compiler/rustc_middle/src/mir/interpret/allocation.rs

   1 //! The virtual memory representation of the MIR interpreter.
   2
   3 use std::borrow::Cow;
   4 use std::convert::TryFrom;
   5 use std::iter;
   6 use std::ops::{Deref, DerefMut, Range};
   7
   8 use rustc_ast::Mutability;
   9 use rustc_data_structures::sorted_map::SortedMap;
  10 use rustc_target::abi::{Align, HasDataLayout, Size};
  11
  12 use super::{
  13     read_target_uint, write_target_uint, AllocId, InterpResult, Pointer, Scalar, ScalarMaybeUninit,
  14     UninitBytesAccess,
  15 };
  16
  17 #[derive(Clone, Debug, Eq, PartialEq, PartialOrd, Ord, Hash, TyEncodable, TyDecodable)]
  18 #[derive(HashStable)]
  19 pub struct Allocation<Tag = (), Extra = ()> {
  20     /// The actual bytes of the allocation.
  21     /// Note that the bytes of a pointer represent the offset of the pointer.
  22     bytes: Vec<u8>,
  23     /// Maps from byte addresses to extra data for each pointer.
  24     /// Only the first byte of a pointer is inserted into the map; i.e.,
  25     /// every entry in this map applies to `pointer_size` consecutive bytes starting
  26     /// at the given offset.
  27     relocations: Relocations<Tag>,
  28     /// Denotes which part of this allocation is initialized.
  29     init_mask: InitMask,
  30     /// The size of the allocation. Currently, must always equal `bytes.len()`.
  31     pub size: Size,
  32     /// The alignment of the allocation to detect unaligned reads.
  33     /// (`Align` guarantees that this is a power of two.)
  34     pub align: Align,
  35     /// `true` if the allocation is mutable.
  36     /// Also used by codegen to determine if a static should be put into mutable memory,
  37     /// which happens for `static mut` and `static` with interior mutability.
  38     pub mutability: Mutability,
  39     /// Extra state for the machine.
  40     pub extra: Extra,
  41 }
  42
  43 pub trait AllocationExtra<Tag>: std::fmt::Debug + Clone {
  44     // There is no constructor in here because the constructor's type depends
  45     // on `MemoryKind`, and making things sufficiently generic leads to painful
  46     // inference failure.
  47
  48     /// Hook for performing extra checks on a memory read access.
  49     ///
  50     /// Takes read-only access to the allocation so we can keep all the memory read
  51     /// operations take `&self`. Use a `RefCell` in `AllocExtra` if you
  52     /// need to mutate.
  53     #[inline(always)]
  54     fn memory_read(
  55         _alloc: &Allocation<Tag, Self>,
  56         _ptr: Pointer<Tag>,
  57         _size: Size,
  58     ) -> InterpResult<'tcx> {
  59         Ok(())
  60     }
  61
  62     /// Hook for performing extra checks on a memory write access.
  63     #[inline(always)]
  64     fn memory_written(
  65         _alloc: &mut Allocation<Tag, Self>,
  66         _ptr: Pointer<Tag>,
  67         _size: Size,
  68     ) -> InterpResult<'tcx> {
  69         Ok(())
  70     }
  71
  72     /// Hook for performing extra checks on a memory deallocation.
  73     /// `size` will be the size of the allocation.
  74     #[inline(always)]
  75     fn memory_deallocated(
  76         _alloc: &mut Allocation<Tag, Self>,
  77         _ptr: Pointer<Tag>,
  78         _size: Size,
  79     ) -> InterpResult<'tcx> {
  80         Ok(())
  81     }
  82 }
  83
  84 // For `Tag = ()` and no extra state, we have a trivial implementation.
  85 impl AllocationExtra<()> for () {}
  86
  87 // The constructors are all without extra; the extra gets added by a machine hook later.
  88 impl<Tag> Allocation<Tag> {
  89     /// Creates a read-only allocation initialized by the given bytes
  90     pub fn from_bytes<'a>(slice: impl Into<Cow<'a, [u8]>>, align: Align) -> Self {
  91         let bytes = slice.into().into_owned();
  92         let size = Size::from_bytes(bytes.len());
  93         Self {
  94             bytes,
  95             relocations: Relocations::new(),
  96             init_mask: InitMask::new(size, true),
  97             size,
  98             align,
  99             mutability: Mutability::Not,
 100             extra: (),
 101         }
 102     }
 103
 104     pub fn from_byte_aligned_bytes<'a>(slice: impl Into<Cow<'a, [u8]>>) -> Self {
 105         Allocation::from_bytes(slice, Align::from_bytes(1).unwrap())
 106     }
 107
 108     pub fn uninit(size: Size, align: Align) -> Self {
 109         Allocation {
 110             bytes: vec![0; size.bytes_usize()],
 111             relocations: Relocations::new(),
 112             init_mask: InitMask::new(size, false),
 113             size,
 114             align,
 115             mutability: Mutability::Mut,
 116             extra: (),
 117         }
 118     }
 119 }
 120
 121 impl Allocation<(), ()> {
 122     /// Add Tag and Extra fields
 123     pub fn with_tags_and_extra<T, E>(
 124         self,
 125         mut tagger: impl FnMut(AllocId) -> T,
 126         extra: E,
 127     ) -> Allocation<T, E> {
 128         Allocation {
 129             bytes: self.bytes,
 130             size: self.size,
 131             relocations: Relocations::from_presorted(
 132                 self.relocations
 133                     .iter()
 134                     // The allocations in the relocations (pointers stored *inside* this allocation)
 135                     // all get the base pointer tag.
 136                     .map(|&(offset, ((), alloc))| {
 137                         let tag = tagger(alloc);
 138                         (offset, (tag, alloc))
 139                     })
 140                     .collect(),
 141             ),
 142             init_mask: self.init_mask,
 143             align: self.align,
 144             mutability: self.mutability,
 145             extra,
 146         }
 147     }
 148 }
 149
 150 /// Raw accessors. Provide access to otherwise private bytes.
 151 impl<Tag, Extra> Allocation<Tag, Extra> {
 152     pub fn len(&self) -> usize {
 153         self.size.bytes_usize()
 154     }
 155
 156     /// Looks at a slice which may describe uninitialized bytes or describe a relocation. This differs
 157     /// from `get_bytes_with_uninit_and_ptr` in that it does no relocation checks (even on the
 158     /// edges) at all. It further ignores `AllocationExtra` callbacks.
 159     /// This must not be used for reads affecting the interpreter execution.
 160     pub fn inspect_with_uninit_and_ptr_outside_interpreter(&self, range: Range<usize>) -> &[u8] {
 161         &self.bytes[range]
 162     }
 163
 164     /// Returns the mask indicating which bytes are initialized.
 165     pub fn init_mask(&self) -> &InitMask {
 166         &self.init_mask
 167     }
 168
 169     /// Returns the relocation list.
 170     pub fn relocations(&self) -> &Relocations<Tag> {
 171         &self.relocations
 172     }
 173 }
 174
 175 /// Byte accessors.
 176 impl<'tcx, Tag: Copy, Extra: AllocationExtra<Tag>> Allocation<Tag, Extra> {
 177     /// Just a small local helper function to avoid a bit of code repetition.
 178     /// Returns the range of this allocation that was meant.
 179     #[inline]
 180     fn check_bounds(&self, offset: Size, size: Size) -> Range<usize> {
 181         let end = offset + size; // This does overflow checking.
 182         let end = usize::try_from(end.bytes()).expect("access too big for this host architecture");
 183         assert!(
 184             end <= self.len(),
 185             "Out-of-bounds access at offset {}, size {} in allocation of size {}",
 186             offset.bytes(),
 187             size.bytes(),
 188             self.len()
 189         );
 190         offset.bytes_usize()..end
 191     }
 192
 193     /// The last argument controls whether we error out when there are uninitialized
 194     /// or pointer bytes. You should never call this, call `get_bytes` or
 195     /// `get_bytes_with_uninit_and_ptr` instead,
 196     ///
 197     /// This function also guarantees that the resulting pointer will remain stable
 198     /// even when new allocations are pushed to the `HashMap`. `copy_repeatedly` relies
 199     /// on that.
 200     ///
 201     /// It is the caller's responsibility to check bounds and alignment beforehand.
 202     fn get_bytes_internal(
 203         &self,
 204         cx: &impl HasDataLayout,
 205         ptr: Pointer<Tag>,
 206         size: Size,
 207         check_init_and_ptr: bool,
 208     ) -> InterpResult<'tcx, &[u8]> {
 209         let range = self.check_bounds(ptr.offset, size);
 210
 211         if check_init_and_ptr {
 212             self.check_init(ptr, size)?;
 213             self.check_relocations(cx, ptr, size)?;
 214         } else {
 215             // We still don't want relocations on the *edges*.
 216             self.check_relocation_edges(cx, ptr, size)?;
 217         }
 218
 219         AllocationExtra::memory_read(self, ptr, size)?;
 220
 221         Ok(&self.bytes[range])
 222     }
 223
 224     /// Checks that these bytes are initialized and not pointer bytes, and then return them
 225     /// as a slice.
 226     ///
 227     /// It is the caller's responsibility to check bounds and alignment beforehand.
 228     /// Most likely, you want to use the `PlaceTy` and `OperandTy`-based methods
 229     /// on `InterpCx` instead.
 230     #[inline]
 231     pub fn get_bytes(
 232         &self,
 233         cx: &impl HasDataLayout,
 234         ptr: Pointer<Tag>,
 235         size: Size,
 236     ) -> InterpResult<'tcx, &[u8]> {
 237         self.get_bytes_internal(cx, ptr, size, true)
 238     }
 239
 240     /// It is the caller's responsibility to handle uninitialized and pointer bytes.
 241     /// However, this still checks that there are no relocations on the *edges*.
 242     ///
 243     /// It is the caller's responsibility to check bounds and alignment beforehand.
 244     #[inline]
 245     pub fn get_bytes_with_uninit_and_ptr(
 246         &self,
 247         cx: &impl HasDataLayout,
 248         ptr: Pointer<Tag>,
 249         size: Size,
 250     ) -> InterpResult<'tcx, &[u8]> {
 251         self.get_bytes_internal(cx, ptr, size, false)
 252     }
 253
 254     /// Just calling this already marks everything as defined and removes relocations,
 255     /// so be sure to actually put data there!
 256     ///
 257     /// It is the caller's responsibility to check bounds and alignment beforehand.
 258     /// Most likely, you want to use the `PlaceTy` and `OperandTy`-based methods
 259     /// on `InterpCx` instead.
 260     pub fn get_bytes_mut(
 261         &mut self,
 262         cx: &impl HasDataLayout,
 263         ptr: Pointer<Tag>,
 264         size: Size,
 265     ) -> InterpResult<'tcx, &mut [u8]> {
 266         let range = self.check_bounds(ptr.offset, size);
 267
 268         self.mark_init(ptr, size, true);
 269         self.clear_relocations(cx, ptr, size);
 270
 271         AllocationExtra::memory_written(self, ptr, size)?;
 272
 273         Ok(&mut self.bytes[range])
 274     }
 275 }
 276
 277 /// Reading and writing.
 278 impl<'tcx, Tag: Copy, Extra: AllocationExtra<Tag>> Allocation<Tag, Extra> {
 279     /// Reads bytes until a `0` is encountered. Will error if the end of the allocation is reached
 280     /// before a `0` is found.
 281     ///
 282     /// Most likely, you want to call `Memory::read_c_str` instead of this method.
 283     pub fn read_c_str(
 284         &self,
 285         cx: &impl HasDataLayout,
 286         ptr: Pointer<Tag>,
 287     ) -> InterpResult<'tcx, &[u8]> {
 288         let offset = ptr.offset.bytes_usize();
 289         Ok(match self.bytes[offset..].iter().position(|&c| c == 0) {
 290             Some(size) => {
 291                 let size_with_null = Size::from_bytes(size) + Size::from_bytes(1);
 292                 // Go through `get_bytes` for checks and AllocationExtra hooks.
 293                 // We read the null, so we include it in the request, but we want it removed
 294                 // from the result, so we do subslicing.
 295                 &self.get_bytes(cx, ptr, size_with_null)?[..size]
 296             }
 297             // This includes the case where `offset` is out-of-bounds to begin with.
 298             None => throw_ub!(UnterminatedCString(ptr.erase_tag())),
 299         })
 300     }
 301
 302     /// Validates that `ptr.offset` and `ptr.offset + size` do not point to the middle of a
 303     /// relocation. If `allow_uninit_and_ptr` is `false`, also enforces that the memory in the
 304     /// given range contains neither relocations nor uninitialized bytes.
 305     pub fn check_bytes(
 306         &self,
 307         cx: &impl HasDataLayout,
 308         ptr: Pointer<Tag>,
 309         size: Size,
 310         allow_uninit_and_ptr: bool,
 311     ) -> InterpResult<'tcx> {
 312         // Check bounds and relocations on the edges.
 313         self.get_bytes_with_uninit_and_ptr(cx, ptr, size)?;
 314         // Check uninit and ptr.
 315         if !allow_uninit_and_ptr {
 316             self.check_init(ptr, size)?;
 317             self.check_relocations(cx, ptr, size)?;
 318         }
 319         Ok(())
 320     }
 321
 322     /// Writes `src` to the memory starting at `ptr.offset`.
 323     ///
 324     /// It is the caller's responsibility to check bounds and alignment beforehand.
 325     /// Most likely, you want to call `Memory::write_bytes` instead of this method.
 326     pub fn write_bytes(
 327         &mut self,
 328         cx: &impl HasDataLayout,
 329         ptr: Pointer<Tag>,
 330         src: impl IntoIterator<Item = u8>,
 331     ) -> InterpResult<'tcx> {
 332         let mut src = src.into_iter();
 333         let (lower, upper) = src.size_hint();
 334         let len = upper.expect("can only write bounded iterators");
 335         assert_eq!(lower, len, "can only write iterators with a precise length");
 336         let bytes = self.get_bytes_mut(cx, ptr, Size::from_bytes(len))?;
 337         // `zip` would stop when the first iterator ends; we want to definitely
 338         // cover all of `bytes`.
 339         for dest in bytes {
 340             *dest = src.next().expect("iterator was shorter than it said it would be");
 341         }
 342         assert_matches!(src.next(), None, "iterator was longer than it said it would be");
 343         Ok(())
 344     }
 345
 346     /// Reads a *non-ZST* scalar.
 347     ///
 348     /// ZSTs can't be read because in order to obtain a `Pointer`, we need to check
 349     /// for ZSTness anyway due to integer pointers being valid for ZSTs.
 350     ///
 351     /// It is the caller's responsibility to check bounds and alignment beforehand.
 352     /// Most likely, you want to call `InterpCx::read_scalar` instead of this method.
 353     pub fn read_scalar(
 354         &self,
 355         cx: &impl HasDataLayout,
 356         ptr: Pointer<Tag>,
 357         size: Size,
 358     ) -> InterpResult<'tcx, ScalarMaybeUninit<Tag>> {
 359         // `get_bytes_unchecked` tests relocation edges.
 360         let bytes = self.get_bytes_with_uninit_and_ptr(cx, ptr, size)?;
 361         // Uninit check happens *after* we established that the alignment is correct.
 362         // We must not return `Ok()` for unaligned pointers!
 363         if self.is_init(ptr, size).is_err() {
 364             // This inflates uninitialized bytes to the entire scalar, even if only a few
 365             // bytes are uninitialized.
 366             return Ok(ScalarMaybeUninit::Uninit);
 367         }
 368         // Now we do the actual reading.
 369         let bits = read_target_uint(cx.data_layout().endian, bytes).unwrap();
 370         // See if we got a pointer.
 371         if size != cx.data_layout().pointer_size {
 372             // *Now*, we better make sure that the inside is free of relocations too.
 373             self.check_relocations(cx, ptr, size)?;
 374         } else {
 375             if let Some(&(tag, alloc_id)) = self.relocations.get(&ptr.offset) {
 376                 let ptr = Pointer::new_with_tag(alloc_id, Size::from_bytes(bits), tag);
 377                 return Ok(ScalarMaybeUninit::Scalar(ptr.into()));
 378             }
 379         }
 380         // We don't. Just return the bits.
 381         Ok(ScalarMaybeUninit::Scalar(Scalar::from_uint(bits, size)))
 382     }
 383
 384     /// Reads a pointer-sized scalar.
 385     ///
 386     /// It is the caller's responsibility to check bounds and alignment beforehand.
 387     /// Most likely, you want to call `InterpCx::read_scalar` instead of this method.
 388     pub fn read_ptr_sized(
 389         &self,
 390         cx: &impl HasDataLayout,
 391         ptr: Pointer<Tag>,
 392     ) -> InterpResult<'tcx, ScalarMaybeUninit<Tag>> {
 393         self.read_scalar(cx, ptr, cx.data_layout().pointer_size)
 394     }
 395
 396     /// Writes a *non-ZST* scalar.
 397     ///
 398     /// ZSTs can't be read because in order to obtain a `Pointer`, we need to check
 399     /// for ZSTness anyway due to integer pointers being valid for ZSTs.
 400     ///
 401     /// It is the caller's responsibility to check bounds and alignment beforehand.
 402     /// Most likely, you want to call `InterpCx::write_scalar` instead of this method.
 403     pub fn write_scalar(
 404         &mut self,
 405         cx: &impl HasDataLayout,
 406         ptr: Pointer<Tag>,
 407         val: ScalarMaybeUninit<Tag>,
 408         type_size: Size,
 409     ) -> InterpResult<'tcx> {
 410         let val = match val {
 411             ScalarMaybeUninit::Scalar(scalar) => scalar,
 412             ScalarMaybeUninit::Uninit => {
 413                 self.mark_init(ptr, type_size, false);
 414                 return Ok(());
 415             }
 416         };
 417
 418         let bytes = match val.to_bits_or_ptr(type_size, cx) {
 419             Err(val) => u128::from(val.offset.bytes()),
 420             Ok(data) => data,
 421         };
 422
 423         let endian = cx.data_layout().endian;
 424         let dst = self.get_bytes_mut(cx, ptr, type_size)?;
 425         write_target_uint(endian, dst, bytes).unwrap();
 426
 427         // See if we have to also write a relocation.
 428         if let Scalar::Ptr(val) = val {
 429             self.relocations.insert(ptr.offset, (val.tag, val.alloc_id));
 430         }
 431
 432         Ok(())
 433     }
 434
 435     /// Writes a pointer-sized scalar.
 436     ///
 437     /// It is the caller's responsibility to check bounds and alignment beforehand.
 438     /// Most likely, you want to call `InterpCx::write_scalar` instead of this method.
 439     pub fn write_ptr_sized(
 440         &mut self,
 441         cx: &impl HasDataLayout,
 442         ptr: Pointer<Tag>,
 443         val: ScalarMaybeUninit<Tag>,
 444     ) -> InterpResult<'tcx> {
 445         let ptr_size = cx.data_layout().pointer_size;
 446         self.write_scalar(cx, ptr, val, ptr_size)
 447     }
 448 }
 449
 450 /// Relocations.
 451 impl<'tcx, Tag: Copy, Extra> Allocation<Tag, Extra> {
 452     /// Returns all relocations overlapping with the given pointer-offset pair.
 453     pub fn get_relocations(
 454         &self,
 455         cx: &impl HasDataLayout,
 456         ptr: Pointer<Tag>,
 457         size: Size,
 458     ) -> &[(Size, (Tag, AllocId))] {
 459         // We have to go back `pointer_size - 1` bytes, as that one would still overlap with
 460         // the beginning of this range.
 461         let start = ptr.offset.bytes().saturating_sub(cx.data_layout().pointer_size.bytes() - 1);
 462         let end = ptr.offset + size; // This does overflow checking.
 463         self.relocations.range(Size::from_bytes(start)..end)
 464     }
 465
 466     /// Checks that there are no relocations overlapping with the given range.
 467     #[inline(always)]
 468     fn check_relocations(
 469         &self,
 470         cx: &impl HasDataLayout,
 471         ptr: Pointer<Tag>,
 472         size: Size,
 473     ) -> InterpResult<'tcx> {
 474         if self.get_relocations(cx, ptr, size).is_empty() {
 475             Ok(())
 476         } else {
 477             throw_unsup!(ReadPointerAsBytes)
 478         }
 479     }
 480
 481     /// Removes all relocations inside the given range.
 482     /// If there are relocations overlapping with the edges, they
 483     /// are removed as well *and* the bytes they cover are marked as
 484     /// uninitialized. This is a somewhat odd "spooky action at a distance",
 485     /// but it allows strictly more code to run than if we would just error
 486     /// immediately in that case.
 487     fn clear_relocations(&mut self, cx: &impl HasDataLayout, ptr: Pointer<Tag>, size: Size) {
 488         // Find the start and end of the given range and its outermost relocations.
 489         let (first, last) = {
 490             // Find all relocations overlapping the given range.
 491             let relocations = self.get_relocations(cx, ptr, size);
 492             if relocations.is_empty() {
 493                 return;
 494             }
 495
 496             (
 497                 relocations.first().unwrap().0,
 498                 relocations.last().unwrap().0 + cx.data_layout().pointer_size,
 499             )
 500         };
 501         let start = ptr.offset;
 502         let end = start + size; // `Size` addition
 503
 504         // Mark parts of the outermost relocations as uninitialized if they partially fall outside the
 505         // given range.
 506         if first < start {
 507             self.init_mask.set_range(first, start, false);
 508         }
 509         if last > end {
 510             self.init_mask.set_range(end, last, false);
 511         }
 512
 513         // Forget all the relocations.
 514         self.relocations.remove_range(first..last);
 515     }
 516
 517     /// Errors if there are relocations overlapping with the edges of the
 518     /// given memory range.
 519     #[inline]
 520     fn check_relocation_edges(
 521         &self,
 522         cx: &impl HasDataLayout,
 523         ptr: Pointer<Tag>,
 524         size: Size,
 525     ) -> InterpResult<'tcx> {
 526         self.check_relocations(cx, ptr, Size::ZERO)?;
 527         self.check_relocations(cx, ptr.offset(size, cx)?, Size::ZERO)?;
 528         Ok(())
 529     }
 530 }
 531
 532 /// Uninitialized bytes.
 533 impl<'tcx, Tag: Copy, Extra> Allocation<Tag, Extra> {
 534     /// Checks whether the given range  is entirely initialized.
 535     ///
 536     /// Returns `Ok(())` if it's initialized. Otherwise returns the range of byte
 537     /// indexes of the first contiguous uninitialized access.
 538     fn is_init(&self, ptr: Pointer<Tag>, size: Size) -> Result<(), Range<Size>> {
 539         self.init_mask.is_range_initialized(ptr.offset, ptr.offset + size) // `Size` addition
 540     }
 541
 542     /// Checks that a range of bytes is initialized. If not, returns the `InvalidUninitBytes`
 543     /// error which will report the first range of bytes which is uninitialized.
 544     fn check_init(&self, ptr: Pointer<Tag>, size: Size) -> InterpResult<'tcx> {
 545         self.is_init(ptr, size).or_else(|idx_range| {
 546             throw_ub!(InvalidUninitBytes(Some(UninitBytesAccess {
 547                 access_ptr: ptr.erase_tag(),
 548                 access_size: size,
 549                 uninit_ptr: Pointer::new(ptr.alloc_id, idx_range.start),
 550                 uninit_size: idx_range.end - idx_range.start, // `Size` subtraction
 551             })))
 552         })
 553     }
 554
 555     pub fn mark_init(&mut self, ptr: Pointer<Tag>, size: Size, is_init: bool) {
 556         if size.bytes() == 0 {
 557             return;
 558         }
 559         self.init_mask.set_range(ptr.offset, ptr.offset + size, is_init);
 560     }
 561 }
 562
 563 /// Run-length encoding of the uninit mask.
 564 /// Used to copy parts of a mask multiple times to another allocation.
 565 pub struct InitMaskCompressed {
 566     /// Whether the first range is initialized.
 567     initial: bool,
 568     /// The lengths of ranges that are run-length encoded.
 569     /// The initialization state of the ranges alternate starting with `initial`.
 570     ranges: smallvec::SmallVec<[u64; 1]>,
 571 }
 572
 573 impl InitMaskCompressed {
 574     pub fn no_bytes_init(&self) -> bool {
 575         // The `ranges` are run-length encoded and of alternating initialization state.
 576         // So if `ranges.len() > 1` then the second block is an initialized range.
 577         !self.initial && self.ranges.len() == 1
 578     }
 579 }
 580
 581 /// Transferring the initialization mask to other allocations.
 582 impl<Tag, Extra> Allocation<Tag, Extra> {
 583     /// Creates a run-length encoding of the initialization mask.
 584     pub fn compress_uninit_range(&self, src: Pointer<Tag>, size: Size) -> InitMaskCompressed {
 585         // Since we are copying `size` bytes from `src` to `dest + i * size` (`for i in 0..repeat`),
 586         // a naive initialization mask copying algorithm would repeatedly have to read the initialization mask from
 587         // the source and write it to the destination. Even if we optimized the memory accesses,
 588         // we'd be doing all of this `repeat` times.
 589         // Therefore we precompute a compressed version of the initialization mask of the source value and
 590         // then write it back `repeat` times without computing any more information from the source.
 591
 592         // A precomputed cache for ranges of initialized / uninitialized bits
 593         // 0000010010001110 will become
 594         // `[5, 1, 2, 1, 3, 3, 1]`,
 595         // where each element toggles the state.
 596
 597         let mut ranges = smallvec::SmallVec::<[u64; 1]>::new();
 598         let initial = self.init_mask.get(src.offset);
 599         let mut cur_len = 1;
 600         let mut cur = initial;
 601
 602         for i in 1..size.bytes() {
 603             // FIXME: optimize to bitshift the current uninitialized block's bits and read the top bit.
 604             if self.init_mask.get(src.offset + Size::from_bytes(i)) == cur {
 605                 cur_len += 1;
 606             } else {
 607                 ranges.push(cur_len);
 608                 cur_len = 1;
 609                 cur = !cur;
 610             }
 611         }
 612
 613         ranges.push(cur_len);
 614
 615         InitMaskCompressed { ranges, initial }
 616     }
 617
 618     /// Applies multiple instances of the run-length encoding to the initialization mask.
 619     pub fn mark_compressed_init_range(
 620         &mut self,
 621         defined: &InitMaskCompressed,
 622         dest: Pointer<Tag>,
 623         size: Size,
 624         repeat: u64,
 625     ) {
 626         // An optimization where we can just overwrite an entire range of initialization
 627         // bits if they are going to be uniformly `1` or `0`.
 628         if defined.ranges.len() <= 1 {
 629             self.init_mask.set_range_inbounds(
 630                 dest.offset,
 631                 dest.offset + size * repeat, // `Size` operations
 632                 defined.initial,
 633             );
 634             return;
 635         }
 636
 637         for mut j in 0..repeat {
 638             j *= size.bytes();
 639             j += dest.offset.bytes();
 640             let mut cur = defined.initial;
 641             for range in &defined.ranges {
 642                 let old_j = j;
 643                 j += range;
 644                 self.init_mask.set_range_inbounds(
 645                     Size::from_bytes(old_j),
 646                     Size::from_bytes(j),
 647                     cur,
 648                 );
 649                 cur = !cur;
 650             }
 651         }
 652     }
 653 }
 654
 655 /// Relocations.
 656 #[derive(Clone, PartialEq, Eq, PartialOrd, Ord, Hash, Debug, TyEncodable, TyDecodable)]
 657 pub struct Relocations<Tag = (), Id = AllocId>(SortedMap<Size, (Tag, Id)>);
 658
 659 impl<Tag, Id> Relocations<Tag, Id> {
 660     pub fn new() -> Self {
 661         Relocations(SortedMap::new())
 662     }
 663
 664     // The caller must guarantee that the given relocations are already sorted
 665     // by address and contain no duplicates.
 666     pub fn from_presorted(r: Vec<(Size, (Tag, Id))>) -> Self {
 667         Relocations(SortedMap::from_presorted_elements(r))
 668     }
 669 }
 670
 671 impl<Tag> Deref for Relocations<Tag> {
 672     type Target = SortedMap<Size, (Tag, AllocId)>;
 673
 674     fn deref(&self) -> &Self::Target {
 675         &self.0
 676     }
 677 }
 678
 679 impl<Tag> DerefMut for Relocations<Tag> {
 680     fn deref_mut(&mut self) -> &mut Self::Target {
 681         &mut self.0
 682     }
 683 }
 684
 685 /// A partial, owned list of relocations to transfer into another allocation.
 686 pub struct AllocationRelocations<Tag> {
 687     relative_relocations: Vec<(Size, (Tag, AllocId))>,
 688 }
 689
 690 impl<Tag: Copy, Extra> Allocation<Tag, Extra> {
 691     pub fn prepare_relocation_copy(
 692         &self,
 693         cx: &impl HasDataLayout,
 694         src: Pointer<Tag>,
 695         size: Size,
 696         dest: Pointer<Tag>,
 697         length: u64,
 698     ) -> AllocationRelocations<Tag> {
 699         let relocations = self.get_relocations(cx, src, size);
 700         if relocations.is_empty() {
 701             return AllocationRelocations { relative_relocations: Vec::new() };
 702         }
 703
 704         let mut new_relocations = Vec::with_capacity(relocations.len() * (length as usize));
 705
 706         for i in 0..length {
 707             new_relocations.extend(relocations.iter().map(|&(offset, reloc)| {
 708                 // compute offset for current repetition
 709                 let dest_offset = dest.offset + size * i; // `Size` operations
 710                 (
 711                     // shift offsets from source allocation to destination allocation
 712                     (offset + dest_offset) - src.offset, // `Size` operations
 713                     reloc,
 714                 )
 715             }));
 716         }
 717
 718         AllocationRelocations { relative_relocations: new_relocations }
 719     }
 720
 721     /// Applies a relocation copy.
 722     /// The affected range, as defined in the parameters to `prepare_relocation_copy` is expected
 723     /// to be clear of relocations.
 724     pub fn mark_relocation_range(&mut self, relocations: AllocationRelocations<Tag>) {
 725         self.relocations.insert_presorted(relocations.relative_relocations);
 726     }
 727 }
 728
 729 ////////////////////////////////////////////////////////////////////////////////
 730 // Uninitialized byte tracking
 731 ////////////////////////////////////////////////////////////////////////////////
 732
 733 type Block = u64;
 734
 735 /// A bitmask where each bit refers to the byte with the same index. If the bit is `true`, the byte
 736 /// is initialized. If it is `false` the byte is uninitialized.
 737 #[derive(Clone, Debug, Eq, PartialEq, PartialOrd, Ord, Hash, TyEncodable, TyDecodable)]
 738 #[derive(HashStable)]
 739 pub struct InitMask {
 740     blocks: Vec<Block>,
 741     len: Size,
 742 }
 743
 744 impl InitMask {
 745     pub const BLOCK_SIZE: u64 = 64;
 746
 747     pub fn new(size: Size, state: bool) -> Self {
 748         let mut m = InitMask { blocks: vec![], len: Size::ZERO };
 749         m.grow(size, state);
 750         m
 751     }
 752
 753     /// Checks whether the range `start..end` (end-exclusive) is entirely initialized.
 754     ///
 755     /// Returns `Ok(())` if it's initialized. Otherwise returns a range of byte
 756     /// indexes for the first contiguous span of the uninitialized access.
 757     #[inline]
 758     pub fn is_range_initialized(&self, start: Size, end: Size) -> Result<(), Range<Size>> {
 759         if end > self.len {
 760             return Err(self.len..end);
 761         }
 762
 763         // FIXME(oli-obk): optimize this for allocations larger than a block.
 764         let idx = (start.bytes()..end.bytes()).map(Size::from_bytes).find(|&i| !self.get(i));
 765
 766         match idx {
 767             Some(idx) => {
 768                 let uninit_end = (idx.bytes()..end.bytes())
 769                     .map(Size::from_bytes)
 770                     .find(|&i| self.get(i))
 771                     .unwrap_or(end);
 772                 Err(idx..uninit_end)
 773             }
 774             None => Ok(()),
 775         }
 776     }
 777
 778     pub fn set_range(&mut self, start: Size, end: Size, new_state: bool) {
 779         let len = self.len;
 780         if end > len {
 781             self.grow(end - len, new_state);
 782         }
 783         self.set_range_inbounds(start, end, new_state);
 784     }
 785
 786     pub fn set_range_inbounds(&mut self, start: Size, end: Size, new_state: bool) {
 787         let (blocka, bita) = bit_index(start);
 788         let (blockb, bitb) = bit_index(end);
 789         if blocka == blockb {
 790             // First set all bits except the first `bita`,
 791             // then unset the last `64 - bitb` bits.
 792             let range = if bitb == 0 {
 793                 u64::MAX << bita
 794             } else {
 795                 (u64::MAX << bita) & (u64::MAX >> (64 - bitb))
 796             };
 797             if new_state {
 798                 self.blocks[blocka] |= range;
 799             } else {
 800                 self.blocks[blocka] &= !range;
 801             }
 802             return;
 803         }
 804         // across block boundaries
 805         if new_state {
 806             // Set `bita..64` to `1`.
 807             self.blocks[blocka] |= u64::MAX << bita;
 808             // Set `0..bitb` to `1`.
 809             if bitb != 0 {
 810                 self.blocks[blockb] |= u64::MAX >> (64 - bitb);
 811             }
 812             // Fill in all the other blocks (much faster than one bit at a time).
 813             for block in (blocka + 1)..blockb {
 814                 self.blocks[block] = u64::MAX;
 815             }
 816         } else {
 817             // Set `bita..64` to `0`.
 818             self.blocks[blocka] &= !(u64::MAX << bita);
 819             // Set `0..bitb` to `0`.
 820             if bitb != 0 {
 821                 self.blocks[blockb] &= !(u64::MAX >> (64 - bitb));
 822             }
 823             // Fill in all the other blocks (much faster than one bit at a time).
 824             for block in (blocka + 1)..blockb {
 825                 self.blocks[block] = 0;
 826             }
 827         }
 828     }
 829
 830     #[inline]
 831     pub fn get(&self, i: Size) -> bool {
 832         let (block, bit) = bit_index(i);
 833         (self.blocks[block] & (1 << bit)) != 0
 834     }
 835
 836     #[inline]
 837     pub fn set(&mut self, i: Size, new_state: bool) {
 838         let (block, bit) = bit_index(i);
 839         self.set_bit(block, bit, new_state);
 840     }
 841
 842     #[inline]
 843     fn set_bit(&mut self, block: usize, bit: usize, new_state: bool) {
 844         if new_state {
 845             self.blocks[block] |= 1 << bit;
 846         } else {
 847             self.blocks[block] &= !(1 << bit);
 848         }
 849     }
 850
 851     pub fn grow(&mut self, amount: Size, new_state: bool) {
 852         if amount.bytes() == 0 {
 853             return;
 854         }
 855         let unused_trailing_bits =
 856             u64::try_from(self.blocks.len()).unwrap() * Self::BLOCK_SIZE - self.len.bytes();
 857         if amount.bytes() > unused_trailing_bits {
 858             let additional_blocks = amount.bytes() / Self::BLOCK_SIZE + 1;
 859             self.blocks.extend(
 860                 // FIXME(oli-obk): optimize this by repeating `new_state as Block`.
 861                 iter::repeat(0).take(usize::try_from(additional_blocks).unwrap()),
 862             );
 863         }
 864         let start = self.len;
 865         self.len += amount;
 866         self.set_range_inbounds(start, start + amount, new_state); // `Size` operation
 867     }
 868 }
 869
 870 #[inline]
 871 fn bit_index(bits: Size) -> (usize, usize) {
 872     let bits = bits.bytes();
 873     let a = bits / InitMask::BLOCK_SIZE;
 874     let b = bits % InitMask::BLOCK_SIZE;
 875     (usize::try_from(a).unwrap(), usize::try_from(b).unwrap())
 876 }