[Tracing] Add allocation register for heap profiling

base/trace_event/memory_profiler_allocation_register.cc

Index: base/trace_event/memory_profiler_allocation_register.cc
diff --git a/base/trace_event/memory_profiler_allocation_register.cc b/base/trace_event/memory_profiler_allocation_register.cc
new file mode 100644
index 0000000000000000000000000000000000000000..49043af5e820420c318df5178ee8d8289dcac195
--- /dev/null
+++ b/base/trace_event/memory_profiler_allocation_register.cc
@@ -0,0 +1,160 @@
+// Copyright 2015 The Chromium Authors. All rights reserved.
+// Use of this source code is governed by a BSD-style license that can be
+// found in the LICENSE file.
+
+#include "base/trace_event/memory_profiler_allocation_register.h"
+
+namespace base {
+namespace trace_event {
+
+AllocationRegister::AllocationRegister()
+    // Reserve enough address space to store |kNumCells| entries if necessary,
+    // with a guard page after it to crash the program when attempting to store
+    // more entries.
+    : cells_(static_cast<Cell*>(AllocateVirtualMemory(kNumCells *
+                                                      sizeof(Cell)))),
+      buckets_(static_cast<CellIndex*>(
+          AllocateVirtualMemory(kNumBuckets * sizeof(CellIndex)))) {
+  // The free list is empty. The first unused cell is cell 1, because index 0 is
+  // used as list terminator.
+  free_list_ = 0;

[Primiano Tucci (use gerrit), 2015/10/06 13:34:44]

why don't you inizialize also these in the ctor?
I think that the clang-checker will complain that free_list_ is not initialized
(not sure if the plugin is smart enough to look in the ctor body)

[Ruud van Asseldonk, 2015/10/06 15:44:34]

Done. The checker did not complain, |free_list_| was initialized perfectly well
in the constructor and all members were initialized in declaration order.

+  next_unused_cell_ = 1;
+}
+
+AllocationRegister::~AllocationRegister() {
+  FreeVirtualMemory(buckets_, kNumBuckets * sizeof(CellIndex));
+  FreeVirtualMemory(cells_, kNumCells * sizeof(Cell));
+}
+
+void AllocationRegister::Insert(void* address,
+                                size_t size,
+                                AllocationContext context) {
+  DCHECK(address != nullptr);
+
+  CellIndex* idx_ptr = Lookup(address);
+
+  // If the index is 0, the address is not yet present, so insert it.
+  if (*idx_ptr == 0) {
+    *idx_ptr = GetFreeCell();
+
+    cells_[*idx_ptr].allocation.address = address;
+    cells_[*idx_ptr].next = 0;
+  }
+
+  cells_[*idx_ptr].allocation.size = size;
+  cells_[*idx_ptr].allocation.context = context;
+}
+
+void AllocationRegister::Remove(void* address) {
+  CellIndex* idx_ptr = Lookup(address);

[Primiano Tucci (use gerrit), 2015/10/06 13:34:44]

I'd probably move the comment that says that idx_ptr is the pointer to either
the bucket or the next member here, as this is the place where I would have
really needed that comment

[Ruud van Asseldonk, 2015/10/06 15:44:35]

Done, though it is not actually relevant where the index is stored.

+
+  // If the index is 0, the address was not there in the first place.
+  if (*idx_ptr == 0)
+    return;
+
+  // The cell at the index is now free, remove it from the linked list for
+  // |Hash(address)|.
+  CellIndex free_idx = *idx_ptr;

[Primiano Tucci (use gerrit), 2015/10/06 13:34:44]

move this up before the if, and use if(free_idx == 0) in the if

[Ruud van Asseldonk, 2015/10/06 15:44:34]

As you wish.

+  *idx_ptr = cells_[free_idx].next;

[Primiano Tucci (use gerrit), 2015/10/06 13:34:44]

I this this code below would be more readable if you add an extra variable:

Cell* freed_cell = &cells_[free_idx];
and use freed_cell below

Also note freed vs free, IMHO that conveys more the information about "this is
the cell/index being freed" and avoids the ambiguity of "was this free already
before the remove or happened to become free after the the remove?"

[Ruud van Asseldonk, 2015/10/06 15:44:34]

Done. Yeah freeing is not atomic so I think it is impossible to be perfectly
accurate with the names.

+
+  // Put the free cell at the front of the free list.
+  cells_[free_idx].next = free_list_;
+  free_list_ = free_idx;
+
+  // Reset the address, so that on iteration the free cell is ignored.
+  cells_[free_idx].allocation.address = nullptr;
+}
+
+AllocationRegister::ConstIterator AllocationRegister::begin() const {
+  // Initialize the iterator's index to 0. Cell 0 never stores an entry.
+  ConstIterator iterator(*this, 0);
+  // Incrementing will advance the iterator to the first used cell.
+  ++iterator;

[Primiano Tucci (use gerrit), 2015/10/06 13:34:44]

Why don't you just construct the iterator passing 1 above?
In essence, why this is not just a return ConstIterator(*this, 1) ?
in which case could be inlined in the header as becomes trivial

[Ruud van Asseldonk, 2015/10/06 15:44:35]

Because cell 1 might be free.

+  return iterator;
+}
+
+AllocationRegister::ConstIterator AllocationRegister::end() const {
+  // Cell |next_unused_cell_ - 1| is the last cell that could contain an entry,
+  // so index |next_unused_cell_| is an iterator past the last element, in line
+  // with the STL iterator conventions.
+  return ConstIterator(*this, next_unused_cell_);
+}
+
+AllocationRegister::ConstIterator::ConstIterator(
+    const AllocationRegister& alloc_register,
+    CellIndex index)
+    : register_(alloc_register), index_(index) {}
+
+void AllocationRegister::ConstIterator::operator++() {
+  // Find the next cell with a non-null address until all cells that could
+  // possibly be used have been iterated. A null address indicates a free cell.
+  do

[Primiano Tucci (use gerrit), 2015/10/06 13:34:44]

I think you still need {braces} as this is a multiline while (i.e. a similar
for() loop would have been similarly {braced})

[Ruud van Asseldonk, 2015/10/06 15:44:35]

As you wish.

+    index_++;
+  while (index_ < register_.next_unused_cell_ &&
+         register_.cells_[index_].allocation.address == nullptr);
+}
+
+bool AllocationRegister::ConstIterator::operator!=(
+    const ConstIterator& other) const {
+  return index_ != other.index_;
+}
+
+const AllocationRegister::Allocation& AllocationRegister::ConstIterator::
+operator*() const {
+  return register_.cells_[index_].allocation;
+}
+
+uint32_t AllocationRegister::Hash(void* address) {
+  // The multiplicative hashing scheme from [Knuth 1998]. The value of |a| has
+  // been chosen carefully based on measurements with real-word data (addresses
+  // recorded from a Chrome trace run). It is the first prime after 2^17. For

[Primiano Tucci (use gerrit), 2015/10/06 13:34:44]

This is worth the entire CL

+  // |shift|, 13, 14 and 15 yield good results. These values are tuned to 2^18
+  // buckets. Microbenchmarks show that this simple scheme outperforms fancy
+  // hashes like Murmur3 by 20 to 40 percent.
+  const uintptr_t key = reinterpret_cast<uintptr_t>(address);
+  const uintptr_t a = 131101;
+  const uintptr_t shift = 14;
+  const uintptr_t h = (key * a) >> shift;
+  return static_cast<uint32_t>(h) & kNumBucketsMask;
+}
+
+AllocationRegister::CellIndex* AllocationRegister::Lookup(void* address) {
+  // The list head is in |buckets_| at the hash offset.
+  CellIndex* idx_ptr = buckets_ + Hash(address);

[Primiano Tucci (use gerrit), 2015/10/06 13:34:44]

IMHO probably slightly more readable if you:
CellIndex* idx_ptr = &buckets_[Hash(address)]

[Ruud van Asseldonk, 2015/10/06 15:44:35]

As you wish. Fun fact: as |&buckets_[Hash(address)]| desugars to |buckets_ +
Hash(address)| and + is commutative, you can also write
|&Hash(address)[buckets_]|.

+
+  // Chase down the list until the cell that holds |key| is found,
+  // or until the list ends.
+  while (*idx_ptr != 0 && cells_[*idx_ptr].allocation.address != address)
+    idx_ptr = &cells_[*idx_ptr].next;
+
+  return idx_ptr;
+}
+
+AllocationRegister::CellIndex AllocationRegister::GetFreeCell() {
+  // First try to re-use a cell from the freelist.
+  if (free_list_) {
+    CellIndex idx = free_list_;
+    free_list_ = cells_[idx].next;
+    return idx;
+  }
+
+  // Otherwise pick the next cell that has not been touched before.
+  CellIndex idx = next_unused_cell_;
+  next_unused_cell_++;
+
+  // If the hash table has too little capacity (when too little address space
+  // was reserved for |cells_|), |next_unused_cell_| can be an index outside of
+  // the allocated storage. A guard page is allocated there to crash the
+  // program in that case. There are alternative solutions:
+  // - Deal with it, increase capacity by reallocating |cells_|.
+  // - Refuse to insert and let the caller deal with it.
+  // Because free cells are re-used before accessing fresh cells with a higher
+  // index, and because reserving address space without touching it is cheap,
+  // the simplest solution is to just allocate a humongous chunk of address
+  // space.
+
+  return idx;

[Primiano Tucci (use gerrit), 2015/10/06 13:34:44]

I'd probably add a DCHECK here, so in debug builds we can figure out the issue
without debugging a more cryptic SEGFAULT later

[Ruud van Asseldonk, 2015/10/06 15:44:35]

Done.

+}
+
+}  // namespace trace_event
+}  // namespace base