diff --git a/lib/Target/ARM/ARMISelDAGToDAG.cpp b/lib/Target/ARM/ARMISelDAGToDAG.cpp
index 8ef541f..a773916 100644
--- a/lib/Target/ARM/ARMISelDAGToDAG.cpp
+++ b/lib/Target/ARM/ARMISelDAGToDAG.cpp
@@ -1284,7 +1284,7 @@
       MVT HalfVT;
       unsigned Opc = 0;
       switch (VT.getVectorElementType().getSimpleVT()) {
-      default: assert(false && "unhandled VDUP splat type");
+      default: llvm_unreachable("unhandled VDUP splat type");
       case MVT::i8:  Opc = ARM::VDUPLN8q;  HalfVT = MVT::v8i8; break;
       case MVT::i16: Opc = ARM::VDUPLN16q; HalfVT = MVT::v4i16; break;
       case MVT::i32: Opc = ARM::VDUPLN32q; HalfVT = MVT::v2i32; break;
@@ -1304,6 +1304,62 @@
 
     break;
   }
+
+  case ARMISD::VLD2D: {
+    MVT VT = Op.getValueType();
+    SDValue MemAddr, MemUpdate, MemOpc;
+    if (!SelectAddrMode6(Op, N->getOperand(1), MemAddr, MemUpdate, MemOpc))
+        return NULL;
+    unsigned Opc;
+    switch (VT.getSimpleVT()) {
+    default: llvm_unreachable("unhandled VLD2D type");
+    case MVT::v8i8:  Opc = ARM::VLD2d8; break;
+    case MVT::v4i16: Opc = ARM::VLD2d16; break;
+    case MVT::v2f32:
+    case MVT::v2i32: Opc = ARM::VLD2d32; break;
+    case MVT::v1i64: Opc = ARM::VLD2d64; break;
+    }
+    const SDValue Ops[] = { MemAddr, MemUpdate, MemOpc };
+    return CurDAG->getTargetNode(Opc, dl, VT, VT, MVT::Other, Ops, 3);
+  }
+
+  case ARMISD::VLD3D: {
+    MVT VT = Op.getValueType();
+    SDValue MemAddr, MemUpdate, MemOpc;
+    if (!SelectAddrMode6(Op, N->getOperand(1), MemAddr, MemUpdate, MemOpc))
+        return NULL;
+    unsigned Opc;
+    switch (VT.getSimpleVT()) {
+    default: llvm_unreachable("unhandled VLD3D type");
+    case MVT::v8i8:  Opc = ARM::VLD3d8; break;
+    case MVT::v4i16: Opc = ARM::VLD3d16; break;
+    case MVT::v2f32:
+    case MVT::v2i32: Opc = ARM::VLD3d32; break;
+    case MVT::v1i64: Opc = ARM::VLD3d64; break;
+    }
+    const SDValue Ops[] = { MemAddr, MemUpdate, MemOpc };
+    return CurDAG->getTargetNode(Opc, dl, VT, VT, VT, MVT::Other, Ops, 3);
+  }
+
+  case ARMISD::VLD4D: {
+    MVT VT = Op.getValueType();
+    SDValue MemAddr, MemUpdate, MemOpc;
+    if (!SelectAddrMode6(Op, N->getOperand(1), MemAddr, MemUpdate, MemOpc))
+        return NULL;
+    unsigned Opc;
+    switch (VT.getSimpleVT()) {
+    default: llvm_unreachable("unhandled VLD4D type");
+    case MVT::v8i8:  Opc = ARM::VLD4d8; break;
+    case MVT::v4i16: Opc = ARM::VLD4d16; break;
+    case MVT::v2f32:
+    case MVT::v2i32: Opc = ARM::VLD4d32; break;
+    case MVT::v1i64: Opc = ARM::VLD4d64; break;
+    }
+    const SDValue Ops[] = { MemAddr, MemUpdate, MemOpc };
+    std::vector<MVT> ResTys(4, VT);
+    ResTys.push_back(MVT::Other);
+    return CurDAG->getTargetNode(Opc, dl, ResTys, Ops, 3);
+  }
   }
 
   return SelectCode(Op);
diff --git a/lib/Target/ARM/ARMISelLowering.cpp b/lib/Target/ARM/ARMISelLowering.cpp
index eb77547..987fc7f 100644
--- a/lib/Target/ARM/ARMISelLowering.cpp
+++ b/lib/Target/ARM/ARMISelLowering.cpp
@@ -1323,7 +1323,7 @@
 }
 
 static SDValue LowerNeonVLDIntrinsic(SDValue Op, SelectionDAG &DAG,
-                                     unsigned Opcode, unsigned NumVecs) {
+                                     unsigned Opcode) {
   SDNode *Node = Op.getNode();
   MVT VT = Node->getValueType(0);
   DebugLoc dl = Op.getDebugLoc();
@@ -1332,25 +1332,8 @@
     return SDValue(); // unimplemented
 
   SDValue Ops[] = { Node->getOperand(0),
-                    Node->getOperand(1) };
-  SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Flag);
-  SDValue Result = DAG.getNode(Opcode, dl, Tys, Ops, 2);
-
-  static const unsigned VLDRegs[] = {
-    ARM::D0, ARM::D1, ARM::D2, ARM::D3
-  };
-
-  SmallVector<SDValue, 4> ResultVals;
-  SDValue Chain = Result.getValue(0);
-  SDValue Flag = Result.getValue(1);
-  for (unsigned N = 0; N < NumVecs; ++N) {
-    Chain = DAG.getCopyFromReg(Chain, dl, VLDRegs[N], VT, Flag).getValue(1);
-    ResultVals.push_back(Chain.getValue(0));
-    Flag = Chain.getValue(2);
-  }
-  ResultVals.push_back(Chain);
-  return DAG.getNode(ISD::MERGE_VALUES, dl, Node->getVTList(),
-                     ResultVals.data(), NumVecs + 1);
+                    Node->getOperand(2) };
+  return DAG.getNode(Opcode, dl, Node->getVTList(), Ops, 2);
 }
 
 SDValue
@@ -1359,13 +1342,13 @@
   switch (IntNo) {
   case Intrinsic::arm_neon_vld2i:
   case Intrinsic::arm_neon_vld2f:
-    return LowerNeonVLDIntrinsic(Op, DAG, ARMISD::VLD2D, 2);
+    return LowerNeonVLDIntrinsic(Op, DAG, ARMISD::VLD2D);
   case Intrinsic::arm_neon_vld3i:
   case Intrinsic::arm_neon_vld3f:
-    return LowerNeonVLDIntrinsic(Op, DAG, ARMISD::VLD3D, 3);
+    return LowerNeonVLDIntrinsic(Op, DAG, ARMISD::VLD3D);
   case Intrinsic::arm_neon_vld4i:
   case Intrinsic::arm_neon_vld4f:
-    return LowerNeonVLDIntrinsic(Op, DAG, ARMISD::VLD4D, 4);
+    return LowerNeonVLDIntrinsic(Op, DAG, ARMISD::VLD4D);
   case Intrinsic::arm_neon_vst2i:
   case Intrinsic::arm_neon_vst2f:
   case Intrinsic::arm_neon_vst3i:
diff --git a/lib/Target/ARM/ARMInstrNEON.td b/lib/Target/ARM/ARMInstrNEON.td
index 540bd2f..e69ae61 100644
--- a/lib/Target/ARM/ARMInstrNEON.td
+++ b/lib/Target/ARM/ARMInstrNEON.td
@@ -68,13 +68,18 @@
 def NEONvduplaneq : SDNode<"ARMISD::VDUPLANEQ",
                            SDTypeProfile<1, 2, [SDTCisVT<2, i32>]>>;
 
-def SDTARMVLD     : SDTypeProfile<0, 1, [SDTCisPtrTy<0>]>;
-def NEONvld2d     : SDNode<"ARMISD::VLD2D", SDTARMVLD,
-                           [SDNPHasChain, SDNPOutFlag, SDNPMayLoad]>;
-def NEONvld3d     : SDNode<"ARMISD::VLD3D", SDTARMVLD,
-                           [SDNPHasChain, SDNPOutFlag, SDNPMayLoad]>;
-def NEONvld4d     : SDNode<"ARMISD::VLD4D", SDTARMVLD,
-                           [SDNPHasChain, SDNPOutFlag, SDNPMayLoad]>;
+def SDTARMVLD2    : SDTypeProfile<2, 1, [SDTCisSameAs<0, 1>, SDTCisPtrTy<2>]>;
+def SDTARMVLD3    : SDTypeProfile<3, 1, [SDTCisSameAs<0, 1>,
+                                         SDTCisSameAs<0, 2>, SDTCisPtrTy<3>]>;
+def SDTARMVLD4    : SDTypeProfile<4, 1, [SDTCisSameAs<0, 1>,
+                                         SDTCisSameAs<0, 2>,
+                                         SDTCisSameAs<0, 3>, SDTCisPtrTy<4>]>;
+def NEONvld2d     : SDNode<"ARMISD::VLD2D", SDTARMVLD2,
+                           [SDNPHasChain, SDNPMayLoad]>;
+def NEONvld3d     : SDNode<"ARMISD::VLD3D", SDTARMVLD3,
+                           [SDNPHasChain, SDNPMayLoad]>;
+def NEONvld4d     : SDNode<"ARMISD::VLD4D", SDTARMVLD4,
+                           [SDNPHasChain, SDNPMayLoad]>;
 
 //===----------------------------------------------------------------------===//
 // NEON operand definitions
@@ -183,6 +188,37 @@
 def  VST1qf   : VST1Q<"vst1.32", v4f32, int_arm_neon_vst1f>;
 def  VST1q64  : VST1Q<"vst1.64", v2i64, int_arm_neon_vst1i>;
 
+//   VLD2     : Vector Load (multiple 2-element structures)
+class VLD2D<string OpcodeStr>
+  : NLdSt<(outs DPR:$dst1, DPR:$dst2), (ins addrmode6:$addr),
+          !strconcat(OpcodeStr, "\t\\{$dst1,$dst2\\}, $addr"), []>;
+
+def  VLD2d8   : VLD2D<"vld2.8">;
+def  VLD2d16  : VLD2D<"vld2.16">;
+def  VLD2d32  : VLD2D<"vld2.32">;
+def  VLD2d64  : VLD2D<"vld2.64">;
+
+//   VLD3     : Vector Load (multiple 3-element structures)
+class VLD3D<string OpcodeStr>
+  : NLdSt<(outs DPR:$dst1, DPR:$dst2, DPR:$dst3), (ins addrmode6:$addr),
+          !strconcat(OpcodeStr, "\t\\{$dst1,$dst2,$dst3\\}, $addr"), []>;
+
+def  VLD3d8   : VLD3D<"vld3.8">;
+def  VLD3d16  : VLD3D<"vld3.16">;
+def  VLD3d32  : VLD3D<"vld3.32">;
+def  VLD3d64  : VLD3D<"vld3.64">;
+
+//   VLD4     : Vector Load (multiple 4-element structures)
+class VLD4D<string OpcodeStr>
+  : NLdSt<(outs DPR:$dst1, DPR:$dst2, DPR:$dst3, DPR:$dst4),
+          (ins addrmode6:$addr),
+          !strconcat(OpcodeStr, "\t\\{$dst1,$dst2,$dst3,$dst4\\}, $addr"), []>;
+
+def  VLD4d8   : VLD4D<"vld4.8">;
+def  VLD4d16  : VLD4D<"vld4.16">;
+def  VLD4d32  : VLD4D<"vld4.32">;
+def  VLD4d64  : VLD4D<"vld4.64">;
+
 
 //===----------------------------------------------------------------------===//
 // NEON pattern fragments
diff --git a/lib/Target/ARM/ARMTargetMachine.cpp b/lib/Target/ARM/ARMTargetMachine.cpp
index a6e987b..ff0b1d3 100644
--- a/lib/Target/ARM/ARMTargetMachine.cpp
+++ b/lib/Target/ARM/ARMTargetMachine.cpp
@@ -93,6 +93,8 @@
 
 bool ARMBaseTargetMachine::addPreRegAlloc(PassManagerBase &PM,
                                           CodeGenOpt::Level OptLevel) {
+  // Call NEON pre-alloc pass here.
+
   // FIXME: temporarily disabling load / store optimization pass for Thumb mode.
   if (OptLevel != CodeGenOpt::None && !DisableLdStOpti && !Subtarget.isThumb())
     PM.add(createARMLoadStoreOptimizationPass(true));