Reapply bottom-up fast-isel, with several fixes for x86-32:

 - Check getBytesToPopOnReturn().
 - Eschew ST0 and ST1 for return values.
 - Fix the PIC base register initialization so that it doesn't ever
   fail to end up the top of the entry block.


git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@108039 91177308-0d34-0410-b5e6-96231b3b80d8

6 files changed, 298 insertions, 89 deletions

diff --git a/lib/Target/X86/X86.h b/lib/Target/X86/X86.h
index 22e89a5..677781d 100644
--- a/lib/Target/X86/X86.h
+++ b/lib/Target/X86/X86.h
@@ -35,6 +35,10 @@ class formatted_raw_ostream;
 FunctionPass *createX86ISelDag(X86TargetMachine &TM,
                                CodeGenOpt::Level OptLevel);
 
+/// createGlobalBaseRegPass - This pass initializes a global base
+/// register for PIC on x86-32.
+FunctionPass* createGlobalBaseRegPass();
+
 /// createX86FloatingPointStackifierPass - This function returns a pass which
 /// converts floating point register references and pseudo instructions into
 /// floating point stack references and physical instructions.
diff --git a/lib/Target/X86/X86FastISel.cpp b/lib/Target/X86/X86FastISel.cpp
index a8fdc76..408b403 100644
--- a/lib/Target/X86/X86FastISel.cpp
+++ b/lib/Target/X86/X86FastISel.cpp
@@ -23,6 +23,7 @@
 #include "llvm/GlobalVariable.h"
 #include "llvm/Instructions.h"
 #include "llvm/IntrinsicInst.h"
+#include "llvm/CodeGen/Analysis.h"
 #include "llvm/CodeGen/FastISel.h"
 #include "llvm/CodeGen/FunctionLoweringInfo.h"
 #include "llvm/CodeGen/MachineConstantPool.h"
@@ -84,6 +85,8 @@ private:
   
   bool X86SelectStore(const Instruction *I);
 
+  bool X86SelectRet(const Instruction *I);
+
   bool X86SelectCmp(const Instruction *I);
 
   bool X86SelectZExt(const Instruction *I);
@@ -105,6 +108,7 @@ private:
   bool X86SelectCall(const Instruction *I);
 
   CCAssignFn *CCAssignFnForCall(CallingConv::ID CC, bool isTailCall = false);
+  CCAssignFn *CCAssignFnForRet(CallingConv::ID CC, bool isTailCall = false);
 
   const X86InstrInfo *getInstrInfo() const {
     return getTargetMachine()->getInstrInfo();
@@ -178,6 +182,20 @@ CCAssignFn *X86FastISel::CCAssignFnForCall(CallingConv::ID CC,
     return CC_X86_32_C;
 }
 
+/// CCAssignFnForRet - Selects the correct CCAssignFn for a given calling
+/// convention.
+CCAssignFn *X86FastISel::CCAssignFnForRet(CallingConv::ID CC,
+                                          bool isTaillCall) {
+  if (Subtarget->is64Bit()) {
+    if (Subtarget->isTargetWin64())
+      return RetCC_X86_Win64_C;
+    else
+      return RetCC_X86_64_C;
+  }
+
+  return RetCC_X86_32_C;
+}
+
 /// X86FastEmitLoad - Emit a machine instruction to load a value of type VT.
 /// The address is either pre-computed, i.e. Ptr, or a GlobalAddress, i.e. GV.
 /// Return true and the result register by reference if it is possible.
@@ -230,7 +248,8 @@ bool X86FastISel::X86FastEmitLoad(EVT VT, const X86AddressMode &AM,
   }
 
   ResultReg = createResultReg(RC);
-  addFullAddress(BuildMI(MBB, DL, TII.get(Opc), ResultReg), AM);
+  addFullAddress(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt,
+                         DL, TII.get(Opc), ResultReg), AM);
   return true;
 }
 
@@ -249,7 +268,7 @@ X86FastISel::X86FastEmitStore(EVT VT, unsigned Val,
   case MVT::i1: {
     // Mask out all but lowest bit.
     unsigned AndResult = createResultReg(X86::GR8RegisterClass);
-    BuildMI(MBB, DL,
+    BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
             TII.get(X86::AND8ri), AndResult).addReg(Val).addImm(1);
     Val = AndResult;
   }
@@ -266,7 +285,8 @@ X86FastISel::X86FastEmitStore(EVT VT, unsigned Val,
     break;
   }
   
-  addFullAddress(BuildMI(MBB, DL, TII.get(Opc)), AM).addReg(Val);
+  addFullAddress(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt,
+                         DL, TII.get(Opc)), AM).addReg(Val);
   return true;
 }
 
@@ -294,7 +314,8 @@ bool X86FastISel::X86FastEmitStore(EVT VT, const Value *Val,
     }
     
     if (Opc) {
-      addFullAddress(BuildMI(MBB, DL, TII.get(Opc)), AM)
+      addFullAddress(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt,
+                             DL, TII.get(Opc)), AM)
                              .addImm(Signed ? (uint64_t) CI->getSExtValue() :
                                               CI->getZExtValue());
       return true;
@@ -333,7 +354,7 @@ bool X86FastISel::X86SelectAddress(const Value *V, X86AddressMode &AM) {
     // Don't walk into other basic blocks; it's possible we haven't
     // visited them yet, so the instructions may not yet be assigned
     // virtual registers.
-    if (FuncInfo.MBBMap[I->getParent()] != MBB)
+    if (FuncInfo.MBBMap[I->getParent()] != FuncInfo.MBB)
       return false;
 
     Opcode = I->getOpcode();
@@ -518,6 +539,9 @@ bool X86FastISel::X86SelectAddress(const Value *V, X86AddressMode &AM) {
       StubAM.GV = GV;
       StubAM.GVOpFlags = GVFlags;
 
+      // Prepare for inserting code in the local-value area.
+      MachineBasicBlock::iterator SaveInsertPt = enterLocalValueArea();
+
       if (TLI.getPointerTy() == MVT::i64) {
         Opc = X86::MOV64rm;
         RC  = X86::GR64RegisterClass;
@@ -530,8 +554,13 @@ bool X86FastISel::X86SelectAddress(const Value *V, X86AddressMode &AM) {
       }
       
       LoadReg = createResultReg(RC);
-      addFullAddress(BuildMI(MBB, DL, TII.get(Opc), LoadReg), StubAM);
-      
+      MachineInstrBuilder LoadMI =
+        BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(Opc), LoadReg);
+      addFullAddress(LoadMI, StubAM);
+
+      // Ok, back to normal mode.
+      leaveLocalValueArea(SaveInsertPt);
+
       // Prevent loading GV stub multiple times in same MBB.
       LocalValueMap[V] = LoadReg;
     }
@@ -656,6 +685,95 @@ bool X86FastISel::X86SelectStore(const Instruction *I) {
   return X86FastEmitStore(VT, I->getOperand(0), AM);
 }
 
+/// X86SelectRet - Select and emit code to implement ret instructions.
+bool X86FastISel::X86SelectRet(const Instruction *I) {
+  const ReturnInst *Ret = cast<ReturnInst>(I);
+  const Function &F = *I->getParent()->getParent();
+
+  if (!FuncInfo.CanLowerReturn)
+    return false;
+
+  CallingConv::ID CC = F.getCallingConv();
+  if (CC != CallingConv::C &&
+      CC != CallingConv::Fast &&
+      CC != CallingConv::X86_FastCall)
+    return false;
+
+  if (Subtarget->isTargetWin64())
+    return false;
+
+  // Don't handle popping bytes on return for now.
+  if (FuncInfo.MF->getInfo<X86MachineFunctionInfo>()
+        ->getBytesToPopOnReturn() != 0)
+    return 0;
+
+  // fastcc with -tailcallopt is intended to provide a guaranteed
+  // tail call optimization. Fastisel doesn't know how to do that.
+  if (CC == CallingConv::Fast && GuaranteedTailCallOpt)
+    return false;
+
+  // Let SDISel handle vararg functions.
+  if (F.isVarArg())
+    return false;
+
+  if (Ret->getNumOperands() > 0) {
+    SmallVector<ISD::OutputArg, 4> Outs;
+    GetReturnInfo(F.getReturnType(), F.getAttributes().getRetAttributes(),
+                  Outs, TLI);
+
+    // Analyze operands of the call, assigning locations to each operand.
+    SmallVector<CCValAssign, 16> ValLocs;
+    CCState CCInfo(CC, F.isVarArg(), TM, ValLocs, I->getContext());
+    CCInfo.AnalyzeReturn(Outs, CCAssignFnForRet(CC));
+
+    const Value *RV = Ret->getOperand(0);
+    unsigned Reg = getRegForValue(RV);
+    if (Reg == 0)
+      return false;
+
+    // Only handle a single return value for now.
+    if (ValLocs.size() != 1)
+      return false;
+
+    CCValAssign &VA = ValLocs[0];
+  
+    // Don't bother handling odd stuff for now.
+    if (VA.getLocInfo() != CCValAssign::Full)
+      return false;
+    // Only handle register returns for now.
+    if (!VA.isRegLoc())
+      return false;
+    // TODO: For now, don't try to handle cases where getLocInfo()
+    // says Full but the types don't match.
+    if (VA.getValVT() != TLI.getValueType(RV->getType()))
+      return false;
+
+    // The calling-convention tables for x87 returns don't tell
+    // the whole story.
+    if (VA.getLocReg() == X86::ST0 || VA.getLocReg() == X86::ST1)
+      return false;
+
+    // Make the copy.
+    unsigned SrcReg = Reg + VA.getValNo();
+    unsigned DstReg = VA.getLocReg();
+    const TargetRegisterClass* SrcRC = MRI.getRegClass(SrcReg);
+    const TargetRegisterClass* DstRC = TRI.getMinimalPhysRegClass(DstReg);
+    bool Emitted = TII.copyRegToReg(*FuncInfo.MBB, FuncInfo.InsertPt,
+                                    DstReg, SrcReg, DstRC, SrcRC, DL);
+
+    // If the target couldn't make the copy for some reason, bail.
+    if (!Emitted)
+      return false;
+
+    // Mark the register as live out of the function.
+    MRI.addLiveOut(VA.getLocReg());
+  }
+
+  // Now emit the RET.
+  BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(X86::RET));
+  return true;
+}
+
 /// X86SelectLoad - Select and emit code to implement load instructions.
 ///
 bool X86FastISel::X86SelectLoad(const Instruction *I)  {
@@ -720,8 +838,9 @@ bool X86FastISel::X86FastEmitCompare(const Value *Op0, const Value *Op1,
   // CMPri, otherwise use CMPrr.
   if (const ConstantInt *Op1C = dyn_cast<ConstantInt>(Op1)) {
     if (unsigned CompareImmOpc = X86ChooseCmpImmediateOpcode(VT, Op1C)) {
-      BuildMI(MBB, DL, TII.get(CompareImmOpc)).addReg(Op0Reg)
-                                          .addImm(Op1C->getSExtValue());
+      BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(CompareImmOpc))
+        .addReg(Op0Reg)
+        .addImm(Op1C->getSExtValue());
       return true;
     }
   }
@@ -731,7 +850,9 @@ bool X86FastISel::X86FastEmitCompare(const Value *Op0, const Value *Op1,
     
   unsigned Op1Reg = getRegForValue(Op1);
   if (Op1Reg == 0) return false;
-  BuildMI(MBB, DL, TII.get(CompareOpc)).addReg(Op0Reg).addReg(Op1Reg);
+  BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(CompareOpc))
+    .addReg(Op0Reg)
+    .addReg(Op1Reg);
   
   return true;
 }
@@ -753,9 +874,10 @@ bool X86FastISel::X86SelectCmp(const Instruction *I) {
     
     unsigned EReg = createResultReg(&X86::GR8RegClass);
     unsigned NPReg = createResultReg(&X86::GR8RegClass);
-    BuildMI(MBB, DL, TII.get(X86::SETEr), EReg);
-    BuildMI(MBB, DL, TII.get(X86::SETNPr), NPReg);
-    BuildMI(MBB, DL, 
+    BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(X86::SETEr), EReg);
+    BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
+            TII.get(X86::SETNPr), NPReg);
+    BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, 
             TII.get(X86::AND8rr), ResultReg).addReg(NPReg).addReg(EReg);
     UpdateValueMap(I, ResultReg);
     return true;
@@ -766,9 +888,13 @@ bool X86FastISel::X86SelectCmp(const Instruction *I) {
 
     unsigned NEReg = createResultReg(&X86::GR8RegClass);
     unsigned PReg = createResultReg(&X86::GR8RegClass);
-    BuildMI(MBB, DL, TII.get(X86::SETNEr), NEReg);
-    BuildMI(MBB, DL, TII.get(X86::SETPr), PReg);
-    BuildMI(MBB, DL, TII.get(X86::OR8rr), ResultReg).addReg(PReg).addReg(NEReg);
+    BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
+            TII.get(X86::SETNEr), NEReg);
+    BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
+            TII.get(X86::SETPr), PReg);
+    BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
+            TII.get(X86::OR8rr), ResultReg)
+      .addReg(PReg).addReg(NEReg);
     UpdateValueMap(I, ResultReg);
     return true;
   }
@@ -807,7 +933,7 @@ bool X86FastISel::X86SelectCmp(const Instruction *I) {
   if (!X86FastEmitCompare(Op0, Op1, VT))
     return false;
   
-  BuildMI(MBB, DL, TII.get(SetCCOpc), ResultReg);
+  BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(SetCCOpc), ResultReg);
   UpdateValueMap(I, ResultReg);
   return true;
 }
@@ -843,7 +969,7 @@ bool X86FastISel::X86SelectBranch(const Instruction *I) {
 
       // Try to take advantage of fallthrough opportunities.
       CmpInst::Predicate Predicate = CI->getPredicate();
-      if (MBB->isLayoutSuccessor(TrueMBB)) {
+      if (FuncInfo.MBB->isLayoutSuccessor(TrueMBB)) {
         std::swap(TrueMBB, FalseMBB);
         Predicate = CmpInst::getInversePredicate(Predicate);
       }
@@ -892,16 +1018,18 @@ bool X86FastISel::X86SelectBranch(const Instruction *I) {
       if (!X86FastEmitCompare(Op0, Op1, VT))
         return false;
       
-      BuildMI(MBB, DL, TII.get(BranchOpc)).addMBB(TrueMBB);
+      BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(BranchOpc))
+        .addMBB(TrueMBB);
 
       if (Predicate == CmpInst::FCMP_UNE) {
         // X86 requires a second branch to handle UNE (and OEQ,
         // which is mapped to UNE above).
-        BuildMI(MBB, DL, TII.get(X86::JP_4)).addMBB(TrueMBB);
+        BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(X86::JP_4))
+          .addMBB(TrueMBB);
       }
 
       FastEmitBranch(FalseMBB, DL);
-      MBB->addSuccessor(TrueMBB);
+      FuncInfo.MBB->addSuccessor(TrueMBB);
       return true;
     }
   } else if (ExtractValueInst *EI =
@@ -927,7 +1055,8 @@ bool X86FastISel::X86SelectBranch(const Instruction *I) {
         unsigned Reg = getRegForValue(EI);
 
         for (MachineBasicBlock::const_reverse_iterator
-               RI = MBB->rbegin(), RE = MBB->rend(); RI != RE; ++RI) {
+               RI = FuncInfo.MBB->rbegin(), RE = FuncInfo.MBB->rend();
+             RI != RE; ++RI) {
           const MachineInstr &MI = *RI;
 
           if (MI.definesRegister(Reg)) {
@@ -952,11 +1081,11 @@ bool X86FastISel::X86SelectBranch(const Instruction *I) {
           unsigned OpCode = SetMI->getOpcode();
 
           if (OpCode == X86::SETOr || OpCode == X86::SETBr) {
-            BuildMI(MBB, DL, TII.get(OpCode == X86::SETOr ?
-                                        X86::JO_4 : X86::JB_4))
+            BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
+                    TII.get(OpCode == X86::SETOr ?  X86::JO_4 : X86::JB_4))
               .addMBB(TrueMBB);
             FastEmitBranch(FalseMBB, DL);
-            MBB->addSuccessor(TrueMBB);
+            FuncInfo.MBB->addSuccessor(TrueMBB);
             return true;
           }
         }
@@ -968,10 +1097,12 @@ bool X86FastISel::X86SelectBranch(const Instruction *I) {
   unsigned OpReg = getRegForValue(BI->getCondition());
   if (OpReg == 0) return false;
 
-  BuildMI(MBB, DL, TII.get(X86::TEST8rr)).addReg(OpReg).addReg(OpReg);
-  BuildMI(MBB, DL, TII.get(X86::JNE_4)).addMBB(TrueMBB);
+  BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(X86::TEST8rr))
+    .addReg(OpReg).addReg(OpReg);
+  BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(X86::JNE_4))
+    .addMBB(TrueMBB);
   FastEmitBranch(FalseMBB, DL);
-  MBB->addSuccessor(TrueMBB);
+  FuncInfo.MBB->addSuccessor(TrueMBB);
   return true;
 }
 
@@ -1028,7 +1159,7 @@ bool X86FastISel::X86SelectShift(const Instruction *I) {
   // Fold immediate in shl(x,3).
   if (const ConstantInt *CI = dyn_cast<ConstantInt>(I->getOperand(1))) {
     unsigned ResultReg = createResultReg(RC);
-    BuildMI(MBB, DL, TII.get(OpImm), 
+    BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(OpImm), 
             ResultReg).addReg(Op0Reg).addImm(CI->getZExtValue() & 0xff);
     UpdateValueMap(I, ResultReg);
     return true;
@@ -1036,16 +1167,19 @@ bool X86FastISel::X86SelectShift(const Instruction *I) {
   
   unsigned Op1Reg = getRegForValue(I->getOperand(1));
   if (Op1Reg == 0) return false;
-  TII.copyRegToReg(*MBB, MBB->end(), CReg, Op1Reg, RC, RC, DL);
+  TII.copyRegToReg(*FuncInfo.MBB, FuncInfo.InsertPt,
+                   CReg, Op1Reg, RC, RC, DL);
 
   // The shift instruction uses X86::CL. If we defined a super-register
   // of X86::CL, emit a subreg KILL to precisely describe what we're doing here.
   if (CReg != X86::CL)
-    BuildMI(MBB, DL, TII.get(TargetOpcode::KILL), X86::CL)
+    BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
+            TII.get(TargetOpcode::KILL), X86::CL)
       .addReg(CReg, RegState::Kill);
 
   unsigned ResultReg = createResultReg(RC);
-  BuildMI(MBB, DL, TII.get(OpReg), ResultReg).addReg(Op0Reg);
+  BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(OpReg), ResultReg)
+    .addReg(Op0Reg);
   UpdateValueMap(I, ResultReg);
   return true;
 }
@@ -1077,9 +1211,11 @@ bool X86FastISel::X86SelectSelect(const Instruction *I) {
   unsigned Op2Reg = getRegForValue(I->getOperand(2));
   if (Op2Reg == 0) return false;
 
-  BuildMI(MBB, DL, TII.get(X86::TEST8rr)).addReg(Op0Reg).addReg(Op0Reg);
+  BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(X86::TEST8rr))
+    .addReg(Op0Reg).addReg(Op0Reg);
   unsigned ResultReg = createResultReg(RC);
-  BuildMI(MBB, DL, TII.get(Opc), ResultReg).addReg(Op1Reg).addReg(Op2Reg);
+  BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(Opc), ResultReg)
+    .addReg(Op1Reg).addReg(Op2Reg);
   UpdateValueMap(I, ResultReg);
   return true;
 }
@@ -1093,7 +1229,9 @@ bool X86FastISel::X86SelectFPExt(const Instruction *I) {
       unsigned OpReg = getRegForValue(V);
       if (OpReg == 0) return false;
       unsigned ResultReg = createResultReg(X86::FR64RegisterClass);
-      BuildMI(MBB, DL, TII.get(X86::CVTSS2SDrr), ResultReg).addReg(OpReg);
+      BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
+              TII.get(X86::CVTSS2SDrr), ResultReg)
+        .addReg(OpReg);
       UpdateValueMap(I, ResultReg);
       return true;
     }
@@ -1110,7 +1248,9 @@ bool X86FastISel::X86SelectFPTrunc(const Instruction *I) {
         unsigned OpReg = getRegForValue(V);
         if (OpReg == 0) return false;
         unsigned ResultReg = createResultReg(X86::FR32RegisterClass);
-        BuildMI(MBB, DL, TII.get(X86::CVTSD2SSrr), ResultReg).addReg(OpReg);
+        BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
+                TII.get(X86::CVTSD2SSrr), ResultReg)
+          .addReg(OpReg);
         UpdateValueMap(I, ResultReg);
         return true;
       }
@@ -1145,7 +1285,8 @@ bool X86FastISel::X86SelectTrunc(const Instruction *I) {
   const TargetRegisterClass *CopyRC = (SrcVT == MVT::i16)
     ? X86::GR16_ABCDRegisterClass : X86::GR32_ABCDRegisterClass;
   unsigned CopyReg = createResultReg(CopyRC);
-  BuildMI(MBB, DL, TII.get(CopyOpc), CopyReg).addReg(InputReg);
+  BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(CopyOpc), CopyReg)
+    .addReg(InputReg);
 
   // Then issue an extract_subreg.
   unsigned ResultReg = FastEmitInst_extractsubreg(MVT::i8,
@@ -1166,14 +1307,18 @@ bool X86FastISel::X86SelectExtractValue(const Instruction *I) {
     switch (CI->getIntrinsicID()) {
     default: break;
     case Intrinsic::sadd_with_overflow:
-    case Intrinsic::uadd_with_overflow:
+    case Intrinsic::uadd_with_overflow: {
       // Cheat a little. We know that the registers for "add" and "seto" are
       // allocated sequentially. However, we only keep track of the register
       // for "add" in the value map. Use extractvalue's index to get the
       // correct register for "seto".
-      UpdateValueMap(I, lookUpRegForValue(Agg) + *EI->idx_begin());
+      unsigned OpReg = getRegForValue(Agg);
+      if (OpReg == 0)
+        return false;
+      UpdateValueMap(I, OpReg + *EI->idx_begin());
       return true;
     }
+    }
   }
 
   return false;
@@ -1217,7 +1362,7 @@ bool X86FastISel::X86VisitIntrinsicCall(const IntrinsicInst &I) {
       return false;
     
     unsigned ResultReg = createResultReg(TLI.getRegClassFor(VT));
-    BuildMI(MBB, DL, TII.get(OpC), ResultReg).
+    BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(OpC), ResultReg).
                                   addImm(CI->isZero() ? -1ULL : 0);
     UpdateValueMap(&I, ResultReg);
     return true;
@@ -1231,12 +1376,12 @@ bool X86FastISel::X86VisitIntrinsicCall(const IntrinsicInst &I) {
     const TargetInstrDesc &II = TII.get(TargetOpcode::DBG_VALUE);
     // FIXME may need to add RegState::Debug to any registers produced,
     // although ESP/EBP should be the only ones at the moment.
-    addFullAddress(BuildMI(MBB, DL, II), AM).addImm(0).
-                                        addMetadata(DI->getVariable());
+    addFullAddress(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, II), AM).
+      addImm(0).addMetadata(DI->getVariable());
     return true;
   }
   case Intrinsic::trap: {
-    BuildMI(MBB, DL, TII.get(X86::TRAP));
+    BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(X86::TRAP));
     return true;
   }
   case Intrinsic::sadd_with_overflow:
@@ -1272,7 +1417,8 @@ bool X86FastISel::X86VisitIntrinsicCall(const IntrinsicInst &I) {
       return false;
 
     unsigned ResultReg = createResultReg(TLI.getRegClassFor(VT));
-    BuildMI(MBB, DL, TII.get(OpC), ResultReg).addReg(Reg1).addReg(Reg2);
+    BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(OpC), ResultReg)
+      .addReg(Reg1).addReg(Reg2);
     unsigned DestReg1 = UpdateValueMap(&I, ResultReg);
 
     // If the add with overflow is an intra-block value then we just want to
@@ -1290,7 +1436,7 @@ bool X86FastISel::X86VisitIntrinsicCall(const IntrinsicInst &I) {
     unsigned Opc = X86::SETBr;
     if (I.getIntrinsicID() == Intrinsic::sadd_with_overflow)
       Opc = X86::SETOr;
-    BuildMI(MBB, DL, TII.get(Opc), ResultReg);
+    BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(Opc), ResultReg);
     return true;
   }
   }
@@ -1417,7 +1563,8 @@ bool X86FastISel::X86SelectCall(const Instruction *I) {
 
   // Issue CALLSEQ_START
   unsigned AdjStackDown = TM.getRegisterInfo()->getCallFrameSetupOpcode();
-  BuildMI(MBB, DL, TII.get(AdjStackDown)).addImm(NumBytes);
+  BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(AdjStackDown))
+    .addImm(NumBytes);
 
   // Process argument: walk the register/memloc assignments, inserting
   // copies / loads.
@@ -1473,8 +1620,8 @@ bool X86FastISel::X86SelectCall(const Instruction *I) {
     
     if (VA.isRegLoc()) {
       TargetRegisterClass* RC = TLI.getRegClassFor(ArgVT);
-      bool Emitted = TII.copyRegToReg(*MBB, MBB->end(), VA.getLocReg(),
-                                      Arg, RC, RC, DL);
+      bool Emitted = TII.copyRegToReg(*FuncInfo.MBB, FuncInfo.InsertPt,
+                                      VA.getLocReg(), Arg, RC, RC, DL);
       assert(Emitted && "Failed to emit a copy instruction!"); Emitted=Emitted;
       Emitted = true;
       RegArgs.push_back(VA.getLocReg());
@@ -1500,8 +1647,8 @@ bool X86FastISel::X86SelectCall(const Instruction *I) {
   if (Subtarget->isPICStyleGOT()) {
     TargetRegisterClass *RC = X86::GR32RegisterClass;
     unsigned Base = getInstrInfo()->getGlobalBaseReg(FuncInfo.MF);
-    bool Emitted = TII.copyRegToReg(*MBB, MBB->end(), X86::EBX, Base, RC, RC,
-                                    DL);
+    bool Emitted = TII.copyRegToReg(*FuncInfo.MBB, FuncInfo.InsertPt,
+                                    X86::EBX, Base, RC, RC, DL);
     assert(Emitted && "Failed to emit a copy instruction!"); Emitted=Emitted;
     Emitted = true;
   }
@@ -1511,7 +1658,8 @@ bool X86FastISel::X86SelectCall(const Instruction *I) {
   if (CalleeOp) {
     // Register-indirect call.
     unsigned CallOpc = Subtarget->is64Bit() ? X86::CALL64r : X86::CALL32r;
-    MIB = BuildMI(MBB, DL, TII.get(CallOpc)).addReg(CalleeOp);
+    MIB = BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(CallOpc))
+      .addReg(CalleeOp);
     
   } else {
     // Direct call.
@@ -1540,7 +1688,8 @@ bool X86FastISel::X86SelectCall(const Instruction *I) {
     }
     
     
-    MIB = BuildMI(MBB, DL, TII.get(CallOpc)).addGlobalAddress(GV, 0, OpFlags);
+    MIB = BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(CallOpc))
+      .addGlobalAddress(GV, 0, OpFlags);
   }
 
   // Add an implicit use GOT pointer in EBX.
@@ -1553,7 +1702,8 @@ bool X86FastISel::X86SelectCall(const Instruction *I) {
 
   // Issue CALLSEQ_END
   unsigned AdjStackUp = TM.getRegisterInfo()->getCallFrameDestroyOpcode();
-  BuildMI(MBB, DL, TII.get(AdjStackUp)).addImm(NumBytes).addImm(0);
+  BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(AdjStackUp))
+    .addImm(NumBytes).addImm(0);
 
   // Now handle call return value (if any).
   SmallVector<unsigned, 4> UsedRegs;
@@ -1580,7 +1730,7 @@ bool X86FastISel::X86SelectCall(const Instruction *I) {
     }
 
     unsigned ResultReg = createResultReg(DstRC);
-    bool Emitted = TII.copyRegToReg(*MBB, MBB->end(), ResultReg,
+    bool Emitted = TII.copyRegToReg(*FuncInfo.MBB, FuncInfo.InsertPt, ResultReg,
                                     RVLocs[0].getLocReg(), DstRC, SrcRC, DL);
     assert(Emitted && "Failed to emit a copy instruction!"); Emitted=Emitted;
     Emitted = true;
@@ -1594,18 +1744,21 @@ bool X86FastISel::X86SelectCall(const Instruction *I) {
       unsigned Opc = ResVT == MVT::f32 ? X86::ST_Fp80m32 : X86::ST_Fp80m64;
       unsigned MemSize = ResVT.getSizeInBits()/8;
       int FI = MFI.CreateStackObject(MemSize, MemSize, false);
-      addFrameReference(BuildMI(MBB, DL, TII.get(Opc)), FI).addReg(ResultReg);
+      addFrameReference(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
+                                TII.get(Opc)), FI)
+        .addReg(ResultReg);
       DstRC = ResVT == MVT::f32
         ? X86::FR32RegisterClass : X86::FR64RegisterClass;
       Opc = ResVT == MVT::f32 ? X86::MOVSSrm : X86::MOVSDrm;
       ResultReg = createResultReg(DstRC);
-      addFrameReference(BuildMI(MBB, DL, TII.get(Opc), ResultReg), FI);
+      addFrameReference(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
+                                TII.get(Opc), ResultReg), FI);
     }
 
     if (AndToI1) {
       // Mask out all but lowest bit for some call which produces an i1.
       unsigned AndResult = createResultReg(X86::GR8RegisterClass);
-      BuildMI(MBB, DL, 
+      BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, 
               TII.get(X86::AND8ri), AndResult).addReg(ResultReg).addImm(1);
       ResultReg = AndResult;
     }
@@ -1628,6 +1781,8 @@ X86FastISel::TargetSelectInstruction(const Instruction *I)  {
     return X86SelectLoad(I);
   case Instruction::Store:
     return X86SelectStore(I);
+  case Instruction::Ret:
+    return X86SelectRet(I);
   case Instruction::ICmp:
   case Instruction::FCmp:
     return X86SelectCmp(I);
@@ -1728,7 +1883,8 @@ unsigned X86FastISel::TargetMaterializeConstant(const Constant *C) {
       else
         Opc = X86::LEA64r;
       unsigned ResultReg = createResultReg(RC);
-      addFullAddress(BuildMI(MBB, DL, TII.get(Opc), ResultReg), AM);
+      addFullAddress(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
+                             TII.get(Opc), ResultReg), AM);
       return ResultReg;
     }
     return 0;
@@ -1758,7 +1914,8 @@ unsigned X86FastISel::TargetMaterializeConstant(const Constant *C) {
   // Create the load from the constant pool.
   unsigned MCPOffset = MCP.getConstantPoolIndex(C, Align);
   unsigned ResultReg = createResultReg(RC);
-  addConstantPoolReference(BuildMI(MBB, DL, TII.get(Opc), ResultReg),
+  addConstantPoolReference(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
+                                   TII.get(Opc), ResultReg),
                            MCPOffset, PICBase, OpFlag);
 
   return ResultReg;
@@ -1781,7 +1938,8 @@ unsigned X86FastISel::TargetMaterializeAlloca(const AllocaInst *C) {
   unsigned Opc = Subtarget->is64Bit() ? X86::LEA64r : X86::LEA32r;
   TargetRegisterClass* RC = TLI.getRegClassFor(TLI.getPointerTy());
   unsigned ResultReg = createResultReg(RC);
-  addFullAddress(BuildMI(MBB, DL, TII.get(Opc), ResultReg), AM);
+  addFullAddress(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
+                         TII.get(Opc), ResultReg), AM);
   return ResultReg;
 }
 
diff --git a/lib/Target/X86/X86ISelLowering.cpp b/lib/Target/X86/X86ISelLowering.cpp
index efdf076..1a63474 100644
--- a/lib/Target/X86/X86ISelLowering.cpp
+++ b/lib/Target/X86/X86ISelLowering.cpp
@@ -1218,13 +1218,12 @@ bool X86TargetLowering::getStackCookieLocation(unsigned &AddressSpace,
 
 bool 
 X86TargetLowering::CanLowerReturn(CallingConv::ID CallConv, bool isVarArg,
-                        const SmallVectorImpl<EVT> &OutTys,
-                        const SmallVectorImpl<ISD::ArgFlagsTy> &ArgsFlags,
+                        const SmallVectorImpl<ISD::OutputArg> &Outs,
                         LLVMContext &Context) const {
   SmallVector<CCValAssign, 16> RVLocs;
   CCState CCInfo(CallConv, isVarArg, getTargetMachine(),
                  RVLocs, Context);
-  return CCInfo.CheckReturn(OutTys, ArgsFlags, RetCC_X86);
+  return CCInfo.CheckReturn(Outs, RetCC_X86);
 }
 
 SDValue
diff --git a/lib/Target/X86/X86ISelLowering.h b/lib/Target/X86/X86ISelLowering.h
index 39bbdac..2d28e5c 100644
--- a/lib/Target/X86/X86ISelLowering.h
+++ b/lib/Target/X86/X86ISelLowering.h
@@ -740,8 +740,7 @@ namespace llvm {
 
     virtual bool
       CanLowerReturn(CallingConv::ID CallConv, bool isVarArg,
-                     const SmallVectorImpl<EVT> &OutTys,
-                     const SmallVectorImpl<ISD::ArgFlagsTy> &ArgsFlags,
+                     const SmallVectorImpl<ISD::OutputArg> &Outs,
                      LLVMContext &Context) const;
 
     void ReplaceATOMIC_BINARY_64(SDNode *N, SmallVectorImpl<SDValue> &Results,
diff --git a/lib/Target/X86/X86InstrInfo.cpp b/lib/Target/X86/X86InstrInfo.cpp
index a9f826c..f277774 100644
--- a/lib/Target/X86/X86InstrInfo.cpp
+++ b/lib/Target/X86/X86InstrInfo.cpp
@@ -2642,7 +2642,7 @@ MachineInstr* X86InstrInfo::foldMemoryOperandImpl(MachineFunction &MF,
       if (TM.getSubtarget<X86Subtarget>().is64Bit())
         PICBase = X86::RIP;
       else
-        // FIXME: PICBase = TM.getInstrInfo()->getGlobalBaseReg(&MF);
+        // FIXME: PICBase = getGlobalBaseReg(&MF);
         // This doesn't work for several reasons.
         // 1. GlobalBaseReg may have been spilled.
         // 2. It may not be live at MI.
@@ -3717,6 +3717,8 @@ unsigned X86InstrInfo::GetInstSizeInBytes(const MachineInstr *MI) const {
 /// the global base register value. Output instructions required to
 /// initialize the register in the function entry block, if necessary.
 ///
+/// TODO: Eliminate this and move the code to X86MachineFunctionInfo.
+///
 unsigned X86InstrInfo::getGlobalBaseReg(MachineFunction *MF) const {
   assert(!TM.getSubtarget<X86Subtarget>().is64Bit() &&
          "X86-64 PIC uses RIP relative addressing");
@@ -3726,30 +3728,10 @@ unsigned X86InstrInfo::getGlobalBaseReg(MachineFunction *MF) const {
   if (GlobalBaseReg != 0)
     return GlobalBaseReg;
 
-  // Insert the set of GlobalBaseReg into the first MBB of the function
-  MachineBasicBlock &FirstMBB = MF->front();
-  MachineBasicBlock::iterator MBBI = FirstMBB.begin();
-  DebugLoc DL = FirstMBB.findDebugLoc(MBBI);
+  // Create the register. The code to initialize it is inserted
+  // later, by the CGBR pass (below).
   MachineRegisterInfo &RegInfo = MF->getRegInfo();
-  unsigned PC = RegInfo.createVirtualRegister(X86::GR32RegisterClass);
-  
-  const TargetInstrInfo *TII = TM.getInstrInfo();
-  // Operand of MovePCtoStack is completely ignored by asm printer. It's
-  // only used in JIT code emission as displacement to pc.
-  BuildMI(FirstMBB, MBBI, DL, TII->get(X86::MOVPC32r), PC).addImm(0);
-  
-  // If we're using vanilla 'GOT' PIC style, we should use relative addressing
-  // not to pc, but to _GLOBAL_OFFSET_TABLE_ external.
-  if (TM.getSubtarget<X86Subtarget>().isPICStyleGOT()) {
-    GlobalBaseReg = RegInfo.createVirtualRegister(X86::GR32RegisterClass);
-    // Generate addl $__GLOBAL_OFFSET_TABLE_ + [.-piclabel], %some_register
-    BuildMI(FirstMBB, MBBI, DL, TII->get(X86::ADD32ri), GlobalBaseReg)
-      .addReg(PC).addExternalSymbol("_GLOBAL_OFFSET_TABLE_",
-                                    X86II::MO_GOT_ABSOLUTE_ADDRESS);
-  } else {
-    GlobalBaseReg = PC;
-  }
-
+  GlobalBaseReg = RegInfo.createVirtualRegister(X86::GR32RegisterClass);
   X86FI->setGlobalBaseReg(GlobalBaseReg);
   return GlobalBaseReg;
 }
@@ -3806,3 +3788,66 @@ void X86InstrInfo::SetSSEDomain(MachineInstr *MI, unsigned Domain) const {
 void X86InstrInfo::getNoopForMachoTarget(MCInst &NopInst) const {
   NopInst.setOpcode(X86::NOOP);
 }
+
+namespace {
+  /// CGBR - Create Global Base Reg pass. This initializes the PIC
+  /// global base register for x86-32.
+  struct CGBR : public MachineFunctionPass {
+    static char ID;
+    CGBR() : MachineFunctionPass(&ID) {}
+
+    virtual bool runOnMachineFunction(MachineFunction &MF) {
+      const X86TargetMachine *TM =
+        static_cast<const X86TargetMachine *>(&MF.getTarget());
+
+      assert(!TM->getSubtarget<X86Subtarget>().is64Bit() &&
+             "X86-64 PIC uses RIP relative addressing");
+
+      // Only emit a global base reg in PIC mode.
+      if (TM->getRelocationModel() != Reloc::PIC_)
+        return false;
+
+      // Insert the set of GlobalBaseReg into the first MBB of the function
+      MachineBasicBlock &FirstMBB = MF.front();
+      MachineBasicBlock::iterator MBBI = FirstMBB.begin();
+      DebugLoc DL = FirstMBB.findDebugLoc(MBBI);
+      MachineRegisterInfo &RegInfo = MF.getRegInfo();
+      const X86InstrInfo *TII = TM->getInstrInfo();
+
+      unsigned PC;
+      if (TM->getSubtarget<X86Subtarget>().isPICStyleGOT())
+        PC = RegInfo.createVirtualRegister(X86::GR32RegisterClass);
+      else
+        PC = TII->getGlobalBaseReg(&MF);
+  
+      // Operand of MovePCtoStack is completely ignored by asm printer. It's
+      // only used in JIT code emission as displacement to pc.
+      BuildMI(FirstMBB, MBBI, DL, TII->get(X86::MOVPC32r), PC).addImm(0);
+  
+      // If we're using vanilla 'GOT' PIC style, we should use relative addressing
+      // not to pc, but to _GLOBAL_OFFSET_TABLE_ external.
+      if (TM->getSubtarget<X86Subtarget>().isPICStyleGOT()) {
+        unsigned GlobalBaseReg = TII->getGlobalBaseReg(&MF);
+        // Generate addl $__GLOBAL_OFFSET_TABLE_ + [.-piclabel], %some_register
+        BuildMI(FirstMBB, MBBI, DL, TII->get(X86::ADD32ri), GlobalBaseReg)
+          .addReg(PC).addExternalSymbol("_GLOBAL_OFFSET_TABLE_",
+                                        X86II::MO_GOT_ABSOLUTE_ADDRESS);
+      }
+
+      return true;
+    }
+
+    virtual const char *getPassName() const {
+      return "X86 PIC Global Base Reg Initialization";
+    }
+
+    virtual void getAnalysisUsage(AnalysisUsage &AU) const {
+      AU.setPreservesCFG();
+      MachineFunctionPass::getAnalysisUsage(AU);
+    }
+  };
+}
+
+char CGBR::ID = 0;
+FunctionPass*
+llvm::createGlobalBaseRegPass() { return new CGBR(); }
diff --git a/lib/Target/X86/X86TargetMachine.cpp b/lib/Target/X86/X86TargetMachine.cpp
index a0d7cf6..df00d3f 100644
--- a/lib/Target/X86/X86TargetMachine.cpp
+++ b/lib/Target/X86/X86TargetMachine.cpp
@@ -173,6 +173,10 @@ bool X86TargetMachine::addInstSelector(PassManagerBase &PM,
   // Install an instruction selector.
   PM.add(createX86ISelDag(*this, OptLevel));
 
+  // For 32-bit, prepend instructions to set the "global base reg" for PIC.
+  if (!Subtarget.is64Bit())
+    PM.add(createGlobalBaseRegPass());
+
   return false;
 }