Revert of "[heap] Switch to 500k pages" (patchset v8#11 id:220001 of h…

…ttps://codereview.chromium.org/2232653003/ )

Reason for revert:
Breaks benchmark with --turbo on avx2

https://build.chromium.org/p/client.v8/builders/V8%20Linux64%20-%20avx2/builds/9895

Original issue's description:
> Reland of "[heap] Switch to 500k pages"
>
> Decrease regular heap object size to 400k. In a follow up, we can now get rid of
> the new space border page while keeping the 1M minimum new space size.
>
> BUG=chromium:636331
>
> This reverts commit 555c961.
>
> Committed: https://crrev.com/20e2ea80e169e85c5b8231adc02901fb6c989609
> Cr-Commit-Position: refs/heads/master@{#38608}

[email protected],[email protected]
# Skipping CQ checks because original CL landed less than 1 days ago.
NOPRESUBMIT=true
NOTREECHECKS=true
NOTRY=true
BUG=chromium:636331

Review-Url: https://codereview.chromium.org/2239323002
Cr-Commit-Position: refs/heads/master@{#38613}

src/base/build_config.h

@@ -200,7 +200,7 @@
 // Bump up for Power Linux due to larger (64K) page size.
 const int kPageSizeBits = 22;
 #else
-const int kPageSizeBits = 19;
+const int kPageSizeBits = 20;
 #endif
 
 #endif  // V8_BASE_BUILD_CONFIG_H_

src/heap/heap.cc

@@ -77,7 +77,7 @@ Heap::Heap()
       // semispace_size_ should be a power of 2 and old_generation_size_ should
       // be a multiple of Page::kPageSize.
       max_semi_space_size_(8 * (kPointerSize / 4) * MB),
-      initial_semispace_size_(MB),
+      initial_semispace_size_(Page::kPageSize),
       max_old_generation_size_(700ul * (kPointerSize / 4) * MB),
       initial_old_generation_size_(max_old_generation_size_ /
                                    kInitalOldGenerationLimitFactor),
@@ -5429,19 +5429,16 @@ void Heap::PrintAlloctionsHash() {
 
 
 void Heap::NotifyDeserializationComplete() {
-  DCHECK_EQ(0, gc_count());
+  deserialization_complete_ = true;
+#ifdef DEBUG
+  // All pages right after bootstrapping must be marked as never-evacuate.
   PagedSpaces spaces(this);
   for (PagedSpace* s = spaces.next(); s != NULL; s = spaces.next()) {
-    if (isolate()->snapshot_available()) s->ShrinkImmortalImmovablePages();
-#ifdef DEBUG
-    // All pages right after bootstrapping must be marked as never-evacuate.
     for (Page* p : *s) {
       CHECK(p->NeverEvacuate());
     }
-#endif  // DEBUG
   }
-
-  deserialization_complete_ = true;
+#endif  // DEBUG
 }
 
 void Heap::SetEmbedderHeapTracer(EmbedderHeapTracer* tracer) {

src/heap/mark-compact.cc

@@ -600,18 +600,18 @@ void MarkCompactCollector::ComputeEvacuationHeuristics(
   // For memory reducing and optimize for memory mode we directly define both
   // constants.
   const int kTargetFragmentationPercentForReduceMemory = 20;
-  const int kMaxEvacuatedBytesForReduceMemory = 12 * MB;
+  const int kMaxEvacuatedBytesForReduceMemory = 12 * Page::kPageSize;
   const int kTargetFragmentationPercentForOptimizeMemory = 20;
   const int kMaxEvacuatedBytesForOptimizeMemory = 6 * MB;
 
   // For regular mode (which is latency critical) we define less aggressive
   // defaults to start and switch to a trace-based (using compaction speed)
   // approach as soon as we have enough samples.
   const int kTargetFragmentationPercent = 70;
-  const int kMaxEvacuatedBytes = 4 * MB;
+  const int kMaxEvacuatedBytes = 4 * Page::kPageSize;
   // Time to take for a single area (=payload of page). Used as soon as there
   // exist enough compaction speed samples.
-  const float kTargetMsPerArea = 0.5;
+  const int kTargetMsPerArea = 1;
 
   if (heap()->ShouldReduceMemory()) {
     *target_fragmentation_percent = kTargetFragmentationPercentForReduceMemory;
@@ -3223,7 +3223,7 @@ int MarkCompactCollector::NumberOfParallelCompactionTasks(int pages,
   // The number of parallel compaction tasks is limited by:
   // - #evacuation pages
   // - (#cores - 1)
-  const double kTargetCompactionTimeInMs = .5;
+  const double kTargetCompactionTimeInMs = 1;
   const int kNumSweepingTasks = 3;
 
   double compaction_speed =

src/heap/spaces.cc

@@ -596,21 +596,6 @@ void MemoryChunk::Unlink() {
   set_next_chunk(NULL);
 }
 
-void MemoryAllocator::ShrinkChunk(MemoryChunk* chunk, size_t bytes_to_shrink) {
-  DCHECK_GE(bytes_to_shrink, static_cast<size_t>(base::OS::CommitPageSize()));
-  DCHECK_EQ(0, bytes_to_shrink % base::OS::CommitPageSize());
-  Address free_start = chunk->area_end_ - bytes_to_shrink;
-  // Don't adjust the size of the page. The area is just uncomitted but not
-  // released.
-  chunk->area_end_ -= bytes_to_shrink;
-  UncommitBlock(free_start, bytes_to_shrink);
-  if (chunk->IsFlagSet(MemoryChunk::IS_EXECUTABLE)) {
-    if (chunk->reservation_.IsReserved())
-      chunk->reservation_.Guard(chunk->area_end_);
-    else
-      base::OS::Guard(chunk->area_end_, base::OS::CommitPageSize());
-  }
-}
 
 MemoryChunk* MemoryAllocator::AllocateChunk(intptr_t reserve_area_size,
                                             intptr_t commit_area_size,
@@ -758,47 +743,6 @@ void Page::ResetFreeListStatistics() {
   available_in_free_list_ = 0;
 }
 
-size_t Page::ShrinkToHighWaterMark() {
-  // Shrink pages to high water mark. The water mark points either to a filler
-  // or the area_end.
-  HeapObject* filler = HeapObject::FromAddress(HighWaterMark());
-  if (filler->address() == area_end()) return 0;
-  CHECK(filler->IsFiller());
-  if (!filler->IsFreeSpace()) return 0;
-
-#ifdef DEBUG
-  // Check the the filler is indeed the last filler on the page.
-  HeapObjectIterator it(this);
-  HeapObject* filler2 = nullptr;
-  for (HeapObject* obj = it.Next(); obj != nullptr; obj = it.Next()) {
-    filler2 = HeapObject::FromAddress(obj->address() + obj->Size());
-  }
-  if (filler2 == nullptr || filler2->address() == area_end()) return 0;
-  DCHECK(filler2->IsFiller());
-  DCHECK_EQ(filler->address(), filler2->address());
-#endif  // DEBUG
-
-  size_t unused = RoundDown(
-      static_cast<size_t>(area_end() - filler->address() - FreeSpace::kSize),
-      base::OS::CommitPageSize());
-  if (unused > 0) {
-    if (FLAG_trace_gc_verbose) {
-      PrintIsolate(heap()->isolate(), "Shrinking page %p: end %p -> %p\n",
-                   reinterpret_cast<void*>(this),
-                   reinterpret_cast<void*>(area_end()),
-                   reinterpret_cast<void*>(area_end() - unused));
-    }
-    heap()->CreateFillerObjectAt(
-        filler->address(),
-        static_cast<int>(area_end() - filler->address() - unused),
-        ClearRecordedSlots::kNo);
-    heap()->memory_allocator()->ShrinkChunk(this, unused);
-    CHECK(filler->IsFiller());
-    CHECK_EQ(filler->address() + filler->Size(), area_end());
-  }
-  return unused;
-}
-
 void MemoryAllocator::PartialFreeMemory(MemoryChunk* chunk,
                                         Address start_free) {
   // We do not allow partial shrink for code.
@@ -1270,25 +1214,17 @@ Object* PagedSpace::FindObject(Address addr) {
   return Smi::FromInt(0);
 }
 
-void PagedSpace::ShrinkImmortalImmovablePages() {
-  DCHECK(!heap()->deserialization_complete());
-  MemoryChunk::UpdateHighWaterMark(allocation_info_.top());
-  EmptyAllocationInfo();
-  ResetFreeList();
-
-  for (Page* page : *this) {
-    DCHECK(page->IsFlagSet(Page::NEVER_EVACUATE));
-    size_t unused = page->ShrinkToHighWaterMark();
-    accounting_stats_.DecreaseCapacity(static_cast<intptr_t>(unused));
-    AccountUncommitted(unused);
+bool PagedSpace::Expand() {
+  int size = AreaSize();
+  if (snapshotable() && !HasPages()) {
+    size = Snapshot::SizeOfFirstPage(heap()->isolate(), identity());
   }
-}
 
-bool PagedSpace::Expand() {
-  const int size = AreaSize();
   if (!heap()->CanExpandOldGeneration(size)) return false;
+
   Page* p = heap()->memory_allocator()->AllocatePage(size, this, executable());
   if (p == nullptr) return false;
+
   AccountCommitted(static_cast<intptr_t>(p->size()));
 
   // Pages created during bootstrapping may contain immortal immovable objects.
@@ -1379,6 +1315,7 @@ void PagedSpace::IncreaseCapacity(int size) {
 
 void PagedSpace::ReleasePage(Page* page) {
   DCHECK_EQ(page->LiveBytes(), 0);
+  DCHECK_EQ(AreaSize(), page->area_size());
   DCHECK_EQ(page->owner(), this);
 
   free_list_.EvictFreeListItems(page);
@@ -1397,8 +1334,10 @@ void PagedSpace::ReleasePage(Page* page) {
   }
 
   AccountUncommitted(static_cast<intptr_t>(page->size()));
-  accounting_stats_.ShrinkSpace(page->area_size());
   heap()->memory_allocator()->Free<MemoryAllocator::kPreFreeAndQueue>(page);
+
+  DCHECK(Capacity() > 0);
+  accounting_stats_.ShrinkSpace(AreaSize());
 }
 
 #ifdef DEBUG

src/heap/spaces.h

@@ -235,10 +235,7 @@ class MemoryChunk {
     IN_TO_SPACE,    // All pages in new space has one of these two set.
     NEW_SPACE_BELOW_AGE_MARK,
     EVACUATION_CANDIDATE,
-
-    // |NEVER_EVACUATE|: A page tagged with this flag will never be selected
-    // for evacuation. Typically used for immortal immovable pages.
-    NEVER_EVACUATE,
+    NEVER_EVACUATE,  // May contain immortal immutables.
 
     // Large objects can have a progress bar in their page header. These object
     // are scanned in increments and will be kept black while being scanned.
@@ -725,7 +722,7 @@ class Page : public MemoryChunk {
   // account.
   // TODO(hpayer): This limit should be way smaller but we currently have
   // short living objects >256K.
-  static const int kMaxRegularHeapObjectSize = 400 * KB;
+  static const int kMaxRegularHeapObjectSize = 600 * KB;
 
   static inline Page* ConvertNewToOld(Page* old_page, PagedSpace* new_owner);
 
@@ -826,8 +823,6 @@ class Page : public MemoryChunk {
     available_in_free_list_.Increment(available);
   }
 
-  size_t ShrinkToHighWaterMark();
-
 #ifdef DEBUG
   void Print();
 #endif  // DEBUG
@@ -1308,8 +1303,6 @@ class MemoryAllocator {
                              intptr_t commit_area_size,
                              Executability executable, Space* space);
 
-  void ShrinkChunk(MemoryChunk* chunk, size_t bytes_to_shrink);
-
   Address ReserveAlignedMemory(size_t requested, size_t alignment,
                                base::VirtualMemory* controller);
   Address AllocateAlignedMemory(size_t reserve_size, size_t commit_size,
@@ -1625,7 +1618,6 @@ class AllocationStats BASE_EMBEDDED {
     capacity_ -= size_in_bytes;
     size_ -= size_in_bytes;
     CHECK_GE(size_, 0);
-    CHECK_GE(capacity_, 0);
   }
 
   // Allocate from available bytes (available -> size).
@@ -2191,10 +2183,6 @@ class PagedSpace : public Space {
   iterator begin() { return iterator(anchor_.next_page()); }
   iterator end() { return iterator(&anchor_); }
 
-  // Shrink immortal immovable pages of the space to be exactly the size needed
-  // using the high water mark.
-  void ShrinkImmortalImmovablePages();
-
  protected:
   // PagedSpaces that should be included in snapshots have different, i.e.,
   // smaller, initial pages.

src/isolate.cc

@@ -2438,19 +2438,13 @@ bool Isolate::Init(Deserializer* des) {
   runtime_profiler_ = new RuntimeProfiler(this);
 
   // If we are deserializing, read the state into the now-empty heap.
-  {
-    AlwaysAllocateScope always_allocate(this);
-
-    if (!create_heap_objects) {
-      des->Deserialize(this);
-    }
-    load_stub_cache_->Initialize();
-    store_stub_cache_->Initialize();
-    if (FLAG_ignition || serializer_enabled()) {
-      interpreter_->Initialize();
-    }
-
-    heap_.NotifyDeserializationComplete();
+  if (!create_heap_objects) {
+    des->Deserialize(this);
+  }
+  load_stub_cache_->Initialize();
+  store_stub_cache_->Initialize();
+  if (FLAG_ignition || serializer_enabled()) {
+    interpreter_->Initialize();
   }
 
   // Finish initialization of ThreadLocal after deserialization is done.
@@ -2481,6 +2475,8 @@ bool Isolate::Init(Deserializer* des) {
 
   time_millis_at_init_ = heap_.MonotonicallyIncreasingTimeInMs();
 
+  heap_.NotifyDeserializationComplete();
+
   if (!create_heap_objects) {
     // Now that the heap is consistent, it's OK to generate the code for the
     // deopt entry table that might have been referred to by optimized code in

src/objects.h

@@ -4712,7 +4712,6 @@ class FreeSpace: public HeapObject {
   // Size is smi tagged when it is stored.
   static const int kSizeOffset = HeapObject::kHeaderSize;
   static const int kNextOffset = POINTER_SIZE_ALIGN(kSizeOffset + kPointerSize);
-  static const int kSize = kNextOffset + kPointerSize;
 
  private:
   DISALLOW_IMPLICIT_CONSTRUCTORS(FreeSpace);
@@ -10468,12 +10467,12 @@ class JSArray: public JSObject {
   static const int kLengthOffset = JSObject::kHeaderSize;
   static const int kSize = kLengthOffset + kPointerSize;
 
-  // 400 * KB is the Page::kMaxRegularHeapObjectSize defined in spaces.h which
+  // 600 * KB is the Page::kMaxRegularHeapObjectSize defined in spaces.h which
   // we do not want to include in objects.h
   // Note that Page::kMaxRegularHeapObjectSize has to be in sync with
   // kInitialMaxFastElementArray which is checked in a DCHECK in heap.cc.
   static const int kInitialMaxFastElementArray =
-      (400 * KB - FixedArray::kHeaderSize - kSize - AllocationMemento::kSize) /
+      (600 * KB - FixedArray::kHeaderSize - kSize - AllocationMemento::kSize) /
       kPointerSize;
 
  private: