|
| 1 | +use aoc_runner_derive::aoc; |
| 2 | + |
| 3 | +// Transforms b, g, r, u ,w to 0, 1, 4, 3, 5 |
| 4 | +#[inline(always)] |
| 5 | +fn to_idx(i: u8) -> usize { |
| 6 | + let o: usize; |
| 7 | + unsafe { |
| 8 | + std::arch::asm!( |
| 9 | + "mov {o:e}, 19", |
| 10 | + "pext {o:e}, {i:e}, {o:e}", |
| 11 | + "add {o:r}, -2", |
| 12 | + i = in(reg) i as u32, |
| 13 | + o = out(reg) o, |
| 14 | + options(pure, nomem, nostack) |
| 15 | + ); |
| 16 | + } |
| 17 | + o |
| 18 | +} |
| 19 | + |
| 20 | +#[derive(Debug, Copy, Clone)] |
| 21 | +enum NfaTrans { |
| 22 | + None, |
| 23 | + Start, |
| 24 | + Next(usize), |
| 25 | + Both(usize), |
| 26 | +} |
| 27 | + |
| 28 | +impl NfaTrans { |
| 29 | + fn add_start(&mut self) { |
| 30 | + match *self { |
| 31 | + NfaTrans::None => *self = NfaTrans::Start, |
| 32 | + NfaTrans::Next(n) => *self = NfaTrans::Both(n), |
| 33 | + _ => {} |
| 34 | + } |
| 35 | + } |
| 36 | + |
| 37 | + fn add_or_foolow(&mut self, new_nfa_node: impl FnOnce() -> usize) -> usize { |
| 38 | + match *self { |
| 39 | + NfaTrans::None => { |
| 40 | + let new = new_nfa_node(); |
| 41 | + *self = NfaTrans::Next(new); |
| 42 | + new |
| 43 | + } |
| 44 | + NfaTrans::Start => { |
| 45 | + let new = new_nfa_node(); |
| 46 | + *self = NfaTrans::Both(new); |
| 47 | + new |
| 48 | + } |
| 49 | + NfaTrans::Next(n) => n, |
| 50 | + NfaTrans::Both(n) => n, |
| 51 | + } |
| 52 | + } |
| 53 | +} |
| 54 | + |
| 55 | +const NFA_SIZE: usize = 1024; |
| 56 | + |
| 57 | +#[aoc(day19, part1)] |
| 58 | +pub fn part1(s: &str) -> u64 { |
| 59 | + #[expect(unused_unsafe)] |
| 60 | + unsafe { |
| 61 | + inner_part1(s.as_bytes()) |
| 62 | + } |
| 63 | +} |
| 64 | + |
| 65 | +// #[target_feature(enable = "avx2,bmi1,bmi2,cmpxchg16b,lzcnt,movbe,popcnt")] |
| 66 | +fn inner_part1(s: &[u8]) -> u64 { |
| 67 | + let mut nfa = heapless::Vec::<[NfaTrans; 6], NFA_SIZE>::new(); |
| 68 | + nfa.push([NfaTrans::None; 6]).unwrap(); |
| 69 | + |
| 70 | + let mut i = 0; |
| 71 | + let mut nfa_node = 0; |
| 72 | + loop { |
| 73 | + let color = to_idx(s[i]); |
| 74 | + let next = s[i + 1]; |
| 75 | + |
| 76 | + if next == b',' || next == b'\n' { |
| 77 | + nfa[nfa_node][color].add_start(); |
| 78 | + |
| 79 | + nfa_node = 0; |
| 80 | + i += 3; |
| 81 | + if next == b'\n' { |
| 82 | + break; |
| 83 | + } |
| 84 | + } else { |
| 85 | + let mut nfa_trans = nfa[nfa_node][color]; |
| 86 | + let next_nfa_node = nfa_trans.add_or_foolow(|| { |
| 87 | + let new_nfa_node = nfa.len(); |
| 88 | + nfa.push([NfaTrans::None; 6]).unwrap(); |
| 89 | + new_nfa_node |
| 90 | + }); |
| 91 | + nfa[nfa_node][color] = nfa_trans; |
| 92 | + |
| 93 | + nfa_node = next_nfa_node; |
| 94 | + i += 1; |
| 95 | + } |
| 96 | + } |
| 97 | + |
| 98 | + let mut sum = 0; |
| 99 | + |
| 100 | + let mut states1 = &mut (true, heapless::Vec::<usize, NFA_SIZE>::new()); |
| 101 | + let mut states2 = &mut (false, heapless::Vec::<usize, NFA_SIZE>::new()); |
| 102 | + |
| 103 | + while i < s.len() { |
| 104 | + if s[i] == b'\n' { |
| 105 | + if states1.0 { |
| 106 | + sum += 1; |
| 107 | + } |
| 108 | + states1.1.clear(); |
| 109 | + states1.0 = true; |
| 110 | + i += 1; |
| 111 | + continue; |
| 112 | + } |
| 113 | + let color = to_idx(s[i]); |
| 114 | + |
| 115 | + states2.1.clear(); |
| 116 | + states2.0 = false; |
| 117 | + |
| 118 | + if states1.0 { |
| 119 | + let next = nfa[0][color]; |
| 120 | + match next { |
| 121 | + NfaTrans::None => {} |
| 122 | + NfaTrans::Start => { |
| 123 | + states2.0 = true; |
| 124 | + } |
| 125 | + NfaTrans::Next(n) => { |
| 126 | + states2.1.push(n).unwrap(); |
| 127 | + } |
| 128 | + NfaTrans::Both(n) => { |
| 129 | + states2.0 = true; |
| 130 | + states2.1.push(n).unwrap(); |
| 131 | + } |
| 132 | + } |
| 133 | + } |
| 134 | + for s in states1.1.iter() { |
| 135 | + let next = nfa[*s][color]; |
| 136 | + match next { |
| 137 | + NfaTrans::None => {} |
| 138 | + NfaTrans::Start => { |
| 139 | + states2.0 = true; |
| 140 | + } |
| 141 | + NfaTrans::Next(n) => { |
| 142 | + states2.1.push(n).unwrap(); |
| 143 | + } |
| 144 | + NfaTrans::Both(n) => { |
| 145 | + states2.0 = true; |
| 146 | + states2.1.push(n).unwrap(); |
| 147 | + } |
| 148 | + } |
| 149 | + } |
| 150 | + std::mem::swap(&mut states2, &mut states1); |
| 151 | + |
| 152 | + if states1.0 == false && states1.1.is_empty() { |
| 153 | + while i < s.len() && s[i] != b'\n' { |
| 154 | + i += 1; |
| 155 | + } |
| 156 | + } else { |
| 157 | + i += 1; |
| 158 | + } |
| 159 | + } |
| 160 | + sum |
| 161 | +} |
| 162 | + |
| 163 | +#[aoc(day19, part2)] |
| 164 | +pub fn part2(s: &str) -> u64 { |
| 165 | + #[expect(unused_unsafe)] |
| 166 | + unsafe { |
| 167 | + inner_part2(s.as_bytes()) |
| 168 | + } |
| 169 | +} |
| 170 | + |
| 171 | +// #[target_feature(enable = "avx2,bmi1,bmi2,cmpxchg16b,lzcnt,movbe,popcnt")] |
| 172 | +fn inner_part2(s: &[u8]) -> u64 { |
| 173 | + let mut nfa = heapless::Vec::<[NfaTrans; 6], NFA_SIZE>::new(); |
| 174 | + nfa.push([NfaTrans::None; 6]).unwrap(); |
| 175 | + |
| 176 | + let mut i = 0; |
| 177 | + let mut nfa_node = 0; |
| 178 | + loop { |
| 179 | + let color = to_idx(s[i]); |
| 180 | + let next = s[i + 1]; |
| 181 | + |
| 182 | + if next == b',' || next == b'\n' { |
| 183 | + nfa[nfa_node][color].add_start(); |
| 184 | + |
| 185 | + nfa_node = 0; |
| 186 | + i += 3; |
| 187 | + if next == b'\n' { |
| 188 | + break; |
| 189 | + } |
| 190 | + } else { |
| 191 | + let mut nfa_trans = nfa[nfa_node][color]; |
| 192 | + let next_nfa_node = nfa_trans.add_or_foolow(|| { |
| 193 | + let new_nfa_node = nfa.len(); |
| 194 | + nfa.push([NfaTrans::None; 6]).unwrap(); |
| 195 | + new_nfa_node |
| 196 | + }); |
| 197 | + nfa[nfa_node][color] = nfa_trans; |
| 198 | + |
| 199 | + nfa_node = next_nfa_node; |
| 200 | + i += 1; |
| 201 | + } |
| 202 | + } |
| 203 | + |
| 204 | + let mut sum = 0; |
| 205 | + |
| 206 | + let mut states1 = &mut (1, heapless::Vec::<(usize, u64), NFA_SIZE>::new()); |
| 207 | + let mut states2 = &mut (0, heapless::Vec::<(usize, u64), NFA_SIZE>::new()); |
| 208 | + |
| 209 | + while i < s.len() { |
| 210 | + if s[i] == b'\n' { |
| 211 | + sum += states1.0; |
| 212 | + states1.1.clear(); |
| 213 | + states1.0 = 1; |
| 214 | + i += 1; |
| 215 | + continue; |
| 216 | + } |
| 217 | + let color = to_idx(s[i]); |
| 218 | + |
| 219 | + states2.1.clear(); |
| 220 | + states2.0 = 0; |
| 221 | + |
| 222 | + if states1.0 > 0 { |
| 223 | + let next = nfa[0][color]; |
| 224 | + match next { |
| 225 | + NfaTrans::None => {} |
| 226 | + NfaTrans::Start => { |
| 227 | + states2.0 += states1.0; |
| 228 | + } |
| 229 | + NfaTrans::Next(n) => { |
| 230 | + states2.1.push((n, states1.0)).unwrap(); |
| 231 | + } |
| 232 | + NfaTrans::Both(n) => { |
| 233 | + states2.0 += states1.0; |
| 234 | + states2.1.push((n, states1.0)).unwrap(); |
| 235 | + } |
| 236 | + } |
| 237 | + } |
| 238 | + for (s, amount) in states1.1.iter() { |
| 239 | + let next = nfa[*s][color]; |
| 240 | + match next { |
| 241 | + NfaTrans::None => {} |
| 242 | + NfaTrans::Start => { |
| 243 | + states2.0 += *amount; |
| 244 | + } |
| 245 | + NfaTrans::Next(n) => { |
| 246 | + states2.1.push((n, *amount)).unwrap(); |
| 247 | + } |
| 248 | + NfaTrans::Both(n) => { |
| 249 | + states2.0 += *amount; |
| 250 | + states2.1.push((n, *amount)).unwrap(); |
| 251 | + } |
| 252 | + } |
| 253 | + } |
| 254 | + |
| 255 | + std::mem::swap(&mut states2, &mut states1); |
| 256 | + if states1.0 == 0 && states1.1.is_empty() { |
| 257 | + while i < s.len() && s[i] != b'\n' { |
| 258 | + i += 1; |
| 259 | + } |
| 260 | + } else { |
| 261 | + i += 1; |
| 262 | + } |
| 263 | + } |
| 264 | + sum |
| 265 | +} |
| 266 | + |
| 267 | +#[cfg(test)] |
| 268 | +mod tests { |
| 269 | + use super::*; |
| 270 | + |
| 271 | + const EXAMPLE: &str = r"r, wr, b, g, bwu, rb, gb, br |
| 272 | +
|
| 273 | +brwrr |
| 274 | +bggr |
| 275 | +gbbr |
| 276 | +rrbgbr |
| 277 | +ubwu |
| 278 | +bwurrg |
| 279 | +brgr |
| 280 | +bbrgwb |
| 281 | +"; |
| 282 | + |
| 283 | + #[test] |
| 284 | + fn example_part1() { |
| 285 | + assert_eq!(part1(EXAMPLE), 6); |
| 286 | + } |
| 287 | + |
| 288 | + #[test] |
| 289 | + fn example_part2() { |
| 290 | + assert_eq!(part2(EXAMPLE), 16); |
| 291 | + } |
| 292 | +} |
0 commit comments