llvm.org GIT mirror llvm / release_38 lib / Target / AMDGPU / R600ControlFlowFinalizer.cpp
release_38

Tree @release_38 (Download .tar.gz)

R600ControlFlowFinalizer.cpp @release_38raw · 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
//===-- R600ControlFlowFinalizer.cpp - Finalize Control Flow Inst----------===//
//
//                     The LLVM Compiler Infrastructure
//
// This file is distributed under the University of Illinois Open Source
// License. See LICENSE.TXT for details.
//
//===----------------------------------------------------------------------===//
//
/// \file
/// This pass compute turns all control flow pseudo instructions into native one
/// computing their address on the fly ; it also sets STACK_SIZE info.
//===----------------------------------------------------------------------===//

#include "llvm/Support/Debug.h"
#include "AMDGPU.h"
#include "AMDGPUSubtarget.h"
#include "R600Defines.h"
#include "R600InstrInfo.h"
#include "R600MachineFunctionInfo.h"
#include "R600RegisterInfo.h"
#include "llvm/CodeGen/MachineFunctionPass.h"
#include "llvm/CodeGen/MachineInstrBuilder.h"
#include "llvm/CodeGen/MachineRegisterInfo.h"
#include "llvm/Support/raw_ostream.h"

using namespace llvm;

#define DEBUG_TYPE "r600cf"

namespace {

struct CFStack {

  enum StackItem {
    ENTRY = 0,
    SUB_ENTRY = 1,
    FIRST_NON_WQM_PUSH = 2,
    FIRST_NON_WQM_PUSH_W_FULL_ENTRY = 3
  };

  const AMDGPUSubtarget *ST;
  std::vector<StackItem> BranchStack;
  std::vector<StackItem> LoopStack;
  unsigned MaxStackSize;
  unsigned CurrentEntries;
  unsigned CurrentSubEntries;

  CFStack(const AMDGPUSubtarget *st, unsigned ShaderType) : ST(st),
      // We need to reserve a stack entry for CALL_FS in vertex shaders.
      MaxStackSize(ShaderType == ShaderType::VERTEX ? 1 : 0),
      CurrentEntries(0), CurrentSubEntries(0) { }

  unsigned getLoopDepth();
  bool branchStackContains(CFStack::StackItem);
  bool requiresWorkAroundForInst(unsigned Opcode);
  unsigned getSubEntrySize(CFStack::StackItem Item);
  void updateMaxStackSize();
  void pushBranch(unsigned Opcode, bool isWQM = false);
  void pushLoop();
  void popBranch();
  void popLoop();
};

unsigned CFStack::getLoopDepth() {
  return LoopStack.size();
}

bool CFStack::branchStackContains(CFStack::StackItem Item) {
  for (std::vector<CFStack::StackItem>::const_iterator I = BranchStack.begin(),
       E = BranchStack.end(); I != E; ++I) {
    if (*I == Item)
      return true;
  }
  return false;
}

bool CFStack::requiresWorkAroundForInst(unsigned Opcode) {
  if (Opcode == AMDGPU::CF_ALU_PUSH_BEFORE && ST->hasCaymanISA() &&
      getLoopDepth() > 1)
    return true;

  if (!ST->hasCFAluBug())
    return false;

  switch(Opcode) {
  default: return false;
  case AMDGPU::CF_ALU_PUSH_BEFORE:
  case AMDGPU::CF_ALU_ELSE_AFTER:
  case AMDGPU::CF_ALU_BREAK:
  case AMDGPU::CF_ALU_CONTINUE:
    if (CurrentSubEntries == 0)
      return false;
    if (ST->getWavefrontSize() == 64) {
      // We are being conservative here.  We only require this work-around if
      // CurrentSubEntries > 3 &&
      // (CurrentSubEntries % 4 == 3 || CurrentSubEntries % 4 == 0)
      //
      // We have to be conservative, because we don't know for certain that
      // our stack allocation algorithm for Evergreen/NI is correct.  Applying this
      // work-around when CurrentSubEntries > 3 allows us to over-allocate stack
      // resources without any problems.
      return CurrentSubEntries > 3;
    } else {
      assert(ST->getWavefrontSize() == 32);
      // We are being conservative here.  We only require the work-around if
      // CurrentSubEntries > 7 &&
      // (CurrentSubEntries % 8 == 7 || CurrentSubEntries % 8 == 0)
      // See the comment on the wavefront size == 64 case for why we are
      // being conservative.
      return CurrentSubEntries > 7;
    }
  }
}

unsigned CFStack::getSubEntrySize(CFStack::StackItem Item) {
  switch(Item) {
  default:
    return 0;
  case CFStack::FIRST_NON_WQM_PUSH:
  assert(!ST->hasCaymanISA());
  if (ST->getGeneration() <= AMDGPUSubtarget::R700) {
    // +1 For the push operation.
    // +2 Extra space required.
    return 3;
  } else {
    // Some documentation says that this is not necessary on Evergreen,
    // but experimentation has show that we need to allocate 1 extra
    // sub-entry for the first non-WQM push.
    // +1 For the push operation.
    // +1 Extra space required.
    return 2;
  }
  case CFStack::FIRST_NON_WQM_PUSH_W_FULL_ENTRY:
    assert(ST->getGeneration() >= AMDGPUSubtarget::EVERGREEN);
    // +1 For the push operation.
    // +1 Extra space required.
    return 2;
  case CFStack::SUB_ENTRY:
    return 1;
  }
}

void CFStack::updateMaxStackSize() {
  unsigned CurrentStackSize = CurrentEntries +
                              (RoundUpToAlignment(CurrentSubEntries, 4) / 4);
  MaxStackSize = std::max(CurrentStackSize, MaxStackSize);
}

void CFStack::pushBranch(unsigned Opcode, bool isWQM) {
  CFStack::StackItem Item = CFStack::ENTRY;
  switch(Opcode) {
  case AMDGPU::CF_PUSH_EG:
  case AMDGPU::CF_ALU_PUSH_BEFORE:
    if (!isWQM) {
      if (!ST->hasCaymanISA() &&
          !branchStackContains(CFStack::FIRST_NON_WQM_PUSH))
        Item = CFStack::FIRST_NON_WQM_PUSH;  // May not be required on Evergreen/NI
                                             // See comment in
                                             // CFStack::getSubEntrySize()
      else if (CurrentEntries > 0 &&
               ST->getGeneration() > AMDGPUSubtarget::EVERGREEN &&
               !ST->hasCaymanISA() &&
               !branchStackContains(CFStack::FIRST_NON_WQM_PUSH_W_FULL_ENTRY))
        Item = CFStack::FIRST_NON_WQM_PUSH_W_FULL_ENTRY;
      else
        Item = CFStack::SUB_ENTRY;
    } else
      Item = CFStack::ENTRY;
    break;
  }
  BranchStack.push_back(Item);
  if (Item == CFStack::ENTRY)
    CurrentEntries++;
  else
    CurrentSubEntries += getSubEntrySize(Item);
  updateMaxStackSize();
}

void CFStack::pushLoop() {
  LoopStack.push_back(CFStack::ENTRY);
  CurrentEntries++;
  updateMaxStackSize();
}

void CFStack::popBranch() {
  CFStack::StackItem Top = BranchStack.back();
  if (Top == CFStack::ENTRY)
    CurrentEntries--;
  else
    CurrentSubEntries-= getSubEntrySize(Top);
  BranchStack.pop_back();
}

void CFStack::popLoop() {
  CurrentEntries--;
  LoopStack.pop_back();
}

class R600ControlFlowFinalizer : public MachineFunctionPass {

private:
  typedef std::pair<MachineInstr *, std::vector<MachineInstr *> > ClauseFile;

  enum ControlFlowInstruction {
    CF_TC,
    CF_VC,
    CF_CALL_FS,
    CF_WHILE_LOOP,
    CF_END_LOOP,
    CF_LOOP_BREAK,
    CF_LOOP_CONTINUE,
    CF_JUMP,
    CF_ELSE,
    CF_POP,
    CF_END
  };

  static char ID;
  const R600InstrInfo *TII;
  const R600RegisterInfo *TRI;
  unsigned MaxFetchInst;
  const AMDGPUSubtarget *ST;

  bool IsTrivialInst(MachineInstr *MI) const {
    switch (MI->getOpcode()) {
    case AMDGPU::KILL:
    case AMDGPU::RETURN:
      return true;
    default:
      return false;
    }
  }

  const MCInstrDesc &getHWInstrDesc(ControlFlowInstruction CFI) const {
    unsigned Opcode = 0;
    bool isEg = (ST->getGeneration() >= AMDGPUSubtarget::EVERGREEN);
    switch (CFI) {
    case CF_TC:
      Opcode = isEg ? AMDGPU::CF_TC_EG : AMDGPU::CF_TC_R600;
      break;
    case CF_VC:
      Opcode = isEg ? AMDGPU::CF_VC_EG : AMDGPU::CF_VC_R600;
      break;
    case CF_CALL_FS:
      Opcode = isEg ? AMDGPU::CF_CALL_FS_EG : AMDGPU::CF_CALL_FS_R600;
      break;
    case CF_WHILE_LOOP:
      Opcode = isEg ? AMDGPU::WHILE_LOOP_EG : AMDGPU::WHILE_LOOP_R600;
      break;
    case CF_END_LOOP:
      Opcode = isEg ? AMDGPU::END_LOOP_EG : AMDGPU::END_LOOP_R600;
      break;
    case CF_LOOP_BREAK:
      Opcode = isEg ? AMDGPU::LOOP_BREAK_EG : AMDGPU::LOOP_BREAK_R600;
      break;
    case CF_LOOP_CONTINUE:
      Opcode = isEg ? AMDGPU::CF_CONTINUE_EG : AMDGPU::CF_CONTINUE_R600;
      break;
    case CF_JUMP:
      Opcode = isEg ? AMDGPU::CF_JUMP_EG : AMDGPU::CF_JUMP_R600;
      break;
    case CF_ELSE:
      Opcode = isEg ? AMDGPU::CF_ELSE_EG : AMDGPU::CF_ELSE_R600;
      break;
    case CF_POP:
      Opcode = isEg ? AMDGPU::POP_EG : AMDGPU::POP_R600;
      break;
    case CF_END:
      if (ST->hasCaymanISA()) {
        Opcode = AMDGPU::CF_END_CM;
        break;
      }
      Opcode = isEg ? AMDGPU::CF_END_EG : AMDGPU::CF_END_R600;
      break;
    }
    assert (Opcode && "No opcode selected");
    return TII->get(Opcode);
  }

  bool isCompatibleWithClause(const MachineInstr *MI,
      std::set<unsigned> &DstRegs) const {
    unsigned DstMI, SrcMI;
    for (MachineInstr::const_mop_iterator I = MI->operands_begin(),
        E = MI->operands_end(); I != E; ++I) {
      const MachineOperand &MO = *I;
      if (!MO.isReg())
        continue;
      if (MO.isDef()) {
        unsigned Reg = MO.getReg();
        if (AMDGPU::R600_Reg128RegClass.contains(Reg))
          DstMI = Reg;
        else
          DstMI = TRI->getMatchingSuperReg(Reg,
              TRI->getSubRegFromChannel(TRI->getHWRegChan(Reg)),
              &AMDGPU::R600_Reg128RegClass);
      }
      if (MO.isUse()) {
        unsigned Reg = MO.getReg();
        if (AMDGPU::R600_Reg128RegClass.contains(Reg))
          SrcMI = Reg;
        else
          SrcMI = TRI->getMatchingSuperReg(Reg,
              TRI->getSubRegFromChannel(TRI->getHWRegChan(Reg)),
              &AMDGPU::R600_Reg128RegClass);
      }
    }
    if ((DstRegs.find(SrcMI) == DstRegs.end())) {
      DstRegs.insert(DstMI);
      return true;
    } else
      return false;
  }

  ClauseFile
  MakeFetchClause(MachineBasicBlock &MBB, MachineBasicBlock::iterator &I)
      const {
    MachineBasicBlock::iterator ClauseHead = I;
    std::vector<MachineInstr *> ClauseContent;
    unsigned AluInstCount = 0;
    bool IsTex = TII->usesTextureCache(ClauseHead);
    std::set<unsigned> DstRegs;
    for (MachineBasicBlock::iterator E = MBB.end(); I != E; ++I) {
      if (IsTrivialInst(I))
        continue;
      if (AluInstCount >= MaxFetchInst)
        break;
      if ((IsTex && !TII->usesTextureCache(I)) ||
          (!IsTex && !TII->usesVertexCache(I)))
        break;
      if (!isCompatibleWithClause(I, DstRegs))
        break;
      AluInstCount ++;
      ClauseContent.push_back(I);
    }
    MachineInstr *MIb = BuildMI(MBB, ClauseHead, MBB.findDebugLoc(ClauseHead),
        getHWInstrDesc(IsTex?CF_TC:CF_VC))
        .addImm(0) // ADDR
        .addImm(AluInstCount - 1); // COUNT
    return ClauseFile(MIb, std::move(ClauseContent));
  }

  void getLiteral(MachineInstr *MI, std::vector<int64_t> &Lits) const {
    static const unsigned LiteralRegs[] = {
      AMDGPU::ALU_LITERAL_X,
      AMDGPU::ALU_LITERAL_Y,
      AMDGPU::ALU_LITERAL_Z,
      AMDGPU::ALU_LITERAL_W
    };
    const SmallVector<std::pair<MachineOperand *, int64_t>, 3 > Srcs =
        TII->getSrcs(MI);
    for (unsigned i = 0, e = Srcs.size(); i < e; ++i) {
      if (Srcs[i].first->getReg() != AMDGPU::ALU_LITERAL_X)
        continue;
      int64_t Imm = Srcs[i].second;
      std::vector<int64_t>::iterator It =
          std::find(Lits.begin(), Lits.end(), Imm);
      if (It != Lits.end()) {
        unsigned Index = It - Lits.begin();
        Srcs[i].first->setReg(LiteralRegs[Index]);
      } else {
        assert(Lits.size() < 4 && "Too many literals in Instruction Group");
        Srcs[i].first->setReg(LiteralRegs[Lits.size()]);
        Lits.push_back(Imm);
      }
    }
  }

  MachineBasicBlock::iterator insertLiterals(
      MachineBasicBlock::iterator InsertPos,
      const std::vector<unsigned> &Literals) const {
    MachineBasicBlock *MBB = InsertPos->getParent();
    for (unsigned i = 0, e = Literals.size(); i < e; i+=2) {
      unsigned LiteralPair0 = Literals[i];
      unsigned LiteralPair1 = (i + 1 < e)?Literals[i + 1]:0;
      InsertPos = BuildMI(MBB, InsertPos->getDebugLoc(),
          TII->get(AMDGPU::LITERALS))
          .addImm(LiteralPair0)
          .addImm(LiteralPair1);
    }
    return InsertPos;
  }

  ClauseFile
  MakeALUClause(MachineBasicBlock &MBB, MachineBasicBlock::iterator &I)
      const {
    MachineBasicBlock::iterator ClauseHead = I;
    std::vector<MachineInstr *> ClauseContent;
    I++;
    for (MachineBasicBlock::instr_iterator E = MBB.instr_end(); I != E;) {
      if (IsTrivialInst(I)) {
        ++I;
        continue;
      }
      if (!I->isBundle() && !TII->isALUInstr(I->getOpcode()))
        break;
      std::vector<int64_t> Literals;
      if (I->isBundle()) {
        MachineInstr *DeleteMI = I;
        MachineBasicBlock::instr_iterator BI = I.getInstrIterator();
        while (++BI != E && BI->isBundledWithPred()) {
          BI->unbundleFromPred();
          for (unsigned i = 0, e = BI->getNumOperands(); i != e; ++i) {
            MachineOperand &MO = BI->getOperand(i);
            if (MO.isReg() && MO.isInternalRead())
              MO.setIsInternalRead(false);
          }
          getLiteral(&*BI, Literals);
          ClauseContent.push_back(&*BI);
        }
        I = BI;
        DeleteMI->eraseFromParent();
      } else {
        getLiteral(I, Literals);
        ClauseContent.push_back(I);
        I++;
      }
      for (unsigned i = 0, e = Literals.size(); i < e; i+=2) {
        unsigned literal0 = Literals[i];
        unsigned literal2 = (i + 1 < e)?Literals[i + 1]:0;
        MachineInstr *MILit = BuildMI(MBB, I, I->getDebugLoc(),
            TII->get(AMDGPU::LITERALS))
            .addImm(literal0)
            .addImm(literal2);
        ClauseContent.push_back(MILit);
      }
    }
    assert(ClauseContent.size() < 128 && "ALU clause is too big");
    ClauseHead->getOperand(7).setImm(ClauseContent.size() - 1);
    return ClauseFile(ClauseHead, std::move(ClauseContent));
  }

  void
  EmitFetchClause(MachineBasicBlock::iterator InsertPos, ClauseFile &Clause,
      unsigned &CfCount) {
    CounterPropagateAddr(Clause.first, CfCount);
    MachineBasicBlock *BB = Clause.first->getParent();
    BuildMI(BB, InsertPos->getDebugLoc(), TII->get(AMDGPU::FETCH_CLAUSE))
        .addImm(CfCount);
    for (unsigned i = 0, e = Clause.second.size(); i < e; ++i) {
      BB->splice(InsertPos, BB, Clause.second[i]);
    }
    CfCount += 2 * Clause.second.size();
  }

  void
  EmitALUClause(MachineBasicBlock::iterator InsertPos, ClauseFile &Clause,
      unsigned &CfCount) {
    Clause.first->getOperand(0).setImm(0);
    CounterPropagateAddr(Clause.first, CfCount);
    MachineBasicBlock *BB = Clause.first->getParent();
    BuildMI(BB, InsertPos->getDebugLoc(), TII->get(AMDGPU::ALU_CLAUSE))
        .addImm(CfCount);
    for (unsigned i = 0, e = Clause.second.size(); i < e; ++i) {
      BB->splice(InsertPos, BB, Clause.second[i]);
    }
    CfCount += Clause.second.size();
  }

  void CounterPropagateAddr(MachineInstr *MI, unsigned Addr) const {
    MI->getOperand(0).setImm(Addr + MI->getOperand(0).getImm());
  }
  void CounterPropagateAddr(const std::set<MachineInstr *> &MIs,
                            unsigned Addr) const {
    for (MachineInstr *MI : MIs) {
      CounterPropagateAddr(MI, Addr);
    }
  }

public:
  R600ControlFlowFinalizer(TargetMachine &tm)
      : MachineFunctionPass(ID), TII(nullptr), TRI(nullptr), ST(nullptr) {}

  bool runOnMachineFunction(MachineFunction &MF) override {
    ST = &MF.getSubtarget<AMDGPUSubtarget>();
    MaxFetchInst = ST->getTexVTXClauseSize();
    TII = static_cast<const R600InstrInfo *>(ST->getInstrInfo());
    TRI = static_cast<const R600RegisterInfo *>(ST->getRegisterInfo());
    R600MachineFunctionInfo *MFI = MF.getInfo<R600MachineFunctionInfo>();

    CFStack CFStack(ST, MFI->getShaderType());
    for (MachineFunction::iterator MB = MF.begin(), ME = MF.end(); MB != ME;
        ++MB) {
      MachineBasicBlock &MBB = *MB;
      unsigned CfCount = 0;
      std::vector<std::pair<unsigned, std::set<MachineInstr *> > > LoopStack;
      std::vector<MachineInstr * > IfThenElseStack;
      if (MFI->getShaderType() == ShaderType::VERTEX) {
        BuildMI(MBB, MBB.begin(), MBB.findDebugLoc(MBB.begin()),
            getHWInstrDesc(CF_CALL_FS));
        CfCount++;
      }
      std::vector<ClauseFile> FetchClauses, AluClauses;
      std::vector<MachineInstr *> LastAlu(1);
      std::vector<MachineInstr *> ToPopAfter;
      
      for (MachineBasicBlock::iterator I = MBB.begin(), E = MBB.end();
          I != E;) {
        if (TII->usesTextureCache(I) || TII->usesVertexCache(I)) {
          DEBUG(dbgs() << CfCount << ":"; I->dump(););
          FetchClauses.push_back(MakeFetchClause(MBB, I));
          CfCount++;
          LastAlu.back() = nullptr;
          continue;
        }

        MachineBasicBlock::iterator MI = I;
        if (MI->getOpcode() != AMDGPU::ENDIF)
          LastAlu.back() = nullptr;
        if (MI->getOpcode() == AMDGPU::CF_ALU)
          LastAlu.back() = MI;
        I++;
        bool RequiresWorkAround =
            CFStack.requiresWorkAroundForInst(MI->getOpcode());
        switch (MI->getOpcode()) {
        case AMDGPU::CF_ALU_PUSH_BEFORE:
          if (RequiresWorkAround) {
            DEBUG(dbgs() << "Applying bug work-around for ALU_PUSH_BEFORE\n");
            BuildMI(MBB, MI, MBB.findDebugLoc(MI), TII->get(AMDGPU::CF_PUSH_EG))
                .addImm(CfCount + 1)
                .addImm(1);
            MI->setDesc(TII->get(AMDGPU::CF_ALU));
            CfCount++;
            CFStack.pushBranch(AMDGPU::CF_PUSH_EG);
          } else
            CFStack.pushBranch(AMDGPU::CF_ALU_PUSH_BEFORE);

        case AMDGPU::CF_ALU:
          I = MI;
          AluClauses.push_back(MakeALUClause(MBB, I));
          DEBUG(dbgs() << CfCount << ":"; MI->dump(););
          CfCount++;
          break;
        case AMDGPU::WHILELOOP: {
          CFStack.pushLoop();
          MachineInstr *MIb = BuildMI(MBB, MI, MBB.findDebugLoc(MI),
              getHWInstrDesc(CF_WHILE_LOOP))
              .addImm(1);
          std::pair<unsigned, std::set<MachineInstr *> > Pair(CfCount,
              std::set<MachineInstr *>());
          Pair.second.insert(MIb);
          LoopStack.push_back(std::move(Pair));
          MI->eraseFromParent();
          CfCount++;
          break;
        }
        case AMDGPU::ENDLOOP: {
          CFStack.popLoop();
          std::pair<unsigned, std::set<MachineInstr *> > Pair =
              std::move(LoopStack.back());
          LoopStack.pop_back();
          CounterPropagateAddr(Pair.second, CfCount);
          BuildMI(MBB, MI, MBB.findDebugLoc(MI), getHWInstrDesc(CF_END_LOOP))
              .addImm(Pair.first + 1);
          MI->eraseFromParent();
          CfCount++;
          break;
        }
        case AMDGPU::IF_PREDICATE_SET: {
          LastAlu.push_back(nullptr);
          MachineInstr *MIb = BuildMI(MBB, MI, MBB.findDebugLoc(MI),
              getHWInstrDesc(CF_JUMP))
              .addImm(0)
              .addImm(0);
          IfThenElseStack.push_back(MIb);
          DEBUG(dbgs() << CfCount << ":"; MIb->dump(););
          MI->eraseFromParent();
          CfCount++;
          break;
        }
        case AMDGPU::ELSE: {
          MachineInstr * JumpInst = IfThenElseStack.back();
          IfThenElseStack.pop_back();
          CounterPropagateAddr(JumpInst, CfCount);
          MachineInstr *MIb = BuildMI(MBB, MI, MBB.findDebugLoc(MI),
              getHWInstrDesc(CF_ELSE))
              .addImm(0)
              .addImm(0);
          DEBUG(dbgs() << CfCount << ":"; MIb->dump(););
          IfThenElseStack.push_back(MIb);
          MI->eraseFromParent();
          CfCount++;
          break;
        }
        case AMDGPU::ENDIF: {
          CFStack.popBranch();
          if (LastAlu.back()) {
            ToPopAfter.push_back(LastAlu.back());
          } else {
            MachineInstr *MIb = BuildMI(MBB, MI, MBB.findDebugLoc(MI),
                getHWInstrDesc(CF_POP))
                .addImm(CfCount + 1)
                .addImm(1);
            (void)MIb;
            DEBUG(dbgs() << CfCount << ":"; MIb->dump(););
            CfCount++;
          }
          
          MachineInstr *IfOrElseInst = IfThenElseStack.back();
          IfThenElseStack.pop_back();
          CounterPropagateAddr(IfOrElseInst, CfCount);
          IfOrElseInst->getOperand(1).setImm(1);
          LastAlu.pop_back();
          MI->eraseFromParent();
          break;
        }
        case AMDGPU::BREAK: {
          CfCount ++;
          MachineInstr *MIb = BuildMI(MBB, MI, MBB.findDebugLoc(MI),
              getHWInstrDesc(CF_LOOP_BREAK))
              .addImm(0);
          LoopStack.back().second.insert(MIb);
          MI->eraseFromParent();
          break;
        }
        case AMDGPU::CONTINUE: {
          MachineInstr *MIb = BuildMI(MBB, MI, MBB.findDebugLoc(MI),
              getHWInstrDesc(CF_LOOP_CONTINUE))
              .addImm(0);
          LoopStack.back().second.insert(MIb);
          MI->eraseFromParent();
          CfCount++;
          break;
        }
        case AMDGPU::RETURN: {
          BuildMI(MBB, MI, MBB.findDebugLoc(MI), getHWInstrDesc(CF_END));
          CfCount++;
          MI->eraseFromParent();
          if (CfCount % 2) {
            BuildMI(MBB, I, MBB.findDebugLoc(MI), TII->get(AMDGPU::PAD));
            CfCount++;
          }
          for (unsigned i = 0, e = FetchClauses.size(); i < e; i++)
            EmitFetchClause(I, FetchClauses[i], CfCount);
          for (unsigned i = 0, e = AluClauses.size(); i < e; i++)
            EmitALUClause(I, AluClauses[i], CfCount);
        }
        default:
          if (TII->isExport(MI->getOpcode())) {
            DEBUG(dbgs() << CfCount << ":"; MI->dump(););
            CfCount++;
          }
          break;
        }
      }
      for (unsigned i = 0, e = ToPopAfter.size(); i < e; ++i) {
        MachineInstr *Alu = ToPopAfter[i];
        BuildMI(MBB, Alu, MBB.findDebugLoc((MachineBasicBlock::iterator)Alu),
            TII->get(AMDGPU::CF_ALU_POP_AFTER))
            .addImm(Alu->getOperand(0).getImm())
            .addImm(Alu->getOperand(1).getImm())
            .addImm(Alu->getOperand(2).getImm())
            .addImm(Alu->getOperand(3).getImm())
            .addImm(Alu->getOperand(4).getImm())
            .addImm(Alu->getOperand(5).getImm())
            .addImm(Alu->getOperand(6).getImm())
            .addImm(Alu->getOperand(7).getImm())
            .addImm(Alu->getOperand(8).getImm());
        Alu->eraseFromParent();
      }
      MFI->StackSize = CFStack.MaxStackSize;
    }

    return false;
  }

  const char *getPassName() const override {
    return "R600 Control Flow Finalizer Pass";
  }
};

char R600ControlFlowFinalizer::ID = 0;

} // end anonymous namespace


llvm::FunctionPass *llvm::createR600ControlFlowFinalizer(TargetMachine &TM) {
  return new R600ControlFlowFinalizer(TM);
}