void G1PostBarrierStub::emit_code(LIR_Assembler* ce) {
  __ bind(_entry);

  assert(addr()->is_register(), "Precondition.");
  assert(new_val()->is_register(), "Precondition.");
  Register addr_reg = addr()->as_pointer_register();
  Register new_val_reg = new_val()->as_register();
  __ br_on_reg_cond(Assembler::rc_z, /*annul*/false, Assembler::pt,
                    new_val_reg, _continuation);
  __ delayed()->nop();

  __ call(Runtime1::entry_for(Runtime1::Runtime1::g1_post_barrier_slow_id));
  __ delayed()->mov(addr_reg, G4);
  __ br(Assembler::always, false, Assembler::pt, _continuation);
  __ delayed()->nop();
}
void G1PreBarrierStub::emit_code(LIR_Assembler* ce) {
  __ bind(_entry);

  assert(pre_val()->is_register(), "Precondition.");

  Register pre_val_reg = pre_val()->as_register();

  ce->mem2reg(addr(), pre_val(), T_OBJECT, patch_code(), info(), false);
  __ br_on_reg_cond(Assembler::rc_z, /*annul*/false, Assembler::pt,
                    pre_val_reg, _continuation);
  __ delayed()->nop();

  __ call(Runtime1::entry_for(Runtime1::Runtime1::g1_pre_barrier_slow_id));
  __ delayed()->mov(pre_val_reg, G4);
  __ br(Assembler::always, false, Assembler::pt, _continuation);
  __ delayed()->nop();

}
Example #3
0
void G1PreBarrierStub::emit_code(LIR_Assembler* ce) {
  __ bind(_entry);

  assert(pre_val()->is_register(), "Precondition.");

  Register pre_val_reg = pre_val()->as_register();

  ce->mem2reg(addr(), pre_val(), T_OBJECT, patch_code(), info(), false /*wide*/, false /*unaligned*/);
  if (__ is_in_wdisp16_range(_continuation)) {
    __ br_on_reg_cond(Assembler::rc_z, /*annul*/false, Assembler::pt,
                      pre_val_reg, _continuation);
  } else {
    __ cmp(pre_val_reg, G0);
    __ brx(Assembler::equal, false, Assembler::pn, _continuation);
  }
  __ delayed()->nop();

  __ call(Runtime1::entry_for(Runtime1::Runtime1::g1_pre_barrier_slow_id));
  __ delayed()->mov(pre_val_reg, G4);
  __ br(Assembler::always, false, Assembler::pt, _continuation);
  __ delayed()->nop();

}
Example #4
0
OopMapSet* Runtime1::generate_code_for(StubID id, StubAssembler* sasm) {

  OopMapSet* oop_maps = NULL;
  // for better readability
  const bool must_gc_arguments = true;
  const bool dont_gc_arguments = false;

  // stub code & info for the different stubs
  switch (id) {
    case forward_exception_id:
      {
        // we're handling an exception in the context of a compiled
        // frame.  The registers have been saved in the standard
        // places.  Perform an exception lookup in the caller and
        // dispatch to the handler if found.  Otherwise unwind and
        // dispatch to the callers exception handler.

        oop_maps = new OopMapSet();
        OopMap* oop_map = generate_oop_map(sasm, true);

        // transfer the pending exception to the exception_oop
        __ ld_ptr(G2_thread, in_bytes(JavaThread::pending_exception_offset()), Oexception);
        __ ld_ptr(Oexception, 0, G0);
        __ st_ptr(G0, G2_thread, in_bytes(JavaThread::pending_exception_offset()));
        __ add(I7, frame::pc_return_offset, Oissuing_pc);

        generate_handle_exception(sasm, oop_maps, oop_map);
        __ should_not_reach_here();
      }
      break;

    case new_instance_id:
    case fast_new_instance_id:
    case fast_new_instance_init_check_id:
      {
        Register G5_klass = G5; // Incoming
        Register O0_obj   = O0; // Outgoing

        if (id == new_instance_id) {
          __ set_info("new_instance", dont_gc_arguments);
        } else if (id == fast_new_instance_id) {
          __ set_info("fast new_instance", dont_gc_arguments);
        } else {
          assert(id == fast_new_instance_init_check_id, "bad StubID");
          __ set_info("fast new_instance init check", dont_gc_arguments);
        }

        if ((id == fast_new_instance_id || id == fast_new_instance_init_check_id) &&
            UseTLAB && FastTLABRefill) {
          Label slow_path;
          Register G1_obj_size = G1;
          Register G3_t1 = G3;
          Register G4_t2 = G4;
          assert_different_registers(G5_klass, G1_obj_size, G3_t1, G4_t2);

          // Push a frame since we may do dtrace notification for the
          // allocation which requires calling out and we don't want
          // to stomp the real return address.
          __ save_frame(0);

          if (id == fast_new_instance_init_check_id) {
            // make sure the klass is initialized
            __ ld(G5_klass, instanceKlass::init_state_offset_in_bytes() + sizeof(oopDesc), G3_t1);
            __ cmp(G3_t1, instanceKlass::fully_initialized);
            __ br(Assembler::notEqual, false, Assembler::pn, slow_path);
            __ delayed()->nop();
          }
#ifdef ASSERT
          // assert object can be fast path allocated
          {
            Label ok, not_ok;
          __ ld(G5_klass, Klass::layout_helper_offset_in_bytes() + sizeof(oopDesc), G1_obj_size);
          __ cmp(G1_obj_size, 0);  // make sure it's an instance (LH > 0)
          __ br(Assembler::lessEqual, false, Assembler::pn, not_ok);
          __ delayed()->nop();
          __ btst(Klass::_lh_instance_slow_path_bit, G1_obj_size);
          __ br(Assembler::zero, false, Assembler::pn, ok);
          __ delayed()->nop();
          __ bind(not_ok);
          __ stop("assert(can be fast path allocated)");
          __ should_not_reach_here();
          __ bind(ok);
          }
#endif // ASSERT
          // if we got here then the TLAB allocation failed, so try
          // refilling the TLAB or allocating directly from eden.
          Label retry_tlab, try_eden;
          __ tlab_refill(retry_tlab, try_eden, slow_path); // preserves G5_klass

          __ bind(retry_tlab);

          // get the instance size
          __ ld(G5_klass, klassOopDesc::header_size() * HeapWordSize + Klass::layout_helper_offset_in_bytes(), G1_obj_size);
          __ tlab_allocate(O0_obj, G1_obj_size, 0, G3_t1, slow_path);
          __ initialize_object(O0_obj, G5_klass, G1_obj_size, 0, G3_t1, G4_t2);
          __ verify_oop(O0_obj);
          __ mov(O0, I0);
          __ ret();
          __ delayed()->restore();

          __ bind(try_eden);
          // get the instance size
          __ ld(G5_klass, klassOopDesc::header_size() * HeapWordSize + Klass::layout_helper_offset_in_bytes(), G1_obj_size);
          __ eden_allocate(O0_obj, G1_obj_size, 0, G3_t1, G4_t2, slow_path);
          __ initialize_object(O0_obj, G5_klass, G1_obj_size, 0, G3_t1, G4_t2);
          __ verify_oop(O0_obj);
          __ mov(O0, I0);
          __ ret();
          __ delayed()->restore();

          __ bind(slow_path);

          // pop this frame so generate_stub_call can push it's own
          __ restore();
        }

        oop_maps = generate_stub_call(sasm, I0, CAST_FROM_FN_PTR(address, new_instance), G5_klass);
        // I0->O0: new instance
      }

      break;

#ifdef TIERED
    case counter_overflow_id:
        // G4 contains bci
      oop_maps = generate_stub_call(sasm, noreg, CAST_FROM_FN_PTR(address, counter_overflow), G4);
      break;
#endif // TIERED

    case new_type_array_id:
    case new_object_array_id:
      {
        Register G5_klass = G5; // Incoming
        Register G4_length = G4; // Incoming
        Register O0_obj   = O0; // Outgoing

        Address klass_lh(G5_klass, ((klassOopDesc::header_size() * HeapWordSize)
                                    + Klass::layout_helper_offset_in_bytes()));
        assert(Klass::_lh_header_size_shift % BitsPerByte == 0, "bytewise");
        assert(Klass::_lh_header_size_mask == 0xFF, "bytewise");
        // Use this offset to pick out an individual byte of the layout_helper:
        const int klass_lh_header_size_offset = ((BytesPerInt - 1)  // 3 - 2 selects byte {0,1,0,0}
                                                 - Klass::_lh_header_size_shift / BitsPerByte);

        if (id == new_type_array_id) {
          __ set_info("new_type_array", dont_gc_arguments);
        } else {
          __ set_info("new_object_array", dont_gc_arguments);
        }

#ifdef ASSERT
        // assert object type is really an array of the proper kind
        {
          Label ok;
          Register G3_t1 = G3;
          __ ld(klass_lh, G3_t1);
          __ sra(G3_t1, Klass::_lh_array_tag_shift, G3_t1);
          int tag = ((id == new_type_array_id)
                     ? Klass::_lh_array_tag_type_value
                     : Klass::_lh_array_tag_obj_value);
          __ cmp(G3_t1, tag);
          __ brx(Assembler::equal, false, Assembler::pt, ok);
          __ delayed()->nop();
          __ stop("assert(is an array klass)");
          __ should_not_reach_here();
          __ bind(ok);
        }
#endif // ASSERT

        if (UseTLAB && FastTLABRefill) {
          Label slow_path;
          Register G1_arr_size = G1;
          Register G3_t1 = G3;
          Register O1_t2 = O1;
          assert_different_registers(G5_klass, G4_length, G1_arr_size, G3_t1, O1_t2);

          // check that array length is small enough for fast path
          __ set(C1_MacroAssembler::max_array_allocation_length, G3_t1);
          __ cmp(G4_length, G3_t1);
          __ br(Assembler::greaterUnsigned, false, Assembler::pn, slow_path);
          __ delayed()->nop();

          // if we got here then the TLAB allocation failed, so try
          // refilling the TLAB or allocating directly from eden.
          Label retry_tlab, try_eden;
          __ tlab_refill(retry_tlab, try_eden, slow_path); // preserves G4_length and G5_klass

          __ bind(retry_tlab);

          // get the allocation size: (length << (layout_helper & 0x1F)) + header_size
          __ ld(klass_lh, G3_t1);
          __ sll(G4_length, G3_t1, G1_arr_size);
          __ srl(G3_t1, Klass::_lh_header_size_shift, G3_t1);
          __ and3(G3_t1, Klass::_lh_header_size_mask, G3_t1);
          __ add(G1_arr_size, G3_t1, G1_arr_size);
          __ add(G1_arr_size, MinObjAlignmentInBytesMask, G1_arr_size);  // align up
          __ and3(G1_arr_size, ~MinObjAlignmentInBytesMask, G1_arr_size);

          __ tlab_allocate(O0_obj, G1_arr_size, 0, G3_t1, slow_path);  // preserves G1_arr_size

          __ initialize_header(O0_obj, G5_klass, G4_length, G3_t1, O1_t2);
          __ ldub(klass_lh, G3_t1, klass_lh_header_size_offset);
          __ sub(G1_arr_size, G3_t1, O1_t2);  // body length
          __ add(O0_obj, G3_t1, G3_t1);       // body start
          __ initialize_body(G3_t1, O1_t2);
          __ verify_oop(O0_obj);
          __ retl();
          __ delayed()->nop();

          __ bind(try_eden);
          // get the allocation size: (length << (layout_helper & 0x1F)) + header_size
          __ ld(klass_lh, G3_t1);
          __ sll(G4_length, G3_t1, G1_arr_size);
          __ srl(G3_t1, Klass::_lh_header_size_shift, G3_t1);
          __ and3(G3_t1, Klass::_lh_header_size_mask, G3_t1);
          __ add(G1_arr_size, G3_t1, G1_arr_size);
          __ add(G1_arr_size, MinObjAlignmentInBytesMask, G1_arr_size);
          __ and3(G1_arr_size, ~MinObjAlignmentInBytesMask, G1_arr_size);

          __ eden_allocate(O0_obj, G1_arr_size, 0, G3_t1, O1_t2, slow_path);  // preserves G1_arr_size

          __ initialize_header(O0_obj, G5_klass, G4_length, G3_t1, O1_t2);
          __ ldub(klass_lh, G3_t1, klass_lh_header_size_offset);
          __ sub(G1_arr_size, G3_t1, O1_t2);  // body length
          __ add(O0_obj, G3_t1, G3_t1);       // body start
          __ initialize_body(G3_t1, O1_t2);
          __ verify_oop(O0_obj);
          __ retl();
          __ delayed()->nop();

          __ bind(slow_path);
        }

        if (id == new_type_array_id) {
          oop_maps = generate_stub_call(sasm, I0, CAST_FROM_FN_PTR(address, new_type_array), G5_klass, G4_length);
        } else {
          oop_maps = generate_stub_call(sasm, I0, CAST_FROM_FN_PTR(address, new_object_array), G5_klass, G4_length);
        }
        // I0 -> O0: new array
      }
      break;

    case new_multi_array_id:
      { // O0: klass
        // O1: rank
        // O2: address of 1st dimension
        __ set_info("new_multi_array", dont_gc_arguments);
        oop_maps = generate_stub_call(sasm, I0, CAST_FROM_FN_PTR(address, new_multi_array), I0, I1, I2);
        // I0 -> O0: new multi array
      }
      break;

    case register_finalizer_id:
      {
        __ set_info("register_finalizer", dont_gc_arguments);

        // load the klass and check the has finalizer flag
        Label register_finalizer;
        Register t = O1;
        __ ld_ptr(O0, oopDesc::klass_offset_in_bytes(), t);
        __ ld(t, Klass::access_flags_offset_in_bytes() + sizeof(oopDesc), t);
        __ set(JVM_ACC_HAS_FINALIZER, G3);
        __ andcc(G3, t, G0);
        __ br(Assembler::notZero, false, Assembler::pt, register_finalizer);
        __ delayed()->nop();

        // do a leaf return
        __ retl();
        __ delayed()->nop();

        __ bind(register_finalizer);
        OopMap* oop_map = save_live_registers(sasm);
        int call_offset = __ call_RT(noreg, noreg,
                                     CAST_FROM_FN_PTR(address, SharedRuntime::register_finalizer), I0);
        oop_maps = new OopMapSet();
        oop_maps->add_gc_map(call_offset, oop_map);

        // Now restore all the live registers
        restore_live_registers(sasm);

        __ ret();
        __ delayed()->restore();
      }
      break;

    case throw_range_check_failed_id:
      { __ set_info("range_check_failed", dont_gc_arguments); // arguments will be discarded
        // G4: index
        oop_maps = generate_exception_throw(sasm, CAST_FROM_FN_PTR(address, throw_range_check_exception), true);
      }
      break;

    case throw_index_exception_id:
      { __ set_info("index_range_check_failed", dont_gc_arguments); // arguments will be discarded
        // G4: index
        oop_maps = generate_exception_throw(sasm, CAST_FROM_FN_PTR(address, throw_index_exception), true);
      }
      break;

    case throw_div0_exception_id:
      { __ set_info("throw_div0_exception", dont_gc_arguments);
        oop_maps = generate_exception_throw(sasm, CAST_FROM_FN_PTR(address, throw_div0_exception), false);
      }
      break;

    case throw_null_pointer_exception_id:
      { __ set_info("throw_null_pointer_exception", dont_gc_arguments);
        oop_maps = generate_exception_throw(sasm, CAST_FROM_FN_PTR(address, throw_null_pointer_exception), false);
      }
      break;

    case handle_exception_id:
      {
        __ set_info("handle_exception", dont_gc_arguments);
        // make a frame and preserve the caller's caller-save registers

        oop_maps = new OopMapSet();
        OopMap* oop_map = save_live_registers(sasm);
        __ mov(Oexception->after_save(),  Oexception);
        __ mov(Oissuing_pc->after_save(), Oissuing_pc);
        generate_handle_exception(sasm, oop_maps, oop_map);
      }
      break;

    case unwind_exception_id:
      {
        // O0: exception
        // I7: address of call to this method

        __ set_info("unwind_exception", dont_gc_arguments);
        __ mov(Oexception, Oexception->after_save());
        __ add(I7, frame::pc_return_offset, Oissuing_pc->after_save());

        __ call_VM_leaf(L7_thread_cache, CAST_FROM_FN_PTR(address, SharedRuntime::exception_handler_for_return_address),
                        Oissuing_pc->after_save());
        __ verify_not_null_oop(Oexception->after_save());
        __ jmp(O0, 0);
        __ delayed()->restore();
      }
      break;

    case throw_array_store_exception_id:
      {
        __ set_info("throw_array_store_exception", dont_gc_arguments);
        oop_maps = generate_exception_throw(sasm, CAST_FROM_FN_PTR(address, throw_array_store_exception), false);
      }
      break;

    case throw_class_cast_exception_id:
      {
        // G4: object
        __ set_info("throw_class_cast_exception", dont_gc_arguments);
        oop_maps = generate_exception_throw(sasm, CAST_FROM_FN_PTR(address, throw_class_cast_exception), true);
      }
      break;

    case throw_incompatible_class_change_error_id:
      {
        __ set_info("throw_incompatible_class_cast_exception", dont_gc_arguments);
        oop_maps = generate_exception_throw(sasm, CAST_FROM_FN_PTR(address, throw_incompatible_class_change_error), false);
      }
      break;

    case slow_subtype_check_id:
      { // Support for uint StubRoutine::partial_subtype_check( Klass sub, Klass super );
        // Arguments :
        //
        //      ret  : G3
        //      sub  : G3, argument, destroyed
        //      super: G1, argument, not changed
        //      raddr: O7, blown by call
        Label miss;

        __ save_frame(0);               // Blow no registers!

        __ check_klass_subtype_slow_path(G3, G1, L0, L1, L2, L4, NULL, &miss);

        __ mov(1, G3);
        __ ret();                       // Result in G5 is 'true'
        __ delayed()->restore();        // free copy or add can go here

        __ bind(miss);
        __ mov(0, G3);
        __ ret();                       // Result in G5 is 'false'
        __ delayed()->restore();        // free copy or add can go here
      }

    case monitorenter_nofpu_id:
    case monitorenter_id:
      { // G4: object
        // G5: lock address
        __ set_info("monitorenter", dont_gc_arguments);

        int save_fpu_registers = (id == monitorenter_id);
        // make a frame and preserve the caller's caller-save registers
        OopMap* oop_map = save_live_registers(sasm, save_fpu_registers);

        int call_offset = __ call_RT(noreg, noreg, CAST_FROM_FN_PTR(address, monitorenter), G4, G5);

        oop_maps = new OopMapSet();
        oop_maps->add_gc_map(call_offset, oop_map);
        restore_live_registers(sasm, save_fpu_registers);

        __ ret();
        __ delayed()->restore();
      }
      break;

    case monitorexit_nofpu_id:
    case monitorexit_id:
      { // G4: lock address
        // note: really a leaf routine but must setup last java sp
        //       => use call_RT for now (speed can be improved by
        //       doing last java sp setup manually)
        __ set_info("monitorexit", dont_gc_arguments);

        int save_fpu_registers = (id == monitorexit_id);
        // make a frame and preserve the caller's caller-save registers
        OopMap* oop_map = save_live_registers(sasm, save_fpu_registers);

        int call_offset = __ call_RT(noreg, noreg, CAST_FROM_FN_PTR(address, monitorexit), G4);

        oop_maps = new OopMapSet();
        oop_maps->add_gc_map(call_offset, oop_map);
        restore_live_registers(sasm, save_fpu_registers);

        __ ret();
        __ delayed()->restore();

      }
      break;

    case access_field_patching_id:
      { __ set_info("access_field_patching", dont_gc_arguments);
        oop_maps = generate_patching(sasm, CAST_FROM_FN_PTR(address, access_field_patching));
      }
      break;

    case load_klass_patching_id:
      { __ set_info("load_klass_patching", dont_gc_arguments);
        oop_maps = generate_patching(sasm, CAST_FROM_FN_PTR(address, move_klass_patching));
      }
      break;

    case jvmti_exception_throw_id:
      { // Oexception : exception
        __ set_info("jvmti_exception_throw", dont_gc_arguments);
        oop_maps = generate_stub_call(sasm, noreg, CAST_FROM_FN_PTR(address, Runtime1::post_jvmti_exception_throw), I0);
      }
      break;

    case dtrace_object_alloc_id:
      { // O0: object
        __ set_info("dtrace_object_alloc", dont_gc_arguments);
        // we can't gc here so skip the oopmap but make sure that all
        // the live registers get saved.
        save_live_registers(sasm);

        __ save_thread(L7_thread_cache);
        __ call(CAST_FROM_FN_PTR(address, SharedRuntime::dtrace_object_alloc),
                relocInfo::runtime_call_type);
        __ delayed()->mov(I0, O0);
        __ restore_thread(L7_thread_cache);

        restore_live_registers(sasm);
        __ ret();
        __ delayed()->restore();
      }
      break;

#ifndef SERIALGC
    case g1_pre_barrier_slow_id:
      { // G4: previous value of memory
        BarrierSet* bs = Universe::heap()->barrier_set();
        if (bs->kind() != BarrierSet::G1SATBCTLogging) {
          __ save_frame(0);
          __ set((int)id, O1);
          __ call_RT(noreg, noreg, CAST_FROM_FN_PTR(address, unimplemented_entry), I0);
          __ should_not_reach_here();
          break;
        }

        __ set_info("g1_pre_barrier_slow_id", dont_gc_arguments);

        Register pre_val = G4;
        Register tmp  = G1_scratch;
        Register tmp2 = G3_scratch;

        Label refill, restart;
        bool with_frame = false; // I don't know if we can do with-frame.
        int satb_q_index_byte_offset =
          in_bytes(JavaThread::satb_mark_queue_offset() +
                   PtrQueue::byte_offset_of_index());
        int satb_q_buf_byte_offset =
          in_bytes(JavaThread::satb_mark_queue_offset() +
                   PtrQueue::byte_offset_of_buf());
        __ bind(restart);
        __ ld_ptr(G2_thread, satb_q_index_byte_offset, tmp);

        __ br_on_reg_cond(Assembler::rc_z, /*annul*/false,
                          Assembler::pn, tmp, refill);

        // If the branch is taken, no harm in executing this in the delay slot.
        __ delayed()->ld_ptr(G2_thread, satb_q_buf_byte_offset, tmp2);
        __ sub(tmp, oopSize, tmp);

        __ st_ptr(pre_val, tmp2, tmp);  // [_buf + index] := <address_of_card>
        // Use return-from-leaf
        __ retl();
        __ delayed()->st_ptr(tmp, G2_thread, satb_q_index_byte_offset);

        __ bind(refill);
        __ save_frame(0);

        __ mov(pre_val, L0);
        __ mov(tmp,     L1);
        __ mov(tmp2,    L2);

        __ call_VM_leaf(L7_thread_cache,
                        CAST_FROM_FN_PTR(address,
                                         SATBMarkQueueSet::handle_zero_index_for_thread),
                                         G2_thread);

        __ mov(L0, pre_val);
        __ mov(L1, tmp);
        __ mov(L2, tmp2);

        __ br(Assembler::always, /*annul*/false, Assembler::pt, restart);
        __ delayed()->restore();
      }
      break;

    case g1_post_barrier_slow_id:
      {
        BarrierSet* bs = Universe::heap()->barrier_set();
        if (bs->kind() != BarrierSet::G1SATBCTLogging) {
          __ save_frame(0);
          __ set((int)id, O1);
          __ call_RT(noreg, noreg, CAST_FROM_FN_PTR(address, unimplemented_entry), I0);
          __ should_not_reach_here();
          break;
        }

        __ set_info("g1_post_barrier_slow_id", dont_gc_arguments);

        Register addr = G4;
        Register cardtable = G5;
        Register tmp  = G1_scratch;
        Register tmp2 = G3_scratch;
        jbyte* byte_map_base = ((CardTableModRefBS*)bs)->byte_map_base;

        Label not_already_dirty, restart, refill;

#ifdef _LP64
        __ srlx(addr, CardTableModRefBS::card_shift, addr);
#else
        __ srl(addr, CardTableModRefBS::card_shift, addr);
#endif

        AddressLiteral rs(byte_map_base);
        __ set(rs, cardtable);         // cardtable := <card table base>
        __ ldub(addr, cardtable, tmp); // tmp := [addr + cardtable]

        __ br_on_reg_cond(Assembler::rc_nz, /*annul*/false, Assembler::pt,
                          tmp, not_already_dirty);
        // Get cardtable + tmp into a reg by itself -- useful in the take-the-branch
        // case, harmless if not.
        __ delayed()->add(addr, cardtable, tmp2);

        // We didn't take the branch, so we're already dirty: return.
        // Use return-from-leaf
        __ retl();
        __ delayed()->nop();

        // Not dirty.
        __ bind(not_already_dirty);
        // First, dirty it.
        __ stb(G0, tmp2, 0);  // [cardPtr] := 0  (i.e., dirty).

        Register tmp3 = cardtable;
        Register tmp4 = tmp;

        // these registers are now dead
        addr = cardtable = tmp = noreg;

        int dirty_card_q_index_byte_offset =
          in_bytes(JavaThread::dirty_card_queue_offset() +
                   PtrQueue::byte_offset_of_index());
        int dirty_card_q_buf_byte_offset =
          in_bytes(JavaThread::dirty_card_queue_offset() +
                   PtrQueue::byte_offset_of_buf());
        __ bind(restart);
        __ ld_ptr(G2_thread, dirty_card_q_index_byte_offset, tmp3);

        __ br_on_reg_cond(Assembler::rc_z, /*annul*/false, Assembler::pn,
                          tmp3, refill);
        // If the branch is taken, no harm in executing this in the delay slot.
        __ delayed()->ld_ptr(G2_thread, dirty_card_q_buf_byte_offset, tmp4);
        __ sub(tmp3, oopSize, tmp3);

        __ st_ptr(tmp2, tmp4, tmp3);  // [_buf + index] := <address_of_card>
        // Use return-from-leaf
        __ retl();
        __ delayed()->st_ptr(tmp3, G2_thread, dirty_card_q_index_byte_offset);

        __ bind(refill);
        __ save_frame(0);

        __ mov(tmp2, L0);
        __ mov(tmp3, L1);
        __ mov(tmp4, L2);

        __ call_VM_leaf(L7_thread_cache,
                        CAST_FROM_FN_PTR(address,
                                         DirtyCardQueueSet::handle_zero_index_for_thread),
                                         G2_thread);

        __ mov(L0, tmp2);
        __ mov(L1, tmp3);
        __ mov(L2, tmp4);

        __ br(Assembler::always, /*annul*/false, Assembler::pt, restart);
        __ delayed()->restore();
      }
      break;
#endif // !SERIALGC

    default:
      { __ set_info("unimplemented entry", dont_gc_arguments);
        __ save_frame(0);
        __ set((int)id, O1);
        __ call_RT(noreg, noreg, CAST_FROM_FN_PTR(address, unimplemented_entry), O1);
        __ should_not_reach_here();
      }
      break;
  }
  return oop_maps;
}