[llvm] Introduce paged vector (PR #66430)

[Jakub Kuderski via llvm-commits]

================
@@ -0,0 +1,322 @@
+//===- 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/Support/Allocator.h"
+#include <cassert>
+#include <iostream>
+#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
+// with the default constructor and elements are initialised later, on first
+// access.
+//
+// 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.
+//
+// Pages are allocated in SLAB_SIZE chunks, using the BumpPtrAllocator.
+template <typename T, std::size_t PAGE_SIZE = 1024 / sizeof(T)>
+class PagedVector {
+  static_assert(PAGE_SIZE > 0, "PAGE_SIZE 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.
+  std::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 std::vector<uintptr_t> PageToDataIdx;
----------------
kuhar wrote:

I wanted to follow up on the first few points from this comment: https://github.com/llvm/llvm-project/pull/66430#issuecomment-1721685355

Have you checked if using `SmallVector` gives a measurable difference? We don't have to micro-optimize this implementation today, but it would be fun to have a benchmark that we could use to iterate on the implementation in the future. 

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