[Hexagon] Add support for nontemporal loads and stores on HVX

[michaelwu]

Summary

This makes __builtin_nontemporal_load and __builtin_nontemporal_store work properly for HVX registers.

[kparzysz]

test/CodeGen/Hexagon/hvx-nontemporal.ll

1

1. Add -march=hexagon.
2. Please use function attributes instead of -mv60 and -enable-hexagon-hvx-double.

[kparzysz]

With those changes, LGTM.

[michaelwu]

Test adjusted per review comments.

[kparzysz]

Perfect! Thanks.

[michaelwu]

I don't have commit access. Can someone land this for me? Thanks!

lib/Target/Hexagon/HexagonISelDAGToDAG.cpp

void HexagonDAGToDAGISel::SelectIndexedLoad(LoadSDNode *LD, const SDLoc &dl) {
   case MVT::i64:
     Opcode = IsValidInc ? Hexagon::L2_loadrd_pi : Hexagon::L2_loadrd_io;
     break;
   // 64B
   case MVT::v64i8:
   case MVT::v32i16:
   case MVT::v16i32:
   case MVT::v8i64:
-    if (isAlignedMemNode(LD))
-      Opcode = IsValidInc ? Hexagon::V6_vL32b_pi : Hexagon::V6_vL32b_ai;
+    if (isAlignedMemNode(LD)) {
+      if (LD->isNonTemporal())
+        Opcode = IsValidInc ? Hexagon::V6_vL32b_nt_pi : Hexagon::V6_vL32b_nt_ai;
       else
+        Opcode = IsValidInc ? Hexagon::V6_vL32b_pi : Hexagon::V6_vL32b_ai;
+    } else {
       Opcode = IsValidInc ? Hexagon::V6_vL32Ub_pi : Hexagon::V6_vL32Ub_ai;
+    }
     break;
   // 128B
   case MVT::v128i8:
   case MVT::v64i16:
   case MVT::v32i32:
   case MVT::v16i64:
-    if (isAlignedMemNode(LD))
+    if (isAlignedMemNode(LD)) {
+      if (LD->isNonTemporal())
+        Opcode = IsValidInc ? Hexagon::V6_vL32b_nt_pi_128B
+                            : Hexagon::V6_vL32b_nt_ai_128B;
+      else
         Opcode = IsValidInc ? Hexagon::V6_vL32b_pi_128B
                             : Hexagon::V6_vL32b_ai_128B;
-    else
+    } else {
       Opcode = IsValidInc ? Hexagon::V6_vL32Ub_pi_128B
                           : Hexagon::V6_vL32Ub_ai_128B;
+    }
     break;
   default:
     llvm_unreachable("Unexpected memory type in indexed load");
   }
 
   SDValue IncV = CurDAG->getTargetConstant(Inc, dl, MVT::i32);
   MachineSDNode::mmo_iterator MemOp = MF->allocateMemRefsArray(1);
   MemOp[0] = LD->getMemOperand();
void HexagonDAGToDAGISel::SelectIndexedStore(StoreSDNode *ST, const SDLoc &dl) {
   case MVT::i64:
     Opcode = IsValidInc ? Hexagon::S2_storerd_pi : Hexagon::S2_storerd_io;
     break;
   // 64B
   case MVT::v64i8:
   case MVT::v32i16:
   case MVT::v16i32:
   case MVT::v8i64:
-    if (isAlignedMemNode(ST))
-      Opcode = IsValidInc ? Hexagon::V6_vS32b_pi : Hexagon::V6_vS32b_ai;
+    if (isAlignedMemNode(ST)) {
+      if (ST->isNonTemporal())
+        Opcode = IsValidInc ? Hexagon::V6_vS32b_nt_pi : Hexagon::V6_vS32b_nt_ai;
       else
+        Opcode = IsValidInc ? Hexagon::V6_vS32b_pi : Hexagon::V6_vS32b_ai;
+    } else {
       Opcode = IsValidInc ? Hexagon::V6_vS32Ub_pi : Hexagon::V6_vS32Ub_ai;
+    }
     break;
   // 128B
   case MVT::v128i8:
   case MVT::v64i16:
   case MVT::v32i32:
   case MVT::v16i64:
-    if (isAlignedMemNode(ST))
+    if (isAlignedMemNode(ST)) {
+      if (ST->isNonTemporal())
+        Opcode = IsValidInc ? Hexagon::V6_vS32b_nt_pi_128B
+                            : Hexagon::V6_vS32b_nt_ai_128B;
+      else
         Opcode = IsValidInc ? Hexagon::V6_vS32b_pi_128B
                             : Hexagon::V6_vS32b_ai_128B;
-    else
+    } else {
       Opcode = IsValidInc ? Hexagon::V6_vS32Ub_pi_128B
                           : Hexagon::V6_vS32Ub_ai_128B;
+    }
     break;
   default:
     llvm_unreachable("Unexpected memory type in indexed store");
   }
 
   if (ST->isTruncatingStore() && ValueVT.getSizeInBits() == 64) {
     assert(StoredVT.getSizeInBits() < 64 && "Not a truncating store");
     Value = CurDAG->getTargetExtractSubreg(Hexagon::isub_lo,

lib/Target/Hexagon/HexagonInstrInfo.cpp

 unsigned HexagonInstrInfo::isLoadFromStackSlot(const MachineInstr &MI,
                                                int &FrameIndex) const {
   switch (MI.getOpcode()) {
   default:
     break;
   case Hexagon::L2_loadri_io:
   case Hexagon::L2_loadrd_io:
   case Hexagon::V6_vL32b_ai:
+  case Hexagon::V6_vL32b_nt_ai:
   case Hexagon::V6_vL32b_ai_128B:
+  case Hexagon::V6_vL32b_nt_ai_128B:
   case Hexagon::V6_vL32Ub_ai:
   case Hexagon::V6_vL32Ub_ai_128B:
   case Hexagon::LDriw_pred:
   case Hexagon::LDriw_mod:
   case Hexagon::PS_vloadrq_ai:
   case Hexagon::PS_vloadrw_ai:
+  case Hexagon::PS_vloadrw_nt_ai:
   case Hexagon::PS_vloadrq_ai_128B:
-  case Hexagon::PS_vloadrw_ai_128B: {
+  case Hexagon::PS_vloadrw_ai_128B:
+  case Hexagon::PS_vloadrw_nt_ai_128B: {
     const MachineOperand OpFI = MI.getOperand(1);
     if (!OpFI.isFI())
       return 0;
     const MachineOperand OpOff = MI.getOperand(2);
     if (!OpOff.isImm() || OpOff.getImm() != 0)
       return 0;
     FrameIndex = OpFI.getIndex();
     return MI.getOperand(0).getReg();
bool HexagonInstrInfo::isValidOffset(unsigned Opcode, int Offset,
   // We used to assert if the offset was not properly aligned, however,
   // there are cases where a misaligned pointer recast can cause this
   // problem, and we need to allow for it. The front end warns of such
   // misaligns with respect to load size.
 
   switch (Opcode) {
   case Hexagon::PS_vstorerq_ai:
   case Hexagon::PS_vstorerw_ai:
+  case Hexagon::PS_vstorerw_nt_ai:
   case Hexagon::PS_vloadrq_ai:
   case Hexagon::PS_vloadrw_ai:
+  case Hexagon::PS_vloadrw_nt_ai:
   case Hexagon::V6_vL32b_ai:
   case Hexagon::V6_vS32b_ai:
+  case Hexagon::V6_vL32b_nt_ai:
+  case Hexagon::V6_vS32b_nt_ai:
   case Hexagon::V6_vL32Ub_ai:
   case Hexagon::V6_vS32Ub_ai:
     return isShiftedInt<4,6>(Offset);
 
   case Hexagon::PS_vstorerq_ai_128B:
   case Hexagon::PS_vstorerw_ai_128B:
+  case Hexagon::PS_vstorerw_nt_ai_128B:
   case Hexagon::PS_vloadrq_ai_128B:
   case Hexagon::PS_vloadrw_ai_128B:
+  case Hexagon::PS_vloadrw_nt_ai_128B:
   case Hexagon::V6_vL32b_ai_128B:
   case Hexagon::V6_vS32b_ai_128B:
+  case Hexagon::V6_vL32b_nt_ai_128B:
+  case Hexagon::V6_vS32b_nt_ai_128B:
   case Hexagon::V6_vL32Ub_ai_128B:
   case Hexagon::V6_vS32Ub_ai_128B:
     return isShiftedInt<4,7>(Offset);
 
   case Hexagon::J2_loop0i:
   case Hexagon::J2_loop1i:
     return isUInt<10>(Offset);
 

 // Return the cur value instruction for a given store.
 int HexagonInstrInfo::getDotCurOp(const MachineInstr &MI) const {
   switch (MI.getOpcode()) {
   default: llvm_unreachable("Unknown .cur type");
   case Hexagon::V6_vL32b_pi:
     return Hexagon::V6_vL32b_cur_pi;
   case Hexagon::V6_vL32b_ai:
     return Hexagon::V6_vL32b_cur_ai;
+  case Hexagon::V6_vL32b_nt_pi:
+    return Hexagon::V6_vL32b_nt_cur_pi;
+  case Hexagon::V6_vL32b_nt_ai:
+    return Hexagon::V6_vL32b_nt_cur_ai;
   //128B
   case Hexagon::V6_vL32b_pi_128B:
     return Hexagon::V6_vL32b_cur_pi_128B;
   case Hexagon::V6_vL32b_ai_128B:
     return Hexagon::V6_vL32b_cur_ai_128B;
+  case Hexagon::V6_vL32b_nt_pi_128B:
+    return Hexagon::V6_vL32b_nt_cur_pi_128B;
+  case Hexagon::V6_vL32b_nt_ai_128B:
+    return Hexagon::V6_vL32b_nt_cur_ai_128B;
   }
   return 0;
 }
 
 // Return the regular version of the .cur instruction.
 int HexagonInstrInfo::getNonDotCurOp(const MachineInstr &MI) const {
   switch (MI.getOpcode()) {
   default: llvm_unreachable("Unknown .cur type");
   case Hexagon::V6_vL32b_cur_pi:
     return Hexagon::V6_vL32b_pi;
   case Hexagon::V6_vL32b_cur_ai:
     return Hexagon::V6_vL32b_ai;
+  case Hexagon::V6_vL32b_nt_cur_pi:
+    return Hexagon::V6_vL32b_nt_pi;
+  case Hexagon::V6_vL32b_nt_cur_ai:
+    return Hexagon::V6_vL32b_nt_ai;
   //128B
   case Hexagon::V6_vL32b_cur_pi_128B:
     return Hexagon::V6_vL32b_pi_128B;
   case Hexagon::V6_vL32b_cur_ai_128B:
     return Hexagon::V6_vL32b_ai_128B;
+  case Hexagon::V6_vL32b_nt_cur_pi_128B:
+    return Hexagon::V6_vL32b_nt_pi_128B;
+  case Hexagon::V6_vL32b_nt_cur_ai_128B:
+    return Hexagon::V6_vL32b_nt_ai_128B;
   }
   return 0;
 }
 
 
 // The diagram below shows the steps involved in the conversion of a predicated
 // store instruction to its .new predicated new-value form.
 //

lib/Target/Hexagon/HexagonPatterns.td

 
 def unalignedstore : PatFrag<(ops node:$val, node:$addr), (store $val, $addr), [{
   return !isAlignedMemNode(dyn_cast<MemSDNode>(N));
 }]>;
 
 
 multiclass vS32b_ai_pats <ValueType VTSgl, ValueType VTDbl> {
   // Aligned stores
+  def : Pat<(alignednontemporalstore (VTSgl VectorRegs:$src1), IntRegs:$addr),
+            (V6_vS32b_nt_ai IntRegs:$addr, 0, (VTSgl VectorRegs:$src1))>,
+            Requires<[UseHVXSgl]>;
   def : Pat<(alignedstore (VTSgl VectorRegs:$src1), IntRegs:$addr),
             (V6_vS32b_ai IntRegs:$addr, 0, (VTSgl VectorRegs:$src1))>,
             Requires<[UseHVXSgl]>;
   def : Pat<(unalignedstore (VTSgl VectorRegs:$src1), IntRegs:$addr),
             (V6_vS32Ub_ai IntRegs:$addr, 0, (VTSgl VectorRegs:$src1))>,
             Requires<[UseHVXSgl]>;
 
   // 128B Aligned stores
+  def : Pat<(alignednontemporalstore (VTDbl VectorRegs128B:$src1), IntRegs:$addr),
+            (V6_vS32b_nt_ai_128B IntRegs:$addr, 0, (VTDbl VectorRegs128B:$src1))>,
+            Requires<[UseHVXDbl]>;
   def : Pat<(alignedstore (VTDbl VectorRegs128B:$src1), IntRegs:$addr),
             (V6_vS32b_ai_128B IntRegs:$addr, 0, (VTDbl VectorRegs128B:$src1))>,
             Requires<[UseHVXDbl]>;
   def : Pat<(unalignedstore (VTDbl VectorRegs128B:$src1), IntRegs:$addr),
             (V6_vS32Ub_ai_128B IntRegs:$addr, 0, (VTDbl VectorRegs128B:$src1))>,
             Requires<[UseHVXDbl]>;
 
   // Fold Add R+OFF into vector store.
   let AddedComplexity = 10 in {
+    def : Pat<(alignednontemporalstore (VTSgl VectorRegs:$src1),
+                     (add IntRegs:$src2, Iss4_6:$offset)),
+              (V6_vS32b_nt_ai IntRegs:$src2, Iss4_6:$offset,
+                           (VTSgl VectorRegs:$src1))>,
+              Requires<[UseHVXSgl]>;
     def : Pat<(alignedstore (VTSgl VectorRegs:$src1),
                      (add IntRegs:$src2, Iss4_6:$offset)),
               (V6_vS32b_ai IntRegs:$src2, Iss4_6:$offset,
                            (VTSgl VectorRegs:$src1))>,
               Requires<[UseHVXSgl]>;
     def : Pat<(unalignedstore (VTSgl VectorRegs:$src1),
                      (add IntRegs:$src2, Iss4_6:$offset)),
               (V6_vS32Ub_ai IntRegs:$src2, Iss4_6:$offset,
                            (VTSgl VectorRegs:$src1))>,
               Requires<[UseHVXSgl]>;
 
     // Fold Add R+OFF into vector store 128B.
+    def : Pat<(alignednontemporalstore (VTDbl VectorRegs128B:$src1),
+                     (add IntRegs:$src2, Iss4_7:$offset)),
+              (V6_vS32b_nt_ai_128B IntRegs:$src2, Iss4_7:$offset,
+                                (VTDbl VectorRegs128B:$src1))>,
+              Requires<[UseHVXDbl]>;
     def : Pat<(alignedstore (VTDbl VectorRegs128B:$src1),
                      (add IntRegs:$src2, Iss4_7:$offset)),
               (V6_vS32b_ai_128B IntRegs:$src2, Iss4_7:$offset,
                                 (VTDbl VectorRegs128B:$src1))>,
               Requires<[UseHVXDbl]>;
     def : Pat<(unalignedstore (VTDbl VectorRegs128B:$src1),
                      (add IntRegs:$src2, Iss4_7:$offset)),
               (V6_vS32Ub_ai_128B IntRegs:$src2, Iss4_7:$offset,
                                 (VTDbl VectorRegs128B:$src1))>,
               Requires<[UseHVXDbl]>;
   }
 }
 
 defm : vS32b_ai_pats <v64i8,  v128i8>;
 defm : vS32b_ai_pats <v32i16, v64i16>;
 defm : vS32b_ai_pats <v16i32, v32i32>;
 defm : vS32b_ai_pats <v8i64,  v16i64>;
 
 
 multiclass vL32b_ai_pats <ValueType VTSgl, ValueType VTDbl> {
   // Aligned loads
+  def : Pat < (VTSgl (alignednontemporalload IntRegs:$addr)),
+              (V6_vL32b_nt_ai IntRegs:$addr, 0) >,
+              Requires<[UseHVXSgl]>;
   def : Pat < (VTSgl (alignedload IntRegs:$addr)),
               (V6_vL32b_ai IntRegs:$addr, 0) >,
               Requires<[UseHVXSgl]>;
   def : Pat < (VTSgl (unalignedload IntRegs:$addr)),
               (V6_vL32Ub_ai IntRegs:$addr, 0) >,
               Requires<[UseHVXSgl]>;
 
   // 128B Load
+  def : Pat < (VTDbl (alignednontemporalload IntRegs:$addr)),
+              (V6_vL32b_nt_ai_128B IntRegs:$addr, 0) >,
+              Requires<[UseHVXDbl]>;
   def : Pat < (VTDbl (alignedload IntRegs:$addr)),
               (V6_vL32b_ai_128B IntRegs:$addr, 0) >,
               Requires<[UseHVXDbl]>;
   def : Pat < (VTDbl (unalignedload IntRegs:$addr)),
               (V6_vL32Ub_ai_128B IntRegs:$addr, 0) >,
               Requires<[UseHVXDbl]>;
 
   // Fold Add R+OFF into vector load.
   let AddedComplexity = 10 in {
+    def : Pat<(VTDbl (alignednontemporalload (add IntRegs:$src2, Iss4_7:$offset))),
+              (V6_vL32b_nt_ai_128B IntRegs:$src2, Iss4_7:$offset)>,
+               Requires<[UseHVXDbl]>;
     def : Pat<(VTDbl (alignedload (add IntRegs:$src2, Iss4_7:$offset))),
               (V6_vL32b_ai_128B IntRegs:$src2, Iss4_7:$offset)>,
                Requires<[UseHVXDbl]>;
     def : Pat<(VTDbl (unalignedload (add IntRegs:$src2, Iss4_7:$offset))),
               (V6_vL32Ub_ai_128B IntRegs:$src2, Iss4_7:$offset)>,
                Requires<[UseHVXDbl]>;
 
+    def : Pat<(VTSgl (alignednontemporalload (add IntRegs:$src2, Iss4_6:$offset))),
+              (V6_vL32b_nt_ai IntRegs:$src2, Iss4_6:$offset)>,
+              Requires<[UseHVXSgl]>;
     def : Pat<(VTSgl (alignedload (add IntRegs:$src2, Iss4_6:$offset))),
               (V6_vL32b_ai IntRegs:$src2, Iss4_6:$offset)>,
               Requires<[UseHVXSgl]>;
     def : Pat<(VTSgl (unalignedload (add IntRegs:$src2, Iss4_6:$offset))),
               (V6_vL32Ub_ai IntRegs:$src2, Iss4_6:$offset)>,
               Requires<[UseHVXSgl]>;
   }
 }
 
 defm : vL32b_ai_pats <v64i8,  v128i8>;
 defm : vL32b_ai_pats <v32i16, v64i16>;
 defm : vL32b_ai_pats <v16i32, v32i32>;
 defm : vL32b_ai_pats <v8i64,  v16i64>;
 
 multiclass STrivv_pats <ValueType VTSgl, ValueType VTDbl> {
+  def : Pat<(alignednontemporalstore (VTSgl VecDblRegs:$src1), IntRegs:$addr),
+            (PS_vstorerw_nt_ai IntRegs:$addr, 0, (VTSgl VecDblRegs:$src1))>,
+           Requires<[UseHVXSgl]>;
   def : Pat<(alignedstore (VTSgl VecDblRegs:$src1), IntRegs:$addr),
             (PS_vstorerw_ai IntRegs:$addr, 0, (VTSgl VecDblRegs:$src1))>,
            Requires<[UseHVXSgl]>;
   def : Pat<(unalignedstore (VTSgl VecDblRegs:$src1), IntRegs:$addr),
             (PS_vstorerwu_ai IntRegs:$addr, 0, (VTSgl VecDblRegs:$src1))>,
            Requires<[UseHVXSgl]>;
 
+  def : Pat<(alignednontemporalstore (VTDbl VecDblRegs128B:$src1), IntRegs:$addr),
+            (PS_vstorerw_nt_ai_128B IntRegs:$addr, 0,
+                  (VTDbl VecDblRegs128B:$src1))>,
+            Requires<[UseHVXDbl]>;
   def : Pat<(alignedstore (VTDbl VecDblRegs128B:$src1), IntRegs:$addr),
             (PS_vstorerw_ai_128B IntRegs:$addr, 0,
                   (VTDbl VecDblRegs128B:$src1))>,
             Requires<[UseHVXDbl]>;
   def : Pat<(unalignedstore (VTDbl VecDblRegs128B:$src1), IntRegs:$addr),
             (PS_vstorerwu_ai_128B IntRegs:$addr, 0,
                   (VTDbl VecDblRegs128B:$src1))>,
             Requires<[UseHVXDbl]>;
 }
 
 defm : STrivv_pats <v128i8, v256i8>;
 defm : STrivv_pats <v64i16, v128i16>;
 defm : STrivv_pats <v32i32, v64i32>;
 defm : STrivv_pats <v16i64, v32i64>;
 
 multiclass LDrivv_pats <ValueType VTSgl, ValueType VTDbl> {
+  def : Pat<(VTSgl (alignednontemporalload I32:$addr)),
+            (PS_vloadrw_nt_ai I32:$addr, 0)>,
+           Requires<[UseHVXSgl]>;
   def : Pat<(VTSgl (alignedload I32:$addr)),
             (PS_vloadrw_ai I32:$addr, 0)>,
            Requires<[UseHVXSgl]>;
   def : Pat<(VTSgl (unalignedload I32:$addr)),
             (PS_vloadrwu_ai I32:$addr, 0)>,
            Requires<[UseHVXSgl]>;
 
+  def : Pat<(VTDbl (alignednontemporalload I32:$addr)),
+            (PS_vloadrw_nt_ai_128B I32:$addr, 0)>,
+           Requires<[UseHVXDbl]>;
   def : Pat<(VTDbl (alignedload I32:$addr)),
             (PS_vloadrw_ai_128B I32:$addr, 0)>,
            Requires<[UseHVXDbl]>;
   def : Pat<(VTDbl (unalignedload I32:$addr)),
             (PS_vloadrwu_ai_128B I32:$addr, 0)>,
            Requires<[UseHVXDbl]>;
 }
 

lib/Target/Hexagon/HexagonPseudo.td

class STrivv_template<RegisterClass RC, InstHexagon rootInst>
   : InstHexagon<(outs), (ins IntRegs:$addr, s32_0Imm:$off, RC:$src),
     "", [], "", rootInst.Itinerary, rootInst.Type>;
 
 def PS_vstorerw_ai: STrivv_template<VecDblRegs, V6_vS32b_ai>,
       Requires<[HasV60T,UseHVXSgl]>;
 def PS_vstorerw_ai_128B: STrivv_template<VecDblRegs128B, V6_vS32b_ai_128B>,
       Requires<[HasV60T,UseHVXDbl]>;
 
+def PS_vstorerw_nt_ai: STrivv_template<VecDblRegs, V6_vS32b_nt_ai>,
+      Requires<[HasV60T,UseHVXSgl]>;
+def PS_vstorerw_nt_ai_128B: STrivv_template<VecDblRegs128B, V6_vS32b_nt_ai_128B>,
+      Requires<[HasV60T,UseHVXDbl]>;
+
 def PS_vstorerwu_ai: STrivv_template<VecDblRegs, V6_vS32Ub_ai>,
       Requires<[HasV60T,UseHVXSgl]>;
 def PS_vstorerwu_ai_128B: STrivv_template<VecDblRegs128B, V6_vS32Ub_ai_128B>,
       Requires<[HasV60T,UseHVXDbl]>;
 
 let isPseudo = 1, isCodeGenOnly = 1, mayStore = 1, hasSideEffects = 0 in {
   def PS_vstorerq_ai: Pseudo<(outs),
         (ins IntRegs:$Rs, s32_0Imm:$Off, VecPredRegs:$Qt), "", []>,
class LDrivv_template<RegisterClass RC, InstHexagon rootInst>
   : InstHexagon<(outs RC:$dst), (ins IntRegs:$addr, s32_0Imm:$off),
     "", [], "", rootInst.Itinerary, rootInst.Type>;
 
 def PS_vloadrw_ai: LDrivv_template<VecDblRegs, V6_vL32b_ai>,
       Requires<[HasV60T,UseHVXSgl]>;
 def PS_vloadrw_ai_128B: LDrivv_template<VecDblRegs128B, V6_vL32b_ai_128B>,
       Requires<[HasV60T,UseHVXDbl]>;
 
+def PS_vloadrw_nt_ai: LDrivv_template<VecDblRegs, V6_vL32b_nt_ai>,
+      Requires<[HasV60T,UseHVXSgl]>;
+def PS_vloadrw_nt_ai_128B: LDrivv_template<VecDblRegs128B, V6_vL32b_nt_ai_128B>,
+      Requires<[HasV60T,UseHVXDbl]>;
+
 def PS_vloadrwu_ai: LDrivv_template<VecDblRegs, V6_vL32Ub_ai>,
       Requires<[HasV60T,UseHVXSgl]>;
 def PS_vloadrwu_ai_128B: LDrivv_template<VecDblRegs128B, V6_vL32Ub_ai_128B>,
       Requires<[HasV60T,UseHVXDbl]>;
 
 let isPseudo = 1, isCodeGenOnly = 1, mayLoad = 1, hasSideEffects = 0 in {
   def PS_vloadrq_ai: Pseudo<(outs VecPredRegs:$Qd),
         (ins IntRegs:$Rs, s32_0Imm:$Off), "", []>,

test/CodeGen/Hexagon/hvx-nontemporal.ll

+; RUN: llc -mv60 -enable-hexagon-hvx-double < %s | FileCheck %s
+target triple = "hexagon"
+
+; Function Attrs: norecurse nounwind
+define void @test(<32 x i32>* nocapture readonly %x, <32 x i32>* nocapture readnone %y, <32 x i32>* nocapture %a, <32 x i32>* nocapture %b) #0 {
+entry:
+; CHECK: v0 = vmem(r0+#7):nt
+  %add.ptr = getelementptr inbounds <32 x i32>, <32 x i32>* %x, i32 7
+  %0 = load <32 x i32>, <32 x i32>* %add.ptr, align 128, !tbaa !1, !nontemporal !4
+
+; CHECK: v1.cur = vmem(r2+#0):nt
+  %1 = load <32 x i32>, <32 x i32>* %a, align 128, !tbaa !1, !nontemporal !4
+
+; CHECK: vmem(r3+#3):nt = v1
+  %add.ptr2 = getelementptr inbounds <32 x i32>, <32 x i32>* %b, i32 3
+  store <32 x i32> %1, <32 x i32>* %add.ptr2, align 128, !tbaa !1, !nontemporal !4
+
+; CHECK: vmem(r2+#0):nt = v0
+  store <32 x i32> %0, <32 x i32>* %a, align 128, !tbaa !1, !nontemporal !4
+  ret void
+}
+
+!1 = !{!2, !2, i64 0}
+!2 = !{!"omnipotent char", !3, i64 0}
+!3 = !{!"Simple C/C++ TBAA"}
+!4 = !{i32 1}