diff --git a/lib/Target/X86/X86FastISel.cpp b/lib/Target/X86/X86FastISel.cpp
index ca127d312c7..658b65dba73 100644
--- a/lib/Target/X86/X86FastISel.cpp
+++ b/lib/Target/X86/X86FastISel.cpp
@@ -14,6 +14,7 @@
 //===----------------------------------------------------------------------===//
 
 #include "X86.h"
+#include "X86InstrBuilder.h"
 #include "X86ISelLowering.h"
 #include "X86RegisterInfo.h"
 #include "X86Subtarget.h"
@@ -28,7 +29,7 @@ class X86FastISel : public FastISel {
   /// make the right decision when generating code for different targets.
   const X86Subtarget *Subtarget;
     
- public:
+public:
   explicit X86FastISel(MachineFunction &mf) : FastISel(mf) {
     Subtarget = &TM.getSubtarget<X86Subtarget>();
   }
@@ -40,8 +41,134 @@ class X86FastISel : public FastISel {
                             MachineBasicBlock *MBB);
 
 #include "X86GenFastISel.inc"
+
+private:
+  bool X86SelectConstAddr(Value *V,
+                          DenseMap<const Value *, unsigned> &ValueMap,
+                          MachineBasicBlock *MBB, unsigned &Op0);
+
+  bool X86SelectLoad(Instruction *I,
+                     DenseMap<const Value *, unsigned> &ValueMap,
+                     MachineBasicBlock *MBB);
 };
 
+/// X86SelectConstAddr - Select and emit code to materialize constant address.
+/// 
+bool X86FastISel::X86SelectConstAddr(Value *V,
+                                    DenseMap<const Value *, unsigned> &ValueMap,
+                                    MachineBasicBlock *MBB,
+                                    unsigned &Op0) {
+  // FIXME: Only GlobalAddress for now.
+  GlobalValue *GV = dyn_cast<GlobalValue>(V);
+  if (!GV)
+    return false;
+
+  if (Subtarget->GVRequiresExtraLoad(GV, TM, false)) {
+    // Issue load from stub if necessary.
+    unsigned Opc = 0;
+    const TargetRegisterClass *RC = NULL;
+    if (TLI.getPointerTy() == MVT::i32) {
+      Opc = X86::MOV32rm;
+      RC  = X86::GR32RegisterClass;
+    } else {
+      Opc = X86::MOV64rm;
+      RC  = X86::GR64RegisterClass;
+    }
+    Op0 = createResultReg(RC);
+    X86AddressMode AM;
+    AM.GV = GV;
+    addFullAddress(BuildMI(MBB, TII.get(Opc), Op0), AM);
+  }
+  return true;
+}
+
+/// X86SelectLoad - Select and emit code to implement load instructions.
+///
+bool X86FastISel::X86SelectLoad(Instruction *I,
+                                DenseMap<const Value *, unsigned> &ValueMap,
+                                MachineBasicBlock *MBB)  {
+  MVT VT = MVT::getMVT(I->getType(), /*HandleUnknown=*/true);
+  if (VT == MVT::Other || !VT.isSimple())
+    // Unhandled type. Halt "fast" selection and bail.
+    return false;
+  if (VT == MVT::iPTR)
+    // Use pointer type.
+    VT = TLI.getPointerTy();
+  // We only handle legal types. For example, on x86-32 the instruction
+  // selector contains all of the 64-bit instructions from x86-64,
+  // under the assumption that i64 won't be used if the target doesn't
+  // support it.
+  if (!TLI.isTypeLegal(VT))
+    return false;
+
+  Value *V = I->getOperand(0);
+  unsigned Op0 = getRegForValue(V, ValueMap);
+  if (Op0 == 0) {
+    // Handle constant load address.
+    if (!isa<Constant>(V) || !X86SelectConstAddr(V, ValueMap, MBB, Op0))
+      // Unhandled operand. Halt "fast" selection and bail.
+      return false;    
+  }
+
+  // Get opcode and regclass of the output for the given load instruction.
+  unsigned Opc = 0;
+  const TargetRegisterClass *RC = NULL;
+  switch (VT.getSimpleVT()) {
+  default: return false;
+  case MVT::i8:
+    Opc = X86::MOV8rm;
+    RC  = X86::GR8RegisterClass;
+    break;
+  case MVT::i16:
+    Opc = X86::MOV16rm;
+    RC  = X86::GR16RegisterClass;
+    break;
+  case MVT::i32:
+    Opc = X86::MOV32rm;
+    RC  = X86::GR32RegisterClass;
+    break;
+  case MVT::i64:
+    // Must be in x86-64 mode.
+    Opc = X86::MOV64rm;
+    RC  = X86::GR64RegisterClass;
+    break;
+  case MVT::f32:
+    if (Subtarget->hasSSE1()) {
+      Opc = X86::MOVSSrm;
+      RC  = X86::FR32RegisterClass;
+    } else {
+      Opc = X86::LD_Fp32m;
+      RC  = X86::RFP32RegisterClass;
+    }
+    break;
+  case MVT::f64:
+    if (Subtarget->hasSSE2()) {
+      Opc = X86::MOVSDrm;
+      RC  = X86::FR64RegisterClass;
+    } else {
+      Opc = X86::LD_Fp64m;
+      RC  = X86::RFP64RegisterClass;
+    }
+    break;
+  case MVT::f80:
+    Opc = X86::LD_Fp80m;
+    RC  = X86::RFP80RegisterClass;
+    break;
+  }
+
+  unsigned ResultReg = createResultReg(RC);
+  X86AddressMode AM;
+  if (Op0)
+    // Address is in register.
+    AM.Base.Reg = Op0;
+  else
+    AM.GV = cast<GlobalValue>(V);
+  addFullAddress(BuildMI(MBB, TII.get(Opc), ResultReg), AM);
+  UpdateValueMap(I, ResultReg, ValueMap);
+  return true;
+}
+
+
 bool
 X86FastISel::TargetSelectInstruction(Instruction *I,
                                   DenseMap<const Value *, unsigned> &ValueMap,
@@ -49,6 +176,8 @@ X86FastISel::TargetSelectInstruction(Instruction *I,
                                   MachineBasicBlock *MBB)  {
   switch (I->getOpcode()) {
   default: break;
+  case Instruction::Load:
+    return X86SelectLoad(I, ValueMap, MBB);
   }
 
   return false;
diff --git a/test/CodeGen/X86/fast-isel.ll b/test/CodeGen/X86/fast-isel.ll
index 1260eed07f5..10b2e1ee42a 100644
--- a/test/CodeGen/X86/fast-isel.ll
+++ b/test/CodeGen/X86/fast-isel.ll
@@ -2,7 +2,7 @@
 
 ; This tests very minimal fast-isel functionality.
 
-define i32* @foo(i32* %p, i32* %q, i32** %z) {
+define i32* @foo(i32* %p, i32* %q, i32** %z) nounwind {
 entry:
   %r = load i32* %p
   %s = load i32* %q
@@ -25,7 +25,7 @@ exit:
   ret i32* %t8
 }
 
-define double @bar(double* %p, double* %q) {
+define double @bar(double* %p, double* %q) nounwind {
 entry:
   %r = load double* %p
   %s = load double* %q
@@ -42,8 +42,16 @@ exit:
   ret double %t3
 }
 
-define i32 @cast(){
+define i32 @cast() nounwind {
 entry:
 	%tmp2 = bitcast i32 0 to i32
 	ret i32 %tmp2
 }
+
+@src = external global i32
+
+define i32 @loadgv() nounwind {
+entry:
+	%0 = load i32* @src, align 4
+	ret i32 %0
+}