--- a/src/hotspot/share/gc/g1/g1ConcurrentMark.cpp	Mon May 14 11:47:03 2018 +0200
+++ b/src/hotspot/share/gc/g1/g1ConcurrentMark.cpp	Mon May 14 11:47:03 2018 +0200
@@ -1675,6 +1675,44 @@
   }
 }
 
+class G1PrecleanYieldClosure : public YieldClosure {
+  G1ConcurrentMark* _cm;
+
+public:
+  G1PrecleanYieldClosure(G1ConcurrentMark* cm) : _cm(cm) { }
+
+  virtual bool should_return() {
+    return _cm->has_aborted();
+  }
+
+  virtual bool should_return_fine_grain() {
+    _cm->do_yield_check();
+    return _cm->has_aborted();
+  }
+};
+
+void G1ConcurrentMark::preclean() {
+  assert(G1UseReferencePrecleaning, "Precleaning must be enabled.");
+
+  SuspendibleThreadSetJoiner joiner;
+
+  G1CMKeepAliveAndDrainClosure keep_alive(this, task(0), true /* is_serial */);
+  G1CMDrainMarkingStackClosure drain_mark_stack(this, task(0), true /* is_serial */);
+
+  set_concurrency_and_phase(1, true);
+
+  G1PrecleanYieldClosure yield_cl(this);
+
+  ReferenceProcessor* rp = _g1h->ref_processor_cm();
+  // Precleaning is single threaded. Temporarily disable MT discovery.
+  ReferenceProcessorMTDiscoveryMutator rp_mut_discovery(rp, false);
+  rp->preclean_discovered_references(rp->is_alive_non_header(),
+                                     &keep_alive,
+                                     &drain_mark_stack,
+                                     &yield_cl,
+                                     _gc_timer_cm);
+}
+
 // When sampling object counts, we already swapped the mark bitmaps, so we need to use
 // the prev bitmap determining liveness.
 class G1ObjectCountIsAliveClosure: public BoolObjectClosure {

--- a/src/hotspot/share/gc/g1/g1ConcurrentMark.hpp	Mon May 14 11:47:03 2018 +0200
+++ b/src/hotspot/share/gc/g1/g1ConcurrentMark.hpp	Mon May 14 11:47:03 2018 +0200
@@ -563,6 +563,9 @@
   // Do concurrent phase of marking, to a tentative transitive closure.
   void mark_from_roots();
 
+  // Do concurrent preclean work.
+  void preclean();
+
   void remark();
 
   void cleanup();

--- a/src/hotspot/share/gc/g1/g1ConcurrentMarkThread.cpp	Mon May 14 11:47:03 2018 +0200
+++ b/src/hotspot/share/gc/g1/g1ConcurrentMarkThread.cpp	Mon May 14 11:47:03 2018 +0200
@@ -57,6 +57,7 @@
   expander(SCAN_ROOT_REGIONS,, "Concurrent Scan Root Regions")             \
   expander(CONCURRENT_MARK,, "Concurrent Mark")                            \
   expander(MARK_FROM_ROOTS,, "Concurrent Mark From Roots")                 \
+  expander(PRECLEAN,, "Concurrent Preclean")                               \
   expander(BEFORE_REMARK,, NULL)                                           \
   expander(REMARK,, NULL)                                                  \
   expander(REBUILD_REMEMBERED_SETS,, "Concurrent Rebuild Remembered Sets") \
@@ -309,7 +310,12 @@
             break;
           }
 
-          // Provide a control point after mark_from_roots.
+          if (G1UseReferencePrecleaning) {
+            G1ConcPhase p(G1ConcurrentPhase::PRECLEAN, this);
+            _cm->preclean();
+          }
+
+          // Provide a control point before remark.
           {
             G1ConcPhaseManager p(G1ConcurrentPhase::BEFORE_REMARK, this);
           }

--- a/src/hotspot/share/gc/g1/g1_globals.hpp	Mon May 14 11:47:03 2018 +0200
+++ b/src/hotspot/share/gc/g1/g1_globals.hpp	Mon May 14 11:47:03 2018 +0200
@@ -79,6 +79,10 @@
           "draining concurrent marking work queues.")                       \
           range(1, INT_MAX)                                                 \
                                                                             \
+  experimental(bool, G1UseReferencePrecleaning, true,                       \
+               "Concurrently preclean java.lang.ref.references instances "  \
+               "before the Remark pause.")                                  \
+                                                                            \
   experimental(double, G1LastPLABAverageOccupancy, 50.0,                    \
                "The expected average occupancy of the last PLAB in "        \
                "percent.")                                                  \

--- a/src/hotspot/share/gc/shared/referenceProcessor.cpp	Mon May 14 11:47:03 2018 +0200
+++ b/src/hotspot/share/gc/shared/referenceProcessor.cpp	Mon May 14 11:47:03 2018 +0200
@@ -594,19 +594,33 @@
   bool _clear_referent;
 };
 
+void ReferenceProcessor::log_reflist(const char* prefix, DiscoveredList list[], uint num_active_queues) {
+  LogTarget(Trace, gc, ref) lt;
+
+  if (!lt.is_enabled()) {
+    return;
+  }
+
+  size_t total = 0;
+
+  LogStream ls(lt);
+  ls.print("%s", prefix);
+  for (uint i = 0; i < num_active_queues; i++) {
+    ls.print(SIZE_FORMAT " ", list[i].length());
+    total += list[i].length();
+  }
+  ls.print_cr("(" SIZE_FORMAT ")", total);
+}
+
 #ifndef PRODUCT
-void ReferenceProcessor::log_reflist_counts(DiscoveredList ref_lists[], uint active_length, size_t total_refs) {
+void ReferenceProcessor::log_reflist_counts(DiscoveredList ref_lists[], uint num_active_queues) {
   if (!log_is_enabled(Trace, gc, ref)) {
     return;
   }
 
-  stringStream st;
-  for (uint i = 0; i < active_length; ++i) {
-    st.print(SIZE_FORMAT " ", ref_lists[i].length());
-  }
-  log_develop_trace(gc, ref)("%s= " SIZE_FORMAT, st.as_string(), total_refs);
+  log_reflist("", ref_lists, num_active_queues);
 #ifdef ASSERT
-  for (uint i = active_length; i < _max_num_queues; i++) {
+  for (uint i = num_active_queues; i < _max_num_queues; i++) {
     assert(ref_lists[i].length() == 0, SIZE_FORMAT " unexpected References in %u",
            ref_lists[i].length(), i);
   }
@@ -629,10 +643,11 @@
   size_t total_refs = 0;
   log_develop_trace(gc, ref)("Balance ref_lists ");
 
+  log_reflist_counts(ref_lists, _max_num_queues);
+
   for (uint i = 0; i < _max_num_queues; ++i) {
     total_refs += ref_lists[i].length();
   }
-  log_reflist_counts(ref_lists, _max_num_queues, total_refs);
   size_t avg_refs = total_refs / _num_queues + 1;
   uint to_idx = 0;
   for (uint from_idx = 0; from_idx < _max_num_queues; from_idx++) {
@@ -693,11 +708,11 @@
     }
   }
 #ifdef ASSERT
+  log_reflist_counts(ref_lists, _num_queues);
   size_t balanced_total_refs = 0;
   for (uint i = 0; i < _num_queues; ++i) {
     balanced_total_refs += ref_lists[i].length();
   }
-  log_reflist_counts(ref_lists, _num_queues, balanced_total_refs);
   assert(total_refs == balanced_total_refs, "Balancing was incomplete");
 #endif
 }
@@ -1011,63 +1026,79 @@
   return false;
 }
 
-// Preclean the discovered references by removing those
-// whose referents are alive, and by marking from those that
-// are not active. These lists can be handled here
-// in any order and, indeed, concurrently.
-void ReferenceProcessor::preclean_discovered_references(
-  BoolObjectClosure* is_alive,
-  OopClosure* keep_alive,
-  VoidClosure* complete_gc,
-  YieldClosure* yield,
-  GCTimer* gc_timer) {
+void ReferenceProcessor::preclean_discovered_references(BoolObjectClosure* is_alive,
+                                                        OopClosure* keep_alive,
+                                                        VoidClosure* complete_gc,
+                                                        YieldClosure* yield,
+                                                        GCTimer* gc_timer) {
+  // These lists can be handled here in any order and, indeed, concurrently.
 
   // Soft references
   {
     GCTraceTime(Debug, gc, ref) tm("Preclean SoftReferences", gc_timer);
+    log_reflist("SoftRef before: ", _discoveredSoftRefs, _max_num_queues);
     for (uint i = 0; i < _max_num_queues; i++) {
       if (yield->should_return()) {
         return;
       }
-      preclean_discovered_reflist(_discoveredSoftRefs[i], is_alive,
-                                  keep_alive, complete_gc, yield);
+      if (preclean_discovered_reflist(_discoveredSoftRefs[i], is_alive,
+                                      keep_alive, complete_gc, yield)) {
+        log_reflist("SoftRef abort: ", _discoveredSoftRefs, _max_num_queues);
+        return;
+      }
     }
+    log_reflist("SoftRef after: ", _discoveredSoftRefs, _max_num_queues);
   }
 
   // Weak references
   {
     GCTraceTime(Debug, gc, ref) tm("Preclean WeakReferences", gc_timer);
+    log_reflist("WeakRef before: ", _discoveredWeakRefs, _max_num_queues);
     for (uint i = 0; i < _max_num_queues; i++) {
       if (yield->should_return()) {
         return;
       }
-      preclean_discovered_reflist(_discoveredWeakRefs[i], is_alive,
-                                  keep_alive, complete_gc, yield);
+      if (preclean_discovered_reflist(_discoveredWeakRefs[i], is_alive,
+                                      keep_alive, complete_gc, yield)) {
+        log_reflist("WeakRef abort: ", _discoveredWeakRefs, _max_num_queues);
+        return;
+      }
     }
+    log_reflist("WeakRef after: ", _discoveredWeakRefs, _max_num_queues);
   }
 
   // Final references
   {
     GCTraceTime(Debug, gc, ref) tm("Preclean FinalReferences", gc_timer);
+    log_reflist("FinalRef before: ", _discoveredFinalRefs, _max_num_queues);
     for (uint i = 0; i < _max_num_queues; i++) {
       if (yield->should_return()) {
         return;
       }
-      preclean_discovered_reflist(_discoveredFinalRefs[i], is_alive,
-                                  keep_alive, complete_gc, yield);
+      if (preclean_discovered_reflist(_discoveredFinalRefs[i], is_alive,
+                                      keep_alive, complete_gc, yield)) {
+        log_reflist("FinalRef abort: ", _discoveredFinalRefs, _max_num_queues);
+        return;
+      }
     }
+    log_reflist("FinalRef after: ", _discoveredFinalRefs, _max_num_queues);
   }
 
   // Phantom references
   {
     GCTraceTime(Debug, gc, ref) tm("Preclean PhantomReferences", gc_timer);
+    log_reflist("PhantomRef before: ", _discoveredPhantomRefs, _max_num_queues);
     for (uint i = 0; i < _max_num_queues; i++) {
       if (yield->should_return()) {
         return;
       }
-      preclean_discovered_reflist(_discoveredPhantomRefs[i], is_alive,
-                                  keep_alive, complete_gc, yield);
+      if (preclean_discovered_reflist(_discoveredPhantomRefs[i], is_alive,
+                                      keep_alive, complete_gc, yield)) {
+        log_reflist("PhantomRef abort: ", _discoveredPhantomRefs, _max_num_queues);
+        return;
+      }
     }
+    log_reflist("PhantomRef after: ", _discoveredPhantomRefs, _max_num_queues);
   }
 }
 
@@ -1079,19 +1110,20 @@
 // java.lang.Reference. As a result, we need to be careful below
 // that ref removal steps interleave safely with ref discovery steps
 // (in this thread).
-void
-ReferenceProcessor::preclean_discovered_reflist(DiscoveredList&    refs_list,
-                                                BoolObjectClosure* is_alive,
-                                                OopClosure*        keep_alive,
-                                                VoidClosure*       complete_gc,
-                                                YieldClosure*      yield) {
+bool ReferenceProcessor::preclean_discovered_reflist(DiscoveredList&    refs_list,
+                                                     BoolObjectClosure* is_alive,
+                                                     OopClosure*        keep_alive,
+                                                     VoidClosure*       complete_gc,
+                                                     YieldClosure*      yield) {
   DiscoveredListIterator iter(refs_list, keep_alive, is_alive);
   while (iter.has_next()) {
+    if (yield->should_return_fine_grain()) {
+      return true;
+    }
     iter.load_ptrs(DEBUG_ONLY(true /* allow_null_referent */));
     oop obj = iter.obj();
     oop next = java_lang_ref_Reference::next(obj);
-    if (iter.referent() == NULL || iter.is_referent_alive() ||
-        next != NULL) {
+    if (iter.referent() == NULL || iter.is_referent_alive() || next != NULL) {
       // The referent has been cleared, or is alive, or the Reference is not
       // active; we need to trace and mark its cohort.
       log_develop_trace(gc, ref)("Precleaning Reference (" INTPTR_FORMAT ": %s)",
@@ -1121,6 +1153,7 @@
         iter.removed(), iter.processed(), p2i(&refs_list));
     }
   )
+  return false;
 }
 
 const char* ReferenceProcessor::list_name(uint i) {

--- a/src/hotspot/share/gc/shared/referenceProcessor.hpp	Mon May 14 11:47:03 2018 +0200
+++ b/src/hotspot/share/gc/shared/referenceProcessor.hpp	Mon May 14 11:47:03 2018 +0200
@@ -279,15 +279,15 @@
                       OopClosure*        keep_alive,
                       VoidClosure*       complete_gc);
 
-  // "Preclean" all the discovered reference lists
-  // by removing references with strongly reachable referents.
+  // "Preclean" all the discovered reference lists by removing references that
+  // are active (e.g. due to the mutator calling enqueue()) or with NULL or
+  // strongly reachable referents.
   // The first argument is a predicate on an oop that indicates
-  // its (strong) reachability and the second is a closure that
+  // its (strong) reachability and the fourth is a closure that
   // may be used to incrementalize or abort the precleaning process.
   // The caller is responsible for taking care of potential
   // interference with concurrent operations on these lists
-  // (or predicates involved) by other threads. Currently
-  // only used by the CMS collector.
+  // (or predicates involved) by other threads.
   void preclean_discovered_references(BoolObjectClosure* is_alive,
                                       OopClosure*        keep_alive,
                                       VoidClosure*       complete_gc,
@@ -298,15 +298,17 @@
   // occupying the i / _num_queues slot.
   const char* list_name(uint i);
 
-  // "Preclean" the given discovered reference list
-  // by removing references with strongly reachable referents.
-  // Currently used in support of CMS only.
-  void preclean_discovered_reflist(DiscoveredList&    refs_list,
+private:
+  // "Preclean" the given discovered reference list by removing references with
+  // the attributes mentioned in preclean_discovered_references().
+  // Supports both normal and fine grain yielding.
+  // Returns whether the operation should be aborted.
+  bool preclean_discovered_reflist(DiscoveredList&    refs_list,
                                    BoolObjectClosure* is_alive,
                                    OopClosure*        keep_alive,
                                    VoidClosure*       complete_gc,
                                    YieldClosure*      yield);
-private:
+
   // round-robin mod _num_queues (not: _not_ mod _max_num_queues)
   uint next_id() {
     uint id = _next_id;
@@ -323,7 +325,8 @@
 
   void clear_discovered_references(DiscoveredList& refs_list);
 
-  void log_reflist_counts(DiscoveredList ref_lists[], uint active_length, size_t total_count) PRODUCT_RETURN;
+  void log_reflist(const char* prefix, DiscoveredList list[], uint num_active_queues);
+  void log_reflist_counts(DiscoveredList ref_lists[], uint num_active_queues) PRODUCT_RETURN;
 
   // Balances reference queues.
   void balance_queues(DiscoveredList ref_lists[]);

--- a/src/hotspot/share/memory/iterator.hpp	Mon May 14 11:47:03 2018 +0200
+++ b/src/hotspot/share/memory/iterator.hpp	Mon May 14 11:47:03 2018 +0200
@@ -318,8 +318,11 @@
 // by means of checking the return value from the polling
 // call.
 class YieldClosure : public StackObj {
-  public:
-   virtual bool should_return() = 0;
+public:
+ virtual bool should_return() = 0;
+
+ // Yield on a fine-grain level. The check in case of not yielding should be very fast.
+ virtual bool should_return_fine_grain() { return false; }
 };
 
 // Abstract closure for serializing data (read or write).

--- a/test/hotspot/jtreg/gc/concurrent_phase_control/TestConcurrentPhaseControlG1.java	Mon May 14 11:47:03 2018 +0200
+++ b/test/hotspot/jtreg/gc/concurrent_phase_control/TestConcurrentPhaseControlG1.java	Mon May 14 11:47:03 2018 +0200
@@ -51,6 +51,7 @@
         {"CONCURRENT_MARK", "Concurrent Mark [^FR]"},
         {"IDLE", null},  // Resume IDLE before testing subphases
         {"MARK_FROM_ROOTS", "Concurrent Mark From Roots"},
+        {"PRECLEAN", "Concurrent Preclean"},
         {"BEFORE_REMARK", null},
         {"REMARK", "Pause Remark"},
         {"REBUILD_REMEMBERED_SETS", "Concurrent Rebuild Remembered Sets"},

--- a/test/hotspot/jtreg/gc/concurrent_phase_control/TestConcurrentPhaseControlG1Basics.java	Mon May 14 11:47:03 2018 +0200
+++ b/test/hotspot/jtreg/gc/concurrent_phase_control/TestConcurrentPhaseControlG1Basics.java	Mon May 14 11:47:03 2018 +0200
@@ -51,6 +51,7 @@
         "SCAN_ROOT_REGIONS",
         "CONCURRENT_MARK",
         "MARK_FROM_ROOTS",
+        "PRECLEAN",
         "BEFORE_REMARK",
         "REMARK",
         "REBUILD_REMEMBERED_SETS",