[RISCV] Peephole optimization to fold merge.vvm and unmasked intrinsics.

[fakepaper56]

Summary

The patch uses peephole method to fold merge.vvm and unmasked intrinsics to
masked intrinsics. Using peephole intead of tablegen patterns is to avoid large
auto gnerated code.

Note: The patch ignores segment loads since I don't know how to test them.

[reames]

Have you looked at doing this as a dag combine instead of a post-isel peephole? I believe we should have SDNodes with masks for all of these, doing the local peephole in a combine and letting it iterate seems more straight forward. (I'm probably missing something, right?)

llvm/lib/Target/RISCV/RISCVISelDAGToDAG.cpp

154

Order of operations - it seems like this transform should probably be done in the prior loop. In particular, we probably want to fold the mask into the prior instruction before doPeepholeMaskedRVV turns it into an unmasked variant.

2530

Extract out a helper function for the opcode matching please.

2559

I believe you have a missing check here. Don't we need the passthrough for True to be False? Otherwise, it seems like the vmerge might be merging two unrelated values?

We could potentially phrase this as an eager demanded lane style push back of the mask which wouldn't require that check, but that doesn't seem consistent with your code below.

2608

This doesn't make any sense. The only way this could succeed is if the mask was all ones; we shouldn't have a merge with an all ones mask to begin with right?

llvm/test/CodeGen/RISCV/rvv/rvv-peephole-vmerge-vops.ll

1

Can you precommit these tests?

Please also add tests for fixed-length vectors with explicit vector length set on the runline. I remember looking at this a bit before and convincing myself there were some complicating issues with fixed for this, but not what those were unfortunately.

[craig.topper]

In D130442#3685850, @reames wrote:

Have you looked at doing this as a dag combine instead of a post-isel peephole? I believe we should have SDNodes with masks for all of these, doing the local peephole in a combine and letting it iterate seems more straight forward. (I'm probably missing something, right?)

Most of the _VL nodes don't have a passthru operand right now. Nor do they have a policy operand. Important since this patch is focusing on the tail undisturbed case. That would need to change before this could be a DAG combine.

llvm/lib/Target/RISCV/RISCVISelDAGToDAG.cpp

2559

This is only handling the case for True being an unmasked, tail agnostic instruction. That's why it calls lookupMaskedIntrinsicByUnmaskedTA below.

2608

A vp.merge intrinsic can be emitted by the BSC vectorizer to represent a tail undisturbed merge. This happens in reduction loops where the last iteration has less elements than the previous iteration. So we need to merge the upper elements from a previous iteration under a tail undisturbed policy. The mask on the vp.merge would be all 1s.

[reames]

In D130442#3685894, @craig.topper wrote:

In D130442#3685850, @reames wrote:

Have you looked at doing this as a dag combine instead of a post-isel peephole? I believe we should have SDNodes with masks for all of these, doing the local peephole in a combine and letting it iterate seems more straight forward. (I'm probably missing something, right?)

Most of the _VL nodes don't have a passthru operand right now. Nor do they have a policy operand. Important since this patch is focusing on the tail undisturbed case. That would need to change before this could be a DAG combine.

Not a strongly held opinion but... maybe we should just do that?

llvm/lib/Target/RISCV/RISCVISelDAGToDAG.cpp

2559

Right, but that's insufficient.

Converting:
vadd v1, v2, v3
vmul v4, v5, v6
vmerge v7, v1, v4, v0

To:
vadd v7 v2, v3, v0
vmul v4, v5, v6

Doesn't seem like what we want?

There's a variant here which would be legal, but it doesn't seem to be the transform implemented here. That would be:
vadd v7 v2, v3, v0
v0 = not v0
vmul v7, v5, v6, v0

2608

Wouldn't an insert_sub_vector be a better representation here? Or is your point that we had an insert_sub_vector, and that got lowered to the vmerge? I'm guessing yes, that's your point because thinking that through seems to make sense.

I guess we don't have a dedicated insert_sub_reg analogous instruction. VADD could work - in analogy to scalar move idioms - but I don't have a strong reason to prefer that over vmerge if that's what we're currently using.

[craig.topper]

llvm/lib/Target/RISCV/RISCVISelDAGToDAG.cpp

2559

Note the earlier check that the Merge operand and False operand of the VMERGE are the same.

2608

The register size didn’t change. Only the VL. There is no statically sized subvector.

[fakepaper56]

llvm/lib/Target/RISCV/RISCVISelDAGToDAG.cpp

154

The patch is try to work on unmasked intrinsics and some unmasked intrinsics are generated by doPeepholeMaskedRVV. Take an example, vp.add with all-one mask need doPeepholeMaskedRVV to be transformed to unmasked intrinsic.

2559

Yeah. I thought about the scenario during writing the patch, but I wanted to decrease the complexity of the patch and I will do those work in the future. Maybe I could add some TODO in the comments?

[fakepaper56]

1. Use a lambda function to identify a node is VMERGE TU node.
2. Add comments.

[fakepaper56]

llvm/lib/Target/RISCV/RISCVISelDAGToDAG.cpp

2530

Done.

[fakepaper56]

Rebase to precommited test revision D130753.

[craig.topper]

llvm/lib/Target/RISCV/RISCVISelDAGToDAG.cpp

2610

This does not handle the chain output of True correctly if it has users. We would need to replace True.getValue(1) with N->getValue(1). ReplaceUses will only replace the direct users of N.

[craig.topper]

llvm/lib/Target/RISCV/RISCVISelDAGToDAG.cpp

2610

Simple test case

define void @vpmerge_vpload_store(<vscale x 2 x i32> %passthru, <vscale x 2 x i32> * %p, <vscale x 2 x i1> %m, i32 zeroext %vl) {
; CHECK-LABEL: vpmerge_vpload_store:                                             
; CHECK:       # %bb.0:                                                          
; CHECK-NEXT:    vsetvli zero, a1, e32, m1, tu, mu                               
; CHECK-NEXT:    vle32.v v8, (a0), v0.t                                          
; CHECK-NEXT:    vs1r.v v8, (a0)                                                 
; CHECK-NEXT:    ret                                                             
  %splat = insertelement <vscale x 2 x i1> poison, i1 -1, i32 0                  
  %mask = shufflevector <vscale x 2 x i1> %splat, <vscale x 2 x i1> poison, <vscale x 2 x i32> zeroinitializer
  %a = call <vscale x 2 x i32> @llvm.vp.load.nxv2i32.p0nxv2i32(<vscale x 2 x i32> * %p, <vscale x 2 x i1> %mask, i32 %vl)
  %b = call <vscale x 2 x i32> @llvm.vp.merge.nxv2i32(<vscale x 2 x i1> %m, <vscale x 2 x i32> %a, <vscale x 2 x i32> %passthru, i32 %vl)
  store <vscale x 2 x i32> %b, <vscale x 2 x i32> * %p                           
  ret void                                                                       
}

Right after isel the MachineIR is

# Machine code for function vpmerge_vpload_store: IsSSA, TracksLiveness          
Function Live Ins: $v8 in %0, $x10 in %1, $v0 in %2, $x11 in %3                  
                                                                                 
bb.0 (%ir-block.0):                                                              
  liveins: $v8, $x10, $v0, $x11                                                  
  %3:gprnox0 = COPY $x11                                                         
  %2:vr = COPY $v0                                                               
  %1:gpr = COPY $x10                                                             
  %0:vrnov0 = COPY $v8                                                           
  %4:vr = PseudoVLE32_V_M1 %1:gpr, %3:gprnox0, 5 :: (load unknown-size from %ir.p, align 8)
  $v0 = COPY %2:vr                                                               
  %5:vrnov0 = PseudoVLE32_V_M1_MASK %0:vrnov0(tied-def 0), %1:gpr, $v0, %3:gprnox0, 5, 0
  VS1R_V killed %5:vrnov0, %1:gpr :: (store unknown-size into %ir.p, align 8)    
  PseudoRET                                                                      
                                                                                 
# End machine code for function vpmerge_vpload_store.

Notice the two VLEs. Dead code elimination will eventually delete the extra one, but it shouldn't have to. For a more complex test we might put loads and stores in the wrong order in the MachineIR.

[fakepaper56]

Address Craig's comment and add machine ir tests.

[fakepaper56]

llvm/lib/Target/RISCV/RISCVISelDAGToDAG.cpp

2610

Done. Thank you find the bug.

[fakepaper56]

Refine commit message.

[craig.topper]

llvm/lib/Target/RISCV/RISCVISelDAGToDAG.cpp

2614

Assert that True.getResNo() == 0

[fakepaper56]

Address Craig's comment by adding assertion about ResNo of True.

[fakepaper56]

Rebase and ping.

[craig.topper]

LGTM

[fakepaper56]

@reames Do you have any concerns?

[craig.topper]

In D130442#3708917, @fakepaper56 wrote:

@reames Do you have any concerns?

I spoke with @reames yesterday and I think he's ok with this. Philip, can you confirm?

[reames]

In D130442#3713243, @craig.topper wrote:

In D130442#3708917, @fakepaper56 wrote:

@reames Do you have any concerns?

I spoke with @reames yesterday and I think he's ok with this. Philip, can you confirm?

Go for it. @craig.topper convinced me I was wrong about the potential correctness issue, and I have no strong opinion on whether this is the right approach. If it's not, we can always change course and remove the code later.

[fakepaper56]

MERGE_VVM -> VMERGE_VVM, merge.vvm -> vmerge.vvm

[fakepaper56]

I am sorry that I updated the revision to fix typos after the agreement and I will land the revision 48 hours later if nobody will give comments.

Go for it. @craig.topper convinced me I was wrong about the potential correctness issue, and I have no strong opinion on whether this is the right approach. If it's not, we can always change course and remove the code later.

[craig.topper]

LGTM

llvm/lib/Target/RISCV/RISCVISelDAGToDAG.h

public:
   }
 
 // Include the pieces autogenerated from the target description.
 #include "RISCVGenDAGISel.inc"
 
 private:
   bool doPeepholeSExtW(SDNode *Node);
   bool doPeepholeMaskedRVV(SDNode *Node);
+  bool doPeepholeMergeVVMFold();
 };
 
 namespace RISCV {
 struct VLSEGPseudo {
   uint16_t NF : 4;
   uint16_t Masked : 1;
   uint16_t IsTU : 1;
   uint16_t Strided : 1;

llvm/lib/Target/RISCV/RISCVISelDAGToDAG.cpp

if (N->use_empty() || !N->isMachineOpcode())
       continue;
 
     MadeChange |= doPeepholeSExtW(N);
     MadeChange |= doPeepholeMaskedRVV(N);
   }
 
   CurDAG->setRoot(Dummy.getValue());
 
+  MadeChange |= doPeepholeMergeVVMFold();
+
   if (MadeChange)
     CurDAG->RemoveDeadNodes();
 }
 
 static SDNode *selectImmSeq(SelectionDAG *CurDAG, const SDLoc &DL, const MVT VT,
                             RISCVMatInt::InstSeq &Seq) {
   SDNode *Result = nullptr;
   SDValue SrcReg = CurDAG->getRegister(RISCV::X0, VT);
if (auto *TGlued = Glued->getGluedNode())
     Ops.push_back(SDValue(TGlued, TGlued->getNumValues() - 1));
 
   SDNode *Result = CurDAG->getMachineNode(Opc, SDLoc(N), N->getVTList(), Ops);
   ReplaceUses(N, Result);
 
   return true;
 }
 
+// Try to fold MERGE_VVM with unmasked intrinsic to masked intrinsic. The
+// peephole only deals with MERGE_VVM which is TU and has false operand same as
+// its true operand now. E.g. (VMERGE_VVM_M1_TU False, False, (VADD_M1 ...),
+// ...) -> (VADD_VV_M1_MASK)
+bool RISCVDAGToDAGISel::doPeepholeMergeVVMFold() {
+  bool MadeChange = false;
+  SelectionDAG::allnodes_iterator Position = CurDAG->allnodes_end();
+
+  while (Position != CurDAG->allnodes_begin()) {
+    SDNode *N = &*--Position;
+    if (N->use_empty() || !N->isMachineOpcode())
+      continue;
+
+    auto IsVMergeTU = [](unsigned Opcode) {
+      return Opcode == RISCV::PseudoVMERGE_VVM_MF8_TU ||
+             Opcode == RISCV::PseudoVMERGE_VVM_MF4_TU ||
+             Opcode == RISCV::PseudoVMERGE_VVM_MF2_TU ||
+             Opcode == RISCV::PseudoVMERGE_VVM_M1_TU ||
+             Opcode == RISCV::PseudoVMERGE_VVM_M2_TU ||
+             Opcode == RISCV::PseudoVMERGE_VVM_M4_TU ||
+             Opcode == RISCV::PseudoVMERGE_VVM_M8_TU;
+    };
+
+    unsigned Opc = N->getMachineOpcode();
+    // TODO: Also deal with TA VMerge nodes.
+    if (!IsVMergeTU(Opc))
+      continue;
+
+    SDValue Merge = N->getOperand(0);
+    SDValue False = N->getOperand(1);
+    SDValue True = N->getOperand(2);
+    SDValue Mask = N->getOperand(3);
+    SDValue VL = N->getOperand(4);
+
+    if (Merge != False)
+      continue;
+
+    assert(True.getResNo() == 0 &&
+           "Expect True is the first output of an instruction.");
+
+    // Need N is the exactly one using True.
+    if (!True.hasOneUse())
+      continue;
+
+    if (!True.isMachineOpcode())
+      continue;
+
+    unsigned TrueOpc = True.getMachineOpcode();
+
+    // Skip if True has merge operand.
+    // TODO: Deal with True having same merge operand with N.
+    if (RISCVII::hasMergeOp(TII->get(TrueOpc).TSFlags))
+      continue;
+
+    // Skip if True has side effect.
+    // TODO: Support velff and vlsegff.
+    if (TII->get(TrueOpc).hasUnmodeledSideEffects())
+      continue;
+
+    // Only deal with True when True is unmasked intrinsic now.
+    const RISCV::RISCVMaskedPseudoInfo *Info =
+        RISCV::lookupMaskedIntrinsicByUnmaskedTA(TrueOpc);
+
+    if (!Info)
+      continue;
+
+    // The last operand of unmasked intrinsic should be sew or chain.
+    bool HasChainOp =
+        True.getOperand(True.getNumOperands() - 1).getValueType() == MVT::Other;
+
+    // Need True has same VL with N.
+    unsigned TrueVLIndex = True.getNumOperands() - HasChainOp - 2;
+    SDValue TrueVL = True.getOperand(TrueVLIndex);
+    if (TrueVL != VL)
+      continue;
+
+    SDLoc DL(N);
+    unsigned MaskedOpc = Info->MaskedPseudo;
+    SmallVector<SDValue, 8> Ops;
+    Ops.push_back(Merge);
+    Ops.append(True->op_begin(), True->op_begin() + TrueVLIndex);
+    Ops.append({Mask, VL, /* SEW */ True.getOperand(TrueVLIndex + 1)});
+
+    if (RISCVII::hasVecPolicyOp(TII->get(MaskedOpc).TSFlags))
+      Ops.push_back(
+          CurDAG->getTargetConstant(/* TUMU */ 0, DL, Subtarget->getXLenVT()));
+
+    // Result node should have chain operand of True.
+    if (HasChainOp)
+      Ops.push_back(True.getOperand(True.getNumOperands() - 1));
+
+    // Result node should take over glued node of N.
+    if (N->getGluedNode())
+      Ops.push_back(N->getOperand(N->getNumOperands() - 1));
+
+    SDNode *Result =
+        CurDAG->getMachineNode(MaskedOpc, DL, True->getVTList(), Ops);
+
+    // Replace vmerge.vvm node by Result.
+    ReplaceUses(SDValue(N, 0), SDValue(Result, 0));
+
+    // Replace another value of True. E.g. chain and VL.
+    for (unsigned Idx = 1; Idx < True->getNumValues(); ++Idx)
+      ReplaceUses(True.getValue(Idx), SDValue(Result, Idx));
+
+    // Try to transform Result to unmasked intrinsic.
+    doPeepholeMaskedRVV(Result);
+    MadeChange = true;
+  }
+  return MadeChange;
+}
+
 // This pass converts a legalized DAG into a RISCV-specific DAG, ready
 // for instruction scheduling.
 FunctionPass *llvm::createRISCVISelDag(RISCVTargetMachine &TM,
                                        CodeGenOpt::Level OptLevel) {
   return new RISCVDAGToDAGISel(TM, OptLevel);
 }

llvm/lib/Target/RISCV/RISCVInstrInfoVPseudos.td

class RISCVVLE<bit M, bit TU, bit Str, bit F, bits<3> S, bits<3> L> {
   bits<1> IsTU = TU;
   bits<1> Strided = Str;
   bits<1> FF = F;
   bits<3> Log2SEW = S;
   bits<3> LMUL = L;
   Pseudo Pseudo = !cast<Pseudo>(NAME);
 }
 
+def lookupMaskedIntrinsicByUnmaskedTA : SearchIndex {
+  let Table = RISCVMaskedPseudosTable;
+  let Key = ["UnmaskedPseudo"];
+}
+
 def RISCVVLETable : GenericTable {
   let FilterClass = "RISCVVLE";
   let CppTypeName = "VLEPseudo";
   let Fields = ["Masked", "IsTU", "Strided", "FF", "Log2SEW", "LMUL", "Pseudo"];
   let PrimaryKey = ["Masked", "IsTU", "Strided", "FF", "Log2SEW", "LMUL"];
   let PrimaryKeyName = "getVLEPseudo";
 }
 
foreach lmul = MxSet<eew>.m in {
         def "E" # eew # "_V_" # LInfo :
           VPseudoUSLoadNoMask<vreg, eew>,
           VLESched<eew>;
         def "E" # eew # "_V_" # LInfo # "_TU":
           VPseudoUSLoadNoMaskTU<vreg, eew>,
           VLESched<eew>;
         def "E" # eew # "_V_" # LInfo # "_MASK" :
           VPseudoUSLoadMask<vreg, eew>,
+          RISCVMaskedPseudo</*MaskOpIdx*/ 2>,
           VLESched<eew>;
       }
     }
   }
 }
 
 multiclass VPseudoFFLoad {
   foreach eew = EEWList in {
     foreach lmul = MxSet<eew>.m in {
       defvar LInfo = lmul.MX;
       defvar vreg = lmul.vrclass;
       let VLMul = lmul.value in {
         def "E" # eew # "FF_V_" # LInfo:
           VPseudoUSLoadFFNoMask<vreg, eew>,
           VLFSched<eew>;
         def "E" # eew # "FF_V_" # LInfo # "_TU":
           VPseudoUSLoadFFNoMaskTU<vreg, eew>,
           VLFSched<eew>;
         def "E" # eew # "FF_V_" # LInfo # "_MASK":
           VPseudoUSLoadFFMask<vreg, eew>,
+          RISCVMaskedPseudo</*MaskOpIdx*/ 2>,
           VLFSched<eew>;
       }
     }
   }
 }
 
 multiclass VPseudoLoadMask {
   foreach mti = AllMasks in {
     let VLMul = mti.LMul.value in {
       def "_V_" # mti.BX : VPseudoUSLoadNoMask<VR, /*EEW*/1, /*DummyMask*/0>;
     }
   }
 }
 
 multiclass VPseudoSLoad {
   foreach eew = EEWList in {
     foreach lmul = MxSet<eew>.m in {
       defvar LInfo = lmul.MX;
       defvar vreg = lmul.vrclass;
       let VLMul = lmul.value in {
         def "E" # eew # "_V_" # LInfo : VPseudoSLoadNoMask<vreg, eew>,
                                         VLSSched<eew>;
         def "E" # eew # "_V_" # LInfo # "_TU": VPseudoSLoadNoMaskTU<vreg, eew>,
                                         VLSSched<eew>;
-        def "E" # eew # "_V_" # LInfo # "_MASK" : VPseudoSLoadMask<vreg, eew>,
-                                                  VLSSched<eew>;
+        def "E" # eew # "_V_" # LInfo # "_MASK" :
+          VPseudoSLoadMask<vreg, eew>,
+          RISCVMaskedPseudo</*MaskOpIdx*/ 3>,
       }
     }
   }
 }
 
 multiclass VPseudoILoad<bit Ordered> {
   foreach eew = EEWList in {
     foreach sew = EEWList in {
foreach sew = EEWList in {
             def "EI" # eew # "_V_" # IdxLInfo # "_" # LInfo :
               VPseudoILoadNoMask<Vreg, IdxVreg, eew, idx_lmul.value, Ordered, HasConstraint>,
               VLXSched<eew, Order>;
             def "EI" # eew # "_V_" # IdxLInfo # "_" # LInfo # "_TU":
               VPseudoILoadNoMaskTU<Vreg, IdxVreg, eew, idx_lmul.value, Ordered, HasConstraint>,
               VLXSched<eew, Order>;
             def "EI" # eew # "_V_" # IdxLInfo # "_" # LInfo # "_MASK" :
               VPseudoILoadMask<Vreg, IdxVreg, eew, idx_lmul.value, Ordered, HasConstraint>,
+              RISCVMaskedPseudo</*MaskOpIdx*/ 3>,
               VLXSched<eew, Order>;
           }
         }
       }
     }
   }
 }
 

 multiclass VPseudoVID_V {
   foreach m = MxList in {
     let VLMul = m.value in {
       def "_V_" # m.MX : VPseudoNullaryNoMask<m.vrclass>,
                          Sched<[WriteVMIdxV, ReadVMask]>;
       def "_V_" # m.MX # "_TU": VPseudoNullaryNoMaskTU<m.vrclass>,
                                 Sched<[WriteVMIdxV, ReadVMask]>;
       def "_V_" # m.MX # "_MASK" : VPseudoNullaryMask<m.vrclass>,
+                                   RISCVMaskedPseudo</*MaskOpIdx*/ 1>,
                                    Sched<[WriteVMIdxV, ReadVMask]>;
     }
   }
 }
 
 multiclass VPseudoNullaryPseudoM <string BaseInst> {
   foreach mti = AllMasks in {
     let VLMul = mti.LMul.value in {
       def "_M_" # mti.BX : VPseudoNullaryPseudoM<BaseInst # "_MM">;
     }
   }
 }
 
 multiclass VPseudoVIOT_M {
   defvar constraint = "@earlyclobber $rd";
   foreach m = MxList in {
     let VLMul = m.value in {
       def "_" # m.MX : VPseudoUnaryNoMask<m.vrclass, VR, constraint>,
                        Sched<[WriteVMIotV, ReadVMIotV, ReadVMask]>;
       def "_" # m.MX # "_TU" : VPseudoUnaryNoMaskTU<m.vrclass, VR, constraint>,
                                Sched<[WriteVMIotV, ReadVMIotV, ReadVMask]>;
       def "_" # m.MX # "_MASK" : VPseudoUnaryMaskTA<m.vrclass, VR, constraint>,
+                                 RISCVMaskedPseudo</*MaskOpIdx*/ 2>,
                                  Sched<[WriteVMIotV, ReadVMIotV, ReadVMask]>;
     }
   }
 }
 
 multiclass VPseudoVCPR_V {
   foreach m = MxList in {
     let VLMul = m.value in

 multiclass VPseudoVCLS_V {
   foreach m = MxListF in {
     let VLMul = m.value in {
       def "_V_" # m.MX : VPseudoUnaryNoMask<m.vrclass, m.vrclass>,
                          Sched<[WriteVFClassV, ReadVFClassV, ReadVMask]>;
       def "_V_" # m.MX # "_TU": VPseudoUnaryNoMaskTU<m.vrclass, m.vrclass>,
                                 Sched<[WriteVFClassV, ReadVFClassV, ReadVMask]>;
       def "_V_" # m.MX # "_MASK" : VPseudoUnaryMaskTA<m.vrclass, m.vrclass>,
+                                   RISCVMaskedPseudo</*MaskOpIdx*/ 2>,
                                    Sched<[WriteVFClassV, ReadVFClassV, ReadVMask]>;
     }
   }
 }
 
 multiclass VPseudoVSQR_V {
   foreach m = MxListF in {
     let VLMul = m.value in {
       def "_V_" # m.MX : VPseudoUnaryNoMask<m.vrclass, m.vrclass>,
                          Sched<[WriteVFSqrtV, ReadVFSqrtV, ReadVMask]>;
       def "_V_" # m.MX # "_TU": VPseudoUnaryNoMaskTU<m.vrclass, m.vrclass>,
                                 Sched<[WriteVFSqrtV, ReadVFSqrtV, ReadVMask]>;
       def "_V_" # m.MX # "_MASK" : VPseudoUnaryMaskTA<m.vrclass, m.vrclass>,
+                                   RISCVMaskedPseudo</*MaskOpIdx*/ 2>,
                                    Sched<[WriteVFSqrtV, ReadVFSqrtV, ReadVMask]>;
     }
   }
 }
 
 multiclass VPseudoVRCP_V {
   foreach m = MxListF in {
     let VLMul = m.value in {
       def "_V_" # m.MX : VPseudoUnaryNoMask<m.vrclass, m.vrclass>,
                          Sched<[WriteVFRecpV, ReadVFRecpV, ReadVMask]>;
       def "_V_" # m.MX # "_TU": VPseudoUnaryNoMaskTU<m.vrclass, m.vrclass>,
                                 Sched<[WriteVFRecpV, ReadVFRecpV, ReadVMask]>;
       def "_V_" # m.MX # "_MASK" : VPseudoUnaryMaskTA<m.vrclass, m.vrclass>,
+                                   RISCVMaskedPseudo</*MaskOpIdx*/ 2>,
                                    Sched<[WriteVFRecpV, ReadVFRecpV, ReadVMask]>;
     }
   }
 }
 
 multiclass PseudoVEXT_VF2 {
   defvar constraints = "@earlyclobber $rd";
   foreach m = MxListVF2 in

llvm/test/CodeGen/RISCV/rvv/fixed-vectors-peephole-vmerge-vops.ll

-; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py
 ; RUN: llc < %s -mtriple=riscv64 -mattr=+v -riscv-v-vector-bits-min=256 | FileCheck %s
+; RUN: llc < %s -mtriple=riscv64 -mattr=+v -riscv-v-vector-bits-min=256 -stop-after=finalize-isel | FileCheck %s --check-prefix=MIR
 
 declare <8 x i16> @llvm.vp.merge.nxv2i16(<8 x i1>, <8 x i16>, <8 x i16>, i32)
 declare <8 x i32> @llvm.vp.merge.nxv2i32(<8 x i1>, <8 x i32>, <8 x i32>, i32)
 declare <8 x float> @llvm.vp.merge.nxv2f32(<8 x i1>, <8 x float>, <8 x float>, i32)
 declare <8 x double> @llvm.vp.merge.nxv2f64(<8 x i1>, <8 x double>, <8 x double>, i32)
 
 ; Test binary operator with vp.merge and vp.smax.
 declare <8 x i32> @llvm.vp.add.nxv2i32(<8 x i32>, <8 x i32>, <8 x i1>, i32)
 define <8 x i32> @vpmerge_vpadd(<8 x i32> %passthru, <8 x i32> %x, <8 x i32> %y, <8 x i1> %m, i32 zeroext %vl) {
 ; CHECK-LABEL: vpmerge_vpadd:
 ; CHECK:       # %bb.0:
-; CHECK-NEXT:    vsetvli zero, a0, e32, m1, ta, mu
-; CHECK-NEXT:    vadd.vv v9, v9, v10
-; CHECK-NEXT:    vsetvli zero, zero, e32, m1, tu, mu
-; CHECK-NEXT:    vmerge.vvm v8, v8, v9, v0
+; CHECK-NEXT:    vsetvli zero, a0, e32, m1, tu, mu
+; CHECK-NEXT:    vadd.vv v8, v9, v10, v0.t
 ; CHECK-NEXT:    ret
+  ; MIR-LABEL: name: vpmerge_vpadd
+  ; MIR: bb.0 (%ir-block.0):
+  ; MIR-NEXT:   liveins: $v8, $v9, $v10, $v0, $x10
+  ; MIR-NEXT: {{  $}}
+  ; MIR-NEXT:   [[COPY:%[0-9]+]]:gprnox0 = COPY $x10
+  ; MIR-NEXT:   [[COPY1:%[0-9]+]]:vr = COPY $v0
+  ; MIR-NEXT:   [[COPY2:%[0-9]+]]:vr = COPY $v10
+  ; MIR-NEXT:   [[COPY3:%[0-9]+]]:vr = COPY $v9
+  ; MIR-NEXT:   [[COPY4:%[0-9]+]]:vrnov0 = COPY $v8
+  ; MIR-NEXT:   $v0 = COPY [[COPY1]]
+  ; MIR-NEXT:   [[PseudoVADD_VV_M1_MASK:%[0-9]+]]:vrnov0 = PseudoVADD_VV_M1_MASK [[COPY4]], [[COPY3]], [[COPY2]], $v0, [[COPY]], 5 /* e32 */, 0
+  ; MIR-NEXT:   $v8 = COPY [[PseudoVADD_VV_M1_MASK]]
+  ; MIR-NEXT:   PseudoRET implicit $v8
   %splat = insertelement <8 x i1> poison, i1 -1, i32 0
   %mask = shufflevector <8 x i1> %splat, <8 x i1> poison, <8 x i32> zeroinitializer
   %a = call <8 x i32> @llvm.vp.add.nxv2i32(<8 x i32> %x, <8 x i32> %y, <8 x i1> %mask, i32 %vl)
   %b = call <8 x i32> @llvm.vp.merge.nxv2i32(<8 x i1> %m, <8 x i32> %a, <8 x i32> %passthru, i32 %vl)
   ret <8 x i32> %b
 }
 
 ; Test glued node of merge should not be deleted.
 declare <8 x i1> @llvm.vp.icmp.nxv2i32(<8 x i32>, <8 x i32>, metadata, <8 x i1>, i32)
 define <8 x i32> @vpmerge_vpadd2(<8 x i32> %passthru, <8 x i32> %x, <8 x i32> %y, i32 zeroext %vl) {
 ; CHECK-LABEL: vpmerge_vpadd2:
 ; CHECK:       # %bb.0:
 ; CHECK-NEXT:    vsetvli zero, a0, e32, m1, ta, mu
-; CHECK-NEXT:    vadd.vv v11, v9, v10
 ; CHECK-NEXT:    vmseq.vv v0, v9, v10
 ; CHECK-NEXT:    vsetvli zero, zero, e32, m1, tu, mu
-; CHECK-NEXT:    vmerge.vvm v8, v8, v11, v0
+; CHECK-NEXT:    vadd.vv v8, v9, v10, v0.t
 ; CHECK-NEXT:    ret
+  ; MIR-LABEL: name: vpmerge_vpadd2
+  ; MIR: bb.0 (%ir-block.0):
+  ; MIR-NEXT:   liveins: $v8, $v9, $v10, $x10
+  ; MIR-NEXT: {{  $}}
+  ; MIR-NEXT:   [[COPY:%[0-9]+]]:gprnox0 = COPY $x10
+  ; MIR-NEXT:   [[COPY1:%[0-9]+]]:vr = COPY $v10
+  ; MIR-NEXT:   [[COPY2:%[0-9]+]]:vr = COPY $v9
+  ; MIR-NEXT:   [[COPY3:%[0-9]+]]:vrnov0 = COPY $v8
+  ; MIR-NEXT:   [[PseudoVMSEQ_VV_M1_:%[0-9]+]]:vr = PseudoVMSEQ_VV_M1 [[COPY2]], [[COPY1]], [[COPY]], 5 /* e32 */
+  ; MIR-NEXT:   $v0 = COPY [[PseudoVMSEQ_VV_M1_]]
+  ; MIR-NEXT:   [[PseudoVADD_VV_M1_MASK:%[0-9]+]]:vrnov0 = PseudoVADD_VV_M1_MASK [[COPY3]], [[COPY2]], [[COPY1]], $v0, [[COPY]], 5 /* e32 */, 0
+  ; MIR-NEXT:   $v8 = COPY [[PseudoVADD_VV_M1_MASK]]
+  ; MIR-NEXT:   PseudoRET implicit $v8
   %splat = insertelement <8 x i1> poison, i1 -1, i32 0
   %mask = shufflevector <8 x i1> %splat, <8 x i1> poison, <8 x i32> zeroinitializer
   %a = call <8 x i32> @llvm.vp.add.nxv2i32(<8 x i32> %x, <8 x i32> %y, <8 x i1> %mask, i32 %vl)
   %m = call <8 x i1> @llvm.vp.icmp.nxv2i32(<8 x i32> %x, <8 x i32> %y, metadata !"eq", <8 x i1> %mask, i32 %vl)
   %b = call <8 x i32> @llvm.vp.merge.nxv2i32(<8 x i1> %m, <8 x i32> %a, <8 x i32> %passthru, i32 %vl)
   ret <8 x i32> %b
 }
 
 ; Test vp.merge have all-ones mask.
 define <8 x i32> @vpmerge_vpadd3(<8 x i32> %passthru, <8 x i32> %x, <8 x i32> %y, i32 zeroext %vl) {
 ; CHECK-LABEL: vpmerge_vpadd3:
 ; CHECK:       # %bb.0:
-; CHECK-NEXT:    vsetivli zero, 8, e8, mf4, ta, mu
-; CHECK-NEXT:    vmset.m v0
-; CHECK-NEXT:    vsetvli zero, a0, e32, m1, ta, mu
-; CHECK-NEXT:    vadd.vv v9, v9, v10
-; CHECK-NEXT:    vsetvli zero, zero, e32, m1, tu, mu
-; CHECK-NEXT:    vmerge.vvm v8, v8, v9, v0
+; CHECK-NEXT:    vsetvli zero, a0, e32, m1, tu, mu
+; CHECK-NEXT:    vadd.vv v8, v9, v10
 ; CHECK-NEXT:    ret
+  ; MIR-LABEL: name: vpmerge_vpadd3
+  ; MIR: bb.0 (%ir-block.0):
+  ; MIR-NEXT:   liveins: $v8, $v9, $v10, $x10
+  ; MIR-NEXT: {{  $}}
+  ; MIR-NEXT:   [[COPY:%[0-9]+]]:gprnox0 = COPY $x10
+  ; MIR-NEXT:   [[COPY1:%[0-9]+]]:vr = COPY $v10
+  ; MIR-NEXT:   [[COPY2:%[0-9]+]]:vr = COPY $v9
+  ; MIR-NEXT:   [[COPY3:%[0-9]+]]:vr = COPY $v8
+  ; MIR-NEXT:   [[PseudoVADD_VV_M1_TU:%[0-9]+]]:vr = PseudoVADD_VV_M1_TU [[COPY3]], [[COPY2]], [[COPY1]], [[COPY]], 5 /* e32 */
+  ; MIR-NEXT:   $v8 = COPY [[PseudoVADD_VV_M1_TU]]
+  ; MIR-NEXT:   PseudoRET implicit $v8
   %splat = insertelement <8 x i1> poison, i1 -1, i32 0
   %mask = shufflevector <8 x i1> %splat, <8 x i1> poison, <8 x i32> zeroinitializer
   %a = call <8 x i32> @llvm.vp.add.nxv2i32(<8 x i32> %x, <8 x i32> %y, <8 x i1> %mask, i32 %vl)
   %b = call <8 x i32> @llvm.vp.merge.nxv2i32(<8 x i1> %mask, <8 x i32> %a, <8 x i32> %passthru, i32 %vl)
   ret <8 x i32> %b
 }
 
 ; Test float binary operator with vp.merge and vp.fadd.
 declare <8 x float> @llvm.vp.fadd.nxv2f32(<8 x float>, <8 x float>, <8 x i1>, i32)
 define <8 x float> @vpmerge_vpfadd(<8 x float> %passthru, <8 x float> %x, <8 x float> %y, <8 x i1> %m, i32 zeroext %vl) {
 ; CHECK-LABEL: vpmerge_vpfadd:
 ; CHECK:       # %bb.0:
-; CHECK-NEXT:    vsetvli zero, a0, e32, m1, ta, mu
-; CHECK-NEXT:    vfadd.vv v9, v9, v10
-; CHECK-NEXT:    vsetvli zero, zero, e32, m1, tu, mu
-; CHECK-NEXT:    vmerge.vvm v8, v8, v9, v0
+; CHECK-NEXT:    vsetvli zero, a0, e32, m1, tu, mu
+; CHECK-NEXT:    vfadd.vv v8, v9, v10, v0.t
 ; CHECK-NEXT:    ret
+  ; MIR-LABEL: name: vpmerge_vpfadd
+  ; MIR: bb.0 (%ir-block.0):
+  ; MIR-NEXT:   liveins: $v8, $v9, $v10, $v0, $x10
+  ; MIR-NEXT: {{  $}}
+  ; MIR-NEXT:   [[COPY:%[0-9]+]]:gprnox0 = COPY $x10
+  ; MIR-NEXT:   [[COPY1:%[0-9]+]]:vr = COPY $v0
+  ; MIR-NEXT:   [[COPY2:%[0-9]+]]:vr = COPY $v10
+  ; MIR-NEXT:   [[COPY3:%[0-9]+]]:vr = COPY $v9
+  ; MIR-NEXT:   [[COPY4:%[0-9]+]]:vrnov0 = COPY $v8
+  ; MIR-NEXT:   $v0 = COPY [[COPY1]]
+  ; MIR-NEXT:   [[PseudoVFADD_VV_M1_MASK:%[0-9]+]]:vrnov0 = PseudoVFADD_VV_M1_MASK [[COPY4]], [[COPY3]], [[COPY2]], $v0, [[COPY]], 5 /* e32 */, 0, implicit $frm
+  ; MIR-NEXT:   $v8 = COPY [[PseudoVFADD_VV_M1_MASK]]
+  ; MIR-NEXT:   PseudoRET implicit $v8
   %splat = insertelement <8 x i1> poison, i1 -1, i32 0
   %mask = shufflevector <8 x i1> %splat, <8 x i1> poison, <8 x i32> zeroinitializer
   %a = call <8 x float> @llvm.vp.fadd.nxv2f32(<8 x float> %x, <8 x float> %y, <8 x i1> %mask, i32 %vl)
   %b = call <8 x float> @llvm.vp.merge.nxv2f32(<8 x i1> %m, <8 x float> %a, <8 x float> %passthru, i32 %vl)
   ret <8 x float> %b
 }
 
 ; Test conversion by fptosi.
 declare <8 x i16> @llvm.vp.fptosi.nxv2i16.nxv2f32(<8 x float>, <8 x i1>, i32)
 define <8 x i16> @vpmerge_vpfptosi(<8 x i16> %passthru, <8 x float> %x, <8 x i1> %m, i32 zeroext %vl) {
 ; CHECK-LABEL: vpmerge_vpfptosi:
 ; CHECK:       # %bb.0:
-; CHECK-NEXT:    vsetvli zero, a0, e16, mf2, ta, mu
-; CHECK-NEXT:    vfncvt.rtz.x.f.w v10, v9
-; CHECK-NEXT:    vsetvli zero, zero, e16, mf2, tu, mu
-; CHECK-NEXT:    vmerge.vvm v8, v8, v10, v0
+; CHECK-NEXT:    vsetvli zero, a0, e16, mf2, tu, mu
+; CHECK-NEXT:    vfncvt.rtz.x.f.w v8, v9, v0.t
 ; CHECK-NEXT:    ret
+  ; MIR-LABEL: name: vpmerge_vpfptosi
+  ; MIR: bb.0 (%ir-block.0):
+  ; MIR-NEXT:   liveins: $v8, $v9, $v0, $x10
+  ; MIR-NEXT: {{  $}}
+  ; MIR-NEXT:   [[COPY:%[0-9]+]]:gprnox0 = COPY $x10
+  ; MIR-NEXT:   [[COPY1:%[0-9]+]]:vr = COPY $v0
+  ; MIR-NEXT:   [[COPY2:%[0-9]+]]:vr = COPY $v9
+  ; MIR-NEXT:   [[COPY3:%[0-9]+]]:vrnov0 = COPY $v8
+  ; MIR-NEXT:   $v0 = COPY [[COPY1]]
+  ; MIR-NEXT:   early-clobber %4:vrnov0 = PseudoVFNCVT_RTZ_X_F_W_MF2_MASK [[COPY3]], [[COPY2]], $v0, [[COPY]], 4 /* e16 */, 0
+  ; MIR-NEXT:   $v8 = COPY %4
+  ; MIR-NEXT:   PseudoRET implicit $v8
   %splat = insertelement <8 x i1> poison, i1 -1, i32 0
   %mask = shufflevector <8 x i1> %splat, <8 x i1> poison, <8 x i32> zeroinitializer
   %a = call <8 x i16> @llvm.vp.fptosi.nxv2i16.nxv2f32(<8 x float> %x, <8 x i1> %mask, i32 %vl)
   %b = call <8 x i16> @llvm.vp.merge.nxv2i16(<8 x i1> %m, <8 x i16> %a, <8 x i16> %passthru, i32 %vl)
   ret <8 x i16> %b
 }
 
 ; Test conversion by sitofp.
 declare <8 x float> @llvm.vp.sitofp.nxv2f32.nxv2i64(<8 x i64>, <8 x i1>, i32)
 define <8 x float> @vpmerge_vpsitofp(<8 x float> %passthru, <8 x i64> %x, <8 x i1> %m, i32 zeroext %vl) {
 ; CHECK-LABEL: vpmerge_vpsitofp:
 ; CHECK:       # %bb.0:
-; CHECK-NEXT:    vsetvli zero, a0, e32, m1, ta, mu
-; CHECK-NEXT:    vfncvt.f.x.w v9, v10
-; CHECK-NEXT:    vsetvli zero, zero, e32, m1, tu, mu
-; CHECK-NEXT:    vmerge.vvm v8, v8, v9, v0
+; CHECK-NEXT:    vsetvli zero, a0, e32, m1, tu, mu
+; CHECK-NEXT:    vfncvt.f.x.w v8, v10, v0.t
 ; CHECK-NEXT:    ret
+  ; MIR-LABEL: name: vpmerge_vpsitofp
+  ; MIR: bb.0 (%ir-block.0):
+  ; MIR-NEXT:   liveins: $v8, $v10m2, $v0, $x10
+  ; MIR-NEXT: {{  $}}
+  ; MIR-NEXT:   [[COPY:%[0-9]+]]:gprnox0 = COPY $x10
+  ; MIR-NEXT:   [[COPY1:%[0-9]+]]:vr = COPY $v0
+  ; MIR-NEXT:   [[COPY2:%[0-9]+]]:vrm2 = COPY $v10m2
+  ; MIR-NEXT:   [[COPY3:%[0-9]+]]:vrnov0 = COPY $v8
+  ; MIR-NEXT:   $v0 = COPY [[COPY1]]
+  ; MIR-NEXT:   early-clobber %4:vrnov0 = PseudoVFNCVT_F_X_W_M1_MASK [[COPY3]], [[COPY2]], $v0, [[COPY]], 5 /* e32 */, 0
+  ; MIR-NEXT:   $v8 = COPY %4
+  ; MIR-NEXT:   PseudoRET implicit $v8
   %splat = insertelement <8 x i1> poison, i1 -1, i32 0
   %mask = shufflevector <8 x i1> %splat, <8 x i1> poison, <8 x i32> zeroinitializer
   %a = call <8 x float> @llvm.vp.sitofp.nxv2f32.nxv2i64(<8 x i64> %x, <8 x i1> %mask, i32 %vl)
   %b = call <8 x float> @llvm.vp.merge.nxv2f32(<8 x i1> %m, <8 x float> %a, <8 x float> %passthru, i32 %vl)
   ret <8 x float> %b
 }
 
 ; Test integer extension by vp.zext.
 declare <8 x i32> @llvm.vp.zext.nxv2i32.nxv2i8(<8 x i8>, <8 x i1>, i32)
 define <8 x i32> @vpmerge_vpzext(<8 x i32> %passthru, <8 x i8> %x, <8 x i1> %m, i32 zeroext %vl) {
 ; CHECK-LABEL: vpmerge_vpzext:
 ; CHECK:       # %bb.0:
-; CHECK-NEXT:    vsetvli zero, a0, e32, m1, ta, mu
-; CHECK-NEXT:    vzext.vf4 v10, v9
-; CHECK-NEXT:    vsetvli zero, zero, e32, m1, tu, mu
-; CHECK-NEXT:    vmerge.vvm v8, v8, v10, v0
+; CHECK-NEXT:    vsetvli zero, a0, e32, m1, tu, mu
+; CHECK-NEXT:    vzext.vf4 v8, v9, v0.t
 ; CHECK-NEXT:    ret
+  ; MIR-LABEL: name: vpmerge_vpzext
+  ; MIR: bb.0 (%ir-block.0):
+  ; MIR-NEXT:   liveins: $v8, $v9, $v0, $x10
+  ; MIR-NEXT: {{  $}}
+  ; MIR-NEXT:   [[COPY:%[0-9]+]]:gprnox0 = COPY $x10
+  ; MIR-NEXT:   [[COPY1:%[0-9]+]]:vr = COPY $v0
+  ; MIR-NEXT:   [[COPY2:%[0-9]+]]:vr = COPY $v9
+  ; MIR-NEXT:   [[COPY3:%[0-9]+]]:vrnov0 = COPY $v8
+  ; MIR-NEXT:   $v0 = COPY [[COPY1]]
+  ; MIR-NEXT:   early-clobber %4:vrnov0 = PseudoVZEXT_VF4_M1_MASK [[COPY3]], [[COPY2]], $v0, [[COPY]], 5 /* e32 */, 0
+  ; MIR-NEXT:   $v8 = COPY %4
+  ; MIR-NEXT:   PseudoRET implicit $v8
   %splat = insertelement <8 x i1> poison, i1 -1, i32 0
   %mask = shufflevector <8 x i1> %splat, <8 x i1> poison, <8 x i32> zeroinitializer
   %a = call <8 x i32> @llvm.vp.zext.nxv2i32.nxv2i8(<8 x i8> %x, <8 x i1> %mask, i32 %vl)
   %b = call <8 x i32> @llvm.vp.merge.nxv2i32(<8 x i1> %m, <8 x i32> %a, <8 x i32> %passthru, i32 %vl)
   ret <8 x i32> %b
 }
 
 ; Test integer truncation by vp.trunc.
 declare <8 x i32> @llvm.vp.trunc.nxv2i32.nxv2i64(<8 x i64>, <8 x i1>, i32)
 define <8 x i32> @vpmerge_vptrunc(<8 x i32> %passthru, <8 x i64> %x, <8 x i1> %m, i32 zeroext %vl) {
 ; CHECK-LABEL: vpmerge_vptrunc:
 ; CHECK:       # %bb.0:
-; CHECK-NEXT:    vsetvli zero, a0, e32, m1, ta, mu
-; CHECK-NEXT:    vncvt.x.x.w v9, v10
-; CHECK-NEXT:    vsetvli zero, zero, e32, m1, tu, mu
-; CHECK-NEXT:    vmerge.vvm v8, v8, v9, v0
+; CHECK-NEXT:    vsetvli zero, a0, e32, m1, tu, mu
+; CHECK-NEXT:    vncvt.x.x.w v8, v10, v0.t
 ; CHECK-NEXT:    ret
+  ; MIR-LABEL: name: vpmerge_vptrunc
+  ; MIR: bb.0 (%ir-block.0):
+  ; MIR-NEXT:   liveins: $v8, $v10m2, $v0, $x10
+  ; MIR-NEXT: {{  $}}
+  ; MIR-NEXT:   [[COPY:%[0-9]+]]:gprnox0 = COPY $x10
+  ; MIR-NEXT:   [[COPY1:%[0-9]+]]:vr = COPY $v0
+  ; MIR-NEXT:   [[COPY2:%[0-9]+]]:vrm2 = COPY $v10m2
+  ; MIR-NEXT:   [[COPY3:%[0-9]+]]:vrnov0 = COPY $v8
+  ; MIR-NEXT:   $v0 = COPY [[COPY1]]
+  ; MIR-NEXT:   early-clobber %4:vrnov0 = PseudoVNSRL_WX_M1_MASK [[COPY3]], [[COPY2]], $x0, $v0, [[COPY]], 5 /* e32 */, 0
+  ; MIR-NEXT:   $v8 = COPY %4
+  ; MIR-NEXT:   PseudoRET implicit $v8
   %splat = insertelement <8 x i1> poison, i1 -1, i32 0
   %mask = shufflevector <8 x i1> %splat, <8 x i1> poison, <8 x i32> zeroinitializer
   %a = call <8 x i32> @llvm.vp.trunc.nxv2i32.nxv2i64(<8 x i64> %x, <8 x i1> %mask, i32 %vl)
   %b = call <8 x i32> @llvm.vp.merge.nxv2i32(<8 x i1> %m, <8 x i32> %a, <8 x i32> %passthru, i32 %vl)
   ret <8 x i32> %b
 }
 
 ; Test integer extension by vp.fpext.
 declare <8 x double> @llvm.vp.fpext.nxv2f64.nxv2f32(<8 x float>, <8 x i1>, i32)
 define <8 x double> @vpmerge_vpfpext(<8 x double> %passthru, <8 x float> %x, <8 x i1> %m, i32 zeroext %vl) {
 ; CHECK-LABEL: vpmerge_vpfpext:
 ; CHECK:       # %bb.0:
-; CHECK-NEXT:    vsetvli zero, a0, e32, m1, ta, mu
-; CHECK-NEXT:    vfwcvt.f.f.v v12, v10
-; CHECK-NEXT:    vsetvli zero, zero, e64, m2, tu, mu
-; CHECK-NEXT:    vmerge.vvm v8, v8, v12, v0
+; CHECK-NEXT:    vsetvli zero, a0, e32, m1, tu, mu
+; CHECK-NEXT:    vfwcvt.f.f.v v8, v10, v0.t
 ; CHECK-NEXT:    ret
+  ; MIR-LABEL: name: vpmerge_vpfpext
+  ; MIR: bb.0 (%ir-block.0):
+  ; MIR-NEXT:   liveins: $v8m2, $v10, $v0, $x10
+  ; MIR-NEXT: {{  $}}
+  ; MIR-NEXT:   [[COPY:%[0-9]+]]:gprnox0 = COPY $x10
+  ; MIR-NEXT:   [[COPY1:%[0-9]+]]:vr = COPY $v0
+  ; MIR-NEXT:   [[COPY2:%[0-9]+]]:vr = COPY $v10
+  ; MIR-NEXT:   [[COPY3:%[0-9]+]]:vrm2nov0 = COPY $v8m2
+  ; MIR-NEXT:   $v0 = COPY [[COPY1]]
+  ; MIR-NEXT:   early-clobber %4:vrm2nov0 = PseudoVFWCVT_F_F_V_M1_MASK [[COPY3]], [[COPY2]], $v0, [[COPY]], 5 /* e32 */, 0
+  ; MIR-NEXT:   $v8m2 = COPY %4
+  ; MIR-NEXT:   PseudoRET implicit $v8m2
   %splat = insertelement <8 x i1> poison, i1 -1, i32 0
   %mask = shufflevector <8 x i1> %splat, <8 x i1> poison, <8 x i32> zeroinitializer
   %a = call <8 x double> @llvm.vp.fpext.nxv2f64.nxv2f32(<8 x float> %x, <8 x i1> %mask, i32 %vl)
   %b = call <8 x double> @llvm.vp.merge.nxv2f64(<8 x i1> %m, <8 x double> %a, <8 x double> %passthru, i32 %vl)
   ret <8 x double> %b
 }
 
 ; Test integer truncation by vp.trunc.
 declare <8 x float> @llvm.vp.fptrunc.nxv2f32.nxv2f64(<8 x double>, <8 x i1>, i32)
 define <8 x float> @vpmerge_vpfptrunc(<8 x float> %passthru, <8 x double> %x, <8 x i1> %m, i32 zeroext %vl) {
 ; CHECK-LABEL: vpmerge_vpfptrunc:
 ; CHECK:       # %bb.0:
-; CHECK-NEXT:    vsetvli zero, a0, e32, m1, ta, mu
-; CHECK-NEXT:    vfncvt.f.f.w v9, v10
-; CHECK-NEXT:    vsetvli zero, zero, e32, m1, tu, mu
-; CHECK-NEXT:    vmerge.vvm v8, v8, v9, v0
+; CHECK-NEXT:    vsetvli zero, a0, e32, m1, tu, mu
+; CHECK-NEXT:    vfncvt.f.f.w v8, v10, v0.t
 ; CHECK-NEXT:    ret
+  ; MIR-LABEL: name: vpmerge_vpfptrunc
+  ; MIR: bb.0 (%ir-block.0):
+  ; MIR-NEXT:   liveins: $v8, $v10m2, $v0, $x10
+  ; MIR-NEXT: {{  $}}
+  ; MIR-NEXT:   [[COPY:%[0-9]+]]:gprnox0 = COPY $x10
+  ; MIR-NEXT:   [[COPY1:%[0-9]+]]:vr = COPY $v0
+  ; MIR-NEXT:   [[COPY2:%[0-9]+]]:vrm2 = COPY $v10m2
+  ; MIR-NEXT:   [[COPY3:%[0-9]+]]:vrnov0 = COPY $v8
+  ; MIR-NEXT:   $v0 = COPY [[COPY1]]
+  ; MIR-NEXT:   early-clobber %4:vrnov0 = PseudoVFNCVT_F_F_W_M1_MASK [[COPY3]], [[COPY2]], $v0, [[COPY]], 5 /* e32 */, 0
+  ; MIR-NEXT:   $v8 = COPY %4
+  ; MIR-NEXT:   PseudoRET implicit $v8
   %splat = insertelement <8 x i1> poison, i1 -1, i32 0
   %mask = shufflevector <8 x i1> %splat, <8 x i1> poison, <8 x i32> zeroinitializer
   %a = call <8 x float> @llvm.vp.fptrunc.nxv2f32.nxv2f64(<8 x double> %x, <8 x i1> %mask, i32 %vl)
   %b = call <8 x float> @llvm.vp.merge.nxv2f32(<8 x i1> %m, <8 x float> %a, <8 x float> %passthru, i32 %vl)
   ret <8 x float> %b
 }
 
 ; Test load operation by vp.load.
 declare <8 x i32> @llvm.vp.load.nxv2i32.p0nxv2i32(<8 x i32> *, <8 x i1>, i32)
 define <8 x i32> @vpmerge_vpload(<8 x i32> %passthru, <8 x i32> * %p, <8 x i1> %m, i32 zeroext %vl) {
 ; CHECK-LABEL: vpmerge_vpload:
 ; CHECK:       # %bb.0:
-; CHECK-NEXT:    vsetvli zero, a1, e32, m1, ta, mu
-; CHECK-NEXT:    vle32.v v9, (a0)
-; CHECK-NEXT:    vsetvli zero, zero, e32, m1, tu, mu
-; CHECK-NEXT:    vmerge.vvm v8, v8, v9, v0
+; CHECK-NEXT:    vsetvli zero, a1, e32, m1, tu, mu
+; CHECK-NEXT:    vle32.v v8, (a0), v0.t
 ; CHECK-NEXT:    ret
+  ; MIR-LABEL: name: vpmerge_vpload
+  ; MIR: bb.0 (%ir-block.0):
+  ; MIR-NEXT:   liveins: $v8, $x10, $v0, $x11
+  ; MIR-NEXT: {{  $}}
+  ; MIR-NEXT:   [[COPY:%[0-9]+]]:gprnox0 = COPY $x11
+  ; MIR-NEXT:   [[COPY1:%[0-9]+]]:vr = COPY $v0
+  ; MIR-NEXT:   [[COPY2:%[0-9]+]]:gpr = COPY $x10
+  ; MIR-NEXT:   [[COPY3:%[0-9]+]]:vrnov0 = COPY $v8
+  ; MIR-NEXT:   $v0 = COPY [[COPY1]]
+  ; MIR-NEXT:   [[PseudoVLE32_V_M1_MASK:%[0-9]+]]:vrnov0 = PseudoVLE32_V_M1_MASK [[COPY3]], [[COPY2]], $v0, [[COPY]], 5 /* e32 */, 0
+  ; MIR-NEXT:   $v8 = COPY [[PseudoVLE32_V_M1_MASK]]
+  ; MIR-NEXT:   PseudoRET implicit $v8
   %splat = insertelement <8 x i1> poison, i1 -1, i32 0
   %mask = shufflevector <8 x i1> %splat, <8 x i1> poison, <8 x i32> zeroinitializer
   %a = call <8 x i32> @llvm.vp.load.nxv2i32.p0nxv2i32(<8 x i32> * %p, <8 x i1> %mask, i32 %vl)
   %b = call <8 x i32> @llvm.vp.merge.nxv2i32(<8 x i1> %m, <8 x i32> %a, <8 x i32> %passthru, i32 %vl)
   ret <8 x i32> %b
 }
 
 ; Test result have chain and glued node.
 define <8 x i32> @vpmerge_vpload2(<8 x i32> %passthru, <8 x i32> * %p, <8 x i32> %x, <8 x i32> %y, i32 zeroext %vl) {
 ; CHECK-LABEL: vpmerge_vpload2:
 ; CHECK:       # %bb.0:
 ; CHECK-NEXT:    vsetvli zero, a1, e32, m1, ta, mu
-; CHECK-NEXT:    vle32.v v11, (a0)
 ; CHECK-NEXT:    vmseq.vv v0, v9, v10
 ; CHECK-NEXT:    vsetvli zero, zero, e32, m1, tu, mu
-; CHECK-NEXT:    vmerge.vvm v8, v8, v11, v0
+; CHECK-NEXT:    vle32.v v8, (a0), v0.t
 ; CHECK-NEXT:    ret
+  ; MIR-LABEL: name: vpmerge_vpload2
+  ; MIR: bb.0 (%ir-block.0):
+  ; MIR-NEXT:   liveins: $v8, $x10, $v9, $v10, $x11
+  ; MIR-NEXT: {{  $}}
+  ; MIR-NEXT:   [[COPY:%[0-9]+]]:gprnox0 = COPY $x11
+  ; MIR-NEXT:   [[COPY1:%[0-9]+]]:vr = COPY $v10
+  ; MIR-NEXT:   [[COPY2:%[0-9]+]]:vr = COPY $v9
+  ; MIR-NEXT:   [[COPY3:%[0-9]+]]:gpr = COPY $x10
+  ; MIR-NEXT:   [[COPY4:%[0-9]+]]:vrnov0 = COPY $v8
+  ; MIR-NEXT:   [[PseudoVMSEQ_VV_M1_:%[0-9]+]]:vr = PseudoVMSEQ_VV_M1 [[COPY2]], [[COPY1]], [[COPY]], 5 /* e32 */
+  ; MIR-NEXT:   $v0 = COPY [[PseudoVMSEQ_VV_M1_]]
+  ; MIR-NEXT:   [[PseudoVLE32_V_M1_MASK:%[0-9]+]]:vrnov0 = PseudoVLE32_V_M1_MASK [[COPY4]], [[COPY3]], $v0, [[COPY]], 5 /* e32 */, 0
+  ; MIR-NEXT:   $v8 = COPY [[PseudoVLE32_V_M1_MASK]]
+  ; MIR-NEXT:   PseudoRET implicit $v8
   %splat = insertelement <8 x i1> poison, i1 -1, i32 0
   %mask = shufflevector <8 x i1> %splat, <8 x i1> poison, <8 x i32> zeroinitializer
   %a = call <8 x i32> @llvm.vp.load.nxv2i32.p0nxv2i32(<8 x i32> * %p, <8 x i1> %mask, i32 %vl)
   %m = call <8 x i1> @llvm.vp.icmp.nxv2i32(<8 x i32> %x, <8 x i32> %y, metadata !"eq", <8 x i1> %mask, i32 %vl)
   %b = call <8 x i32> @llvm.vp.merge.nxv2i32(<8 x i1> %m, <8 x i32> %a, <8 x i32> %passthru, i32 %vl)
   ret <8 x i32> %b
 }

llvm/test/CodeGen/RISCV/rvv/rvv-peephole-vmerge-vops.ll

-; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py
 ; RUN: llc < %s -mtriple=riscv64 -mattr=+v | FileCheck %s
+; RUN: llc < %s -mtriple=riscv64 -mattr=+v -stop-after=finalize-isel | FileCheck %s --check-prefix=MIR
 
 declare <vscale x 2 x i16> @llvm.vp.merge.nxv2i16(<vscale x 2 x i1>, <vscale x 2 x i16>, <vscale x 2 x i16>, i32)
 declare <vscale x 2 x i32> @llvm.vp.merge.nxv2i32(<vscale x 2 x i1>, <vscale x 2 x i32>, <vscale x 2 x i32>, i32)
 declare <vscale x 2 x float> @llvm.vp.merge.nxv2f32(<vscale x 2 x i1>, <vscale x 2 x float>, <vscale x 2 x float>, i32)
 declare <vscale x 2 x double> @llvm.vp.merge.nxv2f64(<vscale x 2 x i1>, <vscale x 2 x double>, <vscale x 2 x double>, i32)
 
 ; Test binary operator with vp.merge and vp.smax.
 declare <vscale x 2 x i32> @llvm.vp.add.nxv2i32(<vscale x 2 x i32>, <vscale x 2 x i32>, <vscale x 2 x i1>, i32)
 define <vscale x 2 x i32> @vpmerge_vpadd(<vscale x 2 x i32> %passthru, <vscale x 2 x i32> %x, <vscale x 2 x i32> %y, <vscale x 2 x i1> %m, i32 zeroext %vl) {
 ; CHECK-LABEL: vpmerge_vpadd:
 ; CHECK:       # %bb.0:
-; CHECK-NEXT:    vsetvli zero, a0, e32, m1, ta, mu
-; CHECK-NEXT:    vadd.vv v9, v9, v10
-; CHECK-NEXT:    vsetvli zero, zero, e32, m1, tu, mu
-; CHECK-NEXT:    vmerge.vvm v8, v8, v9, v0
+; CHECK-NEXT:    vsetvli zero, a0, e32, m1, tu, mu
+; CHECK-NEXT:    vadd.vv v8, v9, v10, v0.t
 ; CHECK-NEXT:    ret
+  ; MIR-LABEL: name: vpmerge_vpadd
+  ; MIR: bb.0 (%ir-block.0):
+  ; MIR-NEXT:   liveins: $v8, $v9, $v10, $v0, $x10
+  ; MIR-NEXT: {{  $}}
+  ; MIR-NEXT:   [[COPY:%[0-9]+]]:gprnox0 = COPY $x10
+  ; MIR-NEXT:   [[COPY1:%[0-9]+]]:vr = COPY $v0
+  ; MIR-NEXT:   [[COPY2:%[0-9]+]]:vr = COPY $v10
+  ; MIR-NEXT:   [[COPY3:%[0-9]+]]:vr = COPY $v9
+  ; MIR-NEXT:   [[COPY4:%[0-9]+]]:vrnov0 = COPY $v8
+  ; MIR-NEXT:   $v0 = COPY [[COPY1]]
+  ; MIR-NEXT:   [[PseudoVADD_VV_M1_MASK:%[0-9]+]]:vrnov0 = PseudoVADD_VV_M1_MASK [[COPY4]], [[COPY3]], [[COPY2]], $v0, [[COPY]], 5 /* e32 */, 0
+  ; MIR-NEXT:   $v8 = COPY [[PseudoVADD_VV_M1_MASK]]
+  ; MIR-NEXT:   PseudoRET implicit $v8
   %splat = insertelement <vscale x 2 x i1> poison, i1 -1, i32 0
   %mask = shufflevector <vscale x 2 x i1> %splat, <vscale x 2 x i1> poison, <vscale x 2 x i32> zeroinitializer
   %a = call <vscale x 2 x i32> @llvm.vp.add.nxv2i32(<vscale x 2 x i32> %x, <vscale x 2 x i32> %y, <vscale x 2 x i1> %mask, i32 %vl)
   %b = call <vscale x 2 x i32> @llvm.vp.merge.nxv2i32(<vscale x 2 x i1> %m, <vscale x 2 x i32> %a, <vscale x 2 x i32> %passthru, i32 %vl)
   ret <vscale x 2 x i32> %b
 }
 
 ; Test glued node of merge should not be deleted.
 declare <vscale x 2 x i1> @llvm.vp.icmp.nxv2i32(<vscale x 2 x i32>, <vscale x 2 x i32>, metadata, <vscale x 2 x i1>, i32)
 define <vscale x 2 x i32> @vpmerge_vpadd2(<vscale x 2 x i32> %passthru, <vscale x 2 x i32> %x, <vscale x 2 x i32> %y, i32 zeroext %vl) {
 ; CHECK-LABEL: vpmerge_vpadd2:
 ; CHECK:       # %bb.0:
 ; CHECK-NEXT:    vsetvli zero, a0, e32, m1, ta, mu
-; CHECK-NEXT:    vadd.vv v11, v9, v10
 ; CHECK-NEXT:    vmseq.vv v0, v9, v10
 ; CHECK-NEXT:    vsetvli zero, zero, e32, m1, tu, mu
-; CHECK-NEXT:    vmerge.vvm v8, v8, v11, v0
+; CHECK-NEXT:    vadd.vv v8, v9, v10, v0.t
 ; CHECK-NEXT:    ret
+  ; MIR-LABEL: name: vpmerge_vpadd2
+  ; MIR: bb.0 (%ir-block.0):
+  ; MIR-NEXT:   liveins: $v8, $v9, $v10, $x10
+  ; MIR-NEXT: {{  $}}
+  ; MIR-NEXT:   [[COPY:%[0-9]+]]:gprnox0 = COPY $x10
+  ; MIR-NEXT:   [[COPY1:%[0-9]+]]:vr = COPY $v10
+  ; MIR-NEXT:   [[COPY2:%[0-9]+]]:vr = COPY $v9
+  ; MIR-NEXT:   [[COPY3:%[0-9]+]]:vrnov0 = COPY $v8
+  ; MIR-NEXT:   [[PseudoVMSEQ_VV_M1_:%[0-9]+]]:vr = PseudoVMSEQ_VV_M1 [[COPY2]], [[COPY1]], [[COPY]], 5 /* e32 */
+  ; MIR-NEXT:   $v0 = COPY [[PseudoVMSEQ_VV_M1_]]
+  ; MIR-NEXT:   [[PseudoVADD_VV_M1_MASK:%[0-9]+]]:vrnov0 = PseudoVADD_VV_M1_MASK [[COPY3]], [[COPY2]], [[COPY1]], $v0, [[COPY]], 5 /* e32 */, 0
+  ; MIR-NEXT:   $v8 = COPY [[PseudoVADD_VV_M1_MASK]]
+  ; MIR-NEXT:   PseudoRET implicit $v8
   %splat = insertelement <vscale x 2 x i1> poison, i1 -1, i32 0
   %mask = shufflevector <vscale x 2 x i1> %splat, <vscale x 2 x i1> poison, <vscale x 2 x i32> zeroinitializer
   %a = call <vscale x 2 x i32> @llvm.vp.add.nxv2i32(<vscale x 2 x i32> %x, <vscale x 2 x i32> %y, <vscale x 2 x i1> %mask, i32 %vl)
   %m = call <vscale x 2 x i1> @llvm.vp.icmp.nxv2i32(<vscale x 2 x i32> %x, <vscale x 2 x i32> %y, metadata !"eq", <vscale x 2 x i1> %mask, i32 %vl)
   %b = call <vscale x 2 x i32> @llvm.vp.merge.nxv2i32(<vscale x 2 x i1> %m, <vscale x 2 x i32> %a, <vscale x 2 x i32> %passthru, i32 %vl)
   ret <vscale x 2 x i32> %b
 }
 
-; Test vp.merge have all-ones mask.
+; Test vp.merge has all-ones mask.
 define <vscale x 2 x i32> @vpmerge_vpadd3(<vscale x 2 x i32> %passthru, <vscale x 2 x i32> %x, <vscale x 2 x i32> %y, i32 zeroext %vl) {
 ; CHECK-LABEL: vpmerge_vpadd3:
 ; CHECK:       # %bb.0:
-; CHECK-NEXT:    vsetvli a1, zero, e8, mf4, ta, mu
-; CHECK-NEXT:    vmset.m v0
-; CHECK-NEXT:    vsetvli zero, a0, e32, m1, ta, mu
-; CHECK-NEXT:    vadd.vv v9, v9, v10
-; CHECK-NEXT:    vsetvli zero, zero, e32, m1, tu, mu
-; CHECK-NEXT:    vmerge.vvm v8, v8, v9, v0
+; CHECK-NEXT:    vsetvli zero, a0, e32, m1, tu, mu
+; CHECK-NEXT:    vadd.vv v8, v9, v10
 ; CHECK-NEXT:    ret
+  ; MIR-LABEL: name: vpmerge_vpadd3
+  ; MIR: bb.0 (%ir-block.0):
+  ; MIR-NEXT:   liveins: $v8, $v9, $v10, $x10
+  ; MIR-NEXT: {{  $}}
+  ; MIR-NEXT:   [[COPY:%[0-9]+]]:gprnox0 = COPY $x10
+  ; MIR-NEXT:   [[COPY1:%[0-9]+]]:vr = COPY $v10
+  ; MIR-NEXT:   [[COPY2:%[0-9]+]]:vr = COPY $v9
+  ; MIR-NEXT:   [[COPY3:%[0-9]+]]:vr = COPY $v8
+  ; MIR-NEXT:   [[PseudoVADD_VV_M1_TU:%[0-9]+]]:vr = PseudoVADD_VV_M1_TU [[COPY3]], [[COPY2]], [[COPY1]], [[COPY]], 5 /* e32 */
+  ; MIR-NEXT:   $v8 = COPY [[PseudoVADD_VV_M1_TU]]
+  ; MIR-NEXT:   PseudoRET implicit $v8
   %splat = insertelement <vscale x 2 x i1> poison, i1 -1, i32 0
   %mask = shufflevector <vscale x 2 x i1> %splat, <vscale x 2 x i1> poison, <vscale x 2 x i32> zeroinitializer
   %a = call <vscale x 2 x i32> @llvm.vp.add.nxv2i32(<vscale x 2 x i32> %x, <vscale x 2 x i32> %y, <vscale x 2 x i1> %mask, i32 %vl)
   %b = call <vscale x 2 x i32> @llvm.vp.merge.nxv2i32(<vscale x 2 x i1> %mask, <vscale x 2 x i32> %a, <vscale x 2 x i32> %passthru, i32 %vl)
   ret <vscale x 2 x i32> %b
 }
 
 ; Test float binary operator with vp.merge and vp.fadd.
 declare <vscale x 2 x float> @llvm.vp.fadd.nxv2f32(<vscale x 2 x float>, <vscale x 2 x float>, <vscale x 2 x i1>, i32)
 define <vscale x 2 x float> @vpmerge_vpfadd(<vscale x 2 x float> %passthru, <vscale x 2 x float> %x, <vscale x 2 x float> %y, <vscale x 2 x i1> %m, i32 zeroext %vl) {
 ; CHECK-LABEL: vpmerge_vpfadd:
 ; CHECK:       # %bb.0:
-; CHECK-NEXT:    vsetvli zero, a0, e32, m1, ta, mu
-; CHECK-NEXT:    vfadd.vv v9, v9, v10
-; CHECK-NEXT:    vsetvli zero, zero, e32, m1, tu, mu
-; CHECK-NEXT:    vmerge.vvm v8, v8, v9, v0
+; CHECK-NEXT:    vsetvli zero, a0, e32, m1, tu, mu
+; CHECK-NEXT:    vfadd.vv v8, v9, v10, v0.t
 ; CHECK-NEXT:    ret
+  ; MIR-LABEL: name: vpmerge_vpfadd
+  ; MIR: bb.0 (%ir-block.0):
+  ; MIR-NEXT:   liveins: $v8, $v9, $v10, $v0, $x10
+  ; MIR-NEXT: {{  $}}
+  ; MIR-NEXT:   [[COPY:%[0-9]+]]:gprnox0 = COPY $x10
+  ; MIR-NEXT:   [[COPY1:%[0-9]+]]:vr = COPY $v0
+  ; MIR-NEXT:   [[COPY2:%[0-9]+]]:vr = COPY $v10
+  ; MIR-NEXT:   [[COPY3:%[0-9]+]]:vr = COPY $v9
+  ; MIR-NEXT:   [[COPY4:%[0-9]+]]:vrnov0 = COPY $v8
+  ; MIR-NEXT:   $v0 = COPY [[COPY1]]
+  ; MIR-NEXT:   [[PseudoVFADD_VV_M1_MASK:%[0-9]+]]:vrnov0 = PseudoVFADD_VV_M1_MASK [[COPY4]], [[COPY3]], [[COPY2]], $v0, [[COPY]], 5 /* e32 */, 0, implicit $frm
+  ; MIR-NEXT:   $v8 = COPY [[PseudoVFADD_VV_M1_MASK]]
+  ; MIR-NEXT:   PseudoRET implicit $v8
   %splat = insertelement <vscale x 2 x i1> poison, i1 -1, i32 0
   %mask = shufflevector <vscale x 2 x i1> %splat, <vscale x 2 x i1> poison, <vscale x 2 x i32> zeroinitializer
   %a = call <vscale x 2 x float> @llvm.vp.fadd.nxv2f32(<vscale x 2 x float> %x, <vscale x 2 x float> %y, <vscale x 2 x i1> %mask, i32 %vl)
   %b = call <vscale x 2 x float> @llvm.vp.merge.nxv2f32(<vscale x 2 x i1> %m, <vscale x 2 x float> %a, <vscale x 2 x float> %passthru, i32 %vl)
   ret <vscale x 2 x float> %b
 }
 
 ; Test for binary operator with specific EEW by riscv.vrgatherei16.
 declare <vscale x 2 x i32> @llvm.riscv.vrgatherei16.vv.nxv2i32.i64(<vscale x 2 x i32>, <vscale x 2 x i32>, <vscale x 2 x i16>, i64)
 define <vscale x 2 x i32> @vpmerge_vrgatherei16(<vscale x 2 x i32> %passthru, <vscale x 2 x i32> %x, <vscale x 2 x i16> %y, <vscale x 2 x i1> %m, i32 zeroext %vl) {
 ; CHECK-LABEL: vpmerge_vrgatherei16:
 ; CHECK:       # %bb.0:
 ; CHECK-NEXT:    vsetvli zero, a0, e32, m1, ta, mu
 ; CHECK-NEXT:    vrgatherei16.vv v8, v9, v10
 ; CHECK-NEXT:    ret
+  ; MIR-LABEL: name: vpmerge_vrgatherei16
+  ; MIR: bb.0 (%ir-block.0):
+  ; MIR-NEXT:   liveins: $v9, $v10, $x10
+  ; MIR-NEXT: {{  $}}
+  ; MIR-NEXT:   [[COPY:%[0-9]+]]:gprnox0 = COPY $x10
+  ; MIR-NEXT:   [[COPY1:%[0-9]+]]:vr = COPY $v10
+  ; MIR-NEXT:   [[COPY2:%[0-9]+]]:vr = COPY $v9
+  ; MIR-NEXT:   early-clobber %5:vr = PseudoVRGATHEREI16_VV_M1_MF2 [[COPY2]], [[COPY1]], [[COPY]], 5 /* e32 */
+  ; MIR-NEXT:   $v8 = COPY %5
+  ; MIR-NEXT:   PseudoRET implicit $v8
   %1 = zext i32 %vl to i64
   %2 = tail call <vscale x 2 x i32> @llvm.riscv.vrgatherei16.vv.nxv2i32.i64(<vscale x 2 x i32> undef, <vscale x 2 x i32> %x, <vscale x 2 x i16> %y, i64 %1)
   %3 = tail call <vscale x 2 x i32> @llvm.vp.merge.nxv2i32(<vscale x 2 x i1> %m, <vscale x 2 x i32> %2, <vscale x 2 x i32> %passthru, i32 %vl)
   ret <vscale x 2 x i32> %2
 }
 
 ; Test conversion by fptosi.
 declare <vscale x 2 x i16> @llvm.vp.fptosi.nxv2i16.nxv2f32(<vscale x 2 x float>, <vscale x 2 x i1>, i32)
 define <vscale x 2 x i16> @vpmerge_vpfptosi(<vscale x 2 x i16> %passthru, <vscale x 2 x float> %x, <vscale x 2 x i1> %m, i32 zeroext %vl) {
 ; CHECK-LABEL: vpmerge_vpfptosi:
 ; CHECK:       # %bb.0:
-; CHECK-NEXT:    vsetvli zero, a0, e16, mf2, ta, mu
-; CHECK-NEXT:    vfncvt.rtz.x.f.w v10, v9
-; CHECK-NEXT:    vsetvli zero, zero, e16, mf2, tu, mu
-; CHECK-NEXT:    vmerge.vvm v8, v8, v10, v0
+; CHECK-NEXT:    vsetvli zero, a0, e16, mf2, tu, mu
+; CHECK-NEXT:    vfncvt.rtz.x.f.w v8, v9, v0.t
 ; CHECK-NEXT:    ret
+  ; MIR-LABEL: name: vpmerge_vpfptosi
+  ; MIR: bb.0 (%ir-block.0):
+  ; MIR-NEXT:   liveins: $v8, $v9, $v0, $x10
+  ; MIR-NEXT: {{  $}}
+  ; MIR-NEXT:   [[COPY:%[0-9]+]]:gprnox0 = COPY $x10
+  ; MIR-NEXT:   [[COPY1:%[0-9]+]]:vr = COPY $v0
+  ; MIR-NEXT:   [[COPY2:%[0-9]+]]:vr = COPY $v9
+  ; MIR-NEXT:   [[COPY3:%[0-9]+]]:vrnov0 = COPY $v8
+  ; MIR-NEXT:   $v0 = COPY [[COPY1]]
+  ; MIR-NEXT:   early-clobber %4:vrnov0 = PseudoVFNCVT_RTZ_X_F_W_MF2_MASK [[COPY3]], [[COPY2]], $v0, [[COPY]], 4 /* e16 */, 0
+  ; MIR-NEXT:   $v8 = COPY %4
+  ; MIR-NEXT:   PseudoRET implicit $v8
   %splat = insertelement <vscale x 2 x i1> poison, i1 -1, i32 0
   %mask = shufflevector <vscale x 2 x i1> %splat, <vscale x 2 x i1> poison, <vscale x 2 x i32> zeroinitializer
   %a = call <vscale x 2 x i16> @llvm.vp.fptosi.nxv2i16.nxv2f32(<vscale x 2 x float> %x, <vscale x 2 x i1> %mask, i32 %vl)
   %b = call <vscale x 2 x i16> @llvm.vp.merge.nxv2i16(<vscale x 2 x i1> %m, <vscale x 2 x i16> %a, <vscale x 2 x i16> %passthru, i32 %vl)
   ret <vscale x 2 x i16> %b
 }
 
 ; Test conversion by sitofp.
 declare <vscale x 2 x float> @llvm.vp.sitofp.nxv2f32.nxv2i64(<vscale x 2 x i64>, <vscale x 2 x i1>, i32)
 define <vscale x 2 x float> @vpmerge_vpsitofp(<vscale x 2 x float> %passthru, <vscale x 2 x i64> %x, <vscale x 2 x i1> %m, i32 zeroext %vl) {
 ; CHECK-LABEL: vpmerge_vpsitofp:
 ; CHECK:       # %bb.0:
-; CHECK-NEXT:    vsetvli zero, a0, e32, m1, ta, mu
-; CHECK-NEXT:    vfncvt.f.x.w v9, v10
-; CHECK-NEXT:    vsetvli zero, zero, e32, m1, tu, mu
-; CHECK-NEXT:    vmerge.vvm v8, v8, v9, v0
+; CHECK-NEXT:    vsetvli zero, a0, e32, m1, tu, mu
+; CHECK-NEXT:    vfncvt.f.x.w v8, v10, v0.t
 ; CHECK-NEXT:    ret
+  ; MIR-LABEL: name: vpmerge_vpsitofp
+  ; MIR: bb.0 (%ir-block.0):
+  ; MIR-NEXT:   liveins: $v8, $v10m2, $v0, $x10
+  ; MIR-NEXT: {{  $}}
+  ; MIR-NEXT:   [[COPY:%[0-9]+]]:gprnox0 = COPY $x10
+  ; MIR-NEXT:   [[COPY1:%[0-9]+]]:vr = COPY $v0
+  ; MIR-NEXT:   [[COPY2:%[0-9]+]]:vrm2 = COPY $v10m2
+  ; MIR-NEXT:   [[COPY3:%[0-9]+]]:vrnov0 = COPY $v8
+  ; MIR-NEXT:   $v0 = COPY [[COPY1]]
+  ; MIR-NEXT:   early-clobber %4:vrnov0 = PseudoVFNCVT_F_X_W_M1_MASK [[COPY3]], [[COPY2]], $v0, [[COPY]], 5 /* e32 */, 0
+  ; MIR-NEXT:   $v8 = COPY %4
+  ; MIR-NEXT:   PseudoRET implicit $v8
   %splat = insertelement <vscale x 2 x i1> poison, i1 -1, i32 0
   %mask = shufflevector <vscale x 2 x i1> %splat, <vscale x 2 x i1> poison, <vscale x 2 x i32> zeroinitializer
   %a = call <vscale x 2 x float> @llvm.vp.sitofp.nxv2f32.nxv2i64(<vscale x 2 x i64> %x, <vscale x 2 x i1> %mask, i32 %vl)
   %b = call <vscale x 2 x float> @llvm.vp.merge.nxv2f32(<vscale x 2 x i1> %m, <vscale x 2 x float> %a, <vscale x 2 x float> %passthru, i32 %vl)
   ret <vscale x 2 x float> %b
 }
 
 ; Test integer extension by vp.zext.
 declare <vscale x 2 x i32> @llvm.vp.zext.nxv2i32.nxv2i8(<vscale x 2 x i8>, <vscale x 2 x i1>, i32)
 define <vscale x 2 x i32> @vpmerge_vpzext(<vscale x 2 x i32> %passthru, <vscale x 2 x i8> %x, <vscale x 2 x i1> %m, i32 zeroext %vl) {
 ; CHECK-LABEL: vpmerge_vpzext:
 ; CHECK:       # %bb.0:
-; CHECK-NEXT:    vsetvli zero, a0, e32, m1, ta, mu
-; CHECK-NEXT:    vzext.vf4 v10, v9
-; CHECK-NEXT:    vsetvli zero, zero, e32, m1, tu, mu
-; CHECK-NEXT:    vmerge.vvm v8, v8, v10, v0
+; CHECK-NEXT:    vsetvli zero, a0, e32, m1, tu, mu
+; CHECK-NEXT:    vzext.vf4 v8, v9, v0.t
 ; CHECK-NEXT:    ret
+  ; MIR-LABEL: name: vpmerge_vpzext
+  ; MIR: bb.0 (%ir-block.0):
+  ; MIR-NEXT:   liveins: $v8, $v9, $v0, $x10
+  ; MIR-NEXT: {{  $}}
+  ; MIR-NEXT:   [[COPY:%[0-9]+]]:gprnox0 = COPY $x10
+  ; MIR-NEXT:   [[COPY1:%[0-9]+]]:vr = COPY $v0
+  ; MIR-NEXT:   [[COPY2:%[0-9]+]]:vr = COPY $v9
+  ; MIR-NEXT:   [[COPY3:%[0-9]+]]:vrnov0 = COPY $v8
+  ; MIR-NEXT:   $v0 = COPY [[COPY1]]
+  ; MIR-NEXT:   early-clobber %4:vrnov0 = PseudoVZEXT_VF4_M1_MASK [[COPY3]], [[COPY2]], $v0, [[COPY]], 5 /* e32 */, 0
+  ; MIR-NEXT:   $v8 = COPY %4
+  ; MIR-NEXT:   PseudoRET implicit $v8
   %splat = insertelement <vscale x 2 x i1> poison, i1 -1, i32 0
   %mask = shufflevector <vscale x 2 x i1> %splat, <vscale x 2 x i1> poison, <vscale x 2 x i32> zeroinitializer
   %a = call <vscale x 2 x i32> @llvm.vp.zext.nxv2i32.nxv2i8(<vscale x 2 x i8> %x, <vscale x 2 x i1> %mask, i32 %vl)
   %b = call <vscale x 2 x i32> @llvm.vp.merge.nxv2i32(<vscale x 2 x i1> %m, <vscale x 2 x i32> %a, <vscale x 2 x i32> %passthru, i32 %vl)
   ret <vscale x 2 x i32> %b
 }
 
 ; Test integer truncation by vp.trunc.
 declare <vscale x 2 x i32> @llvm.vp.trunc.nxv2i32.nxv2i64(<vscale x 2 x i64>, <vscale x 2 x i1>, i32)
 define <vscale x 2 x i32> @vpmerge_vptrunc(<vscale x 2 x i32> %passthru, <vscale x 2 x i64> %x, <vscale x 2 x i1> %m, i32 zeroext %vl) {
 ; CHECK-LABEL: vpmerge_vptrunc:
 ; CHECK:       # %bb.0:
-; CHECK-NEXT:    vsetvli zero, a0, e32, m1, ta, mu
-; CHECK-NEXT:    vncvt.x.x.w v9, v10
-; CHECK-NEXT:    vsetvli zero, zero, e32, m1, tu, mu
-; CHECK-NEXT:    vmerge.vvm v8, v8, v9, v0
+; CHECK-NEXT:    vsetvli zero, a0, e32, m1, tu, mu
+; CHECK-NEXT:    vncvt.x.x.w v8, v10, v0.t
 ; CHECK-NEXT:    ret
+  ; MIR-LABEL: name: vpmerge_vptrunc
+  ; MIR: bb.0 (%ir-block.0):
+  ; MIR-NEXT:   liveins: $v8, $v10m2, $v0, $x10
+  ; MIR-NEXT: {{  $}}
+  ; MIR-NEXT:   [[COPY:%[0-9]+]]:gprnox0 = COPY $x10
+  ; MIR-NEXT:   [[COPY1:%[0-9]+]]:vr = COPY $v0
+  ; MIR-NEXT:   [[COPY2:%[0-9]+]]:vrm2 = COPY $v10m2
+  ; MIR-NEXT:   [[COPY3:%[0-9]+]]:vrnov0 = COPY $v8
+  ; MIR-NEXT:   $v0 = COPY [[COPY1]]
+  ; MIR-NEXT:   early-clobber %4:vrnov0 = PseudoVNSRL_WX_M1_MASK [[COPY3]], [[COPY2]], $x0, $v0, [[COPY]], 5 /* e32 */, 0
+  ; MIR-NEXT:   $v8 = COPY %4
+  ; MIR-NEXT:   PseudoRET implicit $v8
   %splat = insertelement <vscale x 2 x i1> poison, i1 -1, i32 0
   %mask = shufflevector <vscale x 2 x i1> %splat, <vscale x 2 x i1> poison, <vscale x 2 x i32> zeroinitializer
   %a = call <vscale x 2 x i32> @llvm.vp.trunc.nxv2i32.nxv2i64(<vscale x 2 x i64> %x, <vscale x 2 x i1> %mask, i32 %vl)
   %b = call <vscale x 2 x i32> @llvm.vp.merge.nxv2i32(<vscale x 2 x i1> %m, <vscale x 2 x i32> %a, <vscale x 2 x i32> %passthru, i32 %vl)
   ret <vscale x 2 x i32> %b
 }
 
 ; Test integer extension by vp.fpext.
 declare <vscale x 2 x double> @llvm.vp.fpext.nxv2f64.nxv2f32(<vscale x 2 x float>, <vscale x 2 x i1>, i32)
 define <vscale x 2 x double> @vpmerge_vpfpext(<vscale x 2 x double> %passthru, <vscale x 2 x float> %x, <vscale x 2 x i1> %m, i32 zeroext %vl) {
 ; CHECK-LABEL: vpmerge_vpfpext:
 ; CHECK:       # %bb.0:
-; CHECK-NEXT:    vsetvli zero, a0, e32, m1, ta, mu
-; CHECK-NEXT:    vfwcvt.f.f.v v12, v10
-; CHECK-NEXT:    vsetvli zero, zero, e64, m2, tu, mu
-; CHECK-NEXT:    vmerge.vvm v8, v8, v12, v0
+; CHECK-NEXT:    vsetvli zero, a0, e32, m1, tu, mu
+; CHECK-NEXT:    vfwcvt.f.f.v v8, v10, v0.t
 ; CHECK-NEXT:    ret
+  ; MIR-LABEL: name: vpmerge_vpfpext
+  ; MIR: bb.0 (%ir-block.0):
+  ; MIR-NEXT:   liveins: $v8m2, $v10, $v0, $x10
+  ; MIR-NEXT: {{  $}}
+  ; MIR-NEXT:   [[COPY:%[0-9]+]]:gprnox0 = COPY $x10
+  ; MIR-NEXT:   [[COPY1:%[0-9]+]]:vr = COPY $v0
+  ; MIR-NEXT:   [[COPY2:%[0-9]+]]:vr = COPY $v10
+  ; MIR-NEXT:   [[COPY3:%[0-9]+]]:vrm2nov0 = COPY $v8m2
+  ; MIR-NEXT:   $v0 = COPY [[COPY1]]
+  ; MIR-NEXT:   early-clobber %4:vrm2nov0 = PseudoVFWCVT_F_F_V_M1_MASK [[COPY3]], [[COPY2]], $v0, [[COPY]], 5 /* e32 */, 0
+  ; MIR-NEXT:   $v8m2 = COPY %4
+  ; MIR-NEXT:   PseudoRET implicit $v8m2
   %splat = insertelement <vscale x 2 x i1> poison, i1 -1, i32 0
   %mask = shufflevector <vscale x 2 x i1> %splat, <vscale x 2 x i1> poison, <vscale x 2 x i32> zeroinitializer
   %a = call <vscale x 2 x double> @llvm.vp.fpext.nxv2f64.nxv2f32(<vscale x 2 x float> %x, <vscale x 2 x i1> %mask, i32 %vl)
   %b = call <vscale x 2 x double> @llvm.vp.merge.nxv2f64(<vscale x 2 x i1> %m, <vscale x 2 x double> %a, <vscale x 2 x double> %passthru, i32 %vl)
   ret <vscale x 2 x double> %b
 }
 
 ; Test integer truncation by vp.trunc.
 declare <vscale x 2 x float> @llvm.vp.fptrunc.nxv2f32.nxv2f64(<vscale x 2 x double>, <vscale x 2 x i1>, i32)
 define <vscale x 2 x float> @vpmerge_vpfptrunc(<vscale x 2 x float> %passthru, <vscale x 2 x double> %x, <vscale x 2 x i1> %m, i32 zeroext %vl) {
 ; CHECK-LABEL: vpmerge_vpfptrunc:
 ; CHECK:       # %bb.0:
-; CHECK-NEXT:    vsetvli zero, a0, e32, m1, ta, mu
-; CHECK-NEXT:    vfncvt.f.f.w v9, v10
-; CHECK-NEXT:    vsetvli zero, zero, e32, m1, tu, mu
-; CHECK-NEXT:    vmerge.vvm v8, v8, v9, v0
+; CHECK-NEXT:    vsetvli zero, a0, e32, m1, tu, mu
+; CHECK-NEXT:    vfncvt.f.f.w v8, v10, v0.t
 ; CHECK-NEXT:    ret
+  ; MIR-LABEL: name: vpmerge_vpfptrunc
+  ; MIR: bb.0 (%ir-block.0):
+  ; MIR-NEXT:   liveins: $v8, $v10m2, $v0, $x10
+  ; MIR-NEXT: {{  $}}
+  ; MIR-NEXT:   [[COPY:%[0-9]+]]:gprnox0 = COPY $x10
+  ; MIR-NEXT:   [[COPY1:%[0-9]+]]:vr = COPY $v0
+  ; MIR-NEXT:   [[COPY2:%[0-9]+]]:vrm2 = COPY $v10m2
+  ; MIR-NEXT:   [[COPY3:%[0-9]+]]:vrnov0 = COPY $v8
+  ; MIR-NEXT:   $v0 = COPY [[COPY1]]
+  ; MIR-NEXT:   early-clobber %4:vrnov0 = PseudoVFNCVT_F_F_W_M1_MASK [[COPY3]], [[COPY2]], $v0, [[COPY]], 5 /* e32 */, 0
+  ; MIR-NEXT:   $v8 = COPY %4
+  ; MIR-NEXT:   PseudoRET implicit $v8
   %splat = insertelement <vscale x 2 x i1> poison, i1 -1, i32 0
   %mask = shufflevector <vscale x 2 x i1> %splat, <vscale x 2 x i1> poison, <vscale x 2 x i32> zeroinitializer
   %a = call <vscale x 2 x float> @llvm.vp.fptrunc.nxv2f32.nxv2f64(<vscale x 2 x double> %x, <vscale x 2 x i1> %mask, i32 %vl)
   %b = call <vscale x 2 x float> @llvm.vp.merge.nxv2f32(<vscale x 2 x i1> %m, <vscale x 2 x float> %a, <vscale x 2 x float> %passthru, i32 %vl)
   ret <vscale x 2 x float> %b
 }
 
 ; Test load operation by vp.load.
 declare <vscale x 2 x i32> @llvm.vp.load.nxv2i32.p0nxv2i32(<vscale x 2 x i32> *, <vscale x 2 x i1>, i32)
 define <vscale x 2 x i32> @vpmerge_vpload(<vscale x 2 x i32> %passthru, <vscale x 2 x i32> * %p, <vscale x 2 x i1> %m, i32 zeroext %vl) {
 ; CHECK-LABEL: vpmerge_vpload:
 ; CHECK:       # %bb.0:
-; CHECK-NEXT:    vsetvli zero, a1, e32, m1, ta, mu
-; CHECK-NEXT:    vle32.v v9, (a0)
-; CHECK-NEXT:    vsetvli zero, zero, e32, m1, tu, mu
-; CHECK-NEXT:    vmerge.vvm v8, v8, v9, v0
+; CHECK-NEXT:    vsetvli zero, a1, e32, m1, tu, mu
+; CHECK-NEXT:    vle32.v v8, (a0), v0.t
 ; CHECK-NEXT:    ret
+  ; MIR-LABEL: name: vpmerge_vpload
+  ; MIR: bb.0 (%ir-block.0):
+  ; MIR-NEXT:   liveins: $v8, $x10, $v0, $x11
+  ; MIR-NEXT: {{  $}}
+  ; MIR-NEXT:   [[COPY:%[0-9]+]]:gprnox0 = COPY $x11
+  ; MIR-NEXT:   [[COPY1:%[0-9]+]]:vr = COPY $v0
+  ; MIR-NEXT:   [[COPY2:%[0-9]+]]:gpr = COPY $x10
+  ; MIR-NEXT:   [[COPY3:%[0-9]+]]:vrnov0 = COPY $v8
+  ; MIR-NEXT:   $v0 = COPY [[COPY1]]
+  ; MIR-NEXT:   [[PseudoVLE32_V_M1_MASK:%[0-9]+]]:vrnov0 = PseudoVLE32_V_M1_MASK [[COPY3]], [[COPY2]], $v0, [[COPY]], 5 /* e32 */, 0
+  ; MIR-NEXT:   $v8 = COPY [[PseudoVLE32_V_M1_MASK]]
+  ; MIR-NEXT:   PseudoRET implicit $v8
   %splat = insertelement <vscale x 2 x i1> poison, i1 -1, i32 0
   %mask = shufflevector <vscale x 2 x i1> %splat, <vscale x 2 x i1> poison, <vscale x 2 x i32> zeroinitializer
   %a = call <vscale x 2 x i32> @llvm.vp.load.nxv2i32.p0nxv2i32(<vscale x 2 x i32> * %p, <vscale x 2 x i1> %mask, i32 %vl)
   %b = call <vscale x 2 x i32> @llvm.vp.merge.nxv2i32(<vscale x 2 x i1> %m, <vscale x 2 x i32> %a, <vscale x 2 x i32> %passthru, i32 %vl)
   ret <vscale x 2 x i32> %b
 }
 
-; Test result have chain and glued node.
+; Test result has chain and glued node.
 define <vscale x 2 x i32> @vpmerge_vpload2(<vscale x 2 x i32> %passthru, <vscale x 2 x i32> * %p, <vscale x 2 x i32> %x, <vscale x 2 x i32> %y, i32 zeroext %vl) {
 ; CHECK-LABEL: vpmerge_vpload2:
 ; CHECK:       # %bb.0:
 ; CHECK-NEXT:    vsetvli zero, a1, e32, m1, ta, mu
-; CHECK-NEXT:    vle32.v v11, (a0)
 ; CHECK-NEXT:    vmseq.vv v0, v9, v10
 ; CHECK-NEXT:    vsetvli zero, zero, e32, m1, tu, mu
-; CHECK-NEXT:    vmerge.vvm v8, v8, v11, v0
+; CHECK-NEXT:    vle32.v v8, (a0), v0.t
 ; CHECK-NEXT:    ret
+  ; MIR-LABEL: name: vpmerge_vpload2
+  ; MIR: bb.0 (%ir-block.0):
+  ; MIR-NEXT:   liveins: $v8, $x10, $v9, $v10, $x11
+  ; MIR-NEXT: {{  $}}
+  ; MIR-NEXT:   [[COPY:%[0-9]+]]:gprnox0 = COPY $x11
+  ; MIR-NEXT:   [[COPY1:%[0-9]+]]:vr = COPY $v10
+  ; MIR-NEXT:   [[COPY2:%[0-9]+]]:vr = COPY $v9
+  ; MIR-NEXT:   [[COPY3:%[0-9]+]]:gpr = COPY $x10
+  ; MIR-NEXT:   [[COPY4:%[0-9]+]]:vrnov0 = COPY $v8
+  ; MIR-NEXT:   [[PseudoVMSEQ_VV_M1_:%[0-9]+]]:vr = PseudoVMSEQ_VV_M1 [[COPY2]], [[COPY1]], [[COPY]], 5 /* e32 */
+  ; MIR-NEXT:   $v0 = COPY [[PseudoVMSEQ_VV_M1_]]
+  ; MIR-NEXT:   [[PseudoVLE32_V_M1_MASK:%[0-9]+]]:vrnov0 = PseudoVLE32_V_M1_MASK [[COPY4]], [[COPY3]], $v0, [[COPY]], 5 /* e32 */, 0
+  ; MIR-NEXT:   $v8 = COPY [[PseudoVLE32_V_M1_MASK]]
+  ; MIR-NEXT:   PseudoRET implicit $v8
   %splat = insertelement <vscale x 2 x i1> poison, i1 -1, i32 0
   %mask = shufflevector <vscale x 2 x i1> %splat, <vscale x 2 x i1> poison, <vscale x 2 x i32> zeroinitializer
   %a = call <vscale x 2 x i32> @llvm.vp.load.nxv2i32.p0nxv2i32(<vscale x 2 x i32> * %p, <vscale x 2 x i1> %mask, i32 %vl)
   %m = call <vscale x 2 x i1> @llvm.vp.icmp.nxv2i32(<vscale x 2 x i32> %x, <vscale x 2 x i32> %y, metadata !"eq", <vscale x 2 x i1> %mask, i32 %vl)
   %b = call <vscale x 2 x i32> @llvm.vp.merge.nxv2i32(<vscale x 2 x i1> %m, <vscale x 2 x i32> %a, <vscale x 2 x i32> %passthru, i32 %vl)
   ret <vscale x 2 x i32> %b
 }
 
+; Test result has chain output of true operand of merge.vvm.
+define void @vpmerge_vpload_store(<vscale x 2 x i32> %passthru, <vscale x 2 x i32> * %p, <vscale x 2 x i1> %m, i32 zeroext %vl) {
+; CHECK-LABEL: vpmerge_vpload_store:
+; CHECK:       # %bb.0:
+; CHECK-NEXT:    vsetvli zero, a1, e32, m1, tu, mu
+; CHECK-NEXT:    vle32.v v8, (a0), v0.t
+; CHECK-NEXT:    vs1r.v v8, (a0)
+; CHECK-NEXT:    ret
+  ; MIR-LABEL: name: vpmerge_vpload_store
+  ; MIR: bb.0 (%ir-block.0):
+  ; MIR-NEXT:   liveins: $v8, $x10, $v0, $x11
+  ; MIR-NEXT: {{  $}}
+  ; MIR-NEXT:   [[COPY:%[0-9]+]]:gprnox0 = COPY $x11
+  ; MIR-NEXT:   [[COPY1:%[0-9]+]]:vr = COPY $v0
+  ; MIR-NEXT:   [[COPY2:%[0-9]+]]:gpr = COPY $x10
+  ; MIR-NEXT:   [[COPY3:%[0-9]+]]:vrnov0 = COPY $v8
+  ; MIR-NEXT:   $v0 = COPY [[COPY1]]
+  ; MIR-NEXT:   [[PseudoVLE32_V_M1_MASK:%[0-9]+]]:vrnov0 = PseudoVLE32_V_M1_MASK [[COPY3]], [[COPY2]], $v0, [[COPY]], 5 /* e32 */, 0
+  ; MIR-NEXT:   VS1R_V killed [[PseudoVLE32_V_M1_MASK]], [[COPY2]] :: (store unknown-size into %ir.p, align 8)
+  ; MIR-NEXT:   PseudoRET
+  %splat = insertelement <vscale x 2 x i1> poison, i1 -1, i32 0
+  %mask = shufflevector <vscale x 2 x i1> %splat, <vscale x 2 x i1> poison, <vscale x 2 x i32> zeroinitializer
+  %a = call <vscale x 2 x i32> @llvm.vp.load.nxv2i32.p0nxv2i32(<vscale x 2 x i32> * %p, <vscale x 2 x i1> %mask, i32 %vl)
+  %b = call <vscale x 2 x i32> @llvm.vp.merge.nxv2i32(<vscale x 2 x i1> %m, <vscale x 2 x i32> %a, <vscale x 2 x i32> %passthru, i32 %vl)
+  store <vscale x 2 x i32> %b, <vscale x 2 x i32> * %p
+  ret void
+}
+
 ; FIXME: Merge vmerge.vvm and vleffN.v
 declare { <vscale x 2 x i32>, i64 } @llvm.riscv.vleff.nxv2i32(<vscale x 2 x i32>, <vscale x 2 x i32>*, i64)
 define <vscale x 2 x i32> @vpmerge_vleff(<vscale x 2 x i32> %passthru, <vscale x 2 x i32> * %p, <vscale x 2 x i1> %m, i32 zeroext %vl) {
 ; CHECK-LABEL: vpmerge_vleff:
 ; CHECK:       # %bb.0:
 ; CHECK-NEXT:    vsetvli zero, a1, e32, m1, ta, mu
 ; CHECK-NEXT:    vle32ff.v v9, (a0)
 ; CHECK-NEXT:    vsetvli zero, a1, e32, m1, tu, mu
 ; CHECK-NEXT:    vmerge.vvm v8, v8, v9, v0
 ; CHECK-NEXT:    ret
+  ; MIR-LABEL: name: vpmerge_vleff
+  ; MIR: bb.0 (%ir-block.0):
+  ; MIR-NEXT:   liveins: $v8, $x10, $v0, $x11
+  ; MIR-NEXT: {{  $}}
+  ; MIR-NEXT:   [[COPY:%[0-9]+]]:gprnox0 = COPY $x11
+  ; MIR-NEXT:   [[COPY1:%[0-9]+]]:vr = COPY $v0
+  ; MIR-NEXT:   [[COPY2:%[0-9]+]]:gpr = COPY $x10
+  ; MIR-NEXT:   [[COPY3:%[0-9]+]]:vrnov0 = COPY $v8
+  ; MIR-NEXT:   [[PseudoVLE32FF_V_M1_:%[0-9]+]]:vr, [[PseudoVLE32FF_V_M1_1:%[0-9]+]]:gpr = PseudoVLE32FF_V_M1 [[COPY2]], [[COPY]], 5 /* e32 */, implicit-def dead $vl
+  ; MIR-NEXT:   $v0 = COPY [[COPY1]]
+  ; MIR-NEXT:   [[PseudoVMERGE_VVM_M1_TU:%[0-9]+]]:vrnov0 = PseudoVMERGE_VVM_M1_TU [[COPY3]], [[COPY3]], killed [[PseudoVLE32FF_V_M1_]], $v0, [[COPY]], 5 /* e32 */
+  ; MIR-NEXT:   $v8 = COPY [[PseudoVMERGE_VVM_M1_TU]]
+  ; MIR-NEXT:   PseudoRET implicit $v8
   %1 = zext i32 %vl to i64
   %a = call { <vscale x 2 x i32>, i64 } @llvm.riscv.vleff.nxv2i32(<vscale x 2 x i32> undef, <vscale x 2 x i32>* %p, i64 %1)
   %b = extractvalue { <vscale x 2 x i32>, i64 } %a, 0
   %c = call <vscale x 2 x i32> @llvm.vp.merge.nxv2i32(<vscale x 2 x i1> %m, <vscale x 2 x i32> %b, <vscale x 2 x i32> %passthru, i32 %vl)
   ret <vscale x 2 x i32> %c
 }
 
 ; Test strided load by riscv.vlse
 declare <vscale x 2 x i32> @llvm.riscv.vlse.nxv2i32(<vscale x 2 x i32>, <vscale x 2 x i32>*, i64, i64)
 define <vscale x 2 x i32> @vpmerge_vlse(<vscale x 2 x i32> %passthru,  <vscale x 2 x i32> * %p, <vscale x 2 x i1> %m, i64 %s, i32 zeroext %vl) {
 ; CHECK-LABEL: vpmerge_vlse:
 ; CHECK:       # %bb.0:
-; CHECK-NEXT:    vsetvli zero, a2, e32, m1, ta, mu
-; CHECK-NEXT:    vlse32.v v9, (a0), a1
-; CHECK-NEXT:    vsetvli zero, zero, e32, m1, tu, mu
-; CHECK-NEXT:    vmerge.vvm v8, v8, v9, v0
+; CHECK-NEXT:    vsetvli zero, a2, e32, m1, tu, mu
+; CHECK-NEXT:    vlse32.v v8, (a0), a1, v0.t
 ; CHECK-NEXT:    ret
+  ; MIR-LABEL: name: vpmerge_vlse
+  ; MIR: bb.0 (%ir-block.0):
+  ; MIR-NEXT:   liveins: $v8, $x10, $v0, $x11, $x12
+  ; MIR-NEXT: {{  $}}
+  ; MIR-NEXT:   [[COPY:%[0-9]+]]:gprnox0 = COPY $x12
+  ; MIR-NEXT:   [[COPY1:%[0-9]+]]:gpr = COPY $x11
+  ; MIR-NEXT:   [[COPY2:%[0-9]+]]:vr = COPY $v0
+  ; MIR-NEXT:   [[COPY3:%[0-9]+]]:gpr = COPY $x10
+  ; MIR-NEXT:   [[COPY4:%[0-9]+]]:vrnov0 = COPY $v8
+  ; MIR-NEXT:   $v0 = COPY [[COPY2]]
+  ; MIR-NEXT:   [[PseudoVLSE32_V_M1_MASK:%[0-9]+]]:vrnov0 = PseudoVLSE32_V_M1_MASK [[COPY4]], [[COPY3]], [[COPY1]], $v0, [[COPY]], 5 /* e32 */, 0
+  ; MIR-NEXT:   $v8 = COPY [[PseudoVLSE32_V_M1_MASK]]
+  ; MIR-NEXT:   PseudoRET implicit $v8
   %1 = zext i32 %vl to i64
   %a = call <vscale x 2 x i32> @llvm.riscv.vlse.nxv2i32(<vscale x 2 x i32> undef, <vscale x 2 x i32>* %p, i64 %s, i64 %1)
   %b = call <vscale x 2 x i32> @llvm.vp.merge.nxv2i32(<vscale x 2 x i1> %m, <vscale x 2 x i32> %a, <vscale x 2 x i32> %passthru, i32 %vl)
   ret <vscale x 2 x i32> %b
 }
 
 ; Test indexed load by riscv.vluxei
 declare <vscale x 2 x i32> @llvm.riscv.vluxei.nxv2i32.nxv2i64(<vscale x 2 x i32>, <vscale x 2 x i32>*, <vscale x 2 x i64>, i64)
 define <vscale x 2 x i32> @vpmerge_vluxei(<vscale x 2 x i32> %passthru,  <vscale x 2 x i32> * %p, <vscale x 2 x i64> %idx, <vscale x 2 x i1> %m, i64 %s, i32 zeroext %vl) {
 ; CHECK-LABEL: vpmerge_vluxei:
 ; CHECK:       # %bb.0:
-; CHECK-NEXT:    vsetvli zero, a2, e32, m1, ta, mu
-; CHECK-NEXT:    vluxei64.v v9, (a0), v10
-; CHECK-NEXT:    vsetvli zero, zero, e32, m1, tu, mu
-; CHECK-NEXT:    vmerge.vvm v8, v8, v9, v0
+; CHECK-NEXT:    vsetvli zero, a2, e32, m1, tu, mu
+; CHECK-NEXT:    vluxei64.v v8, (a0), v10, v0.t
 ; CHECK-NEXT:    ret
+  ; MIR-LABEL: name: vpmerge_vluxei
+  ; MIR: bb.0 (%ir-block.0):
+  ; MIR-NEXT:   liveins: $v8, $x10, $v10m2, $v0, $x12
+  ; MIR-NEXT: {{  $}}
+  ; MIR-NEXT:   [[COPY:%[0-9]+]]:gprnox0 = COPY $x12
+  ; MIR-NEXT:   [[COPY1:%[0-9]+]]:vr = COPY $v0
+  ; MIR-NEXT:   [[COPY2:%[0-9]+]]:vrm2 = COPY $v10m2
+  ; MIR-NEXT:   [[COPY3:%[0-9]+]]:gpr = COPY $x10
+  ; MIR-NEXT:   [[COPY4:%[0-9]+]]:vrnov0 = COPY $v8
+  ; MIR-NEXT:   $v0 = COPY [[COPY1]]
+  ; MIR-NEXT:   early-clobber %6:vrnov0 = PseudoVLUXEI64_V_M2_M1_MASK [[COPY4]], [[COPY3]], [[COPY2]], $v0, [[COPY]], 5 /* e32 */, 0
+  ; MIR-NEXT:   $v8 = COPY %6
+  ; MIR-NEXT:   PseudoRET implicit $v8
   %1 = zext i32 %vl to i64
   %a = call <vscale x 2 x i32> @llvm.riscv.vluxei.nxv2i32.nxv2i64(<vscale x 2 x i32> undef, <vscale x 2 x i32>* %p, <vscale x 2 x i64> %idx, i64 %1)
   %b = call <vscale x 2 x i32> @llvm.vp.merge.nxv2i32(<vscale x 2 x i1> %m, <vscale x 2 x i32> %a, <vscale x 2 x i32> %passthru, i32 %vl)
   ret <vscale x 2 x i32> %b
 }
 
 ; Test vector index by riscv.vid
 declare <vscale x 2 x i32> @llvm.riscv.vid.nxv2i32(<vscale x 2 x i32>, i64)
 define <vscale x 2 x i32> @vpmerge_vid(<vscale x 2 x i32> %passthru, <vscale x 2 x i1> %m, i32 zeroext %vl) {
 ; CHECK-LABEL: vpmerge_vid:
 ; CHECK:       # %bb.0:
-; CHECK-NEXT:    vsetvli zero, a0, e32, m1, ta, mu
-; CHECK-NEXT:    vid.v v9
-; CHECK-NEXT:    vsetvli zero, zero, e32, m1, tu, mu
-; CHECK-NEXT:    vmerge.vvm v8, v8, v9, v0
+; CHECK-NEXT:    vsetvli zero, a0, e32, m1, tu, mu
+; CHECK-NEXT:    vid.v v8, v0.t
 ; CHECK-NEXT:    ret
+  ; MIR-LABEL: name: vpmerge_vid
+  ; MIR: bb.0 (%ir-block.0):
+  ; MIR-NEXT:   liveins: $v8, $v0, $x10
+  ; MIR-NEXT: {{  $}}
+  ; MIR-NEXT:   [[COPY:%[0-9]+]]:gprnox0 = COPY $x10
+  ; MIR-NEXT:   [[COPY1:%[0-9]+]]:vr = COPY $v0
+  ; MIR-NEXT:   [[COPY2:%[0-9]+]]:vrnov0 = COPY $v8
+  ; MIR-NEXT:   $v0 = COPY [[COPY1]]
+  ; MIR-NEXT:   [[PseudoVID_V_M1_MASK:%[0-9]+]]:vrnov0 = PseudoVID_V_M1_MASK [[COPY2]], $v0, [[COPY]], 5 /* e32 */, 0
+  ; MIR-NEXT:   $v8 = COPY [[PseudoVID_V_M1_MASK]]
+  ; MIR-NEXT:   PseudoRET implicit $v8
   %1 = zext i32 %vl to i64
   %a = call <vscale x 2 x i32> @llvm.riscv.vid.nxv2i32(<vscale x 2 x i32> undef, i64 %1)
   %b = call <vscale x 2 x i32> @llvm.vp.merge.nxv2i32(<vscale x 2 x i1> %m, <vscale x 2 x i32> %a, <vscale x 2 x i32> %passthru, i32 %vl)
   ret <vscale x 2 x i32> %b
 }
 
 ; Test riscv.viota
 declare <vscale x 2 x i32> @llvm.riscv.viota.nxv2i32(<vscale x 2 x i32>, <vscale x 2 x i1>, i64)
 define <vscale x 2 x i32> @vpmerge_viota(<vscale x 2 x i32> %passthru, <vscale x 2 x i1> %m, <vscale x 2 x i1> %vm, i32 zeroext %vl) {
 ; CHECK-LABEL: vpmerge_viota:
 ; CHECK:       # %bb.0:
-; CHECK-NEXT:    vsetvli zero, a0, e32, m1, ta, mu
-; CHECK-NEXT:    viota.m v10, v9
-; CHECK-NEXT:    vsetvli zero, zero, e32, m1, tu, mu
-; CHECK-NEXT:    vmerge.vvm v8, v8, v10, v0
+; CHECK-NEXT:    vsetvli zero, a0, e32, m1, tu, mu
+; CHECK-NEXT:    viota.m v8, v9, v0.t
 ; CHECK-NEXT:    ret
+  ; MIR-LABEL: name: vpmerge_viota
+  ; MIR: bb.0 (%ir-block.0):
+  ; MIR-NEXT:   liveins: $v8, $v0, $v9, $x10
+  ; MIR-NEXT: {{  $}}
+  ; MIR-NEXT:   [[COPY:%[0-9]+]]:gprnox0 = COPY $x10
+  ; MIR-NEXT:   [[COPY1:%[0-9]+]]:vr = COPY $v9
+  ; MIR-NEXT:   [[COPY2:%[0-9]+]]:vr = COPY $v0
+  ; MIR-NEXT:   [[COPY3:%[0-9]+]]:vrnov0 = COPY $v8
+  ; MIR-NEXT:   $v0 = COPY [[COPY2]]
+  ; MIR-NEXT:   early-clobber %4:vrnov0 = PseudoVIOTA_M_M1_MASK [[COPY3]], [[COPY1]], $v0, [[COPY]], 5 /* e32 */, 0
+  ; MIR-NEXT:   $v8 = COPY %4
+  ; MIR-NEXT:   PseudoRET implicit $v8
   %1 = zext i32 %vl to i64
   %a = call <vscale x 2 x i32> @llvm.riscv.viota.nxv2i32(<vscale x 2 x i32> undef, <vscale x 2 x i1> %vm, i64 %1)
   %b = call <vscale x 2 x i32> @llvm.vp.merge.nxv2i32(<vscale x 2 x i1> %m, <vscale x 2 x i32> %a, <vscale x 2 x i32> %passthru, i32 %vl)
   ret <vscale x 2 x i32> %b
 }
 
 ; Test riscv.vfclass
 declare <vscale x 2 x i32> @llvm.riscv.vfclass.nxv2i32(<vscale x 2 x i32>, <vscale x 2 x float>, i64)
 define <vscale x 2 x i32> @vpmerge_vflcass(<vscale x 2 x i32> %passthru, <vscale x 2 x float> %vf, <vscale x 2 x i1> %m, i32 zeroext %vl) {
 ; CHECK-LABEL: vpmerge_vflcass:
 ; CHECK:       # %bb.0:
-; CHECK-NEXT:    vsetvli zero, a0, e32, m1, ta, mu
-; CHECK-NEXT:    vfclass.v v9, v9
-; CHECK-NEXT:    vsetvli zero, zero, e32, m1, tu, mu
-; CHECK-NEXT:    vmerge.vvm v8, v8, v9, v0
+; CHECK-NEXT:    vsetvli zero, a0, e32, m1, tu, mu
+; CHECK-NEXT:    vfclass.v v8, v9, v0.t
 ; CHECK-NEXT:    ret
+  ; MIR-LABEL: name: vpmerge_vflcass
+  ; MIR: bb.0 (%ir-block.0):
+  ; MIR-NEXT:   liveins: $v8, $v9, $v0, $x10
+  ; MIR-NEXT: {{  $}}
+  ; MIR-NEXT:   [[COPY:%[0-9]+]]:gprnox0 = COPY $x10
+  ; MIR-NEXT:   [[COPY1:%[0-9]+]]:vr = COPY $v0
+  ; MIR-NEXT:   [[COPY2:%[0-9]+]]:vr = COPY $v9
+  ; MIR-NEXT:   [[COPY3:%[0-9]+]]:vrnov0 = COPY $v8
+  ; MIR-NEXT:   $v0 = COPY [[COPY1]]
+  ; MIR-NEXT:   [[PseudoVFCLASS_V_M1_MASK:%[0-9]+]]:vrnov0 = PseudoVFCLASS_V_M1_MASK [[COPY3]], [[COPY2]], $v0, [[COPY]], 5 /* e32 */, 0
+  ; MIR-NEXT:   $v8 = COPY [[PseudoVFCLASS_V_M1_MASK]]
+  ; MIR-NEXT:   PseudoRET implicit $v8
   %1 = zext i32 %vl to i64
   %a = call <vscale x 2 x i32> @llvm.riscv.vfclass.nxv2i32(<vscale x 2 x i32> undef, <vscale x 2 x float> %vf, i64 %1)
   %b = call <vscale x 2 x i32> @llvm.vp.merge.nxv2i32(<vscale x 2 x i1> %m, <vscale x 2 x i32> %a, <vscale x 2 x i32> %passthru, i32 %vl)
   ret <vscale x 2 x i32> %b
 }
 
 ; Test riscv.vfsqrt
 declare <vscale x 2 x float> @llvm.riscv.vfsqrt.nxv2f32(<vscale x 2 x float>, <vscale x 2 x float>, i64)
 define <vscale x 2 x float> @vpmerge_vfsqrt(<vscale x 2 x float> %passthru, <vscale x 2 x float> %vf, <vscale x 2 x i1> %m, i32 zeroext %vl) {
 ; CHECK-LABEL: vpmerge_vfsqrt:
 ; CHECK:       # %bb.0:
-; CHECK-NEXT:    vsetvli zero, a0, e32, m1, ta, mu
-; CHECK-NEXT:    vfsqrt.v v9, v9
-; CHECK-NEXT:    vsetvli zero, zero, e32, m1, tu, mu
-; CHECK-NEXT:    vmerge.vvm v8, v8, v9, v0
+; CHECK-NEXT:    vsetvli zero, a0, e32, m1, tu, mu
+; CHECK-NEXT:    vfsqrt.v v8, v9, v0.t
 ; CHECK-NEXT:    ret
+  ; MIR-LABEL: name: vpmerge_vfsqrt
+  ; MIR: bb.0 (%ir-block.0):
+  ; MIR-NEXT:   liveins: $v8, $v9, $v0, $x10
+  ; MIR-NEXT: {{  $}}
+  ; MIR-NEXT:   [[COPY:%[0-9]+]]:gprnox0 = COPY $x10
+  ; MIR-NEXT:   [[COPY1:%[0-9]+]]:vr = COPY $v0
+  ; MIR-NEXT:   [[COPY2:%[0-9]+]]:vr = COPY $v9
+  ; MIR-NEXT:   [[COPY3:%[0-9]+]]:vrnov0 = COPY $v8
+  ; MIR-NEXT:   $v0 = COPY [[COPY1]]
+  ; MIR-NEXT:   [[PseudoVFSQRT_V_M1_MASK:%[0-9]+]]:vrnov0 = PseudoVFSQRT_V_M1_MASK [[COPY3]], [[COPY2]], $v0, [[COPY]], 5 /* e32 */, 0, implicit $frm
+  ; MIR-NEXT:   $v8 = COPY [[PseudoVFSQRT_V_M1_MASK]]
+  ; MIR-NEXT:   PseudoRET implicit $v8
   %1 = zext i32 %vl to i64
   %a = call <vscale x 2 x float> @llvm.riscv.vfsqrt.nxv2f32(<vscale x 2 x float> undef, <vscale x 2 x float> %vf, i64 %1)
   %b = call <vscale x 2 x float> @llvm.vp.merge.nxv2f32(<vscale x 2 x i1> %m, <vscale x 2 x float> %a, <vscale x 2 x float> %passthru, i32 %vl)
   ret <vscale x 2 x float> %b
 }
 
 ; Test reciprocal operation by riscv.vfrec7
 declare <vscale x 2 x float> @llvm.riscv.vfrec7.nxv2f32(<vscale x 2 x float>, <vscale x 2 x float>, i64)
 define <vscale x 2 x float> @vpmerge_vfrec7(<vscale x 2 x float> %passthru, <vscale x 2 x float> %vf, <vscale x 2 x i1> %m, i32 zeroext %vl) {
 ; CHECK-LABEL: vpmerge_vfrec7:
 ; CHECK:       # %bb.0:
-; CHECK-NEXT:    vsetvli zero, a0, e32, m1, ta, mu
-; CHECK-NEXT:    vfrec7.v v9, v9
-; CHECK-NEXT:    vsetvli zero, zero, e32, m1, tu, mu
-; CHECK-NEXT:    vmerge.vvm v8, v8, v9, v0
+; CHECK-NEXT:    vsetvli zero, a0, e32, m1, tu, mu
+; CHECK-NEXT:    vfrec7.v v8, v9, v0.t
 ; CHECK-NEXT:    ret
+  ; MIR-LABEL: name: vpmerge_vfrec7
+  ; MIR: bb.0 (%ir-block.0):
+  ; MIR-NEXT:   liveins: $v8, $v9, $v0, $x10
+  ; MIR-NEXT: {{  $}}
+  ; MIR-NEXT:   [[COPY:%[0-9]+]]:gprnox0 = COPY $x10
+  ; MIR-NEXT:   [[COPY1:%[0-9]+]]:vr = COPY $v0
+  ; MIR-NEXT:   [[COPY2:%[0-9]+]]:vr = COPY $v9
+  ; MIR-NEXT:   [[COPY3:%[0-9]+]]:vrnov0 = COPY $v8
+  ; MIR-NEXT:   $v0 = COPY [[COPY1]]
+  ; MIR-NEXT:   [[PseudoVFREC7_V_M1_MASK:%[0-9]+]]:vrnov0 = PseudoVFREC7_V_M1_MASK [[COPY3]], [[COPY2]], $v0, [[COPY]], 5 /* e32 */, 0, implicit $frm
+  ; MIR-NEXT:   $v8 = COPY [[PseudoVFREC7_V_M1_MASK]]
+  ; MIR-NEXT:   PseudoRET implicit $v8
   %1 = zext i32 %vl to i64
   %a = call <vscale x 2 x float> @llvm.riscv.vfrec7.nxv2f32(<vscale x 2 x float> undef, <vscale x 2 x float> %vf, i64 %1)
   %b = call <vscale x 2 x float> @llvm.vp.merge.nxv2f32(<vscale x 2 x i1> %m, <vscale x 2 x float> %a, <vscale x 2 x float> %passthru, i32 %vl)
   ret <vscale x 2 x float> %b
 }