[AArch64][SVE] Add fixed length codegen for FP_ROUND/FP_EXTEND

[bsmith]

Summary

Depends on D102498

[paulwalker-arm]

I've added a recommendation that I believe will also simplify D102777.

llvm/lib/Target/AArch64/AArch64ISelLowering.cpp

17751–17774

Although correct can I recommend a different approach. Rather than "any_extending" the scalable value you should be able to actually ISD::ANY_EXTEND the fixed length operand directly. This simplifies the lowering and means the any_extend->uunplo transform is kept in one place. The flow can be something like:

src = bit cast(src_as_int, src)
src = any_extend(dst_as_int, src)
src = convertToScalable(src)
src = svesafecast(src)
src = fpextend(src)

The legaliser will then transform the any_extend separately.

17801

Although this is more direct, if you wanted to loose the line wrapping you can use changeTypeToInteger().

17804–17823

As above, by doing an actual ISD::TRUNCATE of the fixed length result you get to reuse the existing mechanism to lower fixed length truncates. I'm thinking:

Val = convertFromScalableVector(DAG, SrcVT.changeTypeToInteger(), Val);           
Val = DAG.getNode(ISD::TRUNCATE, DL, VT.changeTypeToInteger(), Val);              
return DAG.getNode(ISD::BITCAST, DL, VT, Val);

[bsmith]

• Vastly improve code by re-using fixed length extends/truncates

[paulwalker-arm]

llvm/lib/Target/AArch64/AArch64ISelLowering.cpp

3209–3211

This change looks unnecessary.

17755

This can be SrcVT, which when combined with the next comment means ContainerSrcVT is no longer required.

17760

I'm surprised this works as ContainerSrcVT and Val have different element types? Either way it seems safer to use ContainerDstVT.changeTypeToInteger().

17780

This can be VT, which means ContainerDstVT is no longer needed.

llvm/test/CodeGen/AArch64/sve-fixed-length-fp-extend-trunc.ll

62–68

In general for the fixed length arith tests we've added VBITS_EQ_256 CHECK lines to the 512bit tests, just to show sensible type legalisation (see sve-fixed-length-fp-reduce.ll).

[bsmith]

• Variable name cleanups
• Add tests to check type legalisation

llvm/lib/Target/AArch64/AArch64ISelLowering.h

private:
   SDValue LowerFixedLengthVectorStoreToSVE(SDValue Op, SelectionDAG &DAG) const;
   SDValue LowerFixedLengthVectorTruncateToSVE(SDValue Op,
                                               SelectionDAG &DAG) const;
   SDValue LowerFixedLengthExtractVectorElt(SDValue Op, SelectionDAG &DAG) const;
   SDValue LowerFixedLengthInsertVectorElt(SDValue Op, SelectionDAG &DAG) const;
   SDValue LowerFixedLengthBitcastToSVE(SDValue Op, SelectionDAG &DAG) const;
   SDValue LowerFixedLengthConcatVectorsToSVE(SDValue Op,
                                              SelectionDAG &DAG) const;
+  SDValue LowerFixedLengthFPExtendToSVE(SDValue Op, SelectionDAG &DAG) const;
+  SDValue LowerFixedLengthFPRoundToSVE(SDValue Op, SelectionDAG &DAG) const;
 
   SDValue BuildSDIVPow2(SDNode *N, const APInt &Divisor, SelectionDAG &DAG,
                         SmallVectorImpl<SDNode *> &Created) const override;
   SDValue getSqrtEstimate(SDValue Operand, SelectionDAG &DAG, int Enabled,
                           int &ExtraSteps, bool &UseOneConst,
                           bool Reciprocal) const override;
   SDValue getRecipEstimate(SDValue Operand, SelectionDAG &DAG, int Enabled,
                            int &ExtraSteps) const override;
   SDValue getSqrtInputTest(SDValue Operand, SelectionDAG &DAG,

llvm/lib/Target/AArch64/AArch64ISelLowering.cpp

if (Subtarget->useSVEForFixedLengthVectors()) {
       for (auto VT : {MVT::v8i8, MVT::v4i16})
         setOperationAction(ISD::TRUNCATE, VT, Custom);
       setOperationAction(ISD::FP_ROUND, MVT::v4f16, Custom);
 
       // 128bit results imply a bigger than NEON input.
       for (auto VT : {MVT::v16i8, MVT::v8i16, MVT::v4i32})
         setOperationAction(ISD::TRUNCATE, VT, Custom);
       for (auto VT : {MVT::v8f16, MVT::v4f32})
-        setOperationAction(ISD::FP_ROUND, VT, Expand);
+        setOperationAction(ISD::FP_ROUND, VT, Custom);
 
       // These operations are not supported on NEON but SVE can do them.
       setOperationAction(ISD::BITREVERSE, MVT::v1i64, Custom);
       setOperationAction(ISD::CTLZ, MVT::v1i64, Custom);
       setOperationAction(ISD::CTLZ, MVT::v2i64, Custom);
       setOperationAction(ISD::CTTZ, MVT::v1i64, Custom);
       setOperationAction(ISD::MUL, MVT::v1i64, Custom);
       setOperationAction(ISD::MUL, MVT::v2i64, Custom);
void AArch64TargetLowering::addTypeForFixedLengthSVE(MVT VT) {
   setOperationAction(ISD::FMA, VT, Custom);
   setOperationAction(ISD::FMAXIMUM, VT, Custom);
   setOperationAction(ISD::FMAXNUM, VT, Custom);
   setOperationAction(ISD::FMINIMUM, VT, Custom);
   setOperationAction(ISD::FMINNUM, VT, Custom);
   setOperationAction(ISD::FMUL, VT, Custom);
   setOperationAction(ISD::FNEARBYINT, VT, Custom);
   setOperationAction(ISD::FNEG, VT, Custom);
+  setOperationAction(ISD::FP_EXTEND, VT, Custom);
+  setOperationAction(ISD::FP_ROUND, VT, Custom);
   setOperationAction(ISD::FRINT, VT, Custom);
   setOperationAction(ISD::FROUND, VT, Custom);
   setOperationAction(ISD::FROUNDEVEN, VT, Custom);
   setOperationAction(ISD::FSQRT, VT, Custom);
   setOperationAction(ISD::FSUB, VT, Custom);
   setOperationAction(ISD::FTRUNC, VT, Custom);
   setOperationAction(ISD::LOAD, VT, Custom);
   setOperationAction(ISD::MUL, VT, Custom);
unsigned PrfOp = (IsWrite << 4) |     // Load/Store bit
                    (Locality << 1) |    // Cache level bits
                    (unsigned)IsStream;  // Stream bit
   return DAG.getNode(AArch64ISD::PREFETCH, DL, MVT::Other, Op.getOperand(0),
                      DAG.getConstant(PrfOp, DL, MVT::i32), Op.getOperand(1));
 }
 
 SDValue AArch64TargetLowering::LowerFP_EXTEND(SDValue Op,
                                               SelectionDAG &DAG) const {
-  if (Op.getValueType().isScalableVector())
+  EVT VT = Op.getValueType();
+  if (VT.isScalableVector())
     return LowerToPredicatedOp(Op, DAG, AArch64ISD::FP_EXTEND_MERGE_PASSTHRU);
 
+  if (useSVEForFixedLengthVectorVT(VT))
+    return LowerFixedLengthFPExtendToSVE(Op, DAG);
+
   assert(Op.getValueType() == MVT::f128 && "Unexpected lowering");
   return SDValue();
 }
 
 SDValue AArch64TargetLowering::LowerFP_ROUND(SDValue Op,
                                              SelectionDAG &DAG) const {
-  if (Op.getValueType().isScalableVector())
+  EVT VT = Op.getValueType();
+  if (VT.isScalableVector())
     return LowerToPredicatedOp(Op, DAG, AArch64ISD::FP_ROUND_MERGE_PASSTHRU);
 
   bool IsStrict = Op->isStrictFPOpcode();
   SDValue SrcVal = Op.getOperand(IsStrict ? 1 : 0);
   EVT SrcVT = SrcVal.getValueType();
 
+  if (useSVEForFixedLengthVectorVT(SrcVT))
+    return LowerFixedLengthFPRoundToSVE(Op, DAG);
+
   if (SrcVT != MVT::f128) {
     // Expand cases where the input is a vector bigger than NEON.
     if (useSVEForFixedLengthVectorVT(SrcVT))
       return SDValue();
 
     // It's legal except when f128 is involved
     return Op;
   }
if (SrcOp2.isUndef())
     Op = SrcOp1;
   else
     Op =
         DAG.getNode(AArch64ISD::SPLICE, DL, ContainerDstVT, Pg, SrcOp1, SrcOp2);
 
   return convertFromScalableVector(DAG, VT, Op);
 }
 
+SDValue
+AArch64TargetLowering::LowerFixedLengthFPExtendToSVE(SDValue Op,
+                                                     SelectionDAG &DAG) const {
+  EVT VT = Op.getValueType();
+  assert(VT.isFixedLengthVector() && "Expected fixed length vector type!");
+
+  SDLoc DL(Op);
+  SDValue Val = Op.getOperand(0);
+  SDValue Pg = getPredicateForVector(DAG, DL, VT);
+  EVT ContainerDstVT = getContainerForFixedLengthVector(DAG, VT);
+  EVT ContainerSrcVT =
+      getContainerForFixedLengthVector(DAG, Val.getValueType());
+  EVT ExtendVT = ContainerDstVT.changeVectorElementType(
+      ContainerSrcVT.getVectorElementType());
+
+  Val = convertToScalableVector(DAG, ContainerSrcVT, Val);
+  Val = DAG.getNode(ISD::BITCAST, DL,
+                    ContainerSrcVT.changeVectorElementTypeToInteger(), Val);
+
+  // Repeatedly unpack Val until the result is of the desired element type.
+  switch (Val.getValueType().getSimpleVT().SimpleTy) {
+  default:
+    llvm_unreachable("unimplemented container type");
+  case MVT::nxv8i16:
+    Val = DAG.getNode(AArch64ISD::UUNPKLO, DL, MVT::nxv4i32, Val);
+    if (VT.getVectorElementType() == MVT::f32)
+      break;
+    LLVM_FALLTHROUGH;
+  case MVT::nxv4i32:
+    Val = DAG.getNode(AArch64ISD::UUNPKLO, DL, MVT::nxv2i64, Val);
+    assert(VT.getVectorElementType() == MVT::f64 && "Unexpected element type!");
+    break;
+  }
+
+  Val = getSVESafeBitCast(ExtendVT, Val, DAG);
+  Val = DAG.getNode(AArch64ISD::FP_EXTEND_MERGE_PASSTHRU, DL, ContainerDstVT,
+                    Pg, Val, DAG.getUNDEF(ContainerDstVT));
+
+  return convertFromScalableVector(DAG, VT, Val);
+}
+
+SDValue
+AArch64TargetLowering::LowerFixedLengthFPRoundToSVE(SDValue Op,
+                                                    SelectionDAG &DAG) const {
+  EVT VT = Op.getValueType();
+  assert(VT.isFixedLengthVector() && "Expected fixed length vector type!");
+
+  SDLoc DL(Op);
+  SDValue Val = Op.getOperand(0);
+  EVT ContainerDstVT = getContainerForFixedLengthVector(DAG, VT);
+  EVT ContainerSrcVT =
+      getContainerForFixedLengthVector(DAG, Val.getValueType());
+  EVT RoundVT = ContainerSrcVT.changeVectorElementType(
+      ContainerDstVT.getVectorElementType());
+  SDValue Pg = getPredicateForVector(DAG, DL, RoundVT);
+
+  Val = convertToScalableVector(DAG, ContainerSrcVT, Val);
+  Val = DAG.getNode(AArch64ISD::FP_ROUND_MERGE_PASSTHRU, DL, RoundVT, Pg, Val,
+                    Op.getOperand(1), DAG.getUNDEF(RoundVT));
+  Val = getSVESafeBitCast(ContainerSrcVT.changeVectorElementTypeToInteger(),
+                          Val, DAG);
+
+  // Repeatedly truncate Val until the result is of the desired element type.
+  switch (Val.getValueType().getSimpleVT().SimpleTy) {
+  default:
+    llvm_unreachable("unimplemented container type");
+  case MVT::nxv2i64:
+    Val = DAG.getNode(ISD::BITCAST, DL, MVT::nxv4i32, Val);
+    Val = DAG.getNode(AArch64ISD::UZP1, DL, MVT::nxv4i32, Val, Val);
+    if (VT.getVectorElementType() == MVT::f32)
+      break;
+    LLVM_FALLTHROUGH;
+  case MVT::nxv4i32:
+    Val = DAG.getNode(ISD::BITCAST, DL, MVT::nxv8i16, Val);
+    Val = DAG.getNode(AArch64ISD::UZP1, DL, MVT::nxv8i16, Val, Val);
+    if (VT.getVectorElementType() == MVT::f16)
+      break;
+  }
+
+  Val = DAG.getNode(ISD::BITCAST, DL, ContainerDstVT, Val);
+
+  return convertFromScalableVector(DAG, VT, Val);
+}
+
 SDValue AArch64TargetLowering::getSVESafeBitCast(EVT VT, SDValue Op,
                                                  SelectionDAG &DAG) const {
   SDLoc DL(Op);
   EVT InVT = Op.getValueType();
   const TargetLowering &TLI = DAG.getTargetLoweringInfo();
   (void)TLI;
 
   assert(VT.isScalableVector() && TLI.isTypeLegal(VT) &&
          InVT.isScalableVector() && TLI.isTypeLegal(InVT) &&
          "Only expect to cast between legal scalable vector types!");
   assert((VT.getVectorElementType() == MVT::i1) ==
              (InVT.getVectorElementType() == MVT::i1) &&
          "Cannot cast between data and predicate scalable vector types!");
 
   if (InVT == VT)
     return Op;
 
   if (VT.getVectorElementType() == MVT::i1)
     return DAG.getNode(AArch64ISD::REINTERPRET_CAST, DL, VT, Op);
 
   EVT PackedVT = getPackedSVEVectorVT(VT.getVectorElementType());
   EVT PackedInVT = getPackedSVEVectorVT(InVT.getVectorElementType());
   assert((VT == PackedVT || InVT == PackedInVT) &&
          "Cannot cast between unpacked scalable vector types!");
 
   // Pack input if required.
   if (InVT != PackedInVT)

llvm/test/CodeGen/AArch64/sve-fixed-length-fp-converts.ll

 ; NO_SVE-NOT: z{0-9}
 
 ; NOTE: fptrunc operations bigger than NEON are expanded. These tests just
 ; ensure we've correctly set the operation action for fixed length vector types
 ; that require SVE. They'll be updated to protect their expected code generation
 ; when lowering it implemented.
 
 ;
-; fptrunc f32 -> f16
-;
-
-define <8 x half> @fptrunc_v8f32_v8f16(<8 x float>* %in) #0 {
-; CHECK-LABEL: fptrunc_v8f32_v8f16:
-; CHECK-COUNT-8: fcvt h{{[0-9]+}}, s{{[0-9]+}}
-; CHECK-NOT: fcvt
-; CHECK: ret
-  %a = load <8 x float>, <8 x float>* %in
-  %b = fptrunc <8 x float> %a to <8 x half>
-  ret <8 x half> %b
-}
-
-define void @fptrunc_v16f32_v16f16(<16 x float>* %in, <16 x half>* %out) #0 {
-; CHECK-LABEL: fptrunc_v16f32_v16f16:
-; CHECK-COUNT-16: fcvt h{{[0-9]+}}, s{{[0-9]+}}
-; CHECK-NOT: fcvt
-; CHECK: ret
-  %a = load <16 x float>, <16 x float>* %in
-  %b = fptrunc <16 x float> %a to <16 x half>
-  store <16 x half> %b, <16 x half>* %out
-  ret void
-}
-
-define void @fptrunc_v32f32_v32f16(<32 x float>* %in, <32 x half>* %out) #0 {
-; CHECK-LABEL: fptrunc_v32f32_v32f16:
-; CHECK-COUNT-32: fcvt h{{[0-9]+}}, s{{[0-9]+}}
-; CHECK-NOT: fcvt
-; CHECK: ret
-  %a = load <32 x float>, <32 x float>* %in
-  %b = fptrunc <32 x float> %a to <32 x half>
-  store <32 x half> %b, <32 x half>* %out
-  ret void
-}
-
-define void @fptrunc_v64f32_v64f16(<64 x float>* %in, <64 x half>* %out) #0 {
-; CHECK-LABEL: fptrunc_v64f32_v64f16:
-; CHECK-COUNT-64: fcvt h{{[0-9]+}}, s{{[0-9]+}}
-; CHECK-NOT: fcvt
-; CHECK: ret
-  %a = load <64 x float>, <64 x float>* %in
-  %b = fptrunc <64 x float> %a to <64 x half>
-  store <64 x half> %b, <64 x half>* %out
-  ret void
-}
-
-;
-; fptrunc f64 -> f16
-;
-
-define <4 x half> @fptrunc_v4f64_v4f16(<4 x double>* %in) #0 {
-; CHECK-LABEL: fptrunc_v4f64_v4f16:
-; CHECK-COUNT-4: fcvt h{{[0-9]+}}, d{{[0-9]+}}
-; CHECK-NOT: fcvt
-; CHECK: ret
-  %a = load <4 x double>, <4 x double>* %in
-  %b = fptrunc <4 x double> %a to <4 x half>
-  ret <4 x half> %b
-}
-
-define <8 x half> @fptrunc_v8f64_v8f16(<8 x double>* %in) #0 {
-; CHECK-LABEL: fptrunc_v8f64_v8f16:
-; CHECK-COUNT-8: fcvt h{{[0-9]+}}, d{{[0-9]+}}
-; CHECK-NOT: fcvt
-; CHECK: ret
-  %a = load <8 x double>, <8 x double>* %in
-  %b = fptrunc <8 x double> %a to <8 x half>
-  ret <8 x half> %b
-}
-
-define void @fptrunc_v16f64_v16f16(<16 x double>* %in, <16 x half>* %out) #0 {
-; CHECK-LABEL: fptrunc_v16f64_v16f16:
-; CHECK-COUNT-16: fcvt h{{[0-9]+}}, d{{[0-9]+}}
-; CHECK-NOT: fcvt
-; CHECK: ret
-  %a = load <16 x double>, <16 x double>* %in
-  %b = fptrunc <16 x double> %a to <16 x half>
-  store <16 x half> %b, <16 x half>* %out
-  ret void
-}
-
-define void @fptrunc_v32f64_v32f16(<32 x double>* %in, <32 x half>* %out) #0 {
-; CHECK-LABEL: fptrunc_v32f64_v32f16:
-; CHECK-COUNT-32: fcvt h{{[0-9]+}}, d{{[0-9]+}}
-; CHECK-NOT: fcvt
-; CHECK: ret
-  %a = load <32 x double>, <32 x double>* %in
-  %b = fptrunc <32 x double> %a to <32 x half>
-  store <32 x half> %b, <32 x half>* %out
-  ret void
-}
-
-;
-; fptrunc f64 -> f32
-;
-
-define <4 x float> @fptrunc_v4f64_v4f32(<4 x double>* %in) #0 {
-; CHECK-LABEL: fptrunc_v4f64_v4f32:
-; CHECK-COUNT-4: fcvt s{{[0-9]+}}, d{{[0-9]+}}
-; CHECK-NOT: fcvt
-; CHECK: ret
-  %a = load <4 x double>, <4 x double>* %in
-  %b = fptrunc <4 x double> %a to <4 x float>
-  ret <4 x float> %b
-}
-
-define void @fptrunc_v8f64_v8f32(<8 x double>* %in, <8 x float>* %out) #0 {
-; CHECK-LABEL: fptrunc_v8f64_v8f32:
-; CHECK-COUNT-8: fcvt s{{[0-9]+}}, d{{[0-9]+}}
-; CHECK-NOT: fcvt
-; CHECK: ret
-  %a = load <8 x double>, <8 x double>* %in
-  %b = fptrunc <8 x double> %a to <8 x float>
-  store <8 x float> %b, <8 x float>* %out
-  ret void
-}
-
-define void @fptrunc_v16f64_v16f32(<16 x double>* %in, <16 x float>* %out) #0 {
-; CHECK-LABEL: fptrunc_v16f64_v16f32:
-; CHECK-COUNT-16: fcvt s{{[0-9]+}}, d{{[0-9]+}}
-; CHECK-NOT: fcvt
-; CHECK: ret
-  %a = load <16 x double>, <16 x double>* %in
-  %b = fptrunc <16 x double> %a to <16 x float>
-  store <16 x float> %b, <16 x float>* %out
-  ret void
-}
-
-define void @fptrunc_v32f64_v32f32(<32 x double>* %in, <32 x float>* %out) #0 {
-; CHECK-LABEL: fptrunc_v32f64_v32f32:
-; CHECK-COUNT-32: fcvt s{{[0-9]+}}, d{{[0-9]+}}
-; CHECK-NOT: fcvt
-; CHECK: ret
-  %a = load <32 x double>, <32 x double>* %in
-  %b = fptrunc <32 x double> %a to <32 x float>
-  store <32 x float> %b, <32 x float>* %out
-  ret void
-}
-
-;
 ; vector uint_to_fp i8 -> f32
 ; AArch64 doesn't have a direct vector->f32 conversion instructions for
 ; elements smaller than i32, so make sure inputs are promoted to i32 first.
 ;
 
 define void @uitofp_v4i8_v4f32(<4 x i8>* %in, <4 x float>* %out) #0 {
 ; CHECK-LABEL: uitofp_v4i8_v4f32:
 ; CHECK-COUNT-1: ucvt

llvm/test/CodeGen/AArch64/sve-fixed-length-fp-extend-trunc.ll

+; RUN: llc -aarch64-sve-vector-bits-min=128  -asm-verbose=0 < %s | FileCheck %s -check-prefix=NO_SVE
+; RUN: llc -aarch64-sve-vector-bits-min=256  -asm-verbose=0 < %s | FileCheck %s -check-prefixes=CHECK
+; RUN: llc -aarch64-sve-vector-bits-min=384  -asm-verbose=0 < %s | FileCheck %s -check-prefixes=CHECK
+; RUN: llc -aarch64-sve-vector-bits-min=512  -asm-verbose=0 < %s | FileCheck %s -check-prefixes=CHECK,VBITS_GE_512
+; RUN: llc -aarch64-sve-vector-bits-min=640  -asm-verbose=0 < %s | FileCheck %s -check-prefixes=CHECK,VBITS_GE_512
+; RUN: llc -aarch64-sve-vector-bits-min=768  -asm-verbose=0 < %s | FileCheck %s -check-prefixes=CHECK,VBITS_GE_512
+; RUN: llc -aarch64-sve-vector-bits-min=896  -asm-verbose=0 < %s | FileCheck %s -check-prefixes=CHECK,VBITS_GE_512
+; RUN: llc -aarch64-sve-vector-bits-min=1024 -asm-verbose=0 < %s | FileCheck %s -check-prefixes=CHECK,VBITS_GE_512,VBITS_GE_1024
+; RUN: llc -aarch64-sve-vector-bits-min=1152 -asm-verbose=0 < %s | FileCheck %s -check-prefixes=CHECK,VBITS_GE_512,VBITS_GE_1024
+; RUN: llc -aarch64-sve-vector-bits-min=1280 -asm-verbose=0 < %s | FileCheck %s -check-prefixes=CHECK,VBITS_GE_512,VBITS_GE_1024
+; RUN: llc -aarch64-sve-vector-bits-min=1408 -asm-verbose=0 < %s | FileCheck %s -check-prefixes=CHECK,VBITS_GE_512,VBITS_GE_1024
+; RUN: llc -aarch64-sve-vector-bits-min=1536 -asm-verbose=0 < %s | FileCheck %s -check-prefixes=CHECK,VBITS_GE_512,VBITS_GE_1024
+; RUN: llc -aarch64-sve-vector-bits-min=1664 -asm-verbose=0 < %s | FileCheck %s -check-prefixes=CHECK,VBITS_GE_512,VBITS_GE_1024
+; RUN: llc -aarch64-sve-vector-bits-min=1792 -asm-verbose=0 < %s | FileCheck %s -check-prefixes=CHECK,VBITS_GE_512,VBITS_GE_1024
+; RUN: llc -aarch64-sve-vector-bits-min=1920 -asm-verbose=0 < %s | FileCheck %s -check-prefixes=CHECK,VBITS_GE_512,VBITS_GE_1024
+; RUN: llc -aarch64-sve-vector-bits-min=2048 -asm-verbose=0 < %s | FileCheck %s -check-prefixes=CHECK,VBITS_GE_512,VBITS_GE_1024,VBITS_GE_2048
+
+target triple = "aarch64-unknown-linux-gnu"
+
+; Don't use SVE when its registers are no bigger than NEON.
+; NO_SVE-NOT: ptrue
+
+;
+; FCVT H -> S
+;
+
+; Don't use SVE for 64-bit vectors.
+define <2 x float> @fcvt_v2f16_v2f32(<2 x half> %op1) #0 {
+; CHECK-LABEL: fcvt_v2f16_v2f32:
+; CHECK: fcvtl v0.4s, v0.4h
+; CHECK-NEXT: ret
+  %res = fpext <2 x half> %op1 to <2 x float>
+  ret <2 x float> %res
+}
+
+; Don't use SVE for 128-bit vectors.
+define <4 x float> @fcvt_v4f16_v4f32(<4 x half> %op1) #0 {
+; CHECK-LABEL: fcvt_v4f16_v4f32:
+; CHECK: fcvtl v0.4s, v0.4h
+; CHECK-NEXT: ret
+  %res = fpext <4 x half> %op1 to <4 x float>
+  ret <4 x float> %res
+}
+
+define void @fcvt_v8f16_v8f32(<8 x half>* %a, <8 x float>* %b) #0 {
+; CHECK-LABEL: fcvt_v8f16_v8f32:
+; CHECK: ldr q[[OP:[0-9]+]], [x0]
+; CHECK-NEXT: ptrue [[PG:p[0-9]+]].s, vl8
+; CHECK-NEXT: uunpklo [[UPK:z[0-9]+]].s, z[[OP]].h
+; CHECK-NEXT: fcvt [[RES:z[0-9]+]].s, [[PG]]/m, [[UPK]].h
+; CHECK-NEXT: st1w { [[RES]].s }, [[PG]], [x1]
+; CHECK-NEXT: ret
+  %op1 = load <8 x half>, <8 x half>* %a
+  %res = fpext <8 x half> %op1 to <8 x float>
+  store <8 x float> %res, <8 x float>* %b
+  ret void
+}
+
+define void @fcvt_v16f16_v16f32(<16 x half>* %a, <16 x float>* %b) #0 {
+; CHECK-LABEL: fcvt_v16f16_v16f32:
+; VBITS_GE_512: ptrue [[PG1:p[0-9]+]].h, vl16
+; VBITS_GE_512-NEXT: ld1h { [[OP:z[0-9]+]].h }, [[PG1]]/z, [x0]
+; VBITS_GE_512-NEXT: ptrue [[PG2:p[0-9]+]].s, vl16
+; VBITS_GE_512-NEXT: uunpklo [[UPK:z[0-9]+]].s, [[OP]].h
+; VBITS_GE_512-NEXT: fcvt [[RES:z[0-9]+]].s, [[PG2]]/m, [[UPK]].h
+; VBITS_GE_512-NEXT: st1w { [[RES]].s }, [[PG1]], [x1]
+; VBITS_GE_512-NEXT: ret
+  %op1 = load <16 x half>, <16 x half>* %a
+  %res = fpext <16 x half> %op1 to <16 x float>
+  store <16 x float> %res, <16 x float>* %b
+  ret void
+}
+
+define void @fcvt_v32f16_v32f32(<32 x half>* %a, <32 x float>* %b) #0 {
+; CHECK-LABEL: fcvt_v32f16_v32f32:
+; VBITS_GE_1024: ptrue [[PG1:p[0-9]+]].h, vl32
+; VBITS_GE_1024-NEXT: ld1h { [[OP:z[0-9]+]].h }, [[PG1]]/z, [x0]
+; VBITS_GE_1024-NEXT: ptrue [[PG2:p[0-9]+]].s, vl32
+; VBITS_GE_1024-NEXT: uunpklo [[UPK:z[0-9]+]].s, [[OP]].h
+; VBITS_GE_1024-NEXT: fcvt [[RES:z[0-9]+]].s, [[PG2]]/m, [[UPK]].h
+; VBITS_GE_1024-NEXT: st1w { [[RES]].s }, [[PG1]], [x1]
+; VBITS_GE_1024-NEXT: ret
+  %op1 = load <32 x half>, <32 x half>* %a
+  %res = fpext <32 x half> %op1 to <32 x float>
+  store <32 x float> %res, <32 x float>* %b
+  ret void
+}
+
+define void @fcvt_v64f16_v64f32(<64 x half>* %a, <64 x float>* %b) #0 {
+; CHECK-LABEL: fcvt_v64f16_v64f32:
+; VBITS_GE_2048: ptrue [[PG1:p[0-9]+]].h, vl64
+; VBITS_GE_2048-NEXT: ld1h { [[OP:z[0-9]+]].h }, [[PG1]]/z, [x0]
+; VBITS_GE_2048-NEXT: ptrue [[PG2:p[0-9]+]].s, vl64
+; VBITS_GE_2048-NEXT: uunpklo [[UPK:z[0-9]+]].s, [[OP]].h
+; VBITS_GE_2048-NEXT: fcvt [[RES:z[0-9]+]].s, [[PG2]]/m, [[UPK]].h
+; VBITS_GE_2048-NEXT: st1w { [[RES]].s }, [[PG1]], [x1]
+; VBITS_GE_2048-NEXT: ret
+  %op1 = load <64 x half>, <64 x half>* %a
+  %res = fpext <64 x half> %op1 to <64 x float>
+  store <64 x float> %res, <64 x float>* %b
+  ret void
+}
+
+;
+; FCVT H -> D
+;
+
+; Don't use SVE for 64-bit vectors.
+define <1 x double> @fcvt_v1f16_v1f64(<1 x half> %op1) #0 {
+; CHECK-LABEL: fcvt_v1f16_v1f64:
+; CHECK: fcvt d0, h0
+; CHECK-NEXT: ret
+  %res = fpext <1 x half> %op1 to <1 x double>
+  ret <1 x double> %res
+}
+
+; v2f16 is not legal for NEON, so use SVE
+define <2 x double> @fcvt_v2f16_v2f64(<2 x half> %op1) #0 {
+; CHECK-LABEL: fcvt_v2f16_v2f64:
+; CHECK: ptrue [[PG:p[0-9]+]].d, vl4
+; CHECK-NEXT: uunpklo [[UPK1:z[0-9]+]].s, z0.h
+; CHECK-NEXT: uunpklo [[UPK2:z[0-9]+]].d, [[UPK1]].s
+; CHECK-NEXT: fcvt z0.d, [[PG]]/m, [[UPK2]].h
+; CHECK-NEXT: ret
+  %res = fpext <2 x half> %op1 to <2 x double>
+  ret <2 x double> %res
+}
+
+define void @fcvt_v4f16_v4f64(<4 x half>* %a, <4 x double>* %b) #0 {
+; CHECK-LABEL: fcvt_v4f16_v4f64:
+; CHECK: ldr d[[OP:[0-9]+]], [x0]
+; CHECK-NEXT: ptrue [[PG:p[0-9]+]].d, vl4
+; CHECK-NEXT: uunpklo [[UPK1:z[0-9]+]].s, z[[OP]].h
+; CHECK-NEXT: uunpklo [[UPK2:z[0-9]+]].d, [[UPK1]].s
+; CHECK-NEXT: fcvt [[RES:z[0-9]+]].d, [[PG]]/m, [[UPK2]].h
+; CHECK-NEXT: st1d { [[RES]].d }, [[PG]], [x1]
+; CHECK-NEXT: ret
+  %op1 = load <4 x half>, <4 x half>* %a
+  %res = fpext <4 x half> %op1 to <4 x double>
+  store <4 x double> %res, <4 x double>* %b
+  ret void
+}
+
+define void @fcvt_v8f16_v8f64(<8 x half>* %a, <8 x double>* %b) #0 {
+; CHECK-LABEL: fcvt_v8f16_v8f64:
+; VBITS_GE_512: ldr q[[OP:[0-9]+]], [x0]
+; VBITS_GE_512-NEXT: ptrue [[PG:p[0-9]+]].d, vl8
+; VBITS_GE_512-NEXT: uunpklo [[UPK1:z[0-9]+]].s, z[[OP]].h
+; VBITS_GE_512-NEXT: uunpklo [[UPK2:z[0-9]+]].d, [[UPK1]].s
+; VBITS_GE_512-NEXT: fcvt [[RES:z[0-9]+]].d, [[PG]]/m, [[UPK2]].h
+; VBITS_GE_512-NEXT: st1d { [[RES]].d }, [[PG]], [x1]
+; VBITS_GE_512-NEXT: ret
+  %op1 = load <8 x half>, <8 x half>* %a
+  %res = fpext <8 x half> %op1 to <8 x double>
+  store <8 x double> %res, <8 x double>* %b
+  ret void
+}
+
+define void @fcvt_v16f16_v16f64(<16 x half>* %a, <16 x double>* %b) #0 {
+; CHECK-LABEL: fcvt_v16f16_v16f64:
+; VBITS_GE_1024: ptrue [[PG1:p[0-9]+]].h, vl16
+; VBITS_GE_1024-NEXT: ld1h { [[OP:z[0-9]+]].h }, [[PG1]]/z, [x0]
+; VBITS_GE_1024-NEXT: ptrue [[PG2:p[0-9]+]].d, vl16
+; VBITS_GE_1024-NEXT: uunpklo [[UPK1:z[0-9]+]].s, [[OP]].h
+; VBITS_GE_1024-NEXT: uunpklo [[UPK2:z[0-9]+]].d, [[UPK1]].s
+; VBITS_GE_1024-NEXT: fcvt [[RES:z[0-9]+]].d, [[PG2]]/m, [[UPK2]].h
+; VBITS_GE_1024-NEXT: st1d { [[RES]].d }, [[PG1]], [x1]
+; VBITS_GE_1024-NEXT: ret
+  %op1 = load <16 x half>, <16 x half>* %a
+  %res = fpext <16 x half> %op1 to <16 x double>
+  store <16 x double> %res, <16 x double>* %b
+  ret void
+}
+
+define void @fcvt_v32f16_v32f64(<32 x half>* %a, <32 x double>* %b) #0 {
+; CHECK-LABEL: fcvt_v32f16_v32f64:
+; VBITS_GE_2048: ptrue [[PG1:p[0-9]+]].h, vl32
+; VBITS_GE_2048-NEXT: ld1h { [[OP:z[0-9]+]].h }, [[PG1]]/z, [x0]
+; VBITS_GE_2048-NEXT: ptrue [[PG2:p[0-9]+]].d, vl32
+; VBITS_GE_2048-NEXT: uunpklo [[UPK1:z[0-9]+]].s, [[OP]].h
+; VBITS_GE_2048-NEXT: uunpklo [[UPK2:z[0-9]+]].d, [[UPK]].s
+; VBITS_GE_2048-NEXT: fcvt [[RES:z[0-9]+]].d, [[PG2]]/m, [[UPK2]].h
+; VBITS_GE_2048-NEXT: st1d { [[RES]].d }, [[PG1]], [x1]
+; VBITS_GE_2048-NEXT: ret
+  %op1 = load <32 x half>, <32 x half>* %a
+  %res = fpext <32 x half> %op1 to <32 x double>
+  store <32 x double> %res, <32 x double>* %b
+  ret void
+}
+
+;
+; FCVT S -> D
+;
+
+; Don't use SVE for 64-bit vectors.
+define <1 x double> @fcvt_v1f32_v1f64(<1 x float> %op1) #0 {
+; CHECK-LABEL: fcvt_v1f32_v1f64:
+; CHECK: fcvtl v0.2d, v0.2s
+; CHECK-NEXT: ret
+  %res = fpext <1 x float> %op1 to <1 x double>
+  ret <1 x double> %res
+}
+
+; Don't use SVE for 128-bit vectors.
+define <2 x double> @fcvt_v2f32_v2f64(<2 x float> %op1) #0 {
+; CHECK-LABEL: fcvt_v2f32_v2f64:
+; CHECK: fcvtl v0.2d, v0.2s
+; CHECK-NEXT: ret
+  %res = fpext <2 x float> %op1 to <2 x double>
+  ret <2 x double> %res
+}
+
+define void @fcvt_v4f32_v4f64(<4 x float>* %a, <4 x double>* %b) #0 {
+; CHECK-LABEL: fcvt_v4f32_v4f64:
+; CHECK: ldr q[[OP:[0-9]+]], [x0]
+; CHECK-NEXT: ptrue [[PG:p[0-9]+]].d, vl4
+; CHECK-NEXT: uunpklo [[UPK:z[0-9]+]].d, z[[OP]].s
+; CHECK-NEXT: fcvt [[RES:z[0-9]+]].d, [[PG]]/m, [[UPK]].s
+; CHECK-NEXT: st1d { [[RES]].d }, [[PG]], [x1]
+; CHECK-NEXT: ret
+  %op1 = load <4 x float>, <4 x float>* %a
+  %res = fpext <4 x float> %op1 to <4 x double>
+  store <4 x double> %res, <4 x double>* %b
+  ret void
+}
+
+define void @fcvt_v8f32_v8f64(<8 x float>* %a, <8 x double>* %b) #0 {
+; CHECK-LABEL: fcvt_v8f32_v8f64:
+; VBITS_GE_512: ptrue [[PG1:p[0-9]+]].s, vl8
+; VBITS_GE_512-NEXT: ld1w { [[OP:z[0-9]+]].s }, [[PG1]]/z, [x0]
+; VBITS_GE_512-NEXT: ptrue [[PG:p[0-9]+]].d, vl8
+; VBITS_GE_512-NEXT: uunpklo [[UPK:z[0-9]+]].d, [[OP]].s
+; VBITS_GE_512-NEXT: fcvt [[RES:z[0-9]+]].d, [[PG1]]/m, [[UPK]].s
+; VBITS_GE_512-NEXT: st1d { [[RES]].d }, [[PG1]], [x1]
+; VBITS_GE_512-NEXT: ret
+  %op1 = load <8 x float>, <8 x float>* %a
+  %res = fpext <8 x float> %op1 to <8 x double>
+  store <8 x double> %res, <8 x double>* %b
+  ret void
+}
+
+define void @fcvt_v16f32_v16f64(<16 x float>* %a, <16 x double>* %b) #0 {
+; CHECK-LABEL: fcvt_v16f32_v16f64:
+; VBITS_GE_1024: ptrue [[PG1:p[0-9]+]].s, vl16
+; VBITS_GE_1024-NEXT: ld1w { [[OP:z[0-9]+]].s }, [[PG1]]/z, [x0]
+; VBITS_GE_1024-NEXT: ptrue [[PG2:p[0-9]+]].d, vl16
+; VBITS_GE_1024-NEXT: uunpklo [[UPK:z[0-9]+]].d, [[OP]].s
+; VBITS_GE_1024-NEXT: fcvt [[RES:z[0-9]+]].d, [[PG2]]/m, [[UPK]].s
+; VBITS_GE_1024-NEXT: st1d { [[RES]].d }, [[PG1]], [x1]
+; VBITS_GE_1024-NEXT: ret
+  %op1 = load <16 x float>, <16 x float>* %a
+  %res = fpext <16 x float> %op1 to <16 x double>
+  store <16 x double> %res, <16 x double>* %b
+  ret void
+}
+
+define void @fcvt_v32f32_v32f64(<32 x float>* %a, <32 x double>* %b) #0 {
+; CHECK-LABEL: fcvt_v32f32_v32f64:
+; VBITS_GE_2048: ptrue [[PG1:p[0-9]+]].s, vl32
+; VBITS_GE_2048-NEXT: ld1w { [[OP:z[0-9]+]].s }, [[PG1]]/z, [x0]
+; VBITS_GE_2048-NEXT: ptrue [[PG2:p[0-9]+]].d, vl32
+; VBITS_GE_2048-NEXT: uunpklo [[UPK:z[0-9]+]].d, [[OP]].s
+; VBITS_GE_2048-NEXT: fcvt [[RES:z[0-9]+]].d, [[PG2]]/m, [[UPK]].s
+; VBITS_GE_2048-NEXT: st1d { [[RES]].d }, [[PG1]], [x1]
+; VBITS_GE_2048-NEXT: ret
+  %op1 = load <32 x float>, <32 x float>* %a
+  %res = fpext <32 x float> %op1 to <32 x double>
+  store <32 x double> %res, <32 x double>* %b
+  ret void
+}
+
+;
+; FCVT S -> H
+;
+
+; Don't use SVE for 64-bit vectors.
+define <2 x half> @fcvt_v2f32_v2f16(<2 x float> %op1) #0 {
+; CHECK-LABEL: fcvt_v2f32_v2f16:
+; CHECK: fcvtn v0.4h, v0.4s
+; CHECK-NEXT: ret
+  %res = fptrunc <2 x float> %op1 to <2 x half>
+  ret <2 x half> %res
+}
+
+; Don't use SVE for 128-bit vectors.
+define <4 x half> @fcvt_v4f32_v4f16(<4 x float> %op1) #0 {
+; CHECK-LABEL: fcvt_v4f32_v4f16:
+; CHECK: fcvtn v0.4h, v0.4s
+; CHECK-NEXT: ret
+  %res = fptrunc <4 x float> %op1 to <4 x half>
+  ret <4 x half> %res
+}
+
+define <8 x half> @fcvt_v8f32_v8f16(<8 x float>* %a) #0 {
+; CHECK-LABEL: fcvt_v8f32_v8f16:
+; CHECK: ptrue [[PG1:p[0-9]+]].s, vl8
+; CHECK-NEXT: ld1w { [[OP:z[0-9]+]].s }, [[PG1]]/z, [x0]
+; CHECK-NEXT: ptrue [[PG2:p[0-9]+]].s
+; CHECK-NEXT: fcvt [[CVT:z[0-9]+]].h, [[PG2]]/m, [[OP]].s
+; CHECK-NEXT: uzp1 z0.h, [[CVT]].h, [[CVT]].h
+; CHECK-NEXT: ret
+  %op1 = load <8 x float>, <8 x float>* %a
+  %res = fptrunc <8 x float> %op1 to <8 x half>
+  ret <8 x half> %res
+}
+
+define void @fcvt_v16f32_v16f16(<16 x float>* %a, <16 x half>* %b) #0 {
+; CHECK-LABEL: fcvt_v16f32_v16f16:
+; VBITS_GE_512: ptrue [[PG1:p[0-9]+]].s, vl16
+; VBITS_GE_512-NEXT: ld1w { [[OP:z[0-9]+]].s }, [[PG1]]/z, [x0]
+; VBITS_GE_512-NEXT: ptrue [[PG2:p[0-9]+]].s
+; VBITS_GE_512-NEXT: fcvt [[CVT:z[0-9]+]].h, [[PG2]]/m, [[OP]].s
+; VBITS_GE_512-NEXT: uzp1 [[RES:z[0-9]+]].h, [[CVT]].h, [[CVT]].h
+; VBITS_GE_512-NEXT: ptrue [[PG3:p[0-9]+]].h, vl16
+; VBITS_GE_512-NEXT: st1h { [[RES]].h }, [[PG3]], [x1]
+; VBITS_GE_512-NEXT: ret
+  %op1 = load <16 x float>, <16 x float>* %a
+  %res = fptrunc <16 x float> %op1 to <16 x half>
+  store <16 x half> %res, <16 x half>* %b
+  ret void
+}
+
+define void @fcvt_v32f32_v32f16(<32 x float>* %a, <32 x half>* %b) #0 {
+; CHECK-LABEL: fcvt_v32f32_v32f16:
+; VBITS_GE_1024: ptrue [[PG1:p[0-9]+]].s, vl32
+; VBITS_GE_1024-NEXT: ld1w { [[OP:z[0-9]+]].s }, [[PG1]]/z, [x0]
+; VBITS_GE_1024-NEXT: ptrue [[PG2:p[0-9]+]].s
+; VBITS_GE_1024-NEXT: fcvt [[CVT:z[0-9]+]].h, [[PG2]]/m, [[OP]].s
+; VBITS_GE_1024-NEXT: uzp1 [[RES:z[0-9]+]].h, [[CVT]].h, [[CVT]].h
+; VBITS_GE_1024-NEXT: ptrue [[PG3:p[0-9]+]].h, vl32
+; VBITS_GE_1024-NEXT: st1h { [[RES]].h }, [[PG3]], [x1]
+; VBITS_GE_1024-NEXT: ret
+  %op1 = load <32 x float>, <32 x float>* %a
+  %res = fptrunc <32 x float> %op1 to <32 x half>
+  store <32 x half> %res, <32 x half>* %b
+  ret void
+}
+
+define void @fcvt_v64f32_v64f16(<64 x float>* %a, <64 x half>* %b) #0 {
+; CHECK-LABEL: fcvt_v64f32_v64f16:
+; VBITS_GE_2048: ptrue [[PG1:p[0-9]+]].s, vl64
+; VBITS_GE_2048-NEXT: ld1w { [[OP:z[0-9]+]].s }, [[PG1]]/z, [x0]
+; VBITS_GE_2048-NEXT: ptrue [[PG2:p[0-9]+]].s
+; VBITS_GE_2048-NEXT: fcvt [[RES:z[0-9]+]].h, [[PG2]]/m, [[UPK]].s
+; VBITS_GE_2048-NEXT: uzp1 [[RES:z[0-9]+]].h, [[CVT]].h, [[CVT]].h
+; VBITS_GE_2048-NEXT: ptrue [[PG3:p[0-9]+]].h, vl64
+; VBITS_GE_2048-NEXT: st1h { [[RES]].h }, [[PG3]], [x1]
+; VBITS_GE_2048-NEXT: ret
+  %op1 = load <64 x float>, <64 x float>* %a
+  %res = fptrunc <64 x float> %op1 to <64 x half>
+  store <64 x half> %res, <64 x half>* %b
+  ret void
+}
+
+;
+; FCVT D -> H
+;
+
+; Don't use SVE for 64-bit vectors.
+define <1 x half> @fcvt_v1f64_v1f16(<1 x double> %op1) #0 {
+; CHECK-LABEL: fcvt_v1f64_v1f16:
+; CHECK: fcvt h0, d0
+; CHECK-NEXT: ret
+  %res = fptrunc <1 x double> %op1 to <1 x half>
+  ret <1 x half> %res
+}
+
+; v2f16 is not legal for NEON, so use SVE
+define <2 x half> @fcvt_v2f64_v2f16(<2 x double> %op1) #0 {
+; CHECK-LABEL: fcvt_v2f64_v2f16:
+; CHECK: ptrue [[PG:p[0-9]+]].d
+; CHECK-NEXT: fcvt [[CVT:z[0-9]+]].h, [[PG]]/m, z0.d
+; CHECK-NEXT: uzp1 [[UZP:z[0-9]+]].s, [[CVT]].s, [[CVT]].s
+; CHECK-NEXT: uzp1 z0.h, [[UZP]].h, [[UZP]].h
+; CHECK-NEXT: ret
+  %res = fptrunc <2 x double> %op1 to <2 x half>
+  ret <2 x half> %res
+}
+
+define <4 x half> @fcvt_v4f64_v4f16(<4 x double>* %a) #0 {
+; CHECK-LABEL: fcvt_v4f64_v4f16:
+; CHECK: ptrue [[PG1:p[0-9]+]].d, vl4
+; CHECK-NEXT: ld1d { [[OP:z[0-9]+]].d }, [[PG1]]/z, [x0]
+; CHECK-NEXT: ptrue [[PG2:p[0-9]+]].d
+; CHECK-NEXT: fcvt [[CVT:z[0-9]+]].h, [[PG2]]/m, [[OP]].d
+; CHECK-NEXT: uzp1 [[UZP:z[0-9]+]].s, [[CVT]].s, [[CVT]].s
+; CHECK-NEXT: uzp1 z0.h, [[UZP]].h, [[UZP]].h
+; CHECK-NEXT: ret
+  %op1 = load <4 x double>, <4 x double>* %a
+  %res = fptrunc <4 x double> %op1 to <4 x half>
+  ret <4 x half> %res
+}
+
+define <8 x half> @fcvt_v8f64_v8f16(<8 x double>* %a) #0 {
+; CHECK-LABEL: fcvt_v8f64_v8f16:
+; VBITS_GE_512: ptrue [[PG1:p[0-9]+]].d, vl8
+; VBITS_GE_512-NEXT: ld1d { [[OP:z[0-9]+]].d }, [[PG1]]/z, [x0]
+; VBITS_GE_512-NEXT: ptrue [[PG2:p[0-9]+]].d
+; VBITS_GE_512-NEXT: fcvt [[CVT:z[0-9]+]].h, [[PG2]]/m, [[OP]].d
+; VBITS_GE_512-NEXT: uzp1 [[UZP:z[0-9]+]].s, [[CVT]].s, [[CVT]].s
+; VBITS_GE_512-NEXT: uzp1 z0.h, [[UZP]].h, [[UZP]].h
+; VBITS_GE_512-NEXT: ret
+  %op1 = load <8 x double>, <8 x double>* %a
+  %res = fptrunc <8 x double> %op1 to <8 x half>
+  ret <8 x half> %res
+}
+
+define void @fcvt_v16f64_v16f16(<16 x double>* %a, <16 x half>* %b) #0 {
+; CHECK-LABEL: fcvt_v16f64_v16f16:
+; VBITS_GE_1024: ptrue [[PG1:p[0-9]+]].d, vl16
+; VBITS_GE_1024-NEXT: ld1d { [[OP:z[0-9]+]].d }, [[PG1]]/z, [x0]
+; VBITS_GE_1024-NEXT: ptrue [[PG2:p[0-9]+]].d
+; VBITS_GE_1024-NEXT: fcvt [[CVT:z[0-9]+]].h, [[PG2]]/m, [[OP]].d
+; VBITS_GE_1024-NEXT: uzp1 [[UZP:z[0-9]+]].s, [[CVT]].s, [[CVT]].s
+; VBITS_GE_1024-NEXT: uzp1 [[RES:z[0-9]+]].h, [[UZP]].h, [[UZP]].h
+; VBITS_GE_1024-NEXT: ptrue [[PG3:p[0-9]+]].h, vl16
+; VBITS_GE_1024-NEXT: st1h { [[RES]].h }, [[PG3]], [x1]
+; VBITS_GE_1024-NEXT: ret
+  %op1 = load <16 x double>, <16 x double>* %a
+  %res = fptrunc <16 x double> %op1 to <16 x half>
+  store <16 x half> %res, <16 x half>* %b
+  ret void
+}
+
+define void @fcvt_v32f64_v32f16(<32 x double>* %a, <32 x half>* %b) #0 {
+; CHECK-LABEL: fcvt_v32f64_v32f16:
+; VBITS_GE_2048: ptrue [[PG1:p[0-9]+]].d, vl32
+; VBITS_GE_2048-NEXT: ld1d { [[OP:z[0-9]+]].d }, [[PG1]]/z, [x0]
+; VBITS_GE_2048-NEXT: ptrue [[PG2:p[0-9]+]].d
+; VBITS_GE_2048-NEXT: fcvt [[CVT:z[0-9]+]].h, [[PG2]]/m, [[OP]].d
+; VBITS_GE_2048-NEXT: uzp1 [[UZP:z[0-9]+]].s, [[CVT]].s, [[CVT]].s
+; VBITS_GE_2048-NEXT: uzp1 [[RES:z[0-9]+]].h, [[UZP]].h, [[UZP]].h
+; VBITS_GE_2048-NEXT: ptrue [[PG3:p[0-9]+]].h, vl32
+; VBITS_GE_2048-NEXT: st1h { [[RES]].h }, [[PG3]], [x1]
+; VBITS_GE_2048-NEXT: ret
+  %op1 = load <32 x double>, <32 x double>* %a
+  %res = fptrunc <32 x double> %op1 to <32 x half>
+  store <32 x half> %res, <32 x half>* %b
+  ret void
+}
+
+;
+; FCVT D -> S
+;
+
+; Don't use SVE for 64-bit vectors.
+define <1 x float> @fcvt_v1f64_v1f32(<1 x double> %op1) #0 {
+; CHECK-LABEL: fcvt_v1f64_v1f32:
+; CHECK: fcvtn v0.2s, v0.2d
+; CHECK-NEXT: ret
+  %res = fptrunc <1 x double> %op1 to <1 x float>
+  ret <1 x float> %res
+}
+
+; Don't use SVE for 128-bit vectors.
+define <2 x float> @fcvt_v2f64_v2f32(<2 x double> %op1) #0 {
+; CHECK-LABEL: fcvt_v2f64_v2f32:
+; CHECK: fcvtn v0.2s, v0.2d
+; CHECK-NEXT: ret
+  %res = fptrunc <2 x double> %op1 to <2 x float>
+  ret <2 x float> %res
+}
+
+define <4 x float> @fcvt_v4f64_v4f32(<4 x double>* %a) #0 {
+; CHECK-LABEL: fcvt_v4f64_v4f32:
+; CHECK: ptrue [[PG1:p[0-9]+]].d, vl4
+; CHECK-NEXT: ld1d { [[OP:z[0-9]+]].d }, [[PG1]]/z, [x0]
+; CHECK-NEXT: ptrue [[PG2:p[0-9]+]].d
+; CHECK-NEXT: fcvt [[CVT:z[0-9]+]].s, [[PG2]]/m, [[OP]].d
+; CHECK-NEXT: uzp1 z0.s, [[CVT]].s, [[CVT]].s
+; CHECK-NEXT: ret
+  %op1 = load <4 x double>, <4 x double>* %a
+  %res = fptrunc <4 x double> %op1 to <4 x float>
+  ret <4 x float> %res
+}
+
+define void @fcvt_v8f64_v8f32(<8 x double>* %a, <8 x float>* %b) #0 {
+; CHECK-LABEL: fcvt_v8f64_v8f32:
+; VBITS_GE_512: ptrue [[PG1:p[0-9]+]].d, vl8
+; VBITS_GE_512-NEXT: ld1d { [[OP:z[0-9]+]].d }, [[PG1]]/z, [x0]
+; VBITS_GE_512-NEXT: ptrue [[PG2:p[0-9]+]].d
+; VBITS_GE_512-NEXT: fcvt [[CVT:z[0-9]+]].s, [[PG2]]/m, [[OP]].d
+; VBITS_GE_512-NEXT: uzp1 [[RES:z[0-9]+]].s, [[CVT]].s, [[CVT]].s
+; VBITS_GE_512-NEXT: ptrue [[PG3:p[0-9]+]].s, vl8
+; VBITS_GE_512-NEXT: st1w { [[RES]].s }, [[PG3]], [x1]
+; VBITS_GE_512-NEXT: ret
+  %op1 = load <8 x double>, <8 x double>* %a
+  %res = fptrunc <8 x double> %op1 to <8 x float>
+  store <8 x float> %res, <8 x float>* %b
+  ret void
+}
+
+define void @fcvt_v16f64_v16f32(<16 x double>* %a, <16 x float>* %b) #0 {
+; CHECK-LABEL: fcvt_v16f64_v16f32:
+; VBITS_GE_1024: ptrue [[PG1:p[0-9]+]].d, vl16
+; VBITS_GE_1024-NEXT: ld1d { [[OP:z[0-9]+]].d }, [[PG1]]/z, [x0]
+; VBITS_GE_1024-NEXT: ptrue [[PG2:p[0-9]+]].d
+; VBITS_GE_1024-NEXT: fcvt [[CVT:z[0-9]+]].s, [[PG2]]/m, [[OP]].d
+; VBITS_GE_1024-NEXT: uzp1 [[RES:z[0-9]+]].s, [[CVT]].s, [[CVT]].s
+; VBITS_GE_1024-NEXT: ptrue [[PG3:p[0-9]+]].s, vl16
+; VBITS_GE_1024-NEXT: st1w { [[RES]].s }, [[PG3]], [x1]
+; VBITS_GE_1024-NEXT: ret
+  %op1 = load <16 x double>, <16 x double>* %a
+  %res = fptrunc <16 x double> %op1 to <16 x float>
+  store <16 x float> %res, <16 x float>* %b
+  ret void
+}
+
+define void @fcvt_v32f64_v32f32(<32 x double>* %a, <32 x float>* %b) #0 {
+; CHECK-LABEL: fcvt_v32f64_v32f32:
+; VBITS_GE_2048: ptrue [[PG1:p[0-9]+]].d, vl32
+; VBITS_GE_2048-NEXT: ld1d { [[OP:z[0-9]+]].d }, [[PG1]]/z, [x0]
+; VBITS_GE_2048-NEXT: ptrue [[PG2:p[0-9]+]].d
+; VBITS_GE_2048-NEXT: fcvt [[CVT:z[0-9]+]].s, [[PG2]]/m, [[OP]].d
+; VBITS_GE_2048-NEXT: uzp1 [[RES:z[0-9]+]].s, [[CVT]].s, [[CVT]].s
+; VBITS_GE_2048-NEXT: ptrue [[PG3:p[0-9]+]].s, vl32
+; VBITS_GE_2048-NEXT: st1w { [[RES]].s }, [[PG3]], [x1]
+; VBITS_GE_2048-NEXT: ret
+  %op1 = load <32 x double>, <32 x double>* %a
+  %res = fptrunc <32 x double> %op1 to <32 x float>
+  store <32 x float> %res, <32 x float>* %b
+  ret void
+}
+
+attributes #0 = { "target-features"="+sve" }