diff --git a/lib/Target/ARM/ARMISelDAGToDAG.cpp b/lib/Target/ARM/ARMISelDAGToDAG.cpp
index ebff3e1c4e0..bdc29402b45 100644
--- a/lib/Target/ARM/ARMISelDAGToDAG.cpp
+++ b/lib/Target/ARM/ARMISelDAGToDAG.cpp
@@ -444,7 +444,7 @@ bool ARMDAGToDAGISel::SelectAddrMode6(SDValue Op, SDValue N,
                                       SDValue &Addr, SDValue &Update,
                                       SDValue &Opc) {
   Addr = N;
-  // The optional writeback is handled in ARMLoadStoreOpt.
+  // Default to no writeback.
   Update = CurDAG->getRegister(0, MVT::i32);
   Opc = CurDAG->getTargetConstant(ARM_AM::getAM6Opc(false), MVT::i32);
   return true;
@@ -1388,16 +1388,57 @@ SDNode *ARMDAGToDAGISel::Select(SDValue Op) {
       SDValue MemAddr, MemUpdate, MemOpc;
       if (!SelectAddrMode6(Op, N->getOperand(2), MemAddr, MemUpdate, MemOpc))
         return NULL;
+      if (VT.is64BitVector()) {
+        switch (VT.getSimpleVT().SimpleTy) {
+        default: llvm_unreachable("unhandled vld3 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;
+        }
+        SDValue Chain = N->getOperand(0);
+        const SDValue Ops[] = { MemAddr, MemUpdate, MemOpc, Chain };
+        return CurDAG->getMachineNode(Opc, dl, VT, VT, VT, MVT::Other, Ops, 4);
+      }
+      // Quad registers are loaded with two separate instructions, where one
+      // loads the even registers and the other loads the odd registers.
+      EVT RegVT = VT;
+      unsigned Opc2 = 0;
       switch (VT.getSimpleVT().SimpleTy) {
       default: llvm_unreachable("unhandled vld3 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::v16i8:
+        Opc = ARM::VLD3q8a;  Opc2 = ARM::VLD3q8b;  RegVT = MVT::v8i8; break;
+      case MVT::v8i16:
+        Opc = ARM::VLD3q16a; Opc2 = ARM::VLD3q16b; RegVT = MVT::v4i16; break;
+      case MVT::v4f32:
+        Opc = ARM::VLD3q32a; Opc2 = ARM::VLD3q32b; RegVT = MVT::v2f32; break;
+      case MVT::v4i32:
+        Opc = ARM::VLD3q32a; Opc2 = ARM::VLD3q32b; RegVT = MVT::v2i32; break;
       }
       SDValue Chain = N->getOperand(0);
-      const SDValue Ops[] = { MemAddr, MemUpdate, MemOpc, Chain };
-      return CurDAG->getMachineNode(Opc, dl, VT, VT, VT, MVT::Other, Ops, 4);
+      // Enable writeback to the address register.
+      MemOpc = CurDAG->getTargetConstant(ARM_AM::getAM6Opc(true), MVT::i32);
+
+      std::vector<EVT> ResTys(3, RegVT);
+      ResTys.push_back(MemAddr.getValueType());
+      ResTys.push_back(MVT::Other);
+
+      const SDValue OpsA[] = { MemAddr, MemUpdate, MemOpc, Chain };
+      SDNode *VLdA = CurDAG->getMachineNode(Opc, dl, ResTys, OpsA, 4);
+      Chain = SDValue(VLdA, 4);
+
+      const SDValue OpsB[] = { SDValue(VLdA, 3), MemUpdate, MemOpc, Chain };
+      SDNode *VLdB = CurDAG->getMachineNode(Opc2, dl, ResTys, OpsB, 4);
+      Chain = SDValue(VLdB, 4);
+
+      SDNode *Q0 = PairDRegs(VT, SDValue(VLdA, 0), SDValue(VLdB, 0));
+      SDNode *Q1 = PairDRegs(VT, SDValue(VLdA, 1), SDValue(VLdB, 1));
+      SDNode *Q2 = PairDRegs(VT, SDValue(VLdA, 2), SDValue(VLdB, 2));
+      ReplaceUses(SDValue(N, 0), SDValue(Q0, 0));
+      ReplaceUses(SDValue(N, 1), SDValue(Q1, 0));
+      ReplaceUses(SDValue(N, 2), SDValue(Q2, 0));
+      ReplaceUses(SDValue(N, 3), Chain);
+      return NULL;
     }
 
     case Intrinsic::arm_neon_vld4: {
diff --git a/lib/Target/ARM/ARMInstrNEON.td b/lib/Target/ARM/ARMInstrNEON.td
index 3f07d302b03..c7ff523b437 100644
--- a/lib/Target/ARM/ARMInstrNEON.td
+++ b/lib/Target/ARM/ARMInstrNEON.td
@@ -201,11 +201,26 @@ class VLD3D<string OpcodeStr>
   : NLdSt<(outs DPR:$dst1, DPR:$dst2, DPR:$dst3), (ins addrmode6:$addr),
           IIC_VLD3,
           !strconcat(OpcodeStr, "\t\\{$dst1,$dst2,$dst3\\}, $addr"), "", []>;
+class VLD3WB<string OpcodeStr>
+  : NLdSt<(outs DPR:$dst1, DPR:$dst2, DPR:$dst3, GPR:$wb),
+          (ins addrmode6:$addr), IIC_VLD3,
+          !strconcat(OpcodeStr, "\t\\{$dst1,$dst2,$dst3\\}, $addr"),
+          "$addr.addr = $wb", []>;
 
 def  VLD3d8   : VLD3D<"vld3.8">;
 def  VLD3d16  : VLD3D<"vld3.16">;
 def  VLD3d32  : VLD3D<"vld3.32">;
 
+// vld3 to double-spaced even registers.
+def  VLD3q8a  : VLD3WB<"vld3.8">;
+def  VLD3q16a : VLD3WB<"vld3.16">;
+def  VLD3q32a : VLD3WB<"vld3.32">;
+
+// vld3 to double-spaced odd registers.
+def  VLD3q8b  : VLD3WB<"vld3.8">;
+def  VLD3q16b : VLD3WB<"vld3.16">;
+def  VLD3q32b : VLD3WB<"vld3.32">;
+
 //   VLD4     : Vector Load (multiple 4-element structures)
 class VLD4D<string OpcodeStr>
   : NLdSt<(outs DPR:$dst1, DPR:$dst2, DPR:$dst3, DPR:$dst4),
diff --git a/lib/Target/ARM/NEONPreAllocPass.cpp b/lib/Target/ARM/NEONPreAllocPass.cpp
index da1c662edb6..fab62f6cfc2 100644
--- a/lib/Target/ARM/NEONPreAllocPass.cpp
+++ b/lib/Target/ARM/NEONPreAllocPass.cpp
@@ -36,8 +36,12 @@ namespace {
   char NEONPreAllocPass::ID = 0;
 }
 
-static bool isNEONMultiRegOp(int Opcode, unsigned &FirstOpnd,
-                             unsigned &NumRegs) {
+static bool isNEONMultiRegOp(int Opcode, unsigned &FirstOpnd, unsigned &NumRegs,
+                             unsigned &Offset, unsigned &Stride) {
+  // Default to unit stride with no offset.
+  Stride = 1;
+  Offset = 0;
+
   switch (Opcode) {
   default:
     break;
@@ -69,6 +73,24 @@ static bool isNEONMultiRegOp(int Opcode, unsigned &FirstOpnd,
     NumRegs = 3;
     return true;
 
+  case ARM::VLD3q8a:
+  case ARM::VLD3q16a:
+  case ARM::VLD3q32a:
+    FirstOpnd = 0;
+    NumRegs = 3;
+    Offset = 0;
+    Stride = 2;
+    return true;
+
+  case ARM::VLD3q8b:
+  case ARM::VLD3q16b:
+  case ARM::VLD3q32b:
+    FirstOpnd = 0;
+    NumRegs = 3;
+    Offset = 1;
+    Stride = 2;
+    return true;
+
   case ARM::VLD4d8:
   case ARM::VLD4d16:
   case ARM::VLD4d32:
@@ -149,8 +171,8 @@ bool NEONPreAllocPass::PreAllocNEONRegisters(MachineBasicBlock &MBB) {
   MachineBasicBlock::iterator MBBI = MBB.begin(), E = MBB.end();
   for (; MBBI != E; ++MBBI) {
     MachineInstr *MI = &*MBBI;
-    unsigned FirstOpnd, NumRegs;
-    if (!isNEONMultiRegOp(MI->getOpcode(), FirstOpnd, NumRegs))
+    unsigned FirstOpnd, NumRegs, Offset, Stride;
+    if (!isNEONMultiRegOp(MI->getOpcode(), FirstOpnd, NumRegs, Offset, Stride))
       continue;
 
     MachineBasicBlock::iterator NextI = next(MBBI);
@@ -164,9 +186,10 @@ bool NEONPreAllocPass::PreAllocNEONRegisters(MachineBasicBlock &MBB) {
       // For now, just assign a fixed set of adjacent registers.
       // This leaves plenty of room for future improvements.
       static const unsigned NEONDRegs[] = {
-        ARM::D0, ARM::D1, ARM::D2, ARM::D3
+        ARM::D0, ARM::D1, ARM::D2, ARM::D3,
+        ARM::D4, ARM::D5, ARM::D6, ARM::D7
       };
-      MO.setReg(NEONDRegs[R]);
+      MO.setReg(NEONDRegs[Offset + R * Stride]);
 
       if (MO.isUse()) {
         // Insert a copy from VirtReg.
diff --git a/test/CodeGen/ARM/vld3.ll b/test/CodeGen/ARM/vld3.ll
index b03d74ab5dd..4ed53092c3f 100644
--- a/test/CodeGen/ARM/vld3.ll
+++ b/test/CodeGen/ARM/vld3.ll
@@ -5,6 +5,11 @@
 %struct.__neon_int32x2x3_t = type { <2 x i32>, <2 x i32>, <2 x i32> }
 %struct.__neon_float32x2x3_t = type { <2 x float>, <2 x float>, <2 x float> }
 
+%struct.__neon_int8x16x3_t = type { <16 x i8>,  <16 x i8>,  <16 x i8> }
+%struct.__neon_int16x8x3_t = type { <8 x i16>, <8 x i16>, <8 x i16> }
+%struct.__neon_int32x4x3_t = type { <4 x i32>, <4 x i32>, <4 x i32> }
+%struct.__neon_float32x4x3_t = type { <4 x float>, <4 x float>, <4 x float> }
+
 define <8 x i8> @vld3i8(i8* %A) nounwind {
 ;CHECK: vld3i8:
 ;CHECK: vld3.8
@@ -45,7 +50,56 @@ define <2 x float> @vld3f(float* %A) nounwind {
 	ret <2 x float> %tmp4
 }
 
+define <16 x i8> @vld3Qi8(i8* %A) nounwind {
+;CHECK: vld3Qi8:
+;CHECK: vld3.8
+;CHECK: vld3.8
+	%tmp1 = call %struct.__neon_int8x16x3_t @llvm.arm.neon.vld3.v16i8(i8* %A)
+        %tmp2 = extractvalue %struct.__neon_int8x16x3_t %tmp1, 0
+        %tmp3 = extractvalue %struct.__neon_int8x16x3_t %tmp1, 2
+        %tmp4 = add <16 x i8> %tmp2, %tmp3
+	ret <16 x i8> %tmp4
+}
+
+define <8 x i16> @vld3Qi16(i16* %A) nounwind {
+;CHECK: vld3Qi16:
+;CHECK: vld3.16
+;CHECK: vld3.16
+	%tmp1 = call %struct.__neon_int16x8x3_t @llvm.arm.neon.vld3.v8i16(i16* %A)
+        %tmp2 = extractvalue %struct.__neon_int16x8x3_t %tmp1, 0
+        %tmp3 = extractvalue %struct.__neon_int16x8x3_t %tmp1, 2
+        %tmp4 = add <8 x i16> %tmp2, %tmp3
+	ret <8 x i16> %tmp4
+}
+
+define <4 x i32> @vld3Qi32(i32* %A) nounwind {
+;CHECK: vld3Qi32:
+;CHECK: vld3.32
+;CHECK: vld3.32
+	%tmp1 = call %struct.__neon_int32x4x3_t @llvm.arm.neon.vld3.v4i32(i32* %A)
+        %tmp2 = extractvalue %struct.__neon_int32x4x3_t %tmp1, 0
+        %tmp3 = extractvalue %struct.__neon_int32x4x3_t %tmp1, 2
+        %tmp4 = add <4 x i32> %tmp2, %tmp3
+	ret <4 x i32> %tmp4
+}
+
+define <4 x float> @vld3Qf(float* %A) nounwind {
+;CHECK: vld3Qf:
+;CHECK: vld3.32
+;CHECK: vld3.32
+	%tmp1 = call %struct.__neon_float32x4x3_t @llvm.arm.neon.vld3.v4f32(float* %A)
+        %tmp2 = extractvalue %struct.__neon_float32x4x3_t %tmp1, 0
+        %tmp3 = extractvalue %struct.__neon_float32x4x3_t %tmp1, 2
+        %tmp4 = add <4 x float> %tmp2, %tmp3
+	ret <4 x float> %tmp4
+}
+
 declare %struct.__neon_int8x8x3_t @llvm.arm.neon.vld3.v8i8(i8*) nounwind readonly
 declare %struct.__neon_int16x4x3_t @llvm.arm.neon.vld3.v4i16(i8*) nounwind readonly
 declare %struct.__neon_int32x2x3_t @llvm.arm.neon.vld3.v2i32(i8*) nounwind readonly
 declare %struct.__neon_float32x2x3_t @llvm.arm.neon.vld3.v2f32(i8*) nounwind readonly
+
+declare %struct.__neon_int8x16x3_t @llvm.arm.neon.vld3.v16i8(i8*) nounwind readonly
+declare %struct.__neon_int16x8x3_t @llvm.arm.neon.vld3.v8i16(i8*) nounwind readonly
+declare %struct.__neon_int32x4x3_t @llvm.arm.neon.vld3.v4i32(i8*) nounwind readonly
+declare %struct.__neon_float32x4x3_t @llvm.arm.neon.vld3.v4f32(i8*) nounwind readonly