[IR][DAG][RISCV] Allow scalable vector ISD::STRICT_FP_EXTEND and RISC-V supports for vector ISD::STRICT_FP_EXTEND.

[fakepaper56]

Summary

The patch mainly does two things. The first is allowing scalable vector
ISD::STRICT_FP_EXTEND. The second is making RISC-V customized lower
strict_fpextend to riscv_strict_fpextend_vl, the strict version of
riscv_fpextend_vl.

[craig.topper]

llvm/lib/Target/RISCV/RISCVISelLowering.h

252

There's a special place in the enum for STRICT opcodes. Look for STRICT_FCVT_WU_RV64.

llvm/lib/Target/RISCV/RISCVInstrInfoVVLPatterns.td

1852–1855

Can we use any_riscv_fpextend_vl?

[fakepaper56]

Use any_riscv_fpextend_vl to replace riscv_fpextend_vl and riscv_strict_fpextend_vl.

[craig.topper]

llvm/lib/Target/RISCV/RISCVISelLowering.h

252

This was not addressed.

[fakepaper56]

More STRICT_FP_EXTEND_VL after STRICT_FCVT_W_RV64.

[craig.topper]

LGTM

llvm/lib/CodeGen/SelectionDAG/SelectionDAG.cpp

case ISD::STRICT_FP_EXTEND:
     assert(VTList.NumVTs == 2 && Ops.size() == 2 &&
            "Invalid STRICT_FP_EXTEND!");
     assert(VTList.VTs[0].isFloatingPoint() &&
            Ops[1].getValueType().isFloatingPoint() && "Invalid FP cast!");
     assert(VTList.VTs[0].isVector() == Ops[1].getValueType().isVector() &&
            "STRICT_FP_EXTEND result type should be vector iff the operand "
            "type is vector!");
     assert((!VTList.VTs[0].isVector() ||
-            VTList.VTs[0].getVectorNumElements() ==
-            Ops[1].getValueType().getVectorNumElements()) &&
+            VTList.VTs[0].getVectorElementCount() ==
+                Ops[1].getValueType().getVectorElementCount()) &&
            "Vector element count mismatch!");
     assert(Ops[1].getValueType().bitsLT(VTList.VTs[0]) &&
            "Invalid fpext node, dst <= src!");
     break;
   case ISD::STRICT_FP_ROUND:
     assert(VTList.NumVTs == 2 && Ops.size() == 3 && "Invalid STRICT_FP_ROUND!");
     assert(VTList.VTs[0].isVector() == Ops[1].getValueType().isVector() &&
            "STRICT_FP_ROUND result type should be vector iff the operand "

llvm/lib/IR/Verifier.cpp

case Intrinsic::experimental_constrained_fpext: {
     Type *ResultTy = Result->getType();
     Check(OperandTy->isFPOrFPVectorTy(),
           "Intrinsic first argument must be FP or FP vector", &FPI);
     Check(ResultTy->isFPOrFPVectorTy(),
           "Intrinsic result must be FP or FP vector", &FPI);
     Check(OperandTy->isVectorTy() == ResultTy->isVectorTy(),
           "Intrinsic first argument and result disagree on vector use", &FPI);
     if (OperandTy->isVectorTy()) {
-      Check(cast<FixedVectorType>(OperandTy)->getNumElements() ==
-                cast<FixedVectorType>(ResultTy)->getNumElements(),
+      Check(cast<VectorType>(OperandTy)->getElementCount() ==
+                cast<VectorType>(ResultTy)->getElementCount(),
             "Intrinsic first argument and result vector lengths must be equal",
             &FPI);
     }
     if (FPI.getIntrinsicID() == Intrinsic::experimental_constrained_fptrunc) {
       Check(OperandTy->getScalarSizeInBits() > ResultTy->getScalarSizeInBits(),
             "Intrinsic first argument's type must be larger than result type",
             &FPI);
     } else {

llvm/lib/Target/RISCV/RISCVISelLowering.h

 //===-- RISCVISelLowering.h - RISCV DAG Lowering Interface ------*- C++ -*-===//
 //
 // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
 // See https://llvm.org/LICENSE.txt for license information.
 // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
 //
 //===----------------------------------------------------------------------===//
 //
 // This file defines the interfaces that RISCV uses to lower LLVM code into a
enum NodeType : unsigned {
   VFCVT_RM_X_F_VL, // Has a rounding mode operand.
   VFCVT_RM_XU_F_VL, // Has a rounding mode operand.
   SINT_TO_FP_VL,
   UINT_TO_FP_VL,
   VFCVT_RM_F_X_VL, // Has a rounding mode operand.
   VFCVT_RM_F_XU_VL, // Has a rounding mode operand.
   FP_ROUND_VL,
   FP_EXTEND_VL,
+  STRICT_FP_EXTEND_VL,
 
   // Vector FMA ops with a mask as a fourth operand and VL as a fifth operand.
   VFMADD_VL,
   VFNMADD_VL,
   VFMSUB_VL,
   VFNMSUB_VL,
 
   // Widening instructions with a merge value a third operand, a mask as a
private:
   SDValue lowerFixedLengthVectorExtendToRVV(SDValue Op, SelectionDAG &DAG,
                                             unsigned ExtendOpc) const;
   SDValue lowerGET_ROUNDING(SDValue Op, SelectionDAG &DAG) const;
   SDValue lowerSET_ROUNDING(SDValue Op, SelectionDAG &DAG) const;
 
   SDValue lowerEH_DWARF_CFA(SDValue Op, SelectionDAG &DAG) const;
   SDValue lowerCTLZ_CTTZ_ZERO_UNDEF(SDValue Op, SelectionDAG &DAG) const;
 
+  SDValue lowerStrictFPExtend(SDValue Op, SelectionDAG &DAG) const;
+
   SDValue expandUnalignedRVVLoad(SDValue Op, SelectionDAG &DAG) const;
   SDValue expandUnalignedRVVStore(SDValue Op, SelectionDAG &DAG) const;
 
   bool isEligibleForTailCallOptimization(
       CCState &CCInfo, CallLoweringInfo &CLI, MachineFunction &MF,
       const SmallVector<CCValAssign, 16> &ArgLocs) const;
 
   /// Generate error diagnostics if any register used by CC has been marked

llvm/lib/Target/RISCV/RISCVISelLowering.cpp

const auto SetCommonVFPActions = [&](MVT VT) {
           VT, Custom);
 
       setOperationAction(ISD::VECTOR_DEINTERLEAVE, VT, Custom);
       setOperationAction(ISD::VECTOR_INTERLEAVE, VT, Custom);
 
       setOperationAction({ISD::VECTOR_REVERSE, ISD::VECTOR_SPLICE}, VT, Custom);
 
       setOperationAction(FloatingPointVPOps, VT, Custom);
+
+      setOperationAction(ISD::STRICT_FP_EXTEND, VT, Custom);
     };
 
     // Sets common extload/truncstore actions on RVV floating-point vector
     // types.
     const auto SetCommonVFPExtLoadTruncStoreActions =
         [&](MVT VT, ArrayRef<MVT::SimpleValueType> SmallerVTs) {
           for (auto SmallVT : SmallerVTs) {
             setTruncStoreAction(VT, SmallVT, Expand);
if (Subtarget.useRVVForFixedLengthVectors()) {
         setOperationAction({ISD::VSELECT, ISD::SELECT}, VT, Custom);
         setOperationAction(ISD::SELECT_CC, VT, Expand);
 
         setOperationAction(ISD::BITCAST, VT, Custom);
 
         setOperationAction(FloatingPointVecReduceOps, VT, Custom);
 
         setOperationAction(FloatingPointVPOps, VT, Custom);
+
+        setOperationAction(ISD::STRICT_FP_EXTEND, VT, Custom);
       }
 
       // Custom-legalize bitcasts from fixed-length vectors to scalar types.
       setOperationAction(ISD::BITCAST, {MVT::i8, MVT::i16, MVT::i32, MVT::i64},
                          Custom);
       if (Subtarget.hasStdExtZfhOrZfhmin())
         setOperationAction(ISD::BITCAST, MVT::f16, Custom);
       if (Subtarget.hasStdExtF())
case ISD::FPOWI: {
     }
     return SDValue();
   }
   case ISD::FP_EXTEND:
   case ISD::FP_ROUND:
     if (!Op.getValueType().isVector())
       return Op;
     return lowerVectorFPExtendOrRoundLike(Op, DAG);
+  case ISD::STRICT_FP_EXTEND:
+    return lowerStrictFPExtend(Op, DAG);
   case ISD::FP_TO_SINT:
   case ISD::FP_TO_UINT:
   case ISD::SINT_TO_FP:
   case ISD::UINT_TO_FP: {
     // RVV can only do fp<->int conversions to types half/double the size as
     // the source. We custom-lower any conversions that do two hops into
     // sequences.
     MVT VT = Op.getSimpleValueType();
SDValue RISCVTargetLowering::lowerVectorTruncLike(SDValue Op,
   } while (SrcEltVT != DstEltVT);
 
   if (SrcVT.isFixedLengthVector())
     Result = convertFromScalableVector(VT, Result, DAG, Subtarget);
 
   return Result;
 }
 
+SDValue RISCVTargetLowering::lowerStrictFPExtend(SDValue Op,
+                                                 SelectionDAG &DAG) const {
+  SDLoc DL(Op);
+  SDValue Chain = Op.getOperand(0);
+  SDValue Src = Op.getOperand(1);
+  MVT VT = Op.getSimpleValueType();
+  MVT SrcVT = Src.getSimpleValueType();
+
+  MVT ContainerVT = VT;
+  if (VT.isFixedLengthVector()) {
+    MVT SrcContainerVT = getContainerForFixedLengthVector(SrcVT);
+    ContainerVT =
+        SrcContainerVT.changeVectorElementType(VT.getVectorElementType());
+    Src = convertToScalableVector(SrcContainerVT, Src, DAG, Subtarget);
+  }
+
+  auto [Mask, VL] = getDefaultVLOps(SrcVT, ContainerVT, DL, DAG, Subtarget);
+
+  // RVV can only widen fp to types double the size as the source, so it needs
+  // two vfwcvt to achieve extending fp16 to fp64.
+  if (VT.getVectorElementType() == MVT::f64 &&
+      SrcVT.getVectorElementType() == MVT::f16) {
+    MVT InterVT = ContainerVT.changeVectorElementType(MVT::f32);
+    Src = DAG.getNode(RISCVISD::STRICT_FP_EXTEND_VL, DL,
+                      DAG.getVTList(InterVT, MVT::Other), Chain, Src, Mask, VL);
+    Chain = Src.getValue(1);
+  }
+
+  SDValue Res =
+      DAG.getNode(RISCVISD::STRICT_FP_EXTEND_VL, DL,
+                  DAG.getVTList(ContainerVT, MVT::Other), Chain, Src, Mask, VL);
+  if (VT.isFixedLengthVector()) {
+    // StrictFP operations have two result values. Their lowered result should
+    // have same result count.
+    SDValue SubVec = convertFromScalableVector(VT, Res, DAG, Subtarget);
+    Res = DAG.getMergeValues({SubVec, Res.getValue(1)}, DL);
+  }
+  return Res;
+}
+
 SDValue
 RISCVTargetLowering::lowerVectorFPExtendOrRoundLike(SDValue Op,
                                                     SelectionDAG &DAG) const {
   bool IsVP =
       Op.getOpcode() == ISD::VP_FP_ROUND || Op.getOpcode() == ISD::VP_FP_EXTEND;
   bool IsExtend =
       Op.getOpcode() == ISD::VP_FP_EXTEND || Op.getOpcode() == ISD::FP_EXTEND;
   // RVV can only do truncate fp to types half the size as the source. We
#define NODE_NAME_CASE(NODE)                                                   \
   NODE_NAME_CASE(VFCVT_X_F_VL)
   NODE_NAME_CASE(VFCVT_XU_F_VL)
   NODE_NAME_CASE(VFROUND_NOEXCEPT_VL)
   NODE_NAME_CASE(SINT_TO_FP_VL)
   NODE_NAME_CASE(UINT_TO_FP_VL)
   NODE_NAME_CASE(VFCVT_RM_F_XU_VL)
   NODE_NAME_CASE(VFCVT_RM_F_X_VL)
   NODE_NAME_CASE(FP_EXTEND_VL)
+  NODE_NAME_CASE(STRICT_FP_EXTEND_VL)
   NODE_NAME_CASE(FP_ROUND_VL)
   NODE_NAME_CASE(VWMUL_VL)
   NODE_NAME_CASE(VWMULU_VL)
   NODE_NAME_CASE(VWMULSU_VL)
   NODE_NAME_CASE(VWADD_VL)
   NODE_NAME_CASE(VWADDU_VL)
   NODE_NAME_CASE(VWSUB_VL)
   NODE_NAME_CASE(VWSUBU_VL)

llvm/lib/Target/RISCV/RISCVInstrInfoVVLPatterns.td

 ]>;
 def SDT_RISCVFPExtendOp_VL  : SDTypeProfile<1, 3, [
   SDTCisFP<0>, SDTCisFP<1>, SDTCisOpSmallerThanOp<1, 0>, SDTCisSameNumEltsAs<0, 1>,
   SDTCVecEltisVT<2, i1>, SDTCisSameNumEltsAs<1, 2>, SDTCisVT<3, XLenVT>
 ]>;
 
 def riscv_fpround_vl : SDNode<"RISCVISD::FP_ROUND_VL", SDT_RISCVFPRoundOp_VL>;
 def riscv_fpextend_vl : SDNode<"RISCVISD::FP_EXTEND_VL", SDT_RISCVFPExtendOp_VL>;
+def riscv_strict_fpextend_vl : SDNode<"RISCVISD::STRICT_FP_EXTEND_VL", SDT_RISCVFPExtendOp_VL, [SDNPHasChain]>;
 def riscv_fncvt_rod_vl : SDNode<"RISCVISD::VFNCVT_ROD_VL", SDT_RISCVFPRoundOp_VL>;
 
+def any_riscv_fpextend_vl : PatFrags<(ops node:$src, node:$mask, node:$vl),
+                            [(riscv_fpextend_vl node:$src, node:$mask, node:$vl),
+                             (riscv_strict_fpextend_vl node:$src, node:$mask, node:$vl)]>;
+
 def SDT_RISCVFP2IOp_VL  : SDTypeProfile<1, 3, [
   SDTCisInt<0>, SDTCisFP<1>, SDTCisSameNumEltsAs<0, 1>,
   SDTCVecEltisVT<2, i1>, SDTCisSameNumEltsAs<1, 2>, SDTCisVT<3, XLenVT>
 ]>;
 def SDT_RISCVFP2IOp_RM_VL  : SDTypeProfile<1, 4, [
   SDTCisInt<0>, SDTCisFP<1>, SDTCisSameNumEltsAs<0, 1>,
   SDTCVecEltisVT<2, i1>, SDTCisSameNumEltsAs<1, 2>, SDTCisVT<3, XLenVT>,
   SDTCisVT<4, XLenVT> // Rounding mode
foreach fvti = AllFloatVectors in {
   defm : VPatWConvertI2FPVL_V<riscv_sint_to_fp_vl, "PseudoVFWCVT_F_X_V">;
 
   defm : VPatWConvertI2FP_RM_VL_V<riscv_vfcvt_rm_f_xu_vl, "PseudoVFWCVT_RM_F_XU_V">;
   defm : VPatWConvertI2FP_RM_VL_V<riscv_vfcvt_rm_f_x_vl, "PseudoVFWCVT_RM_F_X_V">;
 
   foreach fvtiToFWti = AllWidenableFloatVectors in {
     defvar fvti = fvtiToFWti.Vti;
     defvar fwti = fvtiToFWti.Wti;
-    def : Pat<(fwti.Vector (riscv_fpextend_vl (fvti.Vector fvti.RegClass:$rs1),
-                                              (fvti.Mask V0),
-                                              VLOpFrag)),
+    def : Pat<(fwti.Vector (any_riscv_fpextend_vl
+                               (fvti.Vector fvti.RegClass:$rs1),
               (!cast<Instruction>("PseudoVFWCVT_F_F_V_"#fvti.LMul.MX#"_MASK")
                   (fwti.Vector (IMPLICIT_DEF)), fvti.RegClass:$rs1,
                   (fvti.Mask V0), GPR:$vl, fvti.Log2SEW, TA_MA)>;
   }
 
   // 13.19 Narrowing Floating-Point/Integer Type-Convert Instructions
   defm : VPatNConvertFP2IVL_W<riscv_vfcvt_xu_f_vl, "PseudoVFNCVT_XU_F_W">;
   defm : VPatNConvertFP2IVL_W<riscv_vfcvt_x_f_vl, "PseudoVFNCVT_X_F_W">;

llvm/test/CodeGen/RISCV/rvv/fixed-vectors-vfpext-constrained-sdnode.ll

+; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py
+; RUN: llc -mtriple=riscv32 -mattr=+d,+zfh,+experimental-zvfh,+v -target-abi=ilp32d -riscv-v-vector-bits-min=128\
+; RUN:     -verify-machineinstrs < %s | FileCheck %s
+; RUN: llc -mtriple=riscv64 -mattr=+d,+zfh,+experimental-zvfh,+v -target-abi=lp64d -riscv-v-vector-bits-min=128\
+; RUN:     -verify-machineinstrs < %s | FileCheck %s
+
+declare <2 x float> @llvm.experimental.constrained.fpext.v2f32.v2f16(<2 x half>, metadata)
+define <2 x float> @vfpext_v2f16_v2f32(<2 x half> %va) {
+; CHECK-LABEL: vfpext_v2f16_v2f32:
+; CHECK:       # %bb.0:
+; CHECK-NEXT:    vsetivli zero, 2, e16, mf4, ta, ma
+; CHECK-NEXT:    vfwcvt.f.f.v v9, v8
+; CHECK-NEXT:    vmv1r.v v8, v9
+; CHECK-NEXT:    ret
+  %evec = call <2 x float> @llvm.experimental.constrained.fpext.v2f32.v2f16(<2 x half> %va, metadata !"fpexcept.strict")
+  ret <2 x float> %evec
+}
+
+declare <2 x double> @llvm.experimental.constrained.fpext.v2f64.v2f16(<2 x half>, metadata)
+define <2 x double> @vfpext_v2f16_v2f64(<2 x half> %va) {
+; CHECK-LABEL: vfpext_v2f16_v2f64:
+; CHECK:       # %bb.0:
+; CHECK-NEXT:    vsetivli zero, 2, e16, mf4, ta, ma
+; CHECK-NEXT:    vfwcvt.f.f.v v9, v8
+; CHECK-NEXT:    vsetvli zero, zero, e32, mf2, ta, ma
+; CHECK-NEXT:    vfwcvt.f.f.v v8, v9
+; CHECK-NEXT:    ret
+  %evec = call <2 x double> @llvm.experimental.constrained.fpext.v2f64.v2f16(<2 x half> %va, metadata !"fpexcept.strict")
+  ret <2 x double> %evec
+}
+
+declare <4 x float> @llvm.experimental.constrained.fpext.v4f32.v4f16(<4 x half>, metadata)
+define <4 x float> @vfpext_v4f16_v4f32(<4 x half> %va) {
+; CHECK-LABEL: vfpext_v4f16_v4f32:
+; CHECK:       # %bb.0:
+; CHECK-NEXT:    vsetivli zero, 4, e16, mf2, ta, ma
+; CHECK-NEXT:    vfwcvt.f.f.v v9, v8
+; CHECK-NEXT:    vmv1r.v v8, v9
+; CHECK-NEXT:    ret
+  %evec = call <4 x float> @llvm.experimental.constrained.fpext.v4f32.v4f16(<4 x half> %va, metadata !"fpexcept.strict")
+  ret <4 x float> %evec
+}
+
+declare <4 x double> @llvm.experimental.constrained.fpext.v4f64.v4f16(<4 x half>, metadata)
+define <4 x double> @vfpext_v4f16_v4f64(<4 x half> %va) {
+; CHECK-LABEL: vfpext_v4f16_v4f64:
+; CHECK:       # %bb.0:
+; CHECK-NEXT:    vsetivli zero, 4, e16, mf2, ta, ma
+; CHECK-NEXT:    vfwcvt.f.f.v v10, v8
+; CHECK-NEXT:    vsetvli zero, zero, e32, m1, ta, ma
+; CHECK-NEXT:    vfwcvt.f.f.v v8, v10
+; CHECK-NEXT:    ret
+  %evec = call <4 x double> @llvm.experimental.constrained.fpext.v4f64.v4f16(<4 x half> %va, metadata !"fpexcept.strict")
+  ret <4 x double> %evec
+}
+
+declare <8 x float> @llvm.experimental.constrained.fpext.v8f32.v8f16(<8 x half>, metadata)
+define <8 x float> @vfpext_v8f16_v8f32(<8 x half> %va) {
+; CHECK-LABEL: vfpext_v8f16_v8f32:
+; CHECK:       # %bb.0:
+; CHECK-NEXT:    vsetivli zero, 8, e16, m1, ta, ma
+; CHECK-NEXT:    vfwcvt.f.f.v v10, v8
+; CHECK-NEXT:    vmv2r.v v8, v10
+; CHECK-NEXT:    ret
+  %evec = call <8 x float> @llvm.experimental.constrained.fpext.v8f32.v8f16(<8 x half> %va, metadata !"fpexcept.strict")
+  ret <8 x float> %evec
+}
+
+declare <8 x double> @llvm.experimental.constrained.fpext.v8f64.v8f16(<8 x half>, metadata)
+define <8 x double> @vfpext_v8f16_v8f64(<8 x half> %va) {
+; CHECK-LABEL: vfpext_v8f16_v8f64:
+; CHECK:       # %bb.0:
+; CHECK-NEXT:    vsetivli zero, 8, e16, m1, ta, ma
+; CHECK-NEXT:    vfwcvt.f.f.v v12, v8
+; CHECK-NEXT:    vsetvli zero, zero, e32, m2, ta, ma
+; CHECK-NEXT:    vfwcvt.f.f.v v8, v12
+; CHECK-NEXT:    ret
+  %evec = call <8 x double> @llvm.experimental.constrained.fpext.v8f64.v8f16(<8 x half> %va, metadata !"fpexcept.strict")
+  ret <8 x double> %evec
+}
+
+declare <2 x double> @llvm.experimental.constrained.fpext.v2f64.v2f32(<2 x float>, metadata)
+define <2 x double> @vfpext_v2f32_v2f64(<2 x float> %va) {
+; CHECK-LABEL: vfpext_v2f32_v2f64:
+; CHECK:       # %bb.0:
+; CHECK-NEXT:    vsetivli zero, 2, e32, mf2, ta, ma
+; CHECK-NEXT:    vfwcvt.f.f.v v9, v8
+; CHECK-NEXT:    vmv1r.v v8, v9
+; CHECK-NEXT:    ret
+  %evec = call <2 x double> @llvm.experimental.constrained.fpext.v2f64.v2f32(<2 x float> %va, metadata !"fpexcept.strict")
+  ret <2 x double> %evec
+}
+
+declare <4 x double> @llvm.experimental.constrained.fpext.v4f64.v4f32(<4 x float>, metadata)
+define <4 x double> @vfpext_v4f32_v4f64(<4 x float> %va) {
+; CHECK-LABEL: vfpext_v4f32_v4f64:
+; CHECK:       # %bb.0:
+; CHECK-NEXT:    vsetivli zero, 4, e32, m1, ta, ma
+; CHECK-NEXT:    vfwcvt.f.f.v v10, v8
+; CHECK-NEXT:    vmv2r.v v8, v10
+; CHECK-NEXT:    ret
+  %evec = call <4 x double> @llvm.experimental.constrained.fpext.v4f64.v4f32(<4 x float> %va, metadata !"fpexcept.strict")
+  ret <4 x double> %evec
+}
+
+declare <8 x double> @llvm.experimental.constrained.fpext.v8f64.v8f32(<8 x float>, metadata)
+define <8 x double> @vfpext_v8f32_v8f64(<8 x float> %va) {
+; CHECK-LABEL: vfpext_v8f32_v8f64:
+; CHECK:       # %bb.0:
+; CHECK-NEXT:    vsetivli zero, 8, e32, m2, ta, ma
+; CHECK-NEXT:    vfwcvt.f.f.v v12, v8
+; CHECK-NEXT:    vmv4r.v v8, v12
+; CHECK-NEXT:    ret
+  %evec = call <8 x double> @llvm.experimental.constrained.fpext.v8f64.v8f32(<8 x float> %va, metadata !"fpexcept.strict")
+  ret <8 x double> %evec
+}

llvm/test/CodeGen/RISCV/rvv/vfpext-constrained-sdnode.ll

+; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py
+; RUN: llc -mtriple=riscv32 -mattr=+d,+zfh,+experimental-zvfh,+v -target-abi=ilp32d \
+; RUN:     -verify-machineinstrs < %s | FileCheck %s
+; RUN: llc -mtriple=riscv64 -mattr=+d,+zfh,+experimental-zvfh,+v -target-abi=lp64d \
+; RUN:     -verify-machineinstrs < %s | FileCheck %s
+
+declare <vscale x 1 x float> @llvm.experimental.constrained.fpext.nxv1f32.nxv1f16(<vscale x 1 x half>, metadata)
+define <vscale x 1 x float> @vfpext_nxv1f16_nxv1f32(<vscale x 1 x half> %va) {
+; CHECK-LABEL: vfpext_nxv1f16_nxv1f32:
+; CHECK:       # %bb.0:
+; CHECK-NEXT:    vsetvli a0, zero, e16, mf4, ta, ma
+; CHECK-NEXT:    vfwcvt.f.f.v v9, v8
+; CHECK-NEXT:    vmv1r.v v8, v9
+; CHECK-NEXT:    ret
+  %evec = call <vscale x 1 x float> @llvm.experimental.constrained.fpext.nxv1f32.nxv1f16(<vscale x 1 x half> %va, metadata !"fpexcept.strict")
+  ret <vscale x 1 x float> %evec
+}
+
+declare <vscale x 1 x double> @llvm.experimental.constrained.fpext.nxv1f64.nxv1f16(<vscale x 1 x half>, metadata)
+define <vscale x 1 x double> @vfpext_nxv1f16_nxv1f64(<vscale x 1 x half> %va) {
+; CHECK-LABEL: vfpext_nxv1f16_nxv1f64:
+; CHECK:       # %bb.0:
+; CHECK-NEXT:    vsetvli a0, zero, e16, mf4, ta, ma
+; CHECK-NEXT:    vfwcvt.f.f.v v9, v8
+; CHECK-NEXT:    vsetvli zero, zero, e32, mf2, ta, ma
+; CHECK-NEXT:    vfwcvt.f.f.v v8, v9
+; CHECK-NEXT:    ret
+  %evec = call <vscale x 1 x double> @llvm.experimental.constrained.fpext.nxv1f64.nxv1f16(<vscale x 1 x half> %va, metadata !"fpexcept.strict")
+  ret <vscale x 1 x double> %evec
+}
+
+declare <vscale x 2 x float> @llvm.experimental.constrained.fpext.nxv2f32.nxv2f16(<vscale x 2 x half>, metadata)
+define <vscale x 2 x float> @vfpext_nxv2f16_nxv2f32(<vscale x 2 x half> %va) {
+; CHECK-LABEL: vfpext_nxv2f16_nxv2f32:
+; CHECK:       # %bb.0:
+; CHECK-NEXT:    vsetvli a0, zero, e16, mf2, ta, ma
+; CHECK-NEXT:    vfwcvt.f.f.v v9, v8
+; CHECK-NEXT:    vmv1r.v v8, v9
+; CHECK-NEXT:    ret
+  %evec = call <vscale x 2 x float> @llvm.experimental.constrained.fpext.nxv2f32.nxv2f16(<vscale x 2 x half> %va, metadata !"fpexcept.strict")
+  ret <vscale x 2 x float> %evec
+}
+
+declare <vscale x 2 x double> @llvm.experimental.constrained.fpext.nxv2f64.nxv2f16(<vscale x 2 x half>, metadata)
+define <vscale x 2 x double> @vfpext_nxv2f16_nxv2f64(<vscale x 2 x half> %va) {
+; CHECK-LABEL: vfpext_nxv2f16_nxv2f64:
+; CHECK:       # %bb.0:
+; CHECK-NEXT:    vsetvli a0, zero, e16, mf2, ta, ma
+; CHECK-NEXT:    vfwcvt.f.f.v v10, v8
+; CHECK-NEXT:    vsetvli zero, zero, e32, m1, ta, ma
+; CHECK-NEXT:    vfwcvt.f.f.v v8, v10
+; CHECK-NEXT:    ret
+  %evec = call <vscale x 2 x double> @llvm.experimental.constrained.fpext.nxv2f64.nxv2f16(<vscale x 2 x half> %va, metadata !"fpexcept.strict")
+  ret <vscale x 2 x double> %evec
+}
+
+declare <vscale x 4 x float> @llvm.experimental.constrained.fpext.nxv4f32.nxv4f16(<vscale x 4 x half>, metadata)
+define <vscale x 4 x float> @vfpext_nxv4f16_nxv4f32(<vscale x 4 x half> %va) {
+; CHECK-LABEL: vfpext_nxv4f16_nxv4f32:
+; CHECK:       # %bb.0:
+; CHECK-NEXT:    vsetvli a0, zero, e16, m1, ta, ma
+; CHECK-NEXT:    vfwcvt.f.f.v v10, v8
+; CHECK-NEXT:    vmv2r.v v8, v10
+; CHECK-NEXT:    ret
+  %evec = call <vscale x 4 x float> @llvm.experimental.constrained.fpext.nxv4f32.nxv4f16(<vscale x 4 x half> %va, metadata !"fpexcept.strict")
+  ret <vscale x 4 x float> %evec
+}
+
+declare <vscale x 4 x double> @llvm.experimental.constrained.fpext.nxv4f64.nxv4f16(<vscale x 4 x half>, metadata)
+define <vscale x 4 x double> @vfpext_nxv4f16_nxv4f64(<vscale x 4 x half> %va) {
+; CHECK-LABEL: vfpext_nxv4f16_nxv4f64:
+; CHECK:       # %bb.0:
+; CHECK-NEXT:    vsetvli a0, zero, e16, m1, ta, ma
+; CHECK-NEXT:    vfwcvt.f.f.v v12, v8
+; CHECK-NEXT:    vsetvli zero, zero, e32, m2, ta, ma
+; CHECK-NEXT:    vfwcvt.f.f.v v8, v12
+; CHECK-NEXT:    ret
+  %evec = call <vscale x 4 x double> @llvm.experimental.constrained.fpext.nxv4f64.nxv4f16(<vscale x 4 x half> %va, metadata !"fpexcept.strict")
+  ret <vscale x 4 x double> %evec
+}
+
+declare <vscale x 8 x float> @llvm.experimental.constrained.fpext.nxv8f32.nxv8f16(<vscale x 8 x half>, metadata)
+define <vscale x 8 x float> @vfpext_nxv8f16_nxv8f32(<vscale x 8 x half> %va) {
+; CHECK-LABEL: vfpext_nxv8f16_nxv8f32:
+; CHECK:       # %bb.0:
+; CHECK-NEXT:    vsetvli a0, zero, e16, m2, ta, ma
+; CHECK-NEXT:    vfwcvt.f.f.v v12, v8
+; CHECK-NEXT:    vmv4r.v v8, v12
+; CHECK-NEXT:    ret
+  %evec = call <vscale x 8 x float> @llvm.experimental.constrained.fpext.nxv8f32.nxv8f16(<vscale x 8 x half> %va, metadata !"fpexcept.strict")
+  ret <vscale x 8 x float> %evec
+}
+
+declare <vscale x 8 x double> @llvm.experimental.constrained.fpext.nxv8f64.nxv8f16(<vscale x 8 x half>, metadata)
+define <vscale x 8 x double> @vfpext_nxv8f16_nxv8f64(<vscale x 8 x half> %va) {
+; CHECK-LABEL: vfpext_nxv8f16_nxv8f64:
+; CHECK:       # %bb.0:
+; CHECK-NEXT:    vsetvli a0, zero, e16, m2, ta, ma
+; CHECK-NEXT:    vfwcvt.f.f.v v16, v8
+; CHECK-NEXT:    vsetvli zero, zero, e32, m4, ta, ma
+; CHECK-NEXT:    vfwcvt.f.f.v v8, v16
+; CHECK-NEXT:    ret
+  %evec = call <vscale x 8 x double> @llvm.experimental.constrained.fpext.nxv8f64.nxv8f16(<vscale x 8 x half> %va, metadata !"fpexcept.strict")
+  ret <vscale x 8 x double> %evec
+}
+
+declare <vscale x 1 x double> @llvm.experimental.constrained.fpext.nxv1f64.nxv1f32(<vscale x 1 x float>, metadata)
+define <vscale x 1 x double> @vfpext_nxv1f32_nxv1f64(<vscale x 1 x float> %va) {
+; CHECK-LABEL: vfpext_nxv1f32_nxv1f64:
+; CHECK:       # %bb.0:
+; CHECK-NEXT:    vsetvli a0, zero, e32, mf2, ta, ma
+; CHECK-NEXT:    vfwcvt.f.f.v v9, v8
+; CHECK-NEXT:    vmv1r.v v8, v9
+; CHECK-NEXT:    ret
+  %evec = call <vscale x 1 x double> @llvm.experimental.constrained.fpext.nxv1f64.nxv1f32(<vscale x 1 x float> %va, metadata !"fpexcept.strict")
+  ret <vscale x 1 x double> %evec
+}
+
+declare <vscale x 2 x double> @llvm.experimental.constrained.fpext.nxv2f64.nxv2f32(<vscale x 2 x float>, metadata)
+define <vscale x 2 x double> @vfpext_nxv2f32_nxv2f64(<vscale x 2 x float> %va) {
+; CHECK-LABEL: vfpext_nxv2f32_nxv2f64:
+; CHECK:       # %bb.0:
+; CHECK-NEXT:    vsetvli a0, zero, e32, m1, ta, ma
+; CHECK-NEXT:    vfwcvt.f.f.v v10, v8
+; CHECK-NEXT:    vmv2r.v v8, v10
+; CHECK-NEXT:    ret
+  %evec = call <vscale x 2 x double> @llvm.experimental.constrained.fpext.nxv2f64.nxv2f32(<vscale x 2 x float> %va, metadata !"fpexcept.strict")
+  ret <vscale x 2 x double> %evec
+}
+
+declare <vscale x 4 x double> @llvm.experimental.constrained.fpext.nxv4f64.nxv4f32(<vscale x 4 x float>, metadata)
+define <vscale x 4 x double> @vfpext_nxv4f32_nxv4f64(<vscale x 4 x float> %va) {
+; CHECK-LABEL: vfpext_nxv4f32_nxv4f64:
+; CHECK:       # %bb.0:
+; CHECK-NEXT:    vsetvli a0, zero, e32, m2, ta, ma
+; CHECK-NEXT:    vfwcvt.f.f.v v12, v8
+; CHECK-NEXT:    vmv4r.v v8, v12
+; CHECK-NEXT:    ret
+  %evec = call <vscale x 4 x double> @llvm.experimental.constrained.fpext.nxv4f64.nxv4f32(<vscale x 4 x float> %va, metadata !"fpexcept.strict")
+  ret <vscale x 4 x double> %evec
+}
+
+declare <vscale x 8 x double> @llvm.experimental.constrained.fpext.nxv8f64.nxv8f32(<vscale x 8 x float>, metadata)
+define <vscale x 8 x double> @vfpext_nxv8f32_nxv8f64(<vscale x 8 x float> %va) {
+; CHECK-LABEL: vfpext_nxv8f32_nxv8f64:
+; CHECK:       # %bb.0:
+; CHECK-NEXT:    vsetvli a0, zero, e32, m4, ta, ma
+; CHECK-NEXT:    vfwcvt.f.f.v v16, v8
+; CHECK-NEXT:    vmv8r.v v8, v16
+; CHECK-NEXT:    ret
+  %evec = call <vscale x 8 x double> @llvm.experimental.constrained.fpext.nxv8f64.nxv8f32(<vscale x 8 x float> %va, metadata !"fpexcept.strict")
+  ret <vscale x 8 x double> %evec
+}