Add codegen support for NEON vst3lane intrinsics with 128-bit vectors.
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@83598 91177308-0d34-0410-b5e6-96231b3b80d8
diff --git a/lib/Target/ARM/ARMISelDAGToDAG.cpp b/lib/Target/ARM/ARMISelDAGToDAG.cpp
index 717826d..fd5aac9 100644
--- a/lib/Target/ARM/ARMISelDAGToDAG.cpp
+++ b/lib/Target/ARM/ARMISelDAGToDAG.cpp
@@ -1967,18 +1967,58 @@
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 vst3lane type");
+ case MVT::v8i8: Opc = ARM::VST3LNd8; break;
+ case MVT::v4i16: Opc = ARM::VST3LNd16; break;
+ case MVT::v2f32:
+ case MVT::v2i32: Opc = ARM::VST3LNd32; break;
+ }
+ SDValue Chain = N->getOperand(0);
+ const SDValue Ops[] = { MemAddr, MemUpdate, MemOpc,
+ N->getOperand(3), N->getOperand(4),
+ N->getOperand(5), N->getOperand(6), Chain };
+ return CurDAG->getMachineNode(Opc, dl, MVT::Other, Ops, 8);
+ }
+ // 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 vst3lane type");
- case MVT::v8i8: Opc = ARM::VST3LNd8; break;
- case MVT::v4i16: Opc = ARM::VST3LNd16; break;
- case MVT::v2f32:
- case MVT::v2i32: Opc = ARM::VST3LNd32; break;
+ case MVT::v8i16:
+ Opc = ARM::VST3LNq16a;
+ Opc2 = ARM::VST3LNq16b;
+ RegVT = MVT::v4i16;
+ break;
+ case MVT::v4f32:
+ Opc = ARM::VST3LNq32a;
+ Opc2 = ARM::VST3LNq32b;
+ RegVT = MVT::v2f32;
+ break;
+ case MVT::v4i32:
+ Opc = ARM::VST3LNq32a;
+ Opc2 = ARM::VST3LNq32b;
+ 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), Chain };
- return CurDAG->getMachineNode(Opc, dl, MVT::Other, Ops, 8);
+ unsigned Lane = cast<ConstantSDNode>(N->getOperand(6))->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));
+ const SDValue Ops[] = { MemAddr, MemUpdate, MemOpc, D0, D1, D2,
+ getI32Imm(Lane % NumElts), Chain };
+ return CurDAG->getMachineNode((Lane < NumElts) ? Opc : Opc2,
+ dl, MVT::Other, Ops, 8);
}
case Intrinsic::arm_neon_vst4lane: {