[AArch64][SVE] Implement structured load intrinsics

[c-rhodes]

Summary

This patch adds initial support for the following instrinsics:

• llvm.aarch64.sve.ld2
• llvm.aarch64.sve.ld3
• llvm.aarch64.sve.ld4

For loading two, three and four vectors worth of data. Basic codegen is
implemented with reg+reg and reg+imm addressing modes being addressed
in a later patch.

The types returned by these intrinsics have a number of elements that is
a multiple of the elements in a 128-bit vector for a given type and N,
where N is the number of vectors being loaded, i.e. 2, 3 or 4. Thus, for
32-bit elements the types are:

LD2 : <vscale x 8 x i32>
LD3 : <vscale x 12 x i32>
LD4 : <vscale x 16 x i32>

This is implemented with target-specific intrinsics for each variant
that take the same operands as the IR intrinsic but return N values,
where the type of each value is a full vector, i.e. <vscale x 4 x i32>
in the above example. These values are then concatenated using the
standard concat_vector intrinsic to maintain type legality with the IR.

These intrinsics are intended for use in the Arm C Language
Extension (ACLE).

[fpetrogalli]

Hi @c-rhodes ,

thank you for working on this. I am basing the addressing mode optimization for ldN on this patch, I just wanted to point out a couple of minor remarks!

Grazie,

Francesco

[Nit] Should use vscale instead of n in the commit message:

LD2 : <n x 8 x i32>
LD3 : <n x 12 x i32>
LD4 : <n x 16 x i32>

llvm/lib/Target/AArch64/AArch64ISelDAGToDAG.cpp

1421

SubRegIdx is always set to AArch64::zsub0. Can we remove it from the parameter list of the method and use AArch64::zsub0 directly inside the function?

llvm/lib/Target/AArch64/AArch64ISelLowering.cpp

9791

[Nit] Might be worth asserting that VT.getVectorElementCount() % N == 0.

[fpetrogalli]

llvm/include/llvm/IR/IntrinsicsAArch64.td

817

Question: you have three overloaded operands here. How comes that you need to specify only one of them in the intrinsic name?

If I look at one of your tests:

%res = call <vscale x 32 x i8> @llvm.aarch64.sve.ld2.nxv32i8(<vscale x 16 x i1> %pred,
                                                             <vscale x 16 x i8>* %addr)

By the definition you have specified here, I was expecting to see the following intrinsic: @llvm.aarch64.sve.ld2.nxv32i8.nxv16i8.p0nxv16i8. Am I missing something?

[c-rhodes]

Thanks for the comments @fpetrogalli! I'll update this patch to address your comments.

llvm/include/llvm/IR/IntrinsicsAArch64.td

817

You're right, there should be a suffix for each overloaded type in the intrinsic name. I initially implemented each load intrinsic separately with special types where the predicate and address were based on the return type, so something like LLVMVectorHalfWidth<0, llvm_i1_ty> and LLVMPointerToHalfWidthVector<0> for LD2 args. With this definition only the return type was overloaded and I missed updating the intrinsics in the tests. Interesting how there's no issues without specifying each overloaded type, I'll fix the tests.

[fpetrogalli]

llvm/test/CodeGen/AArch64/sve-intrinsics-loads-with-extract.ll

23 ↗

(On Diff #248735)

I just realized... do we need these tests? I might have missed something, but it seems to me that none of the code you have added here is related to the interaction between ld<N> and tuple.get. For example, the fact that in this specific test the load is loading memory in z31 and z0, in this order, seems to me more of a register allocation problem other than anything related to the fact that %res is coming from a ld<N> intrinsic.

I think that we can simplify these tests by removing the ld<N> and just writing tests that involve using the tuple.get on the input parameter of the test. For example, in the specific case of @ld2b_i8_1, I think you want to run llc on this code:

define <vscale x 16 x i8> @ld2b_i8_1(<vscale x 32 x i8> %res) {
; CHECK-LABEL: ld2b_i8_1:
; CHECK: mov z0.d, z1.d
; CHECK-NEXT: ret
  %v2 = call <vscale x 16 x i8> @llvm.aarch64.sve.tuple.get.nxv32i8(<vscale x 32 x i8> %res, i32 1)
  ret <vscale x 16 x i8> %v2

What you would be testing here is that whatever register is assigned to the input parameter, it is shuffled around by the intrinsic to be returned by the correct output register.

If you agree on this approach though, I think this deserves a separate patch.

One last (minor) extra reason for splitting things like I suggested: if we make these tests ld<N>-free, we won't have to bother updating/extending them when we need to do any work around the ld<N>.

[c-rhodes]

llvm/test/CodeGen/AArch64/sve-intrinsics-loads-with-extract.ll

23 ↗

(On Diff #248735)

I implemented these tests before we had calling convention support for passing and returning tuples of scalable vectors to/from functions (downstream) so the extract was necessary. There are already tests similar to what you suggested in D75674 covering tuple.get that should be sufficient (?), I can remove these if they're not adding anything.

[fpetrogalli]

llvm/test/CodeGen/AArch64/sve-intrinsics-loads-with-extract.ll

23 ↗

(On Diff #248735)

I think you can safely remove them. As you said, they were written when we didn't have the tuple calling conventions, but now we have it, and it is tested in https://reviews.llvm.org/D75674.

Nit: it might not be a problem, but in https://reviews.llvm.org/D75674 I cannot see tests that check when a tuple like <vscale x 8 x i64> is passed to a function definition. I see tests that are getting multiple "one -register" parameters in input, building a tuple with an intrinsic and passing it to a tuple.get intrinsic or to a function call, but not a test like the one I extracted in which the tuple is in input to the define. I'll leave a note there and drag @sdesmalen attention to see whether we should add them.

Thanks,

Francesco

[c-rhodes]

Changes:

• Rebased.
• Move lowering to DAGCombine pre-legalisation.
• Added assert for VT.getVectorElementCount().Min % N == 0.
• Removed test/CodeGen/AArch64/sve-intrinsics-loads-with-extract.ll.
• Added missing overloaded argument types to intrinsics in test/CodeGen/AArch64/sve-intrinsics-loads.ll.
• Removed SubRegIdx from SelectPredicatedLoad since AArch64::zsub0 is always used.

[c-rhodes]

llvm/lib/Target/AArch64/AArch64ISelDAGToDAG.cpp

1421

Done, you'll have to rebase D77251 once this is merged

[c-rhodes]

ping

[sdesmalen]

llvm/lib/Target/AArch64/AArch64ISelDAGToDAG.cpp

1461

Is the above statement correct when having just replaced SDValue(N, NumVecs-1) in the loop above?

llvm/lib/Target/AArch64/AArch64ISelLowering.cpp

9763

nit: call this LowerSVEStructLoad ?

[sdesmalen]

LGTM if you can add the extra comment.

llvm/lib/Target/AArch64/AArch64ISelDAGToDAG.cpp

1461

Sorry, never mind that comment, I now see it is taking the chain from Load, not N.
Can you add a comment clarifying this though? (that it is copying the chain here)?

[c-rhodes]

Changes:

• Rename LowerSVEStructuredLoad -> LowerSVEStructLoad.
• Clarify copy chain in SelectPredicatedLoad.

[c-rhodes]

llvm/lib/Target/AArch64/AArch64ISelDAGToDAG.cpp

1461

Yeah it's just copying the chain, agree it's confusing using NumVecs as the index, I've updated it.

llvm/include/llvm/IR/IntrinsicsAArch64.td

class AdvSIMD_SVE_Set_Vector_Tuple
     : Intrinsic<[llvm_anyvector_ty],
                 [LLVMMatchType<0>, llvm_i32_ty, llvm_anyvector_ty],
                 [IntrReadMem, ImmArg<1>]>;
 
   class AdvSIMD_SVE_Get_Vector_Tuple
     : Intrinsic<[llvm_anyvector_ty], [llvm_anyvector_ty, llvm_i32_ty],
                 [IntrReadMem, IntrArgMemOnly, ImmArg<1>]>;
 
+  class AdvSIMD_ManyVec_PredLoad_Intrinsic
+    : Intrinsic<[llvm_anyvector_ty], [llvm_anyvector_ty, llvm_anyptr_ty],
+                [IntrReadMem, IntrArgMemOnly]>;
+
   class AdvSIMD_1Vec_PredLoad_Intrinsic
     : Intrinsic<[llvm_anyvector_ty],
                 [LLVMScalarOrSameVectorWidth<0, llvm_i1_ty>,
                  LLVMPointerToElt<0>],
                 [IntrReadMem, IntrArgMemOnly]>;
 
   class AdvSIMD_1Vec_PredStore_Intrinsic
     : Intrinsic<[],

 def int_aarch64_sve_tuple_set : AdvSIMD_SVE_Set_Vector_Tuple;
 
 //
 // Loads
 //
 
 def int_aarch64_sve_ld1   : AdvSIMD_1Vec_PredLoad_Intrinsic;
 
+def int_aarch64_sve_ld2 : AdvSIMD_ManyVec_PredLoad_Intrinsic;
+def int_aarch64_sve_ld3 : AdvSIMD_ManyVec_PredLoad_Intrinsic;
+def int_aarch64_sve_ld4 : AdvSIMD_ManyVec_PredLoad_Intrinsic;
+
 def int_aarch64_sve_ldnt1 : AdvSIMD_1Vec_PredLoad_Intrinsic;
 def int_aarch64_sve_ldnf1 : AdvSIMD_1Vec_PredLoad_Intrinsic;
 def int_aarch64_sve_ldff1 : AdvSIMD_1Vec_PredLoad_Intrinsic;
 
 def int_aarch64_sve_ld1rq : AdvSIMD_1Vec_PredLoad_Intrinsic;
 
 //
 // Stores

llvm/lib/Target/AArch64/AArch64ISelDAGToDAG.cpp

public:
   void SelectTagP(SDNode *N);
 
   void SelectLoad(SDNode *N, unsigned NumVecs, unsigned Opc,
                      unsigned SubRegIdx);
   void SelectPostLoad(SDNode *N, unsigned NumVecs, unsigned Opc,
                          unsigned SubRegIdx);
   void SelectLoadLane(SDNode *N, unsigned NumVecs, unsigned Opc);
   void SelectPostLoadLane(SDNode *N, unsigned NumVecs, unsigned Opc);
+  void SelectPredicatedLoad(SDNode *N, unsigned NumVecs, const unsigned Opc);
 
   bool SelectAddrModeFrameIndexSVE(SDValue N, SDValue &Base, SDValue &OffImm);
   /// SVE Reg+Imm addressing mode.
   template <int64_t Min, int64_t Max>
   bool SelectAddrModeIndexedSVE(SDNode *Root, SDValue N, SDValue &Base,
                                 SDValue &OffImm);
   /// SVE Reg+Reg address mode.
   template <unsigned Scale>
void AArch64DAGToDAGISel::SelectPostLoad(SDNode *N, unsigned NumVecs,
 
   // Update the chain
   ReplaceUses(SDValue(N, NumVecs + 1), SDValue(Ld, 2));
   CurDAG->RemoveDeadNode(N);
 }
 
 /// Optimize \param OldBase and \param OldOffset selecting the best addressing
 /// mode. Returns a tuple consisting of an Opcode, an SDValue representing the
 /// new Base and an SDValue representing the new offset.
 template <unsigned Scale>
 std::tuple<unsigned, SDValue, SDValue>
 AArch64DAGToDAGISel::findAddrModeSVELoadStore(SDNode *N, const unsigned Opc_rr,
                                               const unsigned Opc_ri,
                                               const SDValue &OldBase,
                                               const SDValue &OldOffset) {
   SDValue NewBase = OldBase;
   SDValue NewOffset = OldOffset;
   // Detect a possible Reg+Imm addressing mode.
   const bool IsRegImm = SelectAddrModeIndexedSVE</*Min=*/-8, /*Max=*/7>(
       N, OldBase, NewBase, NewOffset);
 
   // Detect a possible reg+reg addressing mode, but only if we haven't already
   // detected a Reg+Imm one.
   const bool IsRegReg =
       !IsRegImm && SelectSVERegRegAddrMode<Scale>(OldBase, NewBase, NewOffset);
 
   // Select the instruction.
   return std::make_tuple(IsRegReg ? Opc_rr : Opc_ri, NewBase, NewOffset);
 }
 
+void AArch64DAGToDAGISel::SelectPredicatedLoad(SDNode *N, unsigned NumVecs,
+                                               const unsigned Opc) {
+  SDLoc DL(N);
+  EVT VT = N->getValueType(0);
+  SDValue Chain = N->getOperand(0);
+
+  SDValue Ops[] = {N->getOperand(1), // Predicate
+                   N->getOperand(2), // Memory operand
+                   CurDAG->getTargetConstant(0, DL, MVT::i64), Chain};
+
+  const EVT ResTys[] = {MVT::Untyped, MVT::Other};
+
+  SDNode *Load = CurDAG->getMachineNode(Opc, DL, ResTys, Ops);
+  SDValue SuperReg = SDValue(Load, 0);
+  for (unsigned i = 0; i < NumVecs; ++i)
+    ReplaceUses(SDValue(N, i), CurDAG->getTargetExtractSubreg(
+                                   AArch64::zsub0 + i, DL, VT, SuperReg));
+
+  // Copy chain
+  unsigned ChainIdx = NumVecs;
+  ReplaceUses(SDValue(N, ChainIdx), SDValue(Load, 1));
+  CurDAG->RemoveDeadNode(N);
+}
+
 void AArch64DAGToDAGISel::SelectStore(SDNode *N, unsigned NumVecs,
                                       unsigned Opc) {
   SDLoc dl(N);
   EVT VT = N->getOperand(2)->getValueType(0);
 
   // Form a REG_SEQUENCE to force register allocation.
   bool Is128Bit = VT.getSizeInBits() == 128;
   SmallVector<SDValue, 4> Regs(N->op_begin() + 2, N->op_begin() + 2 + NumVecs);
if (VT == MVT::v16i8 || VT == MVT::v8i8) {
       return;
     } else if (VT == MVT::v2i64 || VT == MVT::v1i64 || VT == MVT::v2f64 ||
                VT == MVT::v1f64) {
       SelectPostStoreLane(Node, 4, AArch64::ST4i64_POST);
       return;
     }
     break;
   }
+  case AArch64ISD::SVE_LD2: {
+    if (VT == MVT::nxv16i8) {
+      SelectPredicatedLoad(Node, 2, AArch64::LD2B_IMM);
+      return;
+    } else if (VT == MVT::nxv8i16 || VT == MVT::nxv8f16) {
+      SelectPredicatedLoad(Node, 2, AArch64::LD2H_IMM);
+      return;
+    } else if (VT == MVT::nxv4i32 || VT == MVT::nxv4f32) {
+      SelectPredicatedLoad(Node, 2, AArch64::LD2W_IMM);
+      return;
+    } else if (VT == MVT::nxv2i64 || VT == MVT::nxv2f64) {
+      SelectPredicatedLoad(Node, 2, AArch64::LD2D_IMM);
+      return;
+    }
+    break;
+  }
+  case AArch64ISD::SVE_LD3: {
+    if (VT == MVT::nxv16i8) {
+      SelectPredicatedLoad(Node, 3, AArch64::LD3B_IMM);
+      return;
+    } else if (VT == MVT::nxv8i16 || VT == MVT::nxv8f16) {
+      SelectPredicatedLoad(Node, 3, AArch64::LD3H_IMM);
+      return;
+    } else if (VT == MVT::nxv4i32 || VT == MVT::nxv4f32) {
+      SelectPredicatedLoad(Node, 3, AArch64::LD3W_IMM);
+      return;
+    } else if (VT == MVT::nxv2i64 || VT == MVT::nxv2f64) {
+      SelectPredicatedLoad(Node, 3, AArch64::LD3D_IMM);
+      return;
+    }
+    break;
+  }
+  case AArch64ISD::SVE_LD4: {
+    if (VT == MVT::nxv16i8) {
+      SelectPredicatedLoad(Node, 4, AArch64::LD4B_IMM);
+      return;
+    } else if (VT == MVT::nxv8i16 || VT == MVT::nxv8f16) {
+      SelectPredicatedLoad(Node, 4, AArch64::LD4H_IMM);
+      return;
+    } else if (VT == MVT::nxv4i32 || VT == MVT::nxv4f32) {
+      SelectPredicatedLoad(Node, 4, AArch64::LD4W_IMM);
+      return;
+    } else if (VT == MVT::nxv2i64 || VT == MVT::nxv2f64) {
+      SelectPredicatedLoad(Node, 4, AArch64::LD4D_IMM);
+      return;
+    }
+    break;
+  }
   }
 
   // Select the default instruction
   SelectCode(Node);
 }
 
 /// createAArch64ISelDag - This pass converts a legalized DAG into a
 /// AArch64-specific DAG, ready for instruction scheduling.

llvm/lib/Target/AArch64/AArch64ISelLowering.h

enum NodeType : unsigned {
   LD1,
   LD1S,
   LDNF1,
   LDNF1S,
   LDFF1,
   LDFF1S,
   LD1RQ,
 
+  // Structured loads.
+  SVE_LD2,
+  SVE_LD3,
+  SVE_LD4,
+
   // Unsigned gather loads.
   GLD1,
   GLD1_SCALED,
   GLD1_UXTW,
   GLD1_SXTW,
   GLD1_UXTW_SCALED,
   GLD1_SXTW_SCALED,
   GLD1_IMM,
private:
   SDValue LowerVSCALE(SDValue Op, SelectionDAG &DAG) const;
   SDValue LowerVECREDUCE(SDValue Op, SelectionDAG &DAG) const;
   SDValue LowerATOMIC_LOAD_SUB(SDValue Op, SelectionDAG &DAG) const;
   SDValue LowerATOMIC_LOAD_AND(SDValue Op, SelectionDAG &DAG) const;
   SDValue LowerDYNAMIC_STACKALLOC(SDValue Op, SelectionDAG &DAG) const;
   SDValue LowerWindowsDYNAMIC_STACKALLOC(SDValue Op, SDValue Chain,
                                          SDValue &Size,
                                          SelectionDAG &DAG) const;
+  SDValue LowerSVEStructLoad(unsigned Intrinsic, ArrayRef<SDValue> LoadOps,
+                             EVT VT, SelectionDAG &DAG, const SDLoc &DL) const;
 
   SDValue BuildSDIVPow2(SDNode *N, const APInt &Divisor, SelectionDAG &DAG,
                         SmallVectorImpl<SDNode *> &Created) const override;
   SDValue getSqrtEstimate(SDValue Operand, SelectionDAG &DAG, int Enabled,
                           int &ExtraSteps, bool &UseOneConst,
                           bool Reciprocal) const override;
   SDValue getRecipEstimate(SDValue Operand, SelectionDAG &DAG, int Enabled,
                            int &ExtraSteps) const override;

llvm/lib/Target/AArch64/AArch64ISelLowering.cpp

const char *AArch64TargetLowering::getTargetNodeName(unsigned Opcode) const {
   case AArch64ISD::PTRUE:             return "AArch64ISD::PTRUE";
   case AArch64ISD::LD1:               return "AArch64ISD::LD1";
   case AArch64ISD::LD1S:              return "AArch64ISD::LD1S";
   case AArch64ISD::LDNF1:             return "AArch64ISD::LDNF1";
   case AArch64ISD::LDNF1S:            return "AArch64ISD::LDNF1S";
   case AArch64ISD::LDFF1:             return "AArch64ISD::LDFF1";
   case AArch64ISD::LDFF1S:            return "AArch64ISD::LDFF1S";
   case AArch64ISD::LD1RQ:             return "AArch64ISD::LD1RQ";
+  case AArch64ISD::SVE_LD2:           return "AArch64ISD::SVE_LD2";
+  case AArch64ISD::SVE_LD3:           return "AArch64ISD::SVE_LD3";
+  case AArch64ISD::SVE_LD4:           return "AArch64ISD::SVE_LD4";
   case AArch64ISD::GLD1:              return "AArch64ISD::GLD1";
   case AArch64ISD::GLD1_SCALED:       return "AArch64ISD::GLD1_SCALED";
   case AArch64ISD::GLD1_SXTW:         return "AArch64ISD::GLD1_SXTW";
   case AArch64ISD::GLD1_UXTW:         return "AArch64ISD::GLD1_UXTW";
   case AArch64ISD::GLD1_SXTW_SCALED:  return "AArch64ISD::GLD1_SXTW_SCALED";
   case AArch64ISD::GLD1_UXTW_SCALED:  return "AArch64ISD::GLD1_UXTW_SCALED";
   case AArch64ISD::GLD1_IMM:          return "AArch64ISD::GLD1_IMM";
   case AArch64ISD::GLD1S:             return "AArch64ISD::GLD1S";
if (StoreCount > 0)
                                             BaseAddr, LaneLen * Factor);
 
     Ops.push_back(Builder.CreateBitCast(BaseAddr, PtrTy));
     Builder.CreateCall(StNFunc, Ops);
   }
   return true;
 }
 
+// Lower an SVE structured load intrinsic returning a tuple type to target
+// specific intrinsic taking the same input but returning a multi-result value
+// of the split tuple type.
+//
+// E.g. Lowering an LD3:
+//
+//  call <vscale x 12 x i32> @llvm.aarch64.sve.ld3.nxv12i32(
+//                                                    <vscale x 4 x i1> %pred,
+//                                                    <vscale x 4 x i32>* %addr)
+//
+//  Output DAG:
+//
+//    t0: ch = EntryToken
+//        t2: nxv4i1,ch = CopyFromReg t0, Register:nxv4i1 %0
+//        t4: i64,ch = CopyFromReg t0, Register:i64 %1
+//    t5: nxv4i32,nxv4i32,nxv4i32,ch = AArch64ISD::SVE_LD3 t0, t2, t4
+//    t6: nxv12i32 = concat_vectors t5, t5:1, t5:2
+//
+// This is called pre-legalization to avoid widening/splitting issues with
+// non-power-of-2 tuple types used for LD3, such as nxv12i32.
+SDValue AArch64TargetLowering::LowerSVEStructLoad(unsigned Intrinsic,
+                                                  ArrayRef<SDValue> LoadOps,
+                                                  EVT VT, SelectionDAG &DAG,
+                                                  const SDLoc &DL) const {
+  assert(VT.isScalableVector() && "Can only lower scalable vectors");
+
+  unsigned N, Opcode;
+  static std::map<unsigned, std::pair<unsigned, unsigned>> IntrinsicMap = {
+      {Intrinsic::aarch64_sve_ld2, {2, AArch64ISD::SVE_LD2}},
+      {Intrinsic::aarch64_sve_ld3, {3, AArch64ISD::SVE_LD3}},
+      {Intrinsic::aarch64_sve_ld4, {4, AArch64ISD::SVE_LD4}}};
+
+  std::tie(N, Opcode) = IntrinsicMap[Intrinsic];
+  assert(VT.getVectorElementCount().Min % N == 0 &&
+         "invalid tuple vector type!");
+
+  EVT SplitVT = EVT::getVectorVT(*DAG.getContext(), VT.getVectorElementType(),
+                                 VT.getVectorElementCount() / N);
+  assert(isTypeLegal(SplitVT));
+
+  SmallVector<EVT, 5> VTs(N, SplitVT);
+  VTs.push_back(MVT::Other); // Chain
+  SDVTList NodeTys = DAG.getVTList(VTs);
+
+  SDValue PseudoLoad = DAG.getNode(Opcode, DL, NodeTys, LoadOps);
+  SmallVector<SDValue, 4> PseudoLoadOps;
+  for (unsigned I = 0; I < N; ++I)
+    PseudoLoadOps.push_back(SDValue(PseudoLoad.getNode(), I));
+  return DAG.getNode(ISD::CONCAT_VECTORS, DL, VT, PseudoLoadOps);
+}
 
 EVT AArch64TargetLowering::getOptimalMemOpType(
     const MemOp &Op, const AttributeList &FuncAttributes) const {
   bool CanImplicitFloat =
       !FuncAttributes.hasFnAttribute(Attribute::NoImplicitFloat);
   bool CanUseNEON = Subtarget->hasNEON() && CanImplicitFloat;
   bool CanUseFP = Subtarget->hasFPARMv8() && CanImplicitFloat;
   // Only use AdvSIMD to implement memset of 32-byte and above. It would have
case Intrinsic::aarch64_sve_tuple_create4: {
       EVT VT = Opnds[0].getValueType();
       EVT EltVT = VT.getVectorElementType();
       EVT DestVT = EVT::getVectorVT(*DAG.getContext(), EltVT,
                                     VT.getVectorElementCount() *
                                         (N->getNumOperands() - 2));
       SDValue Concat = DAG.getNode(ISD::CONCAT_VECTORS, DL, DestVT, Opnds);
       return DAG.getMergeValues({Concat, Chain}, DL);
     }
+    case Intrinsic::aarch64_sve_ld2:
+    case Intrinsic::aarch64_sve_ld3:
+    case Intrinsic::aarch64_sve_ld4: {
+      SDLoc DL(N);
+      SDValue Chain = N->getOperand(0);
+      SDValue Mask = N->getOperand(2);
+      SDValue BasePtr = N->getOperand(3);
+      SDValue LoadOps[] = {Chain, Mask, BasePtr};
+      unsigned IntrinsicID =
+          cast<ConstantSDNode>(N->getOperand(1))->getZExtValue();
+      SDValue Result =
+          LowerSVEStructLoad(IntrinsicID, LoadOps, N->getValueType(0), DAG, DL);
+      return DAG.getMergeValues({Result, Chain}, DL);
+    }
     default:
       break;
     }
     break;
   case ISD::GlobalAddress:
     return performGlobalAddressCombine(N, DAG, Subtarget, getTargetMachine());
   }
   return SDValue();

llvm/test/CodeGen/AArch64/sve-intrinsics-loads.ll

-; RUN: llc -mtriple=aarch64-linux-gnu -mattr=+sve < %s | FileCheck %s
+; RUN: llc -mtriple=aarch64-linux-gnu -mattr=+sve -asm-verbose=0 < %s | FileCheck %s
 
 ;
 ; LD1RQB
 ;
 
 define <vscale x 16 x i8> @ld1rqb_i8(<vscale x 16 x i1> %pred, i8* %addr) {
 ; CHECK-LABEL: ld1rqb_i8:
 ; CHECK: ld1rqb { z0.b }, p0/z, [x0]

 ; CHECK-LABEL: ldnt1d_f64:
 ; CHECK: ldnt1d { z0.d }, p0/z, [x0]
 ; CHECK-NEXT: ret
   %res = call <vscale x 2 x double> @llvm.aarch64.sve.ldnt1.nxv2f64(<vscale x 2 x i1> %pred,
                                                                     double* %addr)
   ret <vscale x 2 x double> %res
 }
 
+;
+; LD2B
+;
+
+define <vscale x 32 x i8> @ld2b_i8(<vscale x 16 x i1> %pred, <vscale x 16 x i8>* %addr) {
+; CHECK-LABEL: ld2b_i8:
+; CHECK: ld2b { z0.b, z1.b }, p0/z, [x0]
+; CHECK-NEXT: ret
+  %res = call <vscale x 32 x i8> @llvm.aarch64.sve.ld2.nxv32i8.nxv16i1.p0nxv16i8(<vscale x 16 x i1> %pred,
+                                                                                 <vscale x 16 x i8>* %addr)
+  ret <vscale x 32 x i8> %res
+}
+
+;
+; LD2H
+;
+
+define <vscale x 16 x i16> @ld2h_i16(<vscale x 8 x i1> %pred, <vscale x 8 x i16>* %addr) {
+; CHECK-LABEL: ld2h_i16:
+; CHECK: ld2h { z0.h, z1.h }, p0/z, [x0]
+; CHECK-NEXT: ret
+  %res = call <vscale x 16 x i16> @llvm.aarch64.sve.ld2.nxv16i16.nxv8i1.p0nxv8i16(<vscale x 8 x i1> %pred,
+                                                                                  <vscale x 8 x i16>* %addr)
+  ret <vscale x 16 x i16> %res
+}
+
+define <vscale x 16 x half> @ld2h_f16(<vscale x 8 x i1> %pred, <vscale x 8 x half>* %addr) {
+; CHECK-LABEL: ld2h_f16:
+; CHECK: ld2h { z0.h, z1.h }, p0/z, [x0]
+; CHECK-NEXT: ret
+  %res = call <vscale x 16 x half> @llvm.aarch64.sve.ld2.nxv16f16.nxv8i1.p0nxv8f16(<vscale x 8 x i1> %pred,
+                                                                                   <vscale x 8 x half>* %addr)
+  ret <vscale x 16 x half> %res
+}
+
+;
+; LD2W
+;
+
+define <vscale x 8 x i32> @ld2w_i32(<vscale x 4 x i1> %pred, <vscale x 4 x i32>* %addr) {
+; CHECK-LABEL: ld2w_i32:
+; CHECK: ld2w { z0.s, z1.s }, p0/z, [x0]
+; CHECK-NEXT: ret
+  %res = call <vscale x 8 x i32> @llvm.aarch64.sve.ld2.nxv8i32.nxv4i1.p0nxv4i32(<vscale x 4 x i1> %pred,
+                                                                                <vscale x 4 x i32>* %addr)
+  ret <vscale x 8 x i32> %res
+}
+
+define <vscale x 8 x float> @ld2w_f32(<vscale x 4 x i1> %pred, <vscale x 4 x float>* %addr) {
+; CHECK-LABEL: ld2w_f32:
+; CHECK: ld2w { z0.s, z1.s }, p0/z, [x0]
+; CHECK-NEXT: ret
+  %res = call <vscale x 8 x float> @llvm.aarch64.sve.ld2.nxv8f32.nxv4i1.p0nxv4f32(<vscale x 4 x i1> %pred,
+                                                                                  <vscale x 4 x float>* %addr)
+  ret <vscale x 8 x float> %res
+}
+
+;
+; LD2D
+;
+
+define <vscale x 4 x i64> @ld2d_i64(<vscale x 2 x i1> %pred, <vscale x 2 x i64>* %addr) {
+; CHECK-LABEL: ld2d_i64:
+; CHECK: ld2d { z0.d, z1.d }, p0/z, [x0]
+; CHECK-NEXT: ret
+  %res = call <vscale x 4 x i64> @llvm.aarch64.sve.ld2.nxv4i64.nxv2i1.p0nxv2i64(<vscale x 2 x i1> %pred,
+                                                                                <vscale x 2 x i64>* %addr)
+  ret <vscale x 4 x i64> %res
+}
+
+define <vscale x 4 x double> @ld2d_f64(<vscale x 2 x i1> %pred, <vscale x 2 x double>* %addr) {
+; CHECK-LABEL: ld2d_f64:
+; CHECK: ld2d { z0.d, z1.d }, p0/z, [x0]
+; CHECK-NEXT: ret
+  %res = call <vscale x 4 x double> @llvm.aarch64.sve.ld2.nxv4f64.nxv2i1.p0nxv2f64(<vscale x 2 x i1> %pred,
+                                                                                   <vscale x 2 x double>* %addr)
+  ret <vscale x 4 x double> %res
+}
+
+;
+; LD3B
+;
+
+define <vscale x 48 x i8> @ld3b_i8(<vscale x 16 x i1> %pred, <vscale x 16 x i8>* %addr) {
+; CHECK-LABEL: ld3b_i8:
+; CHECK: ld3b { z0.b, z1.b, z2.b }, p0/z, [x0]
+; CHECK-NEXT: ret
+  %res = call <vscale x 48 x i8> @llvm.aarch64.sve.ld3.nxv48i8.nxv16i1.p0nxv16i8(<vscale x 16 x i1> %pred,
+                                                                                 <vscale x 16 x i8>* %addr)
+  ret <vscale x 48 x i8> %res
+}
+
+;
+; LD3H
+;
+
+define <vscale x 24 x i16> @ld3h_i16(<vscale x 8 x i1> %pred, <vscale x 8 x i16>* %addr) {
+; CHECK-LABEL: ld3h_i16:
+; CHECK: ld3h { z0.h, z1.h, z2.h }, p0/z, [x0]
+; CHECK-NEXT: ret
+  %res = call <vscale x 24 x i16> @llvm.aarch64.sve.ld3.nxv24i16.nxv8i1.p0nxv8i16(<vscale x 8 x i1> %pred,
+                                                                                  <vscale x 8 x i16>* %addr)
+  ret <vscale x 24 x i16> %res
+}
+
+define <vscale x 24 x half> @ld3h_f16(<vscale x 8 x i1> %pred, <vscale x 8 x half>* %addr) {
+; CHECK-LABEL: ld3h_f16:
+; CHECK: ld3h { z0.h, z1.h, z2.h }, p0/z, [x0]
+; CHECK-NEXT: ret
+  %res = call <vscale x 24 x half> @llvm.aarch64.sve.ld3.nxv24f16.nxv8i1.p0nxv8f16(<vscale x 8 x i1> %pred,
+                                                                                   <vscale x 8 x half>* %addr)
+  ret <vscale x 24 x half> %res
+}
+
+;
+; LD3W
+;
+
+define <vscale x 12 x i32> @ld3w_i32(<vscale x 4 x i1> %pred, <vscale x 4 x i32>* %addr) {
+; CHECK-LABEL: ld3w_i32:
+; CHECK: ld3w { z0.s, z1.s, z2.s }, p0/z, [x0]
+; CHECK-NEXT: ret
+  %res = call <vscale x 12 x i32> @llvm.aarch64.sve.ld3.nxv12i32.nxv4i1.p0nxv4i32(<vscale x 4 x i1> %pred,
+                                                                                  <vscale x 4 x i32>* %addr)
+  ret <vscale x 12 x i32> %res
+}
+
+define <vscale x 12 x float> @ld3w_f32(<vscale x 4 x i1> %pred, <vscale x 4 x float>* %addr) {
+; CHECK-LABEL: ld3w_f32:
+; CHECK: ld3w { z0.s, z1.s, z2.s }, p0/z, [x0]
+; CHECK-NEXT: ret
+  %res = call <vscale x 12 x float> @llvm.aarch64.sve.ld3.nxv12f32.nxv4i1.p0nxv4f32(<vscale x 4 x i1> %pred,
+                                                                                    <vscale x 4 x float>* %addr)
+  ret <vscale x 12 x float> %res
+}
+
+;
+; LD3D
+;
+
+define <vscale x 6 x i64> @ld3d_i64(<vscale x 2 x i1> %pred, <vscale x 2 x i64>* %addr) {
+; CHECK-LABEL: ld3d_i64:
+; CHECK: ld3d { z0.d, z1.d, z2.d }, p0/z, [x0]
+; CHECK-NEXT: ret
+  %res = call <vscale x 6 x i64> @llvm.aarch64.sve.ld3.nxv6i64.nxv2i1.p0nxv2i64(<vscale x 2 x i1> %pred,
+                                                                                <vscale x 2 x i64>* %addr)
+  ret <vscale x 6 x i64> %res
+}
+
+define <vscale x 6 x double> @ld3d_f64(<vscale x 2 x i1> %pred, <vscale x 2 x double>* %addr) {
+; CHECK-LABEL: ld3d_f64:
+; CHECK: ld3d { z0.d, z1.d, z2.d }, p0/z, [x0]
+; CHECK-NEXT: ret
+  %res = call <vscale x 6 x double> @llvm.aarch64.sve.ld3.nxv6f64.nxv2i1.p0nxv2f64(<vscale x 2 x i1> %pred,
+                                                                                   <vscale x 2 x double>* %addr)
+  ret <vscale x 6 x double> %res
+}
+
+;
+; LD4B
+;
+
+define <vscale x 64 x i8> @ld4b_i8(<vscale x 16 x i1> %pred, <vscale x 16 x i8>* %addr) {
+; CHECK-LABEL: ld4b_i8:
+; CHECK: ld4b { z0.b, z1.b, z2.b, z3.b }, p0/z, [x0]
+; CHECK-NEXT: ret
+  %res = call <vscale x 64 x i8> @llvm.aarch64.sve.ld4.nxv64i8.nxv16i1.p0nxv16i8(<vscale x 16 x i1> %pred,
+                                                                                 <vscale x 16 x i8>* %addr)
+  ret <vscale x 64 x i8> %res
+}
+
+;
+; LD4H
+;
+
+define <vscale x 32 x i16> @ld4h_i16(<vscale x 8 x i1> %pred, <vscale x 8 x i16>* %addr) {
+; CHECK-LABEL: ld4h_i16:
+; CHECK: ld4h { z0.h, z1.h, z2.h, z3.h }, p0/z, [x0]
+; CHECK-NEXT: ret
+  %res = call <vscale x 32 x i16> @llvm.aarch64.sve.ld4.nxv32i16.nxv8i1.p0nxv8i16(<vscale x 8 x i1> %pred,
+                                                                                  <vscale x 8 x i16>* %addr)
+  ret <vscale x 32 x i16> %res
+}
+
+define <vscale x 32 x half> @ld4h_f16(<vscale x 8 x i1> %pred, <vscale x 8 x half>* %addr) {
+; CHECK-LABEL: ld4h_f16:
+; CHECK: ld4h { z0.h, z1.h, z2.h, z3.h }, p0/z, [x0]
+; CHECK-NEXT: ret
+  %res = call <vscale x 32 x half> @llvm.aarch64.sve.ld4.nxv32f16.nxv8i1.p0nxv8f16(<vscale x 8 x i1> %pred,
+                                                                                   <vscale x 8 x half>* %addr)
+  ret <vscale x 32 x half> %res
+}
+
+;
+; LD4W
+;
+
+define <vscale x 16 x i32> @ld4w_i32(<vscale x 4 x i1> %pred, <vscale x 4 x i32>* %addr) {
+; CHECK-LABEL: ld4w_i32:
+; CHECK: ld4w { z0.s, z1.s, z2.s, z3.s }, p0/z, [x0]
+; CHECK-NEXT: ret
+  %res = call <vscale x 16 x i32> @llvm.aarch64.sve.ld4.nxv16i32.nxv4i1.p0nxv4i32(<vscale x 4 x i1> %pred,
+                                                                                  <vscale x 4 x i32>* %addr)
+  ret <vscale x 16 x i32> %res
+}
+
+define <vscale x 16 x float> @ld4w_f32(<vscale x 4 x i1> %pred, <vscale x 4 x float>* %addr) {
+; CHECK-LABEL: ld4w_f32:
+; CHECK: ld4w { z0.s, z1.s, z2.s, z3.s }, p0/z, [x0]
+; CHECK-NEXT: ret
+  %res = call <vscale x 16 x float> @llvm.aarch64.sve.ld4.nxv16f32.nxv4i1.p0nxv4f32(<vscale x 4 x i1> %pred,
+                                                                                    <vscale x 4 x float>* %addr)
+  ret <vscale x 16 x float> %res
+}
+
+;
+; LD4D
+;
+
+define <vscale x 8 x i64> @ld4d_i64(<vscale x 2 x i1> %pred, <vscale x 2 x i64>* %addr) {
+; CHECK-LABEL: ld4d_i64:
+; CHECK: ld4d { z0.d, z1.d, z2.d, z3.d }, p0/z, [x0]
+; CHECK-NEXT: ret
+  %res = call <vscale x 8 x i64> @llvm.aarch64.sve.ld4.nxv8i64.nxv2i1.p0nxv2i64(<vscale x 2 x i1> %pred,
+                                                                                <vscale x 2 x i64>* %addr)
+  ret <vscale x 8 x i64> %res
+}
+
+define <vscale x 8 x double> @ld4d_f64(<vscale x 2 x i1> %pred, <vscale x 2 x double>* %addr) {
+; CHECK-LABEL: ld4d_f64:
+; CHECK: ld4d { z0.d, z1.d, z2.d, z3.d }, p0/z, [x0]
+; CHECK-NEXT: ret
+  %res = call <vscale x 8 x double> @llvm.aarch64.sve.ld4.nxv8f64.nxv2i1.p0nxv2f64(<vscale x 2 x i1> %pred,
+                                                                                   <vscale x 2 x double>* %addr)
+  ret <vscale x 8 x double> %res
+}
+
+
 declare <vscale x 16 x i8> @llvm.aarch64.sve.ld1rq.nxv16i8(<vscale x 16 x i1>, i8*)
 declare <vscale x 8 x i16> @llvm.aarch64.sve.ld1rq.nxv8i16(<vscale x 8 x i1>, i16*)
 declare <vscale x 4 x i32> @llvm.aarch64.sve.ld1rq.nxv4i32(<vscale x 4 x i1>, i32*)
 declare <vscale x 2 x i64> @llvm.aarch64.sve.ld1rq.nxv2i64(<vscale x 2 x i1>, i64*)
 declare <vscale x 8 x half> @llvm.aarch64.sve.ld1rq.nxv8f16(<vscale x 8 x i1>, half*)
 declare <vscale x 4 x float> @llvm.aarch64.sve.ld1rq.nxv4f32(<vscale x 4 x i1>, float*)
 declare <vscale x 2 x double> @llvm.aarch64.sve.ld1rq.nxv2f64(<vscale x 2 x i1>, double*)
 
 declare <vscale x 16 x i8> @llvm.aarch64.sve.ldnt1.nxv16i8(<vscale x 16 x i1>, i8*)
 declare <vscale x 8 x i16> @llvm.aarch64.sve.ldnt1.nxv8i16(<vscale x 8 x i1>, i16*)
 declare <vscale x 4 x i32> @llvm.aarch64.sve.ldnt1.nxv4i32(<vscale x 4 x i1>, i32*)
 declare <vscale x 2 x i64> @llvm.aarch64.sve.ldnt1.nxv2i64(<vscale x 2 x i1>, i64*)
 declare <vscale x 8 x half> @llvm.aarch64.sve.ldnt1.nxv8f16(<vscale x 8 x i1>, half*)
 declare <vscale x 4 x float> @llvm.aarch64.sve.ldnt1.nxv4f32(<vscale x 4 x i1>, float*)
 declare <vscale x 2 x double> @llvm.aarch64.sve.ldnt1.nxv2f64(<vscale x 2 x i1>, double*)
+
+declare <vscale x 32 x i8> @llvm.aarch64.sve.ld2.nxv32i8.nxv16i1.p0nxv16i8(<vscale x 16 x i1>, <vscale x 16 x i8>*)
+declare <vscale x 16 x i16> @llvm.aarch64.sve.ld2.nxv16i16.nxv8i1.p0nxv8i16(<vscale x 8 x i1>, <vscale x 8 x i16>*)
+declare <vscale x 8 x i32> @llvm.aarch64.sve.ld2.nxv8i32.nxv4i1.p0nxv4i32(<vscale x 4 x i1>, <vscale x 4 x i32>*)
+declare <vscale x 4 x i64> @llvm.aarch64.sve.ld2.nxv4i64.nxv2i1.p0nxv2i64(<vscale x 2 x i1>, <vscale x 2 x i64>*)
+declare <vscale x 16 x half> @llvm.aarch64.sve.ld2.nxv16f16.nxv8i1.p0nxv8f16(<vscale x 8 x i1>, <vscale x 8 x half>*)
+declare <vscale x 8 x float> @llvm.aarch64.sve.ld2.nxv8f32.nxv4i1.p0nxv4f32(<vscale x 4 x i1>, <vscale x 4 x float>*)
+declare <vscale x 4 x double> @llvm.aarch64.sve.ld2.nxv4f64.nxv2i1.p0nxv2f64(<vscale x 2 x i1>, <vscale x 2 x double>*)
+
+declare <vscale x 48 x i8> @llvm.aarch64.sve.ld3.nxv48i8.nxv16i1.p0nxv16i8(<vscale x 16 x i1>, <vscale x 16 x i8>*)
+declare <vscale x 24 x i16> @llvm.aarch64.sve.ld3.nxv24i16.nxv8i1.p0nxv8i16(<vscale x 8 x i1>, <vscale x 8 x i16>*)
+declare <vscale x 12 x i32> @llvm.aarch64.sve.ld3.nxv12i32.nxv4i1.p0nxv4i32(<vscale x 4 x i1>, <vscale x 4 x i32>*)
+declare <vscale x 6 x i64> @llvm.aarch64.sve.ld3.nxv6i64.nxv2i1.p0nxv2i64(<vscale x 2 x i1>, <vscale x 2 x i64>*)
+declare <vscale x 24 x half> @llvm.aarch64.sve.ld3.nxv24f16.nxv8i1.p0nxv8f16(<vscale x 8 x i1>, <vscale x 8 x half>*)
+declare <vscale x 12 x float> @llvm.aarch64.sve.ld3.nxv12f32.nxv4i1.p0nxv4f32(<vscale x 4 x i1>, <vscale x 4 x float>*)
+declare <vscale x 6 x double> @llvm.aarch64.sve.ld3.nxv6f64.nxv2i1.p0nxv2f64(<vscale x 2 x i1>, <vscale x 2 x double>*)
+
+declare <vscale x 64 x i8> @llvm.aarch64.sve.ld4.nxv64i8.nxv16i1.p0nxv16i8(<vscale x 16 x i1>, <vscale x 16 x i8>*)
+declare <vscale x 32 x i16> @llvm.aarch64.sve.ld4.nxv32i16.nxv8i1.p0nxv8i16(<vscale x 8 x i1>, <vscale x 8 x i16>*)
+declare <vscale x 16 x i32> @llvm.aarch64.sve.ld4.nxv16i32.nxv4i1.p0nxv4i32(<vscale x 4 x i1>, <vscale x 4 x i32>*)
+declare <vscale x 8 x i64> @llvm.aarch64.sve.ld4.nxv8i64.nxv2i1.p0nxv2i64(<vscale x 2 x i1>, <vscale x 2 x i64>*)
+declare <vscale x 32 x half> @llvm.aarch64.sve.ld4.nxv32f16.nxv8i1.p0nxv8f16(<vscale x 8 x i1>, <vscale x 8 x half>*)
+declare <vscale x 16 x float> @llvm.aarch64.sve.ld4.nxv16f32.nxv4i1.p0nxv4f32(<vscale x 4 x i1>, <vscale x 4 x float>*)
+declare <vscale x 8 x double> @llvm.aarch64.sve.ld4.nxv8f64.nxv2i1.p0nxv2f64(<vscale x 2 x i1>, <vscale x 2 x double>*)