Add codegen support for NEON vst4lane intrinsics with 128-bit vectors.


git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@83600 91177308-0d34-0410-b5e6-96231b3b80d8
diff --git a/lib/Target/ARM/ARMISelDAGToDAG.cpp b/lib/Target/ARM/ARMISelDAGToDAG.cpp
index fd5aac9..be3f343 100644
--- a/lib/Target/ARM/ARMISelDAGToDAG.cpp
+++ b/lib/Target/ARM/ARMISelDAGToDAG.cpp
@@ -2025,19 +2025,61 @@
       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);
     }
     }
   }
diff --git a/lib/Target/ARM/ARMInstrNEON.td b/lib/Target/ARM/ARMInstrNEON.td
index 758e781..cd370aa 100644
--- a/lib/Target/ARM/ARMInstrNEON.td
+++ b/lib/Target/ARM/ARMInstrNEON.td
@@ -489,7 +489,7 @@
 def VST3LNq32b: VST3LN<0b1010, "vst3.32">;
 
 //   VST4LN   : Vector Store (single 4-element structure from one lane)
-class VST4LND<bits<4> op11_8, string OpcodeStr>
+class VST4LN<bits<4> op11_8, string OpcodeStr>
   : NLdSt<1,0b00,op11_8,0b0000, (outs),
           (ins addrmode6:$addr, DPR:$src1, DPR:$src2, DPR:$src3, DPR:$src4,
            nohash_imm:$lane), IIC_VST,
@@ -497,9 +497,18 @@
           "\t\\{$src1[$lane],$src2[$lane],$src3[$lane],$src4[$lane]\\}, $addr"),
           "", []>;
 
-def VST4LNd8  : VST4LND<0b0011, "vst4.8">;
-def VST4LNd16 : VST4LND<0b0111, "vst4.16">;
-def VST4LNd32 : VST4LND<0b1011, "vst4.32">;
+def VST4LNd8  : VST4LN<0b0011, "vst4.8">;
+def VST4LNd16 : VST4LN<0b0111, "vst4.16">;
+def VST4LNd32 : VST4LN<0b1011, "vst4.32">;
+
+// vst4 to double-spaced even registers.
+def VST4LNq16a: VST4LN<0b0111, "vst4.16">;
+def VST4LNq32a: VST4LN<0b1011, "vst4.32">;
+
+// vst4 to double-spaced odd registers.
+def VST4LNq16b: VST4LN<0b0111, "vst4.16">;
+def VST4LNq32b: VST4LN<0b1011, "vst4.32">;
+
 } // mayStore = 1, hasExtraSrcRegAllocReq = 1
 
 
diff --git a/lib/Target/ARM/NEONPreAllocPass.cpp b/lib/Target/ARM/NEONPreAllocPass.cpp
index 4c3a8df..821b872 100644
--- a/lib/Target/ARM/NEONPreAllocPass.cpp
+++ b/lib/Target/ARM/NEONPreAllocPass.cpp
@@ -278,6 +278,22 @@
     Stride = 2;
     return true;
 
+  case ARM::VST4LNq16a:
+  case ARM::VST4LNq32a:
+    FirstOpnd = 3;
+    NumRegs = 4;
+    Offset = 0;
+    Stride = 2;
+    return true;
+
+  case ARM::VST4LNq16b:
+  case ARM::VST4LNq32b:
+    FirstOpnd = 3;
+    NumRegs = 4;
+    Offset = 1;
+    Stride = 2;
+    return true;
+
   case ARM::VTBL2:
     FirstOpnd = 1;
     NumRegs = 2;
diff --git a/test/CodeGen/ARM/vstlane.ll b/test/CodeGen/ARM/vstlane.ll
index 1906705..3bfb14f 100644
--- a/test/CodeGen/ARM/vstlane.ll
+++ b/test/CodeGen/ARM/vstlane.ll
@@ -163,7 +163,35 @@
 	ret void
 }
 
+define void @vst4laneQi16(i16* %A, <8 x i16>* %B) nounwind {
+;CHECK: vst4laneQi16:
+;CHECK: vst4.16
+	%tmp1 = load <8 x i16>* %B
+	call void @llvm.arm.neon.vst4lane.v8i16(i16* %A, <8 x i16> %tmp1, <8 x i16> %tmp1, <8 x i16> %tmp1, <8 x i16> %tmp1, i32 7)
+	ret void
+}
+
+define void @vst4laneQi32(i32* %A, <4 x i32>* %B) nounwind {
+;CHECK: vst4laneQi32:
+;CHECK: vst4.32
+	%tmp1 = load <4 x i32>* %B
+	call void @llvm.arm.neon.vst4lane.v4i32(i32* %A, <4 x i32> %tmp1, <4 x i32> %tmp1, <4 x i32> %tmp1, <4 x i32> %tmp1, i32 2)
+	ret void
+}
+
+define void @vst4laneQf(float* %A, <4 x float>* %B) nounwind {
+;CHECK: vst4laneQf:
+;CHECK: vst4.32
+	%tmp1 = load <4 x float>* %B
+	call void @llvm.arm.neon.vst4lane.v4f32(float* %A, <4 x float> %tmp1, <4 x float> %tmp1, <4 x float> %tmp1, <4 x float> %tmp1, i32 1)
+	ret void
+}
+
 declare void @llvm.arm.neon.vst4lane.v8i8(i8*, <8 x i8>, <8 x i8>, <8 x i8>, <8 x i8>, i32) nounwind
 declare void @llvm.arm.neon.vst4lane.v4i16(i8*, <4 x i16>, <4 x i16>, <4 x i16>, <4 x i16>, i32) nounwind
 declare void @llvm.arm.neon.vst4lane.v2i32(i8*, <2 x i32>, <2 x i32>, <2 x i32>, <2 x i32>, i32) nounwind
 declare void @llvm.arm.neon.vst4lane.v2f32(i8*, <2 x float>, <2 x float>, <2 x float>, <2 x float>, i32) nounwind
+
+declare void @llvm.arm.neon.vst4lane.v8i16(i8*, <8 x i16>, <8 x i16>, <8 x i16>, <8 x i16>, i32) nounwind
+declare void @llvm.arm.neon.vst4lane.v4i32(i8*, <4 x i32>, <4 x i32>, <4 x i32>, <4 x i32>, i32) nounwind
+declare void @llvm.arm.neon.vst4lane.v4f32(i8*, <4 x float>, <4 x float>, <4 x float>, <4 x float>, i32) nounwind