--- a/src/share/vm/gc_implementation/g1/g1StringDedupTable.cpp	Mon Jul 04 14:29:23 2016 -0700
+++ b/src/share/vm/gc_implementation/g1/g1StringDedupTable.cpp	Fri Jul 08 13:59:32 2016 +0100
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2014, Oracle and/or its affiliates. All rights reserved.
+ * Copyright (c) 2014, 2016, 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
@@ -28,22 +28,23 @@
 #include "gc_implementation/g1/g1CollectedHeap.inline.hpp"
 #include "gc_implementation/g1/g1SATBCardTableModRefBS.hpp"
 #include "gc_implementation/g1/g1StringDedupTable.hpp"
+#include "gc_implementation/shared/concurrentGCThread.hpp"
 #include "memory/gcLocker.hpp"
 #include "memory/padded.inline.hpp"
 #include "oops/typeArrayOop.hpp"
 #include "runtime/mutexLocker.hpp"
 
 //
-// Freelist in the deduplication table entry cache. Links table
+// List of deduplication table entries. Links table
 // entries together using their _next fields.
 //
-class G1StringDedupEntryFreeList : public CHeapObj<mtGC> {
+class G1StringDedupEntryList : public CHeapObj<mtGC> {
 private:
   G1StringDedupEntry* _list;
   size_t              _length;
 
 public:
-  G1StringDedupEntryFreeList() :
+  G1StringDedupEntryList() :
     _list(NULL),
     _length(0) {
   }
@@ -63,6 +64,12 @@
     return entry;
   }
 
+  G1StringDedupEntry* remove_all() {
+    G1StringDedupEntry* list = _list;
+    _list = NULL;
+    return list;
+  }
+
   size_t length() {
     return _length;
   }
@@ -84,43 +91,53 @@
 //
 class G1StringDedupEntryCache : public CHeapObj<mtGC> {
 private:
-  // One freelist per GC worker to allow lock less freeing of
-  // entries while doing a parallel scan of the table. Using
-  // PaddedEnd to avoid false sharing.
-  PaddedEnd<G1StringDedupEntryFreeList>* _lists;
-  size_t                                 _nlists;
+  // One cache/overflow list per GC worker to allow lock less freeing of
+  // entries while doing a parallel scan of the table. Using PaddedEnd to
+  // avoid false sharing.
+  size_t                             _nlists;
+  size_t                             _max_list_length;
+  PaddedEnd<G1StringDedupEntryList>* _cached;
+  PaddedEnd<G1StringDedupEntryList>* _overflowed;
 
 public:
-  G1StringDedupEntryCache();
+  G1StringDedupEntryCache(size_t max_size);
   ~G1StringDedupEntryCache();
 
-  // Get a table entry from the cache freelist, or allocate a new
-  // entry if the cache is empty.
+  // Set max number of table entries to cache.
+  void set_max_size(size_t max_size);
+
+  // Get a table entry from the cache, or allocate a new entry if the cache is empty.
   G1StringDedupEntry* alloc();
 
-  // Insert a table entry into the cache freelist.
+  // Insert a table entry into the cache.
   void free(G1StringDedupEntry* entry, uint worker_id);
 
   // Returns current number of entries in the cache.
   size_t size();
 
-  // If the cache has grown above the given max size, trim it down
-  // and deallocate the memory occupied by trimmed of entries.
-  void trim(size_t max_size);
+  // Deletes overflowed entries.
+  void delete_overflowed();
 };
 
-G1StringDedupEntryCache::G1StringDedupEntryCache() {
-  _nlists = MAX2(ParallelGCThreads, (size_t)1);
-  _lists = PaddedArray<G1StringDedupEntryFreeList, mtGC>::create_unfreeable((uint)_nlists);
+G1StringDedupEntryCache::G1StringDedupEntryCache(size_t max_size) :
+  _nlists(MAX2(ParallelGCThreads, (size_t)1)),
+  _max_list_length(0),
+  _cached(PaddedArray<G1StringDedupEntryList, mtGC>::create_unfreeable((uint)_nlists)),
+  _overflowed(PaddedArray<G1StringDedupEntryList, mtGC>::create_unfreeable((uint)_nlists)) {
+  set_max_size(max_size);
 }
 
 G1StringDedupEntryCache::~G1StringDedupEntryCache() {
   ShouldNotReachHere();
 }
 
+void G1StringDedupEntryCache::set_max_size(size_t size) {
+  _max_list_length = size / _nlists;
+}
+
 G1StringDedupEntry* G1StringDedupEntryCache::alloc() {
   for (size_t i = 0; i < _nlists; i++) {
-    G1StringDedupEntry* entry = _lists[i].remove();
+    G1StringDedupEntry* entry = _cached[i].remove();
     if (entry != NULL) {
       return entry;
     }
@@ -131,31 +148,55 @@
 void G1StringDedupEntryCache::free(G1StringDedupEntry* entry, uint worker_id) {
   assert(entry->obj() != NULL, "Double free");
   assert(worker_id < _nlists, "Invalid worker id");
+
   entry->set_obj(NULL);
   entry->set_hash(0);
-  _lists[worker_id].add(entry);
+
+  if (_cached[worker_id].length() < _max_list_length) {
+    // Cache is not full
+    _cached[worker_id].add(entry);
+  } else {
+    // Cache is full, add to overflow list for later deletion
+    _overflowed[worker_id].add(entry);
+  }
 }
 
 size_t G1StringDedupEntryCache::size() {
   size_t size = 0;
   for (size_t i = 0; i < _nlists; i++) {
-    size += _lists[i].length();
+    size += _cached[i].length();
   }
   return size;
 }
 
-void G1StringDedupEntryCache::trim(size_t max_size) {
-  size_t cache_size = 0;
+void G1StringDedupEntryCache::delete_overflowed() {
+  double start = os::elapsedTime();
+  uintx count = 0;
+
   for (size_t i = 0; i < _nlists; i++) {
-    G1StringDedupEntryFreeList* list = &_lists[i];
-    cache_size += list->length();
-    while (cache_size > max_size) {
-      G1StringDedupEntry* entry = list->remove();
-      assert(entry != NULL, "Should not be null");
-      cache_size--;
+    G1StringDedupEntry* entry;
+
+    {
+      // The overflow list can be modified during safepoints, therefore
+      // we temporarily join the suspendible thread set while removing
+      // all entries from the list.
+      SuspendibleThreadSetJoiner sts_join;
+      entry = _overflowed[i].remove_all();
+    }
+
+    // Delete all entries
+    while (entry != NULL) {
+      G1StringDedupEntry* next = entry->next();
       delete entry;
+      entry = next;
+      count++;
     }
   }
+
+  double end = os::elapsedTime();
+  if (PrintStringDeduplicationStatistics) {
+    gclog_or_tty->print_cr("[GC concurrent-string-deduplication, deleted " UINTX_FORMAT " entries, " G1_STRDEDUP_TIME_FORMAT "]", count, end - start);
+  }
 }
 
 G1StringDedupTable*      G1StringDedupTable::_table = NULL;
@@ -192,7 +233,7 @@
 
 void G1StringDedupTable::create() {
   assert(_table == NULL, "One string deduplication table allowed");
-  _entry_cache = new G1StringDedupEntryCache();
+  _entry_cache = new G1StringDedupEntryCache((size_t)(_min_size * _max_cache_factor));
   _table = new G1StringDedupTable(_min_size);
 }
 
@@ -375,6 +416,9 @@
   // Update statistics
   _resize_count++;
 
+  // Update max cache size
+  _entry_cache->set_max_size((size_t)(size * _max_cache_factor));
+
   // Allocate the new table. The new table will be populated by workers
   // calling unlink_or_oops_do() and finally installed by finish_resize().
   return new G1StringDedupTable(size, _table->_hash_seed);
@@ -427,7 +471,7 @@
     removed += unlink_or_oops_do(cl, table_half + partition_begin, table_half + partition_end, worker_id);
   }
 
-  // Delayed update avoid contention on the table lock
+  // Delayed update to avoid contention on the table lock
   if (removed > 0) {
     MutexLockerEx ml(StringDedupTable_lock, Mutex::_no_safepoint_check_flag);
     _table->_entries -= removed;
@@ -545,10 +589,8 @@
   }
 }
 
-void G1StringDedupTable::trim_entry_cache() {
-  MutexLockerEx ml(StringDedupTable_lock, Mutex::_no_safepoint_check_flag);
-  size_t max_cache_size = (size_t)(_table->_size * _max_cache_factor);
-  _entry_cache->trim(max_cache_size);
+void G1StringDedupTable::clean_entry_cache() {
+  _entry_cache->delete_overflowed();
 }
 
 void G1StringDedupTable::print_statistics(outputStream* st) {

--- a/src/share/vm/gc_implementation/g1/g1StringDedupTable.hpp	Mon Jul 04 14:29:23 2016 -0700
+++ b/src/share/vm/gc_implementation/g1/g1StringDedupTable.hpp	Fri Jul 08 13:59:32 2016 +0100
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2014, Oracle and/or its affiliates. All rights reserved.
+ * Copyright (c) 2014, 2016, 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
@@ -218,8 +218,8 @@
   // and deletes the previously active table.
   static void finish_rehash(G1StringDedupTable* rehashed_table);
 
-  // If the table entry cache has grown too large, trim it down according to policy
-  static void trim_entry_cache();
+  // If the table entry cache has grown too large, delete overflowed entries.
+  static void clean_entry_cache();
 
   static void unlink_or_oops_do(G1StringDedupUnlinkOrOopsDoClosure* cl, uint worker_id);
 

--- a/src/share/vm/gc_implementation/g1/g1StringDedupThread.cpp	Mon Jul 04 14:29:23 2016 -0700
+++ b/src/share/vm/gc_implementation/g1/g1StringDedupThread.cpp	Fri Jul 08 13:59:32 2016 +0100
@@ -100,14 +100,14 @@
         }
       }
 
-      G1StringDedupTable::trim_entry_cache();
-
       stat.mark_done();
 
       // Print statistics
       total_stat.add(stat);
       print(gclog_or_tty, stat, total_stat);
     }
+
+    G1StringDedupTable::clean_entry_cache();
   }
 
   terminate();

--- a/src/share/vm/opto/stringopts.cpp	Mon Jul 04 14:29:23 2016 -0700
+++ b/src/share/vm/opto/stringopts.cpp	Fri Jul 08 13:59:32 2016 +0100
@@ -1641,16 +1641,11 @@
     }
     kit.store_String_value(kit.control(), result, char_array);
 
-    // Do not let stores that initialize this object be reordered with
-    // a subsequent store that would make this object accessible by
-    // other threads.
-    // Record what AllocateNode this StoreStore protects so that
-    // escape analysis can go from the MemBarStoreStoreNode to the
-    // AllocateNode and eliminate the MemBarStoreStoreNode if possible
-    // based on the escape status of the AllocateNode.
-    AllocateNode* alloc = AllocateNode::Ideal_allocation(result, _gvn);
-    assert(alloc != NULL, "should be newly allocated");
-    kit.insert_mem_bar(Op_MemBarStoreStore, alloc->proj_out(AllocateNode::RawAddress));
+    // The value field is final. Emit a barrier here to ensure that the effect
+    // of the initialization is committed to memory before any code publishes
+    // a reference to the newly constructed object (see Parse::do_exits()).
+    assert(AllocateNode::Ideal_allocation(result, _gvn) != NULL, "should be newly allocated");
+    kit.insert_mem_bar(Op_MemBarRelease, result);
   } else {
     result = C->top();
   }

--- a/test/compiler/jsr292/VMAnonymousClasses.java	Mon Jul 04 14:29:23 2016 -0700
+++ b/test/compiler/jsr292/VMAnonymousClasses.java	Fri Jul 08 13:59:32 2016 +0100
@@ -24,7 +24,7 @@
 /**
  * @test
  * @bug 8058828
- * @run main/bootclasspath -Xbatch VMAnonymousClasses
+ * @run main/bootclasspath/othervm -Xbatch VMAnonymousClasses
  */
 
 import jdk.internal.org.objectweb.asm.ClassWriter;

--- a/test/compiler/stringopts/TestStringObjectInitialization.java	Mon Jul 04 14:29:23 2016 -0700
+++ b/test/compiler/stringopts/TestStringObjectInitialization.java	Fri Jul 08 13:59:32 2016 +0100
@@ -27,6 +27,7 @@
 /*
  * @test
  * @bug 8159244
+ * @requires vm.gc == "Parallel" | vm.gc == "null"
  * @summary Verifies that no partially initialized String object escapes from
  *          C2's String concat optimization in a highly concurrent setting.
  *          This test triggers the bug in about 1 out of 10 runs.