Line data Source code
1 : // Copyright 2020-2023 Daniel Lemire
2 : // Copyright 2023 Matt Borland
3 : // Distributed under the Boost Software License, Version 1.0.
4 : // https://www.boost.org/LICENSE_1_0.txt
5 : //
6 : // Derivative of: https://github.com/fastfloat/fast_float
7 :
8 : #ifndef BOOST_JSON_DETAIL_CHARCONV_DETAIL_FASTFLOAT_PARSE_NUMBER_HPP
9 : #define BOOST_JSON_DETAIL_CHARCONV_DETAIL_FASTFLOAT_PARSE_NUMBER_HPP
10 :
11 : #include <boost/json/detail/charconv/detail/fast_float/ascii_number.hpp>
12 : #include <boost/json/detail/charconv/detail/fast_float/decimal_to_binary.hpp>
13 : #include <boost/json/detail/charconv/detail/fast_float/digit_comparison.hpp>
14 : #include <boost/json/detail/charconv/detail/fast_float/float_common.hpp>
15 :
16 : #include <cmath>
17 : #include <cstring>
18 : #include <limits>
19 : #include <system_error>
20 :
21 : namespace boost { namespace json { namespace detail { namespace charconv { namespace detail { namespace fast_float {
22 :
23 :
24 : namespace detail {
25 : /**
26 : * Special case +inf, -inf, nan, infinity, -infinity.
27 : * The case comparisons could be made much faster given that we know that the
28 : * strings a null-free and fixed.
29 : **/
30 :
31 : #if defined(__GNUC__) && __GNUC__ < 5 && !defined(__clang__)
32 : # pragma GCC diagnostic push
33 : # pragma GCC diagnostic ignored "-Wmissing-field-initializers"
34 : #endif
35 :
36 : template <typename T, typename UC>
37 : from_chars_result_t<UC> BOOST_JSON_CXX14_CONSTEXPR
38 0 : parse_infnan(UC const * first, UC const * last, T &value) noexcept {
39 0 : from_chars_result_t<UC> answer{};
40 0 : answer.ptr = first;
41 0 : answer.ec = std::errc(); // be optimistic
42 0 : bool minusSign = false;
43 0 : if (*first == UC('-')) { // assume first < last, so dereference without checks; C++17 20.19.3.(7.1) explicitly forbids '+' here
44 0 : minusSign = true;
45 0 : ++first;
46 : }
47 0 : if (last - first >= 3) {
48 0 : if (fastfloat_strncasecmp(first, str_const_nan<UC>(), 3)) {
49 0 : answer.ptr = (first += 3);
50 0 : value = minusSign ? -std::numeric_limits<T>::quiet_NaN() : std::numeric_limits<T>::quiet_NaN();
51 : // 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).
52 0 : if(first != last && *first == UC('(')) {
53 0 : for(UC const * ptr = first + 1; ptr != last; ++ptr) {
54 0 : if (*ptr == UC(')')) {
55 0 : answer.ptr = ptr + 1; // valid nan(n-char-seq-opt)
56 0 : break;
57 : }
58 0 : else if(!((UC('a') <= *ptr && *ptr <= UC('z')) || (UC('A') <= *ptr && *ptr <= UC('Z')) || (UC('0') <= *ptr && *ptr <= UC('9')) || *ptr == UC('_')))
59 0 : break; // forbidden char, not nan(n-char-seq-opt)
60 : }
61 : }
62 0 : return answer;
63 : }
64 0 : if (fastfloat_strncasecmp(first, str_const_inf<UC>(), 3)) {
65 0 : if ((last - first >= 8) && fastfloat_strncasecmp(first + 3, str_const_inf<UC>() + 3, 5)) {
66 0 : answer.ptr = first + 8;
67 : } else {
68 0 : answer.ptr = first + 3;
69 : }
70 0 : value = minusSign ? -std::numeric_limits<T>::infinity() : std::numeric_limits<T>::infinity();
71 0 : return answer;
72 : }
73 : }
74 0 : answer.ec = std::errc::invalid_argument;
75 0 : return answer;
76 : }
77 :
78 : #if defined(__GNUC__) && __GNUC__ < 5 && !defined(__clang__)
79 : # pragma GCC diagnostic pop
80 : #endif
81 :
82 : /**
83 : * Returns true if the floating-pointing rounding mode is to 'nearest'.
84 : * It is the default on most system. This function is meant to be inexpensive.
85 : * Credit : @mwalcott3
86 : */
87 : BOOST_FORCEINLINE bool rounds_to_nearest() noexcept {
88 : // https://lemire.me/blog/2020/06/26/gcc-not-nearest/
89 : #if (FLT_EVAL_METHOD != 1) && (FLT_EVAL_METHOD != 0)
90 : return false;
91 : #endif
92 : // See
93 : // A fast function to check your floating-point rounding mode
94 : // https://lemire.me/blog/2022/11/16/a-fast-function-to-check-your-floating-point-rounding-mode/
95 : //
96 : // This function is meant to be equivalent to :
97 : // prior: #include <cfenv>
98 : // return fegetround() == FE_TONEAREST;
99 : // However, it is expected to be much faster than the fegetround()
100 : // function call.
101 : //
102 : // The volatile keywoard prevents the compiler from computing the function
103 : // at compile-time.
104 : // There might be other ways to prevent compile-time optimizations (e.g., asm).
105 : // The value does not need to be std::numeric_limits<float>::min(), any small
106 : // value so that 1 + x should round to 1 would do (after accounting for excess
107 : // precision, as in 387 instructions).
108 : static volatile float fmin = (std::numeric_limits<float>::min)();
109 5895 : float fmini = fmin; // we copy it so that it gets loaded at most once.
110 : //
111 : // Explanation:
112 : // Only when fegetround() == FE_TONEAREST do we have that
113 : // fmin + 1.0f == 1.0f - fmin.
114 : //
115 : // FE_UPWARD:
116 : // fmin + 1.0f > 1
117 : // 1.0f - fmin == 1
118 : //
119 : // FE_DOWNWARD or FE_TOWARDZERO:
120 : // fmin + 1.0f == 1
121 : // 1.0f - fmin < 1
122 : //
123 : // Note: This may fail to be accurate if fast-math has been
124 : // enabled, as rounding conventions may not apply.
125 : #ifdef BOOST_JSON_FASTFLOAT_VISUAL_STUDIO
126 : # pragma warning(push)
127 : // todo: is there a VS warning?
128 : // see https://stackoverflow.com/questions/46079446/is-there-a-warning-for-floating-point-equality-checking-in-visual-studio-2013
129 : #elif defined(__clang__)
130 : # pragma clang diagnostic push
131 : # pragma clang diagnostic ignored "-Wfloat-equal"
132 : #elif defined(__GNUC__)
133 : # pragma GCC diagnostic push
134 : # pragma GCC diagnostic ignored "-Wfloat-equal"
135 : #endif
136 5895 : return (fmini + 1.0f == 1.0f - fmini);
137 : #ifdef BOOST_JSON_FASTFLOAT_VISUAL_STUDIO
138 : # pragma warning(pop)
139 : #elif defined(__clang__)
140 : # pragma clang diagnostic pop
141 : #elif defined(__GNUC__)
142 : # pragma GCC diagnostic pop
143 : #endif
144 : }
145 :
146 : } // namespace detail
147 :
148 : template<typename T, typename UC>
149 : BOOST_JSON_FASTFLOAT_CONSTEXPR20
150 1009310 : from_chars_result_t<UC> from_chars(UC const * first, UC const * last,
151 : T &value, chars_format fmt /*= chars_format::general*/) noexcept {
152 1009310 : return from_chars_advanced(first, last, value, parse_options_t<UC>{fmt});
153 : }
154 :
155 : template<typename T, typename UC>
156 : BOOST_JSON_FASTFLOAT_CONSTEXPR20
157 1009310 : from_chars_result_t<UC> from_chars_advanced(UC const * first, UC const * last,
158 : T &value, parse_options_t<UC> options) noexcept {
159 :
160 : static_assert (std::is_same<T, double>::value || std::is_same<T, float>::value, "only float and double are supported");
161 : static_assert (std::is_same<UC, char>::value ||
162 : std::is_same<UC, wchar_t>::value ||
163 : std::is_same<UC, char16_t>::value ||
164 : std::is_same<UC, char32_t>::value , "only char, wchar_t, char16_t and char32_t are supported");
165 :
166 : from_chars_result_t<UC> answer;
167 1009310 : if (first == last) {
168 0 : answer.ec = std::errc::invalid_argument;
169 0 : answer.ptr = first;
170 0 : return answer;
171 : }
172 1009310 : parsed_number_string_t<UC> pns = parse_number_string<UC>(first, last, options);
173 1009310 : if (!pns.valid) {
174 0 : return detail::parse_infnan(first, last, value);
175 : }
176 1009310 : answer.ec = std::errc(); // be optimistic
177 1009310 : answer.ptr = pns.lastmatch;
178 : // The implementation of the Clinger's fast path is convoluted because
179 : // we want round-to-nearest in all cases, irrespective of the rounding mode
180 : // selected on the thread.
181 : // We proceed optimistically, assuming that detail::rounds_to_nearest() returns
182 : // true.
183 1009310 : if (binary_format<T>::min_exponent_fast_path() <= pns.exponent && pns.exponent <= binary_format<T>::max_exponent_fast_path() && !pns.too_many_digits) {
184 : // Unfortunately, the conventional Clinger's fast path is only possible
185 : // when the system rounds to the nearest float.
186 : //
187 : // We expect the next branch to almost always be selected.
188 : // We could check it first (before the previous branch), but
189 : // there might be performance advantages at having the check
190 : // be last.
191 11790 : if(!cpp20_and_in_constexpr() && detail::rounds_to_nearest()) {
192 : // We have that fegetround() == FE_TONEAREST.
193 : // Next is Clinger's fast path.
194 5895 : if (pns.mantissa <=binary_format<T>::max_mantissa_fast_path()) {
195 5570 : value = T(pns.mantissa);
196 5570 : if (pns.exponent < 0) { value = value / binary_format<T>::exact_power_of_ten(-pns.exponent); }
197 1756 : else { value = value * binary_format<T>::exact_power_of_ten(pns.exponent); }
198 5570 : if (pns.negative) { value = -value; }
199 5570 : return answer;
200 : }
201 : } else {
202 : // We do not have that fegetround() == FE_TONEAREST.
203 : // Next is a modified Clinger's fast path, inspired by Jakub Jelínek's proposal
204 0 : if (pns.exponent >= 0 && pns.mantissa <=binary_format<T>::max_mantissa_fast_path(pns.exponent)) {
205 : #if defined(__clang__)
206 : // Clang may map 0 to -0.0 when fegetround() == FE_DOWNWARD
207 : if(pns.mantissa == 0) {
208 : value = pns.negative ? -0. : 0.;
209 : return answer;
210 : }
211 : #endif
212 0 : value = T(pns.mantissa) * binary_format<T>::exact_power_of_ten(pns.exponent);
213 0 : if (pns.negative) { value = -value; }
214 0 : return answer;
215 : }
216 : }
217 : }
218 1003740 : adjusted_mantissa am = compute_float<binary_format<T>>(pns.exponent, pns.mantissa);
219 1003740 : if(pns.too_many_digits && am.power2 >= 0) {
220 2001424 : if(am != compute_float<binary_format<T>>(pns.exponent, pns.mantissa + 1)) {
221 5972 : am = compute_error<binary_format<T>>(pns.exponent, pns.mantissa);
222 : }
223 : }
224 : // If we called compute_float<binary_format<T>>(pns.exponent, pns.mantissa) and we have an invalid power (am.power2 < 0),
225 : // then we need to go the long way around again. This is very uncommon.
226 1003740 : if(am.power2 < 0) { am = digit_comp<T>(pns, am); }
227 1003740 : to_float(pns.negative, am, value);
228 : // Test for over/underflow.
229 1003740 : if ((pns.mantissa != 0 && am.mantissa == 0 && am.power2 == 0) || am.power2 == binary_format<T>::infinite_power()) {
230 30608 : answer.ec = std::errc::result_out_of_range;
231 : }
232 1003740 : return answer;
233 : }
234 :
235 : }}}}}} // namespace fast_float
236 :
237 : #endif
|
