am 2d4629c5: Merge branch \'upstream\' into merge_2013_03_18

* commit '2d4629c5d7dcc6582fa7b85a517744f1a3654eba': (424 commits)
  Change NULL to 0.
  Register the flush function for each compile unit.
  Remove trailing spaces.
  Fix PPC unaligned 64-bit loads and stores
  ARM cost model: Make some vector integer to float casts cheaper
  ARM cost model: Correct cost for some cheap float to integer conversions
  Extend global merge pass to optionally consider global constant variables. Also add some checks to not merge globals used within landing pad instructions or marked as "used".
  Change test cases to handle unaligned references.
  Remove unnecessary leading comment characters in lit-only file
  Add SchedRW annotations to most of X86InstrSSE.td.
  Annotate X86 arithmetic instructions with SchedRW lists.
  Check whether a pointer is non-null (isKnownNonNull) in isKnownNonZero.
  TableGen fix for the new machine model.
  Include '.test' suffix in target specific lit configs that need it
  Make the fields in the diagram match the descriptive text above them.
  Update
  Fix 80-col. violations in PPCCTRLoops
  Fix large count and negative constant count handling in PPCCTRLoops
  Cleanup initial-value constants in PPCCTRLoops
  Fix integer comparison in DIEInteger::BestForm.
  ...

12 files changed, 1013 insertions, 203 deletions

diff --git a/lib/Target/ARM/A15SDOptimizer.cpp b/lib/Target/ARM/A15SDOptimizer.cpp
new file mode 100644
index 0000000..f0d4dbe
--- /dev/null
+++ b/lib/Target/ARM/A15SDOptimizer.cpp
@@ -0,0 +1,704 @@
+//=== A15SDOptimizerPass.cpp - Optimize DPR and SPR register accesses on A15==//
+//
+//                     The LLVM Compiler Infrastructure
+//
+// This file is distributed under the University of Illinois Open Source
+// License. See LICENSE.TXT for details.
+//
+//===----------------------------------------------------------------------===//
+//
+// The Cortex-A15 processor employs a tracking scheme in its register renaming
+// in order to process each instruction's micro-ops speculatively and
+// out-of-order with appropriate forwarding. The ARM architecture allows VFP
+// instructions to read and write 32-bit S-registers.  Each S-register
+// corresponds to one half (upper or lower) of an overlaid 64-bit D-register.
+//
+// There are several instruction patterns which can be used to provide this
+// capability which can provide higher performance than other, potentially more
+// direct patterns, specifically around when one micro-op reads a D-register
+// operand that has recently been written as one or more S-register results.
+//
+// This file defines a pre-regalloc pass which looks for SPR producers which
+// are going to be used by a DPR (or QPR) consumers and creates the more
+// optimized access pattern.
+//
+//===----------------------------------------------------------------------===//
+
+#define DEBUG_TYPE "a15-sd-optimizer"
+#include "ARM.h"
+#include "ARMBaseInstrInfo.h"
+#include "ARMSubtarget.h"
+#include "ARMISelLowering.h"
+#include "ARMTargetMachine.h"
+
+#include "llvm/ADT/SmallPtrSet.h"
+#include "llvm/ADT/Statistic.h"
+#include "llvm/CodeGen/MachineFunctionPass.h"
+#include "llvm/CodeGen/MachineInstr.h"
+#include "llvm/CodeGen/MachineInstrBuilder.h"
+#include "llvm/CodeGen/MachineRegisterInfo.h"
+#include "llvm/Support/CommandLine.h"
+#include "llvm/Support/Debug.h"
+#include "llvm/Support/raw_ostream.h"
+#include "llvm/Target/TargetRegisterInfo.h"
+
+#include <set>
+
+using namespace llvm;
+
+namespace {
+  struct A15SDOptimizer : public MachineFunctionPass {
+    static char ID;
+    A15SDOptimizer() : MachineFunctionPass(ID) {}
+
+    virtual bool runOnMachineFunction(MachineFunction &Fn);
+
+    virtual const char *getPassName() const {
+      return "ARM A15 S->D optimizer";
+    }
+
+  private:
+    const ARMBaseInstrInfo *TII;
+    const TargetRegisterInfo *TRI;
+    MachineRegisterInfo *MRI;
+
+    bool runOnInstruction(MachineInstr *MI);
+
+    //
+    // Instruction builder helpers
+    //
+    unsigned createDupLane(MachineBasicBlock &MBB,
+                           MachineBasicBlock::iterator InsertBefore,
+                           DebugLoc DL,
+                           unsigned Reg, unsigned Lane,
+                           bool QPR=false);
+
+    unsigned createExtractSubreg(MachineBasicBlock &MBB,
+                                 MachineBasicBlock::iterator InsertBefore,
+                                 DebugLoc DL,
+                                 unsigned DReg, unsigned Lane,
+                                 const TargetRegisterClass *TRC);
+
+    unsigned createVExt(MachineBasicBlock &MBB,
+                        MachineBasicBlock::iterator InsertBefore,
+                        DebugLoc DL,
+                        unsigned Ssub0, unsigned Ssub1);
+
+    unsigned createRegSequence(MachineBasicBlock &MBB,
+                               MachineBasicBlock::iterator InsertBefore,
+                               DebugLoc DL,
+                               unsigned Reg1, unsigned Reg2);
+
+    unsigned createInsertSubreg(MachineBasicBlock &MBB,
+                                MachineBasicBlock::iterator InsertBefore,
+                                DebugLoc DL, unsigned DReg, unsigned Lane,
+                                unsigned ToInsert);
+
+    unsigned createImplicitDef(MachineBasicBlock &MBB,
+                               MachineBasicBlock::iterator InsertBefore,
+                               DebugLoc DL);
+    
+    //
+    // Various property checkers
+    //
+    bool usesRegClass(MachineOperand &MO, const TargetRegisterClass *TRC);
+    bool hasPartialWrite(MachineInstr *MI);
+    SmallVector<unsigned, 8> getReadDPRs(MachineInstr *MI);
+    unsigned getDPRLaneFromSPR(unsigned SReg);
+
+    //
+    // Methods used for getting the definitions of partial registers
+    //
+
+    MachineInstr *elideCopies(MachineInstr *MI);
+    void elideCopiesAndPHIs(MachineInstr *MI,
+                            SmallVectorImpl<MachineInstr*> &Outs);
+
+    //
+    // Pattern optimization methods
+    //
+    unsigned optimizeAllLanesPattern(MachineInstr *MI, unsigned Reg);
+    unsigned optimizeSDPattern(MachineInstr *MI);
+    unsigned getPrefSPRLane(unsigned SReg);
+
+    //
+    // Sanitizing method - used to make sure if don't leave dead code around.
+    //
+    void eraseInstrWithNoUses(MachineInstr *MI);
+
+    //
+    // A map used to track the changes done by this pass.
+    //
+    std::map<MachineInstr*, unsigned> Replacements;
+    std::set<MachineInstr *> DeadInstr;
+  };
+  char A15SDOptimizer::ID = 0;
+} // end anonymous namespace
+
+// Returns true if this is a use of a SPR register.
+bool A15SDOptimizer::usesRegClass(MachineOperand &MO,
+                                  const TargetRegisterClass *TRC) {
+  if (!MO.isReg())
+    return false;
+  unsigned Reg = MO.getReg();
+
+  if (TargetRegisterInfo::isVirtualRegister(Reg))
+    return MRI->getRegClass(Reg)->hasSuperClassEq(TRC);
+  else
+    return TRC->contains(Reg);
+}
+
+unsigned A15SDOptimizer::getDPRLaneFromSPR(unsigned SReg) {
+  unsigned DReg = TRI->getMatchingSuperReg(SReg, ARM::ssub_1,
+                                           &ARM::DPRRegClass);
+  if (DReg != ARM::NoRegister) return ARM::ssub_1;
+  return ARM::ssub_0;
+}
+
+// Get the subreg type that is most likely to be coalesced
+// for an SPR register that will be used in VDUP32d pseudo.
+unsigned A15SDOptimizer::getPrefSPRLane(unsigned SReg) {
+  if (!TRI->isVirtualRegister(SReg))
+    return getDPRLaneFromSPR(SReg);
+
+  MachineInstr *MI = MRI->getVRegDef(SReg);
+  if (!MI) return ARM::ssub_0;
+  MachineOperand *MO = MI->findRegisterDefOperand(SReg);
+
+  assert(MO->isReg() && "Non register operand found!");
+  if (!MO) return ARM::ssub_0;
+
+  if (MI->isCopy() && usesRegClass(MI->getOperand(1),
+                                    &ARM::SPRRegClass)) {
+    SReg = MI->getOperand(1).getReg();
+  }
+
+  if (TargetRegisterInfo::isVirtualRegister(SReg)) {
+    if (MO->getSubReg() == ARM::ssub_1) return ARM::ssub_1;
+    return ARM::ssub_0;
+  }
+  return getDPRLaneFromSPR(SReg);
+}
+
+// MI is known to be dead. Figure out what instructions
+// are also made dead by this and mark them for removal.
+void A15SDOptimizer::eraseInstrWithNoUses(MachineInstr *MI) {
+  SmallVector<MachineInstr *, 8> Front;
+  DeadInstr.insert(MI);
+
+  DEBUG(dbgs() << "Deleting base instruction " << *MI << "\n");
+  Front.push_back(MI);
+
+  while (Front.size() != 0) {
+    MI = Front.back();
+    Front.pop_back();
+
+    // MI is already known to be dead. We need to see
+    // if other instructions can also be removed.
+    for (unsigned int i = 0; i < MI->getNumOperands(); ++i) {
+      MachineOperand &MO = MI->getOperand(i);
+      if ((!MO.isReg()) || (!MO.isUse()))
+        continue;
+      unsigned Reg = MO.getReg();
+      if (!TRI->isVirtualRegister(Reg))
+        continue;
+      MachineOperand *Op = MI->findRegisterDefOperand(Reg);
+
+      if (!Op)
+        continue;
+
+      MachineInstr *Def = Op->getParent();
+
+      // We don't need to do anything if we have already marked
+      // this instruction as being dead.
+      if (DeadInstr.find(Def) != DeadInstr.end())
+        continue;
+
+      // Check if all the uses of this instruction are marked as
+      // dead. If so, we can also mark this instruction as being
+      // dead.
+      bool IsDead = true;
+      for (unsigned int j = 0; j < Def->getNumOperands(); ++j) {
+        MachineOperand &MODef = Def->getOperand(j);
+        if ((!MODef.isReg()) || (!MODef.isDef()))
+          continue;
+        unsigned DefReg = MODef.getReg();
+        if (!TRI->isVirtualRegister(DefReg)) {
+          IsDead = false;
+          break;
+        }
+        for (MachineRegisterInfo::use_iterator II = MRI->use_begin(Reg),
+                            EE = MRI->use_end();
+                            II != EE; ++II) {
+          // We don't care about self references.
+          if (&*II == Def)
+            continue;
+          if (DeadInstr.find(&*II) == DeadInstr.end()) {
+            IsDead = false;
+            break;
+          }
+        }
+      }
+
+      if (!IsDead) continue;
+
+      DEBUG(dbgs() << "Deleting instruction " << *Def << "\n");
+      DeadInstr.insert(Def);
+    }
+  }
+}
+
+// Creates the more optimized patterns and generally does all the code
+// transformations in this pass.
+unsigned A15SDOptimizer::optimizeSDPattern(MachineInstr *MI) {
+  if (MI->isCopy()) {
+    return optimizeAllLanesPattern(MI, MI->getOperand(1).getReg());
+  }
+
+  if (MI->isInsertSubreg()) {
+    unsigned DPRReg = MI->getOperand(1).getReg();
+    unsigned SPRReg = MI->getOperand(2).getReg();
+
+    if (TRI->isVirtualRegister(DPRReg) && TRI->isVirtualRegister(SPRReg)) {
+      MachineInstr *DPRMI = MRI->getVRegDef(MI->getOperand(1).getReg());
+      MachineInstr *SPRMI = MRI->getVRegDef(MI->getOperand(2).getReg());
+
+      if (DPRMI && SPRMI) {
+        // See if the first operand of this insert_subreg is IMPLICIT_DEF
+        MachineInstr *ECDef = elideCopies(DPRMI);
+        if (ECDef != 0 && ECDef->isImplicitDef()) {
+          // Another corner case - if we're inserting something that is purely
+          // a subreg copy of a DPR, just use that DPR.
+
+          MachineInstr *EC = elideCopies(SPRMI);
+          // Is it a subreg copy of ssub_0?
+          if (EC && EC->isCopy() &&
+              EC->getOperand(1).getSubReg() == ARM::ssub_0) {
+            DEBUG(dbgs() << "Found a subreg copy: " << *SPRMI);
+
+            // Find the thing we're subreg copying out of - is it of the same
+            // regclass as DPRMI? (i.e. a DPR or QPR).
+            unsigned FullReg = SPRMI->getOperand(1).getReg();
+            const TargetRegisterClass *TRC =
+              MRI->getRegClass(MI->getOperand(1).getReg());
+            if (TRC->hasSuperClassEq(MRI->getRegClass(FullReg))) {
+              DEBUG(dbgs() << "Subreg copy is compatible - returning ");
+              DEBUG(dbgs() << PrintReg(FullReg) << "\n");
+              eraseInstrWithNoUses(MI);
+              return FullReg;
+            }
+          }
+
+          return optimizeAllLanesPattern(MI, MI->getOperand(2).getReg());
+        }
+      }
+    }
+    return optimizeAllLanesPattern(MI, MI->getOperand(0).getReg());
+  }
+
+  if (MI->isRegSequence() && usesRegClass(MI->getOperand(1),
+                                          &ARM::SPRRegClass)) {
+    // See if all bar one of the operands are IMPLICIT_DEF and insert the
+    // optimizer pattern accordingly.
+    unsigned NumImplicit = 0, NumTotal = 0;
+    unsigned NonImplicitReg = ~0U;
+
+    for (unsigned I = 1; I < MI->getNumExplicitOperands(); ++I) {
+      if (!MI->getOperand(I).isReg())
+        continue;
+      ++NumTotal;
+      unsigned OpReg = MI->getOperand(I).getReg();
+
+      if (!TRI->isVirtualRegister(OpReg))
+        break;
+
+      MachineInstr *Def = MRI->getVRegDef(OpReg);
+      if (!Def)
+        break;
+      if (Def->isImplicitDef())
+        ++NumImplicit;
+      else
+        NonImplicitReg = MI->getOperand(I).getReg();
+    }
+
+    if (NumImplicit == NumTotal - 1)
+      return optimizeAllLanesPattern(MI, NonImplicitReg);
+    else
+      return optimizeAllLanesPattern(MI, MI->getOperand(0).getReg());
+  }
+
+  assert(0 && "Unhandled update pattern!");
+  return 0;
+}
+
+// Return true if this MachineInstr inserts a scalar (SPR) value into
+// a D or Q register.
+bool A15SDOptimizer::hasPartialWrite(MachineInstr *MI) {
+  // The only way we can do a partial register update is through a COPY,
+  // INSERT_SUBREG or REG_SEQUENCE.
+  if (MI->isCopy() && usesRegClass(MI->getOperand(1), &ARM::SPRRegClass))
+    return true;
+
+  if (MI->isInsertSubreg() && usesRegClass(MI->getOperand(2),
+                                           &ARM::SPRRegClass))
+    return true;
+
+  if (MI->isRegSequence() && usesRegClass(MI->getOperand(1), &ARM::SPRRegClass))
+    return true;
+
+  return false;
+}
+
+// Looks through full copies to get the instruction that defines the input
+// operand for MI.
+MachineInstr *A15SDOptimizer::elideCopies(MachineInstr *MI) {
+  if (!MI->isFullCopy())
+    return MI;
+  if (!TRI->isVirtualRegister(MI->getOperand(1).getReg()))
+    return NULL;
+  MachineInstr *Def = MRI->getVRegDef(MI->getOperand(1).getReg());
+  if (!Def)
+    return NULL;
+  return elideCopies(Def);
+}
+
+// Look through full copies and PHIs to get the set of non-copy MachineInstrs
+// that can produce MI.
+void A15SDOptimizer::elideCopiesAndPHIs(MachineInstr *MI,
+                                        SmallVectorImpl<MachineInstr*> &Outs) {
+   // Looking through PHIs may create loops so we need to track what
+   // instructions we have visited before.
+   std::set<MachineInstr *> Reached;
+   SmallVector<MachineInstr *, 8> Front;
+   Front.push_back(MI);
+   while (Front.size() != 0) {
+     MI = Front.back();
+     Front.pop_back();
+
+     // If we have already explored this MachineInstr, ignore it.
+     if (Reached.find(MI) != Reached.end())
+       continue;
+     Reached.insert(MI);
+     if (MI->isPHI()) {
+       for (unsigned I = 1, E = MI->getNumOperands(); I != E; I += 2) {
+         unsigned Reg = MI->getOperand(I).getReg();
+         if (!TRI->isVirtualRegister(Reg)) {
+           continue;
+         }
+         MachineInstr *NewMI = MRI->getVRegDef(Reg);
+         if (!NewMI)
+           continue;
+         Front.push_back(NewMI);
+       }
+     } else if (MI->isFullCopy()) {
+       if (!TRI->isVirtualRegister(MI->getOperand(1).getReg()))
+         continue;
+       MachineInstr *NewMI = MRI->getVRegDef(MI->getOperand(1).getReg());
+       if (!NewMI)
+         continue;
+       Front.push_back(NewMI);
+     } else {
+       DEBUG(dbgs() << "Found partial copy" << *MI <<"\n");
+       Outs.push_back(MI);
+     }
+   }
+}
+
+// Return the DPR virtual registers that are read by this machine instruction
+// (if any).
+SmallVector<unsigned, 8> A15SDOptimizer::getReadDPRs(MachineInstr *MI) {
+  if (MI->isCopyLike() || MI->isInsertSubreg() || MI->isRegSequence() ||
+      MI->isKill())
+    return SmallVector<unsigned, 8>();
+
+  SmallVector<unsigned, 8> Defs;
+  for (unsigned i = 0; i < MI->getNumOperands(); ++i) {
+    MachineOperand &MO = MI->getOperand(i);
+
+    if (!MO.isReg() || !MO.isUse())
+      continue;
+    if (!usesRegClass(MO, &ARM::DPRRegClass) &&
+        !usesRegClass(MO, &ARM::QPRRegClass))
+      continue;
+
+    Defs.push_back(MO.getReg());
+  }
+  return Defs;
+}
+
+// Creates a DPR register from an SPR one by using a VDUP.
+unsigned
+A15SDOptimizer::createDupLane(MachineBasicBlock &MBB,
+                              MachineBasicBlock::iterator InsertBefore,
+                              DebugLoc DL,
+                              unsigned Reg, unsigned Lane, bool QPR) {
+  unsigned Out = MRI->createVirtualRegister(QPR ? &ARM::QPRRegClass :
+                                                  &ARM::DPRRegClass);
+  AddDefaultPred(BuildMI(MBB,
+                         InsertBefore,
+                         DL,
+                         TII->get(QPR ? ARM::VDUPLN32q : ARM::VDUPLN32d),
+                         Out)
+                   .addReg(Reg)
+                   .addImm(Lane));
+ 
+  return Out;
+}
+
+// Creates a SPR register from a DPR by copying the value in lane 0.
+unsigned
+A15SDOptimizer::createExtractSubreg(MachineBasicBlock &MBB,
+                                    MachineBasicBlock::iterator InsertBefore,
+                                    DebugLoc DL,
+                                    unsigned DReg, unsigned Lane,
+                                    const TargetRegisterClass *TRC) {
+  unsigned Out = MRI->createVirtualRegister(TRC);
+  BuildMI(MBB,
+          InsertBefore,
+          DL,
+          TII->get(TargetOpcode::COPY), Out)
+    .addReg(DReg, 0, Lane);
+
+  return Out;
+}
+
+// Takes two SPR registers and creates a DPR by using a REG_SEQUENCE.
+unsigned
+A15SDOptimizer::createRegSequence(MachineBasicBlock &MBB,
+                                  MachineBasicBlock::iterator InsertBefore,
+                                  DebugLoc DL,
+                                  unsigned Reg1, unsigned Reg2) {
+  unsigned Out = MRI->createVirtualRegister(&ARM::QPRRegClass);
+  BuildMI(MBB,
+          InsertBefore,
+          DL,
+          TII->get(TargetOpcode::REG_SEQUENCE), Out)
+    .addReg(Reg1)
+    .addImm(ARM::dsub_0)
+    .addReg(Reg2)
+    .addImm(ARM::dsub_1);
+  return Out;
+}
+
+// Takes two DPR registers that have previously been VDUPed (Ssub0 and Ssub1)
+// and merges them into one DPR register.
+unsigned
+A15SDOptimizer::createVExt(MachineBasicBlock &MBB,
+                           MachineBasicBlock::iterator InsertBefore,
+                           DebugLoc DL,
+                           unsigned Ssub0, unsigned Ssub1) {
+  unsigned Out = MRI->createVirtualRegister(&ARM::DPRRegClass);
+  AddDefaultPred(BuildMI(MBB,
+                         InsertBefore,
+                         DL,
+                         TII->get(ARM::VEXTd32), Out)
+                   .addReg(Ssub0)
+                   .addReg(Ssub1)
+                   .addImm(1));
+  return Out;
+}
+
+unsigned
+A15SDOptimizer::createInsertSubreg(MachineBasicBlock &MBB,
+                                   MachineBasicBlock::iterator InsertBefore,
+                                   DebugLoc DL, unsigned DReg, unsigned Lane,
+                                   unsigned ToInsert) {
+  unsigned Out = MRI->createVirtualRegister(&ARM::DPR_VFP2RegClass);
+  BuildMI(MBB,
+          InsertBefore,
+          DL,
+          TII->get(TargetOpcode::INSERT_SUBREG), Out)
+    .addReg(DReg)
+    .addReg(ToInsert)
+    .addImm(Lane);
+
+  return Out;
+}
+
+unsigned
+A15SDOptimizer::createImplicitDef(MachineBasicBlock &MBB,
+                                  MachineBasicBlock::iterator InsertBefore,
+                                  DebugLoc DL) {
+  unsigned Out = MRI->createVirtualRegister(&ARM::DPRRegClass);
+  BuildMI(MBB,
+          InsertBefore,
+          DL,
+          TII->get(TargetOpcode::IMPLICIT_DEF), Out);
+  return Out;
+}
+
+// This function inserts instructions in order to optimize interactions between
+// SPR registers and DPR/QPR registers. It does so by performing VDUPs on all
+// lanes, and the using VEXT instructions to recompose the result.
+unsigned
+A15SDOptimizer::optimizeAllLanesPattern(MachineInstr *MI, unsigned Reg) {
+  MachineBasicBlock::iterator InsertPt(MI);
+  DebugLoc DL = MI->getDebugLoc();
+  MachineBasicBlock &MBB = *MI->getParent();
+  InsertPt++;
+  unsigned Out;
+
+  if (MRI->getRegClass(Reg)->hasSuperClassEq(&ARM::QPRRegClass)) {
+    unsigned DSub0 = createExtractSubreg(MBB, InsertPt, DL, Reg,
+                                         ARM::dsub_0, &ARM::DPRRegClass);
+    unsigned DSub1 = createExtractSubreg(MBB, InsertPt, DL, Reg,
+                                         ARM::dsub_1, &ARM::DPRRegClass);
+
+    unsigned Out1 = createDupLane(MBB, InsertPt, DL, DSub0, 0);
+    unsigned Out2 = createDupLane(MBB, InsertPt, DL, DSub0, 1);
+    Out = createVExt(MBB, InsertPt, DL, Out1, Out2);
+
+    unsigned Out3 = createDupLane(MBB, InsertPt, DL, DSub1, 0);
+    unsigned Out4 = createDupLane(MBB, InsertPt, DL, DSub1, 1);
+    Out2 = createVExt(MBB, InsertPt, DL, Out3, Out4);
+
+    Out = createRegSequence(MBB, InsertPt, DL, Out, Out2);
+
+  } else if (MRI->getRegClass(Reg)->hasSuperClassEq(&ARM::DPRRegClass)) {
+    unsigned Out1 = createDupLane(MBB, InsertPt, DL, Reg, 0);
+    unsigned Out2 = createDupLane(MBB, InsertPt, DL, Reg, 1);
+    Out = createVExt(MBB, InsertPt, DL, Out1, Out2);
+
+  } else {
+    assert(MRI->getRegClass(Reg)->hasSuperClassEq(&ARM::SPRRegClass) &&
+           "Found unexpected regclass!");
+
+    unsigned PrefLane = getPrefSPRLane(Reg);
+    unsigned Lane;
+    switch (PrefLane) {
+      case ARM::ssub_0: Lane = 0; break;
+      case ARM::ssub_1: Lane = 1; break;
+      default: llvm_unreachable("Unknown preferred lane!");
+    }
+
+    bool UsesQPR = usesRegClass(MI->getOperand(0), &ARM::QPRRegClass);
+
+    Out = createImplicitDef(MBB, InsertPt, DL);
+    Out = createInsertSubreg(MBB, InsertPt, DL, Out, PrefLane, Reg);
+    Out = createDupLane(MBB, InsertPt, DL, Out, Lane, UsesQPR);
+    eraseInstrWithNoUses(MI);
+  }
+  return Out;
+}
+
+bool A15SDOptimizer::runOnInstruction(MachineInstr *MI) {
+  // We look for instructions that write S registers that are then read as
+  // D/Q registers. These can only be caused by COPY, INSERT_SUBREG and
+  // REG_SEQUENCE pseudos that insert an SPR value into a DPR register or
+  // merge two SPR values to form a DPR register.  In order avoid false
+  // positives we make sure that there is an SPR producer so we look past
+  // COPY and PHI nodes to find it.
+  //
+  // The best code pattern for when an SPR producer is going to be used by a
+  // DPR or QPR consumer depends on whether the other lanes of the
+  // corresponding DPR/QPR are currently defined.
+  //
+  // We can handle these efficiently, depending on the type of
+  // pseudo-instruction that is producing the pattern
+  //
+  //   * COPY:          * VDUP all lanes and merge the results together
+  //                      using VEXTs.
+  //
+  //   * INSERT_SUBREG: * If the SPR value was originally in another DPR/QPR
+  //                      lane, and the other lane(s) of the DPR/QPR register
+  //                      that we are inserting in are undefined, use the
+  //                      original DPR/QPR value. 
+  //                    * Otherwise, fall back on the same stategy as COPY.
+  //
+  //   * REG_SEQUENCE:  * If all except one of the input operands are
+  //                      IMPLICIT_DEFs, insert the VDUP pattern for just the
+  //                      defined input operand
+  //                    * Otherwise, fall back on the same stategy as COPY.
+  //
+
+  // First, get all the reads of D-registers done by this instruction.
+  SmallVector<unsigned, 8> Defs = getReadDPRs(MI);
+  bool Modified = false;
+
+  for (SmallVector<unsigned, 8>::iterator I = Defs.begin(), E = Defs.end();
+     I != E; ++I) {
+    // Follow the def-use chain for this DPR through COPYs, and also through
+    // PHIs (which are essentially multi-way COPYs). It is because of PHIs that
+    // we can end up with multiple defs of this DPR.
+
+    SmallVector<MachineInstr *, 8> DefSrcs;
+    if (!TRI->isVirtualRegister(*I))
+      continue;
+    MachineInstr *Def = MRI->getVRegDef(*I);
+    if (!Def)
+      continue;
+
+    elideCopiesAndPHIs(Def, DefSrcs);
+
+    for (SmallVector<MachineInstr*, 8>::iterator II = DefSrcs.begin(),
+      EE = DefSrcs.end(); II != EE; ++II) {
+      MachineInstr *MI = *II;
+
+      // If we've already analyzed and replaced this operand, don't do
+      // anything.
+      if (Replacements.find(MI) != Replacements.end())
+        continue;
+
+      // Now, work out if the instruction causes a SPR->DPR dependency.
+      if (!hasPartialWrite(MI))
+        continue;
+
+      // Collect all the uses of this MI's DPR def for updating later.
+      SmallVector<MachineOperand*, 8> Uses;
+      unsigned DPRDefReg = MI->getOperand(0).getReg();
+      for (MachineRegisterInfo::use_iterator I = MRI->use_begin(DPRDefReg),
+             E = MRI->use_end(); I != E; ++I)
+        Uses.push_back(&I.getOperand());
+
+      // We can optimize this.
+      unsigned NewReg = optimizeSDPattern(MI);
+
+      if (NewReg != 0) {
+        Modified = true;
+        for (SmallVector<MachineOperand*, 8>::const_iterator I = Uses.begin(),
+               E = Uses.end(); I != E; ++I) {
+          DEBUG(dbgs() << "Replacing operand "
+                       << **I << " with "
+                       << PrintReg(NewReg) << "\n");
+          (*I)->substVirtReg(NewReg, 0, *TRI);
+        }
+      }
+      Replacements[MI] = NewReg;
+    }
+  }
+  return Modified;
+}
+
+bool A15SDOptimizer::runOnMachineFunction(MachineFunction &Fn) {
+  TII = static_cast<const ARMBaseInstrInfo*>(Fn.getTarget().getInstrInfo());
+  TRI = Fn.getTarget().getRegisterInfo();
+  MRI = &Fn.getRegInfo();
+  bool Modified = false;
+
+  DEBUG(dbgs() << "Running on function " << Fn.getName()<< "\n");
+
+  DeadInstr.clear();
+  Replacements.clear();
+
+  for (MachineFunction::iterator MFI = Fn.begin(), E = Fn.end(); MFI != E;
+       ++MFI) {
+
+    for (MachineBasicBlock::iterator MI = MFI->begin(), ME = MFI->end();
+      MI != ME;) {
+      Modified |= runOnInstruction(MI++);
+    }
+ 
+  }
+
+  for (std::set<MachineInstr *>::iterator I = DeadInstr.begin(),
+                                            E = DeadInstr.end();
+                                            I != E; ++I) {
+    (*I)->eraseFromParent();
+  }
+
+  return Modified;
+}
+
+FunctionPass *llvm::createA15SDOptimizerPass() {
+  return new A15SDOptimizer();
+}
diff --git a/lib/Target/ARM/ARM.h b/lib/Target/ARM/ARM.h
index 5faf8c3..80e5f37 100644
--- a/lib/Target/ARM/ARM.h
+++ b/lib/Target/ARM/ARM.h
@@ -35,6 +35,7 @@ FunctionPass *createARMISelDag(ARMBaseTargetMachine &TM,
 FunctionPass *createARMJITCodeEmitterPass(ARMBaseTargetMachine &TM,
                                           JITCodeEmitter &JCE);
 
+FunctionPass *createA15SDOptimizerPass();
 FunctionPass *createARMLoadStoreOptimizationPass(bool PreAlloc = false);
 FunctionPass *createARMExpandPseudoPass();
 FunctionPass *createARMGlobalBaseRegPass();
diff --git a/lib/Target/ARM/ARMAsmPrinter.cpp b/lib/Target/ARM/ARMAsmPrinter.cpp
index 58c7798..13ec208 100644
--- a/lib/Target/ARM/ARMAsmPrinter.cpp
+++ b/lib/Target/ARM/ARMAsmPrinter.cpp
@@ -1357,7 +1357,7 @@ void ARMAsmPrinter::EmitInstruction(const MachineInstr *MI) {
 
     OutStreamer.EmitInstruction(MCInstBuilder(ARM::MOVr)
       .addReg(ARM::PC)
-      .addImm(MI->getOperand(0).getReg())
+      .addReg(MI->getOperand(0).getReg())
       // Add predicate operands.
       .addImm(ARMCC::AL)
       .addReg(0)
diff --git a/lib/Target/ARM/ARMBaseInstrInfo.cpp b/lib/Target/ARM/ARMBaseInstrInfo.cpp
index ed001ea..ed8b9cd 100644
--- a/lib/Target/ARM/ARMBaseInstrInfo.cpp
+++ b/lib/Target/ARM/ARMBaseInstrInfo.cpp
@@ -1125,7 +1125,7 @@ bool ARMBaseInstrInfo::expandPostRAPseudo(MachineBasicBlock::iterator MI) const{
   // copyPhysReg() calls.  Look for VMOVS instructions that can legally be
   // widened to VMOVD.  We prefer the VMOVD when possible because it may be
   // changed into a VORR that can go down the NEON pipeline.
-  if (!WidenVMOVS || !MI->isCopy())
+  if (!WidenVMOVS || !MI->isCopy() || Subtarget.isCortexA15())
     return false;
 
   // Look for a copy between even S-registers.  That is where we keep floats
diff --git a/lib/Target/ARM/ARMFrameLowering.cpp b/lib/Target/ARM/ARMFrameLowering.cpp
index 0ca6450..3b12408 100644
--- a/lib/Target/ARM/ARMFrameLowering.cpp
+++ b/lib/Target/ARM/ARMFrameLowering.cpp
@@ -1038,58 +1038,6 @@ static unsigned GetFunctionSizeInBytes(const MachineFunction &MF,
   return FnSize;
 }
 
-/// estimateStackSize - Estimate and return the size of the frame.
-/// FIXME: Make generic?
-static unsigned estimateStackSize(MachineFunction &MF) {
-  const MachineFrameInfo *MFI = MF.getFrameInfo();
-  const TargetFrameLowering *TFI = MF.getTarget().getFrameLowering();
-  const TargetRegisterInfo *RegInfo = MF.getTarget().getRegisterInfo();
-  unsigned MaxAlign = MFI->getMaxAlignment();
-  int Offset = 0;
-
-  // This code is very, very similar to PEI::calculateFrameObjectOffsets().
-  // It really should be refactored to share code. Until then, changes
-  // should keep in mind that there's tight coupling between the two.
-
-  for (int i = MFI->getObjectIndexBegin(); i != 0; ++i) {
-    int FixedOff = -MFI->getObjectOffset(i);
-    if (FixedOff > Offset) Offset = FixedOff;
-  }
-  for (unsigned i = 0, e = MFI->getObjectIndexEnd(); i != e; ++i) {
-    if (MFI->isDeadObjectIndex(i))
-      continue;
-    Offset += MFI->getObjectSize(i);
-    unsigned Align = MFI->getObjectAlignment(i);
-    // Adjust to alignment boundary
-    Offset = (Offset+Align-1)/Align*Align;
-
-    MaxAlign = std::max(Align, MaxAlign);
-  }
-
-  if (MFI->adjustsStack() && TFI->hasReservedCallFrame(MF))
-    Offset += MFI->getMaxCallFrameSize();
-
-  // Round up the size to a multiple of the alignment.  If the function has
-  // any calls or alloca's, align to the target's StackAlignment value to
-  // ensure that the callee's frame or the alloca data is suitably aligned;
-  // otherwise, for leaf functions, align to the TransientStackAlignment
-  // value.
-  unsigned StackAlign;
-  if (MFI->adjustsStack() || MFI->hasVarSizedObjects() ||
-      (RegInfo->needsStackRealignment(MF) && MFI->getObjectIndexEnd() != 0))
-    StackAlign = TFI->getStackAlignment();
-  else
-    StackAlign = TFI->getTransientStackAlignment();
-
-  // If the frame pointer is eliminated, all frame offsets will be relative to
-  // SP not FP. Align to MaxAlign so this works.
-  StackAlign = std::max(StackAlign, MaxAlign);
-  unsigned AlignMask = StackAlign - 1;
-  Offset = (Offset + AlignMask) & ~uint64_t(AlignMask);
-
-  return (unsigned)Offset;
-}
-
 /// estimateRSStackSizeLimit - Look at each instruction that references stack
 /// frames and return the stack size limit beyond which some of these
 /// instructions will require a scratch register during their expansion later.
@@ -1235,7 +1183,7 @@ ARMFrameLowering::processFunctionBeforeCalleeSavedScan(MachineFunction &MF,
     // we've used all the registers and so R4 is already used, so not marking
     // it here will be OK.
     // FIXME: It will be better just to find spare register here.
-    unsigned StackSize = estimateStackSize(MF);
+    unsigned StackSize = MFI->estimateStackSize(MF);
     if (MFI->hasVarSizedObjects() || StackSize > 508)
       MRI.setPhysRegUsed(ARM::R4);
   }
@@ -1330,7 +1278,8 @@ ARMFrameLowering::processFunctionBeforeCalleeSavedScan(MachineFunction &MF,
   //        worth the effort and added fragility?
   bool BigStack =
     (RS &&
-     (estimateStackSize(MF) + ((hasFP(MF) && AFI->hasStackFrame()) ? 4:0) >=
+     (MFI->estimateStackSize(MF) +
+      ((hasFP(MF) && AFI->hasStackFrame()) ? 4:0) >=
       estimateRSStackSizeLimit(MF, this)))
     || MFI->hasVarSizedObjects()
     || (MFI->adjustsStack() && !canSimplifyCallFramePseudos(MF));
diff --git a/lib/Target/ARM/ARMISelDAGToDAG.cpp b/lib/Target/ARM/ARMISelDAGToDAG.cpp
index a83f052..2c51de2 100644
--- a/lib/Target/ARM/ARMISelDAGToDAG.cpp
+++ b/lib/Target/ARM/ARMISelDAGToDAG.cpp
@@ -3155,7 +3155,7 @@ SDNode *ARMDAGToDAGISel::Select(SDNode *N) {
       cast<MachineSDNode>(Ld)->setMemRefs(MemOp, MemOp + 1);
 
       // Remap uses.
-      SDValue Glue = isThumb ? SDValue(Ld, 2) : SDValue(Ld, 1);
+      SDValue OutChain = isThumb ? SDValue(Ld, 2) : SDValue(Ld, 1);
       if (!SDValue(N, 0).use_empty()) {
         SDValue Result;
         if (isThumb)
@@ -3163,9 +3163,8 @@ SDNode *ARMDAGToDAGISel::Select(SDNode *N) {
         else {
           SDValue SubRegIdx = CurDAG->getTargetConstant(ARM::gsub_0, MVT::i32);
           SDNode *ResNode = CurDAG->getMachineNode(TargetOpcode::EXTRACT_SUBREG,
-              dl, MVT::i32, MVT::Glue, SDValue(Ld, 0), SubRegIdx, Glue);
+              dl, MVT::i32, SDValue(Ld, 0), SubRegIdx);
           Result = SDValue(ResNode,0);
-          Glue = Result.getValue(1);
         }
         ReplaceUses(SDValue(N, 0), Result);
       }
@@ -3176,13 +3175,12 @@ SDNode *ARMDAGToDAGISel::Select(SDNode *N) {
         else {
           SDValue SubRegIdx = CurDAG->getTargetConstant(ARM::gsub_1, MVT::i32);
           SDNode *ResNode = CurDAG->getMachineNode(TargetOpcode::EXTRACT_SUBREG,
-              dl, MVT::i32, MVT::Glue, SDValue(Ld, 0), SubRegIdx, Glue);
+              dl, MVT::i32, SDValue(Ld, 0), SubRegIdx);
           Result = SDValue(ResNode,0);
-          Glue = Result.getValue(1);
         }
         ReplaceUses(SDValue(N, 1), Result);
       }
-      ReplaceUses(SDValue(N, 2), Glue);
+      ReplaceUses(SDValue(N, 2), OutChain);
       return NULL;
     }
 
@@ -3195,9 +3193,7 @@ SDNode *ARMDAGToDAGISel::Select(SDNode *N) {
 
       // Store exclusive double return a i32 value which is the return status
       // of the issued store.
-      std::vector<EVT> ResTys;
-      ResTys.push_back(MVT::i32);
-      ResTys.push_back(MVT::Other);
+      EVT ResTys[] = { MVT::i32, MVT::Other };
 
       bool isThumb = Subtarget->isThumb() && Subtarget->hasThumb2();
       // Place arguments in the right order.
diff --git a/lib/Target/ARM/ARMISelLowering.cpp b/lib/Target/ARM/ARMISelLowering.cpp
index ef96e56..514971f 100644
--- a/lib/Target/ARM/ARMISelLowering.cpp
+++ b/lib/Target/ARM/ARMISelLowering.cpp
@@ -504,6 +504,7 @@ ARMTargetLowering::ARMTargetLowering(TargetMachine &TM)
     setOperationAction(ISD::FRINT, MVT::v2f64, Expand);
     setOperationAction(ISD::FNEARBYINT, MVT::v2f64, Expand);
     setOperationAction(ISD::FFLOOR, MVT::v2f64, Expand);
+    setOperationAction(ISD::FMA, MVT::v2f64, Expand);
 
     setOperationAction(ISD::FSQRT, MVT::v4f32, Expand);
     setOperationAction(ISD::FSIN, MVT::v4f32, Expand);
@@ -521,6 +522,23 @@ ARMTargetLowering::ARMTargetLowering(TargetMachine &TM)
     setOperationAction(ISD::FNEARBYINT, MVT::v4f32, Expand);
     setOperationAction(ISD::FFLOOR, MVT::v4f32, Expand);
 
+    // Mark v2f32 intrinsics.
+    setOperationAction(ISD::FSQRT, MVT::v2f32, Expand);
+    setOperationAction(ISD::FSIN, MVT::v2f32, Expand);
+    setOperationAction(ISD::FCOS, MVT::v2f32, Expand);
+    setOperationAction(ISD::FPOWI, MVT::v2f32, Expand);
+    setOperationAction(ISD::FPOW, MVT::v2f32, Expand);
+    setOperationAction(ISD::FLOG, MVT::v2f32, Expand);
+    setOperationAction(ISD::FLOG2, MVT::v2f32, Expand);
+    setOperationAction(ISD::FLOG10, MVT::v2f32, Expand);
+    setOperationAction(ISD::FEXP, MVT::v2f32, Expand);
+    setOperationAction(ISD::FEXP2, MVT::v2f32, Expand);
+    setOperationAction(ISD::FCEIL, MVT::v2f32, Expand);
+    setOperationAction(ISD::FTRUNC, MVT::v2f32, Expand);
+    setOperationAction(ISD::FRINT, MVT::v2f32, Expand);
+    setOperationAction(ISD::FNEARBYINT, MVT::v2f32, Expand);
+    setOperationAction(ISD::FFLOOR, MVT::v2f32, Expand);
+
     // Neon does not support some operations on v1i64 and v2i64 types.
     setOperationAction(ISD::MUL, MVT::v1i64, Expand);
     // Custom handling for some quad-vector types to detect VMULL.
@@ -554,6 +572,12 @@ ARMTargetLowering::ARMTargetLowering(TargetMachine &TM)
     setOperationAction(ISD::CTPOP,      MVT::v4i16, Custom);
     setOperationAction(ISD::CTPOP,      MVT::v8i16, Custom);
 
+    // NEON only has FMA instructions as of VFP4.
+    if (!Subtarget->hasVFP4()) {
+      setOperationAction(ISD::FMA, MVT::v2f32, Expand);
+      setOperationAction(ISD::FMA, MVT::v4f32, Expand);
+    }
+
     setTargetDAGCombine(ISD::INTRINSIC_VOID);
     setTargetDAGCombine(ISD::INTRINSIC_W_CHAIN);
     setTargetDAGCombine(ISD::INTRINSIC_WO_CHAIN);
@@ -1581,7 +1605,7 @@ ARMTargetLowering::LowerCall(TargetLowering::CallLoweringInfo &CLI,
       // On ELF targets for PIC code, direct calls should go through the PLT
       unsigned OpFlags = 0;
       if (Subtarget->isTargetELF() &&
-                  getTargetMachine().getRelocationModel() == Reloc::PIC_)
+          getTargetMachine().getRelocationModel() == Reloc::PIC_)
         OpFlags = ARMII::MO_PLT;
       Callee = DAG.getTargetGlobalAddress(GV, dl, getPointerTy(), 0, OpFlags);
     }
@@ -2247,8 +2271,7 @@ SDValue ARMTargetLowering::LowerGlobalAddressELF(SDValue Op,
   EVT PtrVT = getPointerTy();
   DebugLoc dl = Op.getDebugLoc();
   const GlobalValue *GV = cast<GlobalAddressSDNode>(Op)->getGlobal();
-  Reloc::Model RelocM = getTargetMachine().getRelocationModel();
-  if (RelocM == Reloc::PIC_) {
+  if (getTargetMachine().getRelocationModel() == Reloc::PIC_) {
     bool UseGOTOFF = GV->hasLocalLinkage() || GV->hasHiddenVisibility();
     ARMConstantPoolValue *CPV =
       ARMConstantPoolConstant::Create(GV,
@@ -2292,8 +2315,6 @@ SDValue ARMTargetLowering::LowerGlobalAddressDarwin(SDValue Op,
   DebugLoc dl = Op.getDebugLoc();
   const GlobalValue *GV = cast<GlobalAddressSDNode>(Op)->getGlobal();
   Reloc::Model RelocM = getTargetMachine().getRelocationModel();
-  MachineFunction &MF = DAG.getMachineFunction();
-  ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
 
   // FIXME: Enable this for static codegen when tool issues are fixed.  Also
   // update ARMFastISel::ARMMaterializeGV.
@@ -2321,6 +2342,7 @@ SDValue ARMTargetLowering::LowerGlobalAddressDarwin(SDValue Op,
   if (RelocM == Reloc::Static) {
     CPAddr = DAG.getTargetConstantPool(GV, PtrVT, 4);
   } else {
+    ARMFunctionInfo *AFI = DAG.getMachineFunction().getInfo<ARMFunctionInfo>();
     ARMPCLabelIndex = AFI->createPICLabelUId();
     unsigned PCAdj = (RelocM != Reloc::PIC_) ? 0 : (Subtarget->isThumb()?4:8);
     ARMConstantPoolValue *CPV =
@@ -2401,7 +2423,6 @@ ARMTargetLowering::LowerINTRINSIC_WO_CHAIN(SDValue Op, SelectionDAG &DAG,
     ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
     unsigned ARMPCLabelIndex = AFI->createPICLabelUId();
     EVT PtrVT = getPointerTy();
-    DebugLoc dl = Op.getDebugLoc();
     Reloc::Model RelocM = getTargetMachine().getRelocationModel();
     SDValue CPAddr;
     unsigned PCAdj = (RelocM != Reloc::PIC_)
@@ -2661,7 +2682,7 @@ ARMTargetLowering::LowerFormalArguments(SDValue Chain,
   CCInfo.AnalyzeFormalArguments(Ins,
                                 CCAssignFnForNode(CallConv, /* Return*/ false,
                                                   isVarArg));
-  
+
   SmallVector<SDValue, 16> ArgValues;
   int lastInsIndex = -1;
   SDValue ArgValue;
@@ -2776,7 +2797,7 @@ ARMTargetLowering::LowerFormalArguments(SDValue Chain,
             } else {
               int FI = MFI->CreateFixedObject(Flags.getByValSize(),
                                               VA.getLocMemOffset(), false);
-              InVals.push_back(DAG.getFrameIndex(FI, getPointerTy()));              
+              InVals.push_back(DAG.getFrameIndex(FI, getPointerTy()));
             }
           } else {
             int FI = MFI->CreateFixedObject(VA.getLocVT().getSizeInBits()/8,
@@ -3573,7 +3594,7 @@ static SDValue LowerCTTZ(SDNode *N, SelectionDAG &DAG,
 /// input    = [v0    v1    v2    v3   ] (vi 16-bit element)
 /// cast: N0 = [w0 w1 w2 w3 w4 w5 w6 w7] (v0 = [w0 w1], wi 8-bit element)
 /// vcnt: N1 = [b0 b1 b2 b3 b4 b5 b6 b7] (bi = bit-count of 8-bit element wi)
-/// vrev: N2 = [b1 b0 b3 b2 b5 b4 b7 b6] 
+/// vrev: N2 = [b1 b0 b3 b2 b5 b4 b7 b6]
 ///            [b0 b1 b2 b3 b4 b5 b6 b7]
 ///           +[b1 b0 b3 b2 b5 b4 b7 b6]
 /// N3=N1+N2 = [k0 k0 k1 k1 k2 k2 k3 k3] (k0 = b0+b1 = bit-count of 16-bit v0,
@@ -3594,7 +3615,7 @@ static SDValue getCTPOP16BitCounts(SDNode *N, SelectionDAG &DAG) {
 /// bit-count for each 16-bit element from the operand.  We need slightly
 /// different sequencing for v4i16 and v8i16 to stay within NEON's available
 /// 64/128-bit registers.
-/// 
+///
 /// Trace for v4i16:
 /// input           = [v0    v1    v2    v3    ] (vi 16-bit element)
 /// v8i8: BitCounts = [k0 k1 k2 k3 k0 k1 k2 k3 ] (ki is the bit-count of vi)
@@ -3625,7 +3646,7 @@ static SDValue lowerCTPOP16BitElements(SDNode *N, SelectionDAG &DAG) {
 /// input    = [v0    v1    ] (vi: 32-bit elements)
 /// Bitcast  = [w0 w1 w2 w3 ] (wi: 16-bit elements, v0 = [w0 w1])
 /// Counts16 = [k0 k1 k2 k3 ] (ki: 16-bit elements, bit-count of wi)
-/// vrev: N0 = [k1 k0 k3 k2 ] 
+/// vrev: N0 = [k1 k0 k3 k2 ]
 ///            [k0 k1 k2 k3 ]
 ///       N1 =+[k1 k0 k3 k2 ]
 ///            [k0 k2 k1 k3 ]
@@ -4403,7 +4424,7 @@ SDValue ARMTargetLowering::LowerBUILD_VECTOR(SDValue Op, SelectionDAG &DAG,
 
     ValueCounts.insert(std::make_pair(V, 0));
     unsigned &Count = ValueCounts[V];
-    
+
     // Is this value dominant? (takes up more than half of the lanes)
     if (++Count > (NumElts / 2)) {
       hasDominantValue = true;
@@ -4431,8 +4452,11 @@ SDValue ARMTargetLowering::LowerBUILD_VECTOR(SDValue Op, SelectionDAG &DAG,
 
       // If we are VDUPing a value that comes directly from a vector, that will
       // cause an unnecessary move to and from a GPR, where instead we could
-      // just use VDUPLANE.
-      if (Value->getOpcode() == ISD::EXTRACT_VECTOR_ELT) {
+      // just use VDUPLANE. We can only do this if the lane being extracted
+      // is at a constant index, as the VDUP from lane instructions only have
+      // constant-index forms.
+      if (Value->getOpcode() == ISD::EXTRACT_VECTOR_ELT &&
+          isa<ConstantSDNode>(Value->getOperand(1))) {
         // We need to create a new undef vector to use for the VDUPLANE if the
         // size of the vector from which we get the value is different than the
         // size of the vector that we need to create. We will insert the element
@@ -4447,12 +4471,10 @@ SDValue ARMTargetLowering::LowerBUILD_VECTOR(SDValue Op, SelectionDAG &DAG,
                  DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, VT, DAG.getUNDEF(VT),
                         Value, DAG.getConstant(index, MVT::i32)),
                            DAG.getConstant(index, MVT::i32));
-        } else {
+        } else
           N = DAG.getNode(ARMISD::VDUPLANE, dl, VT,
                         Value->getOperand(0), Value->getOperand(1));
-        }
-      }
-      else
+      } else
         N = DAG.getNode(ARMISD::VDUP, dl, VT, Value);
 
       if (!usesOnlyOneValue) {
@@ -4484,7 +4506,7 @@ SDValue ARMTargetLowering::LowerBUILD_VECTOR(SDValue Op, SelectionDAG &DAG,
     if (usesOnlyOneValue) {
       SDValue Val = IsSingleInstrConstant(Value, DAG, ST, dl);
       if (isConstant && Val.getNode())
-        return DAG.getNode(ARMISD::VDUP, dl, VT, Val); 
+        return DAG.getNode(ARMISD::VDUP, dl, VT, Val);
     }
   }
 
@@ -6329,6 +6351,7 @@ EmitSjLjDispatchBlock(MachineInstr *MI, MachineBasicBlock *MBB) const {
     MF->getOrCreateJumpTableInfo(MachineJumpTableInfo::EK_Inline);
   unsigned MJTI = JTI->createJumpTableIndex(LPadList);
   unsigned UId = AFI->createJumpTableUId();
+  Reloc::Model RelocM = getTargetMachine().getRelocationModel();
 
   // Create the MBBs for the dispatch code.
 
@@ -6338,14 +6361,11 @@ EmitSjLjDispatchBlock(MachineInstr *MI, MachineBasicBlock *MBB) const {
 
   MachineBasicBlock *TrapBB = MF->CreateMachineBasicBlock();
   unsigned trap_opcode;
-  if (Subtarget->isThumb()) {
+  if (Subtarget->isThumb())
     trap_opcode = ARM::tTRAP;
-  } else {
-    if (Subtarget->useNaClTrap())
-      trap_opcode = ARM::TRAPNaCl;
-    else
-      trap_opcode = ARM::TRAP;
-  }
+  else
+    trap_opcode = Subtarget->useNaClTrap() ? ARM::TRAPNaCl : ARM::TRAP;
+
   BuildMI(TrapBB, dl, TII->get(trap_opcode));
   DispatchBB->addSuccessor(TrapBB);
 
@@ -6492,11 +6512,14 @@ EmitSjLjDispatchBlock(MachineInstr *MI, MachineBasicBlock *MBB) const {
                    .addImm(0)
                    .addMemOperand(JTMMOLd));
 
-    unsigned NewVReg6 = MRI->createVirtualRegister(TRC);
-    AddDefaultPred(BuildMI(DispContBB, dl, TII->get(ARM::tADDrr), NewVReg6)
-                   .addReg(ARM::CPSR, RegState::Define)
-                   .addReg(NewVReg5, RegState::Kill)
-                   .addReg(NewVReg3));
+    unsigned NewVReg6 = NewVReg5;
+    if (RelocM == Reloc::PIC_) {
+      NewVReg6 = MRI->createVirtualRegister(TRC);
+      AddDefaultPred(BuildMI(DispContBB, dl, TII->get(ARM::tADDrr), NewVReg6)
+                     .addReg(ARM::CPSR, RegState::Define)
+                     .addReg(NewVReg5, RegState::Kill)
+                     .addReg(NewVReg3));
+    }
 
     BuildMI(DispContBB, dl, TII->get(ARM::tBR_JTr))
       .addReg(NewVReg6, RegState::Kill)
@@ -6576,11 +6599,18 @@ EmitSjLjDispatchBlock(MachineInstr *MI, MachineBasicBlock *MBB) const {
       .addImm(0)
       .addMemOperand(JTMMOLd));
 
-    BuildMI(DispContBB, dl, TII->get(ARM::BR_JTadd))
-      .addReg(NewVReg5, RegState::Kill)
-      .addReg(NewVReg4)
-      .addJumpTableIndex(MJTI)
-      .addImm(UId);
+    if (RelocM == Reloc::PIC_) {
+      BuildMI(DispContBB, dl, TII->get(ARM::BR_JTadd))
+        .addReg(NewVReg5, RegState::Kill)
+        .addReg(NewVReg4)
+        .addJumpTableIndex(MJTI)
+        .addImm(UId);
+    } else {
+      BuildMI(DispContBB, dl, TII->get(ARM::BR_JTr))
+        .addReg(NewVReg5, RegState::Kill)
+        .addJumpTableIndex(MJTI)
+        .addImm(UId);
+    }
   }
 
   // Add the jump table entries as successors to the MBB.
diff --git a/lib/Target/ARM/ARMTargetMachine.cpp b/lib/Target/ARM/ARMTargetMachine.cpp
index 7745218..3003760 100644
--- a/lib/Target/ARM/ARMTargetMachine.cpp
+++ b/lib/Target/ARM/ARMTargetMachine.cpp
@@ -28,6 +28,11 @@ EnableGlobalMerge("global-merge", cl::Hidden,
                   cl::desc("Enable global merge pass"),
                   cl::init(true));
 
+static cl::opt<bool>
+DisableA15SDOptimization("disable-a15-sd-optimization", cl::Hidden,
+                   cl::desc("Inhibit optimization of S->D register accesses on A15"),
+                   cl::init(false));
+
 extern "C" void LLVMInitializeARMTarget() {
   // Register the target.
   RegisterTargetMachine<ARMTargetMachine> X(TheARMTarget);
@@ -164,6 +169,12 @@ bool ARMPassConfig::addPreRegAlloc() {
     addPass(createARMLoadStoreOptimizationPass(true));
   if (getOptLevel() != CodeGenOpt::None && getARMSubtarget().isLikeA9())
     addPass(createMLxExpansionPass());
+  // Since the A15SDOptimizer pass can insert VDUP instructions, it can only be
+  // enabled when NEON is available.
+  if (getOptLevel() != CodeGenOpt::None && getARMSubtarget().isCortexA15() &&
+    getARMSubtarget().hasNEON() && !DisableA15SDOptimization) {
+    addPass(createA15SDOptimizerPass());
+  }
   return true;
 }
 
@@ -174,7 +185,8 @@ bool ARMPassConfig::addPreSched2() {
       addPass(createARMLoadStoreOptimizationPass());
       printAndVerify("After ARM load / store optimizer");
     }
-    if (getARMSubtarget().hasNEON())
+    if ((DisableA15SDOptimization || !getARMSubtarget().isCortexA15()) &&
+      getARMSubtarget().hasNEON())
       addPass(createExecutionDependencyFixPass(&ARM::DPRRegClass));
   }
 
diff --git a/lib/Target/ARM/ARMTargetTransformInfo.cpp b/lib/Target/ARM/ARMTargetTransformInfo.cpp
index 01c04b4..140a8db 100644
--- a/lib/Target/ARM/ARMTargetTransformInfo.cpp
+++ b/lib/Target/ARM/ARMTargetTransformInfo.cpp
@@ -177,6 +177,23 @@ unsigned ARMTTI::getCastInstrCost(unsigned Opcode, Type *Dst,
   int ISD = TLI->InstructionOpcodeToISD(Opcode);
   assert(ISD && "Invalid opcode");
 
+  // Single to/from double precision conversions.
+  static const CostTblEntry<MVT> NEONFltDblTbl[] = {
+    // Vector fptrunc/fpext conversions.
+    { ISD::FP_ROUND,   MVT::v2f64, 2 },
+    { ISD::FP_EXTEND,  MVT::v2f32, 2 },
+    { ISD::FP_EXTEND,  MVT::v4f32, 4 }
+  };
+
+  if (Src->isVectorTy() && ST->hasNEON() && (ISD == ISD::FP_ROUND ||
+                                          ISD == ISD::FP_EXTEND)) {
+    std::pair<unsigned, MVT> LT = TLI->getTypeLegalizationCost(Src);
+    int Idx = CostTableLookup<MVT>(NEONFltDblTbl, array_lengthof(NEONFltDblTbl),
+                                ISD, LT.second);
+    if (Idx != -1)
+      return LT.first * NEONFltDblTbl[Idx].Cost;
+  }
+
   EVT SrcTy = TLI->getValueType(Src);
   EVT DstTy = TLI->getValueType(Dst);
 
@@ -194,17 +211,63 @@ unsigned ARMTTI::getCastInstrCost(unsigned Opcode, Type *Dst,
     { ISD::TRUNCATE,    MVT::v4i32, MVT::v4i64, 0 },
     { ISD::TRUNCATE,    MVT::v4i16, MVT::v4i32, 1 },
 
+    // Operations that we legalize using load/stores to the stack.
+    { ISD::SIGN_EXTEND, MVT::v16i32, MVT::v16i8, 16*2 + 4*4 },
+    { ISD::ZERO_EXTEND, MVT::v16i32, MVT::v16i8, 16*2 + 4*3 },
+    { ISD::SIGN_EXTEND, MVT::v8i32, MVT::v8i8, 8*2 + 2*4 },
+    { ISD::ZERO_EXTEND, MVT::v8i32, MVT::v8i8, 8*2 + 2*3 },
+    { ISD::TRUNCATE,    MVT::v16i8, MVT::v16i32, 4*1 + 16*2 + 2*1 },
+    { ISD::TRUNCATE,    MVT::v8i8, MVT::v8i32, 2*1 + 8*2 + 1 },
+
     // Vector float <-> i32 conversions.
     { ISD::SINT_TO_FP,  MVT::v4f32, MVT::v4i32, 1 },
     { ISD::UINT_TO_FP,  MVT::v4f32, MVT::v4i32, 1 },
+
+    { ISD::SINT_TO_FP,  MVT::v2f32, MVT::v2i8, 3 },
+    { ISD::UINT_TO_FP,  MVT::v2f32, MVT::v2i8, 3 },
+    { ISD::SINT_TO_FP,  MVT::v2f32, MVT::v2i16, 2 },
+    { ISD::UINT_TO_FP,  MVT::v2f32, MVT::v2i16, 2 },
+    { ISD::SINT_TO_FP,  MVT::v2f32, MVT::v2i32, 1 },
+    { ISD::UINT_TO_FP,  MVT::v2f32, MVT::v2i32, 1 },
+    { ISD::SINT_TO_FP,  MVT::v4f32, MVT::v4i1, 3 },
+    { ISD::UINT_TO_FP,  MVT::v4f32, MVT::v4i1, 3 },
+    { ISD::SINT_TO_FP,  MVT::v4f32, MVT::v4i8, 3 },
+    { ISD::UINT_TO_FP,  MVT::v4f32, MVT::v4i8, 3 },
+    { ISD::SINT_TO_FP,  MVT::v4f32, MVT::v4i16, 2 },
+    { ISD::UINT_TO_FP,  MVT::v4f32, MVT::v4i16, 2 },
+    { ISD::SINT_TO_FP,  MVT::v8f32, MVT::v8i16, 4 },
+    { ISD::UINT_TO_FP,  MVT::v8f32, MVT::v8i16, 4 },
+    { ISD::SINT_TO_FP,  MVT::v8f32, MVT::v8i32, 2 },
+    { ISD::UINT_TO_FP,  MVT::v8f32, MVT::v8i32, 2 },
+    { ISD::SINT_TO_FP,  MVT::v16f32, MVT::v16i16, 8 },
+    { ISD::UINT_TO_FP,  MVT::v16f32, MVT::v16i16, 8 },
+    { ISD::SINT_TO_FP,  MVT::v16f32, MVT::v16i32, 4 },
+    { ISD::UINT_TO_FP,  MVT::v16f32, MVT::v16i32, 4 },
+
     { ISD::FP_TO_SINT,  MVT::v4i32, MVT::v4f32, 1 },
     { ISD::FP_TO_UINT,  MVT::v4i32, MVT::v4f32, 1 },
+    { ISD::FP_TO_SINT,  MVT::v4i8, MVT::v4f32, 3 },
+    { ISD::FP_TO_UINT,  MVT::v4i8, MVT::v4f32, 3 },
+    { ISD::FP_TO_SINT,  MVT::v4i16, MVT::v4f32, 2 },
+    { ISD::FP_TO_UINT,  MVT::v4i16, MVT::v4f32, 2 },
 
     // Vector double <-> i32 conversions.
     { ISD::SINT_TO_FP,  MVT::v2f64, MVT::v2i32, 2 },
     { ISD::UINT_TO_FP,  MVT::v2f64, MVT::v2i32, 2 },
+
+    { ISD::SINT_TO_FP,  MVT::v2f64, MVT::v2i8, 4 },
+    { ISD::UINT_TO_FP,  MVT::v2f64, MVT::v2i8, 4 },
+    { ISD::SINT_TO_FP,  MVT::v2f64, MVT::v2i16, 3 },
+    { ISD::UINT_TO_FP,  MVT::v2f64, MVT::v2i16, 3 },
+    { ISD::SINT_TO_FP,  MVT::v2f64, MVT::v2i32, 2 },
+    { ISD::UINT_TO_FP,  MVT::v2f64, MVT::v2i32, 2 },
+
     { ISD::FP_TO_SINT,  MVT::v2i32, MVT::v2f64, 2 },
-    { ISD::FP_TO_UINT,  MVT::v2i32, MVT::v2f64, 2 }
+    { ISD::FP_TO_UINT,  MVT::v2i32, MVT::v2f64, 2 },
+    { ISD::FP_TO_SINT,  MVT::v8i16, MVT::v8f32, 4 },
+    { ISD::FP_TO_UINT,  MVT::v8i16, MVT::v8f32, 4 },
+    { ISD::FP_TO_SINT,  MVT::v16i16, MVT::v16f32, 8 },
+    { ISD::FP_TO_UINT,  MVT::v16i16, MVT::v16f32, 8 }
   };
 
   if (SrcTy.isVector() && ST->hasNEON()) {
@@ -247,7 +310,6 @@ unsigned ARMTTI::getCastInstrCost(unsigned Opcode, Type *Dst,
         return NEONFloatConversionTbl[Idx].Cost;
   }
 
-
   // Scalar integer to float conversions.
   static const TypeConversionCostTblEntry<MVT> NEONIntegerConversionTbl[] = {
     { ISD::SINT_TO_FP,  MVT::f32, MVT::i1, 2 },
@@ -303,7 +365,6 @@ unsigned ARMTTI::getCastInstrCost(unsigned Opcode, Type *Dst,
       return ARMIntegerConversionTbl[Idx].Cost;
   }
 
-
   return TargetTransformInfo::getCastInstrCost(Opcode, Dst, Src);
 }
 
@@ -326,6 +387,25 @@ unsigned ARMTTI::getCmpSelInstrCost(unsigned Opcode, Type *ValTy,
   int ISD = TLI->InstructionOpcodeToISD(Opcode);
   // On NEON a a vector select gets lowered to vbsl.
   if (ST->hasNEON() && ValTy->isVectorTy() && ISD == ISD::SELECT) {
+    // Lowering of some vector selects is currently far from perfect.
+    static const TypeConversionCostTblEntry<MVT> NEONVectorSelectTbl[] = {
+      { ISD::SELECT, MVT::v16i1, MVT::v16i16, 2*16 + 1 + 3*1 + 4*1 },
+      { ISD::SELECT, MVT::v8i1, MVT::v8i32, 4*8 + 1*3 + 1*4 + 1*2 },
+      { ISD::SELECT, MVT::v16i1, MVT::v16i32, 4*16 + 1*6 + 1*8 + 1*4 },
+      { ISD::SELECT, MVT::v4i1, MVT::v4i64, 4*4 + 1*2 + 1 },
+      { ISD::SELECT, MVT::v8i1, MVT::v8i64, 50 },
+      { ISD::SELECT, MVT::v16i1, MVT::v16i64, 100 }
+    };
+
+    EVT SelCondTy = TLI->getValueType(CondTy);
+    EVT SelValTy = TLI->getValueType(ValTy);
+    int Idx = ConvertCostTableLookup<MVT>(NEONVectorSelectTbl,
+                                          array_lengthof(NEONVectorSelectTbl),
+                                          ISD, SelCondTy.getSimpleVT(),
+                                          SelValTy.getSimpleVT());
+    if (Idx != -1)
+      return NEONVectorSelectTbl[Idx].Cost;
+
     std::pair<unsigned, MVT> LT = TLI->getTypeLegalizationCost(ValTy);
     return LT.first;
   }
diff --git a/lib/Target/ARM/AsmParser/ARMAsmParser.cpp b/lib/Target/ARM/AsmParser/ARMAsmParser.cpp
index 6c678fd..c897efd 100644
--- a/lib/Target/ARM/AsmParser/ARMAsmParser.cpp
+++ b/lib/Target/ARM/AsmParser/ARMAsmParser.cpp
@@ -316,103 +316,127 @@ class ARMOperand : public MCParsedAsmOperand {
   SMLoc StartLoc, EndLoc;
   SmallVector<unsigned, 8> Registers;
 
+  struct CCOp {
+    ARMCC::CondCodes Val;
+  };
+
+  struct CopOp {
+    unsigned Val;
+  };
+
+  struct CoprocOptionOp {
+    unsigned Val;
+  };
+
+  struct ITMaskOp {
+    unsigned Mask:4;
+  };
+
+  struct MBOptOp {
+    ARM_MB::MemBOpt Val;
+  };
+
+  struct IFlagsOp {
+    ARM_PROC::IFlags Val;
+  };
+
+  struct MMaskOp {
+    unsigned Val;
+  };
+
+  struct TokOp {
+    const char *Data;
+    unsigned Length;
+  };
+
+  struct RegOp {
+    unsigned RegNum;
+  };
+
+  // A vector register list is a sequential list of 1 to 4 registers.
+  struct VectorListOp {
+    unsigned RegNum;
+    unsigned Count;
+    unsigned LaneIndex;
+    bool isDoubleSpaced;
+  };
+
+  struct VectorIndexOp {
+    unsigned Val;
+  };
+
+  struct ImmOp {
+    const MCExpr *Val;
+  };
+
+  /// Combined record for all forms of ARM address expressions.
+  struct MemoryOp {
+    unsigned BaseRegNum;
+    // Offset is in OffsetReg or OffsetImm. If both are zero, no offset
+    // was specified.
+    const MCConstantExpr *OffsetImm;  // Offset immediate value
+    unsigned OffsetRegNum;    // Offset register num, when OffsetImm == NULL
+    ARM_AM::ShiftOpc ShiftType; // Shift type for OffsetReg
+    unsigned ShiftImm;        // shift for OffsetReg.
+    unsigned Alignment;       // 0 = no alignment specified
+    // n = alignment in bytes (2, 4, 8, 16, or 32)
+    unsigned isNegative : 1;  // Negated OffsetReg? (~'U' bit)
+  };
+
+  struct PostIdxRegOp {
+    unsigned RegNum;
+    bool isAdd;
+    ARM_AM::ShiftOpc ShiftTy;
+    unsigned ShiftImm;
+  };
+
+  struct ShifterImmOp {
+    bool isASR;
+    unsigned Imm;
+  };
+
+  struct RegShiftedRegOp {
+    ARM_AM::ShiftOpc ShiftTy;
+    unsigned SrcReg;
+    unsigned ShiftReg;
+    unsigned ShiftImm;
+  };
+
+  struct RegShiftedImmOp {
+    ARM_AM::ShiftOpc ShiftTy;
+    unsigned SrcReg;
+    unsigned ShiftImm;
+  };
+
+  struct RotImmOp {
+    unsigned Imm;
+  };
+
+  struct BitfieldOp {
+    unsigned LSB;
+    unsigned Width;
+  };
+
   union {
-    struct {
-      ARMCC::CondCodes Val;
-    } CC;
-
-    struct {
-      unsigned Val;
-    } Cop;
-
-    struct {
-      unsigned Val;
-    } CoprocOption;
-
-    struct {
-      unsigned Mask:4;
-    } ITMask;
-
-    struct {
-      ARM_MB::MemBOpt Val;
-    } MBOpt;
-
-    struct {
-      ARM_PROC::IFlags Val;
-    } IFlags;
-
-    struct {
-      unsigned Val;
-    } MMask;
-
-    struct {
-      const char *Data;
-      unsigned Length;
-    } Tok;
-
-    struct {
-      unsigned RegNum;
-    } Reg;
-
-    // A vector register list is a sequential list of 1 to 4 registers.
-    struct {
-      unsigned RegNum;
-      unsigned Count;
-      unsigned LaneIndex;
-      bool isDoubleSpaced;
-    } VectorList;
-
-    struct {
-      unsigned Val;
-    } VectorIndex;
-
-    struct {
-      const MCExpr *Val;
-    } Imm;
-
-    /// Combined record for all forms of ARM address expressions.
-    struct {
-      unsigned BaseRegNum;
-      // Offset is in OffsetReg or OffsetImm. If both are zero, no offset
-      // was specified.
-      const MCConstantExpr *OffsetImm;  // Offset immediate value
-      unsigned OffsetRegNum;    // Offset register num, when OffsetImm == NULL
-      ARM_AM::ShiftOpc ShiftType; // Shift type for OffsetReg
-      unsigned ShiftImm;        // shift for OffsetReg.
-      unsigned Alignment;       // 0 = no alignment specified
-                                // n = alignment in bytes (2, 4, 8, 16, or 32)
-      unsigned isNegative : 1;  // Negated OffsetReg? (~'U' bit)
-    } Memory;
-
-    struct {
-      unsigned RegNum;
-      bool isAdd;
-      ARM_AM::ShiftOpc ShiftTy;
-      unsigned ShiftImm;
-    } PostIdxReg;
-
-    struct {
-      bool isASR;
-      unsigned Imm;
-    } ShifterImm;
-    struct {
-      ARM_AM::ShiftOpc ShiftTy;
-      unsigned SrcReg;
-      unsigned ShiftReg;
-      unsigned ShiftImm;
-    } RegShiftedReg;
-    struct {
-      ARM_AM::ShiftOpc ShiftTy;
-      unsigned SrcReg;
-      unsigned ShiftImm;
-    } RegShiftedImm;
-    struct {
-      unsigned Imm;
-    } RotImm;
-    struct {
-      unsigned LSB;
-      unsigned Width;
-    } Bitfield;
+    struct CCOp CC;
+    struct CopOp Cop;
+    struct CoprocOptionOp CoprocOption;
+    struct MBOptOp MBOpt;
+    struct ITMaskOp ITMask;
+    struct IFlagsOp IFlags;
+    struct MMaskOp MMask;
+    struct TokOp Tok;
+    struct RegOp Reg;
+    struct VectorListOp VectorList;
+    struct VectorIndexOp VectorIndex;
+    struct ImmOp Imm;
+    struct MemoryOp Memory;
+    struct PostIdxRegOp PostIdxReg;
+    struct ShifterImmOp ShifterImm;
+    struct RegShiftedRegOp RegShiftedReg;
+    struct RegShiftedImmOp RegShiftedImm;
+    struct RotImmOp RotImm;
+    struct BitfieldOp Bitfield;
   };
 
   ARMOperand(KindTy K) : MCParsedAsmOperand(), Kind(K) {}
diff --git a/lib/Target/ARM/CMakeLists.txt b/lib/Target/ARM/CMakeLists.txt
index 586834c..b832508 100644
--- a/lib/Target/ARM/CMakeLists.txt
+++ b/lib/Target/ARM/CMakeLists.txt
@@ -15,6 +15,7 @@ tablegen(LLVM ARMGenDisassemblerTables.inc -gen-disassembler)
 add_public_tablegen_target(ARMCommonTableGen)
 
 add_llvm_target(ARMCodeGen
+  A15SDOptimizer.cpp
   ARMAsmPrinter.cpp
   ARMBaseInstrInfo.cpp
   ARMBaseRegisterInfo.cpp
diff --git a/lib/Target/ARM/MCTargetDesc/ARMMCCodeEmitter.cpp b/lib/Target/ARM/MCTargetDesc/ARMMCCodeEmitter.cpp
index 09e15af..7a59a7d 100644
--- a/lib/Target/ARM/MCTargetDesc/ARMMCCodeEmitter.cpp
+++ b/lib/Target/ARM/MCTargetDesc/ARMMCCodeEmitter.cpp
@@ -655,15 +655,28 @@ getAdrLabelOpValue(const MCInst &MI, unsigned OpIdx,
   int32_t offset = MO.getImm();
   uint32_t Val = 0x2000;
 
+  int SoImmVal;
   if (offset == INT32_MIN) {
     Val = 0x1000;
-    offset = 0;
+    SoImmVal = 0;
   } else if (offset < 0) {
     Val = 0x1000;
     offset *= -1;
+    SoImmVal = ARM_AM::getSOImmVal(offset);
+    if(SoImmVal == -1) {
+      Val = 0x2000;
+      offset *= -1;
+      SoImmVal = ARM_AM::getSOImmVal(offset);
+    }
+  } else {
+    SoImmVal = ARM_AM::getSOImmVal(offset);
+    if(SoImmVal == -1) {
+      Val = 0x1000;
+      offset *= -1;
+      SoImmVal = ARM_AM::getSOImmVal(offset);
+    }
   }
 
-  int SoImmVal = ARM_AM::getSOImmVal(offset);
   assert(SoImmVal != -1 && "Not a valid so_imm value!");
 
   Val |= SoImmVal;