Merge up to r162331, git commit bc363931085587bac42a40653962a3e5acd1ffce

36 files changed, 1911 insertions, 1376 deletions

diff --git a/lib/Target/ARM/ARM.td b/lib/Target/ARM/ARM.td
index 7af8b9d909..06ed89395f 100644
--- a/lib/Target/ARM/ARM.td
+++ b/lib/Target/ARM/ARM.td
@@ -204,17 +204,17 @@ def : Processor<"arm1156t2f-s",     ARMV6Itineraries, [HasV6T2Ops, FeatureVFP2,
                                                        FeatureDSPThumb2]>;
 
 // V7a Processors.
-def : Processor<"cortex-a8",        CortexA8Itineraries,
+def : ProcessorModel<"cortex-a8",   CortexA8Model,
 // @LOCALMOD-BEGIN
 // TODO(pdox): Resolve this mismatch.
                                     [ProcA8, HasV7Ops, FeatureDB]>;
 // FeatureNEON, FeatureDSPThumb2, FeatureHasRAS]>;
 // @LOCALMOD-END
 
-def : Processor<"cortex-a9",        CortexA9Itineraries,
+def : ProcessorModel<"cortex-a9",   CortexA9Model,
                                     [ProcA9, HasV7Ops, FeatureNEON, FeatureDB,
                                      FeatureDSPThumb2, FeatureHasRAS]>;
-def : Processor<"cortex-a9-mp",     CortexA9Itineraries,
+def : ProcessorModel<"cortex-a9-mp", CortexA9Model,
                                     [ProcA9, HasV7Ops, FeatureNEON, FeatureDB,
                                      FeatureDSPThumb2, FeatureMP,
                                      FeatureHasRAS]>;
@@ -228,7 +228,7 @@ def : ProcNoItin<"cortex-m3",       [HasV7Ops,
 def : ProcNoItin<"cortex-m4",       [HasV7Ops,
                                      FeatureThumb2, FeatureNoARM, FeatureDB,
                                      FeatureHWDiv, FeatureDSPThumb2,
-                                     FeatureT2XtPk, FeatureVFP2,
+                                     FeatureT2XtPk, FeatureVFP4,
                                      FeatureVFPOnlySP, FeatureMClass]>;
 
 //===----------------------------------------------------------------------===//
diff --git a/lib/Target/ARM/ARMAsmPrinter.cpp b/lib/Target/ARM/ARMAsmPrinter.cpp
index 76cd0c389d..95caf529d6 100644
--- a/lib/Target/ARM/ARMAsmPrinter.cpp
+++ b/lib/Target/ARM/ARMAsmPrinter.cpp
@@ -622,10 +622,24 @@ bool ARMAsmPrinter::PrintAsmOperand(const MachineInstr *MI, unsigned OpNum,
       return false;
     }
 
-    // These modifiers are not yet supported.
+    // This modifier is not yet supported.
     case 'h': // A range of VFP/NEON registers suitable for VLD1/VST1.
-    case 'H': // The highest-numbered register of a pair.
       return true;
+    case 'H': { // The highest-numbered register of a pair.
+      const MachineOperand &MO = MI->getOperand(OpNum);
+      if (!MO.isReg())
+        return true;
+      const TargetRegisterClass &RC = ARM::GPRRegClass;
+      const MachineFunction &MF = *MI->getParent()->getParent();
+      const TargetRegisterInfo *TRI = MF.getTarget().getRegisterInfo();
+
+      unsigned RegIdx = TRI->getEncodingValue(MO.getReg());
+      RegIdx |= 1; //The odd register is also the higher-numbered one of a pair.
+
+      unsigned Reg = RC.getRegister(RegIdx);
+      O << ARMInstPrinter::getRegisterName(Reg);
+      return false;
+    }
     }
   }
 
@@ -1253,8 +1267,14 @@ void ARMAsmPrinter::EmitUnwindingInstruction(const MachineInstr *MI) {
       assert(SrcReg == ARM::SP &&
              "Only stack pointer as a source reg is supported");
       for (unsigned i = StartOp, NumOps = MI->getNumOperands() - NumOffset;
-           i != NumOps; ++i)
-        RegList.push_back(MI->getOperand(i).getReg());
+           i != NumOps; ++i) {
+        const MachineOperand &MO = MI->getOperand(i);
+        // Actually, there should never be any impdef stuff here. Skip it
+        // temporary to workaround PR11902.
+        if (MO.isImplicit())
+          continue;
+        RegList.push_back(MO.getReg());
+      }
       break;
     case ARM::STR_PRE_IMM:
     case ARM::STR_PRE_REG:
diff --git a/lib/Target/ARM/ARMBaseInstrInfo.cpp b/lib/Target/ARM/ARMBaseInstrInfo.cpp
index 08e55429ce..aab8825444 100644
--- a/lib/Target/ARM/ARMBaseInstrInfo.cpp
+++ b/lib/Target/ARM/ARMBaseInstrInfo.cpp
@@ -795,8 +795,28 @@ storeRegToStackSlot(MachineBasicBlock &MBB, MachineBasicBlock::iterator I,
       } else
         llvm_unreachable("Unknown reg class!");
       break;
+    case 24:
+      if (ARM::DTripleRegClass.hasSubClassEq(RC)) {
+        // Use aligned spills if the stack can be realigned.
+        if (Align >= 16 && getRegisterInfo().canRealignStack(MF)) {
+          AddDefaultPred(BuildMI(MBB, I, DL, get(ARM::VST1d64TPseudo))
+                     .addFrameIndex(FI).addImm(16)
+                     .addReg(SrcReg, getKillRegState(isKill))
+                     .addMemOperand(MMO));
+        } else {
+          MachineInstrBuilder MIB =
+          AddDefaultPred(BuildMI(MBB, I, DL, get(ARM::VSTMDIA))
+                       .addFrameIndex(FI))
+                       .addMemOperand(MMO);
+          MIB = AddDReg(MIB, SrcReg, ARM::dsub_0, getKillRegState(isKill), TRI);
+          MIB = AddDReg(MIB, SrcReg, ARM::dsub_1, 0, TRI);
+          AddDReg(MIB, SrcReg, ARM::dsub_2, 0, TRI);
+        }
+      } else
+        llvm_unreachable("Unknown reg class!");
+      break;
     case 32:
-      if (ARM::QQPRRegClass.hasSubClassEq(RC)) {
+      if (ARM::QQPRRegClass.hasSubClassEq(RC) || ARM::DQuadRegClass.hasSubClassEq(RC)) {
         if (Align >= 16 && getRegisterInfo().canRealignStack(MF)) {
           // FIXME: It's possible to only store part of the QQ register if the
           // spilled def has a sub-register index.
@@ -868,6 +888,8 @@ ARMBaseInstrInfo::isStoreToStackSlot(const MachineInstr *MI,
     }
     break;
   case ARM::VST1q64:
+  case ARM::VST1d64TPseudo:
+  case ARM::VST1d64QPseudo:
     if (MI->getOperand(0).isFI() &&
         MI->getOperand(2).getSubReg() == 0) {
       FrameIndex = MI->getOperand(0).getIndex();
@@ -942,8 +964,28 @@ loadRegFromStackSlot(MachineBasicBlock &MBB, MachineBasicBlock::iterator I,
     } else
       llvm_unreachable("Unknown reg class!");
     break;
-  case 32:
-    if (ARM::QQPRRegClass.hasSubClassEq(RC)) {
+  case 24:
+    if (ARM::DTripleRegClass.hasSubClassEq(RC)) {
+      if (Align >= 16 && getRegisterInfo().canRealignStack(MF)) {
+        AddDefaultPred(BuildMI(MBB, I, DL, get(ARM::VLD1d64TPseudo), DestReg)
+                     .addFrameIndex(FI).addImm(16)
+                     .addMemOperand(MMO));
+      } else {
+        MachineInstrBuilder MIB =
+          AddDefaultPred(BuildMI(MBB, I, DL, get(ARM::VLDMDIA))
+                         .addFrameIndex(FI)
+                         .addMemOperand(MMO));
+        MIB = AddDReg(MIB, DestReg, ARM::dsub_0, RegState::DefineNoRead, TRI);
+        MIB = AddDReg(MIB, DestReg, ARM::dsub_1, RegState::DefineNoRead, TRI);
+        MIB = AddDReg(MIB, DestReg, ARM::dsub_2, RegState::DefineNoRead, TRI);
+        if (TargetRegisterInfo::isPhysicalRegister(DestReg))
+          MIB.addReg(DestReg, RegState::ImplicitDefine);
+      }
+    } else
+      llvm_unreachable("Unknown reg class!");
+    break;
+   case 32:
+    if (ARM::QQPRRegClass.hasSubClassEq(RC) || ARM::DQuadRegClass.hasSubClassEq(RC)) {
       if (Align >= 16 && getRegisterInfo().canRealignStack(MF)) {
         AddDefaultPred(BuildMI(MBB, I, DL, get(ARM::VLD1d64QPseudo), DestReg)
                      .addFrameIndex(FI).addImm(16)
@@ -1016,6 +1058,8 @@ ARMBaseInstrInfo::isLoadFromStackSlot(const MachineInstr *MI,
     }
     break;
   case ARM::VLD1q64:
+  case ARM::VLD1d64TPseudo:
+  case ARM::VLD1d64QPseudo:
     if (MI->getOperand(1).isFI() &&
         MI->getOperand(0).getSubReg() == 0) {
       FrameIndex = MI->getOperand(1).getIndex();
@@ -1524,6 +1568,139 @@ ARMBaseInstrInfo::commuteInstruction(MachineInstr *MI, bool NewMI) const {
   return TargetInstrInfoImpl::commuteInstruction(MI, NewMI);
 }
 
+/// Identify instructions that can be folded into a MOVCC instruction, and
+/// return the corresponding opcode for the predicated pseudo-instruction.
+static unsigned canFoldIntoMOVCC(unsigned Reg, MachineInstr *&MI,
+                                 const MachineRegisterInfo &MRI) {
+  if (!TargetRegisterInfo::isVirtualRegister(Reg))
+    return 0;
+  if (!MRI.hasOneNonDBGUse(Reg))
+    return 0;
+  MI = MRI.getVRegDef(Reg);
+  if (!MI)
+    return 0;
+  // Check if MI has any non-dead defs or physreg uses. This also detects
+  // predicated instructions which will be reading CPSR.
+  for (unsigned i = 1, e = MI->getNumOperands(); i != e; ++i) {
+    const MachineOperand &MO = MI->getOperand(i);
+    // Reject frame index operands, PEI can't handle the predicated pseudos.
+    if (MO.isFI() || MO.isCPI() || MO.isJTI())
+      return 0;
+    if (!MO.isReg())
+      continue;
+    if (TargetRegisterInfo::isPhysicalRegister(MO.getReg()))
+      return 0;
+    if (MO.isDef() && !MO.isDead())
+      return 0;
+  }
+  switch (MI->getOpcode()) {
+  default: return 0;
+  case ARM::ANDri:   return ARM::ANDCCri;
+  case ARM::ANDrr:   return ARM::ANDCCrr;
+  case ARM::ANDrsi:  return ARM::ANDCCrsi;
+  case ARM::ANDrsr:  return ARM::ANDCCrsr;
+  case ARM::t2ANDri: return ARM::t2ANDCCri;
+  case ARM::t2ANDrr: return ARM::t2ANDCCrr;
+  case ARM::t2ANDrs: return ARM::t2ANDCCrs;
+  case ARM::EORri:   return ARM::EORCCri;
+  case ARM::EORrr:   return ARM::EORCCrr;
+  case ARM::EORrsi:  return ARM::EORCCrsi;
+  case ARM::EORrsr:  return ARM::EORCCrsr;
+  case ARM::t2EORri: return ARM::t2EORCCri;
+  case ARM::t2EORrr: return ARM::t2EORCCrr;
+  case ARM::t2EORrs: return ARM::t2EORCCrs;
+  case ARM::ORRri:   return ARM::ORRCCri;
+  case ARM::ORRrr:   return ARM::ORRCCrr;
+  case ARM::ORRrsi:  return ARM::ORRCCrsi;
+  case ARM::ORRrsr:  return ARM::ORRCCrsr;
+  case ARM::t2ORRri: return ARM::t2ORRCCri;
+  case ARM::t2ORRrr: return ARM::t2ORRCCrr;
+  case ARM::t2ORRrs: return ARM::t2ORRCCrs;
+
+  // ARM ADD/SUB
+  case ARM::ADDri:   return ARM::ADDCCri;
+  case ARM::ADDrr:   return ARM::ADDCCrr;
+  case ARM::ADDrsi:  return ARM::ADDCCrsi;
+  case ARM::ADDrsr:  return ARM::ADDCCrsr;
+  case ARM::SUBri:   return ARM::SUBCCri;
+  case ARM::SUBrr:   return ARM::SUBCCrr;
+  case ARM::SUBrsi:  return ARM::SUBCCrsi;
+  case ARM::SUBrsr:  return ARM::SUBCCrsr;
+
+  // Thumb2 ADD/SUB
+  case ARM::t2ADDri:   return ARM::t2ADDCCri;
+  case ARM::t2ADDri12: return ARM::t2ADDCCri12;
+  case ARM::t2ADDrr:   return ARM::t2ADDCCrr;
+  case ARM::t2ADDrs:   return ARM::t2ADDCCrs;
+  case ARM::t2SUBri:   return ARM::t2SUBCCri;
+  case ARM::t2SUBri12: return ARM::t2SUBCCri12;
+  case ARM::t2SUBrr:   return ARM::t2SUBCCrr;
+  case ARM::t2SUBrs:   return ARM::t2SUBCCrs;
+  }
+}
+
+bool ARMBaseInstrInfo::analyzeSelect(const MachineInstr *MI,
+                                     SmallVectorImpl<MachineOperand> &Cond,
+                                     unsigned &TrueOp, unsigned &FalseOp,
+                                     bool &Optimizable) const {
+  assert((MI->getOpcode() == ARM::MOVCCr || MI->getOpcode() == ARM::t2MOVCCr) &&
+         "Unknown select instruction");
+  // MOVCC operands:
+  // 0: Def.
+  // 1: True use.
+  // 2: False use.
+  // 3: Condition code.
+  // 4: CPSR use.
+  TrueOp = 1;
+  FalseOp = 2;
+  Cond.push_back(MI->getOperand(3));
+  Cond.push_back(MI->getOperand(4));
+  // We can always fold a def.
+  Optimizable = true;
+  return false;
+}
+
+MachineInstr *ARMBaseInstrInfo::optimizeSelect(MachineInstr *MI,
+                                               bool PreferFalse) const {
+  assert((MI->getOpcode() == ARM::MOVCCr || MI->getOpcode() == ARM::t2MOVCCr) &&
+         "Unknown select instruction");
+  const MachineRegisterInfo &MRI = MI->getParent()->getParent()->getRegInfo();
+  MachineInstr *DefMI = 0;
+  unsigned Opc = canFoldIntoMOVCC(MI->getOperand(2).getReg(), DefMI, MRI);
+  bool Invert = !Opc;
+  if (!Opc)
+    Opc = canFoldIntoMOVCC(MI->getOperand(1).getReg(), DefMI, MRI);
+  if (!Opc)
+    return 0;
+
+  // Create a new predicated version of DefMI.
+  // Rfalse is the first use.
+  MachineInstrBuilder NewMI = BuildMI(*MI->getParent(), MI, MI->getDebugLoc(),
+                                      get(Opc), MI->getOperand(0).getReg())
+    .addOperand(MI->getOperand(Invert ? 2 : 1));
+
+  // Copy all the DefMI operands, excluding its (null) predicate.
+  const MCInstrDesc &DefDesc = DefMI->getDesc();
+  for (unsigned i = 1, e = DefDesc.getNumOperands();
+       i != e && !DefDesc.OpInfo[i].isPredicate(); ++i)
+    NewMI.addOperand(DefMI->getOperand(i));
+
+  unsigned CondCode = MI->getOperand(3).getImm();
+  if (Invert)
+    NewMI.addImm(ARMCC::getOppositeCondition(ARMCC::CondCodes(CondCode)));
+  else
+    NewMI.addImm(CondCode);
+  NewMI.addOperand(MI->getOperand(4));
+
+  // DefMI is not the -S version that sets CPSR, so add an optional %noreg.
+  if (NewMI->hasOptionalDef())
+    AddDefaultCC(NewMI);
+
+  // The caller will erase MI, but not DefMI.
+  DefMI->eraseFromParent();
+  return NewMI;
+}
+
 /// Map pseudo instructions that imply an 'S' bit onto real opcodes. Whether the
 /// instruction is encoded with an 'S' bit is determined by the optional CPSR
 /// def operand.
@@ -1975,10 +2152,10 @@ optimizeCompareInstr(MachineInstr *CmpInstr, unsigned SrcReg, unsigned SrcReg2,
   case ARM::t2EORri: {
     // Scan forward for the use of CPSR
     // When checking against MI: if it's a conditional code requires
-    // checking of V bit, then this is not safe to do. If we can't find the
-    // CPSR use (i.e. used in another block), then it's not safe to perform
-    // the optimization.
-    // When checking against Sub, we handle the condition codes GE, LT, GT, LE.
+    // checking of V bit, then this is not safe to do.
+    // It is safe to remove CmpInstr if CPSR is redefined or killed.
+    // If we are done with the basic block, we need to check whether CPSR is
+    // live-out.
     SmallVector<std::pair<MachineOperand*, ARMCC::CondCodes>, 4>
         OperandsToUpdate;
     bool isSafe = false;
@@ -2018,7 +2195,7 @@ optimizeCompareInstr(MachineInstr *CmpInstr, unsigned SrcReg, unsigned SrcReg2,
         else
           switch (CC) {
           default:
-            isSafe = true;
+            // CPSR can be used multiple times, we should continue.
             break;
           case ARMCC::VS:
           case ARMCC::VC:
@@ -2031,10 +2208,15 @@ optimizeCompareInstr(MachineInstr *CmpInstr, unsigned SrcReg, unsigned SrcReg2,
       }
     }
 
-    // If the candidate is Sub, we may exit the loop at end of the basic block.
-    // In that case, it is still safe to remove CmpInstr.
-    if (!isSafe && !Sub)
-      return false;
+    // If CPSR is not killed nor re-defined, we should check whether it is
+    // live-out. If it is live-out, do not optimize.
+    if (!isSafe) {
+      MachineBasicBlock *MBB = CmpInstr->getParent();
+      for (MachineBasicBlock::succ_iterator SI = MBB->succ_begin(),
+               SE = MBB->succ_end(); SI != SE; ++SI)
+        if ((*SI)->isLiveIn(ARM::CPSR))
+          return false;
+    }
 
     // Toggle the optional operand to CPSR.
     MI->getOperand(5).setReg(ARM::CPSR);
@@ -2195,7 +2377,7 @@ ARMBaseInstrInfo::getNumMicroOps(const InstrItineraryData *ItinData,
   //
   // On Cortex-A8, each pair of register loads / stores can be scheduled on the
   // same cycle. The scheduling for the first load / store must be done
-  // separately by assuming the the address is not 64-bit aligned.
+  // separately by assuming the address is not 64-bit aligned.
   //
   // On Cortex-A9, the formula is simply (#reg / 2) + (#reg % 2). If the address
   // is not 64-bit aligned, then AGU would take an extra cycle.  For VFP / NEON
@@ -3177,11 +3359,18 @@ enum ARMExeDomain {
 //
 std::pair<uint16_t, uint16_t>
 ARMBaseInstrInfo::getExecutionDomain(const MachineInstr *MI) const {
-  // VMOVD is a VFP instruction, but can be changed to NEON if it isn't
-  // predicated.
+  // VMOVD, VMOVRS and VMOVSR are VFP instructions, but can be changed to NEON
+  // if they are not predicated.
   if (MI->getOpcode() == ARM::VMOVD && !isPredicated(MI))
     return std::make_pair(ExeVFP, (1<<ExeVFP) | (1<<ExeNEON));
 
+  // Cortex-A9 is particularly picky about mixing the two and wants these
+  // converted.
+  if (Subtarget.isCortexA9() && !isPredicated(MI) &&
+      (MI->getOpcode() == ARM::VMOVRS ||
+       MI->getOpcode() == ARM::VMOVSR))
+    return std::make_pair(ExeVFP, (1<<ExeVFP) | (1<<ExeNEON));
+
   // No other instructions can be swizzled, so just determine their domain.
   unsigned Domain = MI->getDesc().TSFlags & ARMII::DomainMask;
 
@@ -3201,22 +3390,95 @@ ARMBaseInstrInfo::getExecutionDomain(const MachineInstr *MI) const {
 
 void
 ARMBaseInstrInfo::setExecutionDomain(MachineInstr *MI, unsigned Domain) const {
-  // We only know how to change VMOVD into VORR.
-  assert(MI->getOpcode() == ARM::VMOVD && "Can only swizzle VMOVD");
-  if (Domain != ExeNEON)
-    return;
+  unsigned DstReg, SrcReg, DReg;
+  unsigned Lane;
+  MachineInstrBuilder MIB(MI);
+  const TargetRegisterInfo *TRI = &getRegisterInfo();
+  bool isKill;
+  switch (MI->getOpcode()) {
+    default:
+      llvm_unreachable("cannot handle opcode!");
+      break;
+    case ARM::VMOVD:
+      if (Domain != ExeNEON)
+        break;
 
-  // Zap the predicate operands.
-  assert(!isPredicated(MI) && "Cannot predicate a VORRd");
-  MI->RemoveOperand(3);
-  MI->RemoveOperand(2);
+      // Zap the predicate operands.
+      assert(!isPredicated(MI) && "Cannot predicate a VORRd");
+      MI->RemoveOperand(3);
+      MI->RemoveOperand(2);
 
-  // Change to a VORRd which requires two identical use operands.
-  MI->setDesc(get(ARM::VORRd));
+      // Change to a VORRd which requires two identical use operands.
+      MI->setDesc(get(ARM::VORRd));
+
+      // Add the extra source operand and new predicates.
+      // This will go before any implicit ops.
+      AddDefaultPred(MachineInstrBuilder(MI).addOperand(MI->getOperand(1)));
+      break;
+    case ARM::VMOVRS:
+      if (Domain != ExeNEON)
+        break;
+      assert(!isPredicated(MI) && "Cannot predicate a VGETLN");
+
+      DstReg = MI->getOperand(0).getReg();
+      SrcReg = MI->getOperand(1).getReg();
+
+      DReg = TRI->getMatchingSuperReg(SrcReg, ARM::ssub_0, &ARM::DPRRegClass);
+      Lane = 0;
+      if (DReg == ARM::NoRegister) {
+        DReg = TRI->getMatchingSuperReg(SrcReg, ARM::ssub_1, &ARM::DPRRegClass);
+        Lane = 1;
+        assert(DReg && "S-register with no D super-register?");
+      }
+
+      MI->RemoveOperand(3);
+      MI->RemoveOperand(2);
+      MI->RemoveOperand(1);
+
+      MI->setDesc(get(ARM::VGETLNi32));
+      MIB.addReg(DReg);
+      MIB.addImm(Lane);
+
+      MIB->getOperand(1).setIsUndef();
+      MIB.addReg(SrcReg, RegState::Implicit);
+
+      AddDefaultPred(MIB);
+      break;
+    case ARM::VMOVSR:
+      if (Domain != ExeNEON)
+        break;
+      assert(!isPredicated(MI) && "Cannot predicate a VSETLN");
+
+      DstReg = MI->getOperand(0).getReg();
+      SrcReg = MI->getOperand(1).getReg();
+      DReg = TRI->getMatchingSuperReg(DstReg, ARM::ssub_0, &ARM::DPRRegClass);
+      Lane = 0;
+      if (DReg == ARM::NoRegister) {
+        DReg = TRI->getMatchingSuperReg(DstReg, ARM::ssub_1, &ARM::DPRRegClass);
+        Lane = 1;
+        assert(DReg && "S-register with no D super-register?");
+      }
+      isKill = MI->getOperand(0).isKill();
+
+      MI->RemoveOperand(3);
+      MI->RemoveOperand(2);
+      MI->RemoveOperand(1);
+      MI->RemoveOperand(0);
+
+      MI->setDesc(get(ARM::VSETLNi32));
+      MIB.addReg(DReg, RegState::Define);
+      MIB.addReg(DReg, RegState::Undef);
+      MIB.addReg(SrcReg);
+      MIB.addImm(Lane);
+
+      if (isKill)
+        MIB->addRegisterKilled(DstReg, TRI, true);
+      MIB->addRegisterDefined(DstReg, TRI);
+
+      AddDefaultPred(MIB);
+      break;
+  }
 
-  // Add the extra source operand and new predicates.
-  // This will go before any implicit ops.
-  AddDefaultPred(MachineInstrBuilder(MI).addOperand(MI->getOperand(1)));
 }
 
 bool ARMBaseInstrInfo::hasNOP() const {
diff --git a/lib/Target/ARM/ARMBaseInstrInfo.h b/lib/Target/ARM/ARMBaseInstrInfo.h
index 1a10a4ab1c..92e5ee8dcb 100644
--- a/lib/Target/ARM/ARMBaseInstrInfo.h
+++ b/lib/Target/ARM/ARMBaseInstrInfo.h
@@ -202,6 +202,13 @@ public:
                                     unsigned SrcReg2, int CmpMask, int CmpValue,
                                     const MachineRegisterInfo *MRI) const;
 
+  virtual bool analyzeSelect(const MachineInstr *MI,
+                             SmallVectorImpl<MachineOperand> &Cond,
+                             unsigned &TrueOp, unsigned &FalseOp,
+                             bool &Optimizable) const;
+
+  virtual MachineInstr *optimizeSelect(MachineInstr *MI, bool) const;
+
   /// FoldImmediate - 'Reg' is known to be defined by a move immediate
   /// instruction, try to fold the immediate into the use instruction.
   virtual bool FoldImmediate(MachineInstr *UseMI, MachineInstr *DefMI,
@@ -352,6 +359,11 @@ ARMCC::CondCodes getInstrPredicate(const MachineInstr *MI, unsigned &PredReg);
 
 int getMatchingCondBranchOpcode(int Opc);
 
+/// Determine if MI can be folded into an ARM MOVCC instruction, and return the
+/// opcode of the SSA instruction representing the conditional MI.
+unsigned canFoldARMInstrIntoMOVCC(unsigned Reg,
+                                  MachineInstr *&MI,
+                                  const MachineRegisterInfo &MRI);
 
 /// Map pseudo instructions that imply an 'S' bit onto real opcodes. Whether
 /// the instruction is encoded with an 'S' bit is determined by the optional
diff --git a/lib/Target/ARM/ARMBaseRegisterInfo.cpp b/lib/Target/ARM/ARMBaseRegisterInfo.cpp
index 533ed8834b..b2ba9703b6 100644
--- a/lib/Target/ARM/ARMBaseRegisterInfo.cpp
+++ b/lib/Target/ARM/ARMBaseRegisterInfo.cpp
@@ -65,8 +65,20 @@ extern cl::opt<bool> ReserveR9; // @LOCALMOD
 const uint16_t*
 ARMBaseRegisterInfo::getCalleeSavedRegs(const MachineFunction *MF) const {
   if (ReserveR9) return CSR_NaCl_SaveList; // @LOCALMOD
+  bool ghcCall = false;
+ 
+  if (MF) {
+    const Function *F = MF->getFunction();
+    ghcCall = (F ? F->getCallingConv() == CallingConv::GHC : false);
+  }
+ 
+  if (ghcCall) {
+      return CSR_GHC_SaveList;
+  }
+  else {
   return (STI.isTargetIOS() && !STI.isAAPCS_ABI())
     ? CSR_iOS_SaveList : CSR_AAPCS_SaveList;
+  }
 }
 
 const uint32_t*
diff --git a/lib/Target/ARM/ARMCallingConv.td b/lib/Target/ARM/ARMCallingConv.td
index 1a79d95f95..e0f3a871ba 100644
--- a/lib/Target/ARM/ARMCallingConv.td
+++ b/lib/Target/ARM/ARMCallingConv.td
@@ -79,6 +79,25 @@ def RetFastCC_ARM_APCS : CallingConv<[
   CCDelegateTo<RetCC_ARM_APCS>
 ]>;
 
+//===----------------------------------------------------------------------===//
+// ARM APCS Calling Convention for GHC
+//===----------------------------------------------------------------------===//
+
+def CC_ARM_APCS_GHC : CallingConv<[
+  // Handle all vector types as either f64 or v2f64.
+  CCIfType<[v1i64, v2i32, v4i16, v8i8, v2f32], CCBitConvertToType<f64>>,
+  CCIfType<[v2i64, v4i32, v8i16, v16i8, v4f32], CCBitConvertToType<v2f64>>,
+
+  CCIfType<[v2f64], CCAssignToReg<[Q4, Q5]>>,
+  CCIfType<[f64], CCAssignToReg<[D8, D9, D10, D11]>>,
+  CCIfType<[f32], CCAssignToReg<[S16, S17, S18, S19, S20, S21, S22, S23]>>,
+
+  // Promote i8/i16 arguments to i32.
+  CCIfType<[i8, i16], CCPromoteToType<i32>>,
+
+  // Pass in STG registers: Base, Sp, Hp, R1, R2, R3, R4, SpLim
+  CCIfType<[i32], CCAssignToReg<[R4, R5, R6, R7, R8, R9, R10, R11]>>
+]>;
 
 //===----------------------------------------------------------------------===//
 // ARM AAPCS (EABI) Calling Convention, common parts
@@ -117,6 +136,9 @@ def RetCC_ARM_AAPCS_Common : CallingConv<[
 //===----------------------------------------------------------------------===//
 
 def CC_ARM_AAPCS : CallingConv<[
+  // Handles byval parameters.
+  CCIfByVal<CCPassByVal<4, 4>>,
+
   // Handle all vector types as either f64 or v2f64.
   CCIfType<[v1i64, v2i32, v4i16, v8i8, v2f32], CCBitConvertToType<f64>>,
   CCIfType<[v2i64, v4i32, v8i16, v16i8, v4f32], CCBitConvertToType<v2f64>>,
@@ -142,6 +164,9 @@ def RetCC_ARM_AAPCS : CallingConv<[
 //===----------------------------------------------------------------------===//
 
 def CC_ARM_AAPCS_VFP : CallingConv<[
+  // Handles byval parameters.
+  CCIfByVal<CCPassByVal<4, 4>>,
+
   // Handle all vector types as either f64 or v2f64.
   CCIfType<[v1i64, v2i32, v4i16, v8i8, v2f32], CCBitConvertToType<f64>>,
   CCIfType<[v2i64, v4i32, v8i16, v16i8, v4f32], CCBitConvertToType<v2f64>>,
@@ -176,6 +201,12 @@ def CSR_AAPCS : CalleeSavedRegs<(add LR, R11, R10, R9, R8, R7, R6, R5, R4,
 // Also save R7-R4 first to match the stack frame fixed spill areas.
 def CSR_iOS : CalleeSavedRegs<(add LR, R7, R6, R5, R4, (sub CSR_AAPCS, R9))>;
 
+// GHC set of callee saved regs is empty as all those regs are
+// used for passing STG regs around
+// add is a workaround for not being able to compile empty list:
+// def CSR_GHC : CalleeSavedRegs<()>;
+def CSR_GHC : CalleeSavedRegs<(add)>;
+
 // @LOCALMOD-START
 // NaCl does not save R9, but otherwise uses the same order as AAPCS
 def CSR_NaCl : CalleeSavedRegs<(add LR, R11, R10, R8, R7, R6, R5, R4,
diff --git a/lib/Target/ARM/ARMCodeEmitter.cpp b/lib/Target/ARM/ARMCodeEmitter.cpp
index fc445a9339..e81b4cc282 100644
--- a/lib/Target/ARM/ARMCodeEmitter.cpp
+++ b/lib/Target/ARM/ARMCodeEmitter.cpp
@@ -254,7 +254,7 @@ namespace {
         emitConstPoolAddress(MO.getIndex(), ARM::reloc_arm_cp_entry);
         return 0;
       }
-      unsigned Reg = getARMRegisterNumbering(MO.getReg());
+      unsigned Reg = II->getRegisterInfo().getEncodingValue(MO.getReg());
       int32_t Imm12 = MO1.getImm();
       uint32_t Binary;
       Binary = Imm12 & 0xfff;
@@ -296,7 +296,7 @@ namespace {
         emitConstPoolAddress(MO.getIndex(), ARM::reloc_arm_cp_entry);
         return 0;
       }
-      unsigned Reg = getARMRegisterNumbering(MO.getReg());
+      unsigned Reg = II->getRegisterInfo().getEncodingValue(MO.getReg());
       int32_t Imm12 = MO1.getImm();
 
       // Special value for #-0
@@ -352,6 +352,12 @@ namespace {
     void emitJumpTableAddress(unsigned JTIndex, unsigned Reloc) const;
     void emitMachineBasicBlock(MachineBasicBlock *BB, unsigned Reloc,
                                intptr_