1 //! The virtual memory representation of the MIR interpreter.
4 use std::convert::{TryFrom, TryInto};
7 use std::ops::{Deref, Range};
10 use rustc_ast::Mutability;
11 use rustc_data_structures::intern::Interned;
12 use rustc_data_structures::sorted_map::SortedMap;
13 use rustc_span::DUMMY_SP;
14 use rustc_target::abi::{Align, HasDataLayout, Size};
17 read_target_uint, write_target_uint, AllocId, InterpError, InterpResult, Pointer, Provenance,
18 ResourceExhaustionInfo, Scalar, ScalarMaybeUninit, UndefinedBehaviorInfo, UninitBytesAccess,
23 /// This type represents an Allocation in the Miri/CTFE core engine.
25 /// Its public API is rather low-level, working directly with allocation offsets and a custom error
26 /// type to account for the lack of an AllocId on this level. The Miri/CTFE core engine `memory`
27 /// module provides higher-level access.
28 #[derive(Clone, Debug, Eq, PartialEq, PartialOrd, Ord, Hash, TyEncodable, TyDecodable)]
30 pub struct Allocation<Tag = AllocId, Extra = ()> {
31 /// The actual bytes of the allocation.
32 /// Note that the bytes of a pointer represent the offset of the pointer.
34 /// Maps from byte addresses to extra data for each pointer.
35 /// Only the first byte of a pointer is inserted into the map; i.e.,
36 /// every entry in this map applies to `pointer_size` consecutive bytes starting
37 /// at the given offset.
38 relocations: Relocations<Tag>,
39 /// Denotes which part of this allocation is initialized.
41 /// The alignment of the allocation to detect unaligned reads.
42 /// (`Align` guarantees that this is a power of two.)
44 /// `true` if the allocation is mutable.
45 /// Also used by codegen to determine if a static should be put into mutable memory,
46 /// which happens for `static mut` and `static` with interior mutability.
47 pub mutability: Mutability,
48 /// Extra state for the machine.
52 /// Interned types generally have an `Outer` type and an `Inner` type, where
53 /// `Outer` is a newtype around `Interned<Inner>`, and all the operations are
54 /// done on `Outer`, because all occurrences are interned. E.g. `Ty` is an
55 /// outer type and `TyS` is its inner type.
57 /// Here things are different because only const allocations are interned. This
58 /// means that both the inner type (`Allocation`) and the outer type
59 /// (`ConstAllocation`) are used quite a bit.
60 #[derive(Copy, Clone, PartialEq, Eq, PartialOrd, Ord, Hash, HashStable)]
61 #[cfg_attr(not(bootstrap), rustc_pass_by_value)]
62 pub struct ConstAllocation<'tcx, Tag = AllocId, Extra = ()>(
63 pub Interned<'tcx, Allocation<Tag, Extra>>,
66 impl<'tcx> fmt::Debug for ConstAllocation<'tcx> {
67 fn fmt(&self, f: &mut fmt::Formatter<'_>) -> fmt::Result {
68 // This matches how `Allocation` is printed. We print it like this to
69 // avoid having to update expected output in a lot of tests.
70 write!(f, "{:?}", self.inner())
74 impl<'tcx, Tag, Extra> ConstAllocation<'tcx, Tag, Extra> {
75 pub fn inner(self) -> &'tcx Allocation<Tag, Extra> {
80 /// We have our own error type that does not know about the `AllocId`; that information
81 /// is added when converting to `InterpError`.
84 /// Encountered a pointer where we needed raw bytes.
86 /// Partially overwriting a pointer.
87 PartialPointerOverwrite(Size),
88 /// Using uninitialized data where it is not allowed.
89 InvalidUninitBytes(Option<UninitBytesAccess>),
91 pub type AllocResult<T = ()> = Result<T, AllocError>;
94 pub fn to_interp_error<'tcx>(self, alloc_id: AllocId) -> InterpError<'tcx> {
97 ReadPointerAsBytes => InterpError::Unsupported(UnsupportedOpInfo::ReadPointerAsBytes),
98 PartialPointerOverwrite(offset) => InterpError::Unsupported(
99 UnsupportedOpInfo::PartialPointerOverwrite(Pointer::new(alloc_id, offset)),
101 InvalidUninitBytes(info) => InterpError::UndefinedBehavior(
102 UndefinedBehaviorInfo::InvalidUninitBytes(info.map(|b| (alloc_id, b))),
108 /// The information that makes up a memory access: offset and size.
109 #[derive(Copy, Clone, Debug)]
110 pub struct AllocRange {
115 /// Free-starting constructor for less syntactic overhead.
117 pub fn alloc_range(start: Size, size: Size) -> AllocRange {
118 AllocRange { start, size }
123 pub fn end(self) -> Size {
124 self.start + self.size // This does overflow checking.
127 /// Returns the `subrange` within this range; panics if it is not a subrange.
129 pub fn subrange(self, subrange: AllocRange) -> AllocRange {
130 let sub_start = self.start + subrange.start;
131 let range = alloc_range(sub_start, subrange.size);
132 assert!(range.end() <= self.end(), "access outside the bounds for given AllocRange");
137 // The constructors are all without extra; the extra gets added by a machine hook later.
138 impl<Tag> Allocation<Tag> {
139 /// Creates an allocation initialized by the given bytes
140 pub fn from_bytes<'a>(
141 slice: impl Into<Cow<'a, [u8]>>,
143 mutability: Mutability,
145 let bytes = Box::<[u8]>::from(slice.into());
146 let size = Size::from_bytes(bytes.len());
149 relocations: Relocations::new(),
150 init_mask: InitMask::new(size, true),
157 pub fn from_bytes_byte_aligned_immutable<'a>(slice: impl Into<Cow<'a, [u8]>>) -> Self {
158 Allocation::from_bytes(slice, Align::ONE, Mutability::Not)
161 /// Try to create an Allocation of `size` bytes, failing if there is not enough memory
162 /// available to the compiler to do so.
163 pub fn uninit(size: Size, align: Align, panic_on_fail: bool) -> InterpResult<'static, Self> {
164 let bytes = Box::<[u8]>::try_new_zeroed_slice(size.bytes_usize()).map_err(|_| {
165 // This results in an error that can happen non-deterministically, since the memory
166 // available to the compiler can change between runs. Normally queries are always
167 // deterministic. However, we can be non-deterministic here because all uses of const
168 // evaluation (including ConstProp!) will make compilation fail (via hard error
169 // or ICE) upon encountering a `MemoryExhausted` error.
171 panic!("Allocation::uninit called with panic_on_fail had allocation failure")
173 ty::tls::with(|tcx| {
174 tcx.sess.delay_span_bug(DUMMY_SP, "exhausted memory during interpretation")
176 InterpError::ResourceExhaustion(ResourceExhaustionInfo::MemoryExhausted)
178 // SAFETY: the box was zero-allocated, which is a valid initial value for Box<[u8]>
179 let bytes = unsafe { bytes.assume_init() };
182 relocations: Relocations::new(),
183 init_mask: InitMask::new(size, false),
185 mutability: Mutability::Mut,
192 /// Convert Tag and add Extra fields
193 pub fn convert_tag_add_extra<Tag, Extra>(
195 cx: &impl HasDataLayout,
197 mut tagger: impl FnMut(Pointer<AllocId>) -> Pointer<Tag>,
198 ) -> Allocation<Tag, Extra> {
199 // Compute new pointer tags, which also adjusts the bytes.
200 let mut bytes = self.bytes;
201 let mut new_relocations = Vec::with_capacity(self.relocations.0.len());
202 let ptr_size = cx.data_layout().pointer_size.bytes_usize();
203 let endian = cx.data_layout().endian;
204 for &(offset, alloc_id) in self.relocations.iter() {
205 let idx = offset.bytes_usize();
206 let ptr_bytes = &mut bytes[idx..idx + ptr_size];
207 let bits = read_target_uint(endian, ptr_bytes).unwrap();
208 let (ptr_tag, ptr_offset) =
209 tagger(Pointer::new(alloc_id, Size::from_bytes(bits))).into_parts();
210 write_target_uint(endian, ptr_bytes, ptr_offset.bytes().into()).unwrap();
211 new_relocations.push((offset, ptr_tag));
213 // Create allocation.
216 relocations: Relocations::from_presorted(new_relocations),
217 init_mask: self.init_mask,
219 mutability: self.mutability,
225 /// Raw accessors. Provide access to otherwise private bytes.
226 impl<Tag, Extra> Allocation<Tag, Extra> {
227 pub fn len(&self) -> usize {
231 pub fn size(&self) -> Size {
232 Size::from_bytes(self.len())
235 /// Looks at a slice which may describe uninitialized bytes or describe a relocation. This differs
236 /// from `get_bytes_with_uninit_and_ptr` in that it does no relocation checks (even on the
238 /// This must not be used for reads affecting the interpreter execution.
239 pub fn inspect_with_uninit_and_ptr_outside_interpreter(&self, range: Range<usize>) -> &[u8] {
243 /// Returns the mask indicating which bytes are initialized.
244 pub fn init_mask(&self) -> &InitMask {
248 /// Returns the relocation list.
249 pub fn relocations(&self) -> &Relocations<Tag> {
255 impl<Tag: Provenance, Extra> Allocation<Tag, Extra> {
256 /// The last argument controls whether we error out when there are uninitialized
257 /// or pointer bytes. You should never call this, call `get_bytes` or
258 /// `get_bytes_with_uninit_and_ptr` instead,
260 /// This function also guarantees that the resulting pointer will remain stable
261 /// even when new allocations are pushed to the `HashMap`. `copy_repeatedly` relies
264 /// It is the caller's responsibility to check bounds and alignment beforehand.
265 fn get_bytes_internal(
267 cx: &impl HasDataLayout,
269 check_init_and_ptr: bool,
270 ) -> AllocResult<&[u8]> {
271 if check_init_and_ptr {
272 self.check_init(range)?;
273 self.check_relocations(cx, range)?;
275 // We still don't want relocations on the *edges*.
276 self.check_relocation_edges(cx, range)?;
279 Ok(&self.bytes[range.start.bytes_usize()..range.end().bytes_usize()])
282 /// Checks that these bytes are initialized and not pointer bytes, and then return them
285 /// It is the caller's responsibility to check bounds and alignment beforehand.
286 /// Most likely, you want to use the `PlaceTy` and `OperandTy`-based methods
287 /// on `InterpCx` instead.
289 pub fn get_bytes(&self, cx: &impl HasDataLayout, range: AllocRange) -> AllocResult<&[u8]> {
290 self.get_bytes_internal(cx, range, true)
293 /// It is the caller's responsibility to handle uninitialized and pointer bytes.
294 /// However, this still checks that there are no relocations on the *edges*.
296 /// It is the caller's responsibility to check bounds and alignment beforehand.
298 pub fn get_bytes_with_uninit_and_ptr(
300 cx: &impl HasDataLayout,
302 ) -> AllocResult<&[u8]> {
303 self.get_bytes_internal(cx, range, false)
306 /// Just calling this already marks everything as defined and removes relocations,
307 /// so be sure to actually put data there!
309 /// It is the caller's responsibility to check bounds and alignment beforehand.
310 /// Most likely, you want to use the `PlaceTy` and `OperandTy`-based methods
311 /// on `InterpCx` instead.
312 pub fn get_bytes_mut(
314 cx: &impl HasDataLayout,
316 ) -> AllocResult<&mut [u8]> {
317 self.mark_init(range, true);
318 self.clear_relocations(cx, range)?;
320 Ok(&mut self.bytes[range.start.bytes_usize()..range.end().bytes_usize()])
323 /// A raw pointer variant of `get_bytes_mut` that avoids invalidating existing aliases into this memory.
324 pub fn get_bytes_mut_ptr(
326 cx: &impl HasDataLayout,
328 ) -> AllocResult<*mut [u8]> {
329 self.mark_init(range, true);
330 self.clear_relocations(cx, range)?;
332 assert!(range.end().bytes_usize() <= self.bytes.len()); // need to do our own bounds-check
333 let begin_ptr = self.bytes.as_mut_ptr().wrapping_add(range.start.bytes_usize());
334 let len = range.end().bytes_usize() - range.start.bytes_usize();
335 Ok(ptr::slice_from_raw_parts_mut(begin_ptr, len))
339 /// Reading and writing.
340 impl<Tag: Provenance, Extra> Allocation<Tag, Extra> {
341 /// Validates that `ptr.offset` and `ptr.offset + size` do not point to the middle of a
342 /// relocation. If `allow_uninit_and_ptr` is `false`, also enforces that the memory in the
343 /// given range contains neither relocations nor uninitialized bytes.
346 cx: &impl HasDataLayout,
348 allow_uninit_and_ptr: bool,
350 // Check bounds and relocations on the edges.
351 self.get_bytes_with_uninit_and_ptr(cx, range)?;
352 // Check uninit and ptr.
353 if !allow_uninit_and_ptr {
354 self.check_init(range)?;
355 self.check_relocations(cx, range)?;
360 /// Reads a *non-ZST* scalar.
362 /// ZSTs can't be read because in order to obtain a `Pointer`, we need to check
363 /// for ZSTness anyway due to integer pointers being valid for ZSTs.
365 /// It is the caller's responsibility to check bounds and alignment beforehand.
366 /// Most likely, you want to call `InterpCx::read_scalar` instead of this method.
369 cx: &impl HasDataLayout,
371 ) -> AllocResult<ScalarMaybeUninit<Tag>> {
372 // `get_bytes_with_uninit_and_ptr` tests relocation edges.
373 // We deliberately error when loading data that partially has provenance, or partially
374 // initialized data (that's the check below), into a scalar. The LLVM semantics of this are
375 // unclear so we are conservative. See <https://github.com/rust-lang/rust/issues/69488> for
376 // further discussion.
377 let bytes = self.get_bytes_with_uninit_and_ptr(cx, range)?;
378 // Uninit check happens *after* we established that the alignment is correct.
379 // We must not return `Ok()` for unaligned pointers!
380 if self.is_init(range).is_err() {
381 // This inflates uninitialized bytes to the entire scalar, even if only a few
382 // bytes are uninitialized.
383 return Ok(ScalarMaybeUninit::Uninit);
385 // Now we do the actual reading.
386 let bits = read_target_uint(cx.data_layout().endian, bytes).unwrap();
387 // See if we got a pointer.
388 if range.size != cx.data_layout().pointer_size {
390 // *Now*, we better make sure that the inside is free of relocations too.
391 self.check_relocations(cx, range)?;
394 if let Some(&prov) = self.relocations.get(&range.start) {
395 let ptr = Pointer::new(prov, Size::from_bytes(bits));
396 return Ok(ScalarMaybeUninit::from_pointer(ptr, cx));
399 // We don't. Just return the bits.
400 Ok(ScalarMaybeUninit::Scalar(Scalar::from_uint(bits, range.size)))
403 /// Writes a *non-ZST* scalar.
405 /// ZSTs can't be read because in order to obtain a `Pointer`, we need to check
406 /// for ZSTness anyway due to integer pointers being valid for ZSTs.
408 /// It is the caller's responsibility to check bounds and alignment beforehand.
409 /// Most likely, you want to call `InterpCx::write_scalar` instead of this method.
412 cx: &impl HasDataLayout,
414 val: ScalarMaybeUninit<Tag>,
416 assert!(self.mutability == Mutability::Mut);
418 let val = match val {
419 ScalarMaybeUninit::Scalar(scalar) => scalar,
420 ScalarMaybeUninit::Uninit => {
421 self.mark_init(range, false);
426 // `to_bits_or_ptr_internal` is the right method because we just want to store this data
427 // as-is into memory.
428 let (bytes, provenance) = match val.to_bits_or_ptr_internal(range.size) {
430 let (provenance, offset) = val.into_parts();
431 (u128::from(offset.bytes()), Some(provenance))
433 Ok(data) => (data, None),
436 let endian = cx.data_layout().endian;
437 let dst = self.get_bytes_mut(cx, range)?;
438 write_target_uint(endian, dst, bytes).unwrap();
440 // See if we have to also write a relocation.
441 if let Some(provenance) = provenance {
442 self.relocations.0.insert(range.start, provenance);
450 impl<Tag: Copy, Extra> Allocation<Tag, Extra> {
451 /// Returns all relocations overlapping with the given pointer-offset pair.
452 pub fn get_relocations(&self, cx: &impl HasDataLayout, range: AllocRange) -> &[(Size, Tag)] {
453 // We have to go back `pointer_size - 1` bytes, as that one would still overlap with
454 // the beginning of this range.
455 let start = range.start.bytes().saturating_sub(cx.data_layout().pointer_size.bytes() - 1);
456 self.relocations.range(Size::from_bytes(start)..range.end())
459 /// Checks that there are no relocations overlapping with the given range.
461 fn check_relocations(&self, cx: &impl HasDataLayout, range: AllocRange) -> AllocResult {
462 if self.get_relocations(cx, range).is_empty() {
465 Err(AllocError::ReadPointerAsBytes)
469 /// Removes all relocations inside the given range.
470 /// If there are relocations overlapping with the edges, they
471 /// are removed as well *and* the bytes they cover are marked as
472 /// uninitialized. This is a somewhat odd "spooky action at a distance",
473 /// but it allows strictly more code to run than if we would just error
474 /// immediately in that case.
475 fn clear_relocations(&mut self, cx: &impl HasDataLayout, range: AllocRange) -> AllocResult
479 // Find the start and end of the given range and its outermost relocations.
480 let (first, last) = {
481 // Find all relocations overlapping the given range.
482 let relocations = self.get_relocations(cx, range);
483 if relocations.is_empty() {
488 relocations.first().unwrap().0,
489 relocations.last().unwrap().0 + cx.data_layout().pointer_size,
492 let start = range.start;
493 let end = range.end();
495 // We need to handle clearing the relocations from parts of a pointer. See
496 // <https://github.com/rust-lang/rust/issues/87184> for details.
498 if Tag::ERR_ON_PARTIAL_PTR_OVERWRITE {
499 return Err(AllocError::PartialPointerOverwrite(first));
501 self.init_mask.set_range(first, start, false);
504 if Tag::ERR_ON_PARTIAL_PTR_OVERWRITE {
505 return Err(AllocError::PartialPointerOverwrite(
506 last - cx.data_layout().pointer_size,
509 self.init_mask.set_range(end, last, false);
512 // Forget all the relocations.
513 self.relocations.0.remove_range(first..last);
518 /// Errors if there are relocations overlapping with the edges of the
519 /// given memory range.
521 fn check_relocation_edges(&self, cx: &impl HasDataLayout, range: AllocRange) -> AllocResult {
522 self.check_relocations(cx, alloc_range(range.start, Size::ZERO))?;
523 self.check_relocations(cx, alloc_range(range.end(), Size::ZERO))?;
528 /// "Relocations" stores the provenance information of pointers stored in memory.
529 #[derive(Clone, PartialEq, Eq, PartialOrd, Ord, Hash, Debug, TyEncodable, TyDecodable)]
530 pub struct Relocations<Tag = AllocId>(SortedMap<Size, Tag>);
532 impl<Tag> Relocations<Tag> {
533 pub fn new() -> Self {
534 Relocations(SortedMap::new())
537 // The caller must guarantee that the given relocations are already sorted
538 // by address and contain no duplicates.
539 pub fn from_presorted(r: Vec<(Size, Tag)>) -> Self {
540 Relocations(SortedMap::from_presorted_elements(r))
544 impl<Tag> Deref for Relocations<Tag> {
545 type Target = SortedMap<Size, Tag>;
547 fn deref(&self) -> &Self::Target {
552 /// A partial, owned list of relocations to transfer into another allocation.
553 pub struct AllocationRelocations<Tag> {
554 relative_relocations: Vec<(Size, Tag)>,
557 impl<Tag: Copy, Extra> Allocation<Tag, Extra> {
558 pub fn prepare_relocation_copy(
560 cx: &impl HasDataLayout,
564 ) -> AllocationRelocations<Tag> {
565 let relocations = self.get_relocations(cx, src);
566 if relocations.is_empty() {
567 return AllocationRelocations { relative_relocations: Vec::new() };
571 let mut new_relocations = Vec::with_capacity(relocations.len() * (count as usize));
574 new_relocations.extend(relocations.iter().map(|&(offset, reloc)| {
575 // compute offset for current repetition
576 let dest_offset = dest + size * i; // `Size` operations
578 // shift offsets from source allocation to destination allocation
579 (offset + dest_offset) - src.start, // `Size` operations
585 AllocationRelocations { relative_relocations: new_relocations }
588 /// Applies a relocation copy.
589 /// The affected range, as defined in the parameters to `prepare_relocation_copy` is expected
590 /// to be clear of relocations.
591 pub fn mark_relocation_range(&mut self, relocations: AllocationRelocations<Tag>) {
592 self.relocations.0.insert_presorted(relocations.relative_relocations);
596 ////////////////////////////////////////////////////////////////////////////////
597 // Uninitialized byte tracking
598 ////////////////////////////////////////////////////////////////////////////////
602 /// A bitmask where each bit refers to the byte with the same index. If the bit is `true`, the byte
603 /// is initialized. If it is `false` the byte is uninitialized.
604 #[derive(Clone, Debug, Eq, PartialEq, PartialOrd, Ord, Hash, TyEncodable, TyDecodable)]
605 #[derive(HashStable)]
606 pub struct InitMask {
612 pub const BLOCK_SIZE: u64 = 64;
615 fn bit_index(bits: Size) -> (usize, usize) {
616 // BLOCK_SIZE is the number of bits that can fit in a `Block`.
617 // Each bit in a `Block` represents the initialization state of one byte of an allocation,
618 // so we use `.bytes()` here.
619 let bits = bits.bytes();
620 let a = bits / InitMask::BLOCK_SIZE;
621 let b = bits % InitMask::BLOCK_SIZE;
622 (usize::try_from(a).unwrap(), usize::try_from(b).unwrap())
626 fn size_from_bit_index(block: impl TryInto<u64>, bit: impl TryInto<u64>) -> Size {
627 let block = block.try_into().ok().unwrap();
628 let bit = bit.try_into().ok().unwrap();
629 Size::from_bytes(block * InitMask::BLOCK_SIZE + bit)
632 pub fn new(size: Size, state: bool) -> Self {
633 let mut m = InitMask { blocks: vec![], len: Size::ZERO };
638 pub fn set_range(&mut self, start: Size, end: Size, new_state: bool) {
641 self.grow(end - len, new_state);
643 self.set_range_inbounds(start, end, new_state);
646 pub fn set_range_inbounds(&mut self, start: Size, end: Size, new_state: bool) {
647 let (blocka, bita) = Self::bit_index(start);
648 let (blockb, bitb) = Self::bit_index(end);
649 if blocka == blockb {
650 // First set all bits except the first `bita`,
651 // then unset the last `64 - bitb` bits.
652 let range = if bitb == 0 {
655 (u64::MAX << bita) & (u64::MAX >> (64 - bitb))
658 self.blocks[blocka] |= range;
660 self.blocks[blocka] &= !range;
664 // across block boundaries
666 // Set `bita..64` to `1`.
667 self.blocks[blocka] |= u64::MAX << bita;
668 // Set `0..bitb` to `1`.
670 self.blocks[blockb] |= u64::MAX >> (64 - bitb);
672 // Fill in all the other blocks (much faster than one bit at a time).
673 for block in (blocka + 1)..blockb {
674 self.blocks[block] = u64::MAX;
677 // Set `bita..64` to `0`.
678 self.blocks[blocka] &= !(u64::MAX << bita);
679 // Set `0..bitb` to `0`.
681 self.blocks[blockb] &= !(u64::MAX >> (64 - bitb));
683 // Fill in all the other blocks (much faster than one bit at a time).
684 for block in (blocka + 1)..blockb {
685 self.blocks[block] = 0;
691 pub fn get(&self, i: Size) -> bool {
692 let (block, bit) = Self::bit_index(i);
693 (self.blocks[block] & (1 << bit)) != 0
697 pub fn set(&mut self, i: Size, new_state: bool) {
698 let (block, bit) = Self::bit_index(i);
699 self.set_bit(block, bit, new_state);
703 fn set_bit(&mut self, block: usize, bit: usize, new_state: bool) {
705 self.blocks[block] |= 1 << bit;
707 self.blocks[block] &= !(1 << bit);
711 pub fn grow(&mut self, amount: Size, new_state: bool) {
712 if amount.bytes() == 0 {
715 let unused_trailing_bits =
716 u64::try_from(self.blocks.len()).unwrap() * Self::BLOCK_SIZE - self.len.bytes();
717 if amount.bytes() > unused_trailing_bits {
718 let additional_blocks = amount.bytes() / Self::BLOCK_SIZE + 1;
720 // FIXME(oli-obk): optimize this by repeating `new_state as Block`.
721 iter::repeat(0).take(usize::try_from(additional_blocks).unwrap()),
724 let start = self.len;
726 self.set_range_inbounds(start, start + amount, new_state); // `Size` operation
729 /// Returns the index of the first bit in `start..end` (end-exclusive) that is equal to is_init.
730 fn find_bit(&self, start: Size, end: Size, is_init: bool) -> Option<Size> {
731 /// A fast implementation of `find_bit`,
732 /// which skips over an entire block at a time if it's all 0s (resp. 1s),
733 /// and finds the first 1 (resp. 0) bit inside a block using `trailing_zeros` instead of a loop.
735 /// Note that all examples below are written with 8 (instead of 64) bit blocks for simplicity,
736 /// and with the least significant bit (and lowest block) first:
738 /// 00000000|00000000
742 /// Also, if not stated, assume that `is_init = true`, that is, we are searching for the first 1 bit.
744 init_mask: &InitMask,
749 /// Search one block, returning the index of the first bit equal to `is_init`.
756 // For the following examples, assume this function was called with:
760 // Note that, for the examples in this function, the most significant bit is written first,
761 // which is backwards compared to the comments in `find_bit`/`find_bit_fast`.
763 // Invert bits so we're always looking for the first set bit.
766 let bits = if is_init { bits } else { !bits };
767 // Mask off unused start bits.
771 let bits = bits & (!0 << start_bit);
772 // Find set bit, if any.
773 // bit = trailing_zeros(0b11000000)
778 let bit = bits.trailing_zeros();
779 Some(InitMask::size_from_bit_index(block, bit))
787 // Convert `start` and `end` to block indexes and bit indexes within each block.
788 // We must convert `end` to an inclusive bound to handle block boundaries correctly.
792 // (a) 00000000|00000000 (b) 00000000|
793 // ^~~~~~~~~~~^ ^~~~~~~~~^
794 // start end start end
796 // In both cases, the block index of `end` is 1.
797 // But we do want to search block 1 in (a), and we don't in (b).
799 // We subtract 1 from both end positions to make them inclusive:
801 // (a) 00000000|00000000 (b) 00000000|
802 // ^~~~~~~~~~^ ^~~~~~~^
803 // start end_inclusive start end_inclusive
805 // For (a), the block index of `end_inclusive` is 1, and for (b), it's 0.
806 // This provides the desired behavior of searching blocks 0 and 1 for (a),
807 // and searching only block 0 for (b).
808 // There is no concern of overflows since we checked for `start >= end` above.
809 let (start_block, start_bit) = InitMask::bit_index(start);
810 let end_inclusive = Size::from_bytes(end.bytes() - 1);
811 let (end_block_inclusive, _) = InitMask::bit_index(end_inclusive);
813 // Handle first block: need to skip `start_bit` bits.
815 // We need to handle the first block separately,
816 // because there may be bits earlier in the block that should be ignored,
817 // such as the bit marked (1) in this example:
821 // (c) 01000000|00000000|00000001
822 // ^~~~~~~~~~~~~~~~~~^
825 search_block(init_mask.blocks[start_block], start_block, start_bit, is_init)
827 // If the range is less than a block, we may find a matching bit after `end`.
829 // For example, we shouldn't successfully find bit (2), because it's after `end`:
833 // (d) 00000001|00000000|00000001
837 // An alternative would be to mask off end bits in the same way as we do for start bits,
838 // but performing this check afterwards is faster and simpler to implement.
846 // Handle remaining blocks.
848 // We can skip over an entire block at once if it's all 0s (resp. 1s).
849 // The block marked (3) in this example is the first block that will be handled by this loop,
850 // and it will be skipped for that reason:
854 // (e) 01000000|00000000|00000001
855 // ^~~~~~~~~~~~~~~~~~^
857 if start_block < end_block_inclusive {
858 // This loop is written in a specific way for performance.
859 // Notably: `..end_block_inclusive + 1` is used for an inclusive range instead of `..=end_block_inclusive`,
860 // and `.zip(start_block + 1..)` is used to track the index instead of `.enumerate().skip().take()`,
861 // because both alternatives result in significantly worse codegen.
862 // `end_block_inclusive + 1` is guaranteed not to wrap, because `end_block_inclusive <= end / BLOCK_SIZE`,
863 // and `BLOCK_SIZE` (the number of bits per block) will always be at least 8 (1 byte).
864 for (&bits, block) in init_mask.blocks[start_block + 1..end_block_inclusive + 1]
866 .zip(start_block + 1..)
868 if let Some(i) = search_block(bits, block, 0, is_init) {
869 // If this is the last block, we may find a matching bit after `end`.
871 // For example, we shouldn't successfully find bit (4), because it's after `end`:
875 // (f) 00000001|00000000|00000001
876 // ^~~~~~~~~~~~~~~~~~^
879 // As above with example (d), we could handle the end block separately and mask off end bits,
880 // but unconditionally searching an entire block at once and performing this check afterwards
881 // is faster and much simpler to implement.
894 #[cfg_attr(not(debug_assertions), allow(dead_code))]
896 init_mask: &InitMask,
901 (start..end).find(|&i| init_mask.get(i) == is_init)
904 let result = find_bit_fast(self, start, end, is_init);
908 find_bit_slow(self, start, end, is_init),
909 "optimized implementation of find_bit is wrong for start={:?} end={:?} is_init={} init_mask={:#?}",
920 /// A contiguous chunk of initialized or uninitialized memory.
928 pub fn is_init(&self) -> bool {
930 Self::Init(_) => true,
931 Self::Uninit(_) => false,
936 pub fn range(&self) -> Range<Size> {
938 Self::Init(r) => r.clone(),
939 Self::Uninit(r) => r.clone(),
945 /// Checks whether the range `start..end` (end-exclusive) is entirely initialized.
947 /// Returns `Ok(())` if it's initialized. Otherwise returns a range of byte
948 /// indexes for the first contiguous span of the uninitialized access.
950 pub fn is_range_initialized(&self, start: Size, end: Size) -> Result<(), Range<Size>> {
952 return Err(self.len..end);
955 let uninit_start = self.find_bit(start, end, false);
958 Some(uninit_start) => {
959 let uninit_end = self.find_bit(uninit_start, end, true).unwrap_or(end);
960 Err(uninit_start..uninit_end)
966 /// Returns an iterator, yielding a range of byte indexes for each contiguous region
967 /// of initialized or uninitialized bytes inside the range `start..end` (end-exclusive).
969 /// The iterator guarantees the following:
970 /// - Chunks are nonempty.
971 /// - Chunks are adjacent (each range's start is equal to the previous range's end).
972 /// - Chunks span exactly `start..end` (the first starts at `start`, the last ends at `end`).
973 /// - Chunks alternate between [`InitChunk::Init`] and [`InitChunk::Uninit`].
975 pub fn range_as_init_chunks(&self, start: Size, end: Size) -> InitChunkIter<'_> {
976 assert!(end <= self.len);
978 let is_init = if start < end {
981 // `start..end` is empty: there are no chunks, so use some arbitrary value
985 InitChunkIter { init_mask: self, is_init, start, end }
989 /// Yields [`InitChunk`]s. See [`InitMask::range_as_init_chunks`].
991 pub struct InitChunkIter<'a> {
992 init_mask: &'a InitMask,
993 /// Whether the next chunk we will return is initialized.
994 /// If there are no more chunks, contains some arbitrary value.
996 /// The current byte index into `init_mask`.
998 /// The end byte index into `init_mask`.
1002 impl<'a> Iterator for InitChunkIter<'a> {
1003 type Item = InitChunk;
1006 fn next(&mut self) -> Option<Self::Item> {
1007 if self.start >= self.end {
1012 self.init_mask.find_bit(self.start, self.end, !self.is_init).unwrap_or(self.end);
1013 let range = self.start..end_of_chunk;
1016 Some(if self.is_init { InitChunk::Init(range) } else { InitChunk::Uninit(range) });
1018 self.is_init = !self.is_init;
1019 self.start = end_of_chunk;
1025 /// Uninitialized bytes.
1026 impl<Tag: Copy, Extra> Allocation<Tag, Extra> {
1027 /// Checks whether the given range is entirely initialized.
1029 /// Returns `Ok(())` if it's initialized. Otherwise returns the range of byte
1030 /// indexes of the first contiguous uninitialized access.
1031 fn is_init(&self, range: AllocRange) -> Result<(), Range<Size>> {
1032 self.init_mask.is_range_initialized(range.start, range.end()) // `Size` addition
1035 /// Checks that a range of bytes is initialized. If not, returns the `InvalidUninitBytes`
1036 /// error which will report the first range of bytes which is uninitialized.
1037 fn check_init(&self, range: AllocRange) -> AllocResult {
1038 self.is_init(range).map_err(|idx_range| {
1039 AllocError::InvalidUninitBytes(Some(UninitBytesAccess {
1040 access_offset: range.start,
1041 access_size: range.size,
1042 uninit_offset: idx_range.start,
1043 uninit_size: idx_range.end - idx_range.start, // `Size` subtraction
1048 pub fn mark_init(&mut self, range: AllocRange, is_init: bool) {
1049 if range.size.bytes() == 0 {
1052 assert!(self.mutability == Mutability::Mut);
1053 self.init_mask.set_range(range.start, range.end(), is_init);
1057 /// Run-length encoding of the uninit mask.
1058 /// Used to copy parts of a mask multiple times to another allocation.
1059 pub struct InitMaskCompressed {
1060 /// Whether the first range is initialized.
1062 /// The lengths of ranges that are run-length encoded.
1063 /// The initialization state of the ranges alternate starting with `initial`.
1064 ranges: smallvec::SmallVec<[u64; 1]>,
1067 impl InitMaskCompressed {
1068 pub fn no_bytes_init(&self) -> bool {
1069 // The `ranges` are run-length encoded and of alternating initialization state.
1070 // So if `ranges.len() > 1` then the second block is an initialized range.
1071 !self.initial && self.ranges.len() == 1
1075 /// Transferring the initialization mask to other allocations.
1076 impl<Tag, Extra> Allocation<Tag, Extra> {
1077 /// Creates a run-length encoding of the initialization mask; panics if range is empty.
1079 /// This is essentially a more space-efficient version of
1080 /// `InitMask::range_as_init_chunks(...).collect::<Vec<_>>()`.
1081 pub fn compress_uninit_range(&self, range: AllocRange) -> InitMaskCompressed {
1082 // Since we are copying `size` bytes from `src` to `dest + i * size` (`for i in 0..repeat`),
1083 // a naive initialization mask copying algorithm would repeatedly have to read the initialization mask from
1084 // the source and write it to the destination. Even if we optimized the memory accesses,
1085 // we'd be doing all of this `repeat` times.
1086 // Therefore we precompute a compressed version of the initialization mask of the source value and
1087 // then write it back `repeat` times without computing any more information from the source.
1089 // A precomputed cache for ranges of initialized / uninitialized bits
1090 // 0000010010001110 will become
1091 // `[5, 1, 2, 1, 3, 3, 1]`,
1092 // where each element toggles the state.
1094 let mut ranges = smallvec::SmallVec::<[u64; 1]>::new();
1096 let mut chunks = self.init_mask.range_as_init_chunks(range.start, range.end()).peekable();
1098 let initial = chunks.peek().expect("range should be nonempty").is_init();
1100 // Here we rely on `range_as_init_chunks` to yield alternating init/uninit chunks.
1101 for chunk in chunks {
1102 let len = chunk.range().end.bytes() - chunk.range().start.bytes();
1106 InitMaskCompressed { ranges, initial }
1109 /// Applies multiple instances of the run-length encoding to the initialization mask.
1110 pub fn mark_compressed_init_range(
1112 defined: &InitMaskCompressed,
1116 // An optimization where we can just overwrite an entire range of initialization
1117 // bits if they are going to be uniformly `1` or `0`.
1118 if defined.ranges.len() <= 1 {
1119 self.init_mask.set_range_inbounds(
1121 range.start + range.size * repeat, // `Size` operations
1127 for mut j in 0..repeat {
1128 j *= range.size.bytes();
1129 j += range.start.bytes();
1130 let mut cur = defined.initial;
1131 for range in &defined.ranges {
1134 self.init_mask.set_range_inbounds(
1135 Size::from_bytes(old_j),
1136 Size::from_bytes(j),