void G1BarrierSetAssembler::g1_write_barrier_pre(MacroAssembler* masm,

                                                 Register obj,

                                                 Register pre_val,

                                                 Register thread,

                                                 Register tmp,

                                                 bool tosca_live,

                                                 bool expand_call) {

  // If expand_call is true then we expand the call_VM_leaf macro

  // directly to skip generating the check by

  // InterpreterMacroAssembler::call_VM_leaf_base that checks _last_sp.

  assert(thread == rthread, "must be");

  Label done;

  Label runtime;

  assert_different_registers(obj, pre_val, tmp, rscratch1);

  assert(pre_val != noreg &&  tmp != noreg, "expecting a register");

  Address in_progress(thread, in_bytes(JavaThread::satb_mark_queue_offset() +

                                       SATBMarkQueue::byte_offset_of_active()));

  Address index(thread, in_bytes(JavaThread::satb_mark_queue_offset() +

                                       SATBMarkQueue::byte_offset_of_index()));

  Address buffer(thread, in_bytes(JavaThread::satb_mark_queue_offset() +

                                       SATBMarkQueue::byte_offset_of_buf()));

  // Is marking active?

  if (in_bytes(SATBMarkQueue::byte_width_of_active()) == 4) {

    __ ldrw(tmp, in_progress);

  } else {

    assert(in_bytes(SATBMarkQueue::byte_width_of_active()) == 1, "Assumption");

    __ ldrb(tmp, in_progress);

  }

  __ cbzw(tmp, done);

  // Do we need to load the previous value?

  if (obj != noreg) {

    __ load_heap_oop(pre_val, Address(obj, 0));

  }

  // Is the previous value null?

  __ cbz(pre_val, done);

  // Can we store original value in the thread's buffer?

  // Is index == 0?

  // (The index field is typed as size_t.)

  __ ldr(tmp, index);                      // tmp := *index_adr

  __ cbz(tmp, runtime);                    // tmp == 0?

                                        // If yes, goto runtime

  __ sub(tmp, tmp, wordSize);              // tmp := tmp - wordSize

  __ str(tmp, index);                      // *index_adr := tmp

  __ ldr(rscratch1, buffer);

  __ add(tmp, tmp, rscratch1);             // tmp := tmp + *buffer_adr

  // Record the previous value

  __ str(pre_val, Address(tmp, 0));

  __ b(done);

  __ bind(runtime);

  // save the live input values

  RegSet saved = RegSet::of(pre_val);

  if (tosca_live) saved += RegSet::of(r0);

  if (obj != noreg) saved += RegSet::of(obj);

  __ push(saved, sp);

  // Calling the runtime using the regular call_VM_leaf mechanism generates

  // code (generated by InterpreterMacroAssember::call_VM_leaf_base)

  // that checks that the *(rfp+frame::interpreter_frame_last_sp) == NULL.

  //

  // If we care generating the pre-barrier without a frame (e.g. in the

  // intrinsified Reference.get() routine) then ebp might be pointing to

  // the caller frame and so this check will most likely fail at runtime.

  //

  // Expanding the call directly bypasses the generation of the check.

  // So when we do not have have a full interpreter frame on the stack

  // expand_call should be passed true.

  if (expand_call) {

    assert(pre_val != c_rarg1, "smashed arg");

    __ super_call_VM_leaf(CAST_FROM_FN_PTR(address, SharedRuntime::g1_wb_pre), pre_val, thread);

  } else {

    __ call_VM_leaf(CAST_FROM_FN_PTR(address, SharedRuntime::g1_wb_pre), pre_val, thread);

  }

  __ pop(saved, sp);

  __ bind(done);

}

void G1BarrierSetAssembler::g1_write_barrier_post(MacroAssembler* masm,

                                                  Register store_addr,

                                                  Register new_val,

                                                  Register thread,

                                                  Register tmp,

                                                  Register tmp2) {

  assert(thread == rthread, "must be");

  assert_different_registers(store_addr, new_val, thread, tmp, tmp2,

                             rscratch1);

  assert(store_addr != noreg && new_val != noreg && tmp != noreg

         && tmp2 != noreg, "expecting a register");

  Address queue_index(thread, in_bytes(JavaThread::dirty_card_queue_offset() +

                                       DirtyCardQueue::byte_offset_of_index()));

  Address buffer(thread, in_bytes(JavaThread::dirty_card_queue_offset() +

                                       DirtyCardQueue::byte_offset_of_buf()));

  BarrierSet* bs = Universe::heap()->barrier_set();

  CardTableBarrierSet* ctbs = barrier_set_cast<CardTableBarrierSet>(bs);

  CardTable* ct = ctbs->card_table();

  assert(sizeof(*ct->byte_map_base()) == sizeof(jbyte), "adjust this code");

  Label done;

  Label runtime;

  // Does store cross heap regions?

  __ eor(tmp, store_addr, new_val);

  __ lsr(tmp, tmp, HeapRegion::LogOfHRGrainBytes);

  __ cbz(tmp, done);

  // crosses regions, storing NULL?

  __ cbz(new_val, done);

  // storing region crossing non-NULL, is card already dirty?

  ExternalAddress cardtable((address) ct->byte_map_base());

  assert(sizeof(*ct->byte_map_base()) == sizeof(jbyte), "adjust this code");

  const Register card_addr = tmp;

  __ lsr(card_addr, store_addr, CardTable::card_shift);

  // get the address of the card

  __ load_byte_map_base(tmp2);

  __ add(card_addr, card_addr, tmp2);

  __ ldrb(tmp2, Address(card_addr));

  __ cmpw(tmp2, (int)G1CardTable::g1_young_card_val());

  __ br(Assembler::EQ, done);

  assert((int)CardTable::dirty_card_val() == 0, "must be 0");

  __ membar(Assembler::StoreLoad);

  __ ldrb(tmp2, Address(card_addr));

  __ cbzw(tmp2, done);

  // storing a region crossing, non-NULL oop, card is clean.

  // dirty card and log.

  __ strb(zr, Address(card_addr));

  __ ldr(rscratch1, queue_index);

  __ cbz(rscratch1, runtime);

  __ sub(rscratch1, rscratch1, wordSize);

  __ str(rscratch1, queue_index);

  __ ldr(tmp2, buffer);

  __ str(card_addr, Address(tmp2, rscratch1));

  __ b(done);

  __ bind(runtime);

  // save the live input values

  RegSet saved = RegSet::of(store_addr, new_val);

  __ push(saved, sp);

  __ call_VM_leaf(CAST_FROM_FN_PTR(address, SharedRuntime::g1_wb_post), card_addr, thread);

  __ pop(saved, sp);

  __ bind(done);

}

void G1BarrierSetAssembler::load_at(MacroAssembler* masm, DecoratorSet decorators, BasicType type,

                                    Register dst, Address src, Register tmp1, Register tmp_thread) {

  bool on_oop = type == T_OBJECT || type == T_ARRAY;

  bool on_weak = (decorators & ON_WEAK_OOP_REF) != 0;

  bool on_phantom = (decorators & ON_PHANTOM_OOP_REF) != 0;

  bool on_reference = on_weak || on_phantom;

  ModRefBarrierSetAssembler::load_at(masm, decorators, type, dst, src, tmp1, tmp_thread);

  if (on_oop && on_reference) {

    // Generate the G1 pre-barrier code to log the value of

    // the referent field in an SATB buffer.

    g1_write_barrier_pre(masm /* masm */,

                         noreg /* obj */,

                         dst /* pre_val */,

                         rthread /* thread */,

                         tmp1 /* tmp */,

                         true /* tosca_live */,

                         true /* expand_call */);

  }

}

void G1BarrierSetAssembler::oop_store_at(MacroAssembler* masm, DecoratorSet decorators, BasicType type,

                                         Address dst, Register val, Register tmp1, Register tmp2) {

  // flatten object address if needed

  if (dst.index() == noreg && dst.offset() == 0) {

    if (dst.base() != r3) {

      __ mov(r3, dst.base());

    }

  } else {

    __ lea(r3, dst);

  }

  g1_write_barrier_pre(masm,

                       r3 /* obj */,

                       tmp2 /* pre_val */,

                       rthread /* thread */,

                       tmp1  /* tmp */,

                       val != noreg /* tosca_live */,

                       false /* expand_call */);

  if (val == noreg) {

    __ store_heap_oop_null(Address(r3, 0));

  } else {

    // G1 barrier needs uncompressed oop for region cross check.

    Register new_val = val;

    if (UseCompressedOops) {

      new_val = rscratch2;

      __ mov(new_val, val);

    }

    __ store_heap_oop(Address(r3, 0), val);

    g1_write_barrier_post(masm,

                          r3 /* store_adr */,

                          new_val /* new_val */,

                          rthread /* thread */,

                          tmp1 /* tmp */,

                          tmp2 /* tmp2 */);

  }

}

#undef __