/rust/registry/src/index.crates.io-6f17d22bba15001f/ryu-1.0.5/src/d2s.rs
Line | Count | Source (jump to first uncovered line) |
1 | | // Translated from C to Rust. The original C code can be found at |
2 | | // https://github.com/ulfjack/ryu and carries the following license: |
3 | | // |
4 | | // Copyright 2018 Ulf Adams |
5 | | // |
6 | | // The contents of this file may be used under the terms of the Apache License, |
7 | | // Version 2.0. |
8 | | // |
9 | | // (See accompanying file LICENSE-Apache or copy at |
10 | | // http://www.apache.org/licenses/LICENSE-2.0) |
11 | | // |
12 | | // Alternatively, the contents of this file may be used under the terms of |
13 | | // the Boost Software License, Version 1.0. |
14 | | // (See accompanying file LICENSE-Boost or copy at |
15 | | // https://www.boost.org/LICENSE_1_0.txt) |
16 | | // |
17 | | // Unless required by applicable law or agreed to in writing, this software |
18 | | // is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY |
19 | | // KIND, either express or implied. |
20 | | |
21 | | use crate::common::*; |
22 | | #[cfg(not(feature = "small"))] |
23 | | pub use crate::d2s_full_table::*; |
24 | | use crate::d2s_intrinsics::*; |
25 | | #[cfg(feature = "small")] |
26 | | pub use crate::d2s_small_table::*; |
27 | | #[cfg(not(maybe_uninit))] |
28 | | use core::mem; |
29 | | #[cfg(maybe_uninit)] |
30 | | use core::mem::MaybeUninit; |
31 | | |
32 | | pub const DOUBLE_MANTISSA_BITS: u32 = 52; |
33 | | pub const DOUBLE_EXPONENT_BITS: u32 = 11; |
34 | | pub const DOUBLE_BIAS: i32 = 1023; |
35 | | pub const DOUBLE_POW5_INV_BITCOUNT: i32 = 125; |
36 | | pub const DOUBLE_POW5_BITCOUNT: i32 = 125; |
37 | | |
38 | | #[cfg_attr(feature = "no-panic", inline)] |
39 | 0 | pub fn decimal_length17(v: u64) -> u32 { |
40 | 0 | // This is slightly faster than a loop. |
41 | 0 | // The average output length is 16.38 digits, so we check high-to-low. |
42 | 0 | // Function precondition: v is not an 18, 19, or 20-digit number. |
43 | 0 | // (17 digits are sufficient for round-tripping.) |
44 | 0 | debug_assert!(v < 100000000000000000); |
45 | | |
46 | 0 | if v >= 10000000000000000 { |
47 | 0 | 17 |
48 | 0 | } else if v >= 1000000000000000 { |
49 | 0 | 16 |
50 | 0 | } else if v >= 100000000000000 { |
51 | 0 | 15 |
52 | 0 | } else if v >= 10000000000000 { |
53 | 0 | 14 |
54 | 0 | } else if v >= 1000000000000 { |
55 | 0 | 13 |
56 | 0 | } else if v >= 100000000000 { |
57 | 0 | 12 |
58 | 0 | } else if v >= 10000000000 { |
59 | 0 | 11 |
60 | 0 | } else if v >= 1000000000 { |
61 | 0 | 10 |
62 | 0 | } else if v >= 100000000 { |
63 | 0 | 9 |
64 | 0 | } else if v >= 10000000 { |
65 | 0 | 8 |
66 | 0 | } else if v >= 1000000 { |
67 | 0 | 7 |
68 | 0 | } else if v >= 100000 { |
69 | 0 | 6 |
70 | 0 | } else if v >= 10000 { |
71 | 0 | 5 |
72 | 0 | } else if v >= 1000 { |
73 | 0 | 4 |
74 | 0 | } else if v >= 100 { |
75 | 0 | 3 |
76 | 0 | } else if v >= 10 { |
77 | 0 | 2 |
78 | | } else { |
79 | 0 | 1 |
80 | | } |
81 | 0 | } |
82 | | |
83 | | // A floating decimal representing m * 10^e. |
84 | | pub struct FloatingDecimal64 { |
85 | | pub mantissa: u64, |
86 | | // Decimal exponent's range is -324 to 308 |
87 | | // inclusive, and can fit in i16 if needed. |
88 | | pub exponent: i32, |
89 | | } |
90 | | |
91 | | #[cfg_attr(feature = "no-panic", inline)] |
92 | 0 | pub fn d2d(ieee_mantissa: u64, ieee_exponent: u32) -> FloatingDecimal64 { |
93 | 0 | let (e2, m2) = if ieee_exponent == 0 { |
94 | 0 | ( |
95 | 0 | // We subtract 2 so that the bounds computation has 2 additional bits. |
96 | 0 | 1 - DOUBLE_BIAS - DOUBLE_MANTISSA_BITS as i32 - 2, |
97 | 0 | ieee_mantissa, |
98 | 0 | ) |
99 | | } else { |
100 | 0 | ( |
101 | 0 | ieee_exponent as i32 - DOUBLE_BIAS - DOUBLE_MANTISSA_BITS as i32 - 2, |
102 | 0 | (1u64 << DOUBLE_MANTISSA_BITS) | ieee_mantissa, |
103 | 0 | ) |
104 | | }; |
105 | 0 | let even = (m2 & 1) == 0; |
106 | 0 | let accept_bounds = even; |
107 | 0 |
|
108 | 0 | // Step 2: Determine the interval of valid decimal representations. |
109 | 0 | let mv = 4 * m2; |
110 | | // Implicit bool -> int conversion. True is 1, false is 0. |
111 | 0 | let mm_shift = (ieee_mantissa != 0 || ieee_exponent <= 1) as u32; |
112 | | // We would compute mp and mm like this: |
113 | | // uint64_t mp = 4 * m2 + 2; |
114 | | // uint64_t mm = mv - 1 - mm_shift; |
115 | | |
116 | | // Step 3: Convert to a decimal power base using 128-bit arithmetic. |
117 | | let mut vr: u64; |
118 | | let mut vp: u64; |
119 | | let mut vm: u64; |
120 | | #[cfg(not(maybe_uninit))] |
121 | | { |
122 | | vp = unsafe { mem::uninitialized() }; |
123 | | vm = unsafe { mem::uninitialized() }; |
124 | | } |
125 | | #[cfg(maybe_uninit)] |
126 | 0 | let mut vp_uninit: MaybeUninit<u64> = MaybeUninit::uninit(); |
127 | 0 | #[cfg(maybe_uninit)] |
128 | 0 | let mut vm_uninit: MaybeUninit<u64> = MaybeUninit::uninit(); |
129 | 0 | let e10: i32; |
130 | 0 | let mut vm_is_trailing_zeros = false; |
131 | 0 | let mut vr_is_trailing_zeros = false; |
132 | 0 | if e2 >= 0 { |
133 | | // I tried special-casing q == 0, but there was no effect on performance. |
134 | | // This expression is slightly faster than max(0, log10_pow2(e2) - 1). |
135 | 0 | let q = log10_pow2(e2) - (e2 > 3) as u32; |
136 | 0 | e10 = q as i32; |
137 | 0 | let k = DOUBLE_POW5_INV_BITCOUNT + pow5bits(q as i32) - 1; |
138 | 0 | let i = -e2 + q as i32 + k; |
139 | 0 | vr = unsafe { |
140 | 0 | mul_shift_all_64( |
141 | 0 | m2, |
142 | 0 | #[cfg(feature = "small")] |
143 | 0 | &compute_inv_pow5(q), |
144 | 0 | #[cfg(not(feature = "small"))] |
145 | 0 | { |
146 | 0 | debug_assert!(q < DOUBLE_POW5_INV_SPLIT.len() as u32); |
147 | 0 | DOUBLE_POW5_INV_SPLIT.get_unchecked(q as usize) |
148 | 0 | }, |
149 | 0 | i as u32, |
150 | 0 | #[cfg(maybe_uninit)] |
151 | 0 | { |
152 | 0 | vp_uninit.as_mut_ptr() |
153 | 0 | }, |
154 | 0 | #[cfg(not(maybe_uninit))] |
155 | 0 | { |
156 | 0 | &mut vp |
157 | 0 | }, |
158 | 0 | #[cfg(maybe_uninit)] |
159 | 0 | { |
160 | 0 | vm_uninit.as_mut_ptr() |
161 | 0 | }, |
162 | 0 | #[cfg(not(maybe_uninit))] |
163 | 0 | { |
164 | 0 | &mut vm |
165 | 0 | }, |
166 | 0 | mm_shift, |
167 | 0 | ) |
168 | 0 | }; |
169 | 0 | #[cfg(maybe_uninit)] |
170 | 0 | { |
171 | 0 | vp = unsafe { vp_uninit.assume_init() }; |
172 | 0 | vm = unsafe { vm_uninit.assume_init() }; |
173 | 0 | } |
174 | 0 | if q <= 21 { |
175 | | // This should use q <= 22, but I think 21 is also safe. Smaller values |
176 | | // may still be safe, but it's more difficult to reason about them. |
177 | | // Only one of mp, mv, and mm can be a multiple of 5, if any. |
178 | 0 | let mv_mod5 = (mv as u32).wrapping_sub(5u32.wrapping_mul(div5(mv) as u32)); |
179 | 0 | if mv_mod5 == 0 { |
180 | 0 | vr_is_trailing_zeros = multiple_of_power_of_5(mv, q); |
181 | 0 | } else if accept_bounds { |
182 | 0 | // Same as min(e2 + (~mm & 1), pow5_factor(mm)) >= q |
183 | 0 | // <=> e2 + (~mm & 1) >= q && pow5_factor(mm) >= q |
184 | 0 | // <=> true && pow5_factor(mm) >= q, since e2 >= q. |
185 | 0 | vm_is_trailing_zeros = multiple_of_power_of_5(mv - 1 - mm_shift as u64, q); |
186 | 0 | } else { |
187 | 0 | // Same as min(e2 + 1, pow5_factor(mp)) >= q. |
188 | 0 | vp -= multiple_of_power_of_5(mv + 2, q) as u64; |
189 | 0 | } |
190 | 0 | } |
191 | | } else { |
192 | | // This expression is slightly faster than max(0, log10_pow5(-e2) - 1). |
193 | 0 | let q = log10_pow5(-e2) - (-e2 > 1) as u32; |
194 | 0 | e10 = q as i32 + e2; |
195 | 0 | let i = -e2 - q as i32; |
196 | 0 | let k = pow5bits(i) - DOUBLE_POW5_BITCOUNT; |
197 | 0 | let j = q as i32 - k; |
198 | 0 | vr = unsafe { |
199 | 0 | mul_shift_all_64( |
200 | 0 | m2, |
201 | 0 | #[cfg(feature = "small")] |
202 | 0 | &compute_pow5(i as u32), |
203 | 0 | #[cfg(not(feature = "small"))] |
204 | 0 | { |
205 | 0 | debug_assert!(i < DOUBLE_POW5_SPLIT.len() as i32); |
206 | 0 | DOUBLE_POW5_SPLIT.get_unchecked(i as usize) |
207 | 0 | }, |
208 | 0 | j as u32, |
209 | 0 | #[cfg(maybe_uninit)] |
210 | 0 | { |
211 | 0 | vp_uninit.as_mut_ptr() |
212 | 0 | }, |
213 | 0 | #[cfg(not(maybe_uninit))] |
214 | 0 | { |
215 | 0 | &mut vp |
216 | 0 | }, |
217 | 0 | #[cfg(maybe_uninit)] |
218 | 0 | { |
219 | 0 | vm_uninit.as_mut_ptr() |
220 | 0 | }, |
221 | 0 | #[cfg(not(maybe_uninit))] |
222 | 0 | { |
223 | 0 | &mut vm |
224 | 0 | }, |
225 | 0 | mm_shift, |
226 | 0 | ) |
227 | 0 | }; |
228 | 0 | #[cfg(maybe_uninit)] |
229 | 0 | { |
230 | 0 | vp = unsafe { vp_uninit.assume_init() }; |
231 | 0 | vm = unsafe { vm_uninit.assume_init() }; |
232 | 0 | } |
233 | 0 | if q <= 1 { |
234 | | // {vr,vp,vm} is trailing zeros if {mv,mp,mm} has at least q trailing 0 bits. |
235 | | // mv = 4 * m2, so it always has at least two trailing 0 bits. |
236 | 0 | vr_is_trailing_zeros = true; |
237 | 0 | if accept_bounds { |
238 | 0 | // mm = mv - 1 - mm_shift, so it has 1 trailing 0 bit iff mm_shift == 1. |
239 | 0 | vm_is_trailing_zeros = mm_shift == 1; |
240 | 0 | } else { |
241 | 0 | // mp = mv + 2, so it always has at least one trailing 0 bit. |
242 | 0 | vp -= 1; |
243 | 0 | } |
244 | 0 | } else if q < 63 { |
245 | 0 | // TODO(ulfjack): Use a tighter bound here. |
246 | 0 | // We want to know if the full product has at least q trailing zeros. |
247 | 0 | // We need to compute min(p2(mv), p5(mv) - e2) >= q |
248 | 0 | // <=> p2(mv) >= q && p5(mv) - e2 >= q |
249 | 0 | // <=> p2(mv) >= q (because -e2 >= q) |
250 | 0 | vr_is_trailing_zeros = multiple_of_power_of_2(mv, q); |
251 | 0 | } |
252 | | } |
253 | | |
254 | | // Step 4: Find the shortest decimal representation in the interval of valid representations. |
255 | 0 | let mut removed = 0i32; |
256 | 0 | let mut last_removed_digit = 0u8; |
257 | | // On average, we remove ~2 digits. |
258 | 0 | let output = if vm_is_trailing_zeros || vr_is_trailing_zeros { |
259 | | // General case, which happens rarely (~0.7%). |
260 | | loop { |
261 | 0 | let vp_div10 = div10(vp); |
262 | 0 | let vm_div10 = div10(vm); |
263 | 0 | if vp_div10 <= vm_div10 { |
264 | 0 | break; |
265 | 0 | } |
266 | 0 | let vm_mod10 = (vm as u32).wrapping_sub(10u32.wrapping_mul(vm_div10 as u32)); |
267 | 0 | let vr_div10 = div10(vr); |
268 | 0 | let vr_mod10 = (vr as u32).wrapping_sub(10u32.wrapping_mul(vr_div10 as u32)); |
269 | 0 | vm_is_trailing_zeros &= vm_mod10 == 0; |
270 | 0 | vr_is_trailing_zeros &= last_removed_digit == 0; |
271 | 0 | last_removed_digit = vr_mod10 as u8; |
272 | 0 | vr = vr_div10; |
273 | 0 | vp = vp_div10; |
274 | 0 | vm = vm_div10; |
275 | 0 | removed += 1; |
276 | | } |
277 | 0 | if vm_is_trailing_zeros { |
278 | | loop { |
279 | 0 | let vm_div10 = div10(vm); |
280 | 0 | let vm_mod10 = (vm as u32).wrapping_sub(10u32.wrapping_mul(vm_div10 as u32)); |
281 | 0 | if vm_mod10 != 0 { |
282 | 0 | break; |
283 | 0 | } |
284 | 0 | let vp_div10 = div10(vp); |
285 | 0 | let vr_div10 = div10(vr); |
286 | 0 | let vr_mod10 = (vr as u32).wrapping_sub(10u32.wrapping_mul(vr_div10 as u32)); |
287 | 0 | vr_is_trailing_zeros &= last_removed_digit == 0; |
288 | 0 | last_removed_digit = vr_mod10 as u8; |
289 | 0 | vr = vr_div10; |
290 | 0 | vp = vp_div10; |
291 | 0 | vm = vm_div10; |
292 | 0 | removed += 1; |
293 | | } |
294 | 0 | } |
295 | 0 | if vr_is_trailing_zeros && last_removed_digit == 5 && vr % 2 == 0 { |
296 | 0 | // Round even if the exact number is .....50..0. |
297 | 0 | last_removed_digit = 4; |
298 | 0 | } |
299 | | // We need to take vr + 1 if vr is outside bounds or we need to round up. |
300 | 0 | vr + ((vr == vm && (!accept_bounds || !vm_is_trailing_zeros)) || last_removed_digit >= 5) |
301 | | as u64 |
302 | | } else { |
303 | | // Specialized for the common case (~99.3%). Percentages below are relative to this. |
304 | 0 | let mut round_up = false; |
305 | 0 | let vp_div100 = div100(vp); |
306 | 0 | let vm_div100 = div100(vm); |
307 | 0 | // Optimization: remove two digits at a time (~86.2%). |
308 | 0 | if vp_div100 > vm_div100 { |
309 | 0 | let vr_div100 = div100(vr); |
310 | 0 | let vr_mod100 = (vr as u32).wrapping_sub(100u32.wrapping_mul(vr_div100 as u32)); |
311 | 0 | round_up = vr_mod100 >= 50; |
312 | 0 | vr = vr_div100; |
313 | 0 | vp = vp_div100; |
314 | 0 | vm = vm_div100; |
315 | 0 | removed += 2; |
316 | 0 | } |
317 | | // Loop iterations below (approximately), without optimization above: |
318 | | // 0: 0.03%, 1: 13.8%, 2: 70.6%, 3: 14.0%, 4: 1.40%, 5: 0.14%, 6+: 0.02% |
319 | | // Loop iterations below (approximately), with optimization above: |
320 | | // 0: 70.6%, 1: 27.8%, 2: 1.40%, 3: 0.14%, 4+: 0.02% |
321 | | loop { |
322 | 0 | let vp_div10 = div10(vp); |
323 | 0 | let vm_div10 = div10(vm); |
324 | 0 | if vp_div10 <= vm_div10 { |
325 | 0 | break; |
326 | 0 | } |
327 | 0 | let vr_div10 = div10(vr); |
328 | 0 | let vr_mod10 = (vr as u32).wrapping_sub(10u32.wrapping_mul(vr_div10 as u32)); |
329 | 0 | round_up = vr_mod10 >= 5; |
330 | 0 | vr = vr_div10; |
331 | 0 | vp = vp_div10; |
332 | 0 | vm = vm_div10; |
333 | 0 | removed += 1; |
334 | | } |
335 | | // We need to take vr + 1 if vr is outside bounds or we need to round up. |
336 | 0 | vr + (vr == vm || round_up) as u64 |
337 | | }; |
338 | 0 | let exp = e10 + removed; |
339 | 0 |
|
340 | 0 | FloatingDecimal64 { |
341 | 0 | exponent: exp, |
342 | 0 | mantissa: output, |
343 | 0 | } |
344 | 0 | } |