[SVE][CodeGen] Lower floating point -> integer conversions

[kmclaughlin]

Summary

This patch adds new ISD nodes, FCVTZS_MERGE_PASSTHRU &
FCVTZU_MERGE_PASSTHRU, which are used to lower the
following intrinsics:

• llvm.aarch64.sve.fcvtzu
• llvm.aarch64.sve.fcvtzs

[paulwalker-arm]

llvm/lib/Target/AArch64/AArch64ISelLowering.cpp

12342–12344

This doesn't look correct. You are dropping the passthru value (i.e. Operand(1)) and always using UNDEF instead. It's a bit troubling that such a change hasn't broken any of the lit tests.

I think you'll want to treat these like the other unary operations, which have matching _MERGE_PASSTHRU ISD nodes rather than _PRED ones.

[kmclaughlin]

• Renamed ISD nodes to FP_TO_[S|U]INT_MERGE_PASSTHRU
• Lower only the llvm.aarch64.sve.fcvtz[u|s] intrinsics to these nodes and not those with a type suffix (e.g. llvm.aarch64.sve.fcvtzs.i32f64)
• Removed combineSVEConversionFP() and added the new nodes to the list in isMergePassthruOpcode()

[efriedma]

llvm/lib/Target/AArch64/AArch64ISelLowering.h

84

Probably should be named FP_TO_UINT_SAT_MERGE_PASSTHRU or something like that, if you're going to use it to lower llvm.aarch64.sve.fcvtzu. (See D54749)

[paulwalker-arm]

llvm/lib/Target/AArch64/AArch64ISelLowering.h

84

I had a feeling you might say this. The problem is this is more than a rename because the SAT nodes take an additional parameter. This is not really a hurdle but there's a question as to where you draw the line. All the intrinsics expect to produce results that match the instructions on which they're based. This would suggest we can never use names based on similar ISD nodes because the common nodes do not define all corner cases. This seems like overkill considering the _PRED/_MERGE nodes are all under the AArch64ISD namespace, which to me suggests they follow AArch64 semantics.

[efriedma]

llvm/lib/Target/AArch64/AArch64ISelLowering.h

84

I think if there's some significant functionality difference, we should call it out in the name somehow. There's a very high likelihood that these nodes, or something like them, will eventually become target-independent. And even before that, people will likely assume they're equivalent based on the name. I don't want something to subtly break because someone didn't realize there was a difference. (It's particularly terrible in cases like fp->int conversions and shifts: if you confuse them, the result appears to mostly work, but you'll get a bug report years later.)

If you're not comfortable naming FP_TO_UINT_SAT_MERGE_PASSTHRU without the matching width parameter, we could just call it FCVTZU_MERGE_PASSTHRU .

[paulwalker-arm]

llvm/lib/Target/AArch64/AArch64ISelLowering.h

84

Yes I much prefer that naming strategy. Given this change I guess there are other nodes that should be renamed.

[kmclaughlin]

• Renamed ISD nodes to FCVTZ[S|U]_MERGE_PASSTHRU
• Renamed AArch64fcvtz[s|u]_p to AArch64fcvtz[s|u]_mt & moved their definitions to the block containing other _mt nodes in AArch64SVEInstrInfo.td

[efriedma]

llvm/lib/CodeGen/SelectionDAG/LegalizeIntegerTypes.cpp

562

Why is isScalableVector() relevant here?

[kmclaughlin]

• Removed isScalableVector() check from PromoteIntRes_FP_TO_XINT
• Updated fptoui tests where the result type is promoted to check for fcvtzu

llvm/lib/CodeGen/SelectionDAG/LegalizeIntegerTypes.cpp

562

It's not really relevant here, when I first created the patch I thought that it was better not to replace FP_TO_UINT with FP_TO_SINT when promoting the result, but I think it's fine to do this for scalable types also.
I've removed the isScalableVector() from here and updated the affected tests (e.g. nxv2f64 -> nxv2i32) to check for fcvtzs.

[paulwalker-arm]

LGTM assuming adding the missing tests don't throw up anything unexpected.

llvm/test/CodeGen/AArch64/sve-fcvt.ll

22

Perhaps worth adding tests for:
fcvtz[s,u]_h_nxv2f32
fcvtz[s,u]_h_nxv2f64
fcvtz[s,u]_h_nxv4f32

127

I know it's in the langref but it cannot hurt to add a comment here along the lines of NOTE: Using fcvtzs is safe because fptoui overflow is considered poison and a 64bit signed value encompasses the entire range of a 16bit unsigned value. What do you think?

[paulwalker-arm]

Oh and the commit message should reference the lowering of FP_TO_SINT/FP_TO_UINT.

[kmclaughlin]

llvm/test/CodeGen/AArch64/sve-fcvt.ll

127

I've added your comment in here, it will definitely help me to remember why this is safe in future :)

llvm/lib/CodeGen/SelectionDAG/LegalizeIntegerTypes.cpp

SDValue DAGTypeLegalizer::PromoteIntRes_FP_TO_XINT(SDNode *N) {
   EVT NVT = TLI.getTypeToTransformTo(*DAG.getContext(), N->getValueType(0));
   unsigned NewOpc = N->getOpcode();
   SDLoc dl(N);
 
   // If we're promoting a UINT to a larger size and the larger FP_TO_UINT is
   // not Legal, check to see if we can use FP_TO_SINT instead.  (If both UINT
   // and SINT conversions are Custom, there is no way to tell which is
   // preferable. We choose SINT because that's the right thing on PPC.)
-  if (N->getOpcode() == ISD::FP_TO_UINT &&
+  if (!NVT.isScalableVector() && N->getOpcode() == ISD::FP_TO_UINT &&
       !TLI.isOperationLegal(ISD::FP_TO_UINT, NVT) &&
       TLI.isOperationLegalOrCustom(ISD::FP_TO_SINT, NVT))
     NewOpc = ISD::FP_TO_SINT;
 
   if (N->getOpcode() == ISD::STRICT_FP_TO_UINT &&
       !TLI.isOperationLegal(ISD::STRICT_FP_TO_UINT, NVT) &&
       TLI.isOperationLegalOrCustom(ISD::STRICT_FP_TO_SINT, NVT))
     NewOpc = ISD::STRICT_FP_TO_SINT;

llvm/lib/Target/AArch64/AArch64ISelLowering.h

enum NodeType : unsigned {
   ADD_PRED,
   FADD_PRED,
   FDIV_PRED,
   FMA_PRED,
   FMAXNM_PRED,
   FMINNM_PRED,
   FMUL_PRED,
   FSUB_PRED,
   MUL_PRED,
   SDIV_PRED,
   SHL_PRED,
   SMAX_PRED,
   SMIN_PRED,
   SRA_PRED,
   SRL_PRED,
   SUB_PRED,
   UDIV_PRED,
   UMAX_PRED,
   UMIN_PRED,
 
   // Predicated instructions with the result of inactive lanes provided by the
   // last operand.
   FCEIL_MERGE_PASSTHRU,
   FFLOOR_MERGE_PASSTHRU,
   FNEARBYINT_MERGE_PASSTHRU,
   FNEG_MERGE_PASSTHRU,
   FRINT_MERGE_PASSTHRU,
   FROUND_MERGE_PASSTHRU,
   FROUNDEVEN_MERGE_PASSTHRU,
   FTRUNC_MERGE_PASSTHRU,
+  FCVTZU_MERGE_PASSTHRU,
+  FCVTZS_MERGE_PASSTHRU,
   SIGN_EXTEND_INREG_MERGE_PASSTHRU,
   ZERO_EXTEND_INREG_MERGE_PASSTHRU,
 
   SETCC_MERGE_ZERO,
 
   // Arithmetic instructions which write flags.
   ADDS,
   SUBS,

llvm/lib/Target/AArch64/AArch64ISelLowering.cpp

static bool isMergePassthruOpcode(unsigned Opc) {
   case AArch64ISD::ZERO_EXTEND_INREG_MERGE_PASSTHRU:
   case AArch64ISD::FCEIL_MERGE_PASSTHRU:
   case AArch64ISD::FFLOOR_MERGE_PASSTHRU:
   case AArch64ISD::FNEARBYINT_MERGE_PASSTHRU:
   case AArch64ISD::FRINT_MERGE_PASSTHRU:
   case AArch64ISD::FROUND_MERGE_PASSTHRU:
   case AArch64ISD::FROUNDEVEN_MERGE_PASSTHRU:
   case AArch64ISD::FTRUNC_MERGE_PASSTHRU:
+  case AArch64ISD::FCVTZU_MERGE_PASSTHRU:
+  case AArch64ISD::FCVTZS_MERGE_PASSTHRU:
     return true;
   }
 }
 
 AArch64TargetLowering::AArch64TargetLowering(const TargetMachine &TM,
                                              const AArch64Subtarget &STI)
     : TargetLowering(TM), Subtarget(&STI) {
   // AArch64 doesn't have comparisons which set GPRs or setcc instructions, so
AArch64TargetLowering::AArch64TargetLowering(const TargetMachine &TM,
 
   if (Subtarget->hasSVE()) {
     // FIXME: Add custom lowering of MLOAD to handle different passthrus (not a
     // splat of 0 or undef) once vector selects supported in SVE codegen. See
     // D68877 for more details.
     for (MVT VT : MVT::integer_scalable_vector_valuetypes()) {
       if (isTypeLegal(VT)) {
         setOperationAction(ISD::INSERT_SUBVECTOR, VT, Custom);
+        setOperationAction(ISD::FP_TO_UINT, VT, Custom);
+        setOperationAction(ISD::FP_TO_SINT, VT, Custom);
         setOperationAction(ISD::MUL, VT, Custom);
         setOperationAction(ISD::SPLAT_VECTOR, VT, Custom);
         setOperationAction(ISD::SELECT, VT, Custom);
         setOperationAction(ISD::SDIV, VT, Custom);
         setOperationAction(ISD::UDIV, VT, Custom);
         setOperationAction(ISD::SMIN, VT, Custom);
         setOperationAction(ISD::UMIN, VT, Custom);
         setOperationAction(ISD::SMAX, VT, Custom);
case AArch64ISD::FIRST_NUMBER:
     MAKE_CASE(AArch64ISD::ZERO_EXTEND_INREG_MERGE_PASSTHRU)
     MAKE_CASE(AArch64ISD::FCEIL_MERGE_PASSTHRU)
     MAKE_CASE(AArch64ISD::FFLOOR_MERGE_PASSTHRU)
     MAKE_CASE(AArch64ISD::FNEARBYINT_MERGE_PASSTHRU)
     MAKE_CASE(AArch64ISD::FRINT_MERGE_PASSTHRU)
     MAKE_CASE(AArch64ISD::FROUND_MERGE_PASSTHRU)
     MAKE_CASE(AArch64ISD::FROUNDEVEN_MERGE_PASSTHRU)
     MAKE_CASE(AArch64ISD::FTRUNC_MERGE_PASSTHRU)
+    MAKE_CASE(AArch64ISD::FCVTZU_MERGE_PASSTHRU)
+    MAKE_CASE(AArch64ISD::FCVTZS_MERGE_PASSTHRU)
     MAKE_CASE(AArch64ISD::SETCC_MERGE_ZERO)
     MAKE_CASE(AArch64ISD::ADC)
     MAKE_CASE(AArch64ISD::SBC)
     MAKE_CASE(AArch64ISD::ADDS)
     MAKE_CASE(AArch64ISD::SUBS)
     MAKE_CASE(AArch64ISD::ADCS)
     MAKE_CASE(AArch64ISD::SBCS)
     MAKE_CASE(AArch64ISD::ANDS)

 
 SDValue AArch64TargetLowering::LowerVectorFP_TO_INT(SDValue Op,
                                                     SelectionDAG &DAG) const {
   // Warning: We maintain cost tables in AArch64TargetTransformInfo.cpp.
   // Any additional optimization in this function should be recorded
   // in the cost tables.
   EVT InVT = Op.getOperand(0).getValueType();
   EVT VT = Op.getValueType();
+
+  if (VT.isScalableVector()) {
+    unsigned Opcode = Op.getOpcode() == ISD::FP_TO_UINT
+                          ? AArch64ISD::FCVTZU_MERGE_PASSTHRU
+                          : AArch64ISD::FCVTZS_MERGE_PASSTHRU;
+    return LowerToPredicatedOp(Op, DAG, Opcode);
+  }
+
   unsigned NumElts = InVT.getVectorNumElements();
 
   // f16 conversions are promoted to f32 when full fp16 is not supported.
   if (InVT.getVectorElementType() == MVT::f16 &&
       !Subtarget->hasFullFP16()) {
     MVT NewVT = MVT::getVectorVT(MVT::f32, NumElts);
     SDLoc dl(Op);
     return DAG.getNode(
case Intrinsic::aarch64_sve_frinta:
     return DAG.getNode(AArch64ISD::FROUND_MERGE_PASSTHRU, dl, Op.getValueType(),
                        Op.getOperand(2), Op.getOperand(3), Op.getOperand(1));
   case Intrinsic::aarch64_sve_frintn:
     return DAG.getNode(AArch64ISD::FROUNDEVEN_MERGE_PASSTHRU, dl, Op.getValueType(),
                        Op.getOperand(2), Op.getOperand(3), Op.getOperand(1));
   case Intrinsic::aarch64_sve_frintz:
     return DAG.getNode(AArch64ISD::FTRUNC_MERGE_PASSTHRU, dl, Op.getValueType(),
                        Op.getOperand(2), Op.getOperand(3), Op.getOperand(1));
+  case Intrinsic::aarch64_sve_fcvtzu:
+    return DAG.getNode(AArch64ISD::FCVTZU_MERGE_PASSTHRU, dl,
+                       Op.getValueType(), Op.getOperand(2), Op.getOperand(3),
+                       Op.getOperand(1));
+  case Intrinsic::aarch64_sve_fcvtzs:
+    return DAG.getNode(AArch64ISD::FCVTZS_MERGE_PASSTHRU, dl,
+                       Op.getValueType(), Op.getOperand(2), Op.getOperand(3),
+                       Op.getOperand(1));
   case Intrinsic::aarch64_sve_convert_to_svbool: {
     EVT OutVT = Op.getValueType();
     EVT InVT = Op.getOperand(1).getValueType();
     // Return the operand if the cast isn't changing type,
     // i.e. <n x 16 x i1> -> <n x 16 x i1>
     if (InVT == OutVT)
       return Op.getOperand(1);
     // Otherwise, zero the newly introduced lanes.

 static SDValue combineSVEReductionOrderedFP(SDNode *N, unsigned Opc,
                                             SelectionDAG &DAG) {
   SDLoc DL(N);
 
   SDValue Pred = N->getOperand(1);
   SDValue InitVal = N->getOperand(2);
   SDValue VecToReduce = N->getOperand(3);
   EVT ReduceVT = VecToReduce.getValueType();
 
   // Ordered reductions use the first lane of the result vector as the
   // reduction's initial value.
   SDValue Zero = DAG.getConstant(0, DL, MVT::i64);
   InitVal = DAG.getNode(ISD::INSERT_VECTOR_ELT, DL, ReduceVT,
                         DAG.getUNDEF(ReduceVT), InitVal, Zero);
 
   SDValue Reduce = DAG.getNode(Opc, DL, ReduceVT, Pred, InitVal, VecToReduce);
 
   // SVE reductions set the whole vector register with the first element
   // containing the reduction result, which we'll now extract.

llvm/lib/Target/AArch64/AArch64SVEInstrInfo.td

 def AArch64frintp_mt : SDNode<"AArch64ISD::FCEIL_MERGE_PASSTHRU", SDT_AArch64Arith>;
 def AArch64frintm_mt : SDNode<"AArch64ISD::FFLOOR_MERGE_PASSTHRU", SDT_AArch64Arith>;
 def AArch64frinti_mt : SDNode<"AArch64ISD::FNEARBYINT_MERGE_PASSTHRU", SDT_AArch64Arith>;
 def AArch64frintx_mt : SDNode<"AArch64ISD::FRINT_MERGE_PASSTHRU", SDT_AArch64Arith>;
 def AArch64frinta_mt : SDNode<"AArch64ISD::FROUND_MERGE_PASSTHRU", SDT_AArch64Arith>;
 def AArch64frintn_mt : SDNode<"AArch64ISD::FROUNDEVEN_MERGE_PASSTHRU", SDT_AArch64Arith>;
 def AArch64frintz_mt : SDNode<"AArch64ISD::FTRUNC_MERGE_PASSTHRU", SDT_AArch64Arith>;
 
+def SDT_AArch64FCVT : SDTypeProfile<1, 3, [
+  SDTCisVec<0>, SDTCisVec<1>, SDTCisVec<2>, SDTCisVec<3>,
+  SDTCVecEltisVT<1,i1>
+]>;
+
+def AArch64fcvtzu_mt : SDNode<"AArch64ISD::FCVTZU_MERGE_PASSTHRU", SDT_AArch64FCVT>;
+def AArch64fcvtzs_mt : SDNode<"AArch64ISD::FCVTZS_MERGE_PASSTHRU", SDT_AArch64FCVT>;
+
 def SDT_AArch64ReduceWithInit : SDTypeProfile<1, 3, [SDTCisVec<1>, SDTCisVec<3>]>;
 def AArch64clasta_n   : SDNode<"AArch64ISD::CLASTA_N",   SDT_AArch64ReduceWithInit>;
 def AArch64clastb_n   : SDNode<"AArch64ISD::CLASTB_N",   SDT_AArch64ReduceWithInit>;
 def AArch64fadda_p    : SDNode<"AArch64ISD::FADDA_PRED", SDT_AArch64ReduceWithInit>;
 
 def SDT_AArch64Rev   : SDTypeProfile<1, 1, [SDTCisVec<0>, SDTCisSameAs<0,1>]>;
 def AArch64rev       : SDNode<"AArch64ISD::REV", SDT_AArch64Rev>;
 
multiclass sve_prefetch<SDPatternOperator prefetch, ValueType PredTy, Instruction RegImmInst, Instruction RegRegInst, int scale, ComplexPattern AddrCP> {
   defm ASR_ZPZZ  : sve_int_bin_pred_bhsd<AArch64asr_p>;
   defm LSR_ZPZZ  : sve_int_bin_pred_bhsd<AArch64lsr_p>;
   defm LSL_ZPZZ  : sve_int_bin_pred_bhsd<AArch64lsl_p>;
 
   defm ASR_WIDE_ZPmZ : sve_int_bin_pred_shift_wide<0b000, "asr", int_aarch64_sve_asr_wide>;
   defm LSR_WIDE_ZPmZ : sve_int_bin_pred_shift_wide<0b001, "lsr", int_aarch64_sve_lsr_wide>;
   defm LSL_WIDE_ZPmZ : sve_int_bin_pred_shift_wide<0b011, "lsl", int_aarch64_sve_lsl_wide>;
 
-  defm FCVT_ZPmZ_StoH   : sve_fp_2op_p_zd<0b1001000, "fcvt",   ZPR32, ZPR16, int_aarch64_sve_fcvt_f16f32,    nxv8f16, nxv4i1, nxv4f32, ElementSizeS>;
-  defm FCVT_ZPmZ_HtoS   : sve_fp_2op_p_zd<0b1001001, "fcvt",   ZPR16, ZPR32, int_aarch64_sve_fcvt_f32f16,    nxv4f32, nxv4i1, nxv8f16, ElementSizeS>;
-  defm SCVTF_ZPmZ_HtoH  : sve_fp_2op_p_zd<0b0110010, "scvtf",  ZPR16, ZPR16, int_aarch64_sve_scvtf,          nxv8f16, nxv8i1, nxv8i16, ElementSizeH>;
-  defm SCVTF_ZPmZ_StoS  : sve_fp_2op_p_zd<0b1010100, "scvtf",  ZPR32, ZPR32, int_aarch64_sve_scvtf,          nxv4f32, nxv4i1, nxv4i32, ElementSizeS>;
-  defm UCVTF_ZPmZ_StoS  : sve_fp_2op_p_zd<0b1010101, "ucvtf",  ZPR32, ZPR32, int_aarch64_sve_ucvtf,          nxv4f32, nxv4i1, nxv4i32, ElementSizeS>;
-  defm UCVTF_ZPmZ_HtoH  : sve_fp_2op_p_zd<0b0110011, "ucvtf",  ZPR16, ZPR16, int_aarch64_sve_ucvtf,          nxv8f16, nxv8i1, nxv8i16, ElementSizeH>;
-  defm FCVTZS_ZPmZ_HtoH : sve_fp_2op_p_zd<0b0111010, "fcvtzs", ZPR16, ZPR16, int_aarch64_sve_fcvtzs,         nxv8i16, nxv8i1, nxv8f16, ElementSizeH>;
-  defm FCVTZS_ZPmZ_StoS : sve_fp_2op_p_zd<0b1011100, "fcvtzs", ZPR32, ZPR32, int_aarch64_sve_fcvtzs,         nxv4i32, nxv4i1, nxv4f32, ElementSizeS>;
-  defm FCVTZU_ZPmZ_HtoH : sve_fp_2op_p_zd<0b0111011, "fcvtzu", ZPR16, ZPR16, int_aarch64_sve_fcvtzu,         nxv8i16, nxv8i1, nxv8f16, ElementSizeH>;
-  defm FCVTZU_ZPmZ_StoS : sve_fp_2op_p_zd<0b1011101, "fcvtzu", ZPR32, ZPR32, int_aarch64_sve_fcvtzu,         nxv4i32, nxv4i1, nxv4f32, ElementSizeS>;
-  defm FCVT_ZPmZ_DtoH   : sve_fp_2op_p_zd<0b1101000, "fcvt",   ZPR64, ZPR16, int_aarch64_sve_fcvt_f16f64,    nxv8f16, nxv2i1, nxv2f64, ElementSizeD>;
-  defm FCVT_ZPmZ_HtoD   : sve_fp_2op_p_zd<0b1101001, "fcvt",   ZPR16, ZPR64, int_aarch64_sve_fcvt_f64f16,    nxv2f64, nxv2i1, nxv8f16, ElementSizeD>;
-  defm FCVT_ZPmZ_DtoS   : sve_fp_2op_p_zd<0b1101010, "fcvt",   ZPR64, ZPR32, int_aarch64_sve_fcvt_f32f64,    nxv4f32, nxv2i1, nxv2f64, ElementSizeD>;
-  defm FCVT_ZPmZ_StoD   : sve_fp_2op_p_zd<0b1101011, "fcvt",   ZPR32, ZPR64, int_aarch64_sve_fcvt_f64f32,    nxv2f64, nxv2i1, nxv4f32, ElementSizeD>;
-  defm SCVTF_ZPmZ_StoD  : sve_fp_2op_p_zd<0b1110000, "scvtf",  ZPR32, ZPR64, int_aarch64_sve_scvtf_f64i32,   nxv2f64, nxv2i1, nxv4i32, ElementSizeD>;
-  defm UCVTF_ZPmZ_StoD  : sve_fp_2op_p_zd<0b1110001, "ucvtf",  ZPR32, ZPR64, int_aarch64_sve_ucvtf_f64i32,   nxv2f64, nxv2i1, nxv4i32, ElementSizeD>;
-  defm UCVTF_ZPmZ_StoH  : sve_fp_2op_p_zd<0b0110101, "ucvtf",  ZPR32, ZPR16, int_aarch64_sve_ucvtf_f16i32,   nxv8f16, nxv4i1, nxv4i32, ElementSizeS>;
-  defm SCVTF_ZPmZ_DtoS  : sve_fp_2op_p_zd<0b1110100, "scvtf",  ZPR64, ZPR32, int_aarch64_sve_scvtf_f32i64,   nxv4f32, nxv2i1, nxv2i64, ElementSizeD>;
-  defm SCVTF_ZPmZ_StoH  : sve_fp_2op_p_zd<0b0110100, "scvtf",  ZPR32, ZPR16, int_aarch64_sve_scvtf_f16i32,   nxv8f16, nxv4i1, nxv4i32, ElementSizeS>;
-  defm SCVTF_ZPmZ_DtoH  : sve_fp_2op_p_zd<0b0110110, "scvtf",  ZPR64, ZPR16, int_aarch64_sve_scvtf_f16i64,   nxv8f16, nxv2i1, nxv2i64, ElementSizeD>;
-  defm UCVTF_ZPmZ_DtoS  : sve_fp_2op_p_zd<0b1110101, "ucvtf",  ZPR64, ZPR32, int_aarch64_sve_ucvtf_f32i64,   nxv4f32, nxv2i1, nxv2i64, ElementSizeD>;
-  defm UCVTF_ZPmZ_DtoH  : sve_fp_2op_p_zd<0b0110111, "ucvtf",  ZPR64, ZPR16, int_aarch64_sve_ucvtf_f16i64,   nxv8f16, nxv2i1, nxv2i64, ElementSizeD>;
-  defm SCVTF_ZPmZ_DtoD  : sve_fp_2op_p_zd<0b1110110, "scvtf",  ZPR64, ZPR64, int_aarch64_sve_scvtf,          nxv2f64, nxv2i1, nxv2i64, ElementSizeD>;
-  defm UCVTF_ZPmZ_DtoD  : sve_fp_2op_p_zd<0b1110111, "ucvtf",  ZPR64, ZPR64, int_aarch64_sve_ucvtf,          nxv2f64, nxv2i1, nxv2i64, ElementSizeD>;
-  defm FCVTZS_ZPmZ_DtoS : sve_fp_2op_p_zd<0b1111000, "fcvtzs", ZPR64, ZPR32, int_aarch64_sve_fcvtzs_i32f64,  nxv4i32, nxv2i1, nxv2f64, ElementSizeD>;
-  defm FCVTZU_ZPmZ_DtoS : sve_fp_2op_p_zd<0b1111001, "fcvtzu", ZPR64, ZPR32, int_aarch64_sve_fcvtzu_i32f64,  nxv4i32, nxv2i1, nxv2f64, ElementSizeD>;
-  defm FCVTZS_ZPmZ_StoD : sve_fp_2op_p_zd<0b1111100, "fcvtzs", ZPR32, ZPR64, int_aarch64_sve_fcvtzs_i64f32,  nxv2i64, nxv2i1, nxv4f32, ElementSizeD>;
-  defm FCVTZS_ZPmZ_HtoS : sve_fp_2op_p_zd<0b0111100, "fcvtzs", ZPR16, ZPR32, int_aarch64_sve_fcvtzs_i32f16,  nxv4i32, nxv4i1, nxv8f16, ElementSizeS>;
-  defm FCVTZS_ZPmZ_HtoD : sve_fp_2op_p_zd<0b0111110, "fcvtzs", ZPR16, ZPR64, int_aarch64_sve_fcvtzs_i64f16,  nxv2i64, nxv2i1, nxv8f16, ElementSizeD>;
-  defm FCVTZU_ZPmZ_HtoS : sve_fp_2op_p_zd<0b0111101, "fcvtzu", ZPR16, ZPR32, int_aarch64_sve_fcvtzu_i32f16,  nxv4i32, nxv4i1, nxv8f16, ElementSizeS>;
-  defm FCVTZU_ZPmZ_HtoD : sve_fp_2op_p_zd<0b0111111, "fcvtzu", ZPR16, ZPR64, int_aarch64_sve_fcvtzu_i64f16,  nxv2i64, nxv2i1, nxv8f16, ElementSizeD>;
-  defm FCVTZU_ZPmZ_StoD : sve_fp_2op_p_zd<0b1111101, "fcvtzu", ZPR32, ZPR64, int_aarch64_sve_fcvtzu_i64f32,  nxv2i64, nxv2i1, nxv4f32, ElementSizeD>;
-  defm FCVTZS_ZPmZ_DtoD : sve_fp_2op_p_zd<0b1111110, "fcvtzs", ZPR64, ZPR64, int_aarch64_sve_fcvtzs,         nxv2i64, nxv2i1, nxv2f64, ElementSizeD>;
-  defm FCVTZU_ZPmZ_DtoD : sve_fp_2op_p_zd<0b1111111, "fcvtzu", ZPR64, ZPR64, int_aarch64_sve_fcvtzu,         nxv2i64, nxv2i1, nxv2f64, ElementSizeD>;
+  defm FCVT_ZPmZ_StoH   : sve_fp_2op_p_zd<0b1001000, "fcvt",   ZPR32, ZPR16, int_aarch64_sve_fcvt_f16f32,   null_frag,        nxv8f16, nxv4i1, nxv4f32, ElementSizeS>;
+  defm FCVT_ZPmZ_HtoS   : sve_fp_2op_p_zd<0b1001001, "fcvt",   ZPR16, ZPR32, int_aarch64_sve_fcvt_f32f16,   null_frag,        nxv4f32, nxv4i1, nxv8f16, ElementSizeS>;
+  defm SCVTF_ZPmZ_HtoH  : sve_fp_2op_p_zd<0b0110010, "scvtf",  ZPR16, ZPR16, int_aarch64_sve_scvtf,         null_frag,        nxv8f16, nxv8i1, nxv8i16, ElementSizeH>;
+  defm SCVTF_ZPmZ_StoS  : sve_fp_2op_p_zd<0b1010100, "scvtf",  ZPR32, ZPR32, int_aarch64_sve_scvtf,         null_frag,        nxv4f32, nxv4i1, nxv4i32, ElementSizeS>;
+  defm UCVTF_ZPmZ_StoS  : sve_fp_2op_p_zd<0b1010101, "ucvtf",  ZPR32, ZPR32, int_aarch64_sve_ucvtf,         null_frag,        nxv4f32, nxv4i1, nxv4i32, ElementSizeS>;
+  defm UCVTF_ZPmZ_HtoH  : sve_fp_2op_p_zd<0b0110011, "ucvtf",  ZPR16, ZPR16, int_aarch64_sve_ucvtf,         null_frag,        nxv8f16, nxv8i1, nxv8i16, ElementSizeH>;
+  defm FCVTZS_ZPmZ_HtoH : sve_fp_2op_p_zd<0b0111010, "fcvtzs", ZPR16, ZPR16, null_frag,                     AArch64fcvtzs_mt, nxv8i16, nxv8i1, nxv8f16, ElementSizeH>;
+  defm FCVTZS_ZPmZ_StoS : sve_fp_2op_p_zd<0b1011100, "fcvtzs", ZPR32, ZPR32, null_frag,                     AArch64fcvtzs_mt, nxv4i32, nxv4i1, nxv4f32, ElementSizeS>;
+  defm FCVTZU_ZPmZ_HtoH : sve_fp_2op_p_zd<0b0111011, "fcvtzu", ZPR16, ZPR16, null_frag,                     AArch64fcvtzu_mt, nxv8i16, nxv8i1, nxv8f16, ElementSizeH>;
+  defm FCVTZU_ZPmZ_StoS : sve_fp_2op_p_zd<0b1011101, "fcvtzu", ZPR32, ZPR32, null_frag,                     AArch64fcvtzu_mt, nxv4i32, nxv4i1, nxv4f32, ElementSizeS>;
+  defm FCVT_ZPmZ_DtoH   : sve_fp_2op_p_zd<0b1101000, "fcvt",   ZPR64, ZPR16, int_aarch64_sve_fcvt_f16f64,   null_frag,        nxv8f16, nxv2i1, nxv2f64, ElementSizeD>;
+  defm FCVT_ZPmZ_HtoD   : sve_fp_2op_p_zd<0b1101001, "fcvt",   ZPR16, ZPR64, int_aarch64_sve_fcvt_f64f16,   null_frag,        nxv2f64, nxv2i1, nxv8f16, ElementSizeD>;
+  defm FCVT_ZPmZ_DtoS   : sve_fp_2op_p_zd<0b1101010, "fcvt",   ZPR64, ZPR32, int_aarch64_sve_fcvt_f32f64,   null_frag,        nxv4f32, nxv2i1, nxv2f64, ElementSizeD>;
+  defm FCVT_ZPmZ_StoD   : sve_fp_2op_p_zd<0b1101011, "fcvt",   ZPR32, ZPR64, int_aarch64_sve_fcvt_f64f32,   null_frag,        nxv2f64, nxv2i1, nxv4f32, ElementSizeD>;
+  defm SCVTF_ZPmZ_StoD  : sve_fp_2op_p_zd<0b1110000, "scvtf",  ZPR32, ZPR64, int_aarch64_sve_scvtf_f64i32,  null_frag,        nxv2f64, nxv2i1, nxv4i32, ElementSizeD>;
+  defm UCVTF_ZPmZ_StoD  : sve_fp_2op_p_zd<0b1110001, "ucvtf",  ZPR32, ZPR64, int_aarch64_sve_ucvtf_f64i32,  null_frag,        nxv2f64, nxv2i1, nxv4i32, ElementSizeD>;
+  defm UCVTF_ZPmZ_StoH  : sve_fp_2op_p_zd<0b0110101, "ucvtf",  ZPR32, ZPR16, int_aarch64_sve_ucvtf_f16i32,  null_frag,        nxv8f16, nxv4i1, nxv4i32, ElementSizeS>;
+  defm SCVTF_ZPmZ_DtoS  : sve_fp_2op_p_zd<0b1110100, "scvtf",  ZPR64, ZPR32, int_aarch64_sve_scvtf_f32i64,  null_frag,        nxv4f32, nxv2i1, nxv2i64, ElementSizeD>;
+  defm SCVTF_ZPmZ_StoH  : sve_fp_2op_p_zd<0b0110100, "scvtf",  ZPR32, ZPR16, int_aarch64_sve_scvtf_f16i32,  null_frag,        nxv8f16, nxv4i1, nxv4i32, ElementSizeS>;
+  defm SCVTF_ZPmZ_DtoH  : sve_fp_2op_p_zd<0b0110110, "scvtf",  ZPR64, ZPR16, int_aarch64_sve_scvtf_f16i64,  null_frag,        nxv8f16, nxv2i1, nxv2i64, ElementSizeD>;
+  defm UCVTF_ZPmZ_DtoS  : sve_fp_2op_p_zd<0b1110101, "ucvtf",  ZPR64, ZPR32, int_aarch64_sve_ucvtf_f32i64,  null_frag,        nxv4f32, nxv2i1, nxv2i64, ElementSizeD>;
+  defm UCVTF_ZPmZ_DtoH  : sve_fp_2op_p_zd<0b0110111, "ucvtf",  ZPR64, ZPR16, int_aarch64_sve_ucvtf_f16i64,  null_frag,        nxv8f16, nxv2i1, nxv2i64, ElementSizeD>;
+  defm SCVTF_ZPmZ_DtoD  : sve_fp_2op_p_zd<0b1110110, "scvtf",  ZPR64, ZPR64, int_aarch64_sve_scvtf,         null_frag,        nxv2f64, nxv2i1, nxv2i64, ElementSizeD>;
+  defm UCVTF_ZPmZ_DtoD  : sve_fp_2op_p_zd<0b1110111, "ucvtf",  ZPR64, ZPR64, int_aarch64_sve_ucvtf,         null_frag,        nxv2f64, nxv2i1, nxv2i64, ElementSizeD>;
+  defm FCVTZS_ZPmZ_DtoS : sve_fp_2op_p_zd<0b1111000, "fcvtzs", ZPR64, ZPR32, int_aarch64_sve_fcvtzs_i32f64, null_frag,        nxv4i32, nxv2i1, nxv2f64, ElementSizeD>;
+  defm FCVTZU_ZPmZ_DtoS : sve_fp_2op_p_zd<0b1111001, "fcvtzu", ZPR64, ZPR32, int_aarch64_sve_fcvtzu_i32f64, null_frag,        nxv4i32, nxv2i1, nxv2f64, ElementSizeD>;
+  defm FCVTZS_ZPmZ_StoD : sve_fp_2op_p_zd<0b1111100, "fcvtzs", ZPR32, ZPR64, int_aarch64_sve_fcvtzs_i64f32, AArch64fcvtzs_mt, nxv2i64, nxv2i1, nxv2f32, ElementSizeD>;
+  defm FCVTZS_ZPmZ_HtoS : sve_fp_2op_p_zd<0b0111100, "fcvtzs", ZPR16, ZPR32, int_aarch64_sve_fcvtzs_i32f16, AArch64fcvtzs_mt, nxv4i32, nxv4i1, nxv4f16, ElementSizeS>;
+  defm FCVTZS_ZPmZ_HtoD : sve_fp_2op_p_zd<0b0111110, "fcvtzs", ZPR16, ZPR64, int_aarch64_sve_fcvtzs_i64f16, AArch64fcvtzs_mt, nxv2i64, nxv2i1, nxv2f16, ElementSizeD>;
+  defm FCVTZU_ZPmZ_HtoS : sve_fp_2op_p_zd<0b0111101, "fcvtzu", ZPR16, ZPR32, int_aarch64_sve_fcvtzu_i32f16, AArch64fcvtzu_mt, nxv4i32, nxv4i1, nxv4f16, ElementSizeS>;
+  defm FCVTZU_ZPmZ_HtoD : sve_fp_2op_p_zd<0b0111111, "fcvtzu", ZPR16, ZPR64, int_aarch64_sve_fcvtzu_i64f16, AArch64fcvtzu_mt, nxv2i64, nxv2i1, nxv2f16, ElementSizeD>;
+  defm FCVTZU_ZPmZ_StoD : sve_fp_2op_p_zd<0b1111101, "fcvtzu", ZPR32, ZPR64, int_aarch64_sve_fcvtzu_i64f32, AArch64fcvtzu_mt, nxv2i64, nxv2i1, nxv2f32, ElementSizeD>;
+  defm FCVTZS_ZPmZ_DtoD : sve_fp_2op_p_zd<0b1111110, "fcvtzs", ZPR64, ZPR64, null_frag,                     AArch64fcvtzs_mt, nxv2i64, nxv2i1, nxv2f64, ElementSizeD>;
+  defm FCVTZU_ZPmZ_DtoD : sve_fp_2op_p_zd<0b1111111, "fcvtzu", ZPR64, ZPR64, null_frag,                     AArch64fcvtzu_mt, nxv2i64, nxv2i1, nxv2f64, ElementSizeD>;
 
   defm FRINTN_ZPmZ : sve_fp_2op_p_zd_HSD<0b00000, "frintn", null_frag, AArch64frintn_mt>;
   defm FRINTP_ZPmZ : sve_fp_2op_p_zd_HSD<0b00001, "frintp", null_frag, AArch64frintp_mt>;
   defm FRINTM_ZPmZ : sve_fp_2op_p_zd_HSD<0b00010, "frintm", null_frag, AArch64frintm_mt>;
   defm FRINTZ_ZPmZ : sve_fp_2op_p_zd_HSD<0b00011, "frintz", null_frag, AArch64frintz_mt>;
   defm FRINTA_ZPmZ : sve_fp_2op_p_zd_HSD<0b00100, "frinta", null_frag, AArch64frinta_mt>;
   defm FRINTX_ZPmZ : sve_fp_2op_p_zd_HSD<0b00110, "frintx", null_frag, AArch64frintx_mt>;
   defm FRINTI_ZPmZ : sve_fp_2op_p_zd_HSD<0b00111, "frinti", null_frag, AArch64frinti_mt>;

llvm/lib/Target/AArch64/SVEInstrFormats.td

: I<(outs o_zprtype:$Zd), (ins i_zprtype:$_Zd, PPR3bAny:$Pg, i_zprtype:$Zn),
   let Constraints = "$Zd = $_Zd";
   let DestructiveInstType = DestructiveOther;
   let ElementSize = size;
 }
 
 multiclass sve_fp_2op_p_zd<bits<7> opc, string asm,
                            RegisterOperand i_zprtype,
                            RegisterOperand o_zprtype,
-                           SDPatternOperator op, ValueType vt1,
+                           SDPatternOperator int_op,
+                           SDPatternOperator ir_op, ValueType vt1,
                            ValueType vt2, ValueType vt3, ElementSizeEnum Sz> {
   def NAME : sve_fp_2op_p_zd<opc, asm, i_zprtype, o_zprtype, Sz>;
 
-  def : SVE_3_Op_Pat<vt1, op, vt1, vt2, vt3, !cast<Instruction>(NAME)>;
+  // convert vt3 to a packed type for the intrinsic patterns
+  defvar packedvt3 = !cond(!eq(!cast<string>(vt3), "nxv2f16"): nxv8f16,
+                           !eq(!cast<string>(vt3), "nxv4f16"): nxv8f16,
+                           !eq(!cast<string>(vt3), "nxv2f32"): nxv4f32,
+                           1 : vt3);
+
+  def : SVE_3_Op_Pat<vt1, int_op, vt1, vt2, packedvt3, !cast<Instruction>(NAME)>;
+
+  def : SVE_1_Op_Passthru_Pat<vt1, ir_op, vt2, vt3, !cast<Instruction>(NAME)>;
 }
 
 multiclass sve_fp_2op_p_zd_HSD<bits<5> opc, string asm, SDPatternOperator op_merge,
                                                 SDPatternOperator op_pt = null_frag> {
   def _H : sve_fp_2op_p_zd<{ 0b01, opc }, asm, ZPR16, ZPR16, ElementSizeH>;
   def _S : sve_fp_2op_p_zd<{ 0b10, opc }, asm, ZPR32, ZPR32, ElementSizeS>;
   def _D : sve_fp_2op_p_zd<{ 0b11, opc }, asm, ZPR64, ZPR64, ElementSizeD>;
 

llvm/test/CodeGen/AArch64/sve-fcvt.ll

+; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py
+; RUN: llc -mtriple=aarch64--linux-gnu -mattr=+sve < %s | FileCheck %s
+
+;
+; FP_TO_SINT
+;
+
+define <vscale x 2 x i16> @fcvtzs_h_nxv2f16(<vscale x 2 x half> %a) {
+; CHECK-LABEL: fcvtzs_h_nxv2f16:
+; CHECK:       // %bb.0:
+; CHECK-NEXT:    ptrue p0.d
+; CHECK-NEXT:    fcvtzs z0.d, p0/m, z0.h
+; CHECK-NEXT:    ret
+  %res = fptosi <vscale x 2 x half> %a to <vscale x 2 x i16>
+  ret <vscale x 2 x i16> %res
+}
+
+define <vscale x 4 x i16> @fcvtzs_h_nxv4f16(<vscale x 4 x half> %a) {
+; CHECK-LABEL: fcvtzs_h_nxv4f16:
+; CHECK:       // %bb.0:
+; CHECK-NEXT:    ptrue p0.s
+; CHECK-NEXT:    fcvtzs z0.s, p0/m, z0.h
+; CHECK-NEXT:    ret
+  %res = fptosi <vscale x 4 x half> %a to <vscale x 4 x i16>
+  ret <vscale x 4 x i16> %res
+}
+
+define <vscale x 8 x i16> @fcvtzs_h_nxv8f16(<vscale x 8 x half> %a) {
+; CHECK-LABEL: fcvtzs_h_nxv8f16:
+; CHECK:       // %bb.0:
+; CHECK-NEXT:    ptrue p0.h
+; CHECK-NEXT:    fcvtzs z0.h, p0/m, z0.h
+; CHECK-NEXT:    ret
+  %res = fptosi <vscale x 8 x half> %a to <vscale x 8 x i16>
+  ret <vscale x 8 x i16> %res
+}
+
+define <vscale x 2 x i32> @fcvtzs_s_nxv2f16(<vscale x 2 x half> %a) {
+; CHECK-LABEL: fcvtzs_s_nxv2f16:
+; CHECK:       // %bb.0:
+; CHECK-NEXT:    ptrue p0.d
+; CHECK-NEXT:    fcvtzs z0.d, p0/m, z0.h
+; CHECK-NEXT:    ret
+  %res = fptosi <vscale x 2 x half> %a to <vscale x 2 x i32>
+  ret <vscale x 2 x i32> %res
+}
+
+define <vscale x 2 x i32> @fcvtzs_s_nxv2f32(<vscale x 2 x float> %a) {
+; CHECK-LABEL: fcvtzs_s_nxv2f32:
+; CHECK:       // %bb.0:
+; CHECK-NEXT:    ptrue p0.d
+; CHECK-NEXT:    fcvtzs z0.d, p0/m, z0.s
+; CHECK-NEXT:    ret
+  %res = fptosi <vscale x 2 x float> %a to <vscale x 2 x i32>
+  ret <vscale x 2 x i32> %res
+}
+
+define <vscale x 2 x i32> @fcvtzs_s_nxv2f64(<vscale x 2 x double> %a) {
+; CHECK-LABEL: fcvtzs_s_nxv2f64:
+; CHECK:       // %bb.0:
+; CHECK-NEXT:    ptrue p0.d
+; CHECK-NEXT:    fcvtzs z0.d, p0/m, z0.d
+; CHECK-NEXT:    ret
+  %res = fptosi <vscale x 2 x double> %a to <vscale x 2 x i32>
+  ret <vscale x 2 x i32> %res
+}
+
+define <vscale x 4 x i32> @fcvtzs_s_nxv4f16(<vscale x 4 x half> %a) {
+; CHECK-LABEL: fcvtzs_s_nxv4f16:
+; CHECK:       // %bb.0:
+; CHECK-NEXT:    ptrue p0.s
+; CHECK-NEXT:    fcvtzs z0.s, p0/m, z0.h
+; CHECK-NEXT:    ret
+  %res = fptosi <vscale x 4 x half> %a to <vscale x 4 x i32>
+  ret <vscale x 4 x i32> %res
+}
+
+define <vscale x 4 x i32> @fcvtzs_s_nxv4f32(<vscale x 4 x float> %a) {
+; CHECK-LABEL: fcvtzs_s_nxv4f32:
+; CHECK:       // %bb.0:
+; CHECK-NEXT:    ptrue p0.s
+; CHECK-NEXT:    fcvtzs z0.s, p0/m, z0.s
+; CHECK-NEXT:    ret
+  %res = fptosi <vscale x 4 x float> %a to <vscale x 4 x i32>
+  ret <vscale x 4 x i32> %res
+}
+
+define <vscale x 2 x i64> @fcvtzs_d_nxv2f16(<vscale x 2 x half> %a) {
+; CHECK-LABEL: fcvtzs_d_nxv2f16:
+; CHECK:       // %bb.0:
+; CHECK-NEXT:    ptrue p0.d
+; CHECK-NEXT:    fcvtzs z0.d, p0/m, z0.h
+; CHECK-NEXT:    ret
+  %res = fptosi <vscale x 2 x half> %a to <vscale x 2 x i64>
+  ret <vscale x 2 x i64> %res
+}
+
+define <vscale x 2 x i64> @fcvtzs_d_nxv2f32(<vscale x 2 x float> %a) {
+; CHECK-LABEL: fcvtzs_d_nxv2f32:
+; CHECK:       // %bb.0:
+; CHECK-NEXT:    ptrue p0.d
+; CHECK-NEXT:    fcvtzs z0.d, p0/m, z0.s
+; CHECK-NEXT:    ret
+  %res = fptosi <vscale x 2 x float> %a to <vscale x 2 x i64>
+  ret <vscale x 2 x i64> %res
+}
+
+define <vscale x 2 x i64> @fcvtzs_d_nxv2f64(<vscale x 2 x double> %a) {
+; CHECK-LABEL: fcvtzs_d_nxv2f64:
+; CHECK:       // %bb.0:
+; CHECK-NEXT:    ptrue p0.d
+; CHECK-NEXT:    fcvtzs z0.d, p0/m, z0.d
+; CHECK-NEXT:    ret
+  %res = fptosi <vscale x 2 x double> %a to <vscale x 2 x i64>
+  ret <vscale x 2 x i64> %res
+}
+
+;
+; FP_TO_UINT
+;
+
+define <vscale x 2 x i16> @fcvtzu_h_nxv2f16(<vscale x 2 x half> %a) {
+; CHECK-LABEL: fcvtzu_h_nxv2f16:
+; CHECK:       // %bb.0:
+; CHECK-NEXT:    ptrue p0.d
+; CHECK-NEXT:    fcvtzu z0.d, p0/m, z0.h
+; CHECK-NEXT:    ret
+  %res = fptoui <vscale x 2 x half> %a to <vscale x 2 x i16>
+  ret <vscale x 2 x i16> %res
+}
+
+define <vscale x 4 x i16> @fcvtzu_h_nxv4f16(<vscale x 4 x half> %a) {
+; CHECK-LABEL: fcvtzu_h_nxv4f16:
+; CHECK:       // %bb.0:
+; CHECK-NEXT:    ptrue p0.s
+; CHECK-NEXT:    fcvtzu z0.s, p0/m, z0.h
+; CHECK-NEXT:    ret
+  %res = fptoui <vscale x 4 x half> %a to <vscale x 4 x i16>
+  ret <vscale x 4 x i16> %res
+}
+
+define <vscale x 8 x i16> @fcvtzu_h_nxv8f16(<vscale x 8 x half> %a) {
+; CHECK-LABEL: fcvtzu_h_nxv8f16:
+; CHECK:       // %bb.0:
+; CHECK-NEXT:    ptrue p0.h
+; CHECK-NEXT:    fcvtzu z0.h, p0/m, z0.h
+; CHECK-NEXT:    ret
+  %res = fptoui <vscale x 8 x half> %a to <vscale x 8 x i16>
+  ret <vscale x 8 x i16> %res
+}
+
+define <vscale x 2 x i32> @fcvtzu_s_nxv2f16(<vscale x 2 x half> %a) {
+; CHECK-LABEL: fcvtzu_s_nxv2f16:
+; CHECK:       // %bb.0:
+; CHECK-NEXT:    ptrue p0.d
+; CHECK-NEXT:    fcvtzu z0.d, p0/m, z0.h
+; CHECK-NEXT:    ret
+  %res = fptoui <vscale x 2 x half> %a to <vscale x 2 x i32>
+  ret <vscale x 2 x i32> %res
+}
+
+define <vscale x 2 x i32> @fcvtzu_s_nxv2f32(<vscale x 2 x float> %a) {
+; CHECK-LABEL: fcvtzu_s_nxv2f32:
+; CHECK:       // %bb.0:
+; CHECK-NEXT:    ptrue p0.d
+; CHECK-NEXT:    fcvtzu z0.d, p0/m, z0.s
+; CHECK-NEXT:    ret
+  %res = fptoui <vscale x 2 x float> %a to <vscale x 2 x i32>
+  ret <vscale x 2 x i32> %res
+}
+
+define <vscale x 2 x i32> @fcvtzu_s_nxv2f64(<vscale x 2 x double> %a) {
+; CHECK-LABEL: fcvtzu_s_nxv2f64:
+; CHECK:       // %bb.0:
+; CHECK-NEXT:    ptrue p0.d
+; CHECK-NEXT:    fcvtzu z0.d, p0/m, z0.d
+; CHECK-NEXT:    ret
+  %res = fptoui <vscale x 2 x double> %a to <vscale x 2 x i32>
+  ret <vscale x 2 x i32> %res
+}
+
+define <vscale x 4 x i32> @fcvtzu_s_nxv4f16(<vscale x 4 x half> %a) {
+; CHECK-LABEL: fcvtzu_s_nxv4f16:
+; CHECK:       // %bb.0:
+; CHECK-NEXT:    ptrue p0.s
+; CHECK-NEXT:    fcvtzu z0.s, p0/m, z0.h
+; CHECK-NEXT:    ret
+  %res = fptoui <vscale x 4 x half> %a to <vscale x 4 x i32>
+  ret <vscale x 4 x i32> %res
+}
+
+define <vscale x 4 x i32> @fcvtzu_s_nxv4f32(<vscale x 4 x float> %a) {
+; CHECK-LABEL: fcvtzu_s_nxv4f32:
+; CHECK:       // %bb.0:
+; CHECK-NEXT:    ptrue p0.s
+; CHECK-NEXT:    fcvtzu z0.s, p0/m, z0.s
+; CHECK-NEXT:    ret
+  %res = fptoui <vscale x 4 x float> %a to <vscale x 4 x i32>
+  ret <vscale x 4 x i32> %res
+}
+
+define <vscale x 2 x i64> @fcvtzu_d_nxv2f16(<vscale x 2 x half> %a) {
+; CHECK-LABEL: fcvtzu_d_nxv2f16:
+; CHECK:       // %bb.0:
+; CHECK-NEXT:    ptrue p0.d
+; CHECK-NEXT:    fcvtzu z0.d, p0/m, z0.h
+; CHECK-NEXT:    ret
+  %res = fptoui <vscale x 2 x half> %a to <vscale x 2 x i64>
+  ret <vscale x 2 x i64> %res
+}
+
+define <vscale x 2 x i64> @fcvtzu_d_nxv2f32(<vscale x 2 x float> %a) {
+; CHECK-LABEL: fcvtzu_d_nxv2f32:
+; CHECK:       // %bb.0:
+; CHECK-NEXT:    ptrue p0.d
+; CHECK-NEXT:    fcvtzu z0.d, p0/m, z0.s
+; CHECK-NEXT:    ret
+  %res = fptoui <vscale x 2 x float> %a to <vscale x 2 x i64>
+  ret <vscale x 2 x i64> %res
+}
+
+define <vscale x 2 x i64> @fcvtzu_d_nxv2f64(<vscale x 2 x double> %a) {
+; CHECK-LABEL: fcvtzu_d_nxv2f64:
+; CHECK:       // %bb.0:
+; CHECK-NEXT:    ptrue p0.d
+; CHECK-NEXT:    fcvtzu z0.d, p0/m, z0.d
+; CHECK-NEXT:    ret
+  %res = fptoui <vscale x 2 x double> %a to <vscale x 2 x i64>
+  ret <vscale x 2 x i64> %res
+}

llvm/test/CodeGen/AArch64/sve-split-fcvt.ll

+; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py
+; RUN: llc -mtriple=aarch64--linux-gnu -mattr=sve < %s | FileCheck %s
+
+; FP_TO_SINT
+
+; Split operand
+define <vscale x 4 x i32> @fcvtzs_s_nxv4f64(<vscale x 4 x double> %a) {
+; CHECK-LABEL: fcvtzs_s_nxv4f64:
+; CHECK:       // %bb.0:
+; CHECK-NEXT:    ptrue p0.d
+; CHECK-NEXT:    fcvtzs z1.d, p0/m, z1.d
+; CHECK-NEXT:    fcvtzs z0.d, p0/m, z0.d
+; CHECK-NEXT:    uzp1 z0.s, z0.s, z1.s
+; CHECK-NEXT:    ret
+  %res = fptosi <vscale x 4 x double> %a to <vscale x 4 x i32>
+  ret <vscale x 4 x i32> %res
+}
+
+define <vscale x 8 x i16> @fcvtzs_h_nxv8f64(<vscale x 8 x double> %a) {
+; CHECK-LABEL: fcvtzs_h_nxv8f64:
+; CHECK:       // %bb.0:
+; CHECK-NEXT:    ptrue p0.d
+; CHECK-NEXT:    fcvtzs z3.d, p0/m, z3.d
+; CHECK-NEXT:    fcvtzs z2.d, p0/m, z2.d
+; CHECK-NEXT:    fcvtzs z1.d, p0/m, z1.d
+; CHECK-NEXT:    fcvtzs z0.d, p0/m, z0.d
+; CHECK-NEXT:    uzp1 z2.s, z2.s, z3.s
+; CHECK-NEXT:    uzp1 z0.s, z0.s, z1.s
+; CHECK-NEXT:    uzp1 z0.h, z0.h, z2.h
+; CHECK-NEXT:    ret
+  %res = fptosi <vscale x 8 x double> %a to <vscale x 8 x i16>
+  ret <vscale x 8 x i16> %res
+}
+
+; Split result
+define <vscale x 4 x i64> @fcvtzs_d_nxv4f32(<vscale x 4 x float> %a) {
+; CHECK-LABEL: fcvtzs_d_nxv4f32:
+; CHECK:       // %bb.0:
+; CHECK-NEXT:    uunpklo z1.d, z0.s
+; CHECK-NEXT:    ptrue p0.d
+; CHECK-NEXT:    uunpkhi z2.d, z0.s
+; CHECK-NEXT:    fcvtzs z0.d, p0/m, z1.s
+; CHECK-NEXT:    fcvtzs z1.d, p0/m, z2.s
+; CHECK-NEXT:    ret
+  %res = fptosi <vscale x 4 x float> %a to <vscale x 4 x i64>
+  ret <vscale x 4 x i64> %res
+}
+
+define <vscale x 16 x i32> @fcvtzs_s_nxv16f16(<vscale x 16 x half> %a) {
+; CHECK-LABEL: fcvtzs_s_nxv16f16:
+; CHECK:       // %bb.0:
+; CHECK-NEXT:    uunpklo z2.s, z0.h
+; CHECK-NEXT:    ptrue p0.s
+; CHECK-NEXT:    uunpkhi z3.s, z0.h
+; CHECK-NEXT:    uunpklo z4.s, z1.h
+; CHECK-NEXT:    uunpkhi z5.s, z1.h
+; CHECK-NEXT:    fcvtzs z0.s, p0/m, z2.h
+; CHECK-NEXT:    fcvtzs z1.s, p0/m, z3.h
+; CHECK-NEXT:    fcvtzs z2.s, p0/m, z4.h
+; CHECK-NEXT:    fcvtzs z3.s, p0/m, z5.h
+; CHECK-NEXT:    ret
+  %res = fptosi <vscale x 16 x half> %a to <vscale x 16 x i32>
+  ret <vscale x 16 x i32> %res
+}
+
+; FP_TO_UINT
+
+; Split operand
+define <vscale x 4 x i32> @fcvtzu_s_nxv4f64(<vscale x 4 x double> %a) {
+; CHECK-LABEL: fcvtzu_s_nxv4f64:
+; CHECK:       // %bb.0:
+; CHECK-NEXT:    ptrue p0.d
+; CHECK-NEXT:    fcvtzu z1.d, p0/m, z1.d
+; CHECK-NEXT:    fcvtzu z0.d, p0/m, z0.d
+; CHECK-NEXT:    uzp1 z0.s, z0.s, z1.s
+; CHECK-NEXT:    ret
+  %res = fptoui <vscale x 4 x double> %a to <vscale x 4 x i32>
+  ret <vscale x 4 x i32> %res
+}
+
+; Split result
+define <vscale x 4 x i64> @fcvtzu_d_nxv4f32(<vscale x 4 x float> %a) {
+; CHECK-LABEL: fcvtzu_d_nxv4f32:
+; CHECK:       // %bb.0:
+; CHECK-NEXT:    uunpklo z1.d, z0.s
+; CHECK-NEXT:    ptrue p0.d
+; CHECK-NEXT:    uunpkhi z2.d, z0.s
+; CHECK-NEXT:    fcvtzu z0.d, p0/m, z1.s
+; CHECK-NEXT:    fcvtzu z1.d, p0/m, z2.s
+; CHECK-NEXT:    ret
+  %res = fptoui <vscale x 4 x float> %a to <vscale x 4 x i64>
+  ret <vscale x 4 x i64> %res
+}