SegmentedVector.h

/* -*- Mode: C++; tab-width: 8; indent-tabs-mode: nil; c-basic-offset: 2 -*- */
/* vim: set ts=8 sts=2 et sw=2 tw=80: */
/* This Source Code Form is subject to the terms of the Mozilla Public
 * License, v. 2.0. If a copy of the MPL was not distributed with this
 * file, You can obtain one at http://mozilla.org/MPL/2.0/. */

// A simple segmented vector class.
//
// This class should be used in preference to mozilla::Vector or nsTArray when
// you are simply gathering items in order to later iterate over them.
//
// - In the case where you don't know the final size in advance, using
//   SegmentedVector avoids the need to repeatedly allocate increasingly large
//   buffers and copy the data into them.
//
// - In the case where you know the final size in advance and so can set the
//   capacity appropriately, using SegmentedVector still avoids the need for
//   large allocations (which can trigger OOMs).

#ifndef mozilla_SegmentedVector_h
#define mozilla_SegmentedVector_h

#include <new>  // for placement new
#include <utility>

#include "mozilla/AllocPolicy.h"
#include "mozilla/Array.h"
#include "mozilla/Attributes.h"
#include "mozilla/LinkedList.h"
#include "mozilla/MemoryReporting.h"
#include "mozilla/OperatorNewExtensions.h"

#ifdef IMPL_LIBXUL
#  include "mozilla/Likely.h"
#  include "mozilla/mozalloc_oom.h"
#endif  // IMPL_LIBXUL

namespace mozilla {

// |IdealSegmentSize| specifies how big each segment will be in bytes (or as
// close as is possible). Use the following guidelines to choose a size.
//
// - It should be a power-of-two, to avoid slop.
//
// - It should not be too small, so that segment allocations are infrequent,
//   and so that per-segment bookkeeping overhead is low. Typically each
//   segment should be able to hold hundreds of elements, at least.
//
// - It should not be too large, so that OOMs are unlikely when allocating
//   segments, and so that not too much space is wasted when the final segment
//   is not full.
//
// The ideal size depends on how the SegmentedVector is used and the size of
// |T|, but reasonable sizes include 1024, 4096 (the default), 8192, and 16384.
//
template <typename T, size_t IdealSegmentSize = 4096,
          typename AllocPolicy = MallocAllocPolicy>
class SegmentedVector : private AllocPolicy {
  template <size_t SegmentCapacity>
  struct SegmentImpl
      : public mozilla::LinkedListElement<SegmentImpl<SegmentCapacity>> {
   private:
    uint32_t mLength;
    alignas(T) MOZ_INIT_OUTSIDE_CTOR
        unsigned char mData[sizeof(T) * SegmentCapacity];

    // Some versions of GCC treat it as a -Wstrict-aliasing violation (ergo a
    // -Werror compile error) to reinterpret_cast<> |mData| to |T*|, even
    // through |void*|.  Placing the latter cast in these separate functions
    // breaks the chain such that affected GCC versions no longer warn/error.
    void* RawData() { return mData; }

   public:
    SegmentImpl() : mLength(0) {}

    ~SegmentImpl() {
      for (uint32_t i = 0; i < mLength; i++) {
        (*this)[i].~T();
      }
    }

    uint32_t Length() const { return mLength; }

    T* Elems() { return reinterpret_cast<T*>(RawData()); }

    T& operator[](size_t aIndex) {
      MOZ_ASSERT(aIndex < mLength);
      return Elems()[aIndex];
    }

    const T& operator[](size_t aIndex) const {
      MOZ_ASSERT(aIndex < mLength);
      return Elems()[aIndex];
    }

    template <typename U>
    void Append(U&& aU) {
      MOZ_ASSERT(mLength < SegmentCapacity);
      // Pre-increment mLength so that the bounds-check in operator[] passes.
      mLength++;
      T* elem = &(*this)[mLength - 1];
      new (KnownNotNull, elem) T(std::forward<U>(aU));
    }

    void PopLast() {
      MOZ_ASSERT(mLength > 0);
      (*this)[mLength - 1].~T();
      mLength--;
    }
  };

  // See how many we elements we can fit in a segment of IdealSegmentSize. If
  // IdealSegmentSize is too small, it'll be just one. The +1 is because
  // kSingleElementSegmentSize already accounts for one element.
  static const size_t kSingleElementSegmentSize = sizeof(SegmentImpl<1>);
  static const size_t kSegmentCapacity =
      kSingleElementSegmentSize <= IdealSegmentSize
          ? (IdealSegmentSize - kSingleElementSegmentSize) / sizeof(T) + 1
          : 1;

 public:
  typedef SegmentImpl<kSegmentCapacity> Segment;

  // The |aIdealSegmentSize| is only for sanity checking. If it's specified, we
  // check that the actual segment size is as close as possible to it. This
  // serves as a sanity check for SegmentedVectorCapacity's capacity
  // computation.
  explicit SegmentedVector(size_t aIdealSegmentSize = 0) {
    // The difference between the actual segment size and the ideal segment
    // size should be less than the size of a single element... unless the
    // ideal size was too small, in which case the capacity should be one.
    MOZ_ASSERT_IF(
        aIdealSegmentSize != 0,
        (sizeof(Segment) > aIdealSegmentSize && kSegmentCapacity == 1) ||
            aIdealSegmentSize - sizeof(Segment) < sizeof(T));
  }

  SegmentedVector(SegmentedVector&& aOther)
      : mSegments(std::move(aOther.mSegments)) {}
  SegmentedVector& operator=(SegmentedVector&& aOther) {
    if (&aOther != this) {
      this->~SegmentedVector();
      new (this) SegmentedVector(std::move(aOther));
    }
    return *this;
  }

  ~SegmentedVector() { Clear(); }

  bool IsEmpty() const { return !mSegments.getFirst(); }

  // Note that this is O(n) rather than O(1), but the constant factor is very
  // small because it only has to do one addition per segment.
  size_t Length() const {
    size_t n = 0;
    for (auto segment = mSegments.getFirst(); segment;
         segment = segment->getNext()) {
      n += segment->Length();
    }
    return n;
  }

  // Returns false if the allocation failed. (If you are using an infallible
  // allocation policy, use InfallibleAppend() instead.)
  template <typename U>
  [[nodiscard]] bool Append(U&& aU) {
    Segment* last = mSegments.getLast();
    if (!last || last->Length() == kSegmentCapacity) {
      last = this->template pod_malloc<Segment>(1);
      if (!last) {
        return false;
      }
      new (KnownNotNull, last) Segment();
      mSegments.insertBack(last);
    }
    last->Append(std::forward<U>(aU));
    return true;
  }

  // You should probably only use this instead of Append() if you are using an
  // infallible allocation policy. It will crash if the allocation fails.
  template <typename U>
  void InfallibleAppend(U&& aU) {
    bool ok = Append(std::forward<U>(aU));

#ifdef IMPL_LIBXUL
    if (MOZ_UNLIKELY(!ok)) {
      mozalloc_handle_oom(sizeof(Segment));
    }
#else
    MOZ_RELEASE_ASSERT(ok);
#endif  // MOZ_INTERNAL_API
  }

  void Clear() {
    Segment* segment;
    while ((segment = mSegments.popFirst())) {
      segment->~Segment();
      this->free_(segment, 1);
    }
  }

  T& GetLast() {
    MOZ_ASSERT(!IsEmpty());
    Segment* last = mSegments.getLast();
    return (*last)[last->Length() - 1];
  }

  const T& GetLast() const {
    MOZ_ASSERT(!IsEmpty());
    Segment* last = mSegments.getLast();
    return (*last)[last->Length() - 1];
  }

  void PopLast() {
    MOZ_ASSERT(!IsEmpty());
    Segment* last = mSegments.getLast();
    last->PopLast();
    if (!last->Length()) {
      mSegments.popLast();
      last->~Segment();
      this->free_(last, 1);
    }
  }

  // Equivalent to calling |PopLast| |aNumElements| times, but potentially
  // more efficient. It is safe to call this even when aNumElements > Length().
  void PopLastN(uint32_t aNumElements) {
    Segment* last;

    // Pop full segments for as long as we can.  Note that this loop
    // cleanly handles the case when the initial last segment is not
    // full and we are popping more elements than said segment contains.
    do {
      last = mSegments.getLast();

      // The list is empty.  We're all done.
      if (!last) {
        return;
      }

      // Check to see if the list contains too many elements.  Handle
      // that in the epilogue.
      uint32_t segmentLen = last->Length();
      if (segmentLen > aNumElements) {
        break;
      }

      // Destroying the segment destroys all elements contained therein.
      mSegments.popLast();
      last->~Segment();
      this->free_(last, 1);

      MOZ_ASSERT(aNumElements >= segmentLen);
      aNumElements -= segmentLen;
      if (aNumElements == 0) {
        return;
      }
    } while (true);

    // Handle the case where the last segment contains more elements
    // than we want to pop.
    MOZ_ASSERT(last);
    MOZ_ASSERT(last == mSegments.getLast());
    MOZ_ASSERT(aNumElements < last->Length());
    for (uint32_t i = 0; i < aNumElements; ++i) {
      last->PopLast();
    }
    MOZ_ASSERT(last->Length() != 0);
  }

  // Use this class to iterate over a SegmentedVector, like so:
  //
  //  for (auto iter = v.Iter(); !iter.Done(); iter.Next()) {
  //    MyElem& elem = iter.Get();
  //    f(elem);
  //  }
  //
  // Note, adding new entries to the SegmentedVector while using iterators
  // is supported, but removing is not!
  // If an iterator has entered Done() state, adding more entries to the
  // vector doesn't affect it.
  class IterImpl {
    friend class SegmentedVector;

    Segment* mSegment;
    size_t mIndex;

    explicit IterImpl(SegmentedVector* aVector, bool aFromFirst)
        : mSegment(aFromFirst ? aVector->mSegments.getFirst()
                              : aVector->mSegments.getLast()),
          mIndex(aFromFirst ? 0 : (mSegment ? mSegment->Length() - 1 : 0)) {
      MOZ_ASSERT_IF(mSegment, mSegment->Length() > 0);
    }

   public:
    bool Done() const {
      MOZ_ASSERT_IF(mSegment, mSegment->isInList());
      MOZ_ASSERT_IF(mSegment, mIndex < mSegment->Length());
      return !mSegment;
    }

    T& Get() {
      MOZ_ASSERT(!Done());
      return (*mSegment)[mIndex];
    }

    const T& Get() const {
      MOZ_ASSERT(!Done());
      return (*mSegment)[mIndex];
    }

    void Next() {
      MOZ_ASSERT(!Done());
      mIndex++;
      if (mIndex == mSegment->Length()) {
        mSegment = mSegment->getNext();
        mIndex = 0;
      }
    }

    void Prev() {
      MOZ_ASSERT(!Done());
      if (mIndex == 0) {
        mSegment = mSegment->getPrevious();
        if (mSegment) {
          mIndex = mSegment->Length() - 1;
        }
      } else {
        --mIndex;
      }
    }
  };

  IterImpl Iter() { return IterImpl(this, true); }
  IterImpl IterFromLast() { return IterImpl(this, false); }

  // Measure the memory consumption of the vector excluding |this|. Note that
  // it only measures the vector itself. If the vector elements contain
  // pointers to other memory blocks, those blocks must be measured separately
  // during a subsequent iteration over the vector.
  size_t SizeOfExcludingThis(mozilla::MallocSizeOf aMallocSizeOf) const {
    return mSegments.sizeOfExcludingThis(aMallocSizeOf);
  }

  // Like sizeOfExcludingThis(), but measures |this| as well.
  size_t SizeOfIncludingThis(mozilla::MallocSizeOf aMallocSizeOf) const {
    return aMallocSizeOf(this) + SizeOfExcludingThis(aMallocSizeOf);
  }

 private:
  mozilla::LinkedList<Segment> mSegments;
};

}  // namespace mozilla

#endif /* mozilla_SegmentedVector_h */