[llvm][CodeGen] Addressing modes for SVE ldN.

[fpetrogalli]

This code is based on https://reviews.llvm.org/D75751, which in turn depends on https://reviews.llvm.org/D75672 and https://reviews.llvm.org/D75674.

I could split the patch in two patches, one for stN and one for ldN. The stN one would not have any dependencies on the aforementioned patches, but I'd rather not do the split because the code uses the common method: AArch64DAGToDAGISel::findAddrModeSVELoadStore.

Of course, I didn't marry this patch. :) If you think I'd better split it so I can submit the stN part without waiting for the dependencies, I'd happily do it.

Thank you!

Francesco

[fpetrogalli]

I have extracted the stores in a separate patch, as it does not rely on any of the dependencies that are specific to the loads. I will update this patch to removed the stores bits.

The stores' patch is at https://reviews.llvm.org/D77435

[fpetrogalli]

I have removed the stores as they are separately implemented in D77435.

[fpetrogalli]

[NFC] Rebased on top of master, after the change for the stN went in (https://github.com/llvm/llvm-project/commit/897fdec586d9ad4c101738caa723bacdda15a769).

[huihuiz]

Hey @fpetrogalli @sdesmalen , any updates on this patch? This should be the last one to commit for structure load? Or are we using different approach?

[fpetrogalli]

I have rebased on top of master and added the bfloat test cases.

[sdesmalen]

llvm/lib/Target/AArch64/AArch64ISelDAGToDAG.cpp

4837

Why is this a template argument instead of a regular argument with default = 1?

4849–4850

remove line?

4874

How come that ST2..ST4 are not covered here?

llvm/test/CodeGen/AArch64/sve-intrinsics-ldN-reg+imm-addr-mode.ll

453

nit: when testing these pairs, can you at least make sure to test the min/max values?

llvm/test/CodeGen/AArch64/sve-intrinsics-ldN-reg+reg-addr-mode.ll

10

nit: s/gp/pg/ (in both tests)

[fpetrogalli]

Thank you for your review @sdesmalen!

llvm/lib/Target/AArch64/AArch64ISelDAGToDAG.cpp

4837

It is a template because I don't foresee NumVec becoming a runtime value. I have used this in other places, for example in setInfoSVEStN. I could set the default to be 1, but I'd rather call it out in the template invocation.

Your comment made me think I should have also made sure that the values of NumVecs are architecturally valid ==> that's why I have added a static assert in the last version of the patch.

4874

We cannot do for the structured load the same we do for the stores because address optimization is done on the intrinsics for structured stores:

case Intrinsic::aarch64_sve_st2: {
  if (VT == MVT::nxv16i8) {
    SelectPredicatedStore</*Scale=*/0>(Node, 2, AArch64::ST2B,
                                       AArch64::ST2B_IMM);
    return;

While for the loads the address optimization is done on the custom ISD nodes:

case AArch64ISD::SVE_LD2_MERGE_ZERO: {
  if (VT == MVT::nxv16i8) {
    SelectPredicatedLoad</*Scale=*/0>(Node, 2, AArch64::LD2B_IMM,
                                      AArch64::LD2B);
    return;

This means that the method getMemVTFromNode determines the memory type with the following generic code for the stores:

if (isa<MemIntrinsicSDNode>(Root))
  return cast<MemIntrinsicSDNode>(Root)->getMemoryVT();

We might do the same for the loads, but we would have to teach to AArch64TargetLowering::getTgtMemIntrinsic how to extract the memory type from the intrinsics call, like we do with the stores:

switch (Intrinsic) {
case Intrinsic::aarch64_sve_st2:
  return setInfoSVEStN<2>(Info, I);
case Intrinsic::aarch64_sve_st3:
  return setInfoSVEStN<3>(Info, I);
case Intrinsic::aarch64_sve_st4:
  return setInfoSVEStN<4>(Info, I);

But that would probably requires changing all the code that converts the intrinsics of the loads into the SVE_LD*_MERGE_ZERO ISD nodes...

If we want to do this cleanup maybe it is better to do it for the whole "structure store intrinsics/ISD nodes" in a separate patch? As far as I understand it, this would be quite an extensive change because we would have to make sure that the nodes of the structure loads are still using the intrinsic and not the custom ISD node when we enter this method, which is not what is going on now. I believe there is a reason for having converted the intrinsics nodes into custom ISD nodes before getting here in address optimization?

llvm/test/CodeGen/AArch64/sve-intrinsics-ldN-reg+imm-addr-mode.ll

453

I can do that, just making sure you didn't miss the comment on the top of file that explains why I haven't add range checks for all cases:

; NOTE: invalid, upper and lower bound immediate values of the regimm
; addressing mode are checked only for the byte version of each
; instruction (`ld<N>b`), as the code for detecting the immediate is
; common to all instructions, and varies only for the number of
; elements of the structure store, which is <N> = 2, 3, 4.

Let me know if you think this is not a sufficient explanation!

[sdesmalen]

llvm/lib/Target/AArch64/AArch64ISelDAGToDAG.cpp

4837

This means the compiler will have to generate two versions of getPackedVectorTypeFromPredicateType because you expect NumVec to be a constant value. It should instead just pass it as a regular parameter.

One of the reasons that this file contains various templated functions is so that you can target them in ComplexPatterns through TableGen by specifying a specific variant of that function, e.g. SelectSVEShiftImm64<0, 64> for ComplexPattern SVEShiftImm64.

4874

Thanks for explaining, I guess I'm happy keeping it as it is now.

llvm/test/CodeGen/AArch64/sve-intrinsics-ldN-reg+imm-addr-mode.ll

453

I meant using #-32 and #28 instead of #16 and #-16 for ld4.nxv8i64 and nxv8f64 respectively.
(and similar for other tests in this file)

[fpetrogalli]

I have addressed your feedback. Thank you @sdesmalen.

[fpetrogalli]

llvm/test/CodeGen/AArch64/sve-intrinsics-ldN-reg+imm-addr-mode.ll

453

Sorry, I don't understand why -32/28 would be better than -16/16. The values you are suggesting are tested to be lower and upper bound for ld4<T> in the byte case, ld4b, between lines 305-323.

The range of the immediate does not depend on the scalar data type being loaded, but only on the number of vector loaded by the instruction, so I think we are goo in testing it only for ld<N>b.

Or am I missing something?

[fpetrogalli]

I have modified the testing of the reg+imm addressing mode to use upper and lower bound instead of strictly in-range values.

Thank you,

Francesco

[sdesmalen]

llvm/lib/Target/AArch64/AArch64ISelDAGToDAG.cpp

248–250

Sorry, one more thing, can you make Scale a normal argument rather than a template argument?
I see that D77435 also uses a template argument for the stores, rather than a regular argument. I guess you can fix that one separately in an NFC patch.

[fpetrogalli]

Address last feedback on template parameters.

llvm/lib/Target/AArch64/AArch64ISelDAGToDAG.cpp

248–250

NFC committed as https://github.com/llvm/llvm-project/commit/dbeb184b7f54db2d3ef20ac153b1c77f81cf0b99

[sdesmalen]

LGTM, thanks!

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);
+  template <unsigned Scale>
+  void SelectPredicatedLoad(SDNode *N, unsigned NumVecs, const unsigned Opc_rr,
+                            const unsigned Opc_ri);
 
   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>
AArch64DAGToDAGISel::findAddrModeSVELoadStore(SDNode *N, const unsigned Opc_rr,
   // 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);
 }
 
+template <unsigned Scale>
 void AArch64DAGToDAGISel::SelectPredicatedLoad(SDNode *N, unsigned NumVecs,
-                                               const unsigned Opc) {
+                                               const unsigned Opc_ri,
+                                               const unsigned Opc_rr) {
+  static_assert(Scale < 4, "Invalid scaling value.");
   SDLoc DL(N);
   EVT VT = N->getValueType(0);
   SDValue Chain = N->getOperand(0);
 
+  // Optimize addressing mode.
+  SDValue Base, Offset;
+  unsigned Opc;
+  std::tie(Opc, Base, Offset) = findAddrModeSVELoadStore<Scale>(
+      N, Opc_rr, Opc_ri, N->getOperand(2),
+      CurDAG->getTargetConstant(0, DL, MVT::i64));
+
   SDValue Ops[] = {N->getOperand(1), // Predicate
-                   N->getOperand(2), // Memory operand
-                   CurDAG->getTargetConstant(0, DL, MVT::i64), Chain};
+                   Base,             // Memory operand
+                   Offset, 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));
if (VT == MVT::v16i8 || VT == MVT::v8i8) {
                VT == MVT::v1f64) {
       SelectPostStoreLane(Node, 4, AArch64::ST4i64_POST);
       return;
     }
     break;
   }
   case AArch64ISD::SVE_LD2_MERGE_ZERO: {
     if (VT == MVT::nxv16i8) {
-      SelectPredicatedLoad(Node, 2, AArch64::LD2B_IMM);
+      SelectPredicatedLoad</*Scale=*/0>(Node, 2, AArch64::LD2B_IMM,
+                                        AArch64::LD2B);
       return;
     } else if (VT == MVT::nxv8i16 || VT == MVT::nxv8f16 ||
                (VT == MVT::nxv8bf16 && Subtarget->hasBF16())) {
-      SelectPredicatedLoad(Node, 2, AArch64::LD2H_IMM);
+      SelectPredicatedLoad</*Scale=*/1>(Node, 2, AArch64::LD2H_IMM,
+                                        AArch64::LD2H);
       return;
     } else if (VT == MVT::nxv4i32 || VT == MVT::nxv4f32) {
-      SelectPredicatedLoad(Node, 2, AArch64::LD2W_IMM);
+      SelectPredicatedLoad</*Scale=*/2>(Node, 2, AArch64::LD2W_IMM,
+                                        AArch64::LD2W);
       return;
     } else if (VT == MVT::nxv2i64 || VT == MVT::nxv2f64) {
-      SelectPredicatedLoad(Node, 2, AArch64::LD2D_IMM);
+      SelectPredicatedLoad</*Scale=*/3>(Node, 2, AArch64::LD2D_IMM,
+                                        AArch64::LD2D);
       return;
     }
     break;
   }
   case AArch64ISD::SVE_LD3_MERGE_ZERO: {
     if (VT == MVT::nxv16i8) {
-      SelectPredicatedLoad(Node, 3, AArch64::LD3B_IMM);
+      SelectPredicatedLoad</*Scale=*/0>(Node, 3, AArch64::LD3B_IMM,
+                                        AArch64::LD3B);
       return;
     } else if (VT == MVT::nxv8i16 || VT == MVT::nxv8f16 ||
                (VT == MVT::nxv8bf16 && Subtarget->hasBF16())) {
-      SelectPredicatedLoad(Node, 3, AArch64::LD3H_IMM);
+      SelectPredicatedLoad</*Scale=*/1>(Node, 3, AArch64::LD3H_IMM,
+                                        AArch64::LD3H);
       return;
     } else if (VT == MVT::nxv4i32 || VT == MVT::nxv4f32) {
-      SelectPredicatedLoad(Node, 3, AArch64::LD3W_IMM);
+      SelectPredicatedLoad</*Scale=*/2>(Node, 3, AArch64::LD3W_IMM,
+                                        AArch64::LD3W);
       return;
     } else if (VT == MVT::nxv2i64 || VT == MVT::nxv2f64) {
-      SelectPredicatedLoad(Node, 3, AArch64::LD3D_IMM);
+      SelectPredicatedLoad</*Scale=*/3>(Node, 3, AArch64::LD3D_IMM,
+                                        AArch64::LD3D);
       return;
     }
     break;
   }
   case AArch64ISD::SVE_LD4_MERGE_ZERO: {
     if (VT == MVT::nxv16i8) {
-      SelectPredicatedLoad(Node, 4, AArch64::LD4B_IMM);
+      SelectPredicatedLoad</*Scale=*/0>(Node, 4, AArch64::LD4B_IMM,
+                                        AArch64::LD4B);
       return;
     } else if (VT == MVT::nxv8i16 || VT == MVT::nxv8f16 ||
                (VT == MVT::nxv8bf16 && Subtarget->hasBF16())) {
-      SelectPredicatedLoad(Node, 4, AArch64::LD4H_IMM);
+      SelectPredicatedLoad</*Scale=*/1>(Node, 4, AArch64::LD4H_IMM,
+                                        AArch64::LD4H);
       return;
     } else if (VT == MVT::nxv4i32 || VT == MVT::nxv4f32) {
-      SelectPredicatedLoad(Node, 4, AArch64::LD4W_IMM);
+      SelectPredicatedLoad</*Scale=*/2>(Node, 4, AArch64::LD4W_IMM,
+                                        AArch64::LD4W);
       return;
     } else if (VT == MVT::nxv2i64 || VT == MVT::nxv2f64) {
-      SelectPredicatedLoad(Node, 4, AArch64::LD4D_IMM);
+      SelectPredicatedLoad</*Scale=*/3>(Node, 4, AArch64::LD4D_IMM,
+                                        AArch64::LD4D);
       return;
     }
     break;
   }
   }
 
   // Select the default instruction
   SelectCode(Node);
 }
 
 /// createAArch64ISelDag - This pass converts a legalized DAG into a
 /// AArch64-specific DAG, ready for instruction scheduling.
 FunctionPass *llvm::createAArch64ISelDag(AArch64TargetMachine &TM,
                                          CodeGenOpt::Level OptLevel) {
   return new AArch64DAGToDAGISel(TM, OptLevel);
 }
 
 /// When \p PredVT is a scalable vector predicate in the form
 /// MVT::nx<M>xi1, it builds the correspondent scalable vector of
-/// integers MVT::nx<M>xi<bits> s.t. M x bits = 128. If the input
+/// integers MVT::nx<M>xi<bits> s.t. M x bits = 128. When targeting
+/// structured vectors (NumVec >1), the output data type is
+/// MVT::nx<M*NumVec>xi<bits> s.t. M x bits = 128. If the input
 /// PredVT is not in the form MVT::nx<M>xi1, it returns an invalid
 /// EVT.
-static EVT getPackedVectorTypeFromPredicateType(LLVMContext &Ctx, EVT PredVT) {
+static EVT getPackedVectorTypeFromPredicateType(LLVMContext &Ctx, EVT PredVT,
+                                                unsigned NumVec) {
+  assert(NumVec > 0 && NumVec < 5 && "Invalid number of vectors.");
   if (!PredVT.isScalableVector() || PredVT.getVectorElementType() != MVT::i1)
     return EVT();
 
   if (PredVT != MVT::nxv16i1 && PredVT != MVT::nxv8i1 &&
       PredVT != MVT::nxv4i1 && PredVT != MVT::nxv2i1)
     return EVT();
 
   ElementCount EC = PredVT.getVectorElementCount();
   EVT ScalarVT = EVT::getIntegerVT(Ctx, AArch64::SVEBitsPerBlock / EC.Min);
-  EVT MemVT = EVT::getVectorVT(Ctx, ScalarVT, EC);
+  EVT MemVT = EVT::getVectorVT(Ctx, ScalarVT, EC * NumVec);
+
   return MemVT;
 }
 
 /// Return the EVT of the data associated to a memory operation in \p
 /// Root. If such EVT cannot be retrived, it returns an invalid EVT.
 static EVT getMemVTFromNode(LLVMContext &Ctx, SDNode *Root) {
   if (isa<MemSDNode>(Root))
     return cast<MemSDNode>(Root)->getMemoryVT();
 
   if (isa<MemIntrinsicSDNode>(Root))
     return cast<MemIntrinsicSDNode>(Root)->getMemoryVT();
 
   const unsigned Opcode = Root->getOpcode();
   // For custom ISD nodes, we have to look at them individually to extract the
   // type of the data moved to/from memory.
   switch (Opcode) {
   case AArch64ISD::LD1_MERGE_ZERO:
   case AArch64ISD::LD1S_MERGE_ZERO:
   case AArch64ISD::LDNF1_MERGE_ZERO:
   case AArch64ISD::LDNF1S_MERGE_ZERO:
     return cast<VTSDNode>(Root->getOperand(3))->getVT();
   case AArch64ISD::ST1_PRED:
     return cast<VTSDNode>(Root->getOperand(4))->getVT();
+  case AArch64ISD::SVE_LD2_MERGE_ZERO:
+    return getPackedVectorTypeFromPredicateType(
+        Ctx, Root->getOperand(1)->getValueType(0), /*NumVec=*/2);
+  case AArch64ISD::SVE_LD3_MERGE_ZERO:
+    return getPackedVectorTypeFromPredicateType(
+        Ctx, Root->getOperand(1)->getValueType(0), /*NumVec=*/3);
+  case AArch64ISD::SVE_LD4_MERGE_ZERO:
+    return getPackedVectorTypeFromPredicateType(
+        Ctx, Root->getOperand(1)->getValueType(0), /*NumVec=*/4);
   default:
     break;
   }
 
   if (Opcode != ISD::INTRINSIC_VOID)
     return EVT();
 
   const unsigned IntNo =
       cast<ConstantSDNode>(Root->getOperand(1))->getZExtValue();
   if (IntNo != Intrinsic::aarch64_sve_prf)
     return EVT();
 
   // We are using an SVE prefetch intrinsic. Type must be inferred
   // from the width of the predicate.
   return getPackedVectorTypeFromPredicateType(
-      Ctx, Root->getOperand(2)->getValueType(0));
+      Ctx, Root->getOperand(2)->getValueType(0), /*NumVec=*/1);
 }
 
 /// SelectAddrModeIndexedSVE - Attempt selection of the addressing mode:
 /// Base + OffImm * sizeof(MemVT) for Min >= OffImm <= Max
 /// where Root is the memory access using N for its address.
 template <int64_t Min, int64_t Max>
 bool AArch64DAGToDAGISel::SelectAddrModeIndexedSVE(SDNode *Root, SDValue N,
                                                    SDValue &Base,

llvm/test/CodeGen/AArch64/sve-intrinsics-ldN-reg+imm-addr-mode.ll

+; RUN: llc -mtriple=aarch64--linux-gnu -mattr=sve < %s | FileCheck %s
+
+; NOTE: invalid, upper and lower bound immediate values of the regimm
+; addressing mode are checked only for the byte version of each
+; instruction (`ld<N>b`), as the code for detecting the immediate is
+; common to all instructions, and varies only for the number of
+; elements of the structure store, which is <N> = 2, 3, 4.
+
+; ld2b
+define <vscale x 32 x i8> @ld2.nxv32i8(<vscale x 16 x i1> %Pg, <vscale x 16 x i8> *%addr) {
+; CHECK-LABEL: ld2.nxv32i8:
+; CHECK: ld2b { z0.b, z1.b }, p0/z, [x0, #2, mul vl]
+; CHECK-NEXT: ret
+%base = getelementptr <vscale x 16 x i8>, <vscale x 16 x i8>* %addr, i64 2
+%base_ptr = bitcast <vscale x 16 x i8>* %base to i8*
+%res = call <vscale x 32 x i8> @llvm.aarch64.sve.ld2.nxv32i8.nxv16i1.p0i8(<vscale x 16 x i1> %Pg, i8 *%base_ptr)
+ret <vscale x 32 x i8> %res
+}
+
+define <vscale x 32 x i8> @ld2.nxv32i8_lower_bound(<vscale x 16 x i1> %Pg, <vscale x 16 x i8> *%addr) {
+; CHECK-LABEL: ld2.nxv32i8_lower_bound:
+; CHECK: ld2b { z0.b, z1.b }, p0/z, [x0, #-16, mul vl]
+; CHECK-NEXT: ret
+%base = getelementptr <vscale x 16 x i8>, <vscale x 16 x i8>* %addr, i64 -16
+%base_ptr = bitcast <vscale x 16 x i8>* %base to i8 *
+%res = call <vscale x 32 x i8> @llvm.aarch64.sve.ld2.nxv32i8.nxv16i1.p0i8(<vscale x 16 x i1> %Pg, i8 *%base_ptr)
+ret <vscale x 32 x i8> %res
+}
+
+define <vscale x 32 x i8> @ld2.nxv32i8_upper_bound(<vscale x 16 x i1> %Pg, <vscale x 16 x i8> *%addr) {
+; CHECK-LABEL: ld2.nxv32i8_upper_bound:
+; CHECK: ld2b { z0.b, z1.b }, p0/z, [x0, #14, mul vl]
+; CHECK-NEXT: ret
+%base = getelementptr <vscale x 16 x i8>, <vscale x 16 x i8>* %addr, i64 14
+%base_ptr = bitcast <vscale x 16 x i8>* %base to i8 *
+%res = call <vscale x 32 x i8> @llvm.aarch64.sve.ld2.nxv32i8.nxv16i1.p0i8(<vscale x 16 x i1> %Pg, i8 *%base_ptr)
+ret <vscale x 32 x i8> %res
+}
+
+define <vscale x 32 x i8> @ld2.nxv32i8_not_multiple_of_2(<vscale x 16 x i1> %Pg, <vscale x 16 x i8> *%addr) {
+; CHECK-LABEL: ld2.nxv32i8_not_multiple_of_2:
+; CHECK:      rdvl x[[OFFSET:[0-9]]], #3
+; CHECK-NEXT: ld2b { z0.b, z1.b }, p0/z, [x0, x[[OFFSET]]]
+; CHECK-NEXT: ret
+%base = getelementptr <vscale x 16 x i8>, <vscale x 16 x i8>* %addr, i64 3
+%base_ptr = bitcast <vscale x 16 x i8>* %base to i8 *
+%res = call <vscale x 32 x i8> @llvm.aarch64.sve.ld2.nxv32i8.nxv16i1.p0i8(<vscale x 16 x i1> %Pg, i8 *%base_ptr)
+ret <vscale x 32 x i8> %res
+}
+
+define <vscale x 32 x i8> @ld2.nxv32i8_outside_lower_bound(<vscale x 16 x i1> %Pg, <vscale x 16 x i8> *%addr) {
+; CHECK-LABEL: ld2.nxv32i8_outside_lower_bound:
+; CHECK:      rdvl x[[OFFSET:[0-9]]], #-18
+; CHECK-NEXT: ld2b { z0.b, z1.b }, p0/z, [x0, x[[OFFSET]]]
+; CHECK-NEXT: ret
+%base = getelementptr <vscale x 16 x i8>, <vscale x 16 x i8>* %addr, i64 -18
+%base_ptr = bitcast <vscale x 16 x i8>* %base to i8 *
+%res = call <vscale x 32 x i8> @llvm.aarch64.sve.ld2.nxv32i8.nxv16i1.p0i8(<vscale x 16 x i1> %Pg, i8 *%base_ptr)
+ret <vscale x 32 x i8> %res
+}
+
+define <vscale x 32 x i8> @ld2.nxv32i8_outside_upper_bound(<vscale x 16 x i1> %Pg, <vscale x 16 x i8> *%addr) {
+; CHECK-LABEL: ld2.nxv32i8_outside_upper_bound:
+; CHECK:      rdvl x[[OFFSET:[0-9]]], #16
+; CHECK-NEXT: ld2b { z0.b, z1.b }, p0/z, [x0, x[[OFFSET]]]
+; CHECK-NEXT: ret
+%base = getelementptr <vscale x 16 x i8>, <vscale x 16 x i8>* %addr, i64 16
+%base_ptr = bitcast <vscale x 16 x i8>* %base to i8 *
+%res = call <vscale x 32 x i8> @llvm.aarch64.sve.ld2.nxv32i8.nxv16i1.p0i8(<vscale x 16 x i1> %Pg, i8 *%base_ptr)
+ret <vscale x 32 x i8> %res
+}
+
+; ld2h
+define <vscale x 16 x i16> @ld2.nxv16i16(<vscale x 8 x i1> %Pg, <vscale x 8 x i16>* %addr) {
+; CHECK-LABEL: ld2.nxv16i16:
+; CHECK: ld2h { z0.h, z1.h }, p0/z, [x0, #12, mul vl]
+; CHECK-NEXT: ret
+%base = getelementptr <vscale x 8 x i16>, <vscale x 8 x i16>* %addr, i64 12
+%base_ptr = bitcast <vscale x 8 x i16>* %base to i16 *
+%res = call <vscale x 16 x i16> @llvm.aarch64.sve.ld2.nxv16i16.nxv8i1.p0i16(<vscale x 8 x i1> %Pg, i16 *%base_ptr)
+ret <vscale x 16 x i16> %res
+}
+
+define <vscale x 16 x half> @ld2.nxv16f16(<vscale x 8 x i1> %Pg, <vscale x 8 x half>* %addr) {
+; CHECK-LABEL: ld2.nxv16f16:
+; CHECK: ld2h { z0.h, z1.h }, p0/z, [x0, #12, mul vl]
+; CHECK-NEXT: ret
+%base = getelementptr <vscale x 8 x half>, <vscale x 8 x half>* %addr, i64 12
+%base_ptr = bitcast <vscale x 8 x half>* %base to half *
+%res = call <vscale x 16 x half> @llvm.aarch64.sve.ld2.nxv16f16.nxv8i1.p0f16(<vscale x 8 x i1> %Pg, half *%base_ptr)
+ret <vscale x 16 x half> %res
+}
+
+define <vscale x 16 x bfloat> @ld2.nxv16bf16(<vscale x 8 x i1> %Pg, <vscale x 8 x bfloat>* %addr) #0 {
+; CHECK-LABEL: ld2.nxv16bf16:
+; CHECK: ld2h { z0.h, z1.h }, p0/z, [x0, #12, mul vl]
+; CHECK-NEXT: ret
+%base = getelementptr <vscale x 8 x bfloat>, <vscale x 8 x bfloat>* %addr, i64 12
+%base_ptr = bitcast <vscale x 8 x bfloat>* %base to bfloat *
+%res = call <vscale x 16 x bfloat> @llvm.aarch64.sve.ld2.nxv16bf16.nxv8i1.p0bf16(<vscale x 8 x i1> %Pg, bfloat *%base_ptr)
+ret <vscale x 16 x bfloat> %res
+}
+
+; ld2w
+define <vscale x 8 x i32> @ld2.nxv8i32(<vscale x 4 x i1> %Pg, <vscale x 4 x i32>* %addr) {
+; CHECK-LABEL: ld2.nxv8i32:
+; CHECK: ld2w { z0.s, z1.s }, p0/z, [x0, #-12, mul vl]
+; CHECK-NEXT: ret
+%base = getelementptr <vscale x 4 x i32>, <vscale x 4 x i32>* %addr, i64 -12
+%base_ptr = bitcast <vscale x 4 x i32>* %base to i32 *
+%res = call <vscale x 8 x i32> @llvm.aarch64.sve.ld2.nxv8i32.nxv4i1.p0i32(<vscale x 4 x i1> %Pg, i32 *%base_ptr)
+ret <vscale x 8 x i32> %res
+}
+
+define <vscale x 8 x float> @ld2.nxv8f32(<vscale x 4 x i1> %Pg, <vscale x 4 x float>* %addr) {
+; CHECK-LABEL: ld2.nxv8f32:
+; CHECK: ld2w { z0.s, z1.s }, p0/z, [x0, #-14, mul vl]
+; CHECK-NEXT: ret
+%base = getelementptr <vscale x 4 x float>, <vscale x 4 x float>* %addr, i64 -14
+%base_ptr = bitcast <vscale x 4 x float>* %base to float *
+%res = call <vscale x 8 x float> @llvm.aarch64.sve.ld2.nxv8f32.nxv4i1.p0f32(<vscale x 4 x i1> %Pg, float *%base_ptr)
+ret <vscale x 8 x float> %res
+}
+
+; ld2d
+define <vscale x 4 x i64> @ld2.nxv4i64(<vscale x 2 x i1> %Pg, <vscale x 2 x i64>* %addr) {
+; CHECK-LABEL: ld2.nxv4i64:
+; CHECK: ld2d { z0.d, z1.d }, p0/z, [x0, #-14, mul vl]
+; CHECK-NEXT: ret
+%base = getelementptr <vscale x 2 x i64>, <vscale x 2 x i64>* %addr, i64 -14
+%base_ptr = bitcast <vscale x 2 x i64>* %base to i64 *
+%res = call <vscale x 4 x i64> @llvm.aarch64.sve.ld2.nxv4i64.nxv2i1.p0i64(<vscale x 2 x i1> %Pg, i64 *%base_ptr)
+ret <vscale x 4 x i64> %res
+}
+
+define <vscale x 4 x double> @ld2.nxv4f64(<vscale x 2 x i1> %Pg, <vscale x 2 x double>* %addr) {
+; CHECK-LABEL: ld2.nxv4f64:
+; CHECK: ld2d { z0.d, z1.d }, p0/z, [x0, #-16, mul vl]
+; CHECK-NEXT: ret
+%base = getelementptr <vscale x 2 x double>, <vscale x 2 x double>* %addr, i64 -16
+%base_ptr = bitcast <vscale x 2 x double>* %base to double *
+%res = call <vscale x 4 x double> @llvm.aarch64.sve.ld2.nxv4f64.nxv2i1.p0f64(<vscale x 2 x i1> %Pg, double *%base_ptr)
+ret <vscale x 4 x double> %res
+}
+
+; ld3b
+define <vscale x 48 x i8> @ld3.nxv48i8(<vscale x 16 x i1> %Pg, <vscale x 16 x i8> *%addr) {
+; CHECK-LABEL: ld3.nxv48i8:
+; CHECK: ld3b { z0.b, z1.b, z2.b }, p0/z, [x0, #3, mul vl]
+; CHECK-NEXT: ret
+%base = getelementptr <vscale x 16 x i8>, <vscale x 16 x i8>* %addr, i64 3
+%base_ptr = bitcast <vscale x 16 x i8>* %base to i8 *
+%res = call <vscale x 48 x i8> @llvm.aarch64.sve.ld3.nxv48i8.nxv16i1.p0i8(<vscale x 16 x i1> %Pg, i8 *%base_ptr)
+ret <vscale x 48 x i8> %res
+}
+
+define <vscale x 48 x i8> @ld3.nxv48i8_lower_bound(<vscale x 16 x i1> %Pg, <vscale x 16 x i8> *%addr) {
+; CHECK-LABEL: ld3.nxv48i8_lower_bound:
+; CHECK: ld3b { z0.b, z1.b, z2.b }, p0/z, [x0, #-24, mul vl]
+; CHECK-NEXT: ret
+%base = getelementptr <vscale x 16 x i8>, <vscale x 16 x i8>* %addr, i64 -24
+%base_ptr = bitcast <vscale x 16 x i8>* %base to i8 *
+%res = call <vscale x 48 x i8> @llvm.aarch64.sve.ld3.nxv48i8.nxv16i1.p0i8(<vscale x 16 x i1> %Pg, i8 *%base_ptr)
+ret <vscale x 48 x i8> %res
+}
+
+define <vscale x 48 x i8> @ld3.nxv48i8_upper_bound(<vscale x 16 x i1> %Pg, <vscale x 16 x i8> *%addr) {
+; CHECK-LABEL: ld3.nxv48i8_upper_bound:
+; CHECK: ld3b { z0.b, z1.b, z2.b }, p0/z, [x0, #21, mul vl]
+; CHECK-NEXT: ret
+%base = getelementptr <vscale x 16 x i8>, <vscale x 16 x i8>* %addr, i64 21
+%base_ptr = bitcast <vscale x 16 x i8>* %base to i8 *
+%res = call <vscale x 48 x i8> @llvm.aarch64.sve.ld3.nxv48i8.nxv16i1.p0i8(<vscale x 16 x i1> %Pg, i8 *%base_ptr)
+ret <vscale x 48 x i8> %res
+}
+
+define <vscale x 48 x i8> @ld3.nxv48i8_not_multiple_of_3_01(<vscale x 16 x i1> %Pg, <vscale x 16 x i8> *%addr) {
+; CHECK-LABEL: ld3.nxv48i8_not_multiple_of_3_01:
+; CHECK:      rdvl x[[OFFSET:[0-9]]], #4
+; CHECK-NEXT: ld3b { z0.b, z1.b, z2.b }, p0/z, [x0, x[[OFFSET]]]
+; CHECK-NEXT: ret
+%base = getelementptr <vscale x 16 x i8>, <vscale x 16 x i8>* %addr, i64 4
+%base_ptr = bitcast <vscale x 16 x i8>* %base to i8 *
+%res = call <vscale x 48 x i8> @llvm.aarch64.sve.ld3.nxv48i8.nxv16i1.p0i8(<vscale x 16 x i1> %Pg, i8 *%base_ptr)
+ret <vscale x 48 x i8> %res
+}
+
+define <vscale x 48 x i8> @ld3.nxv48i8_not_multiple_of_3_02(<vscale x 16 x i1> %Pg, <vscale x 16 x i8> *%addr) {
+; CHECK-LABEL: ld3.nxv48i8_not_multiple_of_3_02:
+; CHECK:      rdvl x[[OFFSET:[0-9]]], #5
+; CHECK-NEXT: ld3b { z0.b, z1.b, z2.b }, p0/z, [x0, x[[OFFSET]]]
+; CHECK-NEXT: ret
+%base = getelementptr <vscale x 16 x i8>, <vscale x 16 x i8>* %addr, i64 5
+%base_ptr = bitcast <vscale x 16 x i8>* %base to i8 *
+%res = call <vscale x 48 x i8> @llvm.aarch64.sve.ld3.nxv48i8.nxv16i1.p0i8(<vscale x 16 x i1> %Pg, i8 *%base_ptr)
+ret <vscale x 48 x i8> %res
+}
+
+define <vscale x 48 x i8> @ld3.nxv48i8_outside_lower_bound(<vscale x 16 x i1> %Pg, <vscale x 16 x i8> *%addr) {
+; CHECK-LABEL: ld3.nxv48i8_outside_lower_bound:
+; CHECK:      rdvl x[[OFFSET:[0-9]]], #-27
+; CHECK-NEXT: ld3b { z0.b, z1.b, z2.b }, p0/z, [x0, x[[OFFSET]]]
+; CHECK-NEXT: ret
+%base = getelementptr <vscale x 16 x i8>, <vscale x 16 x i8>* %addr, i64 -27
+%base_ptr = bitcast <vscale x 16 x i8>* %base to i8 *
+%res = call <vscale x 48 x i8> @llvm.aarch64.sve.ld3.nxv48i8.nxv16i1.p0i8(<vscale x 16 x i1> %Pg, i8 *%base_ptr)
+ret <vscale x 48 x i8> %res
+}
+
+define <vscale x 48 x i8> @ld3.nxv48i8_outside_upper_bound(<vscale x 16 x i1> %Pg, <vscale x 16 x i8> *%addr) {
+; CHECK-LABEL: ld3.nxv48i8_outside_upper_bound:
+; CHECK:      rdvl x[[OFFSET:[0-9]]], #24
+; CHECK-NEXT: ld3b { z0.b, z1.b, z2.b }, p0/z, [x0, x[[OFFSET]]]
+; CHECK-NEXT: ret
+%base = getelementptr <vscale x 16 x i8>, <vscale x 16 x i8>* %addr, i64 24
+%base_ptr = bitcast <vscale x 16 x i8>* %base to i8 *
+%res = call <vscale x 48 x i8> @llvm.aarch64.sve.ld3.nxv48i8.nxv16i1.p0i8(<vscale x 16 x i1> %Pg, i8 *%base_ptr)
+ret <vscale x 48 x i8> %res
+}
+
+; ld3h
+define <vscale x 24 x i16> @ld3.nxv24i16(<vscale x 8 x i1> %Pg, <vscale x 8 x i16> *%addr) {
+; CHECK-LABEL: ld3.nxv24i16:
+; CHECK: ld3h { z0.h, z1.h, z2.h }, p0/z, [x0, #6, mul vl]
+; CHECK-NEXT: ret
+%base = getelementptr <vscale x 8 x i16>, <vscale x 8 x i16>* %addr, i64 6
+%base_ptr = bitcast <vscale x 8 x i16>* %base to i16 *
+%res = call <vscale x 24 x i16> @llvm.aarch64.sve.ld3.nxv24i16.nxv8i1.p0i16(<vscale x 8 x i1> %Pg, i16 *%base_ptr)
+ret <vscale x 24 x i16> %res
+}
+
+define <vscale x 24 x half> @ld3.nxv24f16(<vscale x 8 x i1> %Pg, <vscale x 8 x half> *%addr) {
+; CHECK-LABEL: ld3.nxv24f16:
+; CHECK: ld3h { z0.h, z1.h, z2.h }, p0/z, [x0, #6, mul vl]
+; CHECK-NEXT: ret
+%base = getelementptr <vscale x 8 x half>, <vscale x 8 x half>* %addr, i64 6
+%base_ptr = bitcast <vscale x 8 x half>* %base to half *
+%res = call <vscale x 24 x half> @llvm.aarch64.sve.ld3.nxv24f16.nxv8i1.p0f16(<vscale x 8 x i1> %Pg, half *%base_ptr)
+ret <vscale x 24 x half> %res
+}
+
+define <vscale x 24 x bfloat> @ld3.nxv24bf16(<vscale x 8 x i1> %Pg, <vscale x 8 x bfloat> *%addr) #0 {
+; CHECK-LABEL: ld3.nxv24bf16:
+; CHECK: ld3h { z0.h, z1.h, z2.h }, p0/z, [x0, #6, mul vl]
+; CHECK-NEXT: ret
+%base = getelementptr <vscale x 8 x bfloat>, <vscale x 8 x bfloat>* %addr, i64 6
+%base_ptr = bitcast <vscale x 8 x bfloat>* %base to bfloat *
+%res = call <vscale x 24 x bfloat> @llvm.aarch64.sve.ld3.nxv24bf16.nxv8i1.p0bf16(<vscale x 8 x i1> %Pg, bfloat *%base_ptr)
+ret <vscale x 24 x bfloat> %res
+}
+
+; ld3w
+define <vscale x 12 x i32> @ld3.nxv12i32(<vscale x 4 x i1> %Pg, <vscale x 4 x i32> *%addr) {
+; CHECK-LABEL: ld3.nxv12i32:
+; CHECK: ld3w { z0.s, z1.s, z2.s }, p0/z, [x0, #9, mul vl]
+; CHECK-NEXT: ret
+%base = getelementptr <vscale x 4 x i32>, <vscale x 4 x i32>* %addr, i64 9
+%base_ptr = bitcast <vscale x 4 x i32>* %base to i32 *
+%res = call <vscale x 12 x i32> @llvm.aarch64.sve.ld3.nxv12i32.nxv4i1.p0i32(<vscale x 4 x i1> %Pg, i32 *%base_ptr)
+ret <vscale x 12 x i32> %res
+}
+
+define <vscale x 12 x float> @ld3.nxv12f32(<vscale x 4 x i1> %Pg, <vscale x 4 x float> *%addr) {
+; CHECK-LABEL: ld3.nxv12f32:
+; CHECK: ld3w { z0.s, z1.s, z2.s }, p0/z, [x0, #-9, mul vl]
+; CHECK-NEXT: ret
+%base = getelementptr <vscale x 4 x float>, <vscale x 4 x float>* %addr, i64 -9
+%base_ptr = bitcast <vscale x 4 x float>* %base to float *
+%res = call <vscale x 12 x float> @llvm.aarch64.sve.ld3.nxv12f32.nxv4i1.p0f32(<vscale x 4 x i1> %Pg, float *%base_ptr)
+ret <vscale x 12 x float> %res
+}
+
+; ld3d
+define <vscale x 6 x i64> @ld3.nxv6i64(<vscale x 2 x i1> %Pg, <vscale x 2 x i64> *%addr) {
+; CHECK-LABEL: ld3.nxv6i64:
+; CHECK: ld3d { z0.d, z1.d, z2.d }, p0/z, [x0, #12, mul vl]
+; CHECK-NEXT: ret
+%base = getelementptr <vscale x 2 x i64>, <vscale x 2 x i64>* %addr, i64 12
+%base_ptr = bitcast <vscale x 2 x i64>* %base to i64 *
+%res = call <vscale x 6 x i64> @llvm.aarch64.sve.ld3.nxv6i64.nxv2i1.p0i64(<vscale x 2 x i1> %Pg, i64 *%base_ptr)
+ret <vscale x 6 x i64> %res
+}
+
+define <vscale x 6 x double> @ld3.nxv6f64(<vscale x 2 x i1> %Pg, <vscale x 2 x double> *%addr) {
+; CHECK-LABEL: ld3.nxv6f64:
+; CHECK: ld3d { z0.d, z1.d, z2.d }, p0/z, [x0, #-12, mul vl]
+; CHECK-NEXT: ret
+%base = getelementptr <vscale x 2 x double>, <vscale x 2 x double>* %addr, i64 -12
+%base_ptr = bitcast <vscale x 2 x double>* %base to double *
+%res = call <vscale x 6 x double> @llvm.aarch64.sve.ld3.nxv6f64.nxv2i1.p0f64(<vscale x 2 x i1> %Pg, double *%base_ptr)
+ret <vscale x 6 x double> %res
+}
+
+; ; ld4b
+define <vscale x 64 x i8> @ld4.nxv64i8(<vscale x 16 x i1> %Pg, <vscale x 16 x i8> *%addr) {
+; CHECK-LABEL: ld4.nxv64i8:
+; CHECK: ld4b { z0.b, z1.b, z2.b, z3.b }, p0/z, [x0, #4, mul vl]
+; CHECK-NEXT: ret
+%base = getelementptr <vscale x 16 x i8>, <vscale x 16 x i8>* %addr, i64 4
+%base_ptr = bitcast <vscale x 16 x i8>* %base to i8 *
+%res = call <vscale x 64 x i8> @llvm.aarch64.sve.ld4.nxv64i8.nxv16i1.p0i8(<vscale x 16 x i1> %Pg, i8 *%base_ptr)
+ret <vscale x 64 x i8> %res
+}
+
+define <vscale x 64 x i8> @ld4.nxv64i8_lower_bound(<vscale x 16 x i1> %Pg, <vscale x 16 x i8> *%addr) {
+; CHECK-LABEL: ld4.nxv64i8_lower_bound:
+; CHECK: ld4b { z0.b, z1.b, z2.b, z3.b }, p0/z, [x0, #-32, mul vl]
+; CHECK-NEXT: ret
+%base = getelementptr <vscale x 16 x i8>, <vscale x 16 x i8>* %addr, i64 -32
+%base_ptr = bitcast <vscale x 16 x i8>* %base to i8 *
+%res = call <vscale x 64 x i8> @llvm.aarch64.sve.ld4.nxv64i8.nxv16i1.p0i8(<vscale x 16 x i1> %Pg, i8 *%base_ptr)
+ret <vscale x 64 x i8> %res
+}
+
+define <vscale x 64 x i8> @ld4.nxv64i8_upper_bound(<vscale x 16 x i1> %Pg, <vscale x 16 x i8> *%addr) {
+; CHECK-LABEL: ld4.nxv64i8_upper_bound:
+; CHECK: ld4b { z0.b, z1.b, z2.b, z3.b }, p0/z, [x0, #28, mul vl]
+; CHECK-NEXT: ret
+%base = getelementptr <vscale x 16 x i8>, <vscale x 16 x i8>* %addr, i64 28
+%base_ptr = bitcast <vscale x 16 x i8>* %base to i8 *
+%res = call <vscale x 64 x i8> @llvm.aarch64.sve.ld4.nxv64i8.nxv16i1.p0i8(<vscale x 16 x i1> %Pg, i8 *%base_ptr)
+ret <vscale x 64 x i8> %res
+}
+
+define <vscale x 64 x i8> @ld4.nxv64i8_not_multiple_of_4_01(<vscale x 16 x i1> %Pg, <vscale x 16 x i8> *%addr) {
+; CHECK-LABEL: ld4.nxv64i8_not_multiple_of_4_01:
+; CHECK:      rdvl x[[OFFSET:[0-9]]], #5
+; CHECK-NEXT: ld4b { z0.b, z1.b, z2.b, z3.b }, p0/z, [x0, x[[OFFSET]]]
+; CHECK-NEXT: ret
+%base = getelementptr <vscale x 16 x i8>, <vscale x 16 x i8>* %addr, i64 5
+%base_ptr = bitcast <vscale x 16 x i8>* %base to i8 *
+%res = call <vscale x 64 x i8> @llvm.aarch64.sve.ld4.nxv64i8.nxv16i1.p0i8(<vscale x 16 x i1> %Pg, i8 *%base_ptr)
+ret <vscale x 64 x i8> %res
+}
+
+define <vscale x 64 x i8> @ld4.nxv64i8_not_multiple_of_4_02(<vscale x 16 x i1> %Pg, <vscale x 16 x i8> *%addr) {
+; CHECK-LABEL: ld4.nxv64i8_not_multiple_of_4_02:
+; CHECK:      rdvl x[[OFFSET:[0-9]]], #6
+; CHECK-NEXT: ld4b { z0.b, z1.b, z2.b, z3.b }, p0/z, [x0, x[[OFFSET]]]
+; CHECK-NEXT: ret
+%base = getelementptr <vscale x 16 x i8>, <vscale x 16 x i8>* %addr, i64 6
+%base_ptr = bitcast <vscale x 16 x i8>* %base to i8 *
+%res = call <vscale x 64 x i8> @llvm.aarch64.sve.ld4.nxv64i8.nxv16i1.p0i8(<vscale x 16 x i1> %Pg, i8 *%base_ptr)
+ret <vscale x 64 x i8> %res
+}
+
+define <vscale x 64 x i8> @ld4.nxv64i8_not_multiple_of_4_03(<vscale x 16 x i1> %Pg, <vscale x 16 x i8> *%addr) {
+; CHECK-LABEL: ld4.nxv64i8_not_multiple_of_4_03:
+; CHECK:      rdvl x[[OFFSET:[0-9]]], #7
+; CHECK-NEXT: ld4b { z0.b, z1.b, z2.b, z3.b }, p0/z, [x0, x[[OFFSET]]]
+; CHECK-NEXT: ret
+%base = getelementptr <vscale x 16 x i8>, <vscale x 16 x i8>* %addr, i64 7
+%base_ptr = bitcast <vscale x 16 x i8>* %base to i8 *
+%res = call <vscale x 64 x i8> @llvm.aarch64.sve.ld4.nxv64i8.nxv16i1.p0i8(<vscale x 16 x i1> %Pg, i8 *%base_ptr)
+ret <vscale x 64 x i8> %res
+}
+
+define <vscale x 64 x i8> @ld4.nxv64i8_outside_lower_bound(<vscale x 16 x i1> %Pg, <vscale x 16 x i8> *%addr) {
+; CHECK-LABEL: ld4.nxv64i8_outside_lower_bound:
+; FIXME: optimize OFFSET computation so that xOFFSET = (mul (RDVL #4) #9)
+; xM = -9 * 2^6
+; xP = RDVL * 2^-4 
+; xOFFSET = RDVL * 2^-4 * -9 * 2^6 = RDVL * -36
+; CHECK:      rdvl x[[N:[0-9]]], #1
+; CHECK-DAG:  mov  x[[M:[0-9]]], #-576
+; CHECK-DAG:  lsr  x[[P:[0-9]]], x[[N]], #4
+; CHECK-DAG:  mul  x[[OFFSET:[0-9]]], x[[P]], x[[M]]
+; CHECK-NEXT: ld4b { z0.b, z1.b, z2.b, z3.b }, p0/z, [x0, x[[OFFSET]]]
+; CHECK-NEXT: ret
+%base = getelementptr <vscale x 16 x i8>, <vscale x 16 x i8>* %addr, i64 -36
+%base_ptr = bitcast <vscale x 16 x i8>* %base to i8 *
+%res = call <vscale x 64 x i8> @llvm.aarch64.sve.ld4.nxv64i8.nxv16i1.p0i8(<vscale x 16 x i1> %Pg, i8 *%base_ptr)
+ret <vscale x 64 x i8> %res
+}
+
+define <vscale x 64 x i8> @ld4.nxv64i8_outside_upper_bound(<vscale x 16 x i1> %Pg, <vscale x 16 x i8> *%addr) {
+; CHECK-LABEL: ld4.nxv64i8_outside_upper_bound:
+; FIXME: optimize OFFSET computation so that xOFFSET = (mul (RDVL #16) #2)
+; xM = 2^9
+; xP = RDVL * 2^-4
+; xOFFSET = RDVL * 2^-4 * 2^9 = RDVL * 32
+; CHECK:      rdvl x[[N:[0-9]]], #1
+; CHECK-DAG:  mov  w[[M:[0-9]]], #512
+; CHECK-DAG:  lsr  x[[P:[0-9]]], x[[N]], #4
+; CHECK-DAG:  mul  x[[OFFSET:[0-9]]], x[[P]], x[[M]]
+; CHECK-NEXT: ld4b { z0.b, z1.b, z2.b, z3.b }, p0/z, [x0, x[[OFFSET]]]
+; CHECK-NEXT: ret
+%base = getelementptr <vscale x 16 x i8>, <vscale x 16 x i8>* %addr, i64 32
+%base_ptr = bitcast <vscale x 16 x i8>* %base to i8 *
+%res = call <vscale x 64 x i8> @llvm.aarch64.sve.ld4.nxv64i8.nxv16i1.p0i8(<vscale x 16 x i1> %Pg, i8 *%base_ptr)
+ret <vscale x 64 x i8> %res
+}
+
+; ld4h
+define <vscale x 32 x i16> @ld4.nxv32i16(<vscale x 8 x i1> %Pg, <vscale x 8 x i16> *%addr) {
+; CHECK-LABEL: ld4.nxv32i16:
+; CHECK: ld4h { z0.h, z1.h, z2.h, z3.h }, p0/z, [x0, #8, mul vl]
+; CHECK-NEXT: ret
+%base = getelementptr <vscale x 8 x i16>, <vscale x 8 x i16>* %addr, i64 8
+%base_ptr = bitcast <vscale x 8 x i16>* %base to i16 *
+%res = call <vscale x 32 x i16> @llvm.aarch64.sve.ld4.nxv32i16.nxv8i1.p0i16(<vscale x 8 x i1> %Pg, i16 *%base_ptr)
+ret <vscale x 32 x i16> %res
+}
+
+define <vscale x 32 x half> @ld4.nxv32f16(<vscale x 8 x i1> %Pg, <vscale x 8 x half> *%addr) {
+; CHECK-LABEL: ld4.nxv32f16:
+; CHECK: ld4h { z0.h, z1.h, z2.h, z3.h }, p0/z, [x0, #8, mul vl]
+; CHECK-NEXT: ret
+%base = getelementptr <vscale x 8 x half>, <vscale x 8 x half>* %addr, i64 8
+%base_ptr = bitcast <vscale x 8 x half>* %base to half *
+%res = call <vscale x 32 x half> @llvm.aarch64.sve.ld4.nxv32f16.nxv8i1.p0f16(<vscale x 8 x i1> %Pg, half *%base_ptr)
+ret <vscale x 32 x half> %res
+}
+
+define <vscale x 32 x bfloat> @ld4.nxv32bf16(<vscale x 8 x i1> %Pg, <vscale x 8 x bfloat> *%addr) #0 {
+; CHECK-LABEL: ld4.nxv32bf16:
+; CHECK: ld4h { z0.h, z1.h, z2.h, z3.h }, p0/z, [x0, #8, mul vl]
+; CHECK-NEXT: ret
+%base = getelementptr <vscale x 8 x bfloat>, <vscale x 8 x bfloat>* %addr, i64 8
+%base_ptr = bitcast <vscale x 8 x bfloat>* %base to bfloat *
+%res = call <vscale x 32 x bfloat> @llvm.aarch64.sve.ld4.nxv32bf16.nxv8i1.p0bf16(<vscale x 8 x i1> %Pg, bfloat *%base_ptr)
+ret <vscale x 32 x bfloat> %res
+}
+
+; ld4w
+define <vscale x 16 x i32> @ld4.nxv16i32(<vscale x 4 x i1> %Pg, <vscale x 4 x i32> *%addr) {
+; CHECK-LABEL: ld4.nxv16i32:
+; CHECK: ld4w { z0.s, z1.s, z2.s, z3.s }, p0/z, [x0, #12, mul vl]
+; CHECK-NEXT: ret
+%base = getelementptr <vscale x 4 x i32>, <vscale x 4 x i32>* %addr, i64 12
+%base_ptr = bitcast <vscale x 4 x i32>* %base to i32 *
+%res = call <vscale x 16 x i32> @llvm.aarch64.sve.ld4.nxv16i32.nxv4i1.p0i32(<vscale x 4 x i1> %Pg, i32 *%base_ptr)
+ret <vscale x 16 x i32> %res
+}
+
+define <vscale x 16 x float> @ld4.nxv16f32(<vscale x 4 x i1> %Pg, <vscale x 4 x float>* %addr) {
+; CHECK-LABEL: ld4.nxv16f32:
+; CHECK: ld4w { z0.s, z1.s, z2.s, z3.s }, p0/z, [x0, #-12, mul vl]
+; CHECK-NEXT: ret
+%base = getelementptr <vscale x 4 x float>, <vscale x 4 x float>* %addr, i64 -12
+%base_ptr = bitcast <vscale x 4 x float>* %base to float *
+%res = call <vscale x 16 x float> @llvm.aarch64.sve.ld4.nxv16f32.nxv4i1.p0f32(<vscale x 4 x i1> %Pg, float *%base_ptr)
+ret <vscale x 16 x float> %res
+}
+
+; ld4d
+define <vscale x 8 x i64> @ld4.nxv8i64(<vscale x 2 x i1> %Pg, <vscale x 2 x i64> *%addr) {
+; CHECK-LABEL: ld4.nxv8i64:
+; CHECK: ld4d { z0.d, z1.d, z2.d, z3.d }, p0/z, [x0, #16, mul vl]
+; CHECK-NEXT: ret
+%base = getelementptr <vscale x 2 x i64>, <vscale x 2 x i64>* %addr, i64 16
+%base_ptr = bitcast <vscale x 2 x i64>* %base to i64 *
+%res = call <vscale x 8 x i64> @llvm.aarch64.sve.ld4.nxv8i64.nxv2i1.p0i64(<vscale x 2 x i1> %Pg, i64 *%base_ptr)
+ret <vscale x 8 x i64> %res
+}
+
+define <vscale x 8 x double> @ld4.nxv8f64(<vscale x 2 x i1> %Pg, <vscale x 2 x double> *%addr) {
+; CHECK-LABEL: ld4.nxv8f64:
+; CHECK: ld4d { z0.d, z1.d, z2.d, z3.d }, p0/z, [x0, #-16, mul vl]
+; CHECK-NEXT: ret
+%base = getelementptr <vscale x 2 x double>, <vscale x 2 x double>* %addr, i64 -16
+%base_ptr = bitcast <vscale x 2 x double>* %base to double *
+%res = call <vscale x 8 x double> @llvm.aarch64.sve.ld4.nxv8f64.nxv2i1.p0f64(<vscale x 2 x i1> %Pg, double * %base_ptr)
+ret <vscale x 8 x double> %res
+}
+
+declare <vscale x 32 x i8> @llvm.aarch64.sve.ld2.nxv32i8.nxv16i1.p0i8(<vscale x 16 x i1>, i8*)
+declare <vscale x 16 x i16> @llvm.aarch64.sve.ld2.nxv16i16.nxv8i1.p0i16(<vscale x 8 x i1>, i16*)
+declare <vscale x 8 x i32> @llvm.aarch64.sve.ld2.nxv8i32.nxv4i1.p0i32(<vscale x 4 x i1>, i32*)
+declare <vscale x 4 x i64> @llvm.aarch64.sve.ld2.nxv4i64.nxv2i1.p0i64(<vscale x 2 x i1>, i64*)
+declare <vscale x 16 x half> @llvm.aarch64.sve.ld2.nxv16f16.nxv8i1.p0f16(<vscale x 8 x i1>, half*)
+declare <vscale x 16 x bfloat> @llvm.aarch64.sve.ld2.nxv16bf16.nxv8i1.p0bf16(<vscale x 8 x i1>, bfloat*)
+declare <vscale x 8 x float> @llvm.aarch64.sve.ld2.nxv8f32.nxv4i1.p0f32(<vscale x 4 x i1>, float*)
+declare <vscale x 4 x double> @llvm.aarch64.sve.ld2.nxv4f64.nxv2i1.p0f64(<vscale x 2 x i1>, double*)
+
+declare <vscale x 48 x i8> @llvm.aarch64.sve.ld3.nxv48i8.nxv16i1.p0i8(<vscale x 16 x i1>, i8*)
+declare <vscale x 24 x i16> @llvm.aarch64.sve.ld3.nxv24i16.nxv8i1.p0i16(<vscale x 8 x i1>, i16*)
+declare <vscale x 12 x i32> @llvm.aarch64.sve.ld3.nxv12i32.nxv4i1.p0i32(<vscale x 4 x i1>, i32*)
+declare <vscale x 6 x i64> @llvm.aarch64.sve.ld3.nxv6i64.nxv2i1.p0i64(<vscale x 2 x i1>, i64*)
+declare <vscale x 24 x half> @llvm.aarch64.sve.ld3.nxv24f16.nxv8i1.p0f16(<vscale x 8 x i1>, half*)
+declare <vscale x 24 x bfloat> @llvm.aarch64.sve.ld3.nxv24bf16.nxv8i1.p0bf16(<vscale x 8 x i1>, bfloat*)
+declare <vscale x 12 x float> @llvm.aarch64.sve.ld3.nxv12f32.nxv4i1.p0f32(<vscale x 4 x i1>, float*)
+declare <vscale x 6 x double> @llvm.aarch64.sve.ld3.nxv6f64.nxv2i1.p0f64(<vscale x 2 x i1>, double*)
+
+declare <vscale x 64 x i8> @llvm.aarch64.sve.ld4.nxv64i8.nxv16i1.p0i8(<vscale x 16 x i1>, i8*)
+declare <vscale x 32 x i16> @llvm.aarch64.sve.ld4.nxv32i16.nxv8i1.p0i16(<vscale x 8 x i1>, i16*)
+declare <vscale x 16 x i32> @llvm.aarch64.sve.ld4.nxv16i32.nxv4i1.p0i32(<vscale x 4 x i1>, i32*)
+declare <vscale x 8 x i64> @llvm.aarch64.sve.ld4.nxv8i64.nxv2i1.p0i64(<vscale x 2 x i1>, i64*)
+declare <vscale x 32 x half> @llvm.aarch64.sve.ld4.nxv32f16.nxv8i1.p0f16(<vscale x 8 x i1>, half*)
+declare <vscale x 32 x bfloat> @llvm.aarch64.sve.ld4.nxv32bf16.nxv8i1.p0bf16(<vscale x 8 x i1>, bfloat*)
+declare <vscale x 16 x float> @llvm.aarch64.sve.ld4.nxv16f32.nxv4i1.p0f32(<vscale x 4 x i1>, float*)
+declare <vscale x 8 x double> @llvm.aarch64.sve.ld4.nxv8f64.nxv2i1.p0f64(<vscale x 2 x i1>, double*)
+
+; +bf16 is required for the bfloat version.
+attributes #0 = { "target-features"="+sve,+bf16" }

llvm/test/CodeGen/AArch64/sve-intrinsics-ldN-reg+reg-addr-mode.ll

+; RUN: llc -mtriple=aarch64--linux-gnu -mattr=sve < %s | FileCheck %s
+
+; ld2b
+define <vscale x 32 x i8> @ld2.nxv32i8(<vscale x 16 x i1> %Pg, i8 *%addr, i64 %a) {
+; CHECK-LABEL: ld2.nxv32i8:
+; CHECK: ld2b { z0.b, z1.b }, p0/z, [x0, x1]
+; CHECK-NEXT: ret
+%addr2 = getelementptr i8, i8 *  %addr, i64 %a
+%res = call <vscale x 32 x i8> @llvm.aarch64.sve.ld2.nxv32i8.nxv16i1.p0i8(<vscale x 16 x i1> %Pg, i8 *%addr2)
+ret <vscale x 32 x i8> %res
+}
+
+; ld2h
+define <vscale x 16 x i16> @ld2.nxv16i16(<vscale x 8 x i1> %Pg, i16 *%addr, i64 %a) {
+; CHECK-LABEL: ld2.nxv16i16:
+; CHECK: ld2h { z0.h, z1.h }, p0/z, [x0, x1, lsl #1]
+; CHECK-NEXT: ret
+%addr2 = getelementptr i16, i16 *  %addr, i64 %a
+%res = call <vscale x 16 x i16> @llvm.aarch64.sve.ld2.nxv16i16.nxv8i1.p0i16(<vscale x 8 x i1> %Pg, i16 *%addr2)
+ret <vscale x 16 x i16> %res
+}
+
+define <vscale x 16 x half> @ld2.nxv16f16(<vscale x 8 x i1> %Pg, half *%addr, i64 %a) {
+; CHECK-LABEL: ld2.nxv16f16:
+; CHECK: ld2h { z0.h, z1.h }, p0/z, [x0, x1, lsl #1]
+; CHECK-NEXT: ret
+%addr2 = getelementptr half, half *  %addr, i64 %a
+%res = call <vscale x 16 x half> @llvm.aarch64.sve.ld2.nxv16f16.nxv8i1.p0f16(<vscale x 8 x i1> %Pg, half *%addr2)
+ret <vscale x 16 x half> %res
+}
+
+define <vscale x 16 x bfloat> @ld2.nxv16bf16(<vscale x 8 x i1> %Pg, bfloat *%addr, i64 %a) #0 {
+; CHECK-LABEL: ld2.nxv16bf16:
+; CHECK: ld2h { z0.h, z1.h }, p0/z, [x0, x1, lsl #1]
+; CHECK-NEXT: ret
+%addr2 = getelementptr bfloat, bfloat *  %addr, i64 %a
+%res = call <vscale x 16 x bfloat> @llvm.aarch64.sve.ld2.nxv16bf16.nxv8i1.p0bf16(<vscale x 8 x i1> %Pg, bfloat *%addr2)
+ret <vscale x 16 x bfloat> %res
+}
+
+; ld2w
+define <vscale x 8 x i32> @ld2.nxv8i32(<vscale x 4 x i1> %Pg, i32 *%addr, i64 %a) {
+; CHECK-LABEL: ld2.nxv8i32:
+; CHECK: ld2w { z0.s, z1.s }, p0/z, [x0, x1, lsl #2]
+; CHECK-NEXT: ret
+%addr2 = getelementptr i32, i32 *  %addr, i64 %a
+%res = call <vscale x 8 x i32> @llvm.aarch64.sve.ld2.nxv8i32.nxv4i1.p0i32(<vscale x 4 x i1> %Pg, i32 *%addr2)
+ret <vscale x 8 x i32> %res
+}
+
+define <vscale x 8 x float> @ld2.nxv8f32(<vscale x 4 x i1> %Pg, float *%addr, i64 %a) {
+; CHECK-LABEL: ld2.nxv8f32:
+; CHECK: ld2w { z0.s, z1.s }, p0/z, [x0, x1, lsl #2]
+; CHECK-NEXT: ret
+%addr2 = getelementptr float, float *  %addr, i64 %a
+%res = call <vscale x 8 x float> @llvm.aarch64.sve.ld2.nxv8f32.nxv4i1.p0f32(<vscale x 4 x i1> %Pg, float *%addr2)
+ret <vscale x 8 x float> %res
+}
+
+; ld2d
+define <vscale x 4 x i64> @ld2.nxv4i64(<vscale x 2 x i1> %Pg, i64 *%addr, i64 %a) {
+; CHECK-LABEL: ld2.nxv4i64:
+; CHECK: ld2d { z0.d, z1.d }, p0/z, [x0, x1, lsl #3]
+; CHECK-NEXT: ret
+%addr2 = getelementptr i64, i64 *  %addr, i64 %a
+%res = call <vscale x 4 x i64> @llvm.aarch64.sve.ld2.nxv4i64.nxv2i1.p0i64(<vscale x 2 x i1> %Pg, i64 *%addr2)
+ret <vscale x 4 x i64> %res
+}
+
+define <vscale x 4 x double> @ld2.nxv4f64(<vscale x 2 x i1> %Pg, double *%addr, i64 %a) {
+; CHECK-LABEL: ld2.nxv4f64:
+; CHECK: ld2d { z0.d, z1.d }, p0/z, [x0, x1, lsl #3]
+; CHECK-NEXT: ret
+%addr2 = getelementptr double, double *  %addr, i64 %a
+%res = call <vscale x 4 x double> @llvm.aarch64.sve.ld2.nxv4f64.nxv2i1.p0f64(<vscale x 2 x i1> %Pg, double *%addr2)
+ret <vscale x 4 x double> %res
+}
+
+; ld3b
+define <vscale x 48 x i8> @ld3.nxv48i8(<vscale x 16 x i1> %Pg, i8 *%addr, i64 %a) {
+; CHECK-LABEL: ld3.nxv48i8:
+; CHECK: ld3b { z0.b, z1.b, z2.b }, p0/z, [x0, x1]
+; CHECK-NEXT: ret
+%addr2 = getelementptr i8, i8 *  %addr, i64 %a
+%res = call <vscale x 48 x i8> @llvm.aarch64.sve.ld3.nxv48i8.nxv16i1.p0i8(<vscale x 16 x i1> %Pg, i8 *%addr2)
+ret <vscale x 48 x i8> %res
+}
+
+; ld3h
+define <vscale x 24 x i16> @ld3.nxv24i16(<vscale x 8 x i1> %Pg, i16 *%addr, i64 %a) {
+; CHECK-LABEL: ld3.nxv24i16:
+; CHECK: ld3h { z0.h, z1.h, z2.h }, p0/z, [x0, x1, lsl #1]
+; CHECK-NEXT: ret
+%addr2 = getelementptr i16, i16 *  %addr, i64 %a
+%res = call <vscale x 24 x i16> @llvm.aarch64.sve.ld3.nxv24i16.nxv8i1.p0i16(<vscale x 8 x i1> %Pg, i16 *%addr2)
+ret <vscale x 24 x i16> %res
+}
+
+define <vscale x 24 x half> @ld3.nxv24f16(<vscale x 8 x i1> %Pg, half *%addr, i64 %a) {
+; CHECK-LABEL: ld3.nxv24f16:
+; CHECK: ld3h { z0.h, z1.h, z2.h }, p0/z, [x0, x1, lsl #1]
+; CHECK-NEXT: ret
+%addr2 = getelementptr half, half *  %addr, i64 %a
+%res = call <vscale x 24 x half> @llvm.aarch64.sve.ld3.nxv24f16.nxv8i1.p0f16(<vscale x 8 x i1> %Pg, half *%addr2)
+ret <vscale x 24 x half> %res
+}
+
+define <vscale x 24 x bfloat> @ld3.nxv24bf16(<vscale x 8 x i1> %Pg, bfloat *%addr, i64 %a) #0 {
+; CHECK-LABEL: ld3.nxv24bf16:
+; CHECK: ld3h { z0.h, z1.h, z2.h }, p0/z, [x0, x1, lsl #1]
+; CHECK-NEXT: ret
+%addr2 = getelementptr bfloat, bfloat *  %addr, i64 %a
+%res = call <vscale x 24 x bfloat> @llvm.aarch64.sve.ld3.nxv24bf16.nxv8i1.p0bf16(<vscale x 8 x i1> %Pg, bfloat *%addr2)
+ret <vscale x 24 x bfloat> %res
+}
+
+; ld3w
+define <vscale x 12 x i32> @ld3.nxv12i32(<vscale x 4 x i1> %Pg, i32 *%addr, i64 %a) {
+; CHECK-LABEL: ld3.nxv12i32:
+; CHECK: ld3w { z0.s, z1.s, z2.s }, p0/z, [x0, x1, lsl #2]
+; CHECK-NEXT: ret
+%addr2 = getelementptr i32, i32 *  %addr, i64 %a
+%res = call <vscale x 12 x i32> @llvm.aarch64.sve.ld3.nxv12i32.nxv4i1.p0i32(<vscale x 4 x i1> %Pg, i32 *%addr2)
+ret <vscale x 12 x i32> %res
+}
+
+define <vscale x 12 x float> @ld3.nxv12f32(<vscale x 4 x i1> %Pg, float *%addr, i64 %a) {
+; CHECK-LABEL: ld3.nxv12f32:
+; CHECK: ld3w { z0.s, z1.s, z2.s }, p0/z, [x0, x1, lsl #2]
+; CHECK-NEXT: ret
+%addr2 = getelementptr float, float *  %addr, i64 %a
+%res = call <vscale x 12 x float> @llvm.aarch64.sve.ld3.nxv12f32.nxv4i1.p0f32(<vscale x 4 x i1> %Pg, float *%addr2)
+ret <vscale x 12 x float> %res
+}
+
+; ld3d
+define <vscale x 6 x i64> @ld3.nxv6i64(<vscale x 2 x i1> %Pg, i64 *%addr, i64 %a) {
+; CHECK-LABEL: ld3.nxv6i64:
+; CHECK: ld3d { z0.d, z1.d, z2.d }, p0/z, [x0, x1, lsl #3]
+; CHECK-NEXT: ret
+%addr2 = getelementptr i64, i64 *  %addr, i64 %a
+%res = call <vscale x 6 x i64> @llvm.aarch64.sve.ld3.nxv6i64.nxv2i1.p0i64(<vscale x 2 x i1> %Pg, i64 *%addr2)
+ret <vscale x 6 x i64> %res
+}
+
+define <vscale x 6 x double> @ld3.nxv6f64(<vscale x 2 x i1> %Pg, double *%addr, i64 %a) {
+; CHECK-LABEL: ld3.nxv6f64:
+; CHECK: ld3d { z0.d, z1.d, z2.d }, p0/z, [x0, x1, lsl #3]
+; CHECK-NEXT: ret
+%addr2 = getelementptr double, double *  %addr, i64 %a
+%res = call <vscale x 6 x double> @llvm.aarch64.sve.ld3.nxv6f64.nxv2i1.p0f64(<vscale x 2 x i1> %Pg, double *%addr2)
+ret <vscale x 6 x double> %res
+}
+
+; ld4b
+define <vscale x 64 x i8> @ld4.nxv64i8(<vscale x 16 x i1> %Pg, i8 *%addr, i64 %a) {
+; CHECK-LABEL: ld4.nxv64i8:
+; CHECK: ld4b { z0.b, z1.b, z2.b, z3.b }, p0/z, [x0, x1]
+; CHECK-NEXT: ret
+%addr2 = getelementptr i8, i8 *  %addr, i64 %a
+%res = call <vscale x 64 x i8> @llvm.aarch64.sve.ld4.nxv64i8.nxv16i1.p0i8(<vscale x 16 x i1> %Pg, i8 *%addr2)
+ret <vscale x 64 x i8> %res
+}
+
+; ld4h
+define <vscale x 32 x i16> @ld4.nxv32i16(<vscale x 8 x i1> %Pg, i16 *%addr, i64 %a) {
+; CHECK-LABEL: ld4.nxv32i16:
+; CHECK: ld4h { z0.h, z1.h, z2.h, z3.h }, p0/z, [x0, x1, lsl #1]
+; CHECK-NEXT: ret
+%addr2 = getelementptr i16, i16 *  %addr, i64 %a
+%res = call <vscale x 32 x i16> @llvm.aarch64.sve.ld4.nxv32i16.nxv8i1.p0i16(<vscale x 8 x i1> %Pg, i16 *%addr2)
+ret <vscale x 32 x i16> %res
+}
+
+define <vscale x 32 x half> @ld4.nxv32f16(<vscale x 8 x i1> %Pg, half *%addr, i64 %a) {
+; CHECK-LABEL: ld4.nxv32f16:
+; CHECK: ld4h { z0.h, z1.h, z2.h, z3.h }, p0/z, [x0, x1, lsl #1]
+; CHECK-NEXT: ret
+%addr2 = getelementptr half, half *  %addr, i64 %a
+%res = call <vscale x 32 x half> @llvm.aarch64.sve.ld4.nxv32f16.nxv8i1.p0f16(<vscale x 8 x i1> %Pg, half *%addr2)
+ret <vscale x 32 x half> %res
+}
+
+define <vscale x 32 x bfloat> @ld4.nxv32bf16(<vscale x 8 x i1> %Pg, bfloat *%addr, i64 %a) #0 {
+; CHECK-LABEL: ld4.nxv32bf16:
+; CHECK: ld4h { z0.h, z1.h, z2.h, z3.h }, p0/z, [x0, x1, lsl #1]
+; CHECK-NEXT: ret
+%addr2 = getelementptr bfloat, bfloat *  %addr, i64 %a
+%res = call <vscale x 32 x bfloat> @llvm.aarch64.sve.ld4.nxv32bf16.nxv8i1.p0bf16(<vscale x 8 x i1> %Pg, bfloat *%addr2)
+ret <vscale x 32 x bfloat> %res
+}
+
+; ld4w
+define <vscale x 16 x i32> @ld4.nxv16i32(<vscale x 4 x i1> %Pg, i32 *%addr, i64 %a) {
+; CHECK-LABEL: ld4.nxv16i32:
+; CHECK: ld4w { z0.s, z1.s, z2.s, z3.s }, p0/z, [x0, x1, lsl #2]
+; CHECK-NEXT: ret
+%addr2 = getelementptr i32, i32 *  %addr, i64 %a
+%res = call <vscale x 16 x i32> @llvm.aarch64.sve.ld4.nxv16i32.nxv4i1.p0i32(<vscale x 4 x i1> %Pg, i32 *%addr2)
+ret <vscale x 16 x i32> %res
+}
+
+define <vscale x 16 x float> @ld4.nxv16f32(<vscale x 4 x i1> %Pg, float *%addr, i64 %a) {
+; CHECK-LABEL: ld4.nxv16f32:
+; CHECK: ld4w { z0.s, z1.s, z2.s, z3.s }, p0/z, [x0, x1, lsl #2]
+; CHECK-NEXT: ret
+%addr2 = getelementptr float, float *  %addr, i64 %a
+%res = call <vscale x 16 x float> @llvm.aarch64.sve.ld4.nxv16f32.nxv4i1.p0f32(<vscale x 4 x i1> %Pg, float *%addr2)
+ret <vscale x 16 x float> %res
+}
+
+; ld4d
+define <vscale x 8 x i64> @ld4.nxv8i64(<vscale x 2 x i1> %Pg, i64 *%addr, i64 %a) {
+; CHECK-LABEL: ld4.nxv8i64:
+; CHECK: ld4d { z0.d, z1.d, z2.d, z3.d }, p0/z, [x0, x1, lsl #3]
+; CHECK-NEXT: ret
+%addr2 = getelementptr i64, i64 *  %addr, i64 %a
+%res = call <vscale x 8 x i64> @llvm.aarch64.sve.ld4.nxv8i64.nxv2i1.p0i64(<vscale x 2 x i1> %Pg, i64 *%addr2)
+ret <vscale x 8 x i64> %res
+}
+
+define <vscale x 8 x double> @ld4.nxv8f64(<vscale x 2 x i1> %Pg, double *%addr, i64 %a) {
+; CHECK-LABEL: ld4.nxv8f64:
+; CHECK: ld4d { z0.d, z1.d, z2.d, z3.d }, p0/z, [x0, x1, lsl #3]
+; CHECK-NEXT: ret
+%addr2 = getelementptr double, double *  %addr, i64 %a
+%res = call <vscale x 8 x double> @llvm.aarch64.sve.ld4.nxv8f64.nxv2i1.p0f64(<vscale x 2 x i1> %Pg, double *%addr2)
+ret <vscale x 8 x double> %res
+}
+
+declare <vscale x 32 x i8> @llvm.aarch64.sve.ld2.nxv32i8.nxv16i1.p0i8(<vscale x 16 x i1>, i8*)
+declare <vscale x 16 x i16> @llvm.aarch64.sve.ld2.nxv16i16.nxv8i1.p0i16(<vscale x 8 x i1>, i16*)
+declare <vscale x 8 x i32> @llvm.aarch64.sve.ld2.nxv8i32.nxv4i1.p0i32(<vscale x 4 x i1>, i32*)
+declare <vscale x 4 x i64> @llvm.aarch64.sve.ld2.nxv4i64.nxv2i1.p0i64(<vscale x 2 x i1>, i64*)
+declare <vscale x 16 x half> @llvm.aarch64.sve.ld2.nxv16f16.nxv8i1.p0f16(<vscale x 8 x i1>, half*)
+declare <vscale x 16 x bfloat> @llvm.aarch64.sve.ld2.nxv16bf16.nxv8i1.p0bf16(<vscale x 8 x i1>, bfloat*)
+declare <vscale x 8 x float> @llvm.aarch64.sve.ld2.nxv8f32.nxv4i1.p0f32(<vscale x 4 x i1>, float*)
+declare <vscale x 4 x double> @llvm.aarch64.sve.ld2.nxv4f64.nxv2i1.p0f64(<vscale x 2 x i1>, double*)
+
+declare <vscale x 48 x i8> @llvm.aarch64.sve.ld3.nxv48i8.nxv16i1.p0i8(<vscale x 16 x i1>, i8*)
+declare <vscale x 24 x i16> @llvm.aarch64.sve.ld3.nxv24i16.nxv8i1.p0i16(<vscale x 8 x i1>, i16*)
+declare <vscale x 12 x i32> @llvm.aarch64.sve.ld3.nxv12i32.nxv4i1.p0i32(<vscale x 4 x i1>, i32*)
+declare <vscale x 6 x i64> @llvm.aarch64.sve.ld3.nxv6i64.nxv2i1.p0i64(<vscale x 2 x i1>, i64*)
+declare <vscale x 24 x half> @llvm.aarch64.sve.ld3.nxv24f16.nxv8i1.p0f16(<vscale x 8 x i1>, half*)
+declare <vscale x 24 x bfloat> @llvm.aarch64.sve.ld3.nxv24bf16.nxv8i1.p0bf16(<vscale x 8 x i1>, bfloat*)
+declare <vscale x 12 x float> @llvm.aarch64.sve.ld3.nxv12f32.nxv4i1.p0f32(<vscale x 4 x i1>, float*)
+declare <vscale x 6 x double> @llvm.aarch64.sve.ld3.nxv6f64.nxv2i1.p0f64(<vscale x 2 x i1>, double*)
+
+declare <vscale x 64 x i8> @llvm.aarch64.sve.ld4.nxv64i8.nxv16i1.p0i8(<vscale x 16 x i1>, i8*)
+declare <vscale x 32 x i16> @llvm.aarch64.sve.ld4.nxv32i16.nxv8i1.p0i16(<vscale x 8 x i1>, i16*)
+declare <vscale x 16 x i32> @llvm.aarch64.sve.ld4.nxv16i32.nxv4i1.p0i32(<vscale x 4 x i1>, i32*)
+declare <vscale x 8 x i64> @llvm.aarch64.sve.ld4.nxv8i64.nxv2i1.p0i64(<vscale x 2 x i1>, i64*)
+declare <vscale x 32 x half> @llvm.aarch64.sve.ld4.nxv32f16.nxv8i1.p0f16(<vscale x 8 x i1>, half*)
+declare <vscale x 32 x bfloat> @llvm.aarch64.sve.ld4.nxv32bf16.nxv8i1.p0bf16(<vscale x 8 x i1>, bfloat*)
+declare <vscale x 16 x float> @llvm.aarch64.sve.ld4.nxv16f32.nxv4i1.p0f32(<vscale x 4 x i1>, float*)
+declare <vscale x 8 x double> @llvm.aarch64.sve.ld4.nxv8f64.nxv2i1.p0f64(<vscale x 2 x i1>, double*)
+
+; +bf16 is required for the bfloat version.
+attributes #0 = { "target-features"="+sve,+bf16" }