File _service:tar_scm:8205921.patch of Package openjdk-1.8.0

From 99b38421c0da651ded6d5bd4d39ce1a7f6675188 Mon Sep 17 00:00:00 2001
Date: Fri, 22 Jan 2021 16:14:06 +0800
Subject: 8205921: Optimizing best-of-2 work stealing queue
 selection

Summary: <gc>: Improve gc performance, optimizing best-of-2 work stealing queue selection
LLT: jtreg
Bug url: https://bugs.openjdk.java.net/browse/JDK-8205921
---
 .../concurrentMarkSweepGeneration.cpp         |  29 ++---
 .../concurrentMarkSweepGeneration.hpp         |   3 -
 .../gc_implementation/g1/concurrentMark.cpp   |   4 +-
 .../gc_implementation/g1/concurrentMark.hpp   |   6 +-
 .../g1/g1ParScanThreadState.cpp               |   2 +-
 .../g1/g1ParScanThreadState.hpp               |   2 -
 .../g1/g1ParScanThreadState.inline.hpp        |   2 +-
 .../parNew/parNewGeneration.cpp               |   2 -
 .../parNew/parNewGeneration.hpp               |   2 -
 .../parallelScavenge/pcTasks.cpp              |   8 +-
 .../parallelScavenge/psCompactionManager.hpp  |  12 +-
 .../parallelScavenge/psPromotionManager.hpp   |   4 +-
 .../parallelScavenge/psTasks.cpp              |   3 +-
 hotspot/src/share/vm/utilities/taskqueue.cpp  |  18 ---
 hotspot/src/share/vm/utilities/taskqueue.hpp  | 105 ++++++++++++++----
 16 files changed, 112 insertions(+), 90 deletions(-)

diff --git a/hotspot/src/share/vm/gc_implementation/concurrentMarkSweep/concurrentMarkSweepGeneration.cpp b/hotspot/src/share/vm/gc_implementation/concurrentMarkSweep/concurrentMarkSweepGeneration.cpp
index 02a29c2b0..53b75a4ca 100644
--- a/hotspot/src/share/vm/gc_implementation/concurrentMarkSweep/concurrentMarkSweepGeneration.cpp
+++ b/hotspot/src/share/vm/gc_implementation/concurrentMarkSweep/concurrentMarkSweepGeneration.cpp
@@ -680,11 +680,6 @@ CMSCollector::CMSCollector(ConcurrentMarkSweepGeneration* cmsGen,
         warning("task_queues allocation failure.");
         return;
       }
-      _hash_seed = NEW_C_HEAP_ARRAY(int, num_queues, mtGC);
-      if (_hash_seed == NULL) {
-        warning("_hash_seed array allocation failure");
-        return;
-      }
 
       typedef Padded<OopTaskQueue> PaddedOopTaskQueue;
       for (i = 0; i < num_queues; i++) {
@@ -697,7 +692,6 @@ CMSCollector::CMSCollector(ConcurrentMarkSweepGeneration* cmsGen,
       }
       for (i = 0; i < num_queues; i++) {
         _task_queues->queue(i)->initialize();
-        _hash_seed[i] = 17;  // copied from ParNew
       }
     }
   }
@@ -4391,7 +4385,6 @@ void CMSConcMarkingTask::do_work_steal(int i) {
   oop obj_to_scan;
   CMSBitMap* bm = &(_collector->_markBitMap);
   CMSMarkStack* ovflw = &(_collector->_markStack);
-  int* seed = _collector->hash_seed(i);
   Par_ConcMarkingClosure cl(_collector, this, work_q, bm, ovflw);
   while (true) {
     cl.trim_queue(0);
@@ -4401,7 +4394,7 @@ void CMSConcMarkingTask::do_work_steal(int i) {
       // overflow stack may already have been stolen from us.
       // assert(work_q->size() > 0, "Work from overflow stack");
       continue;
-    } else if (task_queues()->steal(i, seed, /* reference */ obj_to_scan)) {
+    } else if (task_queues()->steal(i, /* reference */ obj_to_scan)) {
       assert(obj_to_scan->is_oop(), "Should be an oop");
       assert(bm->isMarked((HeapWord*)obj_to_scan), "Grey object");
       obj_to_scan->oop_iterate(&cl);
@@ -5373,7 +5366,7 @@ class CMSParRemarkTask: public CMSParMarkTask {
                                   Par_MarkRefsIntoAndScanClosure* cl);
 
   // ... work stealing for the above
-  void do_work_steal(int i, Par_MarkRefsIntoAndScanClosure* cl, int* seed);
+  void do_work_steal(int i, Par_MarkRefsIntoAndScanClosure* cl);
 };
 
 class RemarkKlassClosure : public KlassClosure {
@@ -5539,7 +5532,7 @@ void CMSParRemarkTask::work(uint worker_id) {
   // ---------- ... and drain overflow list.
   _timer.reset();
   _timer.start();
-  do_work_steal(worker_id, &par_mrias_cl, _collector->hash_seed(worker_id));
+  do_work_steal(worker_id, &par_mrias_cl);
   _timer.stop();
   if (PrintCMSStatistics != 0) {
     gclog_or_tty->print_cr(
@@ -5696,8 +5689,7 @@ CMSParRemarkTask::do_dirty_card_rescan_tasks(
 
 // . see if we can share work_queues with ParNew? XXX
 void
-CMSParRemarkTask::do_work_steal(int i, Par_MarkRefsIntoAndScanClosure* cl,
-                                int* seed) {
+CMSParRemarkTask::do_work_steal(int i, Par_MarkRefsIntoAndScanClosure* cl) {
   OopTaskQueue* work_q = work_queue(i);
   NOT_PRODUCT(int num_steals = 0;)
   oop obj_to_scan;
@@ -5728,7 +5720,7 @@ CMSParRemarkTask::do_work_steal(int i, Par_MarkRefsIntoAndScanClosure* cl,
     // Verify that we have no work before we resort to stealing
     assert(work_q->size() == 0, "Have work, shouldn't steal");
     // Try to steal from other queues that have work
-    if (task_queues()->steal(i, seed, /* reference */ obj_to_scan)) {
+    if (task_queues()->steal(i,/* reference */ obj_to_scan)) {
       NOT_PRODUCT(num_steals++;)
       assert(obj_to_scan->is_oop(), "Oops, not an oop!");
       assert(bm->isMarked((HeapWord*)obj_to_scan), "Stole an unmarked oop?");
@@ -6141,8 +6133,7 @@ public:
 
   void do_work_steal(int i,
                      CMSParDrainMarkingStackClosure* drain,
-                     CMSParKeepAliveClosure* keep_alive,
-                     int* seed);
+                     CMSParKeepAliveClosure* keep_alive);
 
   virtual void work(uint worker_id);
 };
@@ -6160,8 +6151,7 @@ void CMSRefProcTaskProxy::work(uint worker_id) {
   CMSIsAliveClosure is_alive_closure(_span, _mark_bit_map);
   _task.work(worker_id, is_alive_closure, par_keep_alive, par_drain_stack);
   if (_task.marks_oops_alive()) {
-    do_work_steal(worker_id, &par_drain_stack, &par_keep_alive,
-                  _collector->hash_seed(worker_id));
+    do_work_steal(worker_id, &par_drain_stack, &par_keep_alive);
   }
   assert(work_queue(worker_id)->size() == 0, "work_queue should be empty");
   assert(_collector->_overflow_list == NULL, "non-empty _overflow_list");
@@ -6196,8 +6186,7 @@ CMSParKeepAliveClosure::CMSParKeepAliveClosure(CMSCollector* collector,
 // . see if we can share work_queues with ParNew? XXX
 void CMSRefProcTaskProxy::do_work_steal(int i,
   CMSParDrainMarkingStackClosure* drain,
-  CMSParKeepAliveClosure* keep_alive,
-  int* seed) {
+  CMSParKeepAliveClosure* keep_alive) {
   OopTaskQueue* work_q = work_queue(i);
   NOT_PRODUCT(int num_steals = 0;)
   oop obj_to_scan;
@@ -6226,7 +6215,7 @@ void CMSRefProcTaskProxy::do_work_steal(int i,
     // Verify that we have no work before we resort to stealing
     assert(work_q->size() == 0, "Have work, shouldn't steal");
     // Try to steal from other queues that have work
-    if (task_queues()->steal(i, seed, /* reference */ obj_to_scan)) {
+    if (task_queues()->steal(i, /* reference */ obj_to_scan)) {
       NOT_PRODUCT(num_steals++;)
       assert(obj_to_scan->is_oop(), "Oops, not an oop!");
       assert(_mark_bit_map->isMarked((HeapWord*)obj_to_scan), "Stole an unmarked oop?");
diff --git a/hotspot/src/share/vm/gc_implementation/concurrentMarkSweep/concurrentMarkSweepGeneration.hpp b/hotspot/src/share/vm/gc_implementation/concurrentMarkSweep/concurrentMarkSweepGeneration.hpp
index 8b65d3426..ca3fee21b 100644
--- a/hotspot/src/share/vm/gc_implementation/concurrentMarkSweep/concurrentMarkSweepGeneration.hpp
+++ b/hotspot/src/share/vm/gc_implementation/concurrentMarkSweep/concurrentMarkSweepGeneration.hpp
@@ -563,8 +563,6 @@ class CMSCollector: public CHeapObj<mtGC> {
   Stack<oop, mtGC>     _preserved_oop_stack;
   Stack<markOop, mtGC> _preserved_mark_stack;
 
-  int*             _hash_seed;
-
   // In support of multi-threaded concurrent phases
   YieldingFlexibleWorkGang* _conc_workers;
 
@@ -741,7 +739,6 @@ class CMSCollector: public CHeapObj<mtGC> {
   bool stop_world_and_do(CMS_op_type op);
 
   OopTaskQueueSet* task_queues() { return _task_queues; }
-  int*             hash_seed(int i) { return &_hash_seed[i]; }
   YieldingFlexibleWorkGang* conc_workers() { return _conc_workers; }
 
   // Support for parallelizing Eden rescan in CMS remark phase
diff --git a/hotspot/src/share/vm/gc_implementation/g1/concurrentMark.cpp b/hotspot/src/share/vm/gc_implementation/g1/concurrentMark.cpp
index f7b64a61b..d782c892d 100644
--- a/hotspot/src/share/vm/gc_implementation/g1/concurrentMark.cpp
+++ b/hotspot/src/share/vm/gc_implementation/g1/concurrentMark.cpp
@@ -4431,7 +4431,7 @@ void CMTask::do_marking_step(double time_target_ms,
       oop obj;
       statsOnly( ++_steal_attempts );
 
-      if (_cm->try_stealing(_worker_id, &_hash_seed, obj)) {
+      if (_cm->try_stealing(_worker_id, obj)) {
         if (_cm->verbose_medium()) {
           gclog_or_tty->print_cr("[%u] stolen " PTR_FORMAT " successfully",
                                  _worker_id, p2i((void*) obj));
@@ -4614,7 +4614,7 @@ CMTask::CMTask(uint worker_id,
     _worker_id(worker_id), _cm(cm),
     _objArray_processor(this),
     _claimed(false),
-    _nextMarkBitMap(NULL), _hash_seed(17),
+    _nextMarkBitMap(NULL),
     _task_queue(task_queue),
     _task_queues(task_queues),
     _cm_oop_closure(NULL),
diff --git a/hotspot/src/share/vm/gc_implementation/g1/concurrentMark.hpp b/hotspot/src/share/vm/gc_implementation/g1/concurrentMark.hpp
index c22c9b601..3404be2a3 100644
--- a/hotspot/src/share/vm/gc_implementation/g1/concurrentMark.hpp
+++ b/hotspot/src/share/vm/gc_implementation/g1/concurrentMark.hpp
@@ -680,8 +680,8 @@ public:
   }
 
   // Attempts to steal an object from the task queues of other tasks
-  bool try_stealing(uint worker_id, int* hash_seed, oop& obj) {
-    return _task_queues->steal(worker_id, hash_seed, obj);
+  bool try_stealing(uint worker_id, oop& obj) {
+    return _task_queues->steal(worker_id, obj);
   }
 
   ConcurrentMark(G1CollectedHeap* g1h,
@@ -1007,8 +1007,6 @@ private:
   // it was decreased).
   size_t                      _real_refs_reached_limit;
 
-  // used by the work stealing stuff
-  int                         _hash_seed;
   // if this is true, then the task has aborted for some reason
   bool                        _has_aborted;
   // set when the task aborts because it has met its time quota
diff --git a/hotspot/src/share/vm/gc_implementation/g1/g1ParScanThreadState.cpp b/hotspot/src/share/vm/gc_implementation/g1/g1ParScanThreadState.cpp
index e765620bc..394f20e82 100644
--- a/hotspot/src/share/vm/gc_implementation/g1/g1ParScanThreadState.cpp
+++ b/hotspot/src/share/vm/gc_implementation/g1/g1ParScanThreadState.cpp
@@ -36,7 +36,7 @@ G1ParScanThreadState::G1ParScanThreadState(G1CollectedHeap* g1h, uint queue_num,
     _dcq(&g1h->dirty_card_queue_set()),
     _ct_bs(g1h->g1_barrier_set()),
     _g1_rem(g1h->g1_rem_set()),
-    _hash_seed(17), _queue_num(queue_num),
+    _queue_num(queue_num),
     _term_attempts(0),
     _tenuring_threshold(g1h->g1_policy()->tenuring_threshold()),
     _age_table(false), _scanner(g1h, rp),
diff --git a/hotspot/src/share/vm/gc_implementation/g1/g1ParScanThreadState.hpp b/hotspot/src/share/vm/gc_implementation/g1/g1ParScanThreadState.hpp
index d9403aa43..990b71d31 100644
--- a/hotspot/src/share/vm/gc_implementation/g1/g1ParScanThreadState.hpp
+++ b/hotspot/src/share/vm/gc_implementation/g1/g1ParScanThreadState.hpp
@@ -59,7 +59,6 @@ class G1ParScanThreadState : public StackObj {
 
   OopsInHeapRegionClosure*      _evac_failure_cl;
 
-  int  _hash_seed;
   uint _queue_num;
 
   size_t _term_attempts;
@@ -129,7 +128,6 @@ class G1ParScanThreadState : public StackObj {
 
   OopsInHeapRegionClosure* evac_failure_closure() { return _evac_failure_cl; }
 
-  int* hash_seed() { return &_hash_seed; }
   uint queue_num() { return _queue_num; }
 
   size_t term_attempts() const  { return _term_attempts; }
diff --git a/hotspot/src/share/vm/gc_implementation/g1/g1ParScanThreadState.inline.hpp b/hotspot/src/share/vm/gc_implementation/g1/g1ParScanThreadState.inline.hpp
index 1b03f8caa..7dedb1517 100644
--- a/hotspot/src/share/vm/gc_implementation/g1/g1ParScanThreadState.inline.hpp
+++ b/hotspot/src/share/vm/gc_implementation/g1/g1ParScanThreadState.inline.hpp
@@ -131,7 +131,7 @@ inline void G1ParScanThreadState::dispatch_reference(StarTask ref) {
 
 void G1ParScanThreadState::steal_and_trim_queue(RefToScanQueueSet *task_queues) {
   StarTask stolen_task;
-  while (task_queues->steal(queue_num(), hash_seed(), stolen_task)) {
+  while (task_queues->steal(queue_num(), stolen_task)) {
     assert(verify_task(stolen_task), "sanity");
     dispatch_reference(stolen_task);
 
diff --git a/hotspot/src/share/vm/gc_implementation/parNew/parNewGeneration.cpp b/hotspot/src/share/vm/gc_implementation/parNew/parNewGeneration.cpp
index 84cd4ed73..c07e9b812 100644
--- a/hotspot/src/share/vm/gc_implementation/parNew/parNewGeneration.cpp
+++ b/hotspot/src/share/vm/gc_implementation/parNew/parNewGeneration.cpp
@@ -92,7 +92,6 @@ ParScanThreadState::ParScanThreadState(Space* to_space_,
 
   _survivor_chunk_array =
     (ChunkArray*) old_gen()->get_data_recorder(thread_num());
-  _hash_seed = 17;  // Might want to take time-based random value.
   _start = os::elapsedTime();
   _old_gen_closure.set_generation(old_gen_);
   _old_gen_root_closure.set_generation(old_gen_);
@@ -560,7 +559,6 @@ void ParEvacuateFollowersClosure::do_void() {
 
     // attempt to steal work from promoted.
     if (task_queues()->steal(par_scan_state()->thread_num(),
-                             par_scan_state()->hash_seed(),
                              obj_to_scan)) {
       bool res = work_q->push(obj_to_scan);
       assert(res, "Empty queue should have room for a push.");
diff --git a/hotspot/src/share/vm/gc_implementation/parNew/parNewGeneration.hpp b/hotspot/src/share/vm/gc_implementation/parNew/parNewGeneration.hpp
index fa4265a2d..ea527fdb5 100644
--- a/hotspot/src/share/vm/gc_implementation/parNew/parNewGeneration.hpp
+++ b/hotspot/src/share/vm/gc_implementation/parNew/parNewGeneration.hpp
@@ -95,7 +95,6 @@ class ParScanThreadState {
 
   HeapWord *_young_old_boundary;
 
-  int _hash_seed;
   int _thread_num;
   ageTable _ageTable;
 
@@ -161,7 +160,6 @@ class ParScanThreadState {
   // Is new_obj a candidate for scan_partial_array_and_push_remainder method.
   inline bool should_be_partially_scanned(oop new_obj, oop old_obj) const;
 
-  int* hash_seed()  { return &_hash_seed; }
   int  thread_num() { return _thread_num; }
 
   // Allocate a to-space block of size "sz", or else return NULL.
diff --git a/hotspot/src/share/vm/gc_implementation/parallelScavenge/pcTasks.cpp b/hotspot/src/share/vm/gc_implementation/parallelScavenge/pcTasks.cpp
index 35ea2992b..37610f3d1 100644
--- a/hotspot/src/share/vm/gc_implementation/parallelScavenge/pcTasks.cpp
+++ b/hotspot/src/share/vm/gc_implementation/parallelScavenge/pcTasks.cpp
@@ -217,14 +217,13 @@ void StealMarkingTask::do_it(GCTaskManager* manager, uint which) {
 
   oop obj = NULL;
   ObjArrayTask task;
-  int random_seed = 17;
   do {
-    while (ParCompactionManager::steal_objarray(which, &random_seed, task)) {
+    while (ParCompactionManager::steal_objarray(which, task)) {
       ObjArrayKlass* k = (ObjArrayKlass*)task.obj()->klass();
       k->oop_follow_contents(cm, task.obj(), task.index());
       cm->follow_marking_stacks();
     }
-    while (ParCompactionManager::steal(which, &random_seed, obj)) {
+    while (ParCompactionManager::steal(which, obj)) {
       obj->follow_contents(cm);
       cm->follow_marking_stacks();
     }
@@ -280,13 +279,12 @@ void StealRegionCompactionTask::do_it(GCTaskManager* manager, uint which) {
   cm->drain_region_stacks();
 
   size_t region_index = 0;
-  int random_seed = 17;
 
   // If we're the termination task, try 10 rounds of stealing before
   // setting the termination flag
 
   while(true) {
-    if (ParCompactionManager::steal(which, &random_seed, region_index)) {
+    if (ParCompactionManager::steal(which, region_index)) {
       PSParallelCompact::fill_and_update_region(cm, region_index);
       cm->drain_region_stacks();
     } else {
diff --git a/hotspot/src/share/vm/gc_implementation/parallelScavenge/psCompactionManager.hpp b/hotspot/src/share/vm/gc_implementation/parallelScavenge/psCompactionManager.hpp
index 7d7a9f495..a16a16762 100644
--- a/hotspot/src/share/vm/gc_implementation/parallelScavenge/psCompactionManager.hpp
+++ b/hotspot/src/share/vm/gc_implementation/parallelScavenge/psCompactionManager.hpp
@@ -177,16 +177,16 @@ private:
   // Access function for compaction managers
   static ParCompactionManager* gc_thread_compaction_manager(int index);
 
-  static bool steal(int queue_num, int* seed, oop& t) {
-    return stack_array()->steal(queue_num, seed, t);
+  static bool steal(int queue_num, oop& t) {
+    return stack_array()->steal(queue_num, t);
   }
 
-  static bool steal_objarray(int queue_num, int* seed, ObjArrayTask& t) {
-    return _objarray_queues->steal(queue_num, seed, t);
+  static bool steal_objarray(int queue_num, ObjArrayTask& t) {
+    return _objarray_queues->steal(queue_num, t);
   }
 
-  static bool steal(int queue_num, int* seed, size_t& region) {
-    return region_array()->steal(queue_num, seed, region);
+  static bool steal(int queue_num, size_t& region) {
+    return region_array()->steal(queue_num, region);
   }
 
   // Process tasks remaining on any marking stack
diff --git a/hotspot/src/share/vm/gc_implementation/parallelScavenge/psPromotionManager.hpp b/hotspot/src/share/vm/gc_implementation/parallelScavenge/psPromotionManager.hpp
index 6eef954b7..542d86943 100644
--- a/hotspot/src/share/vm/gc_implementation/parallelScavenge/psPromotionManager.hpp
+++ b/hotspot/src/share/vm/gc_implementation/parallelScavenge/psPromotionManager.hpp
@@ -164,8 +164,8 @@ class PSPromotionManager VALUE_OBJ_CLASS_SPEC {
   static PSPromotionManager* gc_thread_promotion_manager(int index);
   static PSPromotionManager* vm_thread_promotion_manager();
 
-  static bool steal_depth(int queue_num, int* seed, StarTask& t) {
-    return stack_array_depth()->steal(queue_num, seed, t);
+  static bool steal_depth(int queue_num, StarTask& t) {
+    return stack_array_depth()->steal(queue_num, t);
   }
 
   PSPromotionManager();
diff --git a/hotspot/src/share/vm/gc_implementation/parallelScavenge/psTasks.cpp b/hotspot/src/share/vm/gc_implementation/parallelScavenge/psTasks.cpp
index f829e9344..4fe869fd6 100644
--- a/hotspot/src/share/vm/gc_implementation/parallelScavenge/psTasks.cpp
+++ b/hotspot/src/share/vm/gc_implementation/parallelScavenge/psTasks.cpp
@@ -151,10 +151,9 @@ void StealTask::do_it(GCTaskManager* manager, uint which) {
   guarantee(pm->stacks_empty(),
             "stacks should be empty at this point");
 
-  int random_seed = 17;
   while(true) {
     StarTask p;
-    if (PSPromotionManager::steal_depth(which, &random_seed, p)) {
+    if (PSPromotionManager::steal_depth(which, p)) {
       TASKQUEUE_STATS_ONLY(pm->record_steal(p));
       pm->process_popped_location_depth(p);
       pm->drain_stacks_depth(true);
diff --git a/hotspot/src/share/vm/utilities/taskqueue.cpp b/hotspot/src/share/vm/utilities/taskqueue.cpp
index 0f4dcc90b..37f4066ab 100644
--- a/hotspot/src/share/vm/utilities/taskqueue.cpp
+++ b/hotspot/src/share/vm/utilities/taskqueue.cpp
@@ -112,24 +112,6 @@ void TaskQueueStats::verify() const
 #endif // ASSERT
 #endif // TASKQUEUE_STATS
 
-int TaskQueueSetSuper::randomParkAndMiller(int *seed0) {
-  const int a =      16807;
-  const int m = 2147483647;
-  const int q =     127773;  /* m div a */
-  const int r =       2836;  /* m mod a */
-  assert(sizeof(int) == 4, "I think this relies on that");
-  int seed = *seed0;
-  int hi   = seed / q;
-  int lo   = seed % q;
-  int test = a * lo - r * hi;
-  if (test > 0)
-    seed = test;
-  else
-    seed = test + m;
-  *seed0 = seed;
-  return seed;
-}
-
 ParallelTaskTerminator::
 ParallelTaskTerminator(int n_threads, TaskQueueSetSuper* queue_set) :
   _n_threads(n_threads),
diff --git a/hotspot/src/share/vm/utilities/taskqueue.hpp b/hotspot/src/share/vm/utilities/taskqueue.hpp
index 0f1376b49..77556a7d4 100644
--- a/hotspot/src/share/vm/utilities/taskqueue.hpp
+++ b/hotspot/src/share/vm/utilities/taskqueue.hpp
@@ -309,12 +309,30 @@ public:
   void oops_do(OopClosure* f);
 
 private:
+  DEFINE_PAD_MINUS_SIZE(0, DEFAULT_CACHE_LINE_SIZE, 0);
   // Element array.
   volatile E* _elems;
+
+  DEFINE_PAD_MINUS_SIZE(1, DEFAULT_CACHE_LINE_SIZE, sizeof(E*));
+  // Queue owner local variables. Not to be accessed by other threads.
+
+  static const uint InvalidQueueId = uint(-1);
+  uint _last_stolen_queue_id; // The id of the queue we last stole from
+
+  int _seed; // Current random seed used for selecting a random queue during stealing.
+
+  DEFINE_PAD_MINUS_SIZE(2, DEFAULT_CACHE_LINE_SIZE, sizeof(uint) + sizeof(int));
+public:
+  int next_random_queue_id();
+
+  void set_last_stolen_queue_id(uint id)     { _last_stolen_queue_id = id; }
+  uint last_stolen_queue_id() const          { return _last_stolen_queue_id; }
+  bool is_last_stolen_queue_id_valid() const { return _last_stolen_queue_id != InvalidQueueId; }
+  void invalidate_last_stolen_queue_id()     { _last_stolen_queue_id = InvalidQueueId; }
 };
 
 template<class E, MEMFLAGS F, unsigned int N>
-GenericTaskQueue<E, F, N>::GenericTaskQueue() {
+GenericTaskQueue<E, F, N>::GenericTaskQueue() : _last_stolen_queue_id(InvalidQueueId), _seed(17 /* random number */) {
   assert(sizeof(Age) == sizeof(size_t), "Depends on this.");
 }
 
@@ -429,6 +447,30 @@ bool GenericTaskQueue<E, F, N>::pop_global(volatile E& t) {
   return resAge == oldAge;
 }
 
+inline int randomParkAndMiller(int *seed0) {
+  const int a =      16807;
+  const int m = 2147483647;
+  const int q =     127773;  /* m div a */
+  const int r =       2836;  /* m mod a */
+  STATIC_ASSERT(sizeof(int) == 4);
+  int seed = *seed0;
+  int hi   = seed / q;
+  int lo   = seed % q;
+  int test = a * lo - r * hi;
+  if (test > 0) {
+    seed = test;
+  } else {
+    seed = test + m;
+  }
+  *seed0 = seed;
+  return seed;
+}
+
+template<class E, MEMFLAGS F, unsigned int N>
+int GenericTaskQueue<E, F, N>::next_random_queue_id() {
+  return randomParkAndMiller(&_seed);
+}
+
 template<class E, MEMFLAGS F, unsigned int N>
 GenericTaskQueue<E, F, N>::~GenericTaskQueue() {}
 
@@ -498,8 +540,6 @@ bool OverflowTaskQueue<E, F, N>::try_push_to_taskqueue(E t) {
   return taskqueue_t::push(t);
 }
 class TaskQueueSetSuper {
-protected:
-  static int randomParkAndMiller(int* seed0);
 public:
   // Returns "true" if some TaskQueue in the set contains a task.
   virtual bool peek() = 0;
@@ -520,27 +560,23 @@ private:
 public:
   typedef typename T::element_type E;
 
-  GenericTaskQueueSet(int n) : _n(n) {
+  GenericTaskQueueSet(uint n) : _n(n) {
     typedef T* GenericTaskQueuePtr;
     _queues = NEW_C_HEAP_ARRAY(GenericTaskQueuePtr, n, F);
-    for (int i = 0; i < n; i++) {
+    for (uint i = 0; i < n; i++) {
       _queues[i] = NULL;
     }
   }
 
-  bool steal_best_of_2(uint queue_num, int* seed, E& t);
+  bool steal_best_of_2(uint queue_num, E& t);
 
   void register_queue(uint i, T* q);
 
   T* queue(uint n);
 
-  // The thread with queue number "queue_num" (and whose random number seed is
-  // at "seed") is trying to steal a task from some other queue.  (It may try
-  // several queues, according to some configuration parameter.)  If some steal
-  // succeeds, returns "true" and sets "t" to the stolen task, otherwise returns
-  // false.
-  bool steal(uint queue_num, int* seed, E& t);
-
+  // Try to steal a task from some other queue than queue_num. It may perform several attempts at doing so.
+  // Returns if stealing succeeds, and sets "t" to the stolen task.
+  bool steal(uint queue_num, E& t);
   bool peek();
   uint tasks() const;
   size_t tasks();
@@ -560,9 +596,9 @@ GenericTaskQueueSet<T, F>::queue(uint i) {
 }
 
 template<class T, MEMFLAGS F> bool
-GenericTaskQueueSet<T, F>::steal(uint queue_num, int* seed, E& t) {
+GenericTaskQueueSet<T, F>::steal(uint queue_num, E& t) {
   for (uint i = 0; i < 2 * _n; i++) {
-    if (steal_best_of_2(queue_num, seed, t)) {
+    if (steal_best_of_2(queue_num, t)) {
       TASKQUEUE_STATS_ONLY(queue(queue_num)->stats.record_steal(true));
       return true;
     }
@@ -572,17 +608,46 @@ GenericTaskQueueSet<T, F>::steal(uint queue_num, int* seed, E& t) {
 }
 
 template<class T, MEMFLAGS F> bool
-GenericTaskQueueSet<T, F>::steal_best_of_2(uint queue_num, int* seed, E& t) {
+GenericTaskQueueSet<T, F>::steal_best_of_2(uint queue_num, E& t) {
   if (_n > 2) {
+	T* const local_queue = _queues[queue_num];
     uint k1 = queue_num;
-    while (k1 == queue_num) k1 = TaskQueueSetSuper::randomParkAndMiller(seed) % _n;
+
+	if (local_queue->is_last_stolen_queue_id_valid()) {
+      k1 = local_queue->last_stolen_queue_id();
+      assert(k1 != queue_num, "Should not be the same");
+    } else {
+      while (k1 == queue_num) {
+        k1 = local_queue->next_random_queue_id() % _n;
+      }
+    }
+
     uint k2 = queue_num;
-    while (k2 == queue_num || k2 == k1) k2 = TaskQueueSetSuper::randomParkAndMiller(seed) % _n;
+	while (k2 == queue_num || k2 == k1) {
+	  k2 = local_queue->next_random_queue_id() % _n;
+	}
     // Sample both and try the larger.
     uint sz1 = _queues[k1]->size();
     uint sz2 = _queues[k2]->size();
-    if (sz2 > sz1) return _queues[k2]->pop_global(t);
-    else return _queues[k1]->pop_global(t);
+
+	uint sel_k = 0;
+    bool suc = false;
+
+    if (sz2 > sz1) {
+      sel_k = k2;
+      suc = _queues[k2]->pop_global(t);
+    } else if (sz1 > 0) {
+      sel_k = k1;
+      suc = _queues[k1]->pop_global(t);
+    }
+
+    if (suc) {
+      local_queue->set_last_stolen_queue_id(sel_k);
+    } else {
+      local_queue->invalidate_last_stolen_queue_id();
+    }
+
+    return suc;
   } else if (_n == 2) {
     // Just try the other one.
     uint k = (queue_num + 1) % 2;
-- 
2.19.0