diff options
| -rw-r--r-- | compiler/rustc_middle/src/mir/interpret/allocation.rs | 695 |
1 files changed, 350 insertions, 345 deletions
diff --git a/compiler/rustc_middle/src/mir/interpret/allocation.rs b/compiler/rustc_middle/src/mir/interpret/allocation.rs index e636d7612b4..27a637f2f4f 100644 --- a/compiler/rustc_middle/src/mir/interpret/allocation.rs +++ b/compiler/rustc_middle/src/mir/interpret/allocation.rs @@ -495,122 +495,6 @@ impl<Tag: Copy, Extra> Allocation<Tag, Extra> { } } -/// Uninitialized bytes. -impl<Tag: Copy, Extra> Allocation<Tag, Extra> { - /// Checks whether the given range is entirely initialized. - /// - /// Returns `Ok(())` if it's initialized. Otherwise returns the range of byte - /// indexes of the first contiguous uninitialized access. - fn is_init(&self, range: AllocRange) -> Result<(), Range<Size>> { - self.init_mask.is_range_initialized(range.start, range.end()) // `Size` addition - } - - /// Checks that a range of bytes is initialized. If not, returns the `InvalidUninitBytes` - /// error which will report the first range of bytes which is uninitialized. - fn check_init(&self, range: AllocRange) -> AllocResult { - self.is_init(range).or_else(|idx_range| { - Err(AllocError::InvalidUninitBytes(Some(UninitBytesAccess { - access_offset: range.start, - access_size: range.size, - uninit_offset: idx_range.start, - uninit_size: idx_range.end - idx_range.start, // `Size` subtraction - }))) - }) - } - - pub fn mark_init(&mut self, range: AllocRange, is_init: bool) { - if range.size.bytes() == 0 { - return; - } - assert!(self.mutability == Mutability::Mut); - self.init_mask.set_range(range.start, range.end(), is_init); - } -} - -/// Run-length encoding of the uninit mask. -/// Used to copy parts of a mask multiple times to another allocation. -pub struct InitMaskCompressed { - /// Whether the first range is initialized. - initial: bool, - /// The lengths of ranges that are run-length encoded. - /// The initialization state of the ranges alternate starting with `initial`. - ranges: smallvec::SmallVec<[u64; 1]>, -} - -impl InitMaskCompressed { - pub fn no_bytes_init(&self) -> bool { - // The `ranges` are run-length encoded and of alternating initialization state. - // So if `ranges.len() > 1` then the second block is an initialized range. - !self.initial && self.ranges.len() == 1 - } -} - -/// Transferring the initialization mask to other allocations. -impl<Tag, Extra> Allocation<Tag, Extra> { - /// Creates a run-length encoding of the initialization mask. - /// - /// This is essentially a more space-efficient version of - /// `InitMask::range_as_init_chunks(...).collect::<Vec<_>>()`. - pub fn compress_uninit_range(&self, range: AllocRange) -> InitMaskCompressed { - // Since we are copying `size` bytes from `src` to `dest + i * size` (`for i in 0..repeat`), - // a naive initialization mask copying algorithm would repeatedly have to read the initialization mask from - // the source and write it to the destination. Even if we optimized the memory accesses, - // we'd be doing all of this `repeat` times. - // Therefore we precompute a compressed version of the initialization mask of the source value and - // then write it back `repeat` times without computing any more information from the source. - - // A precomputed cache for ranges of initialized / uninitialized bits - // 0000010010001110 will become - // `[5, 1, 2, 1, 3, 3, 1]`, - // where each element toggles the state. - - let mut ranges = smallvec::SmallVec::<[u64; 1]>::new(); - let initial = self.init_mask.get(range.start); - - for chunk in self.init_mask.range_as_init_chunks(range.start, range.end()) { - let len = chunk.range().end.bytes() - chunk.range().start.bytes(); - ranges.push(len); - } - - InitMaskCompressed { ranges, initial } - } - - /// Applies multiple instances of the run-length encoding to the initialization mask. - pub fn mark_compressed_init_range( - &mut self, - defined: &InitMaskCompressed, - range: AllocRange, - repeat: u64, - ) { - // An optimization where we can just overwrite an entire range of initialization - // bits if they are going to be uniformly `1` or `0`. - if defined.ranges.len() <= 1 { - self.init_mask.set_range_inbounds( - range.start, - range.start + range.size * repeat, // `Size` operations - defined.initial, - ); - return; - } - - for mut j in 0..repeat { - j *= range.size.bytes(); - j += range.start.bytes(); - let mut cur = defined.initial; - for range in &defined.ranges { - let old_j = j; - j += range; - self.init_mask.set_range_inbounds( - Size::from_bytes(old_j), - Size::from_bytes(j), - cur, - ); - cur = !cur; - } - } - } -} - /// "Relocations" stores the provenance information of pointers stored in memory. #[derive(Clone, PartialEq, Eq, PartialOrd, Ord, Hash, Debug, TyEncodable, TyDecodable)] pub struct Relocations<Tag = AllocId>(SortedMap<Size, Tag>); @@ -697,44 +581,25 @@ pub struct InitMask { impl InitMask { pub const BLOCK_SIZE: u64 = 64; - pub fn new(size: Size, state: bool) -> Self { - let mut m = InitMask { blocks: vec![], len: Size::ZERO }; - m.grow(size, state); - m + #[inline] + fn bit_index(bits: Size) -> (usize, usize) { + let bits = bits.bytes(); + let a = bits / InitMask::BLOCK_SIZE; + let b = bits % InitMask::BLOCK_SIZE; + (usize::try_from(a).unwrap(), usize::try_from(b).unwrap()) } - /// Checks whether the range `start..end` (end-exclusive) is entirely initialized. - /// - /// Returns `Ok(())` if it's initialized. Otherwise returns a range of byte - /// indexes for the first contiguous span of the uninitialized access. #[inline] - pub fn is_range_initialized(&self, start: Size, end: Size) -> Result<(), Range<Size>> { - if end > self.len { - return Err(self.len..end); - } - - let uninit_start = find_bit(self, start, end, false); - - match uninit_start { - Some(uninit_start) => { - let uninit_end = find_bit(self, uninit_start, end, true).unwrap_or(end); - Err(uninit_start..uninit_end) - } - None => Ok(()), - } + fn size_from_bit_index(block: impl TryInto<u64>, bit: impl TryInto<u64>) -> Size { + let block = block.try_into().ok().unwrap(); + let bit = bit.try_into().ok().unwrap(); + Size::from_bytes(block * InitMask::BLOCK_SIZE + bit) } - /// Returns an iterator, yielding a range of byte indexes for each contiguous region - /// of initialized or uninitialized bytes inside the range `start..end` (end-exclusive). - /// - /// The iterator guarantees the following: - /// - Chunks are nonempty. - /// - Chunks are adjacent (each range's start is equal to the previous range's end). - /// - Chunks span exactly `start..end` (the first starts at `start`, the last ends at `end`). - /// - Chunks alternate between [`InitChunk::Init`] and [`InitChunk::Uninit`]. - #[inline] - pub fn range_as_init_chunks(&self, start: Size, end: Size) -> InitChunkIter<'_> { - InitChunkIter::new(self, start, end) + pub fn new(size: Size, state: bool) -> Self { + let mut m = InitMask { blocks: vec![], len: Size::ZERO }; + m.grow(size, state); + m } pub fn set_range(&mut self, start: Size, end: Size, new_state: bool) { @@ -746,8 +611,8 @@ impl InitMask { } pub fn set_range_inbounds(&mut self, start: Size, end: Size, new_state: bool) { - let (blocka, bita) = bit_index(start); - let (blockb, bitb) = bit_index(end); + let (blocka, bita) = Self::bit_index(start); + let (blockb, bitb) = Self::bit_index(end); if blocka == blockb { // First set all bits except the first `bita`, // then unset the last `64 - bitb` bits. @@ -791,13 +656,13 @@ impl InitMask { #[inline] pub fn get(&self, i: Size) -> bool { - let (block, bit) = bit_index(i); + let (block, bit) = Self::bit_index(i); (self.blocks[block] & (1 << bit)) != 0 } #[inline] pub fn set(&mut self, i: Size, new_state: bool) { - let (block, bit) = bit_index(i); + let (block, bit) = Self::bit_index(i); self.set_bit(block, bit, new_state); } @@ -827,6 +692,195 @@ impl InitMask { self.len += amount; self.set_range_inbounds(start, start + amount, new_state); // `Size` operation } + + /// Returns the index of the first bit in `start..end` (end-exclusive) that is equal to is_init. + fn find_bit(&self, start: Size, end: Size, is_init: bool) -> Option<Size> { + /// A fast implementation of `find_bit`, + /// which skips over an entire block at a time if it's all 0s (resp. 1s), + /// and finds the first 1 (resp. 0) bit inside a block using `trailing_zeros` instead of a loop. + /// + /// Note that all examples below are written with 8 (instead of 64) bit blocks for simplicity, + /// and with the least significant bit (and lowest block) first: + /// + /// 00000000|00000000 + /// ^ ^ ^ ^ + /// index: 0 7 8 15 + /// + /// Also, if not stated, assume that `is_init = true`, that is, we are searching for the first 1 bit. + fn find_bit_fast( + init_mask: &InitMask, + start: Size, + end: Size, + is_init: bool, + ) -> Option<Size> { + /// Search one block, returning the index of the first bit equal to `is_init`. + fn search_block( + bits: Block, + block: usize, + start_bit: usize, + is_init: bool, + ) -> Option<Size> { + // For the following examples, assume this function was called with: + // bits = 11011100 + // start_bit = 3 + // is_init = false + // Note again that the least significant bit is written first, + // which is backwards compared to how we normally write numbers. + + // Invert bits so we're always looking for the first set bit. + // ! 11011100 + // bits = 00100011 + let bits = if is_init { bits } else { !bits }; + // Mask off unused start bits. + // 00100011 + // & 00011111 + // bits = 00000011 + let bits = bits & (!0 << start_bit); + // Find set bit, if any. + // bit = trailing_zeros(00000011) + // bit = 6 + if bits == 0 { + None + } else { + let bit = bits.trailing_zeros(); + Some(InitMask::size_from_bit_index(block, bit)) + } + } + + if start >= end { + return None; + } + + // Convert `start` and `end` to block indexes and bit indexes within each block. + // We must convert `end` to an inclusive bound to handle block boundaries correctly. + // + // For example: + // + // (a) 00000000|00000000 (b) 00000000| + // ^~~~~~~~~~~^ ^~~~~~~~~^ + // start end start end + // + // In both cases, the block index of `end` is 1. + // But we do want to search block 1 in (a), and we don't in (b). + // + // If we subtract 1 from both end positions to make them inclusive: + // + // (a) 00000000|00000000 (b) 00000000| + // ^~~~~~~~~~^ ^~~~~~~^ + // start end_inclusive start end_inclusive + // + // For (a), the block index of `end_inclusive` is 1, and for (b), it's 0. + // This provides the desired behavior of searching blocks 0 and 1 for (a), + // and searching only block 0 for (b). + let (start_block, start_bit) = InitMask::bit_index(start); + let end_inclusive = Size::from_bytes(end.bytes() - 1); + let (end_block_inclusive, _) = InitMask::bit_index(end_inclusive); + + // Handle first block: need to skip `start_bit` bits. + // + // We need to handle the first block separately, + // because there may be bits earlier in the block that should be ignored, + // such as the bit marked (1) in this example: + // + // (1) + // -|------ + // (c) 01000000|00000000|00000001 + // ^~~~~~~~~~~~~~~~~~^ + // start end + if let Some(i) = + search_block(init_mask.blocks[start_block], start_block, start_bit, is_init) + { + // If the range is less than a block, we may find a matching bit after `end`. + // + // For example, we shouldn't successfully find bit (2), because it's after `end`: + // + // (2) + // -------| + // (d) 00000001|00000000|00000001 + // ^~~~~^ + // start end + // + // An alternative would be to mask off end bits in the same way as we do for start bits, + // but performing this check afterwards is faster and simpler to implement. + if i < end { + return Some(i); + } else { + return None; + } + } + + // Handle remaining blocks. + // + // We can skip over an entire block at once if it's all 0s (resp. 1s). + // The block marked (3) in this example is the first block that will be handled by this loop, + // and it will be skipped for that reason: + // + // (3) + // -------- + // (e) 01000000|00000000|00000001 + // ^~~~~~~~~~~~~~~~~~^ + // start end + if start_block < end_block_inclusive { + // This loop is written in a specific way for performance. + // Notably: `..end_block_inclusive + 1` is used for an inclusive range instead of `..=end_block_inclusive`, + // and `.zip(start_block + 1..)` is used to track the index instead of `.enumerate().skip().take()`, + // because both alternatives result in significantly worse codegen. + // `end_block_inclusive + 1` is guaranteed not to wrap, because `end_block_inclusive <= end / BLOCK_SIZE`, + // and `BLOCK_SIZE` (the number of bits per block) will always be at least 8 (1 byte). + for (&bits, block) in init_mask.blocks[start_block + 1..end_block_inclusive + 1] + .iter() + .zip(start_block + 1..) + { + if let Some(i) = search_block(bits, block, 0, is_init) { + // If this is the last block, we may find a matching bit after `end`. + // + // For example, we shouldn't successfully find bit (4), because it's after `end`: + // + // (4) + // -------| + // (f) 00000001|00000000|00000001 + // ^~~~~~~~~~~~~~~~~~^ + // start end + // + // As above with example (d), we could handle the end block separately and mask off end bits, + // but unconditionally searching an entire block at once and performing this check afterwards + // is faster and much simpler to implement. + if i < end { + return Some(i); + } else { + return None; + } + } + } + } + + None + } + + #[cfg_attr(not(debug_assertions), allow(dead_code))] + fn find_bit_slow( + init_mask: &InitMask, + start: Size, + end: Size, + is_init: bool, + ) -> Option<Size> { + (start..end).find(|&i| init_mask.get(i) == is_init) + } + + let result = find_bit_fast(self, start, end, is_init); + + debug_assert_eq!( + result, + find_bit_slow(self, start, end, is_init), + "optimized implementation of find_bit is wrong for start={:?} end={:?} is_init={} init_mask={:#?}", + start, + end, + is_init, + self + ); + + result + } } /// A contiguous chunk of initialized or uninitialized memory. @@ -845,10 +899,51 @@ impl InitChunk { } } +impl InitMask { + /// Checks whether the range `start..end` (end-exclusive) is entirely initialized. + /// + /// Returns `Ok(())` if it's initialized. Otherwise returns a range of byte + /// indexes for the first contiguous span of the uninitialized access. + #[inline] + pub fn is_range_initialized(&self, start: Size, end: Size) -> Result<(), Range<Size>> { + if end > self.len { + return Err(self.len..end); + } + + let uninit_start = self.find_bit(start, end, false); + + match uninit_start { + Some(uninit_start) => { + let uninit_end = self.find_bit(uninit_start, end, true).unwrap_or(end); + Err(uninit_start..uninit_end) + } + None => Ok(()), + } + } + + /// Returns an iterator, yielding a range of byte indexes for each contiguous region + /// of initialized or uninitialized bytes inside the range `start..end` (end-exclusive). + /// + /// The iterator guarantees the following: + /// - Chunks are nonempty. + /// - Chunks are adjacent (each range's start is equal to the previous range's end). + /// - Chunks span exactly `start..end` (the first starts at `start`, the last ends at `end`). + /// - Chunks alternate between [`InitChunk::Init`] and [`InitChunk::Uninit`]. + #[inline] + pub fn range_as_init_chunks(&self, start: Size, end: Size) -> InitChunkIter<'_> { + assert!(end <= self.len); + + let is_init = if start < end { self.get(start) } else { false }; + + InitChunkIter { init_mask: self, is_init, start, end } + } +} + /// Yields [`InitChunk`]s. See [`InitMask::range_as_init_chunks`]. pub struct InitChunkIter<'a> { init_mask: &'a InitMask, /// Whether the next chunk we will return is initialized. + /// If there are no more chunks, contains some arbitrary value. is_init: bool, /// The current byte index into `init_mask`. start: Size, @@ -856,18 +951,6 @@ pub struct InitChunkIter<'a> { end: Size, } -impl<'a> InitChunkIter<'a> { - #[inline] - fn new(init_mask: &'a InitMask, start: Size, end: Size) -> Self { - assert!(start <= end); - assert!(end <= init_mask.len); - - let is_init = if start < end { init_mask.get(start) } else { false }; - - Self { init_mask, is_init, start, end } - } -} - impl<'a> Iterator for InitChunkIter<'a> { type Item = InitChunk; @@ -878,7 +961,7 @@ impl<'a> Iterator for InitChunkIter<'a> { } let end_of_chunk = - find_bit(&self.init_mask, self.start, self.end, !self.is_init).unwrap_or(self.end); + self.init_mask.find_bit(self.start, self.end, !self.is_init).unwrap_or(self.end); let range = self.start..end_of_chunk; let ret = @@ -891,196 +974,118 @@ impl<'a> Iterator for InitChunkIter<'a> { } } -/// Returns the index of the first bit in `start..end` (end-exclusive) that is equal to is_init. -fn find_bit(init_mask: &InitMask, start: Size, end: Size, is_init: bool) -> Option<Size> { - /// A fast implementation of `find_bit`, - /// which skips over an entire block at a time if it's all 0s (resp. 1s), - /// and finds the first 1 (resp. 0) bit inside a block using `trailing_zeros` instead of a loop. - /// - /// Note that all examples below are written with 8 (instead of 64) bit blocks for simplicity, - /// and with the least significant bit (and lowest block) first: - /// - /// 00000000|00000000 - /// ^ ^ ^ ^ - /// index: 0 7 8 15 +/// Uninitialized bytes. +impl<Tag: Copy, Extra> Allocation<Tag, Extra> { + /// Checks whether the given range is entirely initialized. /// - /// Also, if not stated, assume that `is_init = true`, that is, we are searching for the first 1 bit. - fn find_bit_fast(init_mask: &InitMask, start: Size, end: Size, is_init: bool) -> Option<Size> { - /// Search one block, returning the index of the first bit equal to `is_init`. - fn search_block( - bits: Block, - block: usize, - start_bit: usize, - is_init: bool, - ) -> Option<Size> { - // For the following examples, assume this function was called with: - // bits = 11011100 - // start_bit = 3 - // is_init = false - // Note again that the least significant bit is written first, - // which is backwards compared to how we normally write numbers. - - // Invert bits so we're always looking for the first set bit. - // ! 11011100 - // bits = 00100011 - let bits = if is_init { bits } else { !bits }; - // Mask off unused start bits. - // 00100011 - // & 00011111 - // bits = 00000011 - let bits = bits & (!0 << start_bit); - // Find set bit, if any. - // bit = trailing_zeros(00000011) - // bit = 6 - if bits == 0 { - None - } else { - let bit = bits.trailing_zeros(); - Some(size_from_bit_index(block, bit)) - } - } + /// Returns `Ok(())` if it's initialized. Otherwise returns the range of byte + /// indexes of the first contiguous uninitialized access. + fn is_init(&self, range: AllocRange) -> Result<(), Range<Size>> { + self.init_mask.is_range_initialized(range.start, range.end()) // `Size` addition + } - if start >= end { - return None; - } + /// Checks that a range of bytes is initialized. If not, returns the `InvalidUninitBytes` + /// error which will report the first range of bytes which is uninitialized. + fn check_init(&self, range: AllocRange) -> AllocResult { + self.is_init(range).or_else(|idx_range| { + Err(AllocError::InvalidUninitBytes(Some(UninitBytesAccess { + access_offset: range.start, + access_size: range.size, + uninit_offset: idx_range.start, + uninit_size: idx_range.end - idx_range.start, // `Size` subtraction + }))) + }) + } - // Convert `start` and `end` to block indexes and bit indexes within each block. - // We must convert `end` to an inclusive bound to handle block boundaries correctly. - // - // For example: - // - // (a) 00000000|00000000 (b) 00000000| - // ^~~~~~~~~~~^ ^~~~~~~~~^ - // start end start end - // - // In both cases, the block index of `end` is 1. - // But we do want to search block 1 in (a), and we don't in (b). - // - // If we subtract 1 from both end positions to make them inclusive: - // - // (a) 00000000|00000000 (b) 00000000| - // ^~~~~~~~~~^ ^~~~~~~^ - // start end_inclusive start end_inclusive - // - // For (a), the block index of `end_inclusive` is 1, and for (b), it's 0. - // This provides the desired behavior of searching blocks 0 and 1 for (a), - // and searching only block 0 for (b). - let (start_block, start_bit) = bit_index(start); - let end_inclusive = Size::from_bytes(end.bytes() - 1); - let (end_block_inclusive, _) = bit_index(end_inclusive); - - // Handle first block: need to skip `start_bit` bits. - // - // We need to handle the first block separately, - // because there may be bits earlier in the block that should be ignored, - // such as the bit marked (1) in this example: - // - // (1) - // -|------ - // (c) 01000000|00000000|00000001 - // ^~~~~~~~~~~~~~~~~~^ - // start end - if let Some(i) = - search_block(init_mask.blocks[start_block], start_block, start_bit, is_init) - { - if i < end { - return Some(i); - } else { - // If the range is less than a block, we may find a matching bit after `end`. - // - // For example, we shouldn't successfully find bit (2), because it's after `end`: - // - // (2) - // -------| - // (d) 00000001|00000000|00000001 - // ^~~~~^ - // start end - // - // An alternative would be to mask off end bits in the same way as we do for start bits, - // but performing this check afterwards is faster and simpler to implement. - return None; - } + pub fn mark_init(&mut self, range: AllocRange, is_init: bool) { + if range.size.bytes() == 0 { + return; } + assert!(self.mutability == Mutability::Mut); + self.init_mask.set_range(range.start, range.end(), is_init); + } +} - // Handle remaining blocks. - // - // We can skip over an entire block at once if it's all 0s (resp. 1s). - // The block marked (3) in this example is the first block that will be handled by this loop, - // and it will be skipped for that reason: - // - // (3) - // -------- - // (e) 01000000|00000000|00000001 - // ^~~~~~~~~~~~~~~~~~^ - // start end - if start_block < end_block_inclusive { - // This loop is written in a specific way for performance. - // Notably: `..end_block_inclusive + 1` is used for an inclusive range instead of `..=end_block_inclusive`, - // and `.zip(start_block + 1..)` is used to track the index instead of `.enumerate().skip().take()`, - // because both alternatives result in significantly worse codegen. - // `end_block_inclusive + 1` is guaranteed not to wrap, because `end_block_inclusive <= end / BLOCK_SIZE`, - // and `BLOCK_SIZE` (the number of bits per block) will always be at least 8 (1 byte). - for (&bits, block) in init_mask.blocks[start_block + 1..end_block_inclusive + 1] - .iter() - .zip(start_block + 1..) - { - if let Some(i) = search_block(bits, block, 0, is_init) { - if i < end { - return Some(i); - } else { - // If this is the last block, we may find a matching bit after `end`. - // - // For example, we shouldn't successfully find bit (4), because it's after `end`: - // - // (4) - // -------| - // (f) 00000001|00000000|00000001 - // ^~~~~~~~~~~~~~~~~~^ - // start end - // - // As above with example (d), we could handle the end block separately and mask off end bits, - // but unconditionally searching an entire block at once and performing this check afterwards - // is faster and much simpler to implement. - return None; - } - } - } - } +/// Run-length encoding of the uninit mask. +/// Used to copy parts of a mask multiple times to another allocation. +pub struct InitMaskCompressed { + /// Whether the first range is initialized. + initial: bool, + /// The lengths of ranges that are run-length encoded. + /// The initialization state of the ranges alternate starting with `initial`. + ranges: smallvec::SmallVec<[u64; 1]>, +} - None +impl InitMaskCompressed { + pub fn no_bytes_init(&self) -> bool { + // The `ranges` are run-length encoded and of alternating initialization state. + // So if `ranges.len() > 1` then the second block is an initialized range. + !self.initial && self.ranges.len() == 1 } +} - #[cfg_attr(not(debug_assertions), allow(dead_code))] - fn find_bit_slow(init_mask: &InitMask, start: Size, end: Size, is_init: bool) -> Option<Size> { - (start..end).find(|&i| init_mask.get(i) == is_init) - } +/// Transferring the initialization mask to other allocations. +impl<Tag, Extra> Allocation<Tag, Extra> { + /// Creates a run-length encoding of the initialization mask. + /// + /// This is essentially a more space-efficient version of + /// `InitMask::range_as_init_chunks(...).collect::<Vec<_>>()`. + pub fn compress_uninit_range(&self, range: AllocRange) -> InitMaskCompressed { + // Since we are copying `size` bytes from `src` to `dest + i * size` (`for i in 0..repeat`), + // a naive initialization mask copying algorithm would repeatedly have to read the initialization mask from + // the source and write it to the destination. Even if we optimized the memory accesses, + // we'd be doing all of this `repeat` times. + // Therefore we precompute a compressed version of the initialization mask of the source value and + // then write it back `repeat` times without computing any more information from the source. - let result = find_bit_fast(init_mask, start, end, is_init); + // A precomputed cache for ranges of initialized / uninitialized bits + // 0000010010001110 will become + // `[5, 1, 2, 1, 3, 3, 1]`, + // where each element toggles the state. - debug_assert_eq!( - result, - find_bit_slow(init_mask, start, end, is_init), - "optimized implementation of find_bit is wrong for start={:?} end={:?} is_init={} init_mask={:#?}", - start, - end, - is_init, - init_mask - ); + let mut ranges = smallvec::SmallVec::<[u64; 1]>::new(); + let initial = self.init_mask.get(range.start); - result -} + for chunk in self.init_mask.range_as_init_chunks(range.start, range.end()) { + let len = chunk.range().end.bytes() - chunk.range().start.bytes(); + ranges.push(len); + } -#[inline] -fn bit_index(bits: Size) -> (usize, usize) { - let bits = bits.bytes(); - let a = bits / InitMask::BLOCK_SIZE; - let b = bits % InitMask::BLOCK_SIZE; - (usize::try_from(a).unwrap(), usize::try_from(b).unwrap()) -} + InitMaskCompressed { ranges, initial } + } + + /// Applies multiple instances of the run-length encoding to the initialization mask. + pub fn mark_compressed_init_range( + &mut self, + defined: &InitMaskCompressed, + range: AllocRange, + repeat: u64, + ) { + // An optimization where we can just overwrite an entire range of initialization + // bits if they are going to be uniformly `1` or `0`. + if defined.ranges.len() <= 1 { + self.init_mask.set_range_inbounds( + range.start, + range.start + range.size * repeat, // `Size` operations + defined.initial, + ); + return; + } -#[inline] -fn size_from_bit_index(block: impl TryInto<u64>, bit: impl TryInto<u64>) -> Size { - let block = block.try_into().ok().unwrap(); - let bit = bit.try_into().ok().unwrap(); - Size::from_bytes(block * InitMask::BLOCK_SIZE + bit) + for mut j in 0..repeat { + j *= range.size.bytes(); + j += range.start.bytes(); + let mut cur = defined.initial; + for range in &defined.ranges { + let old_j = j; + j += range; + self.init_mask.set_range_inbounds( + Size::from_bytes(old_j), + Size::from_bytes(j), + cur, + ); + cur = !cur; + } + } + } } |