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