1 //===-- R600ControlFlowFinalizer.cpp - Finalize Control Flow Inst----------===//
2 //
3 //                     The LLVM Compiler Infrastructure
4 //
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
7 //
8 //===----------------------------------------------------------------------===//
9 //
10 /// \file
11 /// This pass compute turns all control flow pseudo instructions into native one
12 /// computing their address on the fly ; it also sets STACK_SIZE info.
13 //===----------------------------------------------------------------------===//
14 
15 #include "llvm/Support/Debug.h"
16 #include "AMDGPU.h"
17 #include "AMDGPUSubtarget.h"
18 #include "R600Defines.h"
19 #include "R600InstrInfo.h"
20 #include "R600MachineFunctionInfo.h"
21 #include "R600RegisterInfo.h"
22 #include "llvm/CodeGen/MachineFunctionPass.h"
23 #include "llvm/CodeGen/MachineInstrBuilder.h"
24 #include "llvm/CodeGen/MachineRegisterInfo.h"
25 #include "llvm/Support/raw_ostream.h"
26 
27 using namespace llvm;
28 
29 #define DEBUG_TYPE "r600cf"
30 
31 namespace {
32 
33 struct CFStack {
34 
35   enum StackItem {
36     ENTRY = 0,
37     SUB_ENTRY = 1,
38     FIRST_NON_WQM_PUSH = 2,
39     FIRST_NON_WQM_PUSH_W_FULL_ENTRY = 3
40   };
41 
42   const R600Subtarget *ST;
43   std::vector<StackItem> BranchStack;
44   std::vector<StackItem> LoopStack;
45   unsigned MaxStackSize;
46   unsigned CurrentEntries;
47   unsigned CurrentSubEntries;
48 
49   CFStack(const R600Subtarget *st, CallingConv::ID cc) : ST(st),
50       // We need to reserve a stack entry for CALL_FS in vertex shaders.
51       MaxStackSize(cc == CallingConv::AMDGPU_VS ? 1 : 0),
52       CurrentEntries(0), CurrentSubEntries(0) { }
53 
54   unsigned getLoopDepth();
55   bool branchStackContains(CFStack::StackItem);
56   bool requiresWorkAroundForInst(unsigned Opcode);
57   unsigned getSubEntrySize(CFStack::StackItem Item);
58   void updateMaxStackSize();
59   void pushBranch(unsigned Opcode, bool isWQM = false);
60   void pushLoop();
61   void popBranch();
62   void popLoop();
63 };
64 
65 unsigned CFStack::getLoopDepth() {
66   return LoopStack.size();
67 }
68 
69 bool CFStack::branchStackContains(CFStack::StackItem Item) {
70   for (std::vector<CFStack::StackItem>::const_iterator I = BranchStack.begin(),
71        E = BranchStack.end(); I != E; ++I) {
72     if (*I == Item)
73       return true;
74   }
75   return false;
76 }
77 
78 bool CFStack::requiresWorkAroundForInst(unsigned Opcode) {
79   if (Opcode == AMDGPU::CF_ALU_PUSH_BEFORE && ST->hasCaymanISA() &&
80       getLoopDepth() > 1)
81     return true;
82 
83   if (!ST->hasCFAluBug())
84     return false;
85 
86   switch(Opcode) {
87   default: return false;
88   case AMDGPU::CF_ALU_PUSH_BEFORE:
89   case AMDGPU::CF_ALU_ELSE_AFTER:
90   case AMDGPU::CF_ALU_BREAK:
91   case AMDGPU::CF_ALU_CONTINUE:
92     if (CurrentSubEntries == 0)
93       return false;
94     if (ST->getWavefrontSize() == 64) {
95       // We are being conservative here.  We only require this work-around if
96       // CurrentSubEntries > 3 &&
97       // (CurrentSubEntries % 4 == 3 || CurrentSubEntries % 4 == 0)
98       //
99       // We have to be conservative, because we don't know for certain that
100       // our stack allocation algorithm for Evergreen/NI is correct.  Applying this
101       // work-around when CurrentSubEntries > 3 allows us to over-allocate stack
102       // resources without any problems.
103       return CurrentSubEntries > 3;
104     } else {
105       assert(ST->getWavefrontSize() == 32);
106       // We are being conservative here.  We only require the work-around if
107       // CurrentSubEntries > 7 &&
108       // (CurrentSubEntries % 8 == 7 || CurrentSubEntries % 8 == 0)
109       // See the comment on the wavefront size == 64 case for why we are
110       // being conservative.
111       return CurrentSubEntries > 7;
112     }
113   }
114 }
115 
116 unsigned CFStack::getSubEntrySize(CFStack::StackItem Item) {
117   switch(Item) {
118   default:
119     return 0;
120   case CFStack::FIRST_NON_WQM_PUSH:
121   assert(!ST->hasCaymanISA());
122   if (ST->getGeneration() <= R600Subtarget::R700) {
123     // +1 For the push operation.
124     // +2 Extra space required.
125     return 3;
126   } else {
127     // Some documentation says that this is not necessary on Evergreen,
128     // but experimentation has show that we need to allocate 1 extra
129     // sub-entry for the first non-WQM push.
130     // +1 For the push operation.
131     // +1 Extra space required.
132     return 2;
133   }
134   case CFStack::FIRST_NON_WQM_PUSH_W_FULL_ENTRY:
135     assert(ST->getGeneration() >= R600Subtarget::EVERGREEN);
136     // +1 For the push operation.
137     // +1 Extra space required.
138     return 2;
139   case CFStack::SUB_ENTRY:
140     return 1;
141   }
142 }
143 
144 void CFStack::updateMaxStackSize() {
145   unsigned CurrentStackSize =
146       CurrentEntries + (alignTo(CurrentSubEntries, 4) / 4);
147   MaxStackSize = std::max(CurrentStackSize, MaxStackSize);
148 }
149 
150 void CFStack::pushBranch(unsigned Opcode, bool isWQM) {
151   CFStack::StackItem Item = CFStack::ENTRY;
152   switch(Opcode) {
153   case AMDGPU::CF_PUSH_EG:
154   case AMDGPU::CF_ALU_PUSH_BEFORE:
155     if (!isWQM) {
156       if (!ST->hasCaymanISA() &&
157           !branchStackContains(CFStack::FIRST_NON_WQM_PUSH))
158         Item = CFStack::FIRST_NON_WQM_PUSH;  // May not be required on Evergreen/NI
159                                              // See comment in
160                                              // CFStack::getSubEntrySize()
161       else if (CurrentEntries > 0 &&
162                ST->getGeneration() > R600Subtarget::EVERGREEN &&
163                !ST->hasCaymanISA() &&
164                !branchStackContains(CFStack::FIRST_NON_WQM_PUSH_W_FULL_ENTRY))
165         Item = CFStack::FIRST_NON_WQM_PUSH_W_FULL_ENTRY;
166       else
167         Item = CFStack::SUB_ENTRY;
168     } else
169       Item = CFStack::ENTRY;
170     break;
171   }
172   BranchStack.push_back(Item);
173   if (Item == CFStack::ENTRY)
174     CurrentEntries++;
175   else
176     CurrentSubEntries += getSubEntrySize(Item);
177   updateMaxStackSize();
178 }
179 
180 void CFStack::pushLoop() {
181   LoopStack.push_back(CFStack::ENTRY);
182   CurrentEntries++;
183   updateMaxStackSize();
184 }
185 
186 void CFStack::popBranch() {
187   CFStack::StackItem Top = BranchStack.back();
188   if (Top == CFStack::ENTRY)
189     CurrentEntries--;
190   else
191     CurrentSubEntries-= getSubEntrySize(Top);
192   BranchStack.pop_back();
193 }
194 
195 void CFStack::popLoop() {
196   CurrentEntries--;
197   LoopStack.pop_back();
198 }
199 
200 class R600ControlFlowFinalizer : public MachineFunctionPass {
201 
202 private:
203   typedef std::pair<MachineInstr *, std::vector<MachineInstr *> > ClauseFile;
204 
205   enum ControlFlowInstruction {
206     CF_TC,
207     CF_VC,
208     CF_CALL_FS,
209     CF_WHILE_LOOP,
210     CF_END_LOOP,
211     CF_LOOP_BREAK,
212     CF_LOOP_CONTINUE,
213     CF_JUMP,
214     CF_ELSE,
215     CF_POP,
216     CF_END
217   };
218 
219   static char ID;
220   const R600InstrInfo *TII;
221   const R600RegisterInfo *TRI;
222   unsigned MaxFetchInst;
223   const R600Subtarget *ST;
224 
225   bool IsTrivialInst(MachineInstr *MI) const {
226     switch (MI->getOpcode()) {
227     case AMDGPU::KILL:
228     case AMDGPU::RETURN:
229       return true;
230     default:
231       return false;
232     }
233   }
234 
235   const MCInstrDesc &getHWInstrDesc(ControlFlowInstruction CFI) const {
236     unsigned Opcode = 0;
237     bool isEg = (ST->getGeneration() >= R600Subtarget::EVERGREEN);
238     switch (CFI) {
239     case CF_TC:
240       Opcode = isEg ? AMDGPU::CF_TC_EG : AMDGPU::CF_TC_R600;
241       break;
242     case CF_VC:
243       Opcode = isEg ? AMDGPU::CF_VC_EG : AMDGPU::CF_VC_R600;
244       break;
245     case CF_CALL_FS:
246       Opcode = isEg ? AMDGPU::CF_CALL_FS_EG : AMDGPU::CF_CALL_FS_R600;
247       break;
248     case CF_WHILE_LOOP:
249       Opcode = isEg ? AMDGPU::WHILE_LOOP_EG : AMDGPU::WHILE_LOOP_R600;
250       break;
251     case CF_END_LOOP:
252       Opcode = isEg ? AMDGPU::END_LOOP_EG : AMDGPU::END_LOOP_R600;
253       break;
254     case CF_LOOP_BREAK:
255       Opcode = isEg ? AMDGPU::LOOP_BREAK_EG : AMDGPU::LOOP_BREAK_R600;
256       break;
257     case CF_LOOP_CONTINUE:
258       Opcode = isEg ? AMDGPU::CF_CONTINUE_EG : AMDGPU::CF_CONTINUE_R600;
259       break;
260     case CF_JUMP:
261       Opcode = isEg ? AMDGPU::CF_JUMP_EG : AMDGPU::CF_JUMP_R600;
262       break;
263     case CF_ELSE:
264       Opcode = isEg ? AMDGPU::CF_ELSE_EG : AMDGPU::CF_ELSE_R600;
265       break;
266     case CF_POP:
267       Opcode = isEg ? AMDGPU::POP_EG : AMDGPU::POP_R600;
268       break;
269     case CF_END:
270       if (ST->hasCaymanISA()) {
271         Opcode = AMDGPU::CF_END_CM;
272         break;
273       }
274       Opcode = isEg ? AMDGPU::CF_END_EG : AMDGPU::CF_END_R600;
275       break;
276     }
277     assert (Opcode && "No opcode selected");
278     return TII->get(Opcode);
279   }
280 
281   bool isCompatibleWithClause(const MachineInstr *MI,
282       std::set<unsigned> &DstRegs) const {
283     unsigned DstMI, SrcMI;
284     for (MachineInstr::const_mop_iterator I = MI->operands_begin(),
285         E = MI->operands_end(); I != E; ++I) {
286       const MachineOperand &MO = *I;
287       if (!MO.isReg())
288         continue;
289       if (MO.isDef()) {
290         unsigned Reg = MO.getReg();
291         if (AMDGPU::R600_Reg128RegClass.contains(Reg))
292           DstMI = Reg;
293         else
294           DstMI = TRI->getMatchingSuperReg(Reg,
295               TRI->getSubRegFromChannel(TRI->getHWRegChan(Reg)),
296               &AMDGPU::R600_Reg128RegClass);
297       }
298       if (MO.isUse()) {
299         unsigned Reg = MO.getReg();
300         if (AMDGPU::R600_Reg128RegClass.contains(Reg))
301           SrcMI = Reg;
302         else
303           SrcMI = TRI->getMatchingSuperReg(Reg,
304               TRI->getSubRegFromChannel(TRI->getHWRegChan(Reg)),
305               &AMDGPU::R600_Reg128RegClass);
306       }
307     }
308     if ((DstRegs.find(SrcMI) == DstRegs.end())) {
309       DstRegs.insert(DstMI);
310       return true;
311     } else
312       return false;
313   }
314 
315   ClauseFile
316   MakeFetchClause(MachineBasicBlock &MBB, MachineBasicBlock::iterator &I)
317       const {
318     MachineBasicBlock::iterator ClauseHead = I;
319     std::vector<MachineInstr *> ClauseContent;
320     unsigned AluInstCount = 0;
321     bool IsTex = TII->usesTextureCache(ClauseHead);
322     std::set<unsigned> DstRegs;
323     for (MachineBasicBlock::iterator E = MBB.end(); I != E; ++I) {
324       if (IsTrivialInst(I))
325         continue;
326       if (AluInstCount >= MaxFetchInst)
327         break;
328       if ((IsTex && !TII->usesTextureCache(I)) ||
329           (!IsTex && !TII->usesVertexCache(I)))
330         break;
331       if (!isCompatibleWithClause(I, DstRegs))
332         break;
333       AluInstCount ++;
334       ClauseContent.push_back(I);
335     }
336     MachineInstr *MIb = BuildMI(MBB, ClauseHead, MBB.findDebugLoc(ClauseHead),
337         getHWInstrDesc(IsTex?CF_TC:CF_VC))
338         .addImm(0) // ADDR
339         .addImm(AluInstCount - 1); // COUNT
340     return ClauseFile(MIb, std::move(ClauseContent));
341   }
342 
343   void getLiteral(MachineInstr *MI, std::vector<MachineOperand *> &Lits) const {
344     static const unsigned LiteralRegs[] = {
345       AMDGPU::ALU_LITERAL_X,
346       AMDGPU::ALU_LITERAL_Y,
347       AMDGPU::ALU_LITERAL_Z,
348       AMDGPU::ALU_LITERAL_W
349     };
350     const SmallVector<std::pair<MachineOperand *, int64_t>, 3 > Srcs =
351         TII->getSrcs(MI);
352     for (const auto &Src:Srcs) {
353       if (Src.first->getReg() != AMDGPU::ALU_LITERAL_X)
354         continue;
355       int64_t Imm = Src.second;
356       std::vector<MachineOperand*>::iterator It =
357           std::find_if(Lits.begin(), Lits.end(),
358                     [&](MachineOperand* val)
359                         { return val->isImm() && (val->getImm() == Imm);});
360 
361       // Get corresponding Operand
362       MachineOperand &Operand = MI->getOperand(
363               TII->getOperandIdx(MI->getOpcode(), AMDGPU::OpName::literal));
364 
365       if (It != Lits.end()) {
366         // Reuse existing literal reg
367         unsigned Index = It - Lits.begin();
368         Src.first->setReg(LiteralRegs[Index]);
369       } else {
370         // Allocate new literal reg
371         assert(Lits.size() < 4 && "Too many literals in Instruction Group");
372         Src.first->setReg(LiteralRegs[Lits.size()]);
373         Lits.push_back(&Operand);
374       }
375     }
376   }
377 
378   MachineBasicBlock::iterator insertLiterals(
379       MachineBasicBlock::iterator InsertPos,
380       const std::vector<unsigned> &Literals) const {
381     MachineBasicBlock *MBB = InsertPos->getParent();
382     for (unsigned i = 0, e = Literals.size(); i < e; i+=2) {
383       unsigned LiteralPair0 = Literals[i];
384       unsigned LiteralPair1 = (i + 1 < e)?Literals[i + 1]:0;
385       InsertPos = BuildMI(MBB, InsertPos->getDebugLoc(),
386           TII->get(AMDGPU::LITERALS))
387           .addImm(LiteralPair0)
388           .addImm(LiteralPair1);
389     }
390     return InsertPos;
391   }
392 
393   ClauseFile
394   MakeALUClause(MachineBasicBlock &MBB, MachineBasicBlock::iterator &I)
395       const {
396     MachineBasicBlock::iterator ClauseHead = I;
397     std::vector<MachineInstr *> ClauseContent;
398     I++;
399     for (MachineBasicBlock::instr_iterator E = MBB.instr_end(); I != E;) {
400       if (IsTrivialInst(I)) {
401         ++I;
402         continue;
403       }
404       if (!I->isBundle() && !TII->isALUInstr(I->getOpcode()))
405         break;
406       std::vector<MachineOperand *>Literals;
407       if (I->isBundle()) {
408         MachineInstr *DeleteMI = I;
409         MachineBasicBlock::instr_iterator BI = I.getInstrIterator();
410         while (++BI != E && BI->isBundledWithPred()) {
411           BI->unbundleFromPred();
412           for (MachineOperand &MO : BI->operands()) {
413             if (MO.isReg() && MO.isInternalRead())
414               MO.setIsInternalRead(false);
415           }
416           getLiteral(&*BI, Literals);
417           ClauseContent.push_back(&*BI);
418         }
419         I = BI;
420         DeleteMI->eraseFromParent();
421       } else {
422         getLiteral(I, Literals);
423         ClauseContent.push_back(I);
424         I++;
425       }
426       for (unsigned i = 0, e = Literals.size(); i < e; i += 2) {
427         MachineInstrBuilder MILit = BuildMI(MBB, I, I->getDebugLoc(),
428             TII->get(AMDGPU::LITERALS));
429         if (Literals[i]->isImm()) {
430             MILit.addImm(Literals[i]->getImm());
431         } else {
432             MILit.addGlobalAddress(Literals[i]->getGlobal(),
433                                    Literals[i]->getOffset());
434         }
435         if (i + 1 < e) {
436           if (Literals[i + 1]->isImm()) {
437             MILit.addImm(Literals[i + 1]->getImm());
438           } else {
439             MILit.addGlobalAddress(Literals[i + 1]->getGlobal(),
440                                    Literals[i + 1]->getOffset());
441           }
442         } else
443           MILit.addImm(0);
444         ClauseContent.push_back(MILit);
445       }
446     }
447     assert(ClauseContent.size() < 128 && "ALU clause is too big");
448     ClauseHead->getOperand(7).setImm(ClauseContent.size() - 1);
449     return ClauseFile(ClauseHead, std::move(ClauseContent));
450   }
451 
452   void
453   EmitFetchClause(MachineBasicBlock::iterator InsertPos, ClauseFile &Clause,
454       unsigned &CfCount) {
455     CounterPropagateAddr(Clause.first, CfCount);
456     MachineBasicBlock *BB = Clause.first->getParent();
457     BuildMI(BB, InsertPos->getDebugLoc(), TII->get(AMDGPU::FETCH_CLAUSE))
458         .addImm(CfCount);
459     for (unsigned i = 0, e = Clause.second.size(); i < e; ++i) {
460       BB->splice(InsertPos, BB, Clause.second[i]);
461     }
462     CfCount += 2 * Clause.second.size();
463   }
464 
465   void
466   EmitALUClause(MachineBasicBlock::iterator InsertPos, ClauseFile &Clause,
467       unsigned &CfCount) {
468     Clause.first->getOperand(0).setImm(0);
469     CounterPropagateAddr(Clause.first, CfCount);
470     MachineBasicBlock *BB = Clause.first->getParent();
471     BuildMI(BB, InsertPos->getDebugLoc(), TII->get(AMDGPU::ALU_CLAUSE))
472         .addImm(CfCount);
473     for (unsigned i = 0, e = Clause.second.size(); i < e; ++i) {
474       BB->splice(InsertPos, BB, Clause.second[i]);
475     }
476     CfCount += Clause.second.size();
477   }
478 
479   void CounterPropagateAddr(MachineInstr *MI, unsigned Addr) const {
480     MI->getOperand(0).setImm(Addr + MI->getOperand(0).getImm());
481   }
482   void CounterPropagateAddr(const std::set<MachineInstr *> &MIs,
483                             unsigned Addr) const {
484     for (MachineInstr *MI : MIs) {
485       CounterPropagateAddr(MI, Addr);
486     }
487   }
488 
489 public:
490   R600ControlFlowFinalizer(TargetMachine &tm)
491       : MachineFunctionPass(ID), TII(nullptr), TRI(nullptr), ST(nullptr) {}
492 
493   bool runOnMachineFunction(MachineFunction &MF) override {
494     ST = &MF.getSubtarget<R600Subtarget>();
495     MaxFetchInst = ST->getTexVTXClauseSize();
496     TII = ST->getInstrInfo();
497     TRI = ST->getRegisterInfo();
498 
499     R600MachineFunctionInfo *MFI = MF.getInfo<R600MachineFunctionInfo>();
500 
501     CFStack CFStack(ST, MF.getFunction()->getCallingConv());
502     for (MachineFunction::iterator MB = MF.begin(), ME = MF.end(); MB != ME;
503         ++MB) {
504       MachineBasicBlock &MBB = *MB;
505       unsigned CfCount = 0;
506       std::vector<std::pair<unsigned, std::set<MachineInstr *> > > LoopStack;
507       std::vector<MachineInstr * > IfThenElseStack;
508       if (MF.getFunction()->getCallingConv() == CallingConv::AMDGPU_VS) {
509         BuildMI(MBB, MBB.begin(), MBB.findDebugLoc(MBB.begin()),
510             getHWInstrDesc(CF_CALL_FS));
511         CfCount++;
512       }
513       std::vector<ClauseFile> FetchClauses, AluClauses;
514       std::vector<MachineInstr *> LastAlu(1);
515       std::vector<MachineInstr *> ToPopAfter;
516 
517       for (MachineBasicBlock::iterator I = MBB.begin(), E = MBB.end();
518           I != E;) {
519         if (TII->usesTextureCache(I) || TII->usesVertexCache(I)) {
520           DEBUG(dbgs() << CfCount << ":"; I->dump(););
521           FetchClauses.push_back(MakeFetchClause(MBB, I));
522           CfCount++;
523           LastAlu.back() = nullptr;
524           continue;
525         }
526 
527         MachineBasicBlock::iterator MI = I;
528         if (MI->getOpcode() != AMDGPU::ENDIF)
529           LastAlu.back() = nullptr;
530         if (MI->getOpcode() == AMDGPU::CF_ALU)
531           LastAlu.back() = MI;
532         I++;
533         bool RequiresWorkAround =
534             CFStack.requiresWorkAroundForInst(MI->getOpcode());
535         switch (MI->getOpcode()) {
536         case AMDGPU::CF_ALU_PUSH_BEFORE:
537           if (RequiresWorkAround) {
538             DEBUG(dbgs() << "Applying bug work-around for ALU_PUSH_BEFORE\n");
539             BuildMI(MBB, MI, MBB.findDebugLoc(MI), TII->get(AMDGPU::CF_PUSH_EG))
540                 .addImm(CfCount + 1)
541                 .addImm(1);
542             MI->setDesc(TII->get(AMDGPU::CF_ALU));
543             CfCount++;
544             CFStack.pushBranch(AMDGPU::CF_PUSH_EG);
545           } else
546             CFStack.pushBranch(AMDGPU::CF_ALU_PUSH_BEFORE);
547 
548         case AMDGPU::CF_ALU:
549           I = MI;
550           AluClauses.push_back(MakeALUClause(MBB, I));
551           DEBUG(dbgs() << CfCount << ":"; MI->dump(););
552           CfCount++;
553           break;
554         case AMDGPU::WHILELOOP: {
555           CFStack.pushLoop();
556           MachineInstr *MIb = BuildMI(MBB, MI, MBB.findDebugLoc(MI),
557               getHWInstrDesc(CF_WHILE_LOOP))
558               .addImm(1);
559           std::pair<unsigned, std::set<MachineInstr *> > Pair(CfCount,
560               std::set<MachineInstr *>());
561           Pair.second.insert(MIb);
562           LoopStack.push_back(std::move(Pair));
563           MI->eraseFromParent();
564           CfCount++;
565           break;
566         }
567         case AMDGPU::ENDLOOP: {
568           CFStack.popLoop();
569           std::pair<unsigned, std::set<MachineInstr *> > Pair =
570               std::move(LoopStack.back());
571           LoopStack.pop_back();
572           CounterPropagateAddr(Pair.second, CfCount);
573           BuildMI(MBB, MI, MBB.findDebugLoc(MI), getHWInstrDesc(CF_END_LOOP))
574               .addImm(Pair.first + 1);
575           MI->eraseFromParent();
576           CfCount++;
577           break;
578         }
579         case AMDGPU::IF_PREDICATE_SET: {
580           LastAlu.push_back(nullptr);
581           MachineInstr *MIb = BuildMI(MBB, MI, MBB.findDebugLoc(MI),
582               getHWInstrDesc(CF_JUMP))
583               .addImm(0)
584               .addImm(0);
585           IfThenElseStack.push_back(MIb);
586           DEBUG(dbgs() << CfCount << ":"; MIb->dump(););
587           MI->eraseFromParent();
588           CfCount++;
589           break;
590         }
591         case AMDGPU::ELSE: {
592           MachineInstr * JumpInst = IfThenElseStack.back();
593           IfThenElseStack.pop_back();
594           CounterPropagateAddr(JumpInst, CfCount);
595           MachineInstr *MIb = BuildMI(MBB, MI, MBB.findDebugLoc(MI),
596               getHWInstrDesc(CF_ELSE))
597               .addImm(0)
598               .addImm(0);
599           DEBUG(dbgs() << CfCount << ":"; MIb->dump(););
600           IfThenElseStack.push_back(MIb);
601           MI->eraseFromParent();
602           CfCount++;
603           break;
604         }
605         case AMDGPU::ENDIF: {
606           CFStack.popBranch();
607           if (LastAlu.back()) {
608             ToPopAfter.push_back(LastAlu.back());
609           } else {
610             MachineInstr *MIb = BuildMI(MBB, MI, MBB.findDebugLoc(MI),
611                 getHWInstrDesc(CF_POP))
612                 .addImm(CfCount + 1)
613                 .addImm(1);
614             (void)MIb;
615             DEBUG(dbgs() << CfCount << ":"; MIb->dump(););
616             CfCount++;
617           }
618 
619           MachineInstr *IfOrElseInst = IfThenElseStack.back();
620           IfThenElseStack.pop_back();
621           CounterPropagateAddr(IfOrElseInst, CfCount);
622           IfOrElseInst->getOperand(1).setImm(1);
623           LastAlu.pop_back();
624           MI->eraseFromParent();
625           break;
626         }
627         case AMDGPU::BREAK: {
628           CfCount ++;
629           MachineInstr *MIb = BuildMI(MBB, MI, MBB.findDebugLoc(MI),
630               getHWInstrDesc(CF_LOOP_BREAK))
631               .addImm(0);
632           LoopStack.back().second.insert(MIb);
633           MI->eraseFromParent();
634           break;
635         }
636         case AMDGPU::CONTINUE: {
637           MachineInstr *MIb = BuildMI(MBB, MI, MBB.findDebugLoc(MI),
638               getHWInstrDesc(CF_LOOP_CONTINUE))
639               .addImm(0);
640           LoopStack.back().second.insert(MIb);
641           MI->eraseFromParent();
642           CfCount++;
643           break;
644         }
645         case AMDGPU::RETURN: {
646           BuildMI(MBB, MI, MBB.findDebugLoc(MI), getHWInstrDesc(CF_END));
647           CfCount++;
648           if (CfCount % 2) {
649             BuildMI(MBB, I, MBB.findDebugLoc(MI), TII->get(AMDGPU::PAD));
650             CfCount++;
651           }
652           MI->eraseFromParent();
653           for (unsigned i = 0, e = FetchClauses.size(); i < e; i++)
654             EmitFetchClause(I, FetchClauses[i], CfCount);
655           for (unsigned i = 0, e = AluClauses.size(); i < e; i++)
656             EmitALUClause(I, AluClauses[i], CfCount);
657           break;
658         }
659         default:
660           if (TII->isExport(MI->getOpcode())) {
661             DEBUG(dbgs() << CfCount << ":"; MI->dump(););
662             CfCount++;
663           }
664           break;
665         }
666       }
667       for (unsigned i = 0, e = ToPopAfter.size(); i < e; ++i) {
668         MachineInstr *Alu = ToPopAfter[i];
669         BuildMI(MBB, Alu, MBB.findDebugLoc((MachineBasicBlock::iterator)Alu),
670             TII->get(AMDGPU::CF_ALU_POP_AFTER))
671             .addImm(Alu->getOperand(0).getImm())
672             .addImm(Alu->getOperand(1).getImm())
673             .addImm(Alu->getOperand(2).getImm())
674             .addImm(Alu->getOperand(3).getImm())
675             .addImm(Alu->getOperand(4).getImm())
676             .addImm(Alu->getOperand(5).getImm())
677             .addImm(Alu->getOperand(6).getImm())
678             .addImm(Alu->getOperand(7).getImm())
679             .addImm(Alu->getOperand(8).getImm());
680         Alu->eraseFromParent();
681       }
682       MFI->StackSize = CFStack.MaxStackSize;
683     }
684 
685     return false;
686   }
687 
688   const char *getPassName() const override {
689     return "R600 Control Flow Finalizer Pass";
690   }
691 };
692 
693 char R600ControlFlowFinalizer::ID = 0;
694 
695 } // end anonymous namespace
696 
697 
698 llvm::FunctionPass *llvm::createR600ControlFlowFinalizer(TargetMachine &TM) {
699   return new R600ControlFlowFinalizer(TM);
700 }
701