1 #[cfg(test)] 2 mod tests; 3 4 use std::hash; 5 use std::iter; 6 use std::ops::Range; 7 8 use rustc_serialize::{Decodable, Encodable}; 9 use rustc_target::abi::Size; 10 use rustc_type_ir::{TyDecoder, TyEncoder}; 11 12 use super::AllocRange; 13 14 type Block = u64; 15 16 /// A bitmask where each bit refers to the byte with the same index. If the bit is `true`, the byte 17 /// is initialized. If it is `false` the byte is uninitialized. 18 /// The actual bits are only materialized when needed, and we try to keep this data lazy as long as 19 /// possible. Currently, if all the blocks have the same value, then the mask represents either a 20 /// fully initialized or fully uninitialized const allocation, so we can only store that single 21 /// value. 22 #[derive(Clone, Debug, Eq, PartialEq, TyEncodable, TyDecodable, Hash, HashStable)] 23 pub struct InitMask { 24 blocks: InitMaskBlocks, 25 len: Size, 26 } 27 28 #[derive(Clone, Debug, Eq, PartialEq, TyEncodable, TyDecodable, Hash, HashStable)] 29 enum InitMaskBlocks { 30 Lazy { 31 /// Whether the lazy init mask is fully initialized or uninitialized. 32 state: bool, 33 }, 34 Materialized(InitMaskMaterialized), 35 } 36 37 impl InitMask { new(size: Size, state: bool) -> Self38 pub fn new(size: Size, state: bool) -> Self { 39 // Blocks start lazily allocated, until we have to materialize them. 40 let blocks = InitMaskBlocks::Lazy { state }; 41 InitMask { len: size, blocks } 42 } 43 44 /// Checks whether the `range` is entirely initialized. 45 /// 46 /// Returns `Ok(())` if it's initialized. Otherwise returns a range of byte 47 /// indexes for the first contiguous span of the uninitialized access. 48 #[inline] is_range_initialized(&self, range: AllocRange) -> Result<(), AllocRange>49 pub fn is_range_initialized(&self, range: AllocRange) -> Result<(), AllocRange> { 50 let end = range.end(); 51 if end > self.len { 52 return Err(AllocRange::from(self.len..end)); 53 } 54 55 match self.blocks { 56 InitMaskBlocks::Lazy { state } => { 57 // Lazily allocated blocks represent the full mask, and cover the requested range by 58 // definition. 59 if state { Ok(()) } else { Err(range) } 60 } 61 InitMaskBlocks::Materialized(ref blocks) => { 62 blocks.is_range_initialized(range.start, end) 63 } 64 } 65 } 66 67 /// Sets a specified range to a value. If the range is out-of-bounds, the mask will grow to 68 /// accommodate it entirely. set_range(&mut self, range: AllocRange, new_state: bool)69 pub fn set_range(&mut self, range: AllocRange, new_state: bool) { 70 let start = range.start; 71 let end = range.end(); 72 73 let is_full_overwrite = start == Size::ZERO && end >= self.len; 74 75 // Optimize the cases of a full init/uninit state, while handling growth if needed. 76 match self.blocks { 77 InitMaskBlocks::Lazy { ref mut state } if is_full_overwrite => { 78 // This is fully overwriting the mask, and we'll still have a single initialization 79 // state: the blocks can stay lazy. 80 *state = new_state; 81 self.len = end; 82 } 83 InitMaskBlocks::Materialized(_) if is_full_overwrite => { 84 // This is also fully overwriting materialized blocks with a single initialization 85 // state: we'll have no need for these blocks anymore and can make them lazy. 86 self.blocks = InitMaskBlocks::Lazy { state: new_state }; 87 self.len = end; 88 } 89 InitMaskBlocks::Lazy { state } if state == new_state => { 90 // Here we're partially overwriting the mask but the initialization state doesn't 91 // change: the blocks can stay lazy. 92 if end > self.len { 93 self.len = end; 94 } 95 } 96 _ => { 97 // Otherwise, we have a partial overwrite that can result in a mix of initialization 98 // states, so we'll need materialized blocks. 99 let len = self.len; 100 let blocks = self.materialize_blocks(); 101 102 // There are 3 cases of interest here, if we have: 103 // 104 // [--------] 105 // ^ ^ 106 // 0 len 107 // 108 // 1) the range to set can be in-bounds: 109 // 110 // xxxx = [start, end] 111 // [--------] 112 // ^ ^ 113 // 0 len 114 // 115 // Here, we'll simply set the single `start` to `end` range. 116 // 117 // 2) the range to set can be partially out-of-bounds: 118 // 119 // xxxx = [start, end] 120 // [--------] 121 // ^ ^ 122 // 0 len 123 // 124 // We have 2 subranges to handle: 125 // - we'll set the existing `start` to `len` range. 126 // - we'll grow and set the `len` to `end` range. 127 // 128 // 3) the range to set can be fully out-of-bounds: 129 // 130 // ---xxxx = [start, end] 131 // [--------] 132 // ^ ^ 133 // 0 len 134 // 135 // Since we're growing the mask to a single `new_state` value, we consider the gap 136 // from `len` to `start` to be part of the range, and have a single subrange to 137 // handle: we'll grow and set the `len` to `end` range. 138 // 139 // Note that we have to materialize, set blocks, and grow the mask. We could 140 // therefore slightly optimize things in situations where these writes overlap. 141 // However, as of writing this, growing the mask doesn't happen in practice yet, so 142 // we don't do this micro-optimization. 143 144 if end <= len { 145 // Handle case 1. 146 blocks.set_range_inbounds(start, end, new_state); 147 } else { 148 if start < len { 149 // Handle the first subrange of case 2. 150 blocks.set_range_inbounds(start, len, new_state); 151 } 152 153 // Handle the second subrange of case 2, and case 3. 154 blocks.grow(len, end - len, new_state); // `Size` operation 155 self.len = end; 156 } 157 } 158 } 159 } 160 161 /// Materializes this mask's blocks when the mask is lazy. 162 #[inline] materialize_blocks(&mut self) -> &mut InitMaskMaterialized163 fn materialize_blocks(&mut self) -> &mut InitMaskMaterialized { 164 if let InitMaskBlocks::Lazy { state } = self.blocks { 165 self.blocks = InitMaskBlocks::Materialized(InitMaskMaterialized::new(self.len, state)); 166 } 167 168 let InitMaskBlocks::Materialized(ref mut blocks) = self.blocks else { 169 bug!("initmask blocks must be materialized here") 170 }; 171 blocks 172 } 173 174 /// Returns the initialization state at the specified in-bounds index. 175 #[inline] get(&self, idx: Size) -> bool176 pub fn get(&self, idx: Size) -> bool { 177 match self.blocks { 178 InitMaskBlocks::Lazy { state } => state, 179 InitMaskBlocks::Materialized(ref blocks) => blocks.get(idx), 180 } 181 } 182 } 183 184 /// The actual materialized blocks of the bitmask, when we can't keep the `InitMask` lazy. 185 // Note: for performance reasons when interning, some of the fields can be partially 186 // hashed. (see the `Hash` impl below for more details), so the impl is not derived. 187 #[derive(Clone, Debug, Eq, PartialEq, HashStable)] 188 struct InitMaskMaterialized { 189 blocks: Vec<Block>, 190 } 191 192 // `Block` is a `u64`, but it is a bitmask not a numeric value. If we were to just derive 193 // Encodable and Decodable we would apply varint encoding to the bitmasks, which is slower 194 // and also produces more output when the high bits of each `u64` are occupied. 195 // Note: There is probably a remaining optimization for masks that do not use an entire 196 // `Block`. 197 impl<E: TyEncoder> Encodable<E> for InitMaskMaterialized { encode(&self, encoder: &mut E)198 fn encode(&self, encoder: &mut E) { 199 encoder.emit_usize(self.blocks.len()); 200 for block in &self.blocks { 201 encoder.emit_raw_bytes(&block.to_le_bytes()); 202 } 203 } 204 } 205 206 // This implementation is deliberately not derived, see the matching `Encodable` impl. 207 impl<D: TyDecoder> Decodable<D> for InitMaskMaterialized { decode(decoder: &mut D) -> Self208 fn decode(decoder: &mut D) -> Self { 209 let num_blocks = decoder.read_usize(); 210 let mut blocks = Vec::with_capacity(num_blocks); 211 for _ in 0..num_blocks { 212 let bytes = decoder.read_raw_bytes(8); 213 let block = u64::from_le_bytes(bytes.try_into().unwrap()); 214 blocks.push(block); 215 } 216 InitMaskMaterialized { blocks } 217 } 218 } 219 220 // Const allocations are only hashed for interning. However, they can be large, making the hashing 221 // expensive especially since it uses `FxHash`: it's better suited to short keys, not potentially 222 // big buffers like the allocation's init mask. We can partially hash some fields when they're 223 // large. 224 impl hash::Hash for InitMaskMaterialized { hash<H: hash::Hasher>(&self, state: &mut H)225 fn hash<H: hash::Hasher>(&self, state: &mut H) { 226 const MAX_BLOCKS_TO_HASH: usize = super::MAX_BYTES_TO_HASH / std::mem::size_of::<Block>(); 227 const MAX_BLOCKS_LEN: usize = super::MAX_HASHED_BUFFER_LEN / std::mem::size_of::<Block>(); 228 229 // Partially hash the `blocks` buffer when it is large. To limit collisions with common 230 // prefixes and suffixes, we hash the length and some slices of the buffer. 231 let block_count = self.blocks.len(); 232 if block_count > MAX_BLOCKS_LEN { 233 // Hash the buffer's length. 234 block_count.hash(state); 235 236 // And its head and tail. 237 self.blocks[..MAX_BLOCKS_TO_HASH].hash(state); 238 self.blocks[block_count - MAX_BLOCKS_TO_HASH..].hash(state); 239 } else { 240 self.blocks.hash(state); 241 } 242 } 243 } 244 245 impl InitMaskMaterialized { 246 pub const BLOCK_SIZE: u64 = 64; 247 new(size: Size, state: bool) -> Self248 fn new(size: Size, state: bool) -> Self { 249 let mut m = InitMaskMaterialized { blocks: vec![] }; 250 m.grow(Size::ZERO, size, state); 251 m 252 } 253 254 #[inline] bit_index(bits: Size) -> (usize, usize)255 fn bit_index(bits: Size) -> (usize, usize) { 256 // BLOCK_SIZE is the number of bits that can fit in a `Block`. 257 // Each bit in a `Block` represents the initialization state of one byte of an allocation, 258 // so we use `.bytes()` here. 259 let bits = bits.bytes(); 260 let a = bits / Self::BLOCK_SIZE; 261 let b = bits % Self::BLOCK_SIZE; 262 (usize::try_from(a).unwrap(), usize::try_from(b).unwrap()) 263 } 264 265 #[inline] size_from_bit_index(block: impl TryInto<u64>, bit: impl TryInto<u64>) -> Size266 fn size_from_bit_index(block: impl TryInto<u64>, bit: impl TryInto<u64>) -> Size { 267 let block = block.try_into().ok().unwrap(); 268 let bit = bit.try_into().ok().unwrap(); 269 Size::from_bytes(block * Self::BLOCK_SIZE + bit) 270 } 271 272 /// Checks whether the `range` is entirely initialized. 273 /// 274 /// Returns `Ok(())` if it's initialized. Otherwise returns a range of byte 275 /// indexes for the first contiguous span of the uninitialized access. 276 #[inline] is_range_initialized(&self, start: Size, end: Size) -> Result<(), AllocRange>277 fn is_range_initialized(&self, start: Size, end: Size) -> Result<(), AllocRange> { 278 let uninit_start = self.find_bit(start, end, false); 279 280 match uninit_start { 281 Some(uninit_start) => { 282 let uninit_end = self.find_bit(uninit_start, end, true).unwrap_or(end); 283 Err(AllocRange::from(uninit_start..uninit_end)) 284 } 285 None => Ok(()), 286 } 287 } 288 set_range_inbounds(&mut self, start: Size, end: Size, new_state: bool)289 fn set_range_inbounds(&mut self, start: Size, end: Size, new_state: bool) { 290 let (block_a, bit_a) = Self::bit_index(start); 291 let (block_b, bit_b) = Self::bit_index(end); 292 if block_a == block_b { 293 // First set all bits except the first `bit_a`, 294 // then unset the last `64 - bit_b` bits. 295 let range = if bit_b == 0 { 296 u64::MAX << bit_a 297 } else { 298 (u64::MAX << bit_a) & (u64::MAX >> (64 - bit_b)) 299 }; 300 if new_state { 301 self.blocks[block_a] |= range; 302 } else { 303 self.blocks[block_a] &= !range; 304 } 305 return; 306 } 307 // across block boundaries 308 if new_state { 309 // Set `bit_a..64` to `1`. 310 self.blocks[block_a] |= u64::MAX << bit_a; 311 // Set `0..bit_b` to `1`. 312 if bit_b != 0 { 313 self.blocks[block_b] |= u64::MAX >> (64 - bit_b); 314 } 315 // Fill in all the other blocks (much faster than one bit at a time). 316 for block in (block_a + 1)..block_b { 317 self.blocks[block] = u64::MAX; 318 } 319 } else { 320 // Set `bit_a..64` to `0`. 321 self.blocks[block_a] &= !(u64::MAX << bit_a); 322 // Set `0..bit_b` to `0`. 323 if bit_b != 0 { 324 self.blocks[block_b] &= !(u64::MAX >> (64 - bit_b)); 325 } 326 // Fill in all the other blocks (much faster than one bit at a time). 327 for block in (block_a + 1)..block_b { 328 self.blocks[block] = 0; 329 } 330 } 331 } 332 333 #[inline] get(&self, i: Size) -> bool334 fn get(&self, i: Size) -> bool { 335 let (block, bit) = Self::bit_index(i); 336 (self.blocks[block] & (1 << bit)) != 0 337 } 338 grow(&mut self, len: Size, amount: Size, new_state: bool)339 fn grow(&mut self, len: Size, amount: Size, new_state: bool) { 340 if amount.bytes() == 0 { 341 return; 342 } 343 let unused_trailing_bits = 344 u64::try_from(self.blocks.len()).unwrap() * Self::BLOCK_SIZE - len.bytes(); 345 346 // If there's not enough capacity in the currently allocated blocks, allocate some more. 347 if amount.bytes() > unused_trailing_bits { 348 let additional_blocks = amount.bytes() / Self::BLOCK_SIZE + 1; 349 350 // We allocate the blocks to the correct value for the requested init state, so we won't 351 // have to manually set them with another write. 352 let block = if new_state { u64::MAX } else { 0 }; 353 self.blocks 354 .extend(iter::repeat(block).take(usize::try_from(additional_blocks).unwrap())); 355 } 356 357 // New blocks have already been set here, so we only need to set the unused trailing bits, 358 // if any. 359 if unused_trailing_bits > 0 { 360 let in_bounds_tail = Size::from_bytes(unused_trailing_bits); 361 self.set_range_inbounds(len, len + in_bounds_tail, new_state); // `Size` operation 362 } 363 } 364 365 /// Returns the index of the first bit in `start..end` (end-exclusive) that is equal to is_init. find_bit(&self, start: Size, end: Size, is_init: bool) -> Option<Size>366 fn find_bit(&self, start: Size, end: Size, is_init: bool) -> Option<Size> { 367 /// A fast implementation of `find_bit`, 368 /// which skips over an entire block at a time if it's all 0s (resp. 1s), 369 /// and finds the first 1 (resp. 0) bit inside a block using `trailing_zeros` instead of a loop. 370 /// 371 /// Note that all examples below are written with 8 (instead of 64) bit blocks for simplicity, 372 /// and with the least significant bit (and lowest block) first: 373 /// ```text 374 /// 00000000|00000000 375 /// ^ ^ ^ ^ 376 /// index: 0 7 8 15 377 /// ``` 378 /// Also, if not stated, assume that `is_init = true`, that is, we are searching for the first 1 bit. 379 fn find_bit_fast( 380 init_mask: &InitMaskMaterialized, 381 start: Size, 382 end: Size, 383 is_init: bool, 384 ) -> Option<Size> { 385 /// Search one block, returning the index of the first bit equal to `is_init`. 386 fn search_block( 387 bits: Block, 388 block: usize, 389 start_bit: usize, 390 is_init: bool, 391 ) -> Option<Size> { 392 // For the following examples, assume this function was called with: 393 // bits = 0b00111011 394 // start_bit = 3 395 // is_init = false 396 // Note that, for the examples in this function, the most significant bit is written first, 397 // which is backwards compared to the comments in `find_bit`/`find_bit_fast`. 398 399 // Invert bits so we're always looking for the first set bit. 400 // ! 0b00111011 401 // bits = 0b11000100 402 let bits = if is_init { bits } else { !bits }; 403 // Mask off unused start bits. 404 // 0b11000100 405 // & 0b11111000 406 // bits = 0b11000000 407 let bits = bits & (!0 << start_bit); 408 // Find set bit, if any. 409 // bit = trailing_zeros(0b11000000) 410 // bit = 6 411 if bits == 0 { 412 None 413 } else { 414 let bit = bits.trailing_zeros(); 415 Some(InitMaskMaterialized::size_from_bit_index(block, bit)) 416 } 417 } 418 419 if start >= end { 420 return None; 421 } 422 423 // Convert `start` and `end` to block indexes and bit indexes within each block. 424 // We must convert `end` to an inclusive bound to handle block boundaries correctly. 425 // 426 // For example: 427 // 428 // (a) 00000000|00000000 (b) 00000000| 429 // ^~~~~~~~~~~^ ^~~~~~~~~^ 430 // start end start end 431 // 432 // In both cases, the block index of `end` is 1. 433 // But we do want to search block 1 in (a), and we don't in (b). 434 // 435 // We subtract 1 from both end positions to make them inclusive: 436 // 437 // (a) 00000000|00000000 (b) 00000000| 438 // ^~~~~~~~~~^ ^~~~~~~^ 439 // start end_inclusive start end_inclusive 440 // 441 // For (a), the block index of `end_inclusive` is 1, and for (b), it's 0. 442 // This provides the desired behavior of searching blocks 0 and 1 for (a), 443 // and searching only block 0 for (b). 444 // There is no concern of overflows since we checked for `start >= end` above. 445 let (start_block, start_bit) = InitMaskMaterialized::bit_index(start); 446 let end_inclusive = Size::from_bytes(end.bytes() - 1); 447 let (end_block_inclusive, _) = InitMaskMaterialized::bit_index(end_inclusive); 448 449 // Handle first block: need to skip `start_bit` bits. 450 // 451 // We need to handle the first block separately, 452 // because there may be bits earlier in the block that should be ignored, 453 // such as the bit marked (1) in this example: 454 // 455 // (1) 456 // -|------ 457 // (c) 01000000|00000000|00000001 458 // ^~~~~~~~~~~~~~~~~~^ 459 // start end 460 if let Some(i) = 461 search_block(init_mask.blocks[start_block], start_block, start_bit, is_init) 462 { 463 // If the range is less than a block, we may find a matching bit after `end`. 464 // 465 // For example, we shouldn't successfully find bit (2), because it's after `end`: 466 // 467 // (2) 468 // -------| 469 // (d) 00000001|00000000|00000001 470 // ^~~~~^ 471 // start end 472 // 473 // An alternative would be to mask off end bits in the same way as we do for start bits, 474 // but performing this check afterwards is faster and simpler to implement. 475 if i < end { 476 return Some(i); 477 } else { 478 return None; 479 } 480 } 481 482 // Handle remaining blocks. 483 // 484 // We can skip over an entire block at once if it's all 0s (resp. 1s). 485 // The block marked (3) in this example is the first block that will be handled by this loop, 486 // and it will be skipped for that reason: 487 // 488 // (3) 489 // -------- 490 // (e) 01000000|00000000|00000001 491 // ^~~~~~~~~~~~~~~~~~^ 492 // start end 493 if start_block < end_block_inclusive { 494 // This loop is written in a specific way for performance. 495 // Notably: `..end_block_inclusive + 1` is used for an inclusive range instead of `..=end_block_inclusive`, 496 // and `.zip(start_block + 1..)` is used to track the index instead of `.enumerate().skip().take()`, 497 // because both alternatives result in significantly worse codegen. 498 // `end_block_inclusive + 1` is guaranteed not to wrap, because `end_block_inclusive <= end / BLOCK_SIZE`, 499 // and `BLOCK_SIZE` (the number of bits per block) will always be at least 8 (1 byte). 500 for (&bits, block) in init_mask.blocks[start_block + 1..end_block_inclusive + 1] 501 .iter() 502 .zip(start_block + 1..) 503 { 504 if let Some(i) = search_block(bits, block, 0, is_init) { 505 // If this is the last block, we may find a matching bit after `end`. 506 // 507 // For example, we shouldn't successfully find bit (4), because it's after `end`: 508 // 509 // (4) 510 // -------| 511 // (f) 00000001|00000000|00000001 512 // ^~~~~~~~~~~~~~~~~~^ 513 // start end 514 // 515 // As above with example (d), we could handle the end block separately and mask off end bits, 516 // but unconditionally searching an entire block at once and performing this check afterwards 517 // is faster and much simpler to implement. 518 if i < end { 519 return Some(i); 520 } else { 521 return None; 522 } 523 } 524 } 525 } 526 527 None 528 } 529 530 #[cfg_attr(not(debug_assertions), allow(dead_code))] 531 fn find_bit_slow( 532 init_mask: &InitMaskMaterialized, 533 start: Size, 534 end: Size, 535 is_init: bool, 536 ) -> Option<Size> { 537 (start..end).find(|&i| init_mask.get(i) == is_init) 538 } 539 540 let result = find_bit_fast(self, start, end, is_init); 541 542 debug_assert_eq!( 543 result, 544 find_bit_slow(self, start, end, is_init), 545 "optimized implementation of find_bit is wrong for start={:?} end={:?} is_init={} init_mask={:#?}", 546 start, 547 end, 548 is_init, 549 self 550 ); 551 552 result 553 } 554 } 555 556 /// A contiguous chunk of initialized or uninitialized memory. 557 pub enum InitChunk { 558 Init(Range<Size>), 559 Uninit(Range<Size>), 560 } 561 562 impl InitChunk { 563 #[inline] is_init(&self) -> bool564 pub fn is_init(&self) -> bool { 565 match self { 566 Self::Init(_) => true, 567 Self::Uninit(_) => false, 568 } 569 } 570 571 #[inline] range(&self) -> Range<Size>572 pub fn range(&self) -> Range<Size> { 573 match self { 574 Self::Init(r) => r.clone(), 575 Self::Uninit(r) => r.clone(), 576 } 577 } 578 } 579 580 impl InitMask { 581 /// Returns an iterator, yielding a range of byte indexes for each contiguous region 582 /// of initialized or uninitialized bytes inside the range `start..end` (end-exclusive). 583 /// 584 /// The iterator guarantees the following: 585 /// - Chunks are nonempty. 586 /// - Chunks are adjacent (each range's start is equal to the previous range's end). 587 /// - Chunks span exactly `start..end` (the first starts at `start`, the last ends at `end`). 588 /// - Chunks alternate between [`InitChunk::Init`] and [`InitChunk::Uninit`]. 589 #[inline] range_as_init_chunks(&self, range: AllocRange) -> InitChunkIter<'_>590 pub fn range_as_init_chunks(&self, range: AllocRange) -> InitChunkIter<'_> { 591 let start = range.start; 592 let end = range.end(); 593 assert!(end <= self.len); 594 595 let is_init = if start < end { 596 self.get(start) 597 } else { 598 // `start..end` is empty: there are no chunks, so use some arbitrary value 599 false 600 }; 601 602 InitChunkIter { init_mask: self, is_init, start, end } 603 } 604 } 605 606 /// Yields [`InitChunk`]s. See [`InitMask::range_as_init_chunks`]. 607 #[derive(Clone)] 608 pub struct InitChunkIter<'a> { 609 init_mask: &'a InitMask, 610 /// Whether the next chunk we will return is initialized. 611 /// If there are no more chunks, contains some arbitrary value. 612 is_init: bool, 613 /// The current byte index into `init_mask`. 614 start: Size, 615 /// The end byte index into `init_mask`. 616 end: Size, 617 } 618 619 impl<'a> Iterator for InitChunkIter<'a> { 620 type Item = InitChunk; 621 622 #[inline] next(&mut self) -> Option<Self::Item>623 fn next(&mut self) -> Option<Self::Item> { 624 if self.start >= self.end { 625 return None; 626 } 627 628 let end_of_chunk = match self.init_mask.blocks { 629 InitMaskBlocks::Lazy { .. } => { 630 // If we're iterating over the chunks of lazy blocks, we just emit a single 631 // full-size chunk. 632 self.end 633 } 634 InitMaskBlocks::Materialized(ref blocks) => { 635 let end_of_chunk = 636 blocks.find_bit(self.start, self.end, !self.is_init).unwrap_or(self.end); 637 end_of_chunk 638 } 639 }; 640 let range = self.start..end_of_chunk; 641 let ret = 642 Some(if self.is_init { InitChunk::Init(range) } else { InitChunk::Uninit(range) }); 643 644 self.is_init = !self.is_init; 645 self.start = end_of_chunk; 646 647 ret 648 } 649 } 650 651 /// Run-length encoding of the uninit mask. 652 /// Used to copy parts of a mask multiple times to another allocation. 653 pub struct InitCopy { 654 /// Whether the first range is initialized. 655 initial: bool, 656 /// The lengths of ranges that are run-length encoded. 657 /// The initialization state of the ranges alternate starting with `initial`. 658 ranges: smallvec::SmallVec<[u64; 1]>, 659 } 660 661 impl InitCopy { no_bytes_init(&self) -> bool662 pub fn no_bytes_init(&self) -> bool { 663 // The `ranges` are run-length encoded and of alternating initialization state. 664 // So if `ranges.len() > 1` then the second block is an initialized range. 665 !self.initial && self.ranges.len() == 1 666 } 667 } 668 669 /// Transferring the initialization mask to other allocations. 670 impl InitMask { 671 /// Creates a run-length encoding of the initialization mask; panics if range is empty. 672 /// 673 /// This is essentially a more space-efficient version of 674 /// `InitMask::range_as_init_chunks(...).collect::<Vec<_>>()`. prepare_copy(&self, range: AllocRange) -> InitCopy675 pub fn prepare_copy(&self, range: AllocRange) -> InitCopy { 676 // Since we are copying `size` bytes from `src` to `dest + i * size` (`for i in 0..repeat`), 677 // a naive initialization mask copying algorithm would repeatedly have to read the initialization mask from 678 // the source and write it to the destination. Even if we optimized the memory accesses, 679 // we'd be doing all of this `repeat` times. 680 // Therefore we precompute a compressed version of the initialization mask of the source value and 681 // then write it back `repeat` times without computing any more information from the source. 682 683 // A precomputed cache for ranges of initialized / uninitialized bits 684 // 0000010010001110 will become 685 // `[5, 1, 2, 1, 3, 3, 1]`, 686 // where each element toggles the state. 687 688 let mut ranges = smallvec::SmallVec::<[u64; 1]>::new(); 689 690 let mut chunks = self.range_as_init_chunks(range).peekable(); 691 692 let initial = chunks.peek().expect("range should be nonempty").is_init(); 693 694 // Here we rely on `range_as_init_chunks` to yield alternating init/uninit chunks. 695 for chunk in chunks { 696 let len = chunk.range().end.bytes() - chunk.range().start.bytes(); 697 ranges.push(len); 698 } 699 700 InitCopy { ranges, initial } 701 } 702 703 /// Applies multiple instances of the run-length encoding to the initialization mask. apply_copy(&mut self, defined: InitCopy, range: AllocRange, repeat: u64)704 pub fn apply_copy(&mut self, defined: InitCopy, range: AllocRange, repeat: u64) { 705 // An optimization where we can just overwrite an entire range of initialization bits if 706 // they are going to be uniformly `1` or `0`. If this happens to be a full-range overwrite, 707 // we won't need materialized blocks either. 708 if defined.ranges.len() <= 1 { 709 let start = range.start; 710 let end = range.start + range.size * repeat; // `Size` operations 711 self.set_range(AllocRange::from(start..end), defined.initial); 712 return; 713 } 714 715 // We're about to do one or more partial writes, so we ensure the blocks are materialized. 716 let blocks = self.materialize_blocks(); 717 718 for mut j in 0..repeat { 719 j *= range.size.bytes(); 720 j += range.start.bytes(); 721 let mut cur = defined.initial; 722 for range in &defined.ranges { 723 let old_j = j; 724 j += range; 725 blocks.set_range_inbounds(Size::from_bytes(old_j), Size::from_bytes(j), cur); 726 cur = !cur; 727 } 728 } 729 } 730 } 731