[AArch64] Custom lower concat_vector patterns with v4i16, v4i32, v8i8, v8i16, v16i8 types.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@200491 91177308-0d34-0410-b5e6-96231b3b80d8

2 files changed, 67 insertions, 0 deletions

diff --git a/lib/Target/AArch64/AArch64ISelLowering.cpp b/lib/Target/AArch64/AArch64ISelLowering.cpp
index 7f101fffa1..769a2bcb62 100644
--- a/lib/Target/AArch64/AArch64ISelLowering.cpp
+++ b/lib/Target/AArch64/AArch64ISelLowering.cpp
@@ -332,6 +332,12 @@ AArch64TargetLowering::AArch64TargetLowering(AArch64TargetMachine &TM)
     setOperationAction(ISD::CONCAT_VECTORS, MVT::v4f32, Legal);
     setOperationAction(ISD::CONCAT_VECTORS, MVT::v2f64, Legal);
 
+    setOperationAction(ISD::CONCAT_VECTORS, MVT::v8i8, Custom);
+    setOperationAction(ISD::CONCAT_VECTORS, MVT::v4i16, Custom);
+    setOperationAction(ISD::CONCAT_VECTORS, MVT::v16i8, Custom);
+    setOperationAction(ISD::CONCAT_VECTORS, MVT::v8i16, Custom);
+    setOperationAction(ISD::CONCAT_VECTORS, MVT::v4i32, Custom);
+
     setOperationAction(ISD::SETCC, MVT::v8i8, Custom);
     setOperationAction(ISD::SETCC, MVT::v16i8, Custom);
     setOperationAction(ISD::SETCC, MVT::v4i16, Custom);
@@ -2259,6 +2265,52 @@ static SDValue LowerVectorFP_TO_INT(SDValue Op, SelectionDAG &DAG,
   return DAG.getNode(Opc, dl, VT, Vec);
 }
 
+static SDValue LowerCONCAT_VECTORS(SDValue Op, SelectionDAG &DAG) {
+  // We custom lower concat_vectors with 4, 8, or 16 operands that are all the
+  // same operand and of type v1* using the DUP instruction.
+  unsigned NumOps = Op->getNumOperands();
+  if (NumOps != 4 && NumOps != 8 && NumOps != 16)
+    return Op;
+
+  // Must be a single value for VDUP.
+  bool isConstant = true;
+  SDValue Op0 = Op.getOperand(0);
+  for (unsigned i = 1; i < NumOps; ++i) {
+    SDValue OpN = Op.getOperand(i);
+    if (Op0 != OpN)
+      return Op;
+
+    if (!isa<ConstantSDNode>(OpN->getOperand(0)))
+      isConstant = false;
+  }
+
+  // Verify the value type.
+  EVT EltVT = Op0.getValueType();
+  switch (NumOps) {
+  default: llvm_unreachable("Unexpected number of operands");
+  case 4:
+    if (EltVT != MVT::v1i16 && EltVT != MVT::v1i32)
+      return Op;
+    break;
+  case 8:
+    if (EltVT != MVT::v1i8 && EltVT != MVT::v1i16)
+      return Op;
+    break;
+  case 16:
+    if (EltVT != MVT::v1i8)
+      return Op;
+    break;
+  }
+
+  SDLoc DL(Op);
+  EVT VT = Op.getValueType();
+  // VDUP produces better code for constants.
+  if (isConstant)
+    return DAG.getNode(AArch64ISD::NEON_VDUP, DL, VT, Op0->getOperand(0));
+  return DAG.getNode(AArch64ISD::NEON_VDUPLANE, DL, VT, Op0,
+                     DAG.getConstant(0, MVT::i64));
+}
+
 SDValue
 AArch64TargetLowering::LowerFP_TO_INT(SDValue Op, SelectionDAG &DAG,
                                       bool IsSigned) const {
@@ -3241,6 +3293,7 @@ AArch64TargetLowering::LowerOperation(SDValue Op, SelectionDAG &DAG) const {
   case ISD::VASTART: return LowerVASTART(Op, DAG);
   case ISD::BUILD_VECTOR:
     return LowerBUILD_VECTOR(Op, DAG, getSubtarget());
+  case ISD::CONCAT_VECTORS: return LowerCONCAT_VECTORS(Op, DAG);
   case ISD::VECTOR_SHUFFLE: return LowerVECTOR_SHUFFLE(Op, DAG);
   }
 
diff --git a/lib/Target/AArch64/AArch64InstrNEON.td b/lib/Target/AArch64/AArch64InstrNEON.td
index 1309bf12b5..2cf27b861b 100644
--- a/lib/Target/AArch64/AArch64InstrNEON.td
+++ b/lib/Target/AArch64/AArch64InstrNEON.td
@@ -7106,6 +7106,20 @@ def : Pat<(v2f64 (Neon_vdup (f64 FPR64:$Rn))),
             (SUBREG_TO_REG (i64 0), FPR64:$Rn, sub_64),
             (i64 0)))>;
 
+multiclass NeonI_DUP_pattern<Instruction DUPELT, ValueType ResTy,
+                             ValueType OpTy, RegisterClass OpRC,
+                             Operand OpNImm, SubRegIndex SubIndex> {
+def : Pat<(ResTy (Neon_vduplane (OpTy OpRC:$Rn), OpNImm:$Imm)),
+          (ResTy (DUPELT
+            (SUBREG_TO_REG (i64 0), OpRC:$Rn, SubIndex), OpNImm:$Imm))>;
+}
+
+defm : NeonI_DUP_pattern<DUPELT4h, v4i16, v1i16, FPR16, neon_uimm2_bare,sub_16>;
+defm : NeonI_DUP_pattern<DUPELT4s, v4i32, v1i32, FPR32, neon_uimm2_bare,sub_32>;
+defm : NeonI_DUP_pattern<DUPELT8b, v8i8, v1i8, FPR8, neon_uimm3_bare, sub_8>;
+defm : NeonI_DUP_pattern<DUPELT8h, v8i16, v1i16, FPR16, neon_uimm3_bare,sub_16>;
+defm : NeonI_DUP_pattern<DUPELT16b, v16i8, v1i8, FPR8, neon_uimm4_bare, sub_8>;
+
 class NeonI_DUP<bit Q, string asmop, string rdlane,
                 RegisterOperand ResVPR, ValueType ResTy,
                 RegisterClass OpGPR, ValueType OpTy>