Fix a ton of comment typos found by codespell. Patch by
[oota-llvm.git] / lib / CodeGen / TwoAddressInstructionPass.cpp
index 048b6698a733da192ed5283e784b0f8f2b7e4bdc..52ea87231ccd3c95379ef249862c10f36f0adbd1 100644 (file)
@@ -105,12 +105,12 @@ namespace {
                             MachineFunction::iterator &mbbi,
                             unsigned RegB, unsigned RegC, unsigned Dist);
 
-    bool isProfitableToConv3Addr(unsigned RegA);
+    bool isProfitableToConv3Addr(unsigned RegA, unsigned RegB);
 
     bool ConvertInstTo3Addr(MachineBasicBlock::iterator &mi,
                             MachineBasicBlock::iterator &nmi,
                             MachineFunction::iterator &mbbi,
-                            unsigned RegB, unsigned Dist);
+                            unsigned RegA, unsigned RegB, unsigned Dist);
 
     typedef std::pair<std::pair<unsigned, bool>, MachineInstr*> NewKill;
     bool canUpdateDeletedKills(SmallVector<unsigned, 4> &Kills,
@@ -124,7 +124,11 @@ namespace {
                                  MachineBasicBlock::iterator &nmi,
                                  MachineFunction::iterator &mbbi,
                                  unsigned SrcIdx, unsigned DstIdx,
-                                 unsigned Dist);
+                                 unsigned Dist,
+                                 SmallPtrSet<MachineInstr*, 8> &Processed);
+
+    void ScanUses(unsigned DstReg, MachineBasicBlock *MBB,
+                  SmallPtrSet<MachineInstr*, 8> &Processed);
 
     void ProcessCopy(MachineInstr *MI, MachineBasicBlock *MBB,
                      SmallPtrSet<MachineInstr*, 8> &Processed);
@@ -138,7 +142,9 @@ namespace {
 
   public:
     static char ID; // Pass identification, replacement for typeid
-    TwoAddressInstructionPass() : MachineFunctionPass(&ID) {}
+    TwoAddressInstructionPass() : MachineFunctionPass(ID) {
+      initializeTwoAddressInstructionPassPass(*PassRegistry::getPassRegistry());
+    }
 
     virtual void getAnalysisUsage(AnalysisUsage &AU) const {
       AU.setPreservesCFG();
@@ -146,10 +152,7 @@ namespace {
       AU.addPreserved<LiveVariables>();
       AU.addPreservedID(MachineLoopInfoID);
       AU.addPreservedID(MachineDominatorsID);
-      if (StrongPHIElim)
-        AU.addPreservedID(StrongPHIEliminationID);
-      else
-        AU.addPreservedID(PHIEliminationID);
+      AU.addPreservedID(PHIEliminationID);
       MachineFunctionPass::getAnalysisUsage(AU);
     }
 
@@ -159,10 +162,13 @@ namespace {
 }
 
 char TwoAddressInstructionPass::ID = 0;
-static RegisterPass<TwoAddressInstructionPass>
-X("twoaddressinstruction", "Two-Address instruction pass");
+INITIALIZE_PASS_BEGIN(TwoAddressInstructionPass, "twoaddressinstruction",
+                "Two-Address instruction pass", false, false)
+INITIALIZE_AG_DEPENDENCY(AliasAnalysis)
+INITIALIZE_PASS_END(TwoAddressInstructionPass, "twoaddressinstruction",
+                "Two-Address instruction pass", false, false)
 
-const PassInfo *const llvm::TwoAddressInstructionPassID = &X;
+char &llvm::TwoAddressInstructionPassID = TwoAddressInstructionPass::ID;
 
 /// Sink3AddrInstruction - A two-address instruction has been converted to a
 /// three-address instruction to avoid clobbering a register. Try to sink it
@@ -380,26 +386,18 @@ static bool isCopyToReg(MachineInstr &MI, const TargetInstrInfo *TII,
                         bool &IsSrcPhys, bool &IsDstPhys) {
   SrcReg = 0;
   DstReg = 0;
-  unsigned SrcSubIdx, DstSubIdx;
-  if (!TII->isMoveInstr(MI, SrcReg, DstReg, SrcSubIdx, DstSubIdx)) {
-    if (MI.isExtractSubreg()) {
-      DstReg = MI.getOperand(0).getReg();
-      SrcReg = MI.getOperand(1).getReg();
-    } else if (MI.isInsertSubreg()) {
-      DstReg = MI.getOperand(0).getReg();
-      SrcReg = MI.getOperand(2).getReg();
-    } else if (MI.isSubregToReg()) {
-      DstReg = MI.getOperand(0).getReg();
-      SrcReg = MI.getOperand(2).getReg();
-    }
-  }
+  if (MI.isCopy()) {
+    DstReg = MI.getOperand(0).getReg();
+    SrcReg = MI.getOperand(1).getReg();
+  } else if (MI.isInsertSubreg() || MI.isSubregToReg()) {
+    DstReg = MI.getOperand(0).getReg();
+    SrcReg = MI.getOperand(2).getReg();
+  } else
+    return false;
 
-  if (DstReg) {
-    IsSrcPhys = TargetRegisterInfo::isPhysicalRegister(SrcReg);
-    IsDstPhys = TargetRegisterInfo::isPhysicalRegister(DstReg);
-    return true;
-  }
-  return false;
+  IsSrcPhys = TargetRegisterInfo::isPhysicalRegister(SrcReg);
+  IsDstPhys = TargetRegisterInfo::isPhysicalRegister(DstReg);
+  return true;
 }
 
 /// isKilled - Test if the given register value, which is used by the given
@@ -556,8 +554,9 @@ TwoAddressInstructionPass::isProfitableToCommute(unsigned regB, unsigned regC,
   unsigned FromRegC = getMappedReg(regC, SrcRegMap);
   unsigned ToRegB = getMappedReg(regB, DstRegMap);
   unsigned ToRegC = getMappedReg(regC, DstRegMap);
-  if (!regsAreCompatible(FromRegB, ToRegB, TRI) &&
-      (regsAreCompatible(FromRegB, ToRegC, TRI) ||
+  if ((FromRegB && ToRegB && !regsAreCompatible(FromRegB, ToRegB, TRI)) &&
+      ((!FromRegC && !ToRegC) ||
+       regsAreCompatible(FromRegB, ToRegC, TRI) ||
        regsAreCompatible(FromRegC, ToRegB, TRI)))
     return true;
 
@@ -620,16 +619,18 @@ TwoAddressInstructionPass::CommuteInstruction(MachineBasicBlock::iterator &mi,
 /// isProfitableToConv3Addr - Return true if it is profitable to convert the
 /// given 2-address instruction to a 3-address one.
 bool
-TwoAddressInstructionPass::isProfitableToConv3Addr(unsigned RegA{
+TwoAddressInstructionPass::isProfitableToConv3Addr(unsigned RegA,unsigned RegB){
   // Look for situations like this:
   // %reg1024<def> = MOV r1
   // %reg1025<def> = MOV r0
   // %reg1026<def> = ADD %reg1024, %reg1025
   // r2            = MOV %reg1026
   // Turn ADD into a 3-address instruction to avoid a copy.
-  unsigned FromRegA = getMappedReg(RegA, SrcRegMap);
+  unsigned FromRegB = getMappedReg(RegB, SrcRegMap);
+  if (!FromRegB)
+    return false;
   unsigned ToRegA = getMappedReg(RegA, DstRegMap);
-  return (FromRegA && ToRegA && !regsAreCompatible(FromRegA, ToRegA, TRI));
+  return (ToRegA && !regsAreCompatible(FromRegB, ToRegA, TRI));
 }
 
 /// ConvertInstTo3Addr - Convert the specified two-address instruction into a
@@ -638,7 +639,8 @@ bool
 TwoAddressInstructionPass::ConvertInstTo3Addr(MachineBasicBlock::iterator &mi,
                                               MachineBasicBlock::iterator &nmi,
                                               MachineFunction::iterator &mbbi,
-                                              unsigned RegB, unsigned Dist) {
+                                              unsigned RegA, unsigned RegB,
+                                              unsigned Dist) {
   MachineInstr *NewMI = TII->convertToThreeAddress(mbbi, mi, LV);
   if (NewMI) {
     DEBUG(dbgs() << "2addr: CONVERTING 2-ADDR: " << *mi);
@@ -658,12 +660,64 @@ TwoAddressInstructionPass::ConvertInstTo3Addr(MachineBasicBlock::iterator &mi,
       mi = NewMI;
       nmi = llvm::next(mi);
     }
+
+    // Update source and destination register maps.
+    SrcRegMap.erase(RegA);
+    DstRegMap.erase(RegB);
     return true;
   }
 
   return false;
 }
 
+/// ScanUses - Scan forward recursively for only uses, update maps if the use
+/// is a copy or a two-address instruction.
+void
+TwoAddressInstructionPass::ScanUses(unsigned DstReg, MachineBasicBlock *MBB,
+                                    SmallPtrSet<MachineInstr*, 8> &Processed) {
+  SmallVector<unsigned, 4> VirtRegPairs;
+  bool IsDstPhys;
+  bool IsCopy = false;
+  unsigned NewReg = 0;
+  unsigned Reg = DstReg;
+  while (MachineInstr *UseMI = findOnlyInterestingUse(Reg, MBB, MRI, TII,IsCopy,
+                                                      NewReg, IsDstPhys)) {
+    if (IsCopy && !Processed.insert(UseMI))
+      break;
+
+    DenseMap<MachineInstr*, unsigned>::iterator DI = DistanceMap.find(UseMI);
+    if (DI != DistanceMap.end())
+      // Earlier in the same MBB.Reached via a back edge.
+      break;
+
+    if (IsDstPhys) {
+      VirtRegPairs.push_back(NewReg);
+      break;
+    }
+    bool isNew = SrcRegMap.insert(std::make_pair(NewReg, Reg)).second;
+    if (!isNew)
+      assert(SrcRegMap[NewReg] == Reg && "Can't map to two src registers!");
+    VirtRegPairs.push_back(NewReg);
+    Reg = NewReg;
+  }
+
+  if (!VirtRegPairs.empty()) {
+    unsigned ToReg = VirtRegPairs.back();
+    VirtRegPairs.pop_back();
+    while (!VirtRegPairs.empty()) {
+      unsigned FromReg = VirtRegPairs.back();
+      VirtRegPairs.pop_back();
+      bool isNew = DstRegMap.insert(std::make_pair(FromReg, ToReg)).second;
+      if (!isNew)
+        assert(DstRegMap[FromReg] == ToReg &&"Can't map to two dst registers!");
+      ToReg = FromReg;
+    }
+    bool isNew = DstRegMap.insert(std::make_pair(DstReg, ToReg)).second;
+    if (!isNew)
+      assert(DstRegMap[DstReg] == ToReg && "Can't map to two dst registers!");
+  }
+}
+
 /// ProcessCopy - If the specified instruction is not yet processed, process it
 /// if it's a copy. For a copy instruction, we find the physical registers the
 /// source and destination registers might be mapped to. These are kept in
@@ -695,49 +749,11 @@ void TwoAddressInstructionPass::ProcessCopy(MachineInstr *MI,
       assert(SrcRegMap[DstReg] == SrcReg &&
              "Can't map to two src physical registers!");
 
-    SmallVector<unsigned, 4> VirtRegPairs;
-    bool IsCopy = false;
-    unsigned NewReg = 0;
-    while (MachineInstr *UseMI = findOnlyInterestingUse(DstReg, MBB, MRI,TII,
-                                                   IsCopy, NewReg, IsDstPhys)) {
-      if (IsCopy) {
-        if (!Processed.insert(UseMI))
-          break;
-      }
-
-      DenseMap<MachineInstr*, unsigned>::iterator DI = DistanceMap.find(UseMI);
-      if (DI != DistanceMap.end())
-        // Earlier in the same MBB.Reached via a back edge.
-        break;
-
-      if (IsDstPhys) {
-        VirtRegPairs.push_back(NewReg);
-        break;
-      }
-      bool isNew = SrcRegMap.insert(std::make_pair(NewReg, DstReg)).second;
-      if (!isNew)
-        assert(SrcRegMap[NewReg] == DstReg &&
-               "Can't map to two src physical registers!");
-      VirtRegPairs.push_back(NewReg);
-      DstReg = NewReg;
-    }
-
-    if (!VirtRegPairs.empty()) {
-      unsigned ToReg = VirtRegPairs.back();
-      VirtRegPairs.pop_back();
-      while (!VirtRegPairs.empty()) {
-        unsigned FromReg = VirtRegPairs.back();
-        VirtRegPairs.pop_back();
-        bool isNew = DstRegMap.insert(std::make_pair(FromReg, ToReg)).second;
-        if (!isNew)
-          assert(DstRegMap[FromReg] == ToReg &&
-                 "Can't map to two dst physical registers!");
-        ToReg = FromReg;
-      }
-    }
+    ScanUses(DstReg, MBB, Processed);
   }
 
   Processed.insert(MI);
+  return;
 }
 
 /// isSafeToDelete - If the specified instruction does not produce any side
@@ -748,7 +764,7 @@ static bool isSafeToDelete(MachineInstr *MI,
   const TargetInstrDesc &TID = MI->getDesc();
   if (TID.mayStore() || TID.isCall())
     return false;
-  if (TID.isTerminator() || TID.hasUnmodeledSideEffects())
+  if (TID.isTerminator() || MI->hasUnmodeledSideEffects())
     return false;
 
   for (unsigned i = 0, e = MI->getNumOperands(); i != e; ++i) {
@@ -836,7 +852,8 @@ bool TwoAddressInstructionPass::
 TryInstructionTransform(MachineBasicBlock::iterator &mi,
                         MachineBasicBlock::iterator &nmi,
                         MachineFunction::iterator &mbbi,
-                        unsigned SrcIdx, unsigned DstIdx, unsigned Dist) {
+                        unsigned SrcIdx, unsigned DstIdx, unsigned Dist,
+                        SmallPtrSet<MachineInstr*, 8> &Processed) {
   const TargetInstrDesc &TID = mi->getDesc();
   unsigned regA = mi->getOperand(DstIdx).getReg();
   unsigned regB = mi->getOperand(SrcIdx).getReg();
@@ -887,17 +904,122 @@ TryInstructionTransform(MachineBasicBlock::iterator &mi,
     return false;
   }
 
+  if (TargetRegisterInfo::isVirtualRegister(regA))
+    ScanUses(regA, &*mbbi, Processed);
+
   if (TID.isConvertibleTo3Addr()) {
     // This instruction is potentially convertible to a true
     // three-address instruction.  Check if it is profitable.
-    if (!regBKilled || isProfitableToConv3Addr(regA)) {
+    if (!regBKilled || isProfitableToConv3Addr(regA, regB)) {
       // Try to convert it.
-      if (ConvertInstTo3Addr(mi, nmi, mbbi, regB, Dist)) {
+      if (ConvertInstTo3Addr(mi, nmi, mbbi, regA, regB, Dist)) {
         ++NumConvertedTo3Addr;
         return true; // Done with this instruction.
       }
     }
   }
+
+  // If this is an instruction with a load folded into it, try unfolding
+  // the load, e.g. avoid this:
+  //   movq %rdx, %rcx
+  //   addq (%rax), %rcx
+  // in favor of this:
+  //   movq (%rax), %rcx
+  //   addq %rdx, %rcx
+  // because it's preferable to schedule a load than a register copy.
+  if (TID.mayLoad() && !regBKilled) {
+    // Determine if a load can be unfolded.
+    unsigned LoadRegIndex;
+    unsigned NewOpc =
+      TII->getOpcodeAfterMemoryUnfold(mi->getOpcode(),
+                                      /*UnfoldLoad=*/true,
+                                      /*UnfoldStore=*/false,
+                                      &LoadRegIndex);
+    if (NewOpc != 0) {
+      const TargetInstrDesc &UnfoldTID = TII->get(NewOpc);
+      if (UnfoldTID.getNumDefs() == 1) {
+        MachineFunction &MF = *mbbi->getParent();
+
+        // Unfold the load.
+        DEBUG(dbgs() << "2addr:   UNFOLDING: " << *mi);
+        const TargetRegisterClass *RC =
+          UnfoldTID.OpInfo[LoadRegIndex].getRegClass(TRI);
+        unsigned Reg = MRI->createVirtualRegister(RC);
+        SmallVector<MachineInstr *, 2> NewMIs;
+        if (!TII->unfoldMemoryOperand(MF, mi, Reg,
+                                      /*UnfoldLoad=*/true,/*UnfoldStore=*/false,
+                                      NewMIs)) {
+          DEBUG(dbgs() << "2addr: ABANDONING UNFOLD\n");
+          return false;
+        }
+        assert(NewMIs.size() == 2 &&
+               "Unfolded a load into multiple instructions!");
+        // The load was previously folded, so this is the only use.
+        NewMIs[1]->addRegisterKilled(Reg, TRI);
+
+        // Tentatively insert the instructions into the block so that they
+        // look "normal" to the transformation logic.
+        mbbi->insert(mi, NewMIs[0]);
+        mbbi->insert(mi, NewMIs[1]);
+
+        DEBUG(dbgs() << "2addr:    NEW LOAD: " << *NewMIs[0]
+                     << "2addr:    NEW INST: " << *NewMIs[1]);
+
+        // Transform the instruction, now that it no longer has a load.
+        unsigned NewDstIdx = NewMIs[1]->findRegisterDefOperandIdx(regA);
+        unsigned NewSrcIdx = NewMIs[1]->findRegisterUseOperandIdx(regB);
+        MachineBasicBlock::iterator NewMI = NewMIs[1];
+        bool TransformSuccess =
+          TryInstructionTransform(NewMI, mi, mbbi,
+                                  NewSrcIdx, NewDstIdx, Dist, Processed);
+        if (TransformSuccess ||
+            NewMIs[1]->getOperand(NewSrcIdx).isKill()) {
+          // Success, or at least we made an improvement. Keep the unfolded
+          // instructions and discard the original.
+          if (LV) {
+            for (unsigned i = 0, e = mi->getNumOperands(); i != e; ++i) {
+              MachineOperand &MO = mi->getOperand(i);
+              if (MO.isReg() && 
+                  TargetRegisterInfo::isVirtualRegister(MO.getReg())) {
+                if (MO.isUse()) {
+                  if (MO.isKill()) {
+                    if (NewMIs[0]->killsRegister(MO.getReg()))
+                      LV->replaceKillInstruction(MO.getReg(), mi, NewMIs[0]);
+                    else {
+                      assert(NewMIs[1]->killsRegister(MO.getReg()) &&
+                             "Kill missing after load unfold!");
+                      LV->replaceKillInstruction(MO.getReg(), mi, NewMIs[1]);
+                    }
+                  }
+                } else if (LV->removeVirtualRegisterDead(MO.getReg(), mi)) {
+                  if (NewMIs[1]->registerDefIsDead(MO.getReg()))
+                    LV->addVirtualRegisterDead(MO.getReg(), NewMIs[1]);
+                  else {
+                    assert(NewMIs[0]->registerDefIsDead(MO.getReg()) &&
+                           "Dead flag missing after load unfold!");
+                    LV->addVirtualRegisterDead(MO.getReg(), NewMIs[0]);
+                  }
+                }
+              }
+            }
+            LV->addVirtualRegisterKilled(Reg, NewMIs[1]);
+          }
+          mi->eraseFromParent();
+          mi = NewMIs[1];
+          if (TransformSuccess)
+            return true;
+        } else {
+          // Transforming didn't eliminate the tie and didn't lead to an
+          // improvement. Clean up the unfolded instructions and keep the
+          // original.
+          DEBUG(dbgs() << "2addr: ABANDONING UNFOLD\n");
+          NewMIs[0]->eraseFromParent();
+          NewMIs[1]->eraseFromParent();
+        }
+      }
+    }
+  }
+
   return false;
 }
 
@@ -919,8 +1041,7 @@ bool TwoAddressInstructionPass::runOnMachineFunction(MachineFunction &MF) {
         << MF.getFunction()->getName() << '\n');
 
   // ReMatRegs - Keep track of the registers whose def's are remat'ed.
-  BitVector ReMatRegs;
-  ReMatRegs.resize(MRI->getLastVirtReg()+1);
+  BitVector ReMatRegs(MRI->getNumVirtRegs());
 
   typedef DenseMap<unsigned, SmallVector<std::pair<unsigned, unsigned>, 4> >
     TiedOperandMap;
@@ -999,7 +1120,8 @@ bool TwoAddressInstructionPass::runOnMachineFunction(MachineFunction &MF) {
               mi->getOperand(DstIdx).getReg())
             break; // Done with this instruction.
 
-          if (TryInstructionTransform(mi, nmi, mbbi, SrcIdx, DstIdx, Dist))
+          if (TryInstructionTransform(mi, nmi, mbbi, SrcIdx, DstIdx, Dist,
+                                      Processed))
             break; // The tied operands have been eliminated.
         }
 
@@ -1049,13 +1171,11 @@ bool TwoAddressInstructionPass::runOnMachineFunction(MachineFunction &MF) {
             DEBUG(dbgs() << "2addr: REMATTING : " << *DefMI << "\n");
             unsigned regASubIdx = mi->getOperand(DstIdx).getSubReg();
             TII->reMaterialize(*mbbi, mi, regA, regASubIdx, DefMI, *TRI);
-            ReMatRegs.set(regB);
+            ReMatRegs.set(TargetRegisterInfo::virtReg2Index(regB));
             ++NumReMats;
           } else {
-            bool Emitted = TII->copyRegToReg(*mbbi, mi, regA, regB, rc, rc,
-                                             mi->getDebugLoc());
-            (void)Emitted;
-            assert(Emitted && "Unable to issue a copy instruction!\n");
+            BuildMI(*mbbi, mi, mi->getDebugLoc(), TII->get(TargetOpcode::COPY),
+                    regA).addReg(regB);
           }
 
           MachineBasicBlock::iterator prevMI = prior(mi);
@@ -1116,6 +1236,19 @@ bool TwoAddressInstructionPass::runOnMachineFunction(MachineFunction &MF) {
         DEBUG(dbgs() << "\t\trewrite to:\t" << *mi);
       }
 
+      // Rewrite INSERT_SUBREG as COPY now that we no longer need SSA form.
+      if (mi->isInsertSubreg()) {
+        // From %reg = INSERT_SUBREG %reg, %subreg, subidx
+        // To   %reg:subidx = COPY %subreg
+        unsigned SubIdx = mi->getOperand(3).getImm();
+        mi->RemoveOperand(3);
+        assert(mi->getOperand(0).getSubReg() == 0 && "Unexpected subreg idx");
+        mi->getOperand(0).setSubReg(SubIdx);
+        mi->RemoveOperand(1);
+        mi->setDesc(TII->get(TargetOpcode::COPY));
+        DEBUG(dbgs() << "\t\tconvert to:\t" << *mi);
+      }
+
       // Clear TiedOperands here instead of at the top of the loop
       // since most instructions do not have tied operands.
       TiedOperands.clear();
@@ -1124,13 +1257,12 @@ bool TwoAddressInstructionPass::runOnMachineFunction(MachineFunction &MF) {
   }
 
   // Some remat'ed instructions are dead.
-  int VReg = ReMatRegs.find_first();
-  while (VReg != -1) {
+  for (int i = ReMatRegs.find_first(); i != -1; i = ReMatRegs.find_next(i)) {
+    unsigned VReg = TargetRegisterInfo::index2VirtReg(i);
     if (MRI->use_nodbg_empty(VReg)) {
       MachineInstr *DefMI = MRI->getVRegDef(VReg);
       DefMI->eraseFromParent();
     }
-    VReg = ReMatRegs.find_next(VReg);
   }
 
   // Eliminate REG_SEQUENCE instructions. Their whole purpose was to preseve
@@ -1176,7 +1308,7 @@ TwoAddressInstructionPass::CoalesceExtSubRegs(SmallVector<unsigned,4> &Srcs,
     if (SrcDefMI->getParent() != DstDefMI->getParent())
       continue;
 
-    // If there are no other uses than extract_subreg which feed into
+    // If there are no other uses than copies which feed into
     // the reg_sequence, then we might be able to coalesce them.
     bool CanCoalesce = true;
     SmallVector<unsigned, 4> SrcSubIndices, DstSubIndices;
@@ -1184,13 +1316,11 @@ TwoAddressInstructionPass::CoalesceExtSubRegs(SmallVector<unsigned,4> &Srcs,
            UI = MRI->use_nodbg_begin(SrcReg),
            UE = MRI->use_nodbg_end(); UI != UE; ++UI) {
       MachineInstr *UseMI = &*UI;
-      if (!UseMI->isExtractSubreg() ||
-          UseMI->getOperand(0).getReg() != DstReg ||
-          UseMI->getOperand(1).getSubReg() != 0) {
+      if (!UseMI->isCopy() || UseMI->getOperand(0).getReg() != DstReg) {
         CanCoalesce = false;
         break;
       }
-      SrcSubIndices.push_back(UseMI->getOperand(2).getImm());
+      SrcSubIndices.push_back(UseMI->getOperand(1).getSubReg());
       DstSubIndices.push_back(UseMI->getOperand(0).getSubReg());
     }
 
@@ -1225,9 +1355,9 @@ TwoAddressInstructionPass::CoalesceExtSubRegs(SmallVector<unsigned,4> &Srcs,
            UI = MRI->use_nodbg_begin(SrcReg),
            UE = MRI->use_nodbg_end(); UI != UE; ++UI) {
       MachineInstr *UseMI = &*UI;
-      assert(UseMI->isExtractSubreg());
+      assert(UseMI->isCopy());
       unsigned DstSubIdx = UseMI->getOperand(0).getSubReg();
-      unsigned SrcSubIdx = UseMI->getOperand(2).getImm();
+      unsigned SrcSubIdx = UseMI->getOperand(1).getSubReg();
       assert(DstSubIdx != 0 && "missing subreg from RegSequence elimination");
       if ((NewDstSubIdx == 0 &&
            TRI->composeSubRegIndices(NewSrcSubIdx, DstSubIdx) != SrcSubIdx) ||
@@ -1242,27 +1372,12 @@ TwoAddressInstructionPass::CoalesceExtSubRegs(SmallVector<unsigned,4> &Srcs,
     if (!CanCoalesce)
       continue;
 
-    // Insert a copy or an extract to replace the original extracts.
-    MachineBasicBlock::iterator InsertLoc = SomeMI;
-    if (NewSrcSubIdx) {
-      // Insert an extract subreg.
-      BuildMI(*SomeMI->getParent(), InsertLoc, SomeMI->getDebugLoc(),
-              TII->get(TargetOpcode::EXTRACT_SUBREG), DstReg)
-        .addReg(SrcReg).addImm(NewSrcSubIdx);
-    } else if (NewDstSubIdx) {
-      // Do a subreg insertion.
-      BuildMI(*SomeMI->getParent(), InsertLoc, SomeMI->getDebugLoc(),
-              TII->get(TargetOpcode::INSERT_SUBREG), DstReg)
-        .addReg(DstReg).addReg(SrcReg).addImm(NewDstSubIdx);
-    } else {
-      // Insert a copy.
-      bool Emitted =
-        TII->copyRegToReg(*SomeMI->getParent(), InsertLoc, DstReg, SrcReg,
-                          MRI->getRegClass(DstReg), MRI->getRegClass(SrcReg),
-                          SomeMI->getDebugLoc());
-      (void)Emitted;
-    }
-    MachineBasicBlock::iterator CopyMI = prior(InsertLoc);
+    // Insert a copy to replace the original.
+    MachineInstr *CopyMI = BuildMI(*SomeMI->getParent(), SomeMI,
+                                   SomeMI->getDebugLoc(),
+                                   TII->get(TargetOpcode::COPY))
+      .addReg(DstReg, RegState::Define, NewDstSubIdx)
+      .addReg(SrcReg, 0, NewSrcSubIdx);
 
     // Remove all the old extract instructions.
     for (MachineRegisterInfo::use_nodbg_iterator
@@ -1272,11 +1387,10 @@ TwoAddressInstructionPass::CoalesceExtSubRegs(SmallVector<unsigned,4> &Srcs,
       ++UI;
       if (UseMI == CopyMI)
         continue;
-      assert(UseMI->isExtractSubreg());
+      assert(UseMI->isCopy());
       // Move any kills to the new copy or extract instruction.
       if (UseMI->getOperand(1).isKill()) {
-        MachineOperand *KillMO = CopyMI->findRegisterUseOperand(SrcReg);
-        KillMO->setIsKill();
+        CopyMI->getOperand(1).setIsKill();
         if (LV)
           // Update live variables
           LV->replaceKillInstruction(SrcReg, UseMI, &*CopyMI);
@@ -1324,6 +1438,7 @@ bool TwoAddressInstructionPass::EliminateRegSequences() {
     SmallSet<unsigned, 4> Seen;
     for (unsigned i = 1, e = MI->getNumOperands(); i < e; i += 2) {
       unsigned SrcReg = MI->getOperand(i).getReg();
+      unsigned SubIdx = MI->getOperand(i+1).getImm();
       if (MI->getOperand(i).getSubReg() ||
           TargetRegisterInfo::isPhysicalRegister(SrcReg)) {
         DEBUG(dbgs() << "Illegal REG_SEQUENCE instruction:" << *MI);
@@ -1337,15 +1452,15 @@ bool TwoAddressInstructionPass::EliminateRegSequences() {
       }
       IsImpDef = false;
 
-      // Remember EXTRACT_SUBREG sources. These might be candidate for
-      // coalescing.
-      if (DefMI->isExtractSubreg())
+      // Remember COPY sources. These might be candidate for coalescing.
+      if (DefMI->isCopy() && DefMI->getOperand(1).getSubReg())
         RealSrcs.push_back(DefMI->getOperand(1).getReg());
 
-      if (!Seen.insert(SrcReg) ||
-          MI->getParent() != DefMI->getParent() ||
-          !MI->getOperand(i).isKill() ||
-          HasOtherRegSequenceUses(SrcReg, MI, MRI)) {
+      bool isKill = MI->getOperand(i).isKill();
+      if (!Seen.insert(SrcReg) || MI->getParent() != DefMI->getParent() ||
+          !isKill || HasOtherRegSequenceUses(SrcReg, MI, MRI) ||
+          !TRI->getMatchingSuperRegClass(MRI->getRegClass(DstReg),
+                                         MRI->getRegClass(SrcReg), SubIdx)) {
         // REG_SEQUENCE cannot have duplicated operands, add a copy.
         // Also add an copy if the source is live-in the block. We don't want
         // to end up with a partial-redef of a livein, e.g.
@@ -1360,29 +1475,32 @@ bool TwoAddressInstructionPass::EliminateRegSequences() {
         //
         // If the REG_SEQUENCE doesn't kill its source, keeping live variables
         // correctly up to date becomes very difficult. Insert a copy.
-        //
-        const TargetRegisterClass *RC = MRI->getRegClass(SrcReg);
-        unsigned NewReg = MRI->createVirtualRegister(RC);
+
+        // Defer any kill flag to the last operand using SrcReg. Otherwise, we
+        // might insert a COPY that uses SrcReg after is was killed.
+        if (isKill)
+          for (unsigned j = i + 2; j < e; j += 2)
+            if (MI->getOperand(j).getReg() == SrcReg) {
+              MI->getOperand(j).setIsKill();
+              isKill = false;
+              break;
+            }
+
         MachineBasicBlock::iterator InsertLoc = MI;
-        bool Emitted =
-          TII->copyRegToReg(*MI->getParent(), InsertLoc, NewReg, SrcReg, RC, RC,
-                            MI->getDebugLoc());
-        (void)Emitted;
-        assert(Emitted && "Unable to issue a copy instruction!\n");
-        MI->getOperand(i).setReg(NewReg);
-        if (MI->getOperand(i).isKill()) {
-          MachineBasicBlock::iterator CopyMI = prior(InsertLoc);
-          MachineOperand *KillMO = CopyMI->findRegisterUseOperand(SrcReg);
-          KillMO->setIsKill();
-          if (LV)
-            // Update live variables
-            LV->replaceKillInstruction(SrcReg, MI, &*CopyMI);
-        }
+        MachineInstr *CopyMI = BuildMI(*MI->getParent(), InsertLoc,
+                                MI->getDebugLoc(), TII->get(TargetOpcode::COPY))
+            .addReg(DstReg, RegState::Define, SubIdx)
+            .addReg(SrcReg, getKillRegState(isKill));
+        MI->getOperand(i).setReg(0);
+        if (LV && isKill)
+          LV->replaceKillInstruction(SrcReg, MI, CopyMI);
+        DEBUG(dbgs() << "Inserted: " << *CopyMI);
       }
     }
 
     for (unsigned i = 1, e = MI->getNumOperands(); i < e; i += 2) {
       unsigned SrcReg = MI->getOperand(i).getReg();
+      if (!SrcReg) continue;
       unsigned SubIdx = MI->getOperand(i+1).getImm();
       UpdateRegSequenceSrcs(SrcReg, DstReg, SubIdx, MRI, *TRI);
     }
@@ -1397,8 +1515,11 @@ bool TwoAddressInstructionPass::EliminateRegSequences() {
       MI->eraseFromParent();
     }
 
-    // Try coalescing some EXTRACT_SUBREG instructions.
-    CoalesceExtSubRegs(RealSrcs, DstReg);
+    // Try coalescing some EXTRACT_SUBREG instructions. This can create
+    // INSERT_SUBREG instructions that must have <undef> flags added by
+    // LiveIntervalAnalysis, so only run it when LiveVariables is available.
+    if (LV)
+      CoalesceExtSubRegs(RealSrcs, DstReg);
   }
 
   RegSequences.clear();