lexical-write-float 1.0.6

Efficient formatting of floats to strings.
Documentation
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
//! Adaptation of the V8 ftoa algorithm with a custom radix.
//!
//! This algorithm is adapted from the V8 codebase,
//! and may be found [here](https://github.com/v8/v8).
//!
//! # Unsupported Features
//!
//! This does not support a few features from the format packed struct,
//! most notably, it will never write numbers in scientific notation.
//! Scientific notation must be disabled.

#![cfg(feature = "radix")]
#![doc(hidden)]

use lexical_util::algorithm::{copy_to_dst, ltrim_char_count, rtrim_char_count};
use lexical_util::constants::{FormattedSize, BUFFER_SIZE};
use lexical_util::digit::{char_to_digit_const, digit_to_char_const};
use lexical_util::format::NumberFormat;
use lexical_util::num::Float;
use lexical_write_integer::write::WriteInteger;

use crate::options::{Options, RoundMode};
use crate::shared;

// ALGORITHM
// ---------

/// Naive float-to-string algorithm for generic radixes.
///
/// This assumes the float is:
///     1). Non-special (NaN or Infinite).
///     2). Non-negative.
///
/// # Panics
///
/// Panics if exponent notation is used.
#[allow(clippy::collapsible_if)] // reason="conditions are different logical concepts"
pub fn write_float<F: Float, const FORMAT: u128>(
    float: F,
    bytes: &mut [u8],
    options: &Options,
) -> usize
where
    <F as Float>::Unsigned: WriteInteger + FormattedSize,
{
    // PRECONDITIONS

    // Assert no special cases remain, no negative numbers, and a valid format.
    let format = NumberFormat::<{ FORMAT }> {};
    assert!(format.is_valid());
    debug_assert!(!float.is_special());
    debug_assert!(float >= F::ZERO);
    debug_assert!(F::BITS <= 64);

    // Validate our options: we don't support different exponent bases here.
    debug_assert!(format.mantissa_radix() == format.exponent_base());

    // Temporary buffer for the result. We start with the decimal point in the
    // middle and write to the left for the integer part and to the right for the
    // fractional part. 1024 characters for the exponent and 52 for the mantissa
    // either way, with additional space for sign, decimal point and string
    // termination should be sufficient.
    const SIZE: usize = 2200;
    let mut buffer = [0u8; SIZE];
    let initial_cursor: usize = SIZE / 2;
    let mut integer_cursor = initial_cursor;
    let mut fraction_cursor = initial_cursor;
    let base = F::as_cast(format.radix());

    // Split the float into an integer part and a fractional part.
    let mut integer = float.floor();
    let mut fraction = float - integer;

    // We only compute fractional digits up to the input double's precision.
    // This fails if the value is at f64::MAX. IF we take the next positive,
    // we'll get literal infinite. We don't care about NaN comparisons, since
    // the float **must** be finite, so do this.
    let mut delta = if float.to_bits() == F::MAX.to_bits() {
        F::as_cast(0.5) * (float - float.prev_positive())
    } else {
        F::as_cast(0.5) * (float.next_positive() - float)
    };
    delta = F::ZERO.next_positive().max_finite(delta);
    debug_assert!(delta > F::ZERO);

    // Write our fraction digits.
    // Won't panic since we have 1100 digits, which is enough for any float f64 or
    // smaller.
    if fraction > delta {
        loop {
            // Shift up by one digit.
            fraction *= base;
            delta *= base;
            // Write digit.
            let digit = fraction.as_u32();
            let c = digit_to_char_const(digit, format.radix());
            buffer[fraction_cursor] = c;
            fraction_cursor += 1;
            // Calculate remainder.
            fraction -= F::as_cast(digit);
            // Round to even.
            if fraction > F::as_cast(0.5) || (fraction == F::as_cast(0.5) && (digit & 1) != 0) {
                if fraction + delta > F::ONE {
                    // We need to back trace already written digits in case of carry-over.
                    loop {
                        fraction_cursor -= 1;
                        if fraction_cursor == initial_cursor - 1 {
                            // Carry over to the integer part.
                            integer += F::ONE;
                            break;
                        }
                        // Reconstruct digit.
                        let c = buffer[fraction_cursor];
                        if let Some(digit) = char_to_digit_const(c, format.radix()) {
                            let idx = digit + 1;
                            let c = digit_to_char_const(idx, format.radix());
                            buffer[fraction_cursor] = c;
                            fraction_cursor += 1;
                            break;
                        }
                    }
                    break;
                }
            }

            if delta >= fraction {
                break;
            }
        }
    }

    // Compute integer digits. Fill unrepresented digits with zero.
    // Won't panic we have 1100 digits, which is enough for any float f64 or
    // smaller. We do this first, so we can do extended precision control later.
    while (integer / base).exponent() > 0 {
        integer /= base;
        integer_cursor -= 1;
        buffer[integer_cursor] = b'0';
    }

    loop {
        let remainder = integer % base;
        integer_cursor -= 1;
        let idx = remainder.as_u32();
        let c = digit_to_char_const(idx, format.radix());
        buffer[integer_cursor] = c;
        integer = (integer - remainder) / base;

        if integer <= F::ZERO {
            break;
        }
    }

    // Get our exponent.
    // We can't use a naive float log algorithm, since rounding issues can
    // cause major issues. For example, `12157665459056928801f64` is `3^40`,
    // but glibc gives us `(f.ln() / 3.0.ln())` of `39.999`, while Android, and
    // MUSL libm, and openlibm give us `40.0`, the correct answer. This of
    // course means we have off-by-1 errors, so the correct way is to trim
    // leading zeros, and then calculate the exponent as the offset.
    let digits = &buffer[integer_cursor..fraction_cursor];
    let zero_count = ltrim_char_count(digits, b'0');
    let sci_exp: i32 = initial_cursor as i32 - integer_cursor as i32 - zero_count as i32 - 1;
    write_float!(
        float,
        FORMAT,
        sci_exp,
        options,
        write_float_scientific,
        write_float_nonscientific,
        write_float_nonscientific,
        bytes => bytes,
        args => sci_exp, &mut buffer, initial_cursor,
                integer_cursor, fraction_cursor, options,
    )
}

/// Write float to string in scientific notation.
///
/// # Preconditions
///
/// The mantissa must be truncated and rounded, prior to calling this,
/// based on the number of maximum digits. In addition, `exponent_base`
/// and `mantissa_radix` in `FORMAT` must be identical.
#[cfg_attr(not(feature = "compact"), inline(always))]
pub fn write_float_scientific<const FORMAT: u128>(
    bytes: &mut [u8],
    sci_exp: i32,
    buffer: &mut [u8],
    initial_cursor: usize,
    integer_cursor: usize,
    fraction_cursor: usize,
    options: &Options,
) -> usize {
    // PRECONDITIONS
    debug_assert!(bytes.len() >= BUFFER_SIZE);

    // Config options.
    let format = NumberFormat::<{ FORMAT }> {};
    assert!(format.is_valid());
    let decimal_point = options.decimal_point();

    // Round and truncate the number of significant digits.
    let start: usize = if sci_exp <= 0 {
        ((initial_cursor as i32) - sci_exp - 1) as usize
    } else {
        integer_cursor
    };
    let end = fraction_cursor.min(start + MAX_DIGIT_LENGTH + 1);
    let (mut digit_count, carried) =
        truncate_and_round(buffer, start, end, format.radix(), options);
    let digits = &buffer[start..start + digit_count];
    // If we carried, just adjust the exponent since we will always have a
    // `digit_count == 1`. This means we don't have to worry about any other
    // digits.
    let sci_exp = sci_exp + carried as i32;

    // Non-exponent portion.
    // Get as many digits as possible, up to `MAX_DIGIT_LENGTH+1`
    // since we are ignoring the digit for the first digit,
    // or the number of written digits.
    bytes[0] = digits[0];
    bytes[1] = decimal_point;
    let src = &digits[1..digit_count];
    let dst = &mut bytes[2..digit_count + 1];
    copy_to_dst(dst, src);
    let zeros = rtrim_char_count(&bytes[2..digit_count + 1], b'0');
    digit_count -= zeros;
    // Extra 1 since we have the decimal point.
    let mut cursor = digit_count + 1;

    // Determine if we need to add more trailing zeros.
    let exact_count = shared::min_exact_digits(digit_count, options);

    // Write any trailing digits to the output.
    // Won't panic since bytes cannot be empty.
    if !format.no_exponent_without_fraction() && cursor == 2 && options.trim_floats() {
        // Need to trim floats from trailing zeros, and we have only a decimal.
        cursor -= 1;
    } else if exact_count < 2 {
        // Need to have at least 1 digit, the trailing `.0`.
        bytes[cursor] = b'0';
        cursor += 1;
    } else if exact_count > digit_count {
        // NOTE: Neither `exact_count >= digit_count >= 2`.
        // We need to write `exact_count - (cursor - 1)` digits, since
        // cursor includes the decimal point.
        let digits_end = exact_count + 1;
        bytes[cursor..digits_end].fill(b'0');
        cursor = digits_end;
    }

    // Now, write our scientific notation.
    shared::write_exponent::<FORMAT>(bytes, &mut cursor, sci_exp, options.exponent());

    cursor
}

/// Write float to string without scientific notation.
#[cfg_attr(not(feature = "compact"), inline(always))]
pub fn write_float_nonscientific<const FORMAT: u128>(
    bytes: &mut [u8],
    _: i32,
    buffer: &mut [u8],
    initial_cursor: usize,
    integer_cursor: usize,
    fraction_cursor: usize,
    options: &Options,
) -> usize {
    // PRECONDITIONS
    debug_assert!(bytes.len() >= BUFFER_SIZE);

    // Config options.
    let format = NumberFormat::<{ FORMAT }> {};
    assert!(format.is_valid());
    let decimal_point = options.decimal_point();

    // Round and truncate the number of significant digits.
    let mut start = integer_cursor;
    let end = fraction_cursor.min(start + MAX_DIGIT_LENGTH + 1);
    let (mut digit_count, carried) =
        truncate_and_round(buffer, start, end, format.radix(), options);

    // Adjust the buffer if we carried.
    // Note that we can **only** carry if it overflowed through the integer
    // component, since we always write at least 1 integer digit.
    if carried {
        debug_assert!(digit_count == 1);
        start -= 1;
        buffer[start] = b'1';
    }

    let digits = &buffer[start..start + digit_count];

    // Write the integer component.
    let integer_length = initial_cursor - start;
    let integer_count = digit_count.min(integer_length);
    let src = &digits[..integer_count];
    let dst = &mut bytes[..integer_count];
    copy_to_dst(dst, src);
    if integer_count < integer_length {
        // We have more leading digits than digits we wrote: can write
        // any additional digits, and then just write the remaining zeros.
        bytes[integer_count..integer_length].fill(b'0');
    }
    let mut cursor = integer_length;
    bytes[cursor] = decimal_point;
    cursor += 1;

    // Write the fraction component.
    // We've only consumed `integer_count` digits, since this input
    // may have been truncated.
    // Won't panic since `integer_count < digits.len()` since `digit_count <
    // digits.len()`.
    let digits = &digits[integer_count..];
    let fraction_count = digit_count.saturating_sub(integer_length);
    if fraction_count > 0 {
        // Need to write additional fraction digits.
        let src = &digits[..fraction_count];
        let end = cursor + fraction_count;
        let dst = &mut bytes[cursor..end];
        copy_to_dst(dst, src);
        let zeros = rtrim_char_count(&bytes[cursor..end], b'0');
        cursor += fraction_count - zeros;
    } else if options.trim_floats() {
        // Remove the decimal point, went too far.
        cursor -= 1;
    } else {
        bytes[cursor] = b'0';
        cursor += 1;
        digit_count += 1;
    }

    // Determine if we need to add more trailing zeros.
    let exact_count = shared::min_exact_digits(digit_count, options);

    // Write any trailing digits to the output.
    // Won't panic since bytes cannot be empty.
    if (fraction_count > 0 || !options.trim_floats()) && exact_count > digit_count {
        // NOTE: Neither `exact_count >= digit_count >= 2`.
        // We need to write `exact_count - (cursor - 1)` digits, since
        // cursor includes the decimal point.
        let digits_end = cursor + exact_count - digit_count;
        bytes[cursor..digits_end].fill(b'0');
        cursor = digits_end;
    }

    cursor
}

// Store the first digit and up to `BUFFER_SIZE - 20` digits
// that occur from left-to-right in the decimal representation.
// For example, for the number 123.45, store the first digit `1`
// and `2345` as the remaining values. Then, decide on-the-fly
// if we need scientific or regular formatting.
//
//   BUFFER_SIZE
// - 1      # first digit
// - 1      # period
// - 1      # +/- sign
// - 2      # e and +/- sign
// - 9      # max exp is 308, in radix2 is 9
// - 1      # null terminator
// = 15 characters of formatting required
// Just pad it a bit, we don't want memory corruption.
const MAX_NONDIGIT_LENGTH: usize = 25;
const MAX_DIGIT_LENGTH: usize = BUFFER_SIZE - MAX_NONDIGIT_LENGTH;

/// Round mantissa to the nearest value, returning only the number
/// of significant digits. Returns the number of digits of the mantissa,
/// and if the rounding did a full carry.
#[cfg_attr(not(feature = "compact"), inline(always))]
#[allow(clippy::comparison_chain)] // reason="conditions are different logical concepts"
pub fn truncate_and_round(
    buffer: &mut [u8],
    start: usize,
    end: usize,
    radix: u32,
    options: &Options,
) -> (usize, bool) {
    debug_assert!(end >= start);
    debug_assert!(end <= buffer.len());

    // Get the number of max digits, and then calculate if we need to round.
    let digit_count = end - start;
    let max_digits = if let Some(digits) = options.max_significant_digits() {
        digits.get()
    } else {
        return (digit_count, false);
    };

    if max_digits >= digit_count {
        return (digit_count, false);
    }
    if options.round_mode() == RoundMode::Truncate {
        // Don't round input, just shorten number of digits emitted.
        return (max_digits, false);
    }

    // Need to add the number of leading zeros to the digits `digit_count`.
    let max_digits = {
        let digits = &mut buffer[start..start + max_digits];
        max_digits + ltrim_char_count(digits, b'0')
    };

    // We need to round-nearest, tie-even, so we need to handle
    // the truncation **here**. If the representation is above
    // halfway at all, we need to round up, even if 1 bit.
    let last = buffer[start + max_digits - 1];
    let first = buffer[start + max_digits];
    let halfway = digit_to_char_const(radix / 2, radix);
    let rem = radix % 2;
    if first < halfway {
        // Just truncate, going to round-down anyway.
        (max_digits, false)
    } else if first > halfway {
        // Round-up always.
        let digits = &mut buffer[start..start + max_digits];
        shared::round_up(digits, max_digits, radix)
    } else if rem == 0 {
        // Even radix, our halfway point `$c00000.....`.
        let truncated = &buffer[start + max_digits + 1..end];
        if truncated.iter().all(|&x| x == b'0') && last & 1 == 0 {
            // At an exact halfway point, and even, round-down.
            (max_digits, false)
        } else {
            // Above halfway or at halfway and even, round-up
            let digits = &mut buffer[start..start + max_digits];
            shared::round_up(digits, max_digits, radix)
        }
    } else {
        // Odd radix, our halfway point is `$c$c$c$c$c$c....`. Cannot halfway points.
        let truncated = &buffer[start + max_digits + 1..end];
        for &c in truncated.iter() {
            if c < halfway {
                return (max_digits, false);
            } else if c > halfway {
                // Above halfway
                let digits = &mut buffer[start..start + max_digits];
                return shared::round_up(digits, max_digits, radix);
            }
        }
        (max_digits, false)
    }
}