1 files changed, 52 insertions, 10 deletions
diff --git a/llvm/lib/Target/ARM/ARMISelDAGToDAG.cpp b/llvm/lib/Target/ARM/ARMISelDAGToDAG.cpp
index fd5aac977b9..be3f343b68f 100644
--- a/llvm/lib/Target/ARM/ARMISelDAGToDAG.cpp
+++ b/llvm/lib/Target/ARM/ARMISelDAGToDAG.cpp
@@ -2025,19 +2025,61 @@ SDNode *ARMDAGToDAGISel::Select(SDValue Op) {
       SDValue MemAddr, MemUpdate, MemOpc;
       if (!SelectAddrMode6(Op, N->getOperand(2), MemAddr, MemUpdate, MemOpc))
         return NULL;
-      switch (N->getOperand(3).getValueType().getSimpleVT().SimpleTy) {
+      VT = N->getOperand(3).getValueType();
+      if (VT.is64BitVector()) {
+        switch (VT.getSimpleVT().SimpleTy) {
+        default: llvm_unreachable("unhandled vst4lane type");
+        case MVT::v8i8:  Opc = ARM::VST4LNd8; break;
+        case MVT::v4i16: Opc = ARM::VST4LNd16; break;
+        case MVT::v2f32:
+        case MVT::v2i32: Opc = ARM::VST4LNd32; break;
+        }
+        SDValue Chain = N->getOperand(0);
+        const SDValue Ops[] = { MemAddr, MemUpdate, MemOpc,
+                                N->getOperand(3), N->getOperand(4),
+                                N->getOperand(5), N->getOperand(6),
+                                N->getOperand(7), Chain };
+        return CurDAG->getMachineNode(Opc, dl, MVT::Other, Ops, 9);
+      }
+      // Quad registers are handled by extracting subregs and then doing
+      // the store.
+      EVT RegVT;
+      unsigned Opc2 = 0;
+      switch (VT.getSimpleVT().SimpleTy) {
       default: llvm_unreachable("unhandled vst4lane type");
-      case MVT::v8i8:  Opc = ARM::VST4LNd8; break;
-      case MVT::v4i16: Opc = ARM::VST4LNd16; break;
-      case MVT::v2f32:
-      case MVT::v2i32: Opc = ARM::VST4LNd32; break;
+      case MVT::v8i16:
+        Opc = ARM::VST4LNq16a;
+        Opc2 = ARM::VST4LNq16b;
+        RegVT = MVT::v4i16;
+        break;
+      case MVT::v4f32:
+        Opc = ARM::VST4LNq32a;
+        Opc2 = ARM::VST4LNq32b;
+        RegVT = MVT::v2f32;
+        break;
+      case MVT::v4i32:
+        Opc = ARM::VST4LNq32a;
+        Opc2 = ARM::VST4LNq32b;
+        RegVT = MVT::v2i32;
+        break;
       }
       SDValue Chain = N->getOperand(0);
-      const SDValue Ops[] = { MemAddr, MemUpdate, MemOpc,
-                              N->getOperand(3), N->getOperand(4),
-                              N->getOperand(5), N->getOperand(6),
-                              N->getOperand(7), Chain };
-      return CurDAG->getMachineNode(Opc, dl, MVT::Other, Ops, 9);
+      unsigned Lane = cast<ConstantSDNode>(N->getOperand(7))->getZExtValue();
+      unsigned NumElts = RegVT.getVectorNumElements();
+      int SubregIdx = (Lane < NumElts) ? ARM::DSUBREG_0 : ARM::DSUBREG_1;
+
+      SDValue D0 = CurDAG->getTargetExtractSubreg(SubregIdx, dl, RegVT,
+                                                  N->getOperand(3));
+      SDValue D1 = CurDAG->getTargetExtractSubreg(SubregIdx, dl, RegVT,
+                                                  N->getOperand(4));
+      SDValue D2 = CurDAG->getTargetExtractSubreg(SubregIdx, dl, RegVT,
+                                                  N->getOperand(5));
+      SDValue D3 = CurDAG->getTargetExtractSubreg(SubregIdx, dl, RegVT,
+                                                  N->getOperand(6));
+      const SDValue Ops[] = { MemAddr, MemUpdate, MemOpc, D0, D1, D2, D3,
+                              getI32Imm(Lane % NumElts), Chain };
+      return CurDAG->getMachineNode((Lane < NumElts) ? Opc : Opc2,
+                                    dl, MVT::Other, Ops, 9);
     }
     }
   }