[RISCV] Merge vmerge.vvm and unmasked intrinsic with VLMAX vector length.

[fakepaper56]

Summary

The motivation of this patch is to lower the IR pattern
(vp.merge mask, (add x, y), false, vl) to
(PseudoVADD_VV_<LMUL>_MASK false, x, y, mask, vl).

[craig.topper]

llvm/lib/Target/RISCV/RISCVISelDAGToDAG.cpp

2678

This is incorrect if the True opcode can raise a floating point exception. Applying the mask could lose an exception that should occur even if the result is discarded.

I think we need to look up the MayRaiseException property from the MCInstrDesc entry for the opcode. If it is has that properly, we need to check that the hasNoFPExcept bit in the SDNodeFlags is set before doing this combine.

[craig.topper]

llvm/lib/Target/RISCV/RISCVISelDAGToDAG.cpp

2707–2708

We also should be copying the NoFPExcept flag here. Can probably just take the SDNodeFlags from the True node.

[fakepaper56]

llvm/lib/Target/RISCV/RISCVISelDAGToDAG.cpp

2678

Good point. Thank you find the bug.

[fakepaper56]

Address Craig's comment and rebase.

[fakepaper56]

Make test case more similiar as previous commit.

[craig.topper]

LGTM

[luke]

llvm/lib/Target/RISCV/RISCVISelDAGToDAG.cpp

2678

@fakepaper56 @craig.topper is there a reason why we only check for fp exceptions when the VL changes and not the mask too?

llvm/lib/Target/RISCV/RISCVISelDAGToDAG.cpp

while (Position != CurDAG->allnodes_begin()) {
 
     // 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)
+
+    auto IsNoFPExcept = [this](SDValue N) {
+      return !this->mayRaiseFPException(N.getNode()) ||
+             N->getFlags().hasNoFPExcept();
+    };
+
+    // Allow the peephole for non-exception True with VLMAX vector length, since
+    // all the values after VL of N are dependent on Merge. VLMAX should be
+    // lowered to (XLenVT -1).
+    if (TrueVL != VL && !(IsNoFPExcept(True) && isAllOnesConstant(TrueVL)))
       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);
     Result->setFlags(True->getFlags());
 
     // 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));
 

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

; CHECK-NEXT:    ret
   ; MIR-NEXT:   [[PseudoVFREC7_V_M1_MASK:%[0-9]+]]:vrnov0 = nofpexcept 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
 }
+
+; Test vector operations with VLMAX vector length.
+
+; Test binary operator with vp.merge and add.
+define <vscale x 2 x i32> @vpmerge_add(<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_add:
+; CHECK:       # %bb.0:
+; 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_add
+  ; 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
+  %a = add <vscale x 2 x i32> %x, %y
+  %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 binary operator with vp.merge and fadd.
+define <vscale x 2 x float> @vpmerge_fadd(<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_fadd:
+; CHECK:       # %bb.0:
+; 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_fadd
+  ; 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:   %5:vrnov0 = nofpexcept PseudoVFADD_VV_M1_MASK [[COPY4]], [[COPY3]], [[COPY2]], $v0, [[COPY]], 5 /* e32 */, 0, implicit $frm
+  ; MIR-NEXT:   $v8 = COPY %5
+  ; MIR-NEXT:   PseudoRET implicit $v8
+  %a = fadd <vscale x 2 x float> %x, %y
+  %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 conversion by fptosi.
+define <vscale x 2 x i16> @vpmerge_fptosi(<vscale x 2 x i16> %passthru, <vscale x 2 x float> %x, <vscale x 2 x i1> %m, i32 zeroext %vl) {
+; CHECK-LABEL: vpmerge_fptosi:
+; CHECK:       # %bb.0:
+; 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_fptosi
+  ; 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
+  %a = fptosi <vscale x 2 x float> %x to <vscale x 2 x i16>
+  %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.
+define <vscale x 2 x float> @vpmerge_sitofp(<vscale x 2 x float> %passthru, <vscale x 2 x i64> %x, <vscale x 2 x i1> %m, i32 zeroext %vl) {
+; CHECK-LABEL: vpmerge_sitofp:
+; CHECK:       # %bb.0:
+; 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_sitofp
+  ; 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
+  %a = sitofp <vscale x 2 x i64> %x to <vscale x 2 x float>
+  %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 float extension by fpext.
+define <vscale x 2 x double> @vpmerge_fpext(<vscale x 2 x double> %passthru, <vscale x 2 x float> %x, <vscale x 2 x i1> %m, i32 zeroext %vl) {
+; CHECK-LABEL: vpmerge_fpext:
+; CHECK:       # %bb.0:
+; 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_fpext
+  ; 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
+  %a = fpext <vscale x 2 x float> %x to <vscale x 2 x double>
+  %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 float truncation by fptrunc.
+define <vscale x 2 x float> @vpmerge_fptrunc(<vscale x 2 x float> %passthru, <vscale x 2 x double> %x, <vscale x 2 x i1> %m, i32 zeroext %vl) {
+; CHECK-LABEL: vpmerge_fptrunc:
+; CHECK:       # %bb.0:
+; 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_fptrunc
+  ; 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
+  %a = fptrunc <vscale x 2 x double> %x to <vscale x 2 x float>
+  %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 zext.
+define <vscale x 2 x i32> @vpmerge_zext(<vscale x 2 x i32> %passthru, <vscale x 2 x i8> %x, <vscale x 2 x i1> %m, i32 zeroext %vl) {
+; CHECK-LABEL: vpmerge_zext:
+; CHECK:       # %bb.0:
+; CHECK-NEXT:    vsetvli zero, a0, e32, m1, tu, mu
+; CHECK-NEXT:    vzext.vf4 v8, v9, v0.t
+; CHECK-NEXT:    ret
+  ; MIR-LABEL: name: vpmerge_zext
+  ; 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
+  %a = zext <vscale x 2 x i8> %x to <vscale x 2 x i32>
+  %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 trunc.
+define <vscale x 2 x i32> @vpmerge_trunc(<vscale x 2 x i32> %passthru, <vscale x 2 x i64> %x, <vscale x 2 x i1> %m, i32 zeroext %vl) {
+; CHECK-LABEL: vpmerge_trunc:
+; CHECK:       # %bb.0:
+; CHECK-NEXT:    vsetvli zero, a0, e32, m1, tu, mu
+; CHECK-NEXT:    vnsrl.wi v8, v10, 0, v0.t
+; CHECK-NEXT:    ret
+  ; MIR-LABEL: name: vpmerge_trunc
+  ; 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_WI_M1_MASK [[COPY3]], [[COPY2]], 0, $v0, [[COPY]], 5 /* e32 */, 0
+  ; MIR-NEXT:   $v8 = COPY %4
+  ; MIR-NEXT:   PseudoRET implicit $v8
+  %a = trunc <vscale x 2 x i64> %x to <vscale x 2 x i32>
+  %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
+}

llvm/test/CodeGen/RISCV/rvv/vpmerge-sdnode.ll

 define <vscale x 1 x i64> @vpmerge_vx_nxv1i64(i64 %a, <vscale x 1 x i64> %vb, <vscale x 1 x i1> %m, i32 zeroext %evl) {
 ; RV32-LABEL: vpmerge_vx_nxv1i64:
 ; RV32:       # %bb.0:
 ; RV32-NEXT:    addi sp, sp, -16
 ; RV32-NEXT:    .cfi_def_cfa_offset 16
 ; RV32-NEXT:    sw a1, 12(sp)
 ; RV32-NEXT:    sw a0, 8(sp)
 ; RV32-NEXT:    addi a0, sp, 8
-; RV32-NEXT:    vsetvli a1, zero, e64, m1, ta, mu
-; RV32-NEXT:    vlse64.v v9, (a0), zero
 ; RV32-NEXT:    vsetvli zero, a2, e64, m1, tu, mu
-; RV32-NEXT:    vmerge.vvm v8, v8, v9, v0
+; RV32-NEXT:    vlse64.v v8, (a0), zero, v0.t
 ; RV32-NEXT:    addi sp, sp, 16
 ; RV32-NEXT:    ret
 ;
 ; RV64-LABEL: vpmerge_vx_nxv1i64:
 ; RV64:       # %bb.0:
 ; RV64-NEXT:    vsetvli zero, a1, e64, m1, tu, mu
 ; RV64-NEXT:    vmerge.vxm v8, v8, a0, v0
 ; RV64-NEXT:    ret

 define <vscale x 2 x i64> @vpmerge_vx_nxv2i64(i64 %a, <vscale x 2 x i64> %vb, <vscale x 2 x i1> %m, i32 zeroext %evl) {
 ; RV32-LABEL: vpmerge_vx_nxv2i64:
 ; RV32:       # %bb.0:
 ; RV32-NEXT:    addi sp, sp, -16
 ; RV32-NEXT:    .cfi_def_cfa_offset 16
 ; RV32-NEXT:    sw a1, 12(sp)
 ; RV32-NEXT:    sw a0, 8(sp)
 ; RV32-NEXT:    addi a0, sp, 8
-; RV32-NEXT:    vsetvli a1, zero, e64, m2, ta, mu
-; RV32-NEXT:    vlse64.v v10, (a0), zero
 ; RV32-NEXT:    vsetvli zero, a2, e64, m2, tu, mu
-; RV32-NEXT:    vmerge.vvm v8, v8, v10, v0
+; RV32-NEXT:    vlse64.v v8, (a0), zero, v0.t
 ; RV32-NEXT:    addi sp, sp, 16
 ; RV32-NEXT:    ret
 ;
 ; RV64-LABEL: vpmerge_vx_nxv2i64:
 ; RV64:       # %bb.0:
 ; RV64-NEXT:    vsetvli zero, a1, e64, m2, tu, mu
 ; RV64-NEXT:    vmerge.vxm v8, v8, a0, v0
 ; RV64-NEXT:    ret

 define <vscale x 4 x i64> @vpmerge_vx_nxv4i64(i64 %a, <vscale x 4 x i64> %vb, <vscale x 4 x i1> %m, i32 zeroext %evl) {
 ; RV32-LABEL: vpmerge_vx_nxv4i64:
 ; RV32:       # %bb.0:
 ; RV32-NEXT:    addi sp, sp, -16
 ; RV32-NEXT:    .cfi_def_cfa_offset 16
 ; RV32-NEXT:    sw a1, 12(sp)
 ; RV32-NEXT:    sw a0, 8(sp)
 ; RV32-NEXT:    addi a0, sp, 8
-; RV32-NEXT:    vsetvli a1, zero, e64, m4, ta, mu
-; RV32-NEXT:    vlse64.v v12, (a0), zero
 ; RV32-NEXT:    vsetvli zero, a2, e64, m4, tu, mu
-; RV32-NEXT:    vmerge.vvm v8, v8, v12, v0
+; RV32-NEXT:    vlse64.v v8, (a0), zero, v0.t
 ; RV32-NEXT:    addi sp, sp, 16
 ; RV32-NEXT:    ret
 ;
 ; RV64-LABEL: vpmerge_vx_nxv4i64:
 ; RV64:       # %bb.0:
 ; RV64-NEXT:    vsetvli zero, a1, e64, m4, tu, mu
 ; RV64-NEXT:    vmerge.vxm v8, v8, a0, v0
 ; RV64-NEXT:    ret

 define <vscale x 8 x i64> @vpmerge_vx_nxv8i64(i64 %a, <vscale x 8 x i64> %vb, <vscale x 8 x i1> %m, i32 zeroext %evl) {
 ; RV32-LABEL: vpmerge_vx_nxv8i64:
 ; RV32:       # %bb.0:
 ; RV32-NEXT:    addi sp, sp, -16
 ; RV32-NEXT:    .cfi_def_cfa_offset 16
 ; RV32-NEXT:    sw a1, 12(sp)
 ; RV32-NEXT:    sw a0, 8(sp)
 ; RV32-NEXT:    addi a0, sp, 8
-; RV32-NEXT:    vsetvli a1, zero, e64, m8, ta, mu
-; RV32-NEXT:    vlse64.v v16, (a0), zero
 ; RV32-NEXT:    vsetvli zero, a2, e64, m8, tu, mu
-; RV32-NEXT:    vmerge.vvm v8, v8, v16, v0
+; RV32-NEXT:    vlse64.v v8, (a0), zero, v0.t
 ; RV32-NEXT:    addi sp, sp, 16
 ; RV32-NEXT:    ret
 ;
 ; RV64-LABEL: vpmerge_vx_nxv8i64:
 ; RV64:       # %bb.0:
 ; RV64-NEXT:    vsetvli zero, a1, e64, m8, tu, mu
 ; RV64-NEXT:    vmerge.vxm v8, v8, a0, v0
 ; RV64-NEXT:    ret