[GC] Add option G1BarrierSimple to use simple g1 post barrier

Summary: Provide option G1BarrierSimple to use simple G1 post barrier for better mutator performance

Testing: CI pipeline

Reviewers: yude, yifeng

Issue: dragonwell-project#828

src/hotspot/cpu/aarch64/gc/g1/g1BarrierSetAssembler_aarch64.cpp

@@ -87,6 +87,29 @@ void G1BarrierSetAssembler::gen_write_ref_array_pre_barrier(MacroAssembler* masm
 
 void G1BarrierSetAssembler::gen_write_ref_array_post_barrier(MacroAssembler* masm, DecoratorSet decorators,
                                                              Register start, Register count, Register scratch, RegSet saved_regs) {
+  if (G1BarrierSimple) {
+
+    Label L_loop, L_done;
+    const Register end = count;
+
+    __ cbz(count, L_done); // zero count - nothing to do
+
+    __ lea(end, Address(start, count, Address::lsl(LogBytesPerHeapOop))); // end = start + count << LogBytesPerHeapOop
+    __ sub(end, end, BytesPerHeapOop); // last element address to make inclusive
+    __ lsr(start, start, CardTable::card_shift);
+    __ lsr(end, end, CardTable::card_shift);
+    __ sub(count, end, start); // number of bytes to copy
+
+    __ load_byte_map_base(scratch);
+    __ add(start, start, scratch);
+    __ bind(L_loop);
+    __ strb(zr, Address(start, count));
+    __ subs(count, count, 1);
+    __ br(Assembler::GE, L_loop);
+    __ bind(L_done);
+    return;
+  }
+
   __ push(saved_regs, sp);
   assert_different_registers(start, count, scratch);
   assert_different_registers(c_rarg0, count);
@@ -209,6 +232,15 @@ void G1BarrierSetAssembler::g1_write_barrier_post(MacroAssembler* masm,
   Label done;
   Label runtime;
 
+  if (G1BarrierSimple) {
+    const Register card_addr = tmp;
+    __ lsr(card_addr, store_addr, CardTable::card_shift);
+    __ load_byte_map_base(tmp2);
+    __ add(card_addr, card_addr, tmp2);
+    __ strb(zr, Address(card_addr));
+    __ bind(done);
+    return;
+  }
   // Does store cross heap regions?
 
   __ eor(tmp, store_addr, new_val);
@@ -449,6 +481,13 @@ void G1BarrierSetAssembler::generate_c1_post_barrier_runtime_stub(StubAssembler*
   __ load_parameter(0, card_offset);
   __ lsr(card_offset, card_offset, CardTable::card_shift);
   __ load_byte_map_base(byte_map_base);
+
+  if (G1BarrierSimple) {
+    __ strb(zr, Address(byte_map_base, card_offset));
+    __ bind(done);
+    __ epilogue();
+    return;
+  }
   __ ldrb(rscratch1, Address(byte_map_base, card_offset));
   __ cmpw(rscratch1, (int)G1CardTable::g1_young_card_val());
   __ br(Assembler::EQ, done);

src/hotspot/cpu/x86/gc/g1/g1BarrierSetAssembler_x86.cpp

@@ -41,6 +41,8 @@
 
 #define __ masm->
 
+#define TIMES_OOP (UseCompressedOops ? Address::times_4 : Address::times_8)
+
 void G1BarrierSetAssembler::gen_write_ref_array_pre_barrier(MacroAssembler* masm, DecoratorSet decorators,
                                                             Register addr, Register count) {
   bool dest_uninitialized = (decorators & IS_DEST_UNINITIALIZED) != 0;
@@ -97,6 +99,35 @@ void G1BarrierSetAssembler::gen_write_ref_array_pre_barrier(MacroAssembler* masm
 
 void G1BarrierSetAssembler::gen_write_ref_array_post_barrier(MacroAssembler* masm, DecoratorSet decorators,
                                                              Register addr, Register count, Register tmp) {
+  if (G1BarrierSimple) {
+    CardTableBarrierSet* ct =
+      barrier_set_cast<CardTableBarrierSet>(BarrierSet::barrier_set());
+    intptr_t disp = (intptr_t) ct->card_table()->byte_map_base();
+
+    Label L_loop, L_done;
+    const Register end = count;
+    assert_different_registers(addr, end);
+
+    __ testl(count, count);
+    __ jcc(Assembler::zero, L_done); // zero count - nothing to do
+
+    __ leaq(end, Address(addr, count, TIMES_OOP, 0));  // end == addr+count*oop_size
+    __ subptr(end, BytesPerHeapOop); // end - 1 to make inclusive
+    __ shrptr(addr, CardTable::card_shift);
+    __ shrptr(end, CardTable::card_shift);
+    __ subptr(end, addr); // end --> cards count
+
+    __ mov64(tmp, disp);
+    __ addptr(addr, tmp);
+    __ bind(L_loop);
+    __ movb(Address(addr, count, Address::times_1), 0);
+    __ decrement(count);
+    __ jcc(Assembler::greaterEqual, L_loop);
+    __ bind(L_done);
+
+    return;
+  }
+
   __ pusha();             // push registers (overkill)
 #ifdef _LP64
   if (c_rarg0 == count) { // On win64 c_rarg0 == rcx
@@ -302,6 +333,12 @@ void G1BarrierSetAssembler::g1_write_barrier_post(MacroAssembler* masm,
   __ movptr(cardtable, (intptr_t)ct->card_table()->byte_map_base());
   __ addptr(card_addr, cardtable);
 
+  if (G1BarrierSimple) {
+    __ movb(Address(card_addr, 0), (int)G1CardTable::dirty_card_val());
+    __ bind(done);
+    return;
+  }
+
   __ cmpb(Address(card_addr, 0), (int)G1CardTable::g1_young_card_val());
   __ jcc(Assembler::equal, done);
 
@@ -551,6 +588,17 @@ void G1BarrierSetAssembler::generate_c1_post_barrier_runtime_stub(StubAssembler*
 
   NOT_LP64(__ get_thread(thread);)
 
+  if (G1BarrierSimple) {
+    __ movb(Address(card_addr, 0), (int)G1CardTable::dirty_card_val());
+    __ bind(done);
+    __ pop(rcx);
+    __ pop(rax);
+
+    __ epilogue();
+
+    return;
+  }
+
   __ cmpb(Address(card_addr, 0), (int)G1CardTable::g1_young_card_val());
   __ jcc(Assembler::equal, done);
 

src/hotspot/share/gc/g1/c2/g1BarrierSetC2.cpp

@@ -447,6 +447,13 @@ void G1BarrierSetC2::post_barrier(GraphKit* kit,
   // Combine card table base and card offset
   Node* card_adr = __ AddP(no_base, byte_map_base_node(kit), card_offset );
 
+  if (G1BarrierSimple) {
+    __ store(__ ctrl(), card_adr, dirty_card, T_BYTE, Compile::AliasIdxRaw, MemNode::unordered);
+    // Final sync IdealKit and GraphKit.
+    kit->final_sync(ideal);
+    return;
+  }
+
   // If we know the value being stored does it cross regions?
 
   if (val != NULL) {
@@ -658,6 +665,39 @@ bool G1BarrierSetC2::is_gc_barrier_node(Node* node) const {
 
 void G1BarrierSetC2::eliminate_gc_barrier(PhaseMacroExpand* macro, Node* node) const {
   assert(node->Opcode() == Op_CastP2X, "ConvP2XNode required");
+
+  if (G1BarrierSimple) {
+    Node* this_region = node->in(0);
+    // Search "if (marking != 0)" check and set it to "false".
+    // There is no G1 pre barrier if previous stored value is NULL
+    // (for example, after initialization).
+    if (this_region->is_Region() && this_region->req() == 3) {
+      int ind = 1;
+      if (!this_region->in(ind)->is_IfFalse()) {
+        ind = 2;
+      }
+      if (this_region->in(ind)->is_IfFalse() &&
+          this_region->in(ind)->in(0)->Opcode() == Op_If) {
+        Node* bol = this_region->in(ind)->in(0)->in(1);
+        assert(bol->is_Bool(), "");
+        Node* cmpx = bol->in(1);
+        if (bol->as_Bool()->_test._test == BoolTest::ne &&
+            cmpx->is_Cmp() && cmpx->in(2) == macro->intcon(0) &&
+            cmpx->in(1)->is_Load()) {
+          Node* adr = cmpx->in(1)->as_Load()->in(MemNode::Address);
+          const int marking_offset = in_bytes(G1ThreadLocalData::satb_mark_queue_active_offset());
+          if (adr->is_AddP() && adr->in(AddPNode::Base) == macro->top() &&
+              adr->in(AddPNode::Address)->Opcode() == Op_ThreadLocal &&
+              adr->in(AddPNode::Offset) == macro->MakeConX(marking_offset)) {
+            macro->replace_node(cmpx, macro->makecon(TypeInt::CC_EQ));
+          }
+        }
+      }
+    }
+    CardTableBarrierSetC2::eliminate_gc_barrier(macro, node);
+    return;
+  }
+
   assert(node->outcnt() <= 2, "expects 1 or 2 users: Xor and URShift nodes");
   // It could be only one user, URShift node, in Object.clone() intrinsic
   // but the new allocation is passed to arraycopy stub and it could not

src/hotspot/share/gc/g1/g1Arguments.cpp

@@ -83,7 +83,16 @@ void G1Arguments::initialize() {
     vm_exit_during_initialization("The flag -XX:+UseG1GC can not be combined with -XX:ParallelGCThreads=0", NULL);
   }
 
-  if (FLAG_IS_DEFAULT(G1ConcRefinementThreads)) {
+  if (G1BarrierSimple) {
+#if !defined(_LP64) || !(defined(X86) || defined(AARCH64))
+    warning("G1BarrierSimple is not supported with current platform"
+            "; ignoring G1BarrierSimple flag.");
+    FLAG_SET_DEFAULT(G1BarrierSimple, false);
+#else
+    FLAG_SET_DEFAULT(G1ConcRefinementThreads, 0);
+    FLAG_SET_DEFAULT(G1ConcRSLogCacheSize, 0);
+#endif
+  } else if (FLAG_IS_DEFAULT(G1ConcRefinementThreads)) {
     FLAG_SET_ERGO(uint, G1ConcRefinementThreads, ParallelGCThreads);
   }
 

src/hotspot/share/gc/g1/g1BarrierSet.cpp

@@ -97,6 +97,7 @@ void G1BarrierSet::write_ref_array_pre(narrowOop* dst, size_t count, bool dest_u
 }
 
 void G1BarrierSet::write_ref_field_post_slow(volatile jbyte* byte) {
+  assert(!G1BarrierSimple, "sanity");
   // In the slow path, we know a card is not young
   assert(*byte != G1CardTable::g1_young_card_val(), "slow path invoked without filtering");
   OrderAccess::storeload();
@@ -120,6 +121,10 @@ void G1BarrierSet::invalidate(MemRegion mr) {
   volatile jbyte* byte = _card_table->byte_for(mr.start());
   jbyte* last_byte = _card_table->byte_for(mr.last());
   Thread* thr = Thread::current();
+  if (G1BarrierSimple) {
+    memset((void*)byte, G1CardTable::dirty_card_val(), last_byte - byte + 1);
+    return;
+  }
     // skip all consecutive young cards
   for (; byte <= last_byte && *byte == G1CardTable::g1_young_card_val(); byte++);
 

src/hotspot/share/gc/g1/g1BarrierSet.inline.hpp

@@ -48,6 +48,10 @@ inline void G1BarrierSet::write_ref_field_pre(T* field) {
 template <DecoratorSet decorators, typename T>
 inline void G1BarrierSet::write_ref_field_post(T* field, oop new_val) {
   volatile jbyte* byte = _card_table->byte_for(field);
+  if (G1BarrierSimple) {
+    *byte = G1CardTable::dirty_card_val();
+    return;
+  }
   if (*byte != G1CardTable::g1_young_card_val()) {
     // Take a slow path for cards in old
     write_ref_field_post_slow(byte);

src/hotspot/share/gc/g1/g1CardTable.cpp

@@ -53,6 +53,8 @@ bool G1CardTable::mark_card_deferred(size_t card_index) {
 }
 
 void G1CardTable::g1_mark_as_young(const MemRegion& mr) {
+  assert(!G1BarrierSimple, "Should not be called with G1BarrierSimple");
+
   jbyte *const first = byte_for(mr.start());
   jbyte *const last = byte_after(mr.last());
 
@@ -61,7 +63,9 @@ void G1CardTable::g1_mark_as_young(const MemRegion& mr) {
 
 #ifndef PRODUCT
 void G1CardTable::verify_g1_young_region(MemRegion mr) {
-  verify_region(mr, g1_young_gen,  true);
+  if (!G1BarrierSimple) {
+    verify_region(mr, g1_young_gen,  true);
+  }
 }
 #endif
 
@@ -97,6 +101,9 @@ void G1CardTable::initialize(G1RegionToSpaceMapper* mapper) {
 }
 
 bool G1CardTable::is_in_young(oop obj) const {
+  if (G1BarrierSimple) {
+    return G1CollectedHeap::heap()->heap_region_containing(obj)->is_young();
+  }
   volatile jbyte* p = byte_for(obj);
   return *p == G1CardTable::g1_young_card_val();
 }

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

@@ -2181,6 +2181,12 @@ void G1CollectedHeap::heap_region_par_iterate_from_start(HeapRegionClosure* cl,
   _hrm.par_iterate(cl, hrclaimer, 0);
 }
 
+void G1CollectedHeap::heap_region_par_iterate_chunk_based(HeapRegionClosure* cl,
+                                                          HeapRegionChunkClaimer* chunk_claimer,
+                                                          uint worker_id) const {
+  _hrm.par_iterate(cl, chunk_claimer, chunk_claimer->offset_for_worker(worker_id));
+}
+
 void G1CollectedHeap::collection_set_iterate(HeapRegionClosure* cl) {
   _collection_set.iterate(cl);
 }
@@ -2671,7 +2677,9 @@ class RegisterHumongousWithInCSetFastTestClosure : public HeapRegionClosure {
           if (g1h->is_in_closed_subset(ct->addr_for(card_ptr))) {
             if (*card_ptr != G1CardTable::dirty_card_val()) {
               *card_ptr = G1CardTable::dirty_card_val();
-              _dcq.enqueue(card_ptr);
+              if (!G1BarrierSimple) {
+                _dcq.enqueue(card_ptr);
+              }
             }
           }
         }
@@ -3188,6 +3196,7 @@ class G1ParTask : public AbstractGangTask {
   G1RootProcessor*         _root_processor;
   TaskTerminator           _terminator;
   uint                     _n_workers;
+  HeapRegionChunkClaimer   _chunk_claimer;
 
 public:
   G1ParTask(G1CollectedHeap* g1h, G1ParScanThreadStateSet* per_thread_states, RefToScanQueueSet *task_queues, G1RootProcessor* root_processor, uint n_workers)
@@ -3197,7 +3206,8 @@ class G1ParTask : public AbstractGangTask {
       _queues(task_queues),
       _root_processor(root_processor),
       _terminator(n_workers, _queues),
-      _n_workers(n_workers)
+      _n_workers(n_workers),
+      _chunk_claimer(n_workers)
   {}
 
   void work(uint worker_id) {
@@ -3223,7 +3233,7 @@ class G1ParTask : public AbstractGangTask {
       // treating the nmethods visited to act as roots for concurrent marking.
       // We only want to make sure that the oops in the nmethods are adjusted with regard to the
       // objects copied by the current evacuation.
-      _g1h->g1_rem_set()->oops_into_collection_set_do(pss, worker_id);
+      _g1h->g1_rem_set()->oops_into_collection_set_do(pss, worker_id, &_chunk_claimer);
 
       double strong_roots_sec = os::elapsedTime() - start_strong_roots_sec;
 
@@ -3718,8 +3728,12 @@ void G1CollectedHeap::redirty_logged_cards() {
   dirty_card_queue_set().reset_for_par_iteration();
   workers()->run_task(&redirty_task);
 
-  DirtyCardQueueSet& dcq = G1BarrierSet::dirty_card_queue_set();
-  dcq.merge_bufferlists(&dirty_card_queue_set());
+  if (G1BarrierSimple) {
+    dirty_card_queue_set().clear();
+  } else {
+    DirtyCardQueueSet& dcq = G1BarrierSet::dirty_card_queue_set();
+    dcq.merge_bufferlists(&dirty_card_queue_set());
+  }
   assert(dirty_card_queue_set().completed_buffers_num() == 0, "All should be consumed");
 
   g1_policy()->phase_times()->record_redirty_logged_cards_time_ms((os::elapsedTime() - redirty_logged_cards_start) * 1000.0);

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

@@ -148,6 +148,7 @@ class G1CollectedHeap : public CollectedHeap {
 
   // Other related classes.
   friend class HeapRegionClaimer;
+  friend class HeapRegionChunkClaimer;
 
   // Testing classes.
   friend class G1CheckCSetFastTableClosure;
@@ -1167,6 +1168,10 @@ class G1CollectedHeap : public CollectedHeap {
   void heap_region_par_iterate_from_start(HeapRegionClosure* cl,
                                           HeapRegionClaimer* hrclaimer) const;
 
+  void heap_region_par_iterate_chunk_based(HeapRegionClosure* cl,
+                                           HeapRegionChunkClaimer* chunk_claimer,
+                                           uint worker_id) const;
+
   // Iterate over the regions (if any) in the current collection set.
   void collection_set_iterate(HeapRegionClosure* blk);
 

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

@@ -112,6 +112,11 @@ inline void
 G1CollectedHeap::dirty_young_block(HeapWord* start, size_t word_size) {
   assert_heap_not_locked();
 
+  if (G1BarrierSimple) {
+    // Not necessary to mark card young
+    return;
+  }
+
   // Assign the containing region to containing_hr so that we don't
   // have to keep calling heap_region_containing() in the
   // asserts below.

src/hotspot/share/gc/g1/g1ConcurrentRefineThread.cpp

@@ -90,6 +90,7 @@ void G1ConcurrentRefineThread::deactivate() {
 }
 
 void G1ConcurrentRefineThread::run_service() {
+  assert(!G1BarrierSimple, "Concurrent Refinement should be disabled");
   _vtime_start = os::elapsedVTime();
 
   while (!should_terminate()) {