xref: /llvm-project/libc/src/__support/high_precision_decimal.h (revision a0c4f854cad2b97e44a1b58dc1fd982e1c4d60f3)
1 //===-- High Precision Decimal ----------------------------------*- C++ -*-===//
2 //
3 // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
4 // See httpss//llvm.org/LICENSE.txt for license information.
5 // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
6 //
7 //===----------------------------------------------------------------------===//
8 
9 // -----------------------------------------------------------------------------
10 //                               **** WARNING ****
11 // This file is shared with libc++. You should also be careful when adding
12 // dependencies to this file, since it needs to build for all libc++ targets.
13 // -----------------------------------------------------------------------------
14 
15 #ifndef LLVM_LIBC_SRC___SUPPORT_HIGH_PRECISION_DECIMAL_H
16 #define LLVM_LIBC_SRC___SUPPORT_HIGH_PRECISION_DECIMAL_H
17 
18 #include "src/__support/CPP/limits.h"
19 #include "src/__support/ctype_utils.h"
20 #include "src/__support/macros/config.h"
21 #include "src/__support/str_to_integer.h"
22 #include <stdint.h>
23 
24 namespace LIBC_NAMESPACE_DECL {
25 namespace internal {
26 
27 struct LShiftTableEntry {
28   uint32_t new_digits;
29   char const *power_of_five;
30 };
31 
32 // -----------------------------------------------------------------------------
33 //                               **** WARNING ****
34 // This interface is shared with libc++, if you change this interface you need
35 // to update it in both libc and libc++.
36 // -----------------------------------------------------------------------------
37 // This is used in both this file and in the main str_to_float.h.
38 // TODO: Figure out where to put this.
39 enum class RoundDirection { Up, Down, Nearest };
40 
41 // This is based on the HPD data structure described as part of the Simple
42 // Decimal Conversion algorithm by Nigel Tao, described at this link:
43 // https://nigeltao.github.io/blog/2020/parse-number-f64-simple.html
44 class HighPrecisionDecimal {
45 
46   // This precomputed table speeds up left shifts by having the number of new
47   // digits that will be added by multiplying 5^i by 2^i. If the number is less
48   // than 5^i then it will add one fewer digit. There are only 60 entries since
49   // that's the max shift amount.
50   // This table was generated by the script at
51   // libc/utils/mathtools/GenerateHPDConstants.py
52   static constexpr LShiftTableEntry LEFT_SHIFT_DIGIT_TABLE[] = {
53       {0, ""},
54       {1, "5"},
55       {1, "25"},
56       {1, "125"},
57       {2, "625"},
58       {2, "3125"},
59       {2, "15625"},
60       {3, "78125"},
61       {3, "390625"},
62       {3, "1953125"},
63       {4, "9765625"},
64       {4, "48828125"},
65       {4, "244140625"},
66       {4, "1220703125"},
67       {5, "6103515625"},
68       {5, "30517578125"},
69       {5, "152587890625"},
70       {6, "762939453125"},
71       {6, "3814697265625"},
72       {6, "19073486328125"},
73       {7, "95367431640625"},
74       {7, "476837158203125"},
75       {7, "2384185791015625"},
76       {7, "11920928955078125"},
77       {8, "59604644775390625"},
78       {8, "298023223876953125"},
79       {8, "1490116119384765625"},
80       {9, "7450580596923828125"},
81       {9, "37252902984619140625"},
82       {9, "186264514923095703125"},
83       {10, "931322574615478515625"},
84       {10, "4656612873077392578125"},
85       {10, "23283064365386962890625"},
86       {10, "116415321826934814453125"},
87       {11, "582076609134674072265625"},
88       {11, "2910383045673370361328125"},
89       {11, "14551915228366851806640625"},
90       {12, "72759576141834259033203125"},
91       {12, "363797880709171295166015625"},
92       {12, "1818989403545856475830078125"},
93       {13, "9094947017729282379150390625"},
94       {13, "45474735088646411895751953125"},
95       {13, "227373675443232059478759765625"},
96       {13, "1136868377216160297393798828125"},
97       {14, "5684341886080801486968994140625"},
98       {14, "28421709430404007434844970703125"},
99       {14, "142108547152020037174224853515625"},
100       {15, "710542735760100185871124267578125"},
101       {15, "3552713678800500929355621337890625"},
102       {15, "17763568394002504646778106689453125"},
103       {16, "88817841970012523233890533447265625"},
104       {16, "444089209850062616169452667236328125"},
105       {16, "2220446049250313080847263336181640625"},
106       {16, "11102230246251565404236316680908203125"},
107       {17, "55511151231257827021181583404541015625"},
108       {17, "277555756156289135105907917022705078125"},
109       {17, "1387778780781445675529539585113525390625"},
110       {18, "6938893903907228377647697925567626953125"},
111       {18, "34694469519536141888238489627838134765625"},
112       {18, "173472347597680709441192448139190673828125"},
113       {19, "867361737988403547205962240695953369140625"},
114   };
115 
116   // The maximum amount we can shift is the number of bits used in the
117   // accumulator, minus the number of bits needed to represent the base (in this
118   // case 4).
119   static constexpr uint32_t MAX_SHIFT_AMOUNT = sizeof(uint64_t) - 4;
120 
121   // 800 is an arbitrary number of digits, but should be
122   // large enough for any practical number.
123   static constexpr uint32_t MAX_NUM_DIGITS = 800;
124 
125   uint32_t num_digits = 0;
126   int32_t decimal_point = 0;
127   bool truncated = false;
128   uint8_t digits[MAX_NUM_DIGITS];
129 
130 private:
131   LIBC_INLINE bool should_round_up(int32_t round_to_digit,
132                                    RoundDirection round) {
133     if (round_to_digit < 0 ||
134         static_cast<uint32_t>(round_to_digit) >= this->num_digits) {
135       return false;
136     }
137 
138     // The above condition handles all cases where all of the trailing digits
139     // are zero. In that case, if the rounding mode is up, then this number
140     // should be rounded up. Similarly, if the rounding mode is down, then it
141     // should always round down.
142     if (round == RoundDirection::Up) {
143       return true;
144     } else if (round == RoundDirection::Down) {
145       return false;
146     }
147     // Else round to nearest.
148 
149     // If we're right in the middle and there are no extra digits
150     if (this->digits[round_to_digit] == 5 &&
151         static_cast<uint32_t>(round_to_digit + 1) == this->num_digits) {
152 
153       // Round up if we've truncated (since that means the result is slightly
154       // higher than what's represented.)
155       if (this->truncated) {
156         return true;
157       }
158 
159       // If this exactly halfway, round to even.
160       if (round_to_digit == 0)
161         // When the input is ".5".
162         return false;
163       return this->digits[round_to_digit - 1] % 2 != 0;
164     }
165     // If there are digits after round_to_digit, they must be non-zero since we
166     // trim trailing zeroes after all operations that change digits.
167     return this->digits[round_to_digit] >= 5;
168   }
169 
170   // Takes an amount to left shift and returns the number of new digits needed
171   // to store the result based on LEFT_SHIFT_DIGIT_TABLE.
172   LIBC_INLINE uint32_t get_num_new_digits(uint32_t lshift_amount) {
173     const char *power_of_five =
174         LEFT_SHIFT_DIGIT_TABLE[lshift_amount].power_of_five;
175     uint32_t new_digits = LEFT_SHIFT_DIGIT_TABLE[lshift_amount].new_digits;
176     uint32_t digit_index = 0;
177     while (power_of_five[digit_index] != 0) {
178       if (digit_index >= this->num_digits) {
179         return new_digits - 1;
180       }
181       if (this->digits[digit_index] !=
182           internal::b36_char_to_int(power_of_five[digit_index])) {
183         return new_digits -
184                ((this->digits[digit_index] <
185                  internal::b36_char_to_int(power_of_five[digit_index]))
186                     ? 1
187                     : 0);
188       }
189       ++digit_index;
190     }
191     return new_digits;
192   }
193 
194   // Trim all trailing 0s
195   LIBC_INLINE void trim_trailing_zeroes() {
196     while (this->num_digits > 0 && this->digits[this->num_digits - 1] == 0) {
197       --this->num_digits;
198     }
199     if (this->num_digits == 0) {
200       this->decimal_point = 0;
201     }
202   }
203 
204   // Perform a digitwise binary non-rounding right shift on this value by
205   // shift_amount. The shift_amount can't be more than MAX_SHIFT_AMOUNT to
206   // prevent overflow.
207   LIBC_INLINE void right_shift(uint32_t shift_amount) {
208     uint32_t read_index = 0;
209     uint32_t write_index = 0;
210 
211     uint64_t accumulator = 0;
212 
213     const uint64_t shift_mask = (uint64_t(1) << shift_amount) - 1;
214 
215     // Warm Up phase: we don't have enough digits to start writing, so just
216     // read them into the accumulator.
217     while (accumulator >> shift_amount == 0) {
218       uint64_t read_digit = 0;
219       // If there are still digits to read, read the next one, else the digit is
220       // assumed to be 0.
221       if (read_index < this->num_digits) {
222         read_digit = this->digits[read_index];
223       }
224       accumulator = accumulator * 10 + read_digit;
225       ++read_index;
226     }
227 
228     // Shift the decimal point by the number of digits it took to fill the
229     // accumulator.
230     this->decimal_point -= read_index - 1;
231 
232     // Middle phase: we have enough digits to write, as well as more digits to
233     // read. Keep reading until we run out of digits.
234     while (read_index < this->num_digits) {
235       uint64_t read_digit = this->digits[read_index];
236       uint64_t write_digit = accumulator >> shift_amount;
237       accumulator &= shift_mask;
238       this->digits[write_index] = static_cast<uint8_t>(write_digit);
239       accumulator = accumulator * 10 + read_digit;
240       ++read_index;
241       ++write_index;
242     }
243 
244     // Cool Down phase: All of the readable digits have been read, so just write
245     // the remainder, while treating any more digits as 0.
246     while (accumulator > 0) {
247       uint64_t write_digit = accumulator >> shift_amount;
248       accumulator &= shift_mask;
249       if (write_index < MAX_NUM_DIGITS) {
250         this->digits[write_index] = static_cast<uint8_t>(write_digit);
251         ++write_index;
252       } else if (write_digit > 0) {
253         this->truncated = true;
254       }
255       accumulator = accumulator * 10;
256     }
257     this->num_digits = write_index;
258     this->trim_trailing_zeroes();
259   }
260 
261   // Perform a digitwise binary non-rounding left shift on this value by
262   // shift_amount. The shift_amount can't be more than MAX_SHIFT_AMOUNT to
263   // prevent overflow.
264   LIBC_INLINE void left_shift(uint32_t shift_amount) {
265     uint32_t new_digits = this->get_num_new_digits(shift_amount);
266 
267     int32_t read_index = this->num_digits - 1;
268     uint32_t write_index = this->num_digits + new_digits;
269 
270     uint64_t accumulator = 0;
271 
272     // No Warm Up phase. Since we're putting digits in at the top and taking
273     // digits from the bottom we don't have to wait for the accumulator to fill.
274 
275     // Middle phase: while we have more digits to read, keep reading as well as
276     // writing.
277     while (read_index >= 0) {
278       accumulator += static_cast<uint64_t>(this->digits[read_index])
279                      << shift_amount;
280       uint64_t next_accumulator = accumulator / 10;
281       uint64_t write_digit = accumulator - (10 * next_accumulator);
282       --write_index;
283       if (write_index < MAX_NUM_DIGITS) {
284         this->digits[write_index] = static_cast<uint8_t>(write_digit);
285       } else if (write_digit != 0) {
286         this->truncated = true;
287       }
288       accumulator = next_accumulator;
289       --read_index;
290     }
291 
292     // Cool Down phase: there are no more digits to read, so just write the
293     // remaining digits in the accumulator.
294     while (accumulator > 0) {
295       uint64_t next_accumulator = accumulator / 10;
296       uint64_t write_digit = accumulator - (10 * next_accumulator);
297       --write_index;
298       if (write_index < MAX_NUM_DIGITS) {
299         this->digits[write_index] = static_cast<uint8_t>(write_digit);
300       } else if (write_digit != 0) {
301         this->truncated = true;
302       }
303       accumulator = next_accumulator;
304     }
305 
306     this->num_digits += new_digits;
307     if (this->num_digits > MAX_NUM_DIGITS) {
308       this->num_digits = MAX_NUM_DIGITS;
309     }
310     this->decimal_point += new_digits;
311     this->trim_trailing_zeroes();
312   }
313 
314 public:
315   // num_string is assumed to be a string of numeric characters. It doesn't
316   // handle leading spaces.
317   LIBC_INLINE
318   HighPrecisionDecimal(
319       const char *__restrict num_string,
320       const size_t num_len = cpp::numeric_limits<size_t>::max()) {
321     bool saw_dot = false;
322     size_t num_cur = 0;
323     // This counts the digits in the number, even if there isn't space to store
324     // them all.
325     uint32_t total_digits = 0;
326     while (num_cur < num_len &&
327            (isdigit(num_string[num_cur]) || num_string[num_cur] == '.')) {
328       if (num_string[num_cur] == '.') {
329         if (saw_dot) {
330           break;
331         }
332         this->decimal_point = total_digits;
333         saw_dot = true;
334       } else {
335         if (num_string[num_cur] == '0' && this->num_digits == 0) {
336           --this->decimal_point;
337           ++num_cur;
338           continue;
339         }
340         ++total_digits;
341         if (this->num_digits < MAX_NUM_DIGITS) {
342           this->digits[this->num_digits] = static_cast<uint8_t>(
343               internal::b36_char_to_int(num_string[num_cur]));
344           ++this->num_digits;
345         } else if (num_string[num_cur] != '0') {
346           this->truncated = true;
347         }
348       }
349       ++num_cur;
350     }
351 
352     if (!saw_dot)
353       this->decimal_point = total_digits;
354 
355     if (num_cur < num_len &&
356         (num_string[num_cur] == 'e' || num_string[num_cur] == 'E')) {
357       ++num_cur;
358       if (isdigit(num_string[num_cur]) || num_string[num_cur] == '+' ||
359           num_string[num_cur] == '-') {
360         auto result =
361             strtointeger<int32_t>(num_string + num_cur, 10, num_len - num_cur);
362         if (result.has_error()) {
363           // TODO: handle error
364         }
365         int32_t add_to_exponent = result.value;
366 
367         // Here we do this operation as int64 to avoid overflow.
368         int64_t temp_exponent = static_cast<int64_t>(this->decimal_point) +
369                                 static_cast<int64_t>(add_to_exponent);
370 
371         // Theoretically these numbers should be MAX_BIASED_EXPONENT for long
372         // double, but that should be ~16,000 which is much less than 1 << 30.
373         if (temp_exponent > (1 << 30)) {
374           temp_exponent = (1 << 30);
375         } else if (temp_exponent < -(1 << 30)) {
376           temp_exponent = -(1 << 30);
377         }
378         this->decimal_point = static_cast<int32_t>(temp_exponent);
379       }
380     }
381 
382     this->trim_trailing_zeroes();
383   }
384 
385   // Binary shift left (shift_amount > 0) or right (shift_amount < 0)
386   LIBC_INLINE void shift(int shift_amount) {
387     if (shift_amount == 0) {
388       return;
389     }
390     // Left
391     else if (shift_amount > 0) {
392       while (static_cast<uint32_t>(shift_amount) > MAX_SHIFT_AMOUNT) {
393         this->left_shift(MAX_SHIFT_AMOUNT);
394         shift_amount -= MAX_SHIFT_AMOUNT;
395       }
396       this->left_shift(shift_amount);
397     }
398     // Right
399     else {
400       while (static_cast<uint32_t>(shift_amount) < -MAX_SHIFT_AMOUNT) {
401         this->right_shift(MAX_SHIFT_AMOUNT);
402         shift_amount += MAX_SHIFT_AMOUNT;
403       }
404       this->right_shift(-shift_amount);
405     }
406   }
407 
408   // Round the number represented to the closest value of unsigned int type T.
409   // This is done ignoring overflow.
410   template <class T>
411   LIBC_INLINE T
412   round_to_integer_type(RoundDirection round = RoundDirection::Nearest) {
413     T result = 0;
414     uint32_t cur_digit = 0;
415 
416     while (static_cast<int32_t>(cur_digit) < this->decimal_point &&
417            cur_digit < this->num_digits) {
418       result = result * 10 + (this->digits[cur_digit]);
419       ++cur_digit;
420     }
421 
422     // If there are implicit 0s at the end of the number, include those.
423     while (static_cast<int32_t>(cur_digit) < this->decimal_point) {
424       result *= 10;
425       ++cur_digit;
426     }
427     return result + static_cast<unsigned int>(
428                         this->should_round_up(this->decimal_point, round));
429   }
430 
431   // Extra functions for testing.
432 
433   LIBC_INLINE uint8_t *get_digits() { return this->digits; }
434   LIBC_INLINE uint32_t get_num_digits() { return this->num_digits; }
435   LIBC_INLINE int32_t get_decimal_point() { return this->decimal_point; }
436   LIBC_INLINE void set_truncated(bool trunc) { this->truncated = trunc; }
437 };
438 
439 } // namespace internal
440 } // namespace LIBC_NAMESPACE_DECL
441 
442 #endif // LLVM_LIBC_SRC___SUPPORT_HIGH_PRECISION_DECIMAL_H
443