from_chars_result_t<char>;
template <typename UC>
struct parse_options_t {
template <typename UC> struct parse_options_t {
constexpr explicit parse_options_t(chars_format fmt = chars_format::general,
UC dot = UC('.'))
: format(fmt), decimal_point(dot) {}
UC dot = UC('.'))
: format(fmt), decimal_point(dot) {}
/** Which number formats are accepted */
chars_format format;
@ -48,39 +50,41 @@ struct parse_options_t {
};
using parse_options = parse_options_t<char>;
}
} // namespace fast_float
#if FASTFLOAT_HAS_BIT_CAST
#include <bit>
#endif
#if (defined(__x86_64) || defined(__x86_64__) || defined(_M_X64) \
|| defined(__amd64) || defined(__aarch64__) || defined(_M_ARM64) \
|| defined(__MINGW64__) \
|| defined(__s390x__) \
|| (defined(__ppc64__) || defined(__PPC64__) || defined(__ppc64le__) || defined(__PPC64LE__)) \
|| defined(__loongarch64) )
#if (defined(__x86_64) || defined(__x86_64__) || defined(_M_X64) || \
defined(__amd64) || defined(__aarch64__) || defined(_M_ARM64) || \
defined(__MINGW64__) || defined(__s390x__) || \
(defined(__ppc64__) || defined(__PPC64__) || defined(__ppc64le__) || \
defined(__PPC64LE__)) || \
defined(__loongarch64))
#define FASTFLOAT_64BIT 1
#elif (defined(__i386) || defined(__i386__) || defined(_M_IX86) \
|| defined(__arm__) || defined(_M_ARM) || defined(__ppc__) \
|| defined(__MINGW32__) || defined(__EMSCRIPTEN__))
#elif (defined(__i386) || defined(__i386__) || defined(_M_IX86) || \
defined(__arm__) || defined(_M_ARM) || defined(__ppc__) || \
defined(__MINGW32__) || defined(__EMSCRIPTEN__))
#define FASTFLOAT_32BIT 1
#else
// Need to check incrementally, since SIZE_MAX is a size_t, avoid overflow.
// We can never tell the register width, but the SIZE_MAX is a good approximation.
// UINTPTR_MAX and INTPTR_MAX are optional, so avoid them for max portability.
#if SIZE_MAX == 0xffff
#error Unknown platform (16-bit, unsupported)
#elif SIZE_MAX == 0xffffffff
#define FASTFLOAT_32BIT 1
#elif SIZE_MAX == 0xffffffffffffffff
#define FASTFLOAT_64BIT 1
#else
#error Unknown platform (not 32-bit, not 64-bit?)
#endif
// We can never tell the register width, but the SIZE_MAX is a good
// approximation. UINTPTR_MAX and INTPTR_MAX are optional, so avoid them for max
// portability.
#if SIZE_MAX == 0xffff
#error Unknown platform (16-bit, unsupported)
#elif SIZE_MAX == 0xffffffff
#define FASTFLOAT_32BIT 1
#elif SIZE_MAX == 0xffffffffffffffff
#define FASTFLOAT_64BIT 1
#else
#error Unknown platform (not 32-bit, not 64-bit?)
#endif
#endif
#if ((defined(_WIN32) || defined(_WIN64)) && !defined(__clang__))
#if ((defined(_WIN32) || defined(_WIN64)) && !defined(__clang__)) || \
(defined(_M_ARM64) && !defined(__MINGW32__))
#include <intrin.h>
#endif
@ -124,9 +128,9 @@ using parse_options = parse_options_t<char>;
#endif
#endif
#if defined(__SSE2__) || \
(defined(FASTFLOAT_VISUAL_STUDIO) && \
(defined(_M_AMD64) || defined(_M_X64) || (defined(_M_IX86_FP) && _M_IX86_FP == 2)))
#if defined(__SSE2__) || (defined(FASTFLOAT_VISUAL_STUDIO) && \
(defined(_M_AMD64) || defined(_M_X64) || \
(defined(_M_IX86_FP) && _M_IX86_FP == 2)))
#define FASTFLOAT_SSE2 1
#endif
@ -134,28 +138,25 @@ using parse_options = parse_options_t<char>;
#define FASTFLOAT_NEON 1
#endif
#if defined(FASTFLOAT_SSE2) || defined(FASTFLOAT_ARM64)
#if defined(FASTFLOAT_SSE2) || defined(FASTFLOAT_NEON)
#define FASTFLOAT_HAS_SIMD 1
#endif
#if defined(__GNUC__)
// disable -Wcast-align=strict (GCC only)
#define FASTFLOAT_SIMD_DISABLE_WARNINGS \
_Pragma("GCC diagnostic push") \
_Pragma("GCC diagnostic ignored \"-Wcast-align\"")
#define FASTFLOAT_SIMD_DISABLE_WARNINGS \
_Pragma("GCC diagnostic push") \
_Pragma("GCC diagnostic ignored \"-Wcast-align\"")
#else
#define FASTFLOAT_SIMD_DISABLE_WARNINGS
#endif
#if defined(__GNUC__)
#define FASTFLOAT_SIMD_RESTORE_WARNINGS \
_Pragma("GCC diagnostic pop")
#define FASTFLOAT_SIMD_RESTORE_WARNINGS _Pragma("GCC diagnostic pop")
#else
#define FASTFLOAT_SIMD_RESTORE_WARNINGS
#endif
#ifdef FASTFLOAT_VISUAL_STUDIO
#define fastfloat_really_inline __forceinline
#else
@ -163,18 +164,24 @@ using parse_options = parse_options_t<char>;
#endif
#ifndef FASTFLOAT_ASSERT
#define FASTFLOAT_ASSERT(x) { ((void)(x)); }
#define FASTFLOAT_ASSERT(x) \
{ ((void)(x)); }
#endif
#ifndef FASTFLOAT_DEBUG_ASSERT
#define FASTFLOAT_DEBUG_ASSERT(x) { ((void)(x)); }
#define FASTFLOAT_DEBUG_ASSERT(x) \
{ ((void)(x)); }
#endif
// rust style `try!()` macro, or `?` operator
#define FASTFLOAT_TRY(x) { if (!(x)) return false; }
#define FASTFLOAT_ENABLE_IF(...) typename std::enable_if<(__VA_ARGS__), int>::type = 0
#define FASTFLOAT_TRY(x) \
{ \
if (!(x)) \
return false; \
}
#define FASTFLOAT_ENABLE_IF(...) \
typename std::enable_if<(__VA_ARGS__), int>::type
namespace fast_float {
@ -186,10 +193,28 @@ fastfloat_really_inline constexpr bool cpp20_and_in_constexpr() {
#endif
}
template <typename T>
fastfloat_really_inline constexpr bool is_supported_float_type() {
return std::is_same<T, float>::value || std::is_same<T, double>::value
#if __STDCPP_FLOAT32_T__
|| std::is_same<T, std::float32_t>::value
#endif
#if __STDCPP_FLOAT64_T__
|| std::is_same<T, std::float64_t>::value
#endif
;
}
template <typename UC>
fastfloat_really_inline constexpr bool is_supported_char_type() {
return std::is_same<UC, char>::value || std::is_same<UC, wchar_t>::value ||
std::is_same<UC, char16_t>::value || std::is_same<UC, char32_t>::value;
}
// Compares two ASCII strings in a case insensitive manner.
template <typename UC>
inline FASTFLOAT_CONSTEXPR14 bool
fastfloat_strncasecmp(UC const * input1, UC const * input2, size_t length) {
fastfloat_strncasecmp(UC const *input1, UC const *input2, size_t length) {
char running_diff{0};
for (size_t i = 0; i < length; ++i) {
running_diff |= (char(input1[i]) ^ char(input2[i]));
@ -202,18 +227,15 @@ fastfloat_strncasecmp(UC const * input1, UC const * input2, size_t length) {
#endif
// a pointer and a length to a contiguous block of memory
template <typename T>
struct span {
const T* ptr;
template <typename T> struct span {
const T *ptr;
size_t length;
constexpr span(const T* _ptr, size_t _length) : ptr(_ptr), length(_length) {}
constexpr span(const T *_ptr, size_t _length) : ptr(_ptr), length(_length) {}
constexpr span() : ptr(nullptr), length(0) {}
constexpr size_t len() const noexcept {
return length;
}
constexpr size_t len() const noexcept { return length; }
FASTFLOAT_CONSTEXPR14 const T& operator[](size_t index) const noexcept {
FASTFLOAT_CONSTEXPR14 const T &operator[](size_t index) const noexcept {
FASTFLOAT_DEBUG_ASSERT(index < length);
return ptr[index];
}
@ -227,34 +249,51 @@ struct value128 {
};
/* Helper C++14 constexpr generic implementation of leading_zeroes */
fastfloat_really_inline FASTFLOAT_CONSTEXPR14
int leading_zeroes_generic(uint64_t input_num, int last_bit = 0) {
if(input_num & uint64_t(0xffffffff00000000)) { input_num >>= 32; last_bit |= 32; }
if(input_num & uint64_t( 0xffff0000)) { input_num >>= 16; last_bit |= 16; }
if(input_num & uint64_t( 0xff00)) { input_num >>= 8; last_bit |= 8; }
if(input_num & uint64_t( 0xf0)) { input_num >>= 4; last_bit |= 4; }
if(input_num & uint64_t( 0xc)) { input_num >>= 2; last_bit |= 2; }
if(input_num & uint64_t( 0x2)) { input_num >>= 1; last_bit |= 1; }
return 63 - last_bit;
fastfloat_really_inline FASTFLOAT_CONSTEXPR14 int
leading_zeroes_generic(uint64_t input_num, int last_bit = 0) {
if (input_num & uint64_t(0xffffffff00000000)) {
input_num >>= 32;
last_bit |= 32;
}
if (input_num & uint64_t(0xffff0000)) {
input_num >>= 16;
last_bit |= 16;
}
if (input_num & uint64_t(0xff00)) {
input_num >>= 8;
last_bit |= 8;
}
if (input_num & uint64_t(0xf0)) {
input_num >>= 4;
last_bit |= 4;
}
if (input_num & uint64_t(0xc)) {
input_num >>= 2;
last_bit |= 2;
}
if (input_num & uint64_t(0x2)) { /* input_num >>= 1; */
last_bit |= 1;
}
return 63 - last_bit;
}
/* result might be undefined when input_num is zero */
fastfloat_really_inline FASTFLOAT_CONSTEXPR20
int leading_zeroes(uint64_t input_num) {
fastfloat_really_inline FASTFLOAT_CONSTEXPR20 int
leading_zeroes(uint64_t input_num) {
assert(input_num > 0);
if (cpp20_and_in_constexpr()) {
return leading_zeroes_generic(input_num);
}
#ifdef FASTFLOAT_VISUAL_STUDIO
#if defined(_M_X64) || defined(_M_ARM64)
#if defined(_M_X64) || defined(_M_ARM64)
unsigned long leading_zero = 0;
// Search the mask data from most significant bit (MSB)
// to least significant bit (LSB) for a set bit (1).
_BitScanReverse64(&leading_zero, input_num);
return (int)(63 - leading_zero);
#else
#else
return leading_zeroes_generic(input_num);
#endif
#endif
#else
return __builtin_clzll(input_num);
#endif
@ -262,18 +301,18 @@ int leading_zeroes(uint64_t input_num) {
// slow emulation routine for 32-bit
fastfloat_really_inline constexpr uint64_t emulu(uint32_t x, uint32_t y) {
return x * (uint64_t)y;
return x * (uint64_t)y;
}
fastfloat_really_inline FASTFLOAT_CONSTEXPR14
uint64_t umul128_generic(uint64_t ab, uint64_t cd, uint64_t *hi) {
fastfloat_really_inline FASTFLOAT_CONSTEXPR14 uint64_t
umul128_generic(uint64_t ab, uint64_t cd, uint64_t *hi) {
uint64_t ad = emulu((uint32_t)(ab >> 32), (uint32_t)cd);
uint64_t bd = emulu((uint32_t)ab, (uint32_t)cd);
uint64_t adbc = ad + emulu((uint32_t)ab, (uint32_t)(cd >> 32));
uint64_t adbc_carry = !!(adbc < ad);
uint64_t adbc_carry = (uint64_t)(adbc < ad);
uint64_t lo = bd + (adbc << 32);
*hi = emulu((uint32_t)(ab >> 32), (uint32_t)(cd >> 32)) + (adbc >> 32) +
(adbc_carry << 32) + !!(lo < bd);
(adbc_carry << 32) + (uint64_t)(lo < bd);
return lo;
}
@ -281,18 +320,18 @@ uint64_t umul128_generic(uint64_t ab, uint64_t cd, uint64_t *hi) {
// slow emulation routine for 32-bit
#if !defined(__MINGW64__)
fastfloat_really_inline FASTFLOAT_CONSTEXPR14
uint64_t _umul128(uint64_t ab, uint64_t cd, uint64_t *hi) {
fastfloat_really_inline FASTFLOAT_CONSTEXPR14 uint64_t _umul128(uint64_t ab,
uint64_t cd,
uint64_t *hi) {
return umul128_generic(ab, cd, hi);
}
#endif // !__MINGW64__
#endif // FASTFLOAT_32BIT
// compute 64-bit a*b
fastfloat_really_inline FASTFLOAT_CONSTEXPR20
value128 full_multiplication(uint64_t a, uint64_t b) {
fastfloat_really_inline FASTFLOAT_CONSTEXPR20 value128
full_multiplication(uint64_t a, uint64_t b) {
if (cpp20_and_in_constexpr()) {
value128 answer;
answer.low = umul128_generic(a, b, &answer.high);
@ -304,9 +343,10 @@ value128 full_multiplication(uint64_t a, uint64_t b) {
// But MinGW on ARM64 doesn't have native support for 64-bit multiplications
answer.high = __umulh(a, b);
answer.low = a * b;
#elif defined(FASTFLOAT_32BIT) || (defined(_WIN64) && !defined(__clang__))
#elif defined(FASTFLOAT_32BIT) || \
(defined(_WIN64) && !defined(__clang__) && !defined(_M_ARM64))
answer.low = _umul128(a, b, &answer.high); // _umul128 not available on ARM64
#elif defined(FASTFLOAT_64BIT)
#elif defined(FASTFLOAT_64BIT) && defined(__SIZEOF_INT128__)
__uint128_t r = ((__uint128_t)a) * b;
answer.low = uint64_t(r);
answer.high = uint64_t(r >> 64);
@ -334,22 +374,24 @@ constexpr static int32_t invalid_am_bias = -0x8000;
// used for binary_format_lookup_tables<T>::max_mantissa
constexpr uint64_t constant_55555 = 5 * 5 * 5 * 5 * 5;
template <typename T, typename U = void>
struct binary_format_lookup_tables;
template <typename T, typename U = void> struct binary_format_lookup_tables;
template <typename T> struct binary_format : binary_format_lookup_tables<T> {
using equiv_uint = typename std::conditional<sizeof(T) == 4, uint32_t, uint64_t>::type;
using equiv_uint =
typename std::conditional<sizeof(T) == 4, uint32_t, uint64_t>::type;
static inline constexpr int mantissa_explicit_bits();
static inline constexpr int minimum_exponent();
static inline constexpr int infinite_power();
static inline constexpr int sign_index();
static inline constexpr int min_exponent_fast_path(); // used when fegetround() == FE_TONEAREST
static inline constexpr int
min_exponent_fast_path(); // used when fegetround() == FE_TONEAREST
static inline constexpr int max_exponent_fast_path();
static inline constexpr int max_exponent_round_to_even();
static inline constexpr int min_exponent_round_to_even();
static inline constexpr uint64_t max_mantissa_fast_path(int64_t power);
static inline constexpr uint64_t max_mantissa_fast_path(); // used when fegetround() == FE_TONEAREST
static inline constexpr uint64_t
max_mantissa_fast_path(); // used when fegetround() == FE_TONEAREST
static inline constexpr int largest_power_of_ten();
static inline constexpr int smallest_power_of_ten();
static inline constexpr T exact_power_of_ten(int64_t power);
@ -359,76 +401,91 @@ template <typename T> struct binary_format : binary_format_lookup_tables<T> {
static inline constexpr equiv_uint hidden_bit_mask();
};
template <typename U>
struct binary_format_lookup_tables<double, U> {
template <typename U> struct binary_format_lookup_tables<double, U> {
static constexpr double powers_of_ten[] = {
1e0, 1e1, 1e2, 1e3, 1e4, 1e5, 1e6, 1e7, 1e8, 1e9, 1e10, 1e11,
1e12, 1e13, 1e14, 1e15, 1e16, 1e17, 1e18, 1e19, 1e20, 1e21, 1e22};
// Largest integer value v so that (5**index * v) <= 1<<53.
// 0x10000000000000 == 1 << 53
// 0x20000000000000 == 1 << 53
static constexpr uint64_t max_mantissa[] = {
0x10000000000000,
0x10000000000000 / 5,
0x10000000000000 / (5 * 5),
0x10000000000000 / (5 * 5 * 5),
0x10000000000000 / (5 * 5 * 5 * 5),
0x10000000000000 / (constant_55555),
0x10000000000000 / (constant_55555 * 5),
0x10000000000000 / (constant_55555 * 5 * 5),
0x10000000000000 / (constant_55555 * 5 * 5 * 5),
0x10000000000000 / (constant_55555 * 5 * 5 * 5 * 5),
0x10000000000000 / (constant_55555 * constant_55555),
0x10000000000000 / (constant_55555 * constant_55555 * 5),
0x10000000000000 / (constant_55555 * constant_55555 * 5 * 5),
0x10000000000000 / (constant_55555 * constant_55555 * 5 * 5 * 5),
0x10000000000000 / (constant_55555 * constant_55555 * constant_55555),
0x10000000000000 / (constant_55555 * constant_55555 * constant_55555 * 5),
0x10000000000000 / (constant_55555 * constant_55555 * constant_55555 * 5 * 5),
0x10000000000000 / (constant_55555 * constant_55555 * constant_55555 * 5 * 5 * 5),
0x10000000000000 / (constant_55555 * constant_55555 * constant_55555 * 5 * 5 * 5 * 5),
0x10000000000000 / (constant_55555 * constant_55555 * constant_55555 * constant_55555),
0x10000000000000 / (constant_55555 * constant_55555 * constant_55555 * constant_55555 * 5),
0x10000000000000 / (constant_55555 * constant_55555 * constant_55555 * constant_55555 * 5 * 5),
0x10000000000000 / (constant_55555 * constant_55555 * constant_55555 * constant_55555 * 5 * 5 * 5),
0x10000000000000 / (constant_55555 * constant_55555 * constant_55555 * constant_55555 * 5 * 5 * 5 * 5)};
0x20000000000000,
0x20000000000000 / 5,
0x20000000000000 / (5 * 5),
0x20000000000000 / (5 * 5 * 5),
0x20000000000000 / (5 * 5 * 5 * 5),
0x20000000000000 / (constant_55555),
0x20000000000000 / (constant_55555 * 5),
0x20000000000000 / (constant_55555 * 5 * 5),
0x20000000000000 / (constant_55555 * 5 * 5 * 5),
0x20000000000000 / (constant_55555 * 5 * 5 * 5 * 5),
0x20000000000000 / (constant_55555 * constant_55555),
0x20000000000000 / (constant_55555 * constant_55555 * 5),
0x20000000000000 / (constant_55555 * constant_55555 * 5 * 5),
0x20000000000000 / (constant_55555 * constant_55555 * 5 * 5 * 5),
0x20000000000000 / (constant_55555 * constant_55555 * constant_55555),
0x20000000000000 / (constant_55555 * constant_55555 * constant_55555 * 5),
0x20000000000000 /
(constant_55555 * constant_55555 * constant_55555 * 5 * 5),
0x20000000000000 /
(constant_55555 * constant_55555 * constant_55555 * 5 * 5 * 5),
0x20000000000000 /
(constant_55555 * constant_55555 * constant_55555 * 5 * 5 * 5 * 5),
0x20000000000000 /
(constant_55555 * constant_55555 * constant_55555 * constant_55555),
0x20000000000000 / (constant_55555 * constant_55555 * constant_55555 *
constant_55555 * 5),
0x20000000000000 / (constant_55555 * constant_55555 * constant_55555 *
constant_55555 * 5 * 5),
0x20000000000000 / (constant_55555 * constant_55555 * constant_55555 *
constant_55555 * 5 * 5 * 5),
0x20000000000000 / (constant_55555 * constant_55555 * constant_55555 *
constant_55555 * 5 * 5 * 5 * 5)};
};
#if FASTFLOAT_DETAIL_MUST_DEFINE_CONSTEXPR_VARIABLE
template <typename U>
constexpr double binary_format_lookup_tables<double, U>::powers_of_ten[];
template <typename U>
constexpr uint64_t binary_format_lookup_tables<double, U>::max_mantissa[];
template <typename U>
struct binary_format_lookup_tables<float, U> {
#endif
template <typename U> struct binary_format_lookup_tables<float, U> {
static constexpr float powers_of_ten[] = {1e0f, 1e1f, 1e2f, 1e3f, 1e4f, 1e5f,
1e6f, 1e7f, 1e8f, 1e9f, 1e10f};
1e6f, 1e7f, 1e8f, 1e9f, 1e10f};
// Largest integer value v so that (5**index * v) <= 1<<24.
// 0x1000000 == 1<<24
static constexpr uint64_t max_mantissa[] = {
0x1000000,
0x1000000 / 5,
0x1000000 / (5 * 5),
0x1000000 / (5 * 5 * 5),
0x1000000 / (5 * 5 * 5 * 5),
0x1000000 / (constant_55555),
0x1000000 / (constant_55555 * 5),
0x1000000 / (constant_55555 * 5 * 5),
0x1000000 / (constant_55555 * 5 * 5 * 5),
0x1000000 / (constant_55555 * 5 * 5 * 5 * 5),
0x1000000 / (constant_55555 * constant_55555),
0x1000000 / (constant_55555 * constant_55555 * 5)};
0x1000000,
0x1000000 / 5,
0x1000000 / (5 * 5),
0x1000000 / (5 * 5 * 5),
0x1000000 / (5 * 5 * 5 * 5),
0x1000000 / (constant_55555),
0x1000000 / (constant_55555 * 5),
0x1000000 / (constant_55555 * 5 * 5),
0x1000000 / (constant_55555 * 5 * 5 * 5),
0x1000000 / (constant_55555 * 5 * 5 * 5 * 5),
0x1000000 / (constant_55555 * constant_55555),
0x1000000 / (constant_55555 * constant_55555 * 5)};
};
#if FASTFLOAT_DETAIL_MUST_DEFINE_CONSTEXPR_VARIABLE
template <typename U>
constexpr float binary_format_lookup_tables<float, U>::powers_of_ten[];
template <typename U>
constexpr uint64_t binary_format_lookup_tables<float, U>::max_mantissa[];
template <> inline constexpr int binary_format<double>::min_exponent_fast_path() {
#endif
template <>
inline constexpr int binary_format<double>::min_exponent_fast_path() {
#if (FLT_EVAL_METHOD != 1) && (FLT_EVAL_METHOD != 0)
return 0;
#else
@ -436,7 +493,8 @@ template <> inline constexpr int binary_format<double>::min_exponent_fast_path()
#endif
}
template <> inline constexpr int binary_format<float>::min_exponent_fast_path() {
template <>
inline constexpr int binary_format<float>::min_exponent_fast_path() {
#if (FLT_EVAL_METHOD != 1) && (FLT_EVAL_METHOD != 0)
return 0;
#else
@ -444,26 +502,32 @@ template <> inline constexpr int binary_format<float>::min_exponent_fast_path()
#endif
}
template <> inline constexpr int binary_format<double>::mantissa_explicit_bits() {
template <>
inline constexpr int binary_format<double>::mantissa_explicit_bits() {
return 52;
}
template <> inline constexpr int binary_format<float>::mantissa_explicit_bits() {
template <>
inline constexpr int binary_format<float>::mantissa_explicit_bits() {
return 23;
}
template <> inline constexpr int binary_format<double>::max_exponent_round_to_even() {
template <>
inline constexpr int binary_format<double>::max_exponent_round_to_even() {
return 23;
}
template <> inline constexpr int binary_format<float>::max_exponent_round_to_even() {
template <>
inline constexpr int binary_format<float>::max_exponent_round_to_even() {
return 10;
}
template <> inline constexpr int binary_format<double>::min_exponent_round_to_even() {
template <>
inline constexpr int binary_format<double>::min_exponent_round_to_even() {
return -4;
}
template <> inline constexpr int binary_format<float>::min_exponent_round_to_even() {
template <>
inline constexpr int binary_format<float>::min_exponent_round_to_even() {
return -17;
}
@ -481,30 +545,42 @@ template <> inline constexpr int binary_format<float>::infinite_power() {
return 0xFF;
}
template <> inline constexpr int binary_format<double>::sign_index() { return 63; }
template <> inline constexpr int binary_format<float>::sign_index() { return 31; }
template <> inline constexpr int binary_format<double>::sign_index() {
return 63;
}
template <> inline constexpr int binary_format<float>::sign_index() {
return 31;
}
template <> inline constexpr int binary_format<double>::max_exponent_fast_path() {
template <>
inline constexpr int binary_format<double>::max_exponent_fast_path() {
return 22;
}
template <> inline constexpr int binary_format<float>::max_exponent_fast_path() {
template <>
inline constexpr int binary_format<float>::max_exponent_fast_path() {
return 10;
}
template <> inline constexpr uint64_t binary_format<double>::max_mantissa_fast_path() {
template <>
inline constexpr uint64_t binary_format<double>::max_mantissa_fast_path() {
return uint64_t(2) << mantissa_explicit_bits();
}
template <> inline constexpr uint64_t binary_format<double>::max_mantissa_fast_path(int64_t power) {
template <>
inline constexpr uint64_t
binary_format<double>::max_mantissa_fast_path(int64_t power) {
// caller is responsible to ensure that
// power >= 0 && power <= 22
//
// Work around clang bug https://godbolt.org/z/zedh7rrhc
return (void)max_mantissa[0], max_mantissa[power];
}
template <> inline constexpr uint64_t binary_format<float>::max_mantissa_fast_path() {
template <>
inline constexpr uint64_t binary_format<float>::max_mantissa_fast_path() {
return uint64_t(2) << mantissa_explicit_bits();
}
template <> inline constexpr uint64_t binary_format<float>::max_mantissa_fast_path(int64_t power) {
template <>
inline constexpr uint64_t
binary_format<float>::max_mantissa_fast_path(int64_t power) {
// caller is responsible to ensure that
// power >= 0 && power <= 10
//
@ -513,7 +589,8 @@ template <> inline constexpr uint64_t binary_format<float>::max_mantissa_fast_pa
}
template <>
inline constexpr double binary_format<double>::exact_power_of_ten(int64_t power) {
inline constexpr double
binary_format<double>::exact_power_of_ten(int64_t power) {
// Work around clang bug https://godbolt.org/z/zedh7rrhc
return (void)powers_of_ten[0], powers_of_ten[power];
}
@ -523,13 +600,10 @@ inline constexpr float binary_format<float>::exact_power_of_ten(int64_t power) {
return (void)powers_of_ten[0], powers_of_ten[power];
}
template <>
inline constexpr int binary_format<double>::largest_power_of_ten() {
template <> inline constexpr int binary_format<double>::largest_power_of_ten() {
return 308;
}
template <>
inline constexpr int binary_format<float>::largest_power_of_ten() {
template <> inline constexpr int binary_format<float>::largest_power_of_ten() {
return 38;
}
@ -537,9 +611,8 @@ template <>
inline constexpr int binary_format<double>::smallest_power_of_ten() {
return -342;
}
template <>
inline constexpr int binary_format<float>::smallest_power_of_ten() {
return -65;
template <> inline constexpr int binary_format<float>::smallest_power_of_ten() {
return -64;
}
template <> inline constexpr size_t binary_format<double>::max_digits() {
@ -549,39 +622,46 @@ template <> inline constexpr size_t binary_format<float>::max_digits() {
return 114;
}
template <> inline constexpr binary_format<float>::equiv_uint
binary_format<float>::exponent_mask() {
template <>
inline constexpr binary_format<float>::equiv_uint
binary_format<float>::exponent_mask() {
return 0x7F800000;
}
template <> inline constexpr binary_format<double>::equiv_uint
binary_format<double>::exponent_mask() {
template <>
inline constexpr binary_format<double>::equiv_uint
binary_format<double>::exponent_mask() {
return 0x7FF0000000000000;
}
template <> inline constexpr binary_format<float>::equiv_uint
binary_format<float>::mantissa_mask() {
template <>
inline constexpr binary_format<float>::equiv_uint
binary_format<float>::mantissa_mask() {
return 0x007FFFFF;
}
template <> inline constexpr binary_format<double>::equiv_uint
binary_format<double>::mantissa_mask() {
template <>
inline constexpr binary_format<double>::equiv_uint
binary_format<double>::mantissa_mask() {
return 0x000FFFFFFFFFFFFF;
}
template <> inline constexpr binary_format<float>::equiv_uint
binary_format<float>::hidden_bit_mask() {
template <>
inline constexpr binary_format<float>::equiv_uint
binary_format<float>::hidden_bit_mask() {
return 0x00800000;
}
template <> inline constexpr binary_format<double>::equiv_uint
binary_format<double>::hidden_bit_mask() {
template <>
inline constexpr binary_format<double>::equiv_uint
binary_format<double>::hidden_bit_mask() {
return 0x0010000000000000;
}
template<typename T>
fastfloat_really_inline FASTFLOAT_CONSTEXPR20
void to_float(bool negative, adjusted_mantissa am, T &value) {
template <typename T>
fastfloat_really_inline FASTFLOAT_CONSTEXPR20 void
to_float(bool negative, adjusted_mantissa am, T &value) {
using fastfloat_uint = typename binary_format<T>::equiv_uint;
fastfloat_uint word = (fastfloat_uint)am.mantissa;
word |= fastfloat_uint(am.power2) << binary_format<T>::mantissa_explicit_bits();
word |= fastfloat_uint(am.power2)
<< binary_format<T>::mantissa_explicit_bits();
word |= fastfloat_uint(negative) << binary_format<T>::sign_index();
#if FASTFLOAT_HAS_BIT_CAST
value = std::bit_cast<T>(word);
@ -591,89 +671,132 @@ void to_float(bool negative, adjusted_mantissa am, T &value) {
}
#ifdef FASTFLOAT_SKIP_WHITE_SPACE // disabled by default
template <typename = void>
struct space_lut {
template <typename = void> struct space_lut {
static constexpr bool value[] = {
0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 1, 1, 1, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0};
0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 1, 1, 1, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0};
};
template <typename T>
constexpr bool space_lut<T>::value[];
#if FASTFLOAT_DETAIL_MUST_DEFINE_CONSTEXPR_VARIABLE
template <typename T> constexpr bool space_lut<T>::value[];
#endif
inline constexpr bool is_space(uint8_t c) { return space_lut<>::value[c]; }
#endif
template<typename UC>
static constexpr uint64_t int_cmp_zeros()
{
static_assert((sizeof(UC) == 1) || (sizeof(UC) == 2) || (sizeof(UC) == 4), "Unsupported character size");
return (sizeof(UC) == 1) ? 0x3030303030303030 : (sizeof(UC) == 2) ? (uint64_t(UC('0')) << 48 | uint64_t(UC('0')) << 32 | uint64_t(UC('0')) << 16 | UC('0')) : (uint64_t(UC('0')) << 32 | UC('0'));
template <typename UC> static constexpr uint64_t int_cmp_zeros() {
static_assert((sizeof(UC) == 1) || (sizeof(UC) == 2) || (sizeof(UC) == 4),
"Unsupported character size");
return (sizeof(UC) == 1) ? 0x3030303030303030
: (sizeof(UC) == 2)
? (uint64_t(UC('0')) << 48 | uint64_t(UC('0')) << 32 |
uint64_t(UC('0')) << 16 | UC('0'))
: (uint64_t(UC('0')) << 32 | UC('0'));
}
template<typename UC>
static constexpr int int_cmp_len()
{
return sizeof(uint64_t) / sizeof(UC);
template <typename UC> static constexpr int int_cmp_len() {
return sizeof(uint64_t) / sizeof(UC);
}
template<typename UC>
static constexpr UC const * str_const_nan()
{
return nullptr;
template <typename UC> static constexpr UC const *str_const_nan() {
return nullptr;
}
template<>
constexpr char const * str_const_nan<char>()
{
return "nan";
template <> constexpr char const *str_const_nan<char>() { return "nan"; }
template <> constexpr wchar_t const *str_const_nan<wchar_t>() { return L"nan"; }
template <> constexpr char16_t const *str_const_nan<char16_t>() {
return u"nan";
}
template<>
constexpr wchar_t const * str_const_nan<wchar_t>()
{
return L"nan";
template <> constexpr char32_t const *str_const_nan<char32_t>() {
return U"nan";
}
template<>
constexpr char16_t const * str_const_nan<char16_t>()
{
return u"nan";
template <typename UC> static constexpr UC const *str_const_inf() {
return nullptr;
}
template<>
constexpr char32_t const * str_const_nan<char32_t>()
{
return U"nan";
template <> constexpr char const *str_const_inf<char>() { return "infinity"; }
template <> constexpr wchar_t const *str_const_inf<wchar_t>() {
return L"infinity";
}
template<typename UC>
static constexpr UC const * str_const_inf()
{
return nullptr;
template <> constexpr char16_t const *str_const_inf<char16_t>() {
return u"infinity";
}
template<>
constexpr char const * str_const_inf<char>()
{
return "infinity";
template <> constexpr char32_t const *str_const_inf<char32_t>() {
return U"infinity";
}
template<>
constexpr wchar_t const * str_const_inf<wchar_t>()
{
return L"infinity";
template <typename = void> struct int_luts {
static constexpr uint8_t chdigit[] = {
255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255,
255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255,
255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255,
255, 255, 255, 0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 255, 255,
255, 255, 255, 255, 255, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19,
20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31, 32, 33, 34,
35, 255, 255, 255, 255, 255, 255, 10, 11, 12, 13, 14, 15, 16, 17,
18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31, 32,
33, 34, 35, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255,
255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255,
255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255,
255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255,
255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255,
255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255,
255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255,
255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255,
255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255, 255,
255};
static constexpr size_t maxdigits_u64[] = {
64, 41, 32, 28, 25, 23, 22, 21, 20, 19, 18, 18, 17, 17, 16, 16, 16, 16,
15, 15, 15, 15, 14, 14, 14, 14, 14, 14, 14, 13, 13, 13, 13, 13, 13};
static constexpr uint64_t min_safe_u64[] = {
9223372036854775808ull, 12157665459056928801ull, 4611686018427387904,
7450580596923828125, 4738381338321616896, 3909821048582988049,
9223372036854775808ull, 12157665459056928801ull, 10000000000000000000ull,
5559917313492231481, 2218611106740436992, 8650415919381337933,
2177953337809371136, 6568408355712890625, 1152921504606846976,
2862423051509815793, 6746640616477458432, 15181127029874798299ull,
1638400000000000000, 3243919932521508681, 6221821273427820544,
11592836324538749809ull, 876488338465357824, 1490116119384765625,
2481152873203736576, 4052555153018976267, 6502111422497947648,
10260628712958602189ull, 15943230000000000000ull, 787662783788549761,
1152921504606846976, 1667889514952984961, 2386420683693101056,
3379220508056640625, 4738381338321616896};
};
#if FASTFLOAT_DETAIL_MUST_DEFINE_CONSTEXPR_VARIABLE
template <typename T> constexpr uint8_t int_luts<T>::chdigit[];
template <typename T> constexpr size_t int_luts<T>::maxdigits_u64[];
template <typename T> constexpr uint64_t int_luts<T>::min_safe_u64[];
#endif
template <typename UC>
fastfloat_really_inline constexpr uint8_t ch_to_digit(UC c) {
return int_luts<>::chdigit[static_cast<unsigned char>(c)];
}
template<>
constexpr char16_t const * str_const_inf<char16_t>()
{
return u"infinity";
fastfloat_really_inline constexpr size_t max_digits_u64(int base) {
return int_luts<>::maxdigits_u64[base - 2];
}
template<>
constexpr char32_t const * str_const_inf<char32_t>()
{
return U"infinity";
// If a u64 is exactly max_digits_u64() in length, this is
// the value below which it has definitely overflowed.
fastfloat_really_inline constexpr uint64_t min_safe_u64(int base) {
return int_luts<>::min_safe_u64[base - 2];
}
} // namespace fast_float
#endif

357
third_party/fast_float/parse_number.h vendored
View File

@ -10,10 +10,8 @@
#include <cstring>
#include <limits>
#include <system_error>
namespace fast_float {
namespace detail {
/**
* Special case +inf, -inf, nan, infinity, -infinity.
@ -21,45 +19,53 @@ namespace detail {
* strings a null-free and fixed.
**/
template <typename T, typename UC>
from_chars_result_t<UC> FASTFLOAT_CONSTEXPR14
parse_infnan(UC const * first, UC const * last, T &value) noexcept {
from_chars_result_t<UC> FASTFLOAT_CONSTEXPR14 parse_infnan(UC const *first,
UC const *last,
T &value) noexcept {
from_chars_result_t<UC> answer{};
answer.ptr = first;
answer.ec = std::errc(); // be optimistic
bool minusSign = false;
if (*first == UC('-')) { // assume first < last, so dereference without checks; C++17 20.19.3.(7.1) explicitly forbids '+' here
minusSign = true;
++first;
if (*first ==
UC('-')) { // assume first < last, so dereference without checks;
// C++17 20.19.3.(7.1) explicitly forbids '+' here
minusSign = true;
++first;
}
#ifdef FASTFLOAT_ALLOWS_LEADING_PLUS // disabled by default
if (*first == UC('+')) {
++first;
++first;
}
#endif
if (last - first >= 3) {
if (fastfloat_strncasecmp(first, str_const_nan<UC>(), 3)) {
answer.ptr = (first += 3);
value = minusSign ? -std::numeric_limits<T>::quiet_NaN() : std::numeric_limits<T>::quiet_NaN();
// Check for possible nan(n-char-seq-opt), C++17 20.19.3.7, C11 7.20.1.3.3. At least MSVC produces nan(ind) and nan(snan).
if(first != last && *first == UC('(')) {
for(UC const * ptr = first + 1; ptr != last; ++ptr) {
value = minusSign ? -std::numeric_limits<T>::quiet_NaN()
: std::numeric_limits<T>::quiet_NaN();
// Check for possible nan(n-char-seq-opt), C++17 20.19.3.7,
// C11 7.20.1.3.3. At least MSVC produces nan(ind) and nan(snan).
if (first != last && *first == UC('(')) {
for (UC const *ptr = first + 1; ptr != last; ++ptr) {
if (*ptr == UC(')')) {
answer.ptr = ptr + 1; // valid nan(n-char-seq-opt)
break;
}
else if(!((UC('a') <= *ptr && *ptr <= UC('z')) || (UC('A') <= *ptr && *ptr <= UC('Z')) || (UC('0') <= *ptr && *ptr <= UC('9')) || *ptr == UC('_')))
} else if (!((UC('a') <= *ptr && *ptr <= UC('z')) ||
(UC('A') <= *ptr && *ptr <= UC('Z')) ||
(UC('0') <= *ptr && *ptr <= UC('9')) || *ptr == UC('_')))
break; // forbidden char, not nan(n-char-seq-opt)
}
}
return answer;
}
if (fastfloat_strncasecmp(first, str_const_inf<UC>(), 3)) {
if ((last - first >= 8) && fastfloat_strncasecmp(first + 3, str_const_inf<UC>() + 3, 5)) {
if ((last - first >= 8) &&
fastfloat_strncasecmp(first + 3, str_const_inf<UC>() + 3, 5)) {
answer.ptr = first + 8;
} else {
answer.ptr = first + 3;
}
value = minusSign ? -std::numeric_limits<T>::infinity() : std::numeric_limits<T>::infinity();
value = minusSign ? -std::numeric_limits<T>::infinity()
: std::numeric_limits<T>::infinity();
return answer;
}
}
@ -89,70 +95,207 @@ fastfloat_really_inline bool rounds_to_nearest() noexcept {
//
// The volatile keywoard prevents the compiler from computing the function
// at compile-time.
// There might be other ways to prevent compile-time optimizations (e.g., asm).
// The value does not need to be std::numeric_limits<float>::min(), any small
// value so that 1 + x should round to 1 would do (after accounting for excess
// precision, as in 387 instructions).
// There might be other ways to prevent compile-time optimizations (e.g.,
// asm). The value does not need to be std::numeric_limits<float>::min(), any
// small value so that 1 + x should round to 1 would do (after accounting for
// excess precision, as in 387 instructions).
static volatile float fmin = std::numeric_limits<float>::min();
float fmini = fmin; // we copy it so that it gets loaded at most once.
//
// Explanation:
// Only when fegetround() == FE_TONEAREST do we have that
// fmin + 1.0f == 1.0f - fmin.
//
// FE_UPWARD:
// fmin + 1.0f > 1
// 1.0f - fmin == 1
//
// FE_DOWNWARD or FE_TOWARDZERO:
// fmin + 1.0f == 1
// 1.0f - fmin < 1
//
// Note: This may fail to be accurate if fast-math has been
// enabled, as rounding conventions may not apply.
#ifdef FASTFLOAT_VISUAL_STUDIO
# pragma warning(push)
// todo: is there a VS warning?
// see https://stackoverflow.com/questions/46079446/is-there-a-warning-for-floating-point-equality-checking-in-visual-studio-2013
#elif defined(__clang__)
# pragma clang diagnostic push
# pragma clang diagnostic ignored "-Wfloat-equal"
#elif defined(__GNUC__)
# pragma GCC diagnostic push
# pragma GCC diagnostic ignored "-Wfloat-equal"
#endif
//
// Explanation:
// Only when fegetround() == FE_TONEAREST do we have that
// fmin + 1.0f == 1.0f - fmin.
//
// FE_UPWARD:
// fmin + 1.0f > 1
// 1.0f - fmin == 1
//
// FE_DOWNWARD or FE_TOWARDZERO:
// fmin + 1.0f == 1
// 1.0f - fmin < 1
//
// Note: This may fail to be accurate if fast-math has been
// enabled, as rounding conventions may not apply.
#ifdef FASTFLOAT_VISUAL_STUDIO
#pragma warning(push)
// todo: is there a VS warning?
// see
// https://stackoverflow.com/questions/46079446/is-there-a-warning-for-floating-point-equality-checking-in-visual-studio-2013
#elif defined(__clang__)
#pragma clang diagnostic push
#pragma clang diagnostic ignored "-Wfloat-equal"
#elif defined(__GNUC__)
#pragma GCC diagnostic push
#pragma GCC diagnostic ignored "-Wfloat-equal"
#endif
return (fmini + 1.0f == 1.0f - fmini);
#ifdef FASTFLOAT_VISUAL_STUDIO
# pragma warning(pop)
#elif defined(__clang__)
# pragma clang diagnostic pop
#elif defined(__GNUC__)
# pragma GCC diagnostic pop
#endif
#ifdef FASTFLOAT_VISUAL_STUDIO
#pragma warning(pop)
#elif defined(__clang__)
#pragma clang diagnostic pop
#elif defined(__GNUC__)
#pragma GCC diagnostic pop
#endif
}
} // namespace detail
template<typename T, typename UC>
FASTFLOAT_CONSTEXPR20
from_chars_result_t<UC> from_chars(UC const * first, UC const * last,
T &value, chars_format fmt /*= chars_format::general*/) noexcept {
return from_chars_advanced(first, last, value, parse_options_t<UC>{fmt});
template <typename T> struct from_chars_caller {
template <typename UC>
FASTFLOAT_CONSTEXPR20 static from_chars_result_t<UC>
call(UC const *first, UC const *last, T &value,
parse_options_t<UC> options) noexcept {
return from_chars_advanced(first, last, value, options);
}
};
#if __STDCPP_FLOAT32_T__ == 1
template <> struct from_chars_caller<std::float32_t> {
template <typename UC>
FASTFLOAT_CONSTEXPR20 static from_chars_result_t<UC>
call(UC const *first, UC const *last, std::float32_t &value,
parse_options_t<UC> options) noexcept {
// if std::float32_t is defined, and we are in C++23 mode; macro set for
// float32; set value to float due to equivalence between float and
// float32_t
float val;
auto ret = from_chars_advanced(first, last, val, options);
value = val;
return ret;
}
};
#endif
#if __STDCPP_FLOAT64_T__ == 1
template <> struct from_chars_caller<std::float64_t> {
template <typename UC>
FASTFLOAT_CONSTEXPR20 static from_chars_result_t<UC>
call(UC const *first, UC const *last, std::float64_t &value,
parse_options_t<UC> options) noexcept {
// if std::float64_t is defined, and we are in C++23 mode; macro set for
// float64; set value as double due to equivalence between double and
// float64_t
double val;
auto ret = from_chars_advanced(first, last, val, options);
value = val;
return ret;
}
};
#endif
template <typename T, typename UC, typename>
FASTFLOAT_CONSTEXPR20 from_chars_result_t<UC>
from_chars(UC const *first, UC const *last, T &value,
chars_format fmt /*= chars_format::general*/) noexcept {
return from_chars_caller<T>::call(first, last, value,
parse_options_t<UC>(fmt));
}
template<typename T, typename UC>
FASTFLOAT_CONSTEXPR20
from_chars_result_t<UC> from_chars_advanced(UC const * first, UC const * last,
T &value, parse_options_t<UC> options) noexcept {
/**
* This function overload takes parsed_number_string_t structure that is created
* and populated either by from_chars_advanced function taking chars range and
* parsing options or other parsing custom function implemented by user.
*/
template <typename T, typename UC>
FASTFLOAT_CONSTEXPR20 from_chars_result_t<UC>
from_chars_advanced(parsed_number_string_t<UC> &pns, T &value) noexcept {
static_assert (std::is_same<T, double>::value || std::is_same<T, float>::value, "only float and double are supported");
static_assert (std::is_same<UC, char>::value ||
std::is_same<UC, wchar_t>::value ||
std::is_same<UC, char16_t>::value ||
std::is_same<UC, char32_t>::value , "only char, wchar_t, char16_t and char32_t are supported");
static_assert(is_supported_float_type<T>(),
"only some floating-point types are supported");
static_assert(is_supported_char_type<UC>(),
"only char, wchar_t, char16_t and char32_t are supported");
from_chars_result_t<UC> answer;
#ifdef FASTFLOAT_SKIP_WHITE_SPACE // disabled by default
answer.ec = std::errc(); // be optimistic
answer.ptr = pns.lastmatch;
// The implementation of the Clinger's fast path is convoluted because
// we want round-to-nearest in all cases, irrespective of the rounding mode
// selected on the thread.
// We proceed optimistically, assuming that detail::rounds_to_nearest()
// returns true.
if (binary_format<T>::min_exponent_fast_path() <= pns.exponent &&
pns.exponent <= binary_format<T>::max_exponent_fast_path() &&
!pns.too_many_digits) {
// Unfortunately, the conventional Clinger's fast path is only possible
// when the system rounds to the nearest float.
//
// We expect the next branch to almost always be selected.
// We could check it first (before the previous branch), but
// there might be performance advantages at having the check
// be last.
if (!cpp20_and_in_constexpr() && detail::rounds_to_nearest()) {
// We have that fegetround() == FE_TONEAREST.
// Next is Clinger's fast path.
if (pns.mantissa <= binary_format<T>::max_mantissa_fast_path()) {
value = T(pns.mantissa);
if (pns.exponent < 0) {
value = value / binary_format<T>::exact_power_of_ten(-pns.exponent);
} else {
value = value * binary_format<T>::exact_power_of_ten(pns.exponent);
}
if (pns.negative) {
value = -value;
}
return answer;
}
} else {
// We do not have that fegetround() == FE_TONEAREST.
// Next is a modified Clinger's fast path, inspired by Jakub Jelínek's
// proposal
if (pns.exponent >= 0 &&
pns.mantissa <=
binary_format<T>::max_mantissa_fast_path(pns.exponent)) {
#if defined(__clang__) || defined(FASTFLOAT_32BIT)
// Clang may map 0 to -0.0 when fegetround() == FE_DOWNWARD
if (pns.mantissa == 0) {
value = pns.negative ? T(-0.) : T(0.);
return answer;
}
#endif
value = T(pns.mantissa) *
binary_format<T>::exact_power_of_ten(pns.exponent);
if (pns.negative) {
value = -value;
}
return answer;
}
}
}
adjusted_mantissa am =
compute_float<binary_format<T>>(pns.exponent, pns.mantissa);
if (pns.too_many_digits && am.power2 >= 0) {
if (am != compute_float<binary_format<T>>(pns.exponent, pns.mantissa + 1)) {
am = compute_error<binary_format<T>>(pns.exponent, pns.mantissa);
}
}
// If we called compute_float<binary_format<T>>(pns.exponent, pns.mantissa)
// and we have an invalid power (am.power2 < 0), then we need to go the long
// way around again. This is very uncommon.
if (am.power2 < 0) {
am = digit_comp<T>(pns, am);
}
to_float(pns.negative, am, value);
// Test for over/underflow.
if ((pns.mantissa != 0 && am.mantissa == 0 && am.power2 == 0) ||
am.power2 == binary_format<T>::infinite_power()) {
answer.ec = std::errc::result_out_of_range;
}
return answer;
}
template <typename T, typename UC>
FASTFLOAT_CONSTEXPR20 from_chars_result_t<UC>
from_chars_advanced(UC const *first, UC const *last, T &value,
parse_options_t<UC> options) noexcept {
static_assert(is_supported_float_type<T>(),
"only some floating-point types are supported");
static_assert(is_supported_char_type<UC>(),
"only char, wchar_t, char16_t and char32_t are supported");
from_chars_result_t<UC> answer;
#ifdef FASTFLOAT_SKIP_WHITE_SPACE // disabled by default
while ((first != last) && fast_float::is_space(uint8_t(*first))) {
first++;
}
@ -162,7 +305,8 @@ from_chars_result_t<UC> from_chars_advanced(UC const * first, UC const * last,
answer.ptr = first;
return answer;
}
parsed_number_string_t<UC> pns = parse_number_string<UC>(first, last, options);
parsed_number_string_t<UC> pns =
parse_number_string<UC>(first, last, options);
if (!pns.valid) {
if (options.format & chars_format::no_infnan) {
answer.ec = std::errc::invalid_argument;
@ -173,63 +317,28 @@ from_chars_result_t<UC> from_chars_advanced(UC const * first, UC const * last,
}
}
answer.ec = std::errc(); // be optimistic
answer.ptr = pns.lastmatch;
// The implementation of the Clinger's fast path is convoluted because
// we want round-to-nearest in all cases, irrespective of the rounding mode
// selected on the thread.
// We proceed optimistically, assuming that detail::rounds_to_nearest() returns
// true.
if (binary_format<T>::min_exponent_fast_path() <= pns.exponent && pns.exponent <= binary_format<T>::max_exponent_fast_path() && !pns.too_many_digits) {
// Unfortunately, the conventional Clinger's fast path is only possible
// when the system rounds to the nearest float.
//
// We expect the next branch to almost always be selected.
// We could check it first (before the previous branch), but
// there might be performance advantages at having the check
// be last.
if(!cpp20_and_in_constexpr() && detail::rounds_to_nearest()) {
// We have that fegetround() == FE_TONEAREST.
// Next is Clinger's fast path.
if (pns.mantissa <=binary_format<T>::max_mantissa_fast_path()) {
value = T(pns.mantissa);
if (pns.exponent < 0) { value = value / binary_format<T>::exact_power_of_ten(-pns.exponent); }
else { value = value * binary_format<T>::exact_power_of_ten(pns.exponent); }
if (pns.negative) { value = -value; }
return answer;
}
} else {
// We do not have that fegetround() == FE_TONEAREST.
// Next is a modified Clinger's fast path, inspired by Jakub Jelínek's proposal
if (pns.exponent >= 0 && pns.mantissa <=binary_format<T>::max_mantissa_fast_path(pns.exponent)) {
#if defined(__clang__)
// Clang may map 0 to -0.0 when fegetround() == FE_DOWNWARD
if(pns.mantissa == 0) {
value = pns.negative ? -0. : 0.;
return answer;
}
// call overload that takes parsed_number_string_t directly.
return from_chars_advanced(pns, value);
}
template <typename T, typename UC, typename>
FASTFLOAT_CONSTEXPR20 from_chars_result_t<UC>
from_chars(UC const *first, UC const *last, T &value, int base) noexcept {
static_assert(is_supported_char_type<UC>(),
"only char, wchar_t, char16_t and char32_t are supported");
from_chars_result_t<UC> answer;
#ifdef FASTFLOAT_SKIP_WHITE_SPACE // disabled by default
while ((first != last) && fast_float::is_space(uint8_t(*first))) {
first++;
}
#endif
value = T(pns.mantissa) * binary_format<T>::exact_power_of_ten(pns.exponent);
if (pns.negative) { value = -value; }
return answer;
}
}
if (first == last || base < 2 || base > 36) {
answer.ec = std::errc::invalid_argument;
answer.ptr = first;
return answer;
}
adjusted_mantissa am = compute_float<binary_format<T>>(pns.exponent, pns.mantissa);
if(pns.too_many_digits && am.power2 >= 0) {
if(am != compute_float<binary_format<T>>(pns.exponent, pns.mantissa + 1)) {
am = compute_error<binary_format<T>>(pns.exponent, pns.mantissa);
}
}
// If we called compute_float<binary_format<T>>(pns.exponent, pns.mantissa) and we have an invalid power (am.power2 < 0),
// then we need to go the long way around again. This is very uncommon.
if(am.power2 < 0) { am = digit_comp<T>(pns, am); }
to_float(pns.negative, am, value);
// Test for over/underflow.
if ((pns.mantissa != 0 && am.mantissa == 0 && am.power2 == 0) || am.power2 == binary_format<T>::infinite_power()) {
answer.ec = std::errc::result_out_of_range;
}
return answer;
return parse_int_string(first, last, value, base);
}
} // namespace fast_float