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