[RISCV] Support insertion of misaligned subvectors

[frasercrmck]

Summary

This patch extends the support for RVV INSERT_SUBVECTOR to cover those
which don't align to a vector register boundary. Like the support for
EXTRACT_SUBVECTOR in D96959, it accomplishes this by extracting the
nearest register-sized subvector (a subregister operation), then sliding
the vector down with VSLIDEDOWN, inserting the subvector to the first
position, and sliding the vector back up again afterwards.

Unlike subvector extraction, for vectors that occupy less than a full
vector register we must preserve the untouched elements. We do this by
lowering to an LMUL=1 INSERT_SUBVECTOR using the above method and
lowering that to a VSLIDEUP with a zero offset. This uses a
tail-undisturbed policy and so has the effect of "sliding in" the
subvector elements while preserving the surrounding ones.

[craig.topper]

llvm/lib/Target/RISCV/RISCVISelDAGToDAG.cpp

966

Should this really be a tail undisturbed vmv.v.v?

[craig.topper]

Can a less than full vector insert be done with a single slideup if you calculate the offset and limit vl to mark the end of the range?

[frasercrmck]

In D96972#2573979, @craig.topper wrote:

Can a less than full vector insert be done with a single slideup if you calculate the offset and limit vl to mark the end of the range?

Probably. That occurred to me slightly too late yesterday and I was going to take a look at that this morning. It would also handle i1 vectors in the same way as others as you're forced to limit VL to account for (some of) those.

llvm/lib/Target/RISCV/RISCVISelDAGToDAG.cpp

966

Yeah, that's what I was searching for originally. I was hesitant to introduce a new pseudo for that purpose in this patch, though. It might be mostly irrelevant if I optimize this to use vslideup everywhere. Though the vmv.v.v would still provide a potential optimization for correctly-sized types at offset 0. Do you think that's worth it?

[frasercrmck]

• rebase
• add float vectors
• lower simply VSLIDEUP of LMUL=1 type (no VSLIDEDOWNs)

[frasercrmck]

• silence unused variable warning
• test inserting into undef subvectors

[frasercrmck]

Sadly this regresses some of the fixed-length tests. Because INSERT_SUBVECTOR is now "custom" (not "legal") it doesn't fold bitcasts around the operation:

Vector/type-legalized selection DAG: %bb.0 'mulhs_vx_v2i64:'
                t27: v4i32 = BUILD_VECTOR Constant:i32<63>, Constant:i32<0>, Constant:i32<63>, Constant:i32<0>
              t28: v2i64 = bitcast t27
            t40: nxv1i64 = insert_subvector undef:nxv1i64, t28, Constant:i32<0>

it used to perform this dag-combine:

Optimized vector-legalized selection DAG: %bb.0 'mulhs_vx_v2i64:'
                t27: v4i32 = BUILD_VECTOR Constant:i32<63>, Constant:i32<0>, Constant:i32<63>, Constant:i32<0>
              t46: nxv2i32 = insert_subvector undef:nxv2i32, t27, Constant:i32<0>
            t47: nxv1i64 = bitcast t46

which for scalable vectors is a no-op, but for fixed-length vectors it still goes through the stack.

Sounds like there's a sort-of dependency on supporting fixed-length bitcasts before we can merge this. @craig.topper, I don't suppose you're looking into it?

[frasercrmck]

• rebase

[craig.topper]

LGTM other than that one question

llvm/lib/Target/RISCV/RISCVISelLowering.cpp

2254

Should this be unsigned instead of auto? I think getConstantOperandVal returns uint64_t but the rest of this code uses unsigned. I probably should have asked this question in the EXTRACT_SUBVECTOR patch.

[frasercrmck]

llvm/lib/Target/RISCV/RISCVISelLowering.cpp

2254

Yes, sounds good. Thanks for spotting that. The EXTRACT_SUBVECTOR code is using unsigned.

[frasercrmck]

• rebase
• address review feedback

llvm/lib/Target/RISCV/RISCVISelDAGToDAG.cpp

case ISD::BITCAST: {
     }
     break;
   }
   case ISD::INSERT_SUBVECTOR: {
     SDValue V = Node->getOperand(0);
     SDValue SubV = Node->getOperand(1);
     SDLoc DL(SubV);
     auto Idx = Node->getConstantOperandVal(2);
+    MVT XLenVT = Subtarget->getXLenVT();
     MVT SubVecVT = Node->getOperand(1).getSimpleValueType();
 
     // TODO: This method of selecting INSERT_SUBVECTOR should work
     // with any type of insertion (fixed <-> scalable) but we don't yet
     // correctly identify the canonical register class for fixed-length types.
     // For now, keep the two paths separate.
     if (VT.isScalableVector() && SubVecVT.isScalableVector()) {
-      bool IsFullVecReg = false;
-      switch (RISCVTargetLowering::getLMUL(SubVecVT)) {
-      default:
-        break;
-      case RISCVVLMUL::LMUL_1:
-      case RISCVVLMUL::LMUL_2:
-      case RISCVVLMUL::LMUL_4:
-      case RISCVVLMUL::LMUL_8:
-        IsFullVecReg = true;
-        break;
-      }
-
-      // If the subvector doesn't occupy a full vector register then we can't
-      // insert it purely using subregister manipulation. We must not clobber
-      // the untouched elements (say, in the upper half of the VR register).
-      if (!IsFullVecReg)
-        break;
-
       const auto *TRI = Subtarget->getRegisterInfo();
       unsigned SubRegIdx;
       std::tie(SubRegIdx, Idx) =
           RISCVTargetLowering::decomposeSubvectorInsertExtractToSubRegs(
               VT, SubVecVT, Idx, TRI);
 
       // If the Idx hasn't been completely eliminated then this is a subvector
       // insert which doesn't naturally align to a vector register. These must
       // be handled using instructions to manipulate the vector registers.
       if (Idx != 0)
         break;
 
+      RISCVVLMUL SubVecLMUL = RISCVTargetLowering::getLMUL(SubVecVT);
+      bool IsSubVecPartReg = SubVecLMUL == RISCVVLMUL::LMUL_F2 ||
+                             SubVecLMUL == RISCVVLMUL::LMUL_F4 ||
+                             SubVecLMUL == RISCVVLMUL::LMUL_F8;
+      (void)IsSubVecPartReg; // Silence unused variable warning without asserts.
+      assert((!IsSubVecPartReg || V.isUndef()) &&
+             "Expecting lowering to have created legal INSERT_SUBVECTORs when "
+             "the subvector is smaller than a full-sized register");
+
+      // If we haven't set a SubRegIdx, then we must be going between LMUL<=1
+      // types (VR -> VR). This can be done as a copy.
+      if (SubRegIdx == RISCV::NoSubRegister) {
+        unsigned InRegClassID = RISCVTargetLowering::getRegClassIDForVecVT(VT);
+        assert(RISCVTargetLowering::getRegClassIDForVecVT(SubVecVT) ==
+                   RISCV::VRRegClassID &&
+               InRegClassID == RISCV::VRRegClassID &&
+               "Unexpected subvector extraction");
+        SDValue RC = CurDAG->getTargetConstant(InRegClassID, DL, XLenVT);
+        SDNode *NewNode = CurDAG->getMachineNode(TargetOpcode::COPY_TO_REGCLASS,
+                                                 DL, VT, SubV, RC);
+        return ReplaceNode(Node, NewNode);
+      }
+
       SDNode *NewNode = CurDAG->getMachineNode(
           TargetOpcode::INSERT_SUBREG, DL, VT, V, SubV,
-          CurDAG->getTargetConstant(SubRegIdx, DL, Subtarget->getXLenVT()));
+          CurDAG->getTargetConstant(SubRegIdx, DL, XLenVT));
       return ReplaceNode(Node, NewNode);
     }
 
     if (VT.isScalableVector() && SubVecVT.isFixedLengthVector()) {
       // Bail when not a "cast" like insert_subvector.
       if (Idx != 0)
         break;
       if (!Node->getOperand(0).isUndef())
         break;
 
       unsigned RegClassID = RISCVTargetLowering::getRegClassIDForVecVT(VT);
 
-      SDValue RC =
-          CurDAG->getTargetConstant(RegClassID, DL, Subtarget->getXLenVT());
+      SDValue RC = CurDAG->getTargetConstant(RegClassID, DL, XLenVT);
       SDNode *NewNode = CurDAG->getMachineNode(TargetOpcode::COPY_TO_REGCLASS,
                                                DL, VT, SubV, RC);
       ReplaceNode(Node, NewNode);
       return;
     }
     break;
   }
   case ISD::EXTRACT_SUBVECTOR: {

llvm/lib/Target/RISCV/RISCVISelLowering.h

SDValue lowerVectorMaskExt(SDValue Op, SelectionDAG &DAG,
                              int64_t ExtTrueVal) const;
   SDValue lowerVectorMaskTrunc(SDValue Op, SelectionDAG &DAG) const;
   SDValue lowerINSERT_VECTOR_ELT(SDValue Op, SelectionDAG &DAG) const;
   SDValue lowerEXTRACT_VECTOR_ELT(SDValue Op, SelectionDAG &DAG) const;
   SDValue LowerINTRINSIC_WO_CHAIN(SDValue Op, SelectionDAG &DAG) const;
   SDValue LowerINTRINSIC_W_CHAIN(SDValue Op, SelectionDAG &DAG) const;
   SDValue lowerVECREDUCE(SDValue Op, SelectionDAG &DAG) const;
   SDValue lowerFPVECREDUCE(SDValue Op, SelectionDAG &DAG) const;
+  SDValue lowerINSERT_SUBVECTOR(SDValue Op, SelectionDAG &DAG) const;
   SDValue lowerEXTRACT_SUBVECTOR(SDValue Op, SelectionDAG &DAG) const;
   SDValue lowerFixedLengthVectorLoadToRVV(SDValue Op, SelectionDAG &DAG) const;
   SDValue lowerFixedLengthVectorStoreToRVV(SDValue Op, SelectionDAG &DAG) const;
   SDValue lowerFixedLengthVectorSetccToRVV(SDValue Op, SelectionDAG &DAG) const;
   SDValue lowerFixedLengthVectorLogicOpToRVV(SDValue Op, SelectionDAG &DAG,
                                              unsigned MaskOpc,
                                              unsigned VecOpc) const;
   SDValue lowerFixedLengthVectorSelectToRVV(SDValue Op,

llvm/lib/Target/RISCV/RISCVISelLowering.cpp

for (MVT VT : IntVecVTs) {
       setOperationAction(ISD::VECREDUCE_AND, VT, Custom);
       setOperationAction(ISD::VECREDUCE_OR, VT, Custom);
       setOperationAction(ISD::VECREDUCE_XOR, VT, Custom);
       setOperationAction(ISD::VECREDUCE_SMAX, VT, Custom);
       setOperationAction(ISD::VECREDUCE_SMIN, VT, Custom);
       setOperationAction(ISD::VECREDUCE_UMAX, VT, Custom);
       setOperationAction(ISD::VECREDUCE_UMIN, VT, Custom);
 
+      setOperationAction(ISD::INSERT_SUBVECTOR, VT, Custom);
       setOperationAction(ISD::EXTRACT_SUBVECTOR, VT, Custom);
     }
 
     // Expand various CCs to best match the RVV ISA, which natively supports UNE
     // but no other unordered comparisons, and supports all ordered comparisons
     // except ONE. Additionally, we expand GT,OGT,GE,OGE for optimization
     // purposes; they are expanded to their swapped-operand CCs (LT,OLT,LE,OLE),
     // and we pattern-match those back to the "original", swapping operands once
const auto SetCommonVFPActions = [&](MVT VT) {
       // Expand various condition codes (explained above).
       for (auto CC : VFPCCToExpand)
         setCondCodeAction(CC, VT, Expand);
 
       setOperationAction(ISD::VECREDUCE_FADD, VT, Custom);
       setOperationAction(ISD::VECREDUCE_SEQ_FADD, VT, Custom);
       setOperationAction(ISD::FCOPYSIGN, VT, Legal);
 
+      setOperationAction(ISD::INSERT_SUBVECTOR, VT, Custom);
       setOperationAction(ISD::EXTRACT_SUBVECTOR, VT, Custom);
     };
 
     if (Subtarget.hasStdExtZfh())
       for (MVT VT : F16VecVTs)
         SetCommonVFPActions(VT);
 
     if (Subtarget.hasStdExtF())
SDValue RISCVTargetLowering::LowerOperation(SDValue Op,
   case ISD::VECREDUCE_SMIN:
   case ISD::VECREDUCE_AND:
   case ISD::VECREDUCE_OR:
   case ISD::VECREDUCE_XOR:
     return lowerVECREDUCE(Op, DAG);
   case ISD::VECREDUCE_FADD:
   case ISD::VECREDUCE_SEQ_FADD:
     return lowerFPVECREDUCE(Op, DAG);
+  case ISD::INSERT_SUBVECTOR:
+    return lowerINSERT_SUBVECTOR(Op, DAG);
   case ISD::EXTRACT_SUBVECTOR:
     return lowerEXTRACT_SUBVECTOR(Op, DAG);
   case ISD::BUILD_VECTOR:
     return lowerBUILD_VECTOR(Op, DAG, Subtarget);
   case ISD::VECTOR_SHUFFLE:
     return lowerVECTOR_SHUFFLE(Op, DAG, Subtarget);
   case ISD::LOAD:
     return lowerFixedLengthVectorLoadToRVV(Op, DAG);

 static MVT getLMUL1VT(MVT VT) {
   assert(VT.getVectorElementType().getSizeInBits() <= 64 &&
          "Unexpected vector MVT");
   return MVT::getScalableVectorVT(
       VT.getVectorElementType(),
       RISCV::RVVBitsPerBlock / VT.getVectorElementType().getSizeInBits());
 }
 
+SDValue RISCVTargetLowering::lowerINSERT_SUBVECTOR(SDValue Op,
+                                                   SelectionDAG &DAG) const {
+  SDValue Vec = Op.getOperand(0);
+  SDValue SubVec = Op.getOperand(1);
+  MVT VecVT = Vec.getSimpleValueType();
+  MVT SubVecVT = SubVec.getSimpleValueType();
+
+  // TODO: Only handle scalable->scalable inserts for now, and revisit this for
+  // fixed-length vectors later.
+  if (!SubVecVT.isScalableVector() || !VecVT.isScalableVector())
+    return Op;
+
+  SDLoc DL(Op);
+  unsigned OrigIdx = Op.getConstantOperandVal(2);
+  const RISCVRegisterInfo *TRI = Subtarget.getRegisterInfo();
+
+  unsigned SubRegIdx, RemIdx;
+  std::tie(SubRegIdx, RemIdx) =
+      RISCVTargetLowering::decomposeSubvectorInsertExtractToSubRegs(
+          VecVT, SubVecVT, OrigIdx, TRI);
+
+  RISCVVLMUL SubVecLMUL = RISCVTargetLowering::getLMUL(SubVecVT);
+  bool IsSubVecPartReg = SubVecLMUL == RISCVVLMUL::LMUL_F2 ||
+                         SubVecLMUL == RISCVVLMUL::LMUL_F4 ||
+                         SubVecLMUL == RISCVVLMUL::LMUL_F8;
+
+  // If the Idx has been completely eliminated and this subvector's size is a
+  // vector register or a multiple thereof, or the surrounding elements are
+  // undef, then this is a subvector insert which naturally aligns to a vector
+  // register. These can easily be handled using subregister manipulation.
+  // If the subvector is smaller than a vector register, then the insertion
+  // must preserve the undisturbed elements of the register. We do this by
+  // lowering to an EXTRACT_SUBVECTOR grabbing the nearest LMUL=1 vector type
+  // (which resolves to a subregister copy), performing a VSLIDEUP to place the
+  // subvector within the vector register, and an INSERT_SUBVECTOR of that
+  // LMUL=1 type back into the larger vector (resolving to another subregister
+  // operation). See below for how our VSLIDEUP works. We go via a LMUL=1 type
+  // to avoid allocating a large register group to hold our subvector.
+  if (RemIdx == 0 && (!IsSubVecPartReg || Vec.isUndef()))
+    return Op;
+
+  // VSLIDEUP works by leaving elements 0<i<OFFSET undisturbed, elements
+  // OFFSET<=i<VL set to the "subvector" and vl<=i<VLMAX set to the tail policy
+  // (in our case undisturbed). This means we can set up a subvector insertion
+  // where OFFSET is the insertion offset, and the VL is the OFFSET plus the
+  // size of the subvector.
+  MVT XLenVT = Subtarget.getXLenVT();
+  MVT InterSubVT = getLMUL1VT(VecVT);
+
+  // Extract a subvector equal to the nearest full vector register type. This
+  // should resolve to a EXTRACT_SUBREG instruction.
+  unsigned AlignedIdx = OrigIdx - RemIdx;
+  SDValue AlignedExtract =
+      DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, InterSubVT, Vec,
+                  DAG.getConstant(AlignedIdx, DL, XLenVT));
+
+  SDValue SlideupAmt = DAG.getConstant(RemIdx, DL, XLenVT);
+  // For scalable vectors this must be further multiplied by vscale.
+  SlideupAmt = DAG.getNode(ISD::VSCALE, DL, XLenVT, SlideupAmt);
+
+  SDValue Mask, VL;
+  std::tie(Mask, VL) = getDefaultScalableVLOps(VecVT, DL, DAG, Subtarget);
+
+  // Construct the vector length corresponding to RemIdx + length(SubVecVT).
+  VL = DAG.getConstant(SubVecVT.getVectorMinNumElements(), DL, XLenVT);
+  VL = DAG.getNode(ISD::VSCALE, DL, XLenVT, VL);
+  VL = DAG.getNode(ISD::ADD, DL, XLenVT, SlideupAmt, VL);
+
+  SubVec = DAG.getNode(ISD::INSERT_SUBVECTOR, DL, InterSubVT,
+                       DAG.getUNDEF(InterSubVT), SubVec,
+                       DAG.getConstant(0, DL, XLenVT));
+
+  SDValue Slideup = DAG.getNode(RISCVISD::VSLIDEUP_VL, DL, InterSubVT,
+                                AlignedExtract, SubVec, SlideupAmt, Mask, VL);
+
+  // Insert this subvector into the correct vector register. This should
+  // resolve to an INSERT_SUBREG instruction.
+  return DAG.getNode(ISD::INSERT_SUBVECTOR, DL, VecVT, Vec, Slideup,
+                     DAG.getConstant(AlignedIdx, DL, XLenVT));
+}
+
 SDValue RISCVTargetLowering::lowerEXTRACT_SUBVECTOR(SDValue Op,
                                                     SelectionDAG &DAG) const {
   SDValue Vec = Op.getOperand(0);
   MVT SubVecVT = Op.getSimpleValueType();
   MVT VecVT = Vec.getSimpleValueType();
 
   // TODO: Only handle scalable->scalable extracts for now, and revisit this
   // for fixed-length vectors later.

llvm/test/CodeGen/RISCV/rvv/insert-subvector.ll

 ; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py
-; RUN: llc -mtriple riscv64 -mattr=+experimental-v -verify-machineinstrs < %s | FileCheck %s
+; RUN: llc -mtriple riscv64 -mattr=+m,+d,+experimental-zfh,+experimental-v -verify-machineinstrs < %s | FileCheck %s
 
 define <vscale x 8 x i32> @insert_nxv8i32_nxv4i32_0(<vscale x 8 x i32> %vec, <vscale x 4 x i32> %subvec) {
 ; CHECK-LABEL: insert_nxv8i32_nxv4i32_0:
 ; CHECK:       # %bb.0:
 ; CHECK-NEXT:    vmv2r.v v8, v12
 ; CHECK-NEXT:    ret
   %v = call <vscale x 8 x i32> @llvm.experimental.vector.insert.nxv4i32.nxv8i32(<vscale x 8 x i32> %vec, <vscale x 4 x i32> %subvec, i64 0)
   ret <vscale x 8 x i32> %v

 ; CHECK-LABEL: insert_nxv16i32_nxv2i32_14:
 ; CHECK:       # %bb.0:
 ; CHECK-NEXT:    vmv1r.v v15, v16
 ; CHECK-NEXT:    ret
   %v = call <vscale x 16 x i32> @llvm.experimental.vector.insert.nxv2i32.nxv16i32(<vscale x 16 x i32> %vec, <vscale x 2 x i32> %subvec, i64 14)
   ret <vscale x 16 x i32> %v
 }
 
-; TODO: Inserts that are less than LMUL=1 are not yet supported. In this case
-; we need mask out the unaffected elements (top half of the VR %subvec
-; register)
-;define <vscale x 16 x i32> @insert_nxv16i32_nxv1i32_0(<vscale x 16 x i32> %vec, <vscale x 1 x i32> %subvec) {
-;  %v = call <vscale x 16 x i32> @llvm.experimental.vector.insert.nxv1i32.nxv16i32(<vscale x 16 x i32> %vec, <vscale x 1 x i32> %subvec, i64 0)
-;  ret <vscale x 16 x i32> %v
-;}
-
-; TODO: Inserts that don't align to a vector register are not yet supported.
-; In this case we want to insert the subvector into the upper half of the
-; lowest VR subregister in the LMUL group.
-;define <vscale x 16 x i32> @insert_nxv16i32_nxv1i32_1(<vscale x 16 x i32> %vec, <vscale x 1 x i32> %subvec) {
-;  %v = call <vscale x 16 x i32> @llvm.experimental.vector.insert.nxv1i32.nxv16i32(<vscale x 16 x i32> %vec, <vscale x 1 x i32> %subvec, i64 1)
-;  ret <vscale x 16 x i32> %v
-;}
+define <vscale x 16 x i32> @insert_nxv16i32_nxv1i32_0(<vscale x 16 x i32> %vec, <vscale x 1 x i32> %subvec) {
+; CHECK-LABEL: insert_nxv16i32_nxv1i32_0:
+; CHECK:       # %bb.0:
+; CHECK-NEXT:    csrr a0, vlenb
+; CHECK-NEXT:    srli a0, a0, 3
+; CHECK-NEXT:    vsetvli a0, a0, e32,m1,tu,mu
+; CHECK-NEXT:    vslideup.vi v8, v16, 0
+; CHECK-NEXT:    ret
+  %v = call <vscale x 16 x i32> @llvm.experimental.vector.insert.nxv1i32.nxv16i32(<vscale x 16 x i32> %vec, <vscale x 1 x i32> %subvec, i64 0)
+  ret <vscale x 16 x i32> %v
+}
+
+define <vscale x 16 x i32> @insert_nxv16i32_nxv1i32_1(<vscale x 16 x i32> %vec, <vscale x 1 x i32> %subvec) {
+; CHECK-LABEL: insert_nxv16i32_nxv1i32_1:
+; CHECK:       # %bb.0:
+; CHECK-NEXT:    csrr a0, vlenb
+; CHECK-NEXT:    srli a0, a0, 3
+; CHECK-NEXT:    add a1, a0, a0
+; CHECK-NEXT:    vsetvli a1, a1, e32,m1,tu,mu
+; CHECK-NEXT:    vslideup.vx v8, v16, a0
+; CHECK-NEXT:    ret
+  %v = call <vscale x 16 x i32> @llvm.experimental.vector.insert.nxv1i32.nxv16i32(<vscale x 16 x i32> %vec, <vscale x 1 x i32> %subvec, i64 1)
+  ret <vscale x 16 x i32> %v
+}
+
+define <vscale x 16 x i32> @insert_nxv16i32_nxv1i32_6(<vscale x 16 x i32> %vec, <vscale x 1 x i32> %subvec) {
+; CHECK-LABEL: insert_nxv16i32_nxv1i32_6:
+; CHECK:       # %bb.0:
+; CHECK-NEXT:    csrr a0, vlenb
+; CHECK-NEXT:    srli a0, a0, 3
+; CHECK-NEXT:    vsetvli a0, a0, e32,m1,tu,mu
+; CHECK-NEXT:    vslideup.vi v11, v16, 0
+; CHECK-NEXT:    ret
+  %v = call <vscale x 16 x i32> @llvm.experimental.vector.insert.nxv1i32.nxv16i32(<vscale x 16 x i32> %vec, <vscale x 1 x i32> %subvec, i64 6)
+  ret <vscale x 16 x i32> %v
+}
+
+define <vscale x 16 x i8> @insert_nxv16i8_nxv1i8_0(<vscale x 16 x i8> %vec, <vscale x 1 x i8> %subvec) {
+; CHECK-LABEL: insert_nxv16i8_nxv1i8_0:
+; CHECK:       # %bb.0:
+; CHECK-NEXT:    csrr a0, vlenb
+; CHECK-NEXT:    srli a0, a0, 3
+; CHECK-NEXT:    vsetvli a0, a0, e8,m1,tu,mu
+; CHECK-NEXT:    vslideup.vi v8, v10, 0
+; CHECK-NEXT:    ret
+  %v = call <vscale x 16 x i8> @llvm.experimental.vector.insert.nxv1i8.nxv16i8(<vscale x 16 x i8> %vec, <vscale x 1 x i8> %subvec, i64 0)
+  ret <vscale x 16 x i8> %v
+}
+
+define <vscale x 16 x i8> @insert_nxv16i8_nxv1i8_1(<vscale x 16 x i8> %vec, <vscale x 1 x i8> %subvec) {
+; CHECK-LABEL: insert_nxv16i8_nxv1i8_1:
+; CHECK:       # %bb.0:
+; CHECK-NEXT:    csrr a0, vlenb
+; CHECK-NEXT:    srli a0, a0, 3
+; CHECK-NEXT:    add a1, a0, a0
+; CHECK-NEXT:    vsetvli a1, a1, e8,m1,tu,mu
+; CHECK-NEXT:    vslideup.vx v8, v10, a0
+; CHECK-NEXT:    ret
+  %v = call <vscale x 16 x i8> @llvm.experimental.vector.insert.nxv1i8.nxv16i8(<vscale x 16 x i8> %vec, <vscale x 1 x i8> %subvec, i64 1)
+  ret <vscale x 16 x i8> %v
+}
+
+define <vscale x 16 x i8> @insert_nxv16i8_nxv1i8_2(<vscale x 16 x i8> %vec, <vscale x 1 x i8> %subvec) {
+; CHECK-LABEL: insert_nxv16i8_nxv1i8_2:
+; CHECK:       # %bb.0:
+; CHECK-NEXT:    csrr a0, vlenb
+; CHECK-NEXT:    srli a0, a0, 3
+; CHECK-NEXT:    slli a1, a0, 1
+; CHECK-NEXT:    add a0, a1, a0
+; CHECK-NEXT:    vsetvli a0, a0, e8,m1,tu,mu
+; CHECK-NEXT:    vslideup.vx v8, v10, a1
+; CHECK-NEXT:    ret
+  %v = call <vscale x 16 x i8> @llvm.experimental.vector.insert.nxv1i8.nxv16i8(<vscale x 16 x i8> %vec, <vscale x 1 x i8> %subvec, i64 2)
+  ret <vscale x 16 x i8> %v
+}
+
+define <vscale x 16 x i8> @insert_nxv16i8_nxv1i8_3(<vscale x 16 x i8> %vec, <vscale x 1 x i8> %subvec) {
+; CHECK-LABEL: insert_nxv16i8_nxv1i8_3:
+; CHECK:       # %bb.0:
+; CHECK-NEXT:    csrr a0, vlenb
+; CHECK-NEXT:    srli a0, a0, 3
+; CHECK-NEXT:    slli a1, a0, 1
+; CHECK-NEXT:    add a1, a1, a0
+; CHECK-NEXT:    add a0, a1, a0
+; CHECK-NEXT:    vsetvli a0, a0, e8,m1,tu,mu
+; CHECK-NEXT:    vslideup.vx v8, v10, a1
+; CHECK-NEXT:    ret
+  %v = call <vscale x 16 x i8> @llvm.experimental.vector.insert.nxv1i8.nxv16i8(<vscale x 16 x i8> %vec, <vscale x 1 x i8> %subvec, i64 3)
+  ret <vscale x 16 x i8> %v
+}
+
+define <vscale x 16 x i8> @insert_nxv16i8_nxv1i8_7(<vscale x 16 x i8> %vec, <vscale x 1 x i8> %subvec) {
+; CHECK-LABEL: insert_nxv16i8_nxv1i8_7:
+; CHECK:       # %bb.0:
+; CHECK-NEXT:    csrr a0, vlenb
+; CHECK-NEXT:    srli a0, a0, 3
+; CHECK-NEXT:    slli a1, a0, 3
+; CHECK-NEXT:    sub a0, a1, a0
+; CHECK-NEXT:    vsetvli a1, a1, e8,m1,tu,mu
+; CHECK-NEXT:    vslideup.vx v8, v10, a0
+; CHECK-NEXT:    ret
+  %v = call <vscale x 16 x i8> @llvm.experimental.vector.insert.nxv1i8.nxv16i8(<vscale x 16 x i8> %vec, <vscale x 1 x i8> %subvec, i64 7)
+  ret <vscale x 16 x i8> %v
+}
+
+define <vscale x 16 x i8> @insert_nxv16i8_nxv1i8_15(<vscale x 16 x i8> %vec, <vscale x 1 x i8> %subvec) {
+; CHECK-LABEL: insert_nxv16i8_nxv1i8_15:
+; CHECK:       # %bb.0:
+; CHECK-NEXT:    csrr a0, vlenb
+; CHECK-NEXT:    srli a0, a0, 3
+; CHECK-NEXT:    slli a1, a0, 3
+; CHECK-NEXT:    sub a0, a1, a0
+; CHECK-NEXT:    vsetvli a1, a1, e8,m1,tu,mu
+; CHECK-NEXT:    vslideup.vx v9, v10, a0
+; CHECK-NEXT:    ret
+  %v = call <vscale x 16 x i8> @llvm.experimental.vector.insert.nxv1i8.nxv16i8(<vscale x 16 x i8> %vec, <vscale x 1 x i8> %subvec, i64 15)
+  ret <vscale x 16 x i8> %v
+}
+
+define <vscale x 32 x half> @insert_nxv32f16_nxv2f16_0(<vscale x 32 x half> %vec, <vscale x 2 x half> %subvec) {
+; CHECK-LABEL: insert_nxv32f16_nxv2f16_0:
+; CHECK:       # %bb.0:
+; CHECK-NEXT:    csrr a0, vlenb
+; CHECK-NEXT:    srli a0, a0, 2
+; CHECK-NEXT:    vsetvli a0, a0, e16,m1,tu,mu
+; CHECK-NEXT:    vslideup.vi v8, v16, 0
+; CHECK-NEXT:    ret
+  %v = call <vscale x 32 x half> @llvm.experimental.vector.insert.nxv2f16.nxv32f16(<vscale x 32 x half> %vec, <vscale x 2 x half> %subvec, i64 0)
+  ret <vscale x 32 x half> %v
+}
+
+define <vscale x 32 x half> @insert_nxv32f16_nxv2f16_2(<vscale x 32 x half> %vec, <vscale x 2 x half> %subvec) {
+; CHECK-LABEL: insert_nxv32f16_nxv2f16_2:
+; CHECK:       # %bb.0:
+; CHECK-NEXT:    csrr a0, vlenb
+; CHECK-NEXT:    srli a0, a0, 2
+; CHECK-NEXT:    add a1, a0, a0
+; CHECK-NEXT:    vsetvli a1, a1, e16,m1,tu,mu
+; CHECK-NEXT:    vslideup.vx v8, v16, a0
+; CHECK-NEXT:    ret
+  %v = call <vscale x 32 x half> @llvm.experimental.vector.insert.nxv2f16.nxv32f16(<vscale x 32 x half> %vec, <vscale x 2 x half> %subvec, i64 2)
+  ret <vscale x 32 x half> %v
+}
+
+define <vscale x 32 x half> @insert_nxv32f16_nxv2f16_26(<vscale x 32 x half> %vec, <vscale x 2 x half> %subvec) {
+; CHECK-LABEL: insert_nxv32f16_nxv2f16_26:
+; CHECK:       # %bb.0:
+; CHECK-NEXT:    csrr a0, vlenb
+; CHECK-NEXT:    srli a0, a0, 2
+; CHECK-NEXT:    add a1, a0, a0
+; CHECK-NEXT:    vsetvli a1, a1, e16,m1,tu,mu
+; CHECK-NEXT:    vslideup.vx v14, v16, a0
+; CHECK-NEXT:    ret
+  %v = call <vscale x 32 x half> @llvm.experimental.vector.insert.nxv2f16.nxv32f16(<vscale x 32 x half> %vec, <vscale x 2 x half> %subvec, i64 26)
+  ret <vscale x 32 x half> %v
+}
+
+define <vscale x 32 x half> @insert_nxv32f16_undef_nxv1f16_0(<vscale x 1 x half> %subvec) {
+; CHECK-LABEL: insert_nxv32f16_undef_nxv1f16_0:
+; CHECK:       # %bb.0:
+; CHECK-NEXT:    # kill: def $v8 killed $v8 def $v8m8
+; CHECK-NEXT:    ret
+  %v = call <vscale x 32 x half> @llvm.experimental.vector.insert.nxv1f16.nxv32f16(<vscale x 32 x half> undef, <vscale x 1 x half> %subvec, i64 0)
+  ret <vscale x 32 x half> %v
+}
+
+define <vscale x 32 x half> @insert_nxv32f16_undef_nxv1f16_26(<vscale x 1 x half> %subvec) {
+; CHECK-LABEL: insert_nxv32f16_undef_nxv1f16_26:
+; CHECK:       # %bb.0:
+; CHECK-NEXT:    csrr a0, vlenb
+; CHECK-NEXT:    srli a0, a0, 3
+; CHECK-NEXT:    slli a1, a0, 1
+; CHECK-NEXT:    add a0, a1, a0
+; CHECK-NEXT:    vsetvli a0, a0, e16,m1,ta,mu
+; CHECK-NEXT:    vslideup.vx v22, v8, a1
+; CHECK-NEXT:    vmv8r.v v8, v16
+; CHECK-NEXT:    ret
+  %v = call <vscale x 32 x half> @llvm.experimental.vector.insert.nxv1f16.nxv32f16(<vscale x 32 x half> undef, <vscale x 1 x half> %subvec, i64 26)
+  ret <vscale x 32 x half> %v
+}
+
+declare <vscale x 16 x i8> @llvm.experimental.vector.insert.nxv1i8.nxv16i8(<vscale x 16 x i8>, <vscale x 1 x i8>, i64)
+
+declare <vscale x 32 x half> @llvm.experimental.vector.insert.nxv1f16.nxv32f16(<vscale x 32 x half>, <vscale x 1 x half>, i64)
+declare <vscale x 32 x half> @llvm.experimental.vector.insert.nxv2f16.nxv32f16(<vscale x 32 x half>, <vscale x 2 x half>, i64)
 
 declare <vscale x 8 x i32> @llvm.experimental.vector.insert.nxv2i32.nxv8i32(<vscale x 8 x i32>, <vscale x 2 x i32>, i64 %idx)
 declare <vscale x 8 x i32> @llvm.experimental.vector.insert.nxv4i32.nxv8i32(<vscale x 8 x i32>, <vscale x 4 x i32>, i64 %idx)
 
 declare <vscale x 16 x i32> @llvm.experimental.vector.insert.nxv1i32.nxv16i32(<vscale x 16 x i32>, <vscale x 1 x i32>, i64 %idx)
 declare <vscale x 16 x i32> @llvm.experimental.vector.insert.nxv2i32.nxv16i32(<vscale x 16 x i32>, <vscale x 2 x i32>, i64 %idx)
 declare <vscale x 16 x i32> @llvm.experimental.vector.insert.nxv4i32.nxv16i32(<vscale x 16 x i32>, <vscale x 4 x i32>, i64 %idx)
 declare <vscale x 16 x i32> @llvm.experimental.vector.insert.nxv8i32.nxv16i32(<vscale x 16 x i32>, <vscale x 8 x i32>, i64 %idx)