[heap] Reland concurrent store buffer processing.

BUG=chromium:648973, chromium:648568

Review-Url: https://codereview.chromium.org/2493083003
Cr-Commit-Position: refs/heads/master@{#40928}

src/heap/heap.cc

@@ -439,7 +439,7 @@ void Heap::GarbageCollectionPrologue() {
   }
   CheckNewSpaceExpansionCriteria();
   UpdateNewSpaceAllocationCounter();
-  store_buffer()->MoveEntriesToRememberedSet();
+  store_buffer()->MoveAllEntriesToRememberedSet();
 }
 
 
@@ -5874,10 +5874,10 @@ void Heap::CheckHandleCount() {
 
 void Heap::ClearRecordedSlot(HeapObject* object, Object** slot) {
   if (!InNewSpace(object)) {
-    store_buffer()->MoveEntriesToRememberedSet();
     Address slot_addr = reinterpret_cast<Address>(slot);
     Page* page = Page::FromAddress(slot_addr);
     DCHECK_EQ(page->owner()->identity(), OLD_SPACE);
+    store_buffer()->MoveAllEntriesToRememberedSet();
     RememberedSet<OLD_TO_NEW>::Remove(page, slot_addr);
     RememberedSet<OLD_TO_OLD>::Remove(page, slot_addr);
   }
@@ -5886,8 +5886,8 @@ void Heap::ClearRecordedSlot(HeapObject* object, Object** slot) {
 void Heap::ClearRecordedSlotRange(Address start, Address end) {
   Page* page = Page::FromAddress(start);
   if (!page->InNewSpace()) {
-    store_buffer()->MoveEntriesToRememberedSet();
     DCHECK_EQ(page->owner()->identity(), OLD_SPACE);
+    store_buffer()->MoveAllEntriesToRememberedSet();
     RememberedSet<OLD_TO_NEW>::RemoveRange(page, start, end,
                                            SlotSet::PREFREE_EMPTY_BUCKETS);
     RememberedSet<OLD_TO_OLD>::RemoveRange(page, start, end,

src/heap/store-buffer.cc

@@ -16,62 +16,135 @@ namespace v8 {
 namespace internal {
 
 StoreBuffer::StoreBuffer(Heap* heap)
-    : heap_(heap),
-      top_(nullptr),
-      start_(nullptr),
-      limit_(nullptr),
-      virtual_memory_(nullptr) {}
+    : heap_(heap), top_(nullptr), current_(0), virtual_memory_(nullptr) {
+  for (int i = 0; i < kStoreBuffers; i++) {
+    start_[i] = nullptr;
+    limit_[i] = nullptr;
+    lazy_top_[i] = nullptr;
+  }
+  task_running_ = false;
+}
 
 void StoreBuffer::SetUp() {
   // Allocate 3x the buffer size, so that we can start the new store buffer
   // aligned to 2x the size.  This lets us use a bit test to detect the end of
   // the area.
-  virtual_memory_ = new base::VirtualMemory(kStoreBufferSize * 2);
+  virtual_memory_ = new base::VirtualMemory(kStoreBufferSize * 3);
   uintptr_t start_as_int =
       reinterpret_cast<uintptr_t>(virtual_memory_->address());
-  start_ = reinterpret_cast<Address*>(RoundUp(start_as_int, kStoreBufferSize));
-  limit_ = start_ + (kStoreBufferSize / kPointerSize);
+  start_[0] =
+      reinterpret_cast<Address*>(RoundUp(start_as_int, kStoreBufferSize));
+  limit_[0] = start_[0] + (kStoreBufferSize / kPointerSize);
+  start_[1] = limit_[0];
+  limit_[1] = start_[1] + (kStoreBufferSize / kPointerSize);
 
-  DCHECK(reinterpret_cast<Address>(start_) >= virtual_memory_->address());
-  DCHECK(reinterpret_cast<Address>(limit_) >= virtual_memory_->address());
   Address* vm_limit = reinterpret_cast<Address*>(
       reinterpret_cast<char*>(virtual_memory_->address()) +
       virtual_memory_->size());
-  DCHECK(start_ <= vm_limit);
-  DCHECK(limit_ <= vm_limit);
+
   USE(vm_limit);
-  DCHECK((reinterpret_cast<uintptr_t>(limit_) & kStoreBufferMask) == 0);
+  for (int i = 0; i < kStoreBuffers; i++) {
+    DCHECK(reinterpret_cast<Address>(start_[i]) >= virtual_memory_->address());
+    DCHECK(reinterpret_cast<Address>(limit_[i]) >= virtual_memory_->address());
+    DCHECK(start_[i] <= vm_limit);
+    DCHECK(limit_[i] <= vm_limit);
+    DCHECK((reinterpret_cast<uintptr_t>(limit_[i]) & kStoreBufferMask) == 0);
+  }
 
-  if (!virtual_memory_->Commit(reinterpret_cast<Address>(start_),
-                               kStoreBufferSize,
+  if (!virtual_memory_->Commit(reinterpret_cast<Address>(start_[0]),
+                               kStoreBufferSize * kStoreBuffers,
                                false)) {  // Not executable.
     V8::FatalProcessOutOfMemory("StoreBuffer::SetUp");
   }
-  top_ = start_;
+  current_ = 0;
+  top_ = start_[current_];
 }
 
 
 void StoreBuffer::TearDown() {
   delete virtual_memory_;
-  top_ = start_ = limit_ = nullptr;
+  top_ = nullptr;
+  for (int i = 0; i < kStoreBuffers; i++) {
+    start_[i] = nullptr;
+    limit_[i] = nullptr;
+    lazy_top_[i] = nullptr;
+  }
 }
 
 
 void StoreBuffer::StoreBufferOverflow(Isolate* isolate) {
-  isolate->heap()->store_buffer()->MoveEntriesToRememberedSet();
+  isolate->heap()->store_buffer()->FlipStoreBuffers();
   isolate->counters()->store_buffer_overflows()->Increment();
 }
 
-void StoreBuffer::MoveEntriesToRememberedSet() {
-  if (top_ == start_) return;
-  DCHECK(top_ <= limit_);
-  for (Address* current = start_; current < top_; current++) {
+void StoreBuffer::FlipStoreBuffers() {
+  base::LockGuard<base::Mutex> guard(&mutex_);
+  int other = (current_ + 1) % kStoreBuffers;
+  MoveEntriesToRememberedSet(other);
+  lazy_top_[current_] = top_;
+  current_ = other;
+  top_ = start_[current_];
+
+  if (!task_running_) {
+    task_running_ = true;
+    Task* task = new Task(heap_->isolate(), this);
+    V8::GetCurrentPlatform()->CallOnBackgroundThread(
+        task, v8::Platform::kShortRunningTask);
+  }
+}
+
+void StoreBuffer::MoveEntriesToRememberedSet(int index) {
+  if (!lazy_top_[index]) return;
+  DCHECK_GE(index, 0);
+  DCHECK_LT(index, kStoreBuffers);
+  for (Address* current = start_[index]; current < lazy_top_[index];
+       current++) {
     DCHECK(!heap_->code_space()->Contains(*current));
     Address addr = *current;
     Page* page = Page::FromAnyPointerAddress(heap_, addr);
-    RememberedSet<OLD_TO_NEW>::Insert(page, addr);
+    if (IsDeletionAddress(addr)) {
+      current++;
+      Address end = *current;
+      DCHECK(!IsDeletionAddress(end));
+      addr = UnmarkDeletionAddress(addr);
+      if (end) {
+        RememberedSet<OLD_TO_NEW>::RemoveRange(page, addr, end,
+                                               SlotSet::PREFREE_EMPTY_BUCKETS);
+      } else {
+        RememberedSet<OLD_TO_NEW>::Remove(page, addr);
+      }
+    } else {
+      DCHECK(!IsDeletionAddress(addr));
+      RememberedSet<OLD_TO_NEW>::Insert(page, addr);
+    }
+  }
+  lazy_top_[index] = nullptr;
+}
+
+void StoreBuffer::MoveAllEntriesToRememberedSet() {
+  base::LockGuard<base::Mutex> guard(&mutex_);
+  int other = (current_ + 1) % kStoreBuffers;
+  MoveEntriesToRememberedSet(other);
+  lazy_top_[current_] = top_;
+  MoveEntriesToRememberedSet(current_);
+  top_ = start_[current_];
+}
+
+void StoreBuffer::ConcurrentlyProcessStoreBuffer() {
+  base::LockGuard<base::Mutex> guard(&mutex_);
+  int other = (current_ + 1) % kStoreBuffers;
+  MoveEntriesToRememberedSet(other);
+  task_running_ = false;
+}
+
+void StoreBuffer::DeleteEntry(Address start, Address end) {
+  if (top_ + sizeof(Address) * 2 > limit_[current_]) {
+    StoreBufferOverflow(heap_->isolate());
   }
-  top_ = start_;
+  *top_ = MarkDeletionAddress(start);
+  top_++;
+  *top_ = end;
+  top_++;
 }
 
 }  // namespace internal

src/heap/store-buffer.h

@@ -8,6 +8,7 @@
 #include "src/allocation.h"
 #include "src/base/logging.h"
 #include "src/base/platform/platform.h"
+#include "src/cancelable-task.h"
 #include "src/globals.h"
 #include "src/heap/slot-set.h"
 
@@ -15,11 +16,17 @@ namespace v8 {
 namespace internal {
 
 // Intermediate buffer that accumulates old-to-new stores from the generated
-// code. On buffer overflow the slots are moved to the remembered set.
+// code. Moreover, it stores invalid old-to-new slots with two entries.
+// The first is a tagged address of the start of the invalid range, the second
+// one is the end address of the invalid range or null if there is just one slot
+// that needs to be removed from the remembered set. On buffer overflow the
+// slots are moved to the remembered set.
 class StoreBuffer {
  public:
   static const int kStoreBufferSize = 1 << (14 + kPointerSizeLog2);
   static const int kStoreBufferMask = kStoreBufferSize - 1;
+  static const int kStoreBuffers = 2;
+  static const intptr_t kDeletionTag = 1;
 
   static void StoreBufferOverflow(Isolate* isolate);
 
@@ -30,17 +37,92 @@ class StoreBuffer {
   // Used to add entries from generated code.
   inline Address* top_address() { return reinterpret_cast<Address*>(&top_); }
 
-  void MoveEntriesToRememberedSet();
+  // Moves entries from a specific store buffer to the remembered set. This
+  // method takes a lock.
+  void MoveEntriesToRememberedSet(int index);
+
+  // This method ensures that all used store buffer entries are transfered to
+  // the remembered set.
+  void MoveAllEntriesToRememberedSet();
+
+  inline bool IsDeletionAddress(Address address) const {
+    return reinterpret_cast<intptr_t>(address) & kDeletionTag;
+  }
+
+  inline Address MarkDeletionAddress(Address address) {
+    return reinterpret_cast<Address>(reinterpret_cast<intptr_t>(address) |
+                                     kDeletionTag);
+  }
+
+  inline Address UnmarkDeletionAddress(Address address) {
+    return reinterpret_cast<Address>(reinterpret_cast<intptr_t>(address) &
+                                     ~kDeletionTag);
+  }
+
+  // If we only want to delete a single slot, end should be set to null which
+  // will be written into the second field. When processing the store buffer
+  // the more efficient Remove method will be called in this case.
+  void DeleteEntry(Address start, Address end = nullptr);
+
+  // Used by the concurrent processing thread to transfer entries from the
+  // store buffer to the remembered set.
+  void ConcurrentlyProcessStoreBuffer();
+
+  bool Empty() {
+    for (int i = 0; i < kStoreBuffers; i++) {
+      if (lazy_top_[i]) {
+        return false;
+      }
+    }
+    return top_ == start_[current_];
+  }
 
  private:
+  // There are two store buffers. If one store buffer fills up, the main thread
+  // publishes the top pointer of the store buffer that needs processing in its
+  // global lazy_top_ field. After that it start the concurrent processing
+  // thread. The concurrent processing thread uses the pointer in lazy_top_.
+  // It will grab the given mutex and transfer its entries to the remembered
+  // set. If the concurrent thread does not make progress, the main thread will
+  // perform the work.
+  // Important: there is an ordering constrained. The store buffer with the
+  // older entries has to be processed first.
+  class Task : public CancelableTask {
+   public:
+    Task(Isolate* isolate, StoreBuffer* store_buffer)
+        : CancelableTask(isolate), store_buffer_(store_buffer) {}
+    virtual ~Task() {}
+
+   private:
+    void RunInternal() override {
+      store_buffer_->ConcurrentlyProcessStoreBuffer();
+    }
+    StoreBuffer* store_buffer_;
+    DISALLOW_COPY_AND_ASSIGN(Task);
+  };
+
+  void FlipStoreBuffers();
+
   Heap* heap_;
 
   Address* top_;
 
   // The start and the limit of the buffer that contains store slots
-  // added from the generated code.
-  Address* start_;
-  Address* limit_;
+  // added from the generated code. We have two chunks of store buffers.
+  // Whenever one fills up, we notify a concurrent processing thread and
+  // use the other empty one in the meantime.
+  Address* start_[kStoreBuffers];
+  Address* limit_[kStoreBuffers];
+
+  // At most one lazy_top_ pointer is set at any time.
+  Address* lazy_top_[kStoreBuffers];
+  base::Mutex mutex_;
+
+  // We only want to have at most one concurrent processing tas running.
+  bool task_running_;
+
+  // Points to the current buffer in use.
+  int current_;
 
   base::VirtualMemory* virtual_memory_;
 };