[DAG][RISCV] Allow scalable vector ISD::STRICT_FP_ROUND and support vector ISD::STRICT_FP_ROUND for RISC-V.

[fakepaper56]

Summary

The patch customized lower vector type ISD::STRICT_FP_ROUND to RISCVISD::STRICT_FP_ROUND.

[craig.topper]

llvm/lib/Target/RISCV/RISCVISelLowering.cpp

5413

comment is stale

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

61

Doesn't this need to be vfncvt.rod.f.f.w?

[fakepaper56]

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

61

You are right. The intermediate conversion should be round-to-odd.

[fakepaper56]

Refine comments and vfncvt.rod.f.f.w for intermediate rounding.

[craig.topper]

LGTM

llvm/lib/CodeGen/SelectionDAG/SelectionDAG.cpp

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 "
            "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(VTList.VTs[0].isFloatingPoint() &&
            Ops[1].getValueType().isFloatingPoint() &&
            VTList.VTs[0].bitsLT(Ops[1].getValueType()) &&
            isa<ConstantSDNode>(Ops[2]) &&
            (cast<ConstantSDNode>(Ops[2])->getZExtValue() == 0 ||
             cast<ConstantSDNode>(Ops[2])->getZExtValue() == 1) &&
            "Invalid STRICT_FP_ROUND!");

llvm/lib/Target/RISCV/RISCVISelLowering.h

 //===-- RISCVISelLowering.h - RISC-V 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 RISC-V uses to lower LLVM code into a
enum NodeType : unsigned {
   STRICT_FSUB_VL,
   STRICT_FMUL_VL,
   STRICT_FDIV_VL,
   STRICT_FSQRT_VL,
   STRICT_VFMADD_VL,
   STRICT_VFNMADD_VL,
   STRICT_VFMSUB_VL,
   STRICT_VFNMSUB_VL,
+  STRICT_FP_ROUND_VL,
   STRICT_FP_EXTEND_VL,
 
   // WARNING: Do not add anything in the end unless you want the node to
   // have memop! In fact, starting from FIRST_TARGET_MEMORY_OPCODE all
   // opcodes will be thought as target memory ops!
 
   // Load address.
   LA = ISD::FIRST_TARGET_MEMORY_OPCODE,
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 lowerStrictFPExtendOrRoundLike(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;
 

llvm/lib/Target/RISCV/RISCVISelLowering.cpp

const auto SetCommonVFPActions = [&](MVT VT) {
 
       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);
+      setOperationAction({ISD::STRICT_FP_EXTEND, ISD::STRICT_FP_ROUND}, VT,
+                         Custom);
       setOperationAction({ISD::STRICT_FADD, ISD::STRICT_FSUB, ISD::STRICT_FMUL,
                           ISD::STRICT_FDIV, ISD::STRICT_FSQRT, ISD::STRICT_FMA},
                          VT, Legal);
     };
 
     // Sets common extload/truncstore actions on RVV floating-point vector
     // types.
     const auto SetCommonVFPExtLoadTruncStoreActions =
if (Subtarget.useRVVForFixedLengthVectors()) {
         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);
+        setOperationAction({ISD::STRICT_FP_EXTEND, ISD::STRICT_FP_ROUND}, VT,
+                           Custom);
         setOperationAction({ISD::STRICT_FADD, ISD::STRICT_FSUB,
                             ISD::STRICT_FMUL, ISD::STRICT_FDIV,
                             ISD::STRICT_FSQRT, ISD::STRICT_FMA},
                            VT, Custom);
       }
 
       // Custom-legalize bitcasts from fixed-length vectors to scalar types.
       setOperationAction(ISD::BITCAST, {MVT::i8, MVT::i16, MVT::i32, MVT::i64},
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_ROUND:
   case ISD::STRICT_FP_EXTEND:
-    return lowerStrictFPExtend(Op, DAG);
+    return lowerStrictFPExtendOrRoundLike(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,
+SDValue
+RISCVTargetLowering::lowerStrictFPExtendOrRoundLike(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();
+  unsigned ConvOpc = Op.getOpcode() == ISD::STRICT_FP_EXTEND
+                         ? RISCVISD::STRICT_FP_EXTEND_VL
+                         : RISCVISD::STRICT_FP_ROUND_VL;
 
   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) {
+  if ((VT.getVectorElementType() == MVT::f64 &&
+       SrcVT.getVectorElementType() == MVT::f16) ||
+      (VT.getVectorElementType() == MVT::f16 &&
+       SrcVT.getVectorElementType() == MVT::f64)) {
     MVT InterVT = ContainerVT.changeVectorElementType(MVT::f32);
-    Src = DAG.getNode(RISCVISD::STRICT_FP_EXTEND_VL, DL,
-                      DAG.getVTList(InterVT, MVT::Other), Chain, Src, Mask, VL);
+    Src = DAG.getNode(ConvOpc, 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);
+  SDValue Res = DAG.getNode(ConvOpc, 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;
 }
#define NODE_NAME_CASE(NODE)                                                   \
   NODE_NAME_CASE(STRICT_FSUB_VL)
   NODE_NAME_CASE(STRICT_FMUL_VL)
   NODE_NAME_CASE(STRICT_FDIV_VL)
   NODE_NAME_CASE(STRICT_FSQRT_VL)
   NODE_NAME_CASE(STRICT_VFMADD_VL)
   NODE_NAME_CASE(STRICT_VFNMADD_VL)
   NODE_NAME_CASE(STRICT_VFMSUB_VL)
   NODE_NAME_CASE(STRICT_VFNMSUB_VL)
+  NODE_NAME_CASE(STRICT_FP_ROUND_VL)
   NODE_NAME_CASE(STRICT_FP_EXTEND_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_RISCVFPRoundOp_VL  : SDTypeProfile<1, 3, [
   SDTCVecEltisVT<2, i1>, SDTCisSameNumEltsAs<1, 2>, SDTCisVT<3, XLenVT>
 ]>;
 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_strict_fpround_vl : SDNode<"RISCVISD::STRICT_FP_ROUND_VL", SDT_RISCVFPRoundOp_VL, [SDNPHasChain]>;
 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_fpround_vl : PatFrags<(ops node:$src, node:$mask, node:$vl),
+                            [(riscv_fpround_vl node:$src, node:$mask, node:$vl),
+                             (riscv_strict_fpround_vl node:$src, node:$mask, node:$vl)]>;
 def any_riscv_fpextend_vl : PatFrags<(ops node:$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>
 ]>;
foreach fvti = AllFloatVectors in {
   defm : VPatNConvertI2FPVL_W<riscv_sint_to_fp_vl, "PseudoVFNCVT_F_X_W">;
 
   defm : VPatNConvertI2FP_RM_VL_W<riscv_vfcvt_rm_f_xu_vl, "PseudoVFNCVT_RM_F_XU_W">;
   defm : VPatNConvertI2FP_RM_VL_W<riscv_vfcvt_rm_f_x_vl, "PseudoVFNCVT_RM_F_X_W">;
 
   foreach fvtiToFWti = AllWidenableFloatVectors in {
     defvar fvti = fvtiToFWti.Vti;
     defvar fwti = fvtiToFWti.Wti;
-    def : Pat<(fvti.Vector (riscv_fpround_vl (fwti.Vector fwti.RegClass:$rs1),
-                                             (fwti.Mask V0),
-                                             VLOpFrag)),
+    def : Pat<(fvti.Vector (any_riscv_fpround_vl
+                               (fwti.Vector fwti.RegClass:$rs1),
+                               (fwti.Mask V0), VLOpFrag)),
               (!cast<Instruction>("PseudoVFNCVT_F_F_W_"#fvti.LMul.MX#"_MASK")
                   (fvti.Vector (IMPLICIT_DEF)), fwti.RegClass:$rs1,
                   (fwti.Mask V0), GPR:$vl, fvti.Log2SEW, TA_MA)>;
 
     def : Pat<(fvti.Vector (riscv_fncvt_rod_vl (fwti.Vector fwti.RegClass:$rs1),
                                                (fwti.Mask V0),
                                                VLOpFrag)),
               (!cast<Instruction>("PseudoVFNCVT_ROD_F_F_W_"#fvti.LMul.MX#"_MASK")

llvm/test/CodeGen/RISCV/rvv/fixed-vectors-vfptrunc-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 <2 x float> @llvm.experimental.constrained.fptrunc.v2f32.v2f64(<2 x double>, metadata, metadata)
+define <2 x float> @vfptrunc_v2f64_v2f32(<2 x double> %va) {
+; CHECK-LABEL: vfptrunc_v2f64_v2f32:
+; CHECK:       # %bb.0:
+; CHECK-NEXT:    vsetivli zero, 2, e32, mf2, ta, ma
+; CHECK-NEXT:    vfncvt.f.f.w v9, v8
+; CHECK-NEXT:    vmv1r.v v8, v9
+; CHECK-NEXT:    ret
+  %evec = call <2 x float> @llvm.experimental.constrained.fptrunc.v2f32.v2f64(<2 x double> %va, metadata !"round.dynamic", metadata !"fpexcept.strict")
+  ret <2 x float> %evec
+}
+
+declare <2 x half> @llvm.experimental.constrained.fptrunc.v2f16.v2f64(<2 x double>, metadata, metadata)
+define <2 x half> @vfptrunc_v2f64_v2f16(<2 x double> %va) {
+; CHECK-LABEL: vfptrunc_v2f64_v2f16:
+; CHECK:       # %bb.0:
+; CHECK-NEXT:    vsetivli zero, 2, e32, mf2, ta, ma
+; CHECK-NEXT:    vfncvt.f.f.w v9, v8
+; CHECK-NEXT:    vsetvli zero, zero, e16, mf4, ta, ma
+; CHECK-NEXT:    vfncvt.f.f.w v8, v9
+; CHECK-NEXT:    ret
+  %evec = call <2 x half> @llvm.experimental.constrained.fptrunc.v2f16.v2f64(<2 x double> %va, metadata !"round.dynamic", metadata !"fpexcept.strict")
+  ret <2 x half> %evec
+}
+
+declare <2 x half> @llvm.experimental.constrained.fptrunc.v2f16.v2f32(<2 x float>, metadata, metadata)
+define <2 x half> @vfptrunc_v2f32_v2f16(<2 x float> %va) {
+; CHECK-LABEL: vfptrunc_v2f32_v2f16:
+; CHECK:       # %bb.0:
+; CHECK-NEXT:    vsetivli zero, 2, e16, mf4, ta, ma
+; CHECK-NEXT:    vfncvt.f.f.w v9, v8
+; CHECK-NEXT:    vmv1r.v v8, v9
+; CHECK-NEXT:    ret
+  %evec = call <2 x half> @llvm.experimental.constrained.fptrunc.v2f16.v2f32(<2 x float> %va, metadata !"round.dynamic", metadata !"fpexcept.strict")
+  ret <2 x half> %evec
+}
+
+declare <4 x float> @llvm.experimental.constrained.fptrunc.v4f32.v4f64(<4 x double>, metadata, metadata)
+define <4 x float> @vfptrunc_v4f64_v4f32(<4 x double> %va) {
+; CHECK-LABEL: vfptrunc_v4f64_v4f32:
+; CHECK:       # %bb.0:
+; CHECK-NEXT:    vsetivli zero, 4, e32, m1, ta, ma
+; CHECK-NEXT:    vfncvt.f.f.w v10, v8
+; CHECK-NEXT:    vmv.v.v v8, v10
+; CHECK-NEXT:    ret
+  %evec = call <4 x float> @llvm.experimental.constrained.fptrunc.v4f32.v4f64(<4 x double> %va, metadata !"round.dynamic", metadata !"fpexcept.strict")
+  ret <4 x float> %evec
+}
+
+declare <4 x half> @llvm.experimental.constrained.fptrunc.v4f16.v4f64(<4 x double>, metadata, metadata)
+define <4 x half> @vfptrunc_v4f64_v4f16(<4 x double> %va) {
+; CHECK-LABEL: vfptrunc_v4f64_v4f16:
+; CHECK:       # %bb.0:
+; CHECK-NEXT:    vsetivli zero, 4, e32, m1, ta, ma
+; CHECK-NEXT:    vfncvt.f.f.w v10, v8
+; CHECK-NEXT:    vsetvli zero, zero, e16, mf2, ta, ma
+; CHECK-NEXT:    vfncvt.f.f.w v8, v10
+; CHECK-NEXT:    ret
+  %evec = call <4 x half> @llvm.experimental.constrained.fptrunc.v4f16.v4f64(<4 x double> %va, metadata !"round.dynamic", metadata !"fpexcept.strict")
+  ret <4 x half> %evec
+}
+
+declare <4 x half> @llvm.experimental.constrained.fptrunc.v4f16.v4f32(<4 x float>, metadata, metadata)
+define <4 x half> @vfptrunc_v4f32_v4f16(<4 x float> %va) {
+; CHECK-LABEL: vfptrunc_v4f32_v4f16:
+; CHECK:       # %bb.0:
+; CHECK-NEXT:    vsetivli zero, 4, e16, mf2, ta, ma
+; CHECK-NEXT:    vfncvt.f.f.w v9, v8
+; CHECK-NEXT:    vmv1r.v v8, v9
+; CHECK-NEXT:    ret
+  %evec = call <4 x half> @llvm.experimental.constrained.fptrunc.v4f16.v4f32(<4 x float> %va, metadata !"round.dynamic", metadata !"fpexcept.strict")
+  ret <4 x half> %evec
+}
+
+declare <8 x float> @llvm.experimental.constrained.fptrunc.v8f32.v8f64(<8 x double>, metadata, metadata)
+define <8 x float> @vfptrunc_v8f64_v8f32(<8 x double> %va) {
+; CHECK-LABEL: vfptrunc_v8f64_v8f32:
+; CHECK:       # %bb.0:
+; CHECK-NEXT:    vsetivli zero, 8, e32, m2, ta, ma
+; CHECK-NEXT:    vfncvt.f.f.w v12, v8
+; CHECK-NEXT:    vmv.v.v v8, v12
+; CHECK-NEXT:    ret
+  %evec = call <8 x float> @llvm.experimental.constrained.fptrunc.v8f32.v8f64(<8 x double> %va, metadata !"round.dynamic", metadata !"fpexcept.strict")
+  ret <8 x float> %evec
+}
+
+declare <8 x half> @llvm.experimental.constrained.fptrunc.v8f16.v8f64(<8 x double>, metadata, metadata)
+define <8 x half> @vfptrunc_v8f64_v8f16(<8 x double> %va) {
+; CHECK-LABEL: vfptrunc_v8f64_v8f16:
+; CHECK:       # %bb.0:
+; CHECK-NEXT:    vsetivli zero, 8, e32, m2, ta, ma
+; CHECK-NEXT:    vfncvt.f.f.w v12, v8
+; CHECK-NEXT:    vsetvli zero, zero, e16, m1, ta, ma
+; CHECK-NEXT:    vfncvt.f.f.w v8, v12
+; CHECK-NEXT:    ret
+  %evec = call <8 x half> @llvm.experimental.constrained.fptrunc.v8f16.v8f64(<8 x double> %va, metadata !"round.dynamic", metadata !"fpexcept.strict")
+  ret <8 x half> %evec
+}
+
+declare <8 x half> @llvm.experimental.constrained.fptrunc.v8f16.v8f32(<8 x float>, metadata, metadata)
+define <8 x half> @vfptrunc_v8f32_v8f16(<8 x float> %va) {
+; CHECK-LABEL: vfptrunc_v8f32_v8f16:
+; CHECK:       # %bb.0:
+; CHECK-NEXT:    vsetivli zero, 8, e16, m1, ta, ma
+; CHECK-NEXT:    vfncvt.f.f.w v10, v8
+; CHECK-NEXT:    vmv.v.v v8, v10
+; CHECK-NEXT:    ret
+  %evec = call <8 x half> @llvm.experimental.constrained.fptrunc.v8f16.v8f32(<8 x float> %va, metadata !"round.dynamic", metadata !"fpexcept.strict")
+  ret <8 x half> %evec
+}

llvm/test/CodeGen/RISCV/rvv/vfptrunc-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.fptrunc.nxv1f32.nxv1f64(<vscale x 1 x double>, metadata, metadata)
+define <vscale x 1 x float> @vfptrunc_nxv1f64_nxv1f32(<vscale x 1 x double> %va) {
+; CHECK-LABEL: vfptrunc_nxv1f64_nxv1f32:
+; CHECK:       # %bb.0:
+; CHECK-NEXT:    vsetvli a0, zero, e32, mf2, ta, ma
+; CHECK-NEXT:    vfncvt.f.f.w v9, v8
+; CHECK-NEXT:    vmv1r.v v8, v9
+; CHECK-NEXT:    ret
+  %evec = call <vscale x 1 x float> @llvm.experimental.constrained.fptrunc.nxv1f32.nxv1f64(<vscale x 1 x double> %va, metadata !"round.dynamic", metadata !"fpexcept.strict")
+  ret <vscale x 1 x float> %evec
+}
+
+declare <vscale x 1 x half> @llvm.experimental.constrained.fptrunc.nxv1f16.nxv1f64(<vscale x 1 x double>, metadata, metadata)
+define <vscale x 1 x half> @vfptrunc_nxv1f64_nxv1f16(<vscale x 1 x double> %va) {
+; CHECK-LABEL: vfptrunc_nxv1f64_nxv1f16:
+; CHECK:       # %bb.0:
+; CHECK-NEXT:    vsetvli a0, zero, e32, mf2, ta, ma
+; CHECK-NEXT:    vfncvt.f.f.w v9, v8
+; CHECK-NEXT:    vsetvli zero, zero, e16, mf4, ta, ma
+; CHECK-NEXT:    vfncvt.f.f.w v8, v9
+; CHECK-NEXT:    ret
+  %evec = call <vscale x 1 x half> @llvm.experimental.constrained.fptrunc.nxv1f16.nxv1f64(<vscale x 1 x double> %va, metadata !"round.dynamic", metadata !"fpexcept.strict")
+  ret <vscale x 1 x half> %evec
+}
+
+declare <vscale x 1 x half> @llvm.experimental.constrained.fptrunc.nxv1f16.nxv1f32(<vscale x 1 x float>, metadata, metadata)
+define <vscale x 1 x half> @vfptrunc_nxv1f32_nxv1f16(<vscale x 1 x float> %va) {
+; CHECK-LABEL: vfptrunc_nxv1f32_nxv1f16:
+; CHECK:       # %bb.0:
+; CHECK-NEXT:    vsetvli a0, zero, e16, mf4, ta, ma
+; CHECK-NEXT:    vfncvt.f.f.w v9, v8
+; CHECK-NEXT:    vmv1r.v v8, v9
+; CHECK-NEXT:    ret
+  %evec = call <vscale x 1 x half> @llvm.experimental.constrained.fptrunc.nxv1f16.nxv1f32(<vscale x 1 x float> %va, metadata !"round.dynamic", metadata !"fpexcept.strict")
+  ret <vscale x 1 x half> %evec
+}
+
+declare <vscale x 2 x float> @llvm.experimental.constrained.fptrunc.nxv2f32.nxv2f64(<vscale x 2 x double>, metadata, metadata)
+define <vscale x 2 x float> @vfptrunc_nxv2f64_nxv2f32(<vscale x 2 x double> %va) {
+; CHECK-LABEL: vfptrunc_nxv2f64_nxv2f32:
+; CHECK:       # %bb.0:
+; CHECK-NEXT:    vsetvli a0, zero, e32, m1, ta, ma
+; CHECK-NEXT:    vfncvt.f.f.w v10, v8
+; CHECK-NEXT:    vmv.v.v v8, v10
+; CHECK-NEXT:    ret
+  %evec = call <vscale x 2 x float> @llvm.experimental.constrained.fptrunc.nxv2f32.nxv2f64(<vscale x 2 x double> %va, metadata !"round.dynamic", metadata !"fpexcept.strict")
+  ret <vscale x 2 x float> %evec
+}
+
+declare <vscale x 2 x half> @llvm.experimental.constrained.fptrunc.nxv2f16.nxv2f64(<vscale x 2 x double>, metadata, metadata)
+define <vscale x 2 x half> @vfptrunc_nxv2f64_nxv2f16(<vscale x 2 x double> %va) {
+; CHECK-LABEL: vfptrunc_nxv2f64_nxv2f16:
+; CHECK:       # %bb.0:
+; CHECK-NEXT:    vsetvli a0, zero, e32, m1, ta, ma
+; CHECK-NEXT:    vfncvt.f.f.w v10, v8
+; CHECK-NEXT:    vsetvli zero, zero, e16, mf2, ta, ma
+; CHECK-NEXT:    vfncvt.f.f.w v8, v10
+; CHECK-NEXT:    ret
+  %evec = call <vscale x 2 x half> @llvm.experimental.constrained.fptrunc.nxv2f16.nxv2f64(<vscale x 2 x double> %va, metadata !"round.dynamic", metadata !"fpexcept.strict")
+  ret <vscale x 2 x half> %evec
+}
+
+declare <vscale x 2 x half> @llvm.experimental.constrained.fptrunc.nxv2f16.nxv2f32(<vscale x 2 x float>, metadata, metadata)
+define <vscale x 2 x half> @vfptrunc_nxv2f32_nxv2f16(<vscale x 2 x float> %va) {
+; CHECK-LABEL: vfptrunc_nxv2f32_nxv2f16:
+; CHECK:       # %bb.0:
+; CHECK-NEXT:    vsetvli a0, zero, e16, mf2, ta, ma
+; CHECK-NEXT:    vfncvt.f.f.w v9, v8
+; CHECK-NEXT:    vmv1r.v v8, v9
+; CHECK-NEXT:    ret
+  %evec = call <vscale x 2 x half> @llvm.experimental.constrained.fptrunc.nxv2f16.nxv2f32(<vscale x 2 x float> %va, metadata !"round.dynamic", metadata !"fpexcept.strict")
+  ret <vscale x 2 x half> %evec
+}
+
+declare <vscale x 4 x float> @llvm.experimental.constrained.fptrunc.nxv4f32.nxv4f64(<vscale x 4 x double>, metadata, metadata)
+define <vscale x 4 x float> @vfptrunc_nxv4f64_nxv4f32(<vscale x 4 x double> %va) {
+; CHECK-LABEL: vfptrunc_nxv4f64_nxv4f32:
+; CHECK:       # %bb.0:
+; CHECK-NEXT:    vsetvli a0, zero, e32, m2, ta, ma
+; CHECK-NEXT:    vfncvt.f.f.w v12, v8
+; CHECK-NEXT:    vmv.v.v v8, v12
+; CHECK-NEXT:    ret
+  %evec = call <vscale x 4 x float> @llvm.experimental.constrained.fptrunc.nxv4f32.nxv4f64(<vscale x 4 x double> %va, metadata !"round.dynamic", metadata !"fpexcept.strict")
+  ret <vscale x 4 x float> %evec
+}
+
+declare <vscale x 4 x half> @llvm.experimental.constrained.fptrunc.nxv4f16.nxv4f64(<vscale x 4 x double>, metadata, metadata)
+define <vscale x 4 x half> @vfptrunc_nxv4f64_nxv4f16(<vscale x 4 x double> %va) {
+; CHECK-LABEL: vfptrunc_nxv4f64_nxv4f16:
+; CHECK:       # %bb.0:
+; CHECK-NEXT:    vsetvli a0, zero, e32, m2, ta, ma
+; CHECK-NEXT:    vfncvt.f.f.w v12, v8
+; CHECK-NEXT:    vsetvli zero, zero, e16, m1, ta, ma
+; CHECK-NEXT:    vfncvt.f.f.w v8, v12
+; CHECK-NEXT:    ret
+  %evec = call <vscale x 4 x half> @llvm.experimental.constrained.fptrunc.nxv4f16.nxv4f64(<vscale x 4 x double> %va, metadata !"round.dynamic", metadata !"fpexcept.strict")
+  ret <vscale x 4 x half> %evec
+}
+
+declare <vscale x 4 x half> @llvm.experimental.constrained.fptrunc.nxv4f16.nxv4f32(<vscale x 4 x float>, metadata, metadata)
+define <vscale x 4 x half> @vfptrunc_nxv4f32_nxv4f16(<vscale x 4 x float> %va) {
+; CHECK-LABEL: vfptrunc_nxv4f32_nxv4f16:
+; CHECK:       # %bb.0:
+; CHECK-NEXT:    vsetvli a0, zero, e16, m1, ta, ma
+; CHECK-NEXT:    vfncvt.f.f.w v10, v8
+; CHECK-NEXT:    vmv.v.v v8, v10
+; CHECK-NEXT:    ret
+  %evec = call <vscale x 4 x half> @llvm.experimental.constrained.fptrunc.nxv4f16.nxv4f32(<vscale x 4 x float> %va, metadata !"round.dynamic", metadata !"fpexcept.strict")
+  ret <vscale x 4 x half> %evec
+}
+
+declare <vscale x 8 x float> @llvm.experimental.constrained.fptrunc.nxv8f32.nxv8f64(<vscale x 8 x double>, metadata, metadata)
+define <vscale x 8 x float> @vfptrunc_nxv8f64_nxv8f32(<vscale x 8 x double> %va) {
+; CHECK-LABEL: vfptrunc_nxv8f64_nxv8f32:
+; CHECK:       # %bb.0:
+; CHECK-NEXT:    vsetvli a0, zero, e32, m4, ta, ma
+; CHECK-NEXT:    vfncvt.f.f.w v16, v8
+; CHECK-NEXT:    vmv.v.v v8, v16
+; CHECK-NEXT:    ret
+  %evec = call <vscale x 8 x float> @llvm.experimental.constrained.fptrunc.nxv8f32.nxv8f64(<vscale x 8 x double> %va, metadata !"round.dynamic", metadata !"fpexcept.strict")
+  ret <vscale x 8 x float> %evec
+}
+
+declare <vscale x 8 x half> @llvm.experimental.constrained.fptrunc.nxv8f16.nxv8f64(<vscale x 8 x double>, metadata, metadata)
+define <vscale x 8 x half> @vfptrunc_nxv8f64_nxv8f16(<vscale x 8 x double> %va) {
+; CHECK-LABEL: vfptrunc_nxv8f64_nxv8f16:
+; CHECK:       # %bb.0:
+; CHECK-NEXT:    vsetvli a0, zero, e32, m4, ta, ma
+; CHECK-NEXT:    vfncvt.f.f.w v16, v8
+; CHECK-NEXT:    vsetvli zero, zero, e16, m2, ta, ma
+; CHECK-NEXT:    vfncvt.f.f.w v8, v16
+; CHECK-NEXT:    ret
+  %evec = call <vscale x 8 x half> @llvm.experimental.constrained.fptrunc.nxv8f16.nxv8f64(<vscale x 8 x double> %va, metadata !"round.dynamic", metadata !"fpexcept.strict")
+  ret <vscale x 8 x half> %evec
+}
+
+declare <vscale x 8 x half> @llvm.experimental.constrained.fptrunc.nxv8f16.nxv8f32(<vscale x 8 x float>, metadata, metadata)
+define <vscale x 8 x half> @vfptrunc_nxv8f32_nxv8f16(<vscale x 8 x float> %va) {
+; CHECK-LABEL: vfptrunc_nxv8f32_nxv8f16:
+; CHECK:       # %bb.0:
+; CHECK-NEXT:    vsetvli a0, zero, e16, m2, ta, ma
+; CHECK-NEXT:    vfncvt.f.f.w v12, v8
+; CHECK-NEXT:    vmv.v.v v8, v12
+; CHECK-NEXT:    ret
+  %evec = call <vscale x 8 x half> @llvm.experimental.constrained.fptrunc.nxv8f16.nxv8f32(<vscale x 8 x float> %va, metadata !"round.dynamic", metadata !"fpexcept.strict")
+  ret <vscale x 8 x half> %evec
+}