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 717826d00c97296cfc1382c243608eca9d1ab82f..fd5aac977b9951d284620bbd94f99d323a6209bd 100644 (file)
--- a/lib/Target/ARM/ARMISelDAGToDAG.cpp
+++ b/lib/Target/ARM/ARMISelDAGToDAG.cpp
@@ -1967,18 +1967,58 @@ 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 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: {

diff --git a/lib/Target/ARM/ARMInstrNEON.td b/lib/Target/ARM/ARMInstrNEON.td
index 1f178f3157dda5487def391a9672715aea470398..758e781b9767aa12a3bc2328acc9f7082c89ec74 100644 (file)
--- a/lib/Target/ARM/ARMInstrNEON.td
+++ b/lib/Target/ARM/ARMInstrNEON.td
@@ -469,16 +469,24 @@ def VST2LNq16b: VST2LN<0b0100, "vst2.16">;
 def VST2LNq32b: VST2LN<0b1000, "vst2.32">;
 
 //   VST3LN   : Vector Store (single 3-element structure from one lane)
-class VST3LND<bits<4> op11_8, string OpcodeStr>
+class VST3LN<bits<4> op11_8, string OpcodeStr>
   : NLdSt<1,0b00,op11_8,0b0000, (outs),
           (ins addrmode6:$addr, DPR:$src1, DPR:$src2, DPR:$src3,
            nohash_imm:$lane), IIC_VST,
           !strconcat(OpcodeStr,
           "\t\\{$src1[$lane],$src2[$lane],$src3[$lane]\\}, $addr"), "", []>;
 
-def VST3LNd8  : VST3LND<0b0010, "vst3.8">;
-def VST3LNd16 : VST3LND<0b0110, "vst3.16">;
-def VST3LNd32 : VST3LND<0b1010, "vst3.32">;
+def VST3LNd8  : VST3LN<0b0010, "vst3.8">;
+def VST3LNd16 : VST3LN<0b0110, "vst3.16">;
+def VST3LNd32 : VST3LN<0b1010, "vst3.32">;
+
+// vst3 to double-spaced even registers.
+def VST3LNq16a: VST3LN<0b0110, "vst3.16">;
+def VST3LNq32a: VST3LN<0b1010, "vst3.32">;
+
+// vst3 to double-spaced odd registers.
+def VST3LNq16b: VST3LN<0b0110, "vst3.16">;
+def VST3LNq32b: VST3LN<0b1010, "vst3.32">;
 
 //   VST4LN   : Vector Store (single 4-element structure from one lane)
 class VST4LND<bits<4> op11_8, string OpcodeStr>

diff --git a/lib/Target/ARM/NEONPreAllocPass.cpp b/lib/Target/ARM/NEONPreAllocPass.cpp
index b8475ff93146b6a4b1791761b8125c0997eaf1e0..4c3a8df33e11f112bee681b5a6a02603d47deb52 100644 (file)
--- a/lib/Target/ARM/NEONPreAllocPass.cpp
+++ b/lib/Target/ARM/NEONPreAllocPass.cpp
@@ -233,6 +233,22 @@ static bool isNEONMultiRegOp(int Opcode, unsigned &FirstOpnd, unsigned &NumRegs,
     Stride = 2;
     return true;
 
+  case ARM::VST3LNq16a:
+  case ARM::VST3LNq32a:
+    FirstOpnd = 3;
+    NumRegs = 3;
+    Offset = 0;
+    Stride = 2;
+    return true;
+
+  case ARM::VST3LNq16b:
+  case ARM::VST3LNq32b:
+    FirstOpnd = 3;
+    NumRegs = 3;
+    Offset = 1;
+    Stride = 2;
+    return true;
+
   case ARM::VST4d8:
   case ARM::VST4d16:
   case ARM::VST4d32:

diff --git a/test/CodeGen/ARM/vstlane.ll b/test/CodeGen/ARM/vstlane.ll
index 715f88de2af9ded804a8033cab0eab60681bb0e0..19067056f68bfcf5333ee3ec0fbb4bc53352e03e 100644 (file)
--- a/test/CodeGen/ARM/vstlane.ll
+++ b/test/CodeGen/ARM/vstlane.ll
@@ -97,11 +97,39 @@ define void @vst3lanef(float* %A, <2 x float>* %B) nounwind {
        ret void
 }
 
+define void @vst3laneQi16(i16* %A, <8 x i16>* %B) nounwind {
+;CHECK: vst3laneQi16:
+;CHECK: vst3.16
+       %tmp1 = load <8 x i16>* %B
+       call void @llvm.arm.neon.vst3lane.v8i16(i16* %A, <8 x i16> %tmp1, <8 x i16> %tmp1, <8 x i16> %tmp1, i32 6)
+       ret void
+}
+
+define void @vst3laneQi32(i32* %A, <4 x i32>* %B) nounwind {
+;CHECK: vst3laneQi32:
+;CHECK: vst3.32
+       %tmp1 = load <4 x i32>* %B
+       call void @llvm.arm.neon.vst3lane.v4i32(i32* %A, <4 x i32> %tmp1, <4 x i32> %tmp1, <4 x i32> %tmp1, i32 0)
+       ret void
+}
+
+define void @vst3laneQf(float* %A, <4 x float>* %B) nounwind {
+;CHECK: vst3laneQf:
+;CHECK: vst3.32
+       %tmp1 = load <4 x float>* %B
+       call void @llvm.arm.neon.vst3lane.v4f32(float* %A, <4 x float> %tmp1, <4 x float> %tmp1, <4 x float> %tmp1, i32 1)
+       ret void
+}
+
 declare void @llvm.arm.neon.vst3lane.v8i8(i8*, <8 x i8>, <8 x i8>, <8 x i8>, i32) nounwind
 declare void @llvm.arm.neon.vst3lane.v4i16(i8*, <4 x i16>, <4 x i16>, <4 x i16>, i32) nounwind
 declare void @llvm.arm.neon.vst3lane.v2i32(i8*, <2 x i32>, <2 x i32>, <2 x i32>, i32) nounwind
 declare void @llvm.arm.neon.vst3lane.v2f32(i8*, <2 x float>, <2 x float>, <2 x float>, i32) nounwind
 
+declare void @llvm.arm.neon.vst3lane.v8i16(i8*, <8 x i16>, <8 x i16>, <8 x i16>, i32) nounwind
+declare void @llvm.arm.neon.vst3lane.v4i32(i8*, <4 x i32>, <4 x i32>, <4 x i32>, i32) nounwind
+declare void @llvm.arm.neon.vst3lane.v4f32(i8*, <4 x float>, <4 x float>, <4 x float>, i32) nounwind
+
 
 define void @vst4lanei8(i8* %A, <8 x i8>* %B) nounwind {
 ;CHECK: vst4lanei8: