Add codegen support for NEON vld2lane intrinsics with 128-bit vectors.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@83568 91177308-0d34-0410-b5e6-96231b3b80d8

lib/Target/ARM/ARMISelDAGToDAG.cpp

@@ -1510,6 +1510,7 @@ 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 vld2lane type");
         case MVT::v8i8:  Opc = ARM::VLD2LNd8; break;
@@ -1523,6 +1524,54 @@ SDNode *ARMDAGToDAGISel::Select(SDValue Op) {
                                 N->getOperand(5), Chain };
         return CurDAG->getMachineNode(Opc, dl, VT, VT, MVT::Other, Ops, 7);
       }
+      // Quad registers are handled by extracting subregs, doing the load,
+      // and then inserting the results as subregs.
+      EVT RegVT;
+      unsigned Opc2 = 0;
+      switch (VT.getSimpleVT().SimpleTy) {
+      default: llvm_unreachable("unhandled vld2lane type");
+      case MVT::v8i16:
+        Opc = ARM::VLD2LNq16a;
+        Opc2 = ARM::VLD2LNq16b;
+        RegVT = MVT::v4i16;
+        break;
+      case MVT::v4f32:
+        Opc = ARM::VLD2LNq32a;
+        Opc2 = ARM::VLD2LNq32b;
+        RegVT = MVT::v2f32;
+        break;
+      case MVT::v4i32:
+        Opc = ARM::VLD2LNq32a;
+        Opc2 = ARM::VLD2LNq32b;
+        RegVT = MVT::v2i32;
+        break;
+      }
+      SDValue Chain = N->getOperand(0);
+      unsigned Lane = cast<ConstantSDNode>(N->getOperand(5))->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));
+      const SDValue Ops[] = { MemAddr, MemUpdate, MemOpc, D0, D1,
+                              getI32Imm(Lane % NumElts), Chain };
+      SDNode *VLdLn = CurDAG->getMachineNode((Lane < NumElts) ? Opc : Opc2,
+                                             dl, RegVT, RegVT, MVT::Other,
+                                             Ops, 7);
+      SDValue Q0 = CurDAG->getTargetInsertSubreg(SubregIdx, dl, VT,
+                                                 N->getOperand(3),
+                                                 SDValue(VLdLn, 0));
+      SDValue Q1 = CurDAG->getTargetInsertSubreg(SubregIdx, dl, VT,
+                                                 N->getOperand(4),
+                                                 SDValue(VLdLn, 1));
+      Chain = SDValue(VLdLn, 2);
+      ReplaceUses(SDValue(N, 0), Q0);
+      ReplaceUses(SDValue(N, 1), Q1);
+      ReplaceUses(SDValue(N, 2), Chain);
+      return NULL;
+    }
 
     case Intrinsic::arm_neon_vld3lane: {
       SDValue MemAddr, MemUpdate, MemOpc;

lib/Target/ARM/ARMInstrNEON.td

@@ -266,16 +266,24 @@ def  VLD4q32b : VLD4WB<0b1000, "vld4.32">;
 //   FIXME: Not yet implemented.
 
 //   VLD2LN   : Vector Load (single 2-element structure to one lane)
-class VLD2LND<bits<4> op11_8, string OpcodeStr>
+class VLD2LN<bits<4> op11_8, string OpcodeStr>
   : NLdSt<1,0b10,op11_8,0b0000, (outs DPR:$dst1, DPR:$dst2),
           (ins addrmode6:$addr, DPR:$src1, DPR:$src2, nohash_imm:$lane),
           IIC_VLD2,
           !strconcat(OpcodeStr, "\t\\{$dst1[$lane],$dst2[$lane]\\}, $addr"),
           "$src1 = $dst1, $src2 = $dst2", []>;
 
-def VLD2LNd8  : VLD2LND<0b0001, "vld2.8">;
+def VLD2LNd8  : VLD2LN<0b0001, "vld2.8">;
-def VLD2LNd16 : VLD2LND<0b0101, "vld2.16">;
+def VLD2LNd16 : VLD2LN<0b0101, "vld2.16">;
-def VLD2LNd32 : VLD2LND<0b1001, "vld2.32">;
+def VLD2LNd32 : VLD2LN<0b1001, "vld2.32">;
+
+// vld2 to double-spaced even registers.
+def VLD2LNq16a: VLD2LN<0b0101, "vld2.16">;
+def VLD2LNq32a: VLD2LN<0b1001, "vld2.32">;
+
+// vld2 to double-spaced odd registers.
+def VLD2LNq16b: VLD2LN<0b0101, "vld2.16">;
+def VLD2LNq32b: VLD2LN<0b1001, "vld2.32">;
 
 //   VLD3LN   : Vector Load (single 3-element structure to one lane)
 class VLD3LND<bits<4> op11_8, string OpcodeStr>

lib/Target/ARM/NEONPreAllocPass.cpp

@@ -57,6 +57,22 @@ static bool isNEONMultiRegOp(int Opcode, unsigned &FirstOpnd, unsigned &NumRegs,
     NumRegs = 2;
     return true;
 
+  case ARM::VLD2LNq16a:
+  case ARM::VLD2LNq32a:
+    FirstOpnd = 0;
+    NumRegs = 2;
+    Offset = 0;
+    Stride = 2;
+    return true;
+
+  case ARM::VLD2LNq16b:
+  case ARM::VLD2LNq32b:
+    FirstOpnd = 0;
+    NumRegs = 2;
+    Offset = 1;
+    Stride = 2;
+    return true;
+
   case ARM::VLD2q8:
   case ARM::VLD2q16:
   case ARM::VLD2q32:

test/CodeGen/ARM/vldlane.ll

@@ -5,6 +5,10 @@
 %struct.__neon_int32x2x2_t = type { <2 x i32>, <2 x i32> }
 %struct.__neon_float32x2x2_t = type { <2 x float>, <2 x float> }
 
+%struct.__neon_int16x8x2_t = type { <8 x i16>, <8 x i16> }
+%struct.__neon_int32x4x2_t = type { <4 x i32>, <4 x i32> }
+%struct.__neon_float32x4x2_t = type { <4 x float>, <4 x float> }
+
 define <8 x i8> @vld2lanei8(i8* %A, <8 x i8>* %B) nounwind {
 ;CHECK: vld2lanei8:
 ;CHECK: vld2.8
@@ -49,11 +53,48 @@ define <2 x float> @vld2lanef(float* %A, <2 x float>* %B) nounwind {
 	ret <2 x float> %tmp5
 }
 
+define <8 x i16> @vld2laneQi16(i16* %A, <8 x i16>* %B) nounwind {
+;CHECK: vld2laneQi16:
+;CHECK: vld2.16
+	%tmp1 = load <8 x i16>* %B
+	%tmp2 = call %struct.__neon_int16x8x2_t @llvm.arm.neon.vld2lane.v8i16(i16* %A, <8 x i16> %tmp1, <8 x i16> %tmp1, i32 1)
+        %tmp3 = extractvalue %struct.__neon_int16x8x2_t %tmp2, 0
+        %tmp4 = extractvalue %struct.__neon_int16x8x2_t %tmp2, 1
+        %tmp5 = add <8 x i16> %tmp3, %tmp4
+	ret <8 x i16> %tmp5
+}
+
+define <4 x i32> @vld2laneQi32(i32* %A, <4 x i32>* %B) nounwind {
+;CHECK: vld2laneQi32:
+;CHECK: vld2.32
+	%tmp1 = load <4 x i32>* %B
+	%tmp2 = call %struct.__neon_int32x4x2_t @llvm.arm.neon.vld2lane.v4i32(i32* %A, <4 x i32> %tmp1, <4 x i32> %tmp1, i32 2)
+        %tmp3 = extractvalue %struct.__neon_int32x4x2_t %tmp2, 0
+        %tmp4 = extractvalue %struct.__neon_int32x4x2_t %tmp2, 1
+        %tmp5 = add <4 x i32> %tmp3, %tmp4
+	ret <4 x i32> %tmp5
+}
+
+define <4 x float> @vld2laneQf(float* %A, <4 x float>* %B) nounwind {
+;CHECK: vld2laneQf:
+;CHECK: vld2.32
+	%tmp1 = load <4 x float>* %B
+	%tmp2 = call %struct.__neon_float32x4x2_t @llvm.arm.neon.vld2lane.v4f32(float* %A, <4 x float> %tmp1, <4 x float> %tmp1, i32 1)
+        %tmp3 = extractvalue %struct.__neon_float32x4x2_t %tmp2, 0
+        %tmp4 = extractvalue %struct.__neon_float32x4x2_t %tmp2, 1
+        %tmp5 = add <4 x float> %tmp3, %tmp4
+	ret <4 x float> %tmp5
+}
+
 declare %struct.__neon_int8x8x2_t @llvm.arm.neon.vld2lane.v8i8(i8*, <8 x i8>, <8 x i8>, i32) nounwind readonly
 declare %struct.__neon_int16x4x2_t @llvm.arm.neon.vld2lane.v4i16(i8*, <4 x i16>, <4 x i16>, i32) nounwind readonly
 declare %struct.__neon_int32x2x2_t @llvm.arm.neon.vld2lane.v2i32(i8*, <2 x i32>, <2 x i32>, i32) nounwind readonly
 declare %struct.__neon_float32x2x2_t @llvm.arm.neon.vld2lane.v2f32(i8*, <2 x float>, <2 x float>, i32) nounwind readonly
 
+declare %struct.__neon_int16x8x2_t @llvm.arm.neon.vld2lane.v8i16(i8*, <8 x i16>, <8 x i16>, i32) nounwind readonly
+declare %struct.__neon_int32x4x2_t @llvm.arm.neon.vld2lane.v4i32(i8*, <4 x i32>, <4 x i32>, i32) nounwind readonly
+declare %struct.__neon_float32x4x2_t @llvm.arm.neon.vld2lane.v4f32(i8*, <4 x float>, <4 x float>, i32) nounwind readonly
+
 %struct.__neon_int8x8x3_t = type { <8 x i8>,  <8 x i8>,  <8 x i8> }
 %struct.__neon_int16x4x3_t = type { <4 x i16>, <4 x i16>, <4 x i16> }
 %struct.__neon_int32x2x3_t = type { <2 x i32>, <2 x i32>, <2 x i32> }