xref: /freebsd/contrib/llvm-project/llvm/lib/CodeGen/TwoAddressInstructionPass.cpp (revision 9e5787d2284e187abb5b654d924394a65772e004)
1 //===- TwoAddressInstructionPass.cpp - Two-Address instruction pass -------===//
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 file implements the TwoAddress instruction pass which is used
10 // by most register allocators. Two-Address instructions are rewritten
11 // from:
12 //
13 //     A = B op C
14 //
15 // to:
16 //
17 //     A = B
18 //     A op= C
19 //
20 // Note that if a register allocator chooses to use this pass, that it
21 // has to be capable of handling the non-SSA nature of these rewritten
22 // virtual registers.
23 //
24 // It is also worth noting that the duplicate operand of the two
25 // address instruction is removed.
26 //
27 //===----------------------------------------------------------------------===//
28 
29 #include "llvm/ADT/DenseMap.h"
30 #include "llvm/ADT/SmallPtrSet.h"
31 #include "llvm/ADT/SmallSet.h"
32 #include "llvm/ADT/SmallVector.h"
33 #include "llvm/ADT/Statistic.h"
34 #include "llvm/ADT/iterator_range.h"
35 #include "llvm/Analysis/AliasAnalysis.h"
36 #include "llvm/CodeGen/LiveInterval.h"
37 #include "llvm/CodeGen/LiveIntervals.h"
38 #include "llvm/CodeGen/LiveVariables.h"
39 #include "llvm/CodeGen/MachineBasicBlock.h"
40 #include "llvm/CodeGen/MachineFunction.h"
41 #include "llvm/CodeGen/MachineFunctionPass.h"
42 #include "llvm/CodeGen/MachineInstr.h"
43 #include "llvm/CodeGen/MachineInstrBuilder.h"
44 #include "llvm/CodeGen/MachineOperand.h"
45 #include "llvm/CodeGen/MachineRegisterInfo.h"
46 #include "llvm/CodeGen/Passes.h"
47 #include "llvm/CodeGen/SlotIndexes.h"
48 #include "llvm/CodeGen/TargetInstrInfo.h"
49 #include "llvm/CodeGen/TargetOpcodes.h"
50 #include "llvm/CodeGen/TargetRegisterInfo.h"
51 #include "llvm/CodeGen/TargetSubtargetInfo.h"
52 #include "llvm/MC/MCInstrDesc.h"
53 #include "llvm/MC/MCInstrItineraries.h"
54 #include "llvm/Pass.h"
55 #include "llvm/Support/CodeGen.h"
56 #include "llvm/Support/CommandLine.h"
57 #include "llvm/Support/Debug.h"
58 #include "llvm/Support/ErrorHandling.h"
59 #include "llvm/Support/raw_ostream.h"
60 #include "llvm/Target/TargetMachine.h"
61 #include <cassert>
62 #include <iterator>
63 #include <utility>
64 
65 using namespace llvm;
66 
67 #define DEBUG_TYPE "twoaddressinstruction"
68 
69 STATISTIC(NumTwoAddressInstrs, "Number of two-address instructions");
70 STATISTIC(NumCommuted        , "Number of instructions commuted to coalesce");
71 STATISTIC(NumAggrCommuted    , "Number of instructions aggressively commuted");
72 STATISTIC(NumConvertedTo3Addr, "Number of instructions promoted to 3-address");
73 STATISTIC(NumReSchedUps,       "Number of instructions re-scheduled up");
74 STATISTIC(NumReSchedDowns,     "Number of instructions re-scheduled down");
75 
76 // Temporary flag to disable rescheduling.
77 static cl::opt<bool>
78 EnableRescheduling("twoaddr-reschedule",
79                    cl::desc("Coalesce copies by rescheduling (default=true)"),
80                    cl::init(true), cl::Hidden);
81 
82 // Limit the number of dataflow edges to traverse when evaluating the benefit
83 // of commuting operands.
84 static cl::opt<unsigned> MaxDataFlowEdge(
85     "dataflow-edge-limit", cl::Hidden, cl::init(3),
86     cl::desc("Maximum number of dataflow edges to traverse when evaluating "
87              "the benefit of commuting operands"));
88 
89 namespace {
90 
91 class TwoAddressInstructionPass : public MachineFunctionPass {
92   MachineFunction *MF;
93   const TargetInstrInfo *TII;
94   const TargetRegisterInfo *TRI;
95   const InstrItineraryData *InstrItins;
96   MachineRegisterInfo *MRI;
97   LiveVariables *LV;
98   LiveIntervals *LIS;
99   AliasAnalysis *AA;
100   CodeGenOpt::Level OptLevel;
101 
102   // The current basic block being processed.
103   MachineBasicBlock *MBB;
104 
105   // Keep track the distance of a MI from the start of the current basic block.
106   DenseMap<MachineInstr*, unsigned> DistanceMap;
107 
108   // Set of already processed instructions in the current block.
109   SmallPtrSet<MachineInstr*, 8> Processed;
110 
111   // A map from virtual registers to physical registers which are likely targets
112   // to be coalesced to due to copies from physical registers to virtual
113   // registers. e.g. v1024 = move r0.
114   DenseMap<unsigned, unsigned> SrcRegMap;
115 
116   // A map from virtual registers to physical registers which are likely targets
117   // to be coalesced to due to copies to physical registers from virtual
118   // registers. e.g. r1 = move v1024.
119   DenseMap<unsigned, unsigned> DstRegMap;
120 
121   bool isRevCopyChain(unsigned FromReg, unsigned ToReg, int Maxlen);
122 
123   bool noUseAfterLastDef(unsigned Reg, unsigned Dist, unsigned &LastDef);
124 
125   bool isProfitableToCommute(unsigned regA, unsigned regB, unsigned regC,
126                              MachineInstr *MI, unsigned Dist);
127 
128   bool commuteInstruction(MachineInstr *MI, unsigned DstIdx,
129                           unsigned RegBIdx, unsigned RegCIdx, unsigned Dist);
130 
131   bool isProfitableToConv3Addr(unsigned RegA, unsigned RegB);
132 
133   bool convertInstTo3Addr(MachineBasicBlock::iterator &mi,
134                           MachineBasicBlock::iterator &nmi,
135                           unsigned RegA, unsigned RegB, unsigned Dist);
136 
137   bool isDefTooClose(unsigned Reg, unsigned Dist, MachineInstr *MI);
138 
139   bool rescheduleMIBelowKill(MachineBasicBlock::iterator &mi,
140                              MachineBasicBlock::iterator &nmi,
141                              unsigned Reg);
142   bool rescheduleKillAboveMI(MachineBasicBlock::iterator &mi,
143                              MachineBasicBlock::iterator &nmi,
144                              unsigned Reg);
145 
146   bool tryInstructionTransform(MachineBasicBlock::iterator &mi,
147                                MachineBasicBlock::iterator &nmi,
148                                unsigned SrcIdx, unsigned DstIdx,
149                                unsigned Dist, bool shouldOnlyCommute);
150 
151   bool tryInstructionCommute(MachineInstr *MI,
152                              unsigned DstOpIdx,
153                              unsigned BaseOpIdx,
154                              bool BaseOpKilled,
155                              unsigned Dist);
156   void scanUses(unsigned DstReg);
157 
158   void processCopy(MachineInstr *MI);
159 
160   using TiedPairList = SmallVector<std::pair<unsigned, unsigned>, 4>;
161   using TiedOperandMap = SmallDenseMap<unsigned, TiedPairList>;
162 
163   bool collectTiedOperands(MachineInstr *MI, TiedOperandMap&);
164   void processTiedPairs(MachineInstr *MI, TiedPairList&, unsigned &Dist);
165   void eliminateRegSequence(MachineBasicBlock::iterator&);
166 
167 public:
168   static char ID; // Pass identification, replacement for typeid
169 
170   TwoAddressInstructionPass() : MachineFunctionPass(ID) {
171     initializeTwoAddressInstructionPassPass(*PassRegistry::getPassRegistry());
172   }
173 
174   void getAnalysisUsage(AnalysisUsage &AU) const override {
175     AU.setPreservesCFG();
176     AU.addUsedIfAvailable<AAResultsWrapperPass>();
177     AU.addUsedIfAvailable<LiveVariables>();
178     AU.addPreserved<LiveVariables>();
179     AU.addPreserved<SlotIndexes>();
180     AU.addPreserved<LiveIntervals>();
181     AU.addPreservedID(MachineLoopInfoID);
182     AU.addPreservedID(MachineDominatorsID);
183     MachineFunctionPass::getAnalysisUsage(AU);
184   }
185 
186   /// Pass entry point.
187   bool runOnMachineFunction(MachineFunction&) override;
188 };
189 
190 } // end anonymous namespace
191 
192 char TwoAddressInstructionPass::ID = 0;
193 
194 char &llvm::TwoAddressInstructionPassID = TwoAddressInstructionPass::ID;
195 
196 INITIALIZE_PASS_BEGIN(TwoAddressInstructionPass, DEBUG_TYPE,
197                 "Two-Address instruction pass", false, false)
198 INITIALIZE_PASS_DEPENDENCY(AAResultsWrapperPass)
199 INITIALIZE_PASS_END(TwoAddressInstructionPass, DEBUG_TYPE,
200                 "Two-Address instruction pass", false, false)
201 
202 static bool isPlainlyKilled(MachineInstr *MI, unsigned Reg, LiveIntervals *LIS);
203 
204 /// Return the MachineInstr* if it is the single def of the Reg in current BB.
205 static MachineInstr *getSingleDef(unsigned Reg, MachineBasicBlock *BB,
206                                   const MachineRegisterInfo *MRI) {
207   MachineInstr *Ret = nullptr;
208   for (MachineInstr &DefMI : MRI->def_instructions(Reg)) {
209     if (DefMI.getParent() != BB || DefMI.isDebugValue())
210       continue;
211     if (!Ret)
212       Ret = &DefMI;
213     else if (Ret != &DefMI)
214       return nullptr;
215   }
216   return Ret;
217 }
218 
219 /// Check if there is a reversed copy chain from FromReg to ToReg:
220 /// %Tmp1 = copy %Tmp2;
221 /// %FromReg = copy %Tmp1;
222 /// %ToReg = add %FromReg ...
223 /// %Tmp2 = copy %ToReg;
224 /// MaxLen specifies the maximum length of the copy chain the func
225 /// can walk through.
226 bool TwoAddressInstructionPass::isRevCopyChain(unsigned FromReg, unsigned ToReg,
227                                                int Maxlen) {
228   unsigned TmpReg = FromReg;
229   for (int i = 0; i < Maxlen; i++) {
230     MachineInstr *Def = getSingleDef(TmpReg, MBB, MRI);
231     if (!Def || !Def->isCopy())
232       return false;
233 
234     TmpReg = Def->getOperand(1).getReg();
235 
236     if (TmpReg == ToReg)
237       return true;
238   }
239   return false;
240 }
241 
242 /// Return true if there are no intervening uses between the last instruction
243 /// in the MBB that defines the specified register and the two-address
244 /// instruction which is being processed. It also returns the last def location
245 /// by reference.
246 bool TwoAddressInstructionPass::noUseAfterLastDef(unsigned Reg, unsigned Dist,
247                                                   unsigned &LastDef) {
248   LastDef = 0;
249   unsigned LastUse = Dist;
250   for (MachineOperand &MO : MRI->reg_operands(Reg)) {
251     MachineInstr *MI = MO.getParent();
252     if (MI->getParent() != MBB || MI->isDebugValue())
253       continue;
254     DenseMap<MachineInstr*, unsigned>::iterator DI = DistanceMap.find(MI);
255     if (DI == DistanceMap.end())
256       continue;
257     if (MO.isUse() && DI->second < LastUse)
258       LastUse = DI->second;
259     if (MO.isDef() && DI->second > LastDef)
260       LastDef = DI->second;
261   }
262 
263   return !(LastUse > LastDef && LastUse < Dist);
264 }
265 
266 /// Return true if the specified MI is a copy instruction or an extract_subreg
267 /// instruction. It also returns the source and destination registers and
268 /// whether they are physical registers by reference.
269 static bool isCopyToReg(MachineInstr &MI, const TargetInstrInfo *TII,
270                         unsigned &SrcReg, unsigned &DstReg,
271                         bool &IsSrcPhys, bool &IsDstPhys) {
272   SrcReg = 0;
273   DstReg = 0;
274   if (MI.isCopy()) {
275     DstReg = MI.getOperand(0).getReg();
276     SrcReg = MI.getOperand(1).getReg();
277   } else if (MI.isInsertSubreg() || MI.isSubregToReg()) {
278     DstReg = MI.getOperand(0).getReg();
279     SrcReg = MI.getOperand(2).getReg();
280   } else
281     return false;
282 
283   IsSrcPhys = Register::isPhysicalRegister(SrcReg);
284   IsDstPhys = Register::isPhysicalRegister(DstReg);
285   return true;
286 }
287 
288 /// Test if the given register value, which is used by the
289 /// given instruction, is killed by the given instruction.
290 static bool isPlainlyKilled(MachineInstr *MI, unsigned Reg,
291                             LiveIntervals *LIS) {
292   if (LIS && Register::isVirtualRegister(Reg) && !LIS->isNotInMIMap(*MI)) {
293     // FIXME: Sometimes tryInstructionTransform() will add instructions and
294     // test whether they can be folded before keeping them. In this case it
295     // sets a kill before recursively calling tryInstructionTransform() again.
296     // If there is no interval available, we assume that this instruction is
297     // one of those. A kill flag is manually inserted on the operand so the
298     // check below will handle it.
299     LiveInterval &LI = LIS->getInterval(Reg);
300     // This is to match the kill flag version where undefs don't have kill
301     // flags.
302     if (!LI.hasAtLeastOneValue())
303       return false;
304 
305     SlotIndex useIdx = LIS->getInstructionIndex(*MI);
306     LiveInterval::const_iterator I = LI.find(useIdx);
307     assert(I != LI.end() && "Reg must be live-in to use.");
308     return !I->end.isBlock() && SlotIndex::isSameInstr(I->end, useIdx);
309   }
310 
311   return MI->killsRegister(Reg);
312 }
313 
314 /// Test if the given register value, which is used by the given
315 /// instruction, is killed by the given instruction. This looks through
316 /// coalescable copies to see if the original value is potentially not killed.
317 ///
318 /// For example, in this code:
319 ///
320 ///   %reg1034 = copy %reg1024
321 ///   %reg1035 = copy killed %reg1025
322 ///   %reg1036 = add killed %reg1034, killed %reg1035
323 ///
324 /// %reg1034 is not considered to be killed, since it is copied from a
325 /// register which is not killed. Treating it as not killed lets the
326 /// normal heuristics commute the (two-address) add, which lets
327 /// coalescing eliminate the extra copy.
328 ///
329 /// If allowFalsePositives is true then likely kills are treated as kills even
330 /// if it can't be proven that they are kills.
331 static bool isKilled(MachineInstr &MI, unsigned Reg,
332                      const MachineRegisterInfo *MRI,
333                      const TargetInstrInfo *TII,
334                      LiveIntervals *LIS,
335                      bool allowFalsePositives) {
336   MachineInstr *DefMI = &MI;
337   while (true) {
338     // All uses of physical registers are likely to be kills.
339     if (Register::isPhysicalRegister(Reg) &&
340         (allowFalsePositives || MRI->hasOneUse(Reg)))
341       return true;
342     if (!isPlainlyKilled(DefMI, Reg, LIS))
343       return false;
344     if (Register::isPhysicalRegister(Reg))
345       return true;
346     MachineRegisterInfo::def_iterator Begin = MRI->def_begin(Reg);
347     // If there are multiple defs, we can't do a simple analysis, so just
348     // go with what the kill flag says.
349     if (std::next(Begin) != MRI->def_end())
350       return true;
351     DefMI = Begin->getParent();
352     bool IsSrcPhys, IsDstPhys;
353     unsigned SrcReg,  DstReg;
354     // If the def is something other than a copy, then it isn't going to
355     // be coalesced, so follow the kill flag.
356     if (!isCopyToReg(*DefMI, TII, SrcReg, DstReg, IsSrcPhys, IsDstPhys))
357       return true;
358     Reg = SrcReg;
359   }
360 }
361 
362 /// Return true if the specified MI uses the specified register as a two-address
363 /// use. If so, return the destination register by reference.
364 static bool isTwoAddrUse(MachineInstr &MI, unsigned Reg, unsigned &DstReg) {
365   for (unsigned i = 0, NumOps = MI.getNumOperands(); i != NumOps; ++i) {
366     const MachineOperand &MO = MI.getOperand(i);
367     if (!MO.isReg() || !MO.isUse() || MO.getReg() != Reg)
368       continue;
369     unsigned ti;
370     if (MI.isRegTiedToDefOperand(i, &ti)) {
371       DstReg = MI.getOperand(ti).getReg();
372       return true;
373     }
374   }
375   return false;
376 }
377 
378 /// Given a register, if has a single in-basic block use, return the use
379 /// instruction if it's a copy or a two-address use.
380 static
381 MachineInstr *findOnlyInterestingUse(unsigned Reg, MachineBasicBlock *MBB,
382                                      MachineRegisterInfo *MRI,
383                                      const TargetInstrInfo *TII,
384                                      bool &IsCopy,
385                                      unsigned &DstReg, bool &IsDstPhys) {
386   if (!MRI->hasOneNonDBGUse(Reg))
387     // None or more than one use.
388     return nullptr;
389   MachineInstr &UseMI = *MRI->use_instr_nodbg_begin(Reg);
390   if (UseMI.getParent() != MBB)
391     return nullptr;
392   unsigned SrcReg;
393   bool IsSrcPhys;
394   if (isCopyToReg(UseMI, TII, SrcReg, DstReg, IsSrcPhys, IsDstPhys)) {
395     IsCopy = true;
396     return &UseMI;
397   }
398   IsDstPhys = false;
399   if (isTwoAddrUse(UseMI, Reg, DstReg)) {
400     IsDstPhys = Register::isPhysicalRegister(DstReg);
401     return &UseMI;
402   }
403   return nullptr;
404 }
405 
406 /// Return the physical register the specified virtual register might be mapped
407 /// to.
408 static unsigned
409 getMappedReg(unsigned Reg, DenseMap<unsigned, unsigned> &RegMap) {
410   while (Register::isVirtualRegister(Reg)) {
411     DenseMap<unsigned, unsigned>::iterator SI = RegMap.find(Reg);
412     if (SI == RegMap.end())
413       return 0;
414     Reg = SI->second;
415   }
416   if (Register::isPhysicalRegister(Reg))
417     return Reg;
418   return 0;
419 }
420 
421 /// Return true if the two registers are equal or aliased.
422 static bool
423 regsAreCompatible(unsigned RegA, unsigned RegB, const TargetRegisterInfo *TRI) {
424   if (RegA == RegB)
425     return true;
426   if (!RegA || !RegB)
427     return false;
428   return TRI->regsOverlap(RegA, RegB);
429 }
430 
431 // Returns true if Reg is equal or aliased to at least one register in Set.
432 static bool regOverlapsSet(const SmallVectorImpl<unsigned> &Set, unsigned Reg,
433                            const TargetRegisterInfo *TRI) {
434   for (unsigned R : Set)
435     if (TRI->regsOverlap(R, Reg))
436       return true;
437 
438   return false;
439 }
440 
441 /// Return true if it's potentially profitable to commute the two-address
442 /// instruction that's being processed.
443 bool
444 TwoAddressInstructionPass::
445 isProfitableToCommute(unsigned regA, unsigned regB, unsigned regC,
446                       MachineInstr *MI, unsigned Dist) {
447   if (OptLevel == CodeGenOpt::None)
448     return false;
449 
450   // Determine if it's profitable to commute this two address instruction. In
451   // general, we want no uses between this instruction and the definition of
452   // the two-address register.
453   // e.g.
454   // %reg1028 = EXTRACT_SUBREG killed %reg1027, 1
455   // %reg1029 = COPY %reg1028
456   // %reg1029 = SHR8ri %reg1029, 7, implicit dead %eflags
457   // insert => %reg1030 = COPY %reg1028
458   // %reg1030 = ADD8rr killed %reg1028, killed %reg1029, implicit dead %eflags
459   // In this case, it might not be possible to coalesce the second COPY
460   // instruction if the first one is coalesced. So it would be profitable to
461   // commute it:
462   // %reg1028 = EXTRACT_SUBREG killed %reg1027, 1
463   // %reg1029 = COPY %reg1028
464   // %reg1029 = SHR8ri %reg1029, 7, implicit dead %eflags
465   // insert => %reg1030 = COPY %reg1029
466   // %reg1030 = ADD8rr killed %reg1029, killed %reg1028, implicit dead %eflags
467 
468   if (!isPlainlyKilled(MI, regC, LIS))
469     return false;
470 
471   // Ok, we have something like:
472   // %reg1030 = ADD8rr killed %reg1028, killed %reg1029, implicit dead %eflags
473   // let's see if it's worth commuting it.
474 
475   // Look for situations like this:
476   // %reg1024 = MOV r1
477   // %reg1025 = MOV r0
478   // %reg1026 = ADD %reg1024, %reg1025
479   // r0            = MOV %reg1026
480   // Commute the ADD to hopefully eliminate an otherwise unavoidable copy.
481   unsigned ToRegA = getMappedReg(regA, DstRegMap);
482   if (ToRegA) {
483     unsigned FromRegB = getMappedReg(regB, SrcRegMap);
484     unsigned FromRegC = getMappedReg(regC, SrcRegMap);
485     bool CompB = FromRegB && regsAreCompatible(FromRegB, ToRegA, TRI);
486     bool CompC = FromRegC && regsAreCompatible(FromRegC, ToRegA, TRI);
487 
488     // Compute if any of the following are true:
489     // -RegB is not tied to a register and RegC is compatible with RegA.
490     // -RegB is tied to the wrong physical register, but RegC is.
491     // -RegB is tied to the wrong physical register, and RegC isn't tied.
492     if ((!FromRegB && CompC) || (FromRegB && !CompB && (!FromRegC || CompC)))
493       return true;
494     // Don't compute if any of the following are true:
495     // -RegC is not tied to a register and RegB is compatible with RegA.
496     // -RegC is tied to the wrong physical register, but RegB is.
497     // -RegC is tied to the wrong physical register, and RegB isn't tied.
498     if ((!FromRegC && CompB) || (FromRegC && !CompC && (!FromRegB || CompB)))
499       return false;
500   }
501 
502   // If there is a use of regC between its last def (could be livein) and this
503   // instruction, then bail.
504   unsigned LastDefC = 0;
505   if (!noUseAfterLastDef(regC, Dist, LastDefC))
506     return false;
507 
508   // If there is a use of regB between its last def (could be livein) and this
509   // instruction, then go ahead and make this transformation.
510   unsigned LastDefB = 0;
511   if (!noUseAfterLastDef(regB, Dist, LastDefB))
512     return true;
513 
514   // Look for situation like this:
515   // %reg101 = MOV %reg100
516   // %reg102 = ...
517   // %reg103 = ADD %reg102, %reg101
518   // ... = %reg103 ...
519   // %reg100 = MOV %reg103
520   // If there is a reversed copy chain from reg101 to reg103, commute the ADD
521   // to eliminate an otherwise unavoidable copy.
522   // FIXME:
523   // We can extend the logic further: If an pair of operands in an insn has
524   // been merged, the insn could be regarded as a virtual copy, and the virtual
525   // copy could also be used to construct a copy chain.
526   // To more generally minimize register copies, ideally the logic of two addr
527   // instruction pass should be integrated with register allocation pass where
528   // interference graph is available.
529   if (isRevCopyChain(regC, regA, MaxDataFlowEdge))
530     return true;
531 
532   if (isRevCopyChain(regB, regA, MaxDataFlowEdge))
533     return false;
534 
535   // Since there are no intervening uses for both registers, then commute
536   // if the def of regC is closer. Its live interval is shorter.
537   return LastDefB && LastDefC && LastDefC > LastDefB;
538 }
539 
540 /// Commute a two-address instruction and update the basic block, distance map,
541 /// and live variables if needed. Return true if it is successful.
542 bool TwoAddressInstructionPass::commuteInstruction(MachineInstr *MI,
543                                                    unsigned DstIdx,
544                                                    unsigned RegBIdx,
545                                                    unsigned RegCIdx,
546                                                    unsigned Dist) {
547   Register RegC = MI->getOperand(RegCIdx).getReg();
548   LLVM_DEBUG(dbgs() << "2addr: COMMUTING  : " << *MI);
549   MachineInstr *NewMI = TII->commuteInstruction(*MI, false, RegBIdx, RegCIdx);
550 
551   if (NewMI == nullptr) {
552     LLVM_DEBUG(dbgs() << "2addr: COMMUTING FAILED!\n");
553     return false;
554   }
555 
556   LLVM_DEBUG(dbgs() << "2addr: COMMUTED TO: " << *NewMI);
557   assert(NewMI == MI &&
558          "TargetInstrInfo::commuteInstruction() should not return a new "
559          "instruction unless it was requested.");
560 
561   // Update source register map.
562   unsigned FromRegC = getMappedReg(RegC, SrcRegMap);
563   if (FromRegC) {
564     Register RegA = MI->getOperand(DstIdx).getReg();
565     SrcRegMap[RegA] = FromRegC;
566   }
567 
568   return true;
569 }
570 
571 /// Return true if it is profitable to convert the given 2-address instruction
572 /// to a 3-address one.
573 bool
574 TwoAddressInstructionPass::isProfitableToConv3Addr(unsigned RegA,unsigned RegB){
575   // Look for situations like this:
576   // %reg1024 = MOV r1
577   // %reg1025 = MOV r0
578   // %reg1026 = ADD %reg1024, %reg1025
579   // r2            = MOV %reg1026
580   // Turn ADD into a 3-address instruction to avoid a copy.
581   unsigned FromRegB = getMappedReg(RegB, SrcRegMap);
582   if (!FromRegB)
583     return false;
584   unsigned ToRegA = getMappedReg(RegA, DstRegMap);
585   return (ToRegA && !regsAreCompatible(FromRegB, ToRegA, TRI));
586 }
587 
588 /// Convert the specified two-address instruction into a three address one.
589 /// Return true if this transformation was successful.
590 bool
591 TwoAddressInstructionPass::convertInstTo3Addr(MachineBasicBlock::iterator &mi,
592                                               MachineBasicBlock::iterator &nmi,
593                                               unsigned RegA, unsigned RegB,
594                                               unsigned Dist) {
595   // FIXME: Why does convertToThreeAddress() need an iterator reference?
596   MachineFunction::iterator MFI = MBB->getIterator();
597   MachineInstr *NewMI = TII->convertToThreeAddress(MFI, *mi, LV);
598   assert(MBB->getIterator() == MFI &&
599          "convertToThreeAddress changed iterator reference");
600   if (!NewMI)
601     return false;
602 
603   LLVM_DEBUG(dbgs() << "2addr: CONVERTING 2-ADDR: " << *mi);
604   LLVM_DEBUG(dbgs() << "2addr:         TO 3-ADDR: " << *NewMI);
605 
606   if (LIS)
607     LIS->ReplaceMachineInstrInMaps(*mi, *NewMI);
608 
609   MBB->erase(mi); // Nuke the old inst.
610 
611   DistanceMap.insert(std::make_pair(NewMI, Dist));
612   mi = NewMI;
613   nmi = std::next(mi);
614 
615   // Update source and destination register maps.
616   SrcRegMap.erase(RegA);
617   DstRegMap.erase(RegB);
618   return true;
619 }
620 
621 /// Scan forward recursively for only uses, update maps if the use is a copy or
622 /// a two-address instruction.
623 void
624 TwoAddressInstructionPass::scanUses(unsigned DstReg) {
625   SmallVector<unsigned, 4> VirtRegPairs;
626   bool IsDstPhys;
627   bool IsCopy = false;
628   unsigned NewReg = 0;
629   unsigned Reg = DstReg;
630   while (MachineInstr *UseMI = findOnlyInterestingUse(Reg, MBB, MRI, TII,IsCopy,
631                                                       NewReg, IsDstPhys)) {
632     if (IsCopy && !Processed.insert(UseMI).second)
633       break;
634 
635     DenseMap<MachineInstr*, unsigned>::iterator DI = DistanceMap.find(UseMI);
636     if (DI != DistanceMap.end())
637       // Earlier in the same MBB.Reached via a back edge.
638       break;
639 
640     if (IsDstPhys) {
641       VirtRegPairs.push_back(NewReg);
642       break;
643     }
644     bool isNew = SrcRegMap.insert(std::make_pair(NewReg, Reg)).second;
645     if (!isNew)
646       assert(SrcRegMap[NewReg] == Reg && "Can't map to two src registers!");
647     VirtRegPairs.push_back(NewReg);
648     Reg = NewReg;
649   }
650 
651   if (!VirtRegPairs.empty()) {
652     unsigned ToReg = VirtRegPairs.back();
653     VirtRegPairs.pop_back();
654     while (!VirtRegPairs.empty()) {
655       unsigned FromReg = VirtRegPairs.back();
656       VirtRegPairs.pop_back();
657       bool isNew = DstRegMap.insert(std::make_pair(FromReg, ToReg)).second;
658       if (!isNew)
659         assert(DstRegMap[FromReg] == ToReg &&"Can't map to two dst registers!");
660       ToReg = FromReg;
661     }
662     bool isNew = DstRegMap.insert(std::make_pair(DstReg, ToReg)).second;
663     if (!isNew)
664       assert(DstRegMap[DstReg] == ToReg && "Can't map to two dst registers!");
665   }
666 }
667 
668 /// If the specified instruction is not yet processed, process it if it's a
669 /// copy. For a copy instruction, we find the physical registers the
670 /// source and destination registers might be mapped to. These are kept in
671 /// point-to maps used to determine future optimizations. e.g.
672 /// v1024 = mov r0
673 /// v1025 = mov r1
674 /// v1026 = add v1024, v1025
675 /// r1    = mov r1026
676 /// If 'add' is a two-address instruction, v1024, v1026 are both potentially
677 /// coalesced to r0 (from the input side). v1025 is mapped to r1. v1026 is
678 /// potentially joined with r1 on the output side. It's worthwhile to commute
679 /// 'add' to eliminate a copy.
680 void TwoAddressInstructionPass::processCopy(MachineInstr *MI) {
681   if (Processed.count(MI))
682     return;
683 
684   bool IsSrcPhys, IsDstPhys;
685   unsigned SrcReg, DstReg;
686   if (!isCopyToReg(*MI, TII, SrcReg, DstReg, IsSrcPhys, IsDstPhys))
687     return;
688 
689   if (IsDstPhys && !IsSrcPhys)
690     DstRegMap.insert(std::make_pair(SrcReg, DstReg));
691   else if (!IsDstPhys && IsSrcPhys) {
692     bool isNew = SrcRegMap.insert(std::make_pair(DstReg, SrcReg)).second;
693     if (!isNew)
694       assert(SrcRegMap[DstReg] == SrcReg &&
695              "Can't map to two src physical registers!");
696 
697     scanUses(DstReg);
698   }
699 
700   Processed.insert(MI);
701 }
702 
703 /// If there is one more local instruction that reads 'Reg' and it kills 'Reg,
704 /// consider moving the instruction below the kill instruction in order to
705 /// eliminate the need for the copy.
706 bool TwoAddressInstructionPass::
707 rescheduleMIBelowKill(MachineBasicBlock::iterator &mi,
708                       MachineBasicBlock::iterator &nmi,
709                       unsigned Reg) {
710   // Bail immediately if we don't have LV or LIS available. We use them to find
711   // kills efficiently.
712   if (!LV && !LIS)
713     return false;
714 
715   MachineInstr *MI = &*mi;
716   DenseMap<MachineInstr*, unsigned>::iterator DI = DistanceMap.find(MI);
717   if (DI == DistanceMap.end())
718     // Must be created from unfolded load. Don't waste time trying this.
719     return false;
720 
721   MachineInstr *KillMI = nullptr;
722   if (LIS) {
723     LiveInterval &LI = LIS->getInterval(Reg);
724     assert(LI.end() != LI.begin() &&
725            "Reg should not have empty live interval.");
726 
727     SlotIndex MBBEndIdx = LIS->getMBBEndIdx(MBB).getPrevSlot();
728     LiveInterval::const_iterator I = LI.find(MBBEndIdx);
729     if (I != LI.end() && I->start < MBBEndIdx)
730       return false;
731 
732     --I;
733     KillMI = LIS->getInstructionFromIndex(I->end);
734   } else {
735     KillMI = LV->getVarInfo(Reg).findKill(MBB);
736   }
737   if (!KillMI || MI == KillMI || KillMI->isCopy() || KillMI->isCopyLike())
738     // Don't mess with copies, they may be coalesced later.
739     return false;
740 
741   if (KillMI->hasUnmodeledSideEffects() || KillMI->isCall() ||
742       KillMI->isBranch() || KillMI->isTerminator())
743     // Don't move pass calls, etc.
744     return false;
745 
746   unsigned DstReg;
747   if (isTwoAddrUse(*KillMI, Reg, DstReg))
748     return false;
749 
750   bool SeenStore = true;
751   if (!MI->isSafeToMove(AA, SeenStore))
752     return false;
753 
754   if (TII->getInstrLatency(InstrItins, *MI) > 1)
755     // FIXME: Needs more sophisticated heuristics.
756     return false;
757 
758   SmallVector<unsigned, 2> Uses;
759   SmallVector<unsigned, 2> Kills;
760   SmallVector<unsigned, 2> Defs;
761   for (const MachineOperand &MO : MI->operands()) {
762     if (!MO.isReg())
763       continue;
764     Register MOReg = MO.getReg();
765     if (!MOReg)
766       continue;
767     if (MO.isDef())
768       Defs.push_back(MOReg);
769     else {
770       Uses.push_back(MOReg);
771       if (MOReg != Reg && (MO.isKill() ||
772                            (LIS && isPlainlyKilled(MI, MOReg, LIS))))
773         Kills.push_back(MOReg);
774     }
775   }
776 
777   // Move the copies connected to MI down as well.
778   MachineBasicBlock::iterator Begin = MI;
779   MachineBasicBlock::iterator AfterMI = std::next(Begin);
780   MachineBasicBlock::iterator End = AfterMI;
781   while (End != MBB->end()) {
782     End = skipDebugInstructionsForward(End, MBB->end());
783     if (End->isCopy() && regOverlapsSet(Defs, End->getOperand(1).getReg(), TRI))
784       Defs.push_back(End->getOperand(0).getReg());
785     else
786       break;
787     ++End;
788   }
789 
790   // Check if the reschedule will not break dependencies.
791   unsigned NumVisited = 0;
792   MachineBasicBlock::iterator KillPos = KillMI;
793   ++KillPos;
794   for (MachineInstr &OtherMI : make_range(End, KillPos)) {
795     // Debug instructions cannot be counted against the limit.
796     if (OtherMI.isDebugInstr())
797       continue;
798     if (NumVisited > 10)  // FIXME: Arbitrary limit to reduce compile time cost.
799       return false;
800     ++NumVisited;
801     if (OtherMI.hasUnmodeledSideEffects() || OtherMI.isCall() ||
802         OtherMI.isBranch() || OtherMI.isTerminator())
803       // Don't move pass calls, etc.
804       return false;
805     for (const MachineOperand &MO : OtherMI.operands()) {
806       if (!MO.isReg())
807         continue;
808       Register MOReg = MO.getReg();
809       if (!MOReg)
810         continue;
811       if (MO.isDef()) {
812         if (regOverlapsSet(Uses, MOReg, TRI))
813           // Physical register use would be clobbered.
814           return false;
815         if (!MO.isDead() && regOverlapsSet(Defs, MOReg, TRI))
816           // May clobber a physical register def.
817           // FIXME: This may be too conservative. It's ok if the instruction
818           // is sunken completely below the use.
819           return false;
820       } else {
821         if (regOverlapsSet(Defs, MOReg, TRI))
822           return false;
823         bool isKill =
824             MO.isKill() || (LIS && isPlainlyKilled(&OtherMI, MOReg, LIS));
825         if (MOReg != Reg && ((isKill && regOverlapsSet(Uses, MOReg, TRI)) ||
826                              regOverlapsSet(Kills, MOReg, TRI)))
827           // Don't want to extend other live ranges and update kills.
828           return false;
829         if (MOReg == Reg && !isKill)
830           // We can't schedule across a use of the register in question.
831           return false;
832         // Ensure that if this is register in question, its the kill we expect.
833         assert((MOReg != Reg || &OtherMI == KillMI) &&
834                "Found multiple kills of a register in a basic block");
835       }
836     }
837   }
838 
839   // Move debug info as well.
840   while (Begin != MBB->begin() && std::prev(Begin)->isDebugInstr())
841     --Begin;
842 
843   nmi = End;
844   MachineBasicBlock::iterator InsertPos = KillPos;
845   if (LIS) {
846     // We have to move the copies first so that the MBB is still well-formed
847     // when calling handleMove().
848     for (MachineBasicBlock::iterator MBBI = AfterMI; MBBI != End;) {
849       auto CopyMI = MBBI++;
850       MBB->splice(InsertPos, MBB, CopyMI);
851       LIS->handleMove(*CopyMI);
852       InsertPos = CopyMI;
853     }
854     End = std::next(MachineBasicBlock::iterator(MI));
855   }
856 
857   // Copies following MI may have been moved as well.
858   MBB->splice(InsertPos, MBB, Begin, End);
859   DistanceMap.erase(DI);
860 
861   // Update live variables
862   if (LIS) {
863     LIS->handleMove(*MI);
864   } else {
865     LV->removeVirtualRegisterKilled(Reg, *KillMI);
866     LV->addVirtualRegisterKilled(Reg, *MI);
867   }
868 
869   LLVM_DEBUG(dbgs() << "\trescheduled below kill: " << *KillMI);
870   return true;
871 }
872 
873 /// Return true if the re-scheduling will put the given instruction too close
874 /// to the defs of its register dependencies.
875 bool TwoAddressInstructionPass::isDefTooClose(unsigned Reg, unsigned Dist,
876                                               MachineInstr *MI) {
877   for (MachineInstr &DefMI : MRI->def_instructions(Reg)) {
878     if (DefMI.getParent() != MBB || DefMI.isCopy() || DefMI.isCopyLike())
879       continue;
880     if (&DefMI == MI)
881       return true; // MI is defining something KillMI uses
882     DenseMap<MachineInstr*, unsigned>::iterator DDI = DistanceMap.find(&DefMI);
883     if (DDI == DistanceMap.end())
884       return true;  // Below MI
885     unsigned DefDist = DDI->second;
886     assert(Dist > DefDist && "Visited def already?");
887     if (TII->getInstrLatency(InstrItins, DefMI) > (Dist - DefDist))
888       return true;
889   }
890   return false;
891 }
892 
893 /// If there is one more local instruction that reads 'Reg' and it kills 'Reg,
894 /// consider moving the kill instruction above the current two-address
895 /// instruction in order to eliminate the need for the copy.
896 bool TwoAddressInstructionPass::
897 rescheduleKillAboveMI(MachineBasicBlock::iterator &mi,
898                       MachineBasicBlock::iterator &nmi,
899                       unsigned Reg) {
900   // Bail immediately if we don't have LV or LIS available. We use them to find
901   // kills efficiently.
902   if (!LV && !LIS)
903     return false;
904 
905   MachineInstr *MI = &*mi;
906   DenseMap<MachineInstr*, unsigned>::iterator DI = DistanceMap.find(MI);
907   if (DI == DistanceMap.end())
908     // Must be created from unfolded load. Don't waste time trying this.
909     return false;
910 
911   MachineInstr *KillMI = nullptr;
912   if (LIS) {
913     LiveInterval &LI = LIS->getInterval(Reg);
914     assert(LI.end() != LI.begin() &&
915            "Reg should not have empty live interval.");
916 
917     SlotIndex MBBEndIdx = LIS->getMBBEndIdx(MBB).getPrevSlot();
918     LiveInterval::const_iterator I = LI.find(MBBEndIdx);
919     if (I != LI.end() && I->start < MBBEndIdx)
920       return false;
921 
922     --I;
923     KillMI = LIS->getInstructionFromIndex(I->end);
924   } else {
925     KillMI = LV->getVarInfo(Reg).findKill(MBB);
926   }
927   if (!KillMI || MI == KillMI || KillMI->isCopy() || KillMI->isCopyLike())
928     // Don't mess with copies, they may be coalesced later.
929     return false;
930 
931   unsigned DstReg;
932   if (isTwoAddrUse(*KillMI, Reg, DstReg))
933     return false;
934 
935   bool SeenStore = true;
936   if (!KillMI->isSafeToMove(AA, SeenStore))
937     return false;
938 
939   SmallSet<unsigned, 2> Uses;
940   SmallSet<unsigned, 2> Kills;
941   SmallSet<unsigned, 2> Defs;
942   SmallSet<unsigned, 2> LiveDefs;
943   for (const MachineOperand &MO : KillMI->operands()) {
944     if (!MO.isReg())
945       continue;
946     Register MOReg = MO.getReg();
947     if (MO.isUse()) {
948       if (!MOReg)
949         continue;
950       if (isDefTooClose(MOReg, DI->second, MI))
951         return false;
952       bool isKill = MO.isKill() || (LIS && isPlainlyKilled(KillMI, MOReg, LIS));
953       if (MOReg == Reg && !isKill)
954         return false;
955       Uses.insert(MOReg);
956       if (isKill && MOReg != Reg)
957         Kills.insert(MOReg);
958     } else if (Register::isPhysicalRegister(MOReg)) {
959       Defs.insert(MOReg);
960       if (!MO.isDead())
961         LiveDefs.insert(MOReg);
962     }
963   }
964 
965   // Check if the reschedule will not break depedencies.
966   unsigned NumVisited = 0;
967   for (MachineInstr &OtherMI :
968        make_range(mi, MachineBasicBlock::iterator(KillMI))) {
969     // Debug instructions cannot be counted against the limit.
970     if (OtherMI.isDebugInstr())
971       continue;
972     if (NumVisited > 10)  // FIXME: Arbitrary limit to reduce compile time cost.
973       return false;
974     ++NumVisited;
975     if (OtherMI.hasUnmodeledSideEffects() || OtherMI.isCall() ||
976         OtherMI.isBranch() || OtherMI.isTerminator())
977       // Don't move pass calls, etc.
978       return false;
979     SmallVector<unsigned, 2> OtherDefs;
980     for (const MachineOperand &MO : OtherMI.operands()) {
981       if (!MO.isReg())
982         continue;
983       Register MOReg = MO.getReg();
984       if (!MOReg)
985         continue;
986       if (MO.isUse()) {
987         if (Defs.count(MOReg))
988           // Moving KillMI can clobber the physical register if the def has
989           // not been seen.
990           return false;
991         if (Kills.count(MOReg))
992           // Don't want to extend other live ranges and update kills.
993           return false;
994         if (&OtherMI != MI && MOReg == Reg &&
995             !(MO.isKill() || (LIS && isPlainlyKilled(&OtherMI, MOReg, LIS))))
996           // We can't schedule across a use of the register in question.
997           return false;
998       } else {
999         OtherDefs.push_back(MOReg);
1000       }
1001     }
1002 
1003     for (unsigned i = 0, e = OtherDefs.size(); i != e; ++i) {
1004       unsigned MOReg = OtherDefs[i];
1005       if (Uses.count(MOReg))
1006         return false;
1007       if (Register::isPhysicalRegister(MOReg) && LiveDefs.count(MOReg))
1008         return false;
1009       // Physical register def is seen.
1010       Defs.erase(MOReg);
1011     }
1012   }
1013 
1014   // Move the old kill above MI, don't forget to move debug info as well.
1015   MachineBasicBlock::iterator InsertPos = mi;
1016   while (InsertPos != MBB->begin() && std::prev(InsertPos)->isDebugInstr())
1017     --InsertPos;
1018   MachineBasicBlock::iterator From = KillMI;
1019   MachineBasicBlock::iterator To = std::next(From);
1020   while (std::prev(From)->isDebugInstr())
1021     --From;
1022   MBB->splice(InsertPos, MBB, From, To);
1023 
1024   nmi = std::prev(InsertPos); // Backtrack so we process the moved instr.
1025   DistanceMap.erase(DI);
1026 
1027   // Update live variables
1028   if (LIS) {
1029     LIS->handleMove(*KillMI);
1030   } else {
1031     LV->removeVirtualRegisterKilled(Reg, *KillMI);
1032     LV->addVirtualRegisterKilled(Reg, *MI);
1033   }
1034 
1035   LLVM_DEBUG(dbgs() << "\trescheduled kill: " << *KillMI);
1036   return true;
1037 }
1038 
1039 /// Tries to commute the operand 'BaseOpIdx' and some other operand in the
1040 /// given machine instruction to improve opportunities for coalescing and
1041 /// elimination of a register to register copy.
1042 ///
1043 /// 'DstOpIdx' specifies the index of MI def operand.
1044 /// 'BaseOpKilled' specifies if the register associated with 'BaseOpIdx'
1045 /// operand is killed by the given instruction.
1046 /// The 'Dist' arguments provides the distance of MI from the start of the
1047 /// current basic block and it is used to determine if it is profitable
1048 /// to commute operands in the instruction.
1049 ///
1050 /// Returns true if the transformation happened. Otherwise, returns false.
1051 bool TwoAddressInstructionPass::tryInstructionCommute(MachineInstr *MI,
1052                                                       unsigned DstOpIdx,
1053                                                       unsigned BaseOpIdx,
1054                                                       bool BaseOpKilled,
1055                                                       unsigned Dist) {
1056   if (!MI->isCommutable())
1057     return false;
1058 
1059   bool MadeChange = false;
1060   Register DstOpReg = MI->getOperand(DstOpIdx).getReg();
1061   Register BaseOpReg = MI->getOperand(BaseOpIdx).getReg();
1062   unsigned OpsNum = MI->getDesc().getNumOperands();
1063   unsigned OtherOpIdx = MI->getDesc().getNumDefs();
1064   for (; OtherOpIdx < OpsNum; OtherOpIdx++) {
1065     // The call of findCommutedOpIndices below only checks if BaseOpIdx
1066     // and OtherOpIdx are commutable, it does not really search for
1067     // other commutable operands and does not change the values of passed
1068     // variables.
1069     if (OtherOpIdx == BaseOpIdx || !MI->getOperand(OtherOpIdx).isReg() ||
1070         !TII->findCommutedOpIndices(*MI, BaseOpIdx, OtherOpIdx))
1071       continue;
1072 
1073     Register OtherOpReg = MI->getOperand(OtherOpIdx).getReg();
1074     bool AggressiveCommute = false;
1075 
1076     // If OtherOp dies but BaseOp does not, swap the OtherOp and BaseOp
1077     // operands. This makes the live ranges of DstOp and OtherOp joinable.
1078     bool OtherOpKilled = isKilled(*MI, OtherOpReg, MRI, TII, LIS, false);
1079     bool DoCommute = !BaseOpKilled && OtherOpKilled;
1080 
1081     if (!DoCommute &&
1082         isProfitableToCommute(DstOpReg, BaseOpReg, OtherOpReg, MI, Dist)) {
1083       DoCommute = true;
1084       AggressiveCommute = true;
1085     }
1086 
1087     // If it's profitable to commute, try to do so.
1088     if (DoCommute && commuteInstruction(MI, DstOpIdx, BaseOpIdx, OtherOpIdx,
1089                                         Dist)) {
1090       MadeChange = true;
1091       ++NumCommuted;
1092       if (AggressiveCommute)
1093         ++NumAggrCommuted;
1094 
1095       // There might be more than two commutable operands, update BaseOp and
1096       // continue scanning.
1097       // FIXME: This assumes that the new instruction's operands are in the
1098       // same positions and were simply swapped.
1099       BaseOpReg = OtherOpReg;
1100       BaseOpKilled = OtherOpKilled;
1101       // Resamples OpsNum in case the number of operands was reduced. This
1102       // happens with X86.
1103       OpsNum = MI->getDesc().getNumOperands();
1104     }
1105   }
1106   return MadeChange;
1107 }
1108 
1109 /// For the case where an instruction has a single pair of tied register
1110 /// operands, attempt some transformations that may either eliminate the tied
1111 /// operands or improve the opportunities for coalescing away the register copy.
1112 /// Returns true if no copy needs to be inserted to untie mi's operands
1113 /// (either because they were untied, or because mi was rescheduled, and will
1114 /// be visited again later). If the shouldOnlyCommute flag is true, only
1115 /// instruction commutation is attempted.
1116 bool TwoAddressInstructionPass::
1117 tryInstructionTransform(MachineBasicBlock::iterator &mi,
1118                         MachineBasicBlock::iterator &nmi,
1119                         unsigned SrcIdx, unsigned DstIdx,
1120                         unsigned Dist, bool shouldOnlyCommute) {
1121   if (OptLevel == CodeGenOpt::None)
1122     return false;
1123 
1124   MachineInstr &MI = *mi;
1125   Register regA = MI.getOperand(DstIdx).getReg();
1126   Register regB = MI.getOperand(SrcIdx).getReg();
1127 
1128   assert(Register::isVirtualRegister(regB) &&
1129          "cannot make instruction into two-address form");
1130   bool regBKilled = isKilled(MI, regB, MRI, TII, LIS, true);
1131 
1132   if (Register::isVirtualRegister(regA))
1133     scanUses(regA);
1134 
1135   bool Commuted = tryInstructionCommute(&MI, DstIdx, SrcIdx, regBKilled, Dist);
1136 
1137   // If the instruction is convertible to 3 Addr, instead
1138   // of returning try 3 Addr transformation aggressively and
1139   // use this variable to check later. Because it might be better.
1140   // For example, we can just use `leal (%rsi,%rdi), %eax` and `ret`
1141   // instead of the following code.
1142   //   addl     %esi, %edi
1143   //   movl     %edi, %eax
1144   //   ret
1145   if (Commuted && !MI.isConvertibleTo3Addr())
1146     return false;
1147 
1148   if (shouldOnlyCommute)
1149     return false;
1150 
1151   // If there is one more use of regB later in the same MBB, consider
1152   // re-schedule this MI below it.
1153   if (!Commuted && EnableRescheduling && rescheduleMIBelowKill(mi, nmi, regB)) {
1154     ++NumReSchedDowns;
1155     return true;
1156   }
1157 
1158   // If we commuted, regB may have changed so we should re-sample it to avoid
1159   // confusing the three address conversion below.
1160   if (Commuted) {
1161     regB = MI.getOperand(SrcIdx).getReg();
1162     regBKilled = isKilled(MI, regB, MRI, TII, LIS, true);
1163   }
1164 
1165   if (MI.isConvertibleTo3Addr()) {
1166     // This instruction is potentially convertible to a true
1167     // three-address instruction.  Check if it is profitable.
1168     if (!regBKilled || isProfitableToConv3Addr(regA, regB)) {
1169       // Try to convert it.
1170       if (convertInstTo3Addr(mi, nmi, regA, regB, Dist)) {
1171         ++NumConvertedTo3Addr;
1172         return true; // Done with this instruction.
1173       }
1174     }
1175   }
1176 
1177   // Return if it is commuted but 3 addr conversion is failed.
1178   if (Commuted)
1179     return false;
1180 
1181   // If there is one more use of regB later in the same MBB, consider
1182   // re-schedule it before this MI if it's legal.
1183   if (EnableRescheduling && rescheduleKillAboveMI(mi, nmi, regB)) {
1184     ++NumReSchedUps;
1185     return true;
1186   }
1187 
1188   // If this is an instruction with a load folded into it, try unfolding
1189   // the load, e.g. avoid this:
1190   //   movq %rdx, %rcx
1191   //   addq (%rax), %rcx
1192   // in favor of this:
1193   //   movq (%rax), %rcx
1194   //   addq %rdx, %rcx
1195   // because it's preferable to schedule a load than a register copy.
1196   if (MI.mayLoad() && !regBKilled) {
1197     // Determine if a load can be unfolded.
1198     unsigned LoadRegIndex;
1199     unsigned NewOpc =
1200       TII->getOpcodeAfterMemoryUnfold(MI.getOpcode(),
1201                                       /*UnfoldLoad=*/true,
1202                                       /*UnfoldStore=*/false,
1203                                       &LoadRegIndex);
1204     if (NewOpc != 0) {
1205       const MCInstrDesc &UnfoldMCID = TII->get(NewOpc);
1206       if (UnfoldMCID.getNumDefs() == 1) {
1207         // Unfold the load.
1208         LLVM_DEBUG(dbgs() << "2addr:   UNFOLDING: " << MI);
1209         const TargetRegisterClass *RC =
1210           TRI->getAllocatableClass(
1211             TII->getRegClass(UnfoldMCID, LoadRegIndex, TRI, *MF));
1212         Register Reg = MRI->createVirtualRegister(RC);
1213         SmallVector<MachineInstr *, 2> NewMIs;
1214         if (!TII->unfoldMemoryOperand(*MF, MI, Reg,
1215                                       /*UnfoldLoad=*/true,
1216                                       /*UnfoldStore=*/false, NewMIs)) {
1217           LLVM_DEBUG(dbgs() << "2addr: ABANDONING UNFOLD\n");
1218           return false;
1219         }
1220         assert(NewMIs.size() == 2 &&
1221                "Unfolded a load into multiple instructions!");
1222         // The load was previously folded, so this is the only use.
1223         NewMIs[1]->addRegisterKilled(Reg, TRI);
1224 
1225         // Tentatively insert the instructions into the block so that they
1226         // look "normal" to the transformation logic.
1227         MBB->insert(mi, NewMIs[0]);
1228         MBB->insert(mi, NewMIs[1]);
1229 
1230         LLVM_DEBUG(dbgs() << "2addr:    NEW LOAD: " << *NewMIs[0]
1231                           << "2addr:    NEW INST: " << *NewMIs[1]);
1232 
1233         // Transform the instruction, now that it no longer has a load.
1234         unsigned NewDstIdx = NewMIs[1]->findRegisterDefOperandIdx(regA);
1235         unsigned NewSrcIdx = NewMIs[1]->findRegisterUseOperandIdx(regB);
1236         MachineBasicBlock::iterator NewMI = NewMIs[1];
1237         bool TransformResult =
1238           tryInstructionTransform(NewMI, mi, NewSrcIdx, NewDstIdx, Dist, true);
1239         (void)TransformResult;
1240         assert(!TransformResult &&
1241                "tryInstructionTransform() should return false.");
1242         if (NewMIs[1]->getOperand(NewSrcIdx).isKill()) {
1243           // Success, or at least we made an improvement. Keep the unfolded
1244           // instructions and discard the original.
1245           if (LV) {
1246             for (unsigned i = 0, e = MI.getNumOperands(); i != e; ++i) {
1247               MachineOperand &MO = MI.getOperand(i);
1248               if (MO.isReg() && Register::isVirtualRegister(MO.getReg())) {
1249                 if (MO.isUse()) {
1250                   if (MO.isKill()) {
1251                     if (NewMIs[0]->killsRegister(MO.getReg()))
1252                       LV->replaceKillInstruction(MO.getReg(), MI, *NewMIs[0]);
1253                     else {
1254                       assert(NewMIs[1]->killsRegister(MO.getReg()) &&
1255                              "Kill missing after load unfold!");
1256                       LV->replaceKillInstruction(MO.getReg(), MI, *NewMIs[1]);
1257                     }
1258                   }
1259                 } else if (LV->removeVirtualRegisterDead(MO.getReg(), MI)) {
1260                   if (NewMIs[1]->registerDefIsDead(MO.getReg()))
1261                     LV->addVirtualRegisterDead(MO.getReg(), *NewMIs[1]);
1262                   else {
1263                     assert(NewMIs[0]->registerDefIsDead(MO.getReg()) &&
1264                            "Dead flag missing after load unfold!");
1265                     LV->addVirtualRegisterDead(MO.getReg(), *NewMIs[0]);
1266                   }
1267                 }
1268               }
1269             }
1270             LV->addVirtualRegisterKilled(Reg, *NewMIs[1]);
1271           }
1272 
1273           SmallVector<Register, 4> OrigRegs;
1274           if (LIS) {
1275             for (const MachineOperand &MO : MI.operands()) {
1276               if (MO.isReg())
1277                 OrigRegs.push_back(MO.getReg());
1278             }
1279           }
1280 
1281           MI.eraseFromParent();
1282 
1283           // Update LiveIntervals.
1284           if (LIS) {
1285             MachineBasicBlock::iterator Begin(NewMIs[0]);
1286             MachineBasicBlock::iterator End(NewMIs[1]);
1287             LIS->repairIntervalsInRange(MBB, Begin, End, OrigRegs);
1288           }
1289 
1290           mi = NewMIs[1];
1291         } else {
1292           // Transforming didn't eliminate the tie and didn't lead to an
1293           // improvement. Clean up the unfolded instructions and keep the
1294           // original.
1295           LLVM_DEBUG(dbgs() << "2addr: ABANDONING UNFOLD\n");
1296           NewMIs[0]->eraseFromParent();
1297           NewMIs[1]->eraseFromParent();
1298         }
1299       }
1300     }
1301   }
1302 
1303   return false;
1304 }
1305 
1306 // Collect tied operands of MI that need to be handled.
1307 // Rewrite trivial cases immediately.
1308 // Return true if any tied operands where found, including the trivial ones.
1309 bool TwoAddressInstructionPass::
1310 collectTiedOperands(MachineInstr *MI, TiedOperandMap &TiedOperands) {
1311   const MCInstrDesc &MCID = MI->getDesc();
1312   bool AnyOps = false;
1313   unsigned NumOps = MI->getNumOperands();
1314 
1315   for (unsigned SrcIdx = 0; SrcIdx < NumOps; ++SrcIdx) {
1316     unsigned DstIdx = 0;
1317     if (!MI->isRegTiedToDefOperand(SrcIdx, &DstIdx))
1318       continue;
1319     AnyOps = true;
1320     MachineOperand &SrcMO = MI->getOperand(SrcIdx);
1321     MachineOperand &DstMO = MI->getOperand(DstIdx);
1322     Register SrcReg = SrcMO.getReg();
1323     Register DstReg = DstMO.getReg();
1324     // Tied constraint already satisfied?
1325     if (SrcReg == DstReg)
1326       continue;
1327 
1328     assert(SrcReg && SrcMO.isUse() && "two address instruction invalid");
1329 
1330     // Deal with undef uses immediately - simply rewrite the src operand.
1331     if (SrcMO.isUndef() && !DstMO.getSubReg()) {
1332       // Constrain the DstReg register class if required.
1333       if (Register::isVirtualRegister(DstReg))
1334         if (const TargetRegisterClass *RC = TII->getRegClass(MCID, SrcIdx,
1335                                                              TRI, *MF))
1336           MRI->constrainRegClass(DstReg, RC);
1337       SrcMO.setReg(DstReg);
1338       SrcMO.setSubReg(0);
1339       LLVM_DEBUG(dbgs() << "\t\trewrite undef:\t" << *MI);
1340       continue;
1341     }
1342     TiedOperands[SrcReg].push_back(std::make_pair(SrcIdx, DstIdx));
1343   }
1344   return AnyOps;
1345 }
1346 
1347 // Process a list of tied MI operands that all use the same source register.
1348 // The tied pairs are of the form (SrcIdx, DstIdx).
1349 void
1350 TwoAddressInstructionPass::processTiedPairs(MachineInstr *MI,
1351                                             TiedPairList &TiedPairs,
1352                                             unsigned &Dist) {
1353   bool IsEarlyClobber = false;
1354   for (unsigned tpi = 0, tpe = TiedPairs.size(); tpi != tpe; ++tpi) {
1355     const MachineOperand &DstMO = MI->getOperand(TiedPairs[tpi].second);
1356     IsEarlyClobber |= DstMO.isEarlyClobber();
1357   }
1358 
1359   bool RemovedKillFlag = false;
1360   bool AllUsesCopied = true;
1361   unsigned LastCopiedReg = 0;
1362   SlotIndex LastCopyIdx;
1363   unsigned RegB = 0;
1364   unsigned SubRegB = 0;
1365   for (unsigned tpi = 0, tpe = TiedPairs.size(); tpi != tpe; ++tpi) {
1366     unsigned SrcIdx = TiedPairs[tpi].first;
1367     unsigned DstIdx = TiedPairs[tpi].second;
1368 
1369     const MachineOperand &DstMO = MI->getOperand(DstIdx);
1370     Register RegA = DstMO.getReg();
1371 
1372     // Grab RegB from the instruction because it may have changed if the
1373     // instruction was commuted.
1374     RegB = MI->getOperand(SrcIdx).getReg();
1375     SubRegB = MI->getOperand(SrcIdx).getSubReg();
1376 
1377     if (RegA == RegB) {
1378       // The register is tied to multiple destinations (or else we would
1379       // not have continued this far), but this use of the register
1380       // already matches the tied destination.  Leave it.
1381       AllUsesCopied = false;
1382       continue;
1383     }
1384     LastCopiedReg = RegA;
1385 
1386     assert(Register::isVirtualRegister(RegB) &&
1387            "cannot make instruction into two-address form");
1388 
1389 #ifndef NDEBUG
1390     // First, verify that we don't have a use of "a" in the instruction
1391     // (a = b + a for example) because our transformation will not
1392     // work. This should never occur because we are in SSA form.
1393     for (unsigned i = 0; i != MI->getNumOperands(); ++i)
1394       assert(i == DstIdx ||
1395              !MI->getOperand(i).isReg() ||
1396              MI->getOperand(i).getReg() != RegA);
1397 #endif
1398 
1399     // Emit a copy.
1400     MachineInstrBuilder MIB = BuildMI(*MI->getParent(), MI, MI->getDebugLoc(),
1401                                       TII->get(TargetOpcode::COPY), RegA);
1402     // If this operand is folding a truncation, the truncation now moves to the
1403     // copy so that the register classes remain valid for the operands.
1404     MIB.addReg(RegB, 0, SubRegB);
1405     const TargetRegisterClass *RC = MRI->getRegClass(RegB);
1406     if (SubRegB) {
1407       if (Register::isVirtualRegister(RegA)) {
1408         assert(TRI->getMatchingSuperRegClass(RC, MRI->getRegClass(RegA),
1409                                              SubRegB) &&
1410                "tied subregister must be a truncation");
1411         // The superreg class will not be used to constrain the subreg class.
1412         RC = nullptr;
1413       } else {
1414         assert(TRI->getMatchingSuperReg(RegA, SubRegB, MRI->getRegClass(RegB))
1415                && "tied subregister must be a truncation");
1416       }
1417     }
1418 
1419     // Update DistanceMap.
1420     MachineBasicBlock::iterator PrevMI = MI;
1421     --PrevMI;
1422     DistanceMap.insert(std::make_pair(&*PrevMI, Dist));
1423     DistanceMap[MI] = ++Dist;
1424 
1425     if (LIS) {
1426       LastCopyIdx = LIS->InsertMachineInstrInMaps(*PrevMI).getRegSlot();
1427 
1428       if (Register::isVirtualRegister(RegA)) {
1429         LiveInterval &LI = LIS->getInterval(RegA);
1430         VNInfo *VNI = LI.getNextValue(LastCopyIdx, LIS->getVNInfoAllocator());
1431         SlotIndex endIdx =
1432             LIS->getInstructionIndex(*MI).getRegSlot(IsEarlyClobber);
1433         LI.addSegment(LiveInterval::Segment(LastCopyIdx, endIdx, VNI));
1434       }
1435     }
1436 
1437     LLVM_DEBUG(dbgs() << "\t\tprepend:\t" << *MIB);
1438 
1439     MachineOperand &MO = MI->getOperand(SrcIdx);
1440     assert(MO.isReg() && MO.getReg() == RegB && MO.isUse() &&
1441            "inconsistent operand info for 2-reg pass");
1442     if (MO.isKill()) {
1443       MO.setIsKill(false);
1444       RemovedKillFlag = true;
1445     }
1446 
1447     // Make sure regA is a legal regclass for the SrcIdx operand.
1448     if (Register::isVirtualRegister(RegA) && Register::isVirtualRegister(RegB))
1449       MRI->constrainRegClass(RegA, RC);
1450     MO.setReg(RegA);
1451     // The getMatchingSuper asserts guarantee that the register class projected
1452     // by SubRegB is compatible with RegA with no subregister. So regardless of
1453     // whether the dest oper writes a subreg, the source oper should not.
1454     MO.setSubReg(0);
1455 
1456     // Propagate SrcRegMap.
1457     SrcRegMap[RegA] = RegB;
1458   }
1459 
1460   if (AllUsesCopied) {
1461     bool ReplacedAllUntiedUses = true;
1462     if (!IsEarlyClobber) {
1463       // Replace other (un-tied) uses of regB with LastCopiedReg.
1464       for (MachineOperand &MO : MI->operands()) {
1465         if (MO.isReg() && MO.getReg() == RegB && MO.isUse()) {
1466           if (MO.getSubReg() == SubRegB) {
1467             if (MO.isKill()) {
1468               MO.setIsKill(false);
1469               RemovedKillFlag = true;
1470             }
1471             MO.setReg(LastCopiedReg);
1472             MO.setSubReg(0);
1473           } else {
1474             ReplacedAllUntiedUses = false;
1475           }
1476         }
1477       }
1478     }
1479 
1480     // Update live variables for regB.
1481     if (RemovedKillFlag && ReplacedAllUntiedUses &&
1482         LV && LV->getVarInfo(RegB).removeKill(*MI)) {
1483       MachineBasicBlock::iterator PrevMI = MI;
1484       --PrevMI;
1485       LV->addVirtualRegisterKilled(RegB, *PrevMI);
1486     }
1487 
1488     // Update LiveIntervals.
1489     if (LIS) {
1490       LiveInterval &LI = LIS->getInterval(RegB);
1491       SlotIndex MIIdx = LIS->getInstructionIndex(*MI);
1492       LiveInterval::const_iterator I = LI.find(MIIdx);
1493       assert(I != LI.end() && "RegB must be live-in to use.");
1494 
1495       SlotIndex UseIdx = MIIdx.getRegSlot(IsEarlyClobber);
1496       if (I->end == UseIdx)
1497         LI.removeSegment(LastCopyIdx, UseIdx);
1498     }
1499   } else if (RemovedKillFlag) {
1500     // Some tied uses of regB matched their destination registers, so
1501     // regB is still used in this instruction, but a kill flag was
1502     // removed from a different tied use of regB, so now we need to add
1503     // a kill flag to one of the remaining uses of regB.
1504     for (MachineOperand &MO : MI->operands()) {
1505       if (MO.isReg() && MO.getReg() == RegB && MO.isUse()) {
1506         MO.setIsKill(true);
1507         break;
1508       }
1509     }
1510   }
1511 }
1512 
1513 /// Reduce two-address instructions to two operands.
1514 bool TwoAddressInstructionPass::runOnMachineFunction(MachineFunction &Func) {
1515   MF = &Func;
1516   const TargetMachine &TM = MF->getTarget();
1517   MRI = &MF->getRegInfo();
1518   TII = MF->getSubtarget().getInstrInfo();
1519   TRI = MF->getSubtarget().getRegisterInfo();
1520   InstrItins = MF->getSubtarget().getInstrItineraryData();
1521   LV = getAnalysisIfAvailable<LiveVariables>();
1522   LIS = getAnalysisIfAvailable<LiveIntervals>();
1523   if (auto *AAPass = getAnalysisIfAvailable<AAResultsWrapperPass>())
1524     AA = &AAPass->getAAResults();
1525   else
1526     AA = nullptr;
1527   OptLevel = TM.getOptLevel();
1528   // Disable optimizations if requested. We cannot skip the whole pass as some
1529   // fixups are necessary for correctness.
1530   if (skipFunction(Func.getFunction()))
1531     OptLevel = CodeGenOpt::None;
1532 
1533   bool MadeChange = false;
1534 
1535   LLVM_DEBUG(dbgs() << "********** REWRITING TWO-ADDR INSTRS **********\n");
1536   LLVM_DEBUG(dbgs() << "********** Function: " << MF->getName() << '\n');
1537 
1538   // This pass takes the function out of SSA form.
1539   MRI->leaveSSA();
1540 
1541   // This pass will rewrite the tied-def to meet the RegConstraint.
1542   MF->getProperties()
1543       .set(MachineFunctionProperties::Property::TiedOpsRewritten);
1544 
1545   TiedOperandMap TiedOperands;
1546   for (MachineFunction::iterator MBBI = MF->begin(), MBBE = MF->end();
1547        MBBI != MBBE; ++MBBI) {
1548     MBB = &*MBBI;
1549     unsigned Dist = 0;
1550     DistanceMap.clear();
1551     SrcRegMap.clear();
1552     DstRegMap.clear();
1553     Processed.clear();
1554     for (MachineBasicBlock::iterator mi = MBB->begin(), me = MBB->end();
1555          mi != me; ) {
1556       MachineBasicBlock::iterator nmi = std::next(mi);
1557       // Skip debug instructions.
1558       if (mi->isDebugInstr()) {
1559         mi = nmi;
1560         continue;
1561       }
1562 
1563       // Expand REG_SEQUENCE instructions. This will position mi at the first
1564       // expanded instruction.
1565       if (mi->isRegSequence())
1566         eliminateRegSequence(mi);
1567 
1568       DistanceMap.insert(std::make_pair(&*mi, ++Dist));
1569 
1570       processCopy(&*mi);
1571 
1572       // First scan through all the tied register uses in this instruction
1573       // and record a list of pairs of tied operands for each register.
1574       if (!collectTiedOperands(&*mi, TiedOperands)) {
1575         mi = nmi;
1576         continue;
1577       }
1578 
1579       ++NumTwoAddressInstrs;
1580       MadeChange = true;
1581       LLVM_DEBUG(dbgs() << '\t' << *mi);
1582 
1583       // If the instruction has a single pair of tied operands, try some
1584       // transformations that may either eliminate the tied operands or
1585       // improve the opportunities for coalescing away the register copy.
1586       if (TiedOperands.size() == 1) {
1587         SmallVectorImpl<std::pair<unsigned, unsigned>> &TiedPairs
1588           = TiedOperands.begin()->second;
1589         if (TiedPairs.size() == 1) {
1590           unsigned SrcIdx = TiedPairs[0].first;
1591           unsigned DstIdx = TiedPairs[0].second;
1592           Register SrcReg = mi->getOperand(SrcIdx).getReg();
1593           Register DstReg = mi->getOperand(DstIdx).getReg();
1594           if (SrcReg != DstReg &&
1595               tryInstructionTransform(mi, nmi, SrcIdx, DstIdx, Dist, false)) {
1596             // The tied operands have been eliminated or shifted further down
1597             // the block to ease elimination. Continue processing with 'nmi'.
1598             TiedOperands.clear();
1599             mi = nmi;
1600             continue;
1601           }
1602         }
1603       }
1604 
1605       // Now iterate over the information collected above.
1606       for (auto &TO : TiedOperands) {
1607         processTiedPairs(&*mi, TO.second, Dist);
1608         LLVM_DEBUG(dbgs() << "\t\trewrite to:\t" << *mi);
1609       }
1610 
1611       // Rewrite INSERT_SUBREG as COPY now that we no longer need SSA form.
1612       if (mi->isInsertSubreg()) {
1613         // From %reg = INSERT_SUBREG %reg, %subreg, subidx
1614         // To   %reg:subidx = COPY %subreg
1615         unsigned SubIdx = mi->getOperand(3).getImm();
1616         mi->RemoveOperand(3);
1617         assert(mi->getOperand(0).getSubReg() == 0 && "Unexpected subreg idx");
1618         mi->getOperand(0).setSubReg(SubIdx);
1619         mi->getOperand(0).setIsUndef(mi->getOperand(1).isUndef());
1620         mi->RemoveOperand(1);
1621         mi->setDesc(TII->get(TargetOpcode::COPY));
1622         LLVM_DEBUG(dbgs() << "\t\tconvert to:\t" << *mi);
1623       }
1624 
1625       // Clear TiedOperands here instead of at the top of the loop
1626       // since most instructions do not have tied operands.
1627       TiedOperands.clear();
1628       mi = nmi;
1629     }
1630   }
1631 
1632   if (LIS)
1633     MF->verify(this, "After two-address instruction pass");
1634 
1635   return MadeChange;
1636 }
1637 
1638 /// Eliminate a REG_SEQUENCE instruction as part of the de-ssa process.
1639 ///
1640 /// The instruction is turned into a sequence of sub-register copies:
1641 ///
1642 ///   %dst = REG_SEQUENCE %v1, ssub0, %v2, ssub1
1643 ///
1644 /// Becomes:
1645 ///
1646 ///   undef %dst:ssub0 = COPY %v1
1647 ///   %dst:ssub1 = COPY %v2
1648 void TwoAddressInstructionPass::
1649 eliminateRegSequence(MachineBasicBlock::iterator &MBBI) {
1650   MachineInstr &MI = *MBBI;
1651   Register DstReg = MI.getOperand(0).getReg();
1652   if (MI.getOperand(0).getSubReg() || Register::isPhysicalRegister(DstReg) ||
1653       !(MI.getNumOperands() & 1)) {
1654     LLVM_DEBUG(dbgs() << "Illegal REG_SEQUENCE instruction:" << MI);
1655     llvm_unreachable(nullptr);
1656   }
1657 
1658   SmallVector<Register, 4> OrigRegs;
1659   if (LIS) {
1660     OrigRegs.push_back(MI.getOperand(0).getReg());
1661     for (unsigned i = 1, e = MI.getNumOperands(); i < e; i += 2)
1662       OrigRegs.push_back(MI.getOperand(i).getReg());
1663   }
1664 
1665   bool DefEmitted = false;
1666   for (unsigned i = 1, e = MI.getNumOperands(); i < e; i += 2) {
1667     MachineOperand &UseMO = MI.getOperand(i);
1668     Register SrcReg = UseMO.getReg();
1669     unsigned SubIdx = MI.getOperand(i+1).getImm();
1670     // Nothing needs to be inserted for undef operands.
1671     if (UseMO.isUndef())
1672       continue;
1673 
1674     // Defer any kill flag to the last operand using SrcReg. Otherwise, we
1675     // might insert a COPY that uses SrcReg after is was killed.
1676     bool isKill = UseMO.isKill();
1677     if (isKill)
1678       for (unsigned j = i + 2; j < e; j += 2)
1679         if (MI.getOperand(j).getReg() == SrcReg) {
1680           MI.getOperand(j).setIsKill();
1681           UseMO.setIsKill(false);
1682           isKill = false;
1683           break;
1684         }
1685 
1686     // Insert the sub-register copy.
1687     MachineInstr *CopyMI = BuildMI(*MI.getParent(), MI, MI.getDebugLoc(),
1688                                    TII->get(TargetOpcode::COPY))
1689                                .addReg(DstReg, RegState::Define, SubIdx)
1690                                .add(UseMO);
1691 
1692     // The first def needs an undef flag because there is no live register
1693     // before it.
1694     if (!DefEmitted) {
1695       CopyMI->getOperand(0).setIsUndef(true);
1696       // Return an iterator pointing to the first inserted instr.
1697       MBBI = CopyMI;
1698     }
1699     DefEmitted = true;
1700 
1701     // Update LiveVariables' kill info.
1702     if (LV && isKill && !Register::isPhysicalRegister(SrcReg))
1703       LV->replaceKillInstruction(SrcReg, MI, *CopyMI);
1704 
1705     LLVM_DEBUG(dbgs() << "Inserted: " << *CopyMI);
1706   }
1707 
1708   MachineBasicBlock::iterator EndMBBI =
1709       std::next(MachineBasicBlock::iterator(MI));
1710 
1711   if (!DefEmitted) {
1712     LLVM_DEBUG(dbgs() << "Turned: " << MI << " into an IMPLICIT_DEF");
1713     MI.setDesc(TII->get(TargetOpcode::IMPLICIT_DEF));
1714     for (int j = MI.getNumOperands() - 1, ee = 0; j > ee; --j)
1715       MI.RemoveOperand(j);
1716   } else {
1717     LLVM_DEBUG(dbgs() << "Eliminated: " << MI);
1718     MI.eraseFromParent();
1719   }
1720 
1721   // Udpate LiveIntervals.
1722   if (LIS)
1723     LIS->repairIntervalsInRange(MBB, MBBI, EndMBBI, OrigRegs);
1724 }
1725