Allow ARM if-converter to be run after post allocation scheduling.

- This fixed a number of bugs in if-converter, tail merging, and post-allocation
  scheduler. If-converter now runs branch folding / tail merging first to
  maximize if-conversion opportunities.
- Also changed the t2IT instruction slightly. It now defines the ITSTATE
  register which is read by instructions in the IT block.
- Added Thumb2 specific hazard recognizer to ensure the scheduler doesn't
  change the instruction ordering in the IT block (since IT mask has been
  finalized). It also ensures no other instructions can be scheduled between
  instructions in the IT block.

This is not yet enabled.


git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@106344 91177308-0d34-0410-b5e6-96231b3b80d8

18 files changed, 302 insertions, 82 deletions

diff --git a/include/llvm/Target/TargetInstrInfo.h b/include/llvm/Target/TargetInstrInfo.h
index 21675dc..90608f1 100644
--- a/include/llvm/Target/TargetInstrInfo.h
+++ b/include/llvm/Target/TargetInstrInfo.h
@@ -321,6 +321,12 @@ public:
     assert(0 && "Target didn't implement TargetInstrInfo::InsertBranch!"); 
     return 0;
   }
+
+  /// ReplaceTailWithBranchTo - Delete the instruction OldInst and everything
+  /// after it, replacing it with an unconditional branch to NewDest. This is
+  /// used by the tail merging pass.
+  virtual void ReplaceTailWithBranchTo(MachineBasicBlock::iterator Tail,
+                                       MachineBasicBlock *NewDest) const = 0;
   
   /// copyRegToReg - Emit instructions to copy between a pair of registers. It
   /// returns false if the target does not how to copy between the specified
@@ -562,6 +568,13 @@ public:
     return true;
   }
 
+  /// isSchedulingBoundary - Test if the given instruction should be
+  /// considered a scheduling boundary. This primarily includes labels and
+  /// terminators.
+  virtual bool isSchedulingBoundary(const MachineInstr *MI,
+                                    const MachineBasicBlock *MBB,
+                                    const MachineFunction &MF) const = 0;
+
   /// GetInstSize - Returns the size of the specified Instruction.
   /// 
   virtual unsigned GetInstSizeInBytes(const MachineInstr *MI) const {
@@ -595,6 +608,8 @@ protected:
   TargetInstrInfoImpl(const TargetInstrDesc *desc, unsigned NumOpcodes)
   : TargetInstrInfo(desc, NumOpcodes) {}
 public:
+  virtual void ReplaceTailWithBranchTo(MachineBasicBlock::iterator OldInst,
+                                       MachineBasicBlock *NewDest) const;
   virtual MachineInstr *commuteInstruction(MachineInstr *MI,
                                            bool NewMI = false) const;
   virtual bool findCommutedOpIndices(MachineInstr *MI, unsigned &SrcOpIdx1,
@@ -610,6 +625,9 @@ public:
                                   MachineFunction &MF) const;
   virtual bool produceSameValue(const MachineInstr *MI0,
                                 const MachineInstr *MI1) const;
+  virtual bool isSchedulingBoundary(const MachineInstr *MI,
+                                    const MachineBasicBlock *MBB,
+                                    const MachineFunction &MF) const;
   virtual unsigned GetFunctionSizeInBytes(const MachineFunction &MF) const;
 
   virtual ScheduleHazardRecognizer *
diff --git a/lib/CodeGen/BranchFolding.cpp b/lib/CodeGen/BranchFolding.cpp
index e440e40..e17e47a 100644
--- a/lib/CodeGen/BranchFolding.cpp
+++ b/lib/CodeGen/BranchFolding.cpp
@@ -358,24 +358,10 @@ static unsigned ComputeCommonTailLength(MachineBasicBlock *MBB1,
 }
 
 /// ReplaceTailWithBranchTo - Delete the instruction OldInst and everything
-/// after it, replacing it with an unconditional branch to NewDest.  This
-/// returns true if OldInst's block is modified, false if NewDest is modified.
+/// after it, replacing it with an unconditional branch to NewDest.
 void BranchFolder::ReplaceTailWithBranchTo(MachineBasicBlock::iterator OldInst,
                                            MachineBasicBlock *NewDest) {
-  MachineBasicBlock *OldBB = OldInst->getParent();
-
-  // Remove all the old successors of OldBB from the CFG.
-  while (!OldBB->succ_empty())
-    OldBB->removeSuccessor(OldBB->succ_begin());
-
-  // Remove all the dead instructions from the end of OldBB.
-  OldBB->erase(OldInst, OldBB->end());
-
-  // If OldBB isn't immediately before OldBB, insert a branch to it.
-  if (++MachineFunction::iterator(OldBB) != MachineFunction::iterator(NewDest))
-    TII->InsertBranch(*OldBB, NewDest, 0, SmallVector<MachineOperand, 0>(),
-                      OldInst->getDebugLoc());
-  OldBB->addSuccessor(NewDest);
+  TII->ReplaceTailWithBranchTo(OldInst, NewDest);
   ++NumTailMerge;
 }
 
diff --git a/lib/CodeGen/IfConversion.cpp b/lib/CodeGen/IfConversion.cpp
index 36e6fc7..16d3247 100644
--- a/lib/CodeGen/IfConversion.cpp
+++ b/lib/CodeGen/IfConversion.cpp
@@ -235,6 +235,12 @@ bool IfConverter::runOnMachineFunction(MachineFunction &MF) {
   TRI = MF.getTarget().getRegisterInfo();
   if (!TII) return false;
 
+  // Tail merge tend to expose more if-conversion opportunities.
+  BranchFolder BF(true);
+  bool BFChange = BF.OptimizeFunction(MF, TII,
+                                   MF.getTarget().getRegisterInfo(),
+                                   getAnalysisIfAvailable<MachineModuleInfo>());
+
   DEBUG(dbgs() << "\nIfcvt: function (" << ++FnNum <<  ") \'"
                << MF.getFunction()->getName() << "\'");
 
@@ -376,6 +382,7 @@ bool IfConverter::runOnMachineFunction(MachineFunction &MF) {
                         getAnalysisIfAvailable<MachineModuleInfo>());
   }
 
+  MadeChange |= BFChange;
   return MadeChange;
 }
 
diff --git a/lib/CodeGen/MachineVerifier.cpp b/lib/CodeGen/MachineVerifier.cpp
index 8baf01c..2297c90 100644
--- a/lib/CodeGen/MachineVerifier.cpp
+++ b/lib/CodeGen/MachineVerifier.cpp
@@ -390,7 +390,8 @@ MachineVerifier::visitMachineBasicBlockBefore(const MachineBasicBlock *MBB) {
         report("MBB exits via unconditional fall-through but its successor "
                "differs from its CFG successor!", MBB);
       }
-      if (!MBB->empty() && MBB->back().getDesc().isBarrier()) {
+      if (!MBB->empty() && MBB->back().getDesc().isBarrier() &&
+          !TII->isPredicated(&MBB->back())) {
         report("MBB exits via unconditional fall-through but ends with a "
                "barrier instruction!", MBB);
       }
diff --git a/lib/CodeGen/PostRASchedulerList.cpp b/lib/CodeGen/PostRASchedulerList.cpp
index f3c725d..4af8e07 100644
--- a/lib/CodeGen/PostRASchedulerList.cpp
+++ b/lib/CodeGen/PostRASchedulerList.cpp
@@ -79,6 +79,7 @@ AntiDepBreaker::~AntiDepBreaker() { }
 namespace {
   class PostRAScheduler : public MachineFunctionPass {
     AliasAnalysis *AA;
+    const TargetInstrInfo *TII;
     CodeGenOpt::Level OptLevel;
 
   public:
@@ -181,30 +182,9 @@ namespace {
   };
 }
 
-/// isSchedulingBoundary - Test if the given instruction should be
-/// considered a scheduling boundary. This primarily includes labels
-/// and terminators.
-///
-static bool isSchedulingBoundary(const MachineInstr *MI,
-                                 const MachineFunction &MF) {
-  // Terminators and labels can't be scheduled around.
-  if (MI->getDesc().isTerminator() || MI->isLabel())
-    return true;
-
-  // Don't attempt to schedule around any instruction that defines
-  // a stack-oriented pointer, as it's unlikely to be profitable. This
-  // saves compile time, because it doesn't require every single
-  // stack slot reference to depend on the instruction that does the
-  // modification.
-  const TargetLowering &TLI = *MF.getTarget().getTargetLowering();
-  if (MI->definesRegister(TLI.getStackPointerRegisterToSaveRestore()))
-    return true;
-
-  return false;
-}
-
 bool PostRAScheduler::runOnMachineFunction(MachineFunction &Fn) {
   AA = &getAnalysis<AliasAnalysis>();
+  TII = Fn.getTarget().getInstrInfo();
 
   // Check for explicit enable/disable of post-ra scheduling.
   TargetSubtarget::AntiDepBreakMode AntiDepMode = TargetSubtarget::ANTIDEP_NONE;
@@ -265,8 +245,8 @@ bool PostRAScheduler::runOnMachineFunction(MachineFunction &Fn) {
     MachineBasicBlock::iterator Current = MBB->end();
     unsigned Count = MBB->size(), CurrentCount = Count;
     for (MachineBasicBlock::iterator I = Current; I != MBB->begin(); ) {
-      MachineInstr *MI = prior(I);
-      if (isSchedulingBoundary(MI, Fn)) {
+      MachineInstr *MI = llvm::prior(I);
+      if (TII->isSchedulingBoundary(MI, MBB, Fn)) {
         Scheduler.Run(MBB, I, Current, CurrentCount);
         Scheduler.EmitSchedule();
         Current = MI;
diff --git a/lib/CodeGen/TargetInstrInfoImpl.cpp b/lib/CodeGen/TargetInstrInfoImpl.cpp
index 1641c71..53f3ee8 100644
--- a/lib/CodeGen/TargetInstrInfoImpl.cpp
+++ b/lib/CodeGen/TargetInstrInfoImpl.cpp
@@ -13,6 +13,7 @@
 //===----------------------------------------------------------------------===//
 
 #include "llvm/Target/TargetInstrInfo.h"
+#include "llvm/Target/TargetLowering.h"
 #include "llvm/Target/TargetMachine.h"
 #include "llvm/Target/TargetRegisterInfo.h"
 #include "llvm/ADT/SmallVector.h"
@@ -27,6 +28,25 @@
 #include "llvm/Support/raw_ostream.h"
 using namespace llvm;
 
+void
+TargetInstrInfoImpl::ReplaceTailWithBranchTo(MachineBasicBlock::iterator Tail,
+                                             MachineBasicBlock *NewDest) const {
+  MachineBasicBlock *MBB = Tail->getParent();
+
+  // Remove all the old successors of MBB from the CFG.
+  while (!MBB->succ_empty())
+    MBB->removeSuccessor(MBB->succ_begin());
+
+  // Remove all the dead instructions from the end of MBB.
+  MBB->erase(Tail, MBB->end());
+
+  // If MBB isn't immediately before MBB, insert a branch to it.
+  if (++MachineFunction::iterator(MBB) != MachineFunction::iterator(NewDest))
+    InsertBranch(*MBB, NewDest, 0, SmallVector<MachineOperand, 0>(),
+                 Tail->getDebugLoc());
+  MBB->addSuccessor(NewDest);
+}
+
 // commuteInstruction - The default implementation of this method just exchanges
 // the two operands returned by findCommutedOpIndices.
 MachineInstr *TargetInstrInfoImpl::commuteInstruction(MachineInstr *MI,
@@ -316,6 +336,28 @@ isReallyTriviallyReMaterializableGeneric(const MachineInstr *MI,
   return true;
 }
 
+/// isSchedulingBoundary - Test if the given instruction should be
+/// considered a scheduling boundary. This primarily includes labels
+/// and terminators.
+bool TargetInstrInfoImpl::isSchedulingBoundary(const MachineInstr *MI,
+                                               const MachineBasicBlock *MBB,
+                                               const MachineFunction &MF) const{
+  // Terminators and labels can't be scheduled around.
+  if (MI->getDesc().isTerminator() || MI->isLabel())
+    return true;
+
+  // Don't attempt to schedule around any instruction that defines
+  // a stack-oriented pointer, as it's unlikely to be profitable. This
+  // saves compile time, because it doesn't require every single
+  // stack slot reference to depend on the instruction that does the
+  // modification.
+  const TargetLowering &TLI = *MF.getTarget().getTargetLowering();
+  if (MI->definesRegister(TLI.getStackPointerRegisterToSaveRestore()))
+    return true;
+
+  return false;
+}
+
 // Default implementation of CreateTargetPostRAHazardRecognizer.
 ScheduleHazardRecognizer *TargetInstrInfoImpl::
 CreateTargetPostRAHazardRecognizer(const InstrItineraryData &II) const {
diff --git a/lib/Target/ARM/ARMBaseInstrInfo.cpp b/lib/Target/ARM/ARMBaseInstrInfo.cpp
index 6eb3fab..8d3ad80 100644
--- a/lib/Target/ARM/ARMBaseInstrInfo.cpp
+++ b/lib/Target/ARM/ARMBaseInstrInfo.cpp
@@ -1306,6 +1306,34 @@ bool ARMBaseInstrInfo::produceSameValue(const MachineInstr *MI0,
   return MI0->isIdenticalTo(MI1, MachineInstr::IgnoreVRegDefs);
 }
 
+bool ARMBaseInstrInfo::isSchedulingBoundary(const MachineInstr *MI,
+                                            const MachineBasicBlock *MBB,
+                                            const MachineFunction &MF) const {
+  // Terminators and labels can't be scheduled around.
+  if (MI->getDesc().isTerminator() || MI->isLabel())
+    return true;
+
+  // Treat the start of the IT block as a scheduling boundary, but schedule
+  // t2IT along with all instructions following it.
+  // FIXME: This is a big hammer. But the alternative is to add all potential
+  // true and anti dependencies to IT block instructions as implicit operands
+  // to the t2IT instruction. The added compile time and complexity does not
+  // seem worth it.
+  MachineBasicBlock::const_iterator I = MI;
+  if (++I != MBB->end() && I->getOpcode() == ARM::t2IT)
+    return true;
+
+  // Don't attempt to schedule around any instruction that defines
+  // a stack-oriented pointer, as it's unlikely to be profitable. This
+  // saves compile time, because it doesn't require every single
+  // stack slot reference to depend on the instruction that does the
+  // modification.
+  if (MI->definesRegister(ARM::SP))
+    return true;
+
+  return false;
+}
+
 /// getInstrPredicate - If instruction is predicated, returns its predicate
 /// condition, otherwise returns AL. It also returns the condition code
 /// register by reference.
diff --git a/lib/Target/ARM/ARMBaseInstrInfo.h b/lib/Target/ARM/ARMBaseInstrInfo.h
index fec26e1..bc82e14 100644
--- a/lib/Target/ARM/ARMBaseInstrInfo.h
+++ b/lib/Target/ARM/ARMBaseInstrInfo.h
@@ -319,6 +319,10 @@ public:
 
   virtual bool produceSameValue(const MachineInstr *MI0,
                                 const MachineInstr *MI1) const;
+
+  virtual bool isSchedulingBoundary(const MachineInstr *MI,
+                                    const MachineBasicBlock *MBB,
+                                    const MachineFunction &MF) const;
 };
 
 static inline
diff --git a/lib/Target/ARM/ARMInstrThumb2.td b/lib/Target/ARM/ARMInstrThumb2.td
index 0172bb4..c2cb05e 100644
--- a/lib/Target/ARM/ARMInstrThumb2.td
+++ b/lib/Target/ARM/ARMInstrThumb2.td
@@ -2528,6 +2528,7 @@ def t2Bcc : T2I<(outs), (ins brtarget:$target), IIC_Br,
 
 
 // IT block
+let Defs = [ITSTATE] in
 def t2IT : Thumb2XI<(outs), (ins it_pred:$cc, it_mask:$mask),
                     AddrModeNone, Size2Bytes,  IIC_iALUx,
                     "it$mask\t$cc", "", []> {
diff --git a/lib/Target/ARM/ARMMachineFunctionInfo.h b/lib/Target/ARM/ARMMachineFunctionInfo.h
index 0134276..7e57a1c 100644
--- a/lib/Target/ARM/ARMMachineFunctionInfo.h
+++ b/lib/Target/ARM/ARMMachineFunctionInfo.h
@@ -88,6 +88,9 @@ class ARMFunctionInfo : public MachineFunctionInfo {
   /// VarArgsFrameIndex - FrameIndex for start of varargs area.
   int VarArgsFrameIndex;
 
+  /// HasITBlocks - True if IT blocks have been inserted.
+  bool HasITBlocks;
+
 public:
   ARMFunctionInfo() :
     isThumb(false),
@@ -97,7 +100,8 @@ public:
     FramePtrSpillOffset(0), GPRCS1Offset(0), GPRCS2Offset(0), DPRCSOffset(0),
     GPRCS1Size(0), GPRCS2Size(0), DPRCSSize(0),
     GPRCS1Frames(0), GPRCS2Frames(0), DPRCSFrames(0),
-    JumpTableUId(0), ConstPoolEntryUId(0), VarArgsFrameIndex(0) {}
+    JumpTableUId(0), ConstPoolEntryUId(0), VarArgsFrameIndex(0),
+    HasITBlocks(false) {}
 
   explicit ARMFunctionInfo(MachineFunction &MF) :
     isThumb(MF.getTarget().getSubtarget<ARMSubtarget>().isThumb()),
@@ -108,7 +112,8 @@ public:
     GPRCS1Size(0), GPRCS2Size(0), DPRCSSize(0),
     GPRCS1Frames(32), GPRCS2Frames(32), DPRCSFrames(32),
     SpilledCSRegs(MF.getTarget().getRegisterInfo()->getNumRegs()),
-    JumpTableUId(0), ConstPoolEntryUId(0), VarArgsFrameIndex(0) {}
+    JumpTableUId(0), ConstPoolEntryUId(0), VarArgsFrameIndex(0),
+    HasITBlocks(false) {}
 
   bool isThumbFunction() const { return isThumb; }
   bool isThumb1OnlyFunction() const { return isThumb && !hasThumb2; }
@@ -229,6 +234,9 @@ public:
 
   int getVarArgsFrameIndex() const { return VarArgsFrameIndex; }
   void setVarArgsFrameIndex(int Index) { VarArgsFrameIndex = Index; }
+
+  bool hasITBlocks() const { return HasITBlocks; }
+  void setHasITBlocks(bool h) { HasITBlocks = h; }
 };
 } // End llvm namespace
 
diff --git a/lib/Target/ARM/ARMRegisterInfo.td b/lib/Target/ARM/ARMRegisterInfo.td
index b9417e3..e41d9fc 100644
--- a/lib/Target/ARM/ARMRegisterInfo.td
+++ b/lib/Target/ARM/ARMRegisterInfo.td
@@ -197,9 +197,9 @@ def QQQQ3 : ARMReg<3, "qqqq3", [QQ6, QQ7]>;
 }
 
 // Current Program Status Register.
-def CPSR  : ARMReg<0, "cpsr">;
-
-def FPSCR : ARMReg<1, "fpscr">;
+def CPSR    : ARMReg<0, "cpsr">;
+def FPSCR   : ARMReg<1, "fpscr">;
+def ITSTATE : ARMReg<2, "itstate">;
 
 // Register classes.
 //
@@ -557,4 +557,3 @@ def QQQQPR : RegisterClass<"ARM", [v8i64],
 
 // Condition code registers.
 def CCR : RegisterClass<"ARM", [i32], 32, [CPSR]>;
-
diff --git a/lib/Target/ARM/ARMTargetMachine.cpp b/lib/Target/ARM/ARMTargetMachine.cpp
index 2101da8..06c893a 100644
--- a/lib/Target/ARM/ARMTargetMachine.cpp
+++ b/lib/Target/ARM/ARMTargetMachine.cpp
@@ -131,8 +131,10 @@ bool ARMBaseTargetMachine::addPreSched2(PassManagerBase &PM,
   PM.add(createARMExpandPseudoPass());
 
   if (EarlyIfConvert && OptLevel != CodeGenOpt::None) {
-    if (!Subtarget.isThumb1Only()) 
+    if (!Subtarget.isThumb1Only())
       PM.add(createIfConverterPass());
+    if (Subtarget.isThumb2())
+      PM.add(createThumb2ITBlockPass());
   }
 
   return true;
@@ -146,7 +148,8 @@ bool ARMBaseTargetMachine::addPreEmitPass(PassManagerBase &PM,
   }
 
   if (Subtarget.isThumb2()) {
-    PM.add(createThumb2ITBlockPass());
+    if (!EarlyIfConvert)
+      PM.add(createThumb2ITBlockPass());
     PM.add(createThumb2SizeReductionPass());
   }
 
diff --git a/lib/Target/ARM/Thumb2ITBlockPass.cpp b/lib/Target/ARM/Thumb2ITBlockPass.cpp
index bd8be96..d72bb5d 100644
--- a/lib/Target/ARM/Thumb2ITBlockPass.cpp
+++ b/lib/Target/ARM/Thumb2ITBlockPass.cpp
@@ -31,6 +31,7 @@ namespace {
       MachineFunctionPass(&ID), PreRegAlloc(PreRA) {}
 
     const Thumb2InstrInfo *TII;
+    const TargetRegisterInfo *TRI;
     ARMFunctionInfo *AFI;
 
     virtual bool runOnMachineFunction(MachineFunction &Fn);
@@ -52,6 +53,10 @@ namespace {
                            SmallVector<MachineInstr*,4> &LastUses);
     bool InsertITBlock(MachineInstr *First, MachineInstr *Last);
     bool InsertITBlocks(MachineBasicBlock &MBB);
+    bool MoveCopyOutOfITBlock(MachineInstr *MI,
+                              ARMCC::CondCodes CC, ARMCC::CondCodes OCC,
+                              SmallSet<unsigned, 4> &Defs,
+                              SmallSet<unsigned, 4> &Uses);
     bool InsertITInstructions(MachineBasicBlock &MBB);
   };
   char Thumb2ITBlockPass::ID = 0;
@@ -249,20 +254,77 @@ bool Thumb2ITBlockPass::InsertITBlocks(MachineBasicBlock &MBB) {
   return Modified;
 }
 
-static void TrackDefUses(MachineInstr *MI, SmallSet<unsigned, 4> &Defs,
-                         SmallSet<unsigned, 4> &Uses) {
+/// TrackDefUses - Tracking what registers are being defined and used by
+/// instructions in the IT block. This also tracks "dependencies", i.e. uses
+/// in the IT block that are defined before the IT instruction.
+static void TrackDefUses(MachineInstr *MI,
+                         SmallSet<unsigned, 4> &Defs,
+                         SmallSet<unsigned, 4> &Uses,
+                         const TargetRegisterInfo *TRI) {
+  SmallVector<unsigned, 4> LocalDefs;
+  SmallVector<unsigned, 4> LocalUses;
+
   for (unsigned i = 0, e = MI->getNumOperands(); i != e; ++i) {
     MachineOperand &MO = MI->getOperand(i);
     if (!MO.isReg())
       continue;
     unsigned Reg = MO.getReg();
-    if (!Reg)
+    if (!Reg || Reg == ARM::ITSTATE || Reg == ARM::SP)
       continue;
-    if (MO.isDef())
-      Defs.insert(Reg);
+    if (MO.isUse())
+      LocalUses.push_back(Reg);
     else
-      Uses.insert(Reg);
+      LocalDefs.push_back(Reg);
+  }
+
+  for (unsigned i = 0, e = LocalUses.size(); i != e; ++i) {
+    unsigned Reg = LocalUses[i];
+    Uses.insert(Reg);
+    for (const unsigned *Subreg = TRI->getSubRegisters(Reg);
+         *Subreg; ++Subreg)
+      Uses.insert(*Subreg);
+  }
+
+  for (unsigned i = 0, e = LocalDefs.size(); i != e; ++i) {
+    unsigned Reg = LocalDefs[i];
+    Defs.insert(Reg);
+    for (const unsigned *Subreg = TRI->getSubRegisters(Reg);
+         *Subreg; ++Subreg)
+      Defs.insert(*Subreg);
+    if (Reg == ARM::CPSR)
+      continue;
+  }
+}
+
+bool
+Thumb2ITBlockPass::MoveCopyOutOfITBlock(MachineInstr *MI,
+                                      ARMCC::CondCodes CC, ARMCC::CondCodes OCC,
+                                        SmallSet<unsigned, 4> &Defs,
+                                        SmallSet<unsigned, 4> &Uses) {
+  unsigned SrcReg, DstReg, SrcSubIdx, DstSubIdx;
+  if (TII->isMoveInstr(*MI, SrcReg, DstReg, SrcSubIdx, DstSubIdx)) {
+    assert(SrcSubIdx == 0 && DstSubIdx == 0 &&
+           "Sub-register indices still around?");
+    // llvm models select's as two-address instructions. That means a copy
+    // is inserted before a t2MOVccr, etc. If the copy is scheduled in
+    // between selects we would end up creating multiple IT blocks.
+
+    // First check if it's safe to move it.
+    if (Uses.count(DstReg) || Defs.count(SrcReg))
+      return false;
+
+    // Then peek at the next instruction to see if it's predicated on CC or OCC.
+    // If not, then there is nothing to be gained by moving the copy.
+    MachineBasicBlock::iterator I = MI; ++I;
+    MachineBasicBlock::iterator E = MI->getParent()->end();
+    while (I != E && I->isDebugValue())
+      ++I;
+    unsigned NPredReg = 0;
+    ARMCC::CondCodes NCC = getPredicate(I, NPredReg);
+    if (NCC == CC || NCC == OCC)
+      return true;
   }
+  return false;
 }
 
 bool Thumb2ITBlockPass::InsertITInstructions(MachineBasicBlock &MBB) {
@@ -283,15 +345,21 @@ bool Thumb2ITBlockPass::InsertITInstructions(MachineBasicBlock &MBB) {
 
     Defs.clear();
     Uses.clear();
-    TrackDefUses(MI, Defs, Uses);
+    TrackDefUses(MI, Defs, Uses, TRI);
 
     // Insert an IT instruction.
     MachineInstrBuilder MIB = BuildMI(MBB, MBBI, dl, TII->get(ARM::t2IT))
       .addImm(CC);
+
+    // Add implicit use of ITSTATE to IT block instructions.
+    MI->addOperand(MachineOperand::CreateReg(ARM::ITSTATE, false/*ifDef*/,
+                                             true/*isImp*/, false/*isKill*/));
+
+    MachineInstr *LastITMI = MI;
     MachineBasicBlock::iterator InsertPos = MIB;
     ++MBBI;
 
-    // Finalize IT mask.
+    // Form IT block.
     ARMCC::CondCodes OCC = ARMCC::getOppositeCondition(CC);
     unsigned Mask = 0, Pos = 3;
     // Branches, including tricky ones like LDM_RET, need to end an IT
@@ -306,35 +374,36 @@ bool Thumb2ITBlockPass::InsertITInstructions(MachineBasicBlock &MBB) {
 
       unsigned NPredReg = 0;
       ARMCC::CondCodes NCC = getPredicate(NMI, NPredReg);
-      if (NCC == CC || NCC == OCC)
+      if (NCC == CC || NCC == OCC) {
         Mask |= (NCC & 1) << Pos;
-      else {
-        unsigned SrcReg, DstReg, SrcSubIdx, DstSubIdx;
+        // Add implicit use of ITSTATE.
+        NMI->addOperand(MachineOperand::CreateReg(ARM::ITSTATE, false/*ifDef*/,
+                                                true/*isImp*/, false/*isKill*/));
+        LastITMI = NMI;
+      } else {
         if (NCC == ARMCC::AL &&
-            TII->isMoveInstr(*NMI, SrcReg, DstReg, SrcSubIdx, DstSubIdx)) {
-          assert(SrcSubIdx == 0 && DstSubIdx == 0 &&
-                 "Sub-register indices still around?");
-          // llvm models select's as two-address instructions. That means a copy
-          // is inserted before a t2MOVccr, etc. If the copy is scheduled in
-          // between selects we would end up creating multiple IT blocks.
-          if (!Uses.count(DstReg) && !Defs.count(SrcReg)) {
-            --MBBI;
-            MBB.remove(NMI);
-            MBB.insert(InsertPos, NMI);
-            ++NumMovedInsts;
-            continue;
-          }
+            MoveCopyOutOfITBlock(NMI, CC, OCC, Defs, Uses)) {
+          --MBBI;
+          MBB.remove(NMI);
+          MBB.insert(InsertPos, NMI);
+          ++NumMovedInsts;
+          continue;
         }
         break;
       }
-      TrackDefUses(NMI, Defs, Uses);
+      TrackDefUses(NMI, Defs, Uses, TRI);
       --Pos;
     }
 
+    // Finalize IT mask.
     Mask |= (1 << Pos);
     // Tag along (firstcond[0] << 4) with the mask.
     Mask |= (CC & 1) << 4;
     MIB.addImm(Mask);
+
+    // Last instruction in IT block kills ITSTATE.
+    LastITMI->findRegisterUseOperand(ARM::ITSTATE)->setIsKill();
+
     Modified = true;
     ++NumITs;
   }
@@ -346,6 +415,7 @@ bool Thumb2ITBlockPass::runOnMachineFunction(MachineFunction &Fn) {
   const TargetMachine &TM = Fn.getTarget();
   AFI = Fn.getInfo<ARMFunctionInfo>();
   TII = static_cast<const Thumb2InstrInfo*>(TM.getInstrInfo());
+  TRI = TM.getRegisterInfo();
 
   if (!AFI->isThumbFunction())
     return false;
@@ -360,6 +430,9 @@ bool Thumb2ITBlockPass::runOnMachineFunction(MachineFunction &Fn) {
       Modified |= InsertITInstructions(MBB);
   }
 
+  if (Modified && !PreRegAlloc)
+    AFI->setHasITBlocks(true);
+
   return Modified;
 }
 
diff --git a/lib/Target/ARM/Thumb2InstrInfo.cpp b/lib/Target/ARM/Thumb2InstrInfo.cpp
index f78111f..866ffb1 100644
--- a/lib/Target/ARM/Thumb2InstrInfo.cpp
+++ b/lib/Target/ARM/Thumb2InstrInfo.cpp
@@ -17,12 +17,13 @@
 #include "ARMAddressingModes.h"
 #include "ARMGenInstrInfo.inc"
 #include "ARMMachineFunctionInfo.h"
+#include "Thumb2HazardRecognizer.h"
+#include "Thumb2InstrInfo.h"
 #include "llvm/CodeGen/MachineFrameInfo.h"
 #include "llvm/CodeGen/MachineInstrBuilder.h"
 #include "llvm/CodeGen/MachineMemOperand.h"
 #include "llvm/CodeGen/PseudoSourceValue.h"
 #include "llvm/ADT/SmallVector.h"
-#include "Thumb2InstrInfo.h"
 
 using namespace llvm;
 
@@ -35,6 +36,57 @@ unsigned Thumb2InstrInfo::getUnindexedOpcode(unsigned Opc) const {
   return 0;
 }
 
+void
+Thumb2InstrInfo::ReplaceTailWithBranchTo(MachineBasicBlock::iterator Tail,
+                                         MachineBasicBlock *NewDest) const {
+  MachineBasicBlock *MBB = Tail->getParent();
+  ARMFunctionInfo *AFI = MBB->getParent()->getInfo<ARMFunctionInfo>();
+  if (!AFI->hasITBlocks()) {
+    TargetInstrInfoImpl::ReplaceTailWithBranchTo(Tail, NewDest);
+    return;
+  }
+
+  // If the first instruction of Tail is predicated, we may have to update
+  // the IT instruction.
+  unsigned PredReg = 0;
+  ARMCC::CondCodes CC = llvm::getInstrPredicate(Tail, PredReg);
+  MachineBasicBlock::iterator MBBI = Tail;
+  if (CC != ARMCC::AL)
+    // Expecting at least the t2IT instruction before it.
+    --MBBI;
+
+  // Actually replace the tail.
+  TargetInstrInfoImpl::ReplaceTailWithBranchTo(Tail, NewDest);
+
+  // Fix up IT.
+  if (CC != ARMCC::AL) {
+    MachineBasicBlock::iterator E = MBB->begin();
+    unsigned Count = 4; // At most 4 instructions in an IT block.
+    while (Count && MBBI != E) {
+      if (MBBI->isDebugValue()) {
+        --MBBI;
+        continue;
+      }
+      if (MBBI->getOpcode() == ARM::t2IT) {
+        unsigned Mask = MBBI->getOperand(1).getImm();
+        if (Count == 4)
+          MBBI->eraseFromParent();
+        else {
+          unsigned MaskOn = 1 << Count;
+          unsigned MaskOff = ~(MaskOn - 1);
+          MBBI->getOperand(1).setImm((Mask & MaskOff) | MaskOn);
+        }
+        return;
+      }
+      --MBBI;
+      --Count;
+    }
+
+    // Ctrl flow can reach here if branch folding is run before IT block
+    // formation pass.
+  }
+}
+
 bool
 Thumb2InstrInfo::copyRegToReg(MachineBasicBlock &MBB,
                               MachineBasicBlock::iterator I,
@@ -116,6 +168,11 @@ loadRegFromStackSlot(MachineBasicBlock &MBB, MachineBasicBlock::iterator I,
   ARMBaseInstrInfo::loadRegFromStackSlot(MBB, I, DestReg, FI, RC, TRI);
 }
 
+ScheduleHazardRecognizer *Thumb2InstrInfo::
+CreateTargetPostRAHazardRecognizer(const InstrItineraryData &II) const {
+  return (ScheduleHazardRecognizer *)new Thumb2HazardRecognizer(II);
+}
+
 void llvm::emitT2RegPlusImmediate(MachineBasicBlock &MBB,
                                MachineBasicBlock::iterator &MBBI, DebugLoc dl,
                                unsigned DestReg, unsigned BaseReg, int NumBytes,
diff --git a/lib/Target/ARM/Thumb2InstrInfo.h b/lib/Target/ARM/Thumb2InstrInfo.h
index 3f233c4..d5fc359 100644
--- a/lib/Target/ARM/Thumb2InstrInfo.h
+++ b/lib/Target/ARM/Thumb2InstrInfo.h
@@ -20,7 +20,8 @@
 #include "Thumb2RegisterInfo.h"
 
 namespace llvm {
-  class ARMSubtarget;
+class ARMSubtarget;
+class ScheduleHazardRecognizer;
 
 class Thumb2InstrInfo : public ARMBaseInstrInfo {
   Thumb2RegisterInfo RI;
@@ -31,6 +32,9 @@ public:
   // if there is not such an opcode.
   unsigned getUnindexedOpcode(unsigned Opc) const;
 
+  void ReplaceTailWithBranchTo(MachineBasicBlock::iterator Tail,
+                               MachineBasicBlock *NewDest) const;
+
   bool copyRegToReg(MachineBasicBlock &MBB,
                     MachineBasicBlock::iterator I,
                     unsigned DestReg, unsigned SrcReg,
@@ -60,6 +64,9 @@ public:
   /// always be able to get register info as well (through this method).
   ///
   const Thumb2RegisterInfo &getRegisterInfo() const { return RI; }
+
+  ScheduleHazardRecognizer *
+  CreateTargetPostRAHazardRecognizer(const InstrItineraryData &II) const;
 };
 }
 
diff --git a/test/CodeGen/ARM/ifcvt2.ll b/test/CodeGen/ARM/ifcvt2.ll
index d9cac80..7b9d0cf 100644
--- a/test/CodeGen/ARM/ifcvt2.ll
+++ b/test/CodeGen/ARM/ifcvt2.ll
@@ -1,10 +1,8 @@
-; RUN: llc < %s -march=arm > %t
-; RUN: grep bxlt %t | count 1
-; RUN: grep bxgt %t | count 1
-; RUN: not grep bxge %t
-; RUN: not grep bxle %t
+; RUN: llc < %s -march=arm | FileCheck %s
 
 define i32 @t1(i32 %a, i32 %b, i32 %c, i32 %d) {
+; CHECK: t1:
+; CHECK: bxlt lr
 	%tmp2 = icmp sgt i32 %c, 10
 	%tmp5 = icmp slt i32 %d, 4
 	%tmp8 = or i1 %tmp5, %tmp2
@@ -21,6 +19,13 @@ UnifiedReturnBlock:
 }
 
 define i32 @t2(i32 %a, i32 %b, i32 %c, i32 %d) {
+; CHECK: t2:
+; CHECK: bxgt lr
+; CHECK: cmp
+; CHECK: addge
+; CHECK: subge
+; CHECK-NOT: bxge lr
+; CHECK: bx lr
 	%tmp2 = icmp sgt i32 %c, 10
 	%tmp5 = icmp slt i32 %d, 4
 	%tmp8 = and i1 %tmp5, %tmp2
diff --git a/test/CodeGen/Thumb2/thumb2-cbnz.ll b/test/CodeGen/Thumb2/thumb2-cbnz.ll
index 4f4c9af..10a4985 100644
--- a/test/CodeGen/Thumb2/thumb2-cbnz.ll
+++ b/test/CodeGen/Thumb2/thumb2-cbnz.ll
@@ -21,7 +21,7 @@ bb7:                                              ; preds = %bb3
 
 bb9:                                              ; preds = %bb7
 ; CHECK:      cmp r0, #0
-; CHECK-NEXT: cmp r0, #0
+; CHECK:      cmp r0, #0
 ; CHECK-NEXT: cbnz
   %0 = tail call  double @floor(double %b) nounwind readnone ; <double> [#uses=0]
   br label %bb11
diff --git a/test/CodeGen/Thumb2/thumb2-ifcvt2.ll b/test/CodeGen/Thumb2/thumb2-ifcvt2.ll
index 4af492c..2c57348 100644
--- a/test/CodeGen/Thumb2/thumb2-ifcvt2.ll
+++ b/test/CodeGen/Thumb2/thumb2-ifcvt2.ll
@@ -32,6 +32,7 @@ entry:
 ; CHECK: it eq
 ; CHECK: cmpeq
 ; CHECK: bne
+; CHECK: cmp
 ; CHECK: itt eq
 ; CHECK: moveq
 ; CHECK: popeq