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