|
| 1 | +/// These functions compute the integer square root of their type, assuming |
| 2 | +/// that someone has already checked that the value is nonnegative. |
| 3 | +
|
| 4 | +const ISQRT_AND_REMAINDER_8_BIT: [(u8, u8); 256] = { |
| 5 | + let mut result = [(0, 0); 256]; |
| 6 | + |
| 7 | + let mut sqrt = 0; |
| 8 | + let mut i = 0; |
| 9 | + 'outer: loop { |
| 10 | + let mut remaining = 2 * sqrt + 1; |
| 11 | + while remaining > 0 { |
| 12 | + result[i as usize] = (sqrt, 2 * sqrt + 1 - remaining); |
| 13 | + i += 1; |
| 14 | + if i >= result.len() { |
| 15 | + break 'outer; |
| 16 | + } |
| 17 | + remaining -= 1; |
| 18 | + } |
| 19 | + sqrt += 1; |
| 20 | + } |
| 21 | + |
| 22 | + result |
| 23 | +}; |
| 24 | + |
| 25 | +// `#[inline(always)]` because the programmer-accessible functions will use |
| 26 | +// this internally and the contents of this should be inlined there. |
| 27 | +#[inline(always)] |
| 28 | +pub const fn u8(n: u8) -> u8 { |
| 29 | + ISQRT_AND_REMAINDER_8_BIT[n as usize].0 |
| 30 | +} |
| 31 | + |
| 32 | +#[inline(always)] |
| 33 | +const fn intermediate_u8(n: u8) -> (u8, u8) { |
| 34 | + ISQRT_AND_REMAINDER_8_BIT[n as usize] |
| 35 | +} |
| 36 | + |
| 37 | +macro_rules! karatsuba_isqrt { |
| 38 | + ($FullBitsT:ty, $fn:ident, $intermediate_fn:ident, $HalfBitsT:ty, $half_fn:ident, $intermediate_half_fn:ident) => { |
| 39 | + // `#[inline(always)]` because the programmer-accessible functions will |
| 40 | + // use this internally and the contents of this should be inlined |
| 41 | + // there. |
| 42 | + #[inline(always)] |
| 43 | + pub const fn $fn(mut n: $FullBitsT) -> $FullBitsT { |
| 44 | + // Performs a Karatsuba square root. |
| 45 | + // https://web.archive.org/web/20230511212802/https://inria.hal.science/inria-00072854v1/file/RR-3805.pdf |
| 46 | + |
| 47 | + const HALF_BITS: u32 = <$FullBitsT>::BITS >> 1; |
| 48 | + const QUARTER_BITS: u32 = <$FullBitsT>::BITS >> 2; |
| 49 | + |
| 50 | + let leading_zeros = n.leading_zeros(); |
| 51 | + let result = if leading_zeros >= HALF_BITS { |
| 52 | + $half_fn(n as $HalfBitsT) as $FullBitsT |
| 53 | + } else { |
| 54 | + // Either the most-significant bit or its neighbor must be a one, so we shift left to make that happen. |
| 55 | + let precondition_shift = leading_zeros & (HALF_BITS - 2); |
| 56 | + n <<= precondition_shift; |
| 57 | + |
| 58 | + let hi = (n >> HALF_BITS) as $HalfBitsT; |
| 59 | + let lo = n & (<$HalfBitsT>::MAX as $FullBitsT); |
| 60 | + |
| 61 | + let (s_prime, r_prime) = $intermediate_half_fn(hi); |
| 62 | + |
| 63 | + let numerator = ((r_prime as $FullBitsT) << QUARTER_BITS) | (lo >> QUARTER_BITS); |
| 64 | + let denominator = (s_prime as $FullBitsT) << 1; |
| 65 | + |
| 66 | + let q = numerator / denominator; |
| 67 | + let u = numerator % denominator; |
| 68 | + |
| 69 | + let mut s = (s_prime << QUARTER_BITS) as $FullBitsT + q; |
| 70 | + if ((u << QUARTER_BITS) | (lo & ((1 << QUARTER_BITS) - 1))) < q * q { |
| 71 | + s -= 1; |
| 72 | + } |
| 73 | + s >> (precondition_shift >> 1) |
| 74 | + }; |
| 75 | + |
| 76 | + result |
| 77 | + } |
| 78 | + |
| 79 | + const fn $intermediate_fn(mut n: $FullBitsT) -> ($FullBitsT, $FullBitsT) { |
| 80 | + // Performs a Karatsuba square root. |
| 81 | + // https://web.archive.org/web/20230511212802/https://inria.hal.science/inria-00072854v1/file/RR-3805.pdf |
| 82 | + |
| 83 | + const HALF_BITS: u32 = <$FullBitsT>::BITS >> 1; |
| 84 | + const QUARTER_BITS: u32 = <$FullBitsT>::BITS >> 2; |
| 85 | + |
| 86 | + let leading_zeros = n.leading_zeros(); |
| 87 | + let result = if leading_zeros >= HALF_BITS { |
| 88 | + let (s, r) = $intermediate_half_fn(n as $HalfBitsT); |
| 89 | + (s as $FullBitsT, r as $FullBitsT) |
| 90 | + } else { |
| 91 | + // Either the most-significant bit or its neighbor must be a one, so we shift left to make that happen. |
| 92 | + let precondition_shift = leading_zeros & (HALF_BITS - 2); |
| 93 | + n <<= precondition_shift; |
| 94 | + |
| 95 | + let hi = (n >> HALF_BITS) as $HalfBitsT; |
| 96 | + let lo = n & (<$HalfBitsT>::MAX as $FullBitsT); |
| 97 | + |
| 98 | + let (s_prime, r_prime) = $intermediate_half_fn(hi); |
| 99 | + |
| 100 | + let numerator = ((r_prime as $FullBitsT) << QUARTER_BITS) | (lo >> QUARTER_BITS); |
| 101 | + let denominator = (s_prime as $FullBitsT) << 1; |
| 102 | + |
| 103 | + let q = numerator / denominator; |
| 104 | + let u = numerator % denominator; |
| 105 | + |
| 106 | + let mut s = (s_prime << QUARTER_BITS) as $FullBitsT + q; |
| 107 | + let (mut r, overflow) = |
| 108 | + ((u << QUARTER_BITS) | (lo & ((1 << QUARTER_BITS) - 1))).overflowing_sub(q * q); |
| 109 | + if overflow { |
| 110 | + r = r.wrapping_add((s << 1) - 1); |
| 111 | + s -= 1; |
| 112 | + } |
| 113 | + (s >> (precondition_shift >> 1), r >> (precondition_shift >> 1)) |
| 114 | + }; |
| 115 | + |
| 116 | + result |
| 117 | + } |
| 118 | + }; |
| 119 | +} |
| 120 | + |
| 121 | +karatsuba_isqrt!(u16, u16, intermediate_u16, u8, u8, intermediate_u8); |
| 122 | +karatsuba_isqrt!(u32, u32, intermediate_u32, u16, u16, intermediate_u16); |
| 123 | +karatsuba_isqrt!(u64, u64, intermediate_u64, u32, u32, intermediate_u32); |
| 124 | +karatsuba_isqrt!(u128, u128, _intermediate_u128, u64, u64, intermediate_u64); |
| 125 | + |
| 126 | +#[cfg(target_pointer_width = "16")] |
| 127 | +#[inline(always)] |
| 128 | +pub const fn usize(n: usize) -> usize { |
| 129 | + u16(n as u16) as usize |
| 130 | +} |
| 131 | + |
| 132 | +#[cfg(target_pointer_width = "32")] |
| 133 | +#[inline(always)] |
| 134 | +pub const fn usize(n: usize) -> usize { |
| 135 | + u32(n as u32) as usize |
| 136 | +} |
| 137 | + |
| 138 | +#[cfg(target_pointer_width = "64")] |
| 139 | +#[inline(always)] |
| 140 | +pub const fn usize(n: usize) -> usize { |
| 141 | + u64(n as u64) as usize |
| 142 | +} |
| 143 | + |
| 144 | +// 0 <= val <= i8::MAX |
| 145 | +#[inline(always)] |
| 146 | +pub const fn i8(n: i8) -> i8 { |
| 147 | + u8(n as u8) as i8 |
| 148 | +} |
| 149 | + |
| 150 | +// 0 <= val <= i16::MAX |
| 151 | +#[inline(always)] |
| 152 | +pub const fn i16(n: i16) -> i16 { |
| 153 | + u16(n as u16) as i16 |
| 154 | +} |
| 155 | + |
| 156 | +// 0 <= val <= i32::MAX |
| 157 | +#[inline(always)] |
| 158 | +pub const fn i32(n: i32) -> i32 { |
| 159 | + u32(n as u32) as i32 |
| 160 | +} |
| 161 | + |
| 162 | +// 0 <= val <= i64::MAX |
| 163 | +#[inline(always)] |
| 164 | +pub const fn i64(n: i64) -> i64 { |
| 165 | + u64(n as u64) as i64 |
| 166 | +} |
| 167 | + |
| 168 | +// 0 <= val <= i128::MAX |
| 169 | +#[inline(always)] |
| 170 | +pub const fn i128(n: i128) -> i128 { |
| 171 | + u128(n as u128) as i128 |
| 172 | +} |
| 173 | + |
| 174 | +/* |
| 175 | +This function is not used. |
| 176 | +
|
| 177 | +// 0 <= val <= isize::MAX |
| 178 | +#[inline(always)] |
| 179 | +pub const fn isize(n: isize) -> isize { |
| 180 | + usize(n as usize) as isize |
| 181 | +} |
| 182 | +*/ |
| 183 | + |
| 184 | +/// Instantiate this panic logic once, rather than for all the ilog methods |
| 185 | +/// on every single primitive type. |
| 186 | +#[cold] |
| 187 | +#[track_caller] |
| 188 | +pub const fn panic_for_negative_argument() -> ! { |
| 189 | + panic!("argument of integer square root cannot be negative") |
| 190 | +} |
0 commit comments