Make the speed of incrmental marking depend also on the rate

src/incremental-marking.cc

 // Copyright 2012 the V8 project authors. All rights reserved.
 // Redistribution and use in source and binary forms, with or without
 // modification, are permitted provided that the following conditions are
 // met:
 //
 //     * Redistributions of source code must retain the above copyright
 //       notice, this list of conditions and the following disclaimer.
 //     * Redistributions in binary form must reproduce the above
 //       copyright notice, this list of conditions and the following
 //       disclaimer in the documentation and/or other materials provided
@@ skipping 34 matching lines @@
       marking_deque_memory_(NULL),
       marking_deque_memory_committed_(false),
       steps_count_(0),
       steps_took_(0),
       longest_step_(0.0),
       old_generation_space_available_at_start_of_incremental_(0),
       old_generation_space_used_at_start_of_incremental_(0),
       steps_count_since_last_gc_(0),
       steps_took_since_last_gc_(0),
       should_hurry_(false),
-      allocation_marking_factor_(0),
+      marking_speed_(0),
       allocated_(0),
       no_marking_scope_depth_(0) {
 }
 
 
 void IncrementalMarking::TearDown() {
   delete marking_deque_memory_;
 }
 
 
 void IncrementalMarking::RecordWriteSlow(HeapObject* obj,
                                          Object** slot,
                                          Object* value) {
   if (BaseRecordWrite(obj, slot, value) && slot != NULL) {
     MarkBit obj_bit = Marking::MarkBitFrom(obj);
     if (Marking::IsBlack(obj_bit)) {
       // Object is not going to be rescanned we need to record the slot.
       heap_->mark_compact_collector()->RecordSlot(
           HeapObject::RawField(obj, 0), slot, value);
     }
   }
 }
 
 
 void IncrementalMarking::RecordWriteFromCode(HeapObject* obj,
                                              Object* value,
                                              Isolate* isolate) {
   ASSERT(obj->IsHeapObject());
-
-  // Fast cases should already be covered by RecordWriteStub.
-  ASSERT(value->IsHeapObject());
-  ASSERT(!value->IsHeapNumber());
-  ASSERT(!value->IsString() ||
-         value->IsConsString() ||
-         value->IsSlicedString());
-  ASSERT(Marking::IsWhite(Marking::MarkBitFrom(HeapObject::cast(value))));
-
   IncrementalMarking* marking = isolate->heap()->incremental_marking();
   ASSERT(!marking->is_compacting_);
+
+  MemoryChunk* chunk = MemoryChunk::FromAddress(obj->address());
+  int counter = chunk->write_barrier_counter();
+  if (counter < (MemoryChunk::kWriteBarrierCounterGranularity / 2)) {
+    marking->write_barriers_invoked_since_last_step_ +=
+        MemoryChunk::kWriteBarrierCounterGranularity -
+            chunk->write_barrier_counter();
+    chunk->set_write_barrier_counter(
+        MemoryChunk::kWriteBarrierCounterGranularity);
+  }
+
   marking->RecordWrite(obj, NULL, value);
 }
 
 
 void IncrementalMarking::RecordWriteForEvacuationFromCode(HeapObject* obj,
                                                           Object** slot,
                                                           Isolate* isolate) {
+  ASSERT(obj->IsHeapObject());
   IncrementalMarking* marking = isolate->heap()->incremental_marking();
   ASSERT(marking->is_compacting_);
+
+  MemoryChunk* chunk = MemoryChunk::FromAddress(obj->address());
+  int counter = chunk->write_barrier_counter();
+  if (counter < (MemoryChunk::kWriteBarrierCounterGranularity / 2)) {
+    marking->write_barriers_invoked_since_last_step_ +=
+        MemoryChunk::kWriteBarrierCounterGranularity -
+            chunk->write_barrier_counter();
+    chunk->set_write_barrier_counter(
+        MemoryChunk::kWriteBarrierCounterGranularity);
+  }
+
   marking->RecordWrite(obj, slot, *slot);
 }
 
 
 void IncrementalMarking::RecordCodeTargetPatch(Code* host,
                                                Address pc,
                                                HeapObject* value) {
   if (IsMarking()) {
     RelocInfo rinfo(pc, RelocInfo::CODE_TARGET, 0, host);
     RecordWriteIntoCode(host, &rinfo, value);
@@ skipping 383 matching lines @@
   if (heap_->old_pointer_space()->IsSweepingComplete() &&
       heap_->old_data_space()->IsSweepingComplete()) {
     StartMarking(ALLOW_COMPACTION);
   } else {
     if (FLAG_trace_incremental_marking) {
       PrintF("[IncrementalMarking] Start sweeping.\n");
     }
     state_ = SWEEPING;
   }
 
-  heap_->new_space()->LowerInlineAllocationLimit(kAllocatedThreshold);
+  heap_->new_space()->LowerInlineAllocationLimit(kIncrementalMarkingThreshold);
 }
 
 
 static void MarkObjectGreyDoNotEnqueue(Object* obj) {
   if (obj->IsHeapObject()) {
     HeapObject* heap_obj = HeapObject::cast(obj);
     MarkBit mark_bit = Marking::MarkBitFrom(HeapObject::cast(obj));
     if (Marking::IsBlack(mark_bit)) {
       MemoryChunk::IncrementLiveBytesFromGC(heap_obj->address(),
                                             -heap_obj->Size());
@@ skipping 248 matching lines @@
                               CompletionAction action) {
   if (heap_->gc_state() != Heap::NOT_IN_GC ||
       !FLAG_incremental_marking ||
       !FLAG_incremental_marking_steps ||
       (state_ != SWEEPING && state_ != MARKING)) {
     return;
   }
 
   allocated_ += allocated_bytes;
 
-  if (allocated_ < kAllocatedThreshold) return;
+  if (allocated_ < kIncrementalMarkingThreshold &&
+      write_barriers_invoked_since_last_step_ < kIncrementalMarkingThreshold) {

[Michael Starzinger, 2012/09/26 08:06:38]

Even if the two thresholds we are comparing against happen to have the same
value, I think we should still use two different constants here (one for
allocation and one for WB counts).

+    return;
+  }
 
   if (state_ == MARKING && no_marking_scope_depth_ > 0) return;
 
-  intptr_t bytes_to_process = allocated_ * allocation_marking_factor_;
+  // The marking speed is driven either by the allocation rate or by the rate
+  // at which we are having to check the color of objects in the write barrier.
+  // It is possible for a tight non-allocating loop to run a lot of write
+  // barriers before we get here and check them (marking can only take place on
+  // allocation), so to reduce the lumpiness we don't use the write barriers
+  // invoked since last step directly to determine the amount of work to do.
+  intptr_t bytes_to_process =
+      marking_speed_ * Max(allocated_, kIncrementalMarkingThreshold);
+  allocated_ = 0;
+  write_barriers_invoked_since_last_step_ = 0;
+
   bytes_scanned_ += bytes_to_process;
 
   double start = 0;
 
   if (FLAG_trace_incremental_marking || FLAG_trace_gc) {
     start = OS::TimeCurrentMillis();
   }
 
   if (state_ == SWEEPING) {
     if (heap_->AdvanceSweepers(static_cast<int>(bytes_to_process))) {
@@ skipping 34 matching lines @@
 
       MarkBit obj_mark_bit = Marking::MarkBitFrom(obj);
       SLOW_ASSERT(Marking::IsGrey(obj_mark_bit) ||
                   (obj->IsFiller() && Marking::IsWhite(obj_mark_bit)));
       Marking::MarkBlack(obj_mark_bit);
       MemoryChunk::IncrementLiveBytesFromGC(obj->address(), size);
     }
     if (marking_deque_.IsEmpty()) MarkingComplete(action);
   }
 
-  allocated_ = 0;
-
   steps_count_++;
   steps_count_since_last_gc_++;
 
   bool speed_up = false;
 
-  if ((steps_count_ % kAllocationMarkingFactorSpeedupInterval) == 0) {
+  if ((steps_count_ % kMarkingSpeedAccellerationInterval) == 0) {
     if (FLAG_trace_gc) {
       PrintPID("Speed up marking after %d steps\n",
-               static_cast<int>(kAllocationMarkingFactorSpeedupInterval));
+               static_cast<int>(kMarkingSpeedAccellerationInterval));
     }
     speed_up = true;
   }
 
   bool space_left_is_very_small =
       (old_generation_space_available_at_start_of_incremental_ < 10 * MB);
 
   bool only_1_nth_of_space_that_was_available_still_left =
-      (SpaceLeftInOldSpace() * (allocation_marking_factor_ + 1) <
+      (SpaceLeftInOldSpace() * (marking_speed_ + 1) <
           old_generation_space_available_at_start_of_incremental_);
 
   if (space_left_is_very_small ||
       only_1_nth_of_space_that_was_available_still_left) {
     if (FLAG_trace_gc) PrintPID("Speed up marking because of low space left\n");
     speed_up = true;
   }
 
   bool size_of_old_space_multiplied_by_n_during_marking =
       (heap_->PromotedTotalSize() >
-       (allocation_marking_factor_ + 1) *
+       (marking_speed_ + 1) *
            old_generation_space_used_at_start_of_incremental_);
   if (size_of_old_space_multiplied_by_n_during_marking) {
     speed_up = true;
     if (FLAG_trace_gc) {
       PrintPID("Speed up marking because of heap size increase\n");
     }
   }
 
   int64_t promoted_during_marking = heap_->PromotedTotalSize()
       - old_generation_space_used_at_start_of_incremental_;
-  intptr_t delay = allocation_marking_factor_ * MB;
+  intptr_t delay = marking_speed_ * MB;
   intptr_t scavenge_slack = heap_->MaxSemiSpaceSize();
 
   // We try to scan at at least twice the speed that we are allocating.
   if (promoted_during_marking > bytes_scanned_ / 2 + scavenge_slack + delay) {
     if (FLAG_trace_gc) {
       PrintPID("Speed up marking because marker was not keeping up\n");
     }
     speed_up = true;
   }
 
   if (speed_up) {
     if (state_ != MARKING) {
       if (FLAG_trace_gc) {
         PrintPID("Postponing speeding up marking until marking starts\n");
       }
     } else {
-      allocation_marking_factor_ += kAllocationMarkingFactorSpeedup;
-      allocation_marking_factor_ = static_cast<int>(
-          Min(kMaxAllocationMarkingFactor,
-              static_cast<intptr_t>(allocation_marking_factor_ * 1.3)));
+      marking_speed_ += kMarkingSpeedAccellerationInterval;
+      marking_speed_ = static_cast<int>(
+          Min(kMaxMarkingSpeed,
+              static_cast<intptr_t>(marking_speed_ * 1.3)));
       if (FLAG_trace_gc) {
-        PrintPID("Marking speed increased to %d\n", allocation_marking_factor_);
+        PrintPID("Marking speed increased to %d\n", marking_speed_);
       }
     }
   }
 
   if (FLAG_trace_incremental_marking || FLAG_trace_gc) {
     double end = OS::TimeCurrentMillis();
     double delta = (end - start);
     longest_step_ = Max(longest_step_, delta);
     steps_took_ += delta;
     steps_took_since_last_gc_ += delta;
   }
 }
 
 
 void IncrementalMarking::ResetStepCounters() {
   steps_count_ = 0;
   steps_took_ = 0;
   longest_step_ = 0.0;
   old_generation_space_available_at_start_of_incremental_ =
       SpaceLeftInOldSpace();
   old_generation_space_used_at_start_of_incremental_ =
       heap_->PromotedTotalSize();
   steps_count_since_last_gc_ = 0;
   steps_took_since_last_gc_ = 0;
   bytes_rescanned_ = 0;
-  allocation_marking_factor_ = kInitialAllocationMarkingFactor;
+  marking_speed_ = kInitialMarkingSpeed;
   bytes_scanned_ = 0;
+  write_barriers_invoked_since_last_step_ = 0;
 }
 
 
 int64_t IncrementalMarking::SpaceLeftInOldSpace() {
   return heap_->MaxOldGenerationSize() - heap_->PromotedSpaceSizeOfObjects();
 }
 
 } }  // namespace v8::internal