src/librustc/mir/interpret/allocation.rs

   1 //! The virtual memory representation of the MIR interpreter.
   2
   3 use super::{
   4     read_target_uint, write_target_uint, AllocId, InterpResult, Pointer, Scalar, ScalarMaybeUndef,
   5 };
   6
   7 use crate::ty::layout::{Align, Size};
   8
   9 use rustc_ast::ast::Mutability;
  10 use rustc_data_structures::sorted_map::SortedMap;
  11 use rustc_target::abi::HasDataLayout;
  12 use std::borrow::Cow;
  13 use std::iter;
  14 use std::ops::{Deref, DerefMut, Range};
  15
  16 // NOTE: When adding new fields, make sure to adjust the `Snapshot` impl in
  17 // `src/librustc_mir/interpret/snapshot.rs`.
  18 #[derive(
  19     Clone,
  20     Debug,
  21     Eq,
  22     PartialEq,
  23     PartialOrd,
  24     Ord,
  25     Hash,
  26     RustcEncodable,
  27     RustcDecodable,
  28     HashStable
  29 )]
  30 pub struct Allocation<Tag = (), Extra = ()> {
  31     /// The actual bytes of the allocation.
  32     /// Note that the bytes of a pointer represent the offset of the pointer.
  33     bytes: Vec<u8>,
  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.
  40     undef_mask: UndefMask,
  41     /// The size of the allocation. Currently, must always equal `bytes.len()`.
  42     pub size: Size,
  43     /// The alignment of the allocation to detect unaligned reads.
  44     pub align: Align,
  45     /// `true` if the allocation is mutable.
  46     /// Also used by codegen to determine if a static should be put into mutable memory,
  47     /// which happens for `static mut` and `static` with interior mutability.
  48     pub mutability: Mutability,
  49     /// Extra state for the machine.
  50     pub extra: Extra,
  51 }
  52
  53 pub trait AllocationExtra<Tag>: ::std::fmt::Debug + Clone {
  54     // There is no constructor in here because the constructor's type depends
  55     // on `MemoryKind`, and making things sufficiently generic leads to painful
  56     // inference failure.
  57
  58     /// Hook for performing extra checks on a memory read access.
  59     ///
  60     /// Takes read-only access to the allocation so we can keep all the memory read
  61     /// operations take `&self`. Use a `RefCell` in `AllocExtra` if you
  62     /// need to mutate.
  63     #[inline(always)]
  64     fn memory_read(
  65         _alloc: &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 write access.
  73     #[inline(always)]
  74     fn memory_written(
  75         _alloc: &mut Allocation<Tag, Self>,
  76         _ptr: Pointer<Tag>,
  77         _size: Size,
  78     ) -> InterpResult<'tcx> {
  79         Ok(())
  80     }
  81
  82     /// Hook for performing extra checks on a memory deallocation.
  83     /// `size` will be the size of the allocation.
  84     #[inline(always)]
  85     fn memory_deallocated(
  86         _alloc: &mut Allocation<Tag, Self>,
  87         _ptr: Pointer<Tag>,
  88         _size: Size,
  89     ) -> InterpResult<'tcx> {
  90         Ok(())
  91     }
  92 }
  93
  94 // For `Tag = ()` and no extra state, we have a trivial implementation.
  95 impl AllocationExtra<()> for () {}
  96
  97 // The constructors are all without extra; the extra gets added by a machine hook later.
  98 impl<Tag> Allocation<Tag> {
  99     /// Creates a read-only allocation initialized by the given bytes
 100     pub fn from_bytes<'a>(slice: impl Into<Cow<'a, [u8]>>, align: Align) -> Self {
 101         let bytes = slice.into().into_owned();
 102         let size = Size::from_bytes(bytes.len() as u64);
 103         Self {
 104             bytes,
 105             relocations: Relocations::new(),
 106             undef_mask: UndefMask::new(size, true),
 107             size,
 108             align,
 109             mutability: Mutability::Not,
 110             extra: (),
 111         }
 112     }
 113
 114     pub fn from_byte_aligned_bytes<'a>(slice: impl Into<Cow<'a, [u8]>>) -> Self {
 115         Allocation::from_bytes(slice, Align::from_bytes(1).unwrap())
 116     }
 117
 118     pub fn undef(size: Size, align: Align) -> Self {
 119         assert_eq!(size.bytes() as usize as u64, size.bytes());
 120         Allocation {
 121             bytes: vec![0; size.bytes() as usize],
 122             relocations: Relocations::new(),
 123             undef_mask: UndefMask::new(size, false),
 124             size,
 125             align,
 126             mutability: Mutability::Mut,
 127             extra: (),
 128         }
 129     }
 130 }
 131
 132 impl Allocation<(), ()> {
 133     /// Add Tag and Extra fields
 134     pub fn with_tags_and_extra<T, E>(
 135         self,
 136         mut tagger: impl FnMut(AllocId) -> T,
 137         extra: E,
 138     ) -> Allocation<T, E> {
 139         Allocation {
 140             bytes: self.bytes,
 141             size: self.size,
 142             relocations: Relocations::from_presorted(
 143                 self.relocations
 144                     .iter()
 145                     // The allocations in the relocations (pointers stored *inside* this allocation)
 146                     // all get the base pointer tag.
 147                     .map(|&(offset, ((), alloc))| {
 148                         let tag = tagger(alloc);
 149                         (offset, (tag, alloc))
 150                     })
 151                     .collect(),
 152             ),
 153             undef_mask: self.undef_mask,
 154             align: self.align,
 155             mutability: self.mutability,
 156             extra,
 157         }
 158     }
 159 }
 160
 161 /// Raw accessors. Provide access to otherwise private bytes.
 162 impl<Tag, Extra> Allocation<Tag, Extra> {
 163     pub fn len(&self) -> usize {
 164         self.size.bytes() as usize
 165     }
 166
 167     /// Looks at a slice which may describe undefined bytes or describe a relocation. This differs
 168     /// from `get_bytes_with_undef_and_ptr` in that it does no relocation checks (even on the
 169     /// edges) at all. It further ignores `AllocationExtra` callbacks.
 170     /// This must not be used for reads affecting the interpreter execution.
 171     pub fn inspect_with_undef_and_ptr_outside_interpreter(&self, range: Range<usize>) -> &[u8] {
 172         &self.bytes[range]
 173     }
 174
 175     /// Returns the undef mask.
 176     pub fn undef_mask(&self) -> &UndefMask {
 177         &self.undef_mask
 178     }
 179
 180     /// Returns the relocation list.
 181     pub fn relocations(&self) -> &Relocations<Tag> {
 182         &self.relocations
 183     }
 184 }
 185
 186 impl<'tcx> rustc_serialize::UseSpecializedDecodable for &'tcx Allocation {}
 187
 188 /// Byte accessors.
 189 impl<'tcx, Tag: Copy, Extra: AllocationExtra<Tag>> Allocation<Tag, Extra> {
 190     /// Just a small local helper function to avoid a bit of code repetition.
 191     /// Returns the range of this allocation that was meant.
 192     #[inline]
 193     fn check_bounds(&self, offset: Size, size: Size) -> Range<usize> {
 194         let end = offset + size; // This does overflow checking.
 195         assert_eq!(
 196             end.bytes() as usize as u64,
 197             end.bytes(),
 198             "cannot handle this access on this host architecture"
 199         );
 200         let end = end.bytes() as usize;
 201         assert!(
 202             end <= self.len(),
 203             "Out-of-bounds access at offset {}, size {} in allocation of size {}",
 204             offset.bytes(),
 205             size.bytes(),
 206             self.len()
 207         );
 208         (offset.bytes() as usize)..end
 209     }
 210
 211     /// The last argument controls whether we error out when there are undefined
 212     /// or pointer bytes. You should never call this, call `get_bytes` or
 213     /// `get_bytes_with_undef_and_ptr` instead,
 214     ///
 215     /// This function also guarantees that the resulting pointer will remain stable
 216     /// even when new allocations are pushed to the `HashMap`. `copy_repeatedly` relies
 217     /// on that.
 218     ///
 219     /// It is the caller's responsibility to check bounds and alignment beforehand.
 220     fn get_bytes_internal(
 221         &self,
 222         cx: &impl HasDataLayout,
 223         ptr: Pointer<Tag>,
 224         size: Size,
 225         check_defined_and_ptr: bool,
 226     ) -> InterpResult<'tcx, &[u8]> {
 227         let range = self.check_bounds(ptr.offset, size);
 228
 229         if check_defined_and_ptr {
 230             self.check_defined(ptr, size)?;
 231             self.check_relocations(cx, ptr, size)?;
 232         } else {
 233             // We still don't want relocations on the *edges*.
 234             self.check_relocation_edges(cx, ptr, size)?;
 235         }
 236
 237         AllocationExtra::memory_read(self, ptr, size)?;
 238
 239         Ok(&self.bytes[range])
 240     }
 241
 242     /// Checks that these bytes are initialized and not pointer bytes, and then return them
 243     /// as a slice.
 244     ///
 245     /// It is the caller's responsibility to check bounds and alignment beforehand.
 246     /// Most likely, you want to use the `PlaceTy` and `OperandTy`-based methods
 247     /// on `InterpCx` instead.
 248     #[inline]
 249     pub fn get_bytes(
 250         &self,
 251         cx: &impl HasDataLayout,
 252         ptr: Pointer<Tag>,
 253         size: Size,
 254     ) -> InterpResult<'tcx, &[u8]> {
 255         self.get_bytes_internal(cx, ptr, size, true)
 256     }
 257
 258     /// It is the caller's responsibility to handle undefined and pointer bytes.
 259     /// However, this still checks that there are no relocations on the *edges*.
 260     ///
 261     /// It is the caller's responsibility to check bounds and alignment beforehand.
 262     #[inline]
 263     pub fn get_bytes_with_undef_and_ptr(
 264         &self,
 265         cx: &impl HasDataLayout,
 266         ptr: Pointer<Tag>,
 267         size: Size,
 268     ) -> InterpResult<'tcx, &[u8]> {
 269         self.get_bytes_internal(cx, ptr, size, false)
 270     }
 271
 272     /// Just calling this already marks everything as defined and removes relocations,
 273     /// so be sure to actually put data there!
 274     ///
 275     /// It is the caller's responsibility to check bounds and alignment beforehand.
 276     /// Most likely, you want to use the `PlaceTy` and `OperandTy`-based methods
 277     /// on `InterpCx` instead.
 278     pub fn get_bytes_mut(
 279         &mut self,
 280         cx: &impl HasDataLayout,
 281         ptr: Pointer<Tag>,
 282         size: Size,
 283     ) -> InterpResult<'tcx, &mut [u8]> {
 284         let range = self.check_bounds(ptr.offset, size);
 285
 286         self.mark_definedness(ptr, size, true);
 287         self.clear_relocations(cx, ptr, size)?;
 288
 289         AllocationExtra::memory_written(self, ptr, size)?;
 290
 291         Ok(&mut self.bytes[range])
 292     }
 293 }
 294
 295 /// Reading and writing.
 296 impl<'tcx, Tag: Copy, Extra: AllocationExtra<Tag>> Allocation<Tag, Extra> {
 297     /// Reads bytes until a `0` is encountered. Will error if the end of the allocation is reached
 298     /// before a `0` is found.
 299     ///
 300     /// Most likely, you want to call `Memory::read_c_str` instead of this method.
 301     pub fn read_c_str(
 302         &self,
 303         cx: &impl HasDataLayout,
 304         ptr: Pointer<Tag>,
 305     ) -> InterpResult<'tcx, &[u8]> {
 306         assert_eq!(ptr.offset.bytes() as usize as u64, ptr.offset.bytes());
 307         let offset = ptr.offset.bytes() as usize;
 308         Ok(match self.bytes[offset..].iter().position(|&c| c == 0) {
 309             Some(size) => {
 310                 let size_with_null = Size::from_bytes((size + 1) as u64);
 311                 // Go through `get_bytes` for checks and AllocationExtra hooks.
 312                 // We read the null, so we include it in the request, but we want it removed
 313                 // from the result, so we do subslicing.
 314                 &self.get_bytes(cx, ptr, size_with_null)?[..size]
 315             }
 316             // This includes the case where `offset` is out-of-bounds to begin with.
 317             None => throw_unsup!(UnterminatedCString(ptr.erase_tag())),
 318         })
 319     }
 320
 321     /// Validates that `ptr.offset` and `ptr.offset + size` do not point to the middle of a
 322     /// relocation. If `allow_ptr_and_undef` is `false`, also enforces that the memory in the
 323     /// given range contains neither relocations nor undef bytes.
 324     pub fn check_bytes(
 325         &self,
 326         cx: &impl HasDataLayout,
 327         ptr: Pointer<Tag>,
 328         size: Size,
 329         allow_ptr_and_undef: bool,
 330     ) -> InterpResult<'tcx> {
 331         // Check bounds and relocations on the edges.
 332         self.get_bytes_with_undef_and_ptr(cx, ptr, size)?;
 333         // Check undef and ptr.
 334         if !allow_ptr_and_undef {
 335             self.check_defined(ptr, size)?;
 336             self.check_relocations(cx, ptr, size)?;
 337         }
 338         Ok(())
 339     }
 340
 341     /// Writes `src` to the memory starting at `ptr.offset`.
 342     ///
 343     /// It is the caller's responsibility to check bounds and alignment beforehand.
 344     /// Most likely, you want to call `Memory::write_bytes` instead of this method.
 345     pub fn write_bytes(
 346         &mut self,
 347         cx: &impl HasDataLayout,
 348         ptr: Pointer<Tag>,
 349         src: impl IntoIterator<Item = u8>,
 350     ) -> InterpResult<'tcx> {
 351         let mut src = src.into_iter();
 352         let (lower, upper) = src.size_hint();
 353         let len = upper.expect("can only write bounded iterators");
 354         assert_eq!(lower, len, "can only write iterators with a precise length");
 355         let bytes = self.get_bytes_mut(cx, ptr, Size::from_bytes(len as u64))?;
 356         // `zip` would stop when the first iterator ends; we want to definitely
 357         // cover all of `bytes`.
 358         for dest in bytes {
 359             *dest = src.next().expect("iterator was shorter than it said it would be");
 360         }
 361         src.next().expect_none("iterator was longer than it said it would be");
 362         Ok(())
 363     }
 364
 365     /// Reads a *non-ZST* scalar.
 366     ///
 367     /// ZSTs can't be read for two reasons:
 368     /// * byte-order cannot work with zero-element buffers;
 369     /// * in order to obtain a `Pointer`, we need to check for ZSTness anyway due to integer
 370     ///   pointers being valid for ZSTs.
 371     ///
 372     /// It is the caller's responsibility to check bounds and alignment beforehand.
 373     /// Most likely, you want to call `InterpCx::read_scalar` instead of this method.
 374     pub fn read_scalar(
 375         &self,
 376         cx: &impl HasDataLayout,
 377         ptr: Pointer<Tag>,
 378         size: Size,
 379     ) -> InterpResult<'tcx, ScalarMaybeUndef<Tag>> {
 380         // `get_bytes_unchecked` tests relocation edges.
 381         let bytes = self.get_bytes_with_undef_and_ptr(cx, ptr, size)?;
 382         // Undef check happens *after* we established that the alignment is correct.
 383         // We must not return `Ok()` for unaligned pointers!
 384         if self.check_defined(ptr, size).is_err() {
 385             // This inflates undefined bytes to the entire scalar, even if only a few
 386             // bytes are undefined.
 387             return Ok(ScalarMaybeUndef::Undef);
 388         }
 389         // Now we do the actual reading.
 390         let bits = read_target_uint(cx.data_layout().endian, bytes).unwrap();
 391         // See if we got a pointer.
 392         if size != cx.data_layout().pointer_size {
 393             // *Now*, we better make sure that the inside is free of relocations too.
 394             self.check_relocations(cx, ptr, size)?;
 395         } else {
 396             match self.relocations.get(&ptr.offset) {
 397                 Some(&(tag, alloc_id)) => {
 398                     let ptr = Pointer::new_with_tag(alloc_id, Size::from_bytes(bits as u64), tag);
 399                     return Ok(ScalarMaybeUndef::Scalar(ptr.into()));
 400                 }
 401                 None => {}
 402             }
 403         }
 404         // We don't. Just return the bits.
 405         Ok(ScalarMaybeUndef::Scalar(Scalar::from_uint(bits, size)))
 406     }
 407
 408     /// Reads a pointer-sized scalar.
 409     ///
 410     /// It is the caller's responsibility to check bounds and alignment beforehand.
 411     /// Most likely, you want to call `InterpCx::read_scalar` instead of this method.
 412     pub fn read_ptr_sized(
 413         &self,
 414         cx: &impl HasDataLayout,
 415         ptr: Pointer<Tag>,
 416     ) -> InterpResult<'tcx, ScalarMaybeUndef<Tag>> {
 417         self.read_scalar(cx, ptr, cx.data_layout().pointer_size)
 418     }
 419
 420     /// Writes a *non-ZST* scalar.
 421     ///
 422     /// ZSTs can't be read for two reasons:
 423     /// * byte-order cannot work with zero-element buffers;
 424     /// * in order to obtain a `Pointer`, we need to check for ZSTness anyway due to integer
 425     ///   pointers being valid for ZSTs.
 426     ///
 427     /// It is the caller's responsibility to check bounds and alignment beforehand.
 428     /// Most likely, you want to call `InterpCx::write_scalar` instead of this method.
 429     pub fn write_scalar(
 430         &mut self,
 431         cx: &impl HasDataLayout,
 432         ptr: Pointer<Tag>,
 433         val: ScalarMaybeUndef<Tag>,
 434         type_size: Size,
 435     ) -> InterpResult<'tcx> {
 436         let val = match val {
 437             ScalarMaybeUndef::Scalar(scalar) => scalar,
 438             ScalarMaybeUndef::Undef => {
 439                 self.mark_definedness(ptr, type_size, false);
 440                 return Ok(());
 441             }
 442         };
 443
 444         let bytes = match val.to_bits_or_ptr(type_size, cx) {
 445             Err(val) => val.offset.bytes() as u128,
 446             Ok(data) => data,
 447         };
 448
 449         let endian = cx.data_layout().endian;
 450         let dst = self.get_bytes_mut(cx, ptr, type_size)?;
 451         write_target_uint(endian, dst, bytes).unwrap();
 452
 453         // See if we have to also write a relocation.
 454         match val {
 455             Scalar::Ptr(val) => {
 456                 self.relocations.insert(ptr.offset, (val.tag, val.alloc_id));
 457             }
 458             _ => {}
 459         }
 460
 461         Ok(())
 462     }
 463
 464     /// Writes a pointer-sized scalar.
 465     ///
 466     /// It is the caller's responsibility to check bounds and alignment beforehand.
 467     /// Most likely, you want to call `InterpCx::write_scalar` instead of this method.
 468     pub fn write_ptr_sized(
 469         &mut self,
 470         cx: &impl HasDataLayout,
 471         ptr: Pointer<Tag>,
 472         val: ScalarMaybeUndef<Tag>,
 473     ) -> InterpResult<'tcx> {
 474         let ptr_size = cx.data_layout().pointer_size;
 475         self.write_scalar(cx, ptr, val, ptr_size)
 476     }
 477 }
 478
 479 /// Relocations.
 480 impl<'tcx, Tag: Copy, Extra> Allocation<Tag, Extra> {
 481     /// Returns all relocations overlapping with the given pointer-offset pair.
 482     pub fn get_relocations(
 483         &self,
 484         cx: &impl HasDataLayout,
 485         ptr: Pointer<Tag>,
 486         size: Size,
 487     ) -> &[(Size, (Tag, AllocId))] {
 488         // We have to go back `pointer_size - 1` bytes, as that one would still overlap with
 489         // the beginning of this range.
 490         let start = ptr.offset.bytes().saturating_sub(cx.data_layout().pointer_size.bytes() - 1);
 491         let end = ptr.offset + size; // This does overflow checking.
 492         self.relocations.range(Size::from_bytes(start)..end)
 493     }
 494
 495     /// Checks that there are no relocations overlapping with the given range.
 496     #[inline(always)]
 497     fn check_relocations(
 498         &self,
 499         cx: &impl HasDataLayout,
 500         ptr: Pointer<Tag>,
 501         size: Size,
 502     ) -> InterpResult<'tcx> {
 503         if self.get_relocations(cx, ptr, size).is_empty() {
 504             Ok(())
 505         } else {
 506             throw_unsup!(ReadPointerAsBytes)
 507         }
 508     }
 509
 510     /// Removes all relocations inside the given range.
 511     /// If there are relocations overlapping with the edges, they
 512     /// are removed as well *and* the bytes they cover are marked as
 513     /// uninitialized. This is a somewhat odd "spooky action at a distance",
 514     /// but it allows strictly more code to run than if we would just error
 515     /// immediately in that case.
 516     fn clear_relocations(
 517         &mut self,
 518         cx: &impl HasDataLayout,
 519         ptr: Pointer<Tag>,
 520         size: Size,
 521     ) -> InterpResult<'tcx> {
 522         // Find the start and end of the given range and its outermost relocations.
 523         let (first, last) = {
 524             // Find all relocations overlapping the given range.
 525             let relocations = self.get_relocations(cx, ptr, size);
 526             if relocations.is_empty() {
 527                 return Ok(());
 528             }
 529
 530             (
 531                 relocations.first().unwrap().0,
 532                 relocations.last().unwrap().0 + cx.data_layout().pointer_size,
 533             )
 534         };
 535         let start = ptr.offset;
 536         let end = start + size;
 537
 538         // Mark parts of the outermost relocations as undefined if they partially fall outside the
 539         // given range.
 540         if first < start {
 541             self.undef_mask.set_range(first, start, false);
 542         }
 543         if last > end {
 544             self.undef_mask.set_range(end, last, false);
 545         }
 546
 547         // Forget all the relocations.
 548         self.relocations.remove_range(first..last);
 549
 550         Ok(())
 551     }
 552
 553     /// Errors if there are relocations overlapping with the edges of the
 554     /// given memory range.
 555     #[inline]
 556     fn check_relocation_edges(
 557         &self,
 558         cx: &impl HasDataLayout,
 559         ptr: Pointer<Tag>,
 560         size: Size,
 561     ) -> InterpResult<'tcx> {
 562         self.check_relocations(cx, ptr, Size::ZERO)?;
 563         self.check_relocations(cx, ptr.offset(size, cx)?, Size::ZERO)?;
 564         Ok(())
 565     }
 566 }
 567
 568 /// Undefined bytes.
 569 impl<'tcx, Tag, Extra> Allocation<Tag, Extra> {
 570     /// Checks that a range of bytes is defined. If not, returns the `ReadUndefBytes`
 571     /// error which will report the first byte which is undefined.
 572     #[inline]
 573     fn check_defined(&self, ptr: Pointer<Tag>, size: Size) -> InterpResult<'tcx> {
 574         self.undef_mask
 575             .is_range_defined(ptr.offset, ptr.offset + size)
 576             .or_else(|idx| throw_unsup!(ReadUndefBytes(idx)))
 577     }
 578
 579     pub fn mark_definedness(&mut self, ptr: Pointer<Tag>, size: Size, new_state: bool) {
 580         if size.bytes() == 0 {
 581             return;
 582         }
 583         self.undef_mask.set_range(ptr.offset, ptr.offset + size, new_state);
 584     }
 585 }
 586
 587 /// Run-length encoding of the undef mask.
 588 /// Used to copy parts of a mask multiple times to another allocation.
 589 pub struct AllocationDefinedness {
 590     /// The definedness of the first range.
 591     initial: bool,
 592     /// The lengths of ranges that are run-length encoded.
 593     /// The definedness of the ranges alternate starting with `initial`.
 594     ranges: smallvec::SmallVec<[u64; 1]>,
 595 }
 596
 597 impl AllocationDefinedness {
 598     pub fn all_bytes_undef(&self) -> bool {
 599         // The `ranges` are run-length encoded and of alternating definedness.
 600         // So if `ranges.len() > 1` then the second block is a range of defined.
 601         !self.initial && self.ranges.len() == 1
 602     }
 603 }
 604
 605 /// Transferring the definedness mask to other allocations.
 606 impl<Tag, Extra> Allocation<Tag, Extra> {
 607     /// Creates a run-length encoding of the undef mask.
 608     pub fn compress_undef_range(&self, src: Pointer<Tag>, size: Size) -> AllocationDefinedness {
 609         // Since we are copying `size` bytes from `src` to `dest + i * size` (`for i in 0..repeat`),
 610         // a naive undef mask copying algorithm would repeatedly have to read the undef mask from
 611         // the source and write it to the destination. Even if we optimized the memory accesses,
 612         // we'd be doing all of this `repeat` times.
 613         // Therefore we precompute a compressed version of the undef mask of the source value and
 614         // then write it back `repeat` times without computing any more information from the source.
 615
 616         // A precomputed cache for ranges of defined/undefined bits
 617         // 0000010010001110 will become
 618         // `[5, 1, 2, 1, 3, 3, 1]`,
 619         // where each element toggles the state.
 620
 621         let mut ranges = smallvec::SmallVec::<[u64; 1]>::new();
 622         let initial = self.undef_mask.get(src.offset);
 623         let mut cur_len = 1;
 624         let mut cur = initial;
 625
 626         for i in 1..size.bytes() {
 627             // FIXME: optimize to bitshift the current undef block's bits and read the top bit.
 628             if self.undef_mask.get(src.offset + Size::from_bytes(i)) == cur {
 629                 cur_len += 1;
 630             } else {
 631                 ranges.push(cur_len);
 632                 cur_len = 1;
 633                 cur = !cur;
 634             }
 635         }
 636
 637         ranges.push(cur_len);
 638
 639         AllocationDefinedness { ranges, initial }
 640     }
 641
 642     /// Applies multiple instances of the run-length encoding to the undef mask.
 643     pub fn mark_compressed_undef_range(
 644         &mut self,
 645         defined: &AllocationDefinedness,
 646         dest: Pointer<Tag>,
 647         size: Size,
 648         repeat: u64,
 649     ) {
 650         // An optimization where we can just overwrite an entire range of definedness bits if
 651         // they are going to be uniformly `1` or `0`.
 652         if defined.ranges.len() <= 1 {
 653             self.undef_mask.set_range_inbounds(
 654                 dest.offset,
 655                 dest.offset + size * repeat,
 656                 defined.initial,
 657             );
 658             return;
 659         }
 660
 661         for mut j in 0..repeat {
 662             j *= size.bytes();
 663             j += dest.offset.bytes();
 664             let mut cur = defined.initial;
 665             for range in &defined.ranges {
 666                 let old_j = j;
 667                 j += range;
 668                 self.undef_mask.set_range_inbounds(
 669                     Size::from_bytes(old_j),
 670                     Size::from_bytes(j),
 671                     cur,
 672                 );
 673                 cur = !cur;
 674             }
 675         }
 676     }
 677 }
 678
 679 /// Relocations.
 680 #[derive(Clone, PartialEq, Eq, PartialOrd, Ord, Hash, Debug, RustcEncodable, RustcDecodable)]
 681 pub struct Relocations<Tag = (), Id = AllocId>(SortedMap<Size, (Tag, Id)>);
 682
 683 impl<Tag, Id> Relocations<Tag, Id> {
 684     pub fn new() -> Self {
 685         Relocations(SortedMap::new())
 686     }
 687
 688     // The caller must guarantee that the given relocations are already sorted
 689     // by address and contain no duplicates.
 690     pub fn from_presorted(r: Vec<(Size, (Tag, Id))>) -> Self {
 691         Relocations(SortedMap::from_presorted_elements(r))
 692     }
 693 }
 694
 695 impl<Tag> Deref for Relocations<Tag> {
 696     type Target = SortedMap<Size, (Tag, AllocId)>;
 697
 698     fn deref(&self) -> &Self::Target {
 699         &self.0
 700     }
 701 }
 702
 703 impl<Tag> DerefMut for Relocations<Tag> {
 704     fn deref_mut(&mut self) -> &mut Self::Target {
 705         &mut self.0
 706     }
 707 }
 708
 709 /// A partial, owned list of relocations to transfer into another allocation.
 710 pub struct AllocationRelocations<Tag> {
 711     relative_relocations: Vec<(Size, (Tag, AllocId))>,
 712 }
 713
 714 impl<Tag: Copy, Extra> Allocation<Tag, Extra> {
 715     pub fn prepare_relocation_copy(
 716         &self,
 717         cx: &impl HasDataLayout,
 718         src: Pointer<Tag>,
 719         size: Size,
 720         dest: Pointer<Tag>,
 721         length: u64,
 722     ) -> AllocationRelocations<Tag> {
 723         let relocations = self.get_relocations(cx, src, size);
 724         if relocations.is_empty() {
 725             return AllocationRelocations { relative_relocations: Vec::new() };
 726         }
 727
 728         let mut new_relocations = Vec::with_capacity(relocations.len() * (length as usize));
 729
 730         for i in 0..length {
 731             new_relocations.extend(relocations.iter().map(|&(offset, reloc)| {
 732                 // compute offset for current repetition
 733                 let dest_offset = dest.offset + (i * size);
 734                 (
 735                     // shift offsets from source allocation to destination allocation
 736                     offset + dest_offset - src.offset,
 737                     reloc,
 738                 )
 739             }));
 740         }
 741
 742         AllocationRelocations { relative_relocations: new_relocations }
 743     }
 744
 745     /// Applies a relocation copy.
 746     /// The affected range, as defined in the parameters to `prepare_relocation_copy` is expected
 747     /// to be clear of relocations.
 748     pub fn mark_relocation_range(&mut self, relocations: AllocationRelocations<Tag>) {
 749         self.relocations.insert_presorted(relocations.relative_relocations);
 750     }
 751 }
 752
 753 ////////////////////////////////////////////////////////////////////////////////
 754 // Undefined byte tracking
 755 ////////////////////////////////////////////////////////////////////////////////
 756
 757 type Block = u64;
 758
 759 /// A bitmask where each bit refers to the byte with the same index. If the bit is `true`, the byte
 760 /// is defined. If it is `false` the byte is undefined.
 761 #[derive(
 762     Clone,
 763     Debug,
 764     Eq,
 765     PartialEq,
 766     PartialOrd,
 767     Ord,
 768     Hash,
 769     RustcEncodable,
 770     RustcDecodable,
 771     HashStable
 772 )]
 773 pub struct UndefMask {
 774     blocks: Vec<Block>,
 775     len: Size,
 776 }
 777
 778 impl UndefMask {
 779     pub const BLOCK_SIZE: u64 = 64;
 780
 781     pub fn new(size: Size, state: bool) -> Self {
 782         let mut m = UndefMask { blocks: vec![], len: Size::ZERO };
 783         m.grow(size, state);
 784         m
 785     }
 786
 787     /// Checks whether the range `start..end` (end-exclusive) is entirely defined.
 788     ///
 789     /// Returns `Ok(())` if it's defined. Otherwise returns the index of the byte
 790     /// at which the first undefined access begins.
 791     #[inline]
 792     pub fn is_range_defined(&self, start: Size, end: Size) -> Result<(), Size> {
 793         if end > self.len {
 794             return Err(self.len);
 795         }
 796
 797         // FIXME(oli-obk): optimize this for allocations larger than a block.
 798         let idx = (start.bytes()..end.bytes()).map(|i| Size::from_bytes(i)).find(|&i| !self.get(i));
 799
 800         match idx {
 801             Some(idx) => Err(idx),
 802             None => Ok(()),
 803         }
 804     }
 805
 806     pub fn set_range(&mut self, start: Size, end: Size, new_state: bool) {
 807         let len = self.len;
 808         if end > len {
 809             self.grow(end - len, new_state);
 810         }
 811         self.set_range_inbounds(start, end, new_state);
 812     }
 813
 814     pub fn set_range_inbounds(&mut self, start: Size, end: Size, new_state: bool) {
 815         let (blocka, bita) = bit_index(start);
 816         let (blockb, bitb) = bit_index(end);
 817         if blocka == blockb {
 818             // First set all bits except the first `bita`,
 819             // then unset the last `64 - bitb` bits.
 820             let range = if bitb == 0 {
 821                 u64::MAX << bita
 822             } else {
 823                 (u64::MAX << bita) & (u64::MAX >> (64 - bitb))
 824             };
 825             if new_state {
 826                 self.blocks[blocka] |= range;
 827             } else {
 828                 self.blocks[blocka] &= !range;
 829             }
 830             return;
 831         }
 832         // across block boundaries
 833         if new_state {
 834             // Set `bita..64` to `1`.
 835             self.blocks[blocka] |= u64::MAX << bita;
 836             // Set `0..bitb` to `1`.
 837             if bitb != 0 {
 838                 self.blocks[blockb] |= u64::MAX >> (64 - bitb);
 839             }
 840             // Fill in all the other blocks (much faster than one bit at a time).
 841             for block in (blocka + 1)..blockb {
 842                 self.blocks[block] = u64::MAX;
 843             }
 844         } else {
 845             // Set `bita..64` to `0`.
 846             self.blocks[blocka] &= !(u64::MAX << bita);
 847             // Set `0..bitb` to `0`.
 848             if bitb != 0 {
 849                 self.blocks[blockb] &= !(u64::MAX >> (64 - bitb));
 850             }
 851             // Fill in all the other blocks (much faster than one bit at a time).
 852             for block in (blocka + 1)..blockb {
 853                 self.blocks[block] = 0;
 854             }
 855         }
 856     }
 857
 858     #[inline]
 859     pub fn get(&self, i: Size) -> bool {
 860         let (block, bit) = bit_index(i);
 861         (self.blocks[block] & (1 << bit)) != 0
 862     }
 863
 864     #[inline]
 865     pub fn set(&mut self, i: Size, new_state: bool) {
 866         let (block, bit) = bit_index(i);
 867         self.set_bit(block, bit, new_state);
 868     }
 869
 870     #[inline]
 871     fn set_bit(&mut self, block: usize, bit: usize, new_state: bool) {
 872         if new_state {
 873             self.blocks[block] |= 1 << bit;
 874         } else {
 875             self.blocks[block] &= !(1 << bit);
 876         }
 877     }
 878
 879     pub fn grow(&mut self, amount: Size, new_state: bool) {
 880         if amount.bytes() == 0 {
 881             return;
 882         }
 883         let unused_trailing_bits = self.blocks.len() as u64 * Self::BLOCK_SIZE - self.len.bytes();
 884         if amount.bytes() > unused_trailing_bits {
 885             let additional_blocks = amount.bytes() / Self::BLOCK_SIZE + 1;
 886             assert_eq!(additional_blocks as usize as u64, additional_blocks);
 887             self.blocks.extend(
 888                 // FIXME(oli-obk): optimize this by repeating `new_state as Block`.
 889                 iter::repeat(0).take(additional_blocks as usize),
 890             );
 891         }
 892         let start = self.len;
 893         self.len += amount;
 894         self.set_range_inbounds(start, start + amount, new_state);
 895     }
 896 }
 897
 898 #[inline]
 899 fn bit_index(bits: Size) -> (usize, usize) {
 900     let bits = bits.bytes();
 901     let a = bits / UndefMask::BLOCK_SIZE;
 902     let b = bits % UndefMask::BLOCK_SIZE;
 903     assert_eq!(a as usize as u64, a);
 904     assert_eq!(b as usize as u64, b);
 905     (a as usize, b as usize)
 906 }