llvm.org GIT mirror llvm / 0392724 lib / Target / AMDGPU / SILowerI1Copies.cpp
0392724

Tree @0392724 (Download .tar.gz)

SILowerI1Copies.cpp @0392724raw · 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
729
730
731
732
733
734
735
736
737
738
739
740
741
742
743
744
745
746
747
748
749
750
751
752
753
754
755
756
757
758
759
760
761
762
763
764
765
766
767
768
769
770
771
772
773
774
775
776
777
778
779
780
781
782
783
784
785
786
787
788
789
790
791
792
793
794
795
796
797
798
799
800
801
802
803
804
805
806
807
808
809
810
811
812
813
814
815
816
817
818
819
820
821
822
823
824
825
826
827
828
829
830
831
832
833
834
835
836
837
838
839
840
841
842
843
844
845
846
847
848
849
850
851
852
853
854
855
856
857
858
859
860
861
862
863
864
865
866
//===-- SILowerI1Copies.cpp - Lower I1 Copies -----------------------------===//
//
// 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 pass lowers all occurrences of i1 values (with a vreg_1 register class)
// to lane masks (32 / 64-bit scalar registers). The pass assumes machine SSA
// form and a wave-level control flow graph.
//
// Before this pass, values that are semantically i1 and are defined and used
// within the same basic block are already represented as lane masks in scalar
// registers. However, values that cross basic blocks are always transferred
// between basic blocks in vreg_1 virtual registers and are lowered by this
// pass.
//
// The only instructions that use or define vreg_1 virtual registers are COPY,
// PHI, and IMPLICIT_DEF.
//
//===----------------------------------------------------------------------===//

#include "AMDGPU.h"
#include "AMDGPUSubtarget.h"
#include "MCTargetDesc/AMDGPUMCTargetDesc.h"
#include "SIInstrInfo.h"
#include "llvm/CodeGen/MachineDominators.h"
#include "llvm/CodeGen/MachineFunctionPass.h"
#include "llvm/CodeGen/MachineInstrBuilder.h"
#include "llvm/CodeGen/MachinePostDominators.h"
#include "llvm/CodeGen/MachineRegisterInfo.h"
#include "llvm/CodeGen/MachineSSAUpdater.h"
#include "llvm/IR/Function.h"
#include "llvm/IR/LLVMContext.h"
#include "llvm/Support/Debug.h"
#include "llvm/Target/TargetMachine.h"

#define DEBUG_TYPE "si-i1-copies"

using namespace llvm;

static unsigned createLaneMaskReg(MachineFunction &MF);
static unsigned insertUndefLaneMask(MachineBasicBlock &MBB);

namespace {

class SILowerI1Copies : public MachineFunctionPass {
public:
  static char ID;

private:
  bool IsWave32 = false;
  MachineFunction *MF = nullptr;
  MachineDominatorTree *DT = nullptr;
  MachinePostDominatorTree *PDT = nullptr;
  MachineRegisterInfo *MRI = nullptr;
  const GCNSubtarget *ST = nullptr;
  const SIInstrInfo *TII = nullptr;

  unsigned ExecReg;
  unsigned MovOp;
  unsigned AndOp;
  unsigned OrOp;
  unsigned XorOp;
  unsigned AndN2Op;
  unsigned OrN2Op;

  DenseSet<unsigned> ConstrainRegs;

public:
  SILowerI1Copies() : MachineFunctionPass(ID) {
    initializeSILowerI1CopiesPass(*PassRegistry::getPassRegistry());
  }

  bool runOnMachineFunction(MachineFunction &MF) override;

  StringRef getPassName() const override { return "SI Lower i1 Copies"; }

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

private:
  void lowerCopiesFromI1();
  void lowerPhis();
  void lowerCopiesToI1();
  bool isConstantLaneMask(unsigned Reg, bool &Val) const;
  void buildMergeLaneMasks(MachineBasicBlock &MBB,
                           MachineBasicBlock::iterator I, const DebugLoc &DL,
                           unsigned DstReg, unsigned PrevReg, unsigned CurReg);
  MachineBasicBlock::iterator
  getSaluInsertionAtEnd(MachineBasicBlock &MBB) const;

  bool isVreg1(unsigned Reg) const {
    return TargetRegisterInfo::isVirtualRegister(Reg) &&
           MRI->getRegClass(Reg) == &AMDGPU::VReg_1RegClass;
  }

  bool isLaneMaskReg(unsigned Reg) const {
    return TII->getRegisterInfo().isSGPRReg(*MRI, Reg) &&
           TII->getRegisterInfo().getRegSizeInBits(Reg, *MRI) ==
               ST->getWavefrontSize();
  }
};

/// Helper class that determines the relationship between incoming values of a
/// phi in the control flow graph to determine where an incoming value can
/// simply be taken as a scalar lane mask as-is, and where it needs to be
/// merged with another, previously defined lane mask.
///
/// The approach is as follows:
///  - Determine all basic blocks which, starting from the incoming blocks,
///    a wave may reach before entering the def block (the block containing the
///    phi).
///  - If an incoming block has no predecessors in this set, we can take the
///    incoming value as a scalar lane mask as-is.
///  -- A special case of this is when the def block has a self-loop.
///  - Otherwise, the incoming value needs to be merged with a previously
///    defined lane mask.
///  - If there is a path into the set of reachable blocks that does _not_ go
///    through an incoming block where we can take the scalar lane mask as-is,
///    we need to invent an available value for the SSAUpdater. Choices are
///    0 and undef, with differing consequences for how to merge values etc.
///
/// TODO: We could use region analysis to quickly skip over SESE regions during
///       the traversal.
///
class PhiIncomingAnalysis {
  MachinePostDominatorTree &PDT;

  // For each reachable basic block, whether it is a source in the induced
  // subgraph of the CFG.
  DenseMap<MachineBasicBlock *, bool> ReachableMap;
  SmallVector<MachineBasicBlock *, 4> ReachableOrdered;
  SmallVector<MachineBasicBlock *, 4> Stack;
  SmallVector<MachineBasicBlock *, 4> Predecessors;

public:
  PhiIncomingAnalysis(MachinePostDominatorTree &PDT) : PDT(PDT) {}

  /// Returns whether \p MBB is a source in the induced subgraph of reachable
  /// blocks.
  bool isSource(MachineBasicBlock &MBB) const {
    return ReachableMap.find(&MBB)->second;
  }

  ArrayRef<MachineBasicBlock *> predecessors() const { return Predecessors; }

  void analyze(MachineBasicBlock &DefBlock,
               ArrayRef<MachineBasicBlock *> IncomingBlocks) {
    assert(Stack.empty());
    ReachableMap.clear();
    ReachableOrdered.clear();
    Predecessors.clear();

    // Insert the def block first, so that it acts as an end point for the
    // traversal.
    ReachableMap.try_emplace(&DefBlock, false);
    ReachableOrdered.push_back(&DefBlock);

    for (MachineBasicBlock *MBB : IncomingBlocks) {
      if (MBB == &DefBlock) {
        ReachableMap[&DefBlock] = true; // self-loop on DefBlock
        continue;
      }

      ReachableMap.try_emplace(MBB, false);
      ReachableOrdered.push_back(MBB);

      // If this block has a divergent terminator and the def block is its
      // post-dominator, the wave may first visit the other successors.
      bool Divergent = false;
      for (MachineInstr &MI : MBB->terminators()) {
        if (MI.getOpcode() == AMDGPU::SI_NON_UNIFORM_BRCOND_PSEUDO ||
            MI.getOpcode() == AMDGPU::SI_IF ||
            MI.getOpcode() == AMDGPU::SI_ELSE ||
            MI.getOpcode() == AMDGPU::SI_LOOP) {
          Divergent = true;
          break;
        }
      }

      if (Divergent && PDT.dominates(&DefBlock, MBB)) {
        for (MachineBasicBlock *Succ : MBB->successors())
          Stack.push_back(Succ);
      }
    }

    while (!Stack.empty()) {
      MachineBasicBlock *MBB = Stack.pop_back_val();
      if (!ReachableMap.try_emplace(MBB, false).second)
        continue;
      ReachableOrdered.push_back(MBB);

      for (MachineBasicBlock *Succ : MBB->successors())
        Stack.push_back(Succ);
    }

    for (MachineBasicBlock *MBB : ReachableOrdered) {
      bool HaveReachablePred = false;
      for (MachineBasicBlock *Pred : MBB->predecessors()) {
        if (ReachableMap.count(Pred)) {
          HaveReachablePred = true;
        } else {
          Stack.push_back(Pred);
        }
      }
      if (!HaveReachablePred)
        ReachableMap[MBB] = true;
      if (HaveReachablePred) {
        for (MachineBasicBlock *UnreachablePred : Stack) {
          if (llvm::find(Predecessors, UnreachablePred) == Predecessors.end())
            Predecessors.push_back(UnreachablePred);
        }
      }
      Stack.clear();
    }
  }
};

/// Helper class that detects loops which require us to lower an i1 COPY into
/// bitwise manipulation.
///
/// Unfortunately, we cannot use LoopInfo because LoopInfo does not distinguish
/// between loops with the same header. Consider this example:
///
///  A-+-+
///  | | |
///  B-+ |
///  |   |
///  C---+
///
/// A is the header of a loop containing A, B, and C as far as LoopInfo is
/// concerned. However, an i1 COPY in B that is used in C must be lowered to
/// bitwise operations to combine results from different loop iterations when
/// B has a divergent branch (since by default we will compile this code such
/// that threads in a wave are merged at the entry of C).
///
/// The following rule is implemented to determine whether bitwise operations
/// are required: use the bitwise lowering for a def in block B if a backward
/// edge to B is reachable without going through the nearest common
/// post-dominator of B and all uses of the def.
///
/// TODO: This rule is conservative because it does not check whether the
///       relevant branches are actually divergent.
///
/// The class is designed to cache the CFG traversal so that it can be re-used
/// for multiple defs within the same basic block.
///
/// TODO: We could use region analysis to quickly skip over SESE regions during
///       the traversal.
///
class LoopFinder {
  MachineDominatorTree &DT;
  MachinePostDominatorTree &PDT;

  // All visited / reachable block, tagged by level (level 0 is the def block,
  // level 1 are all blocks reachable including but not going through the def
  // block's IPDOM, etc.).
  DenseMap<MachineBasicBlock *, unsigned> Visited;

  // Nearest common dominator of all visited blocks by level (level 0 is the
  // def block). Used for seeding the SSAUpdater.
  SmallVector<MachineBasicBlock *, 4> CommonDominators;

  // Post-dominator of all visited blocks.
  MachineBasicBlock *VisitedPostDom = nullptr;

  // Level at which a loop was found: 0 is not possible; 1 = a backward edge is
  // reachable without going through the IPDOM of the def block (if the IPDOM
  // itself has an edge to the def block, the loop level is 2), etc.
  unsigned FoundLoopLevel = ~0u;

  MachineBasicBlock *DefBlock = nullptr;
  SmallVector<MachineBasicBlock *, 4> Stack;
  SmallVector<MachineBasicBlock *, 4> NextLevel;

public:
  LoopFinder(MachineDominatorTree &DT, MachinePostDominatorTree &PDT)
      : DT(DT), PDT(PDT) {}

  void initialize(MachineBasicBlock &MBB) {
    Visited.clear();
    CommonDominators.clear();
    Stack.clear();
    NextLevel.clear();
    VisitedPostDom = nullptr;
    FoundLoopLevel = ~0u;

    DefBlock = &MBB;
  }

  /// Check whether a backward edge can be reached without going through the
  /// given \p PostDom of the def block.
  ///
  /// Return the level of \p PostDom if a loop was found, or 0 otherwise.
  unsigned findLoop(MachineBasicBlock *PostDom) {
    MachineDomTreeNode *PDNode = PDT.getNode(DefBlock);

    if (!VisitedPostDom)
      advanceLevel();

    unsigned Level = 0;
    while (PDNode->getBlock() != PostDom) {
      if (PDNode->getBlock() == VisitedPostDom)
        advanceLevel();
      PDNode = PDNode->getIDom();
      Level++;
      if (FoundLoopLevel == Level)
        return Level;
    }

    return 0;
  }

  /// Add undef values dominating the loop and the optionally given additional
  /// blocks, so that the SSA updater doesn't have to search all the way to the
  /// function entry.
  void addLoopEntries(unsigned LoopLevel, MachineSSAUpdater &SSAUpdater,
                      ArrayRef<MachineBasicBlock *> Blocks = {}) {
    assert(LoopLevel < CommonDominators.size());

    MachineBasicBlock *Dom = CommonDominators[LoopLevel];
    for (MachineBasicBlock *MBB : Blocks)
      Dom = DT.findNearestCommonDominator(Dom, MBB);

    if (!inLoopLevel(*Dom, LoopLevel, Blocks)) {
      SSAUpdater.AddAvailableValue(Dom, insertUndefLaneMask(*Dom));
    } else {
      // The dominator is part of the loop or the given blocks, so add the
      // undef value to unreachable predecessors instead.
      for (MachineBasicBlock *Pred : Dom->predecessors()) {
        if (!inLoopLevel(*Pred, LoopLevel, Blocks))
          SSAUpdater.AddAvailableValue(Pred, insertUndefLaneMask(*Pred));
      }
    }
  }

private:
  bool inLoopLevel(MachineBasicBlock &MBB, unsigned LoopLevel,
                   ArrayRef<MachineBasicBlock *> Blocks) const {
    auto DomIt = Visited.find(&MBB);
    if (DomIt != Visited.end() && DomIt->second <= LoopLevel)
      return true;

    if (llvm::find(Blocks, &MBB) != Blocks.end())
      return true;

    return false;
  }

  void advanceLevel() {
    MachineBasicBlock *VisitedDom;

    if (!VisitedPostDom) {
      VisitedPostDom = DefBlock;
      VisitedDom = DefBlock;
      Stack.push_back(DefBlock);
    } else {
      VisitedPostDom = PDT.getNode(VisitedPostDom)->getIDom()->getBlock();
      VisitedDom = CommonDominators.back();

      for (unsigned i = 0; i < NextLevel.size();) {
        if (PDT.dominates(VisitedPostDom, NextLevel[i])) {
          Stack.push_back(NextLevel[i]);

          NextLevel[i] = NextLevel.back();
          NextLevel.pop_back();
        } else {
          i++;
        }
      }
    }

    unsigned Level = CommonDominators.size();
    while (!Stack.empty()) {
      MachineBasicBlock *MBB = Stack.pop_back_val();
      if (!PDT.dominates(VisitedPostDom, MBB))
        NextLevel.push_back(MBB);

      Visited[MBB] = Level;
      VisitedDom = DT.findNearestCommonDominator(VisitedDom, MBB);

      for (MachineBasicBlock *Succ : MBB->successors()) {
        if (Succ == DefBlock) {
          if (MBB == VisitedPostDom)
            FoundLoopLevel = std::min(FoundLoopLevel, Level + 1);
          else
            FoundLoopLevel = std::min(FoundLoopLevel, Level);
          continue;
        }

        if (Visited.try_emplace(Succ, ~0u).second) {
          if (MBB == VisitedPostDom)
            NextLevel.push_back(Succ);
          else
            Stack.push_back(Succ);
        }
      }
    }

    CommonDominators.push_back(VisitedDom);
  }
};

} // End anonymous namespace.

INITIALIZE_PASS_BEGIN(SILowerI1Copies, DEBUG_TYPE, "SI Lower i1 Copies", false,
                      false)
INITIALIZE_PASS_DEPENDENCY(MachineDominatorTree)
INITIALIZE_PASS_DEPENDENCY(MachinePostDominatorTree)
INITIALIZE_PASS_END(SILowerI1Copies, DEBUG_TYPE, "SI Lower i1 Copies", false,
                    false)

char SILowerI1Copies::ID = 0;

char &llvm::SILowerI1CopiesID = SILowerI1Copies::ID;

FunctionPass *llvm::createSILowerI1CopiesPass() {
  return new SILowerI1Copies();
}

static unsigned createLaneMaskReg(MachineFunction &MF) {
  const GCNSubtarget &ST = MF.getSubtarget<GCNSubtarget>();
  MachineRegisterInfo &MRI = MF.getRegInfo();
  return MRI.createVirtualRegister(ST.isWave32() ? &AMDGPU::SReg_32RegClass
                                                 : &AMDGPU::SReg_64RegClass);
}

static unsigned insertUndefLaneMask(MachineBasicBlock &MBB) {
  MachineFunction &MF = *MBB.getParent();
  const GCNSubtarget &ST = MF.getSubtarget<GCNSubtarget>();
  const SIInstrInfo *TII = ST.getInstrInfo();
  unsigned UndefReg = createLaneMaskReg(MF);
  BuildMI(MBB, MBB.getFirstTerminator(), {}, TII->get(AMDGPU::IMPLICIT_DEF),
          UndefReg);
  return UndefReg;
}

/// Lower all instructions that def or use vreg_1 registers.
///
/// In a first pass, we lower COPYs from vreg_1 to vector registers, as can
/// occur around inline assembly. We do this first, before vreg_1 registers
/// are changed to scalar mask registers.
///
/// Then we lower all defs of vreg_1 registers. Phi nodes are lowered before
/// all others, because phi lowering looks through copies and can therefore
/// often make copy lowering unnecessary.
bool SILowerI1Copies::runOnMachineFunction(MachineFunction &TheMF) {
  MF = &TheMF;
  MRI = &MF->getRegInfo();
  DT = &getAnalysis<MachineDominatorTree>();
  PDT = &getAnalysis<MachinePostDominatorTree>();

  ST = &MF->getSubtarget<GCNSubtarget>();
  TII = ST->getInstrInfo();
  IsWave32 = ST->isWave32();

  if (IsWave32) {
    ExecReg = AMDGPU::EXEC_LO;
    MovOp = AMDGPU::S_MOV_B32;
    AndOp = AMDGPU::S_AND_B32;
    OrOp = AMDGPU::S_OR_B32;
    XorOp = AMDGPU::S_XOR_B32;
    AndN2Op = AMDGPU::S_ANDN2_B32;
    OrN2Op = AMDGPU::S_ORN2_B32;
  } else {
    ExecReg = AMDGPU::EXEC;
    MovOp = AMDGPU::S_MOV_B64;
    AndOp = AMDGPU::S_AND_B64;
    OrOp = AMDGPU::S_OR_B64;
    XorOp = AMDGPU::S_XOR_B64;
    AndN2Op = AMDGPU::S_ANDN2_B64;
    OrN2Op = AMDGPU::S_ORN2_B64;
  }

  lowerCopiesFromI1();
  lowerPhis();
  lowerCopiesToI1();

  for (unsigned Reg : ConstrainRegs)
    MRI->constrainRegClass(Reg, &AMDGPU::SReg_1_XEXECRegClass);
  ConstrainRegs.clear();

  return true;
}

void SILowerI1Copies::lowerCopiesFromI1() {
  SmallVector<MachineInstr *, 4> DeadCopies;

  for (MachineBasicBlock &MBB : *MF) {
    for (MachineInstr &MI : MBB) {
      if (MI.getOpcode() != AMDGPU::COPY)
        continue;

      unsigned DstReg = MI.getOperand(0).getReg();
      unsigned SrcReg = MI.getOperand(1).getReg();
      if (!isVreg1(SrcReg))
        continue;

      if (isLaneMaskReg(DstReg) || isVreg1(DstReg))
        continue;

      // Copy into a 32-bit vector register.
      LLVM_DEBUG(dbgs() << "Lower copy from i1: " << MI);
      DebugLoc DL = MI.getDebugLoc();

      assert(TII->getRegisterInfo().getRegSizeInBits(DstReg, *MRI) == 32);
      assert(!MI.getOperand(0).getSubReg());

      ConstrainRegs.insert(SrcReg);
      BuildMI(MBB, MI, DL, TII->get(AMDGPU::V_CNDMASK_B32_e64), DstReg)
          .addImm(0)
          .addImm(0)
          .addImm(0)
          .addImm(-1)
          .addReg(SrcReg);
      DeadCopies.push_back(&MI);
    }

    for (MachineInstr *MI : DeadCopies)
      MI->eraseFromParent();
    DeadCopies.clear();
  }
}

void SILowerI1Copies::lowerPhis() {
  MachineSSAUpdater SSAUpdater(*MF);
  LoopFinder LF(*DT, *PDT);
  PhiIncomingAnalysis PIA(*PDT);
  SmallVector<MachineInstr *, 4> DeadPhis;
  SmallVector<MachineBasicBlock *, 4> IncomingBlocks;
  SmallVector<unsigned, 4> IncomingRegs;
  SmallVector<unsigned, 4> IncomingUpdated;
#ifndef NDEBUG
  DenseSet<unsigned> PhiRegisters;
#endif

  for (MachineBasicBlock &MBB : *MF) {
    LF.initialize(MBB);

    for (MachineInstr &MI : MBB.phis()) {
      unsigned DstReg = MI.getOperand(0).getReg();
      if (!isVreg1(DstReg))
        continue;

      LLVM_DEBUG(dbgs() << "Lower PHI: " << MI);

      MRI->setRegClass(DstReg, IsWave32 ? &AMDGPU::SReg_32RegClass
                                        : &AMDGPU::SReg_64RegClass);

      // Collect incoming values.
      for (unsigned i = 1; i < MI.getNumOperands(); i += 2) {
        assert(i + 1 < MI.getNumOperands());
        unsigned IncomingReg = MI.getOperand(i).getReg();
        MachineBasicBlock *IncomingMBB = MI.getOperand(i + 1).getMBB();
        MachineInstr *IncomingDef = MRI->getUniqueVRegDef(IncomingReg);

        if (IncomingDef->getOpcode() == AMDGPU::COPY) {
          IncomingReg = IncomingDef->getOperand(1).getReg();
          assert(isLaneMaskReg(IncomingReg) || isVreg1(IncomingReg));
          assert(!IncomingDef->getOperand(1).getSubReg());
        } else if (IncomingDef->getOpcode() == AMDGPU::IMPLICIT_DEF) {
          continue;
        } else {
          assert(IncomingDef->isPHI() || PhiRegisters.count(IncomingReg));
        }

        IncomingBlocks.push_back(IncomingMBB);
        IncomingRegs.push_back(IncomingReg);
      }

#ifndef NDEBUG
      PhiRegisters.insert(DstReg);
#endif

      // Phis in a loop that are observed outside the loop receive a simple but
      // conservatively correct treatment.
      MachineBasicBlock *PostDomBound = &MBB;
      for (MachineInstr &Use : MRI->use_instructions(DstReg)) {
        PostDomBound =
            PDT->findNearestCommonDominator(PostDomBound, Use.getParent());
      }

      unsigned FoundLoopLevel = LF.findLoop(PostDomBound);

      SSAUpdater.Initialize(DstReg);

      if (FoundLoopLevel) {
        LF.addLoopEntries(FoundLoopLevel, SSAUpdater, IncomingBlocks);

        for (unsigned i = 0; i < IncomingRegs.size(); ++i) {
          IncomingUpdated.push_back(createLaneMaskReg(*MF));
          SSAUpdater.AddAvailableValue(IncomingBlocks[i],
                                       IncomingUpdated.back());
        }

        for (unsigned i = 0; i < IncomingRegs.size(); ++i) {
          MachineBasicBlock &IMBB = *IncomingBlocks[i];
          buildMergeLaneMasks(
              IMBB, getSaluInsertionAtEnd(IMBB), {}, IncomingUpdated[i],
              SSAUpdater.GetValueInMiddleOfBlock(&IMBB), IncomingRegs[i]);
        }
      } else {
        // The phi is not observed from outside a loop. Use a more accurate
        // lowering.
        PIA.analyze(MBB, IncomingBlocks);

        for (MachineBasicBlock *MBB : PIA.predecessors())
          SSAUpdater.AddAvailableValue(MBB, insertUndefLaneMask(*MBB));

        for (unsigned i = 0; i < IncomingRegs.size(); ++i) {
          MachineBasicBlock &IMBB = *IncomingBlocks[i];
          if (PIA.isSource(IMBB)) {
            IncomingUpdated.push_back(0);
            SSAUpdater.AddAvailableValue(&IMBB, IncomingRegs[i]);
          } else {
            IncomingUpdated.push_back(createLaneMaskReg(*MF));
            SSAUpdater.AddAvailableValue(&IMBB, IncomingUpdated.back());
          }
        }

        for (unsigned i = 0; i < IncomingRegs.size(); ++i) {
          if (!IncomingUpdated[i])
            continue;

          MachineBasicBlock &IMBB = *IncomingBlocks[i];
          buildMergeLaneMasks(
              IMBB, getSaluInsertionAtEnd(IMBB), {}, IncomingUpdated[i],
              SSAUpdater.GetValueInMiddleOfBlock(&IMBB), IncomingRegs[i]);
        }
      }

      unsigned NewReg = SSAUpdater.GetValueInMiddleOfBlock(&MBB);
      if (NewReg != DstReg) {
        MRI->replaceRegWith(NewReg, DstReg);

        // Ensure that DstReg has a single def and mark the old PHI node for
        // deletion.
        MI.getOperand(0).setReg(NewReg);
        DeadPhis.push_back(&MI);
      }

      IncomingBlocks.clear();
      IncomingRegs.clear();
      IncomingUpdated.clear();
    }

    for (MachineInstr *MI : DeadPhis)
      MI->eraseFromParent();
    DeadPhis.clear();
  }
}

void SILowerI1Copies::lowerCopiesToI1() {
  MachineSSAUpdater SSAUpdater(*MF);
  LoopFinder LF(*DT, *PDT);
  SmallVector<MachineInstr *, 4> DeadCopies;

  for (MachineBasicBlock &MBB : *MF) {
    LF.initialize(MBB);

    for (MachineInstr &MI : MBB) {
      if (MI.getOpcode() != AMDGPU::IMPLICIT_DEF &&
          MI.getOpcode() != AMDGPU::COPY)
        continue;

      unsigned DstReg = MI.getOperand(0).getReg();
      if (!isVreg1(DstReg))
        continue;

      if (MRI->use_empty(DstReg)) {
        DeadCopies.push_back(&MI);
        continue;
      }

      LLVM_DEBUG(dbgs() << "Lower Other: " << MI);

      MRI->setRegClass(DstReg, IsWave32 ? &AMDGPU::SReg_32RegClass
                                        : &AMDGPU::SReg_64RegClass);
      if (MI.getOpcode() == AMDGPU::IMPLICIT_DEF)
        continue;

      DebugLoc DL = MI.getDebugLoc();
      unsigned SrcReg = MI.getOperand(1).getReg();
      assert(!MI.getOperand(1).getSubReg());

      if (!TargetRegisterInfo::isVirtualRegister(SrcReg) ||
          (!isLaneMaskReg(SrcReg) && !isVreg1(SrcReg))) {
        assert(TII->getRegisterInfo().getRegSizeInBits(SrcReg, *MRI) == 32);
        unsigned TmpReg = createLaneMaskReg(*MF);
        BuildMI(MBB, MI, DL, TII->get(AMDGPU::V_CMP_NE_U32_e64), TmpReg)
            .addReg(SrcReg)
            .addImm(0);
        MI.getOperand(1).setReg(TmpReg);
        SrcReg = TmpReg;
      }

      // Defs in a loop that are observed outside the loop must be transformed
      // into appropriate bit manipulation.
      MachineBasicBlock *PostDomBound = &MBB;
      for (MachineInstr &Use : MRI->use_instructions(DstReg)) {
        PostDomBound =
            PDT->findNearestCommonDominator(PostDomBound, Use.getParent());
      }

      unsigned FoundLoopLevel = LF.findLoop(PostDomBound);
      if (FoundLoopLevel) {
        SSAUpdater.Initialize(DstReg);
        SSAUpdater.AddAvailableValue(&MBB, DstReg);
        LF.addLoopEntries(FoundLoopLevel, SSAUpdater);

        buildMergeLaneMasks(MBB, MI, DL, DstReg,
                            SSAUpdater.GetValueInMiddleOfBlock(&MBB), SrcReg);
        DeadCopies.push_back(&MI);
      }
    }

    for (MachineInstr *MI : DeadCopies)
      MI->eraseFromParent();
    DeadCopies.clear();
  }
}

bool SILowerI1Copies::isConstantLaneMask(unsigned Reg, bool &Val) const {
  const MachineInstr *MI;
  for (;;) {
    MI = MRI->getUniqueVRegDef(Reg);
    if (MI->getOpcode() != AMDGPU::COPY)
      break;

    Reg = MI->getOperand(1).getReg();
    if (!TargetRegisterInfo::isVirtualRegister(Reg))
      return false;
    if (!isLaneMaskReg(Reg))
      return false;
  }

  if (MI->getOpcode() != MovOp)
    return false;

  if (!MI->getOperand(1).isImm())
    return false;

  int64_t Imm = MI->getOperand(1).getImm();
  if (Imm == 0) {
    Val = false;
    return true;
  }
  if (Imm == -1) {
    Val = true;
    return true;
  }

  return false;
}

static void instrDefsUsesSCC(const MachineInstr &MI, bool &Def, bool &Use) {
  Def = false;
  Use = false;

  for (const MachineOperand &MO : MI.operands()) {
    if (MO.isReg() && MO.getReg() == AMDGPU::SCC) {
      if (MO.isUse())
        Use = true;
      else
        Def = true;
    }
  }
}

/// Return a point at the end of the given \p MBB to insert SALU instructions
/// for lane mask calculation. Take terminators and SCC into account.
MachineBasicBlock::iterator
SILowerI1Copies::getSaluInsertionAtEnd(MachineBasicBlock &MBB) const {
  auto InsertionPt = MBB.getFirstTerminator();
  bool TerminatorsUseSCC = false;
  for (auto I = InsertionPt, E = MBB.end(); I != E; ++I) {
    bool DefsSCC;
    instrDefsUsesSCC(*I, DefsSCC, TerminatorsUseSCC);
    if (TerminatorsUseSCC || DefsSCC)
      break;
  }

  if (!TerminatorsUseSCC)
    return InsertionPt;

  while (InsertionPt != MBB.begin()) {
    InsertionPt--;

    bool DefSCC, UseSCC;
    instrDefsUsesSCC(*InsertionPt, DefSCC, UseSCC);
    if (DefSCC)
      return InsertionPt;
  }

  // We should have at least seen an IMPLICIT_DEF or COPY
  llvm_unreachable("SCC used by terminator but no def in block");
}

void SILowerI1Copies::buildMergeLaneMasks(MachineBasicBlock &MBB,
                                          MachineBasicBlock::iterator I,
                                          const DebugLoc &DL, unsigned DstReg,
                                          unsigned PrevReg, unsigned CurReg) {
  bool PrevVal;
  bool PrevConstant = isConstantLaneMask(PrevReg, PrevVal);
  bool CurVal;
  bool CurConstant = isConstantLaneMask(CurReg, CurVal);

  if (PrevConstant && CurConstant) {
    if (PrevVal == CurVal) {
      BuildMI(MBB, I, DL, TII->get(AMDGPU::COPY), DstReg).addReg(CurReg);
    } else if (CurVal) {
      BuildMI(MBB, I, DL, TII->get(AMDGPU::COPY), DstReg).addReg(ExecReg);
    } else {
      BuildMI(MBB, I, DL, TII->get(XorOp), DstReg)
          .addReg(ExecReg)
          .addImm(-1);
    }
    return;
  }

  unsigned PrevMaskedReg = 0;
  unsigned CurMaskedReg = 0;
  if (!PrevConstant) {
    if (CurConstant && CurVal) {
      PrevMaskedReg = PrevReg;
    } else {
      PrevMaskedReg = createLaneMaskReg(*MF);
      BuildMI(MBB, I, DL, TII->get(AndN2Op), PrevMaskedReg)
          .addReg(PrevReg)
          .addReg(ExecReg);
    }
  }
  if (!CurConstant) {
    // TODO: check whether CurReg is already masked by EXEC
    if (PrevConstant && PrevVal) {
      CurMaskedReg = CurReg;
    } else {
      CurMaskedReg = createLaneMaskReg(*MF);
      BuildMI(MBB, I, DL, TII->get(AndOp), CurMaskedReg)
          .addReg(CurReg)
          .addReg(ExecReg);
    }
  }

  if (PrevConstant && !PrevVal) {
    BuildMI(MBB, I, DL, TII->get(AMDGPU::COPY), DstReg)
        .addReg(CurMaskedReg);
  } else if (CurConstant && !CurVal) {
    BuildMI(MBB, I, DL, TII->get(AMDGPU::COPY), DstReg)
        .addReg(PrevMaskedReg);
  } else if (PrevConstant && PrevVal) {
    BuildMI(MBB, I, DL, TII->get(OrN2Op), DstReg)
        .addReg(CurMaskedReg)
        .addReg(ExecReg);
  } else {
    BuildMI(MBB, I, DL, TII->get(OrOp), DstReg)
        .addReg(PrevMaskedReg)
        .addReg(CurMaskedReg ? CurMaskedReg : ExecReg);
  }
}