llvm.org GIT mirror llvm / f726aa1 lib / Target / X86 / X86FixupLEAs.cpp
f726aa1

Tree @f726aa1 (Download .tar.gz)

X86FixupLEAs.cpp @f726aa1raw · history · blame

  1
  2
  3
  4
  5
  6
  7
  8
  9
 10
 11
 12
 13
 14
 15
 16
 17
 18
 19
 20
 21
 22
 23
 24
 25
 26
 27
 28
 29
 30
 31
 32
 33
 34
 35
 36
 37
 38
 39
 40
 41
 42
 43
 44
 45
 46
 47
 48
 49
 50
 51
 52
 53
 54
 55
 56
 57
 58
 59
 60
 61
 62
 63
 64
 65
 66
 67
 68
 69
 70
 71
 72
 73
 74
 75
 76
 77
 78
 79
 80
 81
 82
 83
 84
 85
 86
 87
 88
 89
 90
 91
 92
 93
 94
 95
 96
 97
 98
 99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
549
550
551
552
553
554
555
556
557
558
559
560
561
562
563
564
565
566
567
568
569
570
571
572
573
574
575
576
577
578
579
580
581
582
583
584
585
586
587
588
589
590
591
592
593
594
595
596
597
598
599
600
601
602
603
604
605
606
607
608
609
610
611
612
613
614
615
616
617
618
619
620
621
622
623
624
625
626
627
628
629
630
631
632
633
634
635
636
637
638
639
640
641
642
643
644
645
//===-- X86FixupLEAs.cpp - use or replace LEA instructions -----------===//
//
// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
// See https://llvm.org/LICENSE.txt for license information.
// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
//
//===----------------------------------------------------------------------===//
//
// This file defines the pass that finds instructions that can be
// re-written as LEA instructions in order to reduce pipeline delays.
// It replaces LEAs with ADD/INC/DEC when that is better for size/speed.
//
//===----------------------------------------------------------------------===//

#include "X86.h"
#include "X86InstrInfo.h"
#include "X86Subtarget.h"
#include "llvm/ADT/Statistic.h"
#include "llvm/CodeGen/MachineFunctionPass.h"
#include "llvm/CodeGen/MachineInstrBuilder.h"
#include "llvm/CodeGen/Passes.h"
#include "llvm/CodeGen/TargetSchedule.h"
#include "llvm/Support/Debug.h"
#include "llvm/Support/raw_ostream.h"
using namespace llvm;

#define FIXUPLEA_DESC "X86 LEA Fixup"
#define FIXUPLEA_NAME "x86-fixup-LEAs"

#define DEBUG_TYPE FIXUPLEA_NAME

STATISTIC(NumLEAs, "Number of LEA instructions created");

namespace {
class FixupLEAPass : public MachineFunctionPass {
  enum RegUsageState { RU_NotUsed, RU_Write, RU_Read };

  /// Given a machine register, look for the instruction
  /// which writes it in the current basic block. If found,
  /// try to replace it with an equivalent LEA instruction.
  /// If replacement succeeds, then also process the newly created
  /// instruction.
  void seekLEAFixup(MachineOperand &p, MachineBasicBlock::iterator &I,
                    MachineBasicBlock &MBB);

  /// Given a memory access or LEA instruction
  /// whose address mode uses a base and/or index register, look for
  /// an opportunity to replace the instruction which sets the base or index
  /// register with an equivalent LEA instruction.
  void processInstruction(MachineBasicBlock::iterator &I,
                          MachineBasicBlock &MBB);

  /// Given a LEA instruction which is unprofitable
  /// on SlowLEA targets try to replace it with an equivalent ADD instruction.
  void processInstructionForSlowLEA(MachineBasicBlock::iterator &I,
                                    MachineBasicBlock &MBB);

  /// Given a LEA instruction which is unprofitable
  /// on SNB+ try to replace it with other instructions.
  /// According to Intel's Optimization Reference Manual:
  /// " For LEA instructions with three source operands and some specific
  ///   situations, instruction latency has increased to 3 cycles, and must
  ///   dispatch via port 1:
  /// - LEA that has all three source operands: base, index, and offset
  /// - LEA that uses base and index registers where the base is EBP, RBP,
  ///   or R13
  /// - LEA that uses RIP relative addressing mode
  /// - LEA that uses 16-bit addressing mode "
  /// This function currently handles the first 2 cases only.
  MachineInstr *processInstrForSlow3OpLEA(MachineInstr &MI,
                                          MachineBasicBlock &MBB);

  /// Look for LEAs that are really two address LEAs that we might be able to
  /// turn into regular ADD instructions.
  bool optTwoAddrLEA(MachineBasicBlock::iterator &I,
                     MachineBasicBlock &MBB, bool OptIncDec,
                     bool UseLEAForSP) const;

  /// Determine if an instruction references a machine register
  /// and, if so, whether it reads or writes the register.
  RegUsageState usesRegister(MachineOperand &p, MachineBasicBlock::iterator I);

  /// Step backwards through a basic block, looking
  /// for an instruction which writes a register within
  /// a maximum of INSTR_DISTANCE_THRESHOLD instruction latency cycles.
  MachineBasicBlock::iterator searchBackwards(MachineOperand &p,
                                              MachineBasicBlock::iterator &I,
                                              MachineBasicBlock &MBB);

  /// if an instruction can be converted to an
  /// equivalent LEA, insert the new instruction into the basic block
  /// and return a pointer to it. Otherwise, return zero.
  MachineInstr *postRAConvertToLEA(MachineBasicBlock &MBB,
                                   MachineBasicBlock::iterator &MBBI) const;

public:
  static char ID;

  StringRef getPassName() const override { return FIXUPLEA_DESC; }

  FixupLEAPass() : MachineFunctionPass(ID) { }

  /// Loop over all of the basic blocks,
  /// replacing instructions by equivalent LEA instructions
  /// if needed and when possible.
  bool runOnMachineFunction(MachineFunction &MF) override;

  // This pass runs after regalloc and doesn't support VReg operands.
  MachineFunctionProperties getRequiredProperties() const override {
    return MachineFunctionProperties().set(
        MachineFunctionProperties::Property::NoVRegs);
  }

private:
  TargetSchedModel TSM;
  const X86InstrInfo *TII;
  const X86RegisterInfo *TRI;
};
}

char FixupLEAPass::ID = 0;

INITIALIZE_PASS(FixupLEAPass, FIXUPLEA_NAME, FIXUPLEA_DESC, false, false)

MachineInstr *
FixupLEAPass::postRAConvertToLEA(MachineBasicBlock &MBB,
                                 MachineBasicBlock::iterator &MBBI) const {
  MachineInstr &MI = *MBBI;
  switch (MI.getOpcode()) {
  case X86::MOV32rr:
  case X86::MOV64rr: {
    const MachineOperand &Src = MI.getOperand(1);
    const MachineOperand &Dest = MI.getOperand(0);
    MachineInstr *NewMI =
        BuildMI(MBB, MBBI, MI.getDebugLoc(),
                TII->get(MI.getOpcode() == X86::MOV32rr ? X86::LEA32r
                                                        : X86::LEA64r))
            .add(Dest)
            .add(Src)
            .addImm(1)
            .addReg(0)
            .addImm(0)
            .addReg(0);
    return NewMI;
  }
  }

  if (!MI.isConvertibleTo3Addr())
    return nullptr;

  switch (MI.getOpcode()) {
  default:
    // Only convert instructions that we've verified are safe.
    return nullptr;
  case X86::ADD64ri32:
  case X86::ADD64ri8:
  case X86::ADD64ri32_DB:
  case X86::ADD64ri8_DB:
  case X86::ADD32ri:
  case X86::ADD32ri8:
  case X86::ADD32ri_DB:
  case X86::ADD32ri8_DB:
    if (!MI.getOperand(2).isImm()) {
      // convertToThreeAddress will call getImm()
      // which requires isImm() to be true
      return nullptr;
    }
    break;
  case X86::SHL64ri:
  case X86::SHL32ri:
  case X86::INC64r:
  case X86::INC32r:
  case X86::DEC64r:
  case X86::DEC32r:
  case X86::ADD64rr:
  case X86::ADD64rr_DB:
  case X86::ADD32rr:
  case X86::ADD32rr_DB:
    // These instructions are all fine to convert.
    break;
  }
  MachineFunction::iterator MFI = MBB.getIterator();
  return TII->convertToThreeAddress(MFI, MI, nullptr);
}

FunctionPass *llvm::createX86FixupLEAs() { return new FixupLEAPass(); }

static bool isLEA(unsigned Opcode) {
  return Opcode == X86::LEA32r || Opcode == X86::LEA64r ||
         Opcode == X86::LEA64_32r;
}

bool FixupLEAPass::runOnMachineFunction(MachineFunction &MF) {
  if (skipFunction(MF.getFunction()))
    return false;

  const X86Subtarget &ST = MF.getSubtarget<X86Subtarget>();
  bool IsSlowLEA = ST.slowLEA();
  bool IsSlow3OpsLEA = ST.slow3OpsLEA();
  bool LEAUsesAG = ST.LEAusesAG();

  bool OptIncDec = !ST.slowIncDec() || MF.getFunction().hasOptSize();
  bool UseLEAForSP = ST.useLeaForSP();

  TSM.init(&ST);
  TII = ST.getInstrInfo();
  TRI = ST.getRegisterInfo();

  LLVM_DEBUG(dbgs() << "Start X86FixupLEAs\n";);
  for (MachineBasicBlock &MBB : MF) {
    // First pass. Try to remove or optimize existing LEAs.
    for (MachineBasicBlock::iterator I = MBB.begin(); I != MBB.end(); ++I) {
      if (!isLEA(I->getOpcode()))
        continue;

      if (optTwoAddrLEA(I, MBB, OptIncDec, UseLEAForSP))
        continue;

      if (IsSlowLEA) {
        processInstructionForSlowLEA(I, MBB);
      } else if (IsSlow3OpsLEA) {
        if (auto *NewMI = processInstrForSlow3OpLEA(*I, MBB)) {
          MBB.erase(I);
          I = NewMI;
        }
      }
    }

    // Second pass for creating LEAs. This may reverse some of the
    // transformations above.
    if (LEAUsesAG) {
      for (MachineBasicBlock::iterator I = MBB.begin(); I != MBB.end(); ++I)
        processInstruction(I, MBB);
    }
  }

  LLVM_DEBUG(dbgs() << "End X86FixupLEAs\n";);

  return true;
}

FixupLEAPass::RegUsageState
FixupLEAPass::usesRegister(MachineOperand &p, MachineBasicBlock::iterator I) {
  RegUsageState RegUsage = RU_NotUsed;
  MachineInstr &MI = *I;

  for (unsigned i = 0; i < MI.getNumOperands(); ++i) {
    MachineOperand &opnd = MI.getOperand(i);
    if (opnd.isReg() && opnd.getReg() == p.getReg()) {
      if (opnd.isDef())
        return RU_Write;
      RegUsage = RU_Read;
    }
  }
  return RegUsage;
}

/// getPreviousInstr - Given a reference to an instruction in a basic
/// block, return a reference to the previous instruction in the block,
/// wrapping around to the last instruction of the block if the block
/// branches to itself.
static inline bool getPreviousInstr(MachineBasicBlock::iterator &I,
                                    MachineBasicBlock &MBB) {
  if (I == MBB.begin()) {
    if (MBB.isPredecessor(&MBB)) {
      I = --MBB.end();
      return true;
    } else
      return false;
  }
  --I;
  return true;
}

MachineBasicBlock::iterator
FixupLEAPass::searchBackwards(MachineOperand &p, MachineBasicBlock::iterator &I,
                              MachineBasicBlock &MBB) {
  int InstrDistance = 1;
  MachineBasicBlock::iterator CurInst;
  static const int INSTR_DISTANCE_THRESHOLD = 5;

  CurInst = I;
  bool Found;
  Found = getPreviousInstr(CurInst, MBB);
  while (Found && I != CurInst) {
    if (CurInst->isCall() || CurInst->isInlineAsm())
      break;
    if (InstrDistance > INSTR_DISTANCE_THRESHOLD)
      break; // too far back to make a difference
    if (usesRegister(p, CurInst) == RU_Write) {
      return CurInst;
    }
    InstrDistance += TSM.computeInstrLatency(&*CurInst);
    Found = getPreviousInstr(CurInst, MBB);
  }
  return MachineBasicBlock::iterator();
}

static inline bool isInefficientLEAReg(unsigned Reg) {
  return Reg == X86::EBP || Reg == X86::RBP ||
         Reg == X86::R13D || Reg == X86::R13;
}

static inline bool isRegOperand(const MachineOperand &Op) {
  return Op.isReg() && Op.getReg() != X86::NoRegister;
}

/// Returns true if this LEA uses base an index registers, and the base register
/// is known to be inefficient for the subtarget.
// TODO: use a variant scheduling class to model the latency profile
// of LEA instructions, and implement this logic as a scheduling predicate.
static inline bool hasInefficientLEABaseReg(const MachineOperand &Base,
                                            const MachineOperand &Index) {
  return Base.isReg() && isInefficientLEAReg(Base.getReg()) &&
         isRegOperand(Index);
}

static inline bool hasLEAOffset(const MachineOperand &Offset) {
  return (Offset.isImm() && Offset.getImm() != 0) || Offset.isGlobal();
}

static inline unsigned getADDrrFromLEA(unsigned LEAOpcode) {
  switch (LEAOpcode) {
  default:
    llvm_unreachable("Unexpected LEA instruction");
  case X86::LEA32r:
  case X86::LEA64_32r:
    return X86::ADD32rr;
  case X86::LEA64r:
    return X86::ADD64rr;
  }
}

static inline unsigned getADDriFromLEA(unsigned LEAOpcode,
                                       const MachineOperand &Offset) {
  bool IsInt8 = Offset.isImm() && isInt<8>(Offset.getImm());
  switch (LEAOpcode) {
  default:
    llvm_unreachable("Unexpected LEA instruction");
  case X86::LEA32r:
  case X86::LEA64_32r:
    return IsInt8 ? X86::ADD32ri8 : X86::ADD32ri;
  case X86::LEA64r:
    return IsInt8 ? X86::ADD64ri8 : X86::ADD64ri32;
  }
}

static inline unsigned getINCDECFromLEA(unsigned LEAOpcode, bool IsINC) {
  switch (LEAOpcode) {
  default:
    llvm_unreachable("Unexpected LEA instruction");
  case X86::LEA32r:
  case X86::LEA64_32r:
    return IsINC ? X86::INC32r : X86::DEC32r;
  case X86::LEA64r:
    return IsINC ? X86::INC64r : X86::DEC64r;
  }
}

bool FixupLEAPass::optTwoAddrLEA(MachineBasicBlock::iterator &I,
                                 MachineBasicBlock &MBB, bool OptIncDec,
                                 bool UseLEAForSP) const {
  MachineInstr &MI = *I;

  const MachineOperand &Base =    MI.getOperand(1 + X86::AddrBaseReg);
  const MachineOperand &Scale =   MI.getOperand(1 + X86::AddrScaleAmt);
  const MachineOperand &Index =   MI.getOperand(1 + X86::AddrIndexReg);
  const MachineOperand &Disp =    MI.getOperand(1 + X86::AddrDisp);
  const MachineOperand &Segment = MI.getOperand(1 + X86::AddrSegmentReg);

  if (Segment.getReg() != 0 || !Disp.isImm() || Scale.getImm() > 1 ||
      !TII->isSafeToClobberEFLAGS(MBB, I))
    return false;

  Register DestReg = MI.getOperand(0).getReg();
  Register BaseReg = Base.getReg();
  Register IndexReg = Index.getReg();

  // Don't change stack adjustment LEAs.
  if (UseLEAForSP && (DestReg == X86::ESP || DestReg == X86::RSP))
    return false;

  // LEA64_32 has 64-bit operands but 32-bit result.
  if (MI.getOpcode() == X86::LEA64_32r) {
    if (BaseReg != 0)
      BaseReg = TRI->getSubReg(BaseReg, X86::sub_32bit);
    if (IndexReg != 0)
      IndexReg = TRI->getSubReg(IndexReg, X86::sub_32bit);
  }

  MachineInstr *NewMI = nullptr;

  // Look for lea(%reg1, %reg2), %reg1 or lea(%reg2, %reg1), %reg1
  // which can be turned into add %reg2, %reg1
  if (BaseReg != 0 && IndexReg != 0 && Disp.getImm() == 0 &&
      (DestReg == BaseReg || DestReg == IndexReg)) {
    unsigned NewOpcode = getADDrrFromLEA(MI.getOpcode());
    if (DestReg != BaseReg)
      std::swap(BaseReg, IndexReg);

    if (MI.getOpcode() == X86::LEA64_32r) {
      // TODO: Do we need the super register implicit use?
      NewMI = BuildMI(MBB, I, MI.getDebugLoc(), TII->get(NewOpcode), DestReg)
        .addReg(BaseReg).addReg(IndexReg)
        .addReg(Base.getReg(), RegState::Implicit)
        .addReg(Index.getReg(), RegState::Implicit);
    } else {
      NewMI = BuildMI(MBB, I, MI.getDebugLoc(), TII->get(NewOpcode), DestReg)
        .addReg(BaseReg).addReg(IndexReg);
    }
  } else if (DestReg == BaseReg && IndexReg == 0) {
    // This is an LEA with only a base register and a displacement,
    // We can use ADDri or INC/DEC.

    // Does this LEA have one these forms:
    // lea  %reg, 1(%reg)
    // lea  %reg, -1(%reg)
    if (OptIncDec && (Disp.getImm() == 1 || Disp.getImm() == -1)) {
      bool IsINC = Disp.getImm() == 1;
      unsigned NewOpcode = getINCDECFromLEA(MI.getOpcode(), IsINC);

      if (MI.getOpcode() == X86::LEA64_32r) {
        // TODO: Do we need the super register implicit use?
        NewMI = BuildMI(MBB, I, MI.getDebugLoc(), TII->get(NewOpcode), DestReg)
          .addReg(BaseReg).addReg(Base.getReg(), RegState::Implicit);
      } else {
        NewMI = BuildMI(MBB, I, MI.getDebugLoc(), TII->get(NewOpcode), DestReg)
          .addReg(BaseReg);
      }
    } else {
      unsigned NewOpcode = getADDriFromLEA(MI.getOpcode(), Disp);
      if (MI.getOpcode() == X86::LEA64_32r) {
        // TODO: Do we need the super register implicit use?
        NewMI = BuildMI(MBB, I, MI.getDebugLoc(), TII->get(NewOpcode), DestReg)
          .addReg(BaseReg).addImm(Disp.getImm())
          .addReg(Base.getReg(), RegState::Implicit);
      } else {
        NewMI = BuildMI(MBB, I, MI.getDebugLoc(), TII->get(NewOpcode), DestReg)
          .addReg(BaseReg).addImm(Disp.getImm());
      }
    }
  } else
    return false;

  MBB.erase(I);
  I = NewMI;
  return true;
}

void FixupLEAPass::processInstruction(MachineBasicBlock::iterator &I,
                                      MachineBasicBlock &MBB) {
  // Process a load, store, or LEA instruction.
  MachineInstr &MI = *I;
  const MCInstrDesc &Desc = MI.getDesc();
  int AddrOffset = X86II::getMemoryOperandNo(Desc.TSFlags);
  if (AddrOffset >= 0) {
    AddrOffset += X86II::getOperandBias(Desc);
    MachineOperand &p = MI.getOperand(AddrOffset + X86::AddrBaseReg);
    if (p.isReg() && p.getReg() != X86::ESP) {
      seekLEAFixup(p, I, MBB);
    }
    MachineOperand &q = MI.getOperand(AddrOffset + X86::AddrIndexReg);
    if (q.isReg() && q.getReg() != X86::ESP) {
      seekLEAFixup(q, I, MBB);
    }
  }
}

void FixupLEAPass::seekLEAFixup(MachineOperand &p,
                                MachineBasicBlock::iterator &I,
                                MachineBasicBlock &MBB) {
  MachineBasicBlock::iterator MBI = searchBackwards(p, I, MBB);
  if (MBI != MachineBasicBlock::iterator()) {
    MachineInstr *NewMI = postRAConvertToLEA(MBB, MBI);
    if (NewMI) {
      ++NumLEAs;
      LLVM_DEBUG(dbgs() << "FixLEA: Candidate to replace:"; MBI->dump(););
      // now to replace with an equivalent LEA...
      LLVM_DEBUG(dbgs() << "FixLEA: Replaced by: "; NewMI->dump(););
      MBB.erase(MBI);
      MachineBasicBlock::iterator J =
          static_cast<MachineBasicBlock::iterator>(NewMI);
      processInstruction(J, MBB);
    }
  }
}

void FixupLEAPass::processInstructionForSlowLEA(MachineBasicBlock::iterator &I,
                                                MachineBasicBlock &MBB) {
  MachineInstr &MI = *I;
  const unsigned Opcode = MI.getOpcode();

  const MachineOperand &Dst =     MI.getOperand(0);
  const MachineOperand &Base =    MI.getOperand(1 + X86::AddrBaseReg);
  const MachineOperand &Scale =   MI.getOperand(1 + X86::AddrScaleAmt);
  const MachineOperand &Index =   MI.getOperand(1 + X86::AddrIndexReg);
  const MachineOperand &Offset =  MI.getOperand(1 + X86::AddrDisp);
  const MachineOperand &Segment = MI.getOperand(1 + X86::AddrSegmentReg);

  if (Segment.getReg() != 0 || !Offset.isImm() ||
      !TII->isSafeToClobberEFLAGS(MBB, I))
    return;
  const Register DstR = Dst.getReg();
  const Register SrcR1 = Base.getReg();
  const Register SrcR2 = Index.getReg();
  if ((SrcR1 == 0 || SrcR1 != DstR) && (SrcR2 == 0 || SrcR2 != DstR))
    return;
  if (Scale.getImm() > 1)
    return;
  LLVM_DEBUG(dbgs() << "FixLEA: Candidate to replace:"; I->dump(););
  LLVM_DEBUG(dbgs() << "FixLEA: Replaced by: ";);
  MachineInstr *NewMI = nullptr;
  // Make ADD instruction for two registers writing to LEA's destination
  if (SrcR1 != 0 && SrcR2 != 0) {
    const MCInstrDesc &ADDrr = TII->get(getADDrrFromLEA(Opcode));
    const MachineOperand &Src = SrcR1 == DstR ? Index : Base;
    NewMI =
        BuildMI(MBB, I, MI.getDebugLoc(), ADDrr, DstR).addReg(DstR).add(Src);
    LLVM_DEBUG(NewMI->dump(););
  }
  // Make ADD instruction for immediate
  if (Offset.getImm() != 0) {
    const MCInstrDesc &ADDri =
        TII->get(getADDriFromLEA(Opcode, Offset));
    const MachineOperand &SrcR = SrcR1 == DstR ? Base : Index;
    NewMI = BuildMI(MBB, I, MI.getDebugLoc(), ADDri, DstR)
                .add(SrcR)
                .addImm(Offset.getImm());
    LLVM_DEBUG(NewMI->dump(););
  }
  if (NewMI) {
    MBB.erase(I);
    I = NewMI;
  }
}

MachineInstr *
FixupLEAPass::processInstrForSlow3OpLEA(MachineInstr &MI,
                                        MachineBasicBlock &MBB) {
  const unsigned LEAOpcode = MI.getOpcode();

  const MachineOperand &Dst =     MI.getOperand(0);
  const MachineOperand &Base =    MI.getOperand(1 + X86::AddrBaseReg);
  const MachineOperand &Scale =   MI.getOperand(1 + X86::AddrScaleAmt);
  const MachineOperand &Index =   MI.getOperand(1 + X86::AddrIndexReg);
  const MachineOperand &Offset =  MI.getOperand(1 + X86::AddrDisp);
  const MachineOperand &Segment = MI.getOperand(1 + X86::AddrSegmentReg);

  if (!(TII->isThreeOperandsLEA(MI) ||
        hasInefficientLEABaseReg(Base, Index)) ||
      !TII->isSafeToClobberEFLAGS(MBB, MI) ||
      Segment.getReg() != X86::NoRegister)
    return nullptr;

  Register DstR = Dst.getReg();
  Register BaseR = Base.getReg();
  Register IndexR = Index.getReg();
  Register SSDstR =
      (LEAOpcode == X86::LEA64_32r) ? Register(getX86SubSuperRegister(DstR, 64))
                                    : DstR;
  bool IsScale1 = Scale.getImm() == 1;
  bool IsInefficientBase = isInefficientLEAReg(BaseR);
  bool IsInefficientIndex = isInefficientLEAReg(IndexR);

  // Skip these cases since it takes more than 2 instructions
  // to replace the LEA instruction.
  if (IsInefficientBase && SSDstR == BaseR && !IsScale1)
    return nullptr;
  if (LEAOpcode == X86::LEA64_32r && IsInefficientBase &&
      (IsInefficientIndex || !IsScale1))
    return nullptr;

  const DebugLoc DL = MI.getDebugLoc();
  const MCInstrDesc &ADDrr = TII->get(getADDrrFromLEA(LEAOpcode));
  const MCInstrDesc &ADDri = TII->get(getADDriFromLEA(LEAOpcode, Offset));

  LLVM_DEBUG(dbgs() << "FixLEA: Candidate to replace:"; MI.dump(););
  LLVM_DEBUG(dbgs() << "FixLEA: Replaced by: ";);

  // First try to replace LEA with one or two (for the 3-op LEA case)
  // add instructions:
  // 1.lea (%base,%index,1), %base => add %index,%base
  // 2.lea (%base,%index,1), %index => add %base,%index
  if (IsScale1 && (DstR == BaseR || DstR == IndexR)) {
    const MachineOperand &Src = DstR == BaseR ? Index : Base;
    MachineInstr *NewMI =
        BuildMI(MBB, MI, DL, ADDrr, DstR).addReg(DstR).add(Src);
    LLVM_DEBUG(NewMI->dump(););
    // Create ADD instruction for the Offset in case of 3-Ops LEA.
    if (hasLEAOffset(Offset)) {
      NewMI = BuildMI(MBB, MI, DL, ADDri, DstR).addReg(DstR).add(Offset);
      LLVM_DEBUG(NewMI->dump(););
    }
    return NewMI;
  }
  // If the base is inefficient try switching the index and base operands,
  // otherwise just break the 3-Ops LEA inst into 2-Ops LEA + ADD instruction:
  // lea offset(%base,%index,scale),%dst =>
  // lea (%base,%index,scale); add offset,%dst
  if (!IsInefficientBase || (!IsInefficientIndex && IsScale1)) {
    MachineInstr *NewMI = BuildMI(MBB, MI, DL, TII->get(LEAOpcode))
                              .add(Dst)
                              .add(IsInefficientBase ? Index : Base)
                              .add(Scale)
                              .add(IsInefficientBase ? Base : Index)
                              .addImm(0)
                              .add(Segment);
    LLVM_DEBUG(NewMI->dump(););
    // Create ADD instruction for the Offset in case of 3-Ops LEA.
    if (hasLEAOffset(Offset)) {
      NewMI = BuildMI(MBB, MI, DL, ADDri, DstR).addReg(DstR).add(Offset);
      LLVM_DEBUG(NewMI->dump(););
    }
    return NewMI;
  }
  // Handle the rest of the cases with inefficient base register:
  assert(SSDstR != BaseR && "SSDstR == BaseR should be handled already!");
  assert(IsInefficientBase && "efficient base should be handled already!");

  // lea (%base,%index,1), %dst => mov %base,%dst; add %index,%dst
  if (IsScale1 && !hasLEAOffset(Offset)) {
    bool BIK = Base.isKill() && BaseR != IndexR;
    TII->copyPhysReg(MBB, MI, DL, DstR, BaseR, BIK);
    LLVM_DEBUG(MI.getPrevNode()->dump(););

    MachineInstr *NewMI =
        BuildMI(MBB, MI, DL, ADDrr, DstR).addReg(DstR).add(Index);
    LLVM_DEBUG(NewMI->dump(););
    return NewMI;
  }
  // lea offset(%base,%index,scale), %dst =>
  // lea offset( ,%index,scale), %dst; add %base,%dst
  MachineInstr *NewMI = BuildMI(MBB, MI, DL, TII->get(LEAOpcode))
                            .add(Dst)
                            .addReg(0)
                            .add(Scale)
                            .add(Index)
                            .add(Offset)
                            .add(Segment);
  LLVM_DEBUG(NewMI->dump(););

  NewMI = BuildMI(MBB, MI, DL, ADDrr, DstR).addReg(DstR).add(Base);
  LLVM_DEBUG(NewMI->dump(););
  return NewMI;
}