[clang] Introduce paged vector (PR #66430)

Vassil Vassilev via cfe-commits cfe-commits at lists.llvm.org
Tue Sep 26 06:02:35 PDT 2023


================
@@ -0,0 +1,323 @@
+//===- llvm/ADT/PagedVector.h - 'Lazyly allocated' vectors --------*- C++
+//-*-===//
+//
+// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
+// See https://llvm.org/LICENSE.txt for license information.
+// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+//
+//===----------------------------------------------------------------------===//
+//
+// This file defines the PagedVector class.
+//
+//===----------------------------------------------------------------------===//
+#ifndef LLVM_ADT_PAGEDVECTOR_H
+#define LLVM_ADT_PAGEDVECTOR_H
+
+#include "llvm/ADT/PointerIntPair.h"
+#include "llvm/ADT/SmallVector.h"
+#include "llvm/ADT/iterator_range.h"
+#include "llvm/Support/Allocator.h"
+#include <cassert>
+#include <vector>
+
+namespace llvm {
+/// A vector that allocates memory in pages.
+///
+/// Order is kept, but memory is allocated only when one element of the page is
+/// accessed. This introduces a level of indirection, but it is useful when you
+/// have a sparsely initialised vector where the full size is allocated upfront.
+///
+/// As a side effect the elements are initialised later than in a normal vector.
+/// On the first access to one of the elements of a given page all, the elements
+/// of the page are initialised. This also means that the elements of the page
+/// are initialised beyond the size of the vector.
+///
+/// Similarly on destruction the elements are destroyed only when the page is
+/// not needed anymore, delaying invoking the destructor of the elements.
+///
+/// Notice that this does not have iterators, because if you have iterators it
+/// probably means you are going to touch all the memory in any case, so better
+/// use a std::vector in the first place.
+template <typename T, size_t PageSize = 1024 / sizeof(T)> class PagedVector {
+  static_assert(PageSize > 1, "PageSize must be greater than 0. Most likely "
+                              "you want it to be greater than 16.");
+  // The actual number of element in the vector which can be accessed.
+  size_t Size = 0;
+
+  // The position of the initial element of the page in the Data vector.
+  // Pages are allocated contiguously in the Data vector.
+  mutable SmallVector<T *, 0> PageToDataPtrs;
+  // Actual page data. All the page elements are added to this vector on the
+  // first access of any of the elements of the page. Elements default
+  // constructed and elements of the page are stored contiguously. The order of
+  // the elements however depends on the order of access of the pages.
+  PointerIntPair<BumpPtrAllocator *, 1, bool> Allocator;
+
+  constexpr static T *InvalidPage = nullptr;
+
+public:
+  using value_type = T;
+
+  /// Default constructor. We build our own allocator and mark it as such with
+  /// `true` in the second pair element.
+  PagedVector() : Allocator(new BumpPtrAllocator, true) {}
+  PagedVector(BumpPtrAllocator *A) : Allocator(A, false) {
+    assert(A != nullptr && "Allocator cannot be null");
+  }
+
+  ~PagedVector() {
+    clear();
+    // If we own the allocator, delete it.
+    if (Allocator.getInt())
+      delete Allocator.getPointer();
+  }
+
+  /// Look up an element at position `Index`.
+  /// If the associated page is not filled, it will be filled with default
+  /// constructed elements. If the associated page is filled, return the
+  /// element.
+  T &operator[](size_t Index) const {
+    assert(Index < Size);
+    assert(Index / PageSize < PageToDataPtrs.size());
+    T *&PagePtr = PageToDataPtrs[Index / PageSize];
+    // If the page was not yet allocated, allocate it.
+    if (PagePtr == InvalidPage) {
+      T *NewPagePtr = Allocator.getPointer()->template Allocate<T>(PageSize);
+      // We need to invoke the default constructor on all the elements of the
+      // page.
+      for (size_t I = 0; I < PageSize; ++I)
+        new (NewPagePtr + I) T();
+
+      PagePtr = NewPagePtr;
+    }
+    // Dereference the element in the page.
+    return *((Index % PageSize) + PagePtr);
+  }
+
+  /// Return the capacity of the vector. I.e. the maximum size it can be
+  /// expanded to with the resize method without allocating more pages.
+  [[nodiscard]] size_t capacity() const {
+    return PageToDataPtrs.size() * PageSize;
+  }
+
+  /// Return the size of the vector. I.e. the maximum index that can be
+  /// accessed, i.e. the maximum value which was used as argument of the
+  /// resize method.
+  [[nodiscard]] size_t size() const { return Size; }
+
+  /// Resize the vector. Notice that the constructor of the elements will not
+  /// be invoked until an element of a given page is accessed, at which point
+  /// all the elements of the page will be constructed.
+  ///
+  /// If the new size is smaller than the current size, the elements of the
+  /// pages that are not needed anymore will be destroyed, however, elements of
+  /// the last page will not be destroyed.
+  ///
+  /// For these reason the usage of this vector is discouraged if you rely
+  /// on the construction / destructor of the elements to be invoked.
+  void resize(size_t NewSize) {
+    if (NewSize == 0) {
+      clear();
+      return;
+    }
+    // Handle shrink case: destroy the elements in the pages that are not
----------------
vgvassilev wrote:

IIRC @dwblaikie raised this point. @dwblaikie, would you be okay if we assert on and make no-op the resize operation if the size was already set? I am fine keeping it as it is but I feel the current state is not very coherent from the perspective of what we intend to use this data structure for...

https://github.com/llvm/llvm-project/pull/66430


More information about the cfe-commits mailing list