From 5363f350887b1e5b5dd21a86f88c8af9d7fea6da Mon Sep 17 00:00:00 2001 From: Daniel Baumann Date: Wed, 17 Apr 2024 14:18:25 +0200 Subject: Merging upstream version 1.67.1+dfsg1. Signed-off-by: Daniel Baumann --- .../src/mir/interpret/allocation/init_mask.rs | 530 +++++++++++++++++++++ .../src/mir/interpret/allocation/provenance_map.rs | 321 +++++++++++++ .../src/mir/interpret/allocation/tests.rs | 19 + 3 files changed, 870 insertions(+) create mode 100644 compiler/rustc_middle/src/mir/interpret/allocation/init_mask.rs create mode 100644 compiler/rustc_middle/src/mir/interpret/allocation/provenance_map.rs create mode 100644 compiler/rustc_middle/src/mir/interpret/allocation/tests.rs (limited to 'compiler/rustc_middle/src/mir/interpret/allocation') diff --git a/compiler/rustc_middle/src/mir/interpret/allocation/init_mask.rs b/compiler/rustc_middle/src/mir/interpret/allocation/init_mask.rs new file mode 100644 index 000000000..82e9a961a --- /dev/null +++ b/compiler/rustc_middle/src/mir/interpret/allocation/init_mask.rs @@ -0,0 +1,530 @@ +use std::hash; +use std::iter; +use std::ops::Range; + +use rustc_target::abi::Size; + +use super::AllocRange; + +type Block = u64; + +/// A bitmask where each bit refers to the byte with the same index. If the bit is `true`, the byte +/// is initialized. If it is `false` the byte is uninitialized. +// Note: for performance reasons when interning, some of the `InitMask` fields can be partially +// hashed. (see the `Hash` impl below for more details), so the impl is not derived. +#[derive(Clone, Debug, Eq, PartialEq, TyEncodable, TyDecodable)] +#[derive(HashStable)] +pub struct InitMask { + blocks: Vec, + len: Size, +} + +// Const allocations are only hashed for interning. However, they can be large, making the hashing +// expensive especially since it uses `FxHash`: it's better suited to short keys, not potentially +// big buffers like the allocation's init mask. We can partially hash some fields when they're +// large. +impl hash::Hash for InitMask { + fn hash(&self, state: &mut H) { + const MAX_BLOCKS_TO_HASH: usize = super::MAX_BYTES_TO_HASH / std::mem::size_of::(); + const MAX_BLOCKS_LEN: usize = super::MAX_HASHED_BUFFER_LEN / std::mem::size_of::(); + + // Partially hash the `blocks` buffer when it is large. To limit collisions with common + // prefixes and suffixes, we hash the length and some slices of the buffer. + let block_count = self.blocks.len(); + if block_count > MAX_BLOCKS_LEN { + // Hash the buffer's length. + block_count.hash(state); + + // And its head and tail. + self.blocks[..MAX_BLOCKS_TO_HASH].hash(state); + self.blocks[block_count - MAX_BLOCKS_TO_HASH..].hash(state); + } else { + self.blocks.hash(state); + } + + // Hash the other fields as usual. + self.len.hash(state); + } +} + +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) { + // BLOCK_SIZE is the number of bits that can fit in a `Block`. + // Each bit in a `Block` represents the initialization state of one byte of an allocation, + // so we use `.bytes()` here. + 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()) + } + + #[inline] + fn size_from_bit_index(block: impl TryInto, bit: impl TryInto) -> Size { + let block = block.try_into().ok().unwrap(); + let bit = bit.try_into().ok().unwrap(); + Size::from_bytes(block * InitMask::BLOCK_SIZE + bit) + } + + /// Checks whether the `range` 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, range: AllocRange) -> Result<(), AllocRange> { + let end = range.end(); + if end > self.len { + return Err(AllocRange::from(self.len..end)); + } + + let uninit_start = self.find_bit(range.start, end, false); + + match uninit_start { + Some(uninit_start) => { + let uninit_end = self.find_bit(uninit_start, end, true).unwrap_or(end); + Err(AllocRange::from(uninit_start..uninit_end)) + } + None => Ok(()), + } + } + + pub fn set_range(&mut self, range: AllocRange, new_state: bool) { + let end = range.end(); + let len = self.len; + if end > len { + self.grow(end - len, new_state); + } + self.set_range_inbounds(range.start, end, new_state); + } + + fn set_range_inbounds(&mut self, start: Size, end: Size, new_state: bool) { + 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. + let range = if bitb == 0 { + u64::MAX << bita + } else { + (u64::MAX << bita) & (u64::MAX >> (64 - bitb)) + }; + if new_state { + self.blocks[blocka] |= range; + } else { + self.blocks[blocka] &= !range; + } + return; + } + // across block boundaries + if new_state { + // Set `bita..64` to `1`. + self.blocks[blocka] |= u64::MAX << bita; + // Set `0..bitb` to `1`. + if bitb != 0 { + self.blocks[blockb] |= u64::MAX >> (64 - bitb); + } + // Fill in all the other blocks (much faster than one bit at a time). + for block in (blocka + 1)..blockb { + self.blocks[block] = u64::MAX; + } + } else { + // Set `bita..64` to `0`. + self.blocks[blocka] &= !(u64::MAX << bita); + // Set `0..bitb` to `0`. + if bitb != 0 { + self.blocks[blockb] &= !(u64::MAX >> (64 - bitb)); + } + // Fill in all the other blocks (much faster than one bit at a time). + for block in (blocka + 1)..blockb { + self.blocks[block] = 0; + } + } + } + + #[inline] + pub fn get(&self, i: Size) -> bool { + let (block, bit) = Self::bit_index(i); + (self.blocks[block] & (1 << bit)) != 0 + } + + fn grow(&mut self, amount: Size, new_state: bool) { + if amount.bytes() == 0 { + return; + } + let unused_trailing_bits = + u64::try_from(self.blocks.len()).unwrap() * Self::BLOCK_SIZE - self.len.bytes(); + if amount.bytes() > unused_trailing_bits { + let additional_blocks = amount.bytes() / Self::BLOCK_SIZE + 1; + self.blocks.extend( + // FIXME(oli-obk): optimize this by repeating `new_state as Block`. + iter::repeat(0).take(usize::try_from(additional_blocks).unwrap()), + ); + } + let start = self.len; + 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 { + /// 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: + /// ```text + /// 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 { + /// 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 { + // For the following examples, assume this function was called with: + // bits = 0b00111011 + // start_bit = 3 + // is_init = false + // Note that, for the examples in this function, the most significant bit is written first, + // which is backwards compared to the comments in `find_bit`/`find_bit_fast`. + + // Invert bits so we're always looking for the first set bit. + // ! 0b00111011 + // bits = 0b11000100 + let bits = if is_init { bits } else { !bits }; + // Mask off unused start bits. + // 0b11000100 + // & 0b11111000 + // bits = 0b11000000 + let bits = bits & (!0 << start_bit); + // Find set bit, if any. + // bit = trailing_zeros(0b11000000) + // 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). + // + // 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). + // There is no concern of overflows since we checked for `start >= end` above. + 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 { + (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. +pub enum InitChunk { + Init(Range), + Uninit(Range), +} + +impl InitChunk { + #[inline] + pub fn is_init(&self) -> bool { + match self { + Self::Init(_) => true, + Self::Uninit(_) => false, + } + } + + #[inline] + pub fn range(&self) -> Range { + match self { + Self::Init(r) => r.clone(), + Self::Uninit(r) => r.clone(), + } + } +} + +impl InitMask { + /// 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, range: AllocRange) -> InitChunkIter<'_> { + let start = range.start; + let end = range.end(); + assert!(end <= self.len); + + let is_init = if start < end { + self.get(start) + } else { + // `start..end` is empty: there are no chunks, so use some arbitrary value + false + }; + + InitChunkIter { init_mask: self, is_init, start, end } + } +} + +/// Yields [`InitChunk`]s. See [`InitMask::range_as_init_chunks`]. +#[derive(Clone)] +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, + /// The end byte index into `init_mask`. + end: Size, +} + +impl<'a> Iterator for InitChunkIter<'a> { + type Item = InitChunk; + + #[inline] + fn next(&mut self) -> Option { + if self.start >= self.end { + return None; + } + + let end_of_chunk = + 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 = + Some(if self.is_init { InitChunk::Init(range) } else { InitChunk::Uninit(range) }); + + self.is_init = !self.is_init; + self.start = end_of_chunk; + + ret + } +} + +/// Run-length encoding of the uninit mask. +/// Used to copy parts of a mask multiple times to another allocation. +pub struct InitCopy { + /// 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 InitCopy { + 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 InitMask { + /// Creates a run-length encoding of the initialization mask; panics if range is empty. + /// + /// This is essentially a more space-efficient version of + /// `InitMask::range_as_init_chunks(...).collect::>()`. + pub fn prepare_copy(&self, range: AllocRange) -> InitCopy { + // 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 mut chunks = self.range_as_init_chunks(range).peekable(); + + let initial = chunks.peek().expect("range should be nonempty").is_init(); + + // Here we rely on `range_as_init_chunks` to yield alternating init/uninit chunks. + for chunk in chunks { + let len = chunk.range().end.bytes() - chunk.range().start.bytes(); + ranges.push(len); + } + + InitCopy { ranges, initial } + } + + /// Applies multiple instances of the run-length encoding to the initialization mask. + pub fn apply_copy(&mut self, defined: InitCopy, 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.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.set_range_inbounds(Size::from_bytes(old_j), Size::from_bytes(j), cur); + cur = !cur; + } + } + } +} diff --git a/compiler/rustc_middle/src/mir/interpret/allocation/provenance_map.rs b/compiler/rustc_middle/src/mir/interpret/allocation/provenance_map.rs new file mode 100644 index 000000000..ddd3f3943 --- /dev/null +++ b/compiler/rustc_middle/src/mir/interpret/allocation/provenance_map.rs @@ -0,0 +1,321 @@ +//! Store the provenance for each byte in the range, with a more efficient +//! representation for the common case where PTR_SIZE consecutive bytes have the same provenance. + +use std::cmp; + +use rustc_data_structures::sorted_map::SortedMap; +use rustc_target::abi::{HasDataLayout, Size}; + +use super::{alloc_range, AllocError, AllocId, AllocRange, AllocResult, Provenance}; +use rustc_serialize::{Decodable, Decoder, Encodable, Encoder}; + +/// Stores the provenance information of pointers stored in memory. +#[derive(Clone, PartialEq, Eq, Hash, Debug)] +#[derive(HashStable)] +pub struct ProvenanceMap { + /// Provenance in this map applies from the given offset for an entire pointer-size worth of + /// bytes. Two entires in this map are always at least a pointer size apart. + ptrs: SortedMap, + /// Provenance in this map only applies to the given single byte. + /// This map is disjoint from the previous. It will always be empty when + /// `Prov::OFFSET_IS_ADDR` is false. + bytes: Option>>, +} + +impl> Decodable for ProvenanceMap { + fn decode(d: &mut D) -> Self { + assert!(!Prov::OFFSET_IS_ADDR); // only `AllocId` is ever serialized + Self { ptrs: Decodable::decode(d), bytes: None } + } +} + +impl> Encodable for ProvenanceMap { + fn encode(&self, s: &mut S) { + let Self { ptrs, bytes } = self; + assert!(!Prov::OFFSET_IS_ADDR); // only `AllocId` is ever serialized + debug_assert!(bytes.is_none()); + ptrs.encode(s) + } +} + +impl ProvenanceMap { + pub fn new() -> Self { + ProvenanceMap { ptrs: SortedMap::new(), bytes: None } + } + + /// The caller must guarantee that the given provenance list is already sorted + /// by address and contain no duplicates. + pub fn from_presorted_ptrs(r: Vec<(Size, Prov)>) -> Self { + ProvenanceMap { ptrs: SortedMap::from_presorted_elements(r), bytes: None } + } +} + +impl ProvenanceMap { + /// Give access to the ptr-sized provenances (which can also be thought of as relocations, and + /// indeed that is how codegen treats them). + /// + /// Only exposed with `AllocId` provenance, since it panics if there is bytewise provenance. + #[inline] + pub fn ptrs(&self) -> &SortedMap { + debug_assert!(self.bytes.is_none()); // `AllocId::OFFSET_IS_ADDR` is false so this cannot fail + &self.ptrs + } +} + +impl ProvenanceMap { + /// Returns all ptr-sized provenance in the given range. + /// If the range has length 0, returns provenance that crosses the edge between `start-1` and + /// `start`. + fn range_get_ptrs(&self, range: AllocRange, cx: &impl HasDataLayout) -> &[(Size, Prov)] { + // We have to go back `pointer_size - 1` bytes, as that one would still overlap with + // the beginning of this range. + let adjusted_start = Size::from_bytes( + range.start.bytes().saturating_sub(cx.data_layout().pointer_size.bytes() - 1), + ); + self.ptrs.range(adjusted_start..range.end()) + } + + /// Returns all byte-wise provenance in the given range. + fn range_get_bytes(&self, range: AllocRange) -> &[(Size, Prov)] { + if let Some(bytes) = self.bytes.as_ref() { + bytes.range(range.start..range.end()) + } else { + &[] + } + } + + /// Get the provenance of a single byte. + pub fn get(&self, offset: Size, cx: &impl HasDataLayout) -> Option { + let prov = self.range_get_ptrs(alloc_range(offset, Size::from_bytes(1)), cx); + debug_assert!(prov.len() <= 1); + if let Some(entry) = prov.first() { + // If it overlaps with this byte, it is on this byte. + debug_assert!(self.bytes.as_ref().map_or(true, |b| b.get(&offset).is_none())); + Some(entry.1) + } else { + // Look up per-byte provenance. + self.bytes.as_ref().and_then(|b| b.get(&offset).copied()) + } + } + + /// Check if here is ptr-sized provenance at the given index. + /// Does not mean anything for bytewise provenance! But can be useful as an optimization. + pub fn get_ptr(&self, offset: Size) -> Option { + self.ptrs.get(&offset).copied() + } + + /// Returns whether this allocation has provenance overlapping with the given range. + /// + /// Note: this function exists to allow `range_get_provenance` to be private, in order to somewhat + /// limit access to provenance outside of the `Allocation` abstraction. + /// + pub fn range_empty(&self, range: AllocRange, cx: &impl HasDataLayout) -> bool { + self.range_get_ptrs(range, cx).is_empty() && self.range_get_bytes(range).is_empty() + } + + /// Yields all the provenances stored in this map. + pub fn provenances(&self) -> impl Iterator + '_ { + let bytes = self.bytes.iter().flat_map(|b| b.values()); + self.ptrs.values().chain(bytes).copied() + } + + pub fn insert_ptr(&mut self, offset: Size, prov: Prov, cx: &impl HasDataLayout) { + debug_assert!(self.range_empty(alloc_range(offset, cx.data_layout().pointer_size), cx)); + self.ptrs.insert(offset, prov); + } + + /// Removes all provenance inside the given range. + /// If there is provenance overlapping with the edges, might result in an error. + pub fn clear(&mut self, range: AllocRange, cx: &impl HasDataLayout) -> AllocResult { + let start = range.start; + let end = range.end(); + // Clear the bytewise part -- this is easy. + if Prov::OFFSET_IS_ADDR { + if let Some(bytes) = self.bytes.as_mut() { + bytes.remove_range(start..end); + } + } else { + debug_assert!(self.bytes.is_none()); + } + + // For the ptr-sized part, find the first (inclusive) and last (exclusive) byte of + // provenance that overlaps with the given range. + let (first, last) = { + // Find all provenance overlapping the given range. + let provenance = self.range_get_ptrs(range, cx); + if provenance.is_empty() { + // No provenance in this range, we are done. + return Ok(()); + } + + ( + provenance.first().unwrap().0, + provenance.last().unwrap().0 + cx.data_layout().pointer_size, + ) + }; + + // We need to handle clearing the provenance from parts of a pointer. + if first < start { + if !Prov::OFFSET_IS_ADDR { + // We can't split up the provenance into less than a pointer. + return Err(AllocError::PartialPointerOverwrite(first)); + } + // Insert the remaining part in the bytewise provenance. + let prov = self.ptrs[&first]; + let bytes = self.bytes.get_or_insert_with(Box::default); + for offset in first..start { + bytes.insert(offset, prov); + } + } + if last > end { + let begin_of_last = last - cx.data_layout().pointer_size; + if !Prov::OFFSET_IS_ADDR { + // We can't split up the provenance into less than a pointer. + return Err(AllocError::PartialPointerOverwrite(begin_of_last)); + } + // Insert the remaining part in the bytewise provenance. + let prov = self.ptrs[&begin_of_last]; + let bytes = self.bytes.get_or_insert_with(Box::default); + for offset in end..last { + bytes.insert(offset, prov); + } + } + + // Forget all the provenance. + // Since provenance do not overlap, we know that removing until `last` (exclusive) is fine, + // i.e., this will not remove any other provenance just after the ones we care about. + self.ptrs.remove_range(first..last); + + Ok(()) + } +} + +/// A partial, owned list of provenance to transfer into another allocation. +/// +/// Offsets are already adjusted to the destination allocation. +pub struct ProvenanceCopy { + dest_ptrs: Option>, + dest_bytes: Option>, +} + +impl ProvenanceMap { + pub fn prepare_copy( + &self, + src: AllocRange, + dest: Size, + count: u64, + cx: &impl HasDataLayout, + ) -> AllocResult> { + let shift_offset = move |idx, offset| { + // compute offset for current repetition + let dest_offset = dest + src.size * idx; // `Size` operations + // shift offsets from source allocation to destination allocation + (offset - src.start) + dest_offset // `Size` operations + }; + let ptr_size = cx.data_layout().pointer_size; + + // # Pointer-sized provenances + // Get the provenances that are entirely within this range. + // (Different from `range_get_ptrs` which asks if they overlap the range.) + // Only makes sense if we are copying at least one pointer worth of bytes. + let mut dest_ptrs_box = None; + if src.size >= ptr_size { + let adjusted_end = Size::from_bytes(src.end().bytes() - (ptr_size.bytes() - 1)); + let ptrs = self.ptrs.range(src.start..adjusted_end); + // If `count` is large, this is rather wasteful -- we are allocating a big array here, which + // is mostly filled with redundant information since it's just N copies of the same `Prov`s + // at slightly adjusted offsets. The reason we do this is so that in `mark_provenance_range` + // we can use `insert_presorted`. That wouldn't work with an `Iterator` that just produces + // the right sequence of provenance for all N copies. + // Basically, this large array would have to be created anyway in the target allocation. + let mut dest_ptrs = Vec::with_capacity(ptrs.len() * (count as usize)); + for i in 0..count { + dest_ptrs + .extend(ptrs.iter().map(|&(offset, reloc)| (shift_offset(i, offset), reloc))); + } + debug_assert_eq!(dest_ptrs.len(), dest_ptrs.capacity()); + dest_ptrs_box = Some(dest_ptrs.into_boxed_slice()); + }; + + // # Byte-sized provenances + // This includes the existing bytewise provenance in the range, and ptr provenance + // that overlaps with the begin/end of the range. + let mut dest_bytes_box = None; + let begin_overlap = self.range_get_ptrs(alloc_range(src.start, Size::ZERO), cx).first(); + let end_overlap = self.range_get_ptrs(alloc_range(src.end(), Size::ZERO), cx).first(); + if !Prov::OFFSET_IS_ADDR { + // There can't be any bytewise provenance, and we cannot split up the begin/end overlap. + if let Some(entry) = begin_overlap { + return Err(AllocError::PartialPointerCopy(entry.0)); + } + if let Some(entry) = end_overlap { + return Err(AllocError::PartialPointerCopy(entry.0)); + } + debug_assert!(self.bytes.is_none()); + } else { + let mut bytes = Vec::new(); + // First, if there is a part of a pointer at the start, add that. + if let Some(entry) = begin_overlap { + trace!("start overlapping entry: {entry:?}"); + // For really small copies, make sure we don't run off the end of the `src` range. + let entry_end = cmp::min(entry.0 + ptr_size, src.end()); + for offset in src.start..entry_end { + bytes.push((offset, entry.1)); + } + } else { + trace!("no start overlapping entry"); + } + // Then the main part, bytewise provenance from `self.bytes`. + if let Some(all_bytes) = self.bytes.as_ref() { + bytes.extend(all_bytes.range(src.start..src.end())); + } + // And finally possibly parts of a pointer at the end. + if let Some(entry) = end_overlap { + trace!("end overlapping entry: {entry:?}"); + // For really small copies, make sure we don't start before `src` does. + let entry_start = cmp::max(entry.0, src.start); + for offset in entry_start..src.end() { + if bytes.last().map_or(true, |bytes_entry| bytes_entry.0 < offset) { + // The last entry, if it exists, has a lower offset than us. + bytes.push((offset, entry.1)); + } else { + // There already is an entry for this offset in there! This can happen when the + // start and end range checks actually end up hitting the same pointer, so we + // already added this in the "pointer at the start" part above. + assert!(entry.0 <= src.start); + } + } + } else { + trace!("no end overlapping entry"); + } + trace!("byte provenances: {bytes:?}"); + + // And again a buffer for the new list on the target side. + let mut dest_bytes = Vec::with_capacity(bytes.len() * (count as usize)); + for i in 0..count { + dest_bytes + .extend(bytes.iter().map(|&(offset, reloc)| (shift_offset(i, offset), reloc))); + } + debug_assert_eq!(dest_bytes.len(), dest_bytes.capacity()); + dest_bytes_box = Some(dest_bytes.into_boxed_slice()); + } + + Ok(ProvenanceCopy { dest_ptrs: dest_ptrs_box, dest_bytes: dest_bytes_box }) + } + + /// Applies a provenance copy. + /// The affected range, as defined in the parameters to `prepare_copy` is expected + /// to be clear of provenance. + pub fn apply_copy(&mut self, copy: ProvenanceCopy) { + if let Some(dest_ptrs) = copy.dest_ptrs { + self.ptrs.insert_presorted(dest_ptrs.into()); + } + if Prov::OFFSET_IS_ADDR { + if let Some(dest_bytes) = copy.dest_bytes && !dest_bytes.is_empty() { + self.bytes.get_or_insert_with(Box::default).insert_presorted(dest_bytes.into()); + } + } else { + debug_assert!(copy.dest_bytes.is_none()); + } + } +} diff --git a/compiler/rustc_middle/src/mir/interpret/allocation/tests.rs b/compiler/rustc_middle/src/mir/interpret/allocation/tests.rs new file mode 100644 index 000000000..c9c3c50c5 --- /dev/null +++ b/compiler/rustc_middle/src/mir/interpret/allocation/tests.rs @@ -0,0 +1,19 @@ +use super::*; + +#[test] +fn uninit_mask() { + let mut mask = InitMask::new(Size::from_bytes(500), false); + assert!(!mask.get(Size::from_bytes(499))); + mask.set_range(alloc_range(Size::from_bytes(499), Size::from_bytes(1)), true); + assert!(mask.get(Size::from_bytes(499))); + mask.set_range((100..256).into(), true); + for i in 0..100 { + assert!(!mask.get(Size::from_bytes(i)), "{i} should not be set"); + } + for i in 100..256 { + assert!(mask.get(Size::from_bytes(i)), "{i} should be set"); + } + for i in 256..499 { + assert!(!mask.get(Size::from_bytes(i)), "{i} should not be set"); + } +} -- cgit v1.2.3