Home | History | Annotate | Download | only in X86
      1 //===-- X86FixupLEAs.cpp - use or replace LEA instructions -----------===//
      2 //
      3 //                     The LLVM Compiler Infrastructure
      4 //
      5 // This file is distributed under the University of Illinois Open Source
      6 // License. See LICENSE.TXT for details.
      7 //
      8 //===----------------------------------------------------------------------===//
      9 //
     10 // This file defines the pass that finds instructions that can be
     11 // re-written as LEA instructions in order to reduce pipeline delays.
     12 // When optimizing for size it replaces suitable LEAs with INC or DEC.
     13 //
     14 //===----------------------------------------------------------------------===//
     15 
     16 #include "X86.h"
     17 #include "X86InstrInfo.h"
     18 #include "X86Subtarget.h"
     19 #include "llvm/ADT/Statistic.h"
     20 #include "llvm/CodeGen/MachineFunctionPass.h"
     21 #include "llvm/CodeGen/MachineInstrBuilder.h"
     22 #include "llvm/CodeGen/Passes.h"
     23 #include "llvm/CodeGen/TargetSchedule.h"
     24 #include "llvm/Support/Debug.h"
     25 #include "llvm/Support/raw_ostream.h"
     26 using namespace llvm;
     27 
     28 namespace llvm {
     29 void initializeFixupLEAPassPass(PassRegistry &);
     30 }
     31 
     32 #define FIXUPLEA_DESC "X86 LEA Fixup"
     33 #define FIXUPLEA_NAME "x86-fixup-LEAs"
     34 
     35 #define DEBUG_TYPE FIXUPLEA_NAME
     36 
     37 STATISTIC(NumLEAs, "Number of LEA instructions created");
     38 
     39 namespace {
     40 class FixupLEAPass : public MachineFunctionPass {
     41   enum RegUsageState { RU_NotUsed, RU_Write, RU_Read };
     42 
     43   /// Loop over all of the instructions in the basic block
     44   /// replacing applicable instructions with LEA instructions,
     45   /// where appropriate.
     46   bool processBasicBlock(MachineFunction &MF, MachineFunction::iterator MFI);
     47 
     48 
     49   /// Given a machine register, look for the instruction
     50   /// which writes it in the current basic block. If found,
     51   /// try to replace it with an equivalent LEA instruction.
     52   /// If replacement succeeds, then also process the newly created
     53   /// instruction.
     54   void seekLEAFixup(MachineOperand &p, MachineBasicBlock::iterator &I,
     55                     MachineFunction::iterator MFI);
     56 
     57   /// Given a memory access or LEA instruction
     58   /// whose address mode uses a base and/or index register, look for
     59   /// an opportunity to replace the instruction which sets the base or index
     60   /// register with an equivalent LEA instruction.
     61   void processInstruction(MachineBasicBlock::iterator &I,
     62                           MachineFunction::iterator MFI);
     63 
     64   /// Given a LEA instruction which is unprofitable
     65   /// on Silvermont try to replace it with an equivalent ADD instruction
     66   void processInstructionForSLM(MachineBasicBlock::iterator &I,
     67                                 MachineFunction::iterator MFI);
     68 
     69 
     70   /// Given a LEA instruction which is unprofitable
     71   /// on SNB+ try to replace it with other instructions.
     72   /// According to Intel's Optimization Reference Manual:
     73   /// " For LEA instructions with three source operands and some specific
     74   ///   situations, instruction latency has increased to 3 cycles, and must
     75   ///   dispatch via port 1:
     76   /// - LEA that has all three source operands: base, index, and offset
     77   /// - LEA that uses base and index registers where the base is EBP, RBP,
     78   ///   or R13
     79   /// - LEA that uses RIP relative addressing mode
     80   /// - LEA that uses 16-bit addressing mode "
     81   /// This function currently handles the first 2 cases only.
     82   MachineInstr *processInstrForSlow3OpLEA(MachineInstr &MI,
     83                                           MachineFunction::iterator MFI);
     84 
     85   /// Look for LEAs that add 1 to reg or subtract 1 from reg
     86   /// and convert them to INC or DEC respectively.
     87   bool fixupIncDec(MachineBasicBlock::iterator &I,
     88                    MachineFunction::iterator MFI) const;
     89 
     90   /// Determine if an instruction references a machine register
     91   /// and, if so, whether it reads or writes the register.
     92   RegUsageState usesRegister(MachineOperand &p, MachineBasicBlock::iterator I);
     93 
     94   /// Step backwards through a basic block, looking
     95   /// for an instruction which writes a register within
     96   /// a maximum of INSTR_DISTANCE_THRESHOLD instruction latency cycles.
     97   MachineBasicBlock::iterator searchBackwards(MachineOperand &p,
     98                                               MachineBasicBlock::iterator &I,
     99                                               MachineFunction::iterator MFI);
    100 
    101   /// if an instruction can be converted to an
    102   /// equivalent LEA, insert the new instruction into the basic block
    103   /// and return a pointer to it. Otherwise, return zero.
    104   MachineInstr *postRAConvertToLEA(MachineFunction::iterator &MFI,
    105                                    MachineBasicBlock::iterator &MBBI) const;
    106 
    107 public:
    108   static char ID;
    109 
    110   StringRef getPassName() const override { return FIXUPLEA_DESC; }
    111 
    112   FixupLEAPass() : MachineFunctionPass(ID) {
    113     initializeFixupLEAPassPass(*PassRegistry::getPassRegistry());
    114   }
    115 
    116   /// Loop over all of the basic blocks,
    117   /// replacing instructions by equivalent LEA instructions
    118   /// if needed and when possible.
    119   bool runOnMachineFunction(MachineFunction &MF) override;
    120 
    121   // This pass runs after regalloc and doesn't support VReg operands.
    122   MachineFunctionProperties getRequiredProperties() const override {
    123     return MachineFunctionProperties().set(
    124         MachineFunctionProperties::Property::NoVRegs);
    125   }
    126 
    127 private:
    128   TargetSchedModel TSM;
    129   MachineFunction *MF;
    130   const X86InstrInfo *TII; // Machine instruction info.
    131   bool OptIncDec;
    132   bool OptLEA;
    133 };
    134 }
    135 
    136 char FixupLEAPass::ID = 0;
    137 
    138 INITIALIZE_PASS(FixupLEAPass, FIXUPLEA_NAME, FIXUPLEA_DESC, false, false)
    139 
    140 MachineInstr *
    141 FixupLEAPass::postRAConvertToLEA(MachineFunction::iterator &MFI,
    142                                  MachineBasicBlock::iterator &MBBI) const {
    143   MachineInstr &MI = *MBBI;
    144   switch (MI.getOpcode()) {
    145   case X86::MOV32rr:
    146   case X86::MOV64rr: {
    147     const MachineOperand &Src = MI.getOperand(1);
    148     const MachineOperand &Dest = MI.getOperand(0);
    149     MachineInstr *NewMI =
    150         BuildMI(*MF, MI.getDebugLoc(),
    151                 TII->get(MI.getOpcode() == X86::MOV32rr ? X86::LEA32r
    152                                                         : X86::LEA64r))
    153             .add(Dest)
    154             .add(Src)
    155             .addImm(1)
    156             .addReg(0)
    157             .addImm(0)
    158             .addReg(0);
    159     MFI->insert(MBBI, NewMI); // Insert the new inst
    160     return NewMI;
    161   }
    162   case X86::ADD64ri32:
    163   case X86::ADD64ri8:
    164   case X86::ADD64ri32_DB:
    165   case X86::ADD64ri8_DB:
    166   case X86::ADD32ri:
    167   case X86::ADD32ri8:
    168   case X86::ADD32ri_DB:
    169   case X86::ADD32ri8_DB:
    170   case X86::ADD16ri:
    171   case X86::ADD16ri8:
    172   case X86::ADD16ri_DB:
    173   case X86::ADD16ri8_DB:
    174     if (!MI.getOperand(2).isImm()) {
    175       // convertToThreeAddress will call getImm()
    176       // which requires isImm() to be true
    177       return nullptr;
    178     }
    179     break;
    180   case X86::ADD16rr:
    181   case X86::ADD16rr_DB:
    182     if (MI.getOperand(1).getReg() != MI.getOperand(2).getReg()) {
    183       // if src1 != src2, then convertToThreeAddress will
    184       // need to create a Virtual register, which we cannot do
    185       // after register allocation.
    186       return nullptr;
    187     }
    188   }
    189   return TII->convertToThreeAddress(MFI, MI, nullptr);
    190 }
    191 
    192 FunctionPass *llvm::createX86FixupLEAs() { return new FixupLEAPass(); }
    193 
    194 bool FixupLEAPass::runOnMachineFunction(MachineFunction &Func) {
    195   if (skipFunction(Func.getFunction()))
    196     return false;
    197 
    198   MF = &Func;
    199   const X86Subtarget &ST = Func.getSubtarget<X86Subtarget>();
    200   OptIncDec = !ST.slowIncDec() || Func.getFunction().optForMinSize();
    201   OptLEA = ST.LEAusesAG() || ST.slowLEA() || ST.slow3OpsLEA();
    202 
    203   if (!OptLEA && !OptIncDec)
    204     return false;
    205 
    206   TSM.init(&Func.getSubtarget());
    207   TII = ST.getInstrInfo();
    208 
    209   LLVM_DEBUG(dbgs() << "Start X86FixupLEAs\n";);
    210   // Process all basic blocks.
    211   for (MachineFunction::iterator I = Func.begin(), E = Func.end(); I != E; ++I)
    212     processBasicBlock(Func, I);
    213   LLVM_DEBUG(dbgs() << "End X86FixupLEAs\n";);
    214 
    215   return true;
    216 }
    217 
    218 FixupLEAPass::RegUsageState
    219 FixupLEAPass::usesRegister(MachineOperand &p, MachineBasicBlock::iterator I) {
    220   RegUsageState RegUsage = RU_NotUsed;
    221   MachineInstr &MI = *I;
    222 
    223   for (unsigned int i = 0; i < MI.getNumOperands(); ++i) {
    224     MachineOperand &opnd = MI.getOperand(i);
    225     if (opnd.isReg() && opnd.getReg() == p.getReg()) {
    226       if (opnd.isDef())
    227         return RU_Write;
    228       RegUsage = RU_Read;
    229     }
    230   }
    231   return RegUsage;
    232 }
    233 
    234 /// getPreviousInstr - Given a reference to an instruction in a basic
    235 /// block, return a reference to the previous instruction in the block,
    236 /// wrapping around to the last instruction of the block if the block
    237 /// branches to itself.
    238 static inline bool getPreviousInstr(MachineBasicBlock::iterator &I,
    239                                     MachineFunction::iterator MFI) {
    240   if (I == MFI->begin()) {
    241     if (MFI->isPredecessor(&*MFI)) {
    242       I = --MFI->end();
    243       return true;
    244     } else
    245       return false;
    246   }
    247   --I;
    248   return true;
    249 }
    250 
    251 MachineBasicBlock::iterator
    252 FixupLEAPass::searchBackwards(MachineOperand &p, MachineBasicBlock::iterator &I,
    253                               MachineFunction::iterator MFI) {
    254   int InstrDistance = 1;
    255   MachineBasicBlock::iterator CurInst;
    256   static const int INSTR_DISTANCE_THRESHOLD = 5;
    257 
    258   CurInst = I;
    259   bool Found;
    260   Found = getPreviousInstr(CurInst, MFI);
    261   while (Found && I != CurInst) {
    262     if (CurInst->isCall() || CurInst->isInlineAsm())
    263       break;
    264     if (InstrDistance > INSTR_DISTANCE_THRESHOLD)
    265       break; // too far back to make a difference
    266     if (usesRegister(p, CurInst) == RU_Write) {
    267       return CurInst;
    268     }
    269     InstrDistance += TSM.computeInstrLatency(&*CurInst);
    270     Found = getPreviousInstr(CurInst, MFI);
    271   }
    272   return MachineBasicBlock::iterator();
    273 }
    274 
    275 static inline bool isLEA(const int Opcode) {
    276   return Opcode == X86::LEA16r || Opcode == X86::LEA32r ||
    277          Opcode == X86::LEA64r || Opcode == X86::LEA64_32r;
    278 }
    279 
    280 static inline bool isInefficientLEAReg(unsigned int Reg) {
    281   return Reg == X86::EBP || Reg == X86::RBP || Reg == X86::R13;
    282 }
    283 
    284 static inline bool isRegOperand(const MachineOperand &Op) {
    285   return Op.isReg() && Op.getReg() != X86::NoRegister;
    286 }
    287 /// hasIneffecientLEARegs - LEA that uses base and index registers
    288 /// where the base is EBP, RBP, or R13
    289 // TODO: use a variant scheduling class to model the latency profile
    290 // of LEA instructions, and implement this logic as a scheduling predicate.
    291 static inline bool hasInefficientLEABaseReg(const MachineOperand &Base,
    292                                             const MachineOperand &Index) {
    293   return Base.isReg() && isInefficientLEAReg(Base.getReg()) &&
    294          isRegOperand(Index);
    295 }
    296 
    297 static inline bool hasLEAOffset(const MachineOperand &Offset) {
    298   return (Offset.isImm() && Offset.getImm() != 0) || Offset.isGlobal();
    299 }
    300 
    301 static inline int getADDrrFromLEA(int LEAOpcode) {
    302   switch (LEAOpcode) {
    303   default:
    304     llvm_unreachable("Unexpected LEA instruction");
    305   case X86::LEA16r:
    306     return X86::ADD16rr;
    307   case X86::LEA32r:
    308     return X86::ADD32rr;
    309   case X86::LEA64_32r:
    310   case X86::LEA64r:
    311     return X86::ADD64rr;
    312   }
    313 }
    314 
    315 static inline int getADDriFromLEA(int LEAOpcode, const MachineOperand &Offset) {
    316   bool IsInt8 = Offset.isImm() && isInt<8>(Offset.getImm());
    317   switch (LEAOpcode) {
    318   default:
    319     llvm_unreachable("Unexpected LEA instruction");
    320   case X86::LEA16r:
    321     return IsInt8 ? X86::ADD16ri8 : X86::ADD16ri;
    322   case X86::LEA32r:
    323   case X86::LEA64_32r:
    324     return IsInt8 ? X86::ADD32ri8 : X86::ADD32ri;
    325   case X86::LEA64r:
    326     return IsInt8 ? X86::ADD64ri8 : X86::ADD64ri32;
    327   }
    328 }
    329 
    330 /// isLEASimpleIncOrDec - Does this LEA have one these forms:
    331 /// lea  %reg, 1(%reg)
    332 /// lea  %reg, -1(%reg)
    333 static inline bool isLEASimpleIncOrDec(MachineInstr &LEA) {
    334   unsigned SrcReg = LEA.getOperand(1 + X86::AddrBaseReg).getReg();
    335   unsigned DstReg = LEA.getOperand(0).getReg();
    336   unsigned AddrDispOp = 1 + X86::AddrDisp;
    337   return SrcReg == DstReg &&
    338          LEA.getOperand(1 + X86::AddrIndexReg).getReg() == 0 &&
    339          LEA.getOperand(1 + X86::AddrSegmentReg).getReg() == 0 &&
    340          LEA.getOperand(AddrDispOp).isImm() &&
    341          (LEA.getOperand(AddrDispOp).getImm() == 1 ||
    342           LEA.getOperand(AddrDispOp).getImm() == -1);
    343 }
    344 
    345 bool FixupLEAPass::fixupIncDec(MachineBasicBlock::iterator &I,
    346                                MachineFunction::iterator MFI) const {
    347   MachineInstr &MI = *I;
    348   int Opcode = MI.getOpcode();
    349   if (!isLEA(Opcode))
    350     return false;
    351 
    352   if (isLEASimpleIncOrDec(MI) && TII->isSafeToClobberEFLAGS(*MFI, I)) {
    353     int NewOpcode;
    354     bool isINC = MI.getOperand(4).getImm() == 1;
    355     switch (Opcode) {
    356     case X86::LEA16r:
    357       NewOpcode = isINC ? X86::INC16r : X86::DEC16r;
    358       break;
    359     case X86::LEA32r:
    360     case X86::LEA64_32r:
    361       NewOpcode = isINC ? X86::INC32r : X86::DEC32r;
    362       break;
    363     case X86::LEA64r:
    364       NewOpcode = isINC ? X86::INC64r : X86::DEC64r;
    365       break;
    366     }
    367 
    368     MachineInstr *NewMI =
    369         BuildMI(*MFI, I, MI.getDebugLoc(), TII->get(NewOpcode))
    370             .add(MI.getOperand(0))
    371             .add(MI.getOperand(1));
    372     MFI->erase(I);
    373     I = static_cast<MachineBasicBlock::iterator>(NewMI);
    374     return true;
    375   }
    376   return false;
    377 }
    378 
    379 void FixupLEAPass::processInstruction(MachineBasicBlock::iterator &I,
    380                                       MachineFunction::iterator MFI) {
    381   // Process a load, store, or LEA instruction.
    382   MachineInstr &MI = *I;
    383   const MCInstrDesc &Desc = MI.getDesc();
    384   int AddrOffset = X86II::getMemoryOperandNo(Desc.TSFlags);
    385   if (AddrOffset >= 0) {
    386     AddrOffset += X86II::getOperandBias(Desc);
    387     MachineOperand &p = MI.getOperand(AddrOffset + X86::AddrBaseReg);
    388     if (p.isReg() && p.getReg() != X86::ESP) {
    389       seekLEAFixup(p, I, MFI);
    390     }
    391     MachineOperand &q = MI.getOperand(AddrOffset + X86::AddrIndexReg);
    392     if (q.isReg() && q.getReg() != X86::ESP) {
    393       seekLEAFixup(q, I, MFI);
    394     }
    395   }
    396 }
    397 
    398 void FixupLEAPass::seekLEAFixup(MachineOperand &p,
    399                                 MachineBasicBlock::iterator &I,
    400                                 MachineFunction::iterator MFI) {
    401   MachineBasicBlock::iterator MBI = searchBackwards(p, I, MFI);
    402   if (MBI != MachineBasicBlock::iterator()) {
    403     MachineInstr *NewMI = postRAConvertToLEA(MFI, MBI);
    404     if (NewMI) {
    405       ++NumLEAs;
    406       LLVM_DEBUG(dbgs() << "FixLEA: Candidate to replace:"; MBI->dump(););
    407       // now to replace with an equivalent LEA...
    408       LLVM_DEBUG(dbgs() << "FixLEA: Replaced by: "; NewMI->dump(););
    409       MFI->erase(MBI);
    410       MachineBasicBlock::iterator J =
    411           static_cast<MachineBasicBlock::iterator>(NewMI);
    412       processInstruction(J, MFI);
    413     }
    414   }
    415 }
    416 
    417 void FixupLEAPass::processInstructionForSLM(MachineBasicBlock::iterator &I,
    418                                             MachineFunction::iterator MFI) {
    419   MachineInstr &MI = *I;
    420   const int Opcode = MI.getOpcode();
    421   if (!isLEA(Opcode))
    422     return;
    423   if (MI.getOperand(5).getReg() != 0 || !MI.getOperand(4).isImm() ||
    424       !TII->isSafeToClobberEFLAGS(*MFI, I))
    425     return;
    426   const unsigned DstR = MI.getOperand(0).getReg();
    427   const unsigned SrcR1 = MI.getOperand(1).getReg();
    428   const unsigned SrcR2 = MI.getOperand(3).getReg();
    429   if ((SrcR1 == 0 || SrcR1 != DstR) && (SrcR2 == 0 || SrcR2 != DstR))
    430     return;
    431   if (MI.getOperand(2).getImm() > 1)
    432     return;
    433   LLVM_DEBUG(dbgs() << "FixLEA: Candidate to replace:"; I->dump(););
    434   LLVM_DEBUG(dbgs() << "FixLEA: Replaced by: ";);
    435   MachineInstr *NewMI = nullptr;
    436   // Make ADD instruction for two registers writing to LEA's destination
    437   if (SrcR1 != 0 && SrcR2 != 0) {
    438     const MCInstrDesc &ADDrr = TII->get(getADDrrFromLEA(Opcode));
    439     const MachineOperand &Src = MI.getOperand(SrcR1 == DstR ? 3 : 1);
    440     NewMI =
    441         BuildMI(*MFI, I, MI.getDebugLoc(), ADDrr, DstR).addReg(DstR).add(Src);
    442     LLVM_DEBUG(NewMI->dump(););
    443   }
    444   // Make ADD instruction for immediate
    445   if (MI.getOperand(4).getImm() != 0) {
    446     const MCInstrDesc &ADDri =
    447         TII->get(getADDriFromLEA(Opcode, MI.getOperand(4)));
    448     const MachineOperand &SrcR = MI.getOperand(SrcR1 == DstR ? 1 : 3);
    449     NewMI = BuildMI(*MFI, I, MI.getDebugLoc(), ADDri, DstR)
    450                 .add(SrcR)
    451                 .addImm(MI.getOperand(4).getImm());
    452     LLVM_DEBUG(NewMI->dump(););
    453   }
    454   if (NewMI) {
    455     MFI->erase(I);
    456     I = NewMI;
    457   }
    458 }
    459 
    460 MachineInstr *
    461 FixupLEAPass::processInstrForSlow3OpLEA(MachineInstr &MI,
    462                                         MachineFunction::iterator MFI) {
    463 
    464   const int LEAOpcode = MI.getOpcode();
    465   if (!isLEA(LEAOpcode))
    466     return nullptr;
    467 
    468   const MachineOperand &Dst = MI.getOperand(0);
    469   const MachineOperand &Base = MI.getOperand(1);
    470   const MachineOperand &Scale = MI.getOperand(2);
    471   const MachineOperand &Index = MI.getOperand(3);
    472   const MachineOperand &Offset = MI.getOperand(4);
    473   const MachineOperand &Segment = MI.getOperand(5);
    474 
    475   if (!(TII->isThreeOperandsLEA(MI) ||
    476         hasInefficientLEABaseReg(Base, Index)) ||
    477       !TII->isSafeToClobberEFLAGS(*MFI, MI) ||
    478       Segment.getReg() != X86::NoRegister)
    479     return nullptr;
    480 
    481   unsigned int DstR = Dst.getReg();
    482   unsigned int BaseR = Base.getReg();
    483   unsigned int IndexR = Index.getReg();
    484   unsigned SSDstR =
    485       (LEAOpcode == X86::LEA64_32r) ? getX86SubSuperRegister(DstR, 64) : DstR;
    486   bool IsScale1 = Scale.getImm() == 1;
    487   bool IsInefficientBase = isInefficientLEAReg(BaseR);
    488   bool IsInefficientIndex = isInefficientLEAReg(IndexR);
    489 
    490   // Skip these cases since it takes more than 2 instructions
    491   // to replace the LEA instruction.
    492   if (IsInefficientBase && SSDstR == BaseR && !IsScale1)
    493     return nullptr;
    494   if (LEAOpcode == X86::LEA64_32r && IsInefficientBase &&
    495       (IsInefficientIndex || !IsScale1))
    496     return nullptr;
    497 
    498   const DebugLoc DL = MI.getDebugLoc();
    499   const MCInstrDesc &ADDrr = TII->get(getADDrrFromLEA(LEAOpcode));
    500   const MCInstrDesc &ADDri = TII->get(getADDriFromLEA(LEAOpcode, Offset));
    501 
    502   LLVM_DEBUG(dbgs() << "FixLEA: Candidate to replace:"; MI.dump(););
    503   LLVM_DEBUG(dbgs() << "FixLEA: Replaced by: ";);
    504 
    505   // First try to replace LEA with one or two (for the 3-op LEA case)
    506   // add instructions:
    507   // 1.lea (%base,%index,1), %base => add %index,%base
    508   // 2.lea (%base,%index,1), %index => add %base,%index
    509   if (IsScale1 && (DstR == BaseR || DstR == IndexR)) {
    510     const MachineOperand &Src = DstR == BaseR ? Index : Base;
    511     MachineInstr *NewMI =
    512         BuildMI(*MFI, MI, DL, ADDrr, DstR).addReg(DstR).add(Src);
    513     LLVM_DEBUG(NewMI->dump(););
    514     // Create ADD instruction for the Offset in case of 3-Ops LEA.
    515     if (hasLEAOffset(Offset)) {
    516       NewMI = BuildMI(*MFI, MI, DL, ADDri, DstR).addReg(DstR).add(Offset);
    517       LLVM_DEBUG(NewMI->dump(););
    518     }
    519     return NewMI;
    520   }
    521   // If the base is inefficient try switching the index and base operands,
    522   // otherwise just break the 3-Ops LEA inst into 2-Ops LEA + ADD instruction:
    523   // lea offset(%base,%index,scale),%dst =>
    524   // lea (%base,%index,scale); add offset,%dst
    525   if (!IsInefficientBase || (!IsInefficientIndex && IsScale1)) {
    526     MachineInstr *NewMI = BuildMI(*MFI, MI, DL, TII->get(LEAOpcode))
    527                               .add(Dst)
    528                               .add(IsInefficientBase ? Index : Base)
    529                               .add(Scale)
    530                               .add(IsInefficientBase ? Base : Index)
    531                               .addImm(0)
    532                               .add(Segment);
    533     LLVM_DEBUG(NewMI->dump(););
    534     // Create ADD instruction for the Offset in case of 3-Ops LEA.
    535     if (hasLEAOffset(Offset)) {
    536       NewMI = BuildMI(*MFI, MI, DL, ADDri, DstR).addReg(DstR).add(Offset);
    537       LLVM_DEBUG(NewMI->dump(););
    538     }
    539     return NewMI;
    540   }
    541   // Handle the rest of the cases with inefficient base register:
    542   assert(SSDstR != BaseR && "SSDstR == BaseR should be handled already!");
    543   assert(IsInefficientBase && "efficient base should be handled already!");
    544 
    545   // lea (%base,%index,1), %dst => mov %base,%dst; add %index,%dst
    546   if (IsScale1 && !hasLEAOffset(Offset)) {
    547     bool BIK = Base.isKill() && BaseR != IndexR;
    548     TII->copyPhysReg(*MFI, MI, DL, DstR, BaseR, BIK);
    549     LLVM_DEBUG(MI.getPrevNode()->dump(););
    550 
    551     MachineInstr *NewMI =
    552         BuildMI(*MFI, MI, DL, ADDrr, DstR).addReg(DstR).add(Index);
    553     LLVM_DEBUG(NewMI->dump(););
    554     return NewMI;
    555   }
    556   // lea offset(%base,%index,scale), %dst =>
    557   // lea offset( ,%index,scale), %dst; add %base,%dst
    558   MachineInstr *NewMI = BuildMI(*MFI, MI, DL, TII->get(LEAOpcode))
    559                             .add(Dst)
    560                             .addReg(0)
    561                             .add(Scale)
    562                             .add(Index)
    563                             .add(Offset)
    564                             .add(Segment);
    565   LLVM_DEBUG(NewMI->dump(););
    566 
    567   NewMI = BuildMI(*MFI, MI, DL, ADDrr, DstR).addReg(DstR).add(Base);
    568   LLVM_DEBUG(NewMI->dump(););
    569   return NewMI;
    570 }
    571 
    572 bool FixupLEAPass::processBasicBlock(MachineFunction &MF,
    573                                      MachineFunction::iterator MFI) {
    574 
    575   for (MachineBasicBlock::iterator I = MFI->begin(); I != MFI->end(); ++I) {
    576     if (OptIncDec)
    577       if (fixupIncDec(I, MFI))
    578         continue;
    579 
    580     if (OptLEA) {
    581       if (MF.getSubtarget<X86Subtarget>().slowLEA())
    582         processInstructionForSLM(I, MFI);
    583 
    584       else {
    585         if (MF.getSubtarget<X86Subtarget>().slow3OpsLEA()) {
    586           if (auto *NewMI = processInstrForSlow3OpLEA(*I, MFI)) {
    587             MFI->erase(I);
    588             I = NewMI;
    589           }
    590         } else
    591           processInstruction(I, MFI);
    592       }
    593     }
    594   }
    595   return false;
    596 }
    597