Get rid of static constructors for pass registration. Instead, every pass exposes...
[oota-llvm.git] / lib / CodeGen / TwoAddressInstructionPass.cpp
index 31552f89780f7fa8c45963935c3672cf29a9d930..2030f174fe3f4f08a4227d254a15bb3f2c0d29b9 100644 (file)
@@ -33,6 +33,7 @@
 #include "llvm/CodeGen/LiveVariables.h"
 #include "llvm/CodeGen/MachineFunctionPass.h"
 #include "llvm/CodeGen/MachineInstr.h"
+#include "llvm/CodeGen/MachineInstrBuilder.h"
 #include "llvm/CodeGen/MachineRegisterInfo.h"
 #include "llvm/Analysis/AliasAnalysis.h"
 #include "llvm/Target/TargetRegisterInfo.h"
@@ -40,6 +41,7 @@
 #include "llvm/Target/TargetMachine.h"
 #include "llvm/Target/TargetOptions.h"
 #include "llvm/Support/Debug.h"
+#include "llvm/Support/ErrorHandling.h"
 #include "llvm/ADT/BitVector.h"
 #include "llvm/ADT/DenseMap.h"
 #include "llvm/ADT/SmallSet.h"
@@ -77,6 +79,10 @@ namespace {
     // registers from virtual registers. e.g. r1 = move v1024.
     DenseMap<unsigned, unsigned> DstRegMap;
 
+    /// RegSequences - Keep track the list of REG_SEQUENCE instructions seen
+    /// during the initial walk of the machine function.
+    SmallVector<MachineInstr*, 16> RegSequences;
+
     bool Sink3AddrInstruction(MachineBasicBlock *MBB, MachineInstr *MI,
                               unsigned Reg,
                               MachineBasicBlock::iterator OldPos);
@@ -123,9 +129,18 @@ namespace {
     void ProcessCopy(MachineInstr *MI, MachineBasicBlock *MBB,
                      SmallPtrSet<MachineInstr*, 8> &Processed);
 
+    void CoalesceExtSubRegs(SmallVector<unsigned,4> &Srcs, unsigned DstReg);
+
+    /// EliminateRegSequences - Eliminate REG_SEQUENCE instructions as part
+    /// of the de-ssa process. This replaces sources of REG_SEQUENCE as
+    /// sub-register references of the register defined by REG_SEQUENCE.
+    bool EliminateRegSequences();
+
   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 +161,13 @@ 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
@@ -160,7 +178,7 @@ bool TwoAddressInstructionPass::Sink3AddrInstruction(MachineBasicBlock *MBB,
                                            MachineBasicBlock::iterator OldPos) {
   // Check if it's safe to move this instruction.
   bool SeenStore = true; // Be conservative.
-  if (!MI->isSafeToMove(TII, SeenStore, AA))
+  if (!MI->isSafeToMove(TII, AA, SeenStore))
     return false;
 
   unsigned DefReg = 0;
@@ -188,8 +206,9 @@ bool TwoAddressInstructionPass::Sink3AddrInstruction(MachineBasicBlock *MBB,
 
   // Find the instruction that kills SavedReg.
   MachineInstr *KillMI = NULL;
-  for (MachineRegisterInfo::use_iterator UI = MRI->use_begin(SavedReg),
-         UE = MRI->use_end(); UI != UE; ++UI) {
+  for (MachineRegisterInfo::use_nodbg_iterator
+         UI = MRI->use_nodbg_begin(SavedReg),
+         UE = MRI->use_nodbg_end(); UI != UE; ++UI) {
     MachineOperand &UseMO = UI.getOperand();
     if (!UseMO.isKill())
       continue;
@@ -213,6 +232,9 @@ bool TwoAddressInstructionPass::Sink3AddrInstruction(MachineBasicBlock *MBB,
   unsigned NumVisited = 0;
   for (MachineBasicBlock::iterator I = llvm::next(OldPos); I != KillPos; ++I) {
     MachineInstr *OtherMI = I;
+    // DBG_VALUE cannot be counted against the limit.
+    if (OtherMI->isDebugValue())
+      continue;
     if (NumVisited > 30)  // FIXME: Arbitrary limit to reduce compile time cost.
       return false;
     ++NumVisited;
@@ -277,8 +299,8 @@ TwoAddressInstructionPass::isProfitableToReMat(unsigned Reg,
                                          MachineInstr *MI, MachineInstr *DefMI,
                                          MachineBasicBlock *MBB, unsigned Loc) {
   bool OtherUse = false;
-  for (MachineRegisterInfo::use_iterator UI = MRI->use_begin(Reg),
-         UE = MRI->use_end(); UI != UE; ++UI) {
+  for (MachineRegisterInfo::use_nodbg_iterator UI = MRI->use_nodbg_begin(Reg),
+         UE = MRI->use_nodbg_end(); UI != UE; ++UI) {
     MachineOperand &UseMO = UI.getOperand();
     MachineInstr *UseMI = UseMO.getParent();
     MachineBasicBlock *UseMBB = UseMI->getParent();
@@ -363,26 +385,18 @@ 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
@@ -451,13 +465,10 @@ MachineInstr *findOnlyInterestingUse(unsigned Reg, MachineBasicBlock *MBB,
                                      const TargetInstrInfo *TII,
                                      bool &IsCopy,
                                      unsigned &DstReg, bool &IsDstPhys) {
-  MachineRegisterInfo::use_nodbg_iterator UI = MRI->use_nodbg_begin(Reg);
-  if (UI == MRI->use_nodbg_end())
-    return 0;
-  MachineInstr &UseMI = *UI;
-  if (++UI != MRI->use_nodbg_end())
-    // More than one use.
+  if (!MRI->hasOneNonDBGUse(Reg))
+    // None or more than one use.
     return 0;
+  MachineInstr &UseMI = *MRI->use_nodbg_begin(Reg);
   if (UseMI.getParent() != MBB)
     return 0;
   unsigned SrcReg;
@@ -767,7 +778,7 @@ canUpdateDeletedKills(SmallVector<unsigned, 4> &Kills,
     if (!LastKill)
       return false;
 
-    bool isModRef = LastKill->modifiesRegister(Kill);
+    bool isModRef = LastKill->definesRegister(Kill);
     NewKills.push_back(std::make_pair(std::make_pair(Kill, isModRef),
                                       LastKill));
   }
@@ -884,6 +895,108 @@ TryInstructionTransform(MachineBasicBlock::iterator &mi,
       }
     }
   }
+
+  // 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);
+        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() && MO.getReg() != 0 &&
+                  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;
 }
 
@@ -927,6 +1040,11 @@ bool TwoAddressInstructionPass::runOnMachineFunction(MachineFunction &MF) {
         mi = nmi;
         continue;
       }
+
+      // Remember REG_SEQUENCE instructions, we'll deal with them later.
+      if (mi->isRegSequence())
+        RegSequences.push_back(&*mi);
+
       const TargetInstrDesc &TID = mi->getDesc();
       bool FirstTied = true;
 
@@ -1025,17 +1143,16 @@ bool TwoAddressInstructionPass::runOnMachineFunction(MachineFunction &MF) {
           // copying it.
           if (DefMI &&
               DefMI->getDesc().isAsCheapAsAMove() &&
-              DefMI->isSafeToReMat(TII, regB, AA) &&
+              DefMI->isSafeToReMat(TII, AA, regB) &&
               isProfitableToReMat(regB, rc, mi, DefMI, mbbi, Dist)){
             DEBUG(dbgs() << "2addr: REMATTING : " << *DefMI << "\n");
             unsigned regASubIdx = mi->getOperand(DstIdx).getSubReg();
-            TII->reMaterialize(*mbbi, mi, regA, regASubIdx, DefMI, TRI);
+            TII->reMaterialize(*mbbi, mi, regA, regASubIdx, DefMI, *TRI);
             ReMatRegs.set(regB);
             ++NumReMats;
           } else {
-            bool Emitted = TII->copyRegToReg(*mbbi, mi, regA, regB, rc, rc);
-            (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);
@@ -1085,12 +1202,30 @@ bool TwoAddressInstructionPass::runOnMachineFunction(MachineFunction &MF) {
             }
           }
         }
-          
+
+        // Schedule the source copy / remat inserted to form two-address
+        // instruction. FIXME: Does it matter the distance map may not be
+        // accurate after it's scheduled?
+        TII->scheduleTwoAddrSource(prior(mi), mi, *TRI);
+
         MadeChange = true;
 
         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();
@@ -1101,12 +1236,268 @@ bool TwoAddressInstructionPass::runOnMachineFunction(MachineFunction &MF) {
   // Some remat'ed instructions are dead.
   int VReg = ReMatRegs.find_first();
   while (VReg != -1) {
-    if (MRI->use_empty(VReg)) {
+    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
+  // SSA form. It's now safe to de-SSA.
+  MadeChange |= EliminateRegSequences();
+
   return MadeChange;
 }
+
+static void UpdateRegSequenceSrcs(unsigned SrcReg,
+                                  unsigned DstReg, unsigned SubIdx,
+                                  MachineRegisterInfo *MRI,
+                                  const TargetRegisterInfo &TRI) {
+  for (MachineRegisterInfo::reg_iterator RI = MRI->reg_begin(SrcReg),
+         RE = MRI->reg_end(); RI != RE; ) {
+    MachineOperand &MO = RI.getOperand();
+    ++RI;
+    MO.substVirtReg(DstReg, SubIdx, TRI);
+  }
+}
+
+/// CoalesceExtSubRegs - If a number of sources of the REG_SEQUENCE are
+/// EXTRACT_SUBREG from the same register and to the same virtual register
+/// with different sub-register indices, attempt to combine the
+/// EXTRACT_SUBREGs and pre-coalesce them. e.g.
+/// %reg1026<def> = VLDMQ %reg1025<kill>, 260, pred:14, pred:%reg0
+/// %reg1029:6<def> = EXTRACT_SUBREG %reg1026, 6
+/// %reg1029:5<def> = EXTRACT_SUBREG %reg1026<kill>, 5
+/// Since D subregs 5, 6 can combine to a Q register, we can coalesce
+/// reg1026 to reg1029.
+void
+TwoAddressInstructionPass::CoalesceExtSubRegs(SmallVector<unsigned,4> &Srcs,
+                                              unsigned DstReg) {
+  SmallSet<unsigned, 4> Seen;
+  for (unsigned i = 0, e = Srcs.size(); i != e; ++i) {
+    unsigned SrcReg = Srcs[i];
+    if (!Seen.insert(SrcReg))
+      continue;
+
+    // Check that the instructions are all in the same basic block.
+    MachineInstr *SrcDefMI = MRI->getVRegDef(SrcReg);
+    MachineInstr *DstDefMI = MRI->getVRegDef(DstReg);
+    if (SrcDefMI->getParent() != DstDefMI->getParent())
+      continue;
+
+    // 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;
+    for (MachineRegisterInfo::use_nodbg_iterator
+           UI = MRI->use_nodbg_begin(SrcReg),
+           UE = MRI->use_nodbg_end(); UI != UE; ++UI) {
+      MachineInstr *UseMI = &*UI;
+      if (!UseMI->isCopy() || UseMI->getOperand(0).getReg() != DstReg) {
+        CanCoalesce = false;
+        break;
+      }
+      SrcSubIndices.push_back(UseMI->getOperand(1).getSubReg());
+      DstSubIndices.push_back(UseMI->getOperand(0).getSubReg());
+    }
+
+    if (!CanCoalesce || SrcSubIndices.size() < 2)
+      continue;
+
+    // Check that the source subregisters can be combined.
+    std::sort(SrcSubIndices.begin(), SrcSubIndices.end());
+    unsigned NewSrcSubIdx = 0;
+    if (!TRI->canCombineSubRegIndices(MRI->getRegClass(SrcReg), SrcSubIndices,
+                                      NewSrcSubIdx))
+      continue;
+
+    // Check that the destination subregisters can also be combined.
+    std::sort(DstSubIndices.begin(), DstSubIndices.end());
+    unsigned NewDstSubIdx = 0;
+    if (!TRI->canCombineSubRegIndices(MRI->getRegClass(DstReg), DstSubIndices,
+                                      NewDstSubIdx))
+      continue;
+
+    // If neither source nor destination can be combined to the full register,
+    // just give up.  This could be improved if it ever matters.
+    if (NewSrcSubIdx != 0 && NewDstSubIdx != 0)
+      continue;
+
+    // Now that we know that all the uses are extract_subregs and that those
+    // subregs can somehow be combined, scan all the extract_subregs again to
+    // make sure the subregs are in the right order and can be composed.
+    MachineInstr *SomeMI = 0;
+    CanCoalesce = true;
+    for (MachineRegisterInfo::use_nodbg_iterator
+           UI = MRI->use_nodbg_begin(SrcReg),
+           UE = MRI->use_nodbg_end(); UI != UE; ++UI) {
+      MachineInstr *UseMI = &*UI;
+      assert(UseMI->isCopy());
+      unsigned DstSubIdx = UseMI->getOperand(0).getSubReg();
+      unsigned SrcSubIdx = UseMI->getOperand(1).getSubReg();
+      assert(DstSubIdx != 0 && "missing subreg from RegSequence elimination");
+      if ((NewDstSubIdx == 0 &&
+           TRI->composeSubRegIndices(NewSrcSubIdx, DstSubIdx) != SrcSubIdx) ||
+          (NewSrcSubIdx == 0 &&
+           TRI->composeSubRegIndices(NewDstSubIdx, SrcSubIdx) != DstSubIdx)) {
+        CanCoalesce = false;
+        break;
+      }
+      // Keep track of one of the uses.
+      SomeMI = UseMI;
+    }
+    if (!CanCoalesce)
+      continue;
+
+    // Insert a copy to replace the original.
+    MachineBasicBlock::iterator InsertLoc = SomeMI;
+    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
+           UI = MRI->use_nodbg_begin(SrcReg),
+           UE = MRI->use_nodbg_end(); UI != UE; ) {
+      MachineInstr *UseMI = &*UI;
+      ++UI;
+      if (UseMI == CopyMI)
+        continue;
+      assert(UseMI->isCopy());
+      // Move any kills to the new copy or extract instruction.
+      if (UseMI->getOperand(1).isKill()) {
+        CopyMI->getOperand(1).setIsKill();
+        if (LV)
+          // Update live variables
+          LV->replaceKillInstruction(SrcReg, UseMI, &*CopyMI);
+      }
+      UseMI->eraseFromParent();
+    }
+  }
+}
+
+static bool HasOtherRegSequenceUses(unsigned Reg, MachineInstr *RegSeq,
+                                    MachineRegisterInfo *MRI) {
+  for (MachineRegisterInfo::use_iterator UI = MRI->use_begin(Reg),
+         UE = MRI->use_end(); UI != UE; ++UI) {
+    MachineInstr *UseMI = &*UI;
+    if (UseMI != RegSeq && UseMI->isRegSequence())
+      return true;
+  }
+  return false;
+}
+
+/// EliminateRegSequences - Eliminate REG_SEQUENCE instructions as part
+/// of the de-ssa process. This replaces sources of REG_SEQUENCE as
+/// sub-register references of the register defined by REG_SEQUENCE. e.g.
+///
+/// %reg1029<def>, %reg1030<def> = VLD1q16 %reg1024<kill>, ...
+/// %reg1031<def> = REG_SEQUENCE %reg1029<kill>, 5, %reg1030<kill>, 6
+/// =>
+/// %reg1031:5<def>, %reg1031:6<def> = VLD1q16 %reg1024<kill>, ...
+bool TwoAddressInstructionPass::EliminateRegSequences() {
+  if (RegSequences.empty())
+    return false;
+
+  for (unsigned i = 0, e = RegSequences.size(); i != e; ++i) {
+    MachineInstr *MI = RegSequences[i];
+    unsigned DstReg = MI->getOperand(0).getReg();
+    if (MI->getOperand(0).getSubReg() ||
+        TargetRegisterInfo::isPhysicalRegister(DstReg) ||
+        !(MI->getNumOperands() & 1)) {
+      DEBUG(dbgs() << "Illegal REG_SEQUENCE instruction:" << *MI);
+      llvm_unreachable(0);
+    }
+
+    bool IsImpDef = true;
+    SmallVector<unsigned, 4> RealSrcs;
+    SmallSet<unsigned, 4> Seen;
+    for (unsigned i = 1, e = MI->getNumOperands(); i < e; i += 2) {
+      unsigned SrcReg = MI->getOperand(i).getReg();
+      if (MI->getOperand(i).getSubReg() ||
+          TargetRegisterInfo::isPhysicalRegister(SrcReg)) {
+        DEBUG(dbgs() << "Illegal REG_SEQUENCE instruction:" << *MI);
+        llvm_unreachable(0);
+      }
+
+      MachineInstr *DefMI = MRI->getVRegDef(SrcReg);
+      if (DefMI->isImplicitDef()) {
+        DefMI->eraseFromParent();
+        continue;
+      }
+      IsImpDef = false;
+
+      // Remember COPY sources. These might be candidate for coalescing.
+      if (DefMI->isCopy() && DefMI->getOperand(1).getSubReg())
+        RealSrcs.push_back(DefMI->getOperand(1).getReg());
+
+      bool isKill = MI->getOperand(i).isKill();
+      if (!Seen.insert(SrcReg) || MI->getParent() != DefMI->getParent() ||
+          !isKill || HasOtherRegSequenceUses(SrcReg, MI, MRI)) {
+        // 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.
+        // BB0:
+        // reg1051:10<def> =
+        // ...
+        // BB1:
+        // ... = reg1051:10
+        // BB2:
+        // reg1051:9<def> =
+        // LiveIntervalAnalysis won't like it.
+        //
+        // If the REG_SEQUENCE doesn't kill its source, keeping live variables
+        // correctly up to date becomes very difficult. Insert a copy.
+
+        // 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;
+        MachineInstr *CopyMI = BuildMI(*MI->getParent(), InsertLoc,
+                                MI->getDebugLoc(), TII->get(TargetOpcode::COPY))
+            .addReg(DstReg, RegState::Define, MI->getOperand(i+1).getImm())
+            .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);
+    }
+
+    if (IsImpDef) {
+      DEBUG(dbgs() << "Turned: " << *MI << " into an IMPLICIT_DEF");
+      MI->setDesc(TII->get(TargetOpcode::IMPLICIT_DEF));
+      for (int j = MI->getNumOperands() - 1, ee = 0; j > ee; --j)
+        MI->RemoveOperand(j);      
+    } else {
+      DEBUG(dbgs() << "Eliminated: " << *MI);
+      MI->eraseFromParent();
+    }
+
+    // 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();
+  return true;
+}