xref: /openjdk10/hotspot/src/cpu/arm/vm/interp_masm_arm.cpp

/*
 * Copyright (c) 2008, 2017, Oracle and/or its affiliates. All rights reserved.
 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
 *
 * This code is free software; you can redistribute it and/or modify it
 * under the terms of the GNU General Public License version 2 only, as
 * published by the Free Software Foundation.
 *
 * This code is distributed in the hope that it will be useful, but WITHOUT
 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
 * FITNESS FOR A PARTICULAR PURPOSE.  See the GNU General Public License
 * version 2 for more details (a copy is included in the LICENSE file that
 * accompanied this code).
 *
 * You should have received a copy of the GNU General Public License version
 * 2 along with this work; if not, write to the Free Software Foundation,
 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
 *
 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
 * or visit www.oracle.com if you need additional information or have any
 * questions.
 *
 */

#include "precompiled.hpp"
#include "gc/shared/barrierSet.inline.hpp"
#include "gc/shared/cardTableModRefBS.inline.hpp"
#include "gc/shared/collectedHeap.hpp"
#include "interp_masm_arm.hpp"
#include "interpreter/interpreter.hpp"
#include "interpreter/interpreterRuntime.hpp"
#include "logging/log.hpp"
#include "oops/arrayOop.hpp"
#include "oops/markOop.hpp"
#include "oops/method.hpp"
#include "oops/methodData.hpp"
#include "prims/jvmtiExport.hpp"
#include "prims/jvmtiThreadState.hpp"
#include "runtime/basicLock.hpp"
#include "runtime/biasedLocking.hpp"
#include "runtime/sharedRuntime.hpp"

#if INCLUDE_ALL_GCS
#include "gc/g1/g1CollectedHeap.inline.hpp"
#include "gc/g1/g1SATBCardTableModRefBS.hpp"
#include "gc/g1/heapRegion.hpp"
#endif // INCLUDE_ALL_GCS

//--------------------------------------------------------------------
// Implementation of InterpreterMacroAssembler




InterpreterMacroAssembler::InterpreterMacroAssembler(CodeBuffer* code) : MacroAssembler(code) {
}

void InterpreterMacroAssembler::call_VM_helper(Register oop_result, address entry_point, int number_of_arguments, bool check_exceptions) {
#if defined(ASSERT) && !defined(AARCH64)
  // Ensure that last_sp is not filled.
  { Label L;
    ldr(Rtemp, Address(FP, frame::interpreter_frame_last_sp_offset * wordSize));
    cbz(Rtemp, L);
    stop("InterpreterMacroAssembler::call_VM_helper: last_sp != NULL");
    bind(L);
  }
#endif // ASSERT && !AARCH64

  // Rbcp must be saved/restored since it may change due to GC.
  save_bcp();

#ifdef AARCH64
  check_no_cached_stack_top(Rtemp);
  save_stack_top();
  check_extended_sp(Rtemp);
  cut_sp_before_call();
#endif // AARCH64

  // super call
  MacroAssembler::call_VM_helper(oop_result, entry_point, number_of_arguments, check_exceptions);

#ifdef AARCH64
  // Restore SP to extended SP
  restore_sp_after_call(Rtemp);
  check_stack_top();
  clear_cached_stack_top();
#endif // AARCH64

  // Restore interpreter specific registers.
  restore_bcp();
  restore_method();
}

void InterpreterMacroAssembler::jump_to_entry(address entry) {
  assert(entry, "Entry must have been generated by now");
  b(entry);
}

void InterpreterMacroAssembler::check_and_handle_popframe() {
  if (can_pop_frame()) {
    Label L;
    const Register popframe_cond = R2_tmp;

    // Initiate popframe handling only if it is not already being processed.  If the flag
    // has the popframe_processing bit set, it means that this code is called *during* popframe
    // handling - we don't want to reenter.

    ldr_s32(popframe_cond, Address(Rthread, JavaThread::popframe_condition_offset()));
    tbz(popframe_cond, exact_log2(JavaThread::popframe_pending_bit), L);
    tbnz(popframe_cond, exact_log2(JavaThread::popframe_processing_bit), L);

    // Call Interpreter::remove_activation_preserving_args_entry() to get the
    // address of the same-named entrypoint in the generated interpreter code.
    call_VM_leaf(CAST_FROM_FN_PTR(address, Interpreter::remove_activation_preserving_args_entry));

    // Call indirectly to avoid generation ordering problem.
    jump(R0);

    bind(L);
  }
}


// Blows R2, Rtemp. Sets TOS cached value.
void InterpreterMacroAssembler::load_earlyret_value(TosState state) {
  const Register thread_state = R2_tmp;

  ldr(thread_state, Address(Rthread, JavaThread::jvmti_thread_state_offset()));

  const Address tos_addr(thread_state, JvmtiThreadState::earlyret_tos_offset());
  const Address oop_addr(thread_state, JvmtiThreadState::earlyret_oop_offset());
  const Address val_addr(thread_state, JvmtiThreadState::earlyret_value_offset());
#ifndef AARCH64
  const Address val_addr_hi(thread_state, JvmtiThreadState::earlyret_value_offset()
                             + in_ByteSize(wordSize));
#endif // !AARCH64

  Register zero = zero_register(Rtemp);

  switch (state) {
    case atos: ldr(R0_tos, oop_addr);
               str(zero, oop_addr);
               interp_verify_oop(R0_tos, state, __FILE__, __LINE__);
               break;

#ifdef AARCH64
    case ltos: ldr(R0_tos, val_addr);              break;
#else
    case ltos: ldr(R1_tos_hi, val_addr_hi);        // fall through
#endif // AARCH64
    case btos:                                     // fall through
    case ztos:                                     // fall through
    case ctos:                                     // fall through
    case stos:                                     // fall through
    case itos: ldr_s32(R0_tos, val_addr);          break;
#ifdef __SOFTFP__
    case dtos: ldr(R1_tos_hi, val_addr_hi);        // fall through
    case ftos: ldr(R0_tos, val_addr);              break;
#else
    case ftos: ldr_float (S0_tos, val_addr);       break;
    case dtos: ldr_double(D0_tos, val_addr);       break;
#endif // __SOFTFP__
    case vtos: /* nothing to do */                 break;
    default  : ShouldNotReachHere();
  }
  // Clean up tos value in the thread object
  str(zero, val_addr);
#ifndef AARCH64
  str(zero, val_addr_hi);
#endif // !AARCH64

  mov(Rtemp, (int) ilgl);
  str_32(Rtemp, tos_addr);
}


// Blows R2, Rtemp.
void InterpreterMacroAssembler::check_and_handle_earlyret() {
  if (can_force_early_return()) {
    Label L;
    const Register thread_state = R2_tmp;

    ldr(thread_state, Address(Rthread, JavaThread::jvmti_thread_state_offset()));
    cbz(thread_state, L); // if (thread->jvmti_thread_state() == NULL) exit;

    // Initiate earlyret handling only if it is not already being processed.
    // If the flag has the earlyret_processing bit set, it means that this code
    // is called *during* earlyret handling - we don't want to reenter.

    ldr_s32(Rtemp, Address(thread_state, JvmtiThreadState::earlyret_state_offset()));
    cmp(Rtemp, JvmtiThreadState::earlyret_pending);
    b(L, ne);

    // Call Interpreter::remove_activation_early_entry() to get the address of the
    // same-named entrypoint in the generated interpreter code.

    ldr_s32(R0, Address(thread_state, JvmtiThreadState::earlyret_tos_offset()));
    call_VM_leaf(CAST_FROM_FN_PTR(address, Interpreter::remove_activation_early_entry), R0);

    jump(R0);

    bind(L);
  }
}


// Sets reg. Blows Rtemp.
void InterpreterMacroAssembler::get_unsigned_2_byte_index_at_bcp(Register reg, int bcp_offset) {
  assert(bcp_offset >= 0, "bcp is still pointing to start of bytecode");
  assert(reg != Rtemp, "should be different registers");

  ldrb(Rtemp, Address(Rbcp, bcp_offset));
  ldrb(reg, Address(Rbcp, bcp_offset+1));
  orr(reg, reg, AsmOperand(Rtemp, lsl, BitsPerByte));
}

void InterpreterMacroAssembler::get_index_at_bcp(Register index, int bcp_offset, Register tmp_reg, size_t index_size) {
  assert_different_registers(index, tmp_reg);
  if (index_size == sizeof(u2)) {
    // load bytes of index separately to avoid unaligned access
    ldrb(index, Address(Rbcp, bcp_offset+1));
    ldrb(tmp_reg, Address(Rbcp, bcp_offset));
    orr(index, tmp_reg, AsmOperand(index, lsl, BitsPerByte));
  } else if (index_size == sizeof(u4)) {
    // TODO-AARCH64: consider using unaligned access here
    ldrb(index, Address(Rbcp, bcp_offset+3));
    ldrb(tmp_reg, Address(Rbcp, bcp_offset+2));
    orr(index, tmp_reg, AsmOperand(index, lsl, BitsPerByte));
    ldrb(tmp_reg, Address(Rbcp, bcp_offset+1));
    orr(index, tmp_reg, AsmOperand(index, lsl, BitsPerByte));
    ldrb(tmp_reg, Address(Rbcp, bcp_offset));
    orr(index, tmp_reg, AsmOperand(index, lsl, BitsPerByte));
    // Check if the secondary index definition is still ~x, otherwise
    // we have to change the following assembler code to calculate the
    // plain index.
    assert(ConstantPool::decode_invokedynamic_index(~123) == 123, "else change next line");
    mvn_32(index, index);  // convert to plain index
  } else if (index_size == sizeof(u1)) {
    ldrb(index, Address(Rbcp, bcp_offset));
  } else {
    ShouldNotReachHere();
  }
}

// Sets cache, index.
void InterpreterMacroAssembler::get_cache_and_index_at_bcp(Register cache, Register index, int bcp_offset, size_t index_size) {
  assert(bcp_offset > 0, "bcp is still pointing to start of bytecode");
  assert_different_registers(cache, index);

  get_index_at_bcp(index, bcp_offset, cache, index_size);

  // load constant pool cache pointer
  ldr(cache, Address(FP, frame::interpreter_frame_cache_offset * wordSize));

  // convert from field index to ConstantPoolCacheEntry index
  assert(sizeof(ConstantPoolCacheEntry) == 4*wordSize, "adjust code below");
  // TODO-AARCH64 merge this shift with shift "add(..., Rcache, AsmOperand(Rindex, lsl, LogBytesPerWord))" after this method is called
  logical_shift_left(index, index, 2);
}

// Sets cache, index, bytecode.
void InterpreterMacroAssembler::get_cache_and_index_and_bytecode_at_bcp(Register cache, Register index, Register bytecode, int byte_no, int bcp_offset, size_t index_size) {
  get_cache_and_index_at_bcp(cache, index, bcp_offset, index_size);
  // caution index and bytecode can be the same
  add(bytecode, cache, AsmOperand(index, lsl, LogBytesPerWord));
#ifdef AARCH64
  add(bytecode, bytecode, (1 + byte_no) + in_bytes(ConstantPoolCache::base_offset() + ConstantPoolCacheEntry::indices_offset()));
  ldarb(bytecode, bytecode);
#else
  ldrb(bytecode, Address(bytecode, (1 + byte_no) + in_bytes(ConstantPoolCache::base_offset() + ConstantPoolCacheEntry::indices_offset())));
  TemplateTable::volatile_barrier(MacroAssembler::LoadLoad, noreg, true);
#endif // AARCH64
}

// Sets cache. Blows reg_tmp.
void InterpreterMacroAssembler::get_cache_entry_pointer_at_bcp(Register cache, Register reg_tmp, int bcp_offset, size_t index_size) {
  assert(bcp_offset > 0, "bcp is still pointing to start of bytecode");
  assert_different_registers(cache, reg_tmp);

  get_index_at_bcp(reg_tmp, bcp_offset, cache, index_size);

  // load constant pool cache pointer
  ldr(cache, Address(FP, frame::interpreter_frame_cache_offset * wordSize));

  // skip past the header
  add(cache, cache, in_bytes(ConstantPoolCache::base_offset()));
  // convert from field index to ConstantPoolCacheEntry index
  // and from word offset to byte offset
  assert(sizeof(ConstantPoolCacheEntry) == 4*wordSize, "adjust code below");
  add(cache, cache, AsmOperand(reg_tmp, lsl, 2 + LogBytesPerWord));
}

// Load object from cpool->resolved_references(index)
void InterpreterMacroAssembler::load_resolved_reference_at_index(
                                           Register result, Register index) {
  assert_different_registers(result, index);
  get_constant_pool(result);

  Register cache = result;
  // load pointer for resolved_references[] objArray
  ldr(cache, Address(result, ConstantPool::resolved_references_offset_in_bytes()));
  // JNIHandles::resolve(result)
  ldr(cache, Address(cache, 0));
  // Add in the index
  // convert from field index to resolved_references() index and from
  // word index to byte offset. Since this is a java object, it can be compressed
  add(cache, cache, AsmOperand(index, lsl, LogBytesPerHeapOop));
  load_heap_oop(result, Address(cache, arrayOopDesc::base_offset_in_bytes(T_OBJECT)));
}

// Generate a subtype check: branch to not_subtype if sub_klass is
// not a subtype of super_klass.
// Profiling code for the subtype check failure (profile_typecheck_failed)
// should be explicitly generated by the caller in the not_subtype case.
// Blows Rtemp, tmp1, tmp2.
void InterpreterMacroAssembler::gen_subtype_check(Register Rsub_klass,
                                                  Register Rsuper_klass,
                                                  Label &not_subtype,
                                                  Register tmp1,
                                                  Register tmp2) {

  assert_different_registers(Rsub_klass, Rsuper_klass, tmp1, tmp2, Rtemp);
  Label ok_is_subtype, loop, update_cache;

  const Register super_check_offset = tmp1;
  const Register cached_super = tmp2;

  // Profile the not-null value's klass.
  profile_typecheck(tmp1, Rsub_klass);

  // Load the super-klass's check offset into
  ldr_u32(super_check_offset, Address(Rsuper_klass, Klass::super_check_offset_offset()));

  // Check for self
  cmp(Rsub_klass, Rsuper_klass);

  // Load from the sub-klass's super-class display list, or a 1-word cache of
  // the secondary superclass list, or a failing value with a sentinel offset
  // if the super-klass is an interface or exceptionally deep in the Java
  // hierarchy and we have to scan the secondary superclass list the hard way.
  // See if we get an immediate positive hit
  ldr(cached_super, Address(Rsub_klass, super_check_offset));

  cond_cmp(Rsuper_klass, cached_super, ne);
  b(ok_is_subtype, eq);

  // Check for immediate negative hit
  cmp(super_check_offset, in_bytes(Klass::secondary_super_cache_offset()));
  b(not_subtype, ne);

  // Now do a linear scan of the secondary super-klass chain.
  const Register supers_arr = tmp1;
  const Register supers_cnt = tmp2;
  const Register cur_super  = Rtemp;

  // Load objArrayOop of secondary supers.
  ldr(supers_arr, Address(Rsub_klass, Klass::secondary_supers_offset()));

  ldr_u32(supers_cnt, Address(supers_arr, Array<Klass*>::length_offset_in_bytes())); // Load the array length
#ifdef AARCH64
  cbz(supers_cnt, not_subtype);
  add(supers_arr, supers_arr, Array<Klass*>::base_offset_in_bytes());
#else
  cmp(supers_cnt, 0);

  // Skip to the start of array elements and prefetch the first super-klass.
  ldr(cur_super, Address(supers_arr, Array<Klass*>::base_offset_in_bytes(), pre_indexed), ne);
  b(not_subtype, eq);
#endif // AARCH64

  bind(loop);

#ifdef AARCH64
  ldr(cur_super, Address(supers_arr, wordSize, post_indexed));
#endif // AARCH64

  cmp(cur_super, Rsuper_klass);
  b(update_cache, eq);

  subs(supers_cnt, supers_cnt, 1);

#ifndef AARCH64
  ldr(cur_super, Address(supers_arr, wordSize, pre_indexed), ne);
#endif // !AARCH64

  b(loop, ne);

  b(not_subtype);

  bind(update_cache);
  // Must be equal but missed in cache.  Update cache.
  str(Rsuper_klass, Address(Rsub_klass, Klass::secondary_super_cache_offset()));

  bind(ok_is_subtype);
}


// The 1st part of the store check.
// Sets card_table_base register.
void InterpreterMacroAssembler::store_check_part1(Register card_table_base) {
  // Check barrier set type (should be card table) and element size
  BarrierSet* bs = Universe::heap()->barrier_set();
  assert(bs->kind() == BarrierSet::CardTableForRS ||
         bs->kind() == BarrierSet::CardTableExtension,
         "Wrong barrier set kind");

  CardTableModRefBS* ct = barrier_set_cast<CardTableModRefBS>(bs);
  assert(sizeof(*ct->byte_map_base) == sizeof(jbyte), "Adjust store check code");

  // Load card table base address.

  /* Performance note.

     There is an alternative way of loading card table base address
     from thread descriptor, which may look more efficient:

     ldr(card_table_base, Address(Rthread, JavaThread::card_table_base_offset()));

     However, performance measurements of micro benchmarks and specJVM98
     showed that loading of card table base from thread descriptor is
     7-18% slower compared to loading of literal embedded into the code.
     Possible cause is a cache miss (card table base address resides in a
     rarely accessed area of thread descriptor).
  */
  // TODO-AARCH64 Investigate if mov_slow is faster than ldr from Rthread on AArch64
  mov_address(card_table_base, (address)ct->byte_map_base, symbolic_Relocation::card_table_reference);
}

// The 2nd part of the store check.
void InterpreterMacroAssembler::store_check_part2(Register obj, Register card_table_base, Register tmp) {
  assert_different_registers(obj, card_table_base, tmp);

  assert(CardTableModRefBS::dirty_card_val() == 0, "Dirty card value must be 0 due to optimizations.");
#ifdef AARCH64
  add(card_table_base, card_table_base, AsmOperand(obj, lsr, CardTableModRefBS::card_shift));
  Address card_table_addr(card_table_base);
#else
  Address card_table_addr(card_table_base, obj, lsr, CardTableModRefBS::card_shift);
#endif

  if (UseCondCardMark) {
    if (UseConcMarkSweepGC) {
      membar(MacroAssembler::Membar_mask_bits(MacroAssembler::StoreLoad), noreg);
    }
    Label already_dirty;

    ldrb(tmp, card_table_addr);
    cbz(tmp, already_dirty);

    set_card(card_table_base, card_table_addr, tmp);
    bind(already_dirty);

  } else {
    if (UseConcMarkSweepGC && CMSPrecleaningEnabled) {
      membar(MacroAssembler::Membar_mask_bits(MacroAssembler::StoreStore), noreg);
    }
    set_card(card_table_base, card_table_addr, tmp);
  }
}

void InterpreterMacroAssembler::set_card(Register card_table_base, Address card_table_addr, Register tmp) {
#ifdef AARCH64
  strb(ZR, card_table_addr);
#else
  CardTableModRefBS* ct = barrier_set_cast<CardTableModRefBS>(Universe::heap()->barrier_set());
  if ((((uintptr_t)ct->byte_map_base & 0xff) == 0)) {
    // Card table is aligned so the lowest byte of the table address base is zero.
    // This works only if the code is not saved for later use, possibly
    // in a context where the base would no longer be aligned.
    strb(card_table_base, card_table_addr);
  } else {
    mov(tmp, 0);
    strb(tmp, card_table_addr);
  }
#endif // AARCH64
}

//////////////////////////////////////////////////////////////////////////////////


// Java Expression Stack

void InterpreterMacroAssembler::pop_ptr(Register r) {
  assert(r != Rstack_top, "unpredictable instruction");
  ldr(r, Address(Rstack_top, wordSize, post_indexed));
}

void InterpreterMacroAssembler::pop_i(Register r) {
  assert(r != Rstack_top, "unpredictable instruction");
  ldr_s32(r, Address(Rstack_top, wordSize, post_indexed));
  zap_high_non_significant_bits(r);
}

#ifdef AARCH64
void InterpreterMacroAssembler::pop_l(Register r) {
  assert(r != Rstack_top, "unpredictable instruction");
  ldr(r, Address(Rstack_top, 2*wordSize, post_indexed));
}
#else
void InterpreterMacroAssembler::pop_l(Register lo, Register hi) {
  assert_different_registers(lo, hi);
  assert(lo < hi, "lo must be < hi");
  pop(RegisterSet(lo) | RegisterSet(hi));
}
#endif // AARCH64

void InterpreterMacroAssembler::pop_f(FloatRegister fd) {
#ifdef AARCH64
  ldr_s(fd, Address(Rstack_top, wordSize, post_indexed));
#else
  fpops(fd);
#endif // AARCH64
}

void InterpreterMacroAssembler::pop_d(FloatRegister fd) {
#ifdef AARCH64
  ldr_d(fd, Address(Rstack_top, 2*wordSize, post_indexed));
#else
  fpopd(fd);
#endif // AARCH64
}


// Transition vtos -> state. Blows R0, R1. Sets TOS cached value.
void InterpreterMacroAssembler::pop(TosState state) {
  switch (state) {
    case atos: pop_ptr(R0_tos);                              break;
    case btos:                                               // fall through
    case ztos:                                               // fall through
    case ctos:                                               // fall through
    case stos:                                               // fall through
    case itos: pop_i(R0_tos);                                break;
#ifdef AARCH64
    case ltos: pop_l(R0_tos);                                break;
#else
    case ltos: pop_l(R0_tos_lo, R1_tos_hi);                  break;
#endif // AARCH64
#ifdef __SOFTFP__
    case ftos: pop_i(R0_tos);                                break;
    case dtos: pop_l(R0_tos_lo, R1_tos_hi);                  break;
#else
    case ftos: pop_f(S0_tos);                                break;
    case dtos: pop_d(D0_tos);                                break;
#endif // __SOFTFP__
    case vtos: /* nothing to do */                           break;
    default  : ShouldNotReachHere();
  }
  interp_verify_oop(R0_tos, state, __FILE__, __LINE__);
}

void InterpreterMacroAssembler::push_ptr(Register r) {
  assert(r != Rstack_top, "unpredictable instruction");
  str(r, Address(Rstack_top, -wordSize, pre_indexed));
  check_stack_top_on_expansion();
}

void InterpreterMacroAssembler::push_i(Register r) {
  assert(r != Rstack_top, "unpredictable instruction");
  str_32(r, Address(Rstack_top, -wordSize, pre_indexed));
  check_stack_top_on_expansion();
}

#ifdef AARCH64
void InterpreterMacroAssembler::push_l(Register r) {
  assert(r != Rstack_top, "unpredictable instruction");
  stp(r, ZR, Address(Rstack_top, -2*wordSize, pre_indexed));
  check_stack_top_on_expansion();
}
#else
void InterpreterMacroAssembler::push_l(Register lo, Register hi) {
  assert_different_registers(lo, hi);
  assert(lo < hi, "lo must be < hi");
  push(RegisterSet(lo) | RegisterSet(hi));
}
#endif // AARCH64

void InterpreterMacroAssembler::push_f() {
#ifdef AARCH64
  str_s(S0_tos, Address(Rstack_top, -wordSize, pre_indexed));
  check_stack_top_on_expansion();
#else
  fpushs(S0_tos);
#endif // AARCH64
}

void InterpreterMacroAssembler::push_d() {
#ifdef AARCH64
  str_d(D0_tos, Address(Rstack_top, -2*wordSize, pre_indexed));
  check_stack_top_on_expansion();
#else
  fpushd(D0_tos);
#endif // AARCH64
}

// Transition state -> vtos. Blows Rtemp.
void InterpreterMacroAssembler::push(TosState state) {
  interp_verify_oop(R0_tos, state, __FILE__, __LINE__);
  switch (state) {
    case atos: push_ptr(R0_tos);                              break;
    case btos:                                                // fall through
    case ztos:                                                // fall through
    case ctos:                                                // fall through
    case stos:                                                // fall through
    case itos: push_i(R0_tos);                                break;
#ifdef AARCH64
    case ltos: push_l(R0_tos);                                break;
#else
    case ltos: push_l(R0_tos_lo, R1_tos_hi);                  break;
#endif // AARCH64
#ifdef __SOFTFP__
    case ftos: push_i(R0_tos);                                break;
    case dtos: push_l(R0_tos_lo, R1_tos_hi);                  break;
#else
    case ftos: push_f();                                      break;
    case dtos: push_d();                                      break;
#endif // __SOFTFP__
    case vtos: /* nothing to do */                            break;
    default  : ShouldNotReachHere();
  }
}


#ifndef AARCH64

// Converts return value in R0/R1 (interpreter calling conventions) to TOS cached value.
void InterpreterMacroAssembler::convert_retval_to_tos(TosState state) {
#if (!defined __SOFTFP__ && !defined __ABI_HARD__)
  // According to interpreter calling conventions, result is returned in R0/R1,
  // but templates expect ftos in S0, and dtos in D0.
  if (state == ftos) {
    fmsr(S0_tos, R0);
  } else if (state == dtos) {
    fmdrr(D0_tos, R0, R1);
  }
#endif // !__SOFTFP__ && !__ABI_HARD__
}

// Converts TOS cached value to return value in R0/R1 (according to interpreter calling conventions).
void InterpreterMacroAssembler::convert_tos_to_retval(TosState state) {
#if (!defined __SOFTFP__ && !defined __ABI_HARD__)
  // According to interpreter calling conventions, result is returned in R0/R1,
  // so ftos (S0) and dtos (D0) are moved to R0/R1.
  if (state == ftos) {
    fmrs(R0, S0_tos);
  } else if (state == dtos) {
    fmrrd(R0, R1, D0_tos);
  }
#endif // !__SOFTFP__ && !__ABI_HARD__
}

#endif // !AARCH64


// Helpers for swap and dup
void InterpreterMacroAssembler::load_ptr(int n, Register val) {
  ldr(val, Address(Rstack_top, Interpreter::expr_offset_in_bytes(n)));
}

void InterpreterMacroAssembler::store_ptr(int n, Register val) {
  str(val, Address(Rstack_top, Interpreter::expr_offset_in_bytes(n)));
}


void InterpreterMacroAssembler::prepare_to_jump_from_interpreted() {
#ifdef AARCH64
  check_no_cached_stack_top(Rtemp);
  save_stack_top();
  cut_sp_before_call();
  mov(Rparams, Rstack_top);
#endif // AARCH64

  // set sender sp
  mov(Rsender_sp, SP);

#ifndef AARCH64
  // record last_sp
  str(Rsender_sp, Address(FP, frame::interpreter_frame_last_sp_offset * wordSize));
#endif // !AARCH64
}

// Jump to from_interpreted entry of a call unless single stepping is possible
// in this thread in which case we must call the i2i entry
void InterpreterMacroAssembler::jump_from_interpreted(Register method) {
  assert_different_registers(method, Rtemp);

  prepare_to_jump_from_interpreted();

  if (can_post_interpreter_events()) {
    // JVMTI events, such as single-stepping, are implemented partly by avoiding running
    // compiled code in threads for which the event is enabled.  Check here for
    // interp_only_mode if these events CAN be enabled.

    ldr_s32(Rtemp, Address(Rthread, JavaThread::interp_only_mode_offset()));
#ifdef AARCH64
    {
      Label not_interp_only_mode;

      cbz(Rtemp, not_interp_only_mode);
      indirect_jump(Address(method, Method::interpreter_entry_offset()), Rtemp);

      bind(not_interp_only_mode);
    }
#else
    cmp(Rtemp, 0);
    ldr(PC, Address(method, Method::interpreter_entry_offset()), ne);
#endif // AARCH64
  }

  indirect_jump(Address(method, Method::from_interpreted_offset()), Rtemp);
}


void InterpreterMacroAssembler::restore_dispatch() {
  mov_slow(RdispatchTable, (address)Interpreter::dispatch_table(vtos));
}


// The following two routines provide a hook so that an implementation
// can schedule the dispatch in two parts.
void InterpreterMacroAssembler::dispatch_prolog(TosState state, int step) {
  // Nothing ARM-specific to be done here.
}

void InterpreterMacroAssembler::dispatch_epilog(TosState state, int step) {
  dispatch_next(state, step);
}

void InterpreterMacroAssembler::dispatch_base(TosState state,
                                              DispatchTableMode table_mode,
                                              bool verifyoop) {
  if (VerifyActivationFrameSize) {
    Label L;
#ifdef AARCH64
    mov(Rtemp, SP);
    sub(Rtemp, FP, Rtemp);
#else
    sub(Rtemp, FP, SP);
#endif // AARCH64
    int min_frame_size = (frame::link_offset - frame::interpreter_frame_initial_sp_offset) * wordSize;
    cmp(Rtemp, min_frame_size);
    b(L, ge);
    stop("broken stack frame");
    bind(L);
  }

  if (verifyoop) {
    interp_verify_oop(R0_tos, state, __FILE__, __LINE__);
  }

  if((state == itos) || (state == btos) || (state == ztos) || (state == ctos) || (state == stos)) {
    zap_high_non_significant_bits(R0_tos);
  }

#ifdef ASSERT
  Label L;
  mov_slow(Rtemp, (address)Interpreter::dispatch_table(vtos));
  cmp(Rtemp, RdispatchTable);
  b(L, eq);
  stop("invalid RdispatchTable");
  bind(L);
#endif

  if (table_mode == DispatchDefault) {
    if (state == vtos) {
      indirect_jump(Address::indexed_ptr(RdispatchTable, R3_bytecode), Rtemp);
    } else {
#ifdef AARCH64
      sub(Rtemp, R3_bytecode, (Interpreter::distance_from_dispatch_table(vtos) -
                           Interpreter::distance_from_dispatch_table(state)));
      indirect_jump(Address::indexed_ptr(RdispatchTable, Rtemp), Rtemp);
#else
      // on 32-bit ARM this method is faster than the one above.
      sub(Rtemp, RdispatchTable, (Interpreter::distance_from_dispatch_table(vtos) -
                           Interpreter::distance_from_dispatch_table(state)) * wordSize);
      indirect_jump(Address::indexed_ptr(Rtemp, R3_bytecode), Rtemp);
#endif
    }
  } else {
    assert(table_mode == DispatchNormal, "invalid dispatch table mode");
    address table = (address) Interpreter::normal_table(state);
    mov_slow(Rtemp, table);
    indirect_jump(Address::indexed_ptr(Rtemp, R3_bytecode), Rtemp);
  }

  nop(); // to avoid filling CPU pipeline with invalid instructions
  nop();
}

void InterpreterMacroAssembler::dispatch_only(TosState state) {
  dispatch_base(state, DispatchDefault);
}


void InterpreterMacroAssembler::dispatch_only_normal(TosState state) {
  dispatch_base(state, DispatchNormal);
}

void InterpreterMacroAssembler::dispatch_only_noverify(TosState state) {
  dispatch_base(state, DispatchNormal, false);
}

void InterpreterMacroAssembler::dispatch_next(TosState state, int step) {
  // load next bytecode and advance Rbcp
  ldrb(R3_bytecode, Address(Rbcp, step, pre_indexed));
  dispatch_base(state, DispatchDefault);
}

void InterpreterMacroAssembler::narrow(Register result) {
  // mask integer result to narrower return type.
  const Register Rtmp = R2;

  // get method type
  ldr(Rtmp, Address(Rmethod, Method::const_offset()));
  ldrb(Rtmp, Address(Rtmp, ConstMethod::result_type_offset()));

  Label notBool, notByte, notChar, done;
  cmp(Rtmp, T_INT);
  b(done, eq);

  cmp(Rtmp, T_BOOLEAN);
  b(notBool, ne);
  and_32(result, result, 1);
  b(done);

  bind(notBool);
  cmp(Rtmp, T_BYTE);
  b(notByte, ne);
  sign_extend(result, result, 8);
  b(done);

  bind(notByte);
  cmp(Rtmp, T_CHAR);
  b(notChar, ne);
  zero_extend(result, result, 16);
  b(done);

  bind(notChar);
  // cmp(Rtmp, T_SHORT);
  // b(done, ne);
  sign_extend(result, result, 16);

  // Nothing to do
  bind(done);
}

// remove activation
//
// Unlock the receiver if this is a synchronized method.
// Unlock any Java monitors from syncronized blocks.
// Remove the activation from the stack.
//
// If there are locked Java monitors
//    If throw_monitor_exception
//       throws IllegalMonitorStateException
//    Else if install_monitor_exception
//       installs IllegalMonitorStateException
//    Else
//       no error processing
void InterpreterMacroAssembler::remove_activation(TosState state, Register ret_addr,
                                                  bool throw_monitor_exception,
                                                  bool install_monitor_exception,
                                                  bool notify_jvmdi) {
  Label unlock, unlocked, no_unlock;

  // Note: Registers R0, R1, S0 and D0 (TOS cached value) may be in use for the result.

  const Address do_not_unlock_if_synchronized(Rthread,
                         JavaThread::do_not_unlock_if_synchronized_offset());

  const Register Rflag = R2;
  const Register Raccess_flags = R3;

  restore_method();

  ldrb(Rflag, do_not_unlock_if_synchronized);

  // get method access flags
  ldr_u32(Raccess_flags, Address(Rmethod, Method::access_flags_offset()));

  strb(zero_register(Rtemp), do_not_unlock_if_synchronized); // reset the flag

  // check if method is synchronized

  tbz(Raccess_flags, JVM_ACC_SYNCHRONIZED_BIT, unlocked);

  // Don't unlock anything if the _do_not_unlock_if_synchronized flag is set.
  cbnz(Rflag, no_unlock);

  // unlock monitor
  push(state);                                   // save result

  // BasicObjectLock will be first in list, since this is a synchronized method. However, need
  // to check that the object has not been unlocked by an explicit monitorexit bytecode.

  const Register Rmonitor = R1;                  // fixed in unlock_object()
  const Register Robj = R2;

  // address of first monitor
  sub(Rmonitor, FP, - frame::interpreter_frame_monitor_block_bottom_offset * wordSize + (int)sizeof(BasicObjectLock));

  ldr(Robj, Address(Rmonitor, BasicObjectLock::obj_offset_in_bytes()));
  cbnz(Robj, unlock);

  pop(state);

  if (throw_monitor_exception) {
    // Entry already unlocked, need to throw exception
    call_VM(noreg, CAST_FROM_FN_PTR(address, InterpreterRuntime::throw_illegal_monitor_state_exception));
    should_not_reach_here();
  } else {
    // Monitor already unlocked during a stack unroll.
    // If requested, install an illegal_monitor_state_exception.
    // Continue with stack unrolling.
    if (install_monitor_exception) {
      call_VM(noreg, CAST_FROM_FN_PTR(address, InterpreterRuntime::new_illegal_monitor_state_exception));
    }
    b(unlocked);
  }


  // Exception case for the check that all monitors are unlocked.
  const Register Rcur = R2;
  Label restart_check_monitors_unlocked, exception_monitor_is_still_locked;

  bind(exception_monitor_is_still_locked);
  // Monitor entry is still locked, need to throw exception.
  // Rcur: monitor entry.

  if (throw_monitor_exception) {
    // Throw exception
    call_VM(noreg, CAST_FROM_FN_PTR(address, InterpreterRuntime::throw_illegal_monitor_state_exception));
    should_not_reach_here();
  } else {
    // Stack unrolling. Unlock object and install illegal_monitor_exception
    // Unlock does not block, so don't have to worry about the frame

    push(state);
    mov(R1, Rcur);
    unlock_object(R1);

    if (install_monitor_exception) {
      call_VM(noreg, CAST_FROM_FN_PTR(address, InterpreterRuntime::new_illegal_monitor_state_exception));
    }

    pop(state);
    b(restart_check_monitors_unlocked);
  }

  bind(unlock);
  unlock_object(Rmonitor);
  pop(state);

  // Check that for block-structured locking (i.e., that all locked objects has been unlocked)
  bind(unlocked);

  // Check that all monitors are unlocked
  {
    Label loop;

    const int entry_size = frame::interpreter_frame_monitor_size() * wordSize;
    const Register Rbottom = R3;
    const Register Rcur_obj = Rtemp;

    bind(restart_check_monitors_unlocked);

    ldr(Rcur, Address(FP, frame::interpreter_frame_monitor_block_top_offset * wordSize));
                                 // points to current entry, starting with top-most entry
    sub(Rbottom, FP, -frame::interpreter_frame_monitor_block_bottom_offset * wordSize);
                                 // points to word before bottom of monitor block

    cmp(Rcur, Rbottom);          // check if there are no monitors
#ifndef AARCH64
    ldr(Rcur_obj, Address(Rcur, BasicObjectLock::obj_offset_in_bytes()), ne);
                                 // prefetch monitor's object
#endif // !AARCH64
    b(no_unlock, eq);

    bind(loop);
#ifdef AARCH64
    ldr(Rcur_obj, Address(Rcur, BasicObjectLock::obj_offset_in_bytes()));
#endif // AARCH64
    // check if current entry is used
    cbnz(Rcur_obj, exception_monitor_is_still_locked);

    add(Rcur, Rcur, entry_size);      // otherwise advance to next entry
    cmp(Rcur, Rbottom);               // check if bottom reached
#ifndef AARCH64
    ldr(Rcur_obj, Address(Rcur, BasicObjectLock::obj_offset_in_bytes()), ne);
                                      // prefetch monitor's object
#endif // !AARCH64
    b(loop, ne);                      // if not at bottom then check this entry
  }

  bind(no_unlock);

  // jvmti support
  if (notify_jvmdi) {
    notify_method_exit(state, NotifyJVMTI);     // preserve TOSCA
  } else {
    notify_method_exit(state, SkipNotifyJVMTI); // preserve TOSCA
  }

  // remove activation
#ifdef AARCH64
  ldr(Rtemp, Address(FP, frame::interpreter_frame_sender_sp_offset * wordSize));
  ldp(FP, LR, Address(FP));
  mov(SP, Rtemp);
#else
  mov(Rtemp, FP);
  ldmia(FP, RegisterSet(FP) | RegisterSet(LR));
  ldr(SP, Address(Rtemp, frame::interpreter_frame_sender_sp_offset * wordSize));
#endif

  if (ret_addr != LR) {
    mov(ret_addr, LR);
  }
}


// At certain points in the method invocation the monitor of
// synchronized methods hasn't been entered yet.
// To correctly handle exceptions at these points, we set the thread local
// variable _do_not_unlock_if_synchronized to true. The remove_activation will
// check this flag.
void InterpreterMacroAssembler::set_do_not_unlock_if_synchronized(bool flag, Register tmp) {
  const Address do_not_unlock_if_synchronized(Rthread,
                         JavaThread::do_not_unlock_if_synchronized_offset());
  if (flag) {
    mov(tmp, 1);
    strb(tmp, do_not_unlock_if_synchronized);
  } else {
    strb(zero_register(tmp), do_not_unlock_if_synchronized);
  }
}

// Lock object
//
// Argument: R1 : Points to BasicObjectLock to be used for locking.
// Must be initialized with object to lock.
// Blows volatile registers (R0-R3 on 32-bit ARM, R0-R18 on AArch64), Rtemp, LR. Calls VM.
void InterpreterMacroAssembler::lock_object(Register Rlock) {
  assert(Rlock == R1, "the second argument");

  if (UseHeavyMonitors) {
    call_VM(noreg, CAST_FROM_FN_PTR(address, InterpreterRuntime::monitorenter), Rlock);
  } else {
    Label done;

    const Register Robj = R2;
    const Register Rmark = R3;
    assert_different_registers(Robj, Rmark, Rlock, R0, Rtemp);

    const int obj_offset = BasicObjectLock::obj_offset_in_bytes();
    const int lock_offset = BasicObjectLock::lock_offset_in_bytes ();
    const int mark_offset = lock_offset + BasicLock::displaced_header_offset_in_bytes();

    Label already_locked, slow_case;

    // Load object pointer
    ldr(Robj, Address(Rlock, obj_offset));

    if (UseBiasedLocking) {
      biased_locking_enter(Robj, Rmark/*scratched*/, R0, false, Rtemp, done, slow_case);
    }

#ifdef AARCH64
    assert(oopDesc::mark_offset_in_bytes() == 0, "must be");
    ldr(Rmark, Robj);

    // Test if object is already locked
    assert(markOopDesc::unlocked_value == 1, "adjust this code");
    tbz(Rmark, exact_log2(markOopDesc::unlocked_value), already_locked);

#else // AARCH64

    // On MP platforms the next load could return a 'stale' value if the memory location has been modified by another thread.
    // That would be acceptable as ether CAS or slow case path is taken in that case.
    // Exception to that is if the object is locked by the calling thread, then the recursive test will pass (guaranteed as
    // loads are satisfied from a store queue if performed on the same processor).

    assert(oopDesc::mark_offset_in_bytes() == 0, "must be");
    ldr(Rmark, Address(Robj, oopDesc::mark_offset_in_bytes()));

    // Test if object is already locked
    tst(Rmark, markOopDesc::unlocked_value);
    b(already_locked, eq);

#endif // !AARCH64
    // Save old object->mark() into BasicLock's displaced header
    str(Rmark, Address(Rlock, mark_offset));

    cas_for_lock_acquire(Rmark, Rlock, Robj, Rtemp, slow_case);

#ifndef PRODUCT
    if (PrintBiasedLockingStatistics) {
      cond_atomic_inc32(al, BiasedLocking::fast_path_entry_count_addr());
    }
#endif //!PRODUCT

    b(done);

    // If we got here that means the object is locked by ether calling thread or another thread.
    bind(already_locked);
    // Handling of locked objects: recursive locks and slow case.

    // Fast check for recursive lock.
    //
    // Can apply the optimization only if this is a stack lock
    // allocated in this thread. For efficiency, we can focus on
    // recently allocated stack locks (instead of reading the stack
    // base and checking whether 'mark' points inside the current
    // thread stack):
    //  1) (mark & 3) == 0
    //  2) SP <= mark < SP + os::pagesize()
    //
    // Warning: SP + os::pagesize can overflow the stack base. We must
    // neither apply the optimization for an inflated lock allocated
    // just above the thread stack (this is why condition 1 matters)
    // nor apply the optimization if the stack lock is inside the stack
    // of another thread. The latter is avoided even in case of overflow
    // because we have guard pages at the end of all stacks. Hence, if
    // we go over the stack base and hit the stack of another thread,
    // this should not be in a writeable area that could contain a
    // stack lock allocated by that thread. As a consequence, a stack
    // lock less than page size away from SP is guaranteed to be
    // owned by the current thread.
    //
    // Note: assuming SP is aligned, we can check the low bits of
    // (mark-SP) instead of the low bits of mark. In that case,
    // assuming page size is a power of 2, we can merge the two
    // conditions into a single test:
    // => ((mark - SP) & (3 - os::pagesize())) == 0

#ifdef AARCH64
    // Use the single check since the immediate is OK for AARCH64
    sub(R0, Rmark, Rstack_top);
    intptr_t mask = ((intptr_t)3) - ((intptr_t)os::vm_page_size());
    Assembler::LogicalImmediate imm(mask, false);
    ands(R0, R0, imm);

    // For recursive case store 0 into lock record.
    // It is harmless to store it unconditionally as lock record contains some garbage
    // value in its _displaced_header field by this moment.
    str(ZR, Address(Rlock, mark_offset));

#else // AARCH64
    // (3 - os::pagesize()) cannot be encoded as an ARM immediate operand.
    // Check independently the low bits and the distance to SP.
    // -1- test low 2 bits
    movs(R0, AsmOperand(Rmark, lsl, 30));
    // -2- test (mark - SP) if the low two bits are 0
    sub(R0, Rmark, SP, eq);
    movs(R0, AsmOperand(R0, lsr, exact_log2(os::vm_page_size())), eq);
    // If still 'eq' then recursive locking OK: store 0 into lock record
    str(R0, Address(Rlock, mark_offset), eq);

#endif // AARCH64

#ifndef PRODUCT
    if (PrintBiasedLockingStatistics) {
      cond_atomic_inc32(eq, BiasedLocking::fast_path_entry_count_addr());
    }
#endif // !PRODUCT

    b(done, eq);

    bind(slow_case);

    // Call the runtime routine for slow case
    call_VM(noreg, CAST_FROM_FN_PTR(address, InterpreterRuntime::monitorenter), Rlock);

    bind(done);
  }
}


// Unlocks an object. Used in monitorexit bytecode and remove_activation.
//
// Argument: R1: Points to BasicObjectLock structure for lock
// Throw an IllegalMonitorException if object is not locked by current thread
// Blows volatile registers (R0-R3 on 32-bit ARM, R0-R18 on AArch64), Rtemp, LR. Calls VM.
void InterpreterMacroAssembler::unlock_object(Register Rlock) {
  assert(Rlock == R1, "the second argument");

  if (UseHeavyMonitors) {
    call_VM(noreg, CAST_FROM_FN_PTR(address, InterpreterRuntime::monitorexit), Rlock);
  } else {
    Label done, slow_case;

    const Register Robj = R2;
    const Register Rmark = R3;
    const Register Rresult = R0;
    assert_different_registers(Robj, Rmark, Rlock, R0, Rtemp);

    const int obj_offset = BasicObjectLock::obj_offset_in_bytes();
    const int lock_offset = BasicObjectLock::lock_offset_in_bytes ();
    const int mark_offset = lock_offset + BasicLock::displaced_header_offset_in_bytes();

    const Register Rzero = zero_register(Rtemp);

    // Load oop into Robj
    ldr(Robj, Address(Rlock, obj_offset));

    // Free entry
    str(Rzero, Address(Rlock, obj_offset));

    if (UseBiasedLocking) {
      biased_locking_exit(Robj, Rmark, done);
    }

    // Load the old header from BasicLock structure
    ldr(Rmark, Address(Rlock, mark_offset));

    // Test for recursion (zero mark in BasicLock)
    cbz(Rmark, done);

    bool allow_fallthrough_on_failure = true;

    cas_for_lock_release(Rlock, Rmark, Robj, Rtemp, slow_case, allow_fallthrough_on_failure);

    b(done, eq);

    bind(slow_case);

    // Call the runtime routine for slow case.
    str(Robj, Address(Rlock, obj_offset)); // restore obj
    call_VM(noreg, CAST_FROM_FN_PTR(address, InterpreterRuntime::monitorexit), Rlock);

    bind(done);
  }
}


// Test ImethodDataPtr.  If it is null, continue at the specified label
void InterpreterMacroAssembler::test_method_data_pointer(Register mdp, Label& zero_continue) {
  assert(ProfileInterpreter, "must be profiling interpreter");
  ldr(mdp, Address(FP, frame::interpreter_frame_mdp_offset * wordSize));
  cbz(mdp, zero_continue);
}


// Set the method data pointer for the current bcp.
// Blows volatile registers (R0-R3 on 32-bit ARM, R0-R18 on AArch64), Rtemp, LR.
void InterpreterMacroAssembler::set_method_data_pointer_for_bcp() {
  assert(ProfileInterpreter, "must be profiling interpreter");
  Label set_mdp;

  // Test MDO to avoid the call if it is NULL.
  ldr(Rtemp, Address(Rmethod, Method::method_data_offset()));
  cbz(Rtemp, set_mdp);

  mov(R0, Rmethod);
  mov(R1, Rbcp);
  call_VM_leaf(CAST_FROM_FN_PTR(address, InterpreterRuntime::bcp_to_di), R0, R1);
  // R0/W0: mdi

  // mdo is guaranteed to be non-zero here, we checked for it before the call.
  ldr(Rtemp, Address(Rmethod, Method::method_data_offset()));
  add(Rtemp, Rtemp, in_bytes(MethodData::data_offset()));
  add_ptr_scaled_int32(Rtemp, Rtemp, R0, 0);

  bind(set_mdp);
  str(Rtemp, Address(FP, frame::interpreter_frame_mdp_offset * wordSize));
}


void InterpreterMacroAssembler::verify_method_data_pointer() {
  assert(ProfileInterpreter, "must be profiling interpreter");
#ifdef ASSERT
  Label verify_continue;
  save_caller_save_registers();

  const Register Rmdp = R2;
  test_method_data_pointer(Rmdp, verify_continue); // If mdp is zero, continue

  // If the mdp is valid, it will point to a DataLayout header which is
  // consistent with the bcp.  The converse is highly probable also.

  ldrh(R3, Address(Rmdp, DataLayout::bci_offset()));
  ldr(Rtemp, Address(Rmethod, Method::const_offset()));
  add(R3, R3, Rtemp);
  add(R3, R3, in_bytes(ConstMethod::codes_offset()));
  cmp(R3, Rbcp);
  b(verify_continue, eq);

  mov(R0, Rmethod);
  mov(R1, Rbcp);
  call_VM_leaf(CAST_FROM_FN_PTR(address, InterpreterRuntime::verify_mdp), R0, R1, Rmdp);

  bind(verify_continue);
  restore_caller_save_registers();
#endif // ASSERT
}


void InterpreterMacroAssembler::set_mdp_data_at(Register mdp_in, int offset, Register value) {
  assert(ProfileInterpreter, "must be profiling interpreter");
  assert_different_registers(mdp_in, value);
  str(value, Address(mdp_in, offset));
}


// Increments mdp data. Sets bumped_count register to adjusted counter.
void InterpreterMacroAssembler::increment_mdp_data_at(Register mdp_in,
                                                      int offset,
                                                      Register bumped_count,
                                                      bool decrement) {
  assert(ProfileInterpreter, "must be profiling interpreter");

  // Counter address
  Address data(mdp_in, offset);
  assert_different_registers(mdp_in, bumped_count);

  increment_mdp_data_at(data, bumped_count, decrement);
}

void InterpreterMacroAssembler::set_mdp_flag_at(Register mdp_in, int flag_byte_constant) {
  assert_different_registers(mdp_in, Rtemp);
  assert(ProfileInterpreter, "must be profiling interpreter");
  assert((0 < flag_byte_constant) && (flag_byte_constant < (1 << BitsPerByte)), "flag mask is out of range");

  // Set the flag
  ldrb(Rtemp, Address(mdp_in, in_bytes(DataLayout::flags_offset())));
  orr(Rtemp, Rtemp, (unsigned)flag_byte_constant);
  strb(Rtemp, Address(mdp_in, in_bytes(DataLayout::flags_offset())));
}


// Increments mdp data. Sets bumped_count register to adjusted counter.
void InterpreterMacroAssembler::increment_mdp_data_at(Address data,
                                                      Register bumped_count,
                                                      bool decrement) {
  assert(ProfileInterpreter, "must be profiling interpreter");

  ldr(bumped_count, data);
  if (decrement) {
    // Decrement the register. Set condition codes.
    subs(bumped_count, bumped_count, DataLayout::counter_increment);
    // Avoid overflow.
#ifdef AARCH64
    assert(DataLayout::counter_increment == 1, "required for cinc");
    cinc(bumped_count, bumped_count, pl);
#else
    add(bumped_count, bumped_count, DataLayout::counter_increment, pl);
#endif // AARCH64
  } else {
    // Increment the register. Set condition codes.
    adds(bumped_count, bumped_count, DataLayout::counter_increment);
    // Avoid overflow.
#ifdef AARCH64
    assert(DataLayout::counter_increment == 1, "required for cinv");
    cinv(bumped_count, bumped_count, mi); // inverts 0x80..00 back to 0x7f..ff
#else
    sub(bumped_count, bumped_count, DataLayout::counter_increment, mi);
#endif // AARCH64
  }
  str(bumped_count, data);
}


void InterpreterMacroAssembler::test_mdp_data_at(Register mdp_in,
                                                 int offset,
                                                 Register value,
                                                 Register test_value_out,
                                                 Label& not_equal_continue) {
  assert(ProfileInterpreter, "must be profiling interpreter");
  assert_different_registers(mdp_in, test_value_out, value);

  ldr(test_value_out, Address(mdp_in, offset));
  cmp(test_value_out, value);

  b(not_equal_continue, ne);
}


void InterpreterMacroAssembler::update_mdp_by_offset(Register mdp_in, int offset_of_disp, Register reg_temp) {
  assert(ProfileInterpreter, "must be profiling interpreter");
  assert_different_registers(mdp_in, reg_temp);

  ldr(reg_temp, Address(mdp_in, offset_of_disp));
  add(mdp_in, mdp_in, reg_temp);
  str(mdp_in, Address(FP, frame::interpreter_frame_mdp_offset * wordSize));
}


void InterpreterMacroAssembler::update_mdp_by_offset(Register mdp_in, Register reg_offset, Register reg_tmp) {
  assert(ProfileInterpreter, "must be profiling interpreter");
  assert_different_registers(mdp_in, reg_offset, reg_tmp);

  ldr(reg_tmp, Address(mdp_in, reg_offset));
  add(mdp_in, mdp_in, reg_tmp);
  str(mdp_in, Address(FP, frame::interpreter_frame_mdp_offset * wordSize));
}


void InterpreterMacroAssembler::update_mdp_by_constant(Register mdp_in, int constant) {
  assert(ProfileInterpreter, "must be profiling interpreter");
  add(mdp_in, mdp_in, constant);
  str(mdp_in, Address(FP, frame::interpreter_frame_mdp_offset * wordSize));
}


// Blows volatile registers (R0-R3 on 32-bit ARM, R0-R18 on AArch64, Rtemp, LR).
void InterpreterMacroAssembler::update_mdp_for_ret(Register return_bci) {
  assert(ProfileInterpreter, "must be profiling interpreter");
  assert_different_registers(return_bci, R0, R1, R2, R3, Rtemp);

  mov(R1, return_bci);
  call_VM(noreg, CAST_FROM_FN_PTR(address, InterpreterRuntime::update_mdp_for_ret), R1);
}


// Sets mdp, bumped_count registers, blows Rtemp.
void InterpreterMacroAssembler::profile_taken_branch(Register mdp, Register bumped_count) {
  assert_different_registers(mdp, bumped_count);

  if (ProfileInterpreter) {
    Label profile_continue;

    // If no method data exists, go to profile_continue.
    // Otherwise, assign to mdp
    test_method_data_pointer(mdp, profile_continue);

    // We are taking a branch. Increment the taken count.
    increment_mdp_data_at(mdp, in_bytes(JumpData::taken_offset()), bumped_count);

    // The method data pointer needs to be updated to reflect the new target.
    update_mdp_by_offset(mdp, in_bytes(JumpData::displacement_offset()), Rtemp);

    bind (profile_continue);
  }
}


// Sets mdp, blows Rtemp.
void InterpreterMacroAssembler::profile_not_taken_branch(Register mdp) {
  assert_different_registers(mdp, Rtemp);

  if (ProfileInterpreter) {
    Label profile_continue;

    // If no method data exists, go to profile_continue.
    test_method_data_pointer(mdp, profile_continue);

    // We are taking a branch.  Increment the not taken count.
    increment_mdp_data_at(mdp, in_bytes(BranchData::not_taken_offset()), Rtemp);

    // The method data pointer needs to be updated to correspond to the next bytecode
    update_mdp_by_constant(mdp, in_bytes(BranchData::branch_data_size()));

    bind (profile_continue);
  }
}


// Sets mdp, blows Rtemp.
void InterpreterMacroAssembler::profile_call(Register mdp) {
  assert_different_registers(mdp, Rtemp);

  if (ProfileInterpreter) {
    Label profile_continue;

    // If no method data exists, go to profile_continue.
    test_method_data_pointer(mdp, profile_continue);

    // We are making a call.  Increment the count.
    increment_mdp_data_at(mdp, in_bytes(CounterData::count_offset()), Rtemp);

    // The method data pointer needs to be updated to reflect the new target.
    update_mdp_by_constant(mdp, in_bytes(CounterData::counter_data_size()));

    bind (profile_continue);
  }
}


// Sets mdp, blows Rtemp.
void InterpreterMacroAssembler::profile_final_call(Register mdp) {
  if (ProfileInterpreter) {
    Label profile_continue;

    // If no method data exists, go to profile_continue.
    test_method_data_pointer(mdp, profile_continue);

    // We are making a call.  Increment the count.
    increment_mdp_data_at(mdp, in_bytes(CounterData::count_offset()), Rtemp);

    // The method data pointer needs to be updated to reflect the new target.
    update_mdp_by_constant(mdp, in_bytes(VirtualCallData::virtual_call_data_size()));

    bind (profile_continue);
  }
}


// Sets mdp, blows Rtemp.
void InterpreterMacroAssembler::profile_virtual_call(Register mdp, Register receiver, bool receiver_can_be_null) {
  assert_different_registers(mdp, receiver, Rtemp);

  if (ProfileInterpreter) {
    Label profile_continue;

    // If no method data exists, go to profile_continue.
    test_method_data_pointer(mdp, profile_continue);

    Label skip_receiver_profile;
    if (receiver_can_be_null) {
      Label not_null;
      cbnz(receiver, not_null);
      // We are making a call.  Increment the count for null receiver.
      increment_mdp_data_at(mdp, in_bytes(CounterData::count_offset()), Rtemp);
      b(skip_receiver_profile);
      bind(not_null);
    }

    // Record the receiver type.
    record_klass_in_profile(receiver, mdp, Rtemp, true);
    bind(skip_receiver_profile);

    // The method data pointer needs to be updated to reflect the new target.
    update_mdp_by_constant(mdp, in_bytes(VirtualCallData::virtual_call_data_size()));
    bind(profile_continue);
  }
}


void InterpreterMacroAssembler::record_klass_in_profile_helper(
                                        Register receiver, Register mdp,
                                        Register reg_tmp,
                                        int start_row, Label& done, bool is_virtual_call) {
  if (TypeProfileWidth == 0)
    return;

  assert_different_registers(receiver, mdp, reg_tmp);

  int last_row = VirtualCallData::row_limit() - 1;
  assert(start_row <= last_row, "must be work left to do");
  // Test this row for both the receiver and for null.
  // Take any of three different outcomes:
  //   1. found receiver => increment count and goto done
  //   2. found null => keep looking for case 1, maybe allocate this cell
  //   3. found something else => keep looking for cases 1 and 2
  // Case 3 is handled by a recursive call.
  for (int row = start_row; row <= last_row; row++) {
    Label next_test;

    // See if the receiver is receiver[n].
    int recvr_offset = in_bytes(VirtualCallData::receiver_offset(row));

    test_mdp_data_at(mdp, recvr_offset, receiver, reg_tmp, next_test);

    // The receiver is receiver[n].  Increment count[n].
    int count_offset = in_bytes(VirtualCallData::receiver_count_offset(row));
    increment_mdp_data_at(mdp, count_offset, reg_tmp);
    b(done);

    bind(next_test);
    // reg_tmp now contains the receiver from the CallData.

    if (row == start_row) {
      Label found_null;
      // Failed the equality check on receiver[n]...  Test for null.
      if (start_row == last_row) {
        // The only thing left to do is handle the null case.
        if (is_virtual_call) {
          cbz(reg_tmp, found_null);
          // Receiver did not match any saved receiver and there is no empty row for it.
          // Increment total counter to indicate polymorphic case.
          increment_mdp_data_at(mdp, in_bytes(CounterData::count_offset()), reg_tmp);
          b(done);
          bind(found_null);
        } else {
          cbnz(reg_tmp, done);
        }
        break;
      }
      // Since null is rare, make it be the branch-taken case.
      cbz(reg_tmp, found_null);

      // Put all the "Case 3" tests here.
      record_klass_in_profile_helper(receiver, mdp, reg_tmp, start_row + 1, done, is_virtual_call);

      // Found a null.  Keep searching for a matching receiver,
      // but remember that this is an empty (unused) slot.
      bind(found_null);
    }
  }

  // In the fall-through case, we found no matching receiver, but we
  // observed the receiver[start_row] is NULL.

  // Fill in the receiver field and increment the count.
  int recvr_offset = in_bytes(VirtualCallData::receiver_offset(start_row));
  set_mdp_data_at(mdp, recvr_offset, receiver);
  int count_offset = in_bytes(VirtualCallData::receiver_count_offset(start_row));
  mov(reg_tmp, DataLayout::counter_increment);
  set_mdp_data_at(mdp, count_offset, reg_tmp);
  if (start_row > 0) {
    b(done);
  }
}

void InterpreterMacroAssembler::record_klass_in_profile(Register receiver,
                                                        Register mdp,
                                                        Register reg_tmp,
                                                        bool is_virtual_call) {
  assert(ProfileInterpreter, "must be profiling");
  assert_different_registers(receiver, mdp, reg_tmp);

  Label done;

  record_klass_in_profile_helper(receiver, mdp, reg_tmp, 0, done, is_virtual_call);

  bind (done);
}

// Sets mdp, blows volatile registers (R0-R3 on 32-bit ARM, R0-R18 on AArch64, Rtemp, LR).
void InterpreterMacroAssembler::profile_ret(Register mdp, Register return_bci) {
  assert_different_registers(mdp, return_bci, Rtemp, R0, R1, R2, R3);

  if (ProfileInterpreter) {
    Label profile_continue;
    uint row;

    // If no method data exists, go to profile_continue.
    test_method_data_pointer(mdp, profile_continue);

    // Update the total ret count.
    increment_mdp_data_at(mdp, in_bytes(CounterData::count_offset()), Rtemp);

    for (row = 0; row < RetData::row_limit(); row++) {
      Label next_test;

      // See if return_bci is equal to bci[n]:
      test_mdp_data_at(mdp, in_bytes(RetData::bci_offset(row)), return_bci,
                       Rtemp, next_test);

      // return_bci is equal to bci[n].  Increment the count.
      increment_mdp_data_at(mdp, in_bytes(RetData::bci_count_offset(row)), Rtemp);

      // The method data pointer needs to be updated to reflect the new target.
      update_mdp_by_offset(mdp, in_bytes(RetData::bci_displacement_offset(row)), Rtemp);
      b(profile_continue);
      bind(next_test);
    }

    update_mdp_for_ret(return_bci);

    bind(profile_continue);
  }
}


// Sets mdp.
void InterpreterMacroAssembler::profile_null_seen(Register mdp) {
  if (ProfileInterpreter) {
    Label profile_continue;

    // If no method data exists, go to profile_continue.
    test_method_data_pointer(mdp, profile_continue);

    set_mdp_flag_at(mdp, BitData::null_seen_byte_constant());

    // The method data pointer needs to be updated.
    int mdp_delta = in_bytes(BitData::bit_data_size());
    if (TypeProfileCasts) {
      mdp_delta = in_bytes(VirtualCallData::virtual_call_data_size());
    }
    update_mdp_by_constant(mdp, mdp_delta);

    bind (profile_continue);
  }
}


// Sets mdp, blows Rtemp.
void InterpreterMacroAssembler::profile_typecheck_failed(Register mdp) {
  assert_different_registers(mdp, Rtemp);

  if (ProfileInterpreter && TypeProfileCasts) {
    Label profile_continue;

    // If no method data exists, go to profile_continue.
    test_method_data_pointer(mdp, profile_continue);

    int count_offset = in_bytes(CounterData::count_offset());
    // Back up the address, since we have already bumped the mdp.
    count_offset -= in_bytes(VirtualCallData::virtual_call_data_size());

    // *Decrement* the counter.  We expect to see zero or small negatives.
    increment_mdp_data_at(mdp, count_offset, Rtemp, true);

    bind (profile_continue);
  }
}


// Sets mdp, blows Rtemp.
void InterpreterMacroAssembler::profile_typecheck(Register mdp, Register klass)
{
  assert_different_registers(mdp, klass, Rtemp);

  if (ProfileInterpreter) {
    Label profile_continue;

    // If no method data exists, go to profile_continue.
    test_method_data_pointer(mdp, profile_continue);

    // The method data pointer needs to be updated.
    int mdp_delta = in_bytes(BitData::bit_data_size());
    if (TypeProfileCasts) {
      mdp_delta = in_bytes(VirtualCallData::virtual_call_data_size());

      // Record the object type.
      record_klass_in_profile(klass, mdp, Rtemp, false);
    }
    update_mdp_by_constant(mdp, mdp_delta);

    bind(profile_continue);
  }
}


// Sets mdp, blows Rtemp.
void InterpreterMacroAssembler::profile_switch_default(Register mdp) {
  assert_different_registers(mdp, Rtemp);

  if (ProfileInterpreter) {
    Label profile_continue;

    // If no method data exists, go to profile_continue.
    test_method_data_pointer(mdp, profile_continue);

    // Update the default case count
    increment_mdp_data_at(mdp, in_bytes(MultiBranchData::default_count_offset()), Rtemp);

    // The method data pointer needs to be updated.
    update_mdp_by_offset(mdp, in_bytes(MultiBranchData::default_displacement_offset()), Rtemp);

    bind(profile_continue);
  }
}


// Sets mdp. Blows reg_tmp1, reg_tmp2. Index could be the same as reg_tmp2.
void InterpreterMacroAssembler::profile_switch_case(Register mdp, Register index, Register reg_tmp1, Register reg_tmp2) {
  assert_different_registers(mdp, reg_tmp1, reg_tmp2);
  assert_different_registers(mdp, reg_tmp1, index);

  if (ProfileInterpreter) {
    Label profile_continue;

    const int count_offset = in_bytes(MultiBranchData::case_array_offset()) +
                              in_bytes(MultiBranchData::relative_count_offset());

    const int displacement_offset = in_bytes(MultiBranchData::case_array_offset()) +
                              in_bytes(MultiBranchData::relative_displacement_offset());

    // If no method data exists, go to profile_continue.
    test_method_data_pointer(mdp, profile_continue);

    // Build the base (index * per_case_size_in_bytes())
    logical_shift_left(reg_tmp1, index, exact_log2(in_bytes(MultiBranchData::per_case_size())));

    // Update the case count
    add(reg_tmp1, reg_tmp1, count_offset);
    increment_mdp_data_at(Address(mdp, reg_tmp1), reg_tmp2);

    // The method data pointer needs to be updated.
    add(reg_tmp1, reg_tmp1, displacement_offset - count_offset);
    update_mdp_by_offset(mdp, reg_tmp1, reg_tmp2);

    bind (profile_continue);
  }
}


void InterpreterMacroAssembler::byteswap_u32(Register r, Register rtmp1, Register rtmp2) {
#ifdef AARCH64
  rev_w(r, r);
#else
  if (VM_Version::supports_rev()) {
    rev(r, r);
  } else {
    eor(rtmp1, r, AsmOperand(r, ror, 16));
    mvn(rtmp2, 0x0000ff00);
    andr(rtmp1, rtmp2, AsmOperand(rtmp1, lsr, 8));
    eor(r, rtmp1, AsmOperand(r, ror, 8));
  }
#endif // AARCH64
}


void InterpreterMacroAssembler::inc_global_counter(address address_of_counter, int offset, Register tmp1, Register tmp2, bool avoid_overflow) {
  const intx addr = (intx) (address_of_counter + offset);

  assert ((addr & 0x3) == 0, "address of counter should be aligned");
  const intx offset_mask = right_n_bits(AARCH64_ONLY(12 + 2) NOT_AARCH64(12));

  const address base = (address) (addr & ~offset_mask);
  const int offs = (int) (addr & offset_mask);

  const Register addr_base = tmp1;
  const Register val = tmp2;

  mov_slow(addr_base, base);
  ldr_s32(val, Address(addr_base, offs));

  if (avoid_overflow) {
    adds_32(val, val, 1);
#ifdef AARCH64
    Label L;
    b(L, mi);
    str_32(val, Address(addr_base, offs));
    bind(L);
#else
    str(val, Address(addr_base, offs), pl);
#endif // AARCH64
  } else {
    add_32(val, val, 1);
    str_32(val, Address(addr_base, offs));
  }
}

void InterpreterMacroAssembler::interp_verify_oop(Register reg, TosState state, const char *file, int line) {
  if (state == atos) { MacroAssembler::_verify_oop(reg, "broken oop", file, line); }
}

// Inline assembly for:
//
// if (thread is in interp_only_mode) {
//   InterpreterRuntime::post_method_entry();
// }
// if (DTraceMethodProbes) {
//   SharedRuntime::dtrace_method_entry(method, receiver);
// }
// if (RC_TRACE_IN_RANGE(0x00001000, 0x00002000)) {
//   SharedRuntime::rc_trace_method_entry(method, receiver);
// }

void InterpreterMacroAssembler::notify_method_entry() {
  // Whenever JVMTI is interp_only_mode, method entry/exit events are sent to
  // track stack depth.  If it is possible to enter interp_only_mode we add
  // the code to check if the event should be sent.
  if (can_post_interpreter_events()) {
    Label L;

    ldr_s32(Rtemp, Address(Rthread, JavaThread::interp_only_mode_offset()));
    cbz(Rtemp, L);

    call_VM(noreg, CAST_FROM_FN_PTR(address, InterpreterRuntime::post_method_entry));

    bind(L);
  }

  // Note: Disable DTrace runtime check for now to eliminate overhead on each method entry
  if (DTraceMethodProbes) {
    Label Lcontinue;

    ldrb_global(Rtemp, (address)&DTraceMethodProbes);
    cbz(Rtemp, Lcontinue);

    mov(R0, Rthread);
    mov(R1, Rmethod);
    call_VM_leaf(CAST_FROM_FN_PTR(address, SharedRuntime::dtrace_method_entry), R0, R1);

    bind(Lcontinue);
  }
  // RedefineClasses() tracing support for obsolete method entry
  if (log_is_enabled(Trace, redefine, class, obsolete)) {
    mov(R0, Rthread);
    mov(R1, Rmethod);
    call_VM_leaf(CAST_FROM_FN_PTR(address, SharedRuntime::rc_trace_method_entry),
                 R0, R1);
  }
}


void InterpreterMacroAssembler::notify_method_exit(
                 TosState state, NotifyMethodExitMode mode,
                 bool native, Register result_lo, Register result_hi, FloatRegister result_fp) {
  // Whenever JVMTI is interp_only_mode, method entry/exit events are sent to
  // track stack depth.  If it is possible to enter interp_only_mode we add
  // the code to check if the event should be sent.
  if (mode == NotifyJVMTI && can_post_interpreter_events()) {
    Label L;
    // Note: frame::interpreter_frame_result has a dependency on how the
    // method result is saved across the call to post_method_exit. If this
    // is changed then the interpreter_frame_result implementation will
    // need to be updated too.

    ldr_s32(Rtemp, Address(Rthread, JavaThread::interp_only_mode_offset()));
    cbz(Rtemp, L);

    if (native) {
      // For c++ and template interpreter push both result registers on the
      // stack in native, we don't know the state.
      // On AArch64 result registers are stored into the frame at known locations.
      // See frame::interpreter_frame_result for code that gets the result values from here.
      assert(result_lo != noreg, "result registers should be defined");

#ifdef AARCH64
      assert(result_hi == noreg, "result_hi is not used on AArch64");
      assert(result_fp != fnoreg, "FP result register must be defined");

      str_d(result_fp, Address(FP, frame::interpreter_frame_fp_saved_result_offset * wordSize));
      str(result_lo, Address(FP, frame::interpreter_frame_gp_saved_result_offset * wordSize));
#else
      assert(result_hi != noreg, "result registers should be defined");

#ifdef __ABI_HARD__
      assert(result_fp != fnoreg, "FP result register must be defined");
      sub(SP, SP, 2 * wordSize);
      fstd(result_fp, Address(SP));
#endif // __ABI_HARD__

      push(RegisterSet(result_lo) | RegisterSet(result_hi));
#endif // AARCH64

      call_VM(noreg, CAST_FROM_FN_PTR(address, InterpreterRuntime::post_method_exit));

#ifdef AARCH64
      ldr_d(result_fp, Address(FP, frame::interpreter_frame_fp_saved_result_offset * wordSize));
      ldr(result_lo, Address(FP, frame::interpreter_frame_gp_saved_result_offset * wordSize));
#else
      pop(RegisterSet(result_lo) | RegisterSet(result_hi));
#ifdef __ABI_HARD__
      fldd(result_fp, Address(SP));
      add(SP, SP, 2 * wordSize);
#endif // __ABI_HARD__
#endif // AARCH64

    } else {
      // For the template interpreter, the value on tos is the size of the
      // state. (c++ interpreter calls jvmti somewhere else).
      push(state);
      call_VM(noreg, CAST_FROM_FN_PTR(address, InterpreterRuntime::post_method_exit));
      pop(state);
    }

    bind(L);
  }

  // Note: Disable DTrace runtime check for now to eliminate overhead on each method exit
  if (DTraceMethodProbes) {
    Label Lcontinue;

    ldrb_global(Rtemp, (address)&DTraceMethodProbes);
    cbz(Rtemp, Lcontinue);

    push(state);

    mov(R0, Rthread);
    mov(R1, Rmethod);

    call_VM_leaf(CAST_FROM_FN_PTR(address, SharedRuntime::dtrace_method_exit), R0, R1);

    pop(state);

    bind(Lcontinue);
  }
}


#ifndef PRODUCT

void InterpreterMacroAssembler::trace_state(const char* msg) {
  int push_size = save_caller_save_registers();

  Label Lcontinue;
  InlinedString Lmsg0("%s: FP=" INTPTR_FORMAT ", SP=" INTPTR_FORMAT "\n");
  InlinedString Lmsg(msg);
  InlinedAddress Lprintf((address)printf);

  ldr_literal(R0, Lmsg0);
  ldr_literal(R1, Lmsg);
  mov(R2, FP);
  add(R3, SP, push_size);  // original SP (without saved registers)
  ldr_literal(Rtemp, Lprintf);
  call(Rtemp);

  b(Lcontinue);

  bind_literal(Lmsg0);
  bind_literal(Lmsg);
  bind_literal(Lprintf);


  bind(Lcontinue);

  restore_caller_save_registers();
}

#endif

// Jump if ((*counter_addr += increment) & mask) satisfies the condition.
void InterpreterMacroAssembler::increment_mask_and_jump(Address counter_addr,
                                                        int increment, Address mask_addr,
                                                        Register scratch, Register scratch2,
                                                        AsmCondition cond, Label* where) {
  // caution: scratch2 and base address of counter_addr can be the same
  assert_different_registers(scratch, scratch2);
  ldr_u32(scratch, counter_addr);
  add(scratch, scratch, increment);
  str_32(scratch, counter_addr);

#ifdef AARCH64
  ldr_u32(scratch2, mask_addr);
  ands_w(ZR, scratch, scratch2);
#else
  ldr(scratch2, mask_addr);
  andrs(scratch, scratch, scratch2);
#endif // AARCH64
  b(*where, cond);
}

void InterpreterMacroAssembler::get_method_counters(Register method,
                                                    Register Rcounters,
                                                    Label& skip,
                                                    bool saveRegs,
                                                    Register reg1,
                                                    Register reg2,
                                                    Register reg3) {
  const Address method_counters(method, Method::method_counters_offset());
  Label has_counters;

  ldr(Rcounters, method_counters);
  cbnz(Rcounters, has_counters);

  if (saveRegs) {
    // Save and restore in use caller-saved registers since they will be trashed by call_VM
    assert(reg1 != noreg, "must specify reg1");
    assert(reg2 != noreg, "must specify reg2");
#ifdef AARCH64
    assert(reg3 != noreg, "must specify reg3");
    stp(reg1, reg2, Address(Rstack_top, -2*wordSize, pre_indexed));
    stp(reg3, ZR, Address(Rstack_top, -2*wordSize, pre_indexed));
#else
    assert(reg3 == noreg, "must not specify reg3");
    push(RegisterSet(reg1) | RegisterSet(reg2));
#endif
  }

  mov(R1, method);
  call_VM(noreg, CAST_FROM_FN_PTR(address, InterpreterRuntime::build_method_counters), R1);

  if (saveRegs) {
#ifdef AARCH64
    ldp(reg3, ZR, Address(Rstack_top, 2*wordSize, post_indexed));
    ldp(reg1, reg2, Address(Rstack_top, 2*wordSize, post_indexed));
#else
    pop(RegisterSet(reg1) | RegisterSet(reg2));
#endif
  }

  ldr(Rcounters, method_counters);
  cbz(Rcounters, skip); // No MethodCounters created, OutOfMemory

  bind(has_counters);
}