]>
Commit | Line | Data |
---|---|---|
932fbc86 JW |
1 | // std::from_chars implementation for floating-point types -*- C++ -*- |
2 | ||
7adcbafe | 3 | // Copyright (C) 2020-2022 Free Software Foundation, Inc. |
932fbc86 JW |
4 | // |
5 | // This file is part of the GNU ISO C++ Library. This library is free | |
6 | // software; you can redistribute it and/or modify it under the | |
7 | // terms of the GNU General Public License as published by the | |
8 | // Free Software Foundation; either version 3, or (at your option) | |
9 | // any later version. | |
10 | ||
11 | // This library is distributed in the hope that it will be useful, | |
12 | // but WITHOUT ANY WARRANTY; without even the implied warranty of | |
13 | // MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the | |
14 | // GNU General Public License for more details. | |
15 | ||
16 | // Under Section 7 of GPL version 3, you are granted additional | |
17 | // permissions described in the GCC Runtime Library Exception, version | |
18 | // 3.1, as published by the Free Software Foundation. | |
19 | ||
20 | // You should have received a copy of the GNU General Public License and | |
21 | // a copy of the GCC Runtime Library Exception along with this program; | |
22 | // see the files COPYING3 and COPYING.RUNTIME respectively. If not, see | |
23 | // <http://www.gnu.org/licenses/>. | |
24 | ||
25 | // | |
26 | // ISO C++ 14882:2017 | |
27 | // 23.2.9 Primitive numeric input conversion [utility.from.chars] | |
28 | // | |
29 | ||
de77abee FD |
30 | // Prefer to use std::pmr::string if possible, which requires the cxx11 ABI. |
31 | #define _GLIBCXX_USE_CXX11_ABI 1 | |
32 | ||
93dd7f36 | 33 | #include <array> |
932fbc86 | 34 | #include <charconv> |
cc3bf340 | 35 | #include <bit> |
932fbc86 JW |
36 | #include <string> |
37 | #include <memory_resource> | |
2251b4a5 | 38 | #include <cfenv> |
490e2303 | 39 | #include <cfloat> |
932fbc86 JW |
40 | #include <cmath> |
41 | #include <cstdlib> | |
42 | #include <cstring> | |
43 | #include <cctype> | |
44 | #include <locale.h> | |
45 | #include <bits/functexcept.h> | |
46 | #if _GLIBCXX_HAVE_XLOCALE_H | |
47 | # include <xlocale.h> | |
48 | #endif | |
49 | ||
416b6fc7 JW |
50 | #if _GLIBCXX_HAVE_USELOCALE |
51 | // FIXME: This should be reimplemented so it doesn't use strtod and newlocale. | |
52 | // That will avoid the need for any memory allocation, meaning that the | |
53 | // non-conforming errc::not_enough_memory result cannot happen. | |
54 | # define USE_STRTOD_FOR_FROM_CHARS 1 | |
55 | #endif | |
56 | ||
7c1e7eed JW |
57 | #ifdef _GLIBCXX_LONG_DOUBLE_ALT128_COMPAT |
58 | #ifndef __LONG_DOUBLE_IBM128__ | |
59 | #error "floating_from_chars.cc must be compiled with -mabi=ibmlongdouble" | |
60 | #endif | |
61 | // strtold for __ieee128 | |
62 | extern "C" __ieee128 __strtoieee128(const char*, char**); | |
63 | #endif | |
64 | ||
a8db9b90 JW |
65 | #if _GLIBCXX_FLOAT_IS_IEEE_BINARY32 && _GLIBCXX_DOUBLE_IS_IEEE_BINARY64 \ |
66 | && __SIZE_WIDTH__ >= 32 | |
490e2303 | 67 | # define USE_LIB_FAST_FLOAT 1 |
416b6fc7 | 68 | # if __LDBL_MANT_DIG__ == __DBL_MANT_DIG__ |
5a4e2080 | 69 | // No need to use strtold. |
416b6fc7 JW |
70 | # undef USE_STRTOD_FOR_FROM_CHARS |
71 | # endif | |
490e2303 PP |
72 | #endif |
73 | ||
74 | #if USE_LIB_FAST_FLOAT | |
75 | # define FASTFLOAT_DEBUG_ASSERT __glibcxx_assert | |
76 | namespace | |
77 | { | |
78 | # include "fast_float/fast_float.h" | |
79 | } // anon namespace | |
80 | #endif | |
81 | ||
932fbc86 JW |
82 | namespace std _GLIBCXX_VISIBILITY(default) |
83 | { | |
84 | _GLIBCXX_BEGIN_NAMESPACE_VERSION | |
85 | ||
86 | namespace | |
87 | { | |
416b6fc7 | 88 | #if USE_STRTOD_FOR_FROM_CHARS |
932fbc86 JW |
89 | // A memory resource with a static buffer that can be used for small |
90 | // allocations. At most one allocation using the freestore can be done | |
91 | // if the static buffer is insufficient. The callers below only require | |
92 | // a single allocation, so there's no need for anything more complex. | |
93 | struct buffer_resource : pmr::memory_resource | |
94 | { | |
95 | ~buffer_resource() { if (m_ptr) operator delete(m_ptr, m_bytes); } | |
96 | ||
97 | void* | |
98 | do_allocate(size_t bytes, size_t alignment [[maybe_unused]]) override | |
99 | { | |
100 | // Allocate from the buffer if it will fit. | |
101 | if (m_bytes < sizeof(m_buf) && (m_bytes + bytes) <= sizeof(m_buf)) | |
102 | return m_buf + std::__exchange(m_bytes, m_bytes + bytes); | |
103 | ||
104 | __glibcxx_assert(m_ptr == nullptr); | |
932fbc86 JW |
105 | |
106 | m_ptr = operator new(bytes); | |
107 | m_bytes = bytes; | |
108 | return m_ptr; | |
109 | } | |
110 | ||
111 | void | |
112 | do_deallocate(void*, size_t, size_t) noexcept override | |
113 | { /* like pmr::monotonic_buffer_resource, do nothing here */ } | |
114 | ||
115 | bool | |
116 | do_is_equal(const pmr::memory_resource& other) const noexcept override | |
117 | { return &other == this; } | |
118 | ||
119 | static constexpr int guaranteed_capacity() { return sizeof(m_buf); } | |
120 | ||
121 | private: | |
122 | char m_buf[512]; | |
123 | size_t m_bytes = 0; | |
124 | void* m_ptr = nullptr; | |
125 | }; | |
126 | ||
de77abee FD |
127 | #if _GLIBCXX_USE_CXX11_ABI |
128 | using buffered_string = std::pmr::string; | |
129 | #else | |
130 | using buffered_string = std::string; | |
131 | #endif | |
132 | ||
932fbc86 JW |
133 | inline bool valid_fmt(chars_format fmt) |
134 | { | |
135 | return fmt != chars_format{} | |
136 | && ((fmt & chars_format::general) == fmt | |
137 | || (fmt & chars_format::hex) == fmt); | |
138 | } | |
139 | ||
140 | constexpr char hex_digits[] = "abcdefABCDEF0123456789"; | |
141 | constexpr auto dec_digits = hex_digits + 12; | |
142 | ||
143 | // Find initial portion of [first, last) containing a floating-point number. | |
144 | // The string `digits` is either `dec_digits` or `hex_digits` | |
145 | // and `exp` is 'e' or 'p' or '\0'. | |
146 | const char* | |
147 | find_end_of_float(const char* first, const char* last, const char* digits, | |
148 | char exp) | |
149 | { | |
150 | while (first < last && strchr(digits, *first) != nullptr) | |
151 | ++first; | |
152 | if (first < last && *first == '.') | |
153 | { | |
154 | ++first; | |
155 | while (first < last && strchr(digits, *first)) | |
156 | ++first; | |
157 | } | |
158 | if (first < last && exp != 0 && std::tolower((unsigned char)*first) == exp) | |
159 | { | |
160 | ++first; | |
161 | if (first < last && (*first == '-' || *first == '+')) | |
162 | ++first; | |
163 | while (first < last && strchr(dec_digits, *first) != nullptr) | |
164 | ++first; | |
165 | } | |
166 | return first; | |
167 | } | |
168 | ||
169 | // Determine the prefix of [first, last) that matches the pattern | |
170 | // corresponding to `fmt`. | |
171 | // Returns a NTBS containing the pattern, using `buf` to allocate | |
172 | // additional storage if needed. | |
173 | // Returns a nullptr if a valid pattern is not present. | |
174 | const char* | |
175 | pattern(const char* const first, const char* last, | |
de77abee | 176 | chars_format& fmt, buffered_string& buf) |
932fbc86 JW |
177 | { |
178 | // fmt has the value of one of the enumerators of chars_format. | |
179 | __glibcxx_assert(valid_fmt(fmt)); | |
180 | ||
181 | string_view res; | |
182 | ||
183 | if (first == last || *first == '+') [[unlikely]] | |
184 | return nullptr; | |
185 | ||
186 | const int neg = (*first == '-'); | |
187 | ||
188 | if (std::memchr("iInN", (unsigned char)first[neg], 4)) | |
189 | { | |
190 | ptrdiff_t len = last - first; | |
191 | if (len < (3 + neg)) | |
192 | return nullptr; | |
193 | ||
194 | // possible infinity or NaN, let strtod decide | |
195 | if (first[neg] == 'i' || first[neg] == 'I') | |
196 | { | |
197 | // Need at most 9 chars for "-INFINITY", ignore anything after it. | |
198 | len = std::min(len, ptrdiff_t(neg + 8)); | |
199 | } | |
200 | else if (len > (neg + 3) && first[neg + 3] == '(') | |
201 | { | |
202 | // Look for end of "NAN(n-char-sequence)" | |
203 | if (void* p = std::memchr(const_cast<char*>(first)+4, ')', len-4)) | |
204 | len = static_cast<char*>(p) + 1 - first; | |
205 | #ifndef __cpp_exceptions | |
206 | if (len > buffer_resource::guaranteed_capacity()) | |
207 | { | |
208 | // The character sequence is too large for the buffer. | |
209 | // Allocation failure could terminate the process, | |
210 | // so just return an error via the fmt parameter. | |
211 | fmt = chars_format{}; | |
212 | return nullptr; | |
213 | } | |
214 | #endif | |
215 | } | |
216 | else // Only need 4 chars for "-NAN" | |
217 | len = neg + 3; | |
218 | ||
219 | buf.assign(first, 0, len); | |
220 | // prevent make_result correcting for "0x" | |
221 | fmt = chars_format::general; | |
222 | return buf.c_str(); | |
223 | } | |
224 | ||
225 | const char* digits; | |
226 | char* ptr; | |
227 | ||
228 | // Assign [first,last) to a std::string to get a NTBS that can be used | |
229 | // with strspn, strtod etc. | |
230 | // If the string would be longer than the fixed buffer inside the | |
231 | // buffer_resource type use find_end_of_float to try to reduce how | |
232 | // much memory is needed, to reduce the chance of std::bad_alloc. | |
233 | ||
234 | if (fmt == chars_format::hex) | |
235 | { | |
236 | digits = hex_digits; | |
237 | ||
238 | if ((last - first + 2) > buffer_resource::guaranteed_capacity()) | |
239 | { | |
240 | last = find_end_of_float(first + neg, last, digits, 'p'); | |
241 | #ifndef __cpp_exceptions | |
242 | if ((last - first + 2) > buffer_resource::guaranteed_capacity()) | |
243 | { | |
244 | // The character sequence is still too large for the buffer. | |
245 | // Allocation failure could terminate the process, | |
246 | // so just return an error via the fmt parameter. | |
247 | fmt = chars_format{}; | |
248 | return nullptr; | |
249 | } | |
250 | #endif | |
251 | } | |
252 | ||
253 | buf = "-0x" + !neg; | |
254 | buf.append(first + neg, last); | |
255 | ptr = buf.data() + neg + 2; | |
256 | } | |
257 | else | |
258 | { | |
259 | digits = dec_digits; | |
260 | ||
261 | if ((last - first) > buffer_resource::guaranteed_capacity()) | |
262 | { | |
263 | last = find_end_of_float(first + neg, last, digits, | |
264 | "e"[fmt == chars_format::fixed]); | |
265 | #ifndef __cpp_exceptions | |
266 | if ((last - first) > buffer_resource::guaranteed_capacity()) | |
267 | { | |
268 | // The character sequence is still too large for the buffer. | |
269 | // Allocation failure could terminate the process, | |
270 | // so just return an error via the fmt parameter. | |
271 | fmt = chars_format{}; | |
272 | return nullptr; | |
273 | } | |
274 | #endif | |
275 | } | |
276 | buf.assign(first, last); | |
277 | ptr = buf.data() + neg; | |
278 | } | |
279 | ||
280 | // "A non-empty sequence of decimal digits" or | |
281 | // "A non-empty sequence of hexadecimal digits" | |
282 | size_t len = std::strspn(ptr, digits); | |
283 | // "possibly containing a radix character," | |
284 | if (ptr[len] == '.') | |
285 | { | |
286 | const size_t len2 = std::strspn(ptr + len + 1, digits); | |
287 | if (len + len2) | |
288 | ptr += len + 1 + len2; | |
289 | else | |
290 | return nullptr; | |
291 | } | |
292 | else if (len == 0) [[unlikely]] | |
293 | return nullptr; | |
294 | else | |
295 | ptr += len; | |
296 | ||
297 | if (fmt == chars_format::fixed) | |
298 | { | |
299 | // Truncate the string to stop strtod parsing past this point. | |
300 | *ptr = '\0'; | |
301 | } | |
302 | else if (fmt == chars_format::scientific) | |
303 | { | |
304 | // Check for required exponent part which starts with 'e' or 'E' | |
305 | if (*ptr != 'e' && *ptr != 'E') | |
306 | return nullptr; | |
307 | // then an optional plus or minus sign | |
308 | const int sign = (ptr[1] == '-' || ptr[1] == '+'); | |
309 | // then a nonempty sequence of decimal digits | |
310 | if (!std::memchr(dec_digits, (unsigned char)ptr[1+sign], 10)) | |
311 | return nullptr; | |
312 | } | |
313 | else if (fmt == chars_format::general) | |
314 | { | |
315 | if (*ptr == 'x' || *ptr == 'X') | |
316 | *ptr = '\0'; | |
317 | } | |
318 | ||
319 | return buf.c_str(); | |
320 | } | |
321 | ||
322 | // Convert the NTBS `str` to a floating-point value of type `T`. | |
323 | // If `str` cannot be converted, `value` is unchanged and `0` is returned. | |
324 | // Otherwise, let N be the number of characters consumed from `str`. | |
325 | // On success `value` is set to the converted value and N is returned. | |
326 | // If the converted value is out of range, `value` is unchanged and | |
327 | // -N is returned. | |
328 | template<typename T> | |
329 | ptrdiff_t | |
330 | from_chars_impl(const char* str, T& value, errc& ec) noexcept | |
331 | { | |
4143efc1 | 332 | if (locale_t loc = ::newlocale(LC_ALL_MASK, "C", (locale_t)0)) [[likely]] |
932fbc86 JW |
333 | { |
334 | locale_t orig = ::uselocale(loc); | |
335 | ||
266d7464 | 336 | #if _GLIBCXX_USE_C99_FENV_TR1 && defined(FE_TONEAREST) |
2251b4a5 JW |
337 | const int rounding = std::fegetround(); |
338 | if (rounding != FE_TONEAREST) | |
339 | std::fesetround(FE_TONEAREST); | |
340 | #endif | |
341 | ||
932fbc86 JW |
342 | const int save_errno = errno; |
343 | errno = 0; | |
344 | char* endptr; | |
345 | T tmpval; | |
e513e9aa | 346 | #if _GLIBCXX_USE_C99_STDLIB |
932fbc86 JW |
347 | if constexpr (is_same_v<T, float>) |
348 | tmpval = std::strtof(str, &endptr); | |
e513e9aa | 349 | else if constexpr (is_same_v<T, double>) |
932fbc86 JW |
350 | tmpval = std::strtod(str, &endptr); |
351 | else if constexpr (is_same_v<T, long double>) | |
352 | tmpval = std::strtold(str, &endptr); | |
7c1e7eed JW |
353 | # ifdef _GLIBCXX_LONG_DOUBLE_ALT128_COMPAT |
354 | else if constexpr (is_same_v<T, __ieee128>) | |
355 | tmpval = __strtoieee128(str, &endptr); | |
356 | # endif | |
e513e9aa JW |
357 | #else |
358 | tmpval = std::strtod(str, &endptr); | |
359 | #endif | |
932fbc86 JW |
360 | const int conv_errno = std::__exchange(errno, save_errno); |
361 | ||
266d7464 | 362 | #if _GLIBCXX_USE_C99_FENV_TR1 && defined(FE_TONEAREST) |
2251b4a5 JW |
363 | if (rounding != FE_TONEAREST) |
364 | std::fesetround(rounding); | |
365 | #endif | |
366 | ||
932fbc86 JW |
367 | ::uselocale(orig); |
368 | ::freelocale(loc); | |
369 | ||
370 | const ptrdiff_t n = endptr - str; | |
371 | if (conv_errno == ERANGE) [[unlikely]] | |
372 | { | |
7c1e7eed | 373 | if (__builtin_isinf(tmpval)) // overflow |
932fbc86 JW |
374 | ec = errc::result_out_of_range; |
375 | else // underflow (LWG 3081 wants to set value = tmpval here) | |
376 | ec = errc::result_out_of_range; | |
377 | } | |
378 | else if (n) | |
379 | { | |
380 | value = tmpval; | |
381 | ec = errc(); | |
382 | } | |
383 | return n; | |
384 | } | |
385 | else if (errno == ENOMEM) | |
386 | ec = errc::not_enough_memory; | |
387 | ||
388 | return 0; | |
389 | } | |
390 | ||
391 | inline from_chars_result | |
392 | make_result(const char* str, ptrdiff_t n, chars_format fmt, errc ec) noexcept | |
393 | { | |
394 | from_chars_result result = { str, ec }; | |
395 | if (n != 0) | |
396 | { | |
397 | if (fmt == chars_format::hex) | |
398 | n -= 2; // correct for the "0x" inserted into the pattern | |
399 | result.ptr += n; | |
400 | } | |
401 | else if (fmt == chars_format{}) [[unlikely]] | |
402 | { | |
403 | // FIXME: the standard does not allow this result. | |
404 | ec = errc::not_enough_memory; | |
405 | } | |
406 | return result; | |
407 | } | |
408 | ||
de77abee FD |
409 | #if ! _GLIBCXX_USE_CXX11_ABI |
410 | inline bool | |
411 | reserve_string(std::string& s) noexcept | |
412 | { | |
413 | __try | |
414 | { | |
415 | s.reserve(buffer_resource::guaranteed_capacity()); | |
416 | } | |
417 | __catch (const std::bad_alloc&) | |
418 | { | |
419 | return false; | |
420 | } | |
421 | return true; | |
422 | } | |
423 | #endif | |
5a4e2080 JW |
424 | |
425 | template<typename T> | |
426 | from_chars_result | |
427 | from_chars_strtod(const char* first, const char* last, T& value, | |
428 | chars_format fmt) noexcept | |
429 | { | |
430 | errc ec = errc::invalid_argument; | |
431 | #if _GLIBCXX_USE_CXX11_ABI | |
432 | buffer_resource mr; | |
433 | pmr::string buf(&mr); | |
434 | #else | |
435 | string buf; | |
436 | if (!reserve_string(buf)) | |
437 | return make_result(first, 0, {}, ec); | |
438 | #endif | |
439 | size_t len = 0; | |
440 | __try | |
441 | { | |
442 | if (const char* pat = pattern(first, last, fmt, buf)) [[likely]] | |
443 | len = from_chars_impl(pat, value, ec); | |
444 | } | |
445 | __catch (const std::bad_alloc&) | |
446 | { | |
447 | fmt = chars_format{}; | |
448 | } | |
449 | return make_result(first, len, fmt, ec); | |
450 | } | |
416b6fc7 | 451 | #endif // USE_STRTOD_FOR_FROM_CHARS |
de77abee | 452 | |
cc3bf340 | 453 | #if _GLIBCXX_FLOAT_IS_IEEE_BINARY32 && _GLIBCXX_DOUBLE_IS_IEEE_BINARY64 |
cc3bf340 | 454 | // Return true iff [FIRST,LAST) begins with PREFIX, ignoring case. |
93dd7f36 | 455 | // PREFIX is assumed to not contain any uppercase letters. |
cc3bf340 PP |
456 | bool |
457 | starts_with_ci(const char* first, const char* last, string_view prefix) | |
458 | { | |
459 | __glibcxx_requires_valid_range(first, last); | |
460 | ||
93dd7f36 PP |
461 | // A lookup table that maps uppercase letters to lowercase and |
462 | // is otherwise the identity mapping. | |
463 | static constexpr auto upper_to_lower_table = [] { | |
464 | constexpr unsigned char lower_letters[27] = "abcdefghijklmnopqrstuvwxyz"; | |
465 | constexpr unsigned char upper_letters[27] = "ABCDEFGHIJKLMNOPQRSTUVWXYZ"; | |
466 | std::array<unsigned char, (1u << __CHAR_BIT__)> table = {}; | |
467 | for (unsigned i = 0; i < table.size(); ++i) | |
468 | table[i] = i; | |
469 | for (unsigned i = 0; i < 26; ++i) | |
470 | table[upper_letters[i]] = lower_letters[i]; | |
471 | return table; | |
472 | }(); | |
473 | ||
474 | if (last - first < static_cast<ptrdiff_t>(prefix.length())) | |
475 | return false; | |
476 | ||
477 | for (const unsigned char pch : prefix) | |
cc3bf340 | 478 | { |
93dd7f36 PP |
479 | // __glibcxx_assert(pch == upper_to_lower_table[pch]); |
480 | const unsigned char ch = *first; | |
481 | if (ch != pch && upper_to_lower_table[ch] != pch) | |
cc3bf340 PP |
482 | return false; |
483 | ++first; | |
484 | } | |
485 | ||
486 | return true; | |
487 | } | |
488 | ||
489 | // An implementation of hexadecimal float parsing for binary32/64. | |
490 | template<typename T> | |
491 | from_chars_result | |
492 | __floating_from_chars_hex(const char* first, const char* last, T& value) | |
493 | { | |
494 | static_assert(is_same_v<T, float> || is_same_v<T, double>); | |
495 | ||
496 | using uint_t = conditional_t<is_same_v<T, float>, uint32_t, uint64_t>; | |
497 | constexpr int mantissa_bits = is_same_v<T, float> ? 23 : 52; | |
498 | constexpr int exponent_bits = is_same_v<T, float> ? 8 : 11; | |
499 | constexpr int exponent_bias = (1 << (exponent_bits - 1)) - 1; | |
500 | ||
501 | __glibcxx_requires_valid_range(first, last); | |
502 | if (first == last) | |
503 | return {first, errc::invalid_argument}; | |
504 | ||
505 | // Consume the sign bit. | |
506 | const char* const orig_first = first; | |
507 | bool sign_bit = false; | |
508 | if (*first == '-') | |
509 | { | |
510 | sign_bit = true; | |
511 | ++first; | |
512 | } | |
513 | ||
514 | // Handle "inf", "infinity", "NaN" and variants thereof. | |
515 | if (first != last) | |
516 | if (*first == 'i' || *first == 'I' || *first == 'n' || *first == 'N') [[unlikely]] | |
517 | { | |
518 | if (starts_with_ci(first, last, "inf"sv)) | |
519 | { | |
520 | first += strlen("inf"); | |
521 | if (starts_with_ci(first, last, "inity"sv)) | |
522 | first += strlen("inity"); | |
523 | ||
524 | uint_t result = 0; | |
525 | result |= sign_bit; | |
526 | result <<= exponent_bits; | |
527 | result |= (1ull << exponent_bits) - 1; | |
528 | result <<= mantissa_bits; | |
529 | memcpy(&value, &result, sizeof(result)); | |
530 | ||
531 | return {first, errc{}}; | |
532 | } | |
533 | else if (starts_with_ci(first, last, "nan")) | |
534 | { | |
535 | first += strlen("nan"); | |
536 | ||
537 | if (first != last && *first == '(') | |
538 | { | |
539 | // Tentatively consume the '(' as we look for an optional | |
540 | // n-char-sequence followed by a ')'. | |
541 | const char* const fallback_first = first; | |
542 | for (;;) | |
543 | { | |
544 | ++first; | |
545 | if (first == last) | |
546 | { | |
547 | first = fallback_first; | |
548 | break; | |
549 | } | |
550 | ||
551 | char ch = *first; | |
552 | if (ch == ')') | |
553 | { | |
554 | ++first; | |
555 | break; | |
556 | } | |
93dd7f36 PP |
557 | else if (ch == '_' |
558 | || __detail::__from_chars_alnum_to_val(ch) < 127) | |
cc3bf340 PP |
559 | continue; |
560 | else | |
561 | { | |
562 | first = fallback_first; | |
563 | break; | |
564 | } | |
565 | } | |
566 | } | |
567 | ||
568 | // We make the implementation-defined decision of ignoring the | |
569 | // sign bit and the n-char-sequence when assembling the NaN. | |
570 | uint_t result = 0; | |
571 | result <<= exponent_bits; | |
572 | result |= (1ull << exponent_bits) - 1; | |
573 | result <<= mantissa_bits; | |
574 | result |= (1ull << (mantissa_bits - 1)) | 1; | |
575 | memcpy(&value, &result, sizeof(result)); | |
576 | ||
577 | return {first, errc{}}; | |
578 | } | |
579 | } | |
580 | ||
581 | // Consume all insignificant leading zeros in the whole part of the | |
582 | // mantissa. | |
583 | bool seen_hexit = false; | |
584 | while (first != last && *first == '0') | |
585 | { | |
586 | seen_hexit = true; | |
587 | ++first; | |
588 | } | |
589 | ||
590 | // Now consume the rest of the written mantissa, populating MANTISSA with | |
591 | // the first MANTISSA_BITS+k significant bits of the written mantissa, where | |
592 | // 1 <= k <= 4 is the bit width of the leading significant written hexit. | |
593 | // | |
594 | // Examples: | |
595 | // After parsing "1.2f3", MANTISSA is 0x12f30000000000 (bit_width=52+1). | |
596 | // After parsing ".0000f0e", MANTISSA is 0xf0e00000000000 (bit_width=52+4). | |
597 | // After parsing ".1234567890abcd8", MANTISSA is 0x1234567890abcd (bit_width=52+1) | |
598 | // and MIDPOINT_BIT is true (and NONZERO_TAIL is false). | |
599 | uint_t mantissa = 0; | |
600 | int mantissa_idx = mantissa_bits; // The current bit index into MANTISSA | |
601 | // into which we'll write the next hexit. | |
602 | int exponent_adjustment = 0; // How much we'd have to adjust the written | |
603 | // exponent in order to represent the mantissa | |
604 | // in scientific form h.hhhhhhhhhhhhh. | |
605 | bool midpoint_bit = false; // Whether the MANTISSA_BITS+k+1 significant | |
606 | // bit is set in the written mantissa. | |
607 | bool nonzero_tail = false; // Whether some bit thereafter is set in the | |
608 | // written mantissa. | |
609 | bool seen_decimal_point = false; | |
610 | for (; first != last; ++first) | |
611 | { | |
612 | char ch = *first; | |
613 | if (ch == '.' && !seen_decimal_point) | |
614 | { | |
615 | seen_decimal_point = true; | |
616 | continue; | |
617 | } | |
618 | ||
93dd7f36 | 619 | int hexit = __detail::__from_chars_alnum_to_val(ch); |
a54137c8 | 620 | if (hexit >= 16) |
cc3bf340 PP |
621 | break; |
622 | seen_hexit = true; | |
623 | ||
624 | if (!seen_decimal_point && mantissa != 0) | |
625 | exponent_adjustment += 4; | |
626 | else if (seen_decimal_point && mantissa == 0) | |
627 | { | |
628 | exponent_adjustment -= 4; | |
629 | if (hexit == 0x0) | |
630 | continue; | |
631 | } | |
632 | ||
633 | if (mantissa_idx >= 0) | |
634 | mantissa |= uint_t(hexit) << mantissa_idx; | |
635 | else if (mantissa_idx >= -4) | |
636 | { | |
637 | if constexpr (is_same_v<T, float>) | |
638 | { | |
639 | __glibcxx_assert(mantissa_idx == -1); | |
640 | mantissa |= hexit >> 1; | |
641 | midpoint_bit = (hexit & 0b0001) != 0; | |
642 | } | |
643 | else | |
644 | { | |
645 | __glibcxx_assert(mantissa_idx == -4); | |
646 | midpoint_bit = (hexit & 0b1000) != 0; | |
647 | nonzero_tail = (hexit & 0b0111) != 0; | |
648 | } | |
649 | } | |
650 | else | |
651 | nonzero_tail |= (hexit != 0x0); | |
652 | ||
653 | mantissa_idx -= 4; | |
654 | } | |
655 | if (mantissa != 0) | |
656 | __glibcxx_assert(__bit_width(mantissa) >= mantissa_bits + 1 | |
657 | && __bit_width(mantissa) <= mantissa_bits + 4); | |
658 | else | |
659 | __glibcxx_assert(!midpoint_bit && !nonzero_tail); | |
660 | ||
661 | if (!seen_hexit) | |
662 | // If we haven't seen any hexit at this point, the parse failed. | |
663 | return {orig_first, errc::invalid_argument}; | |
664 | ||
665 | // Parse the written exponent. | |
666 | int written_exponent = 0; | |
4a6d7da7 | 667 | if (first != last && (*first == 'p' || *first == 'P')) |
cc3bf340 PP |
668 | { |
669 | // Tentatively consume the 'p' and try to parse a decimal number. | |
670 | const char* const fallback_first = first; | |
671 | ++first; | |
672 | if (first != last && *first == '+') | |
673 | ++first; | |
674 | from_chars_result fcr = from_chars(first, last, written_exponent, 10); | |
675 | if (fcr.ptr == first) | |
676 | // The parse failed, so undo consuming the 'p' and carry on as if the | |
677 | // exponent was omitted (i.e. is 0). | |
678 | first = fallback_first; | |
679 | else | |
680 | { | |
681 | first = fcr.ptr; | |
682 | if (mantissa != 0 && fcr.ec == errc::result_out_of_range) | |
683 | // Punt on very large exponents for now. FIXME | |
684 | return {first, errc::result_out_of_range}; | |
685 | } | |
686 | } | |
687 | int biased_exponent = written_exponent + exponent_bias; | |
688 | if (exponent_adjustment != 0) | |
689 | // The mantissa wasn't written in scientific form. Adjust the exponent | |
690 | // so that we may assume scientific form. | |
691 | // | |
692 | // Examples; | |
693 | // For input "a.bcp5", EXPONENT_ADJUSTMENT would be 0 since this | |
694 | // written mantissa is already in scientific form. | |
695 | // For input "ab.cp5", EXPONENT_ADJUSTMENT would be 4 since the | |
696 | // scientific form is "a.bcp9". | |
697 | // For input 0.0abcp5", EXPONENT_ADJUSTMENT would be -8 since the | |
698 | // scientific form is "a.bcp-3". | |
699 | biased_exponent += exponent_adjustment; | |
700 | ||
701 | // Shifts the mantissa to the right by AMOUNT while updating | |
702 | // BIASED_EXPONENT, MIDPOINT_BIT and NONZERO_TAIL accordingly. | |
703 | auto shift_mantissa = [&] (int amount) { | |
704 | __glibcxx_assert(amount >= 0); | |
705 | if (amount > mantissa_bits + 1) | |
706 | { | |
707 | // Shifting the mantissa by an amount greater than its precision. | |
708 | nonzero_tail |= midpoint_bit; | |
709 | nonzero_tail |= mantissa != 0; | |
710 | midpoint_bit = false; | |
711 | mantissa = 0; | |
712 | biased_exponent += amount; | |
713 | } | |
714 | else if (amount != 0) | |
715 | { | |
716 | nonzero_tail |= midpoint_bit; | |
717 | nonzero_tail |= (mantissa & ((1ull << (amount - 1)) - 1)) != 0; | |
718 | midpoint_bit = (mantissa & (1ull << (amount - 1))) != 0; | |
719 | mantissa >>= amount; | |
720 | biased_exponent += amount; | |
721 | } | |
722 | }; | |
723 | ||
724 | if (mantissa != 0) | |
725 | { | |
726 | // If the leading hexit is not '1', shift MANTISSA to make it so. | |
727 | // This normalizes input like "4.08p0" into "1.02p2". | |
728 | const int leading_hexit = mantissa >> mantissa_bits; | |
729 | const int leading_hexit_width = __bit_width(leading_hexit); // FIXME: optimize? | |
730 | __glibcxx_assert(leading_hexit_width >= 1 && leading_hexit_width <= 4); | |
731 | shift_mantissa(leading_hexit_width - 1); | |
732 | // After this adjustment, we can assume the leading hexit is '1'. | |
733 | __glibcxx_assert((mantissa >> mantissa_bits) == 0x1); | |
734 | } | |
735 | ||
736 | if (biased_exponent <= 0) | |
737 | { | |
738 | // This number is too small to be represented as a normal number, so | |
739 | // try for a subnormal number by shifting the mantissa sufficiently. | |
740 | // We need to shift by 1 more than -BIASED_EXPONENT because the leading | |
741 | // mantissa bit is omitted in the representation of a normal number but | |
742 | // not in a subnormal number. | |
743 | shift_mantissa(-biased_exponent + 1); | |
744 | __glibcxx_assert(!(mantissa & (1ull << mantissa_bits))); | |
745 | __glibcxx_assert(biased_exponent == 1); | |
746 | biased_exponent = 0; | |
747 | } | |
748 | ||
749 | // Perform round-to-nearest, tie-to-even rounding according to | |
750 | // MIDPOINT_BIT and NONZERO_TAIL. | |
751 | if (midpoint_bit && (nonzero_tail || (mantissa % 2) != 0)) | |
752 | { | |
753 | // Rounding away from zero. | |
754 | ++mantissa; | |
755 | midpoint_bit = false; | |
756 | nonzero_tail = false; | |
757 | ||
758 | // Deal with a couple of corner cases after rounding. | |
759 | if (mantissa == (1ull << mantissa_bits)) | |
760 | { | |
761 | // We rounded the subnormal number 1.fffffffffffff...p-1023 | |
762 | // up to the normal number 1p-1022. | |
763 | __glibcxx_assert(biased_exponent == 0); | |
764 | ++biased_exponent; | |
765 | } | |
766 | else if (mantissa & (1ull << (mantissa_bits + 1))) | |
767 | { | |
768 | // We rounded the normal number 1.fffffffffffff8pN (with maximal | |
769 | // mantissa) up to to 1p(N+1). | |
770 | mantissa >>= 1; | |
771 | ++biased_exponent; | |
772 | } | |
773 | } | |
774 | else | |
775 | { | |
776 | // Rounding toward zero. | |
777 | ||
778 | if (mantissa == 0 && (midpoint_bit || nonzero_tail)) | |
779 | { | |
780 | // A nonzero number that rounds to zero is unrepresentable. | |
781 | __glibcxx_assert(biased_exponent == 0); | |
782 | return {first, errc::result_out_of_range}; | |
783 | } | |
784 | ||
785 | midpoint_bit = false; | |
786 | nonzero_tail = false; | |
787 | } | |
788 | ||
789 | if (mantissa != 0 && biased_exponent >= (1 << exponent_bits) - 1) | |
790 | // The exponent of this number is too large to be representable. | |
791 | return {first, errc::result_out_of_range}; | |
792 | ||
793 | uint_t result = 0; | |
794 | if (mantissa == 0) | |
795 | { | |
796 | // Assemble a (possibly signed) zero. | |
797 | if (sign_bit) | |
798 | result |= 1ull << (exponent_bits + mantissa_bits); | |
799 | } | |
800 | else | |
801 | { | |
802 | // Assemble a nonzero normal or subnormal value. | |
803 | result |= sign_bit; | |
804 | result <<= exponent_bits; | |
805 | result |= biased_exponent; | |
806 | result <<= mantissa_bits; | |
807 | result |= mantissa & ((1ull << mantissa_bits) - 1); | |
808 | // The implicit leading mantissa bit is set iff the number is normal. | |
809 | __glibcxx_assert(((mantissa & (1ull << mantissa_bits)) != 0) | |
810 | == (biased_exponent != 0)); | |
811 | } | |
812 | memcpy(&value, &result, sizeof(result)); | |
813 | ||
814 | return {first, errc{}}; | |
815 | } | |
416b6fc7 | 816 | #endif // _GLIBCXX_FLOAT_IS_IEEE_BINARY32 && _GLIBCXX_DOUBLE_IS_IEEE_BINARY64 |
cc3bf340 | 817 | |
932fbc86 JW |
818 | } // namespace |
819 | ||
416b6fc7 | 820 | #if USE_LIB_FAST_FLOAT || USE_STRTOD_FOR_FROM_CHARS |
932fbc86 JW |
821 | |
822 | from_chars_result | |
823 | from_chars(const char* first, const char* last, float& value, | |
824 | chars_format fmt) noexcept | |
825 | { | |
5a4e2080 | 826 | #if USE_LIB_FAST_FLOAT |
cc3bf340 PP |
827 | if (fmt == chars_format::hex) |
828 | return __floating_from_chars_hex(first, last, value); | |
490e2303 PP |
829 | else |
830 | { | |
490e2303 PP |
831 | return fast_float::from_chars(first, last, value, fmt); |
832 | } | |
833 | #else | |
5a4e2080 | 834 | return from_chars_strtod(first, last, value, fmt); |
490e2303 | 835 | #endif |
932fbc86 JW |
836 | } |
837 | ||
838 | from_chars_result | |
839 | from_chars(const char* first, const char* last, double& value, | |
840 | chars_format fmt) noexcept | |
841 | { | |
5a4e2080 | 842 | #if USE_LIB_FAST_FLOAT |
cc3bf340 PP |
843 | if (fmt == chars_format::hex) |
844 | return __floating_from_chars_hex(first, last, value); | |
490e2303 PP |
845 | else |
846 | { | |
490e2303 PP |
847 | return fast_float::from_chars(first, last, value, fmt); |
848 | } | |
849 | #else | |
5a4e2080 | 850 | return from_chars_strtod(first, last, value, fmt); |
490e2303 | 851 | #endif |
932fbc86 JW |
852 | } |
853 | ||
854 | from_chars_result | |
855 | from_chars(const char* first, const char* last, long double& value, | |
856 | chars_format fmt) noexcept | |
857 | { | |
5a4e2080 JW |
858 | #if ! USE_STRTOD_FOR_FROM_CHARS |
859 | // Either long double is the same as double, or we can't use strtold. | |
860 | // In the latter case, this might give an incorrect result (e.g. values | |
861 | // out of range of double give an error, even if they fit in long double). | |
416b6fc7 JW |
862 | double dbl_value; |
863 | from_chars_result result; | |
864 | if (fmt == chars_format::hex) | |
865 | result = __floating_from_chars_hex(first, last, dbl_value); | |
866 | else | |
867 | { | |
416b6fc7 JW |
868 | result = fast_float::from_chars(first, last, dbl_value, fmt); |
869 | } | |
870 | if (result.ec == errc{}) | |
871 | value = dbl_value; | |
872 | return result; | |
873 | #else | |
5a4e2080 | 874 | return from_chars_strtod(first, last, value, fmt); |
416b6fc7 | 875 | #endif |
932fbc86 JW |
876 | } |
877 | ||
878 | #ifdef _GLIBCXX_LONG_DOUBLE_COMPAT | |
7c1e7eed JW |
879 | // Make std::from_chars for 64-bit long double an alias for the overload |
880 | // for double. | |
932fbc86 JW |
881 | extern "C" from_chars_result |
882 | _ZSt10from_charsPKcS0_ReSt12chars_format(const char* first, const char* last, | |
883 | long double& value, | |
884 | chars_format fmt) noexcept | |
885 | __attribute__((alias ("_ZSt10from_charsPKcS0_RdSt12chars_format"))); | |
886 | #endif | |
887 | ||
7c1e7eed JW |
888 | #ifdef _GLIBCXX_LONG_DOUBLE_ALT128_COMPAT |
889 | from_chars_result | |
890 | from_chars(const char* first, const char* last, __ieee128& value, | |
891 | chars_format fmt) noexcept | |
892 | { | |
5a4e2080 JW |
893 | // fast_float doesn't support IEEE binary128 format, but we can use strtold. |
894 | return from_chars_strtod(first, last, value, fmt); | |
7c1e7eed JW |
895 | } |
896 | #endif | |
897 | ||
416b6fc7 JW |
898 | #endif // USE_LIB_FAST_FLOAT || USE_STRTOD_FOR_FROM_CHARS |
899 | ||
932fbc86 JW |
900 | _GLIBCXX_END_NAMESPACE_VERSION |
901 | } // namespace std |