From 4547b622d8d29df964fa2914213088b148c498fc Mon Sep 17 00:00:00 2001 From: Daniel Baumann Date: Wed, 17 Apr 2024 14:18:32 +0200 Subject: Merging upstream version 1.67.1+dfsg1. Signed-off-by: Daniel Baumann --- .../rustc_middle/src/mir/interpret/allocation.rs | 849 ++------------------- 1 file changed, 76 insertions(+), 773 deletions(-) (limited to 'compiler/rustc_middle/src/mir/interpret/allocation.rs') diff --git a/compiler/rustc_middle/src/mir/interpret/allocation.rs b/compiler/rustc_middle/src/mir/interpret/allocation.rs index 37ec04b07..221105ac4 100644 --- a/compiler/rustc_middle/src/mir/interpret/allocation.rs +++ b/compiler/rustc_middle/src/mir/interpret/allocation.rs @@ -1,16 +1,20 @@ //! The virtual memory representation of the MIR interpreter. +mod init_mask; +mod provenance_map; +#[cfg(test)] +mod tests; + use std::borrow::Cow; -use std::convert::{TryFrom, TryInto}; use std::fmt; use std::hash; -use std::iter; -use std::ops::{Deref, Range}; +use std::ops::Range; use std::ptr; +use either::{Left, Right}; + use rustc_ast::Mutability; use rustc_data_structures::intern::Interned; -use rustc_data_structures::sorted_map::SortedMap; use rustc_span::DUMMY_SP; use rustc_target::abi::{Align, HasDataLayout, Size}; @@ -20,6 +24,10 @@ use super::{ UnsupportedOpInfo, }; use crate::ty; +use init_mask::*; +use provenance_map::*; + +pub use init_mask::{InitChunk, InitChunkIter}; /// This type represents an Allocation in the Miri/CTFE core engine. /// @@ -28,9 +36,9 @@ use crate::ty; /// module provides higher-level access. // Note: for performance reasons when interning, some of the `Allocation` fields can be partially // hashed. (see the `Hash` impl below for more details), so the impl is not derived. -#[derive(Clone, Debug, Eq, PartialEq, PartialOrd, Ord, TyEncodable, TyDecodable)] +#[derive(Clone, Eq, PartialEq, TyEncodable, TyDecodable)] #[derive(HashStable)] -pub struct Allocation { +pub struct Allocation { /// The actual bytes of the allocation. /// Note that the bytes of a pointer represent the offset of the pointer. bytes: Box<[u8]>, @@ -95,27 +103,25 @@ impl hash::Hash for Allocation { /// Interned types generally have an `Outer` type and an `Inner` type, where /// `Outer` is a newtype around `Interned`, and all the operations are /// done on `Outer`, because all occurrences are interned. E.g. `Ty` is an -/// outer type and `TyS` is its inner type. +/// outer type and `TyKind` is its inner type. /// /// Here things are different because only const allocations are interned. This /// means that both the inner type (`Allocation`) and the outer type /// (`ConstAllocation`) are used quite a bit. -#[derive(Copy, Clone, PartialEq, Eq, PartialOrd, Ord, Hash, HashStable)] +#[derive(Copy, Clone, PartialEq, Eq, Hash, HashStable)] #[rustc_pass_by_value] -pub struct ConstAllocation<'tcx, Prov = AllocId, Extra = ()>( - pub Interned<'tcx, Allocation>, -); +pub struct ConstAllocation<'tcx>(pub Interned<'tcx, Allocation>); impl<'tcx> fmt::Debug for ConstAllocation<'tcx> { fn fmt(&self, f: &mut fmt::Formatter<'_>) -> fmt::Result { - // This matches how `Allocation` is printed. We print it like this to - // avoid having to update expected output in a lot of tests. - write!(f, "{:?}", self.inner()) + // The debug representation of this is very verbose and basically useless, + // so don't print it. + write!(f, "ConstAllocation {{ .. }}") } } -impl<'tcx, Prov, Extra> ConstAllocation<'tcx, Prov, Extra> { - pub fn inner(self) -> &'tcx Allocation { +impl<'tcx> ConstAllocation<'tcx> { + pub fn inner(self) -> &'tcx Allocation { self.0.0 } } @@ -183,12 +189,21 @@ pub fn alloc_range(start: Size, size: Size) -> AllocRange { AllocRange { start, size } } -impl AllocRange { +impl From> for AllocRange { #[inline] - pub fn from(r: Range) -> Self { + fn from(r: Range) -> Self { alloc_range(r.start, r.end - r.start) // `Size` subtraction (overflow-checked) } +} +impl From> for AllocRange { + #[inline] + fn from(r: Range) -> Self { + AllocRange::from(Size::from_bytes(r.start)..Size::from_bytes(r.end)) + } +} + +impl AllocRange { #[inline(always)] pub fn end(self) -> Size { self.start + self.size // This does overflow checking. @@ -205,7 +220,7 @@ impl AllocRange { } // The constructors are all without extra; the extra gets added by a machine hook later. -impl Allocation { +impl Allocation { /// Creates an allocation initialized by the given bytes pub fn from_bytes<'a>( slice: impl Into>, @@ -263,7 +278,7 @@ impl Allocation { impl Allocation { /// Adjust allocation from the ones in tcx to a custom Machine instance /// with a different Provenance and Extra type. - pub fn adjust_from_tcx( + pub fn adjust_from_tcx( self, cx: &impl HasDataLayout, extra: Extra, @@ -271,10 +286,10 @@ impl Allocation { ) -> Result, Err> { // Compute new pointer provenance, which also adjusts the bytes. let mut bytes = self.bytes; - let mut new_provenance = Vec::with_capacity(self.provenance.0.len()); + let mut new_provenance = Vec::with_capacity(self.provenance.ptrs().len()); let ptr_size = cx.data_layout().pointer_size.bytes_usize(); let endian = cx.data_layout().endian; - for &(offset, alloc_id) in self.provenance.iter() { + for &(offset, alloc_id) in self.provenance.ptrs().iter() { let idx = offset.bytes_usize(); let ptr_bytes = &mut bytes[idx..idx + ptr_size]; let bits = read_target_uint(endian, ptr_bytes).unwrap(); @@ -286,7 +301,7 @@ impl Allocation { // Create allocation. Ok(Allocation { bytes, - provenance: ProvenanceMap::from_presorted(new_provenance), + provenance: ProvenanceMap::from_presorted_ptrs(new_provenance), init_mask: self.init_mask, align: self.align, mutability: self.mutability, @@ -296,7 +311,7 @@ impl Allocation { } /// Raw accessors. Provide access to otherwise private bytes. -impl Allocation { +impl Allocation { pub fn len(&self) -> usize { self.bytes.len() } @@ -349,9 +364,14 @@ impl Allocation { cx: &impl HasDataLayout, range: AllocRange, ) -> AllocResult<&[u8]> { - self.check_init(range)?; + self.init_mask.is_range_initialized(range).map_err(|uninit_range| { + AllocError::InvalidUninitBytes(Some(UninitBytesAccess { + access: range, + uninit: uninit_range, + })) + })?; if !Prov::OFFSET_IS_ADDR { - if self.range_has_provenance(cx, range) { + if !self.provenance.range_empty(range, cx) { return Err(AllocError::ReadPointerAsBytes); } } @@ -370,7 +390,7 @@ impl Allocation { range: AllocRange, ) -> AllocResult<&mut [u8]> { self.mark_init(range, true); - self.clear_provenance(cx, range)?; + self.provenance.clear(range, cx)?; Ok(&mut self.bytes[range.start.bytes_usize()..range.end().bytes_usize()]) } @@ -382,7 +402,7 @@ impl Allocation { range: AllocRange, ) -> AllocResult<*mut [u8]> { self.mark_init(range, true); - self.clear_provenance(cx, range)?; + self.provenance.clear(range, cx)?; assert!(range.end().bytes_usize() <= self.bytes.len()); // need to do our own bounds-check let begin_ptr = self.bytes.as_mut_ptr().wrapping_add(range.start.bytes_usize()); @@ -393,6 +413,15 @@ impl Allocation { /// Reading and writing. impl Allocation { + /// Sets the init bit for the given range. + 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, is_init); + } + /// Reads a *non-ZST* scalar. /// /// If `read_provenance` is `true`, this will also read provenance; otherwise (if the machine @@ -410,7 +439,7 @@ impl Allocation { read_provenance: bool, ) -> AllocResult> { // First and foremost, if anything is uninit, bail. - if self.is_init(range).is_err() { + if self.init_mask.is_range_initialized(range).is_err() { return Err(AllocError::InvalidUninitBytes(None)); } @@ -423,7 +452,7 @@ impl Allocation { // When reading data with provenance, the easy case is finding provenance exactly where we // are reading, then we can put data and provenance back together and return that. - if let Some(&prov) = self.provenance.get(&range.start) { + if let Some(prov) = self.provenance.get_ptr(range.start) { // Now we can return the bits, with their appropriate provenance. let ptr = Pointer::new(prov, Size::from_bytes(bits)); return Ok(Scalar::from_pointer(ptr, cx)); @@ -431,10 +460,9 @@ impl Allocation { // If we can work on pointers byte-wise, join the byte-wise provenances. if Prov::OFFSET_IS_ADDR { - let mut prov = self.offset_get_provenance(cx, range.start); - for offset in 1..range.size.bytes() { - let this_prov = - self.offset_get_provenance(cx, range.start + Size::from_bytes(offset)); + let mut prov = self.provenance.get(range.start, cx); + for offset in Size::from_bytes(1)..range.size { + let this_prov = self.provenance.get(range.start + offset, cx); prov = Prov::join(prov, this_prov); } // Now use this provenance. @@ -452,7 +480,7 @@ impl Allocation { // Fallback path for when we cannot treat provenance bytewise or ignore it. assert!(!Prov::OFFSET_IS_ADDR); - if self.range_has_provenance(cx, range) { + if !self.provenance.range_empty(range, cx) { return Err(AllocError::ReadPointerAsBytes); } // There is no provenance, we can just return the bits. @@ -466,7 +494,6 @@ impl Allocation { /// /// It is the caller's responsibility to check bounds and alignment beforehand. /// Most likely, you want to call `InterpCx::write_scalar` instead of this method. - #[instrument(skip(self, cx), level = "debug")] pub fn write_scalar( &mut self, cx: &impl HasDataLayout, @@ -478,11 +505,11 @@ impl Allocation { // `to_bits_or_ptr_internal` is the right method because we just want to store this data // as-is into memory. let (bytes, provenance) = match val.to_bits_or_ptr_internal(range.size)? { - Err(val) => { - let (provenance, offset) = val.into_parts(); + Right(ptr) => { + let (provenance, offset) = ptr.into_parts(); (u128::from(offset.bytes()), Some(provenance)) } - Ok(data) => (data, None), + Left(data) => (data, None), }; let endian = cx.data_layout().endian; @@ -491,7 +518,8 @@ impl Allocation { // See if we have to also store some provenance. if let Some(provenance) = provenance { - self.provenance.0.insert(range.start, provenance); + assert_eq!(range.size, cx.data_layout().pointer_size); + self.provenance.insert_ptr(range.start, provenance, cx); } Ok(()) @@ -500,750 +528,25 @@ impl Allocation { /// Write "uninit" to the given memory range. pub fn write_uninit(&mut self, cx: &impl HasDataLayout, range: AllocRange) -> AllocResult { self.mark_init(range, false); - self.clear_provenance(cx, range)?; + self.provenance.clear(range, cx)?; return Ok(()); } -} - -/// Provenance. -impl Allocation { - /// Returns all provenance overlapping with the given pointer-offset pair. - fn range_get_provenance(&self, cx: &impl HasDataLayout, range: AllocRange) -> &[(Size, Prov)] { - // We have to go back `pointer_size - 1` bytes, as that one would still overlap with - // the beginning of this range. - let start = range.start.bytes().saturating_sub(cx.data_layout().pointer_size.bytes() - 1); - self.provenance.range(Size::from_bytes(start)..range.end()) - } - - /// Get the provenance of a single byte. - fn offset_get_provenance(&self, cx: &impl HasDataLayout, offset: Size) -> Option { - let prov = self.range_get_provenance(cx, alloc_range(offset, Size::from_bytes(1))); - assert!(prov.len() <= 1); - prov.first().map(|(_offset, prov)| *prov) - } - - /// Returns whether this allocation has progrnance 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_has_provenance(&self, cx: &impl HasDataLayout, range: AllocRange) -> bool { - !self.range_get_provenance(cx, range).is_empty() - } - - /// Removes all provenance inside the given range. - /// If there is provenance overlapping with the edges, it - /// are removed as well *and* the bytes they cover are marked as - /// uninitialized. This is a somewhat odd "spooky action at a distance", - /// but it allows strictly more code to run than if we would just error - /// immediately in that case. - fn clear_provenance(&mut self, cx: &impl HasDataLayout, range: AllocRange) -> AllocResult - where - Prov: Provenance, - { - // Find the start and end of the given range and its outermost provenance. - let (first, last) = { - // Find all provenance overlapping the given range. - let provenance = self.range_get_provenance(cx, range); - if provenance.is_empty() { - return Ok(()); - } - - ( - provenance.first().unwrap().0, - provenance.last().unwrap().0 + cx.data_layout().pointer_size, - ) - }; - let start = range.start; - let end = range.end(); - - // We need to handle clearing the provenance from parts of a pointer. - // FIXME: Miri should preserve partial provenance; see - // https://github.com/rust-lang/miri/issues/2181. - if first < start { - if Prov::ERR_ON_PARTIAL_PTR_OVERWRITE { - return Err(AllocError::PartialPointerOverwrite(first)); - } - warn!( - "Partial pointer overwrite! De-initializing memory at offsets {first:?}..{start:?}." - ); - self.init_mask.set_range(first, start, false); - } - if last > end { - if Prov::ERR_ON_PARTIAL_PTR_OVERWRITE { - return Err(AllocError::PartialPointerOverwrite( - last - cx.data_layout().pointer_size, - )); - } - warn!( - "Partial pointer overwrite! De-initializing memory at offsets {end:?}..{last:?}." - ); - self.init_mask.set_range(end, last, false); - } - - // 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.provenance.0.remove_range(first..last); - - Ok(()) - } -} - -/// Stores the provenance information of pointers stored in memory. -#[derive(Clone, PartialEq, Eq, PartialOrd, Ord, Hash, Debug, TyEncodable, TyDecodable)] -pub struct ProvenanceMap(SortedMap); - -impl ProvenanceMap { - pub fn new() -> Self { - ProvenanceMap(SortedMap::new()) - } - - // The caller must guarantee that the given provenance list is already sorted - // by address and contain no duplicates. - pub fn from_presorted(r: Vec<(Size, Prov)>) -> Self { - ProvenanceMap(SortedMap::from_presorted_elements(r)) - } -} - -impl Deref for ProvenanceMap { - type Target = SortedMap; - - fn deref(&self) -> &Self::Target { - &self.0 - } -} - -/// A partial, owned list of provenance to transfer into another allocation. -/// -/// Offsets are already adjusted to the destination allocation. -pub struct AllocationProvenance { - dest_provenance: Vec<(Size, Prov)>, -} - -impl Allocation { - pub fn prepare_provenance_copy( - &self, - cx: &impl HasDataLayout, - src: AllocRange, - dest: Size, - count: u64, - ) -> AllocationProvenance { - let provenance = self.range_get_provenance(cx, src); - if provenance.is_empty() { - return AllocationProvenance { dest_provenance: Vec::new() }; - } - - let size = src.size; - let mut new_provenance = Vec::with_capacity(provenance.len() * (count as usize)); - - // 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. - for i in 0..count { - new_provenance.extend(provenance.iter().map(|&(offset, reloc)| { - // compute offset for current repetition - let dest_offset = dest + size * i; // `Size` operations - ( - // shift offsets from source allocation to destination allocation - (offset + dest_offset) - src.start, // `Size` operations - reloc, - ) - })); - } - - AllocationProvenance { dest_provenance: new_provenance } - } - /// Applies a provenance copy. - /// The affected range, as defined in the parameters to `prepare_provenance_copy` is expected + /// Applies a previously prepared provenance copy. + /// The affected range, as defined in the parameters to `provenance().prepare_copy` is expected /// to be clear of provenance. /// /// This is dangerous to use as it can violate internal `Allocation` invariants! /// It only exists to support an efficient implementation of `mem_copy_repeatedly`. - pub fn mark_provenance_range(&mut self, provenance: AllocationProvenance) { - self.provenance.0.insert_presorted(provenance.dest_provenance); - } -} - -//////////////////////////////////////////////////////////////////////////////// -// Uninitialized byte tracking -//////////////////////////////////////////////////////////////////////////////// - -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, PartialOrd, Ord, 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 = MAX_BYTES_TO_HASH / std::mem::size_of::(); - const MAX_BLOCKS_LEN: usize = 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; - - #[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) - } - - 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) { - let len = self.len; - if end > len { - self.grow(end - len, new_state); - } - self.set_range_inbounds(start, end, new_state); - } - - pub 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 + pub fn provenance_apply_copy(&mut self, copy: ProvenanceCopy) { + self.provenance.apply_copy(copy) } - #[inline] - pub fn set(&mut self, i: Size, new_state: bool) { - let (block, bit) = Self::bit_index(i); - self.set_bit(block, bit, new_state); - } - - #[inline] - fn set_bit(&mut self, block: usize, bit: usize, new_state: bool) { - if new_state { - self.blocks[block] |= 1 << bit; - } else { - self.blocks[block] &= !(1 << bit); - } - } - - pub 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 { - /// 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<(), AllocRange> { - if end > self.len { - return Err(AllocRange::from(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(AllocRange::from(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 { - // `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 - } -} - -/// Uninitialized bytes. -impl Allocation { - /// 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<(), AllocRange> { - 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).map_err(|uninit_range| { - AllocError::InvalidUninitBytes(Some(UninitBytesAccess { - access: range, - uninit: uninit_range, - })) - }) - } - - 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 Allocation { - /// 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 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 mut chunks = self.init_mask.range_as_init_chunks(range.start, range.end()).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); - } - - InitMaskCompressed { ranges, initial } - } - - /// Applies multiple instances of the run-length encoding to the initialization mask. + /// Applies a previously prepared copy of the init mask. /// /// This is dangerous to use as it can violate internal `Allocation` invariants! /// It only exists to support an efficient implementation of `mem_copy_repeatedly`. - 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; - } - } + pub fn init_mask_apply_copy(&mut self, copy: InitCopy, range: AllocRange, repeat: u64) { + self.init_mask.apply_copy(copy, range, repeat) } } -- cgit v1.2.3