|
5 | 5 | //! |
6 | 6 | //! Unstable sorting is compatible with libcore because it doesn't allocate memory, unlike our |
7 | 7 | //! stable sorting implementation. |
| 8 | +//! |
| 9 | +//! In addition it also contains the core logic of the stable sort used by `slice::sort` based on |
| 10 | +//! TimSort. |
| 11 | +
|
| 12 | +#![allow(unused)] // FIXME debug |
8 | 13 |
|
9 | 14 | use crate::cmp; |
10 | 15 | use crate::mem::{self, MaybeUninit, SizedTypeProperties}; |
@@ -883,6 +888,7 @@ fn partition_at_index_loop<'a, T, F>( |
883 | 888 | } |
884 | 889 | } |
885 | 890 |
|
| 891 | +/// Reorder the slice such that the element at `index` is at its final sorted position. |
886 | 892 | pub fn partition_at_index<T, F>( |
887 | 893 | v: &mut [T], |
888 | 894 | index: usize, |
@@ -927,3 +933,491 @@ where |
927 | 933 | let pivot = &mut pivot[0]; |
928 | 934 | (left, pivot, right) |
929 | 935 | } |
| 936 | + |
| 937 | +/// Inserts `v[0]` into pre-sorted sequence `v[1..]` so that whole `v[..]` becomes sorted. |
| 938 | +/// |
| 939 | +/// This is the integral subroutine of insertion sort. |
| 940 | +fn insert_head<T, F>(v: &mut [T], is_less: &mut F) |
| 941 | +where |
| 942 | + F: FnMut(&T, &T) -> bool, |
| 943 | +{ |
| 944 | + if v.len() >= 2 && is_less(&v[1], &v[0]) { |
| 945 | + unsafe { |
| 946 | + // There are three ways to implement insertion here: |
| 947 | + // |
| 948 | + // 1. Swap adjacent elements until the first one gets to its final destination. |
| 949 | + // However, this way we copy data around more than is necessary. If elements are big |
| 950 | + // structures (costly to copy), this method will be slow. |
| 951 | + // |
| 952 | + // 2. Iterate until the right place for the first element is found. Then shift the |
| 953 | + // elements succeeding it to make room for it and finally place it into the |
| 954 | + // remaining hole. This is a good method. |
| 955 | + // |
| 956 | + // 3. Copy the first element into a temporary variable. Iterate until the right place |
| 957 | + // for it is found. As we go along, copy every traversed element into the slot |
| 958 | + // preceding it. Finally, copy data from the temporary variable into the remaining |
| 959 | + // hole. This method is very good. Benchmarks demonstrated slightly better |
| 960 | + // performance than with the 2nd method. |
| 961 | + // |
| 962 | + // All methods were benchmarked, and the 3rd showed best results. So we chose that one. |
| 963 | + let tmp = mem::ManuallyDrop::new(ptr::read(&v[0])); |
| 964 | + |
| 965 | + // Intermediate state of the insertion process is always tracked by `hole`, which |
| 966 | + // serves two purposes: |
| 967 | + // 1. Protects integrity of `v` from panics in `is_less`. |
| 968 | + // 2. Fills the remaining hole in `v` in the end. |
| 969 | + // |
| 970 | + // Panic safety: |
| 971 | + // |
| 972 | + // If `is_less` panics at any point during the process, `hole` will get dropped and |
| 973 | + // fill the hole in `v` with `tmp`, thus ensuring that `v` still holds every object it |
| 974 | + // initially held exactly once. |
| 975 | + let mut hole = InsertionHole { src: &*tmp, dest: &mut v[1] }; |
| 976 | + ptr::copy_nonoverlapping(&v[1], &mut v[0], 1); |
| 977 | + |
| 978 | + for i in 2..v.len() { |
| 979 | + if !is_less(&v[i], &*tmp) { |
| 980 | + break; |
| 981 | + } |
| 982 | + ptr::copy_nonoverlapping(&v[i], &mut v[i - 1], 1); |
| 983 | + hole.dest = &mut v[i]; |
| 984 | + } |
| 985 | + // `hole` gets dropped and thus copies `tmp` into the remaining hole in `v`. |
| 986 | + } |
| 987 | + } |
| 988 | + |
| 989 | + // When dropped, copies from `src` into `dest`. |
| 990 | + struct InsertionHole<T> { |
| 991 | + src: *const T, |
| 992 | + dest: *mut T, |
| 993 | + } |
| 994 | + |
| 995 | + impl<T> Drop for InsertionHole<T> { |
| 996 | + fn drop(&mut self) { |
| 997 | + unsafe { |
| 998 | + ptr::copy_nonoverlapping(self.src, self.dest, 1); |
| 999 | + } |
| 1000 | + } |
| 1001 | + } |
| 1002 | +} |
| 1003 | + |
| 1004 | +/// Merges non-decreasing runs `v[..mid]` and `v[mid..]` using `buf` as temporary storage, and |
| 1005 | +/// stores the result into `v[..]`. |
| 1006 | +/// |
| 1007 | +/// # Safety |
| 1008 | +/// |
| 1009 | +/// The two slices must be non-empty and `mid` must be in bounds. Buffer `buf` must be long enough |
| 1010 | +/// to hold a copy of the shorter slice. Also, `T` must not be a zero-sized type. |
| 1011 | +unsafe fn merge<T, F>(v: &mut [T], mid: usize, buf: *mut T, is_less: &mut F) |
| 1012 | +where |
| 1013 | + F: FnMut(&T, &T) -> bool, |
| 1014 | +{ |
| 1015 | + let len = v.len(); |
| 1016 | + let v = v.as_mut_ptr(); |
| 1017 | + let (v_mid, v_end) = unsafe { (v.add(mid), v.add(len)) }; |
| 1018 | + |
| 1019 | + // The merge process first copies the shorter run into `buf`. Then it traces the newly copied |
| 1020 | + // run and the longer run forwards (or backwards), comparing their next unconsumed elements and |
| 1021 | + // copying the lesser (or greater) one into `v`. |
| 1022 | + // |
| 1023 | + // As soon as the shorter run is fully consumed, the process is done. If the longer run gets |
| 1024 | + // consumed first, then we must copy whatever is left of the shorter run into the remaining |
| 1025 | + // hole in `v`. |
| 1026 | + // |
| 1027 | + // Intermediate state of the process is always tracked by `hole`, which serves two purposes: |
| 1028 | + // 1. Protects integrity of `v` from panics in `is_less`. |
| 1029 | + // 2. Fills the remaining hole in `v` if the longer run gets consumed first. |
| 1030 | + // |
| 1031 | + // Panic safety: |
| 1032 | + // |
| 1033 | + // If `is_less` panics at any point during the process, `hole` will get dropped and fill the |
| 1034 | + // hole in `v` with the unconsumed range in `buf`, thus ensuring that `v` still holds every |
| 1035 | + // object it initially held exactly once. |
| 1036 | + let mut hole; |
| 1037 | + |
| 1038 | + if mid <= len - mid { |
| 1039 | + // The left run is shorter. |
| 1040 | + unsafe { |
| 1041 | + ptr::copy_nonoverlapping(v, buf, mid); |
| 1042 | + hole = MergeHole { start: buf, end: buf.add(mid), dest: v }; |
| 1043 | + } |
| 1044 | + |
| 1045 | + // Initially, these pointers point to the beginnings of their arrays. |
| 1046 | + let left = &mut hole.start; |
| 1047 | + let mut right = v_mid; |
| 1048 | + let out = &mut hole.dest; |
| 1049 | + |
| 1050 | + while *left < hole.end && right < v_end { |
| 1051 | + // Consume the lesser side. |
| 1052 | + // If equal, prefer the left run to maintain stability. |
| 1053 | + unsafe { |
| 1054 | + let to_copy = if is_less(&*right, &**left) { |
| 1055 | + get_and_increment(&mut right) |
| 1056 | + } else { |
| 1057 | + get_and_increment(left) |
| 1058 | + }; |
| 1059 | + ptr::copy_nonoverlapping(to_copy, get_and_increment(out), 1); |
| 1060 | + } |
| 1061 | + } |
| 1062 | + } else { |
| 1063 | + // The right run is shorter. |
| 1064 | + unsafe { |
| 1065 | + ptr::copy_nonoverlapping(v_mid, buf, len - mid); |
| 1066 | + hole = MergeHole { start: buf, end: buf.add(len - mid), dest: v_mid }; |
| 1067 | + } |
| 1068 | + |
| 1069 | + // Initially, these pointers point past the ends of their arrays. |
| 1070 | + let left = &mut hole.dest; |
| 1071 | + let right = &mut hole.end; |
| 1072 | + let mut out = v_end; |
| 1073 | + |
| 1074 | + while v < *left && buf < *right { |
| 1075 | + // Consume the greater side. |
| 1076 | + // If equal, prefer the right run to maintain stability. |
| 1077 | + unsafe { |
| 1078 | + let to_copy = if is_less(&*right.sub(1), &*left.sub(1)) { |
| 1079 | + decrement_and_get(left) |
| 1080 | + } else { |
| 1081 | + decrement_and_get(right) |
| 1082 | + }; |
| 1083 | + ptr::copy_nonoverlapping(to_copy, decrement_and_get(&mut out), 1); |
| 1084 | + } |
| 1085 | + } |
| 1086 | + } |
| 1087 | + // Finally, `hole` gets dropped. If the shorter run was not fully consumed, whatever remains of |
| 1088 | + // it will now be copied into the hole in `v`. |
| 1089 | + |
| 1090 | + unsafe fn get_and_increment<T>(ptr: &mut *mut T) -> *mut T { |
| 1091 | + let old = *ptr; |
| 1092 | + *ptr = unsafe { ptr.add(1) }; |
| 1093 | + old |
| 1094 | + } |
| 1095 | + |
| 1096 | + unsafe fn decrement_and_get<T>(ptr: &mut *mut T) -> *mut T { |
| 1097 | + *ptr = unsafe { ptr.sub(1) }; |
| 1098 | + *ptr |
| 1099 | + } |
| 1100 | + |
| 1101 | + // When dropped, copies the range `start..end` into `dest..`. |
| 1102 | + struct MergeHole<T> { |
| 1103 | + start: *mut T, |
| 1104 | + end: *mut T, |
| 1105 | + dest: *mut T, |
| 1106 | + } |
| 1107 | + |
| 1108 | + impl<T> Drop for MergeHole<T> { |
| 1109 | + fn drop(&mut self) { |
| 1110 | + // `T` is not a zero-sized type, and these are pointers into a slice's elements. |
| 1111 | + unsafe { |
| 1112 | + let len = self.end.sub_ptr(self.start); |
| 1113 | + ptr::copy_nonoverlapping(self.start, self.dest, len); |
| 1114 | + } |
| 1115 | + } |
| 1116 | + } |
| 1117 | +} |
| 1118 | + |
| 1119 | +/// This merge sort borrows some (but not all) ideas from TimSort, which used to be described in |
| 1120 | +/// detail [here](https://github.com/python/cpython/blob/main/Objects/listsort.txt). However Python |
| 1121 | +/// has switched to a Powersort based implementation. |
| 1122 | +/// |
| 1123 | +/// The algorithm identifies strictly descending and non-descending subsequences, which are called |
| 1124 | +/// natural runs. There is a stack of pending runs yet to be merged. Each newly found run is pushed |
| 1125 | +/// onto the stack, and then some pairs of adjacent runs are merged until these two invariants are |
| 1126 | +/// satisfied: |
| 1127 | +/// |
| 1128 | +/// 1. for every `i` in `1..runs.len()`: `runs[i - 1].len > runs[i].len` |
| 1129 | +/// 2. for every `i` in `2..runs.len()`: `runs[i - 2].len > runs[i - 1].len + runs[i].len` |
| 1130 | +/// |
| 1131 | +/// The invariants ensure that the total running time is *O*(*n* \* log(*n*)) worst-case. |
| 1132 | +pub fn merge_sort<T, CmpF, ElemAllocF, ElemDeallocF, RunAllocF, RunDeallocF>( |
| 1133 | + v: &mut [T], |
| 1134 | + is_less: &mut CmpF, |
| 1135 | + elem_alloc_fn: ElemAllocF, |
| 1136 | + elem_dealloc_fn: ElemDeallocF, |
| 1137 | + run_alloc_fn: RunAllocF, |
| 1138 | + run_dealloc_fn: RunDeallocF, |
| 1139 | +) where |
| 1140 | + CmpF: FnMut(&T, &T) -> bool, |
| 1141 | + ElemAllocF: Fn(usize) -> *mut T, |
| 1142 | + ElemDeallocF: Fn(*mut T, usize), |
| 1143 | + RunAllocF: Fn(usize) -> *mut TimSortRun, |
| 1144 | + RunDeallocF: Fn(*mut TimSortRun, usize), |
| 1145 | +{ |
| 1146 | + // Slices of up to this length get sorted using insertion sort. |
| 1147 | + const MAX_INSERTION: usize = 20; |
| 1148 | + // Very short runs are extended using insertion sort to span at least this many elements. |
| 1149 | + const MIN_RUN: usize = 10; |
| 1150 | + |
| 1151 | + // The caller should have already checked that. |
| 1152 | + debug_assert!(!T::IS_ZST); |
| 1153 | + |
| 1154 | + let len = v.len(); |
| 1155 | + |
| 1156 | + // Short arrays get sorted in-place via insertion sort to avoid allocations. |
| 1157 | + if len <= MAX_INSERTION { |
| 1158 | + if len >= 2 { |
| 1159 | + for i in (0..len - 1).rev() { |
| 1160 | + insert_head(&mut v[i..], is_less); |
| 1161 | + } |
| 1162 | + } |
| 1163 | + return; |
| 1164 | + } |
| 1165 | + |
| 1166 | + // Allocate a buffer to use as scratch memory. We keep the length 0 so we can keep in it |
| 1167 | + // shallow copies of the contents of `v` without risking the dtors running on copies if |
| 1168 | + // `is_less` panics. When merging two sorted runs, this buffer holds a copy of the shorter run, |
| 1169 | + // which will always have length at most `len / 2`. |
| 1170 | + let mut buf = BufGuard::new(len / 2, elem_alloc_fn, elem_dealloc_fn); |
| 1171 | + let buf_ptr = buf.buf_ptr; |
| 1172 | + |
| 1173 | + let mut runs = RunVec::new(run_alloc_fn, run_dealloc_fn); |
| 1174 | + |
| 1175 | + // In order to identify natural runs in `v`, we traverse it backwards. That might seem like a |
| 1176 | + // strange decision, but consider the fact that merges more often go in the opposite direction |
| 1177 | + // (forwards). According to benchmarks, merging forwards is slightly faster than merging |
| 1178 | + // backwards. To conclude, identifying runs by traversing backwards improves performance. |
| 1179 | + let mut end = len; |
| 1180 | + while end > 0 { |
| 1181 | + // Find the next natural run, and reverse it if it's strictly descending. |
| 1182 | + let mut start = end - 1; |
| 1183 | + if start > 0 { |
| 1184 | + start -= 1; |
| 1185 | + unsafe { |
| 1186 | + if is_less(v.get_unchecked(start + 1), v.get_unchecked(start)) { |
| 1187 | + while start > 0 && is_less(v.get_unchecked(start), v.get_unchecked(start - 1)) { |
| 1188 | + start -= 1; |
| 1189 | + } |
| 1190 | + v[start..end].reverse(); |
| 1191 | + } else { |
| 1192 | + while start > 0 && !is_less(v.get_unchecked(start), v.get_unchecked(start - 1)) |
| 1193 | + { |
| 1194 | + start -= 1; |
| 1195 | + } |
| 1196 | + } |
| 1197 | + } |
| 1198 | + } |
| 1199 | + |
| 1200 | + // Insert some more elements into the run if it's too short. Insertion sort is faster than |
| 1201 | + // merge sort on short sequences, so this significantly improves performance. |
| 1202 | + while start > 0 && end - start < MIN_RUN { |
| 1203 | + start -= 1; |
| 1204 | + insert_head(&mut v[start..end], is_less); |
| 1205 | + } |
| 1206 | + |
| 1207 | + // Push this run onto the stack. |
| 1208 | + runs.push(TimSortRun { start, len: end - start }); |
| 1209 | + end = start; |
| 1210 | + |
| 1211 | + // Merge some pairs of adjacent runs to satisfy the invariants. |
| 1212 | + while let Some(r) = collapse(runs.as_slice()) { |
| 1213 | + let left = runs[r + 1]; |
| 1214 | + let right = runs[r]; |
| 1215 | + unsafe { |
| 1216 | + merge(&mut v[left.start..right.start + right.len], left.len, buf_ptr, is_less); |
| 1217 | + } |
| 1218 | + runs[r] = TimSortRun { start: left.start, len: left.len + right.len }; |
| 1219 | + runs.remove(r + 1); |
| 1220 | + } |
| 1221 | + } |
| 1222 | + |
| 1223 | + // Finally, exactly one run must remain in the stack. |
| 1224 | + debug_assert!(runs.len() == 1 && runs[0].start == 0 && runs[0].len == len); |
| 1225 | + |
| 1226 | + // Examines the stack of runs and identifies the next pair of runs to merge. More specifically, |
| 1227 | + // if `Some(r)` is returned, that means `runs[r]` and `runs[r + 1]` must be merged next. If the |
| 1228 | + // algorithm should continue building a new run instead, `None` is returned. |
| 1229 | + // |
| 1230 | + // TimSort is infamous for its buggy implementations, as described here: |
| 1231 | + // http://envisage-project.eu/timsort-specification-and-verification/ |
| 1232 | + // |
| 1233 | + // The gist of the story is: we must enforce the invariants on the top four runs on the stack. |
| 1234 | + // Enforcing them on just top three is not sufficient to ensure that the invariants will still |
| 1235 | + // hold for *all* runs in the stack. |
| 1236 | + // |
| 1237 | + // This function correctly checks invariants for the top four runs. Additionally, if the top |
| 1238 | + // run starts at index 0, it will always demand a merge operation until the stack is fully |
| 1239 | + // collapsed, in order to complete the sort. |
| 1240 | + #[inline] |
| 1241 | + fn collapse(runs: &[TimSortRun]) -> Option<usize> { |
| 1242 | + let n = runs.len(); |
| 1243 | + if n >= 2 |
| 1244 | + && (runs[n - 1].start == 0 |
| 1245 | + || runs[n - 2].len <= runs[n - 1].len |
| 1246 | + || (n >= 3 && runs[n - 3].len <= runs[n - 2].len + runs[n - 1].len) |
| 1247 | + || (n >= 4 && runs[n - 4].len <= runs[n - 3].len + runs[n - 2].len)) |
| 1248 | + { |
| 1249 | + if n >= 3 && runs[n - 3].len < runs[n - 1].len { Some(n - 3) } else { Some(n - 2) } |
| 1250 | + } else { |
| 1251 | + None |
| 1252 | + } |
| 1253 | + } |
| 1254 | + |
| 1255 | + // Extremely basic versions of Vec. |
| 1256 | + // Their use is super limited and by having the code here, it allows reuse between the sort |
| 1257 | + // implementations. |
| 1258 | + struct BufGuard<T, ElemAllocF, ElemDeallocF> |
| 1259 | + where |
| 1260 | + ElemAllocF: Fn(usize) -> *mut T, |
| 1261 | + ElemDeallocF: Fn(*mut T, usize), |
| 1262 | + { |
| 1263 | + buf_ptr: *mut T, |
| 1264 | + capacity: usize, |
| 1265 | + elem_alloc_fn: ElemAllocF, |
| 1266 | + elem_dealloc_fn: ElemDeallocF, |
| 1267 | + } |
| 1268 | + |
| 1269 | + impl<T, ElemAllocF, ElemDeallocF> BufGuard<T, ElemAllocF, ElemDeallocF> |
| 1270 | + where |
| 1271 | + ElemAllocF: Fn(usize) -> *mut T, |
| 1272 | + ElemDeallocF: Fn(*mut T, usize), |
| 1273 | + { |
| 1274 | + fn new(len: usize, elem_alloc_fn: ElemAllocF, elem_dealloc_fn: ElemDeallocF) -> Self { |
| 1275 | + Self { buf_ptr: elem_alloc_fn(len), capacity: len, elem_alloc_fn, elem_dealloc_fn } |
| 1276 | + } |
| 1277 | + } |
| 1278 | + |
| 1279 | + impl<T, ElemAllocF, ElemDeallocF> Drop for BufGuard<T, ElemAllocF, ElemDeallocF> |
| 1280 | + where |
| 1281 | + ElemAllocF: Fn(usize) -> *mut T, |
| 1282 | + ElemDeallocF: Fn(*mut T, usize), |
| 1283 | + { |
| 1284 | + fn drop(&mut self) { |
| 1285 | + (self.elem_dealloc_fn)(self.buf_ptr, self.capacity); |
| 1286 | + } |
| 1287 | + } |
| 1288 | + |
| 1289 | + struct RunVec<RunAllocF, RunDeallocF> |
| 1290 | + where |
| 1291 | + RunAllocF: Fn(usize) -> *mut TimSortRun, |
| 1292 | + RunDeallocF: Fn(*mut TimSortRun, usize), |
| 1293 | + { |
| 1294 | + buf_ptr: *mut TimSortRun, |
| 1295 | + capacity: usize, |
| 1296 | + len: usize, |
| 1297 | + run_alloc_fn: RunAllocF, |
| 1298 | + run_dealloc_fn: RunDeallocF, |
| 1299 | + } |
| 1300 | + |
| 1301 | + impl<RunAllocF, RunDeallocF> RunVec<RunAllocF, RunDeallocF> |
| 1302 | + where |
| 1303 | + RunAllocF: Fn(usize) -> *mut TimSortRun, |
| 1304 | + RunDeallocF: Fn(*mut TimSortRun, usize), |
| 1305 | + { |
| 1306 | + fn new(run_alloc_fn: RunAllocF, run_dealloc_fn: RunDeallocF) -> Self { |
| 1307 | + // Most slices can be sorted with at most 16 runs in-flight. |
| 1308 | + const START_RUN_CAPACITY: usize = 16; |
| 1309 | + |
| 1310 | + Self { |
| 1311 | + buf_ptr: run_alloc_fn(START_RUN_CAPACITY), |
| 1312 | + capacity: START_RUN_CAPACITY, |
| 1313 | + len: 0, |
| 1314 | + run_alloc_fn, |
| 1315 | + run_dealloc_fn, |
| 1316 | + } |
| 1317 | + } |
| 1318 | + |
| 1319 | + fn push(&mut self, val: TimSortRun) { |
| 1320 | + if self.len == self.capacity { |
| 1321 | + let old_capacity = self.capacity; |
| 1322 | + let old_buf_ptr = self.buf_ptr; |
| 1323 | + |
| 1324 | + self.capacity = self.capacity * 2; |
| 1325 | + self.buf_ptr = (self.run_alloc_fn)(self.capacity); |
| 1326 | + |
| 1327 | + // SAFETY: buf_ptr new and old were correctly allocated and old_buf_ptr has |
| 1328 | + // old_capacity valid elements. |
| 1329 | + unsafe { |
| 1330 | + ptr::copy_nonoverlapping(old_buf_ptr, self.buf_ptr, old_capacity); |
| 1331 | + } |
| 1332 | + |
| 1333 | + (self.run_dealloc_fn)(old_buf_ptr, old_capacity); |
| 1334 | + } |
| 1335 | + |
| 1336 | + // SAFETY: The invariant was just checked. |
| 1337 | + unsafe { |
| 1338 | + self.buf_ptr.add(self.len).write(val); |
| 1339 | + } |
| 1340 | + self.len += 1; |
| 1341 | + } |
| 1342 | + |
| 1343 | + fn remove(&mut self, index: usize) { |
| 1344 | + if index >= self.len { |
| 1345 | + panic!("Index out of bounds"); |
| 1346 | + } |
| 1347 | + |
| 1348 | + // SAFETY: buf_ptr needs to be valid and len invariant upheld. |
| 1349 | + unsafe { |
| 1350 | + // the place we are taking from. |
| 1351 | + let ptr = self.buf_ptr.add(index); |
| 1352 | + |
| 1353 | + // Shift everything down to fill in that spot. |
| 1354 | + ptr::copy(ptr.add(1), ptr, self.len - index - 1); |
| 1355 | + } |
| 1356 | + self.len -= 1; |
| 1357 | + } |
| 1358 | + |
| 1359 | + fn as_slice(&self) -> &[TimSortRun] { |
| 1360 | + // SAFETY: Safe as long as buf_ptr is valid and len invariant was upheld. |
| 1361 | + unsafe { &*ptr::slice_from_raw_parts(self.buf_ptr, self.len) } |
| 1362 | + } |
| 1363 | + |
| 1364 | + fn len(&self) -> usize { |
| 1365 | + self.len |
| 1366 | + } |
| 1367 | + } |
| 1368 | + |
| 1369 | + impl<RunAllocF, RunDeallocF> core::ops::Index<usize> for RunVec<RunAllocF, RunDeallocF> |
| 1370 | + where |
| 1371 | + RunAllocF: Fn(usize) -> *mut TimSortRun, |
| 1372 | + RunDeallocF: Fn(*mut TimSortRun, usize), |
| 1373 | + { |
| 1374 | + type Output = TimSortRun; |
| 1375 | + |
| 1376 | + fn index(&self, index: usize) -> &Self::Output { |
| 1377 | + if index < self.len { |
| 1378 | + // SAFETY: buf_ptr and len invariant must be upheld. |
| 1379 | + unsafe { |
| 1380 | + return &*(self.buf_ptr.add(index)); |
| 1381 | + } |
| 1382 | + } |
| 1383 | + |
| 1384 | + panic!("Index out of bounds"); |
| 1385 | + } |
| 1386 | + } |
| 1387 | + |
| 1388 | + impl<RunAllocF, RunDeallocF> core::ops::IndexMut<usize> for RunVec<RunAllocF, RunDeallocF> |
| 1389 | + where |
| 1390 | + RunAllocF: Fn(usize) -> *mut TimSortRun, |
| 1391 | + RunDeallocF: Fn(*mut TimSortRun, usize), |
| 1392 | + { |
| 1393 | + fn index_mut(&mut self, index: usize) -> &mut Self::Output { |
| 1394 | + if index < self.len { |
| 1395 | + // SAFETY: buf_ptr and len invariant must be upheld. |
| 1396 | + unsafe { |
| 1397 | + return &mut *(self.buf_ptr.add(index)); |
| 1398 | + } |
| 1399 | + } |
| 1400 | + |
| 1401 | + panic!("Index out of bounds"); |
| 1402 | + } |
| 1403 | + } |
| 1404 | + |
| 1405 | + impl<RunAllocF, RunDeallocF> Drop for RunVec<RunAllocF, RunDeallocF> |
| 1406 | + where |
| 1407 | + RunAllocF: Fn(usize) -> *mut TimSortRun, |
| 1408 | + RunDeallocF: Fn(*mut TimSortRun, usize), |
| 1409 | + { |
| 1410 | + fn drop(&mut self) { |
| 1411 | + // As long as TimSortRun is Copy we don't need to drop them individually but just the |
| 1412 | + // whole allocation. |
| 1413 | + (self.run_dealloc_fn)(self.buf_ptr, self.capacity); |
| 1414 | + } |
| 1415 | + } |
| 1416 | +} |
| 1417 | + |
| 1418 | +/// Internal type used by merge_sort. |
| 1419 | +#[derive(Clone, Copy, Debug)] |
| 1420 | +pub struct TimSortRun { |
| 1421 | + len: usize, |
| 1422 | + start: usize, |
| 1423 | +} |
0 commit comments