vendor/llvm/llvm-trunk-r308421

14 files changed, 163 insertions, 177 deletions

diff --git a/lib/Target/Hexagon/HexagonBitSimplify.cpp b/lib/Target/Hexagon/HexagonBitSimplify.cpp
index b064778c4bbd..d75d95a6baea 100644
--- a/lib/Target/Hexagon/HexagonBitSimplify.cpp
+++ b/lib/Target/Hexagon/HexagonBitSimplify.cpp
@@ -7,8 +7,6 @@
 //
 //===----------------------------------------------------------------------===//
 
-#define DEBUG_TYPE "hexbit"
-
 #include "HexagonBitTracker.h"
 #include "HexagonTargetMachine.h"
 #include "llvm/ADT/BitVector.h"
@@ -42,6 +40,8 @@
 #include <utility>
 #include <vector>
 
+#define DEBUG_TYPE "hexbit"
+
 using namespace llvm;
 
 static cl::opt<bool> PreserveTiedOps("hexbit-keep-tied", cl::Hidden,
diff --git a/lib/Target/Hexagon/HexagonDepInstrInfo.td b/lib/Target/Hexagon/HexagonDepInstrInfo.td
index 2dc74632e9be..30ebf89c9808 100644
--- a/lib/Target/Hexagon/HexagonDepInstrInfo.td
+++ b/lib/Target/Hexagon/HexagonDepInstrInfo.td
@@ -45863,6 +45863,7 @@ tc_30665cb0, TypeST>, Enc_ecbcc8 {
 let Inst{13-0} = 0b00000000000000;
 let Inst{31-21} = 0b10100000000;
 let isSoloAin1 = 1;
+let hasSideEffects = 1;
 }
 def Y2_dccleaninva : HInst<
 (outs),
@@ -45872,6 +45873,7 @@ tc_30665cb0, TypeST>, Enc_ecbcc8 {
 let Inst{13-0} = 0b00000000000000;
 let Inst{31-21} = 0b10100000010;
 let isSoloAin1 = 1;
+let hasSideEffects = 1;
 }
 def Y2_dcfetch : HInst<
 (outs),
@@ -45900,6 +45902,7 @@ tc_30665cb0, TypeST>, Enc_ecbcc8 {
 let Inst{13-0} = 0b00000000000000;
 let Inst{31-21} = 0b10100000001;
 let isSoloAin1 = 1;
+let hasSideEffects = 1;
 }
 def Y2_dczeroa : HInst<
 (outs),
@@ -45909,6 +45912,7 @@ tc_30665cb0, TypeST>, Enc_ecbcc8 {
 let Inst{13-0} = 0b00000000000000;
 let Inst{31-21} = 0b10100000110;
 let isSoloAin1 = 1;
+let hasSideEffects = 1;
 let mayStore = 1;
 }
 def Y2_icinva : HInst<
diff --git a/lib/Target/Hexagon/HexagonEarlyIfConv.cpp b/lib/Target/Hexagon/HexagonEarlyIfConv.cpp
index 03c4a83594b3..80361015e649 100644
--- a/lib/Target/Hexagon/HexagonEarlyIfConv.cpp
+++ b/lib/Target/Hexagon/HexagonEarlyIfConv.cpp
@@ -59,8 +59,6 @@
 //         J2_jump <BB#6>, %PC<imp-def,dead>
 //     Successors according to CFG: BB#6 BB#3
 
-#define DEBUG_TYPE "hexagon-eif"
-
 #include "Hexagon.h"
 #include "HexagonInstrInfo.h"
 #include "HexagonSubtarget.h"
@@ -90,6 +88,8 @@
 #include <cassert>
 #include <iterator>
 
+#define DEBUG_TYPE "hexagon-eif"
+
 using namespace llvm;
 
 namespace llvm {
diff --git a/lib/Target/Hexagon/HexagonExpandCondsets.cpp b/lib/Target/Hexagon/HexagonExpandCondsets.cpp
index 734f3c6658d9..a2f6dd68c1a1 100644
--- a/lib/Target/Hexagon/HexagonExpandCondsets.cpp
+++ b/lib/Target/Hexagon/HexagonExpandCondsets.cpp
@@ -86,8 +86,6 @@
 // however, is that finding the locations where the implicit uses need
 // to be added, and updating the live ranges will be more involved.
 
-#define DEBUG_TYPE "expand-condsets"
-
 #include "HexagonInstrInfo.h"
 #include "llvm/ADT/DenseMap.h"
 #include "llvm/ADT/SetVector.h"
@@ -116,6 +114,8 @@
 #include <set>
 #include <utility>
 
+#define DEBUG_TYPE "expand-condsets"
+
 using namespace llvm;
 
 static cl::opt<unsigned> OptTfrLimit("expand-condsets-tfr-limit",
diff --git a/lib/Target/Hexagon/HexagonFrameLowering.cpp b/lib/Target/Hexagon/HexagonFrameLowering.cpp
index c790579ccebc..e5e75198b2d1 100644
--- a/lib/Target/Hexagon/HexagonFrameLowering.cpp
+++ b/lib/Target/Hexagon/HexagonFrameLowering.cpp
@@ -8,8 +8,6 @@
 //
 //===----------------------------------------------------------------------===//
 
-#define DEBUG_TYPE "hexagon-pei"
-
 #include "HexagonFrameLowering.h"
 #include "HexagonBlockRanges.h"
 #include "HexagonInstrInfo.h"
@@ -63,6 +61,8 @@
 #include <utility>
 #include <vector>
 
+#define DEBUG_TYPE "hexagon-pei"
+
 // Hexagon stack frame layout as defined by the ABI:
 //
 //                                                       Incoming arguments
diff --git a/lib/Target/Hexagon/HexagonGenInsert.cpp b/lib/Target/Hexagon/HexagonGenInsert.cpp
index bf31e1699284..0a955aedaf1a 100644
--- a/lib/Target/Hexagon/HexagonGenInsert.cpp
+++ b/lib/Target/Hexagon/HexagonGenInsert.cpp
@@ -7,8 +7,6 @@
 //
 //===----------------------------------------------------------------------===//
 
-#define DEBUG_TYPE "hexinsert"
-
 #include "BitTracker.h"
 #include "HexagonBitTracker.h"
 #include "HexagonInstrInfo.h"
@@ -44,6 +42,8 @@
 #include <utility>
 #include <vector>
 
+#define DEBUG_TYPE "hexinsert"
+
 using namespace llvm;
 
 static cl::opt<unsigned> VRegIndexCutoff("insert-vreg-cutoff", cl::init(~0U),
diff --git a/lib/Target/Hexagon/HexagonGenPredicate.cpp b/lib/Target/Hexagon/HexagonGenPredicate.cpp
index 3470480d607d..2da211563e0a 100644
--- a/lib/Target/Hexagon/HexagonGenPredicate.cpp
+++ b/lib/Target/Hexagon/HexagonGenPredicate.cpp
@@ -7,8 +7,6 @@
 //
 //===----------------------------------------------------------------------===//
 
-#define DEBUG_TYPE "gen-pred"
-
 #include "HexagonInstrInfo.h"
 #include "HexagonSubtarget.h"
 #include "llvm/ADT/SetVector.h"
@@ -35,6 +33,8 @@
 #include <set>
 #include <utility>
 
+#define DEBUG_TYPE "gen-pred"
+
 using namespace llvm;
 
 namespace llvm {
diff --git a/lib/Target/Hexagon/HexagonISelLowering.cpp b/lib/Target/Hexagon/HexagonISelLowering.cpp
index 67242764d453..3997702bc962 100644
--- a/lib/Target/Hexagon/HexagonISelLowering.cpp
+++ b/lib/Target/Hexagon/HexagonISelLowering.cpp
@@ -1286,16 +1286,6 @@ HexagonTargetLowering::LowerVASTART(SDValue Op, SelectionDAG &DAG) const {
                       MachinePointerInfo(SV));
 }
 
-// Creates a SPLAT instruction for a constant value VAL.
-static SDValue createSplat(SelectionDAG &DAG, const SDLoc &dl, EVT VT,
-                           SDValue Val) {
-  EVT T = VT.getVectorElementType();
-  if (T == MVT::i8 || T == MVT::i16)
-    return DAG.getNode(HexagonISD::VSPLAT, dl, VT, Val);
-
-  return SDValue();
-}
-
 static bool isSExtFree(SDValue N) {
   // A sign-extend of a truncate of a sign-extend is free.
   if (N.getOpcode() == ISD::TRUNCATE &&
@@ -1374,79 +1364,6 @@ HexagonTargetLowering::LowerVSELECT(SDValue Op, SelectionDAG &DAG) const {
   return SDValue();
 }
 
-// Handle only specific vector loads.
-SDValue HexagonTargetLowering::LowerLOAD(SDValue Op, SelectionDAG &DAG) const {
-  EVT VT = Op.getValueType();
-  SDLoc DL(Op);
-  LoadSDNode *LoadNode = cast<LoadSDNode>(Op);
-  SDValue Chain = LoadNode->getChain();
-  SDValue Ptr = Op.getOperand(1);
-  SDValue LoweredLoad;
-  SDValue Result;
-  SDValue Base = LoadNode->getBasePtr();
-  ISD::LoadExtType Ext = LoadNode->getExtensionType();
-  unsigned Alignment = LoadNode->getAlignment();
-  SDValue LoadChain;
-
-  if(Ext == ISD::NON_EXTLOAD)
-    Ext = ISD::ZEXTLOAD;
-
-  if (VT == MVT::v4i16) {
-    if (Alignment == 2) {
-      SDValue Loads[4];
-      // Base load.
-      Loads[0] = DAG.getExtLoad(Ext, DL, MVT::i32, Chain, Base,
-                                LoadNode->getPointerInfo(), MVT::i16, Alignment,
-                                LoadNode->getMemOperand()->getFlags());
-      // Base+2 load.
-      SDValue Increment = DAG.getConstant(2, DL, MVT::i32);
-      Ptr = DAG.getNode(ISD::ADD, DL, Base.getValueType(), Base, Increment);
-      Loads[1] = DAG.getExtLoad(Ext, DL, MVT::i32, Chain, Ptr,
-                                LoadNode->getPointerInfo(), MVT::i16, Alignment,
-                                LoadNode->getMemOperand()->getFlags());
-      // SHL 16, then OR base and base+2.
-      SDValue ShiftAmount = DAG.getConstant(16, DL, MVT::i32);
-      SDValue Tmp1 = DAG.getNode(ISD::SHL, DL, MVT::i32, Loads[1], ShiftAmount);
-      SDValue Tmp2 = DAG.getNode(ISD::OR, DL, MVT::i32, Tmp1, Loads[0]);
-      // Base + 4.
-      Increment = DAG.getConstant(4, DL, MVT::i32);
-      Ptr = DAG.getNode(ISD::ADD, DL, Base.getValueType(), Base, Increment);
-      Loads[2] = DAG.getExtLoad(Ext, DL, MVT::i32, Chain, Ptr,
-                                LoadNode->getPointerInfo(), MVT::i16, Alignment,
-                                LoadNode->getMemOperand()->getFlags());
-      // Base + 6.
-      Increment = DAG.getConstant(6, DL, MVT::i32);
-      Ptr = DAG.getNode(ISD::ADD, DL, Base.getValueType(), Base, Increment);
-      Loads[3] = DAG.getExtLoad(Ext, DL, MVT::i32, Chain, Ptr,
-                                LoadNode->getPointerInfo(), MVT::i16, Alignment,
-                                LoadNode->getMemOperand()->getFlags());
-      // SHL 16, then OR base+4 and base+6.
-      Tmp1 = DAG.getNode(ISD::SHL, DL, MVT::i32, Loads[3], ShiftAmount);
-      SDValue Tmp4 = DAG.getNode(ISD::OR, DL, MVT::i32, Tmp1, Loads[2]);
-      // Combine to i64. This could be optimised out later if we can
-      // affect reg allocation of this code.
-      Result = DAG.getNode(HexagonISD::COMBINE, DL, MVT::i64, Tmp4, Tmp2);
-      LoadChain = DAG.getNode(ISD::TokenFactor, DL, MVT::Other,
-                              Loads[0].getValue(1), Loads[1].getValue(1),
-                              Loads[2].getValue(1), Loads[3].getValue(1));
-    } else {
-      // Perform default type expansion.
-      Result = DAG.getLoad(MVT::i64, DL, Chain, Ptr, LoadNode->getPointerInfo(),
-                           LoadNode->getAlignment(),
-                           LoadNode->getMemOperand()->getFlags());
-      LoadChain = Result.getValue(1);
-    }
-  } else
-    llvm_unreachable("Custom lowering unsupported load");
-
-  Result = DAG.getNode(ISD::BITCAST, DL, VT, Result);
-  // Since we pretend to lower a load, we need the original chain
-  // info attached to the result.
-  SDValue Ops[] = { Result, LoadChain };
-
-  return DAG.getMergeValues(Ops, DL);
-}
-
 SDValue
 HexagonTargetLowering::LowerConstantPool(SDValue Op, SelectionDAG &DAG) const {
   EVT ValTy = Op.getValueType();
@@ -1971,18 +1888,12 @@ HexagonTargetLowering::HexagonTargetLowering(const TargetMachine &TM,
   // Handling of vector operations.
   //
 
-  // Custom lower v4i16 load only. Let v4i16 store to be
-  // promoted for now.
   promoteLdStType(MVT::v4i8,  MVT::i32);
   promoteLdStType(MVT::v2i16, MVT::i32);
   promoteLdStType(MVT::v8i8,  MVT::i64);
+  promoteLdStType(MVT::v4i16, MVT::i64);
   promoteLdStType(MVT::v2i32, MVT::i64);
 
-  setOperationAction(ISD::LOAD,  MVT::v4i16, Custom);
-  setOperationAction(ISD::STORE, MVT::v4i16, Promote);
-  AddPromotedToType(ISD::LOAD,  MVT::v4i16, MVT::i64);
-  AddPromotedToType(ISD::STORE, MVT::v4i16, MVT::i64);
-
   // Set the action for vector operations to "expand", then override it with
   // either "custom" or "legal" for specific cases.
   static const unsigned VectExpOps[] = {
@@ -2301,7 +2212,8 @@ const char* HexagonTargetLowering::getTargetNodeName(unsigned Opcode) const {
   case HexagonISD::RET_FLAG:      return "HexagonISD::RET_FLAG";
   case HexagonISD::TC_RETURN:     return "HexagonISD::TC_RETURN";
   case HexagonISD::VCOMBINE:      return "HexagonISD::VCOMBINE";
-  case HexagonISD::VPACK:         return "HexagonISD::VPACK";
+  case HexagonISD::VPACKE:        return "HexagonISD::VPACKE";
+  case HexagonISD::VPACKO:        return "HexagonISD::VPACKO";
   case HexagonISD::VASL:          return "HexagonISD::VASL";
   case HexagonISD::VASR:          return "HexagonISD::VASR";
   case HexagonISD::VLSR:          return "HexagonISD::VLSR";
@@ -2394,7 +2306,7 @@ HexagonTargetLowering::LowerVECTOR_SHUFFLE(SDValue Op, SelectionDAG &DAG)
 
     // Test if V1 is a SCALAR_TO_VECTOR.
     if (Lane == 0 && V1.getOpcode() == ISD::SCALAR_TO_VECTOR)
-      return createSplat(DAG, dl, VT, V1.getOperand(0));
+      return DAG.getNode(HexagonISD::VSPLAT, dl, VT, V1.getOperand(0));
 
     // Test if V1 is a BUILD_VECTOR which is equivalent to a SCALAR_TO_VECTOR
     // (and probably will turn into a SCALAR_TO_VECTOR once legalization
@@ -2409,28 +2321,26 @@ HexagonTargetLowering::LowerVECTOR_SHUFFLE(SDValue Op, SelectionDAG &DAG)
         }
       }
       if (IsScalarToVector)
-        return createSplat(DAG, dl, VT, V1.getOperand(0));
+        return DAG.getNode(HexagonISD::VSPLAT, dl, VT, V1.getOperand(0));
     }
-    return createSplat(DAG, dl, VT, DAG.getConstant(Lane, dl, MVT::i32));
+    return DAG.getNode(HexagonISD::VSPLAT, dl, VT,
+                       DAG.getConstant(Lane, dl, MVT::i32));
   }
 
   if (UseHVX) {
     ArrayRef<int> Mask = SVN->getMask();
     size_t MaskLen = Mask.size();
-    int ElemSizeInBits = VT.getScalarSizeInBits();
-    if ((Subtarget.useHVXSglOps() && (ElemSizeInBits * MaskLen) == 64 * 8) ||
-        (Subtarget.useHVXDblOps() && (ElemSizeInBits * MaskLen) == 128 * 8)) {
-      // Return 1 for odd and 2 of even
-      StridedLoadKind Pattern = isStridedLoad(Mask);
+    unsigned SizeInBits = VT.getScalarSizeInBits() * MaskLen;
 
+    if ((Subtarget.useHVXSglOps() && SizeInBits == 64 * 8) ||
+        (Subtarget.useHVXDblOps() && SizeInBits == 128 * 8)) {
+      StridedLoadKind Pattern = isStridedLoad(Mask);
       if (Pattern == StridedLoadKind::NoPattern)
         return SDValue();
 
-      SDValue Vec0 = Op.getOperand(0);
-      SDValue Vec1 = Op.getOperand(1);
-      SDValue StridePattern = DAG.getConstant(Pattern, dl, MVT::i32);
-      SDValue Ops[] = { Vec1, Vec0, StridePattern };
-      return DAG.getNode(HexagonISD::VPACK, dl, VT, Ops);
+      unsigned Opc = Pattern == StridedLoadKind::Even ? HexagonISD::VPACKE
+                                                      : HexagonISD::VPACKO;
+      return DAG.getNode(Opc, dl, VT, {Op.getOperand(1), Op.getOperand(0)});
     }
     // We used to assert in the "else" part here, but that is bad for Halide
     // Halide creates intermediate double registers by interleaving two
@@ -2531,19 +2441,26 @@ HexagonTargetLowering::LowerBUILD_VECTOR(SDValue Op, SelectionDAG &DAG) const {
   if (Size > 64)
     return SDValue();
 
-  APInt APSplatBits, APSplatUndef;
-  unsigned SplatBitSize;
-  bool HasAnyUndefs;
   unsigned NElts = BVN->getNumOperands();
 
   // Try to generate a SPLAT instruction.
-  if ((VT.getSimpleVT() == MVT::v4i8 || VT.getSimpleVT() == MVT::v4i16) &&
-      (BVN->isConstantSplat(APSplatBits, APSplatUndef, SplatBitSize,
-                            HasAnyUndefs, 0, true) && SplatBitSize <= 16)) {
-    unsigned SplatBits = APSplatBits.getZExtValue();
-    int32_t SextVal = ((int32_t) (SplatBits << (32 - SplatBitSize)) >>
-                       (32 - SplatBitSize));
-    return createSplat(DAG, dl, VT, DAG.getConstant(SextVal, dl, MVT::i32));
+  if (VT == MVT::v4i8 || VT == MVT::v4i16 || VT == MVT::v2i32) {
+    APInt APSplatBits, APSplatUndef;
+    unsigned SplatBitSize;
+    bool HasAnyUndefs;
+    if (BVN->isConstantSplat(APSplatBits, APSplatUndef, SplatBitSize,
+                             HasAnyUndefs, 0, false)) {
+      if (SplatBitSize == VT.getVectorElementType().getSizeInBits()) {
+        unsigned ZV = APSplatBits.getZExtValue();
+        assert(SplatBitSize <= 32 && "Can only handle up to i32");
+        // Sign-extend the splat value from SplatBitSize to 32.
+        int32_t SV = SplatBitSize < 32
+                        ? int32_t(ZV << (32-SplatBitSize)) >> (32-SplatBitSize)
+                        : int32_t(ZV);
+        return DAG.getNode(HexagonISD::VSPLAT, dl, VT,
+                           DAG.getConstant(SV, dl, MVT::i32));
+      }
+    }
   }
 
   // Try to generate COMBINE to build v2i32 vectors.
@@ -2974,8 +2891,6 @@ HexagonTargetLowering::LowerOperation(SDValue Op, SelectionDAG &DAG) const {
     case ISD::BlockAddress:         return LowerBlockAddress(Op, DAG);
     case ISD::GLOBAL_OFFSET_TABLE:  return LowerGLOBAL_OFFSET_TABLE(Op, DAG);
     case ISD::VASTART:              return LowerVASTART(Op, DAG);
-    // Custom lower some vector loads.
-    case ISD::LOAD:                 return LowerLOAD(Op, DAG);
     case ISD::DYNAMIC_STACKALLOC:   return LowerDYNAMIC_STACKALLOC(Op, DAG);
     case ISD::SETCC:                return LowerSETCC(Op, DAG);
     case ISD::VSELECT:              return LowerVSELECT(Op, DAG);
diff --git a/lib/Target/Hexagon/HexagonISelLowering.h b/lib/Target/Hexagon/HexagonISelLowering.h
index bfd2c94eeaba..d66cbc95e918 100644
--- a/lib/Target/Hexagon/HexagonISelLowering.h
+++ b/lib/Target/Hexagon/HexagonISelLowering.h
@@ -62,7 +62,8 @@ namespace HexagonISD {
       EXTRACTU,
       EXTRACTURP,
       VCOMBINE,
-      VPACK,
+      VPACKE,
+      VPACKO,
       TC_RETURN,
       EH_RETURN,
       DCFETCH,
@@ -164,7 +165,6 @@ namespace HexagonISD {
     SDValue LowerFRAMEADDR(SDValue Op, SelectionDAG &DAG) const;
     SDValue LowerATOMIC_FENCE(SDValue Op, SelectionDAG& DAG) const;
     SDValue LowerRETURNADDR(SDValue Op, SelectionDAG &DAG) const;
-    SDValue LowerLOAD(SDValue Op, SelectionDAG &DAG) const;
 
     bool CanLowerReturn(CallingConv::ID CallConv,
                         MachineFunction &MF, bool isVarArg,
diff --git a/lib/Target/Hexagon/HexagonIntrinsics.td b/lib/Target/Hexagon/HexagonIntrinsics.td
index c611857ec26a..104a28654dd5 100644
--- a/lib/Target/Hexagon/HexagonIntrinsics.td
+++ b/lib/Target/Hexagon/HexagonIntrinsics.td
@@ -1366,6 +1366,18 @@ defm : MaskedStore <V6_vS32b_nqpred_ai, int_hexagon_V6_vmaskedstorenq>;
 defm : MaskedStore <V6_vS32b_nt_qpred_ai, int_hexagon_V6_vmaskedstorentq>;
 defm : MaskedStore <V6_vS32b_nt_nqpred_ai, int_hexagon_V6_vmaskedstorentnq>;
 
+//*******************************************************************
+//           SYSTEM
+//*******************************************************************
+
+def: T_R_pat<Y2_dccleana,    int_hexagon_Y2_dccleana>;
+def: T_R_pat<Y2_dccleaninva, int_hexagon_Y2_dccleaninva>;
+def: T_R_pat<Y2_dcinva,      int_hexagon_Y2_dcinva>;
+def: T_R_pat<Y2_dczeroa,     int_hexagon_Y2_dczeroa>;
+
+def: T_RR_pat<Y4_l2fetch,    int_hexagon_Y4_l2fetch>;
+def: T_RP_pat<Y5_l2fetch,    int_hexagon_Y5_l2fetch>;
+
 include "HexagonIntrinsicsV3.td"
 include "HexagonIntrinsicsV4.td"
 include "HexagonIntrinsicsV5.td"
diff --git a/lib/Target/Hexagon/HexagonOptAddrMode.cpp b/lib/Target/Hexagon/HexagonOptAddrMode.cpp
index a331c978f59d..374ffa3799b0 100644
--- a/lib/Target/Hexagon/HexagonOptAddrMode.cpp
+++ b/lib/Target/Hexagon/HexagonOptAddrMode.cpp
@@ -10,8 +10,6 @@
 // load/store instructions.
 //===----------------------------------------------------------------------===//
 
-#define DEBUG_TYPE "opt-addr-mode"
-
 #include "HexagonInstrInfo.h"
 #include "HexagonSubtarget.h"
 #include "MCTargetDesc/HexagonBaseInfo.h"
@@ -36,6 +34,8 @@
 #include <cassert>
 #include <cstdint>
 
+#define DEBUG_TYPE "opt-addr-mode"
+
 static cl::opt<int> CodeGrowthLimit("hexagon-amode-growth-limit",
   cl::Hidden, cl::init(0), cl::desc("Code growth limit for address mode "
   "optimization"));
diff --git a/lib/Target/Hexagon/HexagonPatterns.td b/lib/Target/Hexagon/HexagonPatterns.td
index ba98b8994937..804a547d5b33 100644
--- a/lib/Target/Hexagon/HexagonPatterns.td
+++ b/lib/Target/Hexagon/HexagonPatterns.td
@@ -2250,6 +2250,12 @@ def: Storea_pat<SwapSt<atomic_store_16>, I32, addrgp, PS_storerhabs>;
 def: Storea_pat<SwapSt<atomic_store_32>, I32, addrgp, PS_storeriabs>;
 def: Storea_pat<SwapSt<atomic_store_64>, I64, addrgp, PS_storerdabs>;
 
+// Prefer this pattern to S2_asl_i_p_or for the special case of joining
+// two 32-bit words into a 64-bit word.
+let AddedComplexity = 200 in
+def: Pat<(or (shl (Aext64 I32:$a), (i32 32)), (Zext64 I32:$b)),
+         (A2_combinew I32:$a, I32:$b)>;
+
 def: Pat<(or (or (or (shl (i64 (zext (and I32:$b, (i32 65535)))), (i32 16)),
                      (i64 (zext (i32 (and I32:$a, (i32 65535)))))),
                  (shl (i64 (anyext (and I32:$c, (i32 65535)))), (i32 32))),
@@ -2971,45 +2977,40 @@ def: Pat<(v64i32 (HexagonVCOMBINE (v32i32 VecDblRegs:$Vs),
          (V6_vcombine_128B VecDblRegs:$Vs, VecDblRegs:$Vt)>,
          Requires<[UseHVXDbl]>;
 
-def SDTHexagonVPACK: SDTypeProfile<1, 3, [SDTCisSameAs<1, 2>,
-                                          SDTCisInt<3>]>;
-
-def HexagonVPACK: SDNode<"HexagonISD::VPACK", SDTHexagonVPACK>;
-
-// 0 as the last argument denotes vpacke. 1 denotes vpacko
-def: Pat<(v64i8 (HexagonVPACK (v64i8 VectorRegs:$Vs),
-                              (v64i8 VectorRegs:$Vt), (i32 0))),
-         (V6_vpackeb VectorRegs:$Vs, VectorRegs:$Vt)>,
-         Requires<[UseHVXSgl]>;
-def: Pat<(v64i8 (HexagonVPACK (v64i8 VectorRegs:$Vs),
-                              (v64i8 VectorRegs:$Vt), (i32 1))),
-         (V6_vpackob VectorRegs:$Vs, VectorRegs:$Vt)>,
-         Requires<[UseHVXSgl]>;
-def: Pat<(v32i16 (HexagonVPACK (v32i16 VectorRegs:$Vs),
-                               (v32i16 VectorRegs:$Vt), (i32 0))),
-         (V6_vpackeh VectorRegs:$Vs, VectorRegs:$Vt)>,
-         Requires<[UseHVXSgl]>;
-def: Pat<(v32i16 (HexagonVPACK (v32i16 VectorRegs:$Vs),
-                             (v32i16 VectorRegs:$Vt), (i32 1))),
-         (V6_vpackoh VectorRegs:$Vs, VectorRegs:$Vt)>,
-         Requires<[UseHVXSgl]>;
-
-def: Pat<(v128i8 (HexagonVPACK (v128i8 VecDblRegs:$Vs),
-                             (v128i8 VecDblRegs:$Vt), (i32 0))),
-         (V6_vpackeb_128B VecDblRegs:$Vs, VecDblRegs:$Vt)>,
-         Requires<[UseHVXDbl]>;
-def: Pat<(v128i8 (HexagonVPACK (v128i8 VecDblRegs:$Vs),
-                             (v128i8 VecDblRegs:$Vt), (i32 1))),
-         (V6_vpackob_128B VecDblRegs:$Vs, VecDblRegs:$Vt)>,
-         Requires<[UseHVXDbl]>;
-def: Pat<(v64i16 (HexagonVPACK (v64i16 VecDblRegs:$Vs),
-                             (v64i16 VecDblRegs:$Vt), (i32 0))),
-         (V6_vpackeh_128B VecDblRegs:$Vs, VecDblRegs:$Vt)>,
-         Requires<[UseHVXDbl]>;
-def: Pat<(v64i16 (HexagonVPACK (v64i16 VecDblRegs:$Vs),
-                            (v64i16 VecDblRegs:$Vt), (i32 1))),
-        (V6_vpackoh_128B VecDblRegs:$Vs, VecDblRegs:$Vt)>,
-        Requires<[UseHVXDbl]>;
+def SDTHexagonVPACK: SDTypeProfile<1, 2, [SDTCisSameAs<1, 2>, SDTCisVec<1>]>;
+
+def HexagonVPACKE: SDNode<"HexagonISD::VPACKE", SDTHexagonVPACK>;
+def HexagonVPACKO: SDNode<"HexagonISD::VPACKO", SDTHexagonVPACK>;
+
+let Predicates = [UseHVXSgl] in {
+  def: Pat<(v64i8 (HexagonVPACKE (v64i8 VectorRegs:$Vs),
+                                 (v64i8 VectorRegs:$Vt))),
+           (V6_vpackeb VectorRegs:$Vs, VectorRegs:$Vt)>;
+  def: Pat<(v64i8 (HexagonVPACKO (v64i8 VectorRegs:$Vs),
+                                 (v64i8 VectorRegs:$Vt))),
+           (V6_vpackob VectorRegs:$Vs, VectorRegs:$Vt)>;
+  def: Pat<(v32i16 (HexagonVPACKE (v32i16 VectorRegs:$Vs),
+                                  (v32i16 VectorRegs:$Vt))),
+           (V6_vpackeh VectorRegs:$Vs, VectorRegs:$Vt)>;
+  def: Pat<(v32i16 (HexagonVPACKO (v32i16 VectorRegs:$Vs),
+                                  (v32i16 VectorRegs:$Vt))),
+           (V6_vpackoh VectorRegs:$Vs, VectorRegs:$Vt)>;
+}
+
+let Predicates = [UseHVXDbl] in {
+  def: Pat<(v128i8 (HexagonVPACKE (v128i8 VecDblRegs:$Vs),
+                                  (v128i8 VecDblRegs:$Vt))),
+           (V6_vpackeb_128B VecDblRegs:$Vs, VecDblRegs:$Vt)>;
+  def: Pat<(v128i8 (HexagonVPACKO (v128i8 VecDblRegs:$Vs),
+                                  (v128i8 VecDblRegs:$Vt))),
+           (V6_vpackob_128B VecDblRegs:$Vs, VecDblRegs:$Vt)>;
+  def: Pat<(v64i16 (HexagonVPACKE (v64i16 VecDblRegs:$Vs),
+                                  (v64i16 VecDblRegs:$Vt))),
+           (V6_vpackeh_128B VecDblRegs:$Vs, VecDblRegs:$Vt)>;
+  def: Pat<(v64i16 (HexagonVPACKO (v64i16 VecDblRegs:$Vs),
+                                  (v64i16 VecDblRegs:$Vt))),
+          (V6_vpackoh_128B VecDblRegs:$Vs, VecDblRegs:$Vt)>;
+}
 
 def V2I1:  PatLeaf<(v2i1  PredRegs:$R)>;
 def V4I1:  PatLeaf<(v4i1  PredRegs:$R)>;
@@ -3073,6 +3074,10 @@ def: Pat<(v4i8 (HexagonVSPLAT I32:$Rs)), (S2_vsplatrb I32:$Rs)>;
 // four halfwords of 64-bits destination register.
 def: Pat<(v4i16 (HexagonVSPLAT I32:$Rs)), (S2_vsplatrh I32:$Rs)>;
 
+def: Pat<(v2i32 (HexagonVSPLAT s8_0ImmPred:$s8)),
+         (A2_combineii imm:$s8, imm:$s8)>;
+def: Pat<(v2i32 (HexagonVSPLAT I32:$Rs)), (A2_combinew I32:$Rs, I32:$Rs)>;
+
 
 class VArith_pat <InstHexagon MI, SDNode Op, PatFrag Type>
   : Pat <(Op Type:$Rss, Type:$Rtt),
@@ -3099,14 +3104,11 @@ def: VArith_pat <A2_xorp,   xor, V8I8>;
 def: VArith_pat <A2_xorp,   xor, V4I16>;
 def: VArith_pat <A2_xorp,   xor, V2I32>;
 
-def: Pat<(v2i32 (sra V2I32:$b, (i64 (HexagonCOMBINE (i32 u5_0ImmPred:$c),
-                                                    (i32 u5_0ImmPred:$c))))),
+def: Pat<(v2i32 (sra V2I32:$b, (v2i32 (HexagonVSPLAT u5_0ImmPred:$c)))),
          (S2_asr_i_vw V2I32:$b, imm:$c)>;
-def: Pat<(v2i32 (srl V2I32:$b, (i64 (HexagonCOMBINE (i32 u5_0ImmPred:$c),
-                                                    (i32 u5_0ImmPred:$c))))),
+def: Pat<(v2i32 (srl V2I32:$b, (v2i32 (HexagonVSPLAT u5_0ImmPred:$c)))),
          (S2_lsr_i_vw V2I32:$b, imm:$c)>;
-def: Pat<(v2i32 (shl V2I32:$b, (i64 (HexagonCOMBINE (i32 u5_0ImmPred:$c),
-                                                    (i32 u5_0ImmPred:$c))))),
+def: Pat<(v2i32 (shl V2I32:$b, (v2i32 (HexagonVSPLAT u5_0ImmPred:$c)))),
          (S2_asl_i_vw V2I32:$b, imm:$c)>;
 
 def: Pat<(v4i16 (sra V4I16:$b, (v4i16 (HexagonVSPLAT u4_0ImmPred:$c)))),
diff --git a/lib/Target/Hexagon/HexagonTargetObjectFile.cpp b/lib/Target/Hexagon/HexagonTargetObjectFile.cpp
index 34df2ebcc520..ea86c9c42f47 100644
--- a/lib/Target/Hexagon/HexagonTargetObjectFile.cpp
+++ b/lib/Target/Hexagon/HexagonTargetObjectFile.cpp
@@ -53,6 +53,10 @@ static cl::opt<bool>
     EmitJtInText("hexagon-emit-jt-text", cl::Hidden, cl::init(false),
                  cl::desc("Emit hexagon jump tables in function section"));
 
+static cl::opt<bool>
+    EmitLutInText("hexagon-emit-lut-text", cl::Hidden, cl::init(false),
+                 cl::desc("Emit hexagon lookup tables in function section"));
+
 // TraceGVPlacement controls messages for all builds. For builds with assertions
 // (debug or release), messages are also controlled by the usual debug flags
 // (e.g. -debug and -debug-only=globallayout)
@@ -136,6 +140,13 @@ MCSection *HexagonTargetObjectFile::SelectSectionForGlobal(
          << (Kind.isBSS() ? "kind_bss " : "" )
          << (Kind.isBSSLocal() ? "kind_bss_local " : "" ));
 
+  // If the lookup table is used by more than one function, do not place
+  // it in text section.
+  if (EmitLutInText && GO->getName().startswith("switch.table")) {
+    if (const Function *Fn = getLutUsedFunction(GO))
+      return selectSectionForLookupTable(GO, TM, Fn);
+  }
+
   if (isGlobalInSmallSection(GO, TM))
     return selectSmallSectionForGlobal(GO, Kind, TM);
 
@@ -402,3 +413,39 @@ MCSection *HexagonTargetObjectFile::selectSmallSectionForGlobal(
   // Otherwise, we work the same as ELF.
   return TargetLoweringObjectFileELF::SelectSectionForGlobal(GO, Kind, TM);
 }
+
+// Return the function that uses the lookup table. If there are more
+// than one live function that uses this look table, bail out and place
+// the lookup table in default section.
+const Function *
+HexagonTargetObjectFile::getLutUsedFunction(const GlobalObject *GO) const {
+  const Function *ReturnFn = nullptr;
+  for (auto U : GO->users()) {
+    // validate each instance of user to be a live function.
+    auto *I = dyn_cast<Instruction>(U);
+    if (!I)
+      continue;
+    auto *Bb = I->getParent();
+    if (!Bb)
+      continue;
+    auto *UserFn = Bb->getParent();
+    if (!ReturnFn)
+      ReturnFn = UserFn;
+    else if (ReturnFn != UserFn)
+      return nullptr;
+  }
+  return ReturnFn;
+}
+
+MCSection *HexagonTargetObjectFile::selectSectionForLookupTable(
+    const GlobalObject *GO, const TargetMachine &TM, const Function *Fn) const {
+
+  SectionKind Kind = SectionKind::getText();
+  // If the function has explicit section, place the lookup table in this
+  // explicit section.
+  if (Fn->hasSection())
+    return getExplicitSectionGlobal(Fn, Kind, TM);
+
+  const auto *FuncObj = dyn_cast<GlobalObject>(Fn);
+  return SelectSectionForGlobal(FuncObj, Kind, TM);
+}
diff --git a/lib/Target/Hexagon/HexagonTargetObjectFile.h b/lib/Target/Hexagon/HexagonTargetObjectFile.h
index 373d850b53be..eff44f097e03 100644
--- a/lib/Target/Hexagon/HexagonTargetObjectFile.h
+++ b/lib/Target/Hexagon/HexagonTargetObjectFile.h
@@ -36,6 +36,8 @@ namespace llvm {
     bool shouldPutJumpTableInFunctionSection(bool UsesLabelDifference,
                                              const Function &F) const override;
 
+    const Function *getLutUsedFunction(const GlobalObject *GO) const;
+
   private:
     MCSectionELF *SmallDataSection;
     MCSectionELF *SmallBSSSection;
@@ -46,6 +48,10 @@ namespace llvm {
     MCSection *selectSmallSectionForGlobal(const GlobalObject *GO,
                                            SectionKind Kind,
                                            const TargetMachine &TM) const;
+
+    MCSection *selectSectionForLookupTable(const GlobalObject *GO,
+                                           const TargetMachine &TM,
+                                           const Function *Fn) const;
   };
 
 } // namespace llvm