xref: /llvm-project/llvm/lib/Target/PowerPC/PPCMIPeephole.cpp (revision f71cb9dbb739bb58ce7e52e49fe384ff2ff11687)
1 //===-------------- PPCMIPeephole.cpp - MI Peephole Cleanups -------------===//
2 //
3 // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
4 // See https://llvm.org/LICENSE.txt for license information.
5 // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
6 //
7 //===---------------------------------------------------------------------===//
8 //
9 // This pass performs peephole optimizations to clean up ugly code
10 // sequences at the MachineInstruction layer.  It runs at the end of
11 // the SSA phases, following VSX swap removal.  A pass of dead code
12 // elimination follows this one for quick clean-up of any dead
13 // instructions introduced here.  Although we could do this as callbacks
14 // from the generic peephole pass, this would have a couple of bad
15 // effects:  it might remove optimization opportunities for VSX swap
16 // removal, and it would miss cleanups made possible following VSX
17 // swap removal.
18 //
19 // NOTE: We run the verifier after this pass in Asserts/Debug builds so it
20 //       is important to keep the code valid after transformations.
21 //       Common causes of errors stem from violating the contract specified
22 //       by kill flags. Whenever a transformation changes the live range of
23 //       a register, that register should be added to the work list using
24 //       addRegToUpdate(RegsToUpdate, <Reg>). Furthermore, if a transformation
25 //       is changing the definition of a register (i.e. removing the single
26 //       definition of the original vreg), it needs to provide a dummy
27 //       definition of that register using addDummyDef(<MBB>, <Reg>).
28 //===---------------------------------------------------------------------===//
29 
30 #include "MCTargetDesc/PPCMCTargetDesc.h"
31 #include "MCTargetDesc/PPCPredicates.h"
32 #include "PPC.h"
33 #include "PPCInstrInfo.h"
34 #include "PPCMachineFunctionInfo.h"
35 #include "PPCTargetMachine.h"
36 #include "llvm/ADT/Statistic.h"
37 #include "llvm/CodeGen/LiveVariables.h"
38 #include "llvm/CodeGen/MachineBlockFrequencyInfo.h"
39 #include "llvm/CodeGen/MachineDominators.h"
40 #include "llvm/CodeGen/MachineFrameInfo.h"
41 #include "llvm/CodeGen/MachineFunctionPass.h"
42 #include "llvm/CodeGen/MachineInstrBuilder.h"
43 #include "llvm/CodeGen/MachinePostDominators.h"
44 #include "llvm/CodeGen/MachineRegisterInfo.h"
45 #include "llvm/InitializePasses.h"
46 #include "llvm/Support/Debug.h"
47 #include "llvm/Support/DebugCounter.h"
48 
49 using namespace llvm;
50 
51 #define DEBUG_TYPE "ppc-mi-peepholes"
52 
53 STATISTIC(RemoveTOCSave, "Number of TOC saves removed");
54 STATISTIC(MultiTOCSaves,
55           "Number of functions with multiple TOC saves that must be kept");
56 STATISTIC(NumTOCSavesInPrologue, "Number of TOC saves placed in the prologue");
57 STATISTIC(NumEliminatedSExt, "Number of eliminated sign-extensions");
58 STATISTIC(NumEliminatedZExt, "Number of eliminated zero-extensions");
59 STATISTIC(NumOptADDLIs, "Number of optimized ADD instruction fed by LI");
60 STATISTIC(NumConvertedToImmediateForm,
61           "Number of instructions converted to their immediate form");
62 STATISTIC(NumFunctionsEnteredInMIPeephole,
63           "Number of functions entered in PPC MI Peepholes");
64 STATISTIC(NumFixedPointIterations,
65           "Number of fixed-point iterations converting reg-reg instructions "
66           "to reg-imm ones");
67 STATISTIC(NumRotatesCollapsed,
68           "Number of pairs of rotate left, clear left/right collapsed");
69 STATISTIC(NumEXTSWAndSLDICombined,
70           "Number of pairs of EXTSW and SLDI combined as EXTSWSLI");
71 STATISTIC(NumLoadImmZeroFoldedAndRemoved,
72           "Number of LI(8) reg, 0 that are folded to r0 and removed");
73 
74 static cl::opt<bool>
75 FixedPointRegToImm("ppc-reg-to-imm-fixed-point", cl::Hidden, cl::init(true),
76                    cl::desc("Iterate to a fixed point when attempting to "
77                             "convert reg-reg instructions to reg-imm"));
78 
79 static cl::opt<bool>
80 ConvertRegReg("ppc-convert-rr-to-ri", cl::Hidden, cl::init(true),
81               cl::desc("Convert eligible reg+reg instructions to reg+imm"));
82 
83 static cl::opt<bool>
84     EnableSExtElimination("ppc-eliminate-signext",
85                           cl::desc("enable elimination of sign-extensions"),
86                           cl::init(true), cl::Hidden);
87 
88 static cl::opt<bool>
89     EnableZExtElimination("ppc-eliminate-zeroext",
90                           cl::desc("enable elimination of zero-extensions"),
91                           cl::init(true), cl::Hidden);
92 
93 static cl::opt<bool>
94     EnableTrapOptimization("ppc-opt-conditional-trap",
95                            cl::desc("enable optimization of conditional traps"),
96                            cl::init(false), cl::Hidden);
97 
98 DEBUG_COUNTER(
99     PeepholeXToICounter, "ppc-xtoi-peephole",
100     "Controls whether PPC reg+reg to reg+imm peephole is performed on a MI");
101 
102 DEBUG_COUNTER(PeepholePerOpCounter, "ppc-per-op-peephole",
103               "Controls whether PPC per opcode peephole is performed on a MI");
104 
105 namespace {
106 
107 struct PPCMIPeephole : public MachineFunctionPass {
108 
109   static char ID;
110   const PPCInstrInfo *TII;
111   MachineFunction *MF;
112   MachineRegisterInfo *MRI;
113   LiveVariables *LV;
114 
115   PPCMIPeephole() : MachineFunctionPass(ID) {
116     initializePPCMIPeepholePass(*PassRegistry::getPassRegistry());
117   }
118 
119 private:
120   MachineDominatorTree *MDT;
121   MachinePostDominatorTree *MPDT;
122   MachineBlockFrequencyInfo *MBFI;
123   BlockFrequency EntryFreq;
124   SmallSet<Register, 16> RegsToUpdate;
125 
126   // Initialize class variables.
127   void initialize(MachineFunction &MFParm);
128 
129   // Perform peepholes.
130   bool simplifyCode();
131 
132   // Perform peepholes.
133   bool eliminateRedundantCompare();
134   bool eliminateRedundantTOCSaves(std::map<MachineInstr *, bool> &TOCSaves);
135   bool combineSEXTAndSHL(MachineInstr &MI, MachineInstr *&ToErase);
136   bool emitRLDICWhenLoweringJumpTables(MachineInstr &MI,
137                                        MachineInstr *&ToErase);
138   void UpdateTOCSaves(std::map<MachineInstr *, bool> &TOCSaves,
139                       MachineInstr *MI);
140 
141   // A number of transformations will eliminate the definition of a register
142   // as all of its uses will be removed. However, this leaves a register
143   // without a definition for LiveVariables. Such transformations should
144   // use this function to provide a dummy definition of the register that
145   // will simply be removed by DCE.
146   void addDummyDef(MachineBasicBlock &MBB, MachineInstr *At, Register Reg) {
147     BuildMI(MBB, At, At->getDebugLoc(), TII->get(PPC::IMPLICIT_DEF), Reg);
148   }
149   void addRegToUpdateWithLine(Register Reg, int Line);
150   void convertUnprimedAccPHIs(const PPCInstrInfo *TII, MachineRegisterInfo *MRI,
151                               SmallVectorImpl<MachineInstr *> &PHIs,
152                               Register Dst);
153 
154 public:
155 
156   void getAnalysisUsage(AnalysisUsage &AU) const override {
157     AU.addRequired<LiveVariablesWrapperPass>();
158     AU.addRequired<MachineDominatorTreeWrapperPass>();
159     AU.addRequired<MachinePostDominatorTreeWrapperPass>();
160     AU.addRequired<MachineBlockFrequencyInfoWrapperPass>();
161     AU.addPreserved<LiveVariablesWrapperPass>();
162     AU.addPreserved<MachineDominatorTreeWrapperPass>();
163     AU.addPreserved<MachinePostDominatorTreeWrapperPass>();
164     AU.addPreserved<MachineBlockFrequencyInfoWrapperPass>();
165     MachineFunctionPass::getAnalysisUsage(AU);
166   }
167 
168   // Main entry point for this pass.
169   bool runOnMachineFunction(MachineFunction &MF) override {
170     initialize(MF);
171     // At this point, TOC pointer should not be used in a function that uses
172     // PC-Relative addressing.
173     assert((MF.getRegInfo().use_empty(PPC::X2) ||
174             !MF.getSubtarget<PPCSubtarget>().isUsingPCRelativeCalls()) &&
175            "TOC pointer used in a function using PC-Relative addressing!");
176     if (skipFunction(MF.getFunction()))
177       return false;
178     bool Changed = simplifyCode();
179 #ifndef NDEBUG
180     if (Changed)
181       MF.verify(this, "Error in PowerPC MI Peephole optimization, compile with "
182                       "-mllvm -disable-ppc-peephole");
183 #endif
184     return Changed;
185   }
186 };
187 
188 #define addRegToUpdate(R) addRegToUpdateWithLine(R, __LINE__)
189 void PPCMIPeephole::addRegToUpdateWithLine(Register Reg, int Line) {
190   if (!Register::isVirtualRegister(Reg))
191     return;
192   if (RegsToUpdate.insert(Reg).second)
193     LLVM_DEBUG(dbgs() << "Adding register: " << Register::virtReg2Index(Reg)
194                       << " on line " << Line
195                       << " for re-computation of kill flags\n");
196 }
197 
198 // Initialize class variables.
199 void PPCMIPeephole::initialize(MachineFunction &MFParm) {
200   MF = &MFParm;
201   MRI = &MF->getRegInfo();
202   MDT = &getAnalysis<MachineDominatorTreeWrapperPass>().getDomTree();
203   MPDT = &getAnalysis<MachinePostDominatorTreeWrapperPass>().getPostDomTree();
204   MBFI = &getAnalysis<MachineBlockFrequencyInfoWrapperPass>().getMBFI();
205   LV = &getAnalysis<LiveVariablesWrapperPass>().getLV();
206   EntryFreq = MBFI->getEntryFreq();
207   TII = MF->getSubtarget<PPCSubtarget>().getInstrInfo();
208   RegsToUpdate.clear();
209   LLVM_DEBUG(dbgs() << "*** PowerPC MI peephole pass ***\n\n");
210   LLVM_DEBUG(MF->dump());
211 }
212 
213 static MachineInstr *getVRegDefOrNull(MachineOperand *Op,
214                                       MachineRegisterInfo *MRI) {
215   assert(Op && "Invalid Operand!");
216   if (!Op->isReg())
217     return nullptr;
218 
219   Register Reg = Op->getReg();
220   if (!Reg.isVirtual())
221     return nullptr;
222 
223   return MRI->getVRegDef(Reg);
224 }
225 
226 // This function returns number of known zero bits in output of MI
227 // starting from the most significant bit.
228 static unsigned getKnownLeadingZeroCount(const unsigned Reg,
229                                          const PPCInstrInfo *TII,
230                                          const MachineRegisterInfo *MRI) {
231   MachineInstr *MI = MRI->getVRegDef(Reg);
232   unsigned Opcode = MI->getOpcode();
233   if (Opcode == PPC::RLDICL || Opcode == PPC::RLDICL_rec ||
234       Opcode == PPC::RLDCL || Opcode == PPC::RLDCL_rec)
235     return MI->getOperand(3).getImm();
236 
237   if ((Opcode == PPC::RLDIC || Opcode == PPC::RLDIC_rec) &&
238       MI->getOperand(3).getImm() <= 63 - MI->getOperand(2).getImm())
239     return MI->getOperand(3).getImm();
240 
241   if ((Opcode == PPC::RLWINM || Opcode == PPC::RLWINM_rec ||
242        Opcode == PPC::RLWNM || Opcode == PPC::RLWNM_rec ||
243        Opcode == PPC::RLWINM8 || Opcode == PPC::RLWNM8) &&
244       MI->getOperand(3).getImm() <= MI->getOperand(4).getImm())
245     return 32 + MI->getOperand(3).getImm();
246 
247   if (Opcode == PPC::ANDI_rec) {
248     uint16_t Imm = MI->getOperand(2).getImm();
249     return 48 + llvm::countl_zero(Imm);
250   }
251 
252   if (Opcode == PPC::CNTLZW || Opcode == PPC::CNTLZW_rec ||
253       Opcode == PPC::CNTTZW || Opcode == PPC::CNTTZW_rec ||
254       Opcode == PPC::CNTLZW8 || Opcode == PPC::CNTTZW8)
255     // The result ranges from 0 to 32.
256     return 58;
257 
258   if (Opcode == PPC::CNTLZD || Opcode == PPC::CNTLZD_rec ||
259       Opcode == PPC::CNTTZD || Opcode == PPC::CNTTZD_rec)
260     // The result ranges from 0 to 64.
261     return 57;
262 
263   if (Opcode == PPC::LHZ   || Opcode == PPC::LHZX  ||
264       Opcode == PPC::LHZ8  || Opcode == PPC::LHZX8 ||
265       Opcode == PPC::LHZU  || Opcode == PPC::LHZUX ||
266       Opcode == PPC::LHZU8 || Opcode == PPC::LHZUX8)
267     return 48;
268 
269   if (Opcode == PPC::LBZ   || Opcode == PPC::LBZX  ||
270       Opcode == PPC::LBZ8  || Opcode == PPC::LBZX8 ||
271       Opcode == PPC::LBZU  || Opcode == PPC::LBZUX ||
272       Opcode == PPC::LBZU8 || Opcode == PPC::LBZUX8)
273     return 56;
274 
275   if (Opcode == PPC::AND || Opcode == PPC::AND8 || Opcode == PPC::AND_rec ||
276       Opcode == PPC::AND8_rec)
277     return std::max(
278         getKnownLeadingZeroCount(MI->getOperand(1).getReg(), TII, MRI),
279         getKnownLeadingZeroCount(MI->getOperand(2).getReg(), TII, MRI));
280 
281   if (Opcode == PPC::OR || Opcode == PPC::OR8 || Opcode == PPC::XOR ||
282       Opcode == PPC::XOR8 || Opcode == PPC::OR_rec ||
283       Opcode == PPC::OR8_rec || Opcode == PPC::XOR_rec ||
284       Opcode == PPC::XOR8_rec)
285     return std::min(
286         getKnownLeadingZeroCount(MI->getOperand(1).getReg(), TII, MRI),
287         getKnownLeadingZeroCount(MI->getOperand(2).getReg(), TII, MRI));
288 
289   if (TII->isZeroExtended(Reg, MRI))
290     return 32;
291 
292   return 0;
293 }
294 
295 // This function maintains a map for the pairs <TOC Save Instr, Keep>
296 // Each time a new TOC save is encountered, it checks if any of the existing
297 // ones are dominated by the new one. If so, it marks the existing one as
298 // redundant by setting it's entry in the map as false. It then adds the new
299 // instruction to the map with either true or false depending on if any
300 // existing instructions dominated the new one.
301 void PPCMIPeephole::UpdateTOCSaves(
302   std::map<MachineInstr *, bool> &TOCSaves, MachineInstr *MI) {
303   assert(TII->isTOCSaveMI(*MI) && "Expecting a TOC save instruction here");
304   // FIXME: Saving TOC in prologue hasn't been implemented well in AIX ABI part,
305   // here only support it under ELFv2.
306   if (MF->getSubtarget<PPCSubtarget>().isELFv2ABI()) {
307     PPCFunctionInfo *FI = MF->getInfo<PPCFunctionInfo>();
308 
309     MachineBasicBlock *Entry = &MF->front();
310     BlockFrequency CurrBlockFreq = MBFI->getBlockFreq(MI->getParent());
311 
312     // If the block in which the TOC save resides is in a block that
313     // post-dominates Entry, or a block that is hotter than entry (keep in mind
314     // that early MachineLICM has already run so the TOC save won't be hoisted)
315     // we can just do the save in the prologue.
316     if (CurrBlockFreq > EntryFreq || MPDT->dominates(MI->getParent(), Entry))
317       FI->setMustSaveTOC(true);
318 
319     // If we are saving the TOC in the prologue, all the TOC saves can be
320     // removed from the code.
321     if (FI->mustSaveTOC()) {
322       for (auto &TOCSave : TOCSaves)
323         TOCSave.second = false;
324       // Add new instruction to map.
325       TOCSaves[MI] = false;
326       return;
327     }
328   }
329 
330   bool Keep = true;
331   for (auto &I : TOCSaves) {
332     MachineInstr *CurrInst = I.first;
333     // If new instruction dominates an existing one, mark existing one as
334     // redundant.
335     if (I.second && MDT->dominates(MI, CurrInst))
336       I.second = false;
337     // Check if the new instruction is redundant.
338     if (MDT->dominates(CurrInst, MI)) {
339       Keep = false;
340       break;
341     }
342   }
343   // Add new instruction to map.
344   TOCSaves[MI] = Keep;
345 }
346 
347 // This function returns a list of all PHI nodes in the tree starting from
348 // the RootPHI node. We perform a BFS traversal to get an ordered list of nodes.
349 // The list initially only contains the root PHI. When we visit a PHI node, we
350 // add it to the list. We continue to look for other PHI node operands while
351 // there are nodes to visit in the list. The function returns false if the
352 // optimization cannot be applied on this tree.
353 static bool collectUnprimedAccPHIs(MachineRegisterInfo *MRI,
354                                    MachineInstr *RootPHI,
355                                    SmallVectorImpl<MachineInstr *> &PHIs) {
356   PHIs.push_back(RootPHI);
357   unsigned VisitedIndex = 0;
358   while (VisitedIndex < PHIs.size()) {
359     MachineInstr *VisitedPHI = PHIs[VisitedIndex];
360     for (unsigned PHIOp = 1, NumOps = VisitedPHI->getNumOperands();
361          PHIOp != NumOps; PHIOp += 2) {
362       Register RegOp = VisitedPHI->getOperand(PHIOp).getReg();
363       if (!RegOp.isVirtual())
364         return false;
365       MachineInstr *Instr = MRI->getVRegDef(RegOp);
366       // While collecting the PHI nodes, we check if they can be converted (i.e.
367       // all the operands are either copies, implicit defs or PHI nodes).
368       unsigned Opcode = Instr->getOpcode();
369       if (Opcode == PPC::COPY) {
370         Register Reg = Instr->getOperand(1).getReg();
371         if (!Reg.isVirtual() || MRI->getRegClass(Reg) != &PPC::ACCRCRegClass)
372           return false;
373       } else if (Opcode != PPC::IMPLICIT_DEF && Opcode != PPC::PHI)
374         return false;
375       // If we detect a cycle in the PHI nodes, we exit. It would be
376       // possible to change cycles as well, but that would add a lot
377       // of complexity for a case that is unlikely to occur with MMA
378       // code.
379       if (Opcode != PPC::PHI)
380         continue;
381       if (llvm::is_contained(PHIs, Instr))
382         return false;
383       PHIs.push_back(Instr);
384     }
385     VisitedIndex++;
386   }
387   return true;
388 }
389 
390 // This function changes the unprimed accumulator PHI nodes in the PHIs list to
391 // primed accumulator PHI nodes. The list is traversed in reverse order to
392 // change all the PHI operands of a PHI node before changing the node itself.
393 // We keep a map to associate each changed PHI node to its non-changed form.
394 void PPCMIPeephole::convertUnprimedAccPHIs(
395     const PPCInstrInfo *TII, MachineRegisterInfo *MRI,
396     SmallVectorImpl<MachineInstr *> &PHIs, Register Dst) {
397   DenseMap<MachineInstr *, MachineInstr *> ChangedPHIMap;
398   for (MachineInstr *PHI : llvm::reverse(PHIs)) {
399     SmallVector<std::pair<MachineOperand, MachineOperand>, 4> PHIOps;
400     // We check if the current PHI node can be changed by looking at its
401     // operands. If all the operands are either copies from primed
402     // accumulators, implicit definitions or other unprimed accumulator
403     // PHI nodes, we change it.
404     for (unsigned PHIOp = 1, NumOps = PHI->getNumOperands(); PHIOp != NumOps;
405          PHIOp += 2) {
406       Register RegOp = PHI->getOperand(PHIOp).getReg();
407       MachineInstr *PHIInput = MRI->getVRegDef(RegOp);
408       unsigned Opcode = PHIInput->getOpcode();
409       assert((Opcode == PPC::COPY || Opcode == PPC::IMPLICIT_DEF ||
410               Opcode == PPC::PHI) &&
411              "Unexpected instruction");
412       if (Opcode == PPC::COPY) {
413         assert(MRI->getRegClass(PHIInput->getOperand(1).getReg()) ==
414                    &PPC::ACCRCRegClass &&
415                "Unexpected register class");
416         PHIOps.push_back({PHIInput->getOperand(1), PHI->getOperand(PHIOp + 1)});
417       } else if (Opcode == PPC::IMPLICIT_DEF) {
418         Register AccReg = MRI->createVirtualRegister(&PPC::ACCRCRegClass);
419         BuildMI(*PHIInput->getParent(), PHIInput, PHIInput->getDebugLoc(),
420                 TII->get(PPC::IMPLICIT_DEF), AccReg);
421         PHIOps.push_back({MachineOperand::CreateReg(AccReg, false),
422                           PHI->getOperand(PHIOp + 1)});
423       } else if (Opcode == PPC::PHI) {
424         // We found a PHI operand. At this point we know this operand
425         // has already been changed so we get its associated changed form
426         // from the map.
427         assert(ChangedPHIMap.count(PHIInput) == 1 &&
428                "This PHI node should have already been changed.");
429         MachineInstr *PrimedAccPHI = ChangedPHIMap.lookup(PHIInput);
430         PHIOps.push_back({MachineOperand::CreateReg(
431                               PrimedAccPHI->getOperand(0).getReg(), false),
432                           PHI->getOperand(PHIOp + 1)});
433       }
434     }
435     Register AccReg = Dst;
436     // If the PHI node we are changing is the root node, the register it defines
437     // will be the destination register of the original copy (of the PHI def).
438     // For all other PHI's in the list, we need to create another primed
439     // accumulator virtual register as the PHI will no longer define the
440     // unprimed accumulator.
441     if (PHI != PHIs[0])
442       AccReg = MRI->createVirtualRegister(&PPC::ACCRCRegClass);
443     MachineInstrBuilder NewPHI = BuildMI(
444         *PHI->getParent(), PHI, PHI->getDebugLoc(), TII->get(PPC::PHI), AccReg);
445     for (auto RegMBB : PHIOps) {
446       NewPHI.add(RegMBB.first).add(RegMBB.second);
447       if (MRI->isSSA())
448         addRegToUpdate(RegMBB.first.getReg());
449     }
450     // The liveness of old PHI and new PHI have to be updated.
451     addRegToUpdate(PHI->getOperand(0).getReg());
452     addRegToUpdate(AccReg);
453     ChangedPHIMap[PHI] = NewPHI.getInstr();
454     LLVM_DEBUG(dbgs() << "Converting PHI: ");
455     LLVM_DEBUG(PHI->dump());
456     LLVM_DEBUG(dbgs() << "To: ");
457     LLVM_DEBUG(NewPHI.getInstr()->dump());
458   }
459 }
460 
461 // Perform peephole optimizations.
462 bool PPCMIPeephole::simplifyCode() {
463   bool Simplified = false;
464   bool TrapOpt = false;
465   MachineInstr* ToErase = nullptr;
466   std::map<MachineInstr *, bool> TOCSaves;
467   const TargetRegisterInfo *TRI = &TII->getRegisterInfo();
468   NumFunctionsEnteredInMIPeephole++;
469   if (ConvertRegReg) {
470     // Fixed-point conversion of reg/reg instructions fed by load-immediate
471     // into reg/imm instructions. FIXME: This is expensive, control it with
472     // an option.
473     bool SomethingChanged = false;
474     do {
475       NumFixedPointIterations++;
476       SomethingChanged = false;
477       for (MachineBasicBlock &MBB : *MF) {
478         for (MachineInstr &MI : MBB) {
479           if (MI.isDebugInstr())
480             continue;
481 
482           if (!DebugCounter::shouldExecute(PeepholeXToICounter))
483             continue;
484 
485           SmallSet<Register, 4> RRToRIRegsToUpdate;
486           if (!TII->convertToImmediateForm(MI, RRToRIRegsToUpdate))
487             continue;
488           for (Register R : RRToRIRegsToUpdate)
489             addRegToUpdate(R);
490           // The updated instruction may now have new register operands.
491           // Conservatively add them to recompute the flags as well.
492           for (const MachineOperand &MO : MI.operands())
493             if (MO.isReg())
494               addRegToUpdate(MO.getReg());
495           // We don't erase anything in case the def has other uses. Let DCE
496           // remove it if it can be removed.
497           LLVM_DEBUG(dbgs() << "Converted instruction to imm form: ");
498           LLVM_DEBUG(MI.dump());
499           NumConvertedToImmediateForm++;
500           SomethingChanged = true;
501           Simplified = true;
502         }
503       }
504     } while (SomethingChanged && FixedPointRegToImm);
505   }
506 
507   // Since we are deleting this instruction, we need to run LiveVariables
508   // on any of its definitions that are marked as needing an update since
509   // we can't run LiveVariables on a deleted register. This only needs
510   // to be done for defs since uses will have their own defining
511   // instructions so we won't be running LiveVariables on a deleted reg.
512   auto recomputeLVForDyingInstr = [&]() {
513     if (RegsToUpdate.empty())
514       return;
515     for (MachineOperand &MO : ToErase->operands()) {
516       if (!MO.isReg() || !MO.isDef() || !RegsToUpdate.count(MO.getReg()))
517         continue;
518       Register RegToUpdate = MO.getReg();
519       RegsToUpdate.erase(RegToUpdate);
520       // If some transformation has introduced an additional definition of
521       // this register (breaking SSA), we can safely convert this def to
522       // a def of an invalid register as the instruction is going away.
523       if (!MRI->getUniqueVRegDef(RegToUpdate))
524         MO.setReg(PPC::NoRegister);
525       LV->recomputeForSingleDefVirtReg(RegToUpdate);
526     }
527   };
528 
529   for (MachineBasicBlock &MBB : *MF) {
530     for (MachineInstr &MI : MBB) {
531 
532       // If the previous instruction was marked for elimination,
533       // remove it now.
534       if (ToErase) {
535         LLVM_DEBUG(dbgs() << "Deleting instruction: ");
536         LLVM_DEBUG(ToErase->dump());
537         recomputeLVForDyingInstr();
538         ToErase->eraseFromParent();
539         ToErase = nullptr;
540       }
541       // If a conditional trap instruction got optimized to an
542       // unconditional trap, eliminate all the instructions after
543       // the trap.
544       if (EnableTrapOptimization && TrapOpt) {
545         ToErase = &MI;
546         continue;
547       }
548 
549       // Ignore debug instructions.
550       if (MI.isDebugInstr())
551         continue;
552 
553       if (!DebugCounter::shouldExecute(PeepholePerOpCounter))
554         continue;
555 
556       // Per-opcode peepholes.
557       switch (MI.getOpcode()) {
558 
559       default:
560         break;
561       case PPC::COPY: {
562         Register Src = MI.getOperand(1).getReg();
563         Register Dst = MI.getOperand(0).getReg();
564         if (!Src.isVirtual() || !Dst.isVirtual())
565           break;
566         if (MRI->getRegClass(Src) != &PPC::UACCRCRegClass ||
567             MRI->getRegClass(Dst) != &PPC::ACCRCRegClass)
568           break;
569 
570         // We are copying an unprimed accumulator to a primed accumulator.
571         // If the input to the copy is a PHI that is fed only by (i) copies in
572         // the other direction (ii) implicitly defined unprimed accumulators or
573         // (iii) other PHI nodes satisfying (i) and (ii), we can change
574         // the PHI to a PHI on primed accumulators (as long as we also change
575         // its operands). To detect and change such copies, we first get a list
576         // of all the PHI nodes starting from the root PHI node in BFS order.
577         // We then visit all these PHI nodes to check if they can be changed to
578         // primed accumulator PHI nodes and if so, we change them.
579         MachineInstr *RootPHI = MRI->getVRegDef(Src);
580         if (RootPHI->getOpcode() != PPC::PHI)
581           break;
582 
583         SmallVector<MachineInstr *, 4> PHIs;
584         if (!collectUnprimedAccPHIs(MRI, RootPHI, PHIs))
585           break;
586 
587         convertUnprimedAccPHIs(TII, MRI, PHIs, Dst);
588 
589         ToErase = &MI;
590         break;
591       }
592       case PPC::LI:
593       case PPC::LI8: {
594         // If we are materializing a zero, look for any use operands for which
595         // zero means immediate zero. All such operands can be replaced with
596         // PPC::ZERO.
597         if (!MI.getOperand(1).isImm() || MI.getOperand(1).getImm() != 0)
598           break;
599         Register MIDestReg = MI.getOperand(0).getReg();
600         bool Folded = false;
601         for (MachineInstr& UseMI : MRI->use_instructions(MIDestReg))
602           Folded |= TII->onlyFoldImmediate(UseMI, MI, MIDestReg);
603         if (MRI->use_nodbg_empty(MIDestReg)) {
604           ++NumLoadImmZeroFoldedAndRemoved;
605           ToErase = &MI;
606         }
607         if (Folded)
608           addRegToUpdate(MIDestReg);
609         Simplified |= Folded;
610         break;
611       }
612       case PPC::STW:
613       case PPC::STD: {
614         MachineFrameInfo &MFI = MF->getFrameInfo();
615         if (MFI.hasVarSizedObjects() ||
616             (!MF->getSubtarget<PPCSubtarget>().isELFv2ABI() &&
617              !MF->getSubtarget<PPCSubtarget>().isAIXABI()))
618           break;
619         // When encountering a TOC save instruction, call UpdateTOCSaves
620         // to add it to the TOCSaves map and mark any existing TOC saves
621         // it dominates as redundant.
622         if (TII->isTOCSaveMI(MI))
623           UpdateTOCSaves(TOCSaves, &MI);
624         break;
625       }
626       case PPC::XXPERMDI: {
627         // Perform simplifications of 2x64 vector swaps and splats.
628         // A swap is identified by an immediate value of 2, and a splat
629         // is identified by an immediate value of 0 or 3.
630         int Immed = MI.getOperand(3).getImm();
631 
632         if (Immed == 1)
633           break;
634 
635         // For each of these simplifications, we need the two source
636         // regs to match.  Unfortunately, MachineCSE ignores COPY and
637         // SUBREG_TO_REG, so for example we can see
638         //   XXPERMDI t, SUBREG_TO_REG(s), SUBREG_TO_REG(s), immed.
639         // We have to look through chains of COPY and SUBREG_TO_REG
640         // to find the real source values for comparison.
641         Register TrueReg1 =
642           TRI->lookThruCopyLike(MI.getOperand(1).getReg(), MRI);
643         Register TrueReg2 =
644           TRI->lookThruCopyLike(MI.getOperand(2).getReg(), MRI);
645 
646         if (!(TrueReg1 == TrueReg2 && TrueReg1.isVirtual()))
647           break;
648 
649         MachineInstr *DefMI = MRI->getVRegDef(TrueReg1);
650 
651         if (!DefMI)
652           break;
653 
654         unsigned DefOpc = DefMI->getOpcode();
655 
656         // If this is a splat fed by a splatting load, the splat is
657         // redundant. Replace with a copy. This doesn't happen directly due
658         // to code in PPCDAGToDAGISel.cpp, but it can happen when converting
659         // a load of a double to a vector of 64-bit integers.
660         auto isConversionOfLoadAndSplat = [=]() -> bool {
661           if (DefOpc != PPC::XVCVDPSXDS && DefOpc != PPC::XVCVDPUXDS)
662             return false;
663           Register FeedReg1 =
664             TRI->lookThruCopyLike(DefMI->getOperand(1).getReg(), MRI);
665           if (FeedReg1.isVirtual()) {
666             MachineInstr *LoadMI = MRI->getVRegDef(FeedReg1);
667             if (LoadMI && LoadMI->getOpcode() == PPC::LXVDSX)
668               return true;
669           }
670           return false;
671         };
672         if ((Immed == 0 || Immed == 3) &&
673             (DefOpc == PPC::LXVDSX || isConversionOfLoadAndSplat())) {
674           LLVM_DEBUG(dbgs() << "Optimizing load-and-splat/splat "
675                                "to load-and-splat/copy: ");
676           LLVM_DEBUG(MI.dump());
677           BuildMI(MBB, &MI, MI.getDebugLoc(), TII->get(PPC::COPY),
678                   MI.getOperand(0).getReg())
679               .add(MI.getOperand(1));
680           addRegToUpdate(MI.getOperand(1).getReg());
681           ToErase = &MI;
682           Simplified = true;
683         }
684 
685         // If this is a splat or a swap fed by another splat, we
686         // can replace it with a copy.
687         if (DefOpc == PPC::XXPERMDI) {
688           Register DefReg1 = DefMI->getOperand(1).getReg();
689           Register DefReg2 = DefMI->getOperand(2).getReg();
690           unsigned DefImmed = DefMI->getOperand(3).getImm();
691 
692           // If the two inputs are not the same register, check to see if
693           // they originate from the same virtual register after only
694           // copy-like instructions.
695           if (DefReg1 != DefReg2) {
696             Register FeedReg1 = TRI->lookThruCopyLike(DefReg1, MRI);
697             Register FeedReg2 = TRI->lookThruCopyLike(DefReg2, MRI);
698 
699             if (!(FeedReg1 == FeedReg2 && FeedReg1.isVirtual()))
700               break;
701           }
702 
703           if (DefImmed == 0 || DefImmed == 3) {
704             LLVM_DEBUG(dbgs() << "Optimizing splat/swap or splat/splat "
705                                  "to splat/copy: ");
706             LLVM_DEBUG(MI.dump());
707             BuildMI(MBB, &MI, MI.getDebugLoc(), TII->get(PPC::COPY),
708                     MI.getOperand(0).getReg())
709                 .add(MI.getOperand(1));
710             addRegToUpdate(MI.getOperand(1).getReg());
711             ToErase = &MI;
712             Simplified = true;
713           }
714 
715           // If this is a splat fed by a swap, we can simplify modify
716           // the splat to splat the other value from the swap's input
717           // parameter.
718           else if ((Immed == 0 || Immed == 3) && DefImmed == 2) {
719             LLVM_DEBUG(dbgs() << "Optimizing swap/splat => splat: ");
720             LLVM_DEBUG(MI.dump());
721             addRegToUpdate(MI.getOperand(1).getReg());
722             addRegToUpdate(MI.getOperand(2).getReg());
723             MI.getOperand(1).setReg(DefReg1);
724             MI.getOperand(2).setReg(DefReg2);
725             MI.getOperand(3).setImm(3 - Immed);
726             addRegToUpdate(DefReg1);
727             addRegToUpdate(DefReg2);
728             Simplified = true;
729           }
730 
731           // If this is a swap fed by a swap, we can replace it
732           // with a copy from the first swap's input.
733           else if (Immed == 2 && DefImmed == 2) {
734             LLVM_DEBUG(dbgs() << "Optimizing swap/swap => copy: ");
735             LLVM_DEBUG(MI.dump());
736             addRegToUpdate(MI.getOperand(1).getReg());
737             BuildMI(MBB, &MI, MI.getDebugLoc(), TII->get(PPC::COPY),
738                     MI.getOperand(0).getReg())
739                 .add(DefMI->getOperand(1));
740             addRegToUpdate(DefMI->getOperand(0).getReg());
741             addRegToUpdate(DefMI->getOperand(1).getReg());
742             ToErase = &MI;
743             Simplified = true;
744           }
745         } else if ((Immed == 0 || Immed == 3 || Immed == 2) &&
746                    DefOpc == PPC::XXPERMDIs &&
747                    (DefMI->getOperand(2).getImm() == 0 ||
748                     DefMI->getOperand(2).getImm() == 3)) {
749           ToErase = &MI;
750           Simplified = true;
751           // Swap of a splat, convert to copy.
752           if (Immed == 2) {
753             LLVM_DEBUG(dbgs() << "Optimizing swap(splat) => copy(splat): ");
754             LLVM_DEBUG(MI.dump());
755             BuildMI(MBB, &MI, MI.getDebugLoc(), TII->get(PPC::COPY),
756                     MI.getOperand(0).getReg())
757                 .add(MI.getOperand(1));
758             addRegToUpdate(MI.getOperand(1).getReg());
759             break;
760           }
761           // Splat fed by another splat - switch the output of the first
762           // and remove the second.
763           DefMI->getOperand(0).setReg(MI.getOperand(0).getReg());
764           LLVM_DEBUG(dbgs() << "Removing redundant splat: ");
765           LLVM_DEBUG(MI.dump());
766         } else if (Immed == 2 &&
767                    (DefOpc == PPC::VSPLTB || DefOpc == PPC::VSPLTH ||
768                     DefOpc == PPC::VSPLTW || DefOpc == PPC::XXSPLTW ||
769                     DefOpc == PPC::VSPLTISB || DefOpc == PPC::VSPLTISH ||
770                     DefOpc == PPC::VSPLTISW)) {
771           // Swap of various vector splats, convert to copy.
772           ToErase = &MI;
773           Simplified = true;
774           LLVM_DEBUG(dbgs() << "Optimizing swap(vsplt(is)?[b|h|w]|xxspltw) => "
775                                "copy(vsplt(is)?[b|h|w]|xxspltw): ");
776           LLVM_DEBUG(MI.dump());
777           BuildMI(MBB, &MI, MI.getDebugLoc(), TII->get(PPC::COPY),
778                   MI.getOperand(0).getReg())
779               .add(MI.getOperand(1));
780           addRegToUpdate(MI.getOperand(1).getReg());
781         } else if ((Immed == 0 || Immed == 3 || Immed == 2) &&
782                    TII->isLoadFromConstantPool(DefMI)) {
783           const Constant *C = TII->getConstantFromConstantPool(DefMI);
784           if (C && C->getType()->isVectorTy() && C->getSplatValue()) {
785             ToErase = &MI;
786             Simplified = true;
787             LLVM_DEBUG(dbgs()
788                        << "Optimizing swap(splat pattern from constant-pool) "
789                           "=> copy(splat pattern from constant-pool): ");
790             LLVM_DEBUG(MI.dump());
791             BuildMI(MBB, &MI, MI.getDebugLoc(), TII->get(PPC::COPY),
792                     MI.getOperand(0).getReg())
793                 .add(MI.getOperand(1));
794             addRegToUpdate(MI.getOperand(1).getReg());
795           }
796         }
797         break;
798       }
799       case PPC::VSPLTB:
800       case PPC::VSPLTH:
801       case PPC::XXSPLTW: {
802         unsigned MyOpcode = MI.getOpcode();
803         unsigned OpNo = MyOpcode == PPC::XXSPLTW ? 1 : 2;
804         Register TrueReg =
805           TRI->lookThruCopyLike(MI.getOperand(OpNo).getReg(), MRI);
806         if (!TrueReg.isVirtual())
807           break;
808         MachineInstr *DefMI = MRI->getVRegDef(TrueReg);
809         if (!DefMI)
810           break;
811         unsigned DefOpcode = DefMI->getOpcode();
812         auto isConvertOfSplat = [=]() -> bool {
813           if (DefOpcode != PPC::XVCVSPSXWS && DefOpcode != PPC::XVCVSPUXWS)
814             return false;
815           Register ConvReg = DefMI->getOperand(1).getReg();
816           if (!ConvReg.isVirtual())
817             return false;
818           MachineInstr *Splt = MRI->getVRegDef(ConvReg);
819           return Splt && (Splt->getOpcode() == PPC::LXVWSX ||
820             Splt->getOpcode() == PPC::XXSPLTW);
821         };
822         bool AlreadySplat = (MyOpcode == DefOpcode) ||
823           (MyOpcode == PPC::VSPLTB && DefOpcode == PPC::VSPLTBs) ||
824           (MyOpcode == PPC::VSPLTH && DefOpcode == PPC::VSPLTHs) ||
825           (MyOpcode == PPC::XXSPLTW && DefOpcode == PPC::XXSPLTWs) ||
826           (MyOpcode == PPC::XXSPLTW && DefOpcode == PPC::LXVWSX) ||
827           (MyOpcode == PPC::XXSPLTW && DefOpcode == PPC::MTVSRWS)||
828           (MyOpcode == PPC::XXSPLTW && isConvertOfSplat());
829         // If the instruction[s] that feed this splat have already splat
830         // the value, this splat is redundant.
831         if (AlreadySplat) {
832           LLVM_DEBUG(dbgs() << "Changing redundant splat to a copy: ");
833           LLVM_DEBUG(MI.dump());
834           BuildMI(MBB, &MI, MI.getDebugLoc(), TII->get(PPC::COPY),
835                   MI.getOperand(0).getReg())
836               .add(MI.getOperand(OpNo));
837           addRegToUpdate(MI.getOperand(OpNo).getReg());
838           ToErase = &MI;
839           Simplified = true;
840         }
841         // Splat fed by a shift. Usually when we align value to splat into
842         // vector element zero.
843         if (DefOpcode == PPC::XXSLDWI) {
844           Register ShiftRes = DefMI->getOperand(0).getReg();
845           Register ShiftOp1 = DefMI->getOperand(1).getReg();
846           Register ShiftOp2 = DefMI->getOperand(2).getReg();
847           unsigned ShiftImm = DefMI->getOperand(3).getImm();
848           unsigned SplatImm =
849               MI.getOperand(MyOpcode == PPC::XXSPLTW ? 2 : 1).getImm();
850           if (ShiftOp1 == ShiftOp2) {
851             unsigned NewElem = (SplatImm + ShiftImm) & 0x3;
852             if (MRI->hasOneNonDBGUse(ShiftRes)) {
853               LLVM_DEBUG(dbgs() << "Removing redundant shift: ");
854               LLVM_DEBUG(DefMI->dump());
855               ToErase = DefMI;
856             }
857             Simplified = true;
858             LLVM_DEBUG(dbgs() << "Changing splat immediate from " << SplatImm
859                               << " to " << NewElem << " in instruction: ");
860             LLVM_DEBUG(MI.dump());
861             addRegToUpdate(MI.getOperand(OpNo).getReg());
862             addRegToUpdate(ShiftOp1);
863             MI.getOperand(OpNo).setReg(ShiftOp1);
864             MI.getOperand(2).setImm(NewElem);
865           }
866         }
867         break;
868       }
869       case PPC::XVCVDPSP: {
870         // If this is a DP->SP conversion fed by an FRSP, the FRSP is redundant.
871         Register TrueReg =
872           TRI->lookThruCopyLike(MI.getOperand(1).getReg(), MRI);
873         if (!TrueReg.isVirtual())
874           break;
875         MachineInstr *DefMI = MRI->getVRegDef(TrueReg);
876 
877         // This can occur when building a vector of single precision or integer
878         // values.
879         if (DefMI && DefMI->getOpcode() == PPC::XXPERMDI) {
880           Register DefsReg1 =
881             TRI->lookThruCopyLike(DefMI->getOperand(1).getReg(), MRI);
882           Register DefsReg2 =
883             TRI->lookThruCopyLike(DefMI->getOperand(2).getReg(), MRI);
884           if (!DefsReg1.isVirtual() || !DefsReg2.isVirtual())
885             break;
886           MachineInstr *P1 = MRI->getVRegDef(DefsReg1);
887           MachineInstr *P2 = MRI->getVRegDef(DefsReg2);
888 
889           if (!P1 || !P2)
890             break;
891 
892           // Remove the passed FRSP/XSRSP instruction if it only feeds this MI
893           // and set any uses of that FRSP/XSRSP (in this MI) to the source of
894           // the FRSP/XSRSP.
895           auto removeFRSPIfPossible = [&](MachineInstr *RoundInstr) {
896             unsigned Opc = RoundInstr->getOpcode();
897             if ((Opc == PPC::FRSP || Opc == PPC::XSRSP) &&
898                 MRI->hasOneNonDBGUse(RoundInstr->getOperand(0).getReg())) {
899               Simplified = true;
900               Register ConvReg1 = RoundInstr->getOperand(1).getReg();
901               Register FRSPDefines = RoundInstr->getOperand(0).getReg();
902               MachineInstr &Use = *(MRI->use_instr_nodbg_begin(FRSPDefines));
903               for (int i = 0, e = Use.getNumOperands(); i < e; ++i)
904                 if (Use.getOperand(i).isReg() &&
905                     Use.getOperand(i).getReg() == FRSPDefines)
906                   Use.getOperand(i).setReg(ConvReg1);
907               LLVM_DEBUG(dbgs() << "Removing redundant FRSP/XSRSP:\n");
908               LLVM_DEBUG(RoundInstr->dump());
909               LLVM_DEBUG(dbgs() << "As it feeds instruction:\n");
910               LLVM_DEBUG(MI.dump());
911               LLVM_DEBUG(dbgs() << "Through instruction:\n");
912               LLVM_DEBUG(DefMI->dump());
913               addRegToUpdate(ConvReg1);
914               addRegToUpdate(FRSPDefines);
915               ToErase = RoundInstr;
916             }
917           };
918 
919           // If the input to XVCVDPSP is a vector that was built (even
920           // partially) out of FRSP's, the FRSP(s) can safely be removed
921           // since this instruction performs the same operation.
922           if (P1 != P2) {
923             removeFRSPIfPossible(P1);
924             removeFRSPIfPossible(P2);
925             break;
926           }
927           removeFRSPIfPossible(P1);
928         }
929         break;
930       }
931       case PPC::EXTSH:
932       case PPC::EXTSH8:
933       case PPC::EXTSH8_32_64: {
934         if (!EnableSExtElimination) break;
935         Register NarrowReg = MI.getOperand(1).getReg();
936         if (!NarrowReg.isVirtual())
937           break;
938 
939         MachineInstr *SrcMI = MRI->getVRegDef(NarrowReg);
940         unsigned SrcOpcode = SrcMI->getOpcode();
941         // If we've used a zero-extending load that we will sign-extend,
942         // just do a sign-extending load.
943         if (SrcOpcode == PPC::LHZ || SrcOpcode == PPC::LHZX) {
944           if (!MRI->hasOneNonDBGUse(SrcMI->getOperand(0).getReg()))
945             break;
946           // Determine the new opcode. We need to make sure that if the original
947           // instruction has a 64 bit opcode we keep using a 64 bit opcode.
948           // Likewise if the source is X-Form the new opcode should also be
949           // X-Form.
950           unsigned Opc = PPC::LHA;
951           bool SourceIsXForm = SrcOpcode == PPC::LHZX;
952           bool MIIs64Bit = MI.getOpcode() == PPC::EXTSH8 ||
953             MI.getOpcode() == PPC::EXTSH8_32_64;
954 
955           if (SourceIsXForm && MIIs64Bit)
956             Opc = PPC::LHAX8;
957           else if (SourceIsXForm && !MIIs64Bit)
958             Opc = PPC::LHAX;
959           else if (MIIs64Bit)
960             Opc = PPC::LHA8;
961 
962           addRegToUpdate(NarrowReg);
963           addRegToUpdate(MI.getOperand(0).getReg());
964 
965           // We are removing a definition of NarrowReg which will cause
966           // problems in AliveBlocks. Add an implicit def that will be
967           // removed so that AliveBlocks are updated correctly.
968           addDummyDef(MBB, &MI, NarrowReg);
969           LLVM_DEBUG(dbgs() << "Zero-extending load\n");
970           LLVM_DEBUG(SrcMI->dump());
971           LLVM_DEBUG(dbgs() << "and sign-extension\n");
972           LLVM_DEBUG(MI.dump());
973           LLVM_DEBUG(dbgs() << "are merged into sign-extending load\n");
974           SrcMI->setDesc(TII->get(Opc));
975           SrcMI->getOperand(0).setReg(MI.getOperand(0).getReg());
976           ToErase = &MI;
977           Simplified = true;
978           NumEliminatedSExt++;
979         }
980         break;
981       }
982       case PPC::EXTSW:
983       case PPC::EXTSW_32:
984       case PPC::EXTSW_32_64: {
985         if (!EnableSExtElimination) break;
986         Register NarrowReg = MI.getOperand(1).getReg();
987         if (!NarrowReg.isVirtual())
988           break;
989 
990         MachineInstr *SrcMI = MRI->getVRegDef(NarrowReg);
991         unsigned SrcOpcode = SrcMI->getOpcode();
992         // If we've used a zero-extending load that we will sign-extend,
993         // just do a sign-extending load.
994         if (SrcOpcode == PPC::LWZ || SrcOpcode == PPC::LWZX) {
995           if (!MRI->hasOneNonDBGUse(SrcMI->getOperand(0).getReg()))
996             break;
997 
998           // The transformation from a zero-extending load to a sign-extending
999           // load is only legal when the displacement is a multiple of 4.
1000           // If the displacement is not at least 4 byte aligned, don't perform
1001           // the transformation.
1002           bool IsWordAligned = false;
1003           if (SrcMI->getOperand(1).isGlobal()) {
1004             const GlobalObject *GO =
1005                 dyn_cast<GlobalObject>(SrcMI->getOperand(1).getGlobal());
1006             if (GO && GO->getAlign() && *GO->getAlign() >= 4 &&
1007                 (SrcMI->getOperand(1).getOffset() % 4 == 0))
1008               IsWordAligned = true;
1009           } else if (SrcMI->getOperand(1).isImm()) {
1010             int64_t Value = SrcMI->getOperand(1).getImm();
1011             if (Value % 4 == 0)
1012               IsWordAligned = true;
1013           }
1014 
1015           // Determine the new opcode. We need to make sure that if the original
1016           // instruction has a 64 bit opcode we keep using a 64 bit opcode.
1017           // Likewise if the source is X-Form the new opcode should also be
1018           // X-Form.
1019           unsigned Opc = PPC::LWA_32;
1020           bool SourceIsXForm = SrcOpcode == PPC::LWZX;
1021           bool MIIs64Bit = MI.getOpcode() == PPC::EXTSW ||
1022             MI.getOpcode() == PPC::EXTSW_32_64;
1023 
1024           if (SourceIsXForm && MIIs64Bit)
1025             Opc = PPC::LWAX;
1026           else if (SourceIsXForm && !MIIs64Bit)
1027             Opc = PPC::LWAX_32;
1028           else if (MIIs64Bit)
1029             Opc = PPC::LWA;
1030 
1031           if (!IsWordAligned && (Opc == PPC::LWA || Opc == PPC::LWA_32))
1032             break;
1033 
1034           addRegToUpdate(NarrowReg);
1035           addRegToUpdate(MI.getOperand(0).getReg());
1036 
1037           // We are removing a definition of NarrowReg which will cause
1038           // problems in AliveBlocks. Add an implicit def that will be
1039           // removed so that AliveBlocks are updated correctly.
1040           addDummyDef(MBB, &MI, NarrowReg);
1041           LLVM_DEBUG(dbgs() << "Zero-extending load\n");
1042           LLVM_DEBUG(SrcMI->dump());
1043           LLVM_DEBUG(dbgs() << "and sign-extension\n");
1044           LLVM_DEBUG(MI.dump());
1045           LLVM_DEBUG(dbgs() << "are merged into sign-extending load\n");
1046           SrcMI->setDesc(TII->get(Opc));
1047           SrcMI->getOperand(0).setReg(MI.getOperand(0).getReg());
1048           ToErase = &MI;
1049           Simplified = true;
1050           NumEliminatedSExt++;
1051         } else if (MI.getOpcode() == PPC::EXTSW_32_64 &&
1052                    TII->isSignExtended(NarrowReg, MRI)) {
1053           // We can eliminate EXTSW if the input is known to be already
1054           // sign-extended. However, we are not sure whether a spill will occur
1055           // during register allocation. If there is no promotion, it will use
1056           // 'stw' instead of 'std', and 'lwz' instead of 'ld' when spilling,
1057           // since the register class is 32-bits. Consequently, the high 32-bit
1058           // information will be lost. Therefore, all these instructions in the
1059           // chain used to deduce sign extension to eliminate the 'extsw' will
1060           // need to be promoted to 64-bit pseudo instructions when the 'extsw'
1061           // is eliminated.
1062           TII->promoteInstr32To64ForElimEXTSW(NarrowReg, MRI, 0, LV);
1063 
1064           LLVM_DEBUG(dbgs() << "Removing redundant sign-extension\n");
1065           Register TmpReg =
1066               MF->getRegInfo().createVirtualRegister(&PPC::G8RCRegClass);
1067           BuildMI(MBB, &MI, MI.getDebugLoc(), TII->get(PPC::IMPLICIT_DEF),
1068                   TmpReg);
1069           BuildMI(MBB, &MI, MI.getDebugLoc(), TII->get(PPC::INSERT_SUBREG),
1070                   MI.getOperand(0).getReg())
1071               .addReg(TmpReg)
1072               .addReg(NarrowReg)
1073               .addImm(PPC::sub_32);
1074           ToErase = &MI;
1075           Simplified = true;
1076           NumEliminatedSExt++;
1077         }
1078         break;
1079       }
1080       case PPC::RLDICL: {
1081         // We can eliminate RLDICL (e.g. for zero-extension)
1082         // if all bits to clear are already zero in the input.
1083         // This code assume following code sequence for zero-extension.
1084         //   %6 = COPY %5:sub_32; (optional)
1085         //   %8 = IMPLICIT_DEF;
1086         //   %7<def,tied1> = INSERT_SUBREG %8<tied0>, %6, sub_32;
1087         if (!EnableZExtElimination) break;
1088 
1089         if (MI.getOperand(2).getImm() != 0)
1090           break;
1091 
1092         Register SrcReg = MI.getOperand(1).getReg();
1093         if (!SrcReg.isVirtual())
1094           break;
1095 
1096         MachineInstr *SrcMI = MRI->getVRegDef(SrcReg);
1097         if (!(SrcMI && SrcMI->getOpcode() == PPC::INSERT_SUBREG &&
1098               SrcMI->getOperand(0).isReg() && SrcMI->getOperand(1).isReg()))
1099           break;
1100 
1101         MachineInstr *ImpDefMI, *SubRegMI;
1102         ImpDefMI = MRI->getVRegDef(SrcMI->getOperand(1).getReg());
1103         SubRegMI = MRI->getVRegDef(SrcMI->getOperand(2).getReg());
1104         if (ImpDefMI->getOpcode() != PPC::IMPLICIT_DEF) break;
1105 
1106         SrcMI = SubRegMI;
1107         if (SubRegMI->getOpcode() == PPC::COPY) {
1108           Register CopyReg = SubRegMI->getOperand(1).getReg();
1109           if (CopyReg.isVirtual())
1110             SrcMI = MRI->getVRegDef(CopyReg);
1111         }
1112         if (!SrcMI->getOperand(0).isReg())
1113           break;
1114 
1115         unsigned KnownZeroCount =
1116             getKnownLeadingZeroCount(SrcMI->getOperand(0).getReg(), TII, MRI);
1117         if (MI.getOperand(3).getImm() <= KnownZeroCount) {
1118           LLVM_DEBUG(dbgs() << "Removing redundant zero-extension\n");
1119           BuildMI(MBB, &MI, MI.getDebugLoc(), TII->get(PPC::COPY),
1120                   MI.getOperand(0).getReg())
1121               .addReg(SrcReg);
1122           addRegToUpdate(SrcReg);
1123           ToErase = &MI;
1124           Simplified = true;
1125           NumEliminatedZExt++;
1126         }
1127         break;
1128       }
1129 
1130       // TODO: Any instruction that has an immediate form fed only by a PHI
1131       // whose operands are all load immediate can be folded away. We currently
1132       // do this for ADD instructions, but should expand it to arithmetic and
1133       // binary instructions with immediate forms in the future.
1134       case PPC::ADD4:
1135       case PPC::ADD8: {
1136         auto isSingleUsePHI = [&](MachineOperand *PhiOp) {
1137           assert(PhiOp && "Invalid Operand!");
1138           MachineInstr *DefPhiMI = getVRegDefOrNull(PhiOp, MRI);
1139 
1140           return DefPhiMI && (DefPhiMI->getOpcode() == PPC::PHI) &&
1141                  MRI->hasOneNonDBGUse(DefPhiMI->getOperand(0).getReg());
1142         };
1143 
1144         auto dominatesAllSingleUseLIs = [&](MachineOperand *DominatorOp,
1145                                             MachineOperand *PhiOp) {
1146           assert(PhiOp && "Invalid Operand!");
1147           assert(DominatorOp && "Invalid Operand!");
1148           MachineInstr *DefPhiMI = getVRegDefOrNull(PhiOp, MRI);
1149           MachineInstr *DefDomMI = getVRegDefOrNull(DominatorOp, MRI);
1150 
1151           // Note: the vregs only show up at odd indices position of PHI Node,
1152           // the even indices position save the BB info.
1153           for (unsigned i = 1; i < DefPhiMI->getNumOperands(); i += 2) {
1154             MachineInstr *LiMI =
1155                 getVRegDefOrNull(&DefPhiMI->getOperand(i), MRI);
1156             if (!LiMI ||
1157                 (LiMI->getOpcode() != PPC::LI && LiMI->getOpcode() != PPC::LI8)
1158                 || !MRI->hasOneNonDBGUse(LiMI->getOperand(0).getReg()) ||
1159                 !MDT->dominates(DefDomMI, LiMI))
1160               return false;
1161           }
1162 
1163           return true;
1164         };
1165 
1166         MachineOperand Op1 = MI.getOperand(1);
1167         MachineOperand Op2 = MI.getOperand(2);
1168         if (isSingleUsePHI(&Op2) && dominatesAllSingleUseLIs(&Op1, &Op2))
1169           std::swap(Op1, Op2);
1170         else if (!isSingleUsePHI(&Op1) || !dominatesAllSingleUseLIs(&Op2, &Op1))
1171           break; // We don't have an ADD fed by LI's that can be transformed
1172 
1173         // Now we know that Op1 is the PHI node and Op2 is the dominator
1174         Register DominatorReg = Op2.getReg();
1175 
1176         const TargetRegisterClass *TRC = MI.getOpcode() == PPC::ADD8
1177                                              ? &PPC::G8RC_and_G8RC_NOX0RegClass
1178                                              : &PPC::GPRC_and_GPRC_NOR0RegClass;
1179         MRI->setRegClass(DominatorReg, TRC);
1180 
1181         // replace LIs with ADDIs
1182         MachineInstr *DefPhiMI = getVRegDefOrNull(&Op1, MRI);
1183         for (unsigned i = 1; i < DefPhiMI->getNumOperands(); i += 2) {
1184           MachineInstr *LiMI = getVRegDefOrNull(&DefPhiMI->getOperand(i), MRI);
1185           LLVM_DEBUG(dbgs() << "Optimizing LI to ADDI: ");
1186           LLVM_DEBUG(LiMI->dump());
1187 
1188           // There could be repeated registers in the PHI, e.g: %1 =
1189           // PHI %6, <%bb.2>, %8, <%bb.3>, %8, <%bb.6>; So if we've
1190           // already replaced the def instruction, skip.
1191           if (LiMI->getOpcode() == PPC::ADDI || LiMI->getOpcode() == PPC::ADDI8)
1192             continue;
1193 
1194           assert((LiMI->getOpcode() == PPC::LI ||
1195                   LiMI->getOpcode() == PPC::LI8) &&
1196                  "Invalid Opcode!");
1197           auto LiImm = LiMI->getOperand(1).getImm(); // save the imm of LI
1198           LiMI->removeOperand(1);                    // remove the imm of LI
1199           LiMI->setDesc(TII->get(LiMI->getOpcode() == PPC::LI ? PPC::ADDI
1200                                                               : PPC::ADDI8));
1201           MachineInstrBuilder(*LiMI->getParent()->getParent(), *LiMI)
1202               .addReg(DominatorReg)
1203               .addImm(LiImm); // restore the imm of LI
1204           LLVM_DEBUG(LiMI->dump());
1205         }
1206 
1207         // Replace ADD with COPY
1208         LLVM_DEBUG(dbgs() << "Optimizing ADD to COPY: ");
1209         LLVM_DEBUG(MI.dump());
1210         BuildMI(MBB, &MI, MI.getDebugLoc(), TII->get(PPC::COPY),
1211                 MI.getOperand(0).getReg())
1212             .add(Op1);
1213         addRegToUpdate(Op1.getReg());
1214         addRegToUpdate(Op2.getReg());
1215         ToErase = &MI;
1216         Simplified = true;
1217         NumOptADDLIs++;
1218         break;
1219       }
1220       case PPC::RLDICR: {
1221         Simplified |= emitRLDICWhenLoweringJumpTables(MI, ToErase) ||
1222                       combineSEXTAndSHL(MI, ToErase);
1223         break;
1224       }
1225       case PPC::ANDI_rec:
1226       case PPC::ANDI8_rec:
1227       case PPC::ANDIS_rec:
1228       case PPC::ANDIS8_rec: {
1229         Register TrueReg =
1230             TRI->lookThruCopyLike(MI.getOperand(1).getReg(), MRI);
1231         if (!TrueReg.isVirtual() || !MRI->hasOneNonDBGUse(TrueReg))
1232           break;
1233 
1234         MachineInstr *SrcMI = MRI->getVRegDef(TrueReg);
1235         if (!SrcMI)
1236           break;
1237 
1238         unsigned SrcOpCode = SrcMI->getOpcode();
1239         if (SrcOpCode != PPC::RLDICL && SrcOpCode != PPC::RLDICR)
1240           break;
1241 
1242         Register SrcReg, DstReg;
1243         SrcReg = SrcMI->getOperand(1).getReg();
1244         DstReg = MI.getOperand(1).getReg();
1245         const TargetRegisterClass *SrcRC = MRI->getRegClassOrNull(SrcReg);
1246         const TargetRegisterClass *DstRC = MRI->getRegClassOrNull(DstReg);
1247         if (DstRC != SrcRC)
1248           break;
1249 
1250         uint64_t AndImm = MI.getOperand(2).getImm();
1251         if (MI.getOpcode() == PPC::ANDIS_rec ||
1252             MI.getOpcode() == PPC::ANDIS8_rec)
1253           AndImm <<= 16;
1254         uint64_t LZeroAndImm = llvm::countl_zero<uint64_t>(AndImm);
1255         uint64_t RZeroAndImm = llvm::countr_zero<uint64_t>(AndImm);
1256         uint64_t ImmSrc = SrcMI->getOperand(3).getImm();
1257 
1258         // We can transfer `RLDICL/RLDICR + ANDI_rec/ANDIS_rec` to `ANDI_rec 0`
1259         // if all bits to AND are already zero in the input.
1260         bool PatternResultZero =
1261             (SrcOpCode == PPC::RLDICL && (RZeroAndImm + ImmSrc > 63)) ||
1262             (SrcOpCode == PPC::RLDICR && LZeroAndImm > ImmSrc);
1263 
1264         // We can eliminate RLDICL/RLDICR if it's used to clear bits and all
1265         // bits cleared will be ANDed with 0 by ANDI_rec/ANDIS_rec.
1266         bool PatternRemoveRotate =
1267             SrcMI->getOperand(2).getImm() == 0 &&
1268             ((SrcOpCode == PPC::RLDICL && LZeroAndImm >= ImmSrc) ||
1269              (SrcOpCode == PPC::RLDICR && (RZeroAndImm + ImmSrc > 63)));
1270 
1271         if (!PatternResultZero && !PatternRemoveRotate)
1272           break;
1273 
1274         LLVM_DEBUG(dbgs() << "Combining pair: ");
1275         LLVM_DEBUG(SrcMI->dump());
1276         LLVM_DEBUG(MI.dump());
1277         if (PatternResultZero)
1278           MI.getOperand(2).setImm(0);
1279         MI.getOperand(1).setReg(SrcMI->getOperand(1).getReg());
1280         LLVM_DEBUG(dbgs() << "To: ");
1281         LLVM_DEBUG(MI.dump());
1282         addRegToUpdate(MI.getOperand(1).getReg());
1283         addRegToUpdate(SrcMI->getOperand(0).getReg());
1284         Simplified = true;
1285         break;
1286       }
1287       case PPC::RLWINM:
1288       case PPC::RLWINM_rec:
1289       case PPC::RLWINM8:
1290       case PPC::RLWINM8_rec: {
1291         // We might replace operand 1 of the instruction which will
1292         // require we recompute kill flags for it.
1293         Register OrigOp1Reg = MI.getOperand(1).isReg()
1294                                   ? MI.getOperand(1).getReg()
1295                                   : PPC::NoRegister;
1296         Simplified = TII->combineRLWINM(MI, &ToErase);
1297         if (Simplified) {
1298           addRegToUpdate(OrigOp1Reg);
1299           if (MI.getOperand(1).isReg())
1300             addRegToUpdate(MI.getOperand(1).getReg());
1301           if (ToErase && ToErase->getOperand(1).isReg())
1302             for (auto UseReg : ToErase->explicit_uses())
1303               if (UseReg.isReg())
1304                 addRegToUpdate(UseReg.getReg());
1305           ++NumRotatesCollapsed;
1306         }
1307         break;
1308       }
1309       // We will replace TD/TW/TDI/TWI with an unconditional trap if it will
1310       // always trap, we will delete the node if it will never trap.
1311       case PPC::TDI:
1312       case PPC::TWI:
1313       case PPC::TD:
1314       case PPC::TW: {
1315         if (!EnableTrapOptimization) break;
1316         MachineInstr *LiMI1 = getVRegDefOrNull(&MI.getOperand(1), MRI);
1317         MachineInstr *LiMI2 = getVRegDefOrNull(&MI.getOperand(2), MRI);
1318         bool IsOperand2Immediate = MI.getOperand(2).isImm();
1319         // We can only do the optimization if we can get immediates
1320         // from both operands
1321         if (!(LiMI1 && (LiMI1->getOpcode() == PPC::LI ||
1322                         LiMI1->getOpcode() == PPC::LI8)))
1323           break;
1324         if (!IsOperand2Immediate &&
1325             !(LiMI2 && (LiMI2->getOpcode() == PPC::LI ||
1326                         LiMI2->getOpcode() == PPC::LI8)))
1327           break;
1328 
1329         auto ImmOperand0 = MI.getOperand(0).getImm();
1330         auto ImmOperand1 = LiMI1->getOperand(1).getImm();
1331         auto ImmOperand2 = IsOperand2Immediate ? MI.getOperand(2).getImm()
1332                                                : LiMI2->getOperand(1).getImm();
1333 
1334         // We will replace the MI with an unconditional trap if it will always
1335         // trap.
1336         if ((ImmOperand0 == 31) ||
1337             ((ImmOperand0 & 0x10) &&
1338              ((int64_t)ImmOperand1 < (int64_t)ImmOperand2)) ||
1339             ((ImmOperand0 & 0x8) &&
1340              ((int64_t)ImmOperand1 > (int64_t)ImmOperand2)) ||
1341             ((ImmOperand0 & 0x2) &&
1342              ((uint64_t)ImmOperand1 < (uint64_t)ImmOperand2)) ||
1343             ((ImmOperand0 & 0x1) &&
1344              ((uint64_t)ImmOperand1 > (uint64_t)ImmOperand2)) ||
1345             ((ImmOperand0 & 0x4) && (ImmOperand1 == ImmOperand2))) {
1346           BuildMI(MBB, &MI, MI.getDebugLoc(), TII->get(PPC::TRAP));
1347           TrapOpt = true;
1348         }
1349         // We will delete the MI if it will never trap.
1350         ToErase = &MI;
1351         Simplified = true;
1352         break;
1353       }
1354       }
1355     }
1356 
1357     // If the last instruction was marked for elimination,
1358     // remove it now.
1359     if (ToErase) {
1360       recomputeLVForDyingInstr();
1361       ToErase->eraseFromParent();
1362       ToErase = nullptr;
1363     }
1364     // Reset TrapOpt to false at the end of the basic block.
1365     if (EnableTrapOptimization)
1366       TrapOpt = false;
1367   }
1368 
1369   // Eliminate all the TOC save instructions which are redundant.
1370   Simplified |= eliminateRedundantTOCSaves(TOCSaves);
1371   PPCFunctionInfo *FI = MF->getInfo<PPCFunctionInfo>();
1372   if (FI->mustSaveTOC())
1373     NumTOCSavesInPrologue++;
1374 
1375   // We try to eliminate redundant compare instruction.
1376   Simplified |= eliminateRedundantCompare();
1377 
1378   // If we have made any modifications and added any registers to the set of
1379   // registers for which we need to update the kill flags, do so by recomputing
1380   // LiveVariables for those registers.
1381   for (Register Reg : RegsToUpdate) {
1382     if (!MRI->reg_empty(Reg))
1383       LV->recomputeForSingleDefVirtReg(Reg);
1384   }
1385   return Simplified;
1386 }
1387 
1388 // helper functions for eliminateRedundantCompare
1389 static bool isEqOrNe(MachineInstr *BI) {
1390   PPC::Predicate Pred = (PPC::Predicate)BI->getOperand(0).getImm();
1391   unsigned PredCond = PPC::getPredicateCondition(Pred);
1392   return (PredCond == PPC::PRED_EQ || PredCond == PPC::PRED_NE);
1393 }
1394 
1395 static bool isSupportedCmpOp(unsigned opCode) {
1396   return (opCode == PPC::CMPLD  || opCode == PPC::CMPD  ||
1397           opCode == PPC::CMPLW  || opCode == PPC::CMPW  ||
1398           opCode == PPC::CMPLDI || opCode == PPC::CMPDI ||
1399           opCode == PPC::CMPLWI || opCode == PPC::CMPWI);
1400 }
1401 
1402 static bool is64bitCmpOp(unsigned opCode) {
1403   return (opCode == PPC::CMPLD  || opCode == PPC::CMPD ||
1404           opCode == PPC::CMPLDI || opCode == PPC::CMPDI);
1405 }
1406 
1407 static bool isSignedCmpOp(unsigned opCode) {
1408   return (opCode == PPC::CMPD  || opCode == PPC::CMPW ||
1409           opCode == PPC::CMPDI || opCode == PPC::CMPWI);
1410 }
1411 
1412 static unsigned getSignedCmpOpCode(unsigned opCode) {
1413   if (opCode == PPC::CMPLD)  return PPC::CMPD;
1414   if (opCode == PPC::CMPLW)  return PPC::CMPW;
1415   if (opCode == PPC::CMPLDI) return PPC::CMPDI;
1416   if (opCode == PPC::CMPLWI) return PPC::CMPWI;
1417   return opCode;
1418 }
1419 
1420 // We can decrement immediate x in (GE x) by changing it to (GT x-1) or
1421 // (LT x) to (LE x-1)
1422 static unsigned getPredicateToDecImm(MachineInstr *BI, MachineInstr *CMPI) {
1423   uint64_t Imm = CMPI->getOperand(2).getImm();
1424   bool SignedCmp = isSignedCmpOp(CMPI->getOpcode());
1425   if ((!SignedCmp && Imm == 0) || (SignedCmp && Imm == 0x8000))
1426     return 0;
1427 
1428   PPC::Predicate Pred = (PPC::Predicate)BI->getOperand(0).getImm();
1429   unsigned PredCond = PPC::getPredicateCondition(Pred);
1430   unsigned PredHint = PPC::getPredicateHint(Pred);
1431   if (PredCond == PPC::PRED_GE)
1432     return PPC::getPredicate(PPC::PRED_GT, PredHint);
1433   if (PredCond == PPC::PRED_LT)
1434     return PPC::getPredicate(PPC::PRED_LE, PredHint);
1435 
1436   return 0;
1437 }
1438 
1439 // We can increment immediate x in (GT x) by changing it to (GE x+1) or
1440 // (LE x) to (LT x+1)
1441 static unsigned getPredicateToIncImm(MachineInstr *BI, MachineInstr *CMPI) {
1442   uint64_t Imm = CMPI->getOperand(2).getImm();
1443   bool SignedCmp = isSignedCmpOp(CMPI->getOpcode());
1444   if ((!SignedCmp && Imm == 0xFFFF) || (SignedCmp && Imm == 0x7FFF))
1445     return 0;
1446 
1447   PPC::Predicate Pred = (PPC::Predicate)BI->getOperand(0).getImm();
1448   unsigned PredCond = PPC::getPredicateCondition(Pred);
1449   unsigned PredHint = PPC::getPredicateHint(Pred);
1450   if (PredCond == PPC::PRED_GT)
1451     return PPC::getPredicate(PPC::PRED_GE, PredHint);
1452   if (PredCond == PPC::PRED_LE)
1453     return PPC::getPredicate(PPC::PRED_LT, PredHint);
1454 
1455   return 0;
1456 }
1457 
1458 // This takes a Phi node and returns a register value for the specified BB.
1459 static unsigned getIncomingRegForBlock(MachineInstr *Phi,
1460                                        MachineBasicBlock *MBB) {
1461   for (unsigned I = 2, E = Phi->getNumOperands() + 1; I != E; I += 2) {
1462     MachineOperand &MO = Phi->getOperand(I);
1463     if (MO.getMBB() == MBB)
1464       return Phi->getOperand(I-1).getReg();
1465   }
1466   llvm_unreachable("invalid src basic block for this Phi node\n");
1467   return 0;
1468 }
1469 
1470 // This function tracks the source of the register through register copy.
1471 // If BB1 and BB2 are non-NULL, we also track PHI instruction in BB2
1472 // assuming that the control comes from BB1 into BB2.
1473 static unsigned getSrcVReg(unsigned Reg, MachineBasicBlock *BB1,
1474                            MachineBasicBlock *BB2, MachineRegisterInfo *MRI) {
1475   unsigned SrcReg = Reg;
1476   while (true) {
1477     unsigned NextReg = SrcReg;
1478     MachineInstr *Inst = MRI->getVRegDef(SrcReg);
1479     if (BB1 && Inst->getOpcode() == PPC::PHI && Inst->getParent() == BB2) {
1480       NextReg = getIncomingRegForBlock(Inst, BB1);
1481       // We track through PHI only once to avoid infinite loop.
1482       BB1 = nullptr;
1483     }
1484     else if (Inst->isFullCopy())
1485       NextReg = Inst->getOperand(1).getReg();
1486     if (NextReg == SrcReg || !Register::isVirtualRegister(NextReg))
1487       break;
1488     SrcReg = NextReg;
1489   }
1490   return SrcReg;
1491 }
1492 
1493 static bool eligibleForCompareElimination(MachineBasicBlock &MBB,
1494                                           MachineBasicBlock *&PredMBB,
1495                                           MachineBasicBlock *&MBBtoMoveCmp,
1496                                           MachineRegisterInfo *MRI) {
1497 
1498   auto isEligibleBB = [&](MachineBasicBlock &BB) {
1499     auto BII = BB.getFirstInstrTerminator();
1500     // We optimize BBs ending with a conditional branch.
1501     // We check only for BCC here, not BCCLR, because BCCLR
1502     // will be formed only later in the pipeline.
1503     if (BB.succ_size() == 2 &&
1504         BII != BB.instr_end() &&
1505         (*BII).getOpcode() == PPC::BCC &&
1506         (*BII).getOperand(1).isReg()) {
1507       // We optimize only if the condition code is used only by one BCC.
1508       Register CndReg = (*BII).getOperand(1).getReg();
1509       if (!CndReg.isVirtual() || !MRI->hasOneNonDBGUse(CndReg))
1510         return false;
1511 
1512       MachineInstr *CMPI = MRI->getVRegDef(CndReg);
1513       // We assume compare and branch are in the same BB for ease of analysis.
1514       if (CMPI->getParent() != &BB)
1515         return false;
1516 
1517       // We skip this BB if a physical register is used in comparison.
1518       for (MachineOperand &MO : CMPI->operands())
1519         if (MO.isReg() && !MO.getReg().isVirtual())
1520           return false;
1521 
1522       return true;
1523     }
1524     return false;
1525   };
1526 
1527   // If this BB has more than one successor, we can create a new BB and
1528   // move the compare instruction in the new BB.
1529   // So far, we do not move compare instruction to a BB having multiple
1530   // successors to avoid potentially increasing code size.
1531   auto isEligibleForMoveCmp = [](MachineBasicBlock &BB) {
1532     return BB.succ_size() == 1;
1533   };
1534 
1535   if (!isEligibleBB(MBB))
1536     return false;
1537 
1538   unsigned NumPredBBs = MBB.pred_size();
1539   if (NumPredBBs == 1) {
1540     MachineBasicBlock *TmpMBB = *MBB.pred_begin();
1541     if (isEligibleBB(*TmpMBB)) {
1542       PredMBB = TmpMBB;
1543       MBBtoMoveCmp = nullptr;
1544       return true;
1545     }
1546   }
1547   else if (NumPredBBs == 2) {
1548     // We check for partially redundant case.
1549     // So far, we support cases with only two predecessors
1550     // to avoid increasing the number of instructions.
1551     MachineBasicBlock::pred_iterator PI = MBB.pred_begin();
1552     MachineBasicBlock *Pred1MBB = *PI;
1553     MachineBasicBlock *Pred2MBB = *(PI+1);
1554 
1555     if (isEligibleBB(*Pred1MBB) && isEligibleForMoveCmp(*Pred2MBB)) {
1556       // We assume Pred1MBB is the BB containing the compare to be merged and
1557       // Pred2MBB is the BB to which we will append a compare instruction.
1558       // Proceed as is if Pred1MBB is different from MBB.
1559     }
1560     else if (isEligibleBB(*Pred2MBB) && isEligibleForMoveCmp(*Pred1MBB)) {
1561       // We need to swap Pred1MBB and Pred2MBB to canonicalize.
1562       std::swap(Pred1MBB, Pred2MBB);
1563     }
1564     else return false;
1565 
1566     if (Pred1MBB == &MBB)
1567       return false;
1568 
1569     // Here, Pred2MBB is the BB to which we need to append a compare inst.
1570     // We cannot move the compare instruction if operands are not available
1571     // in Pred2MBB (i.e. defined in MBB by an instruction other than PHI).
1572     MachineInstr *BI = &*MBB.getFirstInstrTerminator();
1573     MachineInstr *CMPI = MRI->getVRegDef(BI->getOperand(1).getReg());
1574     for (int I = 1; I <= 2; I++)
1575       if (CMPI->getOperand(I).isReg()) {
1576         MachineInstr *Inst = MRI->getVRegDef(CMPI->getOperand(I).getReg());
1577         if (Inst->getParent() == &MBB && Inst->getOpcode() != PPC::PHI)
1578           return false;
1579       }
1580 
1581     PredMBB = Pred1MBB;
1582     MBBtoMoveCmp = Pred2MBB;
1583     return true;
1584   }
1585 
1586   return false;
1587 }
1588 
1589 // This function will iterate over the input map containing a pair of TOC save
1590 // instruction and a flag. The flag will be set to false if the TOC save is
1591 // proven redundant. This function will erase from the basic block all the TOC
1592 // saves marked as redundant.
1593 bool PPCMIPeephole::eliminateRedundantTOCSaves(
1594     std::map<MachineInstr *, bool> &TOCSaves) {
1595   bool Simplified = false;
1596   int NumKept = 0;
1597   for (auto TOCSave : TOCSaves) {
1598     if (!TOCSave.second) {
1599       TOCSave.first->eraseFromParent();
1600       RemoveTOCSave++;
1601       Simplified = true;
1602     } else {
1603       NumKept++;
1604     }
1605   }
1606 
1607   if (NumKept > 1)
1608     MultiTOCSaves++;
1609 
1610   return Simplified;
1611 }
1612 
1613 // If multiple conditional branches are executed based on the (essentially)
1614 // same comparison, we merge compare instructions into one and make multiple
1615 // conditional branches on this comparison.
1616 // For example,
1617 //   if (a == 0) { ... }
1618 //   else if (a < 0) { ... }
1619 // can be executed by one compare and two conditional branches instead of
1620 // two pairs of a compare and a conditional branch.
1621 //
1622 // This method merges two compare instructions in two MBBs and modifies the
1623 // compare and conditional branch instructions if needed.
1624 // For the above example, the input for this pass looks like:
1625 //   cmplwi r3, 0
1626 //   beq    0, .LBB0_3
1627 //   cmpwi  r3, -1
1628 //   bgt    0, .LBB0_4
1629 // So, before merging two compares, we need to modify these instructions as
1630 //   cmpwi  r3, 0       ; cmplwi and cmpwi yield same result for beq
1631 //   beq    0, .LBB0_3
1632 //   cmpwi  r3, 0       ; greather than -1 means greater or equal to 0
1633 //   bge    0, .LBB0_4
1634 
1635 bool PPCMIPeephole::eliminateRedundantCompare() {
1636   bool Simplified = false;
1637 
1638   for (MachineBasicBlock &MBB2 : *MF) {
1639     MachineBasicBlock *MBB1 = nullptr, *MBBtoMoveCmp = nullptr;
1640 
1641     // For fully redundant case, we select two basic blocks MBB1 and MBB2
1642     // as an optimization target if
1643     // - both MBBs end with a conditional branch,
1644     // - MBB1 is the only predecessor of MBB2, and
1645     // - compare does not take a physical register as a operand in both MBBs.
1646     // In this case, eligibleForCompareElimination sets MBBtoMoveCmp nullptr.
1647     //
1648     // As partially redundant case, we additionally handle if MBB2 has one
1649     // additional predecessor, which has only one successor (MBB2).
1650     // In this case, we move the compare instruction originally in MBB2 into
1651     // MBBtoMoveCmp. This partially redundant case is typically appear by
1652     // compiling a while loop; here, MBBtoMoveCmp is the loop preheader.
1653     //
1654     // Overview of CFG of related basic blocks
1655     // Fully redundant case        Partially redundant case
1656     //   --------                   ----------------  --------
1657     //   | MBB1 | (w/ 2 succ)       | MBBtoMoveCmp |  | MBB1 | (w/ 2 succ)
1658     //   --------                   ----------------  --------
1659     //      |    \                     (w/ 1 succ) \     |    \
1660     //      |     \                                 \    |     \
1661     //      |                                        \   |
1662     //   --------                                     --------
1663     //   | MBB2 | (w/ 1 pred                          | MBB2 | (w/ 2 pred
1664     //   -------- and 2 succ)                         -------- and 2 succ)
1665     //      |    \                                       |    \
1666     //      |     \                                      |     \
1667     //
1668     if (!eligibleForCompareElimination(MBB2, MBB1, MBBtoMoveCmp, MRI))
1669       continue;
1670 
1671     MachineInstr *BI1   = &*MBB1->getFirstInstrTerminator();
1672     MachineInstr *CMPI1 = MRI->getVRegDef(BI1->getOperand(1).getReg());
1673 
1674     MachineInstr *BI2   = &*MBB2.getFirstInstrTerminator();
1675     MachineInstr *CMPI2 = MRI->getVRegDef(BI2->getOperand(1).getReg());
1676     bool IsPartiallyRedundant = (MBBtoMoveCmp != nullptr);
1677 
1678     // We cannot optimize an unsupported compare opcode or
1679     // a mix of 32-bit and 64-bit comparisons
1680     if (!isSupportedCmpOp(CMPI1->getOpcode()) ||
1681         !isSupportedCmpOp(CMPI2->getOpcode()) ||
1682         is64bitCmpOp(CMPI1->getOpcode()) != is64bitCmpOp(CMPI2->getOpcode()))
1683       continue;
1684 
1685     unsigned NewOpCode = 0;
1686     unsigned NewPredicate1 = 0, NewPredicate2 = 0;
1687     int16_t Imm1 = 0, NewImm1 = 0, Imm2 = 0, NewImm2 = 0;
1688     bool SwapOperands = false;
1689 
1690     if (CMPI1->getOpcode() != CMPI2->getOpcode()) {
1691       // Typically, unsigned comparison is used for equality check, but
1692       // we replace it with a signed comparison if the comparison
1693       // to be merged is a signed comparison.
1694       // In other cases of opcode mismatch, we cannot optimize this.
1695 
1696       // We cannot change opcode when comparing against an immediate
1697       // if the most significant bit of the immediate is one
1698       // due to the difference in sign extension.
1699       auto CmpAgainstImmWithSignBit = [](MachineInstr *I) {
1700         if (!I->getOperand(2).isImm())
1701           return false;
1702         int16_t Imm = (int16_t)I->getOperand(2).getImm();
1703         return Imm < 0;
1704       };
1705 
1706       if (isEqOrNe(BI2) && !CmpAgainstImmWithSignBit(CMPI2) &&
1707           CMPI1->getOpcode() == getSignedCmpOpCode(CMPI2->getOpcode()))
1708         NewOpCode = CMPI1->getOpcode();
1709       else if (isEqOrNe(BI1) && !CmpAgainstImmWithSignBit(CMPI1) &&
1710                getSignedCmpOpCode(CMPI1->getOpcode()) == CMPI2->getOpcode())
1711         NewOpCode = CMPI2->getOpcode();
1712       else continue;
1713     }
1714 
1715     if (CMPI1->getOperand(2).isReg() && CMPI2->getOperand(2).isReg()) {
1716       // In case of comparisons between two registers, these two registers
1717       // must be same to merge two comparisons.
1718       unsigned Cmp1Operand1 = getSrcVReg(CMPI1->getOperand(1).getReg(),
1719                                          nullptr, nullptr, MRI);
1720       unsigned Cmp1Operand2 = getSrcVReg(CMPI1->getOperand(2).getReg(),
1721                                          nullptr, nullptr, MRI);
1722       unsigned Cmp2Operand1 = getSrcVReg(CMPI2->getOperand(1).getReg(),
1723                                          MBB1, &MBB2, MRI);
1724       unsigned Cmp2Operand2 = getSrcVReg(CMPI2->getOperand(2).getReg(),
1725                                          MBB1, &MBB2, MRI);
1726 
1727       if (Cmp1Operand1 == Cmp2Operand1 && Cmp1Operand2 == Cmp2Operand2) {
1728         // Same pair of registers in the same order; ready to merge as is.
1729       }
1730       else if (Cmp1Operand1 == Cmp2Operand2 && Cmp1Operand2 == Cmp2Operand1) {
1731         // Same pair of registers in different order.
1732         // We reverse the predicate to merge compare instructions.
1733         PPC::Predicate Pred = (PPC::Predicate)BI2->getOperand(0).getImm();
1734         NewPredicate2 = (unsigned)PPC::getSwappedPredicate(Pred);
1735         // In case of partial redundancy, we need to swap operands
1736         // in another compare instruction.
1737         SwapOperands = true;
1738       }
1739       else continue;
1740     }
1741     else if (CMPI1->getOperand(2).isImm() && CMPI2->getOperand(2).isImm()) {
1742       // In case of comparisons between a register and an immediate,
1743       // the operand register must be same for two compare instructions.
1744       unsigned Cmp1Operand1 = getSrcVReg(CMPI1->getOperand(1).getReg(),
1745                                          nullptr, nullptr, MRI);
1746       unsigned Cmp2Operand1 = getSrcVReg(CMPI2->getOperand(1).getReg(),
1747                                          MBB1, &MBB2, MRI);
1748       if (Cmp1Operand1 != Cmp2Operand1)
1749         continue;
1750 
1751       NewImm1 = Imm1 = (int16_t)CMPI1->getOperand(2).getImm();
1752       NewImm2 = Imm2 = (int16_t)CMPI2->getOperand(2).getImm();
1753 
1754       // If immediate are not same, we try to adjust by changing predicate;
1755       // e.g. GT imm means GE (imm+1).
1756       if (Imm1 != Imm2 && (!isEqOrNe(BI2) || !isEqOrNe(BI1))) {
1757         int Diff = Imm1 - Imm2;
1758         if (Diff < -2 || Diff > 2)
1759           continue;
1760 
1761         unsigned PredToInc1 = getPredicateToIncImm(BI1, CMPI1);
1762         unsigned PredToDec1 = getPredicateToDecImm(BI1, CMPI1);
1763         unsigned PredToInc2 = getPredicateToIncImm(BI2, CMPI2);
1764         unsigned PredToDec2 = getPredicateToDecImm(BI2, CMPI2);
1765         if (Diff == 2) {
1766           if (PredToInc2 && PredToDec1) {
1767             NewPredicate2 = PredToInc2;
1768             NewPredicate1 = PredToDec1;
1769             NewImm2++;
1770             NewImm1--;
1771           }
1772         }
1773         else if (Diff == 1) {
1774           if (PredToInc2) {
1775             NewImm2++;
1776             NewPredicate2 = PredToInc2;
1777           }
1778           else if (PredToDec1) {
1779             NewImm1--;
1780             NewPredicate1 = PredToDec1;
1781           }
1782         }
1783         else if (Diff == -1) {
1784           if (PredToDec2) {
1785             NewImm2--;
1786             NewPredicate2 = PredToDec2;
1787           }
1788           else if (PredToInc1) {
1789             NewImm1++;
1790             NewPredicate1 = PredToInc1;
1791           }
1792         }
1793         else if (Diff == -2) {
1794           if (PredToDec2 && PredToInc1) {
1795             NewPredicate2 = PredToDec2;
1796             NewPredicate1 = PredToInc1;
1797             NewImm2--;
1798             NewImm1++;
1799           }
1800         }
1801       }
1802 
1803       // We cannot merge two compares if the immediates are not same.
1804       if (NewImm2 != NewImm1)
1805         continue;
1806     }
1807 
1808     LLVM_DEBUG(dbgs() << "Optimize two pairs of compare and branch:\n");
1809     LLVM_DEBUG(CMPI1->dump());
1810     LLVM_DEBUG(BI1->dump());
1811     LLVM_DEBUG(CMPI2->dump());
1812     LLVM_DEBUG(BI2->dump());
1813     for (const MachineOperand &MO : CMPI1->operands())
1814       if (MO.isReg())
1815         addRegToUpdate(MO.getReg());
1816     for (const MachineOperand &MO : CMPI2->operands())
1817       if (MO.isReg())
1818         addRegToUpdate(MO.getReg());
1819 
1820     // We adjust opcode, predicates and immediate as we determined above.
1821     if (NewOpCode != 0 && NewOpCode != CMPI1->getOpcode()) {
1822       CMPI1->setDesc(TII->get(NewOpCode));
1823     }
1824     if (NewPredicate1) {
1825       BI1->getOperand(0).setImm(NewPredicate1);
1826     }
1827     if (NewPredicate2) {
1828       BI2->getOperand(0).setImm(NewPredicate2);
1829     }
1830     if (NewImm1 != Imm1) {
1831       CMPI1->getOperand(2).setImm(NewImm1);
1832     }
1833 
1834     if (IsPartiallyRedundant) {
1835       // We touch up the compare instruction in MBB2 and move it to
1836       // a previous BB to handle partially redundant case.
1837       if (SwapOperands) {
1838         Register Op1 = CMPI2->getOperand(1).getReg();
1839         Register Op2 = CMPI2->getOperand(2).getReg();
1840         CMPI2->getOperand(1).setReg(Op2);
1841         CMPI2->getOperand(2).setReg(Op1);
1842       }
1843       if (NewImm2 != Imm2)
1844         CMPI2->getOperand(2).setImm(NewImm2);
1845 
1846       for (int I = 1; I <= 2; I++) {
1847         if (CMPI2->getOperand(I).isReg()) {
1848           MachineInstr *Inst = MRI->getVRegDef(CMPI2->getOperand(I).getReg());
1849           if (Inst->getParent() != &MBB2)
1850             continue;
1851 
1852           assert(Inst->getOpcode() == PPC::PHI &&
1853                  "We cannot support if an operand comes from this BB.");
1854           unsigned SrcReg = getIncomingRegForBlock(Inst, MBBtoMoveCmp);
1855           CMPI2->getOperand(I).setReg(SrcReg);
1856           addRegToUpdate(SrcReg);
1857         }
1858       }
1859       auto I = MachineBasicBlock::iterator(MBBtoMoveCmp->getFirstTerminator());
1860       MBBtoMoveCmp->splice(I, &MBB2, MachineBasicBlock::iterator(CMPI2));
1861 
1862       DebugLoc DL = CMPI2->getDebugLoc();
1863       Register NewVReg = MRI->createVirtualRegister(&PPC::CRRCRegClass);
1864       BuildMI(MBB2, MBB2.begin(), DL,
1865               TII->get(PPC::PHI), NewVReg)
1866         .addReg(BI1->getOperand(1).getReg()).addMBB(MBB1)
1867         .addReg(BI2->getOperand(1).getReg()).addMBB(MBBtoMoveCmp);
1868       BI2->getOperand(1).setReg(NewVReg);
1869       addRegToUpdate(NewVReg);
1870     }
1871     else {
1872       // We finally eliminate compare instruction in MBB2.
1873       // We do not need to treat CMPI2 specially here in terms of re-computing
1874       // live variables even though it is being deleted because:
1875       // - It defines a register that has a single use (already checked in
1876       // eligibleForCompareElimination())
1877       // - The only user (BI2) is no longer using it so the register is dead (no
1878       // def, no uses)
1879       // - We do not attempt to recompute live variables for dead registers
1880       BI2->getOperand(1).setReg(BI1->getOperand(1).getReg());
1881       CMPI2->eraseFromParent();
1882     }
1883 
1884     LLVM_DEBUG(dbgs() << "into a compare and two branches:\n");
1885     LLVM_DEBUG(CMPI1->dump());
1886     LLVM_DEBUG(BI1->dump());
1887     LLVM_DEBUG(BI2->dump());
1888     if (IsPartiallyRedundant) {
1889       LLVM_DEBUG(dbgs() << "The following compare is moved into "
1890                         << printMBBReference(*MBBtoMoveCmp)
1891                         << " to handle partial redundancy.\n");
1892       LLVM_DEBUG(CMPI2->dump());
1893     }
1894     Simplified = true;
1895   }
1896 
1897   return Simplified;
1898 }
1899 
1900 // We miss the opportunity to emit an RLDIC when lowering jump tables
1901 // since ISEL sees only a single basic block. When selecting, the clear
1902 // and shift left will be in different blocks.
1903 bool PPCMIPeephole::emitRLDICWhenLoweringJumpTables(MachineInstr &MI,
1904                                                     MachineInstr *&ToErase) {
1905   if (MI.getOpcode() != PPC::RLDICR)
1906     return false;
1907 
1908   Register SrcReg = MI.getOperand(1).getReg();
1909   if (!SrcReg.isVirtual())
1910     return false;
1911 
1912   MachineInstr *SrcMI = MRI->getVRegDef(SrcReg);
1913   if (SrcMI->getOpcode() != PPC::RLDICL)
1914     return false;
1915 
1916   MachineOperand MOpSHSrc = SrcMI->getOperand(2);
1917   MachineOperand MOpMBSrc = SrcMI->getOperand(3);
1918   MachineOperand MOpSHMI = MI.getOperand(2);
1919   MachineOperand MOpMEMI = MI.getOperand(3);
1920   if (!(MOpSHSrc.isImm() && MOpMBSrc.isImm() && MOpSHMI.isImm() &&
1921         MOpMEMI.isImm()))
1922     return false;
1923 
1924   uint64_t SHSrc = MOpSHSrc.getImm();
1925   uint64_t MBSrc = MOpMBSrc.getImm();
1926   uint64_t SHMI = MOpSHMI.getImm();
1927   uint64_t MEMI = MOpMEMI.getImm();
1928   uint64_t NewSH = SHSrc + SHMI;
1929   uint64_t NewMB = MBSrc - SHMI;
1930   if (NewMB > 63 || NewSH > 63)
1931     return false;
1932 
1933   // The bits cleared with RLDICL are [0, MBSrc).
1934   // The bits cleared with RLDICR are (MEMI, 63].
1935   // After the sequence, the bits cleared are:
1936   // [0, MBSrc-SHMI) and (MEMI, 63).
1937   //
1938   // The bits cleared with RLDIC are [0, NewMB) and (63-NewSH, 63].
1939   if ((63 - NewSH) != MEMI)
1940     return false;
1941 
1942   LLVM_DEBUG(dbgs() << "Converting pair: ");
1943   LLVM_DEBUG(SrcMI->dump());
1944   LLVM_DEBUG(MI.dump());
1945 
1946   MI.setDesc(TII->get(PPC::RLDIC));
1947   MI.getOperand(1).setReg(SrcMI->getOperand(1).getReg());
1948   MI.getOperand(2).setImm(NewSH);
1949   MI.getOperand(3).setImm(NewMB);
1950   addRegToUpdate(MI.getOperand(1).getReg());
1951   addRegToUpdate(SrcMI->getOperand(0).getReg());
1952 
1953   LLVM_DEBUG(dbgs() << "To: ");
1954   LLVM_DEBUG(MI.dump());
1955   NumRotatesCollapsed++;
1956   // If SrcReg has no non-debug use it's safe to delete its def SrcMI.
1957   if (MRI->use_nodbg_empty(SrcReg)) {
1958     assert(!SrcMI->hasImplicitDef() &&
1959            "Not expecting an implicit def with this instr.");
1960     ToErase = SrcMI;
1961   }
1962   return true;
1963 }
1964 
1965 // For case in LLVM IR
1966 // entry:
1967 //   %iconv = sext i32 %index to i64
1968 //   br i1 undef label %true, label %false
1969 // true:
1970 //   %ptr = getelementptr inbounds i32, i32* null, i64 %iconv
1971 // ...
1972 // PPCISelLowering::combineSHL fails to combine, because sext and shl are in
1973 // different BBs when conducting instruction selection. We can do a peephole
1974 // optimization to combine these two instructions into extswsli after
1975 // instruction selection.
1976 bool PPCMIPeephole::combineSEXTAndSHL(MachineInstr &MI,
1977                                       MachineInstr *&ToErase) {
1978   if (MI.getOpcode() != PPC::RLDICR)
1979     return false;
1980 
1981   if (!MF->getSubtarget<PPCSubtarget>().isISA3_0())
1982     return false;
1983 
1984   assert(MI.getNumOperands() == 4 && "RLDICR should have 4 operands");
1985 
1986   MachineOperand MOpSHMI = MI.getOperand(2);
1987   MachineOperand MOpMEMI = MI.getOperand(3);
1988   if (!(MOpSHMI.isImm() && MOpMEMI.isImm()))
1989     return false;
1990 
1991   uint64_t SHMI = MOpSHMI.getImm();
1992   uint64_t MEMI = MOpMEMI.getImm();
1993   if (SHMI + MEMI != 63)
1994     return false;
1995 
1996   Register SrcReg = MI.getOperand(1).getReg();
1997   if (!SrcReg.isVirtual())
1998     return false;
1999 
2000   MachineInstr *SrcMI = MRI->getVRegDef(SrcReg);
2001   if (SrcMI->getOpcode() != PPC::EXTSW &&
2002       SrcMI->getOpcode() != PPC::EXTSW_32_64)
2003     return false;
2004 
2005   // If the register defined by extsw has more than one use, combination is not
2006   // needed.
2007   if (!MRI->hasOneNonDBGUse(SrcReg))
2008     return false;
2009 
2010   assert(SrcMI->getNumOperands() == 2 && "EXTSW should have 2 operands");
2011   assert(SrcMI->getOperand(1).isReg() &&
2012          "EXTSW's second operand should be a register");
2013   if (!SrcMI->getOperand(1).getReg().isVirtual())
2014     return false;
2015 
2016   LLVM_DEBUG(dbgs() << "Combining pair: ");
2017   LLVM_DEBUG(SrcMI->dump());
2018   LLVM_DEBUG(MI.dump());
2019 
2020   MachineInstr *NewInstr =
2021       BuildMI(*MI.getParent(), &MI, MI.getDebugLoc(),
2022               SrcMI->getOpcode() == PPC::EXTSW ? TII->get(PPC::EXTSWSLI)
2023                                                : TII->get(PPC::EXTSWSLI_32_64),
2024               MI.getOperand(0).getReg())
2025           .add(SrcMI->getOperand(1))
2026           .add(MOpSHMI);
2027   (void)NewInstr;
2028 
2029   LLVM_DEBUG(dbgs() << "TO: ");
2030   LLVM_DEBUG(NewInstr->dump());
2031   ++NumEXTSWAndSLDICombined;
2032   ToErase = &MI;
2033   // SrcMI, which is extsw, is of no use now, but we don't erase it here so we
2034   // can recompute its kill flags. We run DCE immediately after this pass
2035   // to clean up dead instructions such as this.
2036   addRegToUpdate(NewInstr->getOperand(1).getReg());
2037   addRegToUpdate(SrcMI->getOperand(0).getReg());
2038   return true;
2039 }
2040 
2041 } // end default namespace
2042 
2043 INITIALIZE_PASS_BEGIN(PPCMIPeephole, DEBUG_TYPE,
2044                       "PowerPC MI Peephole Optimization", false, false)
2045 INITIALIZE_PASS_DEPENDENCY(MachineBlockFrequencyInfoWrapperPass)
2046 INITIALIZE_PASS_DEPENDENCY(MachineDominatorTreeWrapperPass)
2047 INITIALIZE_PASS_DEPENDENCY(MachinePostDominatorTreeWrapperPass)
2048 INITIALIZE_PASS_DEPENDENCY(LiveVariablesWrapperPass)
2049 INITIALIZE_PASS_END(PPCMIPeephole, DEBUG_TYPE,
2050                     "PowerPC MI Peephole Optimization", false, false)
2051 
2052 char PPCMIPeephole::ID = 0;
2053 FunctionPass*
2054 llvm::createPPCMIPeepholePass() { return new PPCMIPeephole(); }
2055