8248314: Parallel: Parallelize parallel full gc Adjust Roots phase

Reviewed-by: tschatzl, iwalulya

Author: albertnetymk

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

@@ -57,7 +57,7 @@
 #include "gc/shared/referenceProcessorPhaseTimes.hpp"
 #include "gc/shared/spaceDecorator.inline.hpp"
 #include "gc/shared/taskTerminator.hpp"
-#include "gc/shared/weakProcessor.hpp"
+#include "gc/shared/weakProcessor.inline.hpp"
 #include "gc/shared/workerPolicy.hpp"
 #include "gc/shared/workgroup.hpp"
 #include "logging/log.hpp"
@@ -780,7 +780,7 @@ bool ParallelCompactData::summarize(SplitInfo& split_info,
   return true;
 }
 
-HeapWord* ParallelCompactData::calc_new_pointer(HeapWord* addr, ParCompactionManager* cm) {
+HeapWord* ParallelCompactData::calc_new_pointer(HeapWord* addr, ParCompactionManager* cm) const {
   assert(addr != NULL, "Should detect NULL oop earlier");
   assert(ParallelScavengeHeap::heap()->is_in(addr), "not in heap");
   assert(PSParallelCompact::mark_bitmap()->is_marked(addr), "not marked");
@@ -1788,8 +1788,6 @@ bool PSParallelCompact::invoke_no_policy(bool maximum_heap_compaction) {
     ParCompactionManager::manager_array(ParallelScavengeHeap::heap()->workers().total_workers());
 
   {
-    ResourceMark rm;
-
     const uint active_workers =
       WorkerPolicy::calc_active_workers(ParallelScavengeHeap::heap()->workers().total_workers(),
                                         ParallelScavengeHeap::heap()->workers().active_workers(),
@@ -1834,7 +1832,7 @@ bool PSParallelCompact::invoke_no_policy(bool maximum_heap_compaction) {
 
     // adjust_roots() updates Universe::_intArrayKlassObj which is
     // needed by the compaction for filling holes in the dense prefix.
-    adjust_roots(vmthread_cm);
+    adjust_roots();
 
     compaction_start.update();
     compact();
@@ -2209,35 +2207,81 @@ void PSParallelCompact::marking_phase(ParCompactionManager* cm,
   _gc_tracer.report_object_count_after_gc(is_alive_closure());
 }
 
-void PSParallelCompact::adjust_roots(ParCompactionManager* cm) {
-  // Adjust the pointers to reflect the new locations
-  GCTraceTime(Info, gc, phases) tm("Adjust Roots", &_gc_timer);
+class PSAdjustTask final : public AbstractGangTask {
+  SubTasksDone                               _sub_tasks;
+  WeakProcessor::Task                        _weak_proc_task;
+  OopStorageSetStrongParState<false, false>  _oop_storage_iter;
+  uint                                       _nworkers;
 
-  // Need new claim bits when tracing through and adjusting pointers.
-  ClassLoaderDataGraph::clear_claimed_marks();
+  enum PSAdjustSubTask {
+    PSAdjustSubTask_code_cache,
+    PSAdjustSubTask_aot,
+    PSAdjustSubTask_old_ref_process,
+    PSAdjustSubTask_young_ref_process,
 
-  PCAdjustPointerClosure oop_closure(cm);
+    PSAdjustSubTask_num_elements
+  };
 
-  // General strong roots.
-  Threads::oops_do(&oop_closure, NULL);
-  OopStorageSet::strong_oops_do(&oop_closure);
-  CLDToOopClosure cld_closure(&oop_closure, ClassLoaderData::_claim_strong);
-  ClassLoaderDataGraph::cld_do(&cld_closure);
+public:
+  PSAdjustTask(uint nworkers) :
+    AbstractGangTask("PSAdjust task"),
+    _sub_tasks(PSAdjustSubTask_num_elements),
+    _weak_proc_task(nworkers),
+    _nworkers(nworkers) {
+    // Need new claim bits when tracing through and adjusting pointers.
+    ClassLoaderDataGraph::clear_claimed_marks();
+    if (nworkers > 1) {
+      Threads::change_thread_claim_token();
+    }
+  }
 
-  // Now adjust pointers in remaining weak roots.  (All of which should
-  // have been cleared if they pointed to non-surviving objects.)
-  WeakProcessor::oops_do(&oop_closure);
+  ~PSAdjustTask() {
+    Threads::assert_all_threads_claimed();
+  }
 
-  CodeBlobToOopClosure adjust_from_blobs(&oop_closure, CodeBlobToOopClosure::FixRelocations);
-  CodeCache::blobs_do(&adjust_from_blobs);
-  AOT_ONLY(AOTLoader::oops_do(&oop_closure);)
+  void work(uint worker_id) {
+    ParCompactionManager* cm = ParCompactionManager::gc_thread_compaction_manager(worker_id);
+    PCAdjustPointerClosure adjust(cm);
+    {
+      ResourceMark rm;
+      Threads::possibly_parallel_oops_do(_nworkers > 1, &adjust, nullptr);
+    }
+    _oop_storage_iter.oops_do(&adjust);
+    {
+      CLDToOopClosure cld_closure(&adjust, ClassLoaderData::_claim_strong);
+      ClassLoaderDataGraph::cld_do(&cld_closure);
+    }
+    {
+      AlwaysTrueClosure always_alive;
+      _weak_proc_task.work(worker_id, &always_alive, &adjust);
+    }
+    if (_sub_tasks.try_claim_task(PSAdjustSubTask_code_cache)) {
+      CodeBlobToOopClosure adjust_code(&adjust, CodeBlobToOopClosure::FixRelocations);
+      CodeCache::blobs_do(&adjust_code);
+    }
+    if (_sub_tasks.try_claim_task(PSAdjustSubTask_aot)) {
+      AOT_ONLY(AOTLoader::oops_do(&adjust);)
+    }
+    if (_sub_tasks.try_claim_task(PSAdjustSubTask_old_ref_process)) {
+      PSParallelCompact::ref_processor()->weak_oops_do(&adjust);
+    }
+    if (_sub_tasks.try_claim_task(PSAdjustSubTask_young_ref_process)) {
+      // Roots were visited so references into the young gen in roots
+      // may have been scanned.  Process them also.
+      // Should the reference processor have a span that excludes
+      // young gen objects?
+      PSScavenge::reference_processor()->weak_oops_do(&adjust);
+    }
+    _sub_tasks.all_tasks_claimed();
+  }
+};
 
-  ref_processor()->weak_oops_do(&oop_closure);
-  // Roots were visited so references into the young gen in roots
-  // may have been scanned.  Process them also.
-  // Should the reference processor have a span that excludes
-  // young gen objects?
-  PSScavenge::reference_processor()->weak_oops_do(&oop_closure);
+void PSParallelCompact::adjust_roots() {
+  // Adjust the pointers to reflect the new locations
+  GCTraceTime(Info, gc, phases) tm("Adjust Roots", &_gc_timer);
+  uint nworkers = ParallelScavengeHeap::heap()->workers().active_workers();
+  PSAdjustTask task(nworkers);
+  ParallelScavengeHeap::heap()->workers().run_task(&task);
 }
 
 // Helper class to print 8 region numbers per line and then print the total at the end.
@@ -2306,7 +2350,7 @@ void PSParallelCompact::prepare_region_draining_tasks(uint parallel_gc_threads)
 
     for (size_t cur = end_region - 1; cur + 1 > beg_region; --cur) {
       if (sd.region(cur)->claim_unsafe()) {
-        ParCompactionManager* cm = ParCompactionManager::manager_array(worker_id);
+        ParCompactionManager* cm = ParCompactionManager::gc_thread_compaction_manager(worker_id);
         bool result = sd.region(cur)->mark_normal();
         assert(result, "Must succeed at this point.");
         cm->region_stack()->push(cur);
@@ -2505,7 +2549,6 @@ static void compaction_with_stealing_work(TaskTerminator* terminator, uint worke
       // Go around again.
     }
   }
-  return;
 }
 
 class UpdateDensePrefixAndCompactionTask: public AbstractGangTask {
@@ -3133,7 +3176,7 @@ void PSParallelCompact::initialize_shadow_regions(uint parallel_gc_threads)
 
   size_t beg_region = sd.addr_to_region_idx(_space_info[old_space_id].dense_prefix());
   for (uint i = 0; i < parallel_gc_threads; i++) {
-    ParCompactionManager *cm = ParCompactionManager::manager_array(i);
+    ParCompactionManager *cm = ParCompactionManager::gc_thread_compaction_manager(i);
     cm->set_next_shadow_region(beg_region + i);
   }
 }

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

@@ -480,9 +480,9 @@ class ParallelCompactData
   HeapWord* partial_obj_end(size_t region_idx) const;
 
   // Return the location of the object after compaction.
-  HeapWord* calc_new_pointer(HeapWord* addr, ParCompactionManager* cm);
+  HeapWord* calc_new_pointer(HeapWord* addr, ParCompactionManager* cm) const;
 
-  HeapWord* calc_new_pointer(oop p, ParCompactionManager* cm) {
+  HeapWord* calc_new_pointer(oop p, ParCompactionManager* cm) const {
     return calc_new_pointer(cast_from_oop<HeapWord*>(p), cm);
   }
 
@@ -1107,7 +1107,7 @@ class PSParallelCompact : AllStatic {
   static void summary_phase(ParCompactionManager* cm, bool maximum_compaction);
 
   // Adjust addresses in roots.  Does not adjust addresses in heap.
-  static void adjust_roots(ParCompactionManager* cm);
+  static void adjust_roots();
 
   DEBUG_ONLY(static void write_block_fill_histogram();)
 

src/hotspot/share/gc/parallel/psParallelCompact.inline.hpp

@@ -113,10 +113,9 @@ inline void PSParallelCompact::adjust_pointer(T* p, ParCompactionManager* cm) {
     assert(ParallelScavengeHeap::heap()->is_in(obj), "should be in heap");
 
     oop new_obj = (oop)summary_data().calc_new_pointer(obj, cm);
-    assert(new_obj != NULL,                    // is forwarding ptr?
-           "should be forwarded");
-    // Just always do the update unconditionally?
-    if (new_obj != NULL) {
+    assert(new_obj != NULL, "non-null address for live objects");
+    // Is it actually relocated at all?
+    if (new_obj != obj) {
       assert(ParallelScavengeHeap::heap()->is_in_reserved(new_obj),
              "should be in object space");
       RawAccess<IS_NOT_NULL>::oop_store(p, new_obj);