8195103: Refactor out card table from CardTableModRefBS to flatten the BarrierSet hierarchy

Reviewed-by: kbarrett, tschatzl

src/hotspot/share/gc/cms/cmsHeap.hpp

@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2017, Oracle and/or its affiliates. All rights reserved.
+ * Copyright (c) 2017, 2018, Oracle and/or its affiliates. All rights reserved.
  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
  *
  * This code is free software; you can redistribute it and/or modify it
@@ -75,10 +75,6 @@ public:
   // supports. Caller does not hold the Heap_lock on entry.
   void collect(GCCause::Cause cause);
 
-  bool card_mark_must_follow_store() const {
-    return true;
-  }
-
   void stop();
   void safepoint_synchronize_begin();
   void safepoint_synchronize_end();

src/hotspot/share/gc/g1/g1CollectedHeap.cpp

@@ -1573,7 +1573,6 @@ jint G1CollectedHeap::initialize_young_gen_sampling_thread() {
 }
 
 jint G1CollectedHeap::initialize() {
-  CollectedHeap::pre_initialize();
   os::enable_vtime();
 
   // Necessary to satisfy locking discipline assertions.

src/hotspot/share/gc/g1/g1CollectedHeap.hpp

@@ -1272,36 +1272,8 @@ public:
   size_t max_tlab_size() const;
   size_t unsafe_max_tlab_alloc(Thread* ignored) const;
 
-  // Can a compiler initialize a new object without store barriers?
-  // This permission only extends from the creation of a new object
-  // via a TLAB up to the first subsequent safepoint. If such permission
-  // is granted for this heap type, the compiler promises to call
-  // defer_store_barrier() below on any slow path allocation of
-  // a new object for which such initializing store barriers will
-  // have been elided. G1, like CMS, allows this, but should be
-  // ready to provide a compensating write barrier as necessary
-  // if that storage came out of a non-young region. The efficiency
-  // of this implementation depends crucially on being able to
-  // answer very efficiently in constant time whether a piece of
-  // storage in the heap comes from a young region or not.
-  // See ReduceInitialCardMarks.
-  virtual bool can_elide_tlab_store_barriers() const {
-    return true;
-  }
-
-  virtual bool card_mark_must_follow_store() const {
-    return true;
-  }
-
   inline bool is_in_young(const oop obj);
 
-  // We don't need barriers for initializing stores to objects
-  // in the young gen: for the SATB pre-barrier, there is no
-  // pre-value that needs to be remembered; for the remembered-set
-  // update logging post-barrier, we don't maintain remembered set
-  // information for young gen objects.
-  virtual inline bool can_elide_initializing_store_barrier(oop new_obj);
-
   // Returns "true" iff the given word_size is "very large".
   static bool is_humongous(size_t word_size) {
     // Note this has to be strictly greater-than as the TLABs

src/hotspot/share/gc/g1/g1CollectedHeap.inline.hpp

@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2001, 2017, Oracle and/or its affiliates. All rights reserved.
+ * Copyright (c) 2001, 2018, Oracle and/or its affiliates. All rights reserved.
  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
  *
  * This code is free software; you can redistribute it and/or modify it
@@ -241,15 +241,6 @@ inline bool G1CollectedHeap::is_in_young(const oop obj) {
   return heap_region_containing(obj)->is_young();
 }
 
-// We don't need barriers for initializing stores to objects
-// in the young gen: for the SATB pre-barrier, there is no
-// pre-value that needs to be remembered; for the remembered-set
-// update logging post-barrier, we don't maintain remembered set
-// information for young gen objects.
-inline bool G1CollectedHeap::can_elide_initializing_store_barrier(oop new_obj) {
-  return is_in_young(new_obj);
-}
-
 inline bool G1CollectedHeap::is_obj_dead(const oop obj) const {
   if (obj == NULL) {
     return false;

src/hotspot/share/gc/g1/g1SATBCardTableModRefBS.cpp

@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2001, 2017, Oracle and/or its affiliates. All rights reserved.
+ * Copyright (c) 2001, 2018, Oracle and/or its affiliates. All rights reserved.
  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
  *
  * This code is free software; you can redistribute it and/or modify it
@@ -131,6 +131,7 @@ G1SATBCardTableLoggingModRefBS(MemRegion whole_heap) :
 }
 
 void G1SATBCardTableLoggingModRefBS::initialize(G1RegionToSpaceMapper* mapper) {
+  initialize_deferred_card_mark_barriers();
   mapper->set_mapping_changed_listener(&_listener);
 
   _byte_map_size = mapper->reserved().byte_size();
@@ -213,3 +214,14 @@ G1SATBCardTableLoggingModRefBS::invalidate(MemRegion mr) {
     }
   }
 }
+
+bool G1SATBCardTableModRefBS::is_in_young(oop obj) const {
+  volatile jbyte* p = byte_for((void*)obj);
+  return *p == g1_young_card_val();
+}
+
+void G1SATBCardTableLoggingModRefBS::flush_deferred_barriers(JavaThread* thread) {
+  CardTableModRefBS::flush_deferred_barriers(thread);
+  thread->satb_mark_queue().flush();
+  thread->dirty_card_queue().flush();
+}

src/hotspot/share/gc/g1/g1SATBCardTableModRefBS.hpp

@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2001, 2017, Oracle and/or its affiliates. All rights reserved.
+ * Copyright (c) 2001, 2018, Oracle and/or its affiliates. All rights reserved.
  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
  *
  * This code is free software; you can redistribute it and/or modify it
@@ -92,6 +92,8 @@ public:
     jbyte val = _byte_map[card_index];
     return (val & (clean_card_mask_val() | deferred_card_val())) == deferred_card_val();
   }
+
+  virtual bool is_in_young(oop obj) const;
 };
 
 template<>
@@ -145,13 +147,19 @@ class G1SATBCardTableLoggingModRefBS: public G1SATBCardTableModRefBS {
   // above no longer applies.
   void invalidate(MemRegion mr);
 
-  void write_region_work(MemRegion mr)    { invalidate(mr); }
+  void write_region(MemRegion mr)         { invalidate(mr); }
   void write_ref_array_work(MemRegion mr) { invalidate(mr); }
 
   template <DecoratorSet decorators, typename T>
   void write_ref_field_post(T* field, oop new_val);
   void write_ref_field_post_slow(volatile jbyte* byte);
 
+  virtual void flush_deferred_barriers(JavaThread* thread);
+
+  virtual bool card_mark_must_follow_store() const {
+    return true;
+  }
+
   // Callbacks for runtime accesses.
   template <DecoratorSet decorators, typename BarrierSetT = G1SATBCardTableLoggingModRefBS>
   class AccessBarrier: public ModRefBarrierSet::AccessBarrier<decorators, BarrierSetT> {

src/hotspot/share/gc/parallel/cardTableExtension.cpp

@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2001, 2017, Oracle and/or its affiliates. All rights reserved.
+ * Copyright (c) 2001, 2018, Oracle and/or its affiliates. All rights reserved.
  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
  *
  * This code is free software; you can redistribute it and/or modify it
@@ -26,7 +26,7 @@
 #include "gc/parallel/cardTableExtension.hpp"
 #include "gc/parallel/gcTaskManager.hpp"
 #include "gc/parallel/objectStartArray.inline.hpp"
-#include "gc/parallel/parallelScavengeHeap.hpp"
+#include "gc/parallel/parallelScavengeHeap.inline.hpp"
 #include "gc/parallel/psPromotionManager.inline.hpp"
 #include "gc/parallel/psScavenge.hpp"
 #include "gc/parallel/psTasks.hpp"
@@ -677,3 +677,7 @@ HeapWord* CardTableExtension::lowest_prev_committed_start(int ind) const {
   }
   return min_start;
 }
+
+bool CardTableExtension::is_in_young(oop obj) const {
+  return ParallelScavengeHeap::heap()->is_in_young(obj);
+}

src/hotspot/share/gc/parallel/cardTableExtension.hpp

@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2001, 2017, Oracle and/or its affiliates. All rights reserved.
+ * Copyright (c) 2001, 2018, Oracle and/or its affiliates. All rights reserved.
  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
  *
  * This code is free software; you can redistribute it and/or modify it
@@ -108,6 +108,13 @@ class CardTableExtension : public CardTableModRefBS {
   }
 
 #endif // ASSERT
+
+  // ReduceInitialCardMarks support
+  virtual bool is_in_young(oop obj) const;
+
+  virtual bool card_mark_must_follow_store() const {
+    return false;
+  }
 };
 
 template<>

src/hotspot/share/gc/parallel/parallelScavengeHeap.cpp

@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2001, 2017, Oracle and/or its affiliates. All rights reserved.
+ * Copyright (c) 2001, 2018, Oracle and/or its affiliates. All rights reserved.
  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
  *
  * This code is free software; you can redistribute it and/or modify it
@@ -57,8 +57,6 @@ PSGCAdaptivePolicyCounters* ParallelScavengeHeap::_gc_policy_counters = NULL;
 GCTaskManager* ParallelScavengeHeap::_gc_task_manager = NULL;
 
 jint ParallelScavengeHeap::initialize() {
-  CollectedHeap::pre_initialize();
-
   const size_t heap_size = _collector_policy->max_heap_byte_size();
 
   ReservedSpace heap_rs = Universe::reserve_heap(heap_size, _collector_policy->heap_alignment());
@@ -490,13 +488,6 @@ void ParallelScavengeHeap::resize_all_tlabs() {
   CollectedHeap::resize_all_tlabs();
 }
 
-bool ParallelScavengeHeap::can_elide_initializing_store_barrier(oop new_obj) {
-  // We don't need barriers for stores to objects in the
-  // young gen and, a fortiori, for initializing stores to
-  // objects therein.
-  return is_in_young(new_obj);
-}
-
 // This method is used by System.gc() and JVMTI.
 void ParallelScavengeHeap::collect(GCCause::Cause cause) {
   assert(!Heap_lock->owned_by_self(),
@@ -719,4 +710,3 @@ GrowableArray<MemoryPool*> ParallelScavengeHeap::memory_pools() {
   memory_pools.append(_old_pool);
   return memory_pools;
 }
-

src/hotspot/share/gc/parallel/parallelScavengeHeap.hpp

@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2001, 2017, Oracle and/or its affiliates. All rights reserved.
+ * Copyright (c) 2001, 2018, Oracle and/or its affiliates. All rights reserved.
  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
  *
  * This code is free software; you can redistribute it and/or modify it
@@ -205,21 +205,6 @@ class ParallelScavengeHeap : public CollectedHeap {
   size_t tlab_used(Thread* thr) const;
   size_t unsafe_max_tlab_alloc(Thread* thr) const;
 
-  // Can a compiler initialize a new object without store barriers?
-  // This permission only extends from the creation of a new object
-  // via a TLAB up to the first subsequent safepoint.
-  virtual bool can_elide_tlab_store_barriers() const {
-    return true;
-  }
-
-  virtual bool card_mark_must_follow_store() const {
-    return false;
-  }
-
-  // Return true if we don't we need a store barrier for
-  // initializing stores to an object at this address.
-  virtual bool can_elide_initializing_store_barrier(oop new_obj);
-
   void object_iterate(ObjectClosure* cl);
   void safe_object_iterate(ObjectClosure* cl) { object_iterate(cl); }
 

src/hotspot/share/gc/serial/serialHeap.hpp

@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2017, Oracle and/or its affiliates. All rights reserved.
+ * Copyright (c) 2017, 2018, Oracle and/or its affiliates. All rights reserved.
  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
  *
  * This code is free software; you can redistribute it and/or modify it
@@ -61,10 +61,6 @@ public:
   virtual bool is_in_closed_subset(const void* p) const {
     return is_in(p);
   }
-
-  virtual bool card_mark_must_follow_store() const {
-    return false;
-  }
 };
 
 #endif // SHARE_VM_GC_CMS_CMSHEAP_HPP

src/hotspot/share/gc/shared/barrierSet.hpp

@@ -32,6 +32,8 @@
 #include "oops/oopsHierarchy.hpp"
 #include "utilities/fakeRttiSupport.hpp"
 
+class JavaThread;
+
 // This class provides the interface between a barrier implementation and
 // the rest of the system.
 
@@ -107,18 +109,18 @@ public:
   static void static_write_ref_array_pre(HeapWord* start, size_t count);
   static void static_write_ref_array_post(HeapWord* start, size_t count);
 
+  // Support for optimizing compilers to call the barrier set on slow path allocations
+  // that did not enter a TLAB. Used for e.g. ReduceInitialCardMarks.
+  // The allocation is safe to use iff it returns true. If not, the slow-path allocation
+  // is redone until it succeeds. This can e.g. prevent allocations from the slow path
+  // to be in old.
+  virtual void on_slowpath_allocation_exit(JavaThread* thread, oop new_obj) {}
+  virtual void flush_deferred_barriers(JavaThread* thread) {}
+  virtual void make_parsable(JavaThread* thread) {}
+
 protected:
   virtual void write_ref_array_work(MemRegion mr) = 0;
 
-public:
-  // (For efficiency reasons, this operation is specialized for certain
-  // barrier types.  Semantically, it should be thought of as a call to the
-  // virtual "_work" function below, which must implement the barrier.)
-  void write_region(MemRegion mr);
-
-protected:
-  virtual void write_region_work(MemRegion mr) = 0;
-
 public:
   // Inform the BarrierSet that the the covered heap region that starts
   // with "base" has been changed to have the given size (possibly from 0,

src/hotspot/share/gc/shared/barrierSet.inline.hpp

@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2001, 2017, Oracle and/or its affiliates. All rights reserved.
+ * Copyright (c) 2001, 2018, Oracle and/or its affiliates. All rights reserved.
  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
  *
  * This code is free software; you can redistribute it and/or modify it
@@ -52,8 +52,4 @@ void BarrierSet::write_ref_array(HeapWord* start, size_t count) {
   write_ref_array_work(MemRegion(aligned_start, aligned_end));
 }
 
-inline void BarrierSet::write_region(MemRegion mr) {
-  write_region_work(mr);
-}
-
 #endif // SHARE_VM_GC_SHARED_BARRIERSET_INLINE_HPP

src/hotspot/share/gc/shared/cardTableModRefBS.cpp

@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2000, 2017, Oracle and/or its affiliates. All rights reserved.
+ * Copyright (c) 2000, 2018, Oracle and/or its affiliates. All rights reserved.
  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
  *
  * This code is free software; you can redistribute it and/or modify it
@@ -30,6 +30,7 @@
 #include "logging/log.hpp"
 #include "memory/virtualspace.hpp"
 #include "oops/oop.inline.hpp"
+#include "runtime/thread.hpp"
 #include "services/memTracker.hpp"
 #include "utilities/align.hpp"
 #include "utilities/macros.hpp"
@@ -61,7 +62,8 @@ CardTableModRefBS::CardTableModRefBS(
   _committed(NULL),
   _cur_covered_regions(0),
   _byte_map(NULL),
-  byte_map_base(NULL)
+  byte_map_base(NULL),
+  _defer_initial_card_mark(false)
 {
   assert((uintptr_t(_whole_heap.start())  & (card_size - 1))  == 0, "heap must start at card boundary");
   assert((uintptr_t(_whole_heap.end()) & (card_size - 1))  == 0, "heap must end at card boundary");
@@ -75,6 +77,7 @@ CardTableModRefBS::CardTableModRefBS(
 }
 
 void CardTableModRefBS::initialize() {
+  initialize_deferred_card_mark_barriers();
   _guard_index = cards_required(_whole_heap.word_size()) - 1;
   _last_valid_index = _guard_index - 1;
 
@@ -521,3 +524,112 @@ void CardTableModRefBS::print_on(outputStream* st) const {
   st->print_cr("Card table byte_map: [" INTPTR_FORMAT "," INTPTR_FORMAT "] byte_map_base: " INTPTR_FORMAT,
                p2i(_byte_map), p2i(_byte_map + _byte_map_size), p2i(byte_map_base));
 }
+
+// Helper for ReduceInitialCardMarks. For performance,
+// compiled code may elide card-marks for initializing stores
+// to a newly allocated object along the fast-path. We
+// compensate for such elided card-marks as follows:
+// (a) Generational, non-concurrent collectors, such as
+//     GenCollectedHeap(ParNew,DefNew,Tenured) and
+//     ParallelScavengeHeap(ParallelGC, ParallelOldGC)
+//     need the card-mark if and only if the region is
+//     in the old gen, and do not care if the card-mark
+//     succeeds or precedes the initializing stores themselves,
+//     so long as the card-mark is completed before the next
+//     scavenge. For all these cases, we can do a card mark
+//     at the point at which we do a slow path allocation
+//     in the old gen, i.e. in this call.
+// (b) GenCollectedHeap(ConcurrentMarkSweepGeneration) requires
+//     in addition that the card-mark for an old gen allocated
+//     object strictly follow any associated initializing stores.
+//     In these cases, the memRegion remembered below is
+//     used to card-mark the entire region either just before the next
+//     slow-path allocation by this thread or just before the next scavenge or
+//     CMS-associated safepoint, whichever of these events happens first.
+//     (The implicit assumption is that the object has been fully
+//     initialized by this point, a fact that we assert when doing the
+//     card-mark.)
+// (c) G1CollectedHeap(G1) uses two kinds of write barriers. When a
+//     G1 concurrent marking is in progress an SATB (pre-write-)barrier
+//     is used to remember the pre-value of any store. Initializing
+//     stores will not need this barrier, so we need not worry about
+//     compensating for the missing pre-barrier here. Turning now
+//     to the post-barrier, we note that G1 needs a RS update barrier
+//     which simply enqueues a (sequence of) dirty cards which may
+//     optionally be refined by the concurrent update threads. Note
+//     that this barrier need only be applied to a non-young write,
+//     but, like in CMS, because of the presence of concurrent refinement
+//     (much like CMS' precleaning), must strictly follow the oop-store.
+//     Thus, using the same protocol for maintaining the intended
+//     invariants turns out, serendepitously, to be the same for both
+//     G1 and CMS.
+//
+// For any future collector, this code should be reexamined with
+// that specific collector in mind, and the documentation above suitably
+// extended and updated.
+void CardTableModRefBS::on_slowpath_allocation_exit(JavaThread* thread, oop new_obj) {
+  if (!ReduceInitialCardMarks) {
+    return;
+  }
+  // If a previous card-mark was deferred, flush it now.
+  flush_deferred_card_mark_barrier(thread);
+  if (new_obj->is_typeArray() || is_in_young(new_obj)) {
+    // Arrays of non-references don't need a post-barrier.
+    // The deferred_card_mark region should be empty
+    // following the flush above.
+    assert(thread->deferred_card_mark().is_empty(), "Error");
+  } else {
+    MemRegion mr((HeapWord*)new_obj, new_obj->size());
+    assert(!mr.is_empty(), "Error");
+    if (_defer_initial_card_mark) {
+      // Defer the card mark
+      thread->set_deferred_card_mark(mr);
+    } else {
+      // Do the card mark
+      write_region(mr);
+    }
+  }
+}
+
+void CardTableModRefBS::initialize_deferred_card_mark_barriers() {
+  // Used for ReduceInitialCardMarks (when COMPILER2 or JVMCI is used);
+  // otherwise remains unused.
+#if defined(COMPILER2) || INCLUDE_JVMCI
+  _defer_initial_card_mark = is_server_compilation_mode_vm() && ReduceInitialCardMarks && can_elide_tlab_store_barriers()
+                             && (DeferInitialCardMark || card_mark_must_follow_store());
+#else
+  assert(_defer_initial_card_mark == false, "Who would set it?");
+#endif
+}
+
+void CardTableModRefBS::flush_deferred_card_mark_barrier(JavaThread* thread) {
+#if defined(COMPILER2) || INCLUDE_JVMCI
+  MemRegion deferred = thread->deferred_card_mark();
+  if (!deferred.is_empty()) {
+    assert(_defer_initial_card_mark, "Otherwise should be empty");
+    {
+      // Verify that the storage points to a parsable object in heap
+      DEBUG_ONLY(oop old_obj = oop(deferred.start());)
+      assert(!is_in_young(old_obj),
+             "Else should have been filtered in on_slowpath_allocation_exit()");
+      assert(oopDesc::is_oop(old_obj, true), "Not an oop");
+      assert(deferred.word_size() == (size_t)(old_obj->size()),
+             "Mismatch: multiple objects?");
+    }
+    write_region(deferred);
+    // "Clear" the deferred_card_mark field
+    thread->set_deferred_card_mark(MemRegion());
+  }
+  assert(thread->deferred_card_mark().is_empty(), "invariant");
+#else
+  assert(!_defer_initial_card_mark, "Should be false");
+  assert(thread->deferred_card_mark().is_empty(), "Should be empty");
+#endif
+}
+
+void CardTableModRefBS::flush_deferred_barriers(JavaThread* thread) {
+  // The deferred store barriers must all have been flushed to the
+  // card-table (or other remembered set structure) before GC starts
+  // processing the card-table (or other remembered set).
+  flush_deferred_card_mark_barrier(thread);
+}

src/hotspot/share/gc/shared/cardTableModRefBS.hpp

@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2000, 2017, Oracle and/or its affiliates. All rights reserved.
+ * Copyright (c) 2000, 2018, Oracle and/or its affiliates. All rights reserved.
  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
  *
  * This code is free software; you can redistribute it and/or modify it
@@ -58,6 +58,10 @@ class CardTableModRefBS: public ModRefBarrierSet {
     CT_MR_BS_last_reserved      = 16
   };
 
+  // Used in support of ReduceInitialCardMarks; only consulted if COMPILER2
+  // or INCLUDE_JVMCI is being used
+  bool _defer_initial_card_mark;
+
   // a word's worth (row) of clean card values
   static const intptr_t clean_card_row = (intptr_t)(-1);
 
@@ -180,8 +184,8 @@ class CardTableModRefBS: public ModRefBarrierSet {
   CardTableModRefBS(MemRegion whole_heap, const BarrierSet::FakeRtti& fake_rtti);
   ~CardTableModRefBS();
 
- protected:
+ public:
-  void write_region_work(MemRegion mr) {
+  void write_region(MemRegion mr) {
     dirty_MemRegion(mr);
   }
 
@@ -314,6 +318,49 @@ class CardTableModRefBS: public ModRefBarrierSet {
   void verify_not_dirty_region(MemRegion mr) PRODUCT_RETURN;
   void verify_dirty_region(MemRegion mr) PRODUCT_RETURN;
 
+  // ReduceInitialCardMarks
+  void initialize_deferred_card_mark_barriers();
+
+  // If the CollectedHeap was asked to defer a store barrier above,
+  // this informs it to flush such a deferred store barrier to the
+  // remembered set.
+  void flush_deferred_card_mark_barrier(JavaThread* thread);
+
+  // Can a compiler initialize a new object without store barriers?
+  // This permission only extends from the creation of a new object
+  // via a TLAB up to the first subsequent safepoint. If such permission
+  // is granted for this heap type, the compiler promises to call
+  // defer_store_barrier() below on any slow path allocation of
+  // a new object for which such initializing store barriers will
+  // have been elided. G1, like CMS, allows this, but should be
+  // ready to provide a compensating write barrier as necessary
+  // if that storage came out of a non-young region. The efficiency
+  // of this implementation depends crucially on being able to
+  // answer very efficiently in constant time whether a piece of
+  // storage in the heap comes from a young region or not.
+  // See ReduceInitialCardMarks.
+  virtual bool can_elide_tlab_store_barriers() const {
+    return true;
+  }
+
+  // If a compiler is eliding store barriers for TLAB-allocated objects,
+  // we will be informed of a slow-path allocation by a call
+  // to on_slowpath_allocation_exit() below. Such a call precedes the
+  // initialization of the object itself, and no post-store-barriers will
+  // be issued. Some heap types require that the barrier strictly follows
+  // the initializing stores. (This is currently implemented by deferring the
+  // barrier until the next slow-path allocation or gc-related safepoint.)
+  // This interface answers whether a particular barrier type needs the card
+  // mark to be thus strictly sequenced after the stores.
+  virtual bool card_mark_must_follow_store() const = 0;
+
+  virtual bool is_in_young(oop obj) const = 0;
+
+  virtual void on_slowpath_allocation_exit(JavaThread* thread, oop new_obj);
+  virtual void flush_deferred_barriers(JavaThread* thread);
+
+  virtual void make_parsable(JavaThread* thread) { flush_deferred_card_mark_barrier(thread); }
+
   template <DecoratorSet decorators, typename BarrierSetT = CardTableModRefBS>
   class AccessBarrier: public ModRefBarrierSet::AccessBarrier<decorators, BarrierSetT> {};
 };

src/hotspot/share/gc/shared/cardTableModRefBSForCTRS.cpp

@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2015, Oracle and/or its affiliates. All rights reserved.
+ * Copyright (c) 2015, 2018, Oracle and/or its affiliates. All rights reserved.
  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
  *
  * This code is free software; you can redistribute it and/or modify it
@@ -121,3 +121,6 @@ void CardTableModRefBSForCTRS::non_clean_card_iterate_possibly_parallel(
   }
 }
 
+bool CardTableModRefBSForCTRS::is_in_young(oop obj) const {
+  return GenCollectedHeap::heap()->is_in_young(obj);
+}

src/hotspot/share/gc/shared/cardTableModRefBSForCTRS.hpp

@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2015, 2017, Oracle and/or its affiliates. All rights reserved.
+ * Copyright (c) 2015, 2018, Oracle and/or its affiliates. All rights reserved.
  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
  *
  * This code is free software; you can redistribute it and/or modify it
@@ -43,6 +43,12 @@ public:
 
   void set_CTRS(CardTableRS* rs) { _rs = rs; }
 
+  virtual bool card_mark_must_follow_store() const {
+    return UseConcMarkSweepGC;
+  }
+
+  virtual bool is_in_young(oop obj) const;
+
 private:
   CardTableRS* _rs;
 

src/hotspot/share/gc/shared/collectedHeap.cpp

@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2001, 2017, Oracle and/or its affiliates. All rights reserved.
+ * Copyright (c) 2001, 2018, Oracle and/or its affiliates. All rights reserved.
  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
  *
  * This code is free software; you can redistribute it and/or modify it
@@ -177,8 +177,7 @@ CollectedHeap::CollectedHeap() :
   _total_collections(0),
   _total_full_collections(0),
   _gc_cause(GCCause::_no_gc),
-  _gc_lastcause(GCCause::_no_gc),
+  _gc_lastcause(GCCause::_no_gc)
-  _defer_initial_card_mark(false) // strengthened by subclass in pre_initialize() below.
 {
   const size_t max_len = size_t(arrayOopDesc::max_array_length(T_INT));
   const size_t elements_per_word = HeapWordSize / sizeof(jint);
@@ -239,17 +238,6 @@ void CollectedHeap::set_barrier_set(BarrierSet* barrier_set) {
   BarrierSet::set_bs(barrier_set);
 }
 
-void CollectedHeap::pre_initialize() {
-  // Used for ReduceInitialCardMarks (when COMPILER2 is used);
-  // otherwise remains unused.
-#if COMPILER2_OR_JVMCI
-  _defer_initial_card_mark = is_server_compilation_mode_vm() &&  ReduceInitialCardMarks && can_elide_tlab_store_barriers()
-                             && (DeferInitialCardMark || card_mark_must_follow_store());
-#else
-  assert(_defer_initial_card_mark == false, "Who would set it?");
-#endif
-}
-
 #ifndef PRODUCT
 void CollectedHeap::check_for_bad_heap_word_value(HeapWord* addr, size_t size) {
   if (CheckMemoryInitialization && ZapUnusedHeapArea) {
@@ -333,28 +321,6 @@ HeapWord* CollectedHeap::allocate_from_tlab_slow(Klass* klass, Thread* thread, s
   return obj;
 }
 
-void CollectedHeap::flush_deferred_store_barrier(JavaThread* thread) {
-  MemRegion deferred = thread->deferred_card_mark();
-  if (!deferred.is_empty()) {
-    assert(_defer_initial_card_mark, "Otherwise should be empty");
-    {
-      // Verify that the storage points to a parsable object in heap
-      DEBUG_ONLY(oop old_obj = oop(deferred.start());)
-      assert(is_in(old_obj), "Not in allocated heap");
-      assert(!can_elide_initializing_store_barrier(old_obj),
-             "Else should have been filtered in new_store_pre_barrier()");
-      assert(oopDesc::is_oop(old_obj, true), "Not an oop");
-      assert(deferred.word_size() == (size_t)(old_obj->size()),
-             "Mismatch: multiple objects?");
-    }
-    BarrierSet* bs = barrier_set();
-    bs->write_region(deferred);
-    // "Clear" the deferred_card_mark field
-    thread->set_deferred_card_mark(MemRegion());
-  }
-  assert(thread->deferred_card_mark().is_empty(), "invariant");
-}
-
 size_t CollectedHeap::max_tlab_size() const {
   // TLABs can't be bigger than we can fill with a int[Integer.MAX_VALUE].
   // This restriction could be removed by enabling filling with multiple arrays.
@@ -370,72 +336,6 @@ size_t CollectedHeap::max_tlab_size() const {
   return align_down(max_int_size, MinObjAlignment);
 }
 
-// Helper for ReduceInitialCardMarks. For performance,
-// compiled code may elide card-marks for initializing stores
-// to a newly allocated object along the fast-path. We
-// compensate for such elided card-marks as follows:
-// (a) Generational, non-concurrent collectors, such as
-//     GenCollectedHeap(ParNew,DefNew,Tenured) and
-//     ParallelScavengeHeap(ParallelGC, ParallelOldGC)
-//     need the card-mark if and only if the region is
-//     in the old gen, and do not care if the card-mark
-//     succeeds or precedes the initializing stores themselves,
-//     so long as the card-mark is completed before the next
-//     scavenge. For all these cases, we can do a card mark
-//     at the point at which we do a slow path allocation
-//     in the old gen, i.e. in this call.
-// (b) GenCollectedHeap(ConcurrentMarkSweepGeneration) requires
-//     in addition that the card-mark for an old gen allocated
-//     object strictly follow any associated initializing stores.
-//     In these cases, the memRegion remembered below is
-//     used to card-mark the entire region either just before the next
-//     slow-path allocation by this thread or just before the next scavenge or
-//     CMS-associated safepoint, whichever of these events happens first.
-//     (The implicit assumption is that the object has been fully
-//     initialized by this point, a fact that we assert when doing the
-//     card-mark.)
-// (c) G1CollectedHeap(G1) uses two kinds of write barriers. When a
-//     G1 concurrent marking is in progress an SATB (pre-write-)barrier
-//     is used to remember the pre-value of any store. Initializing
-//     stores will not need this barrier, so we need not worry about
-//     compensating for the missing pre-barrier here. Turning now
-//     to the post-barrier, we note that G1 needs a RS update barrier
-//     which simply enqueues a (sequence of) dirty cards which may
-//     optionally be refined by the concurrent update threads. Note
-//     that this barrier need only be applied to a non-young write,
-//     but, like in CMS, because of the presence of concurrent refinement
-//     (much like CMS' precleaning), must strictly follow the oop-store.
-//     Thus, using the same protocol for maintaining the intended
-//     invariants turns out, serendepitously, to be the same for both
-//     G1 and CMS.
-//
-// For any future collector, this code should be reexamined with
-// that specific collector in mind, and the documentation above suitably
-// extended and updated.
-oop CollectedHeap::new_store_pre_barrier(JavaThread* thread, oop new_obj) {
-  // If a previous card-mark was deferred, flush it now.
-  flush_deferred_store_barrier(thread);
-  if (can_elide_initializing_store_barrier(new_obj) ||
-      new_obj->is_typeArray()) {
-    // Arrays of non-references don't need a pre-barrier.
-    // The deferred_card_mark region should be empty
-    // following the flush above.
-    assert(thread->deferred_card_mark().is_empty(), "Error");
-  } else {
-    MemRegion mr((HeapWord*)new_obj, new_obj->size());
-    assert(!mr.is_empty(), "Error");
-    if (_defer_initial_card_mark) {
-      // Defer the card mark
-      thread->set_deferred_card_mark(mr);
-    } else {
-      // Do the card mark
-      BarrierSet* bs = barrier_set();
-      bs->write_region(mr);
-    }
-  }
-  return new_obj;
-}
-
 size_t CollectedHeap::filler_array_hdr_size() {
   return align_object_offset(arrayOopDesc::header_size(T_INT)); // align to Long
 }
@@ -538,24 +438,16 @@ void CollectedHeap::ensure_parsability(bool retire_tlabs) {
          " otherwise concurrent mutator activity may make heap "
          " unparsable again");
   const bool use_tlab = UseTLAB;
-  const bool deferred = _defer_initial_card_mark;
   // The main thread starts allocating via a TLAB even before it
   // has added itself to the threads list at vm boot-up.
   JavaThreadIteratorWithHandle jtiwh;
   assert(!use_tlab || jtiwh.length() > 0,
          "Attempt to fill tlabs before main thread has been added"
          " to threads list is doomed to failure!");
+  BarrierSet *bs = barrier_set();
   for (; JavaThread *thread = jtiwh.next(); ) {
      if (use_tlab) thread->tlab().make_parsable(retire_tlabs);
-#if COMPILER2_OR_JVMCI
+     bs->make_parsable(thread);
-     // The deferred store barriers must all have been flushed to the
-     // card-table (or other remembered set structure) before GC starts
-     // processing the card-table (or other remembered set).
-     if (deferred) flush_deferred_store_barrier(thread);
-#else
-     assert(!deferred, "Should be false");
-     assert(thread->deferred_card_mark().is_empty(), "Should be empty");
-#endif
   }
 }
 

src/hotspot/share/gc/shared/collectedHeap.hpp

@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2001, 2017, Oracle and/or its affiliates. All rights reserved.
+ * Copyright (c) 2001, 2018, Oracle and/or its affiliates. All rights reserved.
  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
  *
  * This code is free software; you can redistribute it and/or modify it
@@ -101,10 +101,6 @@ class CollectedHeap : public CHeapObj<mtInternal> {
 
   GCHeapLog* _gc_heap_log;
 
-  // Used in support of ReduceInitialCardMarks; only consulted if COMPILER2
-  // or INCLUDE_JVMCI is being used
-  bool _defer_initial_card_mark;
-
   MemRegion _reserved;
 
  protected:
@@ -129,13 +125,6 @@ class CollectedHeap : public CHeapObj<mtInternal> {
   // Constructor
   CollectedHeap();
 
-  // Do common initializations that must follow instance construction,
-  // for example, those needing virtual calls.
-  // This code could perhaps be moved into initialize() but would
-  // be slightly more awkward because we want the latter to be a
-  // pure virtual.
-  void pre_initialize();
-
   // Create a new tlab. All TLAB allocations must go through this.
   virtual HeapWord* allocate_new_tlab(size_t size);
 
@@ -408,45 +397,6 @@ class CollectedHeap : public CHeapObj<mtInternal> {
     return 0;
   }
 
-  // Can a compiler initialize a new object without store barriers?
-  // This permission only extends from the creation of a new object
-  // via a TLAB up to the first subsequent safepoint. If such permission
-  // is granted for this heap type, the compiler promises to call
-  // defer_store_barrier() below on any slow path allocation of
-  // a new object for which such initializing store barriers will
-  // have been elided.
-  virtual bool can_elide_tlab_store_barriers() const = 0;
-
-  // If a compiler is eliding store barriers for TLAB-allocated objects,
-  // there is probably a corresponding slow path which can produce
-  // an object allocated anywhere.  The compiler's runtime support
-  // promises to call this function on such a slow-path-allocated
-  // object before performing initializations that have elided
-  // store barriers. Returns new_obj, or maybe a safer copy thereof.
-  virtual oop new_store_pre_barrier(JavaThread* thread, oop new_obj);
-
-  // Answers whether an initializing store to a new object currently
-  // allocated at the given address doesn't need a store
-  // barrier. Returns "true" if it doesn't need an initializing
-  // store barrier; answers "false" if it does.
-  virtual bool can_elide_initializing_store_barrier(oop new_obj) = 0;
-
-  // If a compiler is eliding store barriers for TLAB-allocated objects,
-  // we will be informed of a slow-path allocation by a call
-  // to new_store_pre_barrier() above. Such a call precedes the
-  // initialization of the object itself, and no post-store-barriers will
-  // be issued. Some heap types require that the barrier strictly follows
-  // the initializing stores. (This is currently implemented by deferring the
-  // barrier until the next slow-path allocation or gc-related safepoint.)
-  // This interface answers whether a particular heap type needs the card
-  // mark to be thus strictly sequenced after the stores.
-  virtual bool card_mark_must_follow_store() const = 0;
-
-  // If the CollectedHeap was asked to defer a store barrier above,
-  // this informs it to flush such a deferred store barrier to the
-  // remembered set.
-  virtual void flush_deferred_store_barrier(JavaThread* thread);
-
   // Perform a collection of the heap; intended for use in implementing
   // "System.gc".  This probably implies as full a collection as the
   // "CollectedHeap" supports.

src/hotspot/share/gc/shared/genCollectedHeap.cpp

@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2000, 2017, Oracle and/or its affiliates. All rights reserved.
+ * Copyright (c) 2000, 2018, Oracle and/or its affiliates. All rights reserved.
  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
  *
  * This code is free software; you can redistribute it and/or modify it
@@ -71,8 +71,6 @@ GenCollectedHeap::GenCollectedHeap(GenCollectorPolicy *policy) :
 }
 
 jint GenCollectedHeap::initialize() {
-  CollectedHeap::pre_initialize();
-
   // While there are no constraints in the GC code that HeapWordSize
   // be any particular value, there are multiple other areas in the
   // system which believe this to be true (e.g. oop->object_size in some

src/hotspot/share/gc/shared/genCollectedHeap.hpp

@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2000, 2017, Oracle and/or its affiliates. All rights reserved.
+ * Copyright (c) 2000, 2018, Oracle and/or its affiliates. All rights reserved.
  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
  *
  * This code is free software; you can redistribute it and/or modify it
@@ -270,22 +270,6 @@ public:
   virtual size_t unsafe_max_tlab_alloc(Thread* thr) const;
   virtual HeapWord* allocate_new_tlab(size_t size);
 
-  // Can a compiler initialize a new object without store barriers?
-  // This permission only extends from the creation of a new object
-  // via a TLAB up to the first subsequent safepoint.
-  virtual bool can_elide_tlab_store_barriers() const {
-    return true;
-  }
-
-  // We don't need barriers for stores to objects in the
-  // young gen and, a fortiori, for initializing stores to
-  // objects therein. This applies to DefNew+Tenured and ParNew+CMS
-  // only and may need to be re-examined in case other
-  // kinds of collectors are implemented in the future.
-  virtual bool can_elide_initializing_store_barrier(oop new_obj) {
-    return is_in_young(new_obj);
-  }
-
   // The "requestor" generation is performing some garbage collection
   // action for which it would be useful to have scratch space.  The
   // requestor promises to allocate no more than "max_alloc_words" in any

src/hotspot/share/gc/shared/modRefBarrierSet.hpp

@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2000, 2017, Oracle and/or its affiliates. All rights reserved.
+ * Copyright (c) 2000, 2018, Oracle and/or its affiliates. All rights reserved.
  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
  *
  * This code is free software; you can redistribute it and/or modify it
@@ -45,6 +45,7 @@ public:
 
   // Causes all refs in "mr" to be assumed to be modified.
   virtual void invalidate(MemRegion mr) = 0;
+  virtual void write_region(MemRegion mr) = 0;
 
   // The caller guarantees that "mr" contains no references.  (Perhaps it's
   // objects have been moved elsewhere.)

src/hotspot/share/jvmci/jvmciRuntime.cpp

@@ -116,10 +116,7 @@ JRT_BLOCK_ENTRY(void, JVMCIRuntime::new_instance(JavaThread* thread, Klass* klas
   oop obj = ik->allocate_instance(CHECK);
   thread->set_vm_result(obj);
   JRT_BLOCK_END;
-
+  SharedRuntime::on_slowpath_allocation_exit(thread);
-  if (ReduceInitialCardMarks) {
-    new_store_pre_barrier(thread);
-  }
 JRT_END
 
 JRT_BLOCK_ENTRY(void, JVMCIRuntime::new_array(JavaThread* thread, Klass* array_klass, jint length))
@@ -151,29 +148,9 @@ JRT_BLOCK_ENTRY(void, JVMCIRuntime::new_array(JavaThread* thread, Klass* array_k
     }
   }
   JRT_BLOCK_END;
-
+  SharedRuntime::on_slowpath_allocation_exit(thread);
-  if (ReduceInitialCardMarks) {
-    new_store_pre_barrier(thread);
-  }
 JRT_END
 
-void JVMCIRuntime::new_store_pre_barrier(JavaThread* thread) {
-  // After any safepoint, just before going back to compiled code,
-  // we inform the GC that we will be doing initializing writes to
-  // this object in the future without emitting card-marks, so
-  // GC may take any compensating steps.
-  // NOTE: Keep this code consistent with GraphKit::store_barrier.
-
-  oop new_obj = thread->vm_result();
-  if (new_obj == NULL)  return;
-
-  assert(Universe::heap()->can_elide_tlab_store_barriers(),
-         "compiler must check this first");
-  // GC may decide to give back a safer copy of new_obj.
-  new_obj = Universe::heap()->new_store_pre_barrier(thread, new_obj);
-  thread->set_vm_result(new_obj);
-}
-
 JRT_ENTRY(void, JVMCIRuntime::new_multi_array(JavaThread* thread, Klass* klass, int rank, jint* dims))
   assert(klass->is_klass(), "not a class");
   assert(rank >= 1, "rank must be nonzero");

src/hotspot/share/jvmci/jvmciRuntime.hpp

@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2012, 2017, Oracle and/or its affiliates. All rights reserved.
+ * Copyright (c) 2012, 2018, Oracle and/or its affiliates. All rights reserved.
  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
  *
  * This code is free software; you can redistribute it and/or modify it
@@ -154,7 +154,6 @@ class JVMCIRuntime: public AllStatic {
   static void write_barrier_pre(JavaThread* thread, oopDesc* obj);
   static void write_barrier_post(JavaThread* thread, void* card);
   static jboolean validate_object(JavaThread* thread, oopDesc* parent, oopDesc* child);
-  static void new_store_pre_barrier(JavaThread* thread);
 
   // used to throw exceptions from compiled JVMCI code
   static void throw_and_post_jvmti_exception(JavaThread* thread, const char* exception, const char* message);

src/hotspot/share/opto/graphKit.cpp

@@ -3861,7 +3861,7 @@ void GraphKit::write_barrier_post(Node* oop_store,
   if (use_ReduceInitialCardMarks()
       && obj == just_allocated_object(control())) {
     // We can skip marks on a freshly-allocated object in Eden.
-    // Keep this code in sync with new_store_pre_barrier() in runtime.cpp.
+    // Keep this code in sync with new_deferred_store_barrier() in runtime.cpp.
     // That routine informs GC to take appropriate compensating steps,
     // upon a slow-path allocation, so as to make this card-mark
     // elision safe.
@@ -4159,7 +4159,7 @@ void GraphKit::g1_write_barrier_pre(bool do_load,
  * as part of the allocation in the case the allocated object is not located
  * in the nursery, this would happen for humongous objects. This is similar to
  * how CMS is required to handle this case, see the comments for the method
- * CollectedHeap::new_store_pre_barrier and OptoRuntime::new_store_pre_barrier.
+ * CardTableModRefBS::on_allocation_slowpath_exit and OptoRuntime::new_deferred_store_barrier.
  * A deferred card mark is required for these objects and handled in the above
  * mentioned methods.
  *
@@ -4249,7 +4249,7 @@ void GraphKit::g1_write_barrier_post(Node* oop_store,
 
   if (use_ReduceInitialCardMarks() && obj == just_allocated_object(control())) {
     // We can skip marks on a freshly-allocated object in Eden.
-    // Keep this code in sync with new_store_pre_barrier() in runtime.cpp.
+    // Keep this code in sync with new_deferred_store_barrier() in runtime.cpp.
     // That routine informs GC to take appropriate compensating steps,
     // upon a slow-path allocation, so as to make this card-mark
     // elision safe.

src/hotspot/share/opto/graphKit.hpp

@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2001, 2015, Oracle and/or its affiliates. All rights reserved.
+ * Copyright (c) 2001, 2018, Oracle and/or its affiliates. All rights reserved.
  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
  *
  * This code is free software; you can redistribute it and/or modify it
@@ -755,8 +755,10 @@ class GraphKit : public Phase {
   Node* just_allocated_object(Node* current_control);
 
   static bool use_ReduceInitialCardMarks() {
-    return (ReduceInitialCardMarks
+    BarrierSet *bs = Universe::heap()->barrier_set();
-            && Universe::heap()->can_elide_tlab_store_barriers());
+    return bs->is_a(BarrierSet::CardTableModRef)
+           && barrier_set_cast<CardTableModRefBS>(bs)->can_elide_tlab_store_barriers()
+           && ReduceInitialCardMarks;
   }
 
   // Sync Ideal and Graph kits.

src/hotspot/share/opto/runtime.cpp

@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 1998, 2017, Oracle and/or its affiliates. All rights reserved.
+ * Copyright (c) 1998, 2018, Oracle and/or its affiliates. All rights reserved.
  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
  *
  * This code is free software; you can redistribute it and/or modify it
@@ -194,23 +194,6 @@ const char* OptoRuntime::stub_name(address entry) {
 // We failed the fast-path allocation.  Now we need to do a scavenge or GC
 // and try allocation again.
 
-void OptoRuntime::new_store_pre_barrier(JavaThread* thread) {
-  // After any safepoint, just before going back to compiled code,
-  // we inform the GC that we will be doing initializing writes to
-  // this object in the future without emitting card-marks, so
-  // GC may take any compensating steps.
-  // NOTE: Keep this code consistent with GraphKit::store_barrier.
-
-  oop new_obj = thread->vm_result();
-  if (new_obj == NULL)  return;
-
-  assert(Universe::heap()->can_elide_tlab_store_barriers(),
-         "compiler must check this first");
-  // GC may decide to give back a safer copy of new_obj.
-  new_obj = Universe::heap()->new_store_pre_barrier(thread, new_obj);
-  thread->set_vm_result(new_obj);
-}
-
 // object allocation
 JRT_BLOCK_ENTRY(void, OptoRuntime::new_instance_C(Klass* klass, JavaThread* thread))
   JRT_BLOCK;
@@ -244,10 +227,8 @@ JRT_BLOCK_ENTRY(void, OptoRuntime::new_instance_C(Klass* klass, JavaThread* thre
   deoptimize_caller_frame(thread, HAS_PENDING_EXCEPTION);
   JRT_BLOCK_END;
 
-  if (GraphKit::use_ReduceInitialCardMarks()) {
+  // inform GC that we won't do card marks for initializing writes.
-    // inform GC that we won't do card marks for initializing writes.
+  SharedRuntime::on_slowpath_allocation_exit(thread);
-    new_store_pre_barrier(thread);
-  }
 JRT_END
 
 
@@ -284,10 +265,8 @@ JRT_BLOCK_ENTRY(void, OptoRuntime::new_array_C(Klass* array_type, int len, JavaT
   thread->set_vm_result(result);
   JRT_BLOCK_END;
 
-  if (GraphKit::use_ReduceInitialCardMarks()) {
+  // inform GC that we won't do card marks for initializing writes.
-    // inform GC that we won't do card marks for initializing writes.
+  SharedRuntime::on_slowpath_allocation_exit(thread);
-    new_store_pre_barrier(thread);
-  }
 JRT_END
 
 // array allocation without zeroing
@@ -314,10 +293,9 @@ JRT_BLOCK_ENTRY(void, OptoRuntime::new_array_nozero_C(Klass* array_type, int len
   thread->set_vm_result(result);
   JRT_BLOCK_END;
 
-  if (GraphKit::use_ReduceInitialCardMarks()) {
+
-    // inform GC that we won't do card marks for initializing writes.
+  // inform GC that we won't do card marks for initializing writes.
-    new_store_pre_barrier(thread);
+  SharedRuntime::on_slowpath_allocation_exit(thread);
-  }
 
   oop result = thread->vm_result();
   if ((len > 0) && (result != NULL) &&

src/hotspot/share/opto/runtime.hpp

@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 1998, 2015, Oracle and/or its affiliates. All rights reserved.
+ * Copyright (c) 1998, 2018, Oracle and/or its affiliates. All rights reserved.
  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
  *
  * This code is free software; you can redistribute it and/or modify it
@@ -163,10 +163,6 @@ class OptoRuntime : public AllStatic {
   static void new_array_C(Klass* array_klass, int len, JavaThread *thread);
   static void new_array_nozero_C(Klass* array_klass, int len, JavaThread *thread);
 
-  // Post-slow-path-allocation, pre-initializing-stores step for
-  // implementing ReduceInitialCardMarks
-  static void new_store_pre_barrier(JavaThread* thread);
-
   // Allocate storage for a multi-dimensional arrays
   // Note: needs to be fixed for arbitrary number of dimensions
   static void multianewarray2_C(Klass* klass, int len1, int len2, JavaThread *thread);

src/hotspot/share/runtime/sharedRuntime.cpp

@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 1997, 2017, Oracle and/or its affiliates. All rights reserved.
+ * Copyright (c) 1997, 2018, Oracle and/or its affiliates. All rights reserved.
  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
  *
  * This code is free software; you can redistribute it and/or modify it
@@ -3169,3 +3169,16 @@ frame SharedRuntime::look_for_reserved_stack_annotated_method(JavaThread* thread
   }
   return activation;
 }
+
+void SharedRuntime::on_slowpath_allocation_exit(JavaThread* thread) {
+  // After any safepoint, just before going back to compiled code,
+  // we inform the GC that we will be doing initializing writes to
+  // this object in the future without emitting card-marks, so
+  // GC may take any compensating steps.
+
+  oop new_obj = thread->vm_result();
+  if (new_obj == NULL) return;
+
+  BarrierSet *bs = Universe::heap()->barrier_set();
+  bs->on_slowpath_allocation_exit(thread, new_obj);
+}

src/hotspot/share/runtime/sharedRuntime.hpp

@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 1997, 2017, Oracle and/or its affiliates. All rights reserved.
+ * Copyright (c) 1997, 2018, Oracle and/or its affiliates. All rights reserved.
  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
  *
  * This code is free software; you can redistribute it and/or modify it
@@ -211,6 +211,10 @@ class SharedRuntime: AllStatic {
   static address deoptimize_for_implicit_exception(JavaThread* thread, address pc, CompiledMethod* nm, int deopt_reason);
 #endif
 
+  // Post-slow-path-allocation, pre-initializing-stores step for
+  // implementing e.g. ReduceInitialCardMarks
+  static void on_slowpath_allocation_exit(JavaThread* thread);
+
   static void enable_stack_reserved_zone(JavaThread* thread);
   static frame look_for_reserved_stack_annotated_method(JavaThread* thread, frame fr);
 

src/hotspot/share/runtime/thread.cpp

@@ -1994,20 +1994,10 @@ void JavaThread::exit(bool destroy_vm, ExitType exit_type) {
     JvmtiExport::cleanup_thread(this);
   }
 
-  // We must flush any deferred card marks before removing a thread from
+  // We must flush any deferred card marks and other various GC barrier
-  // the list of active threads.
+  // related buffers (e.g. G1 SATB buffer and G1 dirty card queue buffer)
-  Universe::heap()->flush_deferred_store_barrier(this);
+  // before removing a thread from the list of active threads.
-  assert(deferred_card_mark().is_empty(), "Should have been flushed");
+  BarrierSet::barrier_set()->flush_deferred_barriers(this);
-
-#if INCLUDE_ALL_GCS
-  // We must flush the G1-related buffers before removing a thread
-  // from the list of active threads. We must do this after any deferred
-  // card marks have been flushed (above) so that any entries that are
-  // added to the thread's dirty card queue as a result are not lost.
-  if (UseG1GC) {
-    flush_barrier_queues();
-  }
-#endif // INCLUDE_ALL_GCS
 
   log_info(os, thread)("JavaThread %s (tid: " UINTX_FORMAT ").",
     exit_type == JavaThread::normal_exit ? "exiting" : "detaching",

src/hotspot/share/runtime/vmStructs.cpp

@@ -466,6 +466,7 @@ typedef PaddedEnd<ObjectMonitor>              PaddedObjectMonitor;
   nonstatic_field(CardGeneration,              _capacity_at_prologue,                         size_t)                                \
   nonstatic_field(CardGeneration,              _used_at_prologue,                             size_t)                                \
                                                                                                                                      \
+  nonstatic_field(CardTableModRefBS,           _defer_initial_card_mark,                      bool)                                  \
   nonstatic_field(CardTableModRefBS,           _whole_heap,                                   const MemRegion)                       \
   nonstatic_field(CardTableModRefBS,           _guard_index,                                  const size_t)                          \
   nonstatic_field(CardTableModRefBS,           _last_valid_index,                             const size_t)                          \
@@ -482,7 +483,6 @@ typedef PaddedEnd<ObjectMonitor>              PaddedObjectMonitor;
                                                                                                                                      \
   nonstatic_field(CollectedHeap,               _reserved,                                     MemRegion)                             \
   nonstatic_field(CollectedHeap,               _barrier_set,                                  BarrierSet*)                           \
-  nonstatic_field(CollectedHeap,               _defer_initial_card_mark,                      bool)                                  \
   nonstatic_field(CollectedHeap,               _is_gc_active,                                 bool)                                  \
   nonstatic_field(CollectedHeap,               _total_collections,                            unsigned int)                          \
                                                                                                                                      \