xref: /llvm-project/llvm/lib/Target/PowerPC/PPCMIPeephole.cpp (revision be95b4dec2604247d9fbad5f7641b4373b740904)
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 //===---------------------------------------------------------------------===//
20 
21 #include "MCTargetDesc/PPCMCTargetDesc.h"
22 #include "MCTargetDesc/PPCPredicates.h"
23 #include "PPC.h"
24 #include "PPCInstrBuilder.h"
25 #include "PPCInstrInfo.h"
26 #include "PPCMachineFunctionInfo.h"
27 #include "PPCTargetMachine.h"
28 #include "llvm/ADT/Statistic.h"
29 #include "llvm/CodeGen/MachineBlockFrequencyInfo.h"
30 #include "llvm/CodeGen/MachineDominators.h"
31 #include "llvm/CodeGen/MachineFrameInfo.h"
32 #include "llvm/CodeGen/MachineFunctionPass.h"
33 #include "llvm/CodeGen/MachineInstrBuilder.h"
34 #include "llvm/CodeGen/MachinePostDominators.h"
35 #include "llvm/CodeGen/MachineRegisterInfo.h"
36 #include "llvm/InitializePasses.h"
37 #include "llvm/Support/Debug.h"
38 
39 using namespace llvm;
40 
41 #define DEBUG_TYPE "ppc-mi-peepholes"
42 
43 STATISTIC(RemoveTOCSave, "Number of TOC saves removed");
44 STATISTIC(MultiTOCSaves,
45           "Number of functions with multiple TOC saves that must be kept");
46 STATISTIC(NumTOCSavesInPrologue, "Number of TOC saves placed in the prologue");
47 STATISTIC(NumEliminatedSExt, "Number of eliminated sign-extensions");
48 STATISTIC(NumEliminatedZExt, "Number of eliminated zero-extensions");
49 STATISTIC(NumOptADDLIs, "Number of optimized ADD instruction fed by LI");
50 STATISTIC(NumConvertedToImmediateForm,
51           "Number of instructions converted to their immediate form");
52 STATISTIC(NumFunctionsEnteredInMIPeephole,
53           "Number of functions entered in PPC MI Peepholes");
54 STATISTIC(NumFixedPointIterations,
55           "Number of fixed-point iterations converting reg-reg instructions "
56           "to reg-imm ones");
57 STATISTIC(NumRotatesCollapsed,
58           "Number of pairs of rotate left, clear left/right collapsed");
59 STATISTIC(NumEXTSWAndSLDICombined,
60           "Number of pairs of EXTSW and SLDI combined as EXTSWSLI");
61 STATISTIC(NumLoadImmZeroFoldedAndRemoved,
62           "Number of LI(8) reg, 0 that are folded to r0 and removed");
63 
64 static cl::opt<bool>
65 FixedPointRegToImm("ppc-reg-to-imm-fixed-point", cl::Hidden, cl::init(true),
66                    cl::desc("Iterate to a fixed point when attempting to "
67                             "convert reg-reg instructions to reg-imm"));
68 
69 static cl::opt<bool>
70 ConvertRegReg("ppc-convert-rr-to-ri", cl::Hidden, cl::init(true),
71               cl::desc("Convert eligible reg+reg instructions to reg+imm"));
72 
73 static cl::opt<bool>
74     EnableSExtElimination("ppc-eliminate-signext",
75                           cl::desc("enable elimination of sign-extensions"),
76                           cl::init(true), cl::Hidden);
77 
78 static cl::opt<bool>
79     EnableZExtElimination("ppc-eliminate-zeroext",
80                           cl::desc("enable elimination of zero-extensions"),
81                           cl::init(true), cl::Hidden);
82 
83 static cl::opt<bool>
84     EnableTrapOptimization("ppc-opt-conditional-trap",
85                            cl::desc("enable optimization of conditional traps"),
86                            cl::init(false), cl::Hidden);
87 
88 namespace {
89 
90 struct PPCMIPeephole : public MachineFunctionPass {
91 
92   static char ID;
93   const PPCInstrInfo *TII;
94   MachineFunction *MF;
95   MachineRegisterInfo *MRI;
96 
97   PPCMIPeephole() : MachineFunctionPass(ID) {
98     initializePPCMIPeepholePass(*PassRegistry::getPassRegistry());
99   }
100 
101 private:
102   MachineDominatorTree *MDT;
103   MachinePostDominatorTree *MPDT;
104   MachineBlockFrequencyInfo *MBFI;
105   uint64_t EntryFreq;
106 
107   // Initialize class variables.
108   void initialize(MachineFunction &MFParm);
109 
110   // Perform peepholes.
111   bool simplifyCode();
112 
113   // Perform peepholes.
114   bool eliminateRedundantCompare();
115   bool eliminateRedundantTOCSaves(std::map<MachineInstr *, bool> &TOCSaves);
116   bool combineSEXTAndSHL(MachineInstr &MI, MachineInstr *&ToErase);
117   bool emitRLDICWhenLoweringJumpTables(MachineInstr &MI);
118   void UpdateTOCSaves(std::map<MachineInstr *, bool> &TOCSaves,
119                       MachineInstr *MI);
120 
121 public:
122 
123   void getAnalysisUsage(AnalysisUsage &AU) const override {
124     AU.addRequired<MachineDominatorTree>();
125     AU.addRequired<MachinePostDominatorTree>();
126     AU.addRequired<MachineBlockFrequencyInfo>();
127     AU.addPreserved<MachineDominatorTree>();
128     AU.addPreserved<MachinePostDominatorTree>();
129     AU.addPreserved<MachineBlockFrequencyInfo>();
130     MachineFunctionPass::getAnalysisUsage(AU);
131   }
132 
133   // Main entry point for this pass.
134   bool runOnMachineFunction(MachineFunction &MF) override {
135     initialize(MF);
136     // At this point, TOC pointer should not be used in a function that uses
137     // PC-Relative addressing.
138     assert((MF.getRegInfo().use_empty(PPC::X2) ||
139             !MF.getSubtarget<PPCSubtarget>().isUsingPCRelativeCalls()) &&
140            "TOC pointer used in a function using PC-Relative addressing!");
141     if (skipFunction(MF.getFunction()))
142       return false;
143     return simplifyCode();
144   }
145 };
146 
147 // Initialize class variables.
148 void PPCMIPeephole::initialize(MachineFunction &MFParm) {
149   MF = &MFParm;
150   MRI = &MF->getRegInfo();
151   MDT = &getAnalysis<MachineDominatorTree>();
152   MPDT = &getAnalysis<MachinePostDominatorTree>();
153   MBFI = &getAnalysis<MachineBlockFrequencyInfo>();
154   EntryFreq = MBFI->getEntryFreq();
155   TII = MF->getSubtarget<PPCSubtarget>().getInstrInfo();
156   LLVM_DEBUG(dbgs() << "*** PowerPC MI peephole pass ***\n\n");
157   LLVM_DEBUG(MF->dump());
158 }
159 
160 static MachineInstr *getVRegDefOrNull(MachineOperand *Op,
161                                       MachineRegisterInfo *MRI) {
162   assert(Op && "Invalid Operand!");
163   if (!Op->isReg())
164     return nullptr;
165 
166   Register Reg = Op->getReg();
167   if (!Reg.isVirtual())
168     return nullptr;
169 
170   return MRI->getVRegDef(Reg);
171 }
172 
173 // This function returns number of known zero bits in output of MI
174 // starting from the most significant bit.
175 static unsigned getKnownLeadingZeroCount(const unsigned Reg,
176                                          const PPCInstrInfo *TII,
177                                          const MachineRegisterInfo *MRI) {
178   MachineInstr *MI = MRI->getVRegDef(Reg);
179   unsigned Opcode = MI->getOpcode();
180   if (Opcode == PPC::RLDICL || Opcode == PPC::RLDICL_rec ||
181       Opcode == PPC::RLDCL || Opcode == PPC::RLDCL_rec)
182     return MI->getOperand(3).getImm();
183 
184   if ((Opcode == PPC::RLDIC || Opcode == PPC::RLDIC_rec) &&
185       MI->getOperand(3).getImm() <= 63 - MI->getOperand(2).getImm())
186     return MI->getOperand(3).getImm();
187 
188   if ((Opcode == PPC::RLWINM || Opcode == PPC::RLWINM_rec ||
189        Opcode == PPC::RLWNM || Opcode == PPC::RLWNM_rec ||
190        Opcode == PPC::RLWINM8 || Opcode == PPC::RLWNM8) &&
191       MI->getOperand(3).getImm() <= MI->getOperand(4).getImm())
192     return 32 + MI->getOperand(3).getImm();
193 
194   if (Opcode == PPC::ANDI_rec) {
195     uint16_t Imm = MI->getOperand(2).getImm();
196     return 48 + llvm::countl_zero(Imm);
197   }
198 
199   if (Opcode == PPC::CNTLZW || Opcode == PPC::CNTLZW_rec ||
200       Opcode == PPC::CNTTZW || Opcode == PPC::CNTTZW_rec ||
201       Opcode == PPC::CNTLZW8 || Opcode == PPC::CNTTZW8)
202     // The result ranges from 0 to 32.
203     return 58;
204 
205   if (Opcode == PPC::CNTLZD || Opcode == PPC::CNTLZD_rec ||
206       Opcode == PPC::CNTTZD || Opcode == PPC::CNTTZD_rec)
207     // The result ranges from 0 to 64.
208     return 57;
209 
210   if (Opcode == PPC::LHZ   || Opcode == PPC::LHZX  ||
211       Opcode == PPC::LHZ8  || Opcode == PPC::LHZX8 ||
212       Opcode == PPC::LHZU  || Opcode == PPC::LHZUX ||
213       Opcode == PPC::LHZU8 || Opcode == PPC::LHZUX8)
214     return 48;
215 
216   if (Opcode == PPC::LBZ   || Opcode == PPC::LBZX  ||
217       Opcode == PPC::LBZ8  || Opcode == PPC::LBZX8 ||
218       Opcode == PPC::LBZU  || Opcode == PPC::LBZUX ||
219       Opcode == PPC::LBZU8 || Opcode == PPC::LBZUX8)
220     return 56;
221 
222   if (Opcode == PPC::AND || Opcode == PPC::AND8 || Opcode == PPC::AND_rec ||
223       Opcode == PPC::AND8_rec)
224     return std::max(
225         getKnownLeadingZeroCount(MI->getOperand(1).getReg(), TII, MRI),
226         getKnownLeadingZeroCount(MI->getOperand(2).getReg(), TII, MRI));
227 
228   if (Opcode == PPC::OR || Opcode == PPC::OR8 || Opcode == PPC::XOR ||
229       Opcode == PPC::XOR8 || Opcode == PPC::OR_rec ||
230       Opcode == PPC::OR8_rec || Opcode == PPC::XOR_rec ||
231       Opcode == PPC::XOR8_rec)
232     return std::min(
233         getKnownLeadingZeroCount(MI->getOperand(1).getReg(), TII, MRI),
234         getKnownLeadingZeroCount(MI->getOperand(2).getReg(), TII, MRI));
235 
236   if (TII->isZeroExtended(Reg, MRI))
237     return 32;
238 
239   return 0;
240 }
241 
242 // This function maintains a map for the pairs <TOC Save Instr, Keep>
243 // Each time a new TOC save is encountered, it checks if any of the existing
244 // ones are dominated by the new one. If so, it marks the existing one as
245 // redundant by setting it's entry in the map as false. It then adds the new
246 // instruction to the map with either true or false depending on if any
247 // existing instructions dominated the new one.
248 void PPCMIPeephole::UpdateTOCSaves(
249   std::map<MachineInstr *, bool> &TOCSaves, MachineInstr *MI) {
250   assert(TII->isTOCSaveMI(*MI) && "Expecting a TOC save instruction here");
251   // FIXME: Saving TOC in prologue hasn't been implemented well in AIX ABI part,
252   // here only support it under ELFv2.
253   if (MF->getSubtarget<PPCSubtarget>().isELFv2ABI()) {
254     PPCFunctionInfo *FI = MF->getInfo<PPCFunctionInfo>();
255 
256     MachineBasicBlock *Entry = &MF->front();
257     uint64_t CurrBlockFreq = MBFI->getBlockFreq(MI->getParent()).getFrequency();
258 
259     // If the block in which the TOC save resides is in a block that
260     // post-dominates Entry, or a block that is hotter than entry (keep in mind
261     // that early MachineLICM has already run so the TOC save won't be hoisted)
262     // we can just do the save in the prologue.
263     if (CurrBlockFreq > EntryFreq || MPDT->dominates(MI->getParent(), Entry))
264       FI->setMustSaveTOC(true);
265 
266     // If we are saving the TOC in the prologue, all the TOC saves can be
267     // removed from the code.
268     if (FI->mustSaveTOC()) {
269       for (auto &TOCSave : TOCSaves)
270         TOCSave.second = false;
271       // Add new instruction to map.
272       TOCSaves[MI] = false;
273       return;
274     }
275   }
276 
277   bool Keep = true;
278   for (auto &I : TOCSaves) {
279     MachineInstr *CurrInst = I.first;
280     // If new instruction dominates an existing one, mark existing one as
281     // redundant.
282     if (I.second && MDT->dominates(MI, CurrInst))
283       I.second = false;
284     // Check if the new instruction is redundant.
285     if (MDT->dominates(CurrInst, MI)) {
286       Keep = false;
287       break;
288     }
289   }
290   // Add new instruction to map.
291   TOCSaves[MI] = Keep;
292 }
293 
294 // This function returns a list of all PHI nodes in the tree starting from
295 // the RootPHI node. We perform a BFS traversal to get an ordered list of nodes.
296 // The list initially only contains the root PHI. When we visit a PHI node, we
297 // add it to the list. We continue to look for other PHI node operands while
298 // there are nodes to visit in the list. The function returns false if the
299 // optimization cannot be applied on this tree.
300 static bool collectUnprimedAccPHIs(MachineRegisterInfo *MRI,
301                                    MachineInstr *RootPHI,
302                                    SmallVectorImpl<MachineInstr *> &PHIs) {
303   PHIs.push_back(RootPHI);
304   unsigned VisitedIndex = 0;
305   while (VisitedIndex < PHIs.size()) {
306     MachineInstr *VisitedPHI = PHIs[VisitedIndex];
307     for (unsigned PHIOp = 1, NumOps = VisitedPHI->getNumOperands();
308          PHIOp != NumOps; PHIOp += 2) {
309       Register RegOp = VisitedPHI->getOperand(PHIOp).getReg();
310       if (!RegOp.isVirtual())
311         return false;
312       MachineInstr *Instr = MRI->getVRegDef(RegOp);
313       // While collecting the PHI nodes, we check if they can be converted (i.e.
314       // all the operands are either copies, implicit defs or PHI nodes).
315       unsigned Opcode = Instr->getOpcode();
316       if (Opcode == PPC::COPY) {
317         Register Reg = Instr->getOperand(1).getReg();
318         if (!Reg.isVirtual() || MRI->getRegClass(Reg) != &PPC::ACCRCRegClass)
319           return false;
320       } else if (Opcode != PPC::IMPLICIT_DEF && Opcode != PPC::PHI)
321         return false;
322       // If we detect a cycle in the PHI nodes, we exit. It would be
323       // possible to change cycles as well, but that would add a lot
324       // of complexity for a case that is unlikely to occur with MMA
325       // code.
326       if (Opcode != PPC::PHI)
327         continue;
328       if (llvm::is_contained(PHIs, Instr))
329         return false;
330       PHIs.push_back(Instr);
331     }
332     VisitedIndex++;
333   }
334   return true;
335 }
336 
337 // This function changes the unprimed accumulator PHI nodes in the PHIs list to
338 // primed accumulator PHI nodes. The list is traversed in reverse order to
339 // change all the PHI operands of a PHI node before changing the node itself.
340 // We keep a map to associate each changed PHI node to its non-changed form.
341 static void convertUnprimedAccPHIs(const PPCInstrInfo *TII,
342                                    MachineRegisterInfo *MRI,
343                                    SmallVectorImpl<MachineInstr *> &PHIs,
344                                    Register Dst) {
345   DenseMap<MachineInstr *, MachineInstr *> ChangedPHIMap;
346   for (MachineInstr *PHI : llvm::reverse(PHIs)) {
347     SmallVector<std::pair<MachineOperand, MachineOperand>, 4> PHIOps;
348     // We check if the current PHI node can be changed by looking at its
349     // operands. If all the operands are either copies from primed
350     // accumulators, implicit definitions or other unprimed accumulator
351     // PHI nodes, we change it.
352     for (unsigned PHIOp = 1, NumOps = PHI->getNumOperands(); PHIOp != NumOps;
353          PHIOp += 2) {
354       Register RegOp = PHI->getOperand(PHIOp).getReg();
355       MachineInstr *PHIInput = MRI->getVRegDef(RegOp);
356       unsigned Opcode = PHIInput->getOpcode();
357       assert((Opcode == PPC::COPY || Opcode == PPC::IMPLICIT_DEF ||
358               Opcode == PPC::PHI) &&
359              "Unexpected instruction");
360       if (Opcode == PPC::COPY) {
361         assert(MRI->getRegClass(PHIInput->getOperand(1).getReg()) ==
362                    &PPC::ACCRCRegClass &&
363                "Unexpected register class");
364         PHIOps.push_back({PHIInput->getOperand(1), PHI->getOperand(PHIOp + 1)});
365       } else if (Opcode == PPC::IMPLICIT_DEF) {
366         Register AccReg = MRI->createVirtualRegister(&PPC::ACCRCRegClass);
367         BuildMI(*PHIInput->getParent(), PHIInput, PHIInput->getDebugLoc(),
368                 TII->get(PPC::IMPLICIT_DEF), AccReg);
369         PHIOps.push_back({MachineOperand::CreateReg(AccReg, false),
370                           PHI->getOperand(PHIOp + 1)});
371       } else if (Opcode == PPC::PHI) {
372         // We found a PHI operand. At this point we know this operand
373         // has already been changed so we get its associated changed form
374         // from the map.
375         assert(ChangedPHIMap.count(PHIInput) == 1 &&
376                "This PHI node should have already been changed.");
377         MachineInstr *PrimedAccPHI = ChangedPHIMap.lookup(PHIInput);
378         PHIOps.push_back({MachineOperand::CreateReg(
379                               PrimedAccPHI->getOperand(0).getReg(), false),
380                           PHI->getOperand(PHIOp + 1)});
381       }
382     }
383     Register AccReg = Dst;
384     // If the PHI node we are changing is the root node, the register it defines
385     // will be the destination register of the original copy (of the PHI def).
386     // For all other PHI's in the list, we need to create another primed
387     // accumulator virtual register as the PHI will no longer define the
388     // unprimed accumulator.
389     if (PHI != PHIs[0])
390       AccReg = MRI->createVirtualRegister(&PPC::ACCRCRegClass);
391     MachineInstrBuilder NewPHI = BuildMI(
392         *PHI->getParent(), PHI, PHI->getDebugLoc(), TII->get(PPC::PHI), AccReg);
393     for (auto RegMBB : PHIOps)
394       NewPHI.add(RegMBB.first).add(RegMBB.second);
395     ChangedPHIMap[PHI] = NewPHI.getInstr();
396     LLVM_DEBUG(dbgs() << "Converting PHI: ");
397     LLVM_DEBUG(PHI->dump());
398     LLVM_DEBUG(dbgs() << "To: ");
399     LLVM_DEBUG(NewPHI.getInstr()->dump());
400   }
401 }
402 
403 // Perform peephole optimizations.
404 bool PPCMIPeephole::simplifyCode() {
405   bool Simplified = false;
406   bool TrapOpt = false;
407   MachineInstr* ToErase = nullptr;
408   std::map<MachineInstr *, bool> TOCSaves;
409   const TargetRegisterInfo *TRI = &TII->getRegisterInfo();
410   NumFunctionsEnteredInMIPeephole++;
411   if (ConvertRegReg) {
412     // Fixed-point conversion of reg/reg instructions fed by load-immediate
413     // into reg/imm instructions. FIXME: This is expensive, control it with
414     // an option.
415     bool SomethingChanged = false;
416     do {
417       NumFixedPointIterations++;
418       SomethingChanged = false;
419       for (MachineBasicBlock &MBB : *MF) {
420         for (MachineInstr &MI : MBB) {
421           if (MI.isDebugInstr())
422             continue;
423 
424           if (TII->convertToImmediateForm(MI)) {
425             // We don't erase anything in case the def has other uses. Let DCE
426             // remove it if it can be removed.
427             LLVM_DEBUG(dbgs() << "Converted instruction to imm form: ");
428             LLVM_DEBUG(MI.dump());
429             NumConvertedToImmediateForm++;
430             SomethingChanged = true;
431             Simplified = true;
432             continue;
433           }
434         }
435       }
436     } while (SomethingChanged && FixedPointRegToImm);
437   }
438 
439   for (MachineBasicBlock &MBB : *MF) {
440     for (MachineInstr &MI : MBB) {
441 
442       // If the previous instruction was marked for elimination,
443       // remove it now.
444       if (ToErase) {
445         LLVM_DEBUG(dbgs() << "Deleting instruction: ");
446         LLVM_DEBUG(ToErase->dump());
447         ToErase->eraseFromParent();
448         ToErase = nullptr;
449       }
450       // If a conditional trap instruction got optimized to an
451       // unconditional trap, eliminate all the instructions after
452       // the trap.
453       if (EnableTrapOptimization && TrapOpt) {
454         ToErase = &MI;
455         continue;
456       }
457 
458       // Ignore debug instructions.
459       if (MI.isDebugInstr())
460         continue;
461 
462       // Per-opcode peepholes.
463       switch (MI.getOpcode()) {
464 
465       default:
466         break;
467       case PPC::COPY: {
468         Register Src = MI.getOperand(1).getReg();
469         Register Dst = MI.getOperand(0).getReg();
470         if (!Src.isVirtual() || !Dst.isVirtual())
471           break;
472         if (MRI->getRegClass(Src) != &PPC::UACCRCRegClass ||
473             MRI->getRegClass(Dst) != &PPC::ACCRCRegClass)
474           break;
475 
476         // We are copying an unprimed accumulator to a primed accumulator.
477         // If the input to the copy is a PHI that is fed only by (i) copies in
478         // the other direction (ii) implicitly defined unprimed accumulators or
479         // (iii) other PHI nodes satisfying (i) and (ii), we can change
480         // the PHI to a PHI on primed accumulators (as long as we also change
481         // its operands). To detect and change such copies, we first get a list
482         // of all the PHI nodes starting from the root PHI node in BFS order.
483         // We then visit all these PHI nodes to check if they can be changed to
484         // primed accumulator PHI nodes and if so, we change them.
485         MachineInstr *RootPHI = MRI->getVRegDef(Src);
486         if (RootPHI->getOpcode() != PPC::PHI)
487           break;
488 
489         SmallVector<MachineInstr *, 4> PHIs;
490         if (!collectUnprimedAccPHIs(MRI, RootPHI, PHIs))
491           break;
492 
493         convertUnprimedAccPHIs(TII, MRI, PHIs, Dst);
494 
495         ToErase = &MI;
496         break;
497       }
498       case PPC::LI:
499       case PPC::LI8: {
500         // If we are materializing a zero, look for any use operands for which
501         // zero means immediate zero. All such operands can be replaced with
502         // PPC::ZERO.
503         if (!MI.getOperand(1).isImm() || MI.getOperand(1).getImm() != 0)
504           break;
505         Register MIDestReg = MI.getOperand(0).getReg();
506         for (MachineInstr& UseMI : MRI->use_instructions(MIDestReg))
507           Simplified |= TII->onlyFoldImmediate(UseMI, MI, MIDestReg);
508         if (MRI->use_nodbg_empty(MIDestReg)) {
509           ++NumLoadImmZeroFoldedAndRemoved;
510           ToErase = &MI;
511         }
512         break;
513       }
514       case PPC::STW:
515       case PPC::STD: {
516         MachineFrameInfo &MFI = MF->getFrameInfo();
517         if (MFI.hasVarSizedObjects() ||
518             (!MF->getSubtarget<PPCSubtarget>().isELFv2ABI() &&
519              !MF->getSubtarget<PPCSubtarget>().isAIXABI()))
520           break;
521         // When encountering a TOC save instruction, call UpdateTOCSaves
522         // to add it to the TOCSaves map and mark any existing TOC saves
523         // it dominates as redundant.
524         if (TII->isTOCSaveMI(MI))
525           UpdateTOCSaves(TOCSaves, &MI);
526         break;
527       }
528       case PPC::XXPERMDI: {
529         // Perform simplifications of 2x64 vector swaps and splats.
530         // A swap is identified by an immediate value of 2, and a splat
531         // is identified by an immediate value of 0 or 3.
532         int Immed = MI.getOperand(3).getImm();
533 
534         if (Immed == 1)
535           break;
536 
537         // For each of these simplifications, we need the two source
538         // regs to match.  Unfortunately, MachineCSE ignores COPY and
539         // SUBREG_TO_REG, so for example we can see
540         //   XXPERMDI t, SUBREG_TO_REG(s), SUBREG_TO_REG(s), immed.
541         // We have to look through chains of COPY and SUBREG_TO_REG
542         // to find the real source values for comparison.
543         Register TrueReg1 =
544           TRI->lookThruCopyLike(MI.getOperand(1).getReg(), MRI);
545         Register TrueReg2 =
546           TRI->lookThruCopyLike(MI.getOperand(2).getReg(), MRI);
547 
548         if (!(TrueReg1 == TrueReg2 && TrueReg1.isVirtual()))
549           break;
550 
551         MachineInstr *DefMI = MRI->getVRegDef(TrueReg1);
552 
553         if (!DefMI)
554           break;
555 
556         unsigned DefOpc = DefMI->getOpcode();
557 
558         // If this is a splat fed by a splatting load, the splat is
559         // redundant. Replace with a copy. This doesn't happen directly due
560         // to code in PPCDAGToDAGISel.cpp, but it can happen when converting
561         // a load of a double to a vector of 64-bit integers.
562         auto isConversionOfLoadAndSplat = [=]() -> bool {
563           if (DefOpc != PPC::XVCVDPSXDS && DefOpc != PPC::XVCVDPUXDS)
564             return false;
565           Register FeedReg1 =
566             TRI->lookThruCopyLike(DefMI->getOperand(1).getReg(), MRI);
567           if (FeedReg1.isVirtual()) {
568             MachineInstr *LoadMI = MRI->getVRegDef(FeedReg1);
569             if (LoadMI && LoadMI->getOpcode() == PPC::LXVDSX)
570               return true;
571           }
572           return false;
573         };
574         if ((Immed == 0 || Immed == 3) &&
575             (DefOpc == PPC::LXVDSX || isConversionOfLoadAndSplat())) {
576           LLVM_DEBUG(dbgs() << "Optimizing load-and-splat/splat "
577                                "to load-and-splat/copy: ");
578           LLVM_DEBUG(MI.dump());
579           BuildMI(MBB, &MI, MI.getDebugLoc(), TII->get(PPC::COPY),
580                   MI.getOperand(0).getReg())
581               .add(MI.getOperand(1));
582           ToErase = &MI;
583           Simplified = true;
584         }
585 
586         // If this is a splat or a swap fed by another splat, we
587         // can replace it with a copy.
588         if (DefOpc == PPC::XXPERMDI) {
589           Register DefReg1 = DefMI->getOperand(1).getReg();
590           Register DefReg2 = DefMI->getOperand(2).getReg();
591           unsigned DefImmed = DefMI->getOperand(3).getImm();
592 
593           // If the two inputs are not the same register, check to see if
594           // they originate from the same virtual register after only
595           // copy-like instructions.
596           if (DefReg1 != DefReg2) {
597             Register FeedReg1 = TRI->lookThruCopyLike(DefReg1, MRI);
598             Register FeedReg2 = TRI->lookThruCopyLike(DefReg2, MRI);
599 
600             if (!(FeedReg1 == FeedReg2 && FeedReg1.isVirtual()))
601               break;
602           }
603 
604           if (DefImmed == 0 || DefImmed == 3) {
605             LLVM_DEBUG(dbgs() << "Optimizing splat/swap or splat/splat "
606                                  "to splat/copy: ");
607             LLVM_DEBUG(MI.dump());
608             BuildMI(MBB, &MI, MI.getDebugLoc(), TII->get(PPC::COPY),
609                     MI.getOperand(0).getReg())
610                 .add(MI.getOperand(1));
611             ToErase = &MI;
612             Simplified = true;
613           }
614 
615           // If this is a splat fed by a swap, we can simplify modify
616           // the splat to splat the other value from the swap's input
617           // parameter.
618           else if ((Immed == 0 || Immed == 3) && DefImmed == 2) {
619             LLVM_DEBUG(dbgs() << "Optimizing swap/splat => splat: ");
620             LLVM_DEBUG(MI.dump());
621             MI.getOperand(1).setReg(DefReg1);
622             MI.getOperand(2).setReg(DefReg2);
623             MI.getOperand(3).setImm(3 - Immed);
624             Simplified = true;
625           }
626 
627           // If this is a swap fed by a swap, we can replace it
628           // with a copy from the first swap's input.
629           else if (Immed == 2 && DefImmed == 2) {
630             LLVM_DEBUG(dbgs() << "Optimizing swap/swap => copy: ");
631             LLVM_DEBUG(MI.dump());
632             BuildMI(MBB, &MI, MI.getDebugLoc(), TII->get(PPC::COPY),
633                     MI.getOperand(0).getReg())
634                 .add(DefMI->getOperand(1));
635             ToErase = &MI;
636             Simplified = true;
637           }
638         } else if ((Immed == 0 || Immed == 3 || Immed == 2) &&
639                    DefOpc == PPC::XXPERMDIs &&
640                    (DefMI->getOperand(2).getImm() == 0 ||
641                     DefMI->getOperand(2).getImm() == 3)) {
642           ToErase = &MI;
643           Simplified = true;
644           // Swap of a splat, convert to copy.
645           if (Immed == 2) {
646             LLVM_DEBUG(dbgs() << "Optimizing swap(splat) => copy(splat): ");
647             LLVM_DEBUG(MI.dump());
648             BuildMI(MBB, &MI, MI.getDebugLoc(), TII->get(PPC::COPY),
649                     MI.getOperand(0).getReg())
650                 .add(MI.getOperand(1));
651             break;
652           }
653           // Splat fed by another splat - switch the output of the first
654           // and remove the second.
655           DefMI->getOperand(0).setReg(MI.getOperand(0).getReg());
656           LLVM_DEBUG(dbgs() << "Removing redundant splat: ");
657           LLVM_DEBUG(MI.dump());
658         } else if (Immed == 2 &&
659                    (DefOpc == PPC::VSPLTB || DefOpc == PPC::VSPLTH ||
660                     DefOpc == PPC::VSPLTW || DefOpc == PPC::XXSPLTW)) {
661           // Swap of various vector splats, convert to copy.
662           ToErase = &MI;
663           Simplified = true;
664           LLVM_DEBUG(dbgs() << "Optimizing swap(vsplt[b|h|w]|xxspltw) => "
665                                "copy(vsplt[b|h|w]|xxspltw): ");
666           LLVM_DEBUG(MI.dump());
667           BuildMI(MBB, &MI, MI.getDebugLoc(), TII->get(PPC::COPY),
668                   MI.getOperand(0).getReg())
669               .add(MI.getOperand(1));
670         } else if ((Immed == 0 || Immed == 3 || Immed == 2) &&
671                    TII->isLoadFromConstantPool(DefMI)) {
672           const Constant *C = TII->getConstantFromConstantPool(DefMI);
673           if (C && C->getType()->isVectorTy() && C->getSplatValue()) {
674             ToErase = &MI;
675             Simplified = true;
676             LLVM_DEBUG(dbgs()
677                        << "Optimizing swap(splat pattern from constant-pool) "
678                           "=> copy(splat pattern from constant-pool): ");
679             LLVM_DEBUG(MI.dump());
680             BuildMI(MBB, &MI, MI.getDebugLoc(), TII->get(PPC::COPY),
681                     MI.getOperand(0).getReg())
682                 .add(MI.getOperand(1));
683           }
684         }
685         break;
686       }
687       case PPC::VSPLTB:
688       case PPC::VSPLTH:
689       case PPC::XXSPLTW: {
690         unsigned MyOpcode = MI.getOpcode();
691         unsigned OpNo = MyOpcode == PPC::XXSPLTW ? 1 : 2;
692         Register TrueReg =
693           TRI->lookThruCopyLike(MI.getOperand(OpNo).getReg(), MRI);
694         if (!TrueReg.isVirtual())
695           break;
696         MachineInstr *DefMI = MRI->getVRegDef(TrueReg);
697         if (!DefMI)
698           break;
699         unsigned DefOpcode = DefMI->getOpcode();
700         auto isConvertOfSplat = [=]() -> bool {
701           if (DefOpcode != PPC::XVCVSPSXWS && DefOpcode != PPC::XVCVSPUXWS)
702             return false;
703           Register ConvReg = DefMI->getOperand(1).getReg();
704           if (!ConvReg.isVirtual())
705             return false;
706           MachineInstr *Splt = MRI->getVRegDef(ConvReg);
707           return Splt && (Splt->getOpcode() == PPC::LXVWSX ||
708             Splt->getOpcode() == PPC::XXSPLTW);
709         };
710         bool AlreadySplat = (MyOpcode == DefOpcode) ||
711           (MyOpcode == PPC::VSPLTB && DefOpcode == PPC::VSPLTBs) ||
712           (MyOpcode == PPC::VSPLTH && DefOpcode == PPC::VSPLTHs) ||
713           (MyOpcode == PPC::XXSPLTW && DefOpcode == PPC::XXSPLTWs) ||
714           (MyOpcode == PPC::XXSPLTW && DefOpcode == PPC::LXVWSX) ||
715           (MyOpcode == PPC::XXSPLTW && DefOpcode == PPC::MTVSRWS)||
716           (MyOpcode == PPC::XXSPLTW && isConvertOfSplat());
717         // If the instruction[s] that feed this splat have already splat
718         // the value, this splat is redundant.
719         if (AlreadySplat) {
720           LLVM_DEBUG(dbgs() << "Changing redundant splat to a copy: ");
721           LLVM_DEBUG(MI.dump());
722           BuildMI(MBB, &MI, MI.getDebugLoc(), TII->get(PPC::COPY),
723                   MI.getOperand(0).getReg())
724               .add(MI.getOperand(OpNo));
725           ToErase = &MI;
726           Simplified = true;
727         }
728         // Splat fed by a shift. Usually when we align value to splat into
729         // vector element zero.
730         if (DefOpcode == PPC::XXSLDWI) {
731           Register ShiftRes = DefMI->getOperand(0).getReg();
732           Register ShiftOp1 = DefMI->getOperand(1).getReg();
733           Register ShiftOp2 = DefMI->getOperand(2).getReg();
734           unsigned ShiftImm = DefMI->getOperand(3).getImm();
735           unsigned SplatImm =
736               MI.getOperand(MyOpcode == PPC::XXSPLTW ? 2 : 1).getImm();
737           if (ShiftOp1 == ShiftOp2) {
738             unsigned NewElem = (SplatImm + ShiftImm) & 0x3;
739             if (MRI->hasOneNonDBGUse(ShiftRes)) {
740               LLVM_DEBUG(dbgs() << "Removing redundant shift: ");
741               LLVM_DEBUG(DefMI->dump());
742               ToErase = DefMI;
743             }
744             Simplified = true;
745             LLVM_DEBUG(dbgs() << "Changing splat immediate from " << SplatImm
746                               << " to " << NewElem << " in instruction: ");
747             LLVM_DEBUG(MI.dump());
748             MI.getOperand(1).setReg(ShiftOp1);
749             MI.getOperand(2).setImm(NewElem);
750           }
751         }
752         break;
753       }
754       case PPC::XVCVDPSP: {
755         // If this is a DP->SP conversion fed by an FRSP, the FRSP is redundant.
756         Register TrueReg =
757           TRI->lookThruCopyLike(MI.getOperand(1).getReg(), MRI);
758         if (!TrueReg.isVirtual())
759           break;
760         MachineInstr *DefMI = MRI->getVRegDef(TrueReg);
761 
762         // This can occur when building a vector of single precision or integer
763         // values.
764         if (DefMI && DefMI->getOpcode() == PPC::XXPERMDI) {
765           Register DefsReg1 =
766             TRI->lookThruCopyLike(DefMI->getOperand(1).getReg(), MRI);
767           Register DefsReg2 =
768             TRI->lookThruCopyLike(DefMI->getOperand(2).getReg(), MRI);
769           if (!DefsReg1.isVirtual() || !DefsReg2.isVirtual())
770             break;
771           MachineInstr *P1 = MRI->getVRegDef(DefsReg1);
772           MachineInstr *P2 = MRI->getVRegDef(DefsReg2);
773 
774           if (!P1 || !P2)
775             break;
776 
777           // Remove the passed FRSP/XSRSP instruction if it only feeds this MI
778           // and set any uses of that FRSP/XSRSP (in this MI) to the source of
779           // the FRSP/XSRSP.
780           auto removeFRSPIfPossible = [&](MachineInstr *RoundInstr) {
781             unsigned Opc = RoundInstr->getOpcode();
782             if ((Opc == PPC::FRSP || Opc == PPC::XSRSP) &&
783                 MRI->hasOneNonDBGUse(RoundInstr->getOperand(0).getReg())) {
784               Simplified = true;
785               Register ConvReg1 = RoundInstr->getOperand(1).getReg();
786               Register FRSPDefines = RoundInstr->getOperand(0).getReg();
787               MachineInstr &Use = *(MRI->use_instr_nodbg_begin(FRSPDefines));
788               for (int i = 0, e = Use.getNumOperands(); i < e; ++i)
789                 if (Use.getOperand(i).isReg() &&
790                     Use.getOperand(i).getReg() == FRSPDefines)
791                   Use.getOperand(i).setReg(ConvReg1);
792               LLVM_DEBUG(dbgs() << "Removing redundant FRSP/XSRSP:\n");
793               LLVM_DEBUG(RoundInstr->dump());
794               LLVM_DEBUG(dbgs() << "As it feeds instruction:\n");
795               LLVM_DEBUG(MI.dump());
796               LLVM_DEBUG(dbgs() << "Through instruction:\n");
797               LLVM_DEBUG(DefMI->dump());
798               RoundInstr->eraseFromParent();
799             }
800           };
801 
802           // If the input to XVCVDPSP is a vector that was built (even
803           // partially) out of FRSP's, the FRSP(s) can safely be removed
804           // since this instruction performs the same operation.
805           if (P1 != P2) {
806             removeFRSPIfPossible(P1);
807             removeFRSPIfPossible(P2);
808             break;
809           }
810           removeFRSPIfPossible(P1);
811         }
812         break;
813       }
814       case PPC::EXTSH:
815       case PPC::EXTSH8:
816       case PPC::EXTSH8_32_64: {
817         if (!EnableSExtElimination) break;
818         Register NarrowReg = MI.getOperand(1).getReg();
819         if (!NarrowReg.isVirtual())
820           break;
821 
822         MachineInstr *SrcMI = MRI->getVRegDef(NarrowReg);
823         unsigned SrcOpcode = SrcMI->getOpcode();
824         // If we've used a zero-extending load that we will sign-extend,
825         // just do a sign-extending load.
826         if (SrcOpcode == PPC::LHZ || SrcOpcode == PPC::LHZX) {
827           if (!MRI->hasOneNonDBGUse(SrcMI->getOperand(0).getReg()))
828             break;
829           // Determine the new opcode. We need to make sure that if the original
830           // instruction has a 64 bit opcode we keep using a 64 bit opcode.
831           // Likewise if the source is X-Form the new opcode should also be
832           // X-Form.
833           unsigned Opc = PPC::LHA;
834           bool SourceIsXForm = SrcOpcode == PPC::LHZX;
835           bool MIIs64Bit = MI.getOpcode() == PPC::EXTSH8 ||
836             MI.getOpcode() == PPC::EXTSH8_32_64;
837 
838           if (SourceIsXForm && MIIs64Bit)
839             Opc = PPC::LHAX8;
840           else if (SourceIsXForm && !MIIs64Bit)
841             Opc = PPC::LHAX;
842           else if (MIIs64Bit)
843             Opc = PPC::LHA8;
844 
845           LLVM_DEBUG(dbgs() << "Zero-extending load\n");
846           LLVM_DEBUG(SrcMI->dump());
847           LLVM_DEBUG(dbgs() << "and sign-extension\n");
848           LLVM_DEBUG(MI.dump());
849           LLVM_DEBUG(dbgs() << "are merged into sign-extending load\n");
850           SrcMI->setDesc(TII->get(Opc));
851           SrcMI->getOperand(0).setReg(MI.getOperand(0).getReg());
852           ToErase = &MI;
853           Simplified = true;
854           NumEliminatedSExt++;
855         }
856         break;
857       }
858       case PPC::EXTSW:
859       case PPC::EXTSW_32:
860       case PPC::EXTSW_32_64: {
861         if (!EnableSExtElimination) break;
862         Register NarrowReg = MI.getOperand(1).getReg();
863         if (!NarrowReg.isVirtual())
864           break;
865 
866         MachineInstr *SrcMI = MRI->getVRegDef(NarrowReg);
867         unsigned SrcOpcode = SrcMI->getOpcode();
868         // If we've used a zero-extending load that we will sign-extend,
869         // just do a sign-extending load.
870         if (SrcOpcode == PPC::LWZ || SrcOpcode == PPC::LWZX) {
871           if (!MRI->hasOneNonDBGUse(SrcMI->getOperand(0).getReg()))
872             break;
873 
874           // The transformation from a zero-extending load to a sign-extending
875           // load is only legal when the displacement is a multiple of 4.
876           // If the displacement is not at least 4 byte aligned, don't perform
877           // the transformation.
878           bool IsWordAligned = false;
879           if (SrcMI->getOperand(1).isGlobal()) {
880             const GlobalObject *GO =
881                 dyn_cast<GlobalObject>(SrcMI->getOperand(1).getGlobal());
882             if (GO && GO->getAlign() && *GO->getAlign() >= 4 &&
883                 (SrcMI->getOperand(1).getOffset() % 4 == 0))
884               IsWordAligned = true;
885           } else if (SrcMI->getOperand(1).isImm()) {
886             int64_t Value = SrcMI->getOperand(1).getImm();
887             if (Value % 4 == 0)
888               IsWordAligned = true;
889           }
890 
891           // Determine the new opcode. We need to make sure that if the original
892           // instruction has a 64 bit opcode we keep using a 64 bit opcode.
893           // Likewise if the source is X-Form the new opcode should also be
894           // X-Form.
895           unsigned Opc = PPC::LWA_32;
896           bool SourceIsXForm = SrcOpcode == PPC::LWZX;
897           bool MIIs64Bit = MI.getOpcode() == PPC::EXTSW ||
898             MI.getOpcode() == PPC::EXTSW_32_64;
899 
900           if (SourceIsXForm && MIIs64Bit)
901             Opc = PPC::LWAX;
902           else if (SourceIsXForm && !MIIs64Bit)
903             Opc = PPC::LWAX_32;
904           else if (MIIs64Bit)
905             Opc = PPC::LWA;
906 
907           if (!IsWordAligned && (Opc == PPC::LWA || Opc == PPC::LWA_32))
908             break;
909 
910           LLVM_DEBUG(dbgs() << "Zero-extending load\n");
911           LLVM_DEBUG(SrcMI->dump());
912           LLVM_DEBUG(dbgs() << "and sign-extension\n");
913           LLVM_DEBUG(MI.dump());
914           LLVM_DEBUG(dbgs() << "are merged into sign-extending load\n");
915           SrcMI->setDesc(TII->get(Opc));
916           SrcMI->getOperand(0).setReg(MI.getOperand(0).getReg());
917           ToErase = &MI;
918           Simplified = true;
919           NumEliminatedSExt++;
920         } else if (MI.getOpcode() == PPC::EXTSW_32_64 &&
921                    TII->isSignExtended(NarrowReg, MRI)) {
922           // We can eliminate EXTSW if the input is known to be already
923           // sign-extended.
924           LLVM_DEBUG(dbgs() << "Removing redundant sign-extension\n");
925           Register TmpReg =
926               MF->getRegInfo().createVirtualRegister(&PPC::G8RCRegClass);
927           BuildMI(MBB, &MI, MI.getDebugLoc(), TII->get(PPC::IMPLICIT_DEF),
928                   TmpReg);
929           BuildMI(MBB, &MI, MI.getDebugLoc(), TII->get(PPC::INSERT_SUBREG),
930                   MI.getOperand(0).getReg())
931               .addReg(TmpReg)
932               .addReg(NarrowReg)
933               .addImm(PPC::sub_32);
934           ToErase = &MI;
935           Simplified = true;
936           NumEliminatedSExt++;
937         }
938         break;
939       }
940       case PPC::RLDICL: {
941         // We can eliminate RLDICL (e.g. for zero-extension)
942         // if all bits to clear are already zero in the input.
943         // This code assume following code sequence for zero-extension.
944         //   %6 = COPY %5:sub_32; (optional)
945         //   %8 = IMPLICIT_DEF;
946         //   %7<def,tied1> = INSERT_SUBREG %8<tied0>, %6, sub_32;
947         if (!EnableZExtElimination) break;
948 
949         if (MI.getOperand(2).getImm() != 0)
950           break;
951 
952         Register SrcReg = MI.getOperand(1).getReg();
953         if (!SrcReg.isVirtual())
954           break;
955 
956         MachineInstr *SrcMI = MRI->getVRegDef(SrcReg);
957         if (!(SrcMI && SrcMI->getOpcode() == PPC::INSERT_SUBREG &&
958               SrcMI->getOperand(0).isReg() && SrcMI->getOperand(1).isReg()))
959           break;
960 
961         MachineInstr *ImpDefMI, *SubRegMI;
962         ImpDefMI = MRI->getVRegDef(SrcMI->getOperand(1).getReg());
963         SubRegMI = MRI->getVRegDef(SrcMI->getOperand(2).getReg());
964         if (ImpDefMI->getOpcode() != PPC::IMPLICIT_DEF) break;
965 
966         SrcMI = SubRegMI;
967         if (SubRegMI->getOpcode() == PPC::COPY) {
968           Register CopyReg = SubRegMI->getOperand(1).getReg();
969           if (CopyReg.isVirtual())
970             SrcMI = MRI->getVRegDef(CopyReg);
971         }
972         if (!SrcMI->getOperand(0).isReg())
973           break;
974 
975         unsigned KnownZeroCount =
976             getKnownLeadingZeroCount(SrcMI->getOperand(0).getReg(), TII, MRI);
977         if (MI.getOperand(3).getImm() <= KnownZeroCount) {
978           LLVM_DEBUG(dbgs() << "Removing redundant zero-extension\n");
979           BuildMI(MBB, &MI, MI.getDebugLoc(), TII->get(PPC::COPY),
980                   MI.getOperand(0).getReg())
981               .addReg(SrcReg);
982           ToErase = &MI;
983           Simplified = true;
984           NumEliminatedZExt++;
985         }
986         break;
987       }
988 
989       // TODO: Any instruction that has an immediate form fed only by a PHI
990       // whose operands are all load immediate can be folded away. We currently
991       // do this for ADD instructions, but should expand it to arithmetic and
992       // binary instructions with immediate forms in the future.
993       case PPC::ADD4:
994       case PPC::ADD8: {
995         auto isSingleUsePHI = [&](MachineOperand *PhiOp) {
996           assert(PhiOp && "Invalid Operand!");
997           MachineInstr *DefPhiMI = getVRegDefOrNull(PhiOp, MRI);
998 
999           return DefPhiMI && (DefPhiMI->getOpcode() == PPC::PHI) &&
1000                  MRI->hasOneNonDBGUse(DefPhiMI->getOperand(0).getReg());
1001         };
1002 
1003         auto dominatesAllSingleUseLIs = [&](MachineOperand *DominatorOp,
1004                                             MachineOperand *PhiOp) {
1005           assert(PhiOp && "Invalid Operand!");
1006           assert(DominatorOp && "Invalid Operand!");
1007           MachineInstr *DefPhiMI = getVRegDefOrNull(PhiOp, MRI);
1008           MachineInstr *DefDomMI = getVRegDefOrNull(DominatorOp, MRI);
1009 
1010           // Note: the vregs only show up at odd indices position of PHI Node,
1011           // the even indices position save the BB info.
1012           for (unsigned i = 1; i < DefPhiMI->getNumOperands(); i += 2) {
1013             MachineInstr *LiMI =
1014                 getVRegDefOrNull(&DefPhiMI->getOperand(i), MRI);
1015             if (!LiMI ||
1016                 (LiMI->getOpcode() != PPC::LI && LiMI->getOpcode() != PPC::LI8)
1017                 || !MRI->hasOneNonDBGUse(LiMI->getOperand(0).getReg()) ||
1018                 !MDT->dominates(DefDomMI, LiMI))
1019               return false;
1020           }
1021 
1022           return true;
1023         };
1024 
1025         MachineOperand Op1 = MI.getOperand(1);
1026         MachineOperand Op2 = MI.getOperand(2);
1027         if (isSingleUsePHI(&Op2) && dominatesAllSingleUseLIs(&Op1, &Op2))
1028           std::swap(Op1, Op2);
1029         else if (!isSingleUsePHI(&Op1) || !dominatesAllSingleUseLIs(&Op2, &Op1))
1030           break; // We don't have an ADD fed by LI's that can be transformed
1031 
1032         // Now we know that Op1 is the PHI node and Op2 is the dominator
1033         Register DominatorReg = Op2.getReg();
1034 
1035         const TargetRegisterClass *TRC = MI.getOpcode() == PPC::ADD8
1036                                              ? &PPC::G8RC_and_G8RC_NOX0RegClass
1037                                              : &PPC::GPRC_and_GPRC_NOR0RegClass;
1038         MRI->setRegClass(DominatorReg, TRC);
1039 
1040         // replace LIs with ADDIs
1041         MachineInstr *DefPhiMI = getVRegDefOrNull(&Op1, MRI);
1042         for (unsigned i = 1; i < DefPhiMI->getNumOperands(); i += 2) {
1043           MachineInstr *LiMI = getVRegDefOrNull(&DefPhiMI->getOperand(i), MRI);
1044           LLVM_DEBUG(dbgs() << "Optimizing LI to ADDI: ");
1045           LLVM_DEBUG(LiMI->dump());
1046 
1047           // There could be repeated registers in the PHI, e.g: %1 =
1048           // PHI %6, <%bb.2>, %8, <%bb.3>, %8, <%bb.6>; So if we've
1049           // already replaced the def instruction, skip.
1050           if (LiMI->getOpcode() == PPC::ADDI || LiMI->getOpcode() == PPC::ADDI8)
1051             continue;
1052 
1053           assert((LiMI->getOpcode() == PPC::LI ||
1054                   LiMI->getOpcode() == PPC::LI8) &&
1055                  "Invalid Opcode!");
1056           auto LiImm = LiMI->getOperand(1).getImm(); // save the imm of LI
1057           LiMI->removeOperand(1);                    // remove the imm of LI
1058           LiMI->setDesc(TII->get(LiMI->getOpcode() == PPC::LI ? PPC::ADDI
1059                                                               : PPC::ADDI8));
1060           MachineInstrBuilder(*LiMI->getParent()->getParent(), *LiMI)
1061               .addReg(DominatorReg)
1062               .addImm(LiImm); // restore the imm of LI
1063           LLVM_DEBUG(LiMI->dump());
1064         }
1065 
1066         // Replace ADD with COPY
1067         LLVM_DEBUG(dbgs() << "Optimizing ADD to COPY: ");
1068         LLVM_DEBUG(MI.dump());
1069         BuildMI(MBB, &MI, MI.getDebugLoc(), TII->get(PPC::COPY),
1070                 MI.getOperand(0).getReg())
1071             .add(Op1);
1072         ToErase = &MI;
1073         Simplified = true;
1074         NumOptADDLIs++;
1075         break;
1076       }
1077       case PPC::RLDICR: {
1078         Simplified |= emitRLDICWhenLoweringJumpTables(MI) ||
1079                       combineSEXTAndSHL(MI, ToErase);
1080         break;
1081       }
1082       case PPC::RLWINM:
1083       case PPC::RLWINM_rec:
1084       case PPC::RLWINM8:
1085       case PPC::RLWINM8_rec: {
1086         Simplified = TII->combineRLWINM(MI, &ToErase);
1087         if (Simplified)
1088           ++NumRotatesCollapsed;
1089         break;
1090       }
1091       // We will replace TD/TW/TDI/TWI with an unconditional trap if it will
1092       // always trap, we will delete the node if it will never trap.
1093       case PPC::TDI:
1094       case PPC::TWI:
1095       case PPC::TD:
1096       case PPC::TW: {
1097         if (!EnableTrapOptimization) break;
1098         MachineInstr *LiMI1 = getVRegDefOrNull(&MI.getOperand(1), MRI);
1099         MachineInstr *LiMI2 = getVRegDefOrNull(&MI.getOperand(2), MRI);
1100         bool IsOperand2Immediate = MI.getOperand(2).isImm();
1101         // We can only do the optimization if we can get immediates
1102         // from both operands
1103         if (!(LiMI1 && (LiMI1->getOpcode() == PPC::LI ||
1104                         LiMI1->getOpcode() == PPC::LI8)))
1105           break;
1106         if (!IsOperand2Immediate &&
1107             !(LiMI2 && (LiMI2->getOpcode() == PPC::LI ||
1108                         LiMI2->getOpcode() == PPC::LI8)))
1109           break;
1110 
1111         auto ImmOperand0 = MI.getOperand(0).getImm();
1112         auto ImmOperand1 = LiMI1->getOperand(1).getImm();
1113         auto ImmOperand2 = IsOperand2Immediate ? MI.getOperand(2).getImm()
1114                                                : LiMI2->getOperand(1).getImm();
1115 
1116         // We will replace the MI with an unconditional trap if it will always
1117         // trap.
1118         if ((ImmOperand0 == 31) ||
1119             ((ImmOperand0 & 0x10) &&
1120              ((int64_t)ImmOperand1 < (int64_t)ImmOperand2)) ||
1121             ((ImmOperand0 & 0x8) &&
1122              ((int64_t)ImmOperand1 > (int64_t)ImmOperand2)) ||
1123             ((ImmOperand0 & 0x2) &&
1124              ((uint64_t)ImmOperand1 < (uint64_t)ImmOperand2)) ||
1125             ((ImmOperand0 & 0x1) &&
1126              ((uint64_t)ImmOperand1 > (uint64_t)ImmOperand2)) ||
1127             ((ImmOperand0 & 0x4) && (ImmOperand1 == ImmOperand2))) {
1128           BuildMI(MBB, &MI, MI.getDebugLoc(), TII->get(PPC::TRAP));
1129           TrapOpt = true;
1130         }
1131         // We will delete the MI if it will never trap.
1132         ToErase = &MI;
1133         Simplified = true;
1134         break;
1135       }
1136       }
1137     }
1138 
1139     // If the last instruction was marked for elimination,
1140     // remove it now.
1141     if (ToErase) {
1142       ToErase->eraseFromParent();
1143       ToErase = nullptr;
1144     }
1145     // Reset TrapOpt to false at the end of the basic block.
1146     if (EnableTrapOptimization)
1147       TrapOpt = false;
1148   }
1149 
1150   // Eliminate all the TOC save instructions which are redundant.
1151   Simplified |= eliminateRedundantTOCSaves(TOCSaves);
1152   PPCFunctionInfo *FI = MF->getInfo<PPCFunctionInfo>();
1153   if (FI->mustSaveTOC())
1154     NumTOCSavesInPrologue++;
1155 
1156   // We try to eliminate redundant compare instruction.
1157   Simplified |= eliminateRedundantCompare();
1158 
1159   return Simplified;
1160 }
1161 
1162 // helper functions for eliminateRedundantCompare
1163 static bool isEqOrNe(MachineInstr *BI) {
1164   PPC::Predicate Pred = (PPC::Predicate)BI->getOperand(0).getImm();
1165   unsigned PredCond = PPC::getPredicateCondition(Pred);
1166   return (PredCond == PPC::PRED_EQ || PredCond == PPC::PRED_NE);
1167 }
1168 
1169 static bool isSupportedCmpOp(unsigned opCode) {
1170   return (opCode == PPC::CMPLD  || opCode == PPC::CMPD  ||
1171           opCode == PPC::CMPLW  || opCode == PPC::CMPW  ||
1172           opCode == PPC::CMPLDI || opCode == PPC::CMPDI ||
1173           opCode == PPC::CMPLWI || opCode == PPC::CMPWI);
1174 }
1175 
1176 static bool is64bitCmpOp(unsigned opCode) {
1177   return (opCode == PPC::CMPLD  || opCode == PPC::CMPD ||
1178           opCode == PPC::CMPLDI || opCode == PPC::CMPDI);
1179 }
1180 
1181 static bool isSignedCmpOp(unsigned opCode) {
1182   return (opCode == PPC::CMPD  || opCode == PPC::CMPW ||
1183           opCode == PPC::CMPDI || opCode == PPC::CMPWI);
1184 }
1185 
1186 static unsigned getSignedCmpOpCode(unsigned opCode) {
1187   if (opCode == PPC::CMPLD)  return PPC::CMPD;
1188   if (opCode == PPC::CMPLW)  return PPC::CMPW;
1189   if (opCode == PPC::CMPLDI) return PPC::CMPDI;
1190   if (opCode == PPC::CMPLWI) return PPC::CMPWI;
1191   return opCode;
1192 }
1193 
1194 // We can decrement immediate x in (GE x) by changing it to (GT x-1) or
1195 // (LT x) to (LE x-1)
1196 static unsigned getPredicateToDecImm(MachineInstr *BI, MachineInstr *CMPI) {
1197   uint64_t Imm = CMPI->getOperand(2).getImm();
1198   bool SignedCmp = isSignedCmpOp(CMPI->getOpcode());
1199   if ((!SignedCmp && Imm == 0) || (SignedCmp && Imm == 0x8000))
1200     return 0;
1201 
1202   PPC::Predicate Pred = (PPC::Predicate)BI->getOperand(0).getImm();
1203   unsigned PredCond = PPC::getPredicateCondition(Pred);
1204   unsigned PredHint = PPC::getPredicateHint(Pred);
1205   if (PredCond == PPC::PRED_GE)
1206     return PPC::getPredicate(PPC::PRED_GT, PredHint);
1207   if (PredCond == PPC::PRED_LT)
1208     return PPC::getPredicate(PPC::PRED_LE, PredHint);
1209 
1210   return 0;
1211 }
1212 
1213 // We can increment immediate x in (GT x) by changing it to (GE x+1) or
1214 // (LE x) to (LT x+1)
1215 static unsigned getPredicateToIncImm(MachineInstr *BI, MachineInstr *CMPI) {
1216   uint64_t Imm = CMPI->getOperand(2).getImm();
1217   bool SignedCmp = isSignedCmpOp(CMPI->getOpcode());
1218   if ((!SignedCmp && Imm == 0xFFFF) || (SignedCmp && Imm == 0x7FFF))
1219     return 0;
1220 
1221   PPC::Predicate Pred = (PPC::Predicate)BI->getOperand(0).getImm();
1222   unsigned PredCond = PPC::getPredicateCondition(Pred);
1223   unsigned PredHint = PPC::getPredicateHint(Pred);
1224   if (PredCond == PPC::PRED_GT)
1225     return PPC::getPredicate(PPC::PRED_GE, PredHint);
1226   if (PredCond == PPC::PRED_LE)
1227     return PPC::getPredicate(PPC::PRED_LT, PredHint);
1228 
1229   return 0;
1230 }
1231 
1232 // This takes a Phi node and returns a register value for the specified BB.
1233 static unsigned getIncomingRegForBlock(MachineInstr *Phi,
1234                                        MachineBasicBlock *MBB) {
1235   for (unsigned I = 2, E = Phi->getNumOperands() + 1; I != E; I += 2) {
1236     MachineOperand &MO = Phi->getOperand(I);
1237     if (MO.getMBB() == MBB)
1238       return Phi->getOperand(I-1).getReg();
1239   }
1240   llvm_unreachable("invalid src basic block for this Phi node\n");
1241   return 0;
1242 }
1243 
1244 // This function tracks the source of the register through register copy.
1245 // If BB1 and BB2 are non-NULL, we also track PHI instruction in BB2
1246 // assuming that the control comes from BB1 into BB2.
1247 static unsigned getSrcVReg(unsigned Reg, MachineBasicBlock *BB1,
1248                            MachineBasicBlock *BB2, MachineRegisterInfo *MRI) {
1249   unsigned SrcReg = Reg;
1250   while (true) {
1251     unsigned NextReg = SrcReg;
1252     MachineInstr *Inst = MRI->getVRegDef(SrcReg);
1253     if (BB1 && Inst->getOpcode() == PPC::PHI && Inst->getParent() == BB2) {
1254       NextReg = getIncomingRegForBlock(Inst, BB1);
1255       // We track through PHI only once to avoid infinite loop.
1256       BB1 = nullptr;
1257     }
1258     else if (Inst->isFullCopy())
1259       NextReg = Inst->getOperand(1).getReg();
1260     if (NextReg == SrcReg || !Register::isVirtualRegister(NextReg))
1261       break;
1262     SrcReg = NextReg;
1263   }
1264   return SrcReg;
1265 }
1266 
1267 static bool eligibleForCompareElimination(MachineBasicBlock &MBB,
1268                                           MachineBasicBlock *&PredMBB,
1269                                           MachineBasicBlock *&MBBtoMoveCmp,
1270                                           MachineRegisterInfo *MRI) {
1271 
1272   auto isEligibleBB = [&](MachineBasicBlock &BB) {
1273     auto BII = BB.getFirstInstrTerminator();
1274     // We optimize BBs ending with a conditional branch.
1275     // We check only for BCC here, not BCCLR, because BCCLR
1276     // will be formed only later in the pipeline.
1277     if (BB.succ_size() == 2 &&
1278         BII != BB.instr_end() &&
1279         (*BII).getOpcode() == PPC::BCC &&
1280         (*BII).getOperand(1).isReg()) {
1281       // We optimize only if the condition code is used only by one BCC.
1282       Register CndReg = (*BII).getOperand(1).getReg();
1283       if (!CndReg.isVirtual() || !MRI->hasOneNonDBGUse(CndReg))
1284         return false;
1285 
1286       MachineInstr *CMPI = MRI->getVRegDef(CndReg);
1287       // We assume compare and branch are in the same BB for ease of analysis.
1288       if (CMPI->getParent() != &BB)
1289         return false;
1290 
1291       // We skip this BB if a physical register is used in comparison.
1292       for (MachineOperand &MO : CMPI->operands())
1293         if (MO.isReg() && !MO.getReg().isVirtual())
1294           return false;
1295 
1296       return true;
1297     }
1298     return false;
1299   };
1300 
1301   // If this BB has more than one successor, we can create a new BB and
1302   // move the compare instruction in the new BB.
1303   // So far, we do not move compare instruction to a BB having multiple
1304   // successors to avoid potentially increasing code size.
1305   auto isEligibleForMoveCmp = [](MachineBasicBlock &BB) {
1306     return BB.succ_size() == 1;
1307   };
1308 
1309   if (!isEligibleBB(MBB))
1310     return false;
1311 
1312   unsigned NumPredBBs = MBB.pred_size();
1313   if (NumPredBBs == 1) {
1314     MachineBasicBlock *TmpMBB = *MBB.pred_begin();
1315     if (isEligibleBB(*TmpMBB)) {
1316       PredMBB = TmpMBB;
1317       MBBtoMoveCmp = nullptr;
1318       return true;
1319     }
1320   }
1321   else if (NumPredBBs == 2) {
1322     // We check for partially redundant case.
1323     // So far, we support cases with only two predecessors
1324     // to avoid increasing the number of instructions.
1325     MachineBasicBlock::pred_iterator PI = MBB.pred_begin();
1326     MachineBasicBlock *Pred1MBB = *PI;
1327     MachineBasicBlock *Pred2MBB = *(PI+1);
1328 
1329     if (isEligibleBB(*Pred1MBB) && isEligibleForMoveCmp(*Pred2MBB)) {
1330       // We assume Pred1MBB is the BB containing the compare to be merged and
1331       // Pred2MBB is the BB to which we will append a compare instruction.
1332       // Proceed as is if Pred1MBB is different from MBB.
1333     }
1334     else if (isEligibleBB(*Pred2MBB) && isEligibleForMoveCmp(*Pred1MBB)) {
1335       // We need to swap Pred1MBB and Pred2MBB to canonicalize.
1336       std::swap(Pred1MBB, Pred2MBB);
1337     }
1338     else return false;
1339 
1340     if (Pred1MBB == &MBB)
1341       return false;
1342 
1343     // Here, Pred2MBB is the BB to which we need to append a compare inst.
1344     // We cannot move the compare instruction if operands are not available
1345     // in Pred2MBB (i.e. defined in MBB by an instruction other than PHI).
1346     MachineInstr *BI = &*MBB.getFirstInstrTerminator();
1347     MachineInstr *CMPI = MRI->getVRegDef(BI->getOperand(1).getReg());
1348     for (int I = 1; I <= 2; I++)
1349       if (CMPI->getOperand(I).isReg()) {
1350         MachineInstr *Inst = MRI->getVRegDef(CMPI->getOperand(I).getReg());
1351         if (Inst->getParent() == &MBB && Inst->getOpcode() != PPC::PHI)
1352           return false;
1353       }
1354 
1355     PredMBB = Pred1MBB;
1356     MBBtoMoveCmp = Pred2MBB;
1357     return true;
1358   }
1359 
1360   return false;
1361 }
1362 
1363 // This function will iterate over the input map containing a pair of TOC save
1364 // instruction and a flag. The flag will be set to false if the TOC save is
1365 // proven redundant. This function will erase from the basic block all the TOC
1366 // saves marked as redundant.
1367 bool PPCMIPeephole::eliminateRedundantTOCSaves(
1368     std::map<MachineInstr *, bool> &TOCSaves) {
1369   bool Simplified = false;
1370   int NumKept = 0;
1371   for (auto TOCSave : TOCSaves) {
1372     if (!TOCSave.second) {
1373       TOCSave.first->eraseFromParent();
1374       RemoveTOCSave++;
1375       Simplified = true;
1376     } else {
1377       NumKept++;
1378     }
1379   }
1380 
1381   if (NumKept > 1)
1382     MultiTOCSaves++;
1383 
1384   return Simplified;
1385 }
1386 
1387 // If multiple conditional branches are executed based on the (essentially)
1388 // same comparison, we merge compare instructions into one and make multiple
1389 // conditional branches on this comparison.
1390 // For example,
1391 //   if (a == 0) { ... }
1392 //   else if (a < 0) { ... }
1393 // can be executed by one compare and two conditional branches instead of
1394 // two pairs of a compare and a conditional branch.
1395 //
1396 // This method merges two compare instructions in two MBBs and modifies the
1397 // compare and conditional branch instructions if needed.
1398 // For the above example, the input for this pass looks like:
1399 //   cmplwi r3, 0
1400 //   beq    0, .LBB0_3
1401 //   cmpwi  r3, -1
1402 //   bgt    0, .LBB0_4
1403 // So, before merging two compares, we need to modify these instructions as
1404 //   cmpwi  r3, 0       ; cmplwi and cmpwi yield same result for beq
1405 //   beq    0, .LBB0_3
1406 //   cmpwi  r3, 0       ; greather than -1 means greater or equal to 0
1407 //   bge    0, .LBB0_4
1408 
1409 bool PPCMIPeephole::eliminateRedundantCompare() {
1410   bool Simplified = false;
1411 
1412   for (MachineBasicBlock &MBB2 : *MF) {
1413     MachineBasicBlock *MBB1 = nullptr, *MBBtoMoveCmp = nullptr;
1414 
1415     // For fully redundant case, we select two basic blocks MBB1 and MBB2
1416     // as an optimization target if
1417     // - both MBBs end with a conditional branch,
1418     // - MBB1 is the only predecessor of MBB2, and
1419     // - compare does not take a physical register as a operand in both MBBs.
1420     // In this case, eligibleForCompareElimination sets MBBtoMoveCmp nullptr.
1421     //
1422     // As partially redundant case, we additionally handle if MBB2 has one
1423     // additional predecessor, which has only one successor (MBB2).
1424     // In this case, we move the compare instruction originally in MBB2 into
1425     // MBBtoMoveCmp. This partially redundant case is typically appear by
1426     // compiling a while loop; here, MBBtoMoveCmp is the loop preheader.
1427     //
1428     // Overview of CFG of related basic blocks
1429     // Fully redundant case        Partially redundant case
1430     //   --------                   ----------------  --------
1431     //   | MBB1 | (w/ 2 succ)       | MBBtoMoveCmp |  | MBB1 | (w/ 2 succ)
1432     //   --------                   ----------------  --------
1433     //      |    \                     (w/ 1 succ) \     |    \
1434     //      |     \                                 \    |     \
1435     //      |                                        \   |
1436     //   --------                                     --------
1437     //   | MBB2 | (w/ 1 pred                          | MBB2 | (w/ 2 pred
1438     //   -------- and 2 succ)                         -------- and 2 succ)
1439     //      |    \                                       |    \
1440     //      |     \                                      |     \
1441     //
1442     if (!eligibleForCompareElimination(MBB2, MBB1, MBBtoMoveCmp, MRI))
1443       continue;
1444 
1445     MachineInstr *BI1   = &*MBB1->getFirstInstrTerminator();
1446     MachineInstr *CMPI1 = MRI->getVRegDef(BI1->getOperand(1).getReg());
1447 
1448     MachineInstr *BI2   = &*MBB2.getFirstInstrTerminator();
1449     MachineInstr *CMPI2 = MRI->getVRegDef(BI2->getOperand(1).getReg());
1450     bool IsPartiallyRedundant = (MBBtoMoveCmp != nullptr);
1451 
1452     // We cannot optimize an unsupported compare opcode or
1453     // a mix of 32-bit and 64-bit comparisons
1454     if (!isSupportedCmpOp(CMPI1->getOpcode()) ||
1455         !isSupportedCmpOp(CMPI2->getOpcode()) ||
1456         is64bitCmpOp(CMPI1->getOpcode()) != is64bitCmpOp(CMPI2->getOpcode()))
1457       continue;
1458 
1459     unsigned NewOpCode = 0;
1460     unsigned NewPredicate1 = 0, NewPredicate2 = 0;
1461     int16_t Imm1 = 0, NewImm1 = 0, Imm2 = 0, NewImm2 = 0;
1462     bool SwapOperands = false;
1463 
1464     if (CMPI1->getOpcode() != CMPI2->getOpcode()) {
1465       // Typically, unsigned comparison is used for equality check, but
1466       // we replace it with a signed comparison if the comparison
1467       // to be merged is a signed comparison.
1468       // In other cases of opcode mismatch, we cannot optimize this.
1469 
1470       // We cannot change opcode when comparing against an immediate
1471       // if the most significant bit of the immediate is one
1472       // due to the difference in sign extension.
1473       auto CmpAgainstImmWithSignBit = [](MachineInstr *I) {
1474         if (!I->getOperand(2).isImm())
1475           return false;
1476         int16_t Imm = (int16_t)I->getOperand(2).getImm();
1477         return Imm < 0;
1478       };
1479 
1480       if (isEqOrNe(BI2) && !CmpAgainstImmWithSignBit(CMPI2) &&
1481           CMPI1->getOpcode() == getSignedCmpOpCode(CMPI2->getOpcode()))
1482         NewOpCode = CMPI1->getOpcode();
1483       else if (isEqOrNe(BI1) && !CmpAgainstImmWithSignBit(CMPI1) &&
1484                getSignedCmpOpCode(CMPI1->getOpcode()) == CMPI2->getOpcode())
1485         NewOpCode = CMPI2->getOpcode();
1486       else continue;
1487     }
1488 
1489     if (CMPI1->getOperand(2).isReg() && CMPI2->getOperand(2).isReg()) {
1490       // In case of comparisons between two registers, these two registers
1491       // must be same to merge two comparisons.
1492       unsigned Cmp1Operand1 = getSrcVReg(CMPI1->getOperand(1).getReg(),
1493                                          nullptr, nullptr, MRI);
1494       unsigned Cmp1Operand2 = getSrcVReg(CMPI1->getOperand(2).getReg(),
1495                                          nullptr, nullptr, MRI);
1496       unsigned Cmp2Operand1 = getSrcVReg(CMPI2->getOperand(1).getReg(),
1497                                          MBB1, &MBB2, MRI);
1498       unsigned Cmp2Operand2 = getSrcVReg(CMPI2->getOperand(2).getReg(),
1499                                          MBB1, &MBB2, MRI);
1500 
1501       if (Cmp1Operand1 == Cmp2Operand1 && Cmp1Operand2 == Cmp2Operand2) {
1502         // Same pair of registers in the same order; ready to merge as is.
1503       }
1504       else if (Cmp1Operand1 == Cmp2Operand2 && Cmp1Operand2 == Cmp2Operand1) {
1505         // Same pair of registers in different order.
1506         // We reverse the predicate to merge compare instructions.
1507         PPC::Predicate Pred = (PPC::Predicate)BI2->getOperand(0).getImm();
1508         NewPredicate2 = (unsigned)PPC::getSwappedPredicate(Pred);
1509         // In case of partial redundancy, we need to swap operands
1510         // in another compare instruction.
1511         SwapOperands = true;
1512       }
1513       else continue;
1514     }
1515     else if (CMPI1->getOperand(2).isImm() && CMPI2->getOperand(2).isImm()) {
1516       // In case of comparisons between a register and an immediate,
1517       // the operand register must be same for two compare instructions.
1518       unsigned Cmp1Operand1 = getSrcVReg(CMPI1->getOperand(1).getReg(),
1519                                          nullptr, nullptr, MRI);
1520       unsigned Cmp2Operand1 = getSrcVReg(CMPI2->getOperand(1).getReg(),
1521                                          MBB1, &MBB2, MRI);
1522       if (Cmp1Operand1 != Cmp2Operand1)
1523         continue;
1524 
1525       NewImm1 = Imm1 = (int16_t)CMPI1->getOperand(2).getImm();
1526       NewImm2 = Imm2 = (int16_t)CMPI2->getOperand(2).getImm();
1527 
1528       // If immediate are not same, we try to adjust by changing predicate;
1529       // e.g. GT imm means GE (imm+1).
1530       if (Imm1 != Imm2 && (!isEqOrNe(BI2) || !isEqOrNe(BI1))) {
1531         int Diff = Imm1 - Imm2;
1532         if (Diff < -2 || Diff > 2)
1533           continue;
1534 
1535         unsigned PredToInc1 = getPredicateToIncImm(BI1, CMPI1);
1536         unsigned PredToDec1 = getPredicateToDecImm(BI1, CMPI1);
1537         unsigned PredToInc2 = getPredicateToIncImm(BI2, CMPI2);
1538         unsigned PredToDec2 = getPredicateToDecImm(BI2, CMPI2);
1539         if (Diff == 2) {
1540           if (PredToInc2 && PredToDec1) {
1541             NewPredicate2 = PredToInc2;
1542             NewPredicate1 = PredToDec1;
1543             NewImm2++;
1544             NewImm1--;
1545           }
1546         }
1547         else if (Diff == 1) {
1548           if (PredToInc2) {
1549             NewImm2++;
1550             NewPredicate2 = PredToInc2;
1551           }
1552           else if (PredToDec1) {
1553             NewImm1--;
1554             NewPredicate1 = PredToDec1;
1555           }
1556         }
1557         else if (Diff == -1) {
1558           if (PredToDec2) {
1559             NewImm2--;
1560             NewPredicate2 = PredToDec2;
1561           }
1562           else if (PredToInc1) {
1563             NewImm1++;
1564             NewPredicate1 = PredToInc1;
1565           }
1566         }
1567         else if (Diff == -2) {
1568           if (PredToDec2 && PredToInc1) {
1569             NewPredicate2 = PredToDec2;
1570             NewPredicate1 = PredToInc1;
1571             NewImm2--;
1572             NewImm1++;
1573           }
1574         }
1575       }
1576 
1577       // We cannot merge two compares if the immediates are not same.
1578       if (NewImm2 != NewImm1)
1579         continue;
1580     }
1581 
1582     LLVM_DEBUG(dbgs() << "Optimize two pairs of compare and branch:\n");
1583     LLVM_DEBUG(CMPI1->dump());
1584     LLVM_DEBUG(BI1->dump());
1585     LLVM_DEBUG(CMPI2->dump());
1586     LLVM_DEBUG(BI2->dump());
1587 
1588     // We adjust opcode, predicates and immediate as we determined above.
1589     if (NewOpCode != 0 && NewOpCode != CMPI1->getOpcode()) {
1590       CMPI1->setDesc(TII->get(NewOpCode));
1591     }
1592     if (NewPredicate1) {
1593       BI1->getOperand(0).setImm(NewPredicate1);
1594     }
1595     if (NewPredicate2) {
1596       BI2->getOperand(0).setImm(NewPredicate2);
1597     }
1598     if (NewImm1 != Imm1) {
1599       CMPI1->getOperand(2).setImm(NewImm1);
1600     }
1601 
1602     if (IsPartiallyRedundant) {
1603       // We touch up the compare instruction in MBB2 and move it to
1604       // a previous BB to handle partially redundant case.
1605       if (SwapOperands) {
1606         Register Op1 = CMPI2->getOperand(1).getReg();
1607         Register Op2 = CMPI2->getOperand(2).getReg();
1608         CMPI2->getOperand(1).setReg(Op2);
1609         CMPI2->getOperand(2).setReg(Op1);
1610       }
1611       if (NewImm2 != Imm2)
1612         CMPI2->getOperand(2).setImm(NewImm2);
1613 
1614       for (int I = 1; I <= 2; I++) {
1615         if (CMPI2->getOperand(I).isReg()) {
1616           MachineInstr *Inst = MRI->getVRegDef(CMPI2->getOperand(I).getReg());
1617           if (Inst->getParent() != &MBB2)
1618             continue;
1619 
1620           assert(Inst->getOpcode() == PPC::PHI &&
1621                  "We cannot support if an operand comes from this BB.");
1622           unsigned SrcReg = getIncomingRegForBlock(Inst, MBBtoMoveCmp);
1623           CMPI2->getOperand(I).setReg(SrcReg);
1624         }
1625       }
1626       auto I = MachineBasicBlock::iterator(MBBtoMoveCmp->getFirstTerminator());
1627       MBBtoMoveCmp->splice(I, &MBB2, MachineBasicBlock::iterator(CMPI2));
1628 
1629       DebugLoc DL = CMPI2->getDebugLoc();
1630       Register NewVReg = MRI->createVirtualRegister(&PPC::CRRCRegClass);
1631       BuildMI(MBB2, MBB2.begin(), DL,
1632               TII->get(PPC::PHI), NewVReg)
1633         .addReg(BI1->getOperand(1).getReg()).addMBB(MBB1)
1634         .addReg(BI2->getOperand(1).getReg()).addMBB(MBBtoMoveCmp);
1635       BI2->getOperand(1).setReg(NewVReg);
1636     }
1637     else {
1638       // We finally eliminate compare instruction in MBB2.
1639       BI2->getOperand(1).setReg(BI1->getOperand(1).getReg());
1640       CMPI2->eraseFromParent();
1641     }
1642     BI2->getOperand(1).setIsKill(true);
1643     BI1->getOperand(1).setIsKill(false);
1644 
1645     LLVM_DEBUG(dbgs() << "into a compare and two branches:\n");
1646     LLVM_DEBUG(CMPI1->dump());
1647     LLVM_DEBUG(BI1->dump());
1648     LLVM_DEBUG(BI2->dump());
1649     if (IsPartiallyRedundant) {
1650       LLVM_DEBUG(dbgs() << "The following compare is moved into "
1651                         << printMBBReference(*MBBtoMoveCmp)
1652                         << " to handle partial redundancy.\n");
1653       LLVM_DEBUG(CMPI2->dump());
1654     }
1655 
1656     Simplified = true;
1657   }
1658 
1659   return Simplified;
1660 }
1661 
1662 // We miss the opportunity to emit an RLDIC when lowering jump tables
1663 // since ISEL sees only a single basic block. When selecting, the clear
1664 // and shift left will be in different blocks.
1665 bool PPCMIPeephole::emitRLDICWhenLoweringJumpTables(MachineInstr &MI) {
1666   if (MI.getOpcode() != PPC::RLDICR)
1667     return false;
1668 
1669   Register SrcReg = MI.getOperand(1).getReg();
1670   if (!SrcReg.isVirtual())
1671     return false;
1672 
1673   MachineInstr *SrcMI = MRI->getVRegDef(SrcReg);
1674   if (SrcMI->getOpcode() != PPC::RLDICL)
1675     return false;
1676 
1677   MachineOperand MOpSHSrc = SrcMI->getOperand(2);
1678   MachineOperand MOpMBSrc = SrcMI->getOperand(3);
1679   MachineOperand MOpSHMI = MI.getOperand(2);
1680   MachineOperand MOpMEMI = MI.getOperand(3);
1681   if (!(MOpSHSrc.isImm() && MOpMBSrc.isImm() && MOpSHMI.isImm() &&
1682         MOpMEMI.isImm()))
1683     return false;
1684 
1685   uint64_t SHSrc = MOpSHSrc.getImm();
1686   uint64_t MBSrc = MOpMBSrc.getImm();
1687   uint64_t SHMI = MOpSHMI.getImm();
1688   uint64_t MEMI = MOpMEMI.getImm();
1689   uint64_t NewSH = SHSrc + SHMI;
1690   uint64_t NewMB = MBSrc - SHMI;
1691   if (NewMB > 63 || NewSH > 63)
1692     return false;
1693 
1694   // The bits cleared with RLDICL are [0, MBSrc).
1695   // The bits cleared with RLDICR are (MEMI, 63].
1696   // After the sequence, the bits cleared are:
1697   // [0, MBSrc-SHMI) and (MEMI, 63).
1698   //
1699   // The bits cleared with RLDIC are [0, NewMB) and (63-NewSH, 63].
1700   if ((63 - NewSH) != MEMI)
1701     return false;
1702 
1703   LLVM_DEBUG(dbgs() << "Converting pair: ");
1704   LLVM_DEBUG(SrcMI->dump());
1705   LLVM_DEBUG(MI.dump());
1706 
1707   MI.setDesc(TII->get(PPC::RLDIC));
1708   MI.getOperand(1).setReg(SrcMI->getOperand(1).getReg());
1709   MI.getOperand(2).setImm(NewSH);
1710   MI.getOperand(3).setImm(NewMB);
1711   MI.getOperand(1).setIsKill(SrcMI->getOperand(1).isKill());
1712   SrcMI->getOperand(1).setIsKill(false);
1713 
1714   LLVM_DEBUG(dbgs() << "To: ");
1715   LLVM_DEBUG(MI.dump());
1716   NumRotatesCollapsed++;
1717   // If SrcReg has no non-debug use it's safe to delete its def SrcMI.
1718   if (MRI->use_nodbg_empty(SrcReg)) {
1719     assert(!SrcMI->hasImplicitDef() &&
1720            "Not expecting an implicit def with this instr.");
1721     SrcMI->eraseFromParent();
1722   }
1723   return true;
1724 }
1725 
1726 // For case in LLVM IR
1727 // entry:
1728 //   %iconv = sext i32 %index to i64
1729 //   br i1 undef label %true, label %false
1730 // true:
1731 //   %ptr = getelementptr inbounds i32, i32* null, i64 %iconv
1732 // ...
1733 // PPCISelLowering::combineSHL fails to combine, because sext and shl are in
1734 // different BBs when conducting instruction selection. We can do a peephole
1735 // optimization to combine these two instructions into extswsli after
1736 // instruction selection.
1737 bool PPCMIPeephole::combineSEXTAndSHL(MachineInstr &MI,
1738                                       MachineInstr *&ToErase) {
1739   if (MI.getOpcode() != PPC::RLDICR)
1740     return false;
1741 
1742   if (!MF->getSubtarget<PPCSubtarget>().isISA3_0())
1743     return false;
1744 
1745   assert(MI.getNumOperands() == 4 && "RLDICR should have 4 operands");
1746 
1747   MachineOperand MOpSHMI = MI.getOperand(2);
1748   MachineOperand MOpMEMI = MI.getOperand(3);
1749   if (!(MOpSHMI.isImm() && MOpMEMI.isImm()))
1750     return false;
1751 
1752   uint64_t SHMI = MOpSHMI.getImm();
1753   uint64_t MEMI = MOpMEMI.getImm();
1754   if (SHMI + MEMI != 63)
1755     return false;
1756 
1757   Register SrcReg = MI.getOperand(1).getReg();
1758   if (!SrcReg.isVirtual())
1759     return false;
1760 
1761   MachineInstr *SrcMI = MRI->getVRegDef(SrcReg);
1762   if (SrcMI->getOpcode() != PPC::EXTSW &&
1763       SrcMI->getOpcode() != PPC::EXTSW_32_64)
1764     return false;
1765 
1766   // If the register defined by extsw has more than one use, combination is not
1767   // needed.
1768   if (!MRI->hasOneNonDBGUse(SrcReg))
1769     return false;
1770 
1771   assert(SrcMI->getNumOperands() == 2 && "EXTSW should have 2 operands");
1772   assert(SrcMI->getOperand(1).isReg() &&
1773          "EXTSW's second operand should be a register");
1774   if (!SrcMI->getOperand(1).getReg().isVirtual())
1775     return false;
1776 
1777   LLVM_DEBUG(dbgs() << "Combining pair: ");
1778   LLVM_DEBUG(SrcMI->dump());
1779   LLVM_DEBUG(MI.dump());
1780 
1781   MachineInstr *NewInstr =
1782       BuildMI(*MI.getParent(), &MI, MI.getDebugLoc(),
1783               SrcMI->getOpcode() == PPC::EXTSW ? TII->get(PPC::EXTSWSLI)
1784                                                : TII->get(PPC::EXTSWSLI_32_64),
1785               MI.getOperand(0).getReg())
1786           .add(SrcMI->getOperand(1))
1787           .add(MOpSHMI);
1788   (void)NewInstr;
1789 
1790   LLVM_DEBUG(dbgs() << "TO: ");
1791   LLVM_DEBUG(NewInstr->dump());
1792   ++NumEXTSWAndSLDICombined;
1793   ToErase = &MI;
1794   // SrcMI, which is extsw, is of no use now, erase it.
1795   SrcMI->eraseFromParent();
1796   return true;
1797 }
1798 
1799 } // end default namespace
1800 
1801 INITIALIZE_PASS_BEGIN(PPCMIPeephole, DEBUG_TYPE,
1802                       "PowerPC MI Peephole Optimization", false, false)
1803 INITIALIZE_PASS_DEPENDENCY(MachineBlockFrequencyInfo)
1804 INITIALIZE_PASS_DEPENDENCY(MachineDominatorTree)
1805 INITIALIZE_PASS_DEPENDENCY(MachinePostDominatorTree)
1806 INITIALIZE_PASS_END(PPCMIPeephole, DEBUG_TYPE,
1807                     "PowerPC MI Peephole Optimization", false, false)
1808 
1809 char PPCMIPeephole::ID = 0;
1810 FunctionPass*
1811 llvm::createPPCMIPeepholePass() { return new PPCMIPeephole(); }
1812