[SVE][CodeGen] Lower scalable floating-point vector reductions

[kmclaughlin]

Summary

Changes in this patch:

• Minor changes to the LowerVECREDUCE_SEQ_FADD function added by @cameron.mcinally to also work for scalable types
• Added TableGen patterns for FP reductions with unpacked types (nxv2f16, nxv4f16 & nxv2f32)
• Asserts added to expandFMINNUM_FMAXNUM & expandVecReduceSeq for scalable types

[cameron.mcinally]

LGTM with one nit below...

llvm/lib/Target/AArch64/AArch64ISelLowering.cpp

16842

Nit: We could conditionally avoid the call to getPackedSVEVectorVT(...) by inverting this condition and initial value of RdxVT. Something like:

EVT RdxVT = SrcVT; 
if (SrcVT.isFixedVector() || SrcVT.isInteger())
  RdxVT = getPackedSVEVectorVT(ResVT);

I also wonder if we need the isFloatingPoint()/isInteger() check at all. Is just isScalableVector() enough?

[kmclaughlin]

• Reordered the condition added to LowerReductionToSVE which sets RdxVT

[kmclaughlin]

llvm/lib/Target/AArch64/AArch64ISelLowering.cpp

16842

Hi @cameron.mcinally , I've inverted the condition as you suggested here and removed the isFloatingPoint() check. The reason I included the additional check was for UADDV, where the result type should always be nxv2i64. Though as it's the only case for scalable vectors that needs to use getPackedSVEVectorVT, I've added a check for the UADDV_PRED opcode here instead.

[paulwalker-arm]

llvm/lib/Target/AArch64/AArch64ISelLowering.cpp

16773

By getting this far we've already committed to using SVE so can this just be SrcVT.isFixedLengthVector()?

16842

I've nothing against the current approach but there is an argument for unpacked scalable vectors to be treated more like how fixed length vectors are handled. For example we could change the current if (useSVEForFixedLengthVectorVT block to:

if (SrcVT.isFixedLengthVector()) {
    EVT ContainerVT = getContainerForFixedLengthVector(DAG, SrcVT);
    VecOp = convertToScalableVector(DAG, ContainerVT, VecOp);
} else if (!isPackedVectorType(SrcVT)) {
    EVT ContainerVT = getPackedSVEVectorVT(DAG, SrcVT);
    VecOp = DAG.getNode(AArch64ISD::REINTERPRET_CAST, dl, ContainerVT, VecOp);
}

That way the remainder of the function remains as is and you don't need to change SVEInstrFormats.td because these nodes no longer have to worry about unpacked types. The same argument applies to the ordered reductions as well.

As I say, I'm happy either way, so just throwing it out there. What do you think?

[paulwalker-arm]

llvm/lib/Target/AArch64/AArch64ISelLowering.cpp

16842

Actually I retract this. getPredicateForVector doesn't quite work the way we'll need it to. It's something to consider for the future, perhaps as part of a larger effort to remove the need to support unpacked vector types during isel.

[cameron.mcinally]

LGTM

[kmclaughlin]

Thanks @cameron.mcinally & @paulwalker-arm for reviewing this patch!

llvm/lib/CodeGen/SelectionDAG/TargetLowering.cpp

 }
 
 SDValue TargetLowering::expandFMINNUM_FMAXNUM(SDNode *Node,
                                               SelectionDAG &DAG) const {
   SDLoc dl(Node);
   unsigned NewOp = Node->getOpcode() == ISD::FMINNUM ?
     ISD::FMINNUM_IEEE : ISD::FMAXNUM_IEEE;
   EVT VT = Node->getValueType(0);
+
+  if (VT.isScalableVector())
+    report_fatal_error(
+        "Expanding fminnum/fmaxnum for scalable vectors is undefined.");
+
   if (isOperationLegalOrCustom(NewOp, VT)) {
     SDValue Quiet0 = Node->getOperand(0);
     SDValue Quiet1 = Node->getOperand(1);
 
     if (!Node->getFlags().hasNoNaNs()) {
       // Insert canonicalizes if it's possible we need to quiet to get correct
       // sNaN behavior.
       if (!DAG.isKnownNeverSNaN(Quiet0)) {

 SDValue TargetLowering::expandVecReduceSeq(SDNode *Node, SelectionDAG &DAG) const {
   SDLoc dl(Node);
   SDValue AccOp = Node->getOperand(0);
   SDValue VecOp = Node->getOperand(1);
   SDNodeFlags Flags = Node->getFlags();
 
   EVT VT = VecOp.getValueType();
   EVT EltVT = VT.getVectorElementType();
+
+  if (VT.isScalableVector())
+    report_fatal_error(
+        "Expanding reductions for scalable vectors is undefined.");
+
   unsigned NumElts = VT.getVectorNumElements();
 
   SmallVector<SDValue, 8> Ops;
   DAG.ExtractVectorElements(VecOp, Ops, 0, NumElts);
 
   unsigned BaseOpcode = ISD::getVecReduceBaseOpcode(Node->getOpcode());
 
   SDValue Res = AccOp;

llvm/lib/Target/AArch64/AArch64ISelLowering.cpp

for (auto VT : {MVT::nxv2f16, MVT::nxv4f16, MVT::nxv8f16, MVT::nxv2f32,
       setOperationAction(ISD::FRINT, VT, Custom);
       setOperationAction(ISD::FROUND, VT, Custom);
       setOperationAction(ISD::FROUNDEVEN, VT, Custom);
       setOperationAction(ISD::FTRUNC, VT, Custom);
       setOperationAction(ISD::FSQRT, VT, Custom);
       setOperationAction(ISD::FABS, VT, Custom);
       setOperationAction(ISD::FP_EXTEND, VT, Custom);
       setOperationAction(ISD::FP_ROUND, VT, Custom);
+      setOperationAction(ISD::VECREDUCE_FADD, VT, Custom);
+      setOperationAction(ISD::VECREDUCE_FMAX, VT, Custom);
+      setOperationAction(ISD::VECREDUCE_FMIN, VT, Custom);
+      setOperationAction(ISD::VECREDUCE_SEQ_FADD, VT, Custom);
     }
 
     for (auto VT : {MVT::nxv2bf16, MVT::nxv4bf16, MVT::nxv8bf16})
       setOperationAction(ISD::MSCATTER, VT, Custom);
 
     setOperationAction(ISD::SPLAT_VECTOR, MVT::nxv8bf16, Custom);
 
     setOperationAction(ISD::INTRINSIC_WO_CHAIN, MVT::i8, Custom);

 SDValue AArch64TargetLowering::LowerVECREDUCE_SEQ_FADD(SDValue ScalarOp,
     SelectionDAG &DAG) const {
   SDLoc DL(ScalarOp);
   SDValue AccOp = ScalarOp.getOperand(0);
   SDValue VecOp = ScalarOp.getOperand(1);
   EVT SrcVT = VecOp.getValueType();
   EVT ResVT = SrcVT.getVectorElementType();
 
-  // Only fixed length FADDA handled for now.
-  if (!useSVEForFixedLengthVectorVT(SrcVT, /*OverrideNEON=*/true))
-    return SDValue();
+  EVT ContainerVT = SrcVT;
+  if (SrcVT.isFixedLengthVector()) {
+    ContainerVT = getContainerForFixedLengthVector(DAG, SrcVT);
+    VecOp = convertToScalableVector(DAG, ContainerVT, VecOp);
+  }
 
   SDValue Pg = getPredicateForVector(DAG, DL, SrcVT);
-  EVT ContainerVT = getContainerForFixedLengthVector(DAG, SrcVT);
   SDValue Zero = DAG.getConstant(0, DL, MVT::i64);
 
   // Convert operands to Scalable.
   AccOp = DAG.getNode(ISD::INSERT_VECTOR_ELT, DL, ContainerVT,
                       DAG.getUNDEF(ContainerVT), AccOp, Zero);
-  VecOp = convertToScalableVector(DAG, ContainerVT, VecOp);
 
   // Perform reduction.
   SDValue Rdx = DAG.getNode(AArch64ISD::FADDA_PRED, DL, ContainerVT,
                             Pg, AccOp, VecOp);
 
   return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, ResVT, Rdx, Zero);
 }
 
SDValue AArch64TargetLowering::LowerReductionToSVE(unsigned Opcode,
   if (useSVEForFixedLengthVectorVT(SrcVT, true)) {
     EVT ContainerVT = getContainerForFixedLengthVector(DAG, SrcVT);
     VecOp = convertToScalableVector(DAG, ContainerVT, VecOp);
   }
 
   // UADDV always returns an i64 result.
   EVT ResVT = (Opcode == AArch64ISD::UADDV_PRED) ? MVT::i64 :
                                                    SrcVT.getVectorElementType();
+  EVT RdxVT = SrcVT;
+  if (SrcVT.isFixedLengthVector() || Opcode == AArch64ISD::UADDV_PRED)
+    RdxVT = getPackedSVEVectorVT(ResVT);
 
   SDValue Pg = getPredicateForVector(DAG, DL, SrcVT);
-  SDValue Rdx = DAG.getNode(Opcode, DL, getPackedSVEVectorVT(ResVT), Pg, VecOp);
+  SDValue Rdx = DAG.getNode(Opcode, DL, RdxVT, Pg, VecOp);
   SDValue Res = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, ResVT,
                             Rdx, DAG.getConstant(0, DL, MVT::i64));
 
   // The VEC_REDUCE nodes expect an element size result.
   if (ResVT != ScalarOp.getValueType())
     Res = DAG.getAnyExtOrTrunc(Res, DL, ScalarOp.getValueType());
 
   return Res;

llvm/lib/Target/AArch64/SVEInstrFormats.td

: I<(outs dstOpType:$Vd), (ins PPR3bAny:$Pg, zprty:$Zn),
   let Inst{4-0}   = Vd;
 }
 
 multiclass sve_fp_fast_red<bits<3> opc, string asm, SDPatternOperator op> {
   def _H : sve_fp_fast_red<0b01, opc, asm, ZPR16, FPR16asZPR>;
   def _S : sve_fp_fast_red<0b10, opc, asm, ZPR32, FPR32asZPR>;
   def _D : sve_fp_fast_red<0b11, opc, asm, ZPR64, FPR64asZPR>;
 
+  def : SVE_2_Op_Pat<nxv2f16, op, nxv2i1, nxv2f16, !cast<Instruction>(NAME # _H)>;
+  def : SVE_2_Op_Pat<nxv4f16, op, nxv4i1, nxv4f16, !cast<Instruction>(NAME # _H)>;
   def : SVE_2_Op_Pat<nxv8f16, op, nxv8i1, nxv8f16, !cast<Instruction>(NAME # _H)>;
+  def : SVE_2_Op_Pat<nxv2f32, op, nxv2i1, nxv2f32, !cast<Instruction>(NAME # _S)>;
   def : SVE_2_Op_Pat<nxv4f32, op, nxv4i1, nxv4f32, !cast<Instruction>(NAME # _S)>;
   def : SVE_2_Op_Pat<nxv2f64, op, nxv2i1, nxv2f64, !cast<Instruction>(NAME # _D)>;
 }
 
 //===----------------------------------------------------------------------===//
 // SVE Floating Point Accumulating Reduction Group
 //===----------------------------------------------------------------------===//
 
: I<(outs dstOpType:$Vdn), (ins PPR3bAny:$Pg, dstOpType:$_Vdn, zprty:$Zm),
   let Constraints = "$Vdn = $_Vdn";
 }
 
 multiclass sve_fp_2op_p_vd<bits<3> opc, string asm, SDPatternOperator op> {
   def _H : sve_fp_2op_p_vd<0b01, opc, asm, ZPR16, FPR16asZPR>;
   def _S : sve_fp_2op_p_vd<0b10, opc, asm, ZPR32, FPR32asZPR>;
   def _D : sve_fp_2op_p_vd<0b11, opc, asm, ZPR64, FPR64asZPR>;
 
+  def : SVE_3_Op_Pat<nxv2f16, op, nxv2i1, nxv2f16, nxv2f16, !cast<Instruction>(NAME # _H)>;
+  def : SVE_3_Op_Pat<nxv4f16, op, nxv4i1, nxv4f16, nxv4f16, !cast<Instruction>(NAME # _H)>;
   def : SVE_3_Op_Pat<nxv8f16, op, nxv8i1, nxv8f16, nxv8f16, !cast<Instruction>(NAME # _H)>;
+  def : SVE_3_Op_Pat<nxv2f32, op, nxv2i1, nxv2f32, nxv2f32, !cast<Instruction>(NAME # _S)>;
   def : SVE_3_Op_Pat<nxv4f32, op, nxv4i1, nxv4f32, nxv4f32, !cast<Instruction>(NAME # _S)>;
   def : SVE_3_Op_Pat<nxv2f64, op, nxv2i1, nxv2f64, nxv2f64, !cast<Instruction>(NAME # _D)>;
 }
 
 //===----------------------------------------------------------------------===//
 // SVE Floating Point Compare - Vectors Group
 //===----------------------------------------------------------------------===//
 

llvm/test/CodeGen/AArch64/sve-fp-reduce.ll

+; RUN: llc -mtriple=aarch64-linux-gnu -mattr=+sve -asm-verbose=0 < %s 2>%t | FileCheck %s
+; RUN: FileCheck --check-prefix=WARN --allow-empty %s <%t
+
+; If this check fails please read test/CodeGen/AArch64/README for instructions on how to resolve it.
+; WARN-NOT: warning
+
+; FADD
+
+define half @fadda_nxv2f16(half %init, <vscale x 2 x half> %a) {
+; CHECK-LABEL: fadda_nxv2f16:
+; CHECK:      ptrue p0.d
+; CHECK-NEXT: fadda h0, p0, h0, z1.h
+; CHECK-NEXT: ret
+  %res = call half @llvm.vector.reduce.fadd.nxv2f16(half %init, <vscale x 2 x half> %a)
+  ret half %res
+}
+
+define half @fadda_nxv4f16(half %init, <vscale x 4 x half> %a) {
+; CHECK-LABEL: fadda_nxv4f16:
+; CHECK:      ptrue p0.s
+; CHECK-NEXT: fadda h0, p0, h0, z1.h
+; CHECK-NEXT: ret
+  %res = call half @llvm.vector.reduce.fadd.nxv4f16(half %init, <vscale x 4 x half> %a)
+  ret half %res
+}
+
+define half @fadda_nxv8f16(half %init, <vscale x 8 x half> %a) {
+; CHECK-LABEL: fadda_nxv8f16:
+; CHECK:      ptrue p0.h
+; CHECK-NEXT: fadda h0, p0, h0, z1.h
+; CHECK-NEXT: ret
+  %res = call half @llvm.vector.reduce.fadd.nxv8f16(half %init, <vscale x 8 x half> %a)
+  ret half %res
+}
+
+define float @fadda_nxv2f32(float %init, <vscale x 2 x float> %a) {
+; CHECK-LABEL: fadda_nxv2f32:
+; CHECK:      ptrue p0.d
+; CHECK-NEXT: fadda s0, p0, s0, z1.s
+; CHECK-NEXT: ret
+  %res = call float @llvm.vector.reduce.fadd.nxv2f32(float %init, <vscale x 2 x float> %a)
+  ret float %res
+}
+
+define float @fadda_nxv4f32(float %init, <vscale x 4 x float> %a) {
+; CHECK-LABEL: fadda_nxv4f32:
+; CHECK:      ptrue p0.s
+; CHECK-NEXT: fadda s0, p0, s0, z1.s
+; CHECK-NEXT: ret
+  %res = call float @llvm.vector.reduce.fadd.nxv4f32(float %init, <vscale x 4 x float> %a)
+  ret float %res
+}
+
+define double @fadda_nxv2f64(double %init, <vscale x 2 x double> %a) {
+; CHECK-LABEL: fadda_nxv2f64:
+; CHECK:      ptrue p0.d
+; CHECK-NEXT: fadda d0, p0, d0, z1.d
+; CHECK-NEXT: ret
+  %res = call double @llvm.vector.reduce.fadd.nxv2f64(double %init, <vscale x 2 x double> %a)
+  ret double %res
+}
+
+; FADDV
+
+define half @faddv_nxv2f16(half %init, <vscale x 2 x half> %a) {
+; CHECK-LABEL: faddv_nxv2f16:
+; CHECK:      ptrue p0.d
+; CHECK-NEXT: faddv h1, p0, z1.h
+; CHECK-NEXT: fadd h0, h0, h1
+; CHECK-NEXT: ret
+  %res = call fast half @llvm.vector.reduce.fadd.nxv2f16(half %init, <vscale x 2 x half> %a)
+  ret half %res
+}
+
+define half @faddv_nxv4f16(half %init, <vscale x 4 x half> %a) {
+; CHECK-LABEL: faddv_nxv4f16:
+; CHECK:      ptrue p0.s
+; CHECK-NEXT: faddv h1, p0, z1.h
+; CHECK-NEXT: fadd h0, h0, h1
+; CHECK-NEXT: ret
+  %res = call fast half @llvm.vector.reduce.fadd.nxv4f16(half %init, <vscale x 4 x half> %a)
+  ret half %res
+}
+
+define half @faddv_nxv8f16(half %init, <vscale x 8 x half> %a) {
+; CHECK-LABEL: faddv_nxv8f16:
+; CHECK:      ptrue p0.h
+; CHECK-NEXT: faddv h1, p0, z1.h
+; CHECK-NEXT: fadd h0, h0, h1
+; CHECK-NEXT: ret
+  %res = call fast half @llvm.vector.reduce.fadd.nxv8f16(half %init, <vscale x 8 x half> %a)
+  ret half %res
+}
+
+define float @faddv_nxv2f32(float %init, <vscale x 2 x float> %a) {
+; CHECK-LABEL: faddv_nxv2f32:
+; CHECK:      ptrue p0.d
+; CHECK-NEXT: faddv s1, p0, z1.s
+; CHECK-NEXT: fadd s0, s0, s1
+; CHECK-NEXT: ret
+  %res = call fast float @llvm.vector.reduce.fadd.nxv2f32(float %init, <vscale x 2 x float> %a)
+  ret float %res
+}
+
+define float @faddv_nxv4f32(float %init, <vscale x 4 x float> %a) {
+; CHECK-LABEL: faddv_nxv4f32:
+; CHECK:      ptrue p0.s
+; CHECK-NEXT: faddv s1, p0, z1.s
+; CHECK-NEXT: fadd s0, s0, s1
+; CHECK-NEXT: ret
+  %res = call fast float @llvm.vector.reduce.fadd.nxv4f32(float %init, <vscale x 4 x float> %a)
+  ret float %res
+}
+
+define double @faddv_nxv2f64(double %init, <vscale x 2 x double> %a) {
+; CHECK-LABEL: faddv_nxv2f64:
+; CHECK:      ptrue p0.d
+; CHECK-NEXT: faddv d1, p0, z1.d
+; CHECK-NEXT: fadd d0, d0, d1
+; CHECK-NEXT: ret
+  %res = call fast double @llvm.vector.reduce.fadd.nxv2f64(double %init, <vscale x 2 x double> %a)
+  ret double %res
+}
+
+; FMAXV
+
+define half @fmaxv_nxv2f16(<vscale x 2 x half> %a) {
+; CHECK-LABEL: fmaxv_nxv2f16:
+; CHECK:      ptrue p0.d
+; CHECK-NEXT: fmaxnmv h0, p0, z0.h
+; CHECK-NEXT: ret
+  %res = call half @llvm.vector.reduce.fmax.nxv2f16(<vscale x 2 x half> %a)
+  ret half %res
+}
+
+define half @fmaxv_nxv4f16(<vscale x 4 x half> %a) {
+; CHECK-LABEL: fmaxv_nxv4f16:
+; CHECK:      ptrue p0.s
+; CHECK-NEXT: fmaxnmv h0, p0, z0.h
+; CHECK-NEXT: ret
+  %res = call half @llvm.vector.reduce.fmax.nxv4f16(<vscale x 4 x half> %a)
+  ret half %res
+}
+
+define half @fmaxv_nxv8f16(<vscale x 8 x half> %a) {
+; CHECK-LABEL: fmaxv_nxv8f16:
+; CHECK:      ptrue p0.h
+; CHECK-NEXT: fmaxnmv h0, p0, z0.h
+; CHECK-NEXT: ret
+  %res = call half @llvm.vector.reduce.fmax.nxv8f16(<vscale x 8 x half> %a)
+  ret half %res
+}
+
+define float @fmaxv_nxv2f32(<vscale x 2 x float> %a) {
+; CHECK-LABEL: fmaxv_nxv2f32:
+; CHECK:      ptrue p0.d
+; CHECK-NEXT: fmaxnmv s0, p0, z0.s
+; CHECK-NEXT: ret
+  %res = call float @llvm.vector.reduce.fmax.nxv2f32(<vscale x 2 x float> %a)
+  ret float %res
+}
+
+define float @fmaxv_nxv4f32(<vscale x 4 x float> %a) {
+; CHECK-LABEL: fmaxv_nxv4f32:
+; CHECK:      ptrue p0.s
+; CHECK-NEXT: fmaxnmv s0, p0, z0.s
+; CHECK-NEXT: ret
+  %res = call float @llvm.vector.reduce.fmax.nxv4f32(<vscale x 4 x float> %a)
+  ret float %res
+}
+
+define double @fmaxv_nxv2f64(<vscale x 2 x double> %a) {
+; CHECK-LABEL: fmaxv_nxv2f64:
+; CHECK:      ptrue p0.d
+; CHECK-NEXT: fmaxnmv d0, p0, z0.d
+; CHECK-NEXT: ret
+  %res = call double @llvm.vector.reduce.fmax.nxv2f64(<vscale x 2 x double> %a)
+  ret double %res
+}
+
+; FMINV
+
+define half @fminv_nxv2f16(<vscale x 2 x half> %a) {
+; CHECK-LABEL: fminv_nxv2f16:
+; CHECK:      ptrue p0.d
+; CHECK-NEXT: fminnmv h0, p0, z0.h
+; CHECK-NEXT: ret
+  %res = call half @llvm.vector.reduce.fmin.nxv2f16(<vscale x 2 x half> %a)
+  ret half %res
+}
+
+define half @fminv_nxv4f16(<vscale x 4 x half> %a) {
+; CHECK-LABEL: fminv_nxv4f16:
+; CHECK:      ptrue p0.s
+; CHECK-NEXT: fminnmv h0, p0, z0.h
+; CHECK-NEXT: ret
+  %res = call half @llvm.vector.reduce.fmin.nxv4f16(<vscale x 4 x half> %a)
+  ret half %res
+}
+
+define half @fminv_nxv8f16(<vscale x 8 x half> %a) {
+; CHECK-LABEL: fminv_nxv8f16:
+; CHECK:      ptrue p0.h
+; CHECK-NEXT: fminnmv h0, p0, z0.h
+; CHECK-NEXT: ret
+  %res = call half @llvm.vector.reduce.fmin.nxv8f16(<vscale x 8 x half> %a)
+  ret half %res
+}
+
+define float @fminv_nxv2f32(<vscale x 2 x float> %a) {
+; CHECK-LABEL: fminv_nxv2f32:
+; CHECK:      ptrue p0.d
+; CHECK-NEXT: fminnmv s0, p0, z0.s
+; CHECK-NEXT: ret
+  %res = call float @llvm.vector.reduce.fmin.nxv2f32(<vscale x 2 x float> %a)
+  ret float %res
+}
+
+define float @fminv_nxv4f32(<vscale x 4 x float> %a) {
+; CHECK-LABEL: fminv_nxv4f32:
+; CHECK:      ptrue p0.s
+; CHECK-NEXT: fminnmv s0, p0, z0.s
+; CHECK-NEXT: ret
+  %res = call float @llvm.vector.reduce.fmin.nxv4f32(<vscale x 4 x float> %a)
+  ret float %res
+}
+
+define double @fminv_nxv2f64(<vscale x 2 x double> %a) {
+; CHECK-LABEL: fminv_nxv2f64:
+; CHECK:      ptrue p0.d
+; CHECK-NEXT: fminnmv d0, p0, z0.d
+; CHECK-NEXT: ret
+  %res = call double @llvm.vector.reduce.fmin.nxv2f64(<vscale x 2 x double> %a)
+  ret double %res
+}
+
+declare half @llvm.vector.reduce.fadd.nxv2f16(half, <vscale x 2 x half>)
+declare half @llvm.vector.reduce.fadd.nxv4f16(half, <vscale x 4 x half>)
+declare half @llvm.vector.reduce.fadd.nxv8f16(half, <vscale x 8 x half>)
+declare float @llvm.vector.reduce.fadd.nxv2f32(float, <vscale x 2 x float>)
+declare float @llvm.vector.reduce.fadd.nxv4f32(float, <vscale x 4 x float>)
+declare double @llvm.vector.reduce.fadd.nxv2f64(double, <vscale x 2 x double>)
+
+declare half @llvm.vector.reduce.fmax.nxv2f16(<vscale x 2 x half>)
+declare half @llvm.vector.reduce.fmax.nxv4f16(<vscale x 4 x half>)
+declare half @llvm.vector.reduce.fmax.nxv8f16(<vscale x 8 x half>)
+declare float @llvm.vector.reduce.fmax.nxv2f32(<vscale x 2 x float>)
+declare float @llvm.vector.reduce.fmax.nxv4f32(<vscale x 4 x float>)
+declare double @llvm.vector.reduce.fmax.nxv2f64(<vscale x 2 x double>)
+
+declare half @llvm.vector.reduce.fmin.nxv2f16(<vscale x 2 x half>)
+declare half @llvm.vector.reduce.fmin.nxv4f16(<vscale x 4 x half>)
+declare half @llvm.vector.reduce.fmin.nxv8f16(<vscale x 8 x half>)
+declare float @llvm.vector.reduce.fmin.nxv2f32(<vscale x 2 x float>)
+declare float @llvm.vector.reduce.fmin.nxv4f32(<vscale x 4 x float>)
+declare double @llvm.vector.reduce.fmin.nxv2f64(<vscale x 2 x double>)

llvm/test/CodeGen/AArch64/sve-split-fp-reduce.ll

+; RUN: llc -mtriple=aarch64-linux-gnu -mattr=+sve -asm-verbose=0 < %s 2>%t | FileCheck %s
+; RUN: FileCheck --check-prefix=WARN --allow-empty %s <%t
+
+; If this check fails please read test/CodeGen/AArch64/README for instructions on how to resolve it.
+; WARN-NOT: warning
+
+; FADD
+
+define double @fadda_nxv8f64(double %init, <vscale x 8 x double> %a) {
+; CHECK-LABEL: fadda_nxv8f64
+; CHECK: ptrue p0.d
+; CHECK-NEXT: fadda d0, p0, d0, z1.d
+; CHECK-NEXT: fadda d0, p0, d0, z2.d
+; CHECK-NEXT: fadda d0, p0, d0, z3.d
+; CHECK-NEXT: fadda d0, p0, d0, z4.d
+; CHECK-NEXT: ret
+  %res = call double @llvm.vector.reduce.fadd.nxv8f64(double %init, <vscale x 8 x double> %a)
+  ret double %res
+}
+
+; FADDV
+
+define float @faddv_nxv8f32(float %init, <vscale x 8 x float> %a) {
+; CHECK-LABEL: faddv_nxv8f32:
+; CHECK:     fadd z1.s, z1.s, z2.s
+; CHECK-NEXT: ptrue p0.s
+; CHECK-NEXT: faddv s1, p0, z1.s
+; CHECK-NEXT: fadd s0, s0, s1
+; CHECK-NEXT: ret
+  %res = call fast float @llvm.vector.reduce.fadd.nxv8f32(float %init, <vscale x 8 x float> %a)
+  ret float %res
+}
+
+; FMAXV
+
+define double @fmaxv_nxv8f64(<vscale x 8 x double> %a) {
+; CHECK-LABEL: fmaxv_nxv8f64:
+; CHECK:      ptrue p0.d
+; CHECK-NEXT: fmaxnm z1.d, p0/m, z1.d, z3.d
+; CHECK-NEXT: fmaxnm z0.d, p0/m, z0.d, z2.d
+; CHECK-NEXT: fmaxnm z0.d, p0/m, z0.d, z1.d
+; CHECK-NEXT: fmaxnmv d0, p0, z0.d
+; CHECK-NEXT: ret
+  %res = call double @llvm.vector.reduce.fmax.nxv8f64(<vscale x 8 x double> %a)
+  ret double %res
+}
+
+; FMINV
+
+define half @fminv_nxv16f16(<vscale x 16 x half> %a) {
+; CHECK-LABEL: fminv_nxv16f16:
+; CHECK:      ptrue p0.h
+; CHECK-NEXT: fminnm z0.h, p0/m, z0.h, z1.h
+; CHECK-NEXT: fminnmv h0, p0, z0.h
+; CHECK-NEXT: ret
+  %res = call half @llvm.vector.reduce.fmin.nxv16f16(<vscale x 16 x half> %a)
+  ret half %res
+}
+
+declare double @llvm.vector.reduce.fadd.nxv8f64(double, <vscale x 8 x double>)
+declare float @llvm.vector.reduce.fadd.nxv8f32(float, <vscale x 8 x float>)
+
+declare double @llvm.vector.reduce.fmax.nxv8f64(<vscale x 8 x double>)
+
+declare half @llvm.vector.reduce.fmin.nxv16f16(<vscale x 16 x half>)