llvm.org GIT mirror llvm / stable lib / Target / X86 / X86AvoidStoreForwardingBlocks.cpp
stable

Tree @stable (Download .tar.gz)

X86AvoidStoreForwardingBlocks.cpp @stableraw · 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
646
647
648
649
650
651
652
653
654
655
656
657
658
659
660
661
662
663
664
665
666
667
668
669
670
671
672
673
674
675
676
677
678
679
680
681
682
683
684
685
686
687
688
689
690
691
692
693
694
695
696
697
698
699
700
701
702
703
704
705
706
707
708
709
710
711
712
713
714
715
716
717
718
719
720
721
722
723
724
725
726
727
728
//===- X86AvoidStoreForwardingBlockis.cpp - Avoid HW Store Forward Block --===//
//
// 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
//
//===----------------------------------------------------------------------===//
//
// If a load follows a store and reloads data that the store has written to
// memory, Intel microarchitectures can in many cases forward the data directly
// from the store to the load, This "store forwarding" saves cycles by enabling
// the load to directly obtain the data instead of accessing the data from
// cache or memory.
// A "store forward block" occurs in cases that a store cannot be forwarded to
// the load. The most typical case of store forward block on Intel Core
// microarchitecture that a small store cannot be forwarded to a large load.
// The estimated penalty for a store forward block is ~13 cycles.
//
// This pass tries to recognize and handle cases where "store forward block"
// is created by the compiler when lowering memcpy calls to a sequence
// of a load and a store.
//
// The pass currently only handles cases where memcpy is lowered to
// XMM/YMM registers, it tries to break the memcpy into smaller copies.
// breaking the memcpy should be possible since there is no atomicity
// guarantee for loads and stores to XMM/YMM.
//
// It could be better for performance to solve the problem by loading
// to XMM/YMM then inserting the partial store before storing back from XMM/YMM
// to memory, but this will result in a more conservative optimization since it
// requires we prove that all memory accesses between the blocking store and the
// load must alias/don't alias before we can move the store, whereas the
// transformation done here is correct regardless to other memory accesses.
//===----------------------------------------------------------------------===//

#include "X86InstrInfo.h"
#include "X86Subtarget.h"
#include "llvm/CodeGen/MachineBasicBlock.h"
#include "llvm/CodeGen/MachineFunction.h"
#include "llvm/CodeGen/MachineFunctionPass.h"
#include "llvm/CodeGen/MachineInstr.h"
#include "llvm/CodeGen/MachineInstrBuilder.h"
#include "llvm/CodeGen/MachineOperand.h"
#include "llvm/CodeGen/MachineRegisterInfo.h"
#include "llvm/IR/DebugInfoMetadata.h"
#include "llvm/IR/DebugLoc.h"
#include "llvm/IR/Function.h"
#include "llvm/MC/MCInstrDesc.h"

using namespace llvm;

#define DEBUG_TYPE "x86-avoid-SFB"

static cl::opt<bool> DisableX86AvoidStoreForwardBlocks(
    "x86-disable-avoid-SFB", cl::Hidden,
    cl::desc("X86: Disable Store Forwarding Blocks fixup."), cl::init(false));

static cl::opt<unsigned> X86AvoidSFBInspectionLimit(
    "x86-sfb-inspection-limit",
    cl::desc("X86: Number of instructions backward to "
             "inspect for store forwarding blocks."),
    cl::init(20), cl::Hidden);

namespace {

using DisplacementSizeMap = std::map<int64_t, unsigned>;

class X86AvoidSFBPass : public MachineFunctionPass {
public:
  static char ID;
  X86AvoidSFBPass() : MachineFunctionPass(ID) {
    initializeX86AvoidSFBPassPass(*PassRegistry::getPassRegistry());
  }

  StringRef getPassName() const override {
    return "X86 Avoid Store Forwarding Blocks";
  }

  bool runOnMachineFunction(MachineFunction &MF) override;

  void getAnalysisUsage(AnalysisUsage &AU) const override {
    MachineFunctionPass::getAnalysisUsage(AU);
    AU.addRequired<AAResultsWrapperPass>();
  }

private:
  MachineRegisterInfo *MRI;
  const X86InstrInfo *TII;
  const X86RegisterInfo *TRI;
  SmallVector<std::pair<MachineInstr *, MachineInstr *>, 2>
      BlockedLoadsStoresPairs;
  SmallVector<MachineInstr *, 2> ForRemoval;
  AliasAnalysis *AA;

  /// Returns couples of Load then Store to memory which look
  ///  like a memcpy.
  void findPotentiallylBlockedCopies(MachineFunction &MF);
  /// Break the memcpy's load and store into smaller copies
  /// such that each memory load that was blocked by a smaller store
  /// would now be copied separately.
  void breakBlockedCopies(MachineInstr *LoadInst, MachineInstr *StoreInst,
                          const DisplacementSizeMap &BlockingStoresDispSizeMap);
  /// Break a copy of size Size to smaller copies.
  void buildCopies(int Size, MachineInstr *LoadInst, int64_t LdDispImm,
                   MachineInstr *StoreInst, int64_t StDispImm,
                   int64_t LMMOffset, int64_t SMMOffset);

  void buildCopy(MachineInstr *LoadInst, unsigned NLoadOpcode, int64_t LoadDisp,
                 MachineInstr *StoreInst, unsigned NStoreOpcode,
                 int64_t StoreDisp, unsigned Size, int64_t LMMOffset,
                 int64_t SMMOffset);

  bool alias(const MachineMemOperand &Op1, const MachineMemOperand &Op2) const;

  unsigned getRegSizeInBytes(MachineInstr *Inst);
};

} // end anonymous namespace

char X86AvoidSFBPass::ID = 0;

INITIALIZE_PASS_BEGIN(X86AvoidSFBPass, DEBUG_TYPE, "Machine code sinking",
                      false, false)
INITIALIZE_PASS_DEPENDENCY(AAResultsWrapperPass)
INITIALIZE_PASS_END(X86AvoidSFBPass, DEBUG_TYPE, "Machine code sinking", false,
                    false)

FunctionPass *llvm::createX86AvoidStoreForwardingBlocks() {
  return new X86AvoidSFBPass();
}

static bool isXMMLoadOpcode(unsigned Opcode) {
  return Opcode == X86::MOVUPSrm || Opcode == X86::MOVAPSrm ||
         Opcode == X86::VMOVUPSrm || Opcode == X86::VMOVAPSrm ||
         Opcode == X86::VMOVUPDrm || Opcode == X86::VMOVAPDrm ||
         Opcode == X86::VMOVDQUrm || Opcode == X86::VMOVDQArm ||
         Opcode == X86::VMOVUPSZ128rm || Opcode == X86::VMOVAPSZ128rm ||
         Opcode == X86::VMOVUPDZ128rm || Opcode == X86::VMOVAPDZ128rm ||
         Opcode == X86::VMOVDQU64Z128rm || Opcode == X86::VMOVDQA64Z128rm ||
         Opcode == X86::VMOVDQU32Z128rm || Opcode == X86::VMOVDQA32Z128rm;
}
static bool isYMMLoadOpcode(unsigned Opcode) {
  return Opcode == X86::VMOVUPSYrm || Opcode == X86::VMOVAPSYrm ||
         Opcode == X86::VMOVUPDYrm || Opcode == X86::VMOVAPDYrm ||
         Opcode == X86::VMOVDQUYrm || Opcode == X86::VMOVDQAYrm ||
         Opcode == X86::VMOVUPSZ256rm || Opcode == X86::VMOVAPSZ256rm ||
         Opcode == X86::VMOVUPDZ256rm || Opcode == X86::VMOVAPDZ256rm ||
         Opcode == X86::VMOVDQU64Z256rm || Opcode == X86::VMOVDQA64Z256rm ||
         Opcode == X86::VMOVDQU32Z256rm || Opcode == X86::VMOVDQA32Z256rm;
}

static bool isPotentialBlockedMemCpyLd(unsigned Opcode) {
  return isXMMLoadOpcode(Opcode) || isYMMLoadOpcode(Opcode);
}

static bool isPotentialBlockedMemCpyPair(int LdOpcode, int StOpcode) {
  switch (LdOpcode) {
  case X86::MOVUPSrm:
  case X86::MOVAPSrm:
    return StOpcode == X86::MOVUPSmr || StOpcode == X86::MOVAPSmr;
  case X86::VMOVUPSrm:
  case X86::VMOVAPSrm:
    return StOpcode == X86::VMOVUPSmr || StOpcode == X86::VMOVAPSmr;
  case X86::VMOVUPDrm:
  case X86::VMOVAPDrm:
    return StOpcode == X86::VMOVUPDmr || StOpcode == X86::VMOVAPDmr;
  case X86::VMOVDQUrm:
  case X86::VMOVDQArm:
    return StOpcode == X86::VMOVDQUmr || StOpcode == X86::VMOVDQAmr;
  case X86::VMOVUPSZ128rm:
  case X86::VMOVAPSZ128rm:
    return StOpcode == X86::VMOVUPSZ128mr || StOpcode == X86::VMOVAPSZ128mr;
  case X86::VMOVUPDZ128rm:
  case X86::VMOVAPDZ128rm:
    return StOpcode == X86::VMOVUPDZ128mr || StOpcode == X86::VMOVAPDZ128mr;
  case X86::VMOVUPSYrm:
  case X86::VMOVAPSYrm:
    return StOpcode == X86::VMOVUPSYmr || StOpcode == X86::VMOVAPSYmr;
  case X86::VMOVUPDYrm:
  case X86::VMOVAPDYrm:
    return StOpcode == X86::VMOVUPDYmr || StOpcode == X86::VMOVAPDYmr;
  case X86::VMOVDQUYrm:
  case X86::VMOVDQAYrm:
    return StOpcode == X86::VMOVDQUYmr || StOpcode == X86::VMOVDQAYmr;
  case X86::VMOVUPSZ256rm:
  case X86::VMOVAPSZ256rm:
    return StOpcode == X86::VMOVUPSZ256mr || StOpcode == X86::VMOVAPSZ256mr;
  case X86::VMOVUPDZ256rm:
  case X86::VMOVAPDZ256rm:
    return StOpcode == X86::VMOVUPDZ256mr || StOpcode == X86::VMOVAPDZ256mr;
  case X86::VMOVDQU64Z128rm:
  case X86::VMOVDQA64Z128rm:
    return StOpcode == X86::VMOVDQU64Z128mr || StOpcode == X86::VMOVDQA64Z128mr;
  case X86::VMOVDQU32Z128rm:
  case X86::VMOVDQA32Z128rm:
    return StOpcode == X86::VMOVDQU32Z128mr || StOpcode == X86::VMOVDQA32Z128mr;
  case X86::VMOVDQU64Z256rm:
  case X86::VMOVDQA64Z256rm:
    return StOpcode == X86::VMOVDQU64Z256mr || StOpcode == X86::VMOVDQA64Z256mr;
  case X86::VMOVDQU32Z256rm:
  case X86::VMOVDQA32Z256rm:
    return StOpcode == X86::VMOVDQU32Z256mr || StOpcode == X86::VMOVDQA32Z256mr;
  default:
    return false;
  }
}

static bool isPotentialBlockingStoreInst(int Opcode, int LoadOpcode) {
  bool PBlock = false;
  PBlock |= Opcode == X86::MOV64mr || Opcode == X86::MOV64mi32 ||
            Opcode == X86::MOV32mr || Opcode == X86::MOV32mi ||
            Opcode == X86::MOV16mr || Opcode == X86::MOV16mi ||
            Opcode == X86::MOV8mr || Opcode == X86::MOV8mi;
  if (isYMMLoadOpcode(LoadOpcode))
    PBlock |= Opcode == X86::VMOVUPSmr || Opcode == X86::VMOVAPSmr ||
              Opcode == X86::VMOVUPDmr || Opcode == X86::VMOVAPDmr ||
              Opcode == X86::VMOVDQUmr || Opcode == X86::VMOVDQAmr ||
              Opcode == X86::VMOVUPSZ128mr || Opcode == X86::VMOVAPSZ128mr ||
              Opcode == X86::VMOVUPDZ128mr || Opcode == X86::VMOVAPDZ128mr ||
              Opcode == X86::VMOVDQU64Z128mr ||
              Opcode == X86::VMOVDQA64Z128mr ||
              Opcode == X86::VMOVDQU32Z128mr || Opcode == X86::VMOVDQA32Z128mr;
  return PBlock;
}

static const int MOV128SZ = 16;
static const int MOV64SZ = 8;
static const int MOV32SZ = 4;
static const int MOV16SZ = 2;
static const int MOV8SZ = 1;

static unsigned getYMMtoXMMLoadOpcode(unsigned LoadOpcode) {
  switch (LoadOpcode) {
  case X86::VMOVUPSYrm:
  case X86::VMOVAPSYrm:
    return X86::VMOVUPSrm;
  case X86::VMOVUPDYrm:
  case X86::VMOVAPDYrm:
    return X86::VMOVUPDrm;
  case X86::VMOVDQUYrm:
  case X86::VMOVDQAYrm:
    return X86::VMOVDQUrm;
  case X86::VMOVUPSZ256rm:
  case X86::VMOVAPSZ256rm:
    return X86::VMOVUPSZ128rm;
  case X86::VMOVUPDZ256rm:
  case X86::VMOVAPDZ256rm:
    return X86::VMOVUPDZ128rm;
  case X86::VMOVDQU64Z256rm:
  case X86::VMOVDQA64Z256rm:
    return X86::VMOVDQU64Z128rm;
  case X86::VMOVDQU32Z256rm:
  case X86::VMOVDQA32Z256rm:
    return X86::VMOVDQU32Z128rm;
  default:
    llvm_unreachable("Unexpected Load Instruction Opcode");
  }
  return 0;
}

static unsigned getYMMtoXMMStoreOpcode(unsigned StoreOpcode) {
  switch (StoreOpcode) {
  case X86::VMOVUPSYmr:
  case X86::VMOVAPSYmr:
    return X86::VMOVUPSmr;
  case X86::VMOVUPDYmr:
  case X86::VMOVAPDYmr:
    return X86::VMOVUPDmr;
  case X86::VMOVDQUYmr:
  case X86::VMOVDQAYmr:
    return X86::VMOVDQUmr;
  case X86::VMOVUPSZ256mr:
  case X86::VMOVAPSZ256mr:
    return X86::VMOVUPSZ128mr;
  case X86::VMOVUPDZ256mr:
  case X86::VMOVAPDZ256mr:
    return X86::VMOVUPDZ128mr;
  case X86::VMOVDQU64Z256mr:
  case X86::VMOVDQA64Z256mr:
    return X86::VMOVDQU64Z128mr;
  case X86::VMOVDQU32Z256mr:
  case X86::VMOVDQA32Z256mr:
    return X86::VMOVDQU32Z128mr;
  default:
    llvm_unreachable("Unexpected Load Instruction Opcode");
  }
  return 0;
}

static int getAddrOffset(MachineInstr *MI) {
  const MCInstrDesc &Descl = MI->getDesc();
  int AddrOffset = X86II::getMemoryOperandNo(Descl.TSFlags);
  assert(AddrOffset != -1 && "Expected Memory Operand");
  AddrOffset += X86II::getOperandBias(Descl);
  return AddrOffset;
}

static MachineOperand &getBaseOperand(MachineInstr *MI) {
  int AddrOffset = getAddrOffset(MI);
  return MI->getOperand(AddrOffset + X86::AddrBaseReg);
}

static MachineOperand &getDispOperand(MachineInstr *MI) {
  int AddrOffset = getAddrOffset(MI);
  return MI->getOperand(AddrOffset + X86::AddrDisp);
}

// Relevant addressing modes contain only base register and immediate
// displacement or frameindex and immediate displacement.
// TODO: Consider expanding to other addressing modes in the future
static bool isRelevantAddressingMode(MachineInstr *MI) {
  int AddrOffset = getAddrOffset(MI);
  MachineOperand &Base = getBaseOperand(MI);
  MachineOperand &Disp = getDispOperand(MI);
  MachineOperand &Scale = MI->getOperand(AddrOffset + X86::AddrScaleAmt);
  MachineOperand &Index = MI->getOperand(AddrOffset + X86::AddrIndexReg);
  MachineOperand &Segment = MI->getOperand(AddrOffset + X86::AddrSegmentReg);

  if (!((Base.isReg() && Base.getReg() != X86::NoRegister) || Base.isFI()))
    return false;
  if (!Disp.isImm())
    return false;
  if (Scale.getImm() != 1)
    return false;
  if (!(Index.isReg() && Index.getReg() == X86::NoRegister))
    return false;
  if (!(Segment.isReg() && Segment.getReg() == X86::NoRegister))
    return false;
  return true;
}

// Collect potentially blocking stores.
// Limit the number of instructions backwards we want to inspect
// since the effect of store block won't be visible if the store
// and load instructions have enough instructions in between to
// keep the core busy.
static SmallVector<MachineInstr *, 2>
findPotentialBlockers(MachineInstr *LoadInst) {
  SmallVector<MachineInstr *, 2> PotentialBlockers;
  unsigned BlockCount = 0;
  const unsigned InspectionLimit = X86AvoidSFBInspectionLimit;
  for (auto PBInst = std::next(MachineBasicBlock::reverse_iterator(LoadInst)),
            E = LoadInst->getParent()->rend();
       PBInst != E; ++PBInst) {
    BlockCount++;
    if (BlockCount >= InspectionLimit)
      break;
    MachineInstr &MI = *PBInst;
    if (MI.getDesc().isCall())
      return PotentialBlockers;
    PotentialBlockers.push_back(&MI);
  }
  // If we didn't get to the instructions limit try predecessing blocks.
  // Ideally we should traverse the predecessor blocks in depth with some
  // coloring algorithm, but for now let's just look at the first order
  // predecessors.
  if (BlockCount < InspectionLimit) {
    MachineBasicBlock *MBB = LoadInst->getParent();
    int LimitLeft = InspectionLimit - BlockCount;
    for (MachineBasicBlock::pred_iterator PB = MBB->pred_begin(),
                                          PE = MBB->pred_end();
         PB != PE; ++PB) {
      MachineBasicBlock *PMBB = *PB;
      int PredCount = 0;
      for (MachineBasicBlock::reverse_iterator PBInst = PMBB->rbegin(),
                                               PME = PMBB->rend();
           PBInst != PME; ++PBInst) {
        PredCount++;
        if (PredCount >= LimitLeft)
          break;
        if (PBInst->getDesc().isCall())
          break;
        PotentialBlockers.push_back(&*PBInst);
      }
    }
  }
  return PotentialBlockers;
}

void X86AvoidSFBPass::buildCopy(MachineInstr *LoadInst, unsigned NLoadOpcode,
                                int64_t LoadDisp, MachineInstr *StoreInst,
                                unsigned NStoreOpcode, int64_t StoreDisp,
                                unsigned Size, int64_t LMMOffset,
                                int64_t SMMOffset) {
  MachineOperand &LoadBase = getBaseOperand(LoadInst);
  MachineOperand &StoreBase = getBaseOperand(StoreInst);
  MachineBasicBlock *MBB = LoadInst->getParent();
  MachineMemOperand *LMMO = *LoadInst->memoperands_begin();
  MachineMemOperand *SMMO = *StoreInst->memoperands_begin();

  unsigned Reg1 = MRI->createVirtualRegister(
      TII->getRegClass(TII->get(NLoadOpcode), 0, TRI, *(MBB->getParent())));
  MachineInstr *NewLoad =
      BuildMI(*MBB, LoadInst, LoadInst->getDebugLoc(), TII->get(NLoadOpcode),
              Reg1)
          .add(LoadBase)
          .addImm(1)
          .addReg(X86::NoRegister)
          .addImm(LoadDisp)
          .addReg(X86::NoRegister)
          .addMemOperand(
              MBB->getParent()->getMachineMemOperand(LMMO, LMMOffset, Size));
  if (LoadBase.isReg())
    getBaseOperand(NewLoad).setIsKill(false);
  LLVM_DEBUG(NewLoad->dump());
  // If the load and store are consecutive, use the loadInst location to
  // reduce register pressure.
  MachineInstr *StInst = StoreInst;
  if (StoreInst->getPrevNode() == LoadInst)
    StInst = LoadInst;
  MachineInstr *NewStore =
      BuildMI(*MBB, StInst, StInst->getDebugLoc(), TII->get(NStoreOpcode))
          .add(StoreBase)
          .addImm(1)
          .addReg(X86::NoRegister)
          .addImm(StoreDisp)
          .addReg(X86::NoRegister)
          .addReg(Reg1)
          .addMemOperand(
              MBB->getParent()->getMachineMemOperand(SMMO, SMMOffset, Size));
  if (StoreBase.isReg())
    getBaseOperand(NewStore).setIsKill(false);
  MachineOperand &StoreSrcVReg = StoreInst->getOperand(X86::AddrNumOperands);
  assert(StoreSrcVReg.isReg() && "Expected virtual register");
  NewStore->getOperand(X86::AddrNumOperands).setIsKill(StoreSrcVReg.isKill());
  LLVM_DEBUG(NewStore->dump());
}

void X86AvoidSFBPass::buildCopies(int Size, MachineInstr *LoadInst,
                                  int64_t LdDispImm, MachineInstr *StoreInst,
                                  int64_t StDispImm, int64_t LMMOffset,
                                  int64_t SMMOffset) {
  int LdDisp = LdDispImm;
  int StDisp = StDispImm;
  while (Size > 0) {
    if ((Size - MOV128SZ >= 0) && isYMMLoadOpcode(LoadInst->getOpcode())) {
      Size = Size - MOV128SZ;
      buildCopy(LoadInst, getYMMtoXMMLoadOpcode(LoadInst->getOpcode()), LdDisp,
                StoreInst, getYMMtoXMMStoreOpcode(StoreInst->getOpcode()),
                StDisp, MOV128SZ, LMMOffset, SMMOffset);
      LdDisp += MOV128SZ;
      StDisp += MOV128SZ;
      LMMOffset += MOV128SZ;
      SMMOffset += MOV128SZ;
      continue;
    }
    if (Size - MOV64SZ >= 0) {
      Size = Size - MOV64SZ;
      buildCopy(LoadInst, X86::MOV64rm, LdDisp, StoreInst, X86::MOV64mr, StDisp,
                MOV64SZ, LMMOffset, SMMOffset);
      LdDisp += MOV64SZ;
      StDisp += MOV64SZ;
      LMMOffset += MOV64SZ;
      SMMOffset += MOV64SZ;
      continue;
    }
    if (Size - MOV32SZ >= 0) {
      Size = Size - MOV32SZ;
      buildCopy(LoadInst, X86::MOV32rm, LdDisp, StoreInst, X86::MOV32mr, StDisp,
                MOV32SZ, LMMOffset, SMMOffset);
      LdDisp += MOV32SZ;
      StDisp += MOV32SZ;
      LMMOffset += MOV32SZ;
      SMMOffset += MOV32SZ;
      continue;
    }
    if (Size - MOV16SZ >= 0) {
      Size = Size - MOV16SZ;
      buildCopy(LoadInst, X86::MOV16rm, LdDisp, StoreInst, X86::MOV16mr, StDisp,
                MOV16SZ, LMMOffset, SMMOffset);
      LdDisp += MOV16SZ;
      StDisp += MOV16SZ;
      LMMOffset += MOV16SZ;
      SMMOffset += MOV16SZ;
      continue;
    }
    if (Size - MOV8SZ >= 0) {
      Size = Size - MOV8SZ;
      buildCopy(LoadInst, X86::MOV8rm, LdDisp, StoreInst, X86::MOV8mr, StDisp,
                MOV8SZ, LMMOffset, SMMOffset);
      LdDisp += MOV8SZ;
      StDisp += MOV8SZ;
      LMMOffset += MOV8SZ;
      SMMOffset += MOV8SZ;
      continue;
    }
  }
  assert(Size == 0 && "Wrong size division");
}

static void updateKillStatus(MachineInstr *LoadInst, MachineInstr *StoreInst) {
  MachineOperand &LoadBase = getBaseOperand(LoadInst);
  MachineOperand &StoreBase = getBaseOperand(StoreInst);
  if (LoadBase.isReg()) {
    MachineInstr *LastLoad = LoadInst->getPrevNode();
    // If the original load and store to xmm/ymm were consecutive
    // then the partial copies were also created in
    // a consecutive order to reduce register pressure,
    // and the location of the last load is before the last store.
    if (StoreInst->getPrevNode() == LoadInst)
      LastLoad = LoadInst->getPrevNode()->getPrevNode();
    getBaseOperand(LastLoad).setIsKill(LoadBase.isKill());
  }
  if (StoreBase.isReg()) {
    MachineInstr *StInst = StoreInst;
    if (StoreInst->getPrevNode() == LoadInst)
      StInst = LoadInst;
    getBaseOperand(StInst->getPrevNode()).setIsKill(StoreBase.isKill());
  }
}

bool X86AvoidSFBPass::alias(const MachineMemOperand &Op1,
                            const MachineMemOperand &Op2) const {
  if (!Op1.getValue() || !Op2.getValue())
    return true;

  int64_t MinOffset = std::min(Op1.getOffset(), Op2.getOffset());
  int64_t Overlapa = Op1.getSize() + Op1.getOffset() - MinOffset;
  int64_t Overlapb = Op2.getSize() + Op2.getOffset() - MinOffset;

  AliasResult AAResult =
      AA->alias(MemoryLocation(Op1.getValue(), Overlapa, Op1.getAAInfo()),
                MemoryLocation(Op2.getValue(), Overlapb, Op2.getAAInfo()));
  return AAResult != NoAlias;
}

void X86AvoidSFBPass::findPotentiallylBlockedCopies(MachineFunction &MF) {
  for (auto &MBB : MF)
    for (auto &MI : MBB) {
      if (!isPotentialBlockedMemCpyLd(MI.getOpcode()))
        continue;
      int DefVR = MI.getOperand(0).getReg();
      if (!MRI->hasOneUse(DefVR))
        continue;
      for (auto UI = MRI->use_nodbg_begin(DefVR), UE = MRI->use_nodbg_end();
           UI != UE;) {
        MachineOperand &StoreMO = *UI++;
        MachineInstr &StoreMI = *StoreMO.getParent();
        // Skip cases where the memcpy may overlap.
        if (StoreMI.getParent() == MI.getParent() &&
            isPotentialBlockedMemCpyPair(MI.getOpcode(), StoreMI.getOpcode()) &&
            isRelevantAddressingMode(&MI) &&
            isRelevantAddressingMode(&StoreMI)) {
          assert(MI.hasOneMemOperand() &&
                 "Expected one memory operand for load instruction");
          assert(StoreMI.hasOneMemOperand() &&
                 "Expected one memory operand for store instruction");
          if (!alias(**MI.memoperands_begin(), **StoreMI.memoperands_begin()))
            BlockedLoadsStoresPairs.push_back(std::make_pair(&MI, &StoreMI));
        }
      }
    }
}

unsigned X86AvoidSFBPass::getRegSizeInBytes(MachineInstr *LoadInst) {
  auto TRC = TII->getRegClass(TII->get(LoadInst->getOpcode()), 0, TRI,
                              *LoadInst->getParent()->getParent());
  return TRI->getRegSizeInBits(*TRC) / 8;
}

void X86AvoidSFBPass::breakBlockedCopies(
    MachineInstr *LoadInst, MachineInstr *StoreInst,
    const DisplacementSizeMap &BlockingStoresDispSizeMap) {
  int64_t LdDispImm = getDispOperand(LoadInst).getImm();
  int64_t StDispImm = getDispOperand(StoreInst).getImm();
  int64_t LMMOffset = 0;
  int64_t SMMOffset = 0;

  int64_t LdDisp1 = LdDispImm;
  int64_t LdDisp2 = 0;
  int64_t StDisp1 = StDispImm;
  int64_t StDisp2 = 0;
  unsigned Size1 = 0;
  unsigned Size2 = 0;
  int64_t LdStDelta = StDispImm - LdDispImm;

  for (auto DispSizePair : BlockingStoresDispSizeMap) {
    LdDisp2 = DispSizePair.first;
    StDisp2 = DispSizePair.first + LdStDelta;
    Size2 = DispSizePair.second;
    // Avoid copying overlapping areas.
    if (LdDisp2 < LdDisp1) {
      int OverlapDelta = LdDisp1 - LdDisp2;
      LdDisp2 += OverlapDelta;
      StDisp2 += OverlapDelta;
      Size2 -= OverlapDelta;
    }
    Size1 = LdDisp2 - LdDisp1;

    // Build a copy for the point until the current blocking store's
    // displacement.
    buildCopies(Size1, LoadInst, LdDisp1, StoreInst, StDisp1, LMMOffset,
                SMMOffset);
    // Build a copy for the current blocking store.
    buildCopies(Size2, LoadInst, LdDisp2, StoreInst, StDisp2, LMMOffset + Size1,
                SMMOffset + Size1);
    LdDisp1 = LdDisp2 + Size2;
    StDisp1 = StDisp2 + Size2;
    LMMOffset += Size1 + Size2;
    SMMOffset += Size1 + Size2;
  }
  unsigned Size3 = (LdDispImm + getRegSizeInBytes(LoadInst)) - LdDisp1;
  buildCopies(Size3, LoadInst, LdDisp1, StoreInst, StDisp1, LMMOffset,
              LMMOffset);
}

static bool hasSameBaseOpValue(MachineInstr *LoadInst,
                               MachineInstr *StoreInst) {
  MachineOperand &LoadBase = getBaseOperand(LoadInst);
  MachineOperand &StoreBase = getBaseOperand(StoreInst);
  if (LoadBase.isReg() != StoreBase.isReg())
    return false;
  if (LoadBase.isReg())
    return LoadBase.getReg() == StoreBase.getReg();
  return LoadBase.getIndex() == StoreBase.getIndex();
}

static bool isBlockingStore(int64_t LoadDispImm, unsigned LoadSize,
                            int64_t StoreDispImm, unsigned StoreSize) {
  return ((StoreDispImm >= LoadDispImm) &&
          (StoreDispImm <= LoadDispImm + (LoadSize - StoreSize)));
}

// Keep track of all stores blocking a load
static void
updateBlockingStoresDispSizeMap(DisplacementSizeMap &BlockingStoresDispSizeMap,
                                int64_t DispImm, unsigned Size) {
  if (BlockingStoresDispSizeMap.count(DispImm)) {
    // Choose the smallest blocking store starting at this displacement.
    if (BlockingStoresDispSizeMap[DispImm] > Size)
      BlockingStoresDispSizeMap[DispImm] = Size;

  } else
    BlockingStoresDispSizeMap[DispImm] = Size;
}

// Remove blocking stores contained in each other.
static void
removeRedundantBlockingStores(DisplacementSizeMap &BlockingStoresDispSizeMap) {
  if (BlockingStoresDispSizeMap.size() <= 1)
    return;

  SmallVector<std::pair<int64_t, unsigned>, 0> DispSizeStack;
  for (auto DispSizePair : BlockingStoresDispSizeMap) {
    int64_t CurrDisp = DispSizePair.first;
    unsigned CurrSize = DispSizePair.second;
    while (DispSizeStack.size()) {
      int64_t PrevDisp = DispSizeStack.back().first;
      unsigned PrevSize = DispSizeStack.back().second;
      if (CurrDisp + CurrSize > PrevDisp + PrevSize)
        break;
      DispSizeStack.pop_back();
    }
    DispSizeStack.push_back(DispSizePair);
  }
  BlockingStoresDispSizeMap.clear();
  for (auto Disp : DispSizeStack)
    BlockingStoresDispSizeMap.insert(Disp);
}

bool X86AvoidSFBPass::runOnMachineFunction(MachineFunction &MF) {
  bool Changed = false;

  if (DisableX86AvoidStoreForwardBlocks || skipFunction(MF.getFunction()) ||
      !MF.getSubtarget<X86Subtarget>().is64Bit())
    return false;

  MRI = &MF.getRegInfo();
  assert(MRI->isSSA() && "Expected MIR to be in SSA form");
  TII = MF.getSubtarget<X86Subtarget>().getInstrInfo();
  TRI = MF.getSubtarget<X86Subtarget>().getRegisterInfo();
  AA = &getAnalysis<AAResultsWrapperPass>().getAAResults();
  LLVM_DEBUG(dbgs() << "Start X86AvoidStoreForwardBlocks\n";);
  // Look for a load then a store to XMM/YMM which look like a memcpy
  findPotentiallylBlockedCopies(MF);

  for (auto LoadStoreInstPair : BlockedLoadsStoresPairs) {
    MachineInstr *LoadInst = LoadStoreInstPair.first;
    int64_t LdDispImm = getDispOperand(LoadInst).getImm();
    DisplacementSizeMap BlockingStoresDispSizeMap;

    SmallVector<MachineInstr *, 2> PotentialBlockers =
        findPotentialBlockers(LoadInst);
    for (auto PBInst : PotentialBlockers) {
      if (!isPotentialBlockingStoreInst(PBInst->getOpcode(),
                                        LoadInst->getOpcode()) ||
          !isRelevantAddressingMode(PBInst))
        continue;
      int64_t PBstDispImm = getDispOperand(PBInst).getImm();
      assert(PBInst->hasOneMemOperand() && "Expected One Memory Operand");
      unsigned PBstSize = (*PBInst->memoperands_begin())->getSize();
      // This check doesn't cover all cases, but it will suffice for now.
      // TODO: take branch probability into consideration, if the blocking
      // store is in an unreached block, breaking the memcopy could lose
      // performance.
      if (hasSameBaseOpValue(LoadInst, PBInst) &&
          isBlockingStore(LdDispImm, getRegSizeInBytes(LoadInst), PBstDispImm,
                          PBstSize))
        updateBlockingStoresDispSizeMap(BlockingStoresDispSizeMap, PBstDispImm,
                                        PBstSize);
    }

    if (BlockingStoresDispSizeMap.empty())
      continue;

    // We found a store forward block, break the memcpy's load and store
    // into smaller copies such that each smaller store that was causing
    // a store block would now be copied separately.
    MachineInstr *StoreInst = LoadStoreInstPair.second;
    LLVM_DEBUG(dbgs() << "Blocked load and store instructions: \n");
    LLVM_DEBUG(LoadInst->dump());
    LLVM_DEBUG(StoreInst->dump());
    LLVM_DEBUG(dbgs() << "Replaced with:\n");
    removeRedundantBlockingStores(BlockingStoresDispSizeMap);
    breakBlockedCopies(LoadInst, StoreInst, BlockingStoresDispSizeMap);
    updateKillStatus(LoadInst, StoreInst);
    ForRemoval.push_back(LoadInst);
    ForRemoval.push_back(StoreInst);
  }
  for (auto RemovedInst : ForRemoval) {
    RemovedInst->eraseFromParent();
  }
  ForRemoval.clear();
  BlockedLoadsStoresPairs.clear();
  LLVM_DEBUG(dbgs() << "End X86AvoidStoreForwardBlocks\n";);

  return Changed;
}