compiler/dex/quick/arm64/int_arm64.cc

/*
 * Copyright (C) 2011 The Android Open Source Project
 *
 * Licensed under the Apache License, Version 2.0 (the "License");
 * you may not use this file except in compliance with the License.
 * You may obtain a copy of the License at
 *
 *      http://www.apache.org/licenses/LICENSE-2.0
 *
 * Unless required by applicable law or agreed to in writing, software
 * distributed under the License is distributed on an "AS IS" BASIS,
 * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
 * See the License for the specific language governing permissions and
 * limitations under the License.
 */

/* This file contains codegen for the Thumb2 ISA. */

#include "arm64_lir.h"
#include "codegen_arm64.h"
#include "dex/quick/mir_to_lir-inl.h"
#include "entrypoints/quick/quick_entrypoints.h"
#include "mirror/array.h"

namespace art {

LIR* Arm64Mir2Lir::OpCmpBranch(ConditionCode cond, RegStorage src1, RegStorage src2, LIR* target) {
  OpRegReg(kOpCmp, src1, src2);
  return OpCondBranch(cond, target);
}

LIR* Arm64Mir2Lir::OpIT(ConditionCode ccode, const char* guide) {
  LOG(FATAL) << "Unexpected use of OpIT for Arm64";
  return NULL;
}

void Arm64Mir2Lir::OpEndIT(LIR* it) {
  LOG(FATAL) << "Unexpected use of OpEndIT for Arm64";
}

/*
 * 64-bit 3way compare function.
 *     cmp   xA, xB
 *     csinc wC, wzr, wzr, eq  // wC = (xA == xB) ? 0 : 1
 *     csneg wC, wC, wC, ge    // wC = (xA >= xB) ? wC : -wC
 */
void Arm64Mir2Lir::GenCmpLong(RegLocation rl_dest, RegLocation rl_src1,
                              RegLocation rl_src2) {
  RegLocation rl_result;
  rl_src1 = LoadValueWide(rl_src1, kCoreReg);
  rl_src2 = LoadValueWide(rl_src2, kCoreReg);
  rl_result = EvalLoc(rl_dest, kCoreReg, true);

  OpRegReg(kOpCmp, rl_src1.reg, rl_src2.reg);
  NewLIR4(kA64Csinc4rrrc, rl_result.reg.GetReg(), rwzr, rwzr, kArmCondEq);
  NewLIR4(kA64Csneg4rrrc, rl_result.reg.GetReg(), rl_result.reg.GetReg(),
          rl_result.reg.GetReg(), kArmCondGe);
  StoreValue(rl_dest, rl_result);
}

void Arm64Mir2Lir::GenShiftOpLong(Instruction::Code opcode, RegLocation rl_dest,
                             RegLocation rl_src1, RegLocation rl_shift) {
  OpKind op = kOpBkpt;
  switch (opcode) {
  case Instruction::SHL_LONG:
  case Instruction::SHL_LONG_2ADDR:
    op = kOpLsl;
    break;
  case Instruction::SHR_LONG:
  case Instruction::SHR_LONG_2ADDR:
    op = kOpAsr;
    break;
  case Instruction::USHR_LONG:
  case Instruction::USHR_LONG_2ADDR:
    op = kOpLsr;
    break;
  default:
    LOG(FATAL) << "Unexpected case: " << opcode;
  }
  rl_shift = LoadValueWide(rl_shift, kCoreReg);
  rl_src1 = LoadValueWide(rl_src1, kCoreReg);
  RegLocation rl_result = EvalLocWide(rl_dest, kCoreReg, true);
  OpRegRegReg(op, rl_result.reg, rl_src1.reg, rl_shift.reg);
  StoreValueWide(rl_dest, rl_result);
}

void Arm64Mir2Lir::GenSelect(BasicBlock* bb, MIR* mir) {
  RegLocation rl_result;
  RegLocation rl_src = mir_graph_->GetSrc(mir, 0);
  RegLocation rl_dest = mir_graph_->GetDest(mir);
  RegisterClass src_reg_class = rl_src.ref ? kRefReg : kCoreReg;
  RegisterClass result_reg_class = rl_dest.ref ? kRefReg : kCoreReg;
  rl_src = LoadValue(rl_src, src_reg_class);
  ArmConditionCode code = ArmConditionEncoding(mir->meta.ccode);

  RegLocation rl_true = mir_graph_->reg_location_[mir->ssa_rep->uses[1]];
  RegLocation rl_false = mir_graph_->reg_location_[mir->ssa_rep->uses[2]];
  rl_true = LoadValue(rl_true, result_reg_class);
  rl_false = LoadValue(rl_false, result_reg_class);
  rl_result = EvalLoc(rl_dest, result_reg_class, true);
  OpRegImm(kOpCmp, rl_src.reg, 0);
  NewLIR4(kA64Csel4rrrc, rl_result.reg.GetReg(), rl_true.reg.GetReg(),
          rl_false.reg.GetReg(), code);
  StoreValue(rl_dest, rl_result);
}

void Arm64Mir2Lir::GenFusedLongCmpBranch(BasicBlock* bb, MIR* mir) {
  RegLocation rl_src1 = mir_graph_->GetSrcWide(mir, 0);
  RegLocation rl_src2 = mir_graph_->GetSrcWide(mir, 2);
  LIR* taken = &block_label_list_[bb->taken];
  LIR* not_taken = &block_label_list_[bb->fall_through];
  rl_src1 = LoadValueWide(rl_src1, kCoreReg);
  // Normalize such that if either operand is constant, src2 will be constant.
  ConditionCode ccode = mir->meta.ccode;
  if (rl_src1.is_const) {
    std::swap(rl_src1, rl_src2);
    ccode = FlipComparisonOrder(ccode);
  }

  if (rl_src2.is_const) {
    rl_src2 = UpdateLocWide(rl_src2);
    int64_t val = mir_graph_->ConstantValueWide(rl_src2);
    // Special handling using cbz & cbnz.
    if (val == 0 && (ccode == kCondEq || ccode == kCondNe)) {
      OpCmpImmBranch(ccode, rl_src1.reg, 0, taken);
      OpCmpImmBranch(NegateComparison(ccode), rl_src1.reg, 0, not_taken);
      return;
    // Only handle Imm if src2 is not already in a register.
    } else if (rl_src2.location != kLocPhysReg) {
      OpRegImm64(kOpCmp, rl_src1.reg, val);
      OpCondBranch(ccode, taken);
      OpCondBranch(NegateComparison(ccode), not_taken);
      return;
    }
  }

  rl_src2 = LoadValueWide(rl_src2, kCoreReg);
  OpRegReg(kOpCmp, rl_src1.reg, rl_src2.reg);
  OpCondBranch(ccode, taken);
  OpCondBranch(NegateComparison(ccode), not_taken);
}

/*
 * Generate a register comparison to an immediate and branch.  Caller
 * is responsible for setting branch target field.
 */
LIR* Arm64Mir2Lir::OpCmpImmBranch(ConditionCode cond, RegStorage reg, int check_value,
                                  LIR* target) {
  LIR* branch;
  ArmConditionCode arm_cond = ArmConditionEncoding(cond);
  if (check_value == 0 && (arm_cond == kArmCondEq || arm_cond == kArmCondNe)) {
    ArmOpcode opcode = (arm_cond == kArmCondEq) ? kA64Cbz2rt : kA64Cbnz2rt;
    ArmOpcode wide = reg.Is64Bit() ? WIDE(0) : UNWIDE(0);
    branch = NewLIR2(opcode | wide, reg.GetReg(), 0);
  } else {
    OpRegImm(kOpCmp, reg, check_value);
    branch = NewLIR2(kA64B2ct, arm_cond, 0);
  }
  branch->target = target;
  return branch;
}

LIR* Arm64Mir2Lir::OpRegCopyNoInsert(RegStorage r_dest, RegStorage r_src) {
  bool dest_is_fp = r_dest.IsFloat();
  bool src_is_fp = r_src.IsFloat();
  ArmOpcode opcode = kA64Brk1d;
  LIR* res;

  if (LIKELY(dest_is_fp == src_is_fp)) {
    if (LIKELY(!dest_is_fp)) {
      // Core/core copy.
      // Copies involving the sp register require a different instruction.
      opcode = UNLIKELY(A64_REG_IS_SP(r_dest.GetReg())) ? kA64Add4RRdT : kA64Mov2rr;

      // TODO(Arm64): kA64Add4RRdT formally has 4 args, but is used as a 2 args instruction.
      //   This currently works because the other arguments are set to 0 by default. We should
      //   rather introduce an alias kA64Mov2RR.

      // core/core copy. Do a x/x copy only if both registers are x.
      if (r_dest.Is64Bit() && r_src.Is64Bit()) {
        opcode = WIDE(opcode);
      }
    } else {
      // Float/float copy.
      bool dest_is_double = r_dest.IsDouble();
      bool src_is_double = r_src.IsDouble();

      // We do not do float/double or double/float casts here.
      DCHECK_EQ(dest_is_double, src_is_double);

      // Homogeneous float/float copy.
      opcode = (dest_is_double) ? FWIDE(kA64Fmov2ff) : kA64Fmov2ff;
    }
  } else {
    // Inhomogeneous register copy.
    if (dest_is_fp) {
      if (r_dest.IsDouble()) {
        opcode = kA64Fmov2Sx;
      } else {
        DCHECK(r_src.IsSingle());
        opcode = kA64Fmov2sw;
      }
    } else {
      if (r_src.IsDouble()) {
        opcode = kA64Fmov2xS;
      } else {
        DCHECK(r_dest.Is32Bit());
        opcode = kA64Fmov2ws;
      }
    }
  }

  res = RawLIR(current_dalvik_offset_, opcode, r_dest.GetReg(), r_src.GetReg());

  if (!(cu_->disable_opt & (1 << kSafeOptimizations)) && r_dest == r_src) {
    res->flags.is_nop = true;
  }

  return res;
}

void Arm64Mir2Lir::OpRegCopy(RegStorage r_dest, RegStorage r_src) {
  if (r_dest != r_src) {
    LIR* res = OpRegCopyNoInsert(r_dest, r_src);
    AppendLIR(res);
  }
}

void Arm64Mir2Lir::OpRegCopyWide(RegStorage r_dest, RegStorage r_src) {
  OpRegCopy(r_dest, r_src);
}

// Table of magic divisors
struct MagicTable {
  uint32_t magic;
  uint32_t shift;
  DividePattern pattern;
};

static const MagicTable magic_table[] = {
  {0, 0, DivideNone},        // 0
  {0, 0, DivideNone},        // 1
  {0, 0, DivideNone},        // 2
  {0x55555556, 0, Divide3},  // 3
  {0, 0, DivideNone},        // 4
  {0x66666667, 1, Divide5},  // 5
  {0x2AAAAAAB, 0, Divide3},  // 6
  {0x92492493, 2, Divide7},  // 7
  {0, 0, DivideNone},        // 8
  {0x38E38E39, 1, Divide5},  // 9
  {0x66666667, 2, Divide5},  // 10
  {0x2E8BA2E9, 1, Divide5},  // 11
  {0x2AAAAAAB, 1, Divide5},  // 12
  {0x4EC4EC4F, 2, Divide5},  // 13
  {0x92492493, 3, Divide7},  // 14
  {0x88888889, 3, Divide7},  // 15
};

// Integer division by constant via reciprocal multiply (Hacker's Delight, 10-4)
bool Arm64Mir2Lir::SmallLiteralDivRem(Instruction::Code dalvik_opcode, bool is_div,
                                    RegLocation rl_src, RegLocation rl_dest, int lit) {
  // TODO(Arm64): fix this for Arm64. Note: may be worth revisiting the magic table.
  //   It should be possible subtracting one from all its entries, and using smaddl
  //   to counteract this. The advantage is that integers should then be easier to
  //   encode as logical immediates (0x55555555 rather than 0x55555556).
  UNIMPLEMENTED(FATAL);

  if ((lit < 0) || (lit >= static_cast<int>(sizeof(magic_table)/sizeof(magic_table[0])))) {
    return false;
  }
  DividePattern pattern = magic_table[lit].pattern;
  if (pattern == DivideNone) {
    return false;
  }
  // Tuning: add rem patterns
  if (!is_div) {
    return false;
  }

  RegStorage r_magic = AllocTemp();
  LoadConstant(r_magic, magic_table[lit].magic);
  rl_src = LoadValue(rl_src, kCoreReg);
  RegLocation rl_result = EvalLoc(rl_dest, kCoreReg, true);
  RegStorage r_hi = AllocTemp();
  RegStorage r_lo = AllocTemp();
  NewLIR4(kA64Smaddl4xwwx, r_lo.GetReg(), r_magic.GetReg(), rl_src.reg.GetReg(), rxzr);
  switch (pattern) {
    case Divide3:
      OpRegRegRegShift(kOpSub, rl_result.reg, r_hi, rl_src.reg, EncodeShift(kA64Asr, 31));
      break;
    case Divide5:
      OpRegRegImm(kOpAsr, r_lo, rl_src.reg, 31);
      OpRegRegRegShift(kOpRsub, rl_result.reg, r_lo, r_hi, EncodeShift(kA64Asr, magic_table[lit].shift));
      break;
    case Divide7:
      OpRegReg(kOpAdd, r_hi, rl_src.reg);
      OpRegRegImm(kOpAsr, r_lo, rl_src.reg, 31);
      OpRegRegRegShift(kOpRsub, rl_result.reg, r_lo, r_hi, EncodeShift(kA64Asr, magic_table[lit].shift));
      break;
    default:
      LOG(FATAL) << "Unexpected pattern: " << pattern;
  }
  StoreValue(rl_dest, rl_result);
  return true;
}

bool Arm64Mir2Lir::EasyMultiply(RegLocation rl_src, RegLocation rl_dest, int lit) {
  LOG(FATAL) << "Unexpected use of EasyMultiply for Arm64";
  return false;
}

RegLocation Arm64Mir2Lir::GenDivRem(RegLocation rl_dest, RegLocation rl_src1,
                      RegLocation rl_src2, bool is_div, bool check_zero) {
  LOG(FATAL) << "Unexpected use of GenDivRem for Arm64";
  return rl_dest;
}

RegLocation Arm64Mir2Lir::GenDivRemLit(RegLocation rl_dest, RegLocation rl_src1, int lit, bool is_div) {
  LOG(FATAL) << "Unexpected use of GenDivRemLit for Arm64";
  return rl_dest;
}

RegLocation Arm64Mir2Lir::GenDivRemLit(RegLocation rl_dest, RegStorage reg1, int lit, bool is_div) {
  RegLocation rl_result = EvalLoc(rl_dest, kCoreReg, true);

  // Put the literal in a temp.
  RegStorage lit_temp = AllocTemp();
  LoadConstant(lit_temp, lit);
  // Use the generic case for div/rem with arg2 in a register.
  // TODO: The literal temp can be freed earlier during a modulus to reduce reg pressure.
  rl_result = GenDivRem(rl_result, reg1, lit_temp, is_div);
  FreeTemp(lit_temp);

  return rl_result;
}

RegLocation Arm64Mir2Lir::GenDivRem(RegLocation rl_dest, RegStorage r_src1, RegStorage r_src2,
                                  bool is_div) {
  CHECK_EQ(r_src1.Is64Bit(), r_src2.Is64Bit());

  RegLocation rl_result = EvalLoc(rl_dest, kCoreReg, true);
  if (is_div) {
    OpRegRegReg(kOpDiv, rl_result.reg, r_src1, r_src2);
  } else {
    // temp = r_src1 / r_src2
    // dest = r_src1 - temp * r_src2
    RegStorage temp;
    ArmOpcode wide;
    if (rl_result.reg.Is64Bit()) {
      temp = AllocTempWide();
      wide = WIDE(0);
    } else {
      temp = AllocTemp();
      wide = UNWIDE(0);
    }
    OpRegRegReg(kOpDiv, temp, r_src1, r_src2);
    NewLIR4(kA64Msub4rrrr | wide, rl_result.reg.GetReg(), temp.GetReg(),
            r_src1.GetReg(), r_src2.GetReg());
    FreeTemp(temp);
  }
  return rl_result;
}

bool Arm64Mir2Lir::GenInlinedMinMaxInt(CallInfo* info, bool is_min) {
  // TODO(Arm64): implement this.
  UNIMPLEMENTED(FATAL);

  DCHECK_EQ(cu_->instruction_set, kThumb2);
  RegLocation rl_src1 = info->args[0];
  RegLocation rl_src2 = info->args[1];
  rl_src1 = LoadValue(rl_src1, kCoreReg);
  rl_src2 = LoadValue(rl_src2, kCoreReg);
  RegLocation rl_dest = InlineTarget(info);
  RegLocation rl_result = EvalLoc(rl_dest, kCoreReg, true);
  OpRegReg(kOpCmp, rl_src1.reg, rl_src2.reg);
  // OpIT((is_min) ? kCondGt : kCondLt, "E");
  OpRegReg(kOpMov, rl_result.reg, rl_src2.reg);
  OpRegReg(kOpMov, rl_result.reg, rl_src1.reg);
  GenBarrier();
  StoreValue(rl_dest, rl_result);
  return true;
}

bool Arm64Mir2Lir::GenInlinedPeek(CallInfo* info, OpSize size) {
  // TODO(Arm64): implement this.
  UNIMPLEMENTED(WARNING);

  RegLocation rl_src_address = info->args[0];  // long address
  rl_src_address = NarrowRegLoc(rl_src_address);  // ignore high half in info->args[1]
  RegLocation rl_dest = InlineTarget(info);
  RegLocation rl_address = LoadValue(rl_src_address, kCoreReg);
  RegLocation rl_result = EvalLoc(rl_dest, kCoreReg, true);
  if (size == k64) {
    // Fake unaligned LDRD by two unaligned LDR instructions on ARMv7 with SCTLR.A set to 0.
    if (rl_address.reg.GetReg() != rl_result.reg.GetLowReg()) {
      LoadWordDisp(rl_address.reg, 0, rl_result.reg.GetLow());
      LoadWordDisp(rl_address.reg, 4, rl_result.reg.GetHigh());
    } else {
      LoadWordDisp(rl_address.reg, 4, rl_result.reg.GetHigh());
      LoadWordDisp(rl_address.reg, 0, rl_result.reg.GetLow());
    }
    StoreValueWide(rl_dest, rl_result);
  } else {
    DCHECK(size == kSignedByte || size == kSignedHalf || size == k32);
    // Unaligned load with LDR and LDRSH is allowed on ARMv7 with SCTLR.A set to 0.
    LoadBaseDisp(rl_address.reg, 0, rl_result.reg, size);
    StoreValue(rl_dest, rl_result);
  }
  return true;
}

bool Arm64Mir2Lir::GenInlinedPoke(CallInfo* info, OpSize size) {
  // TODO(Arm64): implement this.
  UNIMPLEMENTED(WARNING);

  RegLocation rl_src_address = info->args[0];  // long address
  rl_src_address = NarrowRegLoc(rl_src_address);  // ignore high half in info->args[1]
  RegLocation rl_src_value = info->args[2];  // [size] value
  RegLocation rl_address = LoadValue(rl_src_address, kCoreReg);
  if (size == k64) {
    // Fake unaligned STRD by two unaligned STR instructions on ARMv7 with SCTLR.A set to 0.
    RegLocation rl_value = LoadValueWide(rl_src_value, kCoreReg);
    StoreBaseDisp(rl_address.reg, 0, rl_value.reg.GetLow(), k32);
    StoreBaseDisp(rl_address.reg, 4, rl_value.reg.GetHigh(), k32);
  } else {
    DCHECK(size == kSignedByte || size == kSignedHalf || size == k32);
    // Unaligned store with STR and STRSH is allowed on ARMv7 with SCTLR.A set to 0.
    RegLocation rl_value = LoadValue(rl_src_value, kCoreReg);
    StoreBaseDisp(rl_address.reg, 0, rl_value.reg, size);
  }
  return true;
}

void Arm64Mir2Lir::OpLea(RegStorage r_base, RegStorage reg1, RegStorage reg2, int scale, int offset) {
  LOG(FATAL) << "Unexpected use of OpLea for Arm64";
}

void Arm64Mir2Lir::OpTlsCmp(ThreadOffset<4> offset, int val) {
  UNIMPLEMENTED(FATAL) << "Should not be used.";
}

void Arm64Mir2Lir::OpTlsCmp(ThreadOffset<8> offset, int val) {
  LOG(FATAL) << "Unexpected use of OpTlsCmp for Arm64";
}

bool Arm64Mir2Lir::GenInlinedCas(CallInfo* info, bool is_long, bool is_object) {
  // TODO(Arm64): implement this.
  UNIMPLEMENTED(WARNING);

  DCHECK_EQ(cu_->instruction_set, kThumb2);
  // Unused - RegLocation rl_src_unsafe = info->args[0];
  RegLocation rl_src_obj = info->args[1];  // Object - known non-null
  RegLocation rl_src_offset = info->args[2];  // long low
  rl_src_offset = NarrowRegLoc(rl_src_offset);  // ignore high half in info->args[3]
  RegLocation rl_src_expected = info->args[4];  // int, long or Object
  // If is_long, high half is in info->args[5]
  RegLocation rl_src_new_value = info->args[is_long ? 6 : 5];  // int, long or Object
  // If is_long, high half is in info->args[7]
  RegLocation rl_dest = InlineTarget(info);  // boolean place for result

  // We have only 5 temporary registers available and actually only 4 if the InlineTarget
  // above locked one of the temps. For a straightforward CAS64 we need 7 registers:
  // r_ptr (1), new_value (2), expected(2) and ldrexd result (2). If neither expected nor
  // new_value is in a non-temp core register we shall reload them in the ldrex/strex loop
  // into the same temps, reducing the number of required temps down to 5. We shall work
  // around the potentially locked temp by using LR for r_ptr, unconditionally.
  // TODO: Pass information about the need for more temps to the stack frame generation
  // code so that we can rely on being able to allocate enough temps.
  DCHECK(!GetRegInfo(rs_rA64_LR)->IsTemp());
  MarkTemp(rs_rA64_LR);
  FreeTemp(rs_rA64_LR);
  LockTemp(rs_rA64_LR);
  bool load_early = true;
  if (is_long) {
    RegStorage expected_reg = rl_src_expected.reg.IsPair() ? rl_src_expected.reg.GetLow() :
        rl_src_expected.reg;
    RegStorage new_val_reg = rl_src_new_value.reg.IsPair() ? rl_src_new_value.reg.GetLow() :
        rl_src_new_value.reg;
    bool expected_is_core_reg = rl_src_expected.location == kLocPhysReg && !expected_reg.IsFloat();
    bool new_value_is_core_reg = rl_src_new_value.location == kLocPhysReg && !new_val_reg.IsFloat();
    bool expected_is_good_reg = expected_is_core_reg && !IsTemp(expected_reg);
    bool new_value_is_good_reg = new_value_is_core_reg && !IsTemp(new_val_reg);

    if (!expected_is_good_reg && !new_value_is_good_reg) {
      // None of expected/new_value is non-temp reg, need to load both late
      load_early = false;
      // Make sure they are not in the temp regs and the load will not be skipped.
      if (expected_is_core_reg) {
        FlushRegWide(rl_src_expected.reg);
        ClobberSReg(rl_src_expected.s_reg_low);
        ClobberSReg(GetSRegHi(rl_src_expected.s_reg_low));
        rl_src_expected.location = kLocDalvikFrame;
      }
      if (new_value_is_core_reg) {
        FlushRegWide(rl_src_new_value.reg);
        ClobberSReg(rl_src_new_value.s_reg_low);
        ClobberSReg(GetSRegHi(rl_src_new_value.s_reg_low));
        rl_src_new_value.location = kLocDalvikFrame;
      }
    }
  }

  // Release store semantics, get the barrier out of the way.  TODO: revisit
  GenMemBarrier(kStoreLoad);

  RegLocation rl_object = LoadValue(rl_src_obj, kRefReg);
  RegLocation rl_new_value;
  if (!is_long) {
    rl_new_value = LoadValue(rl_src_new_value);
  } else if (load_early) {
    rl_new_value = LoadValueWide(rl_src_new_value, kCoreReg);
  }

  if (is_object && !mir_graph_->IsConstantNullRef(rl_new_value)) {
    // Mark card for object assuming new value is stored.
    MarkGCCard(rl_new_value.reg, rl_object.reg);
  }

  RegLocation rl_offset = LoadValue(rl_src_offset, kCoreReg);

  RegStorage r_ptr = rs_rA64_LR;
  OpRegRegReg(kOpAdd, r_ptr, rl_object.reg, rl_offset.reg);

  // Free now unneeded rl_object and rl_offset to give more temps.
  ClobberSReg(rl_object.s_reg_low);
  FreeTemp(rl_object.reg);
  ClobberSReg(rl_offset.s_reg_low);
  FreeTemp(rl_offset.reg);

  RegLocation rl_expected;
  if (!is_long) {
    rl_expected = LoadValue(rl_src_expected);
  } else if (load_early) {
    rl_expected = LoadValueWide(rl_src_expected, kCoreReg);
  } else {
    // NOTE: partially defined rl_expected & rl_new_value - but we just want the regs.
    int low_reg = AllocTemp().GetReg();
    int high_reg = AllocTemp().GetReg();
    rl_new_value.reg = RegStorage(RegStorage::k64BitPair, low_reg, high_reg);
    rl_expected = rl_new_value;
  }

  // do {
  //   tmp = [r_ptr] - expected;
  // } while (tmp == 0 && failure([r_ptr] <- r_new_value));
  // result = tmp != 0;

  RegStorage r_tmp = AllocTemp();
  LIR* target = NewLIR0(kPseudoTargetLabel);

  if (is_long) {
    RegStorage r_tmp_high = AllocTemp();
    if (!load_early) {
      LoadValueDirectWide(rl_src_expected, rl_expected.reg);
    }
    NewLIR3(kA64Ldxr2rX, r_tmp.GetReg(), r_tmp_high.GetReg(), r_ptr.GetReg());
    OpRegReg(kOpSub, r_tmp, rl_expected.reg.GetLow());
    OpRegReg(kOpSub, r_tmp_high, rl_expected.reg.GetHigh());
    if (!load_early) {
      LoadValueDirectWide(rl_src_new_value, rl_new_value.reg);
    }

    LIR* branch1 = OpCmpImmBranch(kCondNe, r_tmp, 0, NULL);
    LIR* branch2 = OpCmpImmBranch(kCondNe, r_tmp_high, 0, NULL);
    NewLIR4(WIDE(kA64Stxr3wrX) /* eq */, r_tmp.GetReg(), rl_new_value.reg.GetReg(),
            rl_new_value.reg.GetHighReg(), r_ptr.GetReg());
    LIR* target2 = NewLIR0(kPseudoTargetLabel);
    branch1->target = target2;
    branch2->target = target2;
    FreeTemp(r_tmp_high);  // Now unneeded

  } else {
    NewLIR3(kA64Ldxr2rX, r_tmp.GetReg(), r_ptr.GetReg(), 0);
    OpRegReg(kOpSub, r_tmp, rl_expected.reg);
    DCHECK(last_lir_insn_->u.m.def_mask & ENCODE_CCODE);
    // OpIT(kCondEq, "T");
    NewLIR4(kA64Stxr3wrX /* eq */, r_tmp.GetReg(), rl_new_value.reg.GetReg(), r_ptr.GetReg(), 0);
  }

  // Still one conditional left from OpIT(kCondEq, "T") from either branch
  OpRegImm(kOpCmp /* eq */, r_tmp, 1);
  OpCondBranch(kCondEq, target);

  if (!load_early) {
    FreeTemp(rl_expected.reg);  // Now unneeded.
  }

  // result := (tmp1 != 0) ? 0 : 1;
  RegLocation rl_result = EvalLoc(rl_dest, kCoreReg, true);
  OpRegRegImm(kOpRsub, rl_result.reg, r_tmp, 1);
  DCHECK(last_lir_insn_->u.m.def_mask & ENCODE_CCODE);
  // OpIT(kCondUlt, "");
  LoadConstant(rl_result.reg, 0); /* cc */
  FreeTemp(r_tmp);  // Now unneeded.

  StoreValue(rl_dest, rl_result);

  // Now, restore lr to its non-temp status.
  Clobber(rs_rA64_LR);
  UnmarkTemp(rs_rA64_LR);
  return true;
}

LIR* Arm64Mir2Lir::OpPcRelLoad(RegStorage reg, LIR* target) {
  return RawLIR(current_dalvik_offset_, WIDE(kA64Ldr2rp), reg.GetReg(), 0, 0, 0, 0, target);
}

LIR* Arm64Mir2Lir::OpVldm(RegStorage r_base, int count) {
  LOG(FATAL) << "Unexpected use of OpVldm for Arm64";
  return NULL;
}

LIR* Arm64Mir2Lir::OpVstm(RegStorage r_base, int count) {
  LOG(FATAL) << "Unexpected use of OpVstm for Arm64";
  return NULL;
}

void Arm64Mir2Lir::GenMultiplyByTwoBitMultiplier(RegLocation rl_src,
                                               RegLocation rl_result, int lit,
                                               int first_bit, int second_bit) {
  OpRegRegRegShift(kOpAdd, rl_result.reg, rl_src.reg, rl_src.reg, EncodeShift(kA64Lsl, second_bit - first_bit));
  if (first_bit != 0) {
    OpRegRegImm(kOpLsl, rl_result.reg, rl_result.reg, first_bit);
  }
}

void Arm64Mir2Lir::GenDivZeroCheckWide(RegStorage reg) {
  LOG(FATAL) << "Unexpected use of GenDivZero for Arm64";
}

// Test suspend flag, return target of taken suspend branch
LIR* Arm64Mir2Lir::OpTestSuspend(LIR* target) {
  // FIXME: Define rA64_SUSPEND as w19, when we do not need two copies of reserved register.
  // Note: The opcode is not set as wide, so actually we are using the 32-bit version register.
  NewLIR3(kA64Subs3rRd, rA64_SUSPEND, rA64_SUSPEND, 1);
  return OpCondBranch((target == NULL) ? kCondEq : kCondNe, target);
}

// Decrement register and branch on condition
LIR* Arm64Mir2Lir::OpDecAndBranch(ConditionCode c_code, RegStorage reg, LIR* target) {
  // Combine sub & test using sub setflags encoding here
  OpRegRegImm(kOpSub, reg, reg, 1);  // For value == 1, this should set flags.
  DCHECK(last_lir_insn_->u.m.def_mask & ENCODE_CCODE);
  return OpCondBranch(c_code, target);
}

bool Arm64Mir2Lir::GenMemBarrier(MemBarrierKind barrier_kind) {
#if ANDROID_SMP != 0
  // Start off with using the last LIR as the barrier. If it is not enough, then we will generate one.
  LIR* barrier = last_lir_insn_;

  int dmb_flavor;
  // TODO: revisit Arm barrier kinds
  switch (barrier_kind) {
    case kLoadStore: dmb_flavor = kISH; break;
    case kLoadLoad: dmb_flavor = kISH; break;
    case kStoreStore: dmb_flavor = kISHST; break;
    case kStoreLoad: dmb_flavor = kISH; break;
    default:
      LOG(FATAL) << "Unexpected MemBarrierKind: " << barrier_kind;
      dmb_flavor = kSY;  // quiet gcc.
      break;
  }

  bool ret = false;

  // If the same barrier already exists, don't generate another.
  if (barrier == nullptr
      || (barrier->opcode != kA64Dmb1B || barrier->operands[0] != dmb_flavor)) {
    barrier = NewLIR1(kA64Dmb1B, dmb_flavor);
    ret = true;
  }

  // At this point we must have a memory barrier. Mark it as a scheduling barrier as well.
  DCHECK(!barrier->flags.use_def_invalid);
  barrier->u.m.def_mask = ENCODE_ALL;
  return ret;
#else
  return false;
#endif
}

void Arm64Mir2Lir::GenIntToLong(RegLocation rl_dest, RegLocation rl_src) {
  RegLocation rl_result;

  rl_src = LoadValue(rl_src, kCoreReg);
  rl_result = EvalLocWide(rl_dest, kCoreReg, true);
  NewLIR4(WIDE(kA64Sbfm4rrdd), rl_result.reg.GetReg(), rl_src.reg.GetReg(), 0, 31);
  StoreValueWide(rl_dest, rl_result);
}

void Arm64Mir2Lir::GenDivRemLong(Instruction::Code opcode, RegLocation rl_dest,
                                 RegLocation rl_src1, RegLocation rl_src2, bool is_div) {
  RegLocation rl_result;
  rl_src1 = LoadValueWide(rl_src1, kCoreReg);
  rl_src2 = LoadValueWide(rl_src2, kCoreReg);
  GenDivZeroCheck(rl_src2.reg);
  rl_result = GenDivRem(rl_dest, rl_src1.reg, rl_src2.reg, is_div);
  StoreValueWide(rl_dest, rl_result);
}

void Arm64Mir2Lir::GenLongOp(OpKind op, RegLocation rl_dest, RegLocation rl_src1,
                             RegLocation rl_src2) {
  RegLocation rl_result;

  rl_src1 = LoadValueWide(rl_src1, kCoreReg);
  rl_src2 = LoadValueWide(rl_src2, kCoreReg);
  rl_result = EvalLocWide(rl_dest, kCoreReg, true);
  OpRegRegRegShift(op, rl_result.reg, rl_src1.reg, rl_src2.reg, ENCODE_NO_SHIFT);
  StoreValueWide(rl_dest, rl_result);
}

void Arm64Mir2Lir::GenNegLong(RegLocation rl_dest, RegLocation rl_src) {
  RegLocation rl_result;

  rl_src = LoadValueWide(rl_src, kCoreReg);
  rl_result = EvalLocWide(rl_dest, kCoreReg, true);
  OpRegRegShift(kOpNeg, rl_result.reg, rl_src.reg, ENCODE_NO_SHIFT);
  StoreValueWide(rl_dest, rl_result);
}

void Arm64Mir2Lir::GenNotLong(RegLocation rl_dest, RegLocation rl_src) {
  RegLocation rl_result;

  rl_src = LoadValueWide(rl_src, kCoreReg);
  rl_result = EvalLocWide(rl_dest, kCoreReg, true);
  OpRegRegShift(kOpMvn, rl_result.reg, rl_src.reg, ENCODE_NO_SHIFT);
  StoreValueWide(rl_dest, rl_result);
}

void Arm64Mir2Lir::GenMulLong(Instruction::Code opcode, RegLocation rl_dest,
                              RegLocation rl_src1, RegLocation rl_src2) {
  GenLongOp(kOpMul, rl_dest, rl_src1, rl_src2);
}

void Arm64Mir2Lir::GenAddLong(Instruction::Code opcode, RegLocation rl_dest, RegLocation rl_src1,
                              RegLocation rl_src2) {
  GenLongOp(kOpAdd, rl_dest, rl_src1, rl_src2);
}

void Arm64Mir2Lir::GenSubLong(Instruction::Code opcode, RegLocation rl_dest, RegLocation rl_src1,
                            RegLocation rl_src2) {
  GenLongOp(kOpSub, rl_dest, rl_src1, rl_src2);
}

void Arm64Mir2Lir::GenAndLong(Instruction::Code opcode, RegLocation rl_dest, RegLocation rl_src1,
                            RegLocation rl_src2) {
  GenLongOp(kOpAnd, rl_dest, rl_src1, rl_src2);
}

void Arm64Mir2Lir::GenOrLong(Instruction::Code opcode, RegLocation rl_dest, RegLocation rl_src1,
                           RegLocation rl_src2) {
  GenLongOp(kOpOr, rl_dest, rl_src1, rl_src2);
}

void Arm64Mir2Lir::GenXorLong(Instruction::Code opcode, RegLocation rl_dest, RegLocation rl_src1,
                            RegLocation rl_src2) {
  GenLongOp(kOpXor, rl_dest, rl_src1, rl_src2);
}

/*
 * Generate array load
 */
void Arm64Mir2Lir::GenArrayGet(int opt_flags, OpSize size, RegLocation rl_array,
                             RegLocation rl_index, RegLocation rl_dest, int scale) {
  // TODO(Arm64): check this.
  UNIMPLEMENTED(WARNING);

  RegisterClass reg_class = RegClassBySize(size);
  int len_offset = mirror::Array::LengthOffset().Int32Value();
  int data_offset;
  RegLocation rl_result;
  bool constant_index = rl_index.is_const;
  rl_array = LoadValue(rl_array, kRefReg);
  if (!constant_index) {
    rl_index = LoadValue(rl_index, kCoreReg);
  }

  if (rl_dest.wide) {
    data_offset = mirror::Array::DataOffset(sizeof(int64_t)).Int32Value();
  } else {
    data_offset = mirror::Array::DataOffset(sizeof(int32_t)).Int32Value();
  }

  // If index is constant, just fold it into the data offset
  if (constant_index) {
    data_offset += mir_graph_->ConstantValue(rl_index) << scale;
  }

  /* null object? */
  GenNullCheck(rl_array.reg, opt_flags);

  bool needs_range_check = (!(opt_flags & MIR_IGNORE_RANGE_CHECK));
  RegStorage reg_len;
  if (needs_range_check) {
    reg_len = AllocTemp();
    /* Get len */
    Load32Disp(rl_array.reg, len_offset, reg_len);
    MarkPossibleNullPointerException(opt_flags);
  } else {
    ForceImplicitNullCheck(rl_array.reg, opt_flags);
  }
  if (rl_dest.wide || rl_dest.fp || constant_index) {
    RegStorage reg_ptr;
    if (constant_index) {
      reg_ptr = rl_array.reg;  // NOTE: must not alter reg_ptr in constant case.
    } else {
      // No special indexed operation, lea + load w/ displacement
      reg_ptr = AllocTempRef();
      OpRegRegRegShift(kOpAdd, reg_ptr, rl_array.reg, rl_index.reg, EncodeShift(kA64Lsl, scale));
      FreeTemp(rl_index.reg);
    }
    rl_result = EvalLoc(rl_dest, reg_class, true);

    if (needs_range_check) {
      if (constant_index) {
        GenArrayBoundsCheck(mir_graph_->ConstantValue(rl_index), reg_len);
      } else {
        GenArrayBoundsCheck(rl_index.reg, reg_len);
      }
      FreeTemp(reg_len);
    }
    LoadBaseDisp(reg_ptr, data_offset, rl_result.reg, size);
    MarkPossibleNullPointerException(opt_flags);
    if (!constant_index) {
      FreeTemp(reg_ptr);
    }
    if (rl_dest.wide) {
      StoreValueWide(rl_dest, rl_result);
    } else {
      StoreValue(rl_dest, rl_result);
    }
  } else {
    // Offset base, then use indexed load
    RegStorage reg_ptr = AllocTempRef();
    OpRegRegImm(kOpAdd, reg_ptr, rl_array.reg, data_offset);
    FreeTemp(rl_array.reg);
    rl_result = EvalLoc(rl_dest, reg_class, true);

    if (needs_range_check) {
      GenArrayBoundsCheck(rl_index.reg, reg_len);
      FreeTemp(reg_len);
    }
    LoadBaseIndexed(reg_ptr, rl_index.reg, rl_result.reg, scale, size);
    MarkPossibleNullPointerException(opt_flags);
    FreeTemp(reg_ptr);
    StoreValue(rl_dest, rl_result);
  }
}

/*
 * Generate array store
 *
 */
void Arm64Mir2Lir::GenArrayPut(int opt_flags, OpSize size, RegLocation rl_array,
                             RegLocation rl_index, RegLocation rl_src, int scale, bool card_mark) {
  // TODO(Arm64): check this.
  UNIMPLEMENTED(WARNING);

  RegisterClass reg_class = RegClassBySize(size);
  int len_offset = mirror::Array::LengthOffset().Int32Value();
  bool constant_index = rl_index.is_const;

  int data_offset;
  if (size == k64 || size == kDouble) {
    data_offset = mirror::Array::DataOffset(sizeof(int64_t)).Int32Value();
  } else {
    data_offset = mirror::Array::DataOffset(sizeof(int32_t)).Int32Value();
  }

  // If index is constant, just fold it into the data offset.
  if (constant_index) {
    data_offset += mir_graph_->ConstantValue(rl_index) << scale;
  }

  rl_array = LoadValue(rl_array, kRefReg);
  if (!constant_index) {
    rl_index = LoadValue(rl_index, kCoreReg);
  }

  RegStorage reg_ptr;
  bool allocated_reg_ptr_temp = false;
  if (constant_index) {
    reg_ptr = rl_array.reg;
  } else if (IsTemp(rl_array.reg) && !card_mark) {
    Clobber(rl_array.reg);
    reg_ptr = rl_array.reg;
  } else {
    allocated_reg_ptr_temp = true;
    reg_ptr = AllocTempRef();
  }

  /* null object? */
  GenNullCheck(rl_array.reg, opt_flags);

  bool needs_range_check = (!(opt_flags & MIR_IGNORE_RANGE_CHECK));
  RegStorage reg_len;
  if (needs_range_check) {
    reg_len = AllocTemp();
    // NOTE: max live temps(4) here.
    /* Get len */
    Load32Disp(rl_array.reg, len_offset, reg_len);
    MarkPossibleNullPointerException(opt_flags);
  } else {
    ForceImplicitNullCheck(rl_array.reg, opt_flags);
  }
  /* at this point, reg_ptr points to array, 2 live temps */
  if (rl_src.wide || rl_src.fp || constant_index) {
    if (rl_src.wide) {
      rl_src = LoadValueWide(rl_src, reg_class);
    } else {
      rl_src = LoadValue(rl_src, reg_class);
    }
    if (!constant_index) {
      OpRegRegRegShift(kOpAdd, reg_ptr, rl_array.reg, rl_index.reg, EncodeShift(kA64Lsl, scale));
    }
    if (needs_range_check) {
      if (constant_index) {
        GenArrayBoundsCheck(mir_graph_->ConstantValue(rl_index), reg_len);
      } else {
        GenArrayBoundsCheck(rl_index.reg, reg_len);
      }
      FreeTemp(reg_len);
    }

    StoreBaseDisp(reg_ptr, data_offset, rl_src.reg, size);
    MarkPossibleNullPointerException(opt_flags);
  } else {
    /* reg_ptr -> array data */
    OpRegRegImm(kOpAdd, reg_ptr, rl_array.reg, data_offset);
    rl_src = LoadValue(rl_src, reg_class);
    if (needs_range_check) {
      GenArrayBoundsCheck(rl_index.reg, reg_len);
      FreeTemp(reg_len);
    }
    StoreBaseIndexed(reg_ptr, rl_index.reg, rl_src.reg, scale, size);
    MarkPossibleNullPointerException(opt_flags);
  }
  if (allocated_reg_ptr_temp) {
    FreeTemp(reg_ptr);
  }
  if (card_mark) {
    MarkGCCard(rl_src.reg, rl_array.reg);
  }
}

void Arm64Mir2Lir::GenShiftImmOpLong(Instruction::Code opcode,
                                   RegLocation rl_dest, RegLocation rl_src, RegLocation rl_shift) {
  OpKind op = kOpBkpt;
  // Per spec, we only care about low 6 bits of shift amount.
  int shift_amount = mir_graph_->ConstantValue(rl_shift) & 0x3f;
  rl_src = LoadValueWide(rl_src, kCoreReg);
  if (shift_amount == 0) {
    StoreValueWide(rl_dest, rl_src);
    return;
  }

  RegLocation rl_result = EvalLocWide(rl_dest, kCoreReg, true);
  switch (opcode) {
    case Instruction::SHL_LONG:
    case Instruction::SHL_LONG_2ADDR:
      op = kOpLsl;
      break;
    case Instruction::SHR_LONG:
    case Instruction::SHR_LONG_2ADDR:
      op = kOpAsr;
      break;
    case Instruction::USHR_LONG:
    case Instruction::USHR_LONG_2ADDR:
      op = kOpLsr;
      break;
    default:
      LOG(FATAL) << "Unexpected case";
  }
  OpRegRegImm(op, rl_result.reg, rl_src.reg, shift_amount);
  StoreValueWide(rl_dest, rl_result);
}

void Arm64Mir2Lir::GenArithImmOpLong(Instruction::Code opcode, RegLocation rl_dest,
                                     RegLocation rl_src1, RegLocation rl_src2) {
  if ((opcode == Instruction::SUB_LONG) || (opcode == Instruction::SUB_LONG_2ADDR)) {
    if (!rl_src2.is_const) {
      return GenArithOpLong(opcode, rl_dest, rl_src1, rl_src2);
    }
  } else {
    // Associativity.
    if (!rl_src2.is_const) {
      DCHECK(rl_src1.is_const);
      std::swap(rl_src1, rl_src2);
    }
  }
  DCHECK(rl_src2.is_const);

  OpKind op = kOpBkpt;
  int64_t val = mir_graph_->ConstantValueWide(rl_src2);

  switch (opcode) {
    case Instruction::ADD_LONG:
    case Instruction::ADD_LONG_2ADDR:
      op = kOpAdd;
      break;
    case Instruction::SUB_LONG:
    case Instruction::SUB_LONG_2ADDR:
      op = kOpSub;
      break;
    case Instruction::AND_LONG:
    case Instruction::AND_LONG_2ADDR:
      op = kOpAnd;
      break;
    case Instruction::OR_LONG:
    case Instruction::OR_LONG_2ADDR:
      op = kOpOr;
      break;
    case Instruction::XOR_LONG:
    case Instruction::XOR_LONG_2ADDR:
      op = kOpXor;
      break;
    default:
      LOG(FATAL) << "Unexpected opcode";
  }

  rl_src1 = LoadValueWide(rl_src1, kCoreReg);
  RegLocation rl_result = EvalLocWide(rl_dest, kCoreReg, true);
  OpRegRegImm(op, rl_result.reg, rl_src1.reg, val);
  StoreValueWide(rl_dest, rl_result);
}

/**
 * @brief Split a register list in pairs or registers.
 *
 * Given a list of registers in @p reg_mask, split the list in pairs. Use as follows:
 * @code
 *   int reg1 = -1, reg2 = -1;
 *   while (reg_mask) {
 *     reg_mask = GenPairWise(reg_mask, & reg1, & reg2);
 *     if (UNLIKELY(reg2 < 0)) {
 *       // Single register in reg1.
 *     } else {
 *       // Pair in reg1, reg2.
 *     }
 *   }
 * @endcode
 */
uint32_t Arm64Mir2Lir::GenPairWise(uint32_t reg_mask, int* reg1, int* reg2) {
  // Find first register.
  int first_bit_set = __builtin_ctz(reg_mask) + 1;
  int reg = *reg1 + first_bit_set;
  reg_mask >>= first_bit_set;

  if (LIKELY(reg_mask)) {
    // Save the first register, find the second and use the pair opcode.
    int second_bit_set = __builtin_ctz(reg_mask) + 1;
    *reg2 = reg;
    reg_mask >>= second_bit_set;
    *reg1 = reg + second_bit_set;
    return reg_mask;
  }

  // Use the single opcode, as we just have one register.
  *reg1 = reg;
  *reg2 = -1;
  return reg_mask;
}

void Arm64Mir2Lir::UnSpillCoreRegs(RegStorage base, int offset, uint32_t reg_mask) {
  int reg1 = -1, reg2 = -1;
  const int reg_log2_size = 3;

  for (offset = (offset >> reg_log2_size); reg_mask; offset += 2) {
     reg_mask = GenPairWise(reg_mask, & reg1, & reg2);
    if (UNLIKELY(reg2 < 0)) {
      NewLIR3(WIDE(kA64Ldr3rXD), RegStorage::Solo64(reg1).GetReg(), base.GetReg(), offset);
    } else {
      NewLIR4(WIDE(kA64Ldp4rrXD), RegStorage::Solo64(reg2).GetReg(),
              RegStorage::Solo64(reg1).GetReg(), base.GetReg(), offset);
    }
  }
}

void Arm64Mir2Lir::SpillCoreRegs(RegStorage base, int offset, uint32_t reg_mask) {
  int reg1 = -1, reg2 = -1;
  const int reg_log2_size = 3;

  for (offset = (offset >> reg_log2_size); reg_mask; offset += 2) {
    reg_mask = GenPairWise(reg_mask, & reg1, & reg2);
    if (UNLIKELY(reg2 < 0)) {
      NewLIR3(WIDE(kA64Str3rXD), RegStorage::Solo64(reg1).GetReg(), base.GetReg(), offset);
    } else {
      NewLIR4(WIDE(kA64Stp4rrXD), RegStorage::Solo64(reg2).GetReg(),
              RegStorage::Solo64(reg1).GetReg(), base.GetReg(), offset);
    }
  }
}

void Arm64Mir2Lir::UnSpillFPRegs(RegStorage base, int offset, uint32_t reg_mask) {
  int reg1 = -1, reg2 = -1;
  const int reg_log2_size = 3;

  for (offset = (offset >> reg_log2_size); reg_mask; offset += 2) {
     reg_mask = GenPairWise(reg_mask, & reg1, & reg2);
    if (UNLIKELY(reg2 < 0)) {
      NewLIR3(FWIDE(kA64Ldr3fXD), RegStorage::FloatSolo64(reg1).GetReg(), base.GetReg(), offset);
    } else {
      NewLIR4(WIDE(kA64Ldp4ffXD), RegStorage::FloatSolo64(reg2).GetReg(),
              RegStorage::FloatSolo64(reg1).GetReg(), base.GetReg(), offset);
    }
  }
}

// TODO(Arm64): consider using ld1 and st1?
void Arm64Mir2Lir::SpillFPRegs(RegStorage base, int offset, uint32_t reg_mask) {
  int reg1 = -1, reg2 = -1;
  const int reg_log2_size = 3;

  for (offset = (offset >> reg_log2_size); reg_mask; offset += 2) {
    reg_mask = GenPairWise(reg_mask, & reg1, & reg2);
    if (UNLIKELY(reg2 < 0)) {
      NewLIR3(FWIDE(kA64Str3fXD), RegStorage::FloatSolo64(reg1).GetReg(), base.GetReg(), offset);
    } else {
      NewLIR4(WIDE(kA64Stp4ffXD), RegStorage::FloatSolo64(reg2).GetReg(),
              RegStorage::FloatSolo64(reg1).GetReg(), base.GetReg(), offset);
    }
  }
}

}  // namespace art