[RISCV] Make scalable vector FMA commutable for register allocation.

[craig.topper]

Summary

This adds support for commuting operands and converting between
vfmadd and vfmacc to avoid register copies.

To avoid messing up intrinsic behavior, I've added new pseudo
instructions that have the isCommutable flag set. These pseudos also
force a tail agnostic policy. The intrinsic version still use
the tail undisturbed policy.

For best results it looks like we need to start with fmadd and only
pick fmacc if its beneficial. MachineCSE commutes without contraining
the operands and then commutes back if it didn't help with CSE. So
I've made sure that when the operand choice isn't constrained, we
will keep fmadd for MachineCSE and when it does the second commute,
we get back the original instruction.

[frasercrmck]

There seem to be a lot of linter errors which might make working on this code difficult. At the least it makes the diff harder to read. I know the macro cases confuse clang-tidy but is there something we can do?

llvm/lib/Target/RISCV/RISCVInstrInfo.cpp

945

minued -> minuend.

llvm/lib/Target/RISCV/RISCVInstrInfoVPseudos.td

1784

Is this a FIXME, as in, are we likely to change the policy on the intrinsics?

Also I'm still a bit unclear on what the tail policy has to do with commuting operands. What am I missing?

[craig.topper]

llvm/lib/Target/RISCV/RISCVInstrInfoVPseudos.td

1784

We're currently using tail undisturbed policy on any instruction with a sourced tied to a destination.

I believe the example I was shown where someone expected this to work was something like this.

float foo(float *src1, float *src2, size_t n) {       
  size_t len;                    
  len = vsetvlmax_e32m8();
  vfloat32m8_t v16 = vfmv_v_f_f32m8(0.0, len);
  len = vsetvl_e32m1();
  vfloat32m1_t v24 = vfmv_s_f_f32m1(vundefined_f32m1(), 0.0, len);
  for (; (len = vl_extract(vsetvl_e32m8(n))) > 0; n -= len) {
    vfloat32m8_t v0 = vle32_v_f32m8(src1, len);
    vfloat32m8_t v0 = vle32_v_f32m8(src2, len);
    v16 = vfmacc_vv_f32m1(v16, v0, v8, len);
    src1 += len;
    src2 += len;                                                                                                                                                                         
  }
  len = vsetvlmax_e32m8();                                                                                                                                                                                                                                                                                                       
  vfloat32m1_t result = vfredosum_vs_f32m8_f32m1(v16, v24, len);                                                                                                                                                                                                                                                                
  return vfmv_f_s_f32m1_f32(result);                                                                                                                                                                                                                                                                                       
}

On the last loop iteration, len might be less than vlmax and the code depends on the tail elements of v16 being preserved from the previous iterations. After the loop a reduction is done using vlmax that will access those elements.

If we commute fmacc to fmadd, then the register used for the v16 input will not be tied to the output register used for v16 for the fmacc. This would prevent the tail elements from being preserved.

I'm not sure we should be allowing this code to work, but tail agnostic is a valid implementation of tail undisturbed. So even if we picked tail agnostic this code might work on in order CPUs and then break in the future on an out of order CPU.

[jrtc27]

If you wanted to appease the linter you could do something gross like:

#define CASE_VFMA_OPCODE_COMMON(OP, TYPE, LMUL) \
  RISCV::PseudoV##OP##_##TYPE##_##LMUL##_COMMUTABLE

#define CASE_VFMA_OPCODE_LMULS(OP, TYPE) \
  CASE_VFMA_OPCODE_COMMON(OP, TYPE, MF8): \
  case CASE_VFMA_OPCODE_COMMON(OP, TYPE, MF4): \
  case CASE_VFMA_OPCODE_COMMON(OP, TYPE, MF2): \
  case CASE_VFMA_OPCODE_COMMON(OP, TYPE, M1): \
  case CASE_VFMA_OPCODE_COMMON(OP, TYPE, M2): \
  case CASE_VFMA_OPCODE_COMMON(OP, TYPE, M4): \
  case CASE_VFMA_OPCODE_COMMON(OP, TYPE, M8)

#define CASE_VFMA_SPLATS(OP) \
  CASE_VFMA_OPCODE_LMULS(OP, VF16): \
  case CASE_VFMA_OPCODE_LMULS(OP, VF32): \
  case CASE_VFMA_OPCODE_LMULS(OP, VF64)

(i.e. the leading case and trailing : are omitted) that you then use as, say:

case CASE_VFMA_OPCODE_LMULS(FMADD, VV):
case CASE_VFMA_OPCODE_LMULS(FMSUB, VV):
case CASE_VFMA_OPCODE_LMULS(FNMADD, VV):
case CASE_VFMA_OPCODE_LMULS(FNMSUB, VV): {

which is hopefully enough to get it to recognise the macros as case statements. Maybe you can even keep the case keyword down in CASE_VFMA_OPCODE_COMMON and it's just the colon that's needed in order to make it look label-ish; if so that's rather nicer (and mirrors how statement macros tend to leave off the final semicolon).

[craig.topper]

Attempt to appease clang-format

[jrtc27]

llvm/lib/Target/RISCV/RISCVInstrInfo.cpp

899

Ouch; maybe tactful use of // clang-format [on|off] around these macro definitions would be better?

[jrtc27]

llvm/lib/Target/RISCV/RISCVInstrInfo.cpp

1105

I'd leave a blank line after this given there's one between the macro definitions and the function

[craig.topper]

Disable clang-format around macros

[rogfer01]

llvm/lib/Target/RISCV/RISCVInstrInfoVPseudos.td

1784

I'm not sure we should be allowing this code to work, but tail agnostic is a valid implementation of tail undisturbed.

I think you mean tail undisturbed is a valid implementation of tail agnostic?

I don't think this code should work either, but 4 ops instructions like vfmacc implicitly tie the output to one input.

Whether we should be exposing this detail to intrinsics is not clear to me (I wouldn't) and this code should be using some alternative intrinsic (like the ones sketched https://github.com/riscv/rvv-intrinsic-doc/issues/27#issuecomment-649433549 here).

Most vector codes can work well with just tail agnostic mode. So far only accumulations (or reductions) like the one you show seem to be the ones benefiting from tail undisturbed (as the accumulation happens on the full register due to the varying vl during the execution of the loop). Reading fma as "float multiply accumulate" doesn't help to make the distinction clearer.

[craig.topper]

llvm/lib/Target/RISCV/RISCVInstrInfoVPseudos.td

1784

I think you mean tail undisturbed is a valid implementation of tail agnostic?

Yes. Thanks

[frasercrmck]

llvm/lib/Target/RISCV/RISCVInstrInfoVPseudos.td

1784

Great example, thanks. I think the NOTE is clearer than the FIXME, too.

llvm/lib/Target/RISCV/RISCVInstrInfoVSDPatterns.td

588

You could perhaps put fvti.ScalarSuffix # "_" # fvti.LMul.MX # "_COMMUTABLE" into a defvar to reduce the line length.

[craig.topper]

Use defvar to shorten line length

[craig.topper]

Fix typo minued->minuend

[frasercrmck]

LGTM!

llvm/lib/Target/RISCV/MCTargetDesc/RISCVBaseInfo.h

enum {
 
   VLMulShift = ConstraintShift + 3,
   VLMulMask = 0b111 << VLMulShift,
 
   // Do we need to add a dummy mask op when converting RVV Pseudo to MCInst.
   HasDummyMaskOpShift = VLMulShift + 3,
   HasDummyMaskOpMask = 1 << HasDummyMaskOpShift,
 
-  // Does this instruction only update element 0 the destination register.
-  WritesElement0Shift = HasDummyMaskOpShift + 1,
-  WritesElement0Mask = 1 << WritesElement0Shift,
+  // Force a tail agnostic policy even this instruction has a tied destination.
+  ForceTailAgnosticShift = HasDummyMaskOpShift + 1,
+  ForceTailAgnosticMask = 1 << ForceTailAgnosticShift,
 
   // Does this instruction have a merge operand that must be removed when
   // converting to MCInst. It will be the first explicit use operand. Used by
   // RVV Pseudos.
-  HasMergeOpShift = WritesElement0Shift + 1,
+  HasMergeOpShift = ForceTailAgnosticShift + 1,
   HasMergeOpMask = 1 << HasMergeOpShift,
 
   // Does this instruction have a SEW operand. It will be the last explicit
   // operand. Used by RVV Pseudos.
   HasSEWOpShift = HasMergeOpShift + 1,
   HasSEWOpMask = 1 << HasSEWOpShift,
 
   // Does this instruction have a VL operand. It will be the second to last

llvm/lib/Target/RISCV/RISCVISelLowering.cpp

static MachineBasicBlock *emitSelectPseudo(MachineInstr &MI,
   }
 
   F->getProperties().reset(MachineFunctionProperties::Property::NoPHIs);
   return TailMBB;
 }
 
 static MachineBasicBlock *addVSetVL(MachineInstr &MI, MachineBasicBlock *BB,
                                     int VLIndex, unsigned SEWIndex,
-                                    RISCVVLMUL VLMul, bool WritesElement0) {
+                                    RISCVVLMUL VLMul, bool ForceTailAgnostic) {
   MachineFunction &MF = *BB->getParent();
   DebugLoc DL = MI.getDebugLoc();
   const TargetInstrInfo &TII = *MF.getSubtarget().getInstrInfo();
 
   unsigned SEW = MI.getOperand(SEWIndex).getImm();
   assert(RISCVVType::isValidSEW(SEW) && "Unexpected SEW");
   RISCVVSEW ElementWidth = static_cast<RISCVVSEW>(Log2_32(SEW / 8));
 
static MachineBasicBlock *addVSetVL(MachineInstr &MI, MachineBasicBlock *BB,
   // Default to tail agnostic unless the destination is tied to a source. In
   // that case the user would have some control over the tail values. The tail
   // policy is also ignored on instructions that only update element 0 like
   // vmv.s.x or reductions so use agnostic there to match the common case.
   // FIXME: This is conservatively correct, but we might want to detect that
   // the input is undefined.
   bool TailAgnostic = true;
   unsigned UseOpIdx;
-  if (MI.isRegTiedToUseOperand(0, &UseOpIdx) && !WritesElement0) {
+  if (!ForceTailAgnostic && MI.isRegTiedToUseOperand(0, &UseOpIdx)) {
     TailAgnostic = false;
     // If the tied operand is an IMPLICIT_DEF we can keep TailAgnostic.
     const MachineOperand &UseMO = MI.getOperand(UseOpIdx);
     MachineInstr *UseMI = MRI.getVRegDef(UseMO.getReg());
     if (UseMI && UseMI->isImplicitDef())
       TailAgnostic = true;
   }
 

 RISCVTargetLowering::EmitInstrWithCustomInserter(MachineInstr &MI,
                                                  MachineBasicBlock *BB) const {
   uint64_t TSFlags = MI.getDesc().TSFlags;
 
   if (TSFlags & RISCVII::HasSEWOpMask) {
     unsigned NumOperands = MI.getNumExplicitOperands();
     int VLIndex = (TSFlags & RISCVII::HasVLOpMask) ? NumOperands - 2 : -1;
     unsigned SEWIndex = NumOperands - 1;
-    bool WritesElement0 = TSFlags & RISCVII::WritesElement0Mask;
+    bool ForceTailAgnostic = TSFlags & RISCVII::ForceTailAgnosticMask;
 
     RISCVVLMUL VLMul = static_cast<RISCVVLMUL>((TSFlags & RISCVII::VLMulMask) >>
                                                RISCVII::VLMulShift);
-    return addVSetVL(MI, BB, VLIndex, SEWIndex, VLMul, WritesElement0);
+    return addVSetVL(MI, BB, VLIndex, SEWIndex, VLMul, ForceTailAgnostic);
   }
 
   switch (MI.getOpcode()) {
   default:
     llvm_unreachable("Unexpected instr type to insert");
   case RISCV::ReadCycleWide:
     assert(!Subtarget.is64Bit() &&
            "ReadCycleWrite is only to be used on riscv32");

llvm/lib/Target/RISCV/RISCVInstrFormats.td

class RVInst<dag outs, dag ins, string opcodestr, string argstr,
   let TSFlags{7-5} = RVVConstraint.Value;
 
   bits<3> VLMul = 0;
   let TSFlags{10-8} = VLMul;
 
   bit HasDummyMask = 0;
   let TSFlags{11} = HasDummyMask;
 
-  bit WritesElement0 = 0;
-  let TSFlags{12} = WritesElement0;
+  bit ForceTailAgnostic = false;
+  let TSFlags{12} = ForceTailAgnostic;
 
   bit HasMergeOp = 0;
   let TSFlags{13} = HasMergeOp;
 
   bit HasSEWOp = 0;
   let TSFlags{14} = HasSEWOp;
 
   bit HasVLOp = 0;

llvm/lib/Target/RISCV/RISCVInstrInfo.h

public:
   buildOutlinedFrame(MachineBasicBlock &MBB, MachineFunction &MF,
                      const outliner::OutlinedFunction &OF) const override;
 
   // Insert a call to an outlined function into a given basic block.
   virtual MachineBasicBlock::iterator
   insertOutlinedCall(Module &M, MachineBasicBlock &MBB,
                      MachineBasicBlock::iterator &It, MachineFunction &MF,
                      const outliner::Candidate &C) const override;
+
+  bool findCommutedOpIndices(const MachineInstr &MI, unsigned &SrcOpIdx1,
+                             unsigned &SrcOpIdx2) const override;
+  MachineInstr *commuteInstructionImpl(MachineInstr &MI, bool NewMI,
+                                       unsigned OpIdx1,
+                                       unsigned OpIdx2) const override;
+
 protected:
   const RISCVSubtarget &STI;
 };
 
 namespace RISCVVPseudosTable {
 
 struct PseudoInfo {
   uint16_t Pseudo;

llvm/lib/Target/RISCV/RISCVInstrInfo.cpp

MachineBasicBlock::iterator RISCVInstrInfo::insertOutlinedCall(
 
   // Add in a call instruction to the outlined function at the given location.
   It = MBB.insert(It,
                   BuildMI(MF, DebugLoc(), get(RISCV::PseudoCALLReg), RISCV::X5)
                       .addGlobalAddress(M.getNamedValue(MF.getName()), 0,
                                         RISCVII::MO_CALL));
   return It;
 }
+
+// clang-format off
+#define CASE_VFMA_OPCODE_COMMON(OP, TYPE, LMUL)                                \
+  RISCV::PseudoV##OP##_##TYPE##_##LMUL##_COMMUTABLE
+
+#define CASE_VFMA_OPCODE_LMULS(OP, TYPE)                                       \
+  CASE_VFMA_OPCODE_COMMON(OP, TYPE, MF8):                                      \
+  case CASE_VFMA_OPCODE_COMMON(OP, TYPE, MF4):                                 \
+  case CASE_VFMA_OPCODE_COMMON(OP, TYPE, MF2):                                 \
+  case CASE_VFMA_OPCODE_COMMON(OP, TYPE, M1):                                  \
+  case CASE_VFMA_OPCODE_COMMON(OP, TYPE, M2):                                  \
+  case CASE_VFMA_OPCODE_COMMON(OP, TYPE, M4):                                  \
+  case CASE_VFMA_OPCODE_COMMON(OP, TYPE, M8)
+
+#define CASE_VFMA_SPLATS(OP)                                                   \
+  CASE_VFMA_OPCODE_LMULS(OP, VF16):                                            \
+  case CASE_VFMA_OPCODE_LMULS(OP, VF32):                                       \
+  case CASE_VFMA_OPCODE_LMULS(OP, VF64)
+// clang-format on
+
+bool RISCVInstrInfo::findCommutedOpIndices(const MachineInstr &MI,
+                                           unsigned &SrcOpIdx1,
+                                           unsigned &SrcOpIdx2) const {
+  const MCInstrDesc &Desc = MI.getDesc();
+  if (!Desc.isCommutable())
+    return false;
+
+  switch (MI.getOpcode()) {
+  case CASE_VFMA_SPLATS(FMADD):
+  case CASE_VFMA_SPLATS(FMSUB):
+  case CASE_VFMA_SPLATS(FMACC):
+  case CASE_VFMA_SPLATS(FMSAC):
+  case CASE_VFMA_SPLATS(FNMADD):
+  case CASE_VFMA_SPLATS(FNMSUB):
+  case CASE_VFMA_SPLATS(FNMACC):
+  case CASE_VFMA_SPLATS(FNMSAC):
+  case CASE_VFMA_OPCODE_LMULS(FMACC, VV):
+  case CASE_VFMA_OPCODE_LMULS(FMSAC, VV):
+  case CASE_VFMA_OPCODE_LMULS(FNMACC, VV):
+  case CASE_VFMA_OPCODE_LMULS(FNMSAC, VV): {
+    // For these instructions we can only swap operand 1 and operand 3 by
+    // changing the opcode.
+    unsigned CommutableOpIdx1 = 1;
+    unsigned CommutableOpIdx2 = 3;
+    if (!fixCommutedOpIndices(SrcOpIdx1, SrcOpIdx2, CommutableOpIdx1,
+                              CommutableOpIdx2))
+      return false;
+    return true;
+  }
+  case CASE_VFMA_OPCODE_LMULS(FMADD, VV):
+  case CASE_VFMA_OPCODE_LMULS(FMSUB, VV):
+  case CASE_VFMA_OPCODE_LMULS(FNMADD, VV):
+  case CASE_VFMA_OPCODE_LMULS(FNMSUB, VV): {
+    // For these instructions we have more freedom. We can commute with the
+    // other multiplicand or with the addend/subtrahend/minuend.
+
+    // Any fixed operand must be from source 1, 2 or 3.
+    if (SrcOpIdx1 != CommuteAnyOperandIndex && SrcOpIdx1 > 3)
+      return false;
+    if (SrcOpIdx2 != CommuteAnyOperandIndex && SrcOpIdx2 > 3)
+      return false;
+
+    // It both ops are fixed one must be the tied source.
+    if (SrcOpIdx1 != CommuteAnyOperandIndex &&
+        SrcOpIdx2 != CommuteAnyOperandIndex && SrcOpIdx1 != 1 && SrcOpIdx2 != 1)
+      return false;
+
+    // Look for two different register operands assumed to be commutable
+    // regardless of the FMA opcode. The FMA opcode is adjusted later if
+    // needed.
+    if (SrcOpIdx1 == CommuteAnyOperandIndex ||
+        SrcOpIdx2 == CommuteAnyOperandIndex) {
+      // At least one of operands to be commuted is not specified and
+      // this method is free to choose appropriate commutable operands.
+      unsigned CommutableOpIdx1 = SrcOpIdx1;
+      if (SrcOpIdx1 == SrcOpIdx2) {
+        // Both of operands are not fixed. Set one of commutable
+        // operands to the tied source.
+        CommutableOpIdx1 = 1;
+      } else if (SrcOpIdx1 == CommutableOpIdx1) {
+        // Only one of the operands is not fixed.
+        CommutableOpIdx1 = SrcOpIdx2;
+      }
+
+      // CommutableOpIdx1 is well defined now. Let's choose another commutable
+      // operand and assign its index to CommutableOpIdx2.
+      unsigned CommutableOpIdx2;
+      if (CommutableOpIdx1 != 1) {
+        // If we haven't already used the tied source, we must use it now.
+        CommutableOpIdx2 = 1;
+      } else {
+        Register Op1Reg = MI.getOperand(CommutableOpIdx1).getReg();
+
+        // The commuted operands should have different registers.
+        // Otherwise, the commute transformation does not change anything and
+        // is useless. We use this as a hint to make our decision.
+        if (Op1Reg != MI.getOperand(2).getReg())
+          CommutableOpIdx2 = 2;
+        else
+          CommutableOpIdx2 = 3;
+      }
+
+      // Assign the found pair of commutable indices to SrcOpIdx1 and
+      // SrcOpIdx2 to return those values.
+      if (!fixCommutedOpIndices(SrcOpIdx1, SrcOpIdx2, CommutableOpIdx1,
+                                CommutableOpIdx2))
+        return false;
+    }
+
+    return true;
+  }
+  }
+
+  return TargetInstrInfo::findCommutedOpIndices(MI, SrcOpIdx1, SrcOpIdx2);
+}
+
+#define CASE_VFMA_CHANGE_OPCODE_COMMON(OLDOP, NEWOP, TYPE, LMUL)               \
+  case RISCV::PseudoV##OLDOP##_##TYPE##_##LMUL##_COMMUTABLE:                   \
+    Opc = RISCV::PseudoV##NEWOP##_##TYPE##_##LMUL##_COMMUTABLE;                \
+    break;
+
+#define CASE_VFMA_CHANGE_OPCODE_LMULS(OLDOP, NEWOP, TYPE)                      \
+  CASE_VFMA_CHANGE_OPCODE_COMMON(OLDOP, NEWOP, TYPE, MF8)                      \
+  CASE_VFMA_CHANGE_OPCODE_COMMON(OLDOP, NEWOP, TYPE, MF4)                      \
+  CASE_VFMA_CHANGE_OPCODE_COMMON(OLDOP, NEWOP, TYPE, MF2)                      \
+  CASE_VFMA_CHANGE_OPCODE_COMMON(OLDOP, NEWOP, TYPE, M1)                       \
+  CASE_VFMA_CHANGE_OPCODE_COMMON(OLDOP, NEWOP, TYPE, M2)                       \
+  CASE_VFMA_CHANGE_OPCODE_COMMON(OLDOP, NEWOP, TYPE, M4)                       \
+  CASE_VFMA_CHANGE_OPCODE_COMMON(OLDOP, NEWOP, TYPE, M8)
+
+#define CASE_VFMA_CHANGE_OPCODE_SPLATS(OLDOP, NEWOP)                           \
+  CASE_VFMA_CHANGE_OPCODE_LMULS(OLDOP, NEWOP, VF16)                            \
+  CASE_VFMA_CHANGE_OPCODE_LMULS(OLDOP, NEWOP, VF32)                            \
+  CASE_VFMA_CHANGE_OPCODE_LMULS(OLDOP, NEWOP, VF64)
+
+MachineInstr *RISCVInstrInfo::commuteInstructionImpl(MachineInstr &MI,
+                                                     bool NewMI,
+                                                     unsigned OpIdx1,
+                                                     unsigned OpIdx2) const {
+  auto cloneIfNew = [NewMI](MachineInstr &MI) -> MachineInstr & {
+    if (NewMI)
+      return *MI.getParent()->getParent()->CloneMachineInstr(&MI);
+    return MI;
+  };
+
+  switch (MI.getOpcode()) {
+  case CASE_VFMA_SPLATS(FMACC):
+  case CASE_VFMA_SPLATS(FMADD):
+  case CASE_VFMA_SPLATS(FMSAC):
+  case CASE_VFMA_SPLATS(FMSUB):
+  case CASE_VFMA_SPLATS(FNMACC):
+  case CASE_VFMA_SPLATS(FNMADD):
+  case CASE_VFMA_SPLATS(FNMSAC):
+  case CASE_VFMA_SPLATS(FNMSUB):
+  case CASE_VFMA_OPCODE_LMULS(FMACC, VV):
+  case CASE_VFMA_OPCODE_LMULS(FMSAC, VV):
+  case CASE_VFMA_OPCODE_LMULS(FNMACC, VV):
+  case CASE_VFMA_OPCODE_LMULS(FNMSAC, VV): {
+    // It only make sense to toggle these between clobbering the
+    // addend/subtrahend/minuend one of the multiplicands.
+    assert((OpIdx1 == 1 || OpIdx2 == 1) && "Unexpected opcode index");
+    assert((OpIdx1 == 3 || OpIdx2 == 3) && "Unexpected opcode index");
+    unsigned Opc;
+    switch (MI.getOpcode()) {
+      default:
+        llvm_unreachable("Unexpected opcode");
+      CASE_VFMA_CHANGE_OPCODE_SPLATS(FMACC, FMADD)
+      CASE_VFMA_CHANGE_OPCODE_SPLATS(FMADD, FMACC)
+      CASE_VFMA_CHANGE_OPCODE_SPLATS(FMSAC, FMSUB)
+      CASE_VFMA_CHANGE_OPCODE_SPLATS(FMSUB, FMSAC)
+      CASE_VFMA_CHANGE_OPCODE_SPLATS(FNMACC, FNMADD)
+      CASE_VFMA_CHANGE_OPCODE_SPLATS(FNMADD, FNMACC)
+      CASE_VFMA_CHANGE_OPCODE_SPLATS(FNMSAC, FNMSUB)
+      CASE_VFMA_CHANGE_OPCODE_SPLATS(FNMSUB, FNMSAC)
+      CASE_VFMA_CHANGE_OPCODE_LMULS(FMACC, FMADD, VV)
+      CASE_VFMA_CHANGE_OPCODE_LMULS(FMSAC, FMSUB, VV)
+      CASE_VFMA_CHANGE_OPCODE_LMULS(FNMACC, FNMADD, VV)
+      CASE_VFMA_CHANGE_OPCODE_LMULS(FNMSAC, FNMSUB, VV)
+    }
+
+    auto &WorkingMI = cloneIfNew(MI);
+    WorkingMI.setDesc(get(Opc));
+    return TargetInstrInfo::commuteInstructionImpl(WorkingMI, /*NewMI=*/false,
+                                                   OpIdx1, OpIdx2);
+  }
+  case CASE_VFMA_OPCODE_LMULS(FMADD, VV):
+  case CASE_VFMA_OPCODE_LMULS(FMSUB, VV):
+  case CASE_VFMA_OPCODE_LMULS(FNMADD, VV):
+  case CASE_VFMA_OPCODE_LMULS(FNMSUB, VV): {
+    assert((OpIdx1 == 1 || OpIdx2 == 1) && "Unexpected opcode index");
+    // If one of the operands, is the addend we need to change opcode.
+    // Otherwise we're just swapping 2 of the multiplicands.
+    if (OpIdx1 == 3 || OpIdx2 == 3) {
+      unsigned Opc;
+      switch (MI.getOpcode()) {
+        default:
+          llvm_unreachable("Unexpected opcode");
+        CASE_VFMA_CHANGE_OPCODE_LMULS(FMADD, FMACC, VV)
+        CASE_VFMA_CHANGE_OPCODE_LMULS(FMSUB, FMSAC, VV)
+        CASE_VFMA_CHANGE_OPCODE_LMULS(FNMADD, FNMACC, VV)
+        CASE_VFMA_CHANGE_OPCODE_LMULS(FNMSUB, FNMSAC, VV)
+      }
+
+      auto &WorkingMI = cloneIfNew(MI);
+      WorkingMI.setDesc(get(Opc));
+      return TargetInstrInfo::commuteInstructionImpl(WorkingMI, /*NewMI=*/false,
+                                                     OpIdx1, OpIdx2);
+    }
+    // Let the default code handle it.
+    break;
+  }
+  }
+
+  return TargetInstrInfo::commuteInstructionImpl(MI, NewMI, OpIdx1, OpIdx2);
+}
+
+#undef CASE_VFMA_CHANGE_OPCODE_SPLATS
+#undef CASE_VFMA_CHANGE_OPCODE_LMULS
+#undef CASE_VFMA_CHANGE_OPCODE_COMMON
+#undef CASE_VFMA_SPLATS
+#undef CASE_VFMA_OPCODE_LMULS
+#undef CASE_VFMA_OPCODE_COMMON

llvm/lib/Target/RISCV/RISCVInstrInfoVPseudos.td

string VInst = !subst("_M8", "",
                  !subst("_B1", "",
                  !subst("_B2", "",
                  !subst("_B4", "",
                  !subst("_B8", "",
                  !subst("_B16", "",
                  !subst("_B32", "",
                  !subst("_B64", "",
                  !subst("_MASK", "",
+                 !subst("_COMMUTABLE", "",
                  !subst("F16", "F",
                  !subst("F32", "F",
                  !subst("F64", "F",
-                 !subst("Pseudo", "", PseudoInst)))))))))))))))))));
+                 !subst("Pseudo", "", PseudoInst))))))))))))))))))));
 }
 
 class ToLowerCase<string Upper> {
   string L = !subst("FF", "ff",
              !subst("VLSEG", "vlseg",
              !subst("VLSSEG", "vlsseg",
              !subst("VSSEG", "vsseg",
              !subst("VSSSEG", "vssseg",

 multiclass VPseudoTernaryV_VV_VX_AAXA<string Constraint = ""> {
   defm "" : VPseudoTernaryV_VV<Constraint>;
   defm "" : VPseudoTernaryV_VX_AAXA<Constraint>;
 }
 
 multiclass VPseudoTernaryV_VV_VF_AAXA<string Constraint = ""> {
   defm "" : VPseudoTernaryV_VV<Constraint>;
   defm "" : VPseudoTernaryV_VF_AAXA<Constraint>;
+
+  foreach m = MxList.m in {
+    // Add a commutable version for use by IR fma.
+    // NOTE: We need this because we use a tail undisturbed policy on the
+    // intrinsic version so we can't commute those instructions since it would
+    // change which input operand is tied to the destination. That would
+    // remove user control of the tail elements.
+    let isCommutable = 1, ForceTailAgnostic = true, VLMul = m.value in {
+      def "_VV_" # m.MX # "_COMMUTABLE" : VPseudoTernaryNoMask<m.vrclass,
+                                                               m.vrclass,
+                                                               m.vrclass,
+                                                               Constraint>;
+      foreach f = FPList.fpinfo in
+        def "_V" # f.FX # "_" # m.MX # "_COMMUTABLE" :
+           VPseudoTernaryNoMask<m.vrclass, f.fprclass, m.vrclass, Constraint>;
+    }
+  }
 }
 
 multiclass VPseudoTernaryV_VX_VI<Operand ImmType = simm5, string Constraint = ""> {
   defm "" : VPseudoTernaryV_VX<Constraint>;
   defm "" : VPseudoTernaryV_VI<ImmType, Constraint>;
 }
 
 multiclass VPseudoTernaryW_VV_VX {

 
 multiclass VPseudoBinaryM_VX_VI {
   defm "" : VPseudoBinaryM_VX;
   defm "" : VPseudoBinaryM_VI;
 }
 
 multiclass VPseudoReductionV_VS {
   foreach m = MxList.m in {
-    let WritesElement0 = 1 in
+    let ForceTailAgnostic = true in
     defm _VS : VPseudoTernary<V_M1.vrclass, m.vrclass, V_M1.vrclass, m>;
   }
 }
 
 multiclass VPseudoConversion<VReg RetClass,
                              VReg Op1Class,
                              LMULInfo MInfo,
                              string Constraint = ""> {

 let mayLoad = 0, mayStore = 0, hasSideEffects = 0, usesCustomInserter = 1,
     Uses = [VL, VTYPE] in {
   foreach m = MxList.m in {
     let VLMul = m.value in {
       let HasSEWOp = 1, BaseInstr = VMV_X_S in
       def PseudoVMV_X_S # "_" # m.MX: Pseudo<(outs GPR:$rd),
                                              (ins m.vrclass:$rs2, ixlenimm:$sew),
                                              []>, RISCVVPseudo;
-      let HasVLOp = 1, HasSEWOp = 1, BaseInstr = VMV_S_X, WritesElement0 = 1,
-          Constraints = "$rd = $rs1" in
+      let HasVLOp = 1, HasSEWOp = 1, BaseInstr = VMV_S_X,
+          ForceTailAgnostic = true, Constraints = "$rd = $rs1" in
       def PseudoVMV_S_X # "_" # m.MX: Pseudo<(outs m.vrclass:$rd),
                                              (ins m.vrclass:$rs1, GPR:$rs2,
                                                   GPR:$vl, ixlenimm:$sew),
                                              []>, RISCVVPseudo;
     }
   }
 }
 } // Predicates = [HasStdExtV]
foreach m = MxList.m in {
     foreach f = FPList.fpinfo in {
       let VLMul = m.value in {
         let HasSEWOp = 1, BaseInstr = VFMV_F_S in
         def "PseudoVFMV_" # f.FX # "_S_" # m.MX :
                                           Pseudo<(outs f.fprclass:$rd),
                                                  (ins m.vrclass:$rs2,
                                                       ixlenimm:$sew),
                                                  []>, RISCVVPseudo;
-        let HasVLOp = 1, HasSEWOp = 1, BaseInstr = VFMV_S_F, WritesElement0 = 1,
-            Constraints = "$rd = $rs1" in
+        let HasVLOp = 1, HasSEWOp = 1, BaseInstr = VFMV_S_F,
+            ForceTailAgnostic = true, Constraints = "$rd = $rs1" in
         def "PseudoVFMV_S_" # f.FX # "_" # m.MX :
                                           Pseudo<(outs m.vrclass:$rd),
                                                  (ins m.vrclass:$rs1, f.fprclass:$rs2,
                                                       GPR:$vl, ixlenimm:$sew),
                                                  []>, RISCVVPseudo;
       }
     }
   }

llvm/lib/Target/RISCV/RISCVInstrInfoVSDPatterns.td

 defm "" : VPatBinaryFPSDNode_VV_VF<fmul, "PseudoVFMUL">;
 defm "" : VPatBinaryFPSDNode_VV_VF<fdiv, "PseudoVFDIV">;
 defm "" : VPatBinaryFPSDNode_R_VF<fdiv, "PseudoVFRDIV">;
 
 // 14.6 Vector Single-Width Floating-Point Fused Multiply-Add Instructions.
 foreach fvti = AllFloatVectors in {
   // NOTE: We choose VFMADD because it has the most commuting freedom. So it
   // works best with how TwoAddressInstructionPass tries commuting.
+  defvar suffix = fvti.LMul.MX # "_COMMUTABLE";
   def : Pat<(fvti.Vector (fma fvti.RegClass:$rs1, fvti.RegClass:$rd,
                               fvti.RegClass:$rs2)),
-            (!cast<Instruction>("PseudoVFMADD_VV_"# fvti.LMul.MX)
+            (!cast<Instruction>("PseudoVFMADD_VV_"# suffix)
                  fvti.RegClass:$rd, fvti.RegClass:$rs1, fvti.RegClass:$rs2,
                  fvti.AVL, fvti.SEW)>;
   def : Pat<(fvti.Vector (fma fvti.RegClass:$rs1, fvti.RegClass:$rd,
                               (fneg fvti.RegClass:$rs2))),
-            (!cast<Instruction>("PseudoVFMSUB_VV_"# fvti.LMul.MX)
+            (!cast<Instruction>("PseudoVFMSUB_VV_"# suffix)
                  fvti.RegClass:$rd, fvti.RegClass:$rs1, fvti.RegClass:$rs2,
                  fvti.AVL, fvti.SEW)>;
   def : Pat<(fvti.Vector (fma (fneg fvti.RegClass:$rs1), fvti.RegClass:$rd,
                               (fneg fvti.RegClass:$rs2))),
-            (!cast<Instruction>("PseudoVFNMADD_VV_"# fvti.LMul.MX)
+            (!cast<Instruction>("PseudoVFNMADD_VV_"# suffix)
                  fvti.RegClass:$rd, fvti.RegClass:$rs1, fvti.RegClass:$rs2,
                  fvti.AVL, fvti.SEW)>;
   def : Pat<(fvti.Vector (fma (fneg fvti.RegClass:$rs1), fvti.RegClass:$rd,
                               fvti.RegClass:$rs2)),
-            (!cast<Instruction>("PseudoVFNMSUB_VV_"# fvti.LMul.MX)
+            (!cast<Instruction>("PseudoVFNMSUB_VV_"# suffix)
                  fvti.RegClass:$rd, fvti.RegClass:$rs1, fvti.RegClass:$rs2,
                  fvti.AVL, fvti.SEW)>;
 
   // The choice of VFMADD here is arbitrary, vfmadd.vf and vfmacc.vf are equally
   // commutable.
   def : Pat<(fvti.Vector (fma (splat_vector fvti.ScalarRegClass:$rs1),
                               fvti.RegClass:$rd, fvti.RegClass:$rs2)),
-            (!cast<Instruction>("PseudoVFMADD_V" # fvti.ScalarSuffix # "_" # fvti.LMul.MX)
+            (!cast<Instruction>("PseudoVFMADD_V" # fvti.ScalarSuffix # "_" # suffix)
                  fvti.RegClass:$rd, fvti.ScalarRegClass:$rs1, fvti.RegClass:$rs2,
                  fvti.AVL, fvti.SEW)>;
   def : Pat<(fvti.Vector (fma (splat_vector fvti.ScalarRegClass:$rs1),
                               fvti.RegClass:$rd, (fneg fvti.RegClass:$rs2))),
-            (!cast<Instruction>("PseudoVFMSUB_V" # fvti.ScalarSuffix # "_" # fvti.LMul.MX)
+            (!cast<Instruction>("PseudoVFMSUB_V" # fvti.ScalarSuffix # "_" # suffix)
                  fvti.RegClass:$rd, fvti.ScalarRegClass:$rs1, fvti.RegClass:$rs2,
                  fvti.AVL, fvti.SEW)>;
 
   def : Pat<(fvti.Vector (fma (splat_vector fvti.ScalarRegClass:$rs1),
                               (fneg fvti.RegClass:$rd), (fneg fvti.RegClass:$rs2))),
-            (!cast<Instruction>("PseudoVFNMADD_V" # fvti.ScalarSuffix # "_" # fvti.LMul.MX)
+            (!cast<Instruction>("PseudoVFNMADD_V" # fvti.ScalarSuffix # "_" # suffix)
                  fvti.RegClass:$rd, fvti.ScalarRegClass:$rs1, fvti.RegClass:$rs2,
                  fvti.AVL, fvti.SEW)>;
   def : Pat<(fvti.Vector (fma (splat_vector fvti.ScalarRegClass:$rs1),
                               (fneg fvti.RegClass:$rd), fvti.RegClass:$rs2)),
-            (!cast<Instruction>("PseudoVFNMSUB_V" # fvti.ScalarSuffix # "_" # fvti.LMul.MX)
+            (!cast<Instruction>("PseudoVFNMSUB_V" # fvti.ScalarSuffix # "_" # suffix)
                  fvti.RegClass:$rd, fvti.ScalarRegClass:$rs1, fvti.RegClass:$rs2,
                  fvti.AVL, fvti.SEW)>;
 
   // The splat might be negated.
   def : Pat<(fvti.Vector (fma (fneg (splat_vector fvti.ScalarRegClass:$rs1)),
                               fvti.RegClass:$rd, (fneg fvti.RegClass:$rs2))),
-            (!cast<Instruction>("PseudoVFNMADD_V" # fvti.ScalarSuffix # "_" # fvti.LMul.MX)
+            (!cast<Instruction>("PseudoVFNMADD_V" # fvti.ScalarSuffix # "_" # suffix)
                  fvti.RegClass:$rd, fvti.ScalarRegClass:$rs1, fvti.RegClass:$rs2,
                  fvti.AVL, fvti.SEW)>;
   def : Pat<(fvti.Vector (fma (fneg (splat_vector fvti.ScalarRegClass:$rs1)),
                               fvti.RegClass:$rd, fvti.RegClass:$rs2)),
-            (!cast<Instruction>("PseudoVFNMSUB_V" # fvti.ScalarSuffix # "_" # fvti.LMul.MX)
+            (!cast<Instruction>("PseudoVFNMSUB_V" # fvti.ScalarSuffix # "_" # suffix)
                  fvti.RegClass:$rd, fvti.ScalarRegClass:$rs1, fvti.RegClass:$rs2,
                  fvti.AVL, fvti.SEW)>;
 }
 
 foreach vti = AllFloatVectors in {
   // 14.8. Vector Floating-Point Square-Root Instruction
   def : Pat<(fsqrt (vti.Vector vti.RegClass:$rs2)),
             (!cast<Instruction>("PseudoVFSQRT_V_"# vti.LMul.MX)

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

 ; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py
 ; RUN: llc -mtriple=riscv32 -mattr=+d,+experimental-zfh,+experimental-v -target-abi=ilp32d \
 ; RUN:     -verify-machineinstrs < %s | FileCheck %s
 ; RUN: llc -mtriple=riscv64 -mattr=+d,+experimental-zfh,+experimental-v -target-abi=lp64d \
 ; RUN:     -verify-machineinstrs < %s | FileCheck %s
 
 ; This tests a mix of vfmacc and vfmadd by using different operand orders to
 ; trigger commuting in TwoAddressInstructionPass.
 
 declare <vscale x 1 x half> @llvm.fma.v1f16(<vscale x 1 x half>, <vscale x 1 x half>, <vscale x 1 x half>)
 
 define <vscale x 1 x half> @vfmadd_vv_nxv1f16(<vscale x 1 x half> %va, <vscale x 1 x half> %vb, <vscale x 1 x half> %vc) {
 ; CHECK-LABEL: vfmadd_vv_nxv1f16:
 ; CHECK:       # %bb.0:
-; CHECK-NEXT:    vsetvli a0, zero, e16,mf4,tu,mu
-; CHECK-NEXT:    vfmadd.vv v9, v8, v10
-; CHECK-NEXT:    vmv1r.v v8, v9
+; CHECK-NEXT:    vsetvli a0, zero, e16,mf4,ta,mu
+; CHECK-NEXT:    vfmadd.vv v8, v9, v10
 ; CHECK-NEXT:    ret
   %vd = call <vscale x 1 x half> @llvm.fma.v1f16(<vscale x 1 x half> %va, <vscale x 1 x half> %vb, <vscale x 1 x half> %vc)
   ret <vscale x 1 x half> %vd
 }
 
 define <vscale x 1 x half> @vfmadd_vf_nxv1f16(<vscale x 1 x half> %va, <vscale x 1 x half> %vb, half %c) {
 ; CHECK-LABEL: vfmadd_vf_nxv1f16:
 ; CHECK:       # %bb.0:
-; CHECK-NEXT:    vsetvli a0, zero, e16,mf4,tu,mu
+; CHECK-NEXT:    vsetvli a0, zero, e16,mf4,ta,mu
 ; CHECK-NEXT:    vfmadd.vf v8, fa0, v9
 ; CHECK-NEXT:    ret
   %head = insertelement <vscale x 1 x half> undef, half %c, i32 0
   %splat = shufflevector <vscale x 1 x half> %head, <vscale x 1 x half> undef, <vscale x 1 x i32> zeroinitializer
   %vd = call <vscale x 1 x half> @llvm.fma.v1f16(<vscale x 1 x half> %va, <vscale x 1 x half> %splat, <vscale x 1 x half> %vb)
   ret <vscale x 1 x half> %vd
 }
 
 declare <vscale x 2 x half> @llvm.fma.v2f16(<vscale x 2 x half>, <vscale x 2 x half>, <vscale x 2 x half>)
 
 define <vscale x 2 x half> @vfmadd_vv_nxv2f16(<vscale x 2 x half> %va, <vscale x 2 x half> %vb, <vscale x 2 x half> %vc) {
 ; CHECK-LABEL: vfmadd_vv_nxv2f16:
 ; CHECK:       # %bb.0:
-; CHECK-NEXT:    vsetvli a0, zero, e16,mf2,tu,mu
-; CHECK-NEXT:    vfmadd.vv v10, v8, v9
-; CHECK-NEXT:    vmv1r.v v8, v10
+; CHECK-NEXT:    vsetvli a0, zero, e16,mf2,ta,mu
+; CHECK-NEXT:    vfmadd.vv v8, v10, v9
 ; CHECK-NEXT:    ret
   %vd = call <vscale x 2 x half> @llvm.fma.v2f16(<vscale x 2 x half> %va, <vscale x 2 x half> %vc, <vscale x 2 x half> %vb)
   ret <vscale x 2 x half> %vd
 }
 
 define <vscale x 2 x half> @vfmadd_vf_nxv2f16(<vscale x 2 x half> %va, <vscale x 2 x half> %vb, half %c) {
 ; CHECK-LABEL: vfmadd_vf_nxv2f16:
 ; CHECK:       # %bb.0:
-; CHECK-NEXT:    vsetvli a0, zero, e16,mf2,tu,mu
-; CHECK-NEXT:    vfmadd.vf v9, fa0, v8
-; CHECK-NEXT:    vmv1r.v v8, v9
+; CHECK-NEXT:    vsetvli a0, zero, e16,mf2,ta,mu
+; CHECK-NEXT:    vfmacc.vf v8, fa0, v9
 ; CHECK-NEXT:    ret
   %head = insertelement <vscale x 2 x half> undef, half %c, i32 0
   %splat = shufflevector <vscale x 2 x half> %head, <vscale x 2 x half> undef, <vscale x 2 x i32> zeroinitializer
   %vd = call <vscale x 2 x half> @llvm.fma.v2f16(<vscale x 2 x half> %vb, <vscale x 2 x half> %splat, <vscale x 2 x half> %va)
   ret <vscale x 2 x half> %vd
 }
 
 declare <vscale x 4 x half> @llvm.fma.v4f16(<vscale x 4 x half>, <vscale x 4 x half>, <vscale x 4 x half>)
 
 define <vscale x 4 x half> @vfmadd_vv_nxv4f16(<vscale x 4 x half> %va, <vscale x 4 x half> %vb, <vscale x 4 x half> %vc) {
 ; CHECK-LABEL: vfmadd_vv_nxv4f16:
 ; CHECK:       # %bb.0:
-; CHECK-NEXT:    vsetvli a0, zero, e16,m1,tu,mu
+; CHECK-NEXT:    vsetvli a0, zero, e16,m1,ta,mu
 ; CHECK-NEXT:    vfmadd.vv v8, v9, v10
 ; CHECK-NEXT:    ret
   %vd = call <vscale x 4 x half> @llvm.fma.v4f16(<vscale x 4 x half> %vb, <vscale x 4 x half> %va, <vscale x 4 x half> %vc)
   ret <vscale x 4 x half> %vd
 }
 
 define <vscale x 4 x half> @vfmadd_vf_nxv4f16(<vscale x 4 x half> %va, <vscale x 4 x half> %vb, half %c) {
 ; CHECK-LABEL: vfmadd_vf_nxv4f16:
 ; CHECK:       # %bb.0:
-; CHECK-NEXT:    vsetvli a0, zero, e16,m1,tu,mu
+; CHECK-NEXT:    vsetvli a0, zero, e16,m1,ta,mu
 ; CHECK-NEXT:    vfmadd.vf v8, fa0, v9
 ; CHECK-NEXT:    ret
   %head = insertelement <vscale x 4 x half> undef, half %c, i32 0
   %splat = shufflevector <vscale x 4 x half> %head, <vscale x 4 x half> undef, <vscale x 4 x i32> zeroinitializer
   %vd = call <vscale x 4 x half> @llvm.fma.v4f16(<vscale x 4 x half> %va, <vscale x 4 x half> %splat, <vscale x 4 x half> %vb)
   ret <vscale x 4 x half> %vd
 }
 
 declare <vscale x 8 x half> @llvm.fma.v8f16(<vscale x 8 x half>, <vscale x 8 x half>, <vscale x 8 x half>)
 
 define <vscale x 8 x half> @vfmadd_vv_nxv8f16(<vscale x 8 x half> %va, <vscale x 8 x half> %vb, <vscale x 8 x half> %vc) {
 ; CHECK-LABEL: vfmadd_vv_nxv8f16:
 ; CHECK:       # %bb.0:
-; CHECK-NEXT:    vsetvli a0, zero, e16,m2,tu,mu
-; CHECK-NEXT:    vfmadd.vv v12, v10, v8
-; CHECK-NEXT:    vmv2r.v v8, v12
+; CHECK-NEXT:    vsetvli a0, zero, e16,m2,ta,mu
+; CHECK-NEXT:    vfmacc.vv v8, v12, v10
 ; CHECK-NEXT:    ret
   %vd = call <vscale x 8 x half> @llvm.fma.v8f16(<vscale x 8 x half> %vb, <vscale x 8 x half> %vc, <vscale x 8 x half> %va)
   ret <vscale x 8 x half> %vd
 }
 
 define <vscale x 8 x half> @vfmadd_vf_nxv8f16(<vscale x 8 x half> %va, <vscale x 8 x half> %vb, half %c) {
 ; CHECK-LABEL: vfmadd_vf_nxv8f16:
 ; CHECK:       # %bb.0:
-; CHECK-NEXT:    vsetvli a0, zero, e16,m2,tu,mu
-; CHECK-NEXT:    vfmadd.vf v10, fa0, v8
-; CHECK-NEXT:    vmv2r.v v8, v10
+; CHECK-NEXT:    vsetvli a0, zero, e16,m2,ta,mu
+; CHECK-NEXT:    vfmacc.vf v8, fa0, v10
 ; CHECK-NEXT:    ret
   %head = insertelement <vscale x 8 x half> undef, half %c, i32 0
   %splat = shufflevector <vscale x 8 x half> %head, <vscale x 8 x half> undef, <vscale x 8 x i32> zeroinitializer
   %vd = call <vscale x 8 x half> @llvm.fma.v8f16(<vscale x 8 x half> %vb, <vscale x 8 x half> %splat, <vscale x 8 x half> %va)
   ret <vscale x 8 x half> %vd
 }
 
 declare <vscale x 16 x half> @llvm.fma.v16f16(<vscale x 16 x half>, <vscale x 16 x half>, <vscale x 16 x half>)
 
 define <vscale x 16 x half> @vfmadd_vv_nxv16f16(<vscale x 16 x half> %va, <vscale x 16 x half> %vb, <vscale x 16 x half> %vc) {
 ; CHECK-LABEL: vfmadd_vv_nxv16f16:
 ; CHECK:       # %bb.0:
-; CHECK-NEXT:    vsetvli a0, zero, e16,m4,tu,mu
+; CHECK-NEXT:    vsetvli a0, zero, e16,m4,ta,mu
 ; CHECK-NEXT:    vfmadd.vv v8, v16, v12
 ; CHECK-NEXT:    ret
   %vd = call <vscale x 16 x half> @llvm.fma.v16f16(<vscale x 16 x half> %vc, <vscale x 16 x half> %va, <vscale x 16 x half> %vb)
   ret <vscale x 16 x half> %vd
 }
 
 define <vscale x 16 x half> @vfmadd_vf_nxv16f16(<vscale x 16 x half> %va, <vscale x 16 x half> %vb, half %c) {
 ; CHECK-LABEL: vfmadd_vf_nxv16f16:
 ; CHECK:       # %bb.0:
-; CHECK-NEXT:    vsetvli a0, zero, e16,m4,tu,mu
+; CHECK-NEXT:    vsetvli a0, zero, e16,m4,ta,mu
 ; CHECK-NEXT:    vfmadd.vf v8, fa0, v12
 ; CHECK-NEXT:    ret
   %head = insertelement <vscale x 16 x half> undef, half %c, i32 0
   %splat = shufflevector <vscale x 16 x half> %head, <vscale x 16 x half> undef, <vscale x 16 x i32> zeroinitializer
   %vd = call <vscale x 16 x half> @llvm.fma.v16f16(<vscale x 16 x half> %va, <vscale x 16 x half> %splat, <vscale x 16 x half> %vb)
   ret <vscale x 16 x half> %vd
 }
 
 declare <vscale x 32 x half> @llvm.fma.v32f16(<vscale x 32 x half>, <vscale x 32 x half>, <vscale x 32 x half>)
 
 define <vscale x 32 x half> @vfmadd_vv_nxv32f16(<vscale x 32 x half> %va, <vscale x 32 x half> %vb, <vscale x 32 x half> %vc) {
 ; CHECK-LABEL: vfmadd_vv_nxv32f16:
 ; CHECK:       # %bb.0:
 ; CHECK-NEXT:    vsetvli a1, zero, e16,m8,ta,mu
 ; CHECK-NEXT:    vle16.v v24, (a0)
-; CHECK-NEXT:    vsetvli a0, zero, e16,m8,tu,mu
-; CHECK-NEXT:    vfmadd.vv v16, v24, v8
-; CHECK-NEXT:    vmv8r.v v8, v16
+; CHECK-NEXT:    vfmacc.vv v8, v16, v24
 ; CHECK-NEXT:    ret
   %vd = call <vscale x 32 x half> @llvm.fma.v32f16(<vscale x 32 x half> %vc, <vscale x 32 x half> %vb, <vscale x 32 x half> %va)
   ret <vscale x 32 x half> %vd
 }
 
 define <vscale x 32 x half> @vfmadd_vf_nxv32f16(<vscale x 32 x half> %va, <vscale x 32 x half> %vb, half %c) {
 ; CHECK-LABEL: vfmadd_vf_nxv32f16:
 ; CHECK:       # %bb.0:
-; CHECK-NEXT:    vsetvli a0, zero, e16,m8,tu,mu
-; CHECK-NEXT:    vfmadd.vf v16, fa0, v8
-; CHECK-NEXT:    vmv8r.v v8, v16
+; CHECK-NEXT:    vsetvli a0, zero, e16,m8,ta,mu
+; CHECK-NEXT:    vfmacc.vf v8, fa0, v16
 ; CHECK-NEXT:    ret
   %head = insertelement <vscale x 32 x half> undef, half %c, i32 0
   %splat = shufflevector <vscale x 32 x half> %head, <vscale x 32 x half> undef, <vscale x 32 x i32> zeroinitializer
   %vd = call <vscale x 32 x half> @llvm.fma.v32f16(<vscale x 32 x half> %vb, <vscale x 32 x half> %splat, <vscale x 32 x half> %va)
   ret <vscale x 32 x half> %vd
 }
 
 declare <vscale x 1 x float> @llvm.fma.v1f32(<vscale x 1 x float>, <vscale x 1 x float>, <vscale x 1 x float>)
 
 define <vscale x 1 x float> @vfmadd_vv_nxv1f32(<vscale x 1 x float> %va, <vscale x 1 x float> %vb, <vscale x 1 x float> %vc) {
 ; CHECK-LABEL: vfmadd_vv_nxv1f32:
 ; CHECK:       # %bb.0:
-; CHECK-NEXT:    vsetvli a0, zero, e32,mf2,tu,mu
-; CHECK-NEXT:    vfmadd.vv v9, v8, v10
-; CHECK-NEXT:    vmv1r.v v8, v9
+; CHECK-NEXT:    vsetvli a0, zero, e32,mf2,ta,mu
+; CHECK-NEXT:    vfmadd.vv v8, v9, v10
 ; CHECK-NEXT:    ret
   %vd = call <vscale x 1 x float> @llvm.fma.v1f32(<vscale x 1 x float> %va, <vscale x 1 x float> %vb, <vscale x 1 x float> %vc)
   ret <vscale x 1 x float> %vd
 }
 
 define <vscale x 1 x float> @vfmadd_vf_nxv1f32(<vscale x 1 x float> %va, <vscale x 1 x float> %vb, float %c) {
 ; CHECK-LABEL: vfmadd_vf_nxv1f32:
 ; CHECK:       # %bb.0:
-; CHECK-NEXT:    vsetvli a0, zero, e32,mf2,tu,mu
+; CHECK-NEXT:    vsetvli a0, zero, e32,mf2,ta,mu
 ; CHECK-NEXT:    vfmadd.vf v8, fa0, v9
 ; CHECK-NEXT:    ret
   %head = insertelement <vscale x 1 x float> undef, float %c, i32 0
   %splat = shufflevector <vscale x 1 x float> %head, <vscale x 1 x float> undef, <vscale x 1 x i32> zeroinitializer
   %vd = call <vscale x 1 x float> @llvm.fma.v1f32(<vscale x 1 x float> %va, <vscale x 1 x float> %splat, <vscale x 1 x float> %vb)
   ret <vscale x 1 x float> %vd
 }
 
 declare <vscale x 2 x float> @llvm.fma.v2f32(<vscale x 2 x float>, <vscale x 2 x float>, <vscale x 2 x float>)
 
 define <vscale x 2 x float> @vfmadd_vv_nxv2f32(<vscale x 2 x float> %va, <vscale x 2 x float> %vb, <vscale x 2 x float> %vc) {
 ; CHECK-LABEL: vfmadd_vv_nxv2f32:
 ; CHECK:       # %bb.0:
-; CHECK-NEXT:    vsetvli a0, zero, e32,m1,tu,mu
-; CHECK-NEXT:    vfmadd.vv v10, v8, v9
-; CHECK-NEXT:    vmv1r.v v8, v10
+; CHECK-NEXT:    vsetvli a0, zero, e32,m1,ta,mu
+; CHECK-NEXT:    vfmadd.vv v8, v10, v9
 ; CHECK-NEXT:    ret
   %vd = call <vscale x 2 x float> @llvm.fma.v2f32(<vscale x 2 x float> %va, <vscale x 2 x float> %vc, <vscale x 2 x float> %vb)
   ret <vscale x 2 x float> %vd
 }
 
 define <vscale x 2 x float> @vfmadd_vf_nxv2f32(<vscale x 2 x float> %va, <vscale x 2 x float> %vb, float %c) {
 ; CHECK-LABEL: vfmadd_vf_nxv2f32:
 ; CHECK:       # %bb.0:
-; CHECK-NEXT:    vsetvli a0, zero, e32,m1,tu,mu
-; CHECK-NEXT:    vfmadd.vf v9, fa0, v8
-; CHECK-NEXT:    vmv1r.v v8, v9
+; CHECK-NEXT:    vsetvli a0, zero, e32,m1,ta,mu
+; CHECK-NEXT:    vfmacc.vf v8, fa0, v9
 ; CHECK-NEXT:    ret
   %head = insertelement <vscale x 2 x float> undef, float %c, i32 0
   %splat = shufflevector <vscale x 2 x float> %head, <vscale x 2 x float> undef, <vscale x 2 x i32> zeroinitializer
   %vd = call <vscale x 2 x float> @llvm.fma.v2f32(<vscale x 2 x float> %vb, <vscale x 2 x float> %splat, <vscale x 2 x float> %va)
   ret <vscale x 2 x float> %vd
 }
 
 declare <vscale x 4 x float> @llvm.fma.v4f32(<vscale x 4 x float>, <vscale x 4 x float>, <vscale x 4 x float>)
 
 define <vscale x 4 x float> @vfmadd_vv_nxv4f32(<vscale x 4 x float> %va, <vscale x 4 x float> %vb, <vscale x 4 x float> %vc) {
 ; CHECK-LABEL: vfmadd_vv_nxv4f32:
 ; CHECK:       # %bb.0:
-; CHECK-NEXT:    vsetvli a0, zero, e32,m2,tu,mu
+; CHECK-NEXT:    vsetvli a0, zero, e32,m2,ta,mu
 ; CHECK-NEXT:    vfmadd.vv v8, v10, v12
 ; CHECK-NEXT:    ret
   %vd = call <vscale x 4 x float> @llvm.fma.v4f32(<vscale x 4 x float> %vb, <vscale x 4 x float> %va, <vscale x 4 x float> %vc)
   ret <vscale x 4 x float> %vd
 }
 
 define <vscale x 4 x float> @vfmadd_vf_nxv4f32(<vscale x 4 x float> %va, <vscale x 4 x float> %vb, float %c) {
 ; CHECK-LABEL: vfmadd_vf_nxv4f32:
 ; CHECK:       # %bb.0:
-; CHECK-NEXT:    vsetvli a0, zero, e32,m2,tu,mu
+; CHECK-NEXT:    vsetvli a0, zero, e32,m2,ta,mu
 ; CHECK-NEXT:    vfmadd.vf v8, fa0, v10
 ; CHECK-NEXT:    ret
   %head = insertelement <vscale x 4 x float> undef, float %c, i32 0
   %splat = shufflevector <vscale x 4 x float> %head, <vscale x 4 x float> undef, <vscale x 4 x i32> zeroinitializer
   %vd = call <vscale x 4 x float> @llvm.fma.v4f32(<vscale x 4 x float> %va, <vscale x 4 x float> %splat, <vscale x 4 x float> %vb)
   ret <vscale x 4 x float> %vd
 }
 
 declare <vscale x 8 x float> @llvm.fma.v8f32(<vscale x 8 x float>, <vscale x 8 x float>, <vscale x 8 x float>)
 
 define <vscale x 8 x float> @vfmadd_vv_nxv8f32(<vscale x 8 x float> %va, <vscale x 8 x float> %vb, <vscale x 8 x float> %vc) {
 ; CHECK-LABEL: vfmadd_vv_nxv8f32:
 ; CHECK:       # %bb.0:
-; CHECK-NEXT:    vsetvli a0, zero, e32,m4,tu,mu
-; CHECK-NEXT:    vfmadd.vv v16, v12, v8
-; CHECK-NEXT:    vmv4r.v v8, v16
+; CHECK-NEXT:    vsetvli a0, zero, e32,m4,ta,mu
+; CHECK-NEXT:    vfmacc.vv v8, v16, v12
 ; CHECK-NEXT:    ret
   %vd = call <vscale x 8 x float> @llvm.fma.v8f32(<vscale x 8 x float> %vb, <vscale x 8 x float> %vc, <vscale x 8 x float> %va)
   ret <vscale x 8 x float> %vd
 }
 
 define <vscale x 8 x float> @vfmadd_vf_nxv8f32(<vscale x 8 x float> %va, <vscale x 8 x float> %vb, float %c) {
 ; CHECK-LABEL: vfmadd_vf_nxv8f32:
 ; CHECK:       # %bb.0:
-; CHECK-NEXT:    vsetvli a0, zero, e32,m4,tu,mu
-; CHECK-NEXT:    vfmadd.vf v12, fa0, v8
-; CHECK-NEXT:    vmv4r.v v8, v12
+; CHECK-NEXT:    vsetvli a0, zero, e32,m4,ta,mu
+; CHECK-NEXT:    vfmacc.vf v8, fa0, v12
 ; CHECK-NEXT:    ret
   %head = insertelement <vscale x 8 x float> undef, float %c, i32 0
   %splat = shufflevector <vscale x 8 x float> %head, <vscale x 8 x float> undef, <vscale x 8 x i32> zeroinitializer
   %vd = call <vscale x 8 x float> @llvm.fma.v8f32(<vscale x 8 x float> %vb, <vscale x 8 x float> %splat, <vscale x 8 x float> %va)
   ret <vscale x 8 x float> %vd
 }
 
 declare <vscale x 16 x float> @llvm.fma.v16f32(<vscale x 16 x float>, <vscale x 16 x float>, <vscale x 16 x float>)
 
 define <vscale x 16 x float> @vfmadd_vv_nxv16f32(<vscale x 16 x float> %va, <vscale x 16 x float> %vb, <vscale x 16 x float> %vc) {
 ; CHECK-LABEL: vfmadd_vv_nxv16f32:
 ; CHECK:       # %bb.0:
 ; CHECK-NEXT:    vsetvli a1, zero, e32,m8,ta,mu
 ; CHECK-NEXT:    vle32.v v24, (a0)
-; CHECK-NEXT:    vsetvli a0, zero, e32,m8,tu,mu
 ; CHECK-NEXT:    vfmadd.vv v8, v24, v16
 ; CHECK-NEXT:    ret
   %vd = call <vscale x 16 x float> @llvm.fma.v16f32(<vscale x 16 x float> %vc, <vscale x 16 x float> %va, <vscale x 16 x float> %vb)
   ret <vscale x 16 x float> %vd
 }
 
 define <vscale x 16 x float> @vfmadd_vf_nxv16f32(<vscale x 16 x float> %va, <vscale x 16 x float> %vb, float %c) {
 ; CHECK-LABEL: vfmadd_vf_nxv16f32:
 ; CHECK:       # %bb.0:
-; CHECK-NEXT:    vsetvli a0, zero, e32,m8,tu,mu
+; CHECK-NEXT:    vsetvli a0, zero, e32,m8,ta,mu
 ; CHECK-NEXT:    vfmadd.vf v8, fa0, v16
 ; CHECK-NEXT:    ret
   %head = insertelement <vscale x 16 x float> undef, float %c, i32 0
   %splat = shufflevector <vscale x 16 x float> %head, <vscale x 16 x float> undef, <vscale x 16 x i32> zeroinitializer
   %vd = call <vscale x 16 x float> @llvm.fma.v16f32(<vscale x 16 x float> %va, <vscale x 16 x float> %splat, <vscale x 16 x float> %vb)
   ret <vscale x 16 x float> %vd
 }
 
 declare <vscale x 1 x double> @llvm.fma.v1f64(<vscale x 1 x double>, <vscale x 1 x double>, <vscale x 1 x double>)
 
 define <vscale x 1 x double> @vfmadd_vv_nxv1f64(<vscale x 1 x double> %va, <vscale x 1 x double> %vb, <vscale x 1 x double> %vc) {
 ; CHECK-LABEL: vfmadd_vv_nxv1f64:
 ; CHECK:       # %bb.0:
-; CHECK-NEXT:    vsetvli a0, zero, e64,m1,tu,mu
-; CHECK-NEXT:    vfmadd.vv v9, v8, v10
-; CHECK-NEXT:    vmv1r.v v8, v9
+; CHECK-NEXT:    vsetvli a0, zero, e64,m1,ta,mu
+; CHECK-NEXT:    vfmadd.vv v8, v9, v10
 ; CHECK-NEXT:    ret
   %vd = call <vscale x 1 x double> @llvm.fma.v1f64(<vscale x 1 x double> %va, <vscale x 1 x double> %vb, <vscale x 1 x double> %vc)
   ret <vscale x 1 x double> %vd
 }
 
 define <vscale x 1 x double> @vfmadd_vf_nxv1f64(<vscale x 1 x double> %va, <vscale x 1 x double> %vb, double %c) {
 ; CHECK-LABEL: vfmadd_vf_nxv1f64:
 ; CHECK:       # %bb.0:
-; CHECK-NEXT:    vsetvli a0, zero, e64,m1,tu,mu
+; CHECK-NEXT:    vsetvli a0, zero, e64,m1,ta,mu
 ; CHECK-NEXT:    vfmadd.vf v8, fa0, v9
 ; CHECK-NEXT:    ret
   %head = insertelement <vscale x 1 x double> undef, double %c, i32 0
   %splat = shufflevector <vscale x 1 x double> %head, <vscale x 1 x double> undef, <vscale x 1 x i32> zeroinitializer
   %vd = call <vscale x 1 x double> @llvm.fma.v1f64(<vscale x 1 x double> %va, <vscale x 1 x double> %splat, <vscale x 1 x double> %vb)
   ret <vscale x 1 x double> %vd
 }
 
 declare <vscale x 2 x double> @llvm.fma.v2f64(<vscale x 2 x double>, <vscale x 2 x double>, <vscale x 2 x double>)
 
 define <vscale x 2 x double> @vfmadd_vv_nxv2f64(<vscale x 2 x double> %va, <vscale x 2 x double> %vb, <vscale x 2 x double> %vc) {
 ; CHECK-LABEL: vfmadd_vv_nxv2f64:
 ; CHECK:       # %bb.0:
-; CHECK-NEXT:    vsetvli a0, zero, e64,m2,tu,mu
-; CHECK-NEXT:    vfmadd.vv v12, v8, v10
-; CHECK-NEXT:    vmv2r.v v8, v12
+; CHECK-NEXT:    vsetvli a0, zero, e64,m2,ta,mu
+; CHECK-NEXT:    vfmadd.vv v8, v12, v10
 ; CHECK-NEXT:    ret
   %vd = call <vscale x 2 x double> @llvm.fma.v2f64(<vscale x 2 x double> %va, <vscale x 2 x double> %vc, <vscale x 2 x double> %vb)
   ret <vscale x 2 x double> %vd
 }
 
 define <vscale x 2 x double> @vfmadd_vf_nxv2f64(<vscale x 2 x double> %va, <vscale x 2 x double> %vb, double %c) {
 ; CHECK-LABEL: vfmadd_vf_nxv2f64:
 ; CHECK:       # %bb.0:
-; CHECK-NEXT:    vsetvli a0, zero, e64,m2,tu,mu
-; CHECK-NEXT:    vfmadd.vf v10, fa0, v8
-; CHECK-NEXT:    vmv2r.v v8, v10
+; CHECK-NEXT:    vsetvli a0, zero, e64,m2,ta,mu
+; CHECK-NEXT:    vfmacc.vf v8, fa0, v10
 ; CHECK-NEXT:    ret
   %head = insertelement <vscale x 2 x double> undef, double %c, i32 0
   %splat = shufflevector <vscale x 2 x double> %head, <vscale x 2 x double> undef, <vscale x 2 x i32> zeroinitializer
   %vd = call <vscale x 2 x double> @llvm.fma.v2f64(<vscale x 2 x double> %vb, <vscale x 2 x double> %splat, <vscale x 2 x double> %va)
   ret <vscale x 2 x double> %vd
 }
 
 declare <vscale x 4 x double> @llvm.fma.v4f64(<vscale x 4 x double>, <vscale x 4 x double>, <vscale x 4 x double>)
 
 define <vscale x 4 x double> @vfmadd_vv_nxv4f64(<vscale x 4 x double> %va, <vscale x 4 x double> %vb, <vscale x 4 x double> %vc) {
 ; CHECK-LABEL: vfmadd_vv_nxv4f64:
 ; CHECK:       # %bb.0:
-; CHECK-NEXT:    vsetvli a0, zero, e64,m4,tu,mu
+; CHECK-NEXT:    vsetvli a0, zero, e64,m4,ta,mu
 ; CHECK-NEXT:    vfmadd.vv v8, v12, v16
 ; CHECK-NEXT:    ret
   %vd = call <vscale x 4 x double> @llvm.fma.v4f64(<vscale x 4 x double> %vb, <vscale x 4 x double> %va, <vscale x 4 x double> %vc)
   ret <vscale x 4 x double> %vd
 }
 
 define <vscale x 4 x double> @vfmadd_vf_nxv4f64(<vscale x 4 x double> %va, <vscale x 4 x double> %vb, double %c) {
 ; CHECK-LABEL: vfmadd_vf_nxv4f64:
 ; CHECK:       # %bb.0:
-; CHECK-NEXT:    vsetvli a0, zero, e64,m4,tu,mu
+; CHECK-NEXT:    vsetvli a0, zero, e64,m4,ta,mu
 ; CHECK-NEXT:    vfmadd.vf v8, fa0, v12
 ; CHECK-NEXT:    ret
   %head = insertelement <vscale x 4 x double> undef, double %c, i32 0
   %splat = shufflevector <vscale x 4 x double> %head, <vscale x 4 x double> undef, <vscale x 4 x i32> zeroinitializer
   %vd = call <vscale x 4 x double> @llvm.fma.v4f64(<vscale x 4 x double> %va, <vscale x 4 x double> %splat, <vscale x 4 x double> %vb)
   ret <vscale x 4 x double> %vd
 }
 
 declare <vscale x 8 x double> @llvm.fma.v8f64(<vscale x 8 x double>, <vscale x 8 x double>, <vscale x 8 x double>)
 
 define <vscale x 8 x double> @vfmadd_vv_nxv8f64(<vscale x 8 x double> %va, <vscale x 8 x double> %vb, <vscale x 8 x double> %vc) {
 ; CHECK-LABEL: vfmadd_vv_nxv8f64:
 ; CHECK:       # %bb.0:
 ; CHECK-NEXT:    vsetvli a1, zero, e64,m8,ta,mu
 ; CHECK-NEXT:    vle64.v v24, (a0)
-; CHECK-NEXT:    vsetvli a0, zero, e64,m8,tu,mu
-; CHECK-NEXT:    vfmadd.vv v24, v16, v8
-; CHECK-NEXT:    vmv8r.v v8, v24
+; CHECK-NEXT:    vfmacc.vv v8, v16, v24
 ; CHECK-NEXT:    ret
   %vd = call <vscale x 8 x double> @llvm.fma.v8f64(<vscale x 8 x double> %vb, <vscale x 8 x double> %vc, <vscale x 8 x double> %va)
   ret <vscale x 8 x double> %vd
 }
 
 define <vscale x 8 x double> @vfmadd_vf_nxv8f64(<vscale x 8 x double> %va, <vscale x 8 x double> %vb, double %c) {
 ; CHECK-LABEL: vfmadd_vf_nxv8f64:
 ; CHECK:       # %bb.0:
-; CHECK-NEXT:    vsetvli a0, zero, e64,m8,tu,mu
-; CHECK-NEXT:    vfmadd.vf v16, fa0, v8
-; CHECK-NEXT:    vmv8r.v v8, v16
+; CHECK-NEXT:    vsetvli a0, zero, e64,m8,ta,mu
+; CHECK-NEXT:    vfmacc.vf v8, fa0, v16
 ; CHECK-NEXT:    ret
   %head = insertelement <vscale x 8 x double> undef, double %c, i32 0
   %splat = shufflevector <vscale x 8 x double> %head, <vscale x 8 x double> undef, <vscale x 8 x i32> zeroinitializer
   %vd = call <vscale x 8 x double> @llvm.fma.v8f64(<vscale x 8 x double> %vb, <vscale x 8 x double> %splat, <vscale x 8 x double> %va)
   ret <vscale x 8 x double> %vd
 }

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

 ; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py
 ; RUN: llc -mtriple=riscv32 -mattr=+d,+experimental-zfh,+experimental-v -target-abi=ilp32d \
 ; RUN:     -verify-machineinstrs < %s | FileCheck %s
 ; RUN: llc -mtriple=riscv64 -mattr=+d,+experimental-zfh,+experimental-v -target-abi=lp64d \
 ; RUN:     -verify-machineinstrs < %s | FileCheck %s
 
 ; This tests a mix of vfmsac and vfmsub by using different operand orders to
 ; trigger commuting in TwoAddressInstructionPass.
 
 declare <vscale x 1 x half> @llvm.fma.v1f16(<vscale x 1 x half>, <vscale x 1 x half>, <vscale x 1 x half>)
 
 define <vscale x 1 x half> @vfmsub_vv_nxv1f16(<vscale x 1 x half> %va, <vscale x 1 x half> %vb, <vscale x 1 x half> %vc) {
 ; CHECK-LABEL: vfmsub_vv_nxv1f16:
 ; CHECK:       # %bb.0:
-; CHECK-NEXT:    vsetvli a0, zero, e16,mf4,tu,mu
-; CHECK-NEXT:    vfmsub.vv v9, v8, v10
-; CHECK-NEXT:    vmv1r.v v8, v9
+; CHECK-NEXT:    vsetvli a0, zero, e16,mf4,ta,mu
+; CHECK-NEXT:    vfmsub.vv v8, v9, v10
 ; CHECK-NEXT:    ret
   %neg = fneg <vscale x 1 x half> %vc
   %vd = call <vscale x 1 x half> @llvm.fma.v1f16(<vscale x 1 x half> %va, <vscale x 1 x half> %vb, <vscale x 1 x half> %neg)
   ret <vscale x 1 x half> %vd
 }
 
 define <vscale x 1 x half> @vfmsub_vf_nxv1f16(<vscale x 1 x half> %va, <vscale x 1 x half> %vb, half %c) {
 ; CHECK-LABEL: vfmsub_vf_nxv1f16:
 ; CHECK:       # %bb.0:
-; CHECK-NEXT:    vsetvli a0, zero, e16,mf4,tu,mu
+; CHECK-NEXT:    vsetvli a0, zero, e16,mf4,ta,mu
 ; CHECK-NEXT:    vfmsub.vf v8, fa0, v9
 ; CHECK-NEXT:    ret
   %head = insertelement <vscale x 1 x half> undef, half %c, i32 0
   %splat = shufflevector <vscale x 1 x half> %head, <vscale x 1 x half> undef, <vscale x 1 x i32> zeroinitializer
   %neg = fneg <vscale x 1 x half> %vb
   %vd = call <vscale x 1 x half> @llvm.fma.v1f16(<vscale x 1 x half> %va, <vscale x 1 x half> %splat, <vscale x 1 x half> %neg)
   ret <vscale x 1 x half> %vd
 }
 
 declare <vscale x 2 x half> @llvm.fma.v2f16(<vscale x 2 x half>, <vscale x 2 x half>, <vscale x 2 x half>)
 
 define <vscale x 2 x half> @vfmsub_vv_nxv2f16(<vscale x 2 x half> %va, <vscale x 2 x half> %vb, <vscale x 2 x half> %vc) {
 ; CHECK-LABEL: vfmsub_vv_nxv2f16:
 ; CHECK:       # %bb.0:
-; CHECK-NEXT:    vsetvli a0, zero, e16,mf2,tu,mu
-; CHECK-NEXT:    vfmsub.vv v10, v8, v9
-; CHECK-NEXT:    vmv1r.v v8, v10
+; CHECK-NEXT:    vsetvli a0, zero, e16,mf2,ta,mu
+; CHECK-NEXT:    vfmsub.vv v8, v10, v9
 ; CHECK-NEXT:    ret
   %neg = fneg <vscale x 2 x half> %vb
   %vd = call <vscale x 2 x half> @llvm.fma.v2f16(<vscale x 2 x half> %va, <vscale x 2 x half> %vc, <vscale x 2 x half> %neg)
   ret <vscale x 2 x half> %vd
 }
 
 define <vscale x 2 x half> @vfmsub_vf_nxv2f16(<vscale x 2 x half> %va, <vscale x 2 x half> %vb, half %c) {
 ; CHECK-LABEL: vfmsub_vf_nxv2f16:
 ; CHECK:       # %bb.0:
-; CHECK-NEXT:    vsetvli a0, zero, e16,mf2,tu,mu
-; CHECK-NEXT:    vfmsub.vf v9, fa0, v8
-; CHECK-NEXT:    vmv1r.v v8, v9
+; CHECK-NEXT:    vsetvli a0, zero, e16,mf2,ta,mu
+; CHECK-NEXT:    vfmsac.vf v8, fa0, v9
 ; CHECK-NEXT:    ret
   %head = insertelement <vscale x 2 x half> undef, half %c, i32 0
   %splat = shufflevector <vscale x 2 x half> %head, <vscale x 2 x half> undef, <vscale x 2 x i32> zeroinitializer
   %neg = fneg <vscale x 2 x half> %va
   %vd = call <vscale x 2 x half> @llvm.fma.v2f16(<vscale x 2 x half> %vb, <vscale x 2 x half> %splat, <vscale x 2 x half> %neg)
   ret <vscale x 2 x half> %vd
 }
 
 declare <vscale x 4 x half> @llvm.fma.v4f16(<vscale x 4 x half>, <vscale x 4 x half>, <vscale x 4 x half>)
 
 define <vscale x 4 x half> @vfmsub_vv_nxv4f16(<vscale x 4 x half> %va, <vscale x 4 x half> %vb, <vscale x 4 x half> %vc) {
 ; CHECK-LABEL: vfmsub_vv_nxv4f16:
 ; CHECK:       # %bb.0:
-; CHECK-NEXT:    vsetvli a0, zero, e16,m1,tu,mu
+; CHECK-NEXT:    vsetvli a0, zero, e16,m1,ta,mu
 ; CHECK-NEXT:    vfmsub.vv v8, v9, v10
 ; CHECK-NEXT:    ret
   %neg = fneg <vscale x 4 x half> %vc
   %vd = call <vscale x 4 x half> @llvm.fma.v4f16(<vscale x 4 x half> %vb, <vscale x 4 x half> %va, <vscale x 4 x half> %neg)
   ret <vscale x 4 x half> %vd
 }
 
 define <vscale x 4 x half> @vfmsub_vf_nxv4f16(<vscale x 4 x half> %va, <vscale x 4 x half> %vb, half %c) {
 ; CHECK-LABEL: vfmsub_vf_nxv4f16:
 ; CHECK:       # %bb.0:
-; CHECK-NEXT:    vsetvli a0, zero, e16,m1,tu,mu
+; CHECK-NEXT:    vsetvli a0, zero, e16,m1,ta,mu
 ; CHECK-NEXT:    vfmsub.vf v8, fa0, v9
 ; CHECK-NEXT:    ret
   %head = insertelement <vscale x 4 x half> undef, half %c, i32 0
   %splat = shufflevector <vscale x 4 x half> %head, <vscale x 4 x half> undef, <vscale x 4 x i32> zeroinitializer
   %neg = fneg <vscale x 4 x half> %vb
   %vd = call <vscale x 4 x half> @llvm.fma.v4f16(<vscale x 4 x half> %va, <vscale x 4 x half> %splat, <vscale x 4 x half> %neg)
   ret <vscale x 4 x half> %vd
 }
 
 declare <vscale x 8 x half> @llvm.fma.v8f16(<vscale x 8 x half>, <vscale x 8 x half>, <vscale x 8 x half>)
 
 define <vscale x 8 x half> @vfmsub_vv_nxv8f16(<vscale x 8 x half> %va, <vscale x 8 x half> %vb, <vscale x 8 x half> %vc) {
 ; CHECK-LABEL: vfmsub_vv_nxv8f16:
 ; CHECK:       # %bb.0:
-; CHECK-NEXT:    vsetvli a0, zero, e16,m2,tu,mu
-; CHECK-NEXT:    vfmsub.vv v12, v10, v8
-; CHECK-NEXT:    vmv2r.v v8, v12
+; CHECK-NEXT:    vsetvli a0, zero, e16,m2,ta,mu
+; CHECK-NEXT:    vfmsac.vv v8, v12, v10
 ; CHECK-NEXT:    ret
   %neg = fneg <vscale x 8 x half> %va
   %vd = call <vscale x 8 x half> @llvm.fma.v8f16(<vscale x 8 x half> %vb, <vscale x 8 x half> %vc, <vscale x 8 x half> %neg)
   ret <vscale x 8 x half> %vd
 }
 
 define <vscale x 8 x half> @vfmsub_vf_nxv8f16(<vscale x 8 x half> %va, <vscale x 8 x half> %vb, half %c) {
 ; CHECK-LABEL: vfmsub_vf_nxv8f16:
 ; CHECK:       # %bb.0:
-; CHECK-NEXT:    vsetvli a0, zero, e16,m2,tu,mu
-; CHECK-NEXT:    vfmsub.vf v10, fa0, v8
-; CHECK-NEXT:    vmv2r.v v8, v10
+; CHECK-NEXT:    vsetvli a0, zero, e16,m2,ta,mu
+; CHECK-NEXT:    vfmsac.vf v8, fa0, v10
 ; CHECK-NEXT:    ret
   %head = insertelement <vscale x 8 x half> undef, half %c, i32 0
   %splat = shufflevector <vscale x 8 x half> %head, <vscale x 8 x half> undef, <vscale x 8 x i32> zeroinitializer
   %neg = fneg <vscale x 8 x half> %va
   %vd = call <vscale x 8 x half> @llvm.fma.v8f16(<vscale x 8 x half> %vb, <vscale x 8 x half> %splat, <vscale x 8 x half> %neg)
   ret <vscale x 8 x half> %vd
 }
 
 declare <vscale x 16 x half> @llvm.fma.v16f16(<vscale x 16 x half>, <vscale x 16 x half>, <vscale x 16 x half>)
 
 define <vscale x 16 x half> @vfmsub_vv_nxv16f16(<vscale x 16 x half> %va, <vscale x 16 x half> %vb, <vscale x 16 x half> %vc) {
 ; CHECK-LABEL: vfmsub_vv_nxv16f16:
 ; CHECK:       # %bb.0:
-; CHECK-NEXT:    vsetvli a0, zero, e16,m4,tu,mu
+; CHECK-NEXT:    vsetvli a0, zero, e16,m4,ta,mu
 ; CHECK-NEXT:    vfmsub.vv v8, v16, v12
 ; CHECK-NEXT:    ret
   %neg = fneg <vscale x 16 x half> %vb
   %vd = call <vscale x 16 x half> @llvm.fma.v16f16(<vscale x 16 x half> %vc, <vscale x 16 x half> %va, <vscale x 16 x half> %neg)
   ret <vscale x 16 x half> %vd
 }
 
 define <vscale x 16 x half> @vfmsub_vf_nxv16f16(<vscale x 16 x half> %va, <vscale x 16 x half> %vb, half %c) {
 ; CHECK-LABEL: vfmsub_vf_nxv16f16:
 ; CHECK:       # %bb.0:
-; CHECK-NEXT:    vsetvli a0, zero, e16,m4,tu,mu
+; CHECK-NEXT:    vsetvli a0, zero, e16,m4,ta,mu
 ; CHECK-NEXT:    vfmsub.vf v8, fa0, v12
 ; CHECK-NEXT:    ret
   %head = insertelement <vscale x 16 x half> undef, half %c, i32 0
   %splat = shufflevector <vscale x 16 x half> %head, <vscale x 16 x half> undef, <vscale x 16 x i32> zeroinitializer
   %neg = fneg <vscale x 16 x half> %vb
   %vd = call <vscale x 16 x half> @llvm.fma.v16f16(<vscale x 16 x half> %va, <vscale x 16 x half> %splat, <vscale x 16 x half> %neg)
   ret <vscale x 16 x half> %vd
 }
 
 declare <vscale x 32 x half> @llvm.fma.v32f16(<vscale x 32 x half>, <vscale x 32 x half>, <vscale x 32 x half>)
 
 define <vscale x 32 x half> @vfmsub_vv_nxv32f16(<vscale x 32 x half> %va, <vscale x 32 x half> %vb, <vscale x 32 x half> %vc) {
 ; CHECK-LABEL: vfmsub_vv_nxv32f16:
 ; CHECK:       # %bb.0:
 ; CHECK-NEXT:    vsetvli a1, zero, e16,m8,ta,mu
 ; CHECK-NEXT:    vle16.v v24, (a0)
-; CHECK-NEXT:    vsetvli a0, zero, e16,m8,tu,mu
-; CHECK-NEXT:    vfmsub.vv v16, v24, v8
-; CHECK-NEXT:    vmv8r.v v8, v16
+; CHECK-NEXT:    vfmsac.vv v8, v16, v24
 ; CHECK-NEXT:    ret
   %neg = fneg <vscale x 32 x half> %va
   %vd = call <vscale x 32 x half> @llvm.fma.v32f16(<vscale x 32 x half> %vc, <vscale x 32 x half> %vb, <vscale x 32 x half> %neg)
   ret <vscale x 32 x half> %vd
 }
 
 define <vscale x 32 x half> @vfmsub_vf_nxv32f16(<vscale x 32 x half> %va, <vscale x 32 x half> %vb, half %c) {
 ; CHECK-LABEL: vfmsub_vf_nxv32f16:
 ; CHECK:       # %bb.0:
-; CHECK-NEXT:    vsetvli a0, zero, e16,m8,tu,mu
-; CHECK-NEXT:    vfmsub.vf v16, fa0, v8
-; CHECK-NEXT:    vmv8r.v v8, v16
+; CHECK-NEXT:    vsetvli a0, zero, e16,m8,ta,mu
+; CHECK-NEXT:    vfmsac.vf v8, fa0, v16
 ; CHECK-NEXT:    ret
   %head = insertelement <vscale x 32 x half> undef, half %c, i32 0
   %splat = shufflevector <vscale x 32 x half> %head, <vscale x 32 x half> undef, <vscale x 32 x i32> zeroinitializer
   %neg = fneg <vscale x 32 x half> %va
   %vd = call <vscale x 32 x half> @llvm.fma.v32f16(<vscale x 32 x half> %vb, <vscale x 32 x half> %splat, <vscale x 32 x half> %neg)
   ret <vscale x 32 x half> %vd
 }
 
 declare <vscale x 1 x float> @llvm.fma.v1f32(<vscale x 1 x float>, <vscale x 1 x float>, <vscale x 1 x float>)
 
 define <vscale x 1 x float> @vfmsub_vv_nxv1f32(<vscale x 1 x float> %va, <vscale x 1 x float> %vb, <vscale x 1 x float> %vc) {
 ; CHECK-LABEL: vfmsub_vv_nxv1f32:
 ; CHECK:       # %bb.0:
-; CHECK-NEXT:    vsetvli a0, zero, e32,mf2,tu,mu
-; CHECK-NEXT:    vfmsub.vv v9, v8, v10
-; CHECK-NEXT:    vmv1r.v v8, v9
+; CHECK-NEXT:    vsetvli a0, zero, e32,mf2,ta,mu
+; CHECK-NEXT:    vfmsub.vv v8, v9, v10
 ; CHECK-NEXT:    ret
   %neg = fneg <vscale x 1 x float> %vc
   %vd = call <vscale x 1 x float> @llvm.fma.v1f32(<vscale x 1 x float> %va, <vscale x 1 x float> %vb, <vscale x 1 x float> %neg)
   ret <vscale x 1 x float> %vd
 }
 
 define <vscale x 1 x float> @vfmsub_vf_nxv1f32(<vscale x 1 x float> %va, <vscale x 1 x float> %vb, float %c) {
 ; CHECK-LABEL: vfmsub_vf_nxv1f32:
 ; CHECK:       # %bb.0:
-; CHECK-NEXT:    vsetvli a0, zero, e32,mf2,tu,mu
+; CHECK-NEXT:    vsetvli a0, zero, e32,mf2,ta,mu
 ; CHECK-NEXT:    vfmsub.vf v8, fa0, v9
 ; CHECK-NEXT:    ret
   %head = insertelement <vscale x 1 x float> undef, float %c, i32 0
   %splat = shufflevector <vscale x 1 x float> %head, <vscale x 1 x float> undef, <vscale x 1 x i32> zeroinitializer
   %neg = fneg <vscale x 1 x float> %vb
   %vd = call <vscale x 1 x float> @llvm.fma.v1f32(<vscale x 1 x float> %va, <vscale x 1 x float> %splat, <vscale x 1 x float> %neg)
   ret <vscale x 1 x float> %vd
 }
 
 declare <vscale x 2 x float> @llvm.fma.v2f32(<vscale x 2 x float>, <vscale x 2 x float>, <vscale x 2 x float>)
 
 define <vscale x 2 x float> @vfmsub_vv_nxv2f32(<vscale x 2 x float> %va, <vscale x 2 x float> %vb, <vscale x 2 x float> %vc) {
 ; CHECK-LABEL: vfmsub_vv_nxv2f32:
 ; CHECK:       # %bb.0:
-; CHECK-NEXT:    vsetvli a0, zero, e32,m1,tu,mu
-; CHECK-NEXT:    vfmsub.vv v10, v8, v9
-; CHECK-NEXT:    vmv1r.v v8, v10
+; CHECK-NEXT:    vsetvli a0, zero, e32,m1,ta,mu
+; CHECK-NEXT:    vfmsub.vv v8, v10, v9
 ; CHECK-NEXT:    ret
   %neg = fneg <vscale x 2 x float> %vb
   %vd = call <vscale x 2 x float> @llvm.fma.v2f32(<vscale x 2 x float> %va, <vscale x 2 x float> %vc, <vscale x 2 x float> %neg)
   ret <vscale x 2 x float> %vd
 }
 
 define <vscale x 2 x float> @vfmsub_vf_nxv2f32(<vscale x 2 x float> %va, <vscale x 2 x float> %vb, float %c) {
 ; CHECK-LABEL: vfmsub_vf_nxv2f32:
 ; CHECK:       # %bb.0:
-; CHECK-NEXT:    vsetvli a0, zero, e32,m1,tu,mu
-; CHECK-NEXT:    vfmsub.vf v9, fa0, v8
-; CHECK-NEXT:    vmv1r.v v8, v9
+; CHECK-NEXT:    vsetvli a0, zero, e32,m1,ta,mu
+; CHECK-NEXT:    vfmsac.vf v8, fa0, v9
 ; CHECK-NEXT:    ret
   %head = insertelement <vscale x 2 x float> undef, float %c, i32 0
   %splat = shufflevector <vscale x 2 x float> %head, <vscale x 2 x float> undef, <vscale x 2 x i32> zeroinitializer
   %neg = fneg <vscale x 2 x float> %va
   %vd = call <vscale x 2 x float> @llvm.fma.v2f32(<vscale x 2 x float> %vb, <vscale x 2 x float> %splat, <vscale x 2 x float> %neg)
   ret <vscale x 2 x float> %vd
 }
 
 declare <vscale x 4 x float> @llvm.fma.v4f32(<vscale x 4 x float>, <vscale x 4 x float>, <vscale x 4 x float>)
 
 define <vscale x 4 x float> @vfmsub_vv_nxv4f32(<vscale x 4 x float> %va, <vscale x 4 x float> %vb, <vscale x 4 x float> %vc) {
 ; CHECK-LABEL: vfmsub_vv_nxv4f32:
 ; CHECK:       # %bb.0:
-; CHECK-NEXT:    vsetvli a0, zero, e32,m2,tu,mu
+; CHECK-NEXT:    vsetvli a0, zero, e32,m2,ta,mu
 ; CHECK-NEXT:    vfmsub.vv v8, v10, v12
 ; CHECK-NEXT:    ret
   %neg = fneg <vscale x 4 x float> %vc
   %vd = call <vscale x 4 x float> @llvm.fma.v4f32(<vscale x 4 x float> %vb, <vscale x 4 x float> %va, <vscale x 4 x float> %neg)
   ret <vscale x 4 x float> %vd
 }
 
 define <vscale x 4 x float> @vfmsub_vf_nxv4f32(<vscale x 4 x float> %va, <vscale x 4 x float> %vb, float %c) {
 ; CHECK-LABEL: vfmsub_vf_nxv4f32:
 ; CHECK:       # %bb.0:
-; CHECK-NEXT:    vsetvli a0, zero, e32,m2,tu,mu
+; CHECK-NEXT:    vsetvli a0, zero, e32,m2,ta,mu
 ; CHECK-NEXT:    vfmsub.vf v8, fa0, v10
 ; CHECK-NEXT:    ret
   %head = insertelement <vscale x 4 x float> undef, float %c, i32 0
   %splat = shufflevector <vscale x 4 x float> %head, <vscale x 4 x float> undef, <vscale x 4 x i32> zeroinitializer
   %neg = fneg <vscale x 4 x float> %vb
   %vd = call <vscale x 4 x float> @llvm.fma.v4f32(<vscale x 4 x float> %va, <vscale x 4 x float> %splat, <vscale x 4 x float> %neg)
   ret <vscale x 4 x float> %vd
 }
 
 declare <vscale x 8 x float> @llvm.fma.v8f32(<vscale x 8 x float>, <vscale x 8 x float>, <vscale x 8 x float>)
 
 define <vscale x 8 x float> @vfmsub_vv_nxv8f32(<vscale x 8 x float> %va, <vscale x 8 x float> %vb, <vscale x 8 x float> %vc) {
 ; CHECK-LABEL: vfmsub_vv_nxv8f32:
 ; CHECK:       # %bb.0:
-; CHECK-NEXT:    vsetvli a0, zero, e32,m4,tu,mu
-; CHECK-NEXT:    vfmsub.vv v16, v12, v8
-; CHECK-NEXT:    vmv4r.v v8, v16
+; CHECK-NEXT:    vsetvli a0, zero, e32,m4,ta,mu
+; CHECK-NEXT:    vfmsac.vv v8, v16, v12
 ; CHECK-NEXT:    ret
   %neg = fneg <vscale x 8 x float> %va
   %vd = call <vscale x 8 x float> @llvm.fma.v8f32(<vscale x 8 x float> %vb, <vscale x 8 x float> %vc, <vscale x 8 x float> %neg)
   ret <vscale x 8 x float> %vd
 }
 
 define <vscale x 8 x float> @vfmsub_vf_nxv8f32(<vscale x 8 x float> %va, <vscale x 8 x float> %vb, float %c) {
 ; CHECK-LABEL: vfmsub_vf_nxv8f32:
 ; CHECK:       # %bb.0:
-; CHECK-NEXT:    vsetvli a0, zero, e32,m4,tu,mu
-; CHECK-NEXT:    vfmsub.vf v12, fa0, v8
-; CHECK-NEXT:    vmv4r.v v8, v12
+; CHECK-NEXT:    vsetvli a0, zero, e32,m4,ta,mu
+; CHECK-NEXT:    vfmsac.vf v8, fa0, v12
 ; CHECK-NEXT:    ret
   %head = insertelement <vscale x 8 x float> undef, float %c, i32 0
   %splat = shufflevector <vscale x 8 x float> %head, <vscale x 8 x float> undef, <vscale x 8 x i32> zeroinitializer
   %neg = fneg <vscale x 8 x float> %va
   %vd = call <vscale x 8 x float> @llvm.fma.v8f32(<vscale x 8 x float> %vb, <vscale x 8 x float> %splat, <vscale x 8 x float> %neg)
   ret <vscale x 8 x float> %vd
 }
 
 declare <vscale x 16 x float> @llvm.fma.v16f32(<vscale x 16 x float>, <vscale x 16 x float>, <vscale x 16 x float>)
 
 define <vscale x 16 x float> @vfmsub_vv_nxv16f32(<vscale x 16 x float> %va, <vscale x 16 x float> %vb, <vscale x 16 x float> %vc) {
 ; CHECK-LABEL: vfmsub_vv_nxv16f32:
 ; CHECK:       # %bb.0:
 ; CHECK-NEXT:    vsetvli a1, zero, e32,m8,ta,mu
 ; CHECK-NEXT:    vle32.v v24, (a0)
-; CHECK-NEXT:    vsetvli a0, zero, e32,m8,tu,mu
 ; CHECK-NEXT:    vfmsub.vv v8, v24, v16
 ; CHECK-NEXT:    ret
   %neg = fneg <vscale x 16 x float> %vb
   %vd = call <vscale x 16 x float> @llvm.fma.v16f32(<vscale x 16 x float> %vc, <vscale x 16 x float> %va, <vscale x 16 x float> %neg)
   ret <vscale x 16 x float> %vd
 }
 
 define <vscale x 16 x float> @vfmsub_vf_nxv16f32(<vscale x 16 x float> %va, <vscale x 16 x float> %vb, float %c) {
 ; CHECK-LABEL: vfmsub_vf_nxv16f32:
 ; CHECK:       # %bb.0:
-; CHECK-NEXT:    vsetvli a0, zero, e32,m8,tu,mu
+; CHECK-NEXT:    vsetvli a0, zero, e32,m8,ta,mu
 ; CHECK-NEXT:    vfmsub.vf v8, fa0, v16
 ; CHECK-NEXT:    ret
   %head = insertelement <vscale x 16 x float> undef, float %c, i32 0
   %splat = shufflevector <vscale x 16 x float> %head, <vscale x 16 x float> undef, <vscale x 16 x i32> zeroinitializer
   %neg = fneg <vscale x 16 x float> %vb
   %vd = call <vscale x 16 x float> @llvm.fma.v16f32(<vscale x 16 x float> %va, <vscale x 16 x float> %splat, <vscale x 16 x float> %neg)
   ret <vscale x 16 x float> %vd
 }
 
 declare <vscale x 1 x double> @llvm.fma.v1f64(<vscale x 1 x double>, <vscale x 1 x double>, <vscale x 1 x double>)
 
 define <vscale x 1 x double> @vfmsub_vv_nxv1f64(<vscale x 1 x double> %va, <vscale x 1 x double> %vb, <vscale x 1 x double> %vc) {
 ; CHECK-LABEL: vfmsub_vv_nxv1f64:
 ; CHECK:       # %bb.0:
-; CHECK-NEXT:    vsetvli a0, zero, e64,m1,tu,mu
-; CHECK-NEXT:    vfmsub.vv v9, v8, v10
-; CHECK-NEXT:    vmv1r.v v8, v9
+; CHECK-NEXT:    vsetvli a0, zero, e64,m1,ta,mu
+; CHECK-NEXT:    vfmsub.vv v8, v9, v10
 ; CHECK-NEXT:    ret
   %neg = fneg <vscale x 1 x double> %vc
   %vd = call <vscale x 1 x double> @llvm.fma.v1f64(<vscale x 1 x double> %va, <vscale x 1 x double> %vb, <vscale x 1 x double> %neg)
   ret <vscale x 1 x double> %vd
 }
 
 define <vscale x 1 x double> @vfmsub_vf_nxv1f64(<vscale x 1 x double> %va, <vscale x 1 x double> %vb, double %c) {
 ; CHECK-LABEL: vfmsub_vf_nxv1f64:
 ; CHECK:       # %bb.0:
-; CHECK-NEXT:    vsetvli a0, zero, e64,m1,tu,mu
+; CHECK-NEXT:    vsetvli a0, zero, e64,m1,ta,mu
 ; CHECK-NEXT:    vfmsub.vf v8, fa0, v9
 ; CHECK-NEXT:    ret
   %head = insertelement <vscale x 1 x double> undef, double %c, i32 0
   %splat = shufflevector <vscale x 1 x double> %head, <vscale x 1 x double> undef, <vscale x 1 x i32> zeroinitializer
   %neg = fneg <vscale x 1 x double> %vb
   %vd = call <vscale x 1 x double> @llvm.fma.v1f64(<vscale x 1 x double> %va, <vscale x 1 x double> %splat, <vscale x 1 x double> %neg)
   ret <vscale x 1 x double> %vd
 }
 
 declare <vscale x 2 x double> @llvm.fma.v2f64(<vscale x 2 x double>, <vscale x 2 x double>, <vscale x 2 x double>)
 
 define <vscale x 2 x double> @vfmsub_vv_nxv2f64(<vscale x 2 x double> %va, <vscale x 2 x double> %vb, <vscale x 2 x double> %vc) {
 ; CHECK-LABEL: vfmsub_vv_nxv2f64:
 ; CHECK:       # %bb.0:
-; CHECK-NEXT:    vsetvli a0, zero, e64,m2,tu,mu
-; CHECK-NEXT:    vfmsub.vv v12, v8, v10
-; CHECK-NEXT:    vmv2r.v v8, v12
+; CHECK-NEXT:    vsetvli a0, zero, e64,m2,ta,mu
+; CHECK-NEXT:    vfmsub.vv v8, v12, v10
 ; CHECK-NEXT:    ret
   %neg = fneg <vscale x 2 x double> %vb
   %vd = call <vscale x 2 x double> @llvm.fma.v2f64(<vscale x 2 x double> %va, <vscale x 2 x double> %vc, <vscale x 2 x double> %neg)
   ret <vscale x 2 x double> %vd
 }
 
 define <vscale x 2 x double> @vfmsub_vf_nxv2f64(<vscale x 2 x double> %va, <vscale x 2 x double> %vb, double %c) {
 ; CHECK-LABEL: vfmsub_vf_nxv2f64:
 ; CHECK:       # %bb.0:
-; CHECK-NEXT:    vsetvli a0, zero, e64,m2,tu,mu
-; CHECK-NEXT:    vfmsub.vf v10, fa0, v8
-; CHECK-NEXT:    vmv2r.v v8, v10
+; CHECK-NEXT:    vsetvli a0, zero, e64,m2,ta,mu
+; CHECK-NEXT:    vfmsac.vf v8, fa0, v10
 ; CHECK-NEXT:    ret
   %head = insertelement <vscale x 2 x double> undef, double %c, i32 0
   %splat = shufflevector <vscale x 2 x double> %head, <vscale x 2 x double> undef, <vscale x 2 x i32> zeroinitializer
   %neg = fneg <vscale x 2 x double> %va
   %vd = call <vscale x 2 x double> @llvm.fma.v2f64(<vscale x 2 x double> %vb, <vscale x 2 x double> %splat, <vscale x 2 x double> %neg)
   ret <vscale x 2 x double> %vd
 }
 
 declare <vscale x 4 x double> @llvm.fma.v4f64(<vscale x 4 x double>, <vscale x 4 x double>, <vscale x 4 x double>)
 
 define <vscale x 4 x double> @vfmsub_vv_nxv4f64(<vscale x 4 x double> %va, <vscale x 4 x double> %vb, <vscale x 4 x double> %vc) {
 ; CHECK-LABEL: vfmsub_vv_nxv4f64:
 ; CHECK:       # %bb.0:
-; CHECK-NEXT:    vsetvli a0, zero, e64,m4,tu,mu
+; CHECK-NEXT:    vsetvli a0, zero, e64,m4,ta,mu
 ; CHECK-NEXT:    vfmsub.vv v8, v12, v16
 ; CHECK-NEXT:    ret
   %neg = fneg <vscale x 4 x double> %vc
   %vd = call <vscale x 4 x double> @llvm.fma.v4f64(<vscale x 4 x double> %vb, <vscale x 4 x double> %va, <vscale x 4 x double> %neg)
   ret <vscale x 4 x double> %vd
 }
 
 define <vscale x 4 x double> @vfmsub_vf_nxv4f64(<vscale x 4 x double> %va, <vscale x 4 x double> %vb, double %c) {
 ; CHECK-LABEL: vfmsub_vf_nxv4f64:
 ; CHECK:       # %bb.0:
-; CHECK-NEXT:    vsetvli a0, zero, e64,m4,tu,mu
+; CHECK-NEXT:    vsetvli a0, zero, e64,m4,ta,mu
 ; CHECK-NEXT:    vfmsub.vf v8, fa0, v12
 ; CHECK-NEXT:    ret
   %head = insertelement <vscale x 4 x double> undef, double %c, i32 0
   %splat = shufflevector <vscale x 4 x double> %head, <vscale x 4 x double> undef, <vscale x 4 x i32> zeroinitializer
   %neg = fneg <vscale x 4 x double> %vb
   %vd = call <vscale x 4 x double> @llvm.fma.v4f64(<vscale x 4 x double> %va, <vscale x 4 x double> %splat, <vscale x 4 x double> %neg)
   ret <vscale x 4 x double> %vd
 }
 
 declare <vscale x 8 x double> @llvm.fma.v8f64(<vscale x 8 x double>, <vscale x 8 x double>, <vscale x 8 x double>)
 
 define <vscale x 8 x double> @vfmsub_vv_nxv8f64(<vscale x 8 x double> %va, <vscale x 8 x double> %vb, <vscale x 8 x double> %vc) {
 ; CHECK-LABEL: vfmsub_vv_nxv8f64:
 ; CHECK:       # %bb.0:
 ; CHECK-NEXT:    vsetvli a1, zero, e64,m8,ta,mu
 ; CHECK-NEXT:    vle64.v v24, (a0)
-; CHECK-NEXT:    vsetvli a0, zero, e64,m8,tu,mu
-; CHECK-NEXT:    vfmsub.vv v24, v16, v8
-; CHECK-NEXT:    vmv8r.v v8, v24
+; CHECK-NEXT:    vfmsac.vv v8, v16, v24
 ; CHECK-NEXT:    ret
   %neg = fneg <vscale x 8 x double> %va
   %vd = call <vscale x 8 x double> @llvm.fma.v8f64(<vscale x 8 x double> %vb, <vscale x 8 x double> %vc, <vscale x 8 x double> %neg)
   ret <vscale x 8 x double> %vd
 }
 
 define <vscale x 8 x double> @vfmsub_vf_nxv8f64(<vscale x 8 x double> %va, <vscale x 8 x double> %vb, double %c) {
 ; CHECK-LABEL: vfmsub_vf_nxv8f64:
 ; CHECK:       # %bb.0:
-; CHECK-NEXT:    vsetvli a0, zero, e64,m8,tu,mu
-; CHECK-NEXT:    vfmsub.vf v16, fa0, v8
-; CHECK-NEXT:    vmv8r.v v8, v16
+; CHECK-NEXT:    vsetvli a0, zero, e64,m8,ta,mu
+; CHECK-NEXT:    vfmsac.vf v8, fa0, v16
 ; CHECK-NEXT:    ret
   %head = insertelement <vscale x 8 x double> undef, double %c, i32 0
   %splat = shufflevector <vscale x 8 x double> %head, <vscale x 8 x double> undef, <vscale x 8 x i32> zeroinitializer
   %neg = fneg <vscale x 8 x double> %va
   %vd = call <vscale x 8 x double> @llvm.fma.v8f64(<vscale x 8 x double> %vb, <vscale x 8 x double> %splat, <vscale x 8 x double> %neg)
   ret <vscale x 8 x double> %vd
 }

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

 ; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py
 ; RUN: llc -mtriple=riscv32 -mattr=+d,+experimental-zfh,+experimental-v -target-abi=ilp32d \
 ; RUN:     -verify-machineinstrs < %s | FileCheck %s
 ; RUN: llc -mtriple=riscv64 -mattr=+d,+experimental-zfh,+experimental-v -target-abi=lp64d \
 ; RUN:     -verify-machineinstrs < %s | FileCheck %s
 
 ; This tests a mix of vfnmacc and vfnmadd by using different operand orders to
 ; trigger commuting in TwoAddressInstructionPass.
 
 declare <vscale x 1 x half> @llvm.fma.v1f16(<vscale x 1 x half>, <vscale x 1 x half>, <vscale x 1 x half>)
 
 define <vscale x 1 x half> @vfnmsub_vv_nxv1f16(<vscale x 1 x half> %va, <vscale x 1 x half> %vb, <vscale x 1 x half> %vc) {
 ; CHECK-LABEL: vfnmsub_vv_nxv1f16:
 ; CHECK:       # %bb.0:
-; CHECK-NEXT:    vsetvli a0, zero, e16,mf4,tu,mu
-; CHECK-NEXT:    vfnmadd.vv v9, v8, v10
-; CHECK-NEXT:    vmv1r.v v8, v9
+; CHECK-NEXT:    vsetvli a0, zero, e16,mf4,ta,mu
+; CHECK-NEXT:    vfnmadd.vv v8, v9, v10
 ; CHECK-NEXT:    ret
   %neg = fneg <vscale x 1 x half> %va
   %neg2 = fneg <vscale x 1 x half> %vc
   %vd = call <vscale x 1 x half> @llvm.fma.v1f16(<vscale x 1 x half> %neg, <vscale x 1 x half> %vb, <vscale x 1 x half> %neg2)
   ret <vscale x 1 x half> %vd
 }
 
 define <vscale x 1 x half> @vfnmsub_vf_nxv1f16(<vscale x 1 x half> %va, <vscale x 1 x half> %vb, half %c) {
 ; CHECK-LABEL: vfnmsub_vf_nxv1f16:
 ; CHECK:       # %bb.0:
-; CHECK-NEXT:    vsetvli a0, zero, e16,mf4,tu,mu
+; CHECK-NEXT:    vsetvli a0, zero, e16,mf4,ta,mu
 ; CHECK-NEXT:    vfnmadd.vf v8, fa0, v9
 ; CHECK-NEXT:    ret
   %head = insertelement <vscale x 1 x half> undef, half %c, i32 0
   %splat = shufflevector <vscale x 1 x half> %head, <vscale x 1 x half> undef, <vscale x 1 x i32> zeroinitializer
   %neg = fneg <vscale x 1 x half> %va
   %neg2 = fneg <vscale x 1 x half> %vb
   %vd = call <vscale x 1 x half> @llvm.fma.v1f16(<vscale x 1 x half> %neg, <vscale x 1 x half> %splat, <vscale x 1 x half> %neg2)
   ret <vscale x 1 x half> %vd
 }
 
 declare <vscale x 2 x half> @llvm.fma.v2f16(<vscale x 2 x half>, <vscale x 2 x half>, <vscale x 2 x half>)
 
 define <vscale x 2 x half> @vfnmsub_vv_nxv2f16(<vscale x 2 x half> %va, <vscale x 2 x half> %vb, <vscale x 2 x half> %vc) {
 ; CHECK-LABEL: vfnmsub_vv_nxv2f16:
 ; CHECK:       # %bb.0:
-; CHECK-NEXT:    vsetvli a0, zero, e16,mf2,tu,mu
-; CHECK-NEXT:    vfnmadd.vv v10, v8, v9
-; CHECK-NEXT:    vmv1r.v v8, v10
+; CHECK-NEXT:    vsetvli a0, zero, e16,mf2,ta,mu
+; CHECK-NEXT:    vfnmadd.vv v8, v10, v9
 ; CHECK-NEXT:    ret
   %neg = fneg <vscale x 2 x half> %va
   %neg2 = fneg <vscale x 2 x half> %vb
   %vd = call <vscale x 2 x half> @llvm.fma.v2f16(<vscale x 2 x half> %neg, <vscale x 2 x half> %vc, <vscale x 2 x half> %neg2)
   ret <vscale x 2 x half> %vd
 }
 
 define <vscale x 2 x half> @vfnmsub_vf_nxv2f16(<vscale x 2 x half> %va, <vscale x 2 x half> %vb, half %c) {
 ; CHECK-LABEL: vfnmsub_vf_nxv2f16:
 ; CHECK:       # %bb.0:
-; CHECK-NEXT:    vsetvli a0, zero, e16,mf2,tu,mu
+; CHECK-NEXT:    vsetvli a0, zero, e16,mf2,ta,mu
 ; CHECK-NEXT:    vfnmadd.vf v8, fa0, v9
 ; CHECK-NEXT:    ret
   %head = insertelement <vscale x 2 x half> undef, half %c, i32 0
   %splat = shufflevector <vscale x 2 x half> %head, <vscale x 2 x half> undef, <vscale x 2 x i32> zeroinitializer
   %neg = fneg <vscale x 2 x half> %va
   %neg2 = fneg <vscale x 2 x half> %vb
   %vd = call <vscale x 2 x half> @llvm.fma.v2f16(<vscale x 2 x half> %splat, <vscale x 2 x half> %neg, <vscale x 2 x half> %neg2)
   ret <vscale x 2 x half> %vd
 }
 
 declare <vscale x 4 x half> @llvm.fma.v4f16(<vscale x 4 x half>, <vscale x 4 x half>, <vscale x 4 x half>)
 
 define <vscale x 4 x half> @vfnmsub_vv_nxv4f16(<vscale x 4 x half> %va, <vscale x 4 x half> %vb, <vscale x 4 x half> %vc) {
 ; CHECK-LABEL: vfnmsub_vv_nxv4f16:
 ; CHECK:       # %bb.0:
-; CHECK-NEXT:    vsetvli a0, zero, e16,m1,tu,mu
+; CHECK-NEXT:    vsetvli a0, zero, e16,m1,ta,mu
 ; CHECK-NEXT:    vfnmadd.vv v8, v9, v10
 ; CHECK-NEXT:    ret
   %neg = fneg <vscale x 4 x half> %vb
   %neg2 = fneg <vscale x 4 x half> %vc
   %vd = call <vscale x 4 x half> @llvm.fma.v4f16(<vscale x 4 x half> %neg, <vscale x 4 x half> %va, <vscale x 4 x half> %neg2)
   ret <vscale x 4 x half> %vd
 }
 
 define <vscale x 4 x half> @vfnmsub_vf_nxv4f16(<vscale x 4 x half> %va, <vscale x 4 x half> %vb, half %c) {
 ; CHECK-LABEL: vfnmsub_vf_nxv4f16:
 ; CHECK:       # %bb.0:
-; CHECK-NEXT:    vsetvli a0, zero, e16,m1,tu,mu
+; CHECK-NEXT:    vsetvli a0, zero, e16,m1,ta,mu
 ; CHECK-NEXT:    vfnmadd.vf v8, fa0, v9
 ; CHECK-NEXT:    ret
   %head = insertelement <vscale x 4 x half> undef, half %c, i32 0
   %splat = shufflevector <vscale x 4 x half> %head, <vscale x 4 x half> undef, <vscale x 4 x i32> zeroinitializer
   %neg = fneg <vscale x 4 x half> %splat
   %neg2 = fneg <vscale x 4 x half> %vb
   %vd = call <vscale x 4 x half> @llvm.fma.v4f16(<vscale x 4 x half> %va, <vscale x 4 x half> %neg, <vscale x 4 x half> %neg2)
   ret <vscale x 4 x half> %vd
 }
 
 declare <vscale x 8 x half> @llvm.fma.v8f16(<vscale x 8 x half>, <vscale x 8 x half>, <vscale x 8 x half>)
 
 define <vscale x 8 x half> @vfnmsub_vv_nxv8f16(<vscale x 8 x half> %va, <vscale x 8 x half> %vb, <vscale x 8 x half> %vc) {
 ; CHECK-LABEL: vfnmsub_vv_nxv8f16:
 ; CHECK:       # %bb.0:
-; CHECK-NEXT:    vsetvli a0, zero, e16,m2,tu,mu
-; CHECK-NEXT:    vfnmadd.vv v12, v10, v8
-; CHECK-NEXT:    vmv2r.v v8, v12
+; CHECK-NEXT:    vsetvli a0, zero, e16,m2,ta,mu
+; CHECK-NEXT:    vfnmacc.vv v8, v12, v10
 ; CHECK-NEXT:    ret
   %neg = fneg <vscale x 8 x half> %vb
   %neg2 = fneg <vscale x 8 x half> %va
   %vd = call <vscale x 8 x half> @llvm.fma.v8f16(<vscale x 8 x half> %neg, <vscale x 8 x half> %vc, <vscale x 8 x half> %neg2)
   ret <vscale x 8 x half> %vd
 }
 
 define <vscale x 8 x half> @vfnmsub_vf_nxv8f16(<vscale x 8 x half> %va, <vscale x 8 x half> %vb, half %c) {
 ; CHECK-LABEL: vfnmsub_vf_nxv8f16:
 ; CHECK:       # %bb.0:
-; CHECK-NEXT:    vsetvli a0, zero, e16,m2,tu,mu
-; CHECK-NEXT:    vfnmadd.vf v10, fa0, v8
-; CHECK-NEXT:    vmv2r.v v8, v10
+; CHECK-NEXT:    vsetvli a0, zero, e16,m2,ta,mu
+; CHECK-NEXT:    vfnmacc.vf v8, fa0, v10
 ; CHECK-NEXT:    ret
   %head = insertelement <vscale x 8 x half> undef, half %c, i32 0
   %splat = shufflevector <vscale x 8 x half> %head, <vscale x 8 x half> undef, <vscale x 8 x i32> zeroinitializer
   %neg = fneg <vscale x 8 x half> %splat
   %neg2 = fneg <vscale x 8 x half> %va
   %vd = call <vscale x 8 x half> @llvm.fma.v8f16(<vscale x 8 x half> %vb, <vscale x 8 x half> %neg, <vscale x 8 x half> %neg2)
   ret <vscale x 8 x half> %vd
 }
 
 declare <vscale x 16 x half> @llvm.fma.v16f16(<vscale x 16 x half>, <vscale x 16 x half>, <vscale x 16 x half>)
 
 define <vscale x 16 x half> @vfnmsub_vv_nxv16f16(<vscale x 16 x half> %va, <vscale x 16 x half> %vb, <vscale x 16 x half> %vc) {
 ; CHECK-LABEL: vfnmsub_vv_nxv16f16:
 ; CHECK:       # %bb.0:
-; CHECK-NEXT:    vsetvli a0, zero, e16,m4,tu,mu
+; CHECK-NEXT:    vsetvli a0, zero, e16,m4,ta,mu
 ; CHECK-NEXT:    vfnmadd.vv v8, v16, v12
 ; CHECK-NEXT:    ret
   %neg = fneg <vscale x 16 x half> %vc
   %neg2 = fneg <vscale x 16 x half> %vb
   %vd = call <vscale x 16 x half> @llvm.fma.v16f16(<vscale x 16 x half> %neg, <vscale x 16 x half> %va, <vscale x 16 x half> %neg2)
   ret <vscale x 16 x half> %vd
 }
 
 define <vscale x 16 x half> @vfnmsub_vf_nxv16f16(<vscale x 16 x half> %va, <vscale x 16 x half> %vb, half %c) {
 ; CHECK-LABEL: vfnmsub_vf_nxv16f16:
 ; CHECK:       # %bb.0:
-; CHECK-NEXT:    vsetvli a0, zero, e16,m4,tu,mu
+; CHECK-NEXT:    vsetvli a0, zero, e16,m4,ta,mu
 ; CHECK-NEXT:    vfnmadd.vf v8, fa0, v12
 ; CHECK-NEXT:    ret
   %head = insertelement <vscale x 16 x half> undef, half %c, i32 0
   %splat = shufflevector <vscale x 16 x half> %head, <vscale x 16 x half> undef, <vscale x 16 x i32> zeroinitializer
   %neg = fneg <vscale x 16 x half> %splat
   %neg2 = fneg <vscale x 16 x half> %vb
   %vd = call <vscale x 16 x half> @llvm.fma.v16f16(<vscale x 16 x half> %neg, <vscale x 16 x half> %va, <vscale x 16 x half> %neg2)
   ret <vscale x 16 x half> %vd
 }
 
 declare <vscale x 32 x half> @llvm.fma.v32f16(<vscale x 32 x half>, <vscale x 32 x half>, <vscale x 32 x half>)
 
 define <vscale x 32 x half> @vfnmsub_vv_nxv32f16(<vscale x 32 x half> %va, <vscale x 32 x half> %vb, <vscale x 32 x half> %vc) {
 ; CHECK-LABEL: vfnmsub_vv_nxv32f16:
 ; CHECK:       # %bb.0:
 ; CHECK-NEXT:    vsetvli a1, zero, e16,m8,ta,mu
 ; CHECK-NEXT:    vle16.v v24, (a0)
-; CHECK-NEXT:    vsetvli a0, zero, e16,m8,tu,mu
 ; CHECK-NEXT:    vfnmadd.vv v8, v24, v16
 ; CHECK-NEXT:    ret
   %neg = fneg <vscale x 32 x half> %vc
   %neg2 = fneg <vscale x 32 x half> %vb
   %vd = call <vscale x 32 x half> @llvm.fma.v32f16(<vscale x 32 x half> %neg, <vscale x 32 x half> %va, <vscale x 32 x half> %neg2)
   ret <vscale x 32 x half> %vd
 }
 
 define <vscale x 32 x half> @vfnmsub_vf_nxv32f16(<vscale x 32 x half> %va, <vscale x 32 x half> %vb, half %c) {
 ; CHECK-LABEL: vfnmsub_vf_nxv32f16:
 ; CHECK:       # %bb.0:
-; CHECK-NEXT:    vsetvli a0, zero, e16,m8,tu,mu
-; CHECK-NEXT:    vfnmadd.vf v16, fa0, v8
-; CHECK-NEXT:    vmv8r.v v8, v16
+; CHECK-NEXT:    vsetvli a0, zero, e16,m8,ta,mu
+; CHECK-NEXT:    vfnmacc.vf v8, fa0, v16
 ; CHECK-NEXT:    ret
   %head = insertelement <vscale x 32 x half> undef, half %c, i32 0
   %splat = shufflevector <vscale x 32 x half> %head, <vscale x 32 x half> undef, <vscale x 32 x i32> zeroinitializer
   %neg = fneg <vscale x 32 x half> %splat
   %neg2 = fneg <vscale x 32 x half> %va
   %vd = call <vscale x 32 x half> @llvm.fma.v32f16(<vscale x 32 x half> %neg, <vscale x 32 x half> %vb, <vscale x 32 x half> %neg2)
   ret <vscale x 32 x half> %vd
 }
 
 declare <vscale x 1 x float> @llvm.fma.v1f32(<vscale x 1 x float>, <vscale x 1 x float>, <vscale x 1 x float>)
 
 define <vscale x 1 x float> @vfnmsub_vv_nxv1f32(<vscale x 1 x float> %va, <vscale x 1 x float> %vb, <vscale x 1 x float> %vc) {
 ; CHECK-LABEL: vfnmsub_vv_nxv1f32:
 ; CHECK:       # %bb.0:
-; CHECK-NEXT:    vsetvli a0, zero, e32,mf2,tu,mu
+; CHECK-NEXT:    vsetvli a0, zero, e32,mf2,ta,mu
 ; CHECK-NEXT:    vfnmadd.vv v8, v9, v10
 ; CHECK-NEXT:    ret
   %neg = fneg <vscale x 1 x float> %vb
   %neg2 = fneg <vscale x 1 x float> %vc
   %vd = call <vscale x 1 x float> @llvm.fma.v1f32(<vscale x 1 x float> %va, <vscale x 1 x float> %neg, <vscale x 1 x float> %neg2)
   ret <vscale x 1 x float> %vd
 }
 
 define <vscale x 1 x float> @vfnmsub_vf_nxv1f32(<vscale x 1 x float> %va, <vscale x 1 x float> %vb, float %c) {
 ; CHECK-LABEL: vfnmsub_vf_nxv1f32:
 ; CHECK:       # %bb.0:
-; CHECK-NEXT:    vsetvli a0, zero, e32,mf2,tu,mu
+; CHECK-NEXT:    vsetvli a0, zero, e32,mf2,ta,mu
 ; CHECK-NEXT:    vfnmadd.vf v8, fa0, v9
 ; CHECK-NEXT:    ret
   %head = insertelement <vscale x 1 x float> undef, float %c, i32 0
   %splat = shufflevector <vscale x 1 x float> %head, <vscale x 1 x float> undef, <vscale x 1 x i32> zeroinitializer
   %neg = fneg <vscale x 1 x float> %va
   %neg2 = fneg <vscale x 1 x float> %vb
   %vd = call <vscale x 1 x float> @llvm.fma.v1f32(<vscale x 1 x float> %neg, <vscale x 1 x float> %splat, <vscale x 1 x float> %neg2)
   ret <vscale x 1 x float> %vd
 }
 
 declare <vscale x 2 x float> @llvm.fma.v2f32(<vscale x 2 x float>, <vscale x 2 x float>, <vscale x 2 x float>)
 
 define <vscale x 2 x float> @vfnmsub_vv_nxv2f32(<vscale x 2 x float> %va, <vscale x 2 x float> %vb, <vscale x 2 x float> %vc) {
 ; CHECK-LABEL: vfnmsub_vv_nxv2f32:
 ; CHECK:       # %bb.0:
-; CHECK-NEXT:    vsetvli a0, zero, e32,m1,tu,mu
+; CHECK-NEXT:    vsetvli a0, zero, e32,m1,ta,mu
 ; CHECK-NEXT:    vfnmadd.vv v8, v10, v9
 ; CHECK-NEXT:    ret
   %neg = fneg <vscale x 2 x float> %vc
   %neg2 = fneg <vscale x 2 x float> %vb
   %vd = call <vscale x 2 x float> @llvm.fma.v2f32(<vscale x 2 x float> %va, <vscale x 2 x float> %neg, <vscale x 2 x float> %neg2)
   ret <vscale x 2 x float> %vd
 }
 
 define <vscale x 2 x float> @vfnmsub_vf_nxv2f32(<vscale x 2 x float> %va, <vscale x 2 x float> %vb, float %c) {
 ; CHECK-LABEL: vfnmsub_vf_nxv2f32:
 ; CHECK:       # %bb.0:
-; CHECK-NEXT:    vsetvli a0, zero, e32,m1,tu,mu
+; CHECK-NEXT:    vsetvli a0, zero, e32,m1,ta,mu
 ; CHECK-NEXT:    vfnmadd.vf v8, fa0, v9
 ; CHECK-NEXT:    ret
   %head = insertelement <vscale x 2 x float> undef, float %c, i32 0
   %splat = shufflevector <vscale x 2 x float> %head, <vscale x 2 x float> undef, <vscale x 2 x i32> zeroinitializer
   %neg = fneg <vscale x 2 x float> %va
   %neg2 = fneg <vscale x 2 x float> %vb
   %vd = call <vscale x 2 x float> @llvm.fma.v2f32(<vscale x 2 x float> %splat, <vscale x 2 x float> %neg, <vscale x 2 x float> %neg2)
   ret <vscale x 2 x float> %vd
 }
 
 declare <vscale x 4 x float> @llvm.fma.v4f32(<vscale x 4 x float>, <vscale x 4 x float>, <vscale x 4 x float>)
 
 define <vscale x 4 x float> @vfnmsub_vv_nxv4f32(<vscale x 4 x float> %va, <vscale x 4 x float> %vb, <vscale x 4 x float> %vc) {
 ; CHECK-LABEL: vfnmsub_vv_nxv4f32:
 ; CHECK:       # %bb.0:
-; CHECK-NEXT:    vsetvli a0, zero, e32,m2,tu,mu
-; CHECK-NEXT:    vfnmadd.vv v10, v8, v12
-; CHECK-NEXT:    vmv2r.v v8, v10
+; CHECK-NEXT:    vsetvli a0, zero, e32,m2,ta,mu
+; CHECK-NEXT:    vfnmadd.vv v8, v10, v12
 ; CHECK-NEXT:    ret
   %neg = fneg <vscale x 4 x float> %va
   %neg2 = fneg <vscale x 4 x float> %vc
   %vd = call <vscale x 4 x float> @llvm.fma.v4f32(<vscale x 4 x float> %vb, <vscale x 4 x float> %neg, <vscale x 4 x float> %neg2)
   ret <vscale x 4 x float> %vd
 }
 
 define <vscale x 4 x float> @vfnmsub_vf_nxv4f32(<vscale x 4 x float> %va, <vscale x 4 x float> %vb, float %c) {
 ; CHECK-LABEL: vfnmsub_vf_nxv4f32:
 ; CHECK:       # %bb.0:
-; CHECK-NEXT:    vsetvli a0, zero, e32,m2,tu,mu
+; CHECK-NEXT:    vsetvli a0, zero, e32,m2,ta,mu
 ; CHECK-NEXT:    vfnmadd.vf v8, fa0, v10
 ; CHECK-NEXT:    ret
   %head = insertelement <vscale x 4 x float> undef, float %c, i32 0
   %splat = shufflevector <vscale x 4 x float> %head, <vscale x 4 x float> undef, <vscale x 4 x i32> zeroinitializer
   %neg = fneg <vscale x 4 x float> %splat
   %neg2 = fneg <vscale x 4 x float> %vb
   %vd = call <vscale x 4 x float> @llvm.fma.v4f32(<vscale x 4 x float> %va, <vscale x 4 x float> %neg, <vscale x 4 x float> %neg2)
   ret <vscale x 4 x float> %vd
 }
 
 declare <vscale x 8 x float> @llvm.fma.v8f32(<vscale x 8 x float>, <vscale x 8 x float>, <vscale x 8 x float>)
 
 define <vscale x 8 x float> @vfnmsub_vv_nxv8f32(<vscale x 8 x float> %va, <vscale x 8 x float> %vb, <vscale x 8 x float> %vc) {
 ; CHECK-LABEL: vfnmsub_vv_nxv8f32:
 ; CHECK:       # %bb.0:
-; CHECK-NEXT:    vsetvli a0, zero, e32,m4,tu,mu
-; CHECK-NEXT:    vfnmadd.vv v12, v16, v8
-; CHECK-NEXT:    vmv4r.v v8, v12
+; CHECK-NEXT:    vsetvli a0, zero, e32,m4,ta,mu
+; CHECK-NEXT:    vfnmacc.vv v8, v16, v12
 ; CHECK-NEXT:    ret
   %neg = fneg <vscale x 8 x float> %vc
   %neg2 = fneg <vscale x 8 x float> %va
   %vd = call <vscale x 8 x float> @llvm.fma.v8f32(<vscale x 8 x float> %vb, <vscale x 8 x float> %neg, <vscale x 8 x float> %neg2)
   ret <vscale x 8 x float> %vd
 }
 
 define <vscale x 8 x float> @vfnmsub_vf_nxv8f32(<vscale x 8 x float> %va, <vscale x 8 x float> %vb, float %c) {
 ; CHECK-LABEL: vfnmsub_vf_nxv8f32:
 ; CHECK:       # %bb.0:
-; CHECK-NEXT:    vsetvli a0, zero, e32,m4,tu,mu
-; CHECK-NEXT:    vfnmadd.vf v12, fa0, v8
-; CHECK-NEXT:    vmv4r.v v8, v12
+; CHECK-NEXT:    vsetvli a0, zero, e32,m4,ta,mu
+; CHECK-NEXT:    vfnmacc.vf v8, fa0, v12
 ; CHECK-NEXT:    ret
   %head = insertelement <vscale x 8 x float> undef, float %c, i32 0
   %splat = shufflevector <vscale x 8 x float> %head, <vscale x 8 x float> undef, <vscale x 8 x i32> zeroinitializer
   %neg = fneg <vscale x 8 x float> %splat
   %neg2 = fneg <vscale x 8 x float> %va
   %vd = call <vscale x 8 x float> @llvm.fma.v8f32(<vscale x 8 x float> %vb, <vscale x 8 x float> %neg, <vscale x 8 x float> %neg2)
   ret <vscale x 8 x float> %vd
 }
 
 declare <vscale x 16 x float> @llvm.fma.v16f32(<vscale x 16 x float>, <vscale x 16 x float>, <vscale x 16 x float>)
 
 define <vscale x 16 x float> @vfnmsub_vv_nxv16f32(<vscale x 16 x float> %va, <vscale x 16 x float> %vb, <vscale x 16 x float> %vc) {
 ; CHECK-LABEL: vfnmsub_vv_nxv16f32:
 ; CHECK:       # %bb.0:
 ; CHECK-NEXT:    vsetvli a1, zero, e32,m8,ta,mu
 ; CHECK-NEXT:    vle32.v v24, (a0)
-; CHECK-NEXT:    vsetvli a0, zero, e32,m8,tu,mu
-; CHECK-NEXT:    vfnmadd.vv v24, v8, v16
-; CHECK-NEXT:    vmv8r.v v8, v24
+; CHECK-NEXT:    vfnmadd.vv v8, v24, v16
 ; CHECK-NEXT:    ret
   %neg = fneg <vscale x 16 x float> %va
   %neg2 = fneg <vscale x 16 x float> %vb
   %vd = call <vscale x 16 x float> @llvm.fma.v16f32(<vscale x 16 x float> %vc, <vscale x 16 x float> %neg, <vscale x 16 x float> %neg2)
   ret <vscale x 16 x float> %vd
 }
 
 define <vscale x 16 x float> @vfnmsub_vf_nxv16f32(<vscale x 16 x float> %va, <vscale x 16 x float> %vb, float %c) {
 ; CHECK-LABEL: vfnmsub_vf_nxv16f32:
 ; CHECK:       # %bb.0:
-; CHECK-NEXT:    vsetvli a0, zero, e32,m8,tu,mu
+; CHECK-NEXT:    vsetvli a0, zero, e32,m8,ta,mu
 ; CHECK-NEXT:    vfnmadd.vf v8, fa0, v16
 ; CHECK-NEXT:    ret
   %head = insertelement <vscale x 16 x float> undef, float %c, i32 0
   %splat = shufflevector <vscale x 16 x float> %head, <vscale x 16 x float> undef, <vscale x 16 x i32> zeroinitializer
   %neg = fneg <vscale x 16 x float> %splat
   %neg2 = fneg <vscale x 16 x float> %vb
   %vd = call <vscale x 16 x float> @llvm.fma.v16f32(<vscale x 16 x float> %neg, <vscale x 16 x float> %va, <vscale x 16 x float> %neg2)
   ret <vscale x 16 x float> %vd
 }
 
 declare <vscale x 1 x double> @llvm.fma.v1f64(<vscale x 1 x double>, <vscale x 1 x double>, <vscale x 1 x double>)
 
 define <vscale x 1 x double> @vfnmsub_vv_nxv1f64(<vscale x 1 x double> %va, <vscale x 1 x double> %vb, <vscale x 1 x double> %vc) {
 ; CHECK-LABEL: vfnmsub_vv_nxv1f64:
 ; CHECK:       # %bb.0:
-; CHECK-NEXT:    vsetvli a0, zero, e64,m1,tu,mu
-; CHECK-NEXT:    vfnmadd.vv v10, v9, v8
-; CHECK-NEXT:    vmv1r.v v8, v10
+; CHECK-NEXT:    vsetvli a0, zero, e64,m1,ta,mu
+; CHECK-NEXT:    vfnmacc.vv v8, v10, v9
 ; CHECK-NEXT:    ret
   %neg = fneg <vscale x 1 x double> %vb
   %neg2 = fneg <vscale x 1 x double> %va
   %vd = call <vscale x 1 x double> @llvm.fma.v1f64(<vscale x 1 x double> %vc, <vscale x 1 x double> %neg, <vscale x 1 x double> %neg2)
   ret <vscale x 1 x double> %vd
 }
 
 define <vscale x 1 x double> @vfnmsub_vf_nxv1f64(<vscale x 1 x double> %va, <vscale x 1 x double> %vb, double %c) {
 ; CHECK-LABEL: vfnmsub_vf_nxv1f64:
 ; CHECK:       # %bb.0:
-; CHECK-NEXT:    vsetvli a0, zero, e64,m1,tu,mu
+; CHECK-NEXT:    vsetvli a0, zero, e64,m1,ta,mu
 ; CHECK-NEXT:    vfnmadd.vf v8, fa0, v9
 ; CHECK-NEXT:    ret
   %head = insertelement <vscale x 1 x double> undef, double %c, i32 0
   %splat = shufflevector <vscale x 1 x double> %head, <vscale x 1 x double> undef, <vscale x 1 x i32> zeroinitializer
   %neg = fneg <vscale x 1 x double> %va
   %neg2 = fneg <vscale x 1 x double> %vb
   %vd = call <vscale x 1 x double> @llvm.fma.v1f64(<vscale x 1 x double> %neg, <vscale x 1 x double> %splat, <vscale x 1 x double> %neg2)
   ret <vscale x 1 x double> %vd
 }
 
 declare <vscale x 2 x double> @llvm.fma.v2f64(<vscale x 2 x double>, <vscale x 2 x double>, <vscale x 2 x double>)
 
 define <vscale x 2 x double> @vfnmsub_vv_nxv2f64(<vscale x 2 x double> %va, <vscale x 2 x double> %vb, <vscale x 2 x double> %vc) {
 ; CHECK-LABEL: vfnmsub_vv_nxv2f64:
 ; CHECK:       # %bb.0:
-; CHECK-NEXT:    vsetvli a0, zero, e64,m2,tu,mu
-; CHECK-NEXT:    vfnmadd.vv v12, v8, v10
-; CHECK-NEXT:    vmv2r.v v8, v12
+; CHECK-NEXT:    vsetvli a0, zero, e64,m2,ta,mu
+; CHECK-NEXT:    vfnmadd.vv v8, v12, v10
 ; CHECK-NEXT:    ret
   %neg = fneg <vscale x 2 x double> %va
   %neg2 = fneg <vscale x 2 x double> %vb
   %vd = call <vscale x 2 x double> @llvm.fma.v2f64(<vscale x 2 x double> %neg, <vscale x 2 x double> %vc, <vscale x 2 x double> %neg2)
   ret <vscale x 2 x double> %vd
 }
 
 define <vscale x 2 x double> @vfnmsub_vf_nxv2f64(<vscale x 2 x double> %va, <vscale x 2 x double> %vb, double %c) {
 ; CHECK-LABEL: vfnmsub_vf_nxv2f64:
 ; CHECK:       # %bb.0:
-; CHECK-NEXT:    vsetvli a0, zero, e64,m2,tu,mu
+; CHECK-NEXT:    vsetvli a0, zero, e64,m2,ta,mu
 ; CHECK-NEXT:    vfnmadd.vf v8, fa0, v10
 ; CHECK-NEXT:    ret
   %head = insertelement <vscale x 2 x double> undef, double %c, i32 0
   %splat = shufflevector <vscale x 2 x double> %head, <vscale x 2 x double> undef, <vscale x 2 x i32> zeroinitializer
   %neg = fneg <vscale x 2 x double> %va
   %neg2 = fneg <vscale x 2 x double> %vb
   %vd = call <vscale x 2 x double> @llvm.fma.v2f64(<vscale x 2 x double> %splat, <vscale x 2 x double> %neg, <vscale x 2 x double> %neg2)
   ret <vscale x 2 x double> %vd
 }
 
 declare <vscale x 4 x double> @llvm.fma.v4f64(<vscale x 4 x double>, <vscale x 4 x double>, <vscale x 4 x double>)
 
 define <vscale x 4 x double> @vfnmsub_vv_nxv4f64(<vscale x 4 x double> %va, <vscale x 4 x double> %vb, <vscale x 4 x double> %vc) {
 ; CHECK-LABEL: vfnmsub_vv_nxv4f64:
 ; CHECK:       # %bb.0:
-; CHECK-NEXT:    vsetvli a0, zero, e64,m4,tu,mu
+; CHECK-NEXT:    vsetvli a0, zero, e64,m4,ta,mu
 ; CHECK-NEXT:    vfnmadd.vv v8, v12, v16
 ; CHECK-NEXT:    ret
   %neg = fneg <vscale x 4 x double> %vb
   %neg2 = fneg <vscale x 4 x double> %vc
   %vd = call <vscale x 4 x double> @llvm.fma.v4f64(<vscale x 4 x double> %neg, <vscale x 4 x double> %va, <vscale x 4 x double> %neg2)
   ret <vscale x 4 x double> %vd
 }
 
 define <vscale x 4 x double> @vfnmsub_vf_nxv4f64(<vscale x 4 x double> %va, <vscale x 4 x double> %vb, double %c) {
 ; CHECK-LABEL: vfnmsub_vf_nxv4f64:
 ; CHECK:       # %bb.0:
-; CHECK-NEXT:    vsetvli a0, zero, e64,m4,tu,mu
+; CHECK-NEXT:    vsetvli a0, zero, e64,m4,ta,mu
 ; CHECK-NEXT:    vfnmadd.vf v8, fa0, v12
 ; CHECK-NEXT:    ret
   %head = insertelement <vscale x 4 x double> undef, double %c, i32 0
   %splat = shufflevector <vscale x 4 x double> %head, <vscale x 4 x double> undef, <vscale x 4 x i32> zeroinitializer
   %neg = fneg <vscale x 4 x double> %splat
   %neg2 = fneg <vscale x 4 x double> %vb
   %vd = call <vscale x 4 x double> @llvm.fma.v4f64(<vscale x 4 x double> %va, <vscale x 4 x double> %neg, <vscale x 4 x double> %neg2)
   ret <vscale x 4 x double> %vd
 }
 
 declare <vscale x 8 x double> @llvm.fma.v8f64(<vscale x 8 x double>, <vscale x 8 x double>, <vscale x 8 x double>)
 
 define <vscale x 8 x double> @vfnmsub_vv_nxv8f64(<vscale x 8 x double> %va, <vscale x 8 x double> %vb, <vscale x 8 x double> %vc) {
 ; CHECK-LABEL: vfnmsub_vv_nxv8f64:
 ; CHECK:       # %bb.0:
 ; CHECK-NEXT:    vsetvli a1, zero, e64,m8,ta,mu
 ; CHECK-NEXT:    vle64.v v24, (a0)
-; CHECK-NEXT:    vsetvli a0, zero, e64,m8,tu,mu
-; CHECK-NEXT:    vfnmadd.vv v24, v16, v8
-; CHECK-NEXT:    vmv8r.v v8, v24
+; CHECK-NEXT:    vfnmacc.vv v8, v16, v24
 ; CHECK-NEXT:    ret
   %neg = fneg <vscale x 8 x double> %vb
   %neg2 = fneg <vscale x 8 x double> %va
   %vd = call <vscale x 8 x double> @llvm.fma.v8f64(<vscale x 8 x double> %neg, <vscale x 8 x double> %vc, <vscale x 8 x double> %neg2)
   ret <vscale x 8 x double> %vd
 }
 
 define <vscale x 8 x double> @vfnmsub_vf_nxv8f64(<vscale x 8 x double> %va, <vscale x 8 x double> %vb, double %c) {
 ; CHECK-LABEL: vfnmsub_vf_nxv8f64:
 ; CHECK:       # %bb.0:
-; CHECK-NEXT:    vsetvli a0, zero, e64,m8,tu,mu
-; CHECK-NEXT:    vfnmadd.vf v16, fa0, v8
-; CHECK-NEXT:    vmv8r.v v8, v16
+; CHECK-NEXT:    vsetvli a0, zero, e64,m8,ta,mu
+; CHECK-NEXT:    vfnmacc.vf v8, fa0, v16
 ; CHECK-NEXT:    ret
   %head = insertelement <vscale x 8 x double> undef, double %c, i32 0
   %splat = shufflevector <vscale x 8 x double> %head, <vscale x 8 x double> undef, <vscale x 8 x i32> zeroinitializer
   %neg = fneg <vscale x 8 x double> %splat
   %neg2 = fneg <vscale x 8 x double> %va
   %vd = call <vscale x 8 x double> @llvm.fma.v8f64(<vscale x 8 x double> %vb, <vscale x 8 x double> %neg, <vscale x 8 x double> %neg2)
   ret <vscale x 8 x double> %vd
 }

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

 ; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py
 ; RUN: llc -mtriple=riscv32 -mattr=+d,+experimental-zfh,+experimental-v -target-abi=ilp32d \
 ; RUN:     -verify-machineinstrs < %s | FileCheck %s
 ; RUN: llc -mtriple=riscv64 -mattr=+d,+experimental-zfh,+experimental-v -target-abi=lp64d \
 ; RUN:     -verify-machineinstrs < %s | FileCheck %s
 
 ; This tests a mix of vfnmsac and vfnmsub by using different operand orders to
 ; trigger commuting in TwoAddressInstructionPass.
 
 declare <vscale x 1 x half> @llvm.fma.v1f16(<vscale x 1 x half>, <vscale x 1 x half>, <vscale x 1 x half>)
 
 define <vscale x 1 x half> @vfnmsub_vv_nxv1f16(<vscale x 1 x half> %va, <vscale x 1 x half> %vb, <vscale x 1 x half> %vc) {
 ; CHECK-LABEL: vfnmsub_vv_nxv1f16:
 ; CHECK:       # %bb.0:
-; CHECK-NEXT:    vsetvli a0, zero, e16,mf4,tu,mu
-; CHECK-NEXT:    vfnmsub.vv v9, v8, v10
-; CHECK-NEXT:    vmv1r.v v8, v9
+; CHECK-NEXT:    vsetvli a0, zero, e16,mf4,ta,mu
+; CHECK-NEXT:    vfnmsub.vv v8, v9, v10
 ; CHECK-NEXT:    ret
   %neg = fneg <vscale x 1 x half> %va
   %vd = call <vscale x 1 x half> @llvm.fma.v1f16(<vscale x 1 x half> %neg, <vscale x 1 x half> %vb, <vscale x 1 x half> %vc)
   ret <vscale x 1 x half> %vd
 }
 
 define <vscale x 1 x half> @vfnmsub_vf_nxv1f16(<vscale x 1 x half> %va, <vscale x 1 x half> %vb, half %c) {
 ; CHECK-LABEL: vfnmsub_vf_nxv1f16:
 ; CHECK:       # %bb.0:
-; CHECK-NEXT:    vsetvli a0, zero, e16,mf4,tu,mu
+; CHECK-NEXT:    vsetvli a0, zero, e16,mf4,ta,mu
 ; CHECK-NEXT:    vfnmsub.vf v8, fa0, v9
 ; CHECK-NEXT:    ret
   %head = insertelement <vscale x 1 x half> undef, half %c, i32 0
   %splat = shufflevector <vscale x 1 x half> %head, <vscale x 1 x half> undef, <vscale x 1 x i32> zeroinitializer
   %neg = fneg <vscale x 1 x half> %va
   %vd = call <vscale x 1 x half> @llvm.fma.v1f16(<vscale x 1 x half> %neg, <vscale x 1 x half> %splat, <vscale x 1 x half> %vb)
   ret <vscale x 1 x half> %vd
 }
 
 declare <vscale x 2 x half> @llvm.fma.v2f16(<vscale x 2 x half>, <vscale x 2 x half>, <vscale x 2 x half>)
 
 define <vscale x 2 x half> @vfnmsub_vv_nxv2f16(<vscale x 2 x half> %va, <vscale x 2 x half> %vb, <vscale x 2 x half> %vc) {
 ; CHECK-LABEL: vfnmsub_vv_nxv2f16:
 ; CHECK:       # %bb.0:
-; CHECK-NEXT:    vsetvli a0, zero, e16,mf2,tu,mu
-; CHECK-NEXT:    vfnmsub.vv v10, v8, v9
-; CHECK-NEXT:    vmv1r.v v8, v10
+; CHECK-NEXT:    vsetvli a0, zero, e16,mf2,ta,mu
+; CHECK-NEXT:    vfnmsub.vv v8, v10, v9
 ; CHECK-NEXT:    ret
   %neg = fneg <vscale x 2 x half> %va
   %vd = call <vscale x 2 x half> @llvm.fma.v2f16(<vscale x 2 x half> %neg, <vscale x 2 x half> %vc, <vscale x 2 x half> %vb)
   ret <vscale x 2 x half> %vd
 }
 
 define <vscale x 2 x half> @vfnmsub_vf_nxv2f16(<vscale x 2 x half> %va, <vscale x 2 x half> %vb, half %c) {
 ; CHECK-LABEL: vfnmsub_vf_nxv2f16:
 ; CHECK:       # %bb.0:
-; CHECK-NEXT:    vsetvli a0, zero, e16,mf2,tu,mu
+; CHECK-NEXT:    vsetvli a0, zero, e16,mf2,ta,mu
 ; CHECK-NEXT:    vfnmsub.vf v8, fa0, v9
 ; CHECK-NEXT:    ret
   %head = insertelement <vscale x 2 x half> undef, half %c, i32 0
   %splat = shufflevector <vscale x 2 x half> %head, <vscale x 2 x half> undef, <vscale x 2 x i32> zeroinitializer
   %neg = fneg <vscale x 2 x half> %va
   %vd = call <vscale x 2 x half> @llvm.fma.v2f16(<vscale x 2 x half> %splat, <vscale x 2 x half> %neg, <vscale x 2 x half> %vb)
   ret <vscale x 2 x half> %vd
 }
 
 declare <vscale x 4 x half> @llvm.fma.v4f16(<vscale x 4 x half>, <vscale x 4 x half>, <vscale x 4 x half>)
 
 define <vscale x 4 x half> @vfnmsub_vv_nxv4f16(<vscale x 4 x half> %va, <vscale x 4 x half> %vb, <vscale x 4 x half> %vc) {
 ; CHECK-LABEL: vfnmsub_vv_nxv4f16:
 ; CHECK:       # %bb.0:
-; CHECK-NEXT:    vsetvli a0, zero, e16,m1,tu,mu
+; CHECK-NEXT:    vsetvli a0, zero, e16,m1,ta,mu
 ; CHECK-NEXT:    vfnmsub.vv v8, v9, v10
 ; CHECK-NEXT:    ret
   %neg = fneg <vscale x 4 x half> %vb
   %vd = call <vscale x 4 x half> @llvm.fma.v4f16(<vscale x 4 x half> %neg, <vscale x 4 x half> %va, <vscale x 4 x half> %vc)
   ret <vscale x 4 x half> %vd
 }
 
 define <vscale x 4 x half> @vfnmsub_vf_nxv4f16(<vscale x 4 x half> %va, <vscale x 4 x half> %vb, half %c) {
 ; CHECK-LABEL: vfnmsub_vf_nxv4f16:
 ; CHECK:       # %bb.0:
-; CHECK-NEXT:    vsetvli a0, zero, e16,m1,tu,mu
+; CHECK-NEXT:    vsetvli a0, zero, e16,m1,ta,mu
 ; CHECK-NEXT:    vfnmsub.vf v8, fa0, v9
 ; CHECK-NEXT:    ret
   %head = insertelement <vscale x 4 x half> undef, half %c, i32 0
   %splat = shufflevector <vscale x 4 x half> %head, <vscale x 4 x half> undef, <vscale x 4 x i32> zeroinitializer
   %neg = fneg <vscale x 4 x half> %splat
   %vd = call <vscale x 4 x half> @llvm.fma.v4f16(<vscale x 4 x half> %va, <vscale x 4 x half> %neg, <vscale x 4 x half> %vb)
   ret <vscale x 4 x half> %vd
 }
 
 declare <vscale x 8 x half> @llvm.fma.v8f16(<vscale x 8 x half>, <vscale x 8 x half>, <vscale x 8 x half>)
 
 define <vscale x 8 x half> @vfnmsub_vv_nxv8f16(<vscale x 8 x half> %va, <vscale x 8 x half> %vb, <vscale x 8 x half> %vc) {
 ; CHECK-LABEL: vfnmsub_vv_nxv8f16:
 ; CHECK:       # %bb.0:
-; CHECK-NEXT:    vsetvli a0, zero, e16,m2,tu,mu
-; CHECK-NEXT:    vfnmsub.vv v12, v10, v8
-; CHECK-NEXT:    vmv2r.v v8, v12
+; CHECK-NEXT:    vsetvli a0, zero, e16,m2,ta,mu
+; CHECK-NEXT:    vfnmsac.vv v8, v12, v10
 ; CHECK-NEXT:    ret
   %neg = fneg <vscale x 8 x half> %vb
   %vd = call <vscale x 8 x half> @llvm.fma.v8f16(<vscale x 8 x half> %neg, <vscale x 8 x half> %vc, <vscale x 8 x half> %va)
   ret <vscale x 8 x half> %vd
 }
 
 define <vscale x 8 x half> @vfnmsub_vf_nxv8f16(<vscale x 8 x half> %va, <vscale x 8 x half> %vb, half %c) {
 ; CHECK-LABEL: vfnmsub_vf_nxv8f16:
 ; CHECK:       # %bb.0:
-; CHECK-NEXT:    vsetvli a0, zero, e16,m2,tu,mu
-; CHECK-NEXT:    vfnmsub.vf v10, fa0, v8
-; CHECK-NEXT:    vmv2r.v v8, v10
+; CHECK-NEXT:    vsetvli a0, zero, e16,m2,ta,mu
+; CHECK-NEXT:    vfnmsac.vf v8, fa0, v10
 ; CHECK-NEXT:    ret
   %head = insertelement <vscale x 8 x half> undef, half %c, i32 0
   %splat = shufflevector <vscale x 8 x half> %head, <vscale x 8 x half> undef, <vscale x 8 x i32> zeroinitializer
   %neg = fneg <vscale x 8 x half> %splat
   %vd = call <vscale x 8 x half> @llvm.fma.v8f16(<vscale x 8 x half> %vb, <vscale x 8 x half> %neg, <vscale x 8 x half> %va)
   ret <vscale x 8 x half> %vd
 }
 
 declare <vscale x 16 x half> @llvm.fma.v16f16(<vscale x 16 x half>, <vscale x 16 x half>, <vscale x 16 x half>)
 
 define <vscale x 16 x half> @vfnmsub_vv_nxv16f16(<vscale x 16 x half> %va, <vscale x 16 x half> %vb, <vscale x 16 x half> %vc) {
 ; CHECK-LABEL: vfnmsub_vv_nxv16f16:
 ; CHECK:       # %bb.0:
-; CHECK-NEXT:    vsetvli a0, zero, e16,m4,tu,mu
+; CHECK-NEXT:    vsetvli a0, zero, e16,m4,ta,mu
 ; CHECK-NEXT:    vfnmsub.vv v8, v16, v12
 ; CHECK-NEXT:    ret
   %neg = fneg <vscale x 16 x half> %vc
   %vd = call <vscale x 16 x half> @llvm.fma.v16f16(<vscale x 16 x half> %neg, <vscale x 16 x half> %va, <vscale x 16 x half> %vb)
   ret <vscale x 16 x half> %vd
 }
 
 define <vscale x 16 x half> @vfnmsub_vf_nxv16f16(<vscale x 16 x half> %va, <vscale x 16 x half> %vb, half %c) {
 ; CHECK-LABEL: vfnmsub_vf_nxv16f16:
 ; CHECK:       # %bb.0:
-; CHECK-NEXT:    vsetvli a0, zero, e16,m4,tu,mu
+; CHECK-NEXT:    vsetvli a0, zero, e16,m4,ta,mu
 ; CHECK-NEXT:    vfnmsub.vf v8, fa0, v12
 ; CHECK-NEXT:    ret
   %head = insertelement <vscale x 16 x half> undef, half %c, i32 0
   %splat = shufflevector <vscale x 16 x half> %head, <vscale x 16 x half> undef, <vscale x 16 x i32> zeroinitializer
   %neg = fneg <vscale x 16 x half> %splat
   %vd = call <vscale x 16 x half> @llvm.fma.v16f16(<vscale x 16 x half> %neg, <vscale x 16 x half> %va, <vscale x 16 x half> %vb)
   ret <vscale x 16 x half> %vd
 }
 
 declare <vscale x 32 x half> @llvm.fma.v32f16(<vscale x 32 x half>, <vscale x 32 x half>, <vscale x 32 x half>)
 
 define <vscale x 32 x half> @vfnmsub_vv_nxv32f16(<vscale x 32 x half> %va, <vscale x 32 x half> %vb, <vscale x 32 x half> %vc) {
 ; CHECK-LABEL: vfnmsub_vv_nxv32f16:
 ; CHECK:       # %bb.0:
 ; CHECK-NEXT:    vsetvli a1, zero, e16,m8,ta,mu
 ; CHECK-NEXT:    vle16.v v24, (a0)
-; CHECK-NEXT:    vsetvli a0, zero, e16,m8,tu,mu
 ; CHECK-NEXT:    vfnmsub.vv v8, v24, v16
 ; CHECK-NEXT:    ret
   %neg = fneg <vscale x 32 x half> %vc
   %vd = call <vscale x 32 x half> @llvm.fma.v32f16(<vscale x 32 x half> %neg, <vscale x 32 x half> %va, <vscale x 32 x half> %vb)
   ret <vscale x 32 x half> %vd
 }
 
 define <vscale x 32 x half> @vfnmsub_vf_nxv32f16(<vscale x 32 x half> %va, <vscale x 32 x half> %vb, half %c) {
 ; CHECK-LABEL: vfnmsub_vf_nxv32f16:
 ; CHECK:       # %bb.0:
-; CHECK-NEXT:    vsetvli a0, zero, e16,m8,tu,mu
-; CHECK-NEXT:    vfnmsub.vf v16, fa0, v8
-; CHECK-NEXT:    vmv8r.v v8, v16
+; CHECK-NEXT:    vsetvli a0, zero, e16,m8,ta,mu
+; CHECK-NEXT:    vfnmsac.vf v8, fa0, v16
 ; CHECK-NEXT:    ret
   %head = insertelement <vscale x 32 x half> undef, half %c, i32 0
   %splat = shufflevector <vscale x 32 x half> %head, <vscale x 32 x half> undef, <vscale x 32 x i32> zeroinitializer
   %neg = fneg <vscale x 32 x half> %splat
   %vd = call <vscale x 32 x half> @llvm.fma.v32f16(<vscale x 32 x half> %neg, <vscale x 32 x half> %vb, <vscale x 32 x half> %va)
   ret <vscale x 32 x half> %vd
 }
 
 declare <vscale x 1 x float> @llvm.fma.v1f32(<vscale x 1 x float>, <vscale x 1 x float>, <vscale x 1 x float>)
 
 define <vscale x 1 x float> @vfnmsub_vv_nxv1f32(<vscale x 1 x float> %va, <vscale x 1 x float> %vb, <vscale x 1 x float> %vc) {
 ; CHECK-LABEL: vfnmsub_vv_nxv1f32:
 ; CHECK:       # %bb.0:
-; CHECK-NEXT:    vsetvli a0, zero, e32,mf2,tu,mu
+; CHECK-NEXT:    vsetvli a0, zero, e32,mf2,ta,mu
 ; CHECK-NEXT:    vfnmsub.vv v8, v9, v10
 ; CHECK-NEXT:    ret
   %neg = fneg <vscale x 1 x float> %vb
   %vd = call <vscale x 1 x float> @llvm.fma.v1f32(<vscale x 1 x float> %va, <vscale x 1 x float> %neg, <vscale x 1 x float> %vc)
   ret <vscale x 1 x float> %vd
 }
 
 define <vscale x 1 x float> @vfnmsub_vf_nxv1f32(<vscale x 1 x float> %va, <vscale x 1 x float> %vb, float %c) {
 ; CHECK-LABEL: vfnmsub_vf_nxv1f32:
 ; CHECK:       # %bb.0:
-; CHECK-NEXT:    vsetvli a0, zero, e32,mf2,tu,mu
+; CHECK-NEXT:    vsetvli a0, zero, e32,mf2,ta,mu
 ; CHECK-NEXT:    vfnmsub.vf v8, fa0, v9
 ; CHECK-NEXT:    ret
   %head = insertelement <vscale x 1 x float> undef, float %c, i32 0
   %splat = shufflevector <vscale x 1 x float> %head, <vscale x 1 x float> undef, <vscale x 1 x i32> zeroinitializer
   %neg = fneg <vscale x 1 x float> %va
   %vd = call <vscale x 1 x float> @llvm.fma.v1f32(<vscale x 1 x float> %neg, <vscale x 1 x float> %splat, <vscale x 1 x float> %vb)
   ret <vscale x 1 x float> %vd
 }
 
 declare <vscale x 2 x float> @llvm.fma.v2f32(<vscale x 2 x float>, <vscale x 2 x float>, <vscale x 2 x float>)
 
 define <vscale x 2 x float> @vfnmsub_vv_nxv2f32(<vscale x 2 x float> %va, <vscale x 2 x float> %vb, <vscale x 2 x float> %vc) {
 ; CHECK-LABEL: vfnmsub_vv_nxv2f32:
 ; CHECK:       # %bb.0:
-; CHECK-NEXT:    vsetvli a0, zero, e32,m1,tu,mu
+; CHECK-NEXT:    vsetvli a0, zero, e32,m1,ta,mu
 ; CHECK-NEXT:    vfnmsub.vv v8, v10, v9
 ; CHECK-NEXT:    ret
   %neg = fneg <vscale x 2 x float> %vc
   %vd = call <vscale x 2 x float> @llvm.fma.v2f32(<vscale x 2 x float> %va, <vscale x 2 x float> %neg, <vscale x 2 x float> %vb)
   ret <vscale x 2 x float> %vd
 }
 
 define <vscale x 2 x float> @vfnmsub_vf_nxv2f32(<vscale x 2 x float> %va, <vscale x 2 x float> %vb, float %c) {
 ; CHECK-LABEL: vfnmsub_vf_nxv2f32:
 ; CHECK:       # %bb.0:
-; CHECK-NEXT:    vsetvli a0, zero, e32,m1,tu,mu
+; CHECK-NEXT:    vsetvli a0, zero, e32,m1,ta,mu
 ; CHECK-NEXT:    vfnmsub.vf v8, fa0, v9
 ; CHECK-NEXT:    ret
   %head = insertelement <vscale x 2 x float> undef, float %c, i32 0
   %splat = shufflevector <vscale x 2 x float> %head, <vscale x 2 x float> undef, <vscale x 2 x i32> zeroinitializer
   %neg = fneg <vscale x 2 x float> %va
   %vd = call <vscale x 2 x float> @llvm.fma.v2f32(<vscale x 2 x float> %splat, <vscale x 2 x float> %neg, <vscale x 2 x float> %vb)
   ret <vscale x 2 x float> %vd
 }
 
 declare <vscale x 4 x float> @llvm.fma.v4f32(<vscale x 4 x float>, <vscale x 4 x float>, <vscale x 4 x float>)
 
 define <vscale x 4 x float> @vfnmsub_vv_nxv4f32(<vscale x 4 x float> %va, <vscale x 4 x float> %vb, <vscale x 4 x float> %vc) {
 ; CHECK-LABEL: vfnmsub_vv_nxv4f32:
 ; CHECK:       # %bb.0:
-; CHECK-NEXT:    vsetvli a0, zero, e32,m2,tu,mu
-; CHECK-NEXT:    vfnmsub.vv v10, v8, v12
-; CHECK-NEXT:    vmv2r.v v8, v10
+; CHECK-NEXT:    vsetvli a0, zero, e32,m2,ta,mu
+; CHECK-NEXT:    vfnmsub.vv v8, v10, v12
 ; CHECK-NEXT:    ret
   %neg = fneg <vscale x 4 x float> %va
   %vd = call <vscale x 4 x float> @llvm.fma.v4f32(<vscale x 4 x float> %vb, <vscale x 4 x float> %neg, <vscale x 4 x float> %vc)
   ret <vscale x 4 x float> %vd
 }
 
 define <vscale x 4 x float> @vfnmsub_vf_nxv4f32(<vscale x 4 x float> %va, <vscale x 4 x float> %vb, float %c) {
 ; CHECK-LABEL: vfnmsub_vf_nxv4f32:
 ; CHECK:       # %bb.0:
-; CHECK-NEXT:    vsetvli a0, zero, e32,m2,tu,mu
+; CHECK-NEXT:    vsetvli a0, zero, e32,m2,ta,mu
 ; CHECK-NEXT:    vfnmsub.vf v8, fa0, v10
 ; CHECK-NEXT:    ret
   %head = insertelement <vscale x 4 x float> undef, float %c, i32 0
   %splat = shufflevector <vscale x 4 x float> %head, <vscale x 4 x float> undef, <vscale x 4 x i32> zeroinitializer
   %neg = fneg <vscale x 4 x float> %splat
   %vd = call <vscale x 4 x float> @llvm.fma.v4f32(<vscale x 4 x float> %va, <vscale x 4 x float> %neg, <vscale x 4 x float> %vb)
   ret <vscale x 4 x float> %vd
 }
 
 declare <vscale x 8 x float> @llvm.fma.v8f32(<vscale x 8 x float>, <vscale x 8 x float>, <vscale x 8 x float>)
 
 define <vscale x 8 x float> @vfnmsub_vv_nxv8f32(<vscale x 8 x float> %va, <vscale x 8 x float> %vb, <vscale x 8 x float> %vc) {
 ; CHECK-LABEL: vfnmsub_vv_nxv8f32:
 ; CHECK:       # %bb.0:
-; CHECK-NEXT:    vsetvli a0, zero, e32,m4,tu,mu
-; CHECK-NEXT:    vfnmsub.vv v12, v16, v8
-; CHECK-NEXT:    vmv4r.v v8, v12
+; CHECK-NEXT:    vsetvli a0, zero, e32,m4,ta,mu
+; CHECK-NEXT:    vfnmsac.vv v8, v16, v12
 ; CHECK-NEXT:    ret
   %neg = fneg <vscale x 8 x float> %vc
   %vd = call <vscale x 8 x float> @llvm.fma.v8f32(<vscale x 8 x float> %vb, <vscale x 8 x float> %neg, <vscale x 8 x float> %va)
   ret <vscale x 8 x float> %vd
 }
 
 define <vscale x 8 x float> @vfnmsub_vf_nxv8f32(<vscale x 8 x float> %va, <vscale x 8 x float> %vb, float %c) {
 ; CHECK-LABEL: vfnmsub_vf_nxv8f32:
 ; CHECK:       # %bb.0:
-; CHECK-NEXT:    vsetvli a0, zero, e32,m4,tu,mu
-; CHECK-NEXT:    vfnmsub.vf v12, fa0, v8
-; CHECK-NEXT:    vmv4r.v v8, v12
+; CHECK-NEXT:    vsetvli a0, zero, e32,m4,ta,mu
+; CHECK-NEXT:    vfnmsac.vf v8, fa0, v12
 ; CHECK-NEXT:    ret
   %head = insertelement <vscale x 8 x float> undef, float %c, i32 0
   %splat = shufflevector <vscale x 8 x float> %head, <vscale x 8 x float> undef, <vscale x 8 x i32> zeroinitializer
   %neg = fneg <vscale x 8 x float> %splat
   %vd = call <vscale x 8 x float> @llvm.fma.v8f32(<vscale x 8 x float> %vb, <vscale x 8 x float> %neg, <vscale x 8 x float> %va)
   ret <vscale x 8 x float> %vd
 }
 
 declare <vscale x 16 x float> @llvm.fma.v16f32(<vscale x 16 x float>, <vscale x 16 x float>, <vscale x 16 x float>)
 
 define <vscale x 16 x float> @vfnmsub_vv_nxv16f32(<vscale x 16 x float> %va, <vscale x 16 x float> %vb, <vscale x 16 x float> %vc) {
 ; CHECK-LABEL: vfnmsub_vv_nxv16f32:
 ; CHECK:       # %bb.0:
 ; CHECK-NEXT:    vsetvli a1, zero, e32,m8,ta,mu
 ; CHECK-NEXT:    vle32.v v24, (a0)
-; CHECK-NEXT:    vsetvli a0, zero, e32,m8,tu,mu
-; CHECK-NEXT:    vfnmsub.vv v24, v8, v16
-; CHECK-NEXT:    vmv8r.v v8, v24
+; CHECK-NEXT:    vfnmsub.vv v8, v24, v16
 ; CHECK-NEXT:    ret
   %neg = fneg <vscale x 16 x float> %va
   %vd = call <vscale x 16 x float> @llvm.fma.v16f32(<vscale x 16 x float> %vc, <vscale x 16 x float> %neg, <vscale x 16 x float> %vb)
   ret <vscale x 16 x float> %vd
 }
 
 define <vscale x 16 x float> @vfnmsub_vf_nxv16f32(<vscale x 16 x float> %va, <vscale x 16 x float> %vb, float %c) {
 ; CHECK-LABEL: vfnmsub_vf_nxv16f32:
 ; CHECK:       # %bb.0:
-; CHECK-NEXT:    vsetvli a0, zero, e32,m8,tu,mu
+; CHECK-NEXT:    vsetvli a0, zero, e32,m8,ta,mu
 ; CHECK-NEXT:    vfnmsub.vf v8, fa0, v16
 ; CHECK-NEXT:    ret
   %head = insertelement <vscale x 16 x float> undef, float %c, i32 0
   %splat = shufflevector <vscale x 16 x float> %head, <vscale x 16 x float> undef, <vscale x 16 x i32> zeroinitializer
   %neg = fneg <vscale x 16 x float> %splat
   %vd = call <vscale x 16 x float> @llvm.fma.v16f32(<vscale x 16 x float> %neg, <vscale x 16 x float> %va, <vscale x 16 x float> %vb)
   ret <vscale x 16 x float> %vd
 }
 
 declare <vscale x 1 x double> @llvm.fma.v1f64(<vscale x 1 x double>, <vscale x 1 x double>, <vscale x 1 x double>)
 
 define <vscale x 1 x double> @vfnmsub_vv_nxv1f64(<vscale x 1 x double> %va, <vscale x 1 x double> %vb, <vscale x 1 x double> %vc) {
 ; CHECK-LABEL: vfnmsub_vv_nxv1f64:
 ; CHECK:       # %bb.0:
-; CHECK-NEXT:    vsetvli a0, zero, e64,m1,tu,mu
-; CHECK-NEXT:    vfnmsub.vv v10, v9, v8
-; CHECK-NEXT:    vmv1r.v v8, v10
+; CHECK-NEXT:    vsetvli a0, zero, e64,m1,ta,mu
+; CHECK-NEXT:    vfnmsac.vv v8, v10, v9
 ; CHECK-NEXT:    ret
   %neg = fneg <vscale x 1 x double> %vb
   %vd = call <vscale x 1 x double> @llvm.fma.v1f64(<vscale x 1 x double> %vc, <vscale x 1 x double> %neg, <vscale x 1 x double> %va)
   ret <vscale x 1 x double> %vd
 }
 
 define <vscale x 1 x double> @vfnmsub_vf_nxv1f64(<vscale x 1 x double> %va, <vscale x 1 x double> %vb, double %c) {
 ; CHECK-LABEL: vfnmsub_vf_nxv1f64:
 ; CHECK:       # %bb.0:
-; CHECK-NEXT:    vsetvli a0, zero, e64,m1,tu,mu
+; CHECK-NEXT:    vsetvli a0, zero, e64,m1,ta,mu
 ; CHECK-NEXT:    vfnmsub.vf v8, fa0, v9
 ; CHECK-NEXT:    ret
   %head = insertelement <vscale x 1 x double> undef, double %c, i32 0
   %splat = shufflevector <vscale x 1 x double> %head, <vscale x 1 x double> undef, <vscale x 1 x i32> zeroinitializer
   %neg = fneg <vscale x 1 x double> %va
   %vd = call <vscale x 1 x double> @llvm.fma.v1f64(<vscale x 1 x double> %neg, <vscale x 1 x double> %splat, <vscale x 1 x double> %vb)
   ret <vscale x 1 x double> %vd
 }
 
 declare <vscale x 2 x double> @llvm.fma.v2f64(<vscale x 2 x double>, <vscale x 2 x double>, <vscale x 2 x double>)
 
 define <vscale x 2 x double> @vfnmsub_vv_nxv2f64(<vscale x 2 x double> %va, <vscale x 2 x double> %vb, <vscale x 2 x double> %vc) {
 ; CHECK-LABEL: vfnmsub_vv_nxv2f64:
 ; CHECK:       # %bb.0:
-; CHECK-NEXT:    vsetvli a0, zero, e64,m2,tu,mu
-; CHECK-NEXT:    vfnmsub.vv v12, v8, v10
-; CHECK-NEXT:    vmv2r.v v8, v12
+; CHECK-NEXT:    vsetvli a0, zero, e64,m2,ta,mu
+; CHECK-NEXT:    vfnmsub.vv v8, v12, v10
 ; CHECK-NEXT:    ret
   %neg = fneg <vscale x 2 x double> %va
   %vd = call <vscale x 2 x double> @llvm.fma.v2f64(<vscale x 2 x double> %neg, <vscale x 2 x double> %vc, <vscale x 2 x double> %vb)
   ret <vscale x 2 x double> %vd
 }
 
 define <vscale x 2 x double> @vfnmsub_vf_nxv2f64(<vscale x 2 x double> %va, <vscale x 2 x double> %vb, double %c) {
 ; CHECK-LABEL: vfnmsub_vf_nxv2f64:
 ; CHECK:       # %bb.0:
-; CHECK-NEXT:    vsetvli a0, zero, e64,m2,tu,mu
+; CHECK-NEXT:    vsetvli a0, zero, e64,m2,ta,mu
 ; CHECK-NEXT:    vfnmsub.vf v8, fa0, v10
 ; CHECK-NEXT:    ret
   %head = insertelement <vscale x 2 x double> undef, double %c, i32 0
   %splat = shufflevector <vscale x 2 x double> %head, <vscale x 2 x double> undef, <vscale x 2 x i32> zeroinitializer
   %neg = fneg <vscale x 2 x double> %va
   %vd = call <vscale x 2 x double> @llvm.fma.v2f64(<vscale x 2 x double> %splat, <vscale x 2 x double> %neg, <vscale x 2 x double> %vb)
   ret <vscale x 2 x double> %vd
 }
 
 declare <vscale x 4 x double> @llvm.fma.v4f64(<vscale x 4 x double>, <vscale x 4 x double>, <vscale x 4 x double>)
 
 define <vscale x 4 x double> @vfnmsub_vv_nxv4f64(<vscale x 4 x double> %va, <vscale x 4 x double> %vb, <vscale x 4 x double> %vc) {
 ; CHECK-LABEL: vfnmsub_vv_nxv4f64:
 ; CHECK:       # %bb.0:
-; CHECK-NEXT:    vsetvli a0, zero, e64,m4,tu,mu
+; CHECK-NEXT:    vsetvli a0, zero, e64,m4,ta,mu
 ; CHECK-NEXT:    vfnmsub.vv v8, v12, v16
 ; CHECK-NEXT:    ret
   %neg = fneg <vscale x 4 x double> %vb
   %vd = call <vscale x 4 x double> @llvm.fma.v4f64(<vscale x 4 x double> %neg, <vscale x 4 x double> %va, <vscale x 4 x double> %vc)
   ret <vscale x 4 x double> %vd
 }
 
 define <vscale x 4 x double> @vfnmsub_vf_nxv4f64(<vscale x 4 x double> %va, <vscale x 4 x double> %vb, double %c) {
 ; CHECK-LABEL: vfnmsub_vf_nxv4f64:
 ; CHECK:       # %bb.0:
-; CHECK-NEXT:    vsetvli a0, zero, e64,m4,tu,mu
+; CHECK-NEXT:    vsetvli a0, zero, e64,m4,ta,mu
 ; CHECK-NEXT:    vfnmsub.vf v8, fa0, v12
 ; CHECK-NEXT:    ret
   %head = insertelement <vscale x 4 x double> undef, double %c, i32 0
   %splat = shufflevector <vscale x 4 x double> %head, <vscale x 4 x double> undef, <vscale x 4 x i32> zeroinitializer
   %neg = fneg <vscale x 4 x double> %splat
   %vd = call <vscale x 4 x double> @llvm.fma.v4f64(<vscale x 4 x double> %va, <vscale x 4 x double> %neg, <vscale x 4 x double> %vb)
   ret <vscale x 4 x double> %vd
 }
 
 declare <vscale x 8 x double> @llvm.fma.v8f64(<vscale x 8 x double>, <vscale x 8 x double>, <vscale x 8 x double>)
 
 define <vscale x 8 x double> @vfnmsub_vv_nxv8f64(<vscale x 8 x double> %va, <vscale x 8 x double> %vb, <vscale x 8 x double> %vc) {
 ; CHECK-LABEL: vfnmsub_vv_nxv8f64:
 ; CHECK:       # %bb.0:
 ; CHECK-NEXT:    vsetvli a1, zero, e64,m8,ta,mu
 ; CHECK-NEXT:    vle64.v v24, (a0)
-; CHECK-NEXT:    vsetvli a0, zero, e64,m8,tu,mu
-; CHECK-NEXT:    vfnmsub.vv v24, v16, v8
-; CHECK-NEXT:    vmv8r.v v8, v24
+; CHECK-NEXT:    vfnmsac.vv v8, v16, v24
 ; CHECK-NEXT:    ret
   %neg = fneg <vscale x 8 x double> %vb
   %vd = call <vscale x 8 x double> @llvm.fma.v8f64(<vscale x 8 x double> %neg, <vscale x 8 x double> %vc, <vscale x 8 x double> %va)
   ret <vscale x 8 x double> %vd
 }
 
 define <vscale x 8 x double> @vfnmsub_vf_nxv8f64(<vscale x 8 x double> %va, <vscale x 8 x double> %vb, double %c) {
 ; CHECK-LABEL: vfnmsub_vf_nxv8f64:
 ; CHECK:       # %bb.0:
-; CHECK-NEXT:    vsetvli a0, zero, e64,m8,tu,mu
-; CHECK-NEXT:    vfnmsub.vf v16, fa0, v8
-; CHECK-NEXT:    vmv8r.v v8, v16
+; CHECK-NEXT:    vsetvli a0, zero, e64,m8,ta,mu
+; CHECK-NEXT:    vfnmsac.vf v8, fa0, v16
 ; CHECK-NEXT:    ret
   %head = insertelement <vscale x 8 x double> undef, double %c, i32 0
   %splat = shufflevector <vscale x 8 x double> %head, <vscale x 8 x double> undef, <vscale x 8 x i32> zeroinitializer
   %neg = fneg <vscale x 8 x double> %splat
   %vd = call <vscale x 8 x double> @llvm.fma.v8f64(<vscale x 8 x double> %vb, <vscale x 8 x double> %neg, <vscale x 8 x double> %va)
   ret <vscale x 8 x double> %vd
 }