]>
Commit | Line | Data |
---|---|---|
0f5d89e8 A |
1 | // © 2018 and later: Unicode, Inc. and others. |
2 | // License & terms of use: http://www.unicode.org/copyright.html | |
3 | // | |
4 | // From the double-conversion library. Original license: | |
5 | // | |
6 | // Copyright 2012 the V8 project authors. All rights reserved. | |
7 | // Redistribution and use in source and binary forms, with or without | |
8 | // modification, are permitted provided that the following conditions are | |
9 | // met: | |
10 | // | |
11 | // * Redistributions of source code must retain the above copyright | |
12 | // notice, this list of conditions and the following disclaimer. | |
13 | // * Redistributions in binary form must reproduce the above | |
14 | // copyright notice, this list of conditions and the following | |
15 | // disclaimer in the documentation and/or other materials provided | |
16 | // with the distribution. | |
17 | // * Neither the name of Google Inc. nor the names of its | |
18 | // contributors may be used to endorse or promote products derived | |
19 | // from this software without specific prior written permission. | |
20 | // | |
21 | // THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS | |
22 | // "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT | |
23 | // LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR | |
24 | // A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT | |
25 | // OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, | |
26 | // SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT | |
27 | // LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, | |
28 | // DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY | |
29 | // THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT | |
30 | // (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE | |
31 | // OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE. | |
32 | ||
33 | // ICU PATCH: ifdef around UCONFIG_NO_FORMATTING | |
34 | #include "unicode/utypes.h" | |
35 | #if !UCONFIG_NO_FORMATTING | |
36 | ||
37 | // ICU PATCH: Customize header file paths for ICU. | |
38 | ||
39 | #include "double-conversion-fast-dtoa.h" | |
40 | ||
41 | #include "double-conversion-cached-powers.h" | |
42 | #include "double-conversion-diy-fp.h" | |
43 | #include "double-conversion-ieee.h" | |
44 | ||
45 | // ICU PATCH: Wrap in ICU namespace | |
46 | U_NAMESPACE_BEGIN | |
47 | ||
48 | namespace double_conversion { | |
49 | ||
50 | // The minimal and maximal target exponent define the range of w's binary | |
51 | // exponent, where 'w' is the result of multiplying the input by a cached power | |
52 | // of ten. | |
53 | // | |
54 | // A different range might be chosen on a different platform, to optimize digit | |
55 | // generation, but a smaller range requires more powers of ten to be cached. | |
56 | static const int kMinimalTargetExponent = -60; | |
57 | static const int kMaximalTargetExponent = -32; | |
58 | ||
59 | ||
60 | // Adjusts the last digit of the generated number, and screens out generated | |
61 | // solutions that may be inaccurate. A solution may be inaccurate if it is | |
62 | // outside the safe interval, or if we cannot prove that it is closer to the | |
63 | // input than a neighboring representation of the same length. | |
64 | // | |
65 | // Input: * buffer containing the digits of too_high / 10^kappa | |
66 | // * the buffer's length | |
67 | // * distance_too_high_w == (too_high - w).f() * unit | |
68 | // * unsafe_interval == (too_high - too_low).f() * unit | |
69 | // * rest = (too_high - buffer * 10^kappa).f() * unit | |
70 | // * ten_kappa = 10^kappa * unit | |
71 | // * unit = the common multiplier | |
72 | // Output: returns true if the buffer is guaranteed to contain the closest | |
73 | // representable number to the input. | |
74 | // Modifies the generated digits in the buffer to approach (round towards) w. | |
75 | static bool RoundWeed(Vector<char> buffer, | |
76 | int length, | |
77 | uint64_t distance_too_high_w, | |
78 | uint64_t unsafe_interval, | |
79 | uint64_t rest, | |
80 | uint64_t ten_kappa, | |
81 | uint64_t unit) { | |
82 | uint64_t small_distance = distance_too_high_w - unit; | |
83 | uint64_t big_distance = distance_too_high_w + unit; | |
84 | // Let w_low = too_high - big_distance, and | |
85 | // w_high = too_high - small_distance. | |
86 | // Note: w_low < w < w_high | |
87 | // | |
88 | // The real w (* unit) must lie somewhere inside the interval | |
89 | // ]w_low; w_high[ (often written as "(w_low; w_high)") | |
90 | ||
91 | // Basically the buffer currently contains a number in the unsafe interval | |
92 | // ]too_low; too_high[ with too_low < w < too_high | |
93 | // | |
94 | // too_high - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - | |
95 | // ^v 1 unit ^ ^ ^ ^ | |
96 | // boundary_high --------------------- . . . . | |
97 | // ^v 1 unit . . . . | |
98 | // - - - - - - - - - - - - - - - - - - - + - - + - - - - - - . . | |
99 | // . . ^ . . | |
100 | // . big_distance . . . | |
101 | // . . . . rest | |
102 | // small_distance . . . . | |
103 | // v . . . . | |
104 | // w_high - - - - - - - - - - - - - - - - - - . . . . | |
105 | // ^v 1 unit . . . . | |
106 | // w ---------------------------------------- . . . . | |
107 | // ^v 1 unit v . . . | |
108 | // w_low - - - - - - - - - - - - - - - - - - - - - . . . | |
109 | // . . v | |
110 | // buffer --------------------------------------------------+-------+-------- | |
111 | // . . | |
112 | // safe_interval . | |
113 | // v . | |
114 | // - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - . | |
115 | // ^v 1 unit . | |
116 | // boundary_low ------------------------- unsafe_interval | |
117 | // ^v 1 unit v | |
118 | // too_low - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - | |
119 | // | |
120 | // | |
121 | // Note that the value of buffer could lie anywhere inside the range too_low | |
122 | // to too_high. | |
123 | // | |
124 | // boundary_low, boundary_high and w are approximations of the real boundaries | |
125 | // and v (the input number). They are guaranteed to be precise up to one unit. | |
126 | // In fact the error is guaranteed to be strictly less than one unit. | |
127 | // | |
128 | // Anything that lies outside the unsafe interval is guaranteed not to round | |
129 | // to v when read again. | |
130 | // Anything that lies inside the safe interval is guaranteed to round to v | |
131 | // when read again. | |
132 | // If the number inside the buffer lies inside the unsafe interval but not | |
133 | // inside the safe interval then we simply do not know and bail out (returning | |
134 | // false). | |
135 | // | |
136 | // Similarly we have to take into account the imprecision of 'w' when finding | |
137 | // the closest representation of 'w'. If we have two potential | |
138 | // representations, and one is closer to both w_low and w_high, then we know | |
139 | // it is closer to the actual value v. | |
140 | // | |
141 | // By generating the digits of too_high we got the largest (closest to | |
142 | // too_high) buffer that is still in the unsafe interval. In the case where | |
143 | // w_high < buffer < too_high we try to decrement the buffer. | |
144 | // This way the buffer approaches (rounds towards) w. | |
145 | // There are 3 conditions that stop the decrementation process: | |
146 | // 1) the buffer is already below w_high | |
147 | // 2) decrementing the buffer would make it leave the unsafe interval | |
148 | // 3) decrementing the buffer would yield a number below w_high and farther | |
149 | // away than the current number. In other words: | |
150 | // (buffer{-1} < w_high) && w_high - buffer{-1} > buffer - w_high | |
151 | // Instead of using the buffer directly we use its distance to too_high. | |
152 | // Conceptually rest ~= too_high - buffer | |
153 | // We need to do the following tests in this order to avoid over- and | |
154 | // underflows. | |
155 | ASSERT(rest <= unsafe_interval); | |
156 | while (rest < small_distance && // Negated condition 1 | |
157 | unsafe_interval - rest >= ten_kappa && // Negated condition 2 | |
158 | (rest + ten_kappa < small_distance || // buffer{-1} > w_high | |
159 | small_distance - rest >= rest + ten_kappa - small_distance)) { | |
160 | buffer[length - 1]--; | |
161 | rest += ten_kappa; | |
162 | } | |
163 | ||
164 | // We have approached w+ as much as possible. We now test if approaching w- | |
165 | // would require changing the buffer. If yes, then we have two possible | |
166 | // representations close to w, but we cannot decide which one is closer. | |
167 | if (rest < big_distance && | |
168 | unsafe_interval - rest >= ten_kappa && | |
169 | (rest + ten_kappa < big_distance || | |
170 | big_distance - rest > rest + ten_kappa - big_distance)) { | |
171 | return false; | |
172 | } | |
173 | ||
174 | // Weeding test. | |
175 | // The safe interval is [too_low + 2 ulp; too_high - 2 ulp] | |
176 | // Since too_low = too_high - unsafe_interval this is equivalent to | |
177 | // [too_high - unsafe_interval + 4 ulp; too_high - 2 ulp] | |
178 | // Conceptually we have: rest ~= too_high - buffer | |
179 | return (2 * unit <= rest) && (rest <= unsafe_interval - 4 * unit); | |
180 | } | |
181 | ||
182 | ||
183 | // Rounds the buffer upwards if the result is closer to v by possibly adding | |
184 | // 1 to the buffer. If the precision of the calculation is not sufficient to | |
185 | // round correctly, return false. | |
186 | // The rounding might shift the whole buffer in which case the kappa is | |
187 | // adjusted. For example "99", kappa = 3 might become "10", kappa = 4. | |
188 | // | |
189 | // If 2*rest > ten_kappa then the buffer needs to be round up. | |
190 | // rest can have an error of +/- 1 unit. This function accounts for the | |
191 | // imprecision and returns false, if the rounding direction cannot be | |
192 | // unambiguously determined. | |
193 | // | |
194 | // Precondition: rest < ten_kappa. | |
195 | static bool RoundWeedCounted(Vector<char> buffer, | |
196 | int length, | |
197 | uint64_t rest, | |
198 | uint64_t ten_kappa, | |
199 | uint64_t unit, | |
200 | int* kappa) { | |
201 | ASSERT(rest < ten_kappa); | |
202 | // The following tests are done in a specific order to avoid overflows. They | |
203 | // will work correctly with any uint64 values of rest < ten_kappa and unit. | |
204 | // | |
205 | // If the unit is too big, then we don't know which way to round. For example | |
206 | // a unit of 50 means that the real number lies within rest +/- 50. If | |
207 | // 10^kappa == 40 then there is no way to tell which way to round. | |
208 | if (unit >= ten_kappa) return false; | |
209 | // Even if unit is just half the size of 10^kappa we are already completely | |
210 | // lost. (And after the previous test we know that the expression will not | |
211 | // over/underflow.) | |
212 | if (ten_kappa - unit <= unit) return false; | |
213 | // If 2 * (rest + unit) <= 10^kappa we can safely round down. | |
214 | if ((ten_kappa - rest > rest) && (ten_kappa - 2 * rest >= 2 * unit)) { | |
215 | return true; | |
216 | } | |
217 | // If 2 * (rest - unit) >= 10^kappa, then we can safely round up. | |
218 | if ((rest > unit) && (ten_kappa - (rest - unit) <= (rest - unit))) { | |
219 | // Increment the last digit recursively until we find a non '9' digit. | |
220 | buffer[length - 1]++; | |
221 | for (int i = length - 1; i > 0; --i) { | |
222 | if (buffer[i] != '0' + 10) break; | |
223 | buffer[i] = '0'; | |
224 | buffer[i - 1]++; | |
225 | } | |
226 | // If the first digit is now '0'+ 10 we had a buffer with all '9's. With the | |
227 | // exception of the first digit all digits are now '0'. Simply switch the | |
228 | // first digit to '1' and adjust the kappa. Example: "99" becomes "10" and | |
229 | // the power (the kappa) is increased. | |
230 | if (buffer[0] == '0' + 10) { | |
231 | buffer[0] = '1'; | |
232 | (*kappa) += 1; | |
233 | } | |
234 | return true; | |
235 | } | |
236 | return false; | |
237 | } | |
238 | ||
239 | // Returns the biggest power of ten that is less than or equal to the given | |
240 | // number. We furthermore receive the maximum number of bits 'number' has. | |
241 | // | |
242 | // Returns power == 10^(exponent_plus_one-1) such that | |
243 | // power <= number < power * 10. | |
244 | // If number_bits == 0 then 0^(0-1) is returned. | |
245 | // The number of bits must be <= 32. | |
246 | // Precondition: number < (1 << (number_bits + 1)). | |
247 | ||
248 | // Inspired by the method for finding an integer log base 10 from here: | |
249 | // http://graphics.stanford.edu/~seander/bithacks.html#IntegerLog10 | |
250 | static unsigned int const kSmallPowersOfTen[] = | |
251 | {0, 1, 10, 100, 1000, 10000, 100000, 1000000, 10000000, 100000000, | |
252 | 1000000000}; | |
253 | ||
254 | static void BiggestPowerTen(uint32_t number, | |
255 | int number_bits, | |
256 | uint32_t* power, | |
257 | int* exponent_plus_one) { | |
258 | ASSERT(number < (1u << (number_bits + 1))); | |
259 | // 1233/4096 is approximately 1/lg(10). | |
260 | int exponent_plus_one_guess = ((number_bits + 1) * 1233 >> 12); | |
261 | // We increment to skip over the first entry in the kPowersOf10 table. | |
262 | // Note: kPowersOf10[i] == 10^(i-1). | |
263 | exponent_plus_one_guess++; | |
264 | // We don't have any guarantees that 2^number_bits <= number. | |
265 | if (number < kSmallPowersOfTen[exponent_plus_one_guess]) { | |
266 | exponent_plus_one_guess--; | |
267 | } | |
268 | *power = kSmallPowersOfTen[exponent_plus_one_guess]; | |
269 | *exponent_plus_one = exponent_plus_one_guess; | |
270 | } | |
271 | ||
272 | // Generates the digits of input number w. | |
273 | // w is a floating-point number (DiyFp), consisting of a significand and an | |
274 | // exponent. Its exponent is bounded by kMinimalTargetExponent and | |
275 | // kMaximalTargetExponent. | |
276 | // Hence -60 <= w.e() <= -32. | |
277 | // | |
278 | // Returns false if it fails, in which case the generated digits in the buffer | |
279 | // should not be used. | |
280 | // Preconditions: | |
281 | // * low, w and high are correct up to 1 ulp (unit in the last place). That | |
282 | // is, their error must be less than a unit of their last digits. | |
283 | // * low.e() == w.e() == high.e() | |
284 | // * low < w < high, and taking into account their error: low~ <= high~ | |
285 | // * kMinimalTargetExponent <= w.e() <= kMaximalTargetExponent | |
286 | // Postconditions: returns false if procedure fails. | |
287 | // otherwise: | |
288 | // * buffer is not null-terminated, but len contains the number of digits. | |
289 | // * buffer contains the shortest possible decimal digit-sequence | |
290 | // such that LOW < buffer * 10^kappa < HIGH, where LOW and HIGH are the | |
291 | // correct values of low and high (without their error). | |
292 | // * if more than one decimal representation gives the minimal number of | |
293 | // decimal digits then the one closest to W (where W is the correct value | |
294 | // of w) is chosen. | |
295 | // Remark: this procedure takes into account the imprecision of its input | |
296 | // numbers. If the precision is not enough to guarantee all the postconditions | |
297 | // then false is returned. This usually happens rarely (~0.5%). | |
298 | // | |
299 | // Say, for the sake of example, that | |
300 | // w.e() == -48, and w.f() == 0x1234567890abcdef | |
301 | // w's value can be computed by w.f() * 2^w.e() | |
302 | // We can obtain w's integral digits by simply shifting w.f() by -w.e(). | |
303 | // -> w's integral part is 0x1234 | |
304 | // w's fractional part is therefore 0x567890abcdef. | |
305 | // Printing w's integral part is easy (simply print 0x1234 in decimal). | |
306 | // In order to print its fraction we repeatedly multiply the fraction by 10 and | |
307 | // get each digit. Example the first digit after the point would be computed by | |
308 | // (0x567890abcdef * 10) >> 48. -> 3 | |
309 | // The whole thing becomes slightly more complicated because we want to stop | |
310 | // once we have enough digits. That is, once the digits inside the buffer | |
311 | // represent 'w' we can stop. Everything inside the interval low - high | |
312 | // represents w. However we have to pay attention to low, high and w's | |
313 | // imprecision. | |
314 | static bool DigitGen(DiyFp low, | |
315 | DiyFp w, | |
316 | DiyFp high, | |
317 | Vector<char> buffer, | |
318 | int* length, | |
319 | int* kappa) { | |
320 | ASSERT(low.e() == w.e() && w.e() == high.e()); | |
321 | ASSERT(low.f() + 1 <= high.f() - 1); | |
322 | ASSERT(kMinimalTargetExponent <= w.e() && w.e() <= kMaximalTargetExponent); | |
323 | // low, w and high are imprecise, but by less than one ulp (unit in the last | |
324 | // place). | |
325 | // If we remove (resp. add) 1 ulp from low (resp. high) we are certain that | |
326 | // the new numbers are outside of the interval we want the final | |
327 | // representation to lie in. | |
328 | // Inversely adding (resp. removing) 1 ulp from low (resp. high) would yield | |
329 | // numbers that are certain to lie in the interval. We will use this fact | |
330 | // later on. | |
331 | // We will now start by generating the digits within the uncertain | |
332 | // interval. Later we will weed out representations that lie outside the safe | |
333 | // interval and thus _might_ lie outside the correct interval. | |
334 | uint64_t unit = 1; | |
335 | DiyFp too_low = DiyFp(low.f() - unit, low.e()); | |
336 | DiyFp too_high = DiyFp(high.f() + unit, high.e()); | |
337 | // too_low and too_high are guaranteed to lie outside the interval we want the | |
338 | // generated number in. | |
339 | DiyFp unsafe_interval = DiyFp::Minus(too_high, too_low); | |
340 | // We now cut the input number into two parts: the integral digits and the | |
341 | // fractionals. We will not write any decimal separator though, but adapt | |
342 | // kappa instead. | |
343 | // Reminder: we are currently computing the digits (stored inside the buffer) | |
344 | // such that: too_low < buffer * 10^kappa < too_high | |
345 | // We use too_high for the digit_generation and stop as soon as possible. | |
346 | // If we stop early we effectively round down. | |
347 | DiyFp one = DiyFp(static_cast<uint64_t>(1) << -w.e(), w.e()); | |
348 | // Division by one is a shift. | |
349 | uint32_t integrals = static_cast<uint32_t>(too_high.f() >> -one.e()); | |
350 | // Modulo by one is an and. | |
351 | uint64_t fractionals = too_high.f() & (one.f() - 1); | |
352 | uint32_t divisor; | |
353 | int divisor_exponent_plus_one; | |
354 | BiggestPowerTen(integrals, DiyFp::kSignificandSize - (-one.e()), | |
355 | &divisor, &divisor_exponent_plus_one); | |
356 | *kappa = divisor_exponent_plus_one; | |
357 | *length = 0; | |
358 | // Loop invariant: buffer = too_high / 10^kappa (integer division) | |
359 | // The invariant holds for the first iteration: kappa has been initialized | |
360 | // with the divisor exponent + 1. And the divisor is the biggest power of ten | |
361 | // that is smaller than integrals. | |
362 | while (*kappa > 0) { | |
363 | int digit = integrals / divisor; | |
364 | ASSERT(digit <= 9); | |
365 | buffer[*length] = static_cast<char>('0' + digit); | |
366 | (*length)++; | |
367 | integrals %= divisor; | |
368 | (*kappa)--; | |
369 | // Note that kappa now equals the exponent of the divisor and that the | |
370 | // invariant thus holds again. | |
371 | uint64_t rest = | |
372 | (static_cast<uint64_t>(integrals) << -one.e()) + fractionals; | |
373 | // Invariant: too_high = buffer * 10^kappa + DiyFp(rest, one.e()) | |
374 | // Reminder: unsafe_interval.e() == one.e() | |
375 | if (rest < unsafe_interval.f()) { | |
376 | // Rounding down (by not emitting the remaining digits) yields a number | |
377 | // that lies within the unsafe interval. | |
378 | return RoundWeed(buffer, *length, DiyFp::Minus(too_high, w).f(), | |
379 | unsafe_interval.f(), rest, | |
380 | static_cast<uint64_t>(divisor) << -one.e(), unit); | |
381 | } | |
382 | divisor /= 10; | |
383 | } | |
384 | ||
385 | // The integrals have been generated. We are at the point of the decimal | |
386 | // separator. In the following loop we simply multiply the remaining digits by | |
387 | // 10 and divide by one. We just need to pay attention to multiply associated | |
388 | // data (like the interval or 'unit'), too. | |
389 | // Note that the multiplication by 10 does not overflow, because w.e >= -60 | |
390 | // and thus one.e >= -60. | |
391 | ASSERT(one.e() >= -60); | |
392 | ASSERT(fractionals < one.f()); | |
393 | ASSERT(UINT64_2PART_C(0xFFFFFFFF, FFFFFFFF) / 10 >= one.f()); | |
394 | for (;;) { | |
395 | fractionals *= 10; | |
396 | unit *= 10; | |
397 | unsafe_interval.set_f(unsafe_interval.f() * 10); | |
398 | // Integer division by one. | |
399 | int digit = static_cast<int>(fractionals >> -one.e()); | |
400 | ASSERT(digit <= 9); | |
401 | buffer[*length] = static_cast<char>('0' + digit); | |
402 | (*length)++; | |
403 | fractionals &= one.f() - 1; // Modulo by one. | |
404 | (*kappa)--; | |
405 | if (fractionals < unsafe_interval.f()) { | |
406 | return RoundWeed(buffer, *length, DiyFp::Minus(too_high, w).f() * unit, | |
407 | unsafe_interval.f(), fractionals, one.f(), unit); | |
408 | } | |
409 | } | |
410 | } | |
411 | ||
412 | ||
413 | ||
414 | // Generates (at most) requested_digits digits of input number w. | |
415 | // w is a floating-point number (DiyFp), consisting of a significand and an | |
416 | // exponent. Its exponent is bounded by kMinimalTargetExponent and | |
417 | // kMaximalTargetExponent. | |
418 | // Hence -60 <= w.e() <= -32. | |
419 | // | |
420 | // Returns false if it fails, in which case the generated digits in the buffer | |
421 | // should not be used. | |
422 | // Preconditions: | |
423 | // * w is correct up to 1 ulp (unit in the last place). That | |
424 | // is, its error must be strictly less than a unit of its last digit. | |
425 | // * kMinimalTargetExponent <= w.e() <= kMaximalTargetExponent | |
426 | // | |
427 | // Postconditions: returns false if procedure fails. | |
428 | // otherwise: | |
429 | // * buffer is not null-terminated, but length contains the number of | |
430 | // digits. | |
431 | // * the representation in buffer is the most precise representation of | |
432 | // requested_digits digits. | |
433 | // * buffer contains at most requested_digits digits of w. If there are less | |
434 | // than requested_digits digits then some trailing '0's have been removed. | |
435 | // * kappa is such that | |
436 | // w = buffer * 10^kappa + eps with |eps| < 10^kappa / 2. | |
437 | // | |
438 | // Remark: This procedure takes into account the imprecision of its input | |
439 | // numbers. If the precision is not enough to guarantee all the postconditions | |
440 | // then false is returned. This usually happens rarely, but the failure-rate | |
441 | // increases with higher requested_digits. | |
442 | static bool DigitGenCounted(DiyFp w, | |
443 | int requested_digits, | |
444 | Vector<char> buffer, | |
445 | int* length, | |
446 | int* kappa) { | |
447 | ASSERT(kMinimalTargetExponent <= w.e() && w.e() <= kMaximalTargetExponent); | |
448 | ASSERT(kMinimalTargetExponent >= -60); | |
449 | ASSERT(kMaximalTargetExponent <= -32); | |
450 | // w is assumed to have an error less than 1 unit. Whenever w is scaled we | |
451 | // also scale its error. | |
452 | uint64_t w_error = 1; | |
453 | // We cut the input number into two parts: the integral digits and the | |
454 | // fractional digits. We don't emit any decimal separator, but adapt kappa | |
455 | // instead. Example: instead of writing "1.2" we put "12" into the buffer and | |
456 | // increase kappa by 1. | |
457 | DiyFp one = DiyFp(static_cast<uint64_t>(1) << -w.e(), w.e()); | |
458 | // Division by one is a shift. | |
459 | uint32_t integrals = static_cast<uint32_t>(w.f() >> -one.e()); | |
460 | // Modulo by one is an and. | |
461 | uint64_t fractionals = w.f() & (one.f() - 1); | |
462 | uint32_t divisor; | |
463 | int divisor_exponent_plus_one; | |
464 | BiggestPowerTen(integrals, DiyFp::kSignificandSize - (-one.e()), | |
465 | &divisor, &divisor_exponent_plus_one); | |
466 | *kappa = divisor_exponent_plus_one; | |
467 | *length = 0; | |
468 | ||
469 | // Loop invariant: buffer = w / 10^kappa (integer division) | |
470 | // The invariant holds for the first iteration: kappa has been initialized | |
471 | // with the divisor exponent + 1. And the divisor is the biggest power of ten | |
472 | // that is smaller than 'integrals'. | |
473 | while (*kappa > 0) { | |
474 | int digit = integrals / divisor; | |
475 | ASSERT(digit <= 9); | |
476 | buffer[*length] = static_cast<char>('0' + digit); | |
477 | (*length)++; | |
478 | requested_digits--; | |
479 | integrals %= divisor; | |
480 | (*kappa)--; | |
481 | // Note that kappa now equals the exponent of the divisor and that the | |
482 | // invariant thus holds again. | |
483 | if (requested_digits == 0) break; | |
484 | divisor /= 10; | |
485 | } | |
486 | ||
487 | if (requested_digits == 0) { | |
488 | uint64_t rest = | |
489 | (static_cast<uint64_t>(integrals) << -one.e()) + fractionals; | |
490 | return RoundWeedCounted(buffer, *length, rest, | |
491 | static_cast<uint64_t>(divisor) << -one.e(), w_error, | |
492 | kappa); | |
493 | } | |
494 | ||
495 | // The integrals have been generated. We are at the point of the decimal | |
496 | // separator. In the following loop we simply multiply the remaining digits by | |
497 | // 10 and divide by one. We just need to pay attention to multiply associated | |
498 | // data (the 'unit'), too. | |
499 | // Note that the multiplication by 10 does not overflow, because w.e >= -60 | |
500 | // and thus one.e >= -60. | |
501 | ASSERT(one.e() >= -60); | |
502 | ASSERT(fractionals < one.f()); | |
503 | ASSERT(UINT64_2PART_C(0xFFFFFFFF, FFFFFFFF) / 10 >= one.f()); | |
504 | while (requested_digits > 0 && fractionals > w_error) { | |
505 | fractionals *= 10; | |
506 | w_error *= 10; | |
507 | // Integer division by one. | |
508 | int digit = static_cast<int>(fractionals >> -one.e()); | |
509 | ASSERT(digit <= 9); | |
510 | buffer[*length] = static_cast<char>('0' + digit); | |
511 | (*length)++; | |
512 | requested_digits--; | |
513 | fractionals &= one.f() - 1; // Modulo by one. | |
514 | (*kappa)--; | |
515 | } | |
516 | if (requested_digits != 0) return false; | |
517 | return RoundWeedCounted(buffer, *length, fractionals, one.f(), w_error, | |
518 | kappa); | |
519 | } | |
520 | ||
521 | ||
522 | // Provides a decimal representation of v. | |
523 | // Returns true if it succeeds, otherwise the result cannot be trusted. | |
524 | // There will be *length digits inside the buffer (not null-terminated). | |
525 | // If the function returns true then | |
526 | // v == (double) (buffer * 10^decimal_exponent). | |
527 | // The digits in the buffer are the shortest representation possible: no | |
528 | // 0.09999999999999999 instead of 0.1. The shorter representation will even be | |
529 | // chosen even if the longer one would be closer to v. | |
530 | // The last digit will be closest to the actual v. That is, even if several | |
531 | // digits might correctly yield 'v' when read again, the closest will be | |
532 | // computed. | |
533 | static bool Grisu3(double v, | |
534 | FastDtoaMode mode, | |
535 | Vector<char> buffer, | |
536 | int* length, | |
537 | int* decimal_exponent) { | |
538 | DiyFp w = Double(v).AsNormalizedDiyFp(); | |
539 | // boundary_minus and boundary_plus are the boundaries between v and its | |
540 | // closest floating-point neighbors. Any number strictly between | |
541 | // boundary_minus and boundary_plus will round to v when convert to a double. | |
542 | // Grisu3 will never output representations that lie exactly on a boundary. | |
543 | DiyFp boundary_minus, boundary_plus; | |
544 | if (mode == FAST_DTOA_SHORTEST) { | |
545 | Double(v).NormalizedBoundaries(&boundary_minus, &boundary_plus); | |
546 | } else { | |
547 | ASSERT(mode == FAST_DTOA_SHORTEST_SINGLE); | |
548 | float single_v = static_cast<float>(v); | |
549 | Single(single_v).NormalizedBoundaries(&boundary_minus, &boundary_plus); | |
550 | } | |
551 | ASSERT(boundary_plus.e() == w.e()); | |
552 | DiyFp ten_mk; // Cached power of ten: 10^-k | |
553 | int mk; // -k | |
554 | int ten_mk_minimal_binary_exponent = | |
555 | kMinimalTargetExponent - (w.e() + DiyFp::kSignificandSize); | |
556 | int ten_mk_maximal_binary_exponent = | |
557 | kMaximalTargetExponent - (w.e() + DiyFp::kSignificandSize); | |
558 | PowersOfTenCache::GetCachedPowerForBinaryExponentRange( | |
559 | ten_mk_minimal_binary_exponent, | |
560 | ten_mk_maximal_binary_exponent, | |
561 | &ten_mk, &mk); | |
562 | ASSERT((kMinimalTargetExponent <= w.e() + ten_mk.e() + | |
563 | DiyFp::kSignificandSize) && | |
564 | (kMaximalTargetExponent >= w.e() + ten_mk.e() + | |
565 | DiyFp::kSignificandSize)); | |
566 | // Note that ten_mk is only an approximation of 10^-k. A DiyFp only contains a | |
567 | // 64 bit significand and ten_mk is thus only precise up to 64 bits. | |
568 | ||
569 | // The DiyFp::Times procedure rounds its result, and ten_mk is approximated | |
570 | // too. The variable scaled_w (as well as scaled_boundary_minus/plus) are now | |
571 | // off by a small amount. | |
572 | // In fact: scaled_w - w*10^k < 1ulp (unit in the last place) of scaled_w. | |
573 | // In other words: let f = scaled_w.f() and e = scaled_w.e(), then | |
574 | // (f-1) * 2^e < w*10^k < (f+1) * 2^e | |
575 | DiyFp scaled_w = DiyFp::Times(w, ten_mk); | |
576 | ASSERT(scaled_w.e() == | |
577 | boundary_plus.e() + ten_mk.e() + DiyFp::kSignificandSize); | |
578 | // In theory it would be possible to avoid some recomputations by computing | |
579 | // the difference between w and boundary_minus/plus (a power of 2) and to | |
580 | // compute scaled_boundary_minus/plus by subtracting/adding from | |
581 | // scaled_w. However the code becomes much less readable and the speed | |
582 | // enhancements are not terriffic. | |
583 | DiyFp scaled_boundary_minus = DiyFp::Times(boundary_minus, ten_mk); | |
584 | DiyFp scaled_boundary_plus = DiyFp::Times(boundary_plus, ten_mk); | |
585 | ||
586 | // DigitGen will generate the digits of scaled_w. Therefore we have | |
587 | // v == (double) (scaled_w * 10^-mk). | |
588 | // Set decimal_exponent == -mk and pass it to DigitGen. If scaled_w is not an | |
589 | // integer than it will be updated. For instance if scaled_w == 1.23 then | |
590 | // the buffer will be filled with "123" und the decimal_exponent will be | |
591 | // decreased by 2. | |
592 | int kappa; | |
593 | bool result = DigitGen(scaled_boundary_minus, scaled_w, scaled_boundary_plus, | |
594 | buffer, length, &kappa); | |
595 | *decimal_exponent = -mk + kappa; | |
596 | return result; | |
597 | } | |
598 | ||
599 | ||
600 | // The "counted" version of grisu3 (see above) only generates requested_digits | |
601 | // number of digits. This version does not generate the shortest representation, | |
602 | // and with enough requested digits 0.1 will at some point print as 0.9999999... | |
603 | // Grisu3 is too imprecise for real halfway cases (1.5 will not work) and | |
604 | // therefore the rounding strategy for halfway cases is irrelevant. | |
605 | static bool Grisu3Counted(double v, | |
606 | int requested_digits, | |
607 | Vector<char> buffer, | |
608 | int* length, | |
609 | int* decimal_exponent) { | |
610 | DiyFp w = Double(v).AsNormalizedDiyFp(); | |
611 | DiyFp ten_mk; // Cached power of ten: 10^-k | |
612 | int mk; // -k | |
613 | int ten_mk_minimal_binary_exponent = | |
614 | kMinimalTargetExponent - (w.e() + DiyFp::kSignificandSize); | |
615 | int ten_mk_maximal_binary_exponent = | |
616 | kMaximalTargetExponent - (w.e() + DiyFp::kSignificandSize); | |
617 | PowersOfTenCache::GetCachedPowerForBinaryExponentRange( | |
618 | ten_mk_minimal_binary_exponent, | |
619 | ten_mk_maximal_binary_exponent, | |
620 | &ten_mk, &mk); | |
621 | ASSERT((kMinimalTargetExponent <= w.e() + ten_mk.e() + | |
622 | DiyFp::kSignificandSize) && | |
623 | (kMaximalTargetExponent >= w.e() + ten_mk.e() + | |
624 | DiyFp::kSignificandSize)); | |
625 | // Note that ten_mk is only an approximation of 10^-k. A DiyFp only contains a | |
626 | // 64 bit significand and ten_mk is thus only precise up to 64 bits. | |
627 | ||
628 | // The DiyFp::Times procedure rounds its result, and ten_mk is approximated | |
629 | // too. The variable scaled_w (as well as scaled_boundary_minus/plus) are now | |
630 | // off by a small amount. | |
631 | // In fact: scaled_w - w*10^k < 1ulp (unit in the last place) of scaled_w. | |
632 | // In other words: let f = scaled_w.f() and e = scaled_w.e(), then | |
633 | // (f-1) * 2^e < w*10^k < (f+1) * 2^e | |
634 | DiyFp scaled_w = DiyFp::Times(w, ten_mk); | |
635 | ||
636 | // We now have (double) (scaled_w * 10^-mk). | |
637 | // DigitGen will generate the first requested_digits digits of scaled_w and | |
638 | // return together with a kappa such that scaled_w ~= buffer * 10^kappa. (It | |
639 | // will not always be exactly the same since DigitGenCounted only produces a | |
640 | // limited number of digits.) | |
641 | int kappa; | |
642 | bool result = DigitGenCounted(scaled_w, requested_digits, | |
643 | buffer, length, &kappa); | |
644 | *decimal_exponent = -mk + kappa; | |
645 | return result; | |
646 | } | |
647 | ||
648 | ||
649 | bool FastDtoa(double v, | |
650 | FastDtoaMode mode, | |
651 | int requested_digits, | |
652 | Vector<char> buffer, | |
653 | int* length, | |
654 | int* decimal_point) { | |
655 | ASSERT(v > 0); | |
656 | ASSERT(!Double(v).IsSpecial()); | |
657 | ||
658 | bool result = false; | |
659 | int decimal_exponent = 0; | |
660 | switch (mode) { | |
661 | case FAST_DTOA_SHORTEST: | |
662 | case FAST_DTOA_SHORTEST_SINGLE: | |
663 | result = Grisu3(v, mode, buffer, length, &decimal_exponent); | |
664 | break; | |
665 | case FAST_DTOA_PRECISION: | |
666 | result = Grisu3Counted(v, requested_digits, | |
667 | buffer, length, &decimal_exponent); | |
668 | break; | |
669 | default: | |
670 | UNREACHABLE(); | |
671 | } | |
672 | if (result) { | |
673 | *decimal_point = *length + decimal_exponent; | |
674 | buffer[*length] = '\0'; | |
675 | } | |
676 | return result; | |
677 | } | |
678 | ||
679 | } // namespace double_conversion | |
680 | ||
681 | // ICU PATCH: Close ICU namespace | |
682 | U_NAMESPACE_END | |
683 | #endif // ICU PATCH: close #if !UCONFIG_NO_FORMATTING |