vendor/llvm/llvm-release_33-r183502

47 files changed, 1507 insertions, 627 deletions

diff --git a/lib/CodeGen/Analysis.cpp b/lib/CodeGen/Analysis.cpp
index dd7282c0ad97..4731af5089ac 100644
--- a/lib/CodeGen/Analysis.cpp
+++ b/lib/CodeGen/Analysis.cpp
@@ -201,62 +201,161 @@ ISD::CondCode llvm::getICmpCondCode(ICmpInst::Predicate Pred) {
   }
 }
 
+static bool isNoopBitcast(Type *T1, Type *T2,
+                          const TargetLowering& TLI) {
+  return T1 == T2 || (T1->isPointerTy() && T2->isPointerTy()) ||
+         (isa<VectorType>(T1) && isa<VectorType>(T2) &&
+          TLI.isTypeLegal(EVT::getEVT(T1)) && TLI.isTypeLegal(EVT::getEVT(T2)));
+}
 
-/// getNoopInput - If V is a noop (i.e., lowers to no machine code), look
-/// through it (and any transitive noop operands to it) and return its input
-/// value.  This is used to determine if a tail call can be formed.
-///
-static const Value *getNoopInput(const Value *V, const TargetLowering &TLI) {
-  // If V is not an instruction, it can't be looked through.
-  const Instruction *I = dyn_cast<Instruction>(V);
-  if (I == 0 || !I->hasOneUse() || I->getNumOperands() == 0) return V;
-  
-  Value *Op = I->getOperand(0);
+/// sameNoopInput - Return true if V1 == V2, else if either V1 or V2 is a noop
+/// (i.e., lowers to no machine code), look through it (and any transitive noop
+/// operands to it) and check if it has the same noop input value.  This is
+/// used to determine if a tail call can be formed.
+static bool sameNoopInput(const Value *V1, const Value *V2,
+                          SmallVectorImpl<unsigned> &Els1,
+                          SmallVectorImpl<unsigned> &Els2,
+                          const TargetLowering &TLI) {
+  using std::swap;
+  bool swapParity = false;
+  bool equalEls = Els1 == Els2;
+  while (true) {
+    if ((equalEls && V1 == V2) || isa<UndefValue>(V1) || isa<UndefValue>(V2)) {
+      if (swapParity)
+        // Revert to original Els1 and Els2 to avoid confusing recursive calls
+        swap(Els1, Els2);
+      return true;
+    }
 
-  // Look through truly no-op truncates.
-  if (isa<TruncInst>(I) &&
-      TLI.isTruncateFree(I->getOperand(0)->getType(), I->getType()))
-    return getNoopInput(I->getOperand(0), TLI);
-  
-  // Look through truly no-op bitcasts.
-  if (isa<BitCastInst>(I)) {
-    // No type change at all.
-    if (Op->getType() == I->getType())
-      return getNoopInput(Op, TLI);
+    // Try to look through V1; if V1 is not an instruction, it can't be looked
+    // through.
+    const Instruction *I = dyn_cast<Instruction>(V1);
+    const Value *NoopInput = 0;
+    if (I != 0 && I->getNumOperands() > 0) {
+     Value *Op = I->getOperand(0);
+      if (isa<TruncInst>(I)) {
+        // Look through truly no-op truncates.
+        if (TLI.isTruncateFree(Op->getType(), I->getType()))
+          NoopInput = Op;
+      } else if (isa<BitCastInst>(I)) {
+        // Look through truly no-op bitcasts.
+        if (isNoopBitcast(Op->getType(), I->getType(), TLI))
+          NoopInput = Op;
+      } else if (isa<GetElementPtrInst>(I)) {
+        // Look through getelementptr
+        if (cast<GetElementPtrInst>(I)->hasAllZeroIndices())
+          NoopInput = Op;
+      } else if (isa<IntToPtrInst>(I)) {
+        // Look through inttoptr.
+        // Make sure this isn't a truncating or extending cast.  We could
+        // support this eventually, but don't bother for now.
+        if (!isa<VectorType>(I->getType()) &&
+            TLI.getPointerTy().getSizeInBits() == 
+              cast<IntegerType>(Op->getType())->getBitWidth())
+          NoopInput = Op;
+      } else if (isa<PtrToIntInst>(I)) {
+        // Look through ptrtoint.
+        // Make sure this isn't a truncating or extending cast.  We could
+        // support this eventually, but don't bother for now.
+        if (!isa<VectorType>(I->getType()) &&
+            TLI.getPointerTy().getSizeInBits() == 
+              cast<IntegerType>(I->getType())->getBitWidth())
+          NoopInput = Op;
+      } else if (isa<CallInst>(I)) {
+        // Look through call
+        for (User::const_op_iterator i = I->op_begin(),
+                                     // Skip Callee
+                                     e = I->op_end() - 1;
+             i != e; ++i) {
+          unsigned attrInd = i - I->op_begin() + 1;
+          if (cast<CallInst>(I)->paramHasAttr(attrInd, Attribute::Returned) &&
+              isNoopBitcast((*i)->getType(), I->getType(), TLI)) {
+            NoopInput = *i;
+            break;
+          }
+        }
+      } else if (isa<InvokeInst>(I)) {
+        // Look through invoke
+        for (User::const_op_iterator i = I->op_begin(),
+                                     // Skip BB, BB, Callee
+                                     e = I->op_end() - 3;
+             i != e; ++i) {
+          unsigned attrInd = i - I->op_begin() + 1;
+          if (cast<InvokeInst>(I)->paramHasAttr(attrInd, Attribute::Returned) &&
+              isNoopBitcast((*i)->getType(), I->getType(), TLI)) {
+            NoopInput = *i;
+            break;
+          }
+        }
+      }
+    }
 
-    // Pointer to pointer cast.
-    if (Op->getType()->isPointerTy() && I->getType()->isPointerTy())
-      return getNoopInput(Op, TLI);
-    
-    if (isa<VectorType>(Op->getType()) && isa<VectorType>(I->getType()) &&
-        TLI.isTypeLegal(EVT::getEVT(Op->getType())) &&
-        TLI.isTypeLegal(EVT::getEVT(I->getType())))
-      return getNoopInput(Op, TLI);
-  }
-  
-  // Look through inttoptr.
-  if (isa<IntToPtrInst>(I) && !isa<VectorType>(I->getType())) {
-    // Make sure this isn't a truncating or extending cast.  We could support
-    // this eventually, but don't bother for now.
-    if (TLI.getPointerTy().getSizeInBits() == 
-          cast<IntegerType>(Op->getType())->getBitWidth())
-      return getNoopInput(Op, TLI);
-  }
+    if (NoopInput) {
+      V1 = NoopInput;
+      continue;
+    }
 
-  // Look through ptrtoint.
-  if (isa<PtrToIntInst>(I) && !isa<VectorType>(I->getType())) {
-    // Make sure this isn't a truncating or extending cast.  We could support
-    // this eventually, but don't bother for now.
-    if (TLI.getPointerTy().getSizeInBits() == 
-        cast<IntegerType>(I->getType())->getBitWidth())
-      return getNoopInput(Op, TLI);
+    // If we already swapped, avoid infinite loop
+    if (swapParity)
+      break;
+
+    // Otherwise, swap V1<->V2, Els1<->Els2
+    swap(V1, V2);
+    swap(Els1, Els2);
+    swapParity = !swapParity;
   }
 
+  for (unsigned n = 0; n < 2; ++n) {
+    if (isa<InsertValueInst>(V1)) {
+      if (isa<StructType>(V1->getType())) {
+        // Look through insertvalue
+        unsigned i, e;
+        for (i = 0, e = cast<StructType>(V1->getType())->getNumElements();
+             i != e; ++i) {
+          const Value *InScalar = FindInsertedValue(const_cast<Value*>(V1), i);
+          if (InScalar == 0)
+            break;
+          Els1.push_back(i);
+          if (!sameNoopInput(InScalar, V2, Els1, Els2, TLI)) {
+            Els1.pop_back();
+            break;
+          }
+          Els1.pop_back();
+        }
+        if (i == e) {
+          if (swapParity)
+            swap(Els1, Els2);
+          return true;
+        }
+      }
+    } else if (!Els1.empty() && isa<ExtractValueInst>(V1)) {
+      const ExtractValueInst *EVI = cast<ExtractValueInst>(V1);
+      unsigned i = Els1.back();
+      // If the scalar value being inserted is an extractvalue of the right
+      // index from the call, then everything is good.
+      if (isa<StructType>(EVI->getOperand(0)->getType()) &&
+          EVI->getNumIndices() == 1 && EVI->getIndices()[0] == i) {
+        // Look through extractvalue
+        Els1.pop_back();
+        if (sameNoopInput(EVI->getOperand(0), V2, Els1, Els2, TLI)) {
+          Els1.push_back(i);
+          if (swapParity)
+            swap(Els1, Els2);
+          return true;
+        }
+        Els1.push_back(i);
+      }
+    }
 
-  // Otherwise it's not something we can look through.
-  return V;
-}
+    swap(V1, V2);
+    swap(Els1, Els2);
+    swapParity = !swapParity;
+  }
 
+  if (swapParity)
+    swap(Els1, Els2);
+  return false;
+}
 
 /// Test if the given instruction is in a position to be optimized
 /// with a tail-call. This roughly means that it's in a block with
@@ -264,7 +363,8 @@ static const Value *getNoopInput(const Value *V, const TargetLowering &TLI) {
 /// between it and the return.
 ///
 /// This function only tests target-independent requirements.
-bool llvm::isInTailCallPosition(ImmutableCallSite CS,const TargetLowering &TLI){
+bool llvm::isInTailCallPosition(ImmutableCallSite CS,
+                                const TargetLowering &TLI) {
   const Instruction *I = CS.getInstruction();
   const BasicBlock *ExitBB = I->getParent();
   const TerminatorInst *Term = ExitBB->getTerminator();
@@ -322,28 +422,7 @@ bool llvm::isInTailCallPosition(ImmutableCallSite CS,const TargetLowering &TLI){
       CallerAttrs.hasAttribute(AttributeSet::ReturnIndex, Attribute::SExt))
     return false;
 
-  // Otherwise, make sure the unmodified return value of I is the return value.
-  // We handle two cases: multiple return values + scalars.
-  Value *RetVal = Ret->getOperand(0);
-  if (!isa<InsertValueInst>(RetVal) || !isa<StructType>(RetVal->getType()))
-    // Handle scalars first.
-    return getNoopInput(Ret->getOperand(0), TLI) == I;
-  
-  // If this is an aggregate return, look through the insert/extract values and
-  // see if each is transparent.
-  for (unsigned i = 0, e =cast<StructType>(RetVal->getType())->getNumElements();
-       i != e; ++i) {
-    const Value *InScalar = FindInsertedValue(RetVal, i);
-    if (InScalar == 0) return false;
-    InScalar = getNoopInput(InScalar, TLI);
-    
-    // If the scalar value being inserted is an extractvalue of the right index
-    // from the call, then everything is good.
-    const ExtractValueInst *EVI = dyn_cast<ExtractValueInst>(InScalar);
-    if (EVI == 0 || EVI->getOperand(0) != I || EVI->getNumIndices() != 1 ||
-        EVI->getIndices()[0] != i)
-      return false;
-  }
-  
-  return true;
+  // Otherwise, make sure the return value and I have the same value
+  SmallVector<unsigned, 4> Els1, Els2;
+  return sameNoopInput(Ret->getOperand(0), I, Els1, Els2, TLI);
 }
diff --git a/lib/CodeGen/AsmPrinter/AsmPrinter.cpp b/lib/CodeGen/AsmPrinter/AsmPrinter.cpp
index d4a745d985e8..84162ace4188 100644
--- a/lib/CodeGen/AsmPrinter/AsmPrinter.cpp
+++ b/lib/CodeGen/AsmPrinter/AsmPrinter.cpp
@@ -133,9 +133,13 @@ const DataLayout &AsmPrinter::getDataLayout() const {
   return *TM.getDataLayout();
 }
 
+StringRef AsmPrinter::getTargetTriple() const {
+  return TM.getTargetTriple();
+}
+
 /// getCurrentSection() - Return the current section we are emitting to.
 const MCSection *AsmPrinter::getCurrentSection() const {
-  return OutStreamer.getCurrentSection();
+  return OutStreamer.getCurrentSection().first;
 }
 
 
@@ -813,7 +817,7 @@ void AsmPrinter::EmitDwarfRegOp(const MachineLocation &MLoc) const {
   // caller might be in the middle of an dwarf expression. We should
   // probably assert that Reg >= 0 once debug info generation is more mature.
 
-  if (int Offset =  MLoc.getOffset()) {
+  if (MLoc.isIndirect()) {
     if (Reg < 32) {
       OutStreamer.AddComment(
         dwarf::OperationEncodingString(dwarf::DW_OP_breg0 + Reg));
@@ -824,7 +828,7 @@ void AsmPrinter::EmitDwarfRegOp(const MachineLocation &MLoc) const {
       OutStreamer.AddComment(Twine(Reg));
       EmitULEB128(Reg);
     }
-    EmitSLEB128(Offset);
+    EmitSLEB128(MLoc.getOffset());
   } else {
     if (Reg < 32) {
       OutStreamer.AddComment(
@@ -1213,7 +1217,7 @@ void AsmPrinter::EmitJumpTableEntry(const MachineJumpTableInfo *MJTI,
 bool AsmPrinter::EmitSpecialLLVMGlobal(const GlobalVariable *GV) {
   if (GV->getName() == "llvm.used") {
     if (MAI->hasNoDeadStrip())    // No need to emit this at all.
-      EmitLLVMUsedList(GV->getInitializer());
+      EmitLLVMUsedList(cast<ConstantArray>(GV->getInitializer()));
     return true;
   }
 
@@ -1256,11 +1260,8 @@ bool AsmPrinter::EmitSpecialLLVMGlobal(const GlobalVariable *GV) {
 /// EmitLLVMUsedList - For targets that define a MAI::UsedDirective, mark each
 /// global in the specified llvm.used list for which emitUsedDirectiveFor
 /// is true, as being used with this directive.
-void AsmPrinter::EmitLLVMUsedList(const Constant *List) {
+void AsmPrinter::EmitLLVMUsedList(const ConstantArray *InitList) {
   // Should be an array of 'i8*'.
-  const ConstantArray *InitList = dyn_cast<ConstantArray>(List);
-  if (InitList == 0) return;
-
   for (unsigned i = 0, e = InitList->getNumOperands(); i != e; ++i) {
     const GlobalValue *GV =
       dyn_cast<GlobalValue>(InitList->getOperand(i)->stripPointerCasts());
diff --git a/lib/CodeGen/AsmPrinter/AsmPrinterDwarf.cpp b/lib/CodeGen/AsmPrinter/AsmPrinterDwarf.cpp
index 156acace553d..31e42d47cf46 100644
--- a/lib/CodeGen/AsmPrinter/AsmPrinterDwarf.cpp
+++ b/lib/CodeGen/AsmPrinter/AsmPrinterDwarf.cpp
@@ -141,7 +141,7 @@ void AsmPrinter::EmitTTypeReference(const GlobalValue *GV,
 void AsmPrinter::EmitSectionOffset(const MCSymbol *Label,
                                    const MCSymbol *SectionLabel) const {
   // On COFF targets, we have to emit the special .secrel32 directive.
-  if (MAI->getDwarfSectionOffsetDirective()) {
+  if (MAI->needsDwarfSectionOffsetDirective()) {
     OutStreamer.EmitCOFFSecRel32(Label);
     return;
   }
diff --git a/lib/CodeGen/AsmPrinter/DIE.cpp b/lib/CodeGen/AsmPrinter/DIE.cpp
index 57e0acda890f..673867ada1a2 100644
--- a/lib/CodeGen/AsmPrinter/DIE.cpp
+++ b/lib/CodeGen/AsmPrinter/DIE.cpp
@@ -112,8 +112,9 @@ DIE::~DIE() {
     delete Children[i];
 }
 
-/// Climb up the parent chain to get the compile unit DIE this DIE belongs to.
-DIE *DIE::getCompileUnit() const{
+/// Climb up the parent chain to get the compile unit DIE to which this DIE
+/// belongs.
+DIE *DIE::getCompileUnit() const {
   DIE *p = getParent();
   while (p) {
     if (p->getTag() == dwarf::DW_TAG_compile_unit)
@@ -124,8 +125,7 @@ DIE *DIE::getCompileUnit() const{
 }
 
 #ifndef NDEBUG
-void DIE::print(raw_ostream &O, unsigned IncIndent) {
-  IndentCount += IncIndent;
+void DIE::print(raw_ostream &O, unsigned IndentCount) const {
   const std::string Indent(IndentCount, ' ');
   bool isBlock = Abbrev.getTag() == 0;
 
@@ -164,11 +164,10 @@ void DIE::print(raw_ostream &O, unsigned IncIndent) {
   IndentCount -= 2;
 
   for (unsigned j = 0, M = Children.size(); j < M; ++j) {
-    Children[j]->print(O, 4);
+    Children[j]->print(O, IndentCount+4);
   }
 
   if (!isBlock) O << "\n";
-  IndentCount -= IncIndent;
 }
 
 void DIE::dump() {
diff --git a/lib/CodeGen/AsmPrinter/DIE.h b/lib/CodeGen/AsmPrinter/DIE.h
index c332aa2a7db6..3c06001686f4 100644
--- a/lib/CodeGen/AsmPrinter/DIE.h
+++ b/lib/CodeGen/AsmPrinter/DIE.h
@@ -139,8 +139,7 @@ namespace llvm {
     mutable unsigned IndentCount;
   public:
     explicit DIE(unsigned Tag)
-      : Offset(0), Size(0), Abbrev(Tag, dwarf::DW_CHILDREN_no), Parent(0),
-        IndentCount(0) {}
+      : Offset(0), Size(0), Abbrev(Tag, dwarf::DW_CHILDREN_no), Parent(0) {}
     virtual ~DIE();
 
     // Accessors.
@@ -179,7 +178,7 @@ namespace llvm {
     }
 
 #ifndef NDEBUG
-    void print(raw_ostream &O, unsigned IncIndent = 0);
+    void print(raw_ostream &O, unsigned IndentCount = 0) const;
     void dump();
 #endif
   };
diff --git a/lib/CodeGen/AsmPrinter/DwarfCompileUnit.cpp b/lib/CodeGen/AsmPrinter/DwarfCompileUnit.cpp
index f9b6f9472141..89abcffd938a 100644
--- a/lib/CodeGen/AsmPrinter/DwarfCompileUnit.cpp
+++ b/lib/CodeGen/AsmPrinter/DwarfCompileUnit.cpp
@@ -685,7 +685,7 @@ bool CompileUnit::addConstantValue(DIE *Die, const APInt &Val,
   return true;
 }
 
-/// addTemplateParams - Add template parameters in buffer.
+/// addTemplateParams - Add template parameters into buffer.
 void CompileUnit::addTemplateParams(DIE &Buffer, DIArray TParams) {
   // Add template parameters.
   for (unsigned i = 0, e = TParams.getNumElements(); i != e; ++i) {
@@ -707,7 +707,7 @@ DIE *CompileUnit::getOrCreateContextDIE(DIDescriptor Context) {
     return getOrCreateNameSpace(DINameSpace(Context));
   else if (Context.isSubprogram())
     return getOrCreateSubprogramDIE(DISubprogram(Context));
-  else 
+  else
     return getDIE(Context);
 }
 
diff --git a/lib/CodeGen/AsmPrinter/DwarfCompileUnit.h b/lib/CodeGen/AsmPrinter/DwarfCompileUnit.h
index 2b180c6cc3f4..8f08c63e6498 100644
--- a/lib/CodeGen/AsmPrinter/DwarfCompileUnit.h
+++ b/lib/CodeGen/AsmPrinter/DwarfCompileUnit.h
@@ -94,9 +94,6 @@ class CompileUnit {
   /// DWARF version doesn't handle the language, return -1.
   int64_t getDefaultLowerBound() const;
 
-  /// getOrCreateContextDIE - Get context owner's DIE.
-  DIE *getOrCreateContextDIE(DIDescriptor Context);
-
 public:
   CompileUnit(unsigned UID, unsigned L, DIE *D, AsmPrinter *A, DwarfDebug *DW,
               DwarfUnits *);
@@ -372,6 +369,9 @@ public:
   /// createStaticMemberDIE - Create new static data member DIE.
   DIE *createStaticMemberDIE(DIDerivedType DT);
 
+  /// getOrCreateContextDIE - Get context owner's DIE.
+  DIE *getOrCreateContextDIE(DIDescriptor Context);
+
 private:
 
   // DIEValueAllocator - All DIEValues are allocated through this allocator.
diff --git a/lib/CodeGen/AsmPrinter/DwarfDebug.cpp b/lib/CodeGen/AsmPrinter/DwarfDebug.cpp
index d3cb4f9c1c0e..73bba6989faa 100644
--- a/lib/CodeGen/AsmPrinter/DwarfDebug.cpp
+++ b/lib/CodeGen/AsmPrinter/DwarfDebug.cpp
@@ -94,6 +94,12 @@ static cl::opt<DefaultOnOff> SplitDwarf("split-dwarf", cl::Hidden,
 namespace {
   const char *DWARFGroupName = "DWARF Emission";
   const char *DbgTimerName = "DWARF Debug Writer";
+
+  struct CompareFirst {
+    template <typename T> bool operator()(const T &lhs, const T &rhs) const {
+      return lhs.first < rhs.first;
+    }
+  };
 } // end anonymous namespace
 
 //===----------------------------------------------------------------------===//
@@ -170,12 +176,13 @@ DwarfDebug::DwarfDebug(AsmPrinter *A, Module *M)
   DwarfInfoSectionSym = DwarfAbbrevSectionSym = 0;
   DwarfStrSectionSym = TextSectionSym = 0;
   DwarfDebugRangeSectionSym = DwarfDebugLocSectionSym = DwarfLineSectionSym = 0;
+  DwarfAddrSectionSym = 0;
   DwarfAbbrevDWOSectionSym = DwarfStrDWOSectionSym = 0;
   FunctionBeginSym = FunctionEndSym = 0;
 
   // Turn on accelerator tables and older gdb compatibility
   // for Darwin.
-  bool IsDarwin = Triple(M->getTargetTriple()).isOSDarwin();
+  bool IsDarwin = Triple(A->getTargetTriple()).isOSDarwin();
   if (DarwinGDBCompat == Default) {
     if (IsDarwin)
       IsDarwinGDBCompat = true;
@@ -596,9 +603,16 @@ DIE *DwarfDebug::constructScopeDIE(CompileUnit *TheCU, LexicalScope *Scope) {
   }
   else {
     // There is no need to emit empty lexical block DIE.
-    if (Children.empty())
+    std::pair<ImportedEntityMap::const_iterator,
+              ImportedEntityMap::const_iterator> Range = std::equal_range(
+        ScopesWithImportedEntities.begin(), ScopesWithImportedEntities.end(),
+        std::pair<const MDNode *, const MDNode *>(DS, (const MDNode*)0),
+        CompareFirst());
+    if (Children.empty() && Range.first == Range.second)
       return NULL;
     ScopeDIE = constructLexicalScopeDIE(TheCU, Scope);
+    for (ImportedEntityMap::const_iterator i = Range.first; i != Range.second; ++i)
+      constructImportedModuleDIE(TheCU, i->second, ScopeDIE);
   }
 
   if (!ScopeDIE) return NULL;
@@ -643,7 +657,7 @@ unsigned DwarfDebug::getOrCreateSourceID(StringRef FileName,
 
   // We look up the CUID/file/dir by concatenating them with a zero byte.
   SmallString<128> NamePair;
-  NamePair += CUID;
+  NamePair += utostr(CUID);
   NamePair += '\0';
   NamePair += DirName;
   NamePair += '\0'; // Zero bytes are not allowed in paths.
@@ -681,9 +695,12 @@ CompileUnit *DwarfDebug::constructCompileUnit(const MDNode *N) {
   NewCU->addUInt(Die, dwarf::DW_AT_language, dwarf::DW_FORM_data2,
                  DIUnit.getLanguage());
   NewCU->addString(Die, dwarf::DW_AT_name, FN);
+
   // 2.17.1 requires that we use DW_AT_low_pc for a single entry point
-  // into an entity. We're using 0 (or a NULL label) for this.
-  NewCU->addLabelAddress(Die, dwarf::DW_AT_low_pc, NULL);
+  // into an entity. We're using 0 (or a NULL label) for this. For
+  // split dwarf it's in the skeleton CU so omit it here.
+  if (!useSplitDwarf())
+    NewCU->addLabelAddress(Die, dwarf::DW_AT_low_pc, NULL);
 
   // Define start line table label for each Compile Unit.
   MCSymbol *LineTableStartSym = Asm->GetTempSymbol("line_table_start",
@@ -691,21 +708,32 @@ CompileUnit *DwarfDebug::constructCompileUnit(const MDNode *N) {
   Asm->OutStreamer.getContext().setMCLineTableSymbol(LineTableStartSym,
                                                      NewCU->getUniqueID());
 
+  // Use a single line table if we are using .loc and generating assembly.
+  bool UseTheFirstCU =
+    (Asm->TM.hasMCUseLoc() &&
+     Asm->OutStreamer.getKind() == MCStreamer::SK_AsmStreamer) ||
+    (NewCU->getUniqueID() == 0);
+
   // DW_AT_stmt_list is a offset of line number information for this
-  // compile unit in debug_line section.
+  // compile unit in debug_line section. For split dwarf this is
+  // left in the skeleton CU and so not included.
   // The line table entries are not always emitted in assembly, so it
   // is not okay to use line_table_start here.
-  if (Asm->MAI->doesDwarfUseRelocationsAcrossSections())
-    NewCU->addLabel(Die, dwarf::DW_AT_stmt_list, dwarf::DW_FORM_data4,
-                    NewCU->getUniqueID() == 0 ?
-                    Asm->GetTempSymbol("section_line") : LineTableStartSym);
-  else if (NewCU->getUniqueID() == 0)
-    NewCU->addUInt(Die, dwarf::DW_AT_stmt_list, dwarf::DW_FORM_data4, 0);
-  else
-    NewCU->addDelta(Die, dwarf::DW_AT_stmt_list, dwarf::DW_FORM_data4,
-                    LineTableStartSym, DwarfLineSectionSym);
+  if (!useSplitDwarf()) {
+    if (Asm->MAI->doesDwarfUseRelocationsAcrossSections())
+      NewCU->addLabel(Die, dwarf::DW_AT_stmt_list, dwarf::DW_FORM_data4,
+                      UseTheFirstCU ?
+                      Asm->GetTempSymbol("section_line") : LineTableStartSym);
+    else if (UseTheFirstCU)
+      NewCU->addUInt(Die, dwarf::DW_AT_stmt_list, dwarf::DW_FORM_data4, 0);
+    else
+      NewCU->addDelta(Die, dwarf::DW_AT_stmt_list, dwarf::DW_FORM_data4,
+                      LineTableStartSym, DwarfLineSectionSym);
+  }
 
-  if (!CompilationDir.empty())
+  // If we're using split dwarf the compilation dir is going to be in the
+  // skeleton CU and so we don't need to duplicate it here.
+  if (!useSplitDwarf() && !CompilationDir.empty())
     NewCU->addString(Die, dwarf::DW_AT_comp_dir, CompilationDir);
   if (DIUnit.isOptimized())
     NewCU->addFlag(Die, dwarf::DW_AT_APPLE_optimized);
@@ -754,6 +782,41 @@ void DwarfDebug::constructSubprogramDIE(CompileUnit *TheCU,
     TheCU->addGlobalName(SP.getName(), SubprogramDie);
 }
 
+void DwarfDebug::constructImportedModuleDIE(CompileUnit *TheCU,
+                                            const MDNode *N) {
+  DIImportedModule Module(N);
+  if (!Module.Verify())
+    return;
+  if (DIE *D = TheCU->getOrCreateContextDIE(Module.getContext()))
+    constructImportedModuleDIE(TheCU, Module, D);
+}
+
+void DwarfDebug::constructImportedModuleDIE(CompileUnit *TheCU, const MDNode *N,
+                                            DIE *Context) {
+  DIImportedModule Module(N);
+  if (!Module.Verify())
+    return;
+  return constructImportedModuleDIE(TheCU, Module, Context);
+}
+
+void DwarfDebug::constructImportedModuleDIE(CompileUnit *TheCU,
+                                            const DIImportedModule &Module,
+                                            DIE *Context) {
+  assert(Module.Verify() &&
+         "Use one of the MDNode * overloads to handle invalid metadata");
+  assert(Context && "Should always have a context for an imported_module");
+  DIE *IMDie = new DIE(dwarf::DW_TAG_imported_module);
+  TheCU->insertDIE(Module, IMDie);
+  DIE *NSDie = TheCU->getOrCreateNameSpace(Module.getNameSpace());
+  unsigned FileID = getOrCreateSourceID(Module.getContext().getFilename(),
+                                        Module.getContext().getDirectory(),
+                                        TheCU->getUniqueID());
+  TheCU->addUInt(IMDie, dwarf::DW_AT_decl_file, 0, FileID);
+  TheCU->addUInt(IMDie, dwarf::DW_AT_decl_line, 0, Module.getLineNumber());
+  TheCU->addDIEEntry(IMDie, dwarf::DW_AT_import, dwarf::DW_FORM_ref4, NSDie);
+  Context->addChild(IMDie);
+}
+
 // Emit all Dwarf sections that should come prior to the content. Create
 // global DIEs and emit initial debug info sections. This is invoked by
 // the target AsmPrinter.
@@ -775,6 +838,13 @@ void DwarfDebug::beginModule() {
   for (unsigned i = 0, e = CU_Nodes->getNumOperands(); i != e; ++i) {
     DICompileUnit CUNode(CU_Nodes->getOperand(i));
     CompileUnit *CU = constructCompileUnit(CUNode);
+    DIArray ImportedModules = CUNode.getImportedModules();
+    for (unsigned i = 0, e = ImportedModules.getNumElements(); i != e; ++i)
+      ScopesWithImportedEntities.push_back(std::make_pair(
+          DIImportedModule(ImportedModules.getElement(i)).getContext(),
+          ImportedModules.getElement(i)));
+    std::sort(ScopesWithImportedEntities.begin(),
+              ScopesWithImportedEntities.end(), CompareFirst());
     DIArray GVs = CUNode.getGlobalVariables();
     for (unsigned i = 0, e = GVs.getNumElements(); i != e; ++i)
       CU->createGlobalVariableDIE(GVs.getElement(i));
@@ -787,11 +857,16 @@ void DwarfDebug::beginModule() {
     DIArray RetainedTypes = CUNode.getRetainedTypes();
     for (unsigned i = 0, e = RetainedTypes.getNumElements(); i != e; ++i)
       CU->getOrCreateTypeDIE(RetainedTypes.getElement(i));
+    // Emit imported_modules last so that the relevant context is already
+    // available.
+    for (unsigned i = 0, e = ImportedModules.getNumElements(); i != e; ++i)
+      constructImportedModuleDIE(CU, ImportedModules.getElement(i));
     // If we're splitting the dwarf out now that we've got the entire
     // CU then construct a skeleton CU based upon it.
     if (useSplitDwarf()) {
-    // This should be a unique identifier when we want to build .dwp files.
-      CU->addUInt(CU->getCUDie(), dwarf::DW_AT_GNU_dwo_id, dwarf::DW_FORM_data8, 0);
+      // This should be a unique identifier when we want to build .dwp files.
+      CU->addUInt(CU->getCUDie(), dwarf::DW_AT_GNU_dwo_id,
+                  dwarf::DW_FORM_data8, 0);
       // Now construct the skeleton CU associated.
       constructSkeletonCU(CUNode);
     }
@@ -1099,7 +1174,13 @@ static DotDebugLocEntry getDebugLocEntry(AsmPrinter *Asm,
   }
   if (MI->getOperand(0).isReg() && MI->getOperand(1).isImm()) {
     MachineLocation MLoc;
-    MLoc.set(MI->getOperand(0).getReg(), MI->getOperand(1).getImm());
+    // TODO: Currently an offset of 0 in a DBG_VALUE means
+    // we need to generate a direct register value.
+    // There is no way to specify an indirect value with offset 0.
+    if (MI->getOperand(1).getImm() == 0)
+      MLoc.set(MI->getOperand(0).getReg());
+    else
+      MLoc.set(MI->getOperand(0).getReg(), MI->getOperand(1).getImm());
     return DotDebugLocEntry(FLabel, SLabel, MLoc, Var);
   }
   if (MI->getOperand(0).isImm())
@@ -1366,7 +1447,12 @@ void DwarfDebug::beginFunction(const MachineFunction *MF) {
   LexicalScope *FnScope = LScopes.getCurrentFunctionScope();
   CompileUnit *TheCU = SPMap.lookup(FnScope->getScopeNode());
   assert(TheCU && "Unable to find compile unit!");
-  Asm->OutStreamer.getContext().setDwarfCompileUnitID(TheCU->getUniqueID());
+  if (Asm->TM.hasMCUseLoc() &&
+      Asm->OutStreamer.getKind() == MCStreamer::SK_AsmStreamer)
+    // Use a single line table if we are using .loc and generating assembly.
+    Asm->OutStreamer.getContext().setDwarfCompileUnitID(0);
+  else
+    Asm->OutStreamer.getContext().setDwarfCompileUnitID(TheCU->getUniqueID());
 
   FunctionBeginSym = Asm->GetTempSymbol("func_begin",
                                         Asm->getFunctionNumber());
@@ -1740,9 +1826,12 @@ void DwarfDebug::emitSectionLabels() {
   emitSectionSym(Asm, TLOF.getDwarfPubTypesSection());
   DwarfStrSectionSym =
     emitSectionSym(Asm, TLOF.getDwarfStrSection(), "info_string");
-  if (useSplitDwarf())
+  if (useSplitDwarf()) {
     DwarfStrDWOSectionSym =
       emitSectionSym(Asm, TLOF.getDwarfStrDWOSection(), "skel_string");
+    DwarfAddrSectionSym =
+      emitSectionSym(Asm, TLOF.getDwarfAddrSection(), "addr_sec");
+  }
   DwarfDebugRangeSectionSym = emitSectionSym(Asm, TLOF.getDwarfRangesSection(),
                                              "debug_range");
 
@@ -2510,9 +2599,14 @@ CompileUnit *DwarfDebug::constructSkeletonCU(const MDNode *N) {
   // This should be a unique identifier when we want to build .dwp files.
   NewCU->addUInt(Die, dwarf::DW_AT_GNU_dwo_id, dwarf::DW_FORM_data8, 0);
 
-  // FIXME: The addr base should be relative for each compile unit, however,
-  // this one is going to be 0 anyhow.
-  NewCU->addUInt(Die, dwarf::DW_AT_GNU_addr_base, dwarf::DW_FORM_sec_offset, 0);
+  // Relocate to the beginning of the addr_base section, else 0 for the
+  // beginning of the one for this compile unit.
+  if (Asm->MAI->doesDwarfUseRelocationsAcrossSections())
+    NewCU->addLabel(Die, dwarf::DW_AT_GNU_addr_base, dwarf::DW_FORM_sec_offset,
+                    DwarfAddrSectionSym);
+  else
+    NewCU->addUInt(Die, dwarf::DW_AT_GNU_addr_base,
+                   dwarf::DW_FORM_sec_offset, 0);
 
   // 2.17.1 requires that we use DW_AT_low_pc for a single entry point
   // into an entity. We're using 0, or a NULL label for this.
@@ -2520,6 +2614,7 @@ CompileUnit *DwarfDebug::constructSkeletonCU(const MDNode *N) {
 
   // DW_AT_stmt_list is a offset of line number information for this
   // compile unit in debug_line section.
+  // FIXME: Should handle multiple compile units.
   if (Asm->MAI->doesDwarfUseRelocationsAcrossSections())
     NewCU->addLabel(Die, dwarf::DW_AT_stmt_list, dwarf::DW_FORM_sec_offset,
                     DwarfLineSectionSym);
diff --git a/lib/CodeGen/AsmPrinter/DwarfDebug.h b/lib/CodeGen/AsmPrinter/DwarfDebug.h
index 81e345e6281d..24f758dda9c8 100644
--- a/lib/CodeGen/AsmPrinter/DwarfDebug.h
+++ b/lib/CodeGen/AsmPrinter/DwarfDebug.h
@@ -392,7 +392,7 @@ class DwarfDebug {
   // section offsets and are created by EmitSectionLabels.
   MCSymbol *DwarfInfoSectionSym, *DwarfAbbrevSectionSym;
   MCSymbol *DwarfStrSectionSym, *TextSectionSym, *DwarfDebugRangeSectionSym;
-  MCSymbol *DwarfDebugLocSectionSym, *DwarfLineSectionSym;
+  MCSymbol *DwarfDebugLocSectionSym, *DwarfLineSectionSym, *DwarfAddrSectionSym;
   MCSymbol *FunctionBeginSym, *FunctionEndSym;
   MCSymbol *DwarfAbbrevDWOSectionSym, *DwarfStrDWOSectionSym;
 
@@ -433,6 +433,10 @@ class DwarfDebug {
   // Holder for the skeleton information.
   DwarfUnits SkeletonHolder;
 
+  typedef SmallVector<std::pair<const MDNode *, const MDNode *>, 32>
+    ImportedEntityMap;
+  ImportedEntityMap ScopesWithImportedEntities;
+
 private:
 
   void addScopeVariable(LexicalScope *LS, DbgVariable *Var);
@@ -555,6 +559,18 @@ private:
   /// \brief Construct subprogram DIE.
   void constructSubprogramDIE(CompileUnit *TheCU, const MDNode *N);
 
+  /// \brief Construct import_module DIE.
+  void constructImportedModuleDIE(CompileUnit *TheCU, const MDNode *N);
+
+  /// \brief Construct import_module DIE.
+  void constructImportedModuleDIE(CompileUnit *TheCU, const MDNode *N,
+                                  DIE *Context);
+
+  /// \brief Construct import_module DIE.
+  void constructImportedModuleDIE(CompileUnit *TheCU,
+                                  const DIImportedModule &Module,
+                                  DIE *Context);
+
   /// \brief Register a source line with debug info. Returns the unique
   /// label that was emitted and which provides correspondence to the
   /// source line list.
diff --git a/lib/CodeGen/BasicTargetTransformInfo.cpp b/lib/CodeGen/BasicTargetTransformInfo.cpp
index 012ff8ad8339..4a99184f5eec 100644
--- a/lib/CodeGen/BasicTargetTransformInfo.cpp
+++ b/lib/CodeGen/BasicTargetTransformInfo.cpp
@@ -204,20 +204,25 @@ unsigned BasicTTI::getArithmeticInstrCost(unsigned Opcode, Type *Ty,
 
   std::pair<unsigned, MVT> LT = TLI->getTypeLegalizationCost(Ty);
 
+  bool IsFloat = Ty->getScalarType()->isFloatingPointTy();
+  // Assume that floating point arithmetic operations cost twice as much as
+  // integer operations.
+  unsigned OpCost = (IsFloat ? 2 : 1);
+
   if (TLI->isOperationLegalOrPromote(ISD, LT.second)) {
     // The operation is legal. Assume it costs 1.
-    // If the type is split to multiple registers, assume that thre is some
+    // If the type is split to multiple registers, assume that there is some
     // overhead to this.
     // TODO: Once we have extract/insert subvector cost we need to use them.
     if (LT.first > 1)
-      return LT.first * 2;
-    return LT.first * 1;
+      return LT.first * 2 * OpCost;
+    return LT.first * 1 * OpCost;
   }
 
   if (!TLI->isOperationExpand(ISD, LT.second)) {
     // If the operation is custom lowered then assume
     // thare the code is twice as expensive.
-    return LT.first * 2;
+    return LT.first * 2 * OpCost;
   }
 
   // Else, assume that we need to scalarize this op.
@@ -230,7 +235,7 @@ unsigned BasicTTI::getArithmeticInstrCost(unsigned Opcode, Type *Ty,
   }
 
   // We don't know anything about this scalar instruction.
-  return 1;
+  return OpCost;
 }
 
 unsigned BasicTTI::getShuffleCost(ShuffleKind Kind, Type *Tp, int Index,
diff --git a/lib/CodeGen/CalcSpillWeights.cpp b/lib/CodeGen/CalcSpillWeights.cpp
index dee339a45863..38ae17d23156 100644
--- a/lib/CodeGen/CalcSpillWeights.cpp
+++ b/lib/CodeGen/CalcSpillWeights.cpp
@@ -117,7 +117,7 @@ void VirtRegAuxInfo::CalculateWeightAndHint(LiveInterval &li) {
   float totalWeight = 0;
   SmallPtrSet<MachineInstr*, 8> visited;
 
-  // Find the best physreg hist and the best virtreg hint.
+  // Find the best physreg hint and the best virtreg hint.
   float bestPhys = 0, bestVirt = 0;
   unsigned hintPhys = 0, hintVirt = 0;
 
diff --git a/lib/CodeGen/CallingConvLower.cpp b/lib/CodeGen/CallingConvLower.cpp
index f1d4ace92273..75f4b96e3b03 100644
--- a/lib/CodeGen/CallingConvLower.cpp
+++ b/lib/CodeGen/CallingConvLower.cpp
@@ -32,7 +32,7 @@ CCState::CCState(CallingConv::ID CC, bool isVarArg, MachineFunction &mf,
   // No stack is used.
   StackOffset = 0;
 
-  clearFirstByValReg();
+  clearByValRegsInfo();
   UsedRegs.resize((TRI.getNumRegs()+31)/32);
 }
 
diff --git a/lib/CodeGen/CodeGen.cpp b/lib/CodeGen/CodeGen.cpp
index 35ec68d00cec..c641991d408d 100644
--- a/lib/CodeGen/CodeGen.cpp
+++ b/lib/CodeGen/CodeGen.cpp
@@ -13,6 +13,7 @@
 //===----------------------------------------------------------------------===//
 
 #include "llvm/InitializePasses.h"
+#include "llvm/PassRegistry.h"
 #include "llvm-c/Initialization.h"
 
 using namespace llvm;
diff --git a/lib/CodeGen/IfConversion.cpp b/lib/CodeGen/IfConversion.cpp
index 9958d7daada8..8264d6dbab81 100644
--- a/lib/CodeGen/IfConversion.cpp
+++ b/lib/CodeGen/IfConversion.cpp
@@ -1039,6 +1039,10 @@ bool IfConverter::IfConvertSimple(BBInfo &BBI, IfcvtKind Kind) {
     return false;
   }
 
+  if (CvtBBI->BB->hasAddressTaken())
+    // Conservatively abort if-conversion if BB's address is taken.
+    return false;
+
   if (Kind == ICSimpleFalse)
     if (TII->ReverseBranchCondition(Cond))
       llvm_unreachable("Unable to reverse branch condition!");
@@ -1054,6 +1058,10 @@ bool IfConverter::IfConvertSimple(BBInfo &BBI, IfcvtKind Kind) {
     // Copy instructions in the true block, predicate them, and add them to
     // the entry block.
     CopyAndPredicateBlock(BBI, *CvtBBI, Cond, Redefs);
+
+    // RemoveExtraEdges won't work if the block has an unanalyzable branch, so
+    // explicitly remove CvtBBI as a successor.
+    BBI.BB->removeSuccessor(CvtBBI->BB);
   } else {
     PredicateBlock(*CvtBBI, CvtBBI->BB->end(), Cond, Redefs);
 
@@ -1112,6 +1120,10 @@ bool IfConverter::IfConvertTriangle(BBInfo &BBI, IfcvtKind Kind) {
     return false;
   }
 
+  if (CvtBBI->BB->hasAddressTaken())
+    // Conservatively abort if-conversion if BB's address is taken.
+    return false;
+
   if (Kind == ICTriangleFalse || Kind == ICTriangleFRev)
     if (TII->ReverseBranchCondition(Cond))
       llvm_unreachable("Unable to reverse branch condition!");
@@ -1146,6 +1158,10 @@ bool IfConverter::IfConvertTriangle(BBInfo &BBI, IfcvtKind Kind) {
     // Copy instructions in the true block, predicate them, and add them to
     // the entry block.
     CopyAndPredicateBlock(BBI, *CvtBBI, Cond, Redefs, true);
+
+    // RemoveExtraEdges won't work if the block has an unanalyzable branch, so
+    // explicitly remove CvtBBI as a successor.
+    BBI.BB->removeSuccessor(CvtBBI->BB);
   } else {
     // Predicate the 'true' block after removing its branch.
     CvtBBI->NonPredSize -= TII->RemoveBranch(*CvtBBI->BB);
@@ -1176,7 +1192,8 @@ bool IfConverter::IfConvertTriangle(BBInfo &BBI, IfcvtKind Kind) {
     // block. By not merging them, we make it possible to iteratively
     // ifcvt the blocks.
     if (!HasEarlyExit &&
-        NextBBI->BB->pred_size() == 1 && !NextBBI->HasFallThrough) {
+        NextBBI->BB->pred_size() == 1 && !NextBBI->HasFallThrough &&
+        !NextBBI->BB->hasAddressTaken()) {
       MergeBlocks(BBI, *NextBBI);
       FalseBBDead = true;
     } else {
@@ -1226,6 +1243,10 @@ bool IfConverter::IfConvertDiamond(BBInfo &BBI, IfcvtKind Kind,
     return false;
   }
 
+  if (TrueBBI.BB->hasAddressTaken() || FalseBBI.BB->hasAddressTaken())
+    // Conservatively abort if-conversion if either BB has its address taken.
+    return false;
+
   // Put the predicated instructions from the 'true' block before the
   // instructions from the 'false' block, unless the true block would clobber
   // the predicate, in which case, do the opposite.
@@ -1374,7 +1395,8 @@ bool IfConverter::IfConvertDiamond(BBInfo &BBI, IfcvtKind Kind,
   // tail, add a unconditional branch to it.
   if (TailBB) {
     BBInfo &TailBBI = BBAnalysis[TailBB->getNumber()];
-    bool CanMergeTail = !TailBBI.HasFallThrough;
+    bool CanMergeTail = !TailBBI.HasFallThrough &&
+      !TailBBI.BB->hasAddressTaken();
     // There may still be a fall-through edge from BBI1 or BBI2 to TailBB;
     // check if there are any other predecessors besides those.
     unsigned NumPreds = TailBB->pred_size();
@@ -1543,6 +1565,9 @@ void IfConverter::CopyAndPredicateBlock(BBInfo &ToBBI, BBInfo &FromBBI,
 /// i.e., when FromBBI's branch is being moved, add those successor edges to
 /// ToBBI.
 void IfConverter::MergeBlocks(BBInfo &ToBBI, BBInfo &FromBBI, bool AddEdges) {
+  assert(!FromBBI.BB->hasAddressTaken() &&
+         "Removing a BB whose address is taken!");
+
   ToBBI.BB->splice(ToBBI.BB->end(),
                    FromBBI.BB, FromBBI.BB->begin(), FromBBI.BB->end());
 
diff --git a/lib/CodeGen/InlineSpiller.cpp b/lib/CodeGen/InlineSpiller.cpp
index c6d1a18dbd06..35295fe85886 100644
--- a/lib/CodeGen/InlineSpiller.cpp
+++ b/lib/CodeGen/InlineSpiller.cpp
@@ -955,18 +955,21 @@ void InlineSpiller::reMaterializeAll() {
   Edit->eliminateDeadDefs(DeadDefs, RegsToSpill);
 
   // Get rid of deleted and empty intervals.
-  for (unsigned i = RegsToSpill.size(); i != 0; --i) {
-    unsigned Reg = RegsToSpill[i-1];
-    if (!LIS.hasInterval(Reg)) {
-      RegsToSpill.erase(RegsToSpill.begin() + (i - 1));
+  unsigned ResultPos = 0;
+  for (unsigned i = 0, e = RegsToSpill.size(); i != e; ++i) {
+    unsigned Reg = RegsToSpill[i];
+    if (!LIS.hasInterval(Reg))
       continue;
-    }
+
     LiveInterval &LI = LIS.getInterval(Reg);
-    if (!LI.empty())
+    if (LI.empty()) {
+      Edit->eraseVirtReg(Reg);
       continue;
-    Edit->eraseVirtReg(Reg);
-    RegsToSpill.erase(RegsToSpill.begin() + (i - 1));
+    }
+
+    RegsToSpill[ResultPos++] = Reg;
   }
+  RegsToSpill.erase(RegsToSpill.begin() + ResultPos, RegsToSpill.end());
   DEBUG(dbgs() << RegsToSpill.size() << " registers to spill after remat.\n");
 }
 
diff --git a/lib/CodeGen/IntrinsicLowering.cpp b/lib/CodeGen/IntrinsicLowering.cpp
index 07f0ccf52f8c..d894f664dcb3 100644
--- a/lib/CodeGen/IntrinsicLowering.cpp
+++ b/lib/CodeGen/IntrinsicLowering.cpp
@@ -453,6 +453,12 @@ void IntrinsicLowering::LowerIntrinsicCall(CallInst *CI) {
     CI->replaceAllUsesWith(ConstantInt::get(CI->getType(), 1));
     break;
 
+  case Intrinsic::annotation:
+  case Intrinsic::ptr_annotation:
+    // Just drop the annotation, but forward the value
+    CI->replaceAllUsesWith(CI->getOperand(0));
+    break;
+
   case Intrinsic::var_annotation:
     break;   // Strip out annotate intrinsic
     
diff --git a/lib/CodeGen/LocalStackSlotAllocation.cpp b/lib/CodeGen/LocalStackSlotAllocation.cpp
index 352ef942591f..26a117652b08 100644
--- a/lib/CodeGen/LocalStackSlotAllocation.cpp
+++ b/lib/CodeGen/LocalStackSlotAllocation.cpp
@@ -46,13 +46,16 @@ namespace {
   class FrameRef {
     MachineBasicBlock::iterator MI; // Instr referencing the frame
     int64_t LocalOffset;            // Local offset of the frame idx referenced
+    int FrameIdx;                   // The frame index
   public:
-    FrameRef(MachineBasicBlock::iterator I, int64_t Offset) :
-      MI(I), LocalOffset(Offset) {}
+    FrameRef(MachineBasicBlock::iterator I, int64_t Offset, int Idx) :
+      MI(I), LocalOffset(Offset), FrameIdx(Idx) {}
     bool operator<(const FrameRef &RHS) const {
       return LocalOffset < RHS.LocalOffset;
     }
-    MachineBasicBlock::iterator getMachineInstr() { return MI; }
+    MachineBasicBlock::iterator getMachineInstr() const { return MI; }
+    int64_t getLocalOffset() const { return LocalOffset; }
+    int getFrameIndex() const { return FrameIdx; }
   };
 
   class LocalStackSlotPass: public MachineFunctionPass {
@@ -194,22 +197,15 @@ void LocalStackSlotPass::calculateFrameObjectOffsets(MachineFunction &Fn) {
 }
 
 static inline bool
-lookupCandidateBaseReg(const SmallVector<std::pair<unsigned, int64_t>, 8> &Regs,
-                       std::pair<unsigned, int64_t> &RegOffset,
+lookupCandidateBaseReg(int64_t BaseOffset,
                        int64_t FrameSizeAdjust,
                        int64_t LocalFrameOffset,
                        const MachineInstr *MI,
                        const TargetRegisterInfo *TRI) {
-  unsigned e = Regs.size();
-  for (unsigned i = 0; i < e; ++i) {
-    RegOffset = Regs[i];
-    // Check if the relative offset from the where the base register references
-    // to the target address is in range for the instruction.
-    int64_t Offset = FrameSizeAdjust + LocalFrameOffset - RegOffset.second;
-    if (TRI->isFrameOffsetLegal(MI, Offset))
-      return true;
-  }
-  return false;
+  // Check if the relative offset from the where the base register references
+  // to the target address is in range for the instruction.
+  int64_t Offset = FrameSizeAdjust + LocalFrameOffset - BaseOffset;
+  return TRI->isFrameOffsetLegal(MI, Offset);
 }
 
 bool LocalStackSlotPass::insertFrameReferenceRegisters(MachineFunction &Fn) {
@@ -233,9 +229,6 @@ bool LocalStackSlotPass::insertFrameReferenceRegisters(MachineFunction &Fn) {
   // choose the first one).
   SmallVector<FrameRef, 64> FrameReferenceInsns;
 
-  // A base register definition is a register + offset pair.
-  SmallVector<std::pair<unsigned, int64_t>, 8> BaseRegisters;
-
   for (MachineFunction::iterator BB = Fn.begin(), E = Fn.end(); BB != E; ++BB) {
     for (MachineBasicBlock::iterator I = BB->begin(); I != BB->end(); ++I) {
       MachineInstr *MI = I;
@@ -258,8 +251,12 @@ bool LocalStackSlotPass::insertFrameReferenceRegisters(MachineFunction &Fn) {
           // Don't try this with values not in the local block.
           if (!MFI->isObjectPreAllocated(MI->getOperand(i).getIndex()))
             break;
+          int Idx = MI->getOperand(i).getIndex();
+          int64_t LocalOffset = LocalOffsets[Idx];
+          if (!TRI->needsFrameBaseReg(MI, LocalOffset))
+            break;
           FrameReferenceInsns.
-            push_back(FrameRef(MI, LocalOffsets[MI->getOperand(i).getIndex()]));
+            push_back(FrameRef(MI, LocalOffset, Idx));
           break;
         }
       }
@@ -271,86 +268,106 @@ bool LocalStackSlotPass::insertFrameReferenceRegisters(MachineFunction &Fn) {
 
   MachineBasicBlock *Entry = Fn.begin();
 
+  unsigned BaseReg = 0;
+  int64_t BaseOffset = 0;
+
   // Loop through the frame references and allocate for them as necessary.
   for (int ref = 0, e = FrameReferenceInsns.size(); ref < e ; ++ref) {
-    MachineBasicBlock::iterator I =
-      FrameReferenceInsns[ref].getMachineInstr();
+    FrameRef &FR = FrameReferenceInsns[ref];
+    MachineBasicBlock::iterator I = FR.getMachineInstr();
     MachineInstr *MI = I;
-    for (unsigned idx = 0, e = MI->getNumOperands(); idx != e; ++idx) {
-      // Consider replacing all frame index operands that reference
-      // an object allocated in the local block.
-      if (MI->getOperand(idx).isFI()) {
-        int FrameIdx = MI->getOperand(idx).getIndex();
-
-        assert(MFI->isObjectPreAllocated(FrameIdx) &&
-               "Only pre-allocated locals expected!");
-
-        DEBUG(dbgs() << "Considering: " << *MI);
-        if (TRI->needsFrameBaseReg(MI, LocalOffsets[FrameIdx])) {
-          unsigned BaseReg = 0;
-          int64_t Offset = 0;
-          int64_t FrameSizeAdjust =
-            StackGrowsDown ? MFI->getLocalFrameSize() : 0;
-
-          DEBUG(dbgs() << "  Replacing FI in: " << *MI);
-
-          // If we have a suitable base register available, use it; otherwise
-          // create a new one. Note that any offset encoded in the
-          // instruction itself will be taken into account by the target,
-          // so we don't have to adjust for it here when reusing a base
-          // register.
-          std::pair<unsigned, int64_t> RegOffset;
-          if (lookupCandidateBaseReg(BaseRegisters, RegOffset,
-                                     FrameSizeAdjust,
-                                     LocalOffsets[FrameIdx],
-                                     MI, TRI)) {
-            DEBUG(dbgs() << "  Reusing base register " <<
-                  RegOffset.first << "\n");
-            // We found a register to reuse.
-            BaseReg = RegOffset.first;
-            Offset = FrameSizeAdjust + LocalOffsets[FrameIdx] -
-              RegOffset.second;
-          } else {
-            // No previously defined register was in range, so create a
-            // new one.
-            int64_t InstrOffset = TRI->getFrameIndexInstrOffset(MI, idx);
-            const MachineFunction *MF = MI->getParent()->getParent();
-            const TargetRegisterClass *RC = TRI->getPointerRegClass(*MF);
-            BaseReg = Fn.getRegInfo().createVirtualRegister(RC);
-
-            DEBUG(dbgs() << "  Materializing base register " << BaseReg <<
-                  " at frame local offset " <<
-                  LocalOffsets[FrameIdx] + InstrOffset << "\n");
-
-            // Tell the target to insert the instruction to initialize
-            // the base register.
-            //            MachineBasicBlock::iterator InsertionPt = Entry->begin();
-            TRI->materializeFrameBaseRegister(Entry, BaseReg, FrameIdx,
-                                              InstrOffset);
-
-            // The base register already includes any offset specified
-            // by the instruction, so account for that so it doesn't get
-            // applied twice.
-            Offset = -InstrOffset;
-
-            int64_t BaseOffset = FrameSizeAdjust + LocalOffsets[FrameIdx] +
-              InstrOffset;
-            BaseRegisters.push_back(
-              std::pair<unsigned, int64_t>(BaseReg, BaseOffset));
-            ++NumBaseRegisters;
-            UsedBaseReg = true;
-          }
-          assert(BaseReg != 0 && "Unable to allocate virtual base register!");
-
-          // Modify the instruction to use the new base register rather
-          // than the frame index operand.
-          TRI->resolveFrameIndex(I, BaseReg, Offset);
-          DEBUG(dbgs() << "Resolved: " << *MI);
-
-          ++NumReplacements;
-        }
+    int64_t LocalOffset = FR.getLocalOffset();
+    int FrameIdx = FR.getFrameIndex();
+    assert(MFI->isObjectPreAllocated(FrameIdx) &&
+           "Only pre-allocated locals expected!");
+
+    DEBUG(dbgs() << "Considering: " << *MI);
+
+    unsigned idx = 0;
+    for (unsigned f = MI->getNumOperands(); idx != f; ++idx) {
+      if (!MI->getOperand(idx).isFI())
+        continue;
+
+      if (FrameIdx == I->getOperand(idx).getIndex())
+        break;
+    }
+
+    assert(idx < MI->getNumOperands() && "Cannot find FI operand");
+
+    int64_t Offset = 0;
+    int64_t FrameSizeAdjust = StackGrowsDown ? MFI->getLocalFrameSize() : 0;
+
+    DEBUG(dbgs() << "  Replacing FI in: " << *MI);
+
+    // If we have a suitable base register available, use it; otherwise
+    // create a new one. Note that any offset encoded in the
+    // instruction itself will be taken into account by the target,
+    // so we don't have to adjust for it here when reusing a base
+    // register.
+    if (UsedBaseReg && lookupCandidateBaseReg(BaseOffset, FrameSizeAdjust,
+                                              LocalOffset, MI, TRI)) {
+      DEBUG(dbgs() << "  Reusing base register " << BaseReg << "\n");
+      // We found a register to reuse.
+      Offset = FrameSizeAdjust + LocalOffset - BaseOffset;
+    } else {
+      // No previously defined register was in range, so create a // new one.
+ 
+      int64_t InstrOffset = TRI->getFrameIndexInstrOffset(MI, idx);
+
+      int64_t PrevBaseOffset = BaseOffset;
+      BaseOffset = FrameSizeAdjust + LocalOffset + InstrOffset;
+
+      // We'd like to avoid creating single-use virtual base registers.
+      // Because the FrameRefs are in sorted order, and we've already
+      // processed all FrameRefs before this one, just check whether or not
+      // the next FrameRef will be able to reuse this new register. If not,
+      // then don't bother creating it.
+      bool CanReuse = false;
+      for (int refn = ref + 1; refn < e; ++refn) {
+        FrameRef &FRN = FrameReferenceInsns[refn];
+        MachineBasicBlock::iterator J = FRN.getMachineInstr();
+        MachineInstr *MIN = J;
+
+        CanReuse = lookupCandidateBaseReg(BaseOffset, FrameSizeAdjust,
+                                          FRN.getLocalOffset(), MIN, TRI);
+        break;
       }
+
+      if (!CanReuse) {
+        BaseOffset = PrevBaseOffset;
+        continue;
+      }
+
+      const MachineFunction *MF = MI->getParent()->getParent();
+      const TargetRegisterClass *RC = TRI->getPointerRegClass(*MF);
+      BaseReg = Fn.getRegInfo().createVirtualRegister(RC);
+
+      DEBUG(dbgs() << "  Materializing base register " << BaseReg <<
+            " at frame local offset " << LocalOffset + InstrOffset << "\n");
+
+      // Tell the target to insert the instruction to initialize
+      // the base register.
+      //            MachineBasicBlock::iterator InsertionPt = Entry->begin();
+      TRI->materializeFrameBaseRegister(Entry, BaseReg, FrameIdx,
+                                        InstrOffset);
+
+      // The base register already includes any offset specified
+      // by the instruction, so account for that so it doesn't get
+      // applied twice.
+      Offset = -InstrOffset;
+
+      ++NumBaseRegisters;
+      UsedBaseReg = true;
     }
+    assert(BaseReg != 0 && "Unable to allocate virtual base register!");
+
+    // Modify the instruction to use the new base register rather
+    // than the frame index operand.
+    TRI->resolveFrameIndex(I, BaseReg, Offset);
+    DEBUG(dbgs() << "Resolved: " << *MI);
+
+    ++NumReplacements;
   }
+
   return UsedBaseReg;
 }
diff --git a/lib/CodeGen/MachineBasicBlock.cpp b/lib/CodeGen/MachineBasicBlock.cpp
index 898e165feeab..78e9950e5ef5 100644
--- a/lib/CodeGen/MachineBasicBlock.cpp
+++ b/lib/CodeGen/MachineBasicBlock.cpp
@@ -37,7 +37,7 @@ using namespace llvm;
 
 MachineBasicBlock::MachineBasicBlock(MachineFunction &mf, const BasicBlock *bb)
   : BB(bb), Number(-1), xParent(&mf), Alignment(0), IsLandingPad(false),
-    AddressTaken(false) {
+    AddressTaken(false), CachedMCSymbol(NULL) {
   Insts.Parent = this;
 }
 
@@ -48,12 +48,16 @@ MachineBasicBlock::~MachineBasicBlock() {
 /// getSymbol - Return the MCSymbol for this basic block.
 ///
 MCSymbol *MachineBasicBlock::getSymbol() const {
-  const MachineFunction *MF = getParent();
-  MCContext &Ctx = MF->getContext();
-  const char *Prefix = Ctx.getAsmInfo().getPrivateGlobalPrefix();
-  return Ctx.GetOrCreateSymbol(Twine(Prefix) + "BB" +
-                               Twine(MF->getFunctionNumber()) + "_" +
-                               Twine(getNumber()));
+  if (!CachedMCSymbol) {
+    const MachineFunction *MF = getParent();
+    MCContext &Ctx = MF->getContext();
+    const char *Prefix = Ctx.getAsmInfo().getPrivateGlobalPrefix();
+    CachedMCSymbol = Ctx.GetOrCreateSymbol(Twine(Prefix) + "BB" +
+                                           Twine(MF->getFunctionNumber()) +
+                                           "_" + Twine(getNumber()));
+  }
+
+  return CachedMCSymbol;
 }
 
 
diff --git a/lib/CodeGen/MachineBlockPlacement.cpp b/lib/CodeGen/MachineBlockPlacement.cpp
index cd948e24a6b2..bfba503b351e 100644
--- a/lib/CodeGen/MachineBlockPlacement.cpp
+++ b/lib/CodeGen/MachineBlockPlacement.cpp
@@ -39,6 +39,7 @@
 #include "llvm/CodeGen/MachineLoopInfo.h"
 #include "llvm/CodeGen/MachineModuleInfo.h"
 #include "llvm/Support/Allocator.h"
+#include "llvm/Support/CommandLine.h"
 #include "llvm/Support/Debug.h"
 #include "llvm/Target/TargetInstrInfo.h"
 #include "llvm/Target/TargetLowering.h"
@@ -52,6 +53,11 @@ STATISTIC(CondBranchTakenFreq,
 STATISTIC(UncondBranchTakenFreq,
           "Potential frequency of taking unconditional branches");
 
+static cl::opt<unsigned> AlignAllBlock("align-all-blocks",
+                                       cl::desc("Force the alignment of all "
+                                                "blocks in the function."),
+                                       cl::init(0), cl::Hidden);
+
 namespace {
 class BlockChain;
 /// \brief Type for our function-wide basic block -> block chain mapping.
@@ -1088,6 +1094,12 @@ bool MachineBlockPlacement::runOnMachineFunction(MachineFunction &F) {
   BlockToChain.clear();
   ChainAllocator.DestroyAll();
 
+  if (AlignAllBlock)
+    // Align all of the blocks in the function to a specific alignment.
+    for (MachineFunction::iterator FI = F.begin(), FE = F.end();
+         FI != FE; ++FI)
+      FI->setAlignment(AlignAllBlock);
+
   // We always return true as we have no way to track whether the final order
   // differs from the original order.
   return true;
diff --git a/lib/CodeGen/MachineModuleInfo.cpp b/lib/CodeGen/MachineModuleInfo.cpp
index 0ea9ae0fcc89..8af9d053b12a 100644
--- a/lib/CodeGen/MachineModuleInfo.cpp
+++ b/lib/CodeGen/MachineModuleInfo.cpp
@@ -326,8 +326,7 @@ void MachineModuleInfo::AnalyzeModule(const Module &M) {
   if (!GV || !GV->hasInitializer()) return;
 
   // Should be an array of 'i8*'.
-  const ConstantArray *InitList = dyn_cast<ConstantArray>(GV->getInitializer());
-  if (InitList == 0) return;
+  const ConstantArray *InitList = cast<ConstantArray>(GV->getInitializer());
 
   for (unsigned i = 0, e = InitList->getNumOperands(); i != e; ++i)
     if (const Function *F =
diff --git a/lib/CodeGen/MachineRegisterInfo.cpp b/lib/CodeGen/MachineRegisterInfo.cpp
index 1af00e84a6ed..68372f6c9065 100644
--- a/lib/CodeGen/MachineRegisterInfo.cpp
+++ b/lib/CodeGen/MachineRegisterInfo.cpp
@@ -15,6 +15,8 @@
 #include "llvm/CodeGen/MachineInstrBuilder.h"
 #include "llvm/Target/TargetInstrInfo.h"
 #include "llvm/Target/TargetMachine.h"
+#include "llvm/Support/raw_os_ostream.h"
+
 using namespace llvm;
 
 MachineRegisterInfo::MachineRegisterInfo(const TargetRegisterInfo &TRI)
@@ -106,13 +108,59 @@ MachineRegisterInfo::createVirtualRegister(const TargetRegisterClass *RegClass){
 /// clearVirtRegs - Remove all virtual registers (after physreg assignment).
 void MachineRegisterInfo::clearVirtRegs() {
 #ifndef NDEBUG
-  for (unsigned i = 0, e = getNumVirtRegs(); i != e; ++i)
-    assert(VRegInfo[TargetRegisterInfo::index2VirtReg(i)].second == 0 &&
-           "Vreg use list non-empty still?");
+  for (unsigned i = 0, e = getNumVirtRegs(); i != e; ++i) {
+    unsigned Reg = TargetRegisterInfo::index2VirtReg(i);
+    if (!VRegInfo[Reg].second)
+      continue;
+    verifyUseList(Reg);
+    llvm_unreachable("Remaining virtual register operands");
+  }
 #endif
   VRegInfo.clear();
 }
 
+void MachineRegisterInfo::verifyUseList(unsigned Reg) const {
+#ifndef NDEBUG
+  bool Valid = true;
+  for (reg_iterator I = reg_begin(Reg), E = reg_end(); I != E; ++I) {
+    MachineOperand *MO = &I.getOperand();
+    MachineInstr *MI = MO->getParent();
+    if (!MI) {
+      errs() << PrintReg(Reg, TRI) << " use list MachineOperand " << MO
+             << " has no parent instruction.\n";
+      Valid = false;
+    }
+    MachineOperand *MO0 = &MI->getOperand(0);
+    unsigned NumOps = MI->getNumOperands();
+    if (!(MO >= MO0 && MO < MO0+NumOps)) {
+      errs() << PrintReg(Reg, TRI) << " use list MachineOperand " << MO
+             << " doesn't belong to parent MI: " << *MI;
+      Valid = false;
+    }
+    if (!MO->isReg()) {
+      errs() << PrintReg(Reg, TRI) << " MachineOperand " << MO << ": " << *MO
+             << " is not a register\n";
+      Valid = false;
+    }
+    if (MO->getReg() != Reg) {
+      errs() << PrintReg(Reg, TRI) << " use-list MachineOperand " << MO << ": "
+             << *MO << " is the wrong register\n";
+      Valid = false;
+    }
+  }
+  assert(Valid && "Invalid use list");
+#endif
+}
+
+void MachineRegisterInfo::verifyUseLists() const {
+#ifndef NDEBUG
+  for (unsigned i = 0, e = getNumVirtRegs(); i != e; ++i)
+    verifyUseList(TargetRegisterInfo::index2VirtReg(i));
+  for (unsigned i = 1, e = TRI->getNumRegs(); i != e; ++i)
+    verifyUseList(i);
+#endif
+}
+
 /// Add MO to the linked list of operands for its register.
 void MachineRegisterInfo::addRegOperandToUseList(MachineOperand *MO) {
   assert(!MO->isOnRegUseList() && "Already on list");
diff --git a/lib/CodeGen/MachineScheduler.cpp b/lib/CodeGen/MachineScheduler.cpp
index 5bd2349b50f6..fff6b2b4c062 100644
--- a/lib/CodeGen/MachineScheduler.cpp
+++ b/lib/CodeGen/MachineScheduler.cpp
@@ -51,7 +51,11 @@ static cl::opt<unsigned> MISchedCutoff("misched-cutoff", cl::Hidden,
 static bool ViewMISchedDAGs = false;
 #endif // NDEBUG
 
-// Experimental heuristics
+// FIXME: remove this flag after initial testing. It should always be a good
+// thing.
+static cl::opt<bool> EnableCopyConstrain("misched-vcopy", cl::Hidden,
+    cl::desc("Constrain vreg copies."), cl::init(true));
+
 static cl::opt<bool> EnableLoadCluster("misched-cluster", cl::Hidden,
   cl::desc("Enable load clustering."), cl::init(true));
 
@@ -323,6 +327,10 @@ ScheduleDAGMI::~ScheduleDAGMI() {
   delete SchedImpl;
 }
 
+bool ScheduleDAGMI::canAddEdge(SUnit *SuccSU, SUnit *PredSU) {
+  return SuccSU == &ExitSU || !Topo.IsReachable(PredSU, SuccSU);
+}
+
 bool ScheduleDAGMI::addEdge(SUnit *SuccSU, const SDep &PredDep) {
   if (SuccSU != &ExitSU) {
     // Do not use WillCreateCycle, it assumes SD scheduling.
@@ -404,6 +412,8 @@ void ScheduleDAGMI::releasePredecessors(SUnit *SU) {
   }
 }
 
+/// This is normally called from the main scheduler loop but may also be invoked
+/// by the scheduling strategy to perform additional code motion.
 void ScheduleDAGMI::moveInstruction(MachineInstr *MI,
                                     MachineBasicBlock::iterator InsertPos) {
   // Advance RegionBegin if the first instruction moves down.
@@ -505,6 +515,14 @@ updateScheduledPressure(const std::vector<unsigned> &NewMaxPressure) {
     if ((int)NewMaxPressure[ID] > MaxUnits)
       MaxUnits = NewMaxPressure[ID];
   }
+  DEBUG(
+    for (unsigned i = 0, e = NewMaxPressure.size(); i < e; ++i) {
+      unsigned Limit = TRI->getRegPressureSetLimit(i);
+      if (NewMaxPressure[i] > Limit ) {
+        dbgs() << "  " << TRI->getRegPressureSetName(i) << ": "
+               << NewMaxPressure[i] << " > " << Limit << "\n";
+      }
+    });
 }
 
 /// schedule - Called back from MachineScheduler::runOnMachineFunction
@@ -905,6 +923,184 @@ void MacroFusion::apply(ScheduleDAGMI *DAG) {
 }
 
 //===----------------------------------------------------------------------===//
+// CopyConstrain - DAG post-processing to encourage copy elimination.
+//===----------------------------------------------------------------------===//
+
+namespace {
+/// \brief Post-process the DAG to create weak edges from all uses of a copy to
+/// the one use that defines the copy's source vreg, most likely an induction
+/// variable increment.
+class CopyConstrain : public ScheduleDAGMutation {
+  // Transient state.
+  SlotIndex RegionBeginIdx;
+  // RegionEndIdx is the slot index of the last non-debug instruction in the
+  // scheduling region. So we may have RegionBeginIdx == RegionEndIdx.
+  SlotIndex RegionEndIdx;
+public:
+  CopyConstrain(const TargetInstrInfo *, const TargetRegisterInfo *) {}
+
+  virtual void apply(ScheduleDAGMI *DAG);
+
+protected:
+  void constrainLocalCopy(SUnit *CopySU, ScheduleDAGMI *DAG);
+};
+} // anonymous
+
+/// constrainLocalCopy handles two possibilities:
+/// 1) Local src:
+/// I0:     = dst
+/// I1: src = ...
+/// I2:     = dst
+/// I3: dst = src (copy)
+/// (create pred->succ edges I0->I1, I2->I1)
+///
+/// 2) Local copy:
+/// I0: dst = src (copy)
+/// I1:     = dst
+/// I2: src = ...
+/// I3:     = dst
+/// (create pred->succ edges I1->I2, I3->I2)
+///
+/// Although the MachineScheduler is currently constrained to single blocks,
+/// this algorithm should handle extended blocks. An EBB is a set of
+/// contiguously numbered blocks such that the previous block in the EBB is
+/// always the single predecessor.
+void CopyConstrain::constrainLocalCopy(SUnit *CopySU, ScheduleDAGMI *DAG) {
+  LiveIntervals *LIS = DAG->getLIS();
+  MachineInstr *Copy = CopySU->getInstr();
+
+  // Check for pure vreg copies.
+  unsigned SrcReg = Copy->getOperand(1).getReg();
+  if (!TargetRegisterInfo::isVirtualRegister(SrcReg))
+    return;
+
+  unsigned DstReg = Copy->getOperand(0).getReg();
+  if (!TargetRegisterInfo::isVirtualRegister(DstReg))
+    return;
+
+  // Check if either the dest or source is local. If it's live across a back
+  // edge, it's not local. Note that if both vregs are live across the back
+  // edge, we cannot successfully contrain the copy without cyclic scheduling.
+  unsigned LocalReg = DstReg;
+  unsigned GlobalReg = SrcReg;
+  LiveInterval *LocalLI = &LIS->getInterval(LocalReg);
+  if (!LocalLI->isLocal(RegionBeginIdx, RegionEndIdx)) {
+    LocalReg = SrcReg;
+    GlobalReg = DstReg;
+    LocalLI = &LIS->getInterval(LocalReg);
+    if (!LocalLI->isLocal(RegionBeginIdx, RegionEndIdx))
+      return;
+  }
+  LiveInterval *GlobalLI = &LIS->getInterval(GlobalReg);
+
+  // Find the global segment after the start of the local LI.
+  LiveInterval::iterator GlobalSegment = GlobalLI->find(LocalLI->beginIndex());
+  // If GlobalLI does not overlap LocalLI->start, then a copy directly feeds a
+  // local live range. We could create edges from other global uses to the local
+  // start, but the coalescer should have already eliminated these cases, so
+  // don't bother dealing with it.
+  if (GlobalSegment == GlobalLI->end())
+    return;
+
+  // If GlobalSegment is killed at the LocalLI->start, the call to find()
+  // returned the next global segment. But if GlobalSegment overlaps with
+  // LocalLI->start, then advance to the next segement. If a hole in GlobalLI
+  // exists in LocalLI's vicinity, GlobalSegment will be the end of the hole.
+  if (GlobalSegment->contains(LocalLI->beginIndex()))
+    ++GlobalSegment;
+
+  if (GlobalSegment == GlobalLI->end())
+    return;
+
+  // Check if GlobalLI contains a hole in the vicinity of LocalLI.
+  if (GlobalSegment != GlobalLI->begin()) {
+    // Two address defs have no hole.
+    if (SlotIndex::isSameInstr(llvm::prior(GlobalSegment)->end,
+                               GlobalSegment->start)) {
+      return;
+    }
+    // If GlobalLI has a prior segment, it must be live into the EBB. Otherwise
+    // it would be a disconnected component in the live range.
+    assert(llvm::prior(GlobalSegment)->start < LocalLI->beginIndex() &&
+           "Disconnected LRG within the scheduling region.");
+  }
+  MachineInstr *GlobalDef = LIS->getInstructionFromIndex(GlobalSegment->start);
+  if (!GlobalDef)
+    return;
+
+  SUnit *GlobalSU = DAG->getSUnit(GlobalDef);
+  if (!GlobalSU)
+    return;
+
+  // GlobalDef is the bottom of the GlobalLI hole. Open the hole by
+  // constraining the uses of the last local def to precede GlobalDef.
+  SmallVector<SUnit*,8> LocalUses;
+  const VNInfo *LastLocalVN = LocalLI->getVNInfoBefore(LocalLI->endIndex());
+  MachineInstr *LastLocalDef = LIS->getInstructionFromIndex(LastLocalVN->def);
+  SUnit *LastLocalSU = DAG->getSUnit(LastLocalDef);
+  for (SUnit::const_succ_iterator
+         I = LastLocalSU->Succs.begin(), E = LastLocalSU->Succs.end();
+       I != E; ++I) {
+    if (I->getKind() != SDep::Data || I->getReg() != LocalReg)
+      continue;
+    if (I->getSUnit() == GlobalSU)
+      continue;
+    if (!DAG->canAddEdge(GlobalSU, I->getSUnit()))
+      return;
+    LocalUses.push_back(I->getSUnit());
+  }
+  // Open the top of the GlobalLI hole by constraining any earlier global uses
+  // to precede the start of LocalLI.
+  SmallVector<SUnit*,8> GlobalUses;
+  MachineInstr *FirstLocalDef =
+    LIS->getInstructionFromIndex(LocalLI->beginIndex());
+  SUnit *FirstLocalSU = DAG->getSUnit(FirstLocalDef);
+  for (SUnit::const_pred_iterator
+         I = GlobalSU->Preds.begin(), E = GlobalSU->Preds.end(); I != E; ++I) {
+    if (I->getKind() != SDep::Anti || I->getReg() != GlobalReg)
+      continue;
+    if (I->getSUnit() == FirstLocalSU)
+      continue;
+    if (!DAG->canAddEdge(FirstLocalSU, I->getSUnit()))
+      return;
+    GlobalUses.push_back(I->getSUnit());
+  }
+  DEBUG(dbgs() << "Constraining copy SU(" << CopySU->NodeNum << ")\n");
+  // Add the weak edges.
+  for (SmallVectorImpl<SUnit*>::const_iterator
+         I = LocalUses.begin(), E = LocalUses.end(); I != E; ++I) {
+    DEBUG(dbgs() << "  Local use SU(" << (*I)->NodeNum << ") -> SU("
+          << GlobalSU->NodeNum << ")\n");
+    DAG->addEdge(GlobalSU, SDep(*I, SDep::Weak));
+  }
+  for (SmallVectorImpl<SUnit*>::const_iterator
+         I = GlobalUses.begin(), E = GlobalUses.end(); I != E; ++I) {
+    DEBUG(dbgs() << "  Global use SU(" << (*I)->NodeNum << ") -> SU("
+          << FirstLocalSU->NodeNum << ")\n");
+    DAG->addEdge(FirstLocalSU, SDep(*I, SDep::Weak));
+  }
+}
+
+/// \brief Callback from DAG postProcessing to create weak edges to encourage
+/// copy elimination.
+void CopyConstrain::apply(ScheduleDAGMI *DAG) {
+  MachineBasicBlock::iterator FirstPos = nextIfDebug(DAG->begin(), DAG->end());
+  if (FirstPos == DAG->end())
+    return;
+  RegionBeginIdx = DAG->getLIS()->getInstructionIndex(&*FirstPos);
+  RegionEndIdx = DAG->getLIS()->getInstructionIndex(
+    &*priorNonDebug(DAG->end(), DAG->begin()));
+
+  for (unsigned Idx = 0, End = DAG->SUnits.size(); Idx != End; ++Idx) {
+    SUnit *SU = &DAG->SUnits[Idx];
+    if (!SU->getInstr()->isCopy())
+      continue;
+
+    constrainLocalCopy(SU, DAG);
+  }
+}
+
+//===----------------------------------------------------------------------===//
 // ConvergingScheduler - Implementation of the standard MachineSchedStrategy.
 //===----------------------------------------------------------------------===//
 
@@ -916,7 +1112,7 @@ public:
   /// Represent the type of SchedCandidate found within a single queue.
   /// pickNodeBidirectional depends on these listed by decreasing priority.
   enum CandReason {
-    NoCand, SingleExcess, SingleCritical, Cluster,
+    NoCand, PhysRegCopy, SingleExcess, SingleCritical, Cluster, Weak,
     ResourceReduce, ResourceDemand, BotHeightReduce, BotPathReduce,
     TopDepthReduce, TopPathReduce, SingleMax, MultiPressure, NextDefUse,
     NodeOrder};
@@ -1191,6 +1387,8 @@ protected:
                          const RegPressureTracker &RPTracker,
                          SchedCandidate &Candidate);
 
+  void reschedulePhysRegCopies(SUnit *SU, bool isTop);
+
 #ifndef NDEBUG
   void traceCandidate(const SchedCandidate &Cand);
 #endif
@@ -1339,6 +1537,8 @@ void ConvergingScheduler::SchedBoundary::setLatencyPolicy(CandPolicy &Policy) {
   for (ReadyQueue::iterator I = Available.begin(), E = Available.end();
        I != E; ++I) {
     unsigned L = getUnscheduledLatency(*I);
+    DEBUG(dbgs() << "  " << Available.getName()
+          << " RemLatency SU(" << (*I)->NodeNum << ") " << L << '\n');
     if (L > RemLatency)
       RemLatency = L;
   }
@@ -1349,10 +1549,13 @@ void ConvergingScheduler::SchedBoundary::setLatencyPolicy(CandPolicy &Policy) {
       RemLatency = L;
   }
   unsigned CriticalPathLimit = Rem->CriticalPath + SchedModel->getILPWindow();
+  DEBUG(dbgs() << "  " << Available.getName()
+        << " ExpectedLatency " << ExpectedLatency
+        << " CP Limit " << CriticalPathLimit << '\n');
   if (RemLatency + ExpectedLatency >= CriticalPathLimit
       && RemLatency > Rem->getMaxRemainingCount(SchedModel)) {
     Policy.ReduceLatency = true;
-    DEBUG(dbgs() << "Increase ILP: " << Available.getName() << '\n');
+    DEBUG(dbgs() << "  Increase ILP: " << Available.getName() << '\n');
   }
 }
 
@@ -1569,7 +1772,8 @@ void ConvergingScheduler::balanceZones(
   if ((int)(Rem->getMaxRemainingCount(SchedModel) - RemainingCritCount)
       > (int)SchedModel->getLatencyFactor()) {
     CriticalCand.Policy.ReduceResIdx = CriticalZone.CritResIdx;
-    DEBUG(dbgs() << "Balance " << CriticalZone.Available.getName() << " reduce "
+    DEBUG(dbgs() << "  Balance " << CriticalZone.Available.getName()
+          << " reduce "
           << SchedModel->getProcResource(CriticalZone.CritResIdx)->Name
           << '\n');
   }
@@ -1580,7 +1784,8 @@ void ConvergingScheduler::balanceZones(
   if ((int)(OppositeZone.ExpectedCount - OppositeCount)
       > (int)SchedModel->getLatencyFactor()) {
     OppositeCand.Policy.DemandResIdx = CriticalZone.CritResIdx;
-    DEBUG(dbgs() << "Balance " << OppositeZone.Available.getName() << " demand "
+    DEBUG(dbgs() << "  Balance " << OppositeZone.Available.getName()
+          << " demand "
           << SchedModel->getProcResource(OppositeZone.CritResIdx)->Name
           << '\n');
   }
@@ -1604,7 +1809,7 @@ void ConvergingScheduler::checkResourceLimits(
     if (Top.CritResIdx != Rem.CritResIdx) {
       TopCand.Policy.ReduceResIdx = Top.CritResIdx;
       BotCand.Policy.ReduceResIdx = Bot.CritResIdx;
-      DEBUG(dbgs() << "Reduce scheduled "
+      DEBUG(dbgs() << "  Reduce scheduled "
             << SchedModel->getProcResource(Top.CritResIdx)->Name << '\n');
     }
     return;
@@ -1621,7 +1826,7 @@ void ConvergingScheduler::checkResourceLimits(
         && (Rem.CriticalPath > Top.CurrCycle + Bot.CurrCycle)) {
       TopCand.Policy.ReduceLatency = true;
       BotCand.Policy.ReduceLatency = true;
-      DEBUG(dbgs() << "Reduce scheduled latency " << Top.ExpectedLatency
+      DEBUG(dbgs() << "  Reduce scheduled latency " << Top.ExpectedLatency
             << " + " << Bot.ExpectedLatency << '\n');
     }
     return;
@@ -1696,6 +1901,34 @@ static unsigned getWeakLeft(const SUnit *SU, bool isTop) {
   return (isTop) ? SU->WeakPredsLeft : SU->WeakSuccsLeft;
 }
 
+/// Minimize physical register live ranges. Regalloc wants them adjacent to
+/// their physreg def/use.
+///
+/// FIXME: This is an unnecessary check on the critical path. Most are root/leaf
+/// copies which can be prescheduled. The rest (e.g. x86 MUL) could be bundled
+/// with the operation that produces or consumes the physreg. We'll do this when
+/// regalloc has support for parallel copies.
+static int biasPhysRegCopy(const SUnit *SU, bool isTop) {
+  const MachineInstr *MI = SU->getInstr();
+  if (!MI->isCopy())
+    return 0;
+
+  unsigned ScheduledOper = isTop ? 1 : 0;
+  unsigned UnscheduledOper = isTop ? 0 : 1;
+  // If we have already scheduled the physreg produce/consumer, immediately
+  // schedule the copy.
+  if (TargetRegisterInfo::isPhysicalRegister(
+        MI->getOperand(ScheduledOper).getReg()))
+    return 1;
+  // If the physreg is at the boundary, defer it. Otherwise schedule it
+  // immediately to free the dependent. We can hoist the copy later.
+  bool AtBoundary = isTop ? !SU->NumSuccsLeft : !SU->NumPredsLeft;
+  if (TargetRegisterInfo::isPhysicalRegister(
+        MI->getOperand(UnscheduledOper).getReg()))
+    return AtBoundary ? -1 : 1;
+  return 0;
+}
+
 /// Apply a set of heursitics to a new candidate. Heuristics are currently
 /// hierarchical. This may be more efficient than a graduated cost model because
 /// we don't need to evaluate all aspects of the model for each node in the
@@ -1723,6 +1956,12 @@ void ConvergingScheduler::tryCandidate(SchedCandidate &Cand,
     TryCand.Reason = NodeOrder;
     return;
   }
+
+  if (tryGreater(biasPhysRegCopy(TryCand.SU, Zone.isTop()),
+                 biasPhysRegCopy(Cand.SU, Zone.isTop()),
+                 TryCand, Cand, PhysRegCopy))
+    return;
+
   // Avoid exceeding the target's limit.
   if (tryLess(TryCand.RPDelta.Excess.UnitIncrease,
               Cand.RPDelta.Excess.UnitIncrease, TryCand, Cand, SingleExcess))
@@ -1749,12 +1988,16 @@ void ConvergingScheduler::tryCandidate(SchedCandidate &Cand,
   if (tryGreater(TryCand.SU == NextClusterSU, Cand.SU == NextClusterSU,
                  TryCand, Cand, Cluster))
     return;
-  // Currently, weak edges are for clustering, so we hard-code that reason.
-  // However, deferring the current TryCand will not change Cand's reason.
+
+  // Weak edges are for clustering and other constraints.
+  //
+  // Deferring TryCand here does not change Cand's reason. This is good in the
+  // sense that a bad candidate shouldn't affect a previous candidate's
+  // goodness, but bad in that it is assymetric and depends on queue order.
   CandReason OrigReason = Cand.Reason;
   if (tryLess(getWeakLeft(TryCand.SU, Zone.isTop()),
               getWeakLeft(Cand.SU, Zone.isTop()),
-              TryCand, Cand, Cluster)) {
+              TryCand, Cand, Weak)) {
     Cand.Reason = OrigReason;
     return;
   }
@@ -1825,20 +2068,20 @@ static bool compareRPDelta(const RegPressureDelta &LHS,
 
   // Avoid increasing the max critical pressure in the scheduled region.
   if (LHS.Excess.UnitIncrease != RHS.Excess.UnitIncrease) {
-    DEBUG(dbgs() << "RP excess top - bot: "
+    DEBUG(dbgs() << "  RP excess top - bot: "
           << (LHS.Excess.UnitIncrease - RHS.Excess.UnitIncrease) << '\n');
     return LHS.Excess.UnitIncrease < RHS.Excess.UnitIncrease;
   }
   // Avoid increasing the max critical pressure in the scheduled region.
   if (LHS.CriticalMax.UnitIncrease != RHS.CriticalMax.UnitIncrease) {
-    DEBUG(dbgs() << "RP critical top - bot: "
+    DEBUG(dbgs() << "  RP critical top - bot: "
           << (LHS.CriticalMax.UnitIncrease - RHS.CriticalMax.UnitIncrease)
           << '\n');
     return LHS.CriticalMax.UnitIncrease < RHS.CriticalMax.UnitIncrease;
   }
   // Avoid increasing the max pressure of the entire region.
   if (LHS.CurrentMax.UnitIncrease != RHS.CurrentMax.UnitIncrease) {
-    DEBUG(dbgs() << "RP current top - bot: "
+    DEBUG(dbgs() << "  RP current top - bot: "
           << (LHS.CurrentMax.UnitIncrease - RHS.CurrentMax.UnitIncrease)
           << '\n');
     return LHS.CurrentMax.UnitIncrease < RHS.CurrentMax.UnitIncrease;
@@ -1851,9 +2094,11 @@ const char *ConvergingScheduler::getReasonStr(
   ConvergingScheduler::CandReason Reason) {
   switch (Reason) {
   case NoCand:         return "NOCAND    ";
+  case PhysRegCopy:    return "PREG-COPY";
   case SingleExcess:   return "REG-EXCESS";
   case SingleCritical: return "REG-CRIT  ";
   case Cluster:        return "CLUSTER   ";
+  case Weak:           return "WEAK      ";
   case SingleMax:      return "REG-MAX   ";
   case MultiPressure:  return "REG-MULTI ";
   case ResourceReduce: return "RES-REDUCE";
@@ -1953,8 +2198,7 @@ void ConvergingScheduler::pickNodeFromQueue(SchedBoundary &Zone,
 
 static void tracePick(const ConvergingScheduler::SchedCandidate &Cand,
                       bool IsTop) {
-  DEBUG(dbgs() << "Pick " << (IsTop ? "Top" : "Bot")
-        << " SU(" << Cand.SU->NodeNum << ") "
+  DEBUG(dbgs() << "Pick " << (IsTop ? "Top " : "Bot ")
         << ConvergingScheduler::getReasonStr(Cand.Reason) << '\n');
 }
 
@@ -1964,10 +2208,12 @@ SUnit *ConvergingScheduler::pickNodeBidirectional(bool &IsTopNode) {
   // efficient, but also provides the best heuristics for CriticalPSets.
   if (SUnit *SU = Bot.pickOnlyChoice()) {
     IsTopNode = false;
+    DEBUG(dbgs() << "Pick Top NOCAND\n");
     return SU;
   }
   if (SUnit *SU = Top.pickOnlyChoice()) {
     IsTopNode = true;
+    DEBUG(dbgs() << "Pick Bot NOCAND\n");
     return SU;
   }
   CandPolicy NoPolicy;
@@ -2065,21 +2311,53 @@ SUnit *ConvergingScheduler::pickNode(bool &IsTopNode) {
   if (SU->isBottomReady())
     Bot.removeReady(SU);
 
-  DEBUG(dbgs() << "Scheduling " << *SU->getInstr());
+  DEBUG(dbgs() << "Scheduling SU(" << SU->NodeNum << ") " << *SU->getInstr());
   return SU;
 }
 
+void ConvergingScheduler::reschedulePhysRegCopies(SUnit *SU, bool isTop) {
+
+  MachineBasicBlock::iterator InsertPos = SU->getInstr();
+  if (!isTop)
+    ++InsertPos;
+  SmallVectorImpl<SDep> &Deps = isTop ? SU->Preds : SU->Succs;
+
+  // Find already scheduled copies with a single physreg dependence and move
+  // them just above the scheduled instruction.
+  for (SmallVectorImpl<SDep>::iterator I = Deps.begin(), E = Deps.end();
+       I != E; ++I) {
+    if (I->getKind() != SDep::Data || !TRI->isPhysicalRegister(I->getReg()))
+      continue;
+    SUnit *DepSU = I->getSUnit();
+    if (isTop ? DepSU->Succs.size() > 1 : DepSU->Preds.size() > 1)
+      continue;
+    MachineInstr *Copy = DepSU->getInstr();
+    if (!Copy->isCopy())
+      continue;
+    DEBUG(dbgs() << "  Rescheduling physreg copy ";
+          I->getSUnit()->dump(DAG));
+    DAG->moveInstruction(Copy, InsertPos);
+  }
+}
+
 /// Update the scheduler's state after scheduling a node. This is the same node
 /// that was just returned by pickNode(). However, ScheduleDAGMI needs to update
 /// it's state based on the current cycle before MachineSchedStrategy does.
+///
+/// FIXME: Eventually, we may bundle physreg copies rather than rescheduling
+/// them here. See comments in biasPhysRegCopy.
 void ConvergingScheduler::schedNode(SUnit *SU, bool IsTopNode) {
   if (IsTopNode) {
     SU->TopReadyCycle = Top.CurrCycle;
     Top.bumpNode(SU);
+    if (SU->hasPhysRegUses)
+      reschedulePhysRegCopies(SU, true);
   }
   else {
     SU->BotReadyCycle = Bot.CurrCycle;
     Bot.bumpNode(SU);
+    if (SU->hasPhysRegDefs)
+      reschedulePhysRegCopies(SU, false);
   }
 }
 
@@ -2090,6 +2368,12 @@ static ScheduleDAGInstrs *createConvergingSched(MachineSchedContext *C) {
          "-misched-topdown incompatible with -misched-bottomup");
   ScheduleDAGMI *DAG = new ScheduleDAGMI(C, new ConvergingScheduler());
   // Register DAG post-processors.
+  //
+  // FIXME: extend the mutation API to allow earlier mutations to instantiate
+  // data and pass it to later mutations. Have a single mutation that gathers
+  // the interesting nodes in one pass.
+  if (EnableCopyConstrain)
+    DAG->addMutation(new CopyConstrain(DAG->TII, DAG->TRI));
   if (EnableLoadCluster)
     DAG->addMutation(new LoadClusterMutation(DAG->TII, DAG->TRI));
   if (EnableMacroFusion)
@@ -2179,12 +2463,12 @@ public:
     SUnit *SU = ReadyQ.back();
     ReadyQ.pop_back();
     IsTopNode = false;
-    DEBUG(dbgs() << "*** Scheduling " << "SU(" << SU->NodeNum << "): "
-          << *SU->getInstr()
+    DEBUG(dbgs() << "Pick node " << "SU(" << SU->NodeNum << ") "
           << " ILP: " << DAG->getDFSResult()->getILP(SU)
           << " Tree: " << DAG->getDFSResult()->getSubtreeID(SU) << " @"
           << DAG->getDFSResult()->getSubtreeLevel(
-            DAG->getDFSResult()->getSubtreeID(SU)) << '\n');
+            DAG->getDFSResult()->getSubtreeID(SU)) << '\n'
+          << "Scheduling " << *SU->getInstr());
     return SU;
   }
 
diff --git a/lib/CodeGen/MachineTraceMetrics.cpp b/lib/CodeGen/MachineTraceMetrics.cpp
index 49d8c4e9470d..00f702c846c5 100644
--- a/lib/CodeGen/MachineTraceMetrics.cpp
+++ b/lib/CodeGen/MachineTraceMetrics.cpp
@@ -1200,8 +1200,10 @@ unsigned MachineTraceMetrics::Trace::getResourceDepth(bool Bottom) const {
   return std::max(Instrs, PRMax);
 }
 
+
 unsigned MachineTraceMetrics::Trace::
-getResourceLength(ArrayRef<const MachineBasicBlock*> Extrablocks) const {
+getResourceLength(ArrayRef<const MachineBasicBlock*> Extrablocks,
+                  ArrayRef<const MCSchedClassDesc*> ExtraInstrs) const {
   // Add up resources above and below the center block.
   ArrayRef<unsigned> PRDepths = TE.getProcResourceDepths(getBlockNum());
   ArrayRef<unsigned> PRHeights = TE.getProcResourceHeights(getBlockNum());
@@ -1210,6 +1212,18 @@ getResourceLength(ArrayRef<const MachineBasicBlock*> Extrablocks) const {
     unsigned PRCycles = PRDepths[K] + PRHeights[K];
     for (unsigned I = 0; I != Extrablocks.size(); ++I)
       PRCycles += TE.MTM.getProcResourceCycles(Extrablocks[I]->getNumber())[K];
+    for (unsigned I = 0; I != ExtraInstrs.size(); ++I) {
+      const MCSchedClassDesc* SC = ExtraInstrs[I];
+      if (!SC->isValid())
+        continue;
+      for (TargetSchedModel::ProcResIter
+             PI = TE.MTM.SchedModel.getWriteProcResBegin(SC),
+             PE = TE.MTM.SchedModel.getWriteProcResEnd(SC); PI != PE; ++PI) {
+        if (PI->ProcResourceIdx != K)
+          continue;
+        PRCycles += (PI->Cycles * TE.MTM.SchedModel.getResourceFactor(K));
+      }
+    }
     PRMax = std::max(PRMax, PRCycles);
   }
   // Convert to cycle count.
diff --git a/lib/CodeGen/MachineVerifier.cpp b/lib/CodeGen/MachineVerifier.cpp
index 4b1230029a74..037043f6419c 100644
--- a/lib/CodeGen/MachineVerifier.cpp
+++ b/lib/CodeGen/MachineVerifier.cpp
@@ -472,6 +472,9 @@ void MachineVerifier::visitMachineFunctionBefore() {
     if (MInfo.Succs.size() != I->succ_size())
       report("MBB has duplicate entries in its successor list.", I);
   }
+
+  // Check that the register use lists are sane.
+  MRI->verifyUseLists();
 }
 
 // Does iterator point to a and b as the first two elements?
diff --git a/lib/CodeGen/Passes.cpp b/lib/CodeGen/Passes.cpp
index 1af65c88abeb..bfbc0623f9cc 100644
--- a/lib/CodeGen/Passes.cpp
+++ b/lib/CodeGen/Passes.cpp
@@ -93,9 +93,10 @@ static cl::opt<bool> EarlyLiveIntervals("early-live-intervals", cl::Hidden,
 /// simple binary flags that either suppress the pass or do nothing.
 /// i.e. -disable-mypass=false has no effect.
 /// These should be converted to boolOrDefault in order to use applyOverride.
-static AnalysisID applyDisable(AnalysisID PassID, bool Override) {
+static IdentifyingPassPtr applyDisable(IdentifyingPassPtr PassID,
+                                       bool Override) {
   if (Override)
-    return 0;
+    return IdentifyingPassPtr();
   return PassID;
 }
 
@@ -103,19 +104,20 @@ static AnalysisID applyDisable(AnalysisID PassID, bool Override) {
 /// flags with ternary conditions. TargetID is passed through by default. The
 /// pass is suppressed when the option is false. When the option is true, the
 /// StandardID is selected if the target provides no default.
-static AnalysisID applyOverride(AnalysisID TargetID, cl::boolOrDefault Override,
-                                AnalysisID StandardID) {
+static IdentifyingPassPtr applyOverride(IdentifyingPassPtr TargetID,
+                                        cl::boolOrDefault Override,
+                                        AnalysisID StandardID) {
   switch (Override) {
   case cl::BOU_UNSET:
     return TargetID;
   case cl::BOU_TRUE:
-    if (TargetID)
+    if (TargetID.isValid())
       return TargetID;
     if (StandardID == 0)
       report_fatal_error("Target cannot enable pass");
     return StandardID;
   case cl::BOU_FALSE:
-    return 0;
+    return IdentifyingPassPtr();
   }
   llvm_unreachable("Invalid command line option state");
 }
@@ -132,7 +134,8 @@ static AnalysisID applyOverride(AnalysisID TargetID, cl::boolOrDefault Override,
 /// StandardID may be a pseudo ID. In that case TargetID is the name of the real
 /// pass to run. This allows multiple options to control a single pass depending
 /// on where in the pipeline that pass is added.
-static AnalysisID overridePass(AnalysisID StandardID, AnalysisID TargetID) {
+static IdentifyingPassPtr overridePass(AnalysisID StandardID,
+                                       IdentifyingPassPtr TargetID) {
   if (StandardID == &PostRASchedulerID)
     return applyDisable(TargetID, DisablePostRA);
 
@@ -200,11 +203,11 @@ public:
   // user interface. For example, a target may disable a standard pass by
   // default by substituting a pass ID of zero, and the user may still enable
   // that standard pass with an explicit command line option.
-  DenseMap<AnalysisID,AnalysisID> TargetPasses;
+  DenseMap<AnalysisID,IdentifyingPassPtr> TargetPasses;
 
   /// Store the pairs of <AnalysisID, AnalysisID> of which the second pass
   /// is inserted after each instance of the first one.
-  SmallVector<std::pair<AnalysisID, AnalysisID>, 4> InsertedPasses;
+  SmallVector<std::pair<AnalysisID, IdentifyingPassPtr>, 4> InsertedPasses;
 };
 } // namespace llvm
 
@@ -239,9 +242,13 @@ TargetPassConfig::TargetPassConfig(TargetMachine *tm, PassManagerBase &pm)
 
 /// Insert InsertedPassID pass after TargetPassID.
 void TargetPassConfig::insertPass(AnalysisID TargetPassID,
-                                  AnalysisID InsertedPassID) {
-  assert(TargetPassID != InsertedPassID && "Insert a pass after itself!");
-  std::pair<AnalysisID, AnalysisID> P(TargetPassID, InsertedPassID);
+                                  IdentifyingPassPtr InsertedPassID) {
+  assert(((!InsertedPassID.isInstance() &&
+           TargetPassID != InsertedPassID.getID()) ||
+          (InsertedPassID.isInstance() &&
+           TargetPassID != InsertedPassID.getInstance()->getPassID())) &&
+         "Insert a pass after itself!");
+  std::pair<AnalysisID, IdentifyingPassPtr> P(TargetPassID, InsertedPassID);
   Impl->InsertedPasses.push_back(P);
 }
 
@@ -265,12 +272,12 @@ void TargetPassConfig::setOpt(bool &Opt, bool Val) {
 }
 
 void TargetPassConfig::substitutePass(AnalysisID StandardID,
-                                      AnalysisID TargetID) {
+                                      IdentifyingPassPtr TargetID) {
   Impl->TargetPasses[StandardID] = TargetID;
 }
 
-AnalysisID TargetPassConfig::getPassSubstitution(AnalysisID ID) const {
-  DenseMap<AnalysisID, AnalysisID>::const_iterator
+IdentifyingPassPtr TargetPassConfig::getPassSubstitution(AnalysisID ID) const {
+  DenseMap<AnalysisID, IdentifyingPassPtr>::const_iterator
     I = Impl->TargetPasses.find(ID);
   if (I == Impl->TargetPasses.end())
     return ID;
@@ -303,24 +310,39 @@ void TargetPassConfig::addPass(Pass *P) {
 
 /// Add a CodeGen pass at this point in the pipeline after checking for target
 /// and command line overrides.
+///
+/// addPass cannot return a pointer to the pass instance because is internal the
+/// PassManager and the instance we create here may already be freed.
 AnalysisID TargetPassConfig::addPass(AnalysisID PassID) {
-  AnalysisID TargetID = getPassSubstitution(PassID);
-  AnalysisID FinalID = overridePass(PassID, TargetID);
-  if (FinalID == 0)
-    return FinalID;
-
-  Pass *P = Pass::createPass(FinalID);
-  if (!P)
-    llvm_unreachable("Pass ID not registered");
-  addPass(P);
+  IdentifyingPassPtr TargetID = getPassSubstitution(PassID);
+  IdentifyingPassPtr FinalPtr = overridePass(PassID, TargetID);
+  if (!FinalPtr.isValid())
+    return 0;
+
+  Pass *P;
+  if (FinalPtr.isInstance())
+    P = FinalPtr.getInstance();
+  else {
+    P = Pass::createPass(FinalPtr.getID());
+    if (!P)
+      llvm_unreachable("Pass ID not registered");
+  }
+  AnalysisID FinalID = P->getPassID();
+  addPass(P); // Ends the lifetime of P.
+
   // Add the passes after the pass P if there is any.
-  for (SmallVector<std::pair<AnalysisID, AnalysisID>, 4>::iterator
+  for (SmallVector<std::pair<AnalysisID, IdentifyingPassPtr>, 4>::iterator
          I = Impl->InsertedPasses.begin(), E = Impl->InsertedPasses.end();
        I != E; ++I) {
     if ((*I).first == PassID) {
-      assert((*I).second && "Illegal Pass ID!");
-      Pass *NP = Pass::createPass((*I).second);
-      assert(NP && "Pass ID not registered");
+      assert((*I).second.isValid() && "Illegal Pass ID!");
+      Pass *NP;
+      if ((*I).second.isInstance())
+        NP = (*I).second.getInstance();
+      else {
+        NP = Pass::createPass((*I).second.getID());
+        assert(NP && "Pass ID not registered");
+      }
       addPass(NP);
     }
   }
@@ -687,14 +709,6 @@ void TargetPassConfig::addOptimizedRegAlloc(FunctionPass *RegAllocPass) {
   addPass(&VirtRegRewriterID);
   printAndVerify("After Virtual Register Rewriter");
 
-  // FinalizeRegAlloc is convenient until MachineInstrBundles is more mature,
-  // but eventually, all users of it should probably be moved to addPostRA and
-  // it can go away.  Currently, it's the intended place for targets to run
-  // FinalizeMachineBundles, because passes other than MachineScheduling an
-  // RegAlloc itself may not be aware of bundles.
-  if (addFinalizeRegAlloc())
-    printAndVerify("After RegAlloc finalization");
-
   // Perform stack slot coloring and post-ra machine LICM.
   //
   // FIXME: Re-enable coloring with register when it's capable of adding
diff --git a/lib/CodeGen/PrologEpilogInserter.cpp b/lib/CodeGen/PrologEpilogInserter.cpp
index e5872df731a0..959dd7df58ee 100644
--- a/lib/CodeGen/PrologEpilogInserter.cpp
+++ b/lib/CodeGen/PrologEpilogInserter.cpp
@@ -824,6 +824,12 @@ void PEI::scavengeFrameVirtualRegs(MachineFunction &Fn) {
     // The instruction stream may change in the loop, so check BB->end()
     // directly.
     for (MachineBasicBlock::iterator I = BB->begin(); I != BB->end(); ) {
+      // We might end up here again with a NULL iterator if we scavenged a
+      // register for which we inserted spill code for definition by what was
+      // originally the first instruction in BB.
+      if (I == MachineBasicBlock::iterator(NULL))
+        I = BB->begin();
+
       MachineInstr *MI = I;
       MachineBasicBlock::iterator J = llvm::next(I);
       MachineBasicBlock::iterator P = I == BB->begin() ?
@@ -883,8 +889,6 @@ void PEI::scavengeFrameVirtualRegs(MachineFunction &Fn) {
           "The register scavenger has an unexpected position");
         I = P;
         RS->unprocess(P);
-
-        // RS->skipTo(I == BB->begin() ? NULL : llvm::prior(I));
       } else
         ++I;
     }
diff --git a/lib/CodeGen/RegAllocBasic.cpp b/lib/CodeGen/RegAllocBasic.cpp
index 0b6dc68cdf09..7fcfe9e88bef 100644
--- a/lib/CodeGen/RegAllocBasic.cpp
+++ b/lib/CodeGen/RegAllocBasic.cpp
@@ -63,7 +63,7 @@ class RABasic : public MachineFunctionPass, public RegAllocBase
   MachineFunction *MF;
 
   // state
-  std::auto_ptr<Spiller> SpillerInstance;
+  OwningPtr<Spiller> SpillerInstance;
   std::priority_queue<LiveInterval*, std::vector<LiveInterval*>,
                       CompSpillWeight> Queue;
 
diff --git a/lib/CodeGen/RegAllocGreedy.cpp b/lib/CodeGen/RegAllocGreedy.cpp
index 6d84176af261..9eed1fc62acc 100644
--- a/lib/CodeGen/RegAllocGreedy.cpp
+++ b/lib/CodeGen/RegAllocGreedy.cpp
@@ -78,7 +78,7 @@ class RAGreedy : public MachineFunctionPass,
   LiveDebugVariables *DebugVars;
 
   // state
-  std::auto_ptr<Spiller> SpillerInstance;
+  OwningPtr<Spiller> SpillerInstance;
   std::priority_queue<std::pair<unsigned, unsigned> > Queue;
   unsigned NextCascade;
 
@@ -166,8 +166,8 @@ class RAGreedy : public MachineFunctionPass,
   };
 
   // splitting state.
-  std::auto_ptr<SplitAnalysis> SA;
-  std::auto_ptr<SplitEditor> SE;
+  OwningPtr<SplitAnalysis> SA;
+  OwningPtr<SplitEditor> SE;
 
   /// Cached per-block interference maps
   InterferenceCache IntfCache;
diff --git a/lib/CodeGen/RegAllocPBQP.cpp b/lib/CodeGen/RegAllocPBQP.cpp
index 607edac24bd2..15a88e224faa 100644
--- a/lib/CodeGen/RegAllocPBQP.cpp
+++ b/lib/CodeGen/RegAllocPBQP.cpp
@@ -34,6 +34,7 @@
 #include "llvm/CodeGen/RegAllocPBQP.h"
 #include "RegisterCoalescer.h"
 #include "Spiller.h"
+#include "llvm/ADT/OwningPtr.h"
 #include "llvm/Analysis/AliasAnalysis.h"
 #include "llvm/CodeGen/CalcSpillWeights.h"
 #include "llvm/CodeGen/LiveIntervalAnalysis.h"
@@ -89,8 +90,8 @@ public:
   static char ID;
 
   /// Construct a PBQP register allocator.
-  RegAllocPBQP(std::auto_ptr<PBQPBuilder> b, char *cPassID=0)
-      : MachineFunctionPass(ID), builder(b), customPassID(cPassID) {
+  RegAllocPBQP(OwningPtr<PBQPBuilder> &b, char *cPassID=0)
+      : MachineFunctionPass(ID), builder(b.take()), customPassID(cPassID) {
     initializeSlotIndexesPass(*PassRegistry::getPassRegistry());
     initializeLiveIntervalsPass(*PassRegistry::getPassRegistry());
     initializeCalculateSpillWeightsPass(*PassRegistry::getPassRegistry());
@@ -121,7 +122,7 @@ private:
   typedef std::set<unsigned> RegSet;
 
 
-  std::auto_ptr<PBQPBuilder> builder;
+  OwningPtr<PBQPBuilder> builder;
 
   char *customPassID;
 
@@ -132,7 +133,7 @@ private:
   const MachineLoopInfo *loopInfo;
   MachineRegisterInfo *mri;
 
-  std::auto_ptr<Spiller> spiller;
+  OwningPtr<Spiller> spiller;
   LiveIntervals *lis;
   LiveStacks *lss;
   VirtRegMap *vrm;
@@ -186,16 +187,15 @@ unsigned PBQPRAProblem::getPRegForOption(unsigned vreg, unsigned option) const {
   return allowedSet[option - 1];
 }
 
-std::auto_ptr<PBQPRAProblem> PBQPBuilder::build(MachineFunction *mf,
-                                                const LiveIntervals *lis,
-                                                const MachineLoopInfo *loopInfo,
-                                                const RegSet &vregs) {
+PBQPRAProblem *PBQPBuilder::build(MachineFunction *mf, const LiveIntervals *lis,
+                                  const MachineLoopInfo *loopInfo,
+                                  const RegSet &vregs) {
 
   LiveIntervals *LIS = const_cast<LiveIntervals*>(lis);
   MachineRegisterInfo *mri = &mf->getRegInfo();
   const TargetRegisterInfo *tri = mf->getTarget().getRegisterInfo();
 
-  std::auto_ptr<PBQPRAProblem> p(new PBQPRAProblem());
+  OwningPtr<PBQPRAProblem> p(new PBQPRAProblem());
   PBQP::Graph &g = p->getGraph();
   RegSet pregs;
 
@@ -282,7 +282,7 @@ std::auto_ptr<PBQPRAProblem> PBQPBuilder::build(MachineFunction *mf,
     }
   }
 
-  return p;
+  return p.take();
 }
 
 void PBQPBuilder::addSpillCosts(PBQP::Vector &costVec,
@@ -311,13 +311,12 @@ void PBQPBuilder::addInterferenceCosts(
   }
 }
 
-std::auto_ptr<PBQPRAProblem> PBQPBuilderWithCoalescing::build(
-                                                MachineFunction *mf,
+PBQPRAProblem *PBQPBuilderWithCoalescing::build(MachineFunction *mf,
                                                 const LiveIntervals *lis,
                                                 const MachineLoopInfo *loopInfo,
                                                 const RegSet &vregs) {
 
-  std::auto_ptr<PBQPRAProblem> p = PBQPBuilder::build(mf, lis, loopInfo, vregs);
+  OwningPtr<PBQPRAProblem> p(PBQPBuilder::build(mf, lis, loopInfo, vregs));
   PBQP::Graph &g = p->getGraph();
 
   const TargetMachine &tm = mf->getTarget();
@@ -391,7 +390,7 @@ std::auto_ptr<PBQPRAProblem> PBQPBuilderWithCoalescing::build(
     }
   }
 
-  return p;
+  return p.take();
 }
 
 void PBQPBuilderWithCoalescing::addPhysRegCoalesce(PBQP::Vector &costVec,
@@ -584,8 +583,8 @@ bool RegAllocPBQP::runOnMachineFunction(MachineFunction &MF) {
     while (!pbqpAllocComplete) {
       DEBUG(dbgs() << "  PBQP Regalloc round " << round << ":\n");
 
-      std::auto_ptr<PBQPRAProblem> problem =
-        builder->build(mf, lis, loopInfo, vregsToAlloc);
+      OwningPtr<PBQPRAProblem> problem(
+        builder->build(mf, lis, loopInfo, vregsToAlloc));
 
 #ifndef NDEBUG
       if (pbqpDumpGraphs) {
@@ -621,18 +620,18 @@ bool RegAllocPBQP::runOnMachineFunction(MachineFunction &MF) {
 }
 
 FunctionPass* llvm::createPBQPRegisterAllocator(
-                                           std::auto_ptr<PBQPBuilder> builder,
+                                           OwningPtr<PBQPBuilder> &builder,
                                            char *customPassID) {
   return new RegAllocPBQP(builder, customPassID);
 }
 
 FunctionPass* llvm::createDefaultPBQPRegisterAllocator() {
-  if (pbqpCoalescing) {
-    return createPBQPRegisterAllocator(
-             std::auto_ptr<PBQPBuilder>(new PBQPBuilderWithCoalescing()));
-  } // else
-  return createPBQPRegisterAllocator(
-           std::auto_ptr<PBQPBuilder>(new PBQPBuilder()));
+  OwningPtr<PBQPBuilder> Builder;
+  if (pbqpCoalescing)
+    Builder.reset(new PBQPBuilderWithCoalescing());
+  else
+    Builder.reset(new PBQPBuilder());
+  return createPBQPRegisterAllocator(Builder);
 }
 
 #undef DEBUG_TYPE
diff --git a/lib/CodeGen/RegisterScavenging.cpp b/lib/CodeGen/RegisterScavenging.cpp
index 07ace7a436c7..f82ccbe84da3 100644
--- a/lib/CodeGen/RegisterScavenging.cpp
+++ b/lib/CodeGen/RegisterScavenging.cpp
@@ -154,14 +154,13 @@ void RegScavenger::unprocess() {
   assert(Tracking && "Cannot unprocess because we're not tracking");
 
   MachineInstr *MI = MBBI;
-  if (MI->isDebugValue())
-    return;
-
-  determineKillsAndDefs();
+  if (!MI->isDebugValue()) {
+    determineKillsAndDefs();
 
-  // Commit the changes.
-  setUsed(KillRegs);
-  setUnused(DefRegs);
+    // Commit the changes.
+    setUsed(KillRegs);
+    setUnused(DefRegs);
+  }
 
   if (MBBI == MBB->begin()) {
     MBBI = MachineBasicBlock::iterator(NULL);
diff --git a/lib/CodeGen/ScheduleDAGInstrs.cpp b/lib/CodeGen/ScheduleDAGInstrs.cpp
index 71e7a21ef2bc..e4da6a41eead 100644
--- a/lib/CodeGen/ScheduleDAGInstrs.cpp
+++ b/lib/CodeGen/ScheduleDAGInstrs.cpp
@@ -262,6 +262,9 @@ void ScheduleDAGInstrs::addPhysRegDataDeps(SUnit *SU, unsigned OperIdx) {
       if (UseOp < 0)
         Dep = SDep(SU, SDep::Artificial);
       else {
+        // Set the hasPhysRegDefs only for physreg defs that have a use within
+        // the scheduling region.
+        SU->hasPhysRegDefs = true;
         Dep = SDep(SU, SDep::Data, *Alias);
         RegUse = UseSU->getInstr();
         Dep.setMinLatency(
@@ -318,6 +321,7 @@ void ScheduleDAGInstrs::addPhysRegDeps(SUnit *SU, unsigned OperIdx) {
   }
 
   if (!MO.isDef()) {
+    SU->hasPhysRegUses = true;
     // Either insert a new Reg2SUnits entry with an empty SUnits list, or
     // retrieve the existing SUnits list for this register's uses.
     // Push this SUnit on the use list.
diff --git a/lib/CodeGen/SelectionDAG/DAGCombiner.cpp b/lib/CodeGen/SelectionDAG/DAGCombiner.cpp
index eb1609575016..2e09ec08fdb1 100644
--- a/lib/CodeGen/SelectionDAG/DAGCombiner.cpp
+++ b/lib/CodeGen/SelectionDAG/DAGCombiner.cpp
@@ -205,6 +205,7 @@ namespace {
     SDValue visitCTTZ_ZERO_UNDEF(SDNode *N);
     SDValue visitCTPOP(SDNode *N);
     SDValue visitSELECT(SDNode *N);
+    SDValue visitVSELECT(SDNode *N);
     SDValue visitSELECT_CC(SDNode *N);
     SDValue visitSETCC(SDNode *N);
     SDValue visitSIGN_EXTEND(SDNode *N);
@@ -243,7 +244,6 @@ namespace {
     SDValue visitCONCAT_VECTORS(SDNode *N);
     SDValue visitEXTRACT_SUBVECTOR(SDNode *N);
     SDValue visitVECTOR_SHUFFLE(SDNode *N);
-    SDValue visitMEMBARRIER(SDNode *N);
 
     SDValue XformToShuffleWithZero(SDNode *N);
     SDValue ReassociateOps(unsigned Opc, DebugLoc DL, SDValue LHS, SDValue RHS);
@@ -1127,6 +1127,7 @@ SDValue DAGCombiner::visit(SDNode *N) {
   case ISD::CTTZ_ZERO_UNDEF:    return visitCTTZ_ZERO_UNDEF(N);
   case ISD::CTPOP:              return visitCTPOP(N);
   case ISD::SELECT:             return visitSELECT(N);
+  case ISD::VSELECT:            return visitVSELECT(N);
   case ISD::SELECT_CC:          return visitSELECT_CC(N);
   case ISD::SETCC:              return visitSETCC(N);
   case ISD::SIGN_EXTEND:        return visitSIGN_EXTEND(N);
@@ -1165,7 +1166,6 @@ SDValue DAGCombiner::visit(SDNode *N) {
   case ISD::CONCAT_VECTORS:     return visitCONCAT_VECTORS(N);
   case ISD::EXTRACT_SUBVECTOR:  return visitEXTRACT_SUBVECTOR(N);
   case ISD::VECTOR_SHUFFLE:     return visitVECTOR_SHUFFLE(N);
-  case ISD::MEMBARRIER:         return visitMEMBARRIER(N);
   }
   return SDValue();
 }
@@ -4164,6 +4164,46 @@ SDValue DAGCombiner::visitSELECT(SDNode *N) {
   return SDValue();
 }
 
+SDValue DAGCombiner::visitVSELECT(SDNode *N) {
+  SDValue N0 = N->getOperand(0);
+  SDValue N1 = N->getOperand(1);
+  SDValue N2 = N->getOperand(2);
+  DebugLoc DL = N->getDebugLoc();
+
+  // Canonicalize integer abs.
+  // vselect (setg[te] X,  0),  X, -X ->
+  // vselect (setgt    X, -1),  X, -X ->
+  // vselect (setl[te] X,  0), -X,  X ->
+  // Y = sra (X, size(X)-1); xor (add (X, Y), Y)
+  if (N0.getOpcode() == ISD::SETCC) {
+    SDValue LHS = N0.getOperand(0), RHS = N0.getOperand(1);
+    ISD::CondCode CC = cast<CondCodeSDNode>(N0.getOperand(2))->get();
+    bool isAbs = false;
+    bool RHSIsAllZeros = ISD::isBuildVectorAllZeros(RHS.getNode());
+
+    if (((RHSIsAllZeros && (CC == ISD::SETGT || CC == ISD::SETGE)) ||
+         (ISD::isBuildVectorAllOnes(RHS.getNode()) && CC == ISD::SETGT)) &&
+        N1 == LHS && N2.getOpcode() == ISD::SUB && N1 == N2.getOperand(1))
+      isAbs = ISD::isBuildVectorAllZeros(N2.getOperand(0).getNode());
+    else if ((RHSIsAllZeros && (CC == ISD::SETLT || CC == ISD::SETLE)) &&
+             N2 == LHS && N1.getOpcode() == ISD::SUB && N2 == N1.getOperand(1))
+      isAbs = ISD::isBuildVectorAllZeros(N1.getOperand(0).getNode());
+
+    if (isAbs) {
+      EVT VT = LHS.getValueType();
+      SDValue Shift = DAG.getNode(
+          ISD::SRA, DL, VT, LHS,
+          DAG.getConstant(VT.getScalarType().getSizeInBits() - 1, VT));
+      SDValue Add = DAG.getNode(ISD::ADD, DL, VT, LHS, Shift);
+      AddToWorkList(Shift.getNode());
+      AddToWorkList(Add.getNode());
+      return DAG.getNode(ISD::XOR, DL, VT, Add, Shift);
+    }
+  }
+
+  return SDValue();
+}
+
 SDValue DAGCombiner::visitSELECT_CC(SDNode *N) {
   SDValue N0 = N->getOperand(0);
   SDValue N1 = N->getOperand(1);
@@ -4453,7 +4493,9 @@ SDValue DAGCombiner::visitSIGN_EXTEND(SDNode *N) {
   if (N0.getOpcode() == ISD::SETCC) {
     // sext(setcc) -> sext_in_reg(vsetcc) for vectors.
     // Only do this before legalize for now.
-    if (VT.isVector() && !LegalOperations) {
+    if (VT.isVector() && !LegalOperations &&
+        TLI.getBooleanContents(true) == 
+          TargetLowering::ZeroOrNegativeOneBooleanContent) {
       EVT N0VT = N0.getOperand(0).getValueType();
       // On some architectures (such as SSE/NEON/etc) the SETCC result type is
       // of the same size as the compared operands. Only optimize sext(setcc())
@@ -7110,25 +7152,40 @@ bool DAGCombiner::CombineToPreIndexedLoadStore(SDNode *N) {
     assert(OtherUses[i]->getOperand(!OffsetIdx).getNode() ==
            BasePtr.getNode() && "Expected BasePtr operand");
 
-    APInt OV =
-      cast<ConstantSDNode>(Offset)->getAPIntValue();
-    if (AM == ISD::PRE_DEC)
-      OV = -OV;
+    // We need to replace ptr0 in the following expression:
+    //   x0 * offset0 + y0 * ptr0 = t0
+    // knowing that
+    //   x1 * offset1 + y1 * ptr0 = t1 (the indexed load/store)
+    // 
+    // where x0, x1, y0 and y1 in {-1, 1} are given by the types of the
+    // indexed load/store and the expresion that needs to be re-written.
+    //
+    // Therefore, we have:
+    //   t0 = (x0 * offset0 - x1 * y0 * y1 *offset1) + (y0 * y1) * t1
 
     ConstantSDNode *CN =
       cast<ConstantSDNode>(OtherUses[i]->getOperand(OffsetIdx));
-    APInt CNV = CN->getAPIntValue();
-    if (OtherUses[i]->getOpcode() == ISD::SUB && OffsetIdx == 1)
-      CNV += OV;
-    else
-      CNV -= OV;
+    int X0, X1, Y0, Y1;
+    APInt Offset0 = CN->getAPIntValue();
+    APInt Offset1 = cast<ConstantSDNode>(Offset)->getAPIntValue();
 
-    SDValue NewOp1 = Result.getValue(isLoad ? 1 : 0);
-    SDValue NewOp2 = DAG.getConstant(CNV, CN->getValueType(0));
-    if (OffsetIdx == 0)
-      std::swap(NewOp1, NewOp2);
+    X0 = (OtherUses[i]->getOpcode() == ISD::SUB && OffsetIdx == 1) ? -1 : 1;
+    Y0 = (OtherUses[i]->getOpcode() == ISD::SUB && OffsetIdx == 0) ? -1 : 1;
+    X1 = (AM == ISD::PRE_DEC && !Swapped) ? -1 : 1;
+    Y1 = (AM == ISD::PRE_DEC && Swapped) ? -1 : 1;
 
-    SDValue NewUse = DAG.getNode(OtherUses[i]->getOpcode(),
+    unsigned Opcode = (Y0 * Y1 < 0) ? ISD::SUB : ISD::ADD;
+
+    APInt CNV = Offset0;
+    if (X0 < 0) CNV = -CNV;
+    if (X1 * Y0 * Y1 < 0) CNV = CNV + Offset1;
+    else CNV = CNV - Offset1;
+
+    // We can now generate the new expression.
+    SDValue NewOp1 = DAG.getConstant(CNV, CN->getValueType(0));
+    SDValue NewOp2 = Result.getValue(isLoad ? 1 : 0);
+
+    SDValue NewUse = DAG.getNode(Opcode,
                                  OtherUses[i]->getDebugLoc(),
                                  OtherUses[i]->getValueType(0), NewOp1, NewOp2);
     DAG.ReplaceAllUsesOfValueWith(SDValue(OtherUses[i], 0), NewUse);
@@ -9065,6 +9122,51 @@ SDValue DAGCombiner::visitCONCAT_VECTORS(SDNode *N) {
   if (ISD::allOperandsUndef(N))
     return DAG.getUNDEF(N->getValueType(0));
 
+  // Type legalization of vectors and DAG canonicalization of SHUFFLE_VECTOR
+  // nodes often generate nop CONCAT_VECTOR nodes.
+  // Scan the CONCAT_VECTOR operands and look for a CONCAT operations that
+  // place the incoming vectors at the exact same location.
+  SDValue SingleSource = SDValue();
+  unsigned PartNumElem = N->getOperand(0).getValueType().getVectorNumElements();
+
+  for (unsigned i = 0, e = N->getNumOperands(); i != e; ++i) {
+    SDValue Op = N->getOperand(i);
+
+    if (Op.getOpcode() == ISD::UNDEF)
+      continue;
+
+    // Check if this is the identity extract:
+    if (Op.getOpcode() != ISD::EXTRACT_SUBVECTOR)
+      return SDValue();
+
+    // Find the single incoming vector for the extract_subvector.
+    if (SingleSource.getNode()) {
+      if (Op.getOperand(0) != SingleSource)
+        return SDValue();
+    } else {
+      SingleSource = Op.getOperand(0);
+
+      // Check the source type is the same as the type of the result.
+      // If not, this concat may extend the vector, so we can not
+      // optimize it away.
+      if (SingleSource.getValueType() != N->getValueType(0))
+        return SDValue();
+    }
+
+    unsigned IdentityIndex = i * PartNumElem;
+    ConstantSDNode *CS = dyn_cast<ConstantSDNode>(Op.getOperand(1));
+    // The extract index must be constant.
+    if (!CS)
+      return SDValue();
+    
+    // Check that we are reading from the identity index.
+    if (CS->getZExtValue() != IdentityIndex)
+      return SDValue();
+  }
+
+  if (SingleSource.getNode())
+    return SingleSource;
+  
   return SDValue();
 }
 
@@ -9125,6 +9227,44 @@ SDValue DAGCombiner::visitEXTRACT_SUBVECTOR(SDNode* N) {
   return SDValue();
 }
 
+// Tries to turn a shuffle of two CONCAT_VECTORS into a single concat.
+static SDValue partitionShuffleOfConcats(SDNode *N, SelectionDAG &DAG) {
+  EVT VT = N->getValueType(0);
+  unsigned NumElts = VT.getVectorNumElements();
+
+  SDValue N0 = N->getOperand(0);
+  SDValue N1 = N->getOperand(1);
+  ShuffleVectorSDNode *SVN = cast<ShuffleVectorSDNode>(N);
+
+  SmallVector<SDValue, 4> Ops;
+  EVT ConcatVT = N0.getOperand(0).getValueType();
+  unsigned NumElemsPerConcat = ConcatVT.getVectorNumElements();
+  unsigned NumConcats = NumElts / NumElemsPerConcat;
+
+  // Look at every vector that's inserted. We're looking for exact
+  // subvector-sized copies from a concatenated vector
+  for (unsigned I = 0; I != NumConcats; ++I) {
+    // Make sure we're dealing with a copy.
+    unsigned Begin = I * NumElemsPerConcat;
+    if (SVN->getMaskElt(Begin) % NumElemsPerConcat != 0)
+      return SDValue();
+
+    for (unsigned J = 1; J != NumElemsPerConcat; ++J) {
+      if (SVN->getMaskElt(Begin + J - 1) + 1 != SVN->getMaskElt(Begin + J))
+        return SDValue();
+    }
+
+    unsigned FirstElt = SVN->getMaskElt(Begin) / NumElemsPerConcat;
+    if (FirstElt < N0.getNumOperands())
+      Ops.push_back(N0.getOperand(FirstElt));
+    else
+      Ops.push_back(N1.getOperand(FirstElt - N0.getNumOperands()));
+  }
+
+  return DAG.getNode(ISD::CONCAT_VECTORS, N->getDebugLoc(), VT, Ops.data(),
+                     Ops.size());
+}
+
 SDValue DAGCombiner::visitVECTOR_SHUFFLE(SDNode *N) {
   EVT VT = N->getValueType(0);
   unsigned NumElts = VT.getVectorNumElements();
@@ -9226,6 +9366,17 @@ SDValue DAGCombiner::visitVECTOR_SHUFFLE(SDNode *N) {
     }
   }
 
+  if (N0.getOpcode() == ISD::CONCAT_VECTORS &&
+      Level < AfterLegalizeVectorOps &&
+      (N1.getOpcode() == ISD::UNDEF ||
+      (N1.getOpcode() == ISD::CONCAT_VECTORS &&
+       N0.getOperand(0).getValueType() == N1.getOperand(0).getValueType()))) {
+    SDValue V = partitionShuffleOfConcats(N, DAG);
+
+    if (V.getNode())
+      return V;
+  }
+
   // If this shuffle node is simply a swizzle of another shuffle node,
   // and it reverses the swizzle of the previous shuffle then we can
   // optimize shuffle(shuffle(x, undef), undef) -> x.
@@ -9262,59 +9413,6 @@ SDValue DAGCombiner::visitVECTOR_SHUFFLE(SDNode *N) {
   return SDValue();
 }
 
-SDValue DAGCombiner::visitMEMBARRIER(SDNode* N) {
-  if (!TLI.getShouldFoldAtomicFences())
-    return SDValue();
-
-  SDValue atomic = N->getOperand(0);
-  switch (atomic.getOpcode()) {
-    case ISD::ATOMIC_CMP_SWAP:
-    case ISD::ATOMIC_SWAP:
-    case ISD::ATOMIC_LOAD_ADD:
-    case ISD::ATOMIC_LOAD_SUB:
-    case ISD::ATOMIC_LOAD_AND:
-    case ISD::ATOMIC_LOAD_OR:
-    case ISD::ATOMIC_LOAD_XOR:
-    case ISD::ATOMIC_LOAD_NAND:
-    case ISD::ATOMIC_LOAD_MIN:
-    case ISD::ATOMIC_LOAD_MAX:
-    case ISD::ATOMIC_LOAD_UMIN:
-    case ISD::ATOMIC_LOAD_UMAX:
-      break;
-    default:
-      return SDValue();
-  }
-
-  SDValue fence = atomic.getOperand(0);
-  if (fence.getOpcode() != ISD::MEMBARRIER)
-    return SDValue();
-
-  switch (atomic.getOpcode()) {
-    case ISD::ATOMIC_CMP_SWAP:
-      return SDValue(DAG.UpdateNodeOperands(atomic.getNode(),
-                                    fence.getOperand(0),
-                                    atomic.getOperand(1), atomic.getOperand(2),
-                                    atomic.getOperand(3)), atomic.getResNo());
-    case ISD::ATOMIC_SWAP:
-    case ISD::ATOMIC_LOAD_ADD:
-    case ISD::ATOMIC_LOAD_SUB:
-    case ISD::ATOMIC_LOAD_AND:
-    case ISD::ATOMIC_LOAD_OR:
-    case ISD::ATOMIC_LOAD_XOR:
-    case ISD::ATOMIC_LOAD_NAND:
-    case ISD::ATOMIC_LOAD_MIN:
-    case ISD::ATOMIC_LOAD_MAX:
-    case ISD::ATOMIC_LOAD_UMIN:
-    case ISD::ATOMIC_LOAD_UMAX:
-      return SDValue(DAG.UpdateNodeOperands(atomic.getNode(),
-                                    fence.getOperand(0),
-                                    atomic.getOperand(1), atomic.getOperand(2)),
-                     atomic.getResNo());
-    default:
-      return SDValue();
-  }
-}
-
 /// XformToShuffleWithZero - Returns a vector_shuffle if it able to transform
 /// an AND to a vector_shuffle with the destination vector and a zero vector.
 /// e.g. AND V, <0xffffffff, 0, 0xffffffff, 0>. ==>
diff --git a/lib/CodeGen/SelectionDAG/FastISel.cpp b/lib/CodeGen/SelectionDAG/FastISel.cpp
index 9ac738e50726..288499ac6f32 100644
--- a/lib/CodeGen/SelectionDAG/FastISel.cpp
+++ b/lib/CodeGen/SelectionDAG/FastISel.cpp
@@ -1505,3 +1505,61 @@ bool FastISel::HandlePHINodesInSuccessorBlocks(const BasicBlock *LLVMBB) {
 
   return true;
 }
+
+bool FastISel::tryToFoldLoad(const LoadInst *LI, const Instruction *FoldInst) {
+  assert(LI->hasOneUse() &&
+      "tryToFoldLoad expected a LoadInst with a single use");
+  // We know that the load has a single use, but don't know what it is.  If it
+  // isn't one of the folded instructions, then we can't succeed here.  Handle
+  // this by scanning the single-use users of the load until we get to FoldInst.
+  unsigned MaxUsers = 6;  // Don't scan down huge single-use chains of instrs.
+
+  const Instruction *TheUser = LI->use_back();
+  while (TheUser != FoldInst &&   // Scan up until we find FoldInst.
+         // Stay in the right block.
+         TheUser->getParent() == FoldInst->getParent() &&
+         --MaxUsers) {  // Don't scan too far.
+    // If there are multiple or no uses of this instruction, then bail out.
+    if (!TheUser->hasOneUse())
+      return false;
+
+    TheUser = TheUser->use_back();
+  }
+
+  // If we didn't find the fold instruction, then we failed to collapse the
+  // sequence.
+  if (TheUser != FoldInst)
+    return false;
+
+  // Don't try to fold volatile loads.  Target has to deal with alignment
+  // constraints.
+  if (LI->isVolatile())
+    return false;
+
+  // Figure out which vreg this is going into.  If there is no assigned vreg yet
+  // then there actually was no reference to it.  Perhaps the load is referenced
+  // by a dead instruction.
+  unsigned LoadReg = getRegForValue(LI);
+  if (LoadReg == 0)
+    return false;
+
+  // We can't fold if this vreg has no uses or more than one use.  Multiple uses
+  // may mean that the instruction got lowered to multiple MIs, or the use of
+  // the loaded value ended up being multiple operands of the result.
+  if (!MRI.hasOneUse(LoadReg))
+    return false;
+
+  MachineRegisterInfo::reg_iterator RI = MRI.reg_begin(LoadReg);
+  MachineInstr *User = &*RI;
+
+  // Set the insertion point properly.  Folding the load can cause generation of
+  // other random instructions (like sign extends) for addressing modes; make
+  // sure they get inserted in a logical place before the new instruction.
+  FuncInfo.InsertPt = User;
+  FuncInfo.MBB = User->getParent();
+
+  // Ask the target to try folding the load.
+  return tryToFoldLoadIntoMI(User, RI.getOperandNo(), LI);
+}
+
+
diff --git a/lib/CodeGen/SelectionDAG/LegalizeDAG.cpp b/lib/CodeGen/SelectionDAG/LegalizeDAG.cpp
index 51cc254b2c82..2a1d8c2819ea 100644
--- a/lib/CodeGen/SelectionDAG/LegalizeDAG.cpp
+++ b/lib/CodeGen/SelectionDAG/LegalizeDAG.cpp
@@ -2759,8 +2759,7 @@ void SelectionDAGLegalize::ExpandNode(SDNode *Node) {
     Results.push_back(DAG.getConstant(0, MVT::i32));
     Results.push_back(Node->getOperand(0));
     break;
-  case ISD::ATOMIC_FENCE:
-  case ISD::MEMBARRIER: {
+  case ISD::ATOMIC_FENCE: {
     // If the target didn't lower this, lower it to '__sync_synchronize()' call
     // FIXME: handle "fence singlethread" more efficiently.
     TargetLowering::ArgListTy Args;
diff --git a/lib/CodeGen/SelectionDAG/LegalizeIntegerTypes.cpp b/lib/CodeGen/SelectionDAG/LegalizeIntegerTypes.cpp
index d19c13b8ff13..cd2f060ce0f3 100644
--- a/lib/CodeGen/SelectionDAG/LegalizeIntegerTypes.cpp
+++ b/lib/CodeGen/SelectionDAG/LegalizeIntegerTypes.cpp
@@ -777,7 +777,6 @@ bool DAGTypeLegalizer::PromoteIntegerOperand(SDNode *N, unsigned OpNo) {
                           Res = PromoteIntOp_CONVERT_RNDSAT(N); break;
   case ISD::INSERT_VECTOR_ELT:
                           Res = PromoteIntOp_INSERT_VECTOR_ELT(N, OpNo);break;
-  case ISD::MEMBARRIER:   Res = PromoteIntOp_MEMBARRIER(N); break;
   case ISD::SCALAR_TO_VECTOR:
                           Res = PromoteIntOp_SCALAR_TO_VECTOR(N); break;
   case ISD::VSELECT:
@@ -961,17 +960,6 @@ SDValue DAGTypeLegalizer::PromoteIntOp_INSERT_VECTOR_ELT(SDNode *N,
                                 N->getOperand(1), Idx), 0);
 }
 
-SDValue DAGTypeLegalizer::PromoteIntOp_MEMBARRIER(SDNode *N) {
-  SDValue NewOps[6];
-  DebugLoc dl = N->getDebugLoc();
-  NewOps[0] = N->getOperand(0);
-  for (unsigned i = 1; i < array_lengthof(NewOps); ++i) {
-    SDValue Flag = GetPromotedInteger(N->getOperand(i));
-    NewOps[i] = DAG.getZeroExtendInReg(Flag, dl, MVT::i1);
-  }
-  return SDValue(DAG.UpdateNodeOperands(N, NewOps, array_lengthof(NewOps)), 0);
-}
-
 SDValue DAGTypeLegalizer::PromoteIntOp_SCALAR_TO_VECTOR(SDNode *N) {
   // Integer SCALAR_TO_VECTOR operands are implicitly truncated, so just promote
   // the operand in place.
diff --git a/lib/CodeGen/SelectionDAG/LegalizeTypes.h b/lib/CodeGen/SelectionDAG/LegalizeTypes.h
index 54ea926241cf..1c4274a91089 100644
--- a/lib/CodeGen/SelectionDAG/LegalizeTypes.h
+++ b/lib/CodeGen/SelectionDAG/LegalizeTypes.h
@@ -270,7 +270,6 @@ private:
   SDValue PromoteIntOp_EXTRACT_ELEMENT(SDNode *N);
   SDValue PromoteIntOp_EXTRACT_VECTOR_ELT(SDNode *N);
   SDValue PromoteIntOp_CONCAT_VECTORS(SDNode *N);
-  SDValue PromoteIntOp_MEMBARRIER(SDNode *N);
   SDValue PromoteIntOp_SCALAR_TO_VECTOR(SDNode *N);
   SDValue PromoteIntOp_SELECT(SDNode *N, unsigned OpNo);
   SDValue PromoteIntOp_SELECT_CC(SDNode *N, unsigned OpNo);
@@ -582,6 +581,7 @@ private:
   SDValue SplitVecOp_EXTRACT_VECTOR_ELT(SDNode *N);
   SDValue SplitVecOp_STORE(StoreSDNode *N, unsigned OpNo);
   SDValue SplitVecOp_CONCAT_VECTORS(SDNode *N);
+  SDValue SplitVecOp_TRUNCATE(SDNode *N);
   SDValue SplitVecOp_VSETCC(SDNode *N);
   SDValue SplitVecOp_FP_ROUND(SDNode *N);
 
diff --git a/lib/CodeGen/SelectionDAG/LegalizeVectorTypes.cpp b/lib/CodeGen/SelectionDAG/LegalizeVectorTypes.cpp
index 5ec853563888..04c6bfd0c23a 100644
--- a/lib/CodeGen/SelectionDAG/LegalizeVectorTypes.cpp
+++ b/lib/CodeGen/SelectionDAG/LegalizeVectorTypes.cpp
@@ -1046,6 +1046,7 @@ bool DAGTypeLegalizer::SplitVectorOperand(SDNode *N, unsigned OpNo) {
     case ISD::EXTRACT_SUBVECTOR: Res = SplitVecOp_EXTRACT_SUBVECTOR(N); break;
     case ISD::EXTRACT_VECTOR_ELT:Res = SplitVecOp_EXTRACT_VECTOR_ELT(N); break;
     case ISD::CONCAT_VECTORS:    Res = SplitVecOp_CONCAT_VECTORS(N); break;
+    case ISD::TRUNCATE:          Res = SplitVecOp_TRUNCATE(N); break;
     case ISD::FP_ROUND:          Res = SplitVecOp_FP_ROUND(N); break;
     case ISD::STORE:
       Res = SplitVecOp_STORE(cast<StoreSDNode>(N), OpNo);
@@ -1062,7 +1063,6 @@ bool DAGTypeLegalizer::SplitVectorOperand(SDNode *N, unsigned OpNo) {
     case ISD::SINT_TO_FP:
     case ISD::UINT_TO_FP:
     case ISD::FTRUNC:
-    case ISD::TRUNCATE:
     case ISD::SIGN_EXTEND:
     case ISD::ZERO_EXTEND:
     case ISD::ANY_EXTEND:
@@ -1272,8 +1272,8 @@ SDValue DAGTypeLegalizer::SplitVecOp_STORE(StoreSDNode *N, unsigned OpNo) {
 SDValue DAGTypeLegalizer::SplitVecOp_CONCAT_VECTORS(SDNode *N) {
   DebugLoc DL = N->getDebugLoc();
 
-  // The input operands all must have the same type, and we know the result the
-  // result type is valid.  Convert this to a buildvector which extracts all the
+  // The input operands all must have the same type, and we know the result
+  // type is valid.  Convert this to a buildvector which extracts all the
   // input elements.
   // TODO: If the input elements are power-two vectors, we could convert this to
   // a new CONCAT_VECTORS node with elements that are half-wide.
@@ -1293,6 +1293,66 @@ SDValue DAGTypeLegalizer::SplitVecOp_CONCAT_VECTORS(SDNode *N) {
                      &Elts[0], Elts.size());
 }
 
+SDValue DAGTypeLegalizer::SplitVecOp_TRUNCATE(SDNode *N) {
+  // The result type is legal, but the input type is illegal.  If splitting
+  // ends up with the result type of each half still being legal, just
+  // do that.  If, however, that would result in an illegal result type,
+  // we can try to get more clever with power-two vectors. Specifically,
+  // split the input type, but also widen the result element size, then
+  // concatenate the halves and truncate again.  For example, consider a target
+  // where v8i8 is legal and v8i32 is not (ARM, which doesn't have 256-bit
+  // vectors). To perform a "%res = v8i8 trunc v8i32 %in" we do:
+  //   %inlo = v4i32 extract_subvector %in, 0
+  //   %inhi = v4i32 extract_subvector %in, 4
+  //   %lo16 = v4i16 trunc v4i32 %inlo
+  //   %hi16 = v4i16 trunc v4i32 %inhi
+  //   %in16 = v8i16 concat_vectors v4i16 %lo16, v4i16 %hi16
+  //   %res = v8i8 trunc v8i16 %in16
+  //
+  // Without this transform, the original truncate would end up being
+  // scalarized, which is pretty much always a last resort.
+  SDValue InVec = N->getOperand(0);
+  EVT InVT = InVec->getValueType(0);
+  EVT OutVT = N->getValueType(0);
+  unsigned NumElements = OutVT.getVectorNumElements();
+  // Widening should have already made sure this is a power-two vector
+  // if we're trying to split it at all. assert() that's true, just in case.
+  assert(!(NumElements & 1) && "Splitting vector, but not in half!");
+
+  unsigned InElementSize = InVT.getVectorElementType().getSizeInBits();
+  unsigned OutElementSize = OutVT.getVectorElementType().getSizeInBits();
+
+  // If the input elements are only 1/2 the width of the result elements,
+  // just use the normal splitting. Our trick only work if there's room
+  // to split more than once.
+  if (InElementSize <= OutElementSize * 2)
+    return SplitVecOp_UnaryOp(N);
+  DebugLoc DL = N->getDebugLoc();
+
+  // Extract the halves of the input via extract_subvector.
+  EVT SplitVT = EVT::getVectorVT(*DAG.getContext(),
+                                 InVT.getVectorElementType(), NumElements/2);
+  SDValue InLoVec = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, SplitVT, InVec,
+                                DAG.getIntPtrConstant(0));
+  SDValue InHiVec = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, SplitVT, InVec,
+                                DAG.getIntPtrConstant(NumElements/2));
+  // Truncate them to 1/2 the element size.
+  EVT HalfElementVT = EVT::getIntegerVT(*DAG.getContext(), InElementSize/2);
+  EVT HalfVT = EVT::getVectorVT(*DAG.getContext(), HalfElementVT,
+                                NumElements/2);
+  SDValue HalfLo = DAG.getNode(ISD::TRUNCATE, DL, HalfVT, InLoVec);
+  SDValue HalfHi = DAG.getNode(ISD::TRUNCATE, DL, HalfVT, InHiVec);
+  // Concatenate them to get the full intermediate truncation result.
+  EVT InterVT = EVT::getVectorVT(*DAG.getContext(), HalfElementVT, NumElements);
+  SDValue InterVec = DAG.getNode(ISD::CONCAT_VECTORS, DL, InterVT, HalfLo,
+                                 HalfHi);
+  // Now finish up by truncating all the way down to the original result
+  // type. This should normally be something that ends up being legal directly,
+  // but in theory if a target has very wide vectors and an annoyingly
+  // restricted set of legal types, this split can chain to build things up.
+  return DAG.getNode(ISD::TRUNCATE, DL, OutVT, InterVec);
+}
+
 SDValue DAGTypeLegalizer::SplitVecOp_VSETCC(SDNode *N) {
   assert(N->getValueType(0).isVector() &&
          N->getOperand(0).getValueType().isVector() &&
diff --git a/lib/CodeGen/SelectionDAG/SelectionDAG.cpp b/lib/CodeGen/SelectionDAG/SelectionDAG.cpp
index 64244313a326..15235c8ac307 100644
--- a/lib/CodeGen/SelectionDAG/SelectionDAG.cpp
+++ b/lib/CodeGen/SelectionDAG/SelectionDAG.cpp
@@ -2785,7 +2785,7 @@ SDValue SelectionDAG::FoldConstantArithmetic(unsigned Opcode, EVT VT,
   }
 
   // Handle the scalar case first.
-  if (Outputs.size() == 1)
+  if (Scalar1 && Scalar2)
     return Outputs.back();
 
   // Otherwise build a big vector out of the scalar elements we generated.
@@ -5252,14 +5252,14 @@ SDNode *SelectionDAG::MorphNodeTo(SDNode *N, unsigned Opc,
 MachineSDNode *
 SelectionDAG::getMachineNode(unsigned Opcode, DebugLoc dl, EVT VT) {
   SDVTList VTs = getVTList(VT);
-  return getMachineNode(Opcode, dl, VTs, 0, 0);
+  return getMachineNode(Opcode, dl, VTs, None);
 }
 
 MachineSDNode *
 SelectionDAG::getMachineNode(unsigned Opcode, DebugLoc dl, EVT VT, SDValue Op1) {
   SDVTList VTs = getVTList(VT);
   SDValue Ops[] = { Op1 };
-  return getMachineNode(Opcode, dl, VTs, Ops, array_lengthof(Ops));
+  return getMachineNode(Opcode, dl, VTs, Ops);
 }
 
 MachineSDNode *
@@ -5267,7 +5267,7 @@ SelectionDAG::getMachineNode(unsigned Opcode, DebugLoc dl, EVT VT,
                              SDValue Op1, SDValue Op2) {
   SDVTList VTs = getVTList(VT);
   SDValue Ops[] = { Op1, Op2 };
-  return getMachineNode(Opcode, dl, VTs, Ops, array_lengthof(Ops));
+  return getMachineNode(Opcode, dl, VTs, Ops);
 }
 
 MachineSDNode *
@@ -5275,20 +5275,20 @@ SelectionDAG::getMachineNode(unsigned Opcode, DebugLoc dl, EVT VT,
                              SDValue Op1, SDValue Op2, SDValue Op3) {
   SDVTList VTs = getVTList(VT);
   SDValue Ops[] = { Op1, Op2, Op3 };
-  return getMachineNode(Opcode, dl, VTs, Ops, array_lengthof(Ops));
+  return getMachineNode(Opcode, dl, VTs, Ops);
 }
 
 MachineSDNode *
 SelectionDAG::getMachineNode(unsigned Opcode, DebugLoc dl, EVT VT,
-                             const SDValue *Ops, unsigned NumOps) {
+                             ArrayRef<SDValue> Ops) {
   SDVTList VTs = getVTList(VT);
-  return getMachineNode(Opcode, dl, VTs, Ops, NumOps);
+  return getMachineNode(Opcode, dl, VTs, Ops);
 }
 
 MachineSDNode *
 SelectionDAG::getMachineNode(unsigned Opcode, DebugLoc dl, EVT VT1, EVT VT2) {
   SDVTList VTs = getVTList(VT1, VT2);
-  return getMachineNode(Opcode, dl, VTs, 0, 0);
+  return getMachineNode(Opcode, dl, VTs, None);
 }
 
 MachineSDNode *
@@ -5296,7 +5296,7 @@ SelectionDAG::getMachineNode(unsigned Opcode, DebugLoc dl,
                              EVT VT1, EVT VT2, SDValue Op1) {
   SDVTList VTs = getVTList(VT1, VT2);
   SDValue Ops[] = { Op1 };
-  return getMachineNode(Opcode, dl, VTs, Ops, array_lengthof(Ops));
+  return getMachineNode(Opcode, dl, VTs, Ops);
 }
 
 MachineSDNode *
@@ -5304,7 +5304,7 @@ SelectionDAG::getMachineNode(unsigned Opcode, DebugLoc dl,
                              EVT VT1, EVT VT2, SDValue Op1, SDValue Op2) {
   SDVTList VTs = getVTList(VT1, VT2);
   SDValue Ops[] = { Op1, Op2 };
-  return getMachineNode(Opcode, dl, VTs, Ops, array_lengthof(Ops));
+  return getMachineNode(Opcode, dl, VTs, Ops);
 }
 
 MachineSDNode *
@@ -5313,15 +5313,15 @@ SelectionDAG::getMachineNode(unsigned Opcode, DebugLoc dl,
                              SDValue Op2, SDValue Op3) {
   SDVTList VTs = getVTList(VT1, VT2);
   SDValue Ops[] = { Op1, Op2, Op3 };
-  return getMachineNode(Opcode, dl, VTs, Ops, array_lengthof(Ops));
+  return getMachineNode(Opcode, dl, VTs, Ops);
 }
 
 MachineSDNode *
 SelectionDAG::getMachineNode(unsigned Opcode, DebugLoc dl,
                              EVT VT1, EVT VT2,
-                             const SDValue *Ops, unsigned NumOps) {
+                             ArrayRef<SDValue> Ops) {
   SDVTList VTs = getVTList(VT1, VT2);
-  return getMachineNode(Opcode, dl, VTs, Ops, NumOps);
+  return getMachineNode(Opcode, dl, VTs, Ops);
 }
 
 MachineSDNode *
@@ -5330,7 +5330,7 @@ SelectionDAG::getMachineNode(unsigned Opcode, DebugLoc dl,
                              SDValue Op1, SDValue Op2) {
   SDVTList VTs = getVTList(VT1, VT2, VT3);
   SDValue Ops[] = { Op1, Op2 };
-  return getMachineNode(Opcode, dl, VTs, Ops, array_lengthof(Ops));
+  return getMachineNode(Opcode, dl, VTs, Ops);
 }
 
 MachineSDNode *
@@ -5339,39 +5339,41 @@ SelectionDAG::getMachineNode(unsigned Opcode, DebugLoc dl,
                              SDValue Op1, SDValue Op2, SDValue Op3) {
   SDVTList VTs = getVTList(VT1, VT2, VT3);
   SDValue Ops[] = { Op1, Op2, Op3 };
-  return getMachineNode(Opcode, dl, VTs, Ops, array_lengthof(Ops));
+  return getMachineNode(Opcode, dl, VTs, Ops);
 }
 
 MachineSDNode *
 SelectionDAG::getMachineNode(unsigned Opcode, DebugLoc dl,
                              EVT VT1, EVT VT2, EVT VT3,
-                             const SDValue *Ops, unsigned NumOps) {
+                             ArrayRef<SDValue> Ops) {
   SDVTList VTs = getVTList(VT1, VT2, VT3);
-  return getMachineNode(Opcode, dl, VTs, Ops, NumOps);
+  return getMachineNode(Opcode, dl, VTs, Ops);
 }
 
 MachineSDNode *
 SelectionDAG::getMachineNode(unsigned Opcode, DebugLoc dl, EVT VT1,
                              EVT VT2, EVT VT3, EVT VT4,
-                             const SDValue *Ops, unsigned NumOps) {
+                             ArrayRef<SDValue> Ops) {
   SDVTList VTs = getVTList(VT1, VT2, VT3, VT4);
-  return getMachineNode(Opcode, dl, VTs, Ops, NumOps);
+  return getMachineNode(Opcode, dl, VTs, Ops);
 }
 
 MachineSDNode *
 SelectionDAG::getMachineNode(unsigned Opcode, DebugLoc dl,
                              ArrayRef<EVT> ResultTys,
-                             const SDValue *Ops, unsigned NumOps) {
+                             ArrayRef<SDValue> Ops) {
   SDVTList VTs = getVTList(&ResultTys[0], ResultTys.size());
-  return getMachineNode(Opcode, dl, VTs, Ops, NumOps);
+  return getMachineNode(Opcode, dl, VTs, Ops);
 }
 
 MachineSDNode *
 SelectionDAG::getMachineNode(unsigned Opcode, DebugLoc DL, SDVTList VTs,
-                             const SDValue *Ops, unsigned NumOps) {
+                             ArrayRef<SDValue> OpsArray) {
   bool DoCSE = VTs.VTs[VTs.NumVTs-1] != MVT::Glue;
   MachineSDNode *N;
   void *IP = 0;
+  const SDValue *Ops = OpsArray.data();
+  unsigned NumOps = OpsArray.size();
 
   if (DoCSE) {
     FoldingSetNodeID ID;
diff --git a/lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp b/lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp
index ce40cd6a0c9c..67db211ec427 100644
--- a/lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp
+++ b/lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp
@@ -314,7 +314,7 @@ static SDValue getCopyFromPartsVector(SelectionDAG &DAG, DebugLoc DL,
     } else {
       Ctx.emitError(ErrMsg);
     }
-    report_fatal_error("Cannot handle scalar-to-vector conversion!");
+    return DAG.getUNDEF(ValueVT);
   }
 
   if (ValueVT.getVectorNumElements() == 1 &&
@@ -5034,6 +5034,11 @@ SelectionDAGBuilder::visitIntrinsicCall(const CallInst &I, unsigned Intrinsic) {
     setValue(&I, Res);
     return 0;
   }
+  case Intrinsic::annotation:
+  case Intrinsic::ptr_annotation:
+    // Drop the intrinsic, but forward the value
+    setValue(&I, getValue(I.getOperand(0)));
+    return 0;
   case Intrinsic::var_annotation:
     // Discard annotate attributes
     return 0;
@@ -5232,6 +5237,7 @@ void SelectionDAGBuilder::LowerCallTo(ImmutableCallSite CS, SDValue Callee,
     Entry.isSRet = true;
     Entry.isNest = false;
     Entry.isByVal = false;
+    Entry.isReturned = false;
     Entry.Alignment = Align;
     Args.push_back(Entry);
     RetTy = Type::getVoidTy(FTy->getContext());
@@ -5249,13 +5255,14 @@ void SelectionDAGBuilder::LowerCallTo(ImmutableCallSite CS, SDValue Callee,
     Entry.Node = ArgNode; Entry.Ty = V->getType();
 
     unsigned attrInd = i - CS.arg_begin() + 1;
-    Entry.isSExt  = CS.paramHasAttr(attrInd, Attribute::SExt);
-    Entry.isZExt  = CS.paramHasAttr(attrInd, Attribute::ZExt);
-    Entry.isInReg = CS.paramHasAttr(attrInd, Attribute::InReg);
-    Entry.isSRet  = CS.paramHasAttr(attrInd, Attribute::StructRet);
-    Entry.isNest  = CS.paramHasAttr(attrInd, Attribute::Nest);
-    Entry.isByVal = CS.paramHasAttr(attrInd, Attribute::ByVal);
-    Entry.Alignment = CS.getParamAlignment(attrInd);
+    Entry.isSExt     = CS.paramHasAttr(attrInd, Attribute::SExt);
+    Entry.isZExt     = CS.paramHasAttr(attrInd, Attribute::ZExt);
+    Entry.isInReg    = CS.paramHasAttr(attrInd, Attribute::InReg);
+    Entry.isSRet     = CS.paramHasAttr(attrInd, Attribute::StructRet);
+    Entry.isNest     = CS.paramHasAttr(attrInd, Attribute::Nest);
+    Entry.isByVal    = CS.paramHasAttr(attrInd, Attribute::ByVal);
+    Entry.isReturned = CS.paramHasAttr(attrInd, Attribute::Returned);
+    Entry.Alignment  = CS.getParamAlignment(attrInd);
     Args.push_back(Entry);
   }
 
@@ -6169,10 +6176,17 @@ void SelectionDAGBuilder::visitInlineAsm(ImmutableCallSite CS) {
           MatchedRegs.RegVTs.push_back(RegVT);
           MachineRegisterInfo &RegInfo = DAG.getMachineFunction().getRegInfo();
           for (unsigned i = 0, e = InlineAsm::getNumOperandRegisters(OpFlag);
-               i != e; ++i)
-            MatchedRegs.Regs.push_back
-              (RegInfo.createVirtualRegister(TLI.getRegClassFor(RegVT)));
-
+               i != e; ++i) {
+            if (const TargetRegisterClass *RC = TLI.getRegClassFor(RegVT))
+              MatchedRegs.Regs.push_back(RegInfo.createVirtualRegister(RC));
+            else {
+              LLVMContext &Ctx = *DAG.getContext();
+              Ctx.emitError(CS.getInstruction(), "inline asm error: This value"
+                            " type register class is not natively supported!");
+              report_fatal_error("inline asm error: This value type register "
+                                 "class is not natively supported!");
+            }
+          }
           // Use the produced MatchedRegs object to
           MatchedRegs.getCopyToRegs(InOperandVal, DAG, getCurDebugLoc(),
                                     Chain, &Flag, CS.getInstruction());
@@ -6389,6 +6403,28 @@ void SelectionDAGBuilder::visitVACopy(const CallInst &I) {
 /// migrated to using LowerCall, this hook should be integrated into SDISel.
 std::pair<SDValue, SDValue>
 TargetLowering::LowerCallTo(TargetLowering::CallLoweringInfo &CLI) const {
+  // Handle the incoming return values from the call.
+  CLI.Ins.clear();
+  SmallVector<EVT, 4> RetTys;
+  ComputeValueVTs(*this, CLI.RetTy, RetTys);
+  for (unsigned I = 0, E = RetTys.size(); I != E; ++I) {
+    EVT VT = RetTys[I];
+    MVT RegisterVT = getRegisterType(CLI.RetTy->getContext(), VT);
+    unsigned NumRegs = getNumRegisters(CLI.RetTy->getContext(), VT);
+    for (unsigned i = 0; i != NumRegs; ++i) {
+      ISD::InputArg MyFlags;
+      MyFlags.VT = RegisterVT;
+      MyFlags.Used = CLI.IsReturnValueUsed;
+      if (CLI.RetSExt)
+        MyFlags.Flags.setSExt();
+      if (CLI.RetZExt)
+        MyFlags.Flags.setZExt();
+      if (CLI.IsInReg)
+        MyFlags.Flags.setInReg();
+      CLI.Ins.push_back(MyFlags);
+    }
+  }
+
   // Handle all of the outgoing arguments.
   CLI.Outs.clear();
   CLI.OutVals.clear();
@@ -6442,6 +6478,26 @@ TargetLowering::LowerCallTo(TargetLowering::CallLoweringInfo &CLI) const {
       else if (Args[i].isZExt)
         ExtendKind = ISD::ZERO_EXTEND;
 
+      // Conservatively only handle 'returned' on non-vectors for now
+      if (Args[i].isReturned && !Op.getValueType().isVector()) {
+        assert(CLI.RetTy == Args[i].Ty && RetTys.size() == NumValues &&
+               "unexpected use of 'returned'");
+        // Before passing 'returned' to the target lowering code, ensure that
+        // either the register MVT and the actual EVT are the same size or that
+        // the return value and argument are extended in the same way; in these
+        // cases it's safe to pass the argument register value unchanged as the
+        // return register value (although it's at the target's option whether
+        // to do so)
+        // TODO: allow code generation to take advantage of partially preserved
+        // registers rather than clobbering the entire register when the
+        // parameter extension method is not compatible with the return
+        // extension method
+        if ((NumParts * PartVT.getSizeInBits() == VT.getSizeInBits()) ||
+            (ExtendKind != ISD::ANY_EXTEND &&
+             CLI.RetSExt == Args[i].isSExt && CLI.RetZExt == Args[i].isZExt))
+        Flags.setReturned();
+      }
+
       getCopyToParts(CLI.DAG, CLI.DL, Op, &Parts[0], NumParts,
                      PartVT, CLI.CS ? CLI.CS->getInstruction() : 0, ExtendKind);
 
@@ -6461,28 +6517,6 @@ TargetLowering::LowerCallTo(TargetLowering::CallLoweringInfo &CLI) const {
     }
   }
 
-  // Handle the incoming return values from the call.
-  CLI.Ins.clear();
-  SmallVector<EVT, 4> RetTys;
-  ComputeValueVTs(*this, CLI.RetTy, RetTys);
-  for (unsigned I = 0, E = RetTys.size(); I != E; ++I) {
-    EVT VT = RetTys[I];
-    MVT RegisterVT = getRegisterType(CLI.RetTy->getContext(), VT);
-    unsigned NumRegs = getNumRegisters(CLI.RetTy->getContext(), VT);
-    for (unsigned i = 0; i != NumRegs; ++i) {
-      ISD::InputArg MyFlags;
-      MyFlags.VT = RegisterVT;
-      MyFlags.Used = CLI.IsReturnValueUsed;
-      if (CLI.RetSExt)
-        MyFlags.Flags.setSExt();
-      if (CLI.RetZExt)
-        MyFlags.Flags.setZExt();
-      if (CLI.IsInReg)
-        MyFlags.Flags.setInReg();
-      CLI.Ins.push_back(MyFlags);
-    }
-  }
-
   SmallVector<SDValue, 4> InVals;
   CLI.Chain = LowerCall(CLI, InVals);
 
diff --git a/lib/CodeGen/SelectionDAG/SelectionDAGDumper.cpp b/lib/CodeGen/SelectionDAG/SelectionDAGDumper.cpp
index 3b5823bfb277..47b03918503c 100644
--- a/lib/CodeGen/SelectionDAG/SelectionDAGDumper.cpp
+++ b/lib/CodeGen/SelectionDAG/SelectionDAGDumper.cpp
@@ -54,7 +54,6 @@ std::string SDNode::getOperationName(const SelectionDAG *G) const {
   case ISD::DELETED_NODE:               return "<<Deleted Node!>>";
 #endif
   case ISD::PREFETCH:                   return "Prefetch";
-  case ISD::MEMBARRIER:                 return "MemBarrier";
   case ISD::ATOMIC_FENCE:               return "AtomicFence";
   case ISD::ATOMIC_CMP_SWAP:            return "AtomicCmpSwap";
   case ISD::ATOMIC_SWAP:                return "AtomicSwap";
diff --git a/lib/CodeGen/SelectionDAG/SelectionDAGISel.cpp b/lib/CodeGen/SelectionDAG/SelectionDAGISel.cpp
index eeea9e4cfcff..e21f26e91ceb 100644
--- a/lib/CodeGen/SelectionDAG/SelectionDAGISel.cpp
+++ b/lib/CodeGen/SelectionDAG/SelectionDAGISel.cpp
@@ -63,12 +63,16 @@ STATISTIC(NumFastIselSuccess, "Number of instructions fast isel selected");
 STATISTIC(NumFastIselBlocks, "Number of blocks selected entirely by fast isel");
 STATISTIC(NumDAGBlocks, "Number of blocks selected using DAG");
 STATISTIC(NumDAGIselRetries,"Number of times dag isel has to try another path");
+STATISTIC(NumEntryBlocks, "Number of entry blocks encountered");
+STATISTIC(NumFastIselFailLowerArguments,
+          "Number of entry blocks where fast isel failed to lower arguments");
 
 #ifndef NDEBUG
 static cl::opt<bool>
 EnableFastISelVerbose2("fast-isel-verbose2", cl::Hidden,
           cl::desc("Enable extra verbose messages in the \"fast\" "
                    "instruction selector"));
+
   // Terminators
 STATISTIC(NumFastIselFailRet,"Fast isel fails on Ret");
 STATISTIC(NumFastIselFailBr,"Fast isel fails on Br");
@@ -742,7 +746,7 @@ public:
 } // end anonymous namespace
 
 void SelectionDAGISel::DoInstructionSelection() {
-  DEBUG(errs() << "===== Instruction selection begins: BB#"
+  DEBUG(dbgs() << "===== Instruction selection begins: BB#"
         << FuncInfo->MBB->getNumber()
         << " '" << FuncInfo->MBB->getName() << "'\n");
 
@@ -801,7 +805,7 @@ void SelectionDAGISel::DoInstructionSelection() {
     CurDAG->setRoot(Dummy.getValue());
   }
 
-  DEBUG(errs() << "===== Instruction selection ends:\n");
+  DEBUG(dbgs() << "===== Instruction selection ends:\n");
 
   PostprocessISelDAG();
 }
@@ -831,84 +835,6 @@ void SelectionDAGISel::PrepareEHLandingPad() {
   if (Reg) MBB->addLiveIn(Reg);
 }
 
-/// TryToFoldFastISelLoad - We're checking to see if we can fold the specified
-/// load into the specified FoldInst.  Note that we could have a sequence where
-/// multiple LLVM IR instructions are folded into the same machineinstr.  For
-/// example we could have:
-///   A: x = load i32 *P
-///   B: y = icmp A, 42
-///   C: br y, ...
-///
-/// In this scenario, LI is "A", and FoldInst is "C".  We know about "B" (and
-/// any other folded instructions) because it is between A and C.
-///
-/// If we succeed in folding the load into the operation, return true.
-///
-bool SelectionDAGISel::TryToFoldFastISelLoad(const LoadInst *LI,
-                                             const Instruction *FoldInst,
-                                             FastISel *FastIS) {
-  // We know that the load has a single use, but don't know what it is.  If it
-  // isn't one of the folded instructions, then we can't succeed here.  Handle
-  // this by scanning the single-use users of the load until we get to FoldInst.
-  unsigned MaxUsers = 6;  // Don't scan down huge single-use chains of instrs.
-
-  const Instruction *TheUser = LI->use_back();
-  while (TheUser != FoldInst &&   // Scan up until we find FoldInst.
-         // Stay in the right block.
-         TheUser->getParent() == FoldInst->getParent() &&
-         --MaxUsers) {  // Don't scan too far.
-    // If there are multiple or no uses of this instruction, then bail out.
-    if (!TheUser->hasOneUse())
-      return false;
-
-    TheUser = TheUser->use_back();
-  }
-
-  // If we didn't find the fold instruction, then we failed to collapse the
-  // sequence.
-  if (TheUser != FoldInst)
-    return false;
-
-  // Don't try to fold volatile loads.  Target has to deal with alignment
-  // constraints.
-  if (LI->isVolatile()) return false;
-
-  // Figure out which vreg this is going into.  If there is no assigned vreg yet
-  // then there actually was no reference to it.  Perhaps the load is referenced
-  // by a dead instruction.
-  unsigned LoadReg = FastIS->getRegForValue(LI);
-  if (LoadReg == 0)
-    return false;
-
-  // Check to see what the uses of this vreg are.  If it has no uses, or more
-  // than one use (at the machine instr level) then we can't fold it.
-  MachineRegisterInfo::reg_iterator RI = RegInfo->reg_begin(LoadReg);
-  if (RI == RegInfo->reg_end())
-    return false;
-
-  // See if there is exactly one use of the vreg.  If there are multiple uses,
-  // then the instruction got lowered to multiple machine instructions or the
-  // use of the loaded value ended up being multiple operands of the result, in
-  // either case, we can't fold this.
-  MachineRegisterInfo::reg_iterator PostRI = RI; ++PostRI;
-  if (PostRI != RegInfo->reg_end())
-    return false;
-
-  assert(RI.getOperand().isUse() &&
-         "The only use of the vreg must be a use, we haven't emitted the def!");
-
-  MachineInstr *User = &*RI;
-
-  // Set the insertion point properly.  Folding the load can cause generation of
-  // other random instructions (like sign extends) for addressing modes, make
-  // sure they get inserted in a logical place before the new instruction.
-  FuncInfo->InsertPt = User;
-  FuncInfo->MBB = User->getParent();
-
-  // Ask the target to try folding the load.
-  return FastIS->TryToFoldLoad(User, RI.getOperandNo(), LI);
-}
-
 /// isFoldedOrDeadInstruction - Return true if the specified instruction is
 /// side-effect free and is either dead or folded into a generated instruction.
 /// Return false if it needs to be emitted.
@@ -1054,9 +980,12 @@ void SelectionDAGISel::SelectAllBasicBlocks(const Function &Fn) {
       // Emit code for any incoming arguments. This must happen before
       // beginning FastISel on the entry block.
       if (LLVMBB == &Fn.getEntryBlock()) {
+        ++NumEntryBlocks;
+
         // Lower any arguments needed in this block if this is the entry block.
         if (!FastIS->LowerArguments()) {
           // Fast isel failed to lower these arguments
+          ++NumFastIselFailLowerArguments;
           if (EnableFastISelAbortArgs)
             llvm_unreachable("FastISel didn't lower all arguments");
 
@@ -1106,7 +1035,7 @@ void SelectionDAGISel::SelectAllBasicBlocks(const Function &Fn) {
           }
           if (BeforeInst != Inst && isa<LoadInst>(BeforeInst) &&
               BeforeInst->hasOneUse() &&
-              TryToFoldFastISelLoad(cast<LoadInst>(BeforeInst), Inst, FastIS)) {
+              FastIS->tryToFoldLoad(cast<LoadInst>(BeforeInst), Inst)) {
             // If we succeeded, don't re-select the load.
             BI = llvm::next(BasicBlock::const_iterator(BeforeInst));
             --NumFastIselRemaining;
@@ -1178,8 +1107,10 @@ void SelectionDAGISel::SelectAllBasicBlocks(const Function &Fn) {
       FastIS->recomputeInsertPt();
     } else {
       // Lower any arguments needed in this block if this is the entry block.
-      if (LLVMBB == &Fn.getEntryBlock())
+      if (LLVMBB == &Fn.getEntryBlock()) {
+        ++NumEntryBlocks;
         LowerArguments(Fn);
+      }
     }
 
     if (Begin != BI)
@@ -1771,7 +1702,7 @@ UpdateChainsAndGlue(SDNode *NodeToMatch, SDValue InputChain,
   if (!NowDeadNodes.empty())
     CurDAG->RemoveDeadNodes(NowDeadNodes);
 
-  DEBUG(errs() << "ISEL: Match complete!\n");
+  DEBUG(dbgs() << "ISEL: Match complete!\n");
 }
 
 enum ChainResult {
@@ -2276,9 +2207,9 @@ SelectCodeCommon(SDNode *NodeToMatch, const unsigned char *MatcherTable,
   SmallVector<SDNode*, 3> ChainNodesMatched;
   SmallVector<SDNode*, 3> GlueResultNodesMatched;
 
-  DEBUG(errs() << "ISEL: Starting pattern match on root node: ";
+  DEBUG(dbgs() << "ISEL: Starting pattern match on root node: ";
         NodeToMatch->dump(CurDAG);
-        errs() << '\n');
+        dbgs() << '\n');
 
   // Determine where to start the interpreter.  Normally we start at opcode #0,
   // but if the state machine starts with an OPC_SwitchOpcode, then we
@@ -2290,7 +2221,7 @@ SelectCodeCommon(SDNode *NodeToMatch, const unsigned char *MatcherTable,
     // Already computed the OpcodeOffset table, just index into it.
     if (N.getOpcode() < OpcodeOffset.size())
       MatcherIndex = OpcodeOffset[N.getOpcode()];
-    DEBUG(errs() << "  Initial Opcode index to " << MatcherIndex << "\n");
+    DEBUG(dbgs() << "  Initial Opcode index to " << MatcherIndex << "\n");
 
   } else if (MatcherTable[0] == OPC_SwitchOpcode) {
     // Otherwise, the table isn't computed, but the state machine does start
@@ -2357,7 +2288,7 @@ SelectCodeCommon(SDNode *NodeToMatch, const unsigned char *MatcherTable,
         if (!Result)
           break;
 
-        DEBUG(errs() << "  Skipped scope entry (due to false predicate) at "
+        DEBUG(dbgs() << "  Skipped scope entry (due to false predicate) at "
                      << "index " << MatcherIndexOfPredicate
                      << ", continuing at " << FailIndex << "\n");
         ++NumDAGIselRetries;
@@ -2487,7 +2418,7 @@ SelectCodeCommon(SDNode *NodeToMatch, const unsigned char *MatcherTable,
       if (CaseSize == 0) break;
 
       // Otherwise, execute the case we found.
-      DEBUG(errs() << "  OpcodeSwitch from " << SwitchStart
+      DEBUG(dbgs() << "  OpcodeSwitch from " << SwitchStart
                    << " to " << MatcherIndex << "\n");
       continue;
     }
@@ -2519,7 +2450,7 @@ SelectCodeCommon(SDNode *NodeToMatch, const unsigned char *MatcherTable,
       if (CaseSize == 0) break;
 
       // Otherwise, execute the case we found.
-      DEBUG(errs() << "  TypeSwitch[" << EVT(CurNodeVT).getEVTString()
+      DEBUG(dbgs() << "  TypeSwitch[" << EVT(CurNodeVT).getEVTString()
                    << "] from " << SwitchStart << " to " << MatcherIndex<<'\n');
       continue;
     }
@@ -2787,7 +2718,7 @@ SelectCodeCommon(SDNode *NodeToMatch, const unsigned char *MatcherTable,
         // If this is a normal EmitNode command, just create the new node and
         // add the results to the RecordedNodes list.
         Res = CurDAG->getMachineNode(TargetOpc, NodeToMatch->getDebugLoc(),
-                                     VTList, Ops.data(), Ops.size());
+                                     VTList, Ops);
 
         // Add all the non-glue/non-chain results to the RecordedNodes list.
         for (unsigned i = 0, e = VTs.size(); i != e; ++i) {
@@ -2863,9 +2794,9 @@ SelectCodeCommon(SDNode *NodeToMatch, const unsigned char *MatcherTable,
           ->setMemRefs(MemRefs, MemRefs + NumMemRefs);
       }
 
-      DEBUG(errs() << "  "
+      DEBUG(dbgs() << "  "
                    << (Opcode == OPC_MorphNodeTo ? "Morphed" : "Created")
-                   << " node: "; Res->dump(CurDAG); errs() << "\n");
+                   << " node: "; Res->dump(CurDAG); dbgs() << "\n");
 
       // If this was a MorphNodeTo then we're completely done!
       if (Opcode == OPC_MorphNodeTo) {
@@ -2940,7 +2871,7 @@ SelectCodeCommon(SDNode *NodeToMatch, const unsigned char *MatcherTable,
     // If the code reached this point, then the match failed.  See if there is
     // another child to try in the current 'Scope', otherwise pop it until we
     // find a case to check.
-    DEBUG(errs() << "  Match failed at index " << CurrentOpcodeIndex << "\n");
+    DEBUG(dbgs() << "  Match failed at index " << CurrentOpcodeIndex << "\n");
     ++NumDAGIselRetries;
     while (1) {
       if (MatchScopes.empty()) {
@@ -2960,7 +2891,7 @@ SelectCodeCommon(SDNode *NodeToMatch, const unsigned char *MatcherTable,
         MatchedMemRefs.resize(LastScope.NumMatchedMemRefs);
       MatcherIndex = LastScope.FailIndex;
 
-      DEBUG(errs() << "  Continuing at " << MatcherIndex << "\n");
+      DEBUG(dbgs() << "  Continuing at " << MatcherIndex << "\n");
 
       InputChain = LastScope.InputChain;
       InputGlue = LastScope.InputGlue;
diff --git a/lib/CodeGen/ShrinkWrapping.cpp b/lib/CodeGen/ShrinkWrapping.cpp
index 9ab491808fe5..2feea59c036a 100644
--- a/lib/CodeGen/ShrinkWrapping.cpp
+++ b/lib/CodeGen/ShrinkWrapping.cpp
@@ -70,14 +70,14 @@ ShrinkWrapFunc("shrink-wrap-func", cl::Hidden,
 
 // Debugging level for shrink wrapping.
 enum ShrinkWrapDebugLevel {
-  None, BasicInfo, Iterations, Details
+  Disabled, BasicInfo, Iterations, Details
 };
 
 static cl::opt<enum ShrinkWrapDebugLevel>
 ShrinkWrapDebugging("shrink-wrap-dbg", cl::Hidden,
   cl::desc("Print shrink wrapping debugging information"),
   cl::values(
-    clEnumVal(None      , "disable debug output"),
+    clEnumVal(Disabled  , "disable debug output"),
     clEnumVal(BasicInfo , "print basic DF sets"),
     clEnumVal(Iterations, "print SR sets for each iteration"),
     clEnumVal(Details   , "print all DF sets"),
diff --git a/lib/CodeGen/TargetLoweringBase.cpp b/lib/CodeGen/TargetLoweringBase.cpp
index f42bdbd27643..8074d167f44c 100644
--- a/lib/CodeGen/TargetLoweringBase.cpp
+++ b/lib/CodeGen/TargetLoweringBase.cpp
@@ -620,12 +620,55 @@ static void InitCmpLibcallCCs(ISD::CondCode *CCs) {
 TargetLoweringBase::TargetLoweringBase(const TargetMachine &tm,
                                        const TargetLoweringObjectFile *tlof)
   : TM(tm), TD(TM.getDataLayout()), TLOF(*tlof) {
+  initActions();
+
+  // Perform these initializations only once.
+  IsLittleEndian = TD->isLittleEndian();
+  PointerTy = MVT::getIntegerVT(8*TD->getPointerSize(0));
+  MaxStoresPerMemset = MaxStoresPerMemcpy = MaxStoresPerMemmove = 8;
+  MaxStoresPerMemsetOptSize = MaxStoresPerMemcpyOptSize
+    = MaxStoresPerMemmoveOptSize = 4;
+  UseUnderscoreSetJmp = false;
+  UseUnderscoreLongJmp = false;
+  SelectIsExpensive = false;
+  IntDivIsCheap = false;
+  Pow2DivIsCheap = false;
+  JumpIsExpensive = false;
+  PredictableSelectIsExpensive = false;
+  StackPointerRegisterToSaveRestore = 0;
+  ExceptionPointerRegister = 0;
+  ExceptionSelectorRegister = 0;
+  BooleanContents = UndefinedBooleanContent;
+  BooleanVectorContents = UndefinedBooleanContent;
+  SchedPreferenceInfo = Sched::ILP;
+  JumpBufSize = 0;
+  JumpBufAlignment = 0;
+  MinFunctionAlignment = 0;
+  PrefFunctionAlignment = 0;
+  PrefLoopAlignment = 0;
+  MinStackArgumentAlignment = 1;
+  InsertFencesForAtomic = false;
+  SupportJumpTables = true;
+  MinimumJumpTableEntries = 4;
+
+  InitLibcallNames(LibcallRoutineNames, TM);
+  InitCmpLibcallCCs(CmpLibcallCCs);
+  InitLibcallCallingConvs(LibcallCallingConvs);
+}
+
+TargetLoweringBase::~TargetLoweringBase() {
+  delete &TLOF;
+}
+
+void TargetLoweringBase::initActions() {
   // All operations default to being supported.
   memset(OpActions, 0, sizeof(OpActions));
   memset(LoadExtActions, 0, sizeof(LoadExtActions));
   memset(TruncStoreActions, 0, sizeof(TruncStoreActions));
   memset(IndexedModeActions, 0, sizeof(IndexedModeActions));
   memset(CondCodeActions, 0, sizeof(CondCodeActions));
+  memset(RegClassForVT, 0,MVT::LAST_VALUETYPE*sizeof(TargetRegisterClass*));
+  memset(TargetDAGCombineArray, 0, array_lengthof(TargetDAGCombineArray));
 
   // Set default actions for various operations.
   for (unsigned VT = 0; VT != (unsigned)MVT::LAST_VALUETYPE; ++VT) {
@@ -702,45 +745,6 @@ TargetLoweringBase::TargetLoweringBase(const TargetMachine &tm,
   // here is to inform DAG Legalizer to replace DEBUGTRAP with TRAP.
   //
   setOperationAction(ISD::DEBUGTRAP, MVT::Other, Expand);
-
-  IsLittleEndian = TD->isLittleEndian();
-  PointerTy = MVT::getIntegerVT(8*TD->getPointerSize(0));
-  memset(RegClassForVT, 0,MVT::LAST_VALUETYPE*sizeof(TargetRegisterClass*));
-  memset(TargetDAGCombineArray, 0, array_lengthof(TargetDAGCombineArray));
-  MaxStoresPerMemset = MaxStoresPerMemcpy = MaxStoresPerMemmove = 8;
-  MaxStoresPerMemsetOptSize = MaxStoresPerMemcpyOptSize
-    = MaxStoresPerMemmoveOptSize = 4;
-  UseUnderscoreSetJmp = false;
-  UseUnderscoreLongJmp = false;
-  SelectIsExpensive = false;
-  IntDivIsCheap = false;
-  Pow2DivIsCheap = false;
-  JumpIsExpensive = false;
-  PredictableSelectIsExpensive = false;
-  StackPointerRegisterToSaveRestore = 0;
-  ExceptionPointerRegister = 0;
-  ExceptionSelectorRegister = 0;
-  BooleanContents = UndefinedBooleanContent;
-  BooleanVectorContents = UndefinedBooleanContent;
-  SchedPreferenceInfo = Sched::ILP;
-  JumpBufSize = 0;
-  JumpBufAlignment = 0;
-  MinFunctionAlignment = 0;
-  PrefFunctionAlignment = 0;
-  PrefLoopAlignment = 0;
-  MinStackArgumentAlignment = 1;
-  ShouldFoldAtomicFences = false;
-  InsertFencesForAtomic = false;
-  SupportJumpTables = true;
-  MinimumJumpTableEntries = 4;
-
-  InitLibcallNames(LibcallRoutineNames, TM);
-  InitCmpLibcallCCs(CmpLibcallCCs);
-  InitLibcallCallingConvs(LibcallCallingConvs);
-}
-
-TargetLoweringBase::~TargetLoweringBase() {
-  delete &TLOF;
 }
 
 MVT TargetLoweringBase::getScalarShiftAmountTy(EVT LHSTy) const {
diff --git a/lib/CodeGen/TargetLoweringObjectFileImpl.cpp b/lib/CodeGen/TargetLoweringObjectFileImpl.cpp
index 3bdca4c64078..7e7359a8feed 100644
--- a/lib/CodeGen/TargetLoweringObjectFileImpl.cpp
+++ b/lib/CodeGen/TargetLoweringObjectFileImpl.cpp
@@ -523,11 +523,6 @@ getExplicitSectionGlobal(const GlobalValue *GV, SectionKind Kind,
 const MCSection *TargetLoweringObjectFileMachO::
 SelectSectionForGlobal(const GlobalValue *GV, SectionKind Kind,
                        Mangler *Mang, const TargetMachine &TM) const {
-
-  // Handle thread local data.
-  if (Kind.isThreadBSS()) return TLSBSSSection;
-  if (Kind.isThreadData()) return TLSDataSection;
-
   if (Kind.isText())
     return GV->isWeakForLinker() ? TextCoalSection : TextSection;
 
@@ -580,6 +575,10 @@ SelectSectionForGlobal(const GlobalValue *GV, SectionKind Kind,
   if (Kind.isBSSLocal())
     return DataBSSSection;
 
+  // Handle thread local data.
+  if (Kind.isThreadBSS()) return TLSBSSSection;
+  if (Kind.isThreadData()) return TLSDataSection;
+
   // Otherwise, just drop the variable in the normal data section.
   return DataSection;
 }
@@ -782,3 +781,49 @@ SelectSectionForGlobal(const GlobalValue *GV, SectionKind Kind,
   return getDataSection();
 }
 
+void TargetLoweringObjectFileCOFF::
+emitModuleFlags(MCStreamer &Streamer,
+                ArrayRef<Module::ModuleFlagEntry> ModuleFlags,
+                Mangler *Mang, const TargetMachine &TM) const {
+  MDNode *LinkerOptions = 0;
+
+  // Look for the "Linker Options" flag, since it's the only one we support.
+  for (ArrayRef<Module::ModuleFlagEntry>::iterator
+       i = ModuleFlags.begin(), e = ModuleFlags.end(); i != e; ++i) {
+    const Module::ModuleFlagEntry &MFE = *i;
+    StringRef Key = MFE.Key->getString();
+    Value *Val = MFE.Val;
+    if (Key == "Linker Options") {
+      LinkerOptions = cast<MDNode>(Val);
+      break;
+    }
+  }
+  if (!LinkerOptions)
+    return;
+
+  // Emit the linker options to the linker .drectve section.  According to the
+  // spec, this section is a space-separated string containing flags for linker.
+  const MCSection *Sec = getDrectveSection();
+  Streamer.SwitchSection(Sec);
+  for (unsigned i = 0, e = LinkerOptions->getNumOperands(); i != e; ++i) {
+    MDNode *MDOptions = cast<MDNode>(LinkerOptions->getOperand(i));
+    for (unsigned ii = 0, ie = MDOptions->getNumOperands(); ii != ie; ++ii) {
+      MDString *MDOption = cast<MDString>(MDOptions->getOperand(ii));
+      StringRef Op = MDOption->getString();
+      // Lead with a space for consistency with our dllexport implementation.
+      std::string Escaped(" ");
+      if (Op.find(" ") != StringRef::npos) {
+        // The PE-COFF spec says args with spaces must be quoted.  It doesn't say
+        // how to escape quotes, but it probably uses this algorithm:
+        // http://msdn.microsoft.com/en-us/library/17w5ykft(v=vs.85).aspx
+        // FIXME: Reuse escaping code from Support/Windows/Program.inc
+        Escaped.push_back('\"');
+        Escaped.append(Op);
+        Escaped.push_back('\"');
+      } else {
+        Escaped.append(Op);
+      }
+      Streamer.EmitBytes(Escaped);
+    }
+  }
+}
diff --git a/lib/CodeGen/TargetOptionsImpl.cpp b/lib/CodeGen/TargetOptionsImpl.cpp
index 0f59d0169e18..435a5e7e0bf9 100644
--- a/lib/CodeGen/TargetOptionsImpl.cpp
+++ b/lib/CodeGen/TargetOptionsImpl.cpp
@@ -50,3 +50,29 @@ StringRef TargetOptions::getTrapFunctionName() const {
   return TrapFuncName;
 }
 
+bool TargetOptions::operator==(const TargetOptions &TO) {
+#define ARE_EQUAL(X) X == TO.X
+  return
+    ARE_EQUAL(UnsafeFPMath) &&
+    ARE_EQUAL(NoInfsFPMath) &&
+    ARE_EQUAL(NoNaNsFPMath) &&
+    ARE_EQUAL(HonorSignDependentRoundingFPMathOption) &&
+    ARE_EQUAL(UseSoftFloat) &&
+    ARE_EQUAL(NoZerosInBSS) &&
+    ARE_EQUAL(JITExceptionHandling) &&
+    ARE_EQUAL(JITEmitDebugInfo) &&
+    ARE_EQUAL(JITEmitDebugInfoToDisk) &&
+    ARE_EQUAL(GuaranteedTailCallOpt) &&
+    ARE_EQUAL(DisableTailCalls) &&
+    ARE_EQUAL(StackAlignmentOverride) &&
+    ARE_EQUAL(RealignStack) &&
+    ARE_EQUAL(SSPBufferSize) &&
+    ARE_EQUAL(EnableFastISel) &&
+    ARE_EQUAL(PositionIndependentExecutable) &&
+    ARE_EQUAL(EnableSegmentedStacks) &&
+    ARE_EQUAL(UseInitArray) &&
+    ARE_EQUAL(TrapFuncName) &&
+    ARE_EQUAL(FloatABIType) &&
+    ARE_EQUAL(AllowFPOpFusion);
+#undef ARE_EQUAL
+}
diff --git a/lib/CodeGen/TargetSchedule.cpp b/lib/CodeGen/TargetSchedule.cpp
index 783bfa1c1a18..1bf14dbcef2c 100644
--- a/lib/CodeGen/TargetSchedule.cpp
+++ b/lib/CodeGen/TargetSchedule.cpp
@@ -128,6 +128,8 @@ resolveSchedClass(const MachineInstr *MI) const {
   // Get the definition's scheduling class descriptor from this machine model.
   unsigned SchedClass = MI->getDesc().getSchedClass();
   const MCSchedClassDesc *SCDesc = SchedModel.getSchedClassDesc(SchedClass);
+  if (!SCDesc->isValid())
+    return SCDesc;
 
 #ifndef NDEBUG
   unsigned NIter = 0;
diff --git a/lib/CodeGen/TwoAddressInstructionPass.cpp b/lib/CodeGen/TwoAddressInstructionPass.cpp
index e6dfe104c82f..7ca2beef6515 100644
--- a/lib/CodeGen/TwoAddressInstructionPass.cpp
+++ b/lib/CodeGen/TwoAddressInstructionPass.cpp
@@ -43,6 +43,7 @@
 #include "llvm/CodeGen/MachineRegisterInfo.h"
 #include "llvm/IR/Function.h"
 #include "llvm/MC/MCInstrItineraries.h"
+#include "llvm/Support/CommandLine.h"
 #include "llvm/Support/Debug.h"
 #include "llvm/Support/ErrorHandling.h"
 #include "llvm/Target/TargetInstrInfo.h"
@@ -58,6 +59,12 @@ STATISTIC(Num3AddrSunk,        "Number of 3-address instructions sunk");
 STATISTIC(NumReSchedUps,       "Number of instructions re-scheduled up");
 STATISTIC(NumReSchedDowns,     "Number of instructions re-scheduled down");
 
+// Temporary flag to disable rescheduling.
+static cl::opt<bool>
+EnableRescheduling("twoaddr-reschedule",
+                   cl::desc("Coalesce copies by rescheduling (default=true)"),
+                   cl::init(true), cl::Hidden);
+
 namespace {
 class TwoAddressInstructionPass : public MachineFunctionPass {
   MachineFunction *MF;
@@ -426,10 +433,7 @@ static bool isKilled(MachineInstr &MI, unsigned Reg,
 /// isTwoAddrUse - Return true if the specified MI uses the specified register
 /// as a two-address use. If so, return the destination register by reference.
 static bool isTwoAddrUse(MachineInstr &MI, unsigned Reg, unsigned &DstReg) {
-  const MCInstrDesc &MCID = MI.getDesc();
-  unsigned NumOps = MI.isInlineAsm()
-    ? MI.getNumOperands() : MCID.getNumOperands();
-  for (unsigned i = 0; i != NumOps; ++i) {
+  for (unsigned i = 0, NumOps = MI.getNumOperands(); i != NumOps; ++i) {
     const MachineOperand &MO = MI.getOperand(i);
     if (!MO.isReg() || !MO.isUse() || MO.getReg() != Reg)
       continue;
@@ -1144,7 +1148,7 @@ tryInstructionTransform(MachineBasicBlock::iterator &mi,
 
   // If there is one more use of regB later in the same MBB, consider
   // re-schedule this MI below it.
-  if (rescheduleMIBelowKill(mi, nmi, regB)) {
+  if (EnableRescheduling && rescheduleMIBelowKill(mi, nmi, regB)) {
     ++NumReSchedDowns;
     return true;
   }
@@ -1163,7 +1167,7 @@ tryInstructionTransform(MachineBasicBlock::iterator &mi,
 
   // If there is one more use of regB later in the same MBB, consider
   // re-schedule it before this MI if it's legal.
-  if (rescheduleKillAboveMI(mi, nmi, regB)) {
+  if (EnableRescheduling && rescheduleKillAboveMI(mi, nmi, regB)) {
     ++NumReSchedUps;
     return true;
   }