[AArch64][SVE] Mark fixed-type FP extending/truncating loads/stores as custom

[bsmith]

Summary

This allows the generic DAG combine to fold fp_extend/fp_trunc into
loads/stores which we can then lower into a integer extending
load/truncating store plus an FP_EXTEND/FP_ROUND.

The nuance here is that fixed-type FP_EXTEND/FP_ROUND require unpacked
types hence lowering them introduces an unpack/zip. By allowing these
nodes to be combined with loads/store we make it much easier to have
this unpack/zip combined into the load/store by our custom lowering.

[paulwalker-arm]

llvm/lib/Target/AArch64/AArch64ISelLowering.cpp

1545

This looks weird to me, shouldn't InnerVT be floating point? I guess the reason this works is because the i8 case is essentially bogus and you end up with the necessary floating point types from the second iteration. Either way I think this wants to be MVT::f16.

18600

Purely to help with formatting I believe you can just do VT.isFloatingPoint() here.

18601

Can you be specific here as in use == ISD::EXTLOAD as we should need to support any other extension types.

18611

I think it'll be safer to mirror the other if condition as the two are linked (i.e. we need to do both or none based on the same requirement).

[bsmith]

llvm/lib/Target/AArch64/AArch64ISelLowering.cpp

1545

I meant to change this but clearly forgot!

[bsmith]

• Fix incorrect type used in addTypeForFixedLengthSVE
• Cleanup conditions in LowerFixedLengthVectorLoadToSVE

[paulwalker-arm]

llvm/lib/Target/AArch64/AArch64ISelLowering.cpp

18593–18594

You could move this into the if block where it's used.

18690–18691

Perhaps TruncVT? Plus as above this could be moved into the if block where it is used.

18697

As above this can be just VT.

llvm/lib/CodeGen/SelectionDAG/DAGCombiner.cpp

if (N0.getOpcode() == ISD::FP_ROUND
     if (VT.bitsLT(In.getValueType()))
       return DAG.getNode(ISD::FP_ROUND, SDLoc(N), VT,
                          In, N0.getOperand(1));
     return DAG.getNode(ISD::FP_EXTEND, SDLoc(N), VT, In);
   }
 
   // fold (fpext (load x)) -> (fpext (fptrunc (extload x)))
   if (ISD::isNormalLoad(N0.getNode()) && N0.hasOneUse() &&
-       TLI.isLoadExtLegal(ISD::EXTLOAD, VT, N0.getValueType())) {
+      TLI.isLoadExtLegalOrCustom(ISD::EXTLOAD, VT, N0.getValueType())) {
     LoadSDNode *LN0 = cast<LoadSDNode>(N0);
     SDValue ExtLoad = DAG.getExtLoad(ISD::EXTLOAD, SDLoc(N), VT,
                                      LN0->getChain(),
                                      LN0->getBasePtr(), N0.getValueType(),
                                      LN0->getMemOperand());
     CombineTo(N, ExtLoad);
     CombineTo(N0.getNode(),
               DAG.getNode(ISD::FP_ROUND, SDLoc(N0),

llvm/lib/Target/AArch64/AArch64ISelLowering.cpp

if (VT.isFloatingPoint()) {
     setCondCodeAction(ISD::SETULT, VT, Expand);
     setCondCodeAction(ISD::SETULE, VT, Expand);
     setCondCodeAction(ISD::SETUGE, VT, Expand);
     setCondCodeAction(ISD::SETUGT, VT, Expand);
     setCondCodeAction(ISD::SETUEQ, VT, Expand);
     setCondCodeAction(ISD::SETUNE, VT, Expand);
   }
 
-  // Mark integer truncating stores as having custom lowering
+  // Mark integer truncating stores/extending loads as having custom lowering
   if (VT.isInteger()) {
     MVT InnerVT = VT.changeVectorElementType(MVT::i8);
     while (InnerVT != VT) {
       setTruncStoreAction(VT, InnerVT, Custom);
       setLoadExtAction(ISD::ZEXTLOAD, VT, InnerVT, Custom);
       setLoadExtAction(ISD::SEXTLOAD, VT, InnerVT, Custom);
       InnerVT = InnerVT.changeVectorElementType(
           MVT::getIntegerVT(2 * InnerVT.getScalarSizeInBits()));
     }
   }
 
+  // Mark floating-point truncating stores/extending loads as having custom
+  // lowering
+  if (VT.isFloatingPoint()) {
+    MVT InnerVT = VT.changeVectorElementType(MVT::i8);
+    while (InnerVT != VT) {
+      setTruncStoreAction(VT, InnerVT, Custom);
+      setLoadExtAction(ISD::EXTLOAD, VT, InnerVT, Custom);
+      InnerVT = InnerVT.changeVectorElementType(
+          MVT::getFloatingPointVT(2 * InnerVT.getScalarSizeInBits()));
+    }
+  }
+
   // Lower fixed length vector operations to scalable equivalents.
   setOperationAction(ISD::ABS, VT, Custom);
   setOperationAction(ISD::ADD, VT, Custom);
   setOperationAction(ISD::AND, VT, Custom);
   setOperationAction(ISD::ANY_EXTEND, VT, Custom);
   setOperationAction(ISD::BITCAST, VT, Custom);
   setOperationAction(ISD::BITREVERSE, VT, Custom);
   setOperationAction(ISD::BSWAP, VT, Custom);

 // Convert all fixed length vector loads larger than NEON to masked_loads.
 SDValue AArch64TargetLowering::LowerFixedLengthVectorLoadToSVE(
     SDValue Op, SelectionDAG &DAG) const {
   auto Load = cast<LoadSDNode>(Op);
 
   SDLoc DL(Op);
   EVT VT = Op.getValueType();
   EVT ContainerVT = getContainerForFixedLengthVector(DAG, VT);
+  EVT ExtendVT = ContainerVT.changeVectorElementType(
+      Load->getMemoryVT().getVectorElementType());
+  EVT LoadVT = ContainerVT;
+  EVT MemVT = Load->getMemoryVT();
+
+  auto Pg = getPredicateForFixedLengthVector(DAG, DL, VT);
+
+  if (ContainerVT.getVectorElementType().isFloatingPoint() &&
+      Load->getExtensionType() != ISD::NON_EXTLOAD) {
+    LoadVT = ContainerVT.changeTypeToInteger();
+    MemVT = MemVT.changeTypeToInteger();
+  }
 
   auto NewLoad = DAG.getMaskedLoad(
-      ContainerVT, DL, Load->getChain(), Load->getBasePtr(), Load->getOffset(),
-      getPredicateForFixedLengthVector(DAG, DL, VT), DAG.getUNDEF(ContainerVT),
-      Load->getMemoryVT(), Load->getMemOperand(), Load->getAddressingMode(),
-      Load->getExtensionType());
+      LoadVT, DL, Load->getChain(), Load->getBasePtr(), Load->getOffset(), Pg,
+      DAG.getUNDEF(LoadVT), MemVT, Load->getMemOperand(),
+      Load->getAddressingMode(), Load->getExtensionType());
+
+  if (LoadVT != ContainerVT) {
+    NewLoad = getSVESafeBitCast(ExtendVT, NewLoad, DAG);
+    NewLoad = DAG.getNode(AArch64ISD::FP_EXTEND_MERGE_PASSTHRU, DL, ContainerVT,
+                          Pg, NewLoad, DAG.getUNDEF(ContainerVT));
+  }
 
   auto Result = convertFromScalableVector(DAG, VT, NewLoad);
   SDValue MergedValues[2] = {Result, Load->getChain()};
   return DAG.getMergeValues(MergedValues, DL);
 }
 
 static SDValue convertFixedMaskToScalableVector(SDValue Mask,
                                                 SelectionDAG &DAG) {

 // Convert all fixed length vector stores larger than NEON to masked_stores.
 SDValue AArch64TargetLowering::LowerFixedLengthVectorStoreToSVE(
     SDValue Op, SelectionDAG &DAG) const {
   auto Store = cast<StoreSDNode>(Op);
 
   SDLoc DL(Op);
   EVT VT = Store->getValue().getValueType();
   EVT ContainerVT = getContainerForFixedLengthVector(DAG, VT);
+  EVT ExtendVT = ContainerVT.changeVectorElementType(
+      Store->getMemoryVT().getVectorElementType());
+  EVT MemVT = Store->getMemoryVT();
 
+  auto Pg = getPredicateForFixedLengthVector(DAG, DL, VT);
   auto NewValue = convertToScalableVector(DAG, ContainerVT, Store->getValue());
-  return DAG.getMaskedStore(
-      Store->getChain(), DL, NewValue, Store->getBasePtr(), Store->getOffset(),
-      getPredicateForFixedLengthVector(DAG, DL, VT), Store->getMemoryVT(),
+
+  if (ContainerVT.getVectorElementType().isFloatingPoint() &&
+      Store->isTruncatingStore()) {
+    MemVT = MemVT.changeTypeToInteger();
+    NewValue = DAG.getNode(AArch64ISD::FP_ROUND_MERGE_PASSTHRU, DL, ExtendVT,
+                           Pg, NewValue, DAG.getTargetConstant(0, DL, MVT::i64),
+                           DAG.getUNDEF(ExtendVT));
+    NewValue =
+        getSVESafeBitCast(ContainerVT.changeTypeToInteger(), NewValue, DAG);
+  }
+
+  return DAG.getMaskedStore(Store->getChain(), DL, NewValue,
+                            Store->getBasePtr(), Store->getOffset(), Pg, MemVT,
                             Store->getMemOperand(), Store->getAddressingMode(),
                             Store->isTruncatingStore());
 }
 
 SDValue AArch64TargetLowering::LowerFixedLengthVectorMStoreToSVE(
     SDValue Op, SelectionDAG &DAG) const {
   auto Store = cast<MaskedStoreSDNode>(Op);
 
   if (Store->isTruncatingStore())
     return SDValue();

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

 ; 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:       // %bb.0:
-; CHECK-NEXT:    ldr q0, [x0]
 ; CHECK-NEXT:    ptrue p0.s, vl8
-; CHECK-NEXT:    uunpklo z0.s, z0.h
+; CHECK-NEXT:    ld1sh { z0.s }, p0/z, [x0]
 ; CHECK-NEXT:    fcvt z0.s, p0/m, z0.h
 ; CHECK-NEXT:    st1w { z0.s }, p0, [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
 }

 ; VBITS_EQ_256-NEXT:    uunpklo z1.s, z0.h
 ; VBITS_EQ_256-NEXT:    ext z0.b, z0.b, z0.b, #16
 ; VBITS_EQ_256-NEXT:    uunpklo z0.s, z0.h
 ; VBITS_EQ_256-NEXT:    fcvt z1.s, p0/m, z1.h
 ; VBITS_EQ_256-NEXT:    fcvt z0.s, p0/m, z0.h
 ; VBITS_EQ_256-NEXT:    st1w { z1.s }, p0, [x1]
 ; VBITS_EQ_256-NEXT:    st1w { z0.s }, p0, [x1, x8, lsl #2]
 ; VBITS_EQ_256-NEXT:    ret
-
+;
 ; VBITS_GE_512-LABEL: fcvt_v16f16_v16f32:
 ; VBITS_GE_512:       // %bb.0:
-; VBITS_GE_512-NEXT:    ptrue p0.h, vl16
-; VBITS_GE_512-NEXT:    ld1h { z0.h }, p0/z, [x0]
 ; VBITS_GE_512-NEXT:    ptrue p0.s, vl16
-; VBITS_GE_512-NEXT:    uunpklo z0.s, z0.h
+; VBITS_GE_512-NEXT:    ld1sh { z0.s }, p0/z, [x0]
 ; VBITS_GE_512-NEXT:    fcvt z0.s, p0/m, z0.h
 ; VBITS_GE_512-NEXT:    st1w { z0.s }, p0, [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 {
 ; VBITS_GE_1024-LABEL: fcvt_v32f16_v32f32:
 ; VBITS_GE_1024:       // %bb.0:
-; VBITS_GE_1024-NEXT:    ptrue p0.h, vl32
-; VBITS_GE_1024-NEXT:    ld1h { z0.h }, p0/z, [x0]
 ; VBITS_GE_1024-NEXT:    ptrue p0.s, vl32
-; VBITS_GE_1024-NEXT:    uunpklo z0.s, z0.h
+; VBITS_GE_1024-NEXT:    ld1sh { z0.s }, p0/z, [x0]
 ; VBITS_GE_1024-NEXT:    fcvt z0.s, p0/m, z0.h
 ; VBITS_GE_1024-NEXT:    st1w { z0.s }, p0, [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 {
 ; VBITS_GE_2048-LABEL: fcvt_v64f16_v64f32:
 ; VBITS_GE_2048:       // %bb.0:
-; VBITS_GE_2048-NEXT:    ptrue p0.h, vl64
-; VBITS_GE_2048-NEXT:    ld1h { z0.h }, p0/z, [x0]
 ; VBITS_GE_2048-NEXT:    ptrue p0.s, vl64
-; VBITS_GE_2048-NEXT:    uunpklo z0.s, z0.h
+; VBITS_GE_2048-NEXT:    ld1sh { z0.s }, p0/z, [x0]
 ; VBITS_GE_2048-NEXT:    fcvt z0.s, p0/m, z0.h
 ; VBITS_GE_2048-NEXT:    st1w { z0.s }, p0, [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
 }

 ; 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:       // %bb.0:
-; CHECK-NEXT:    ldr d0, [x0]
 ; CHECK-NEXT:    ptrue p0.d, vl4
-; CHECK-NEXT:    uunpklo z0.s, z0.h
-; CHECK-NEXT:    uunpklo z0.d, z0.s
+; CHECK-NEXT:    ld1sh { z0.d }, p0/z, [x0]
 ; CHECK-NEXT:    fcvt z0.d, p0/m, z0.h
 ; CHECK-NEXT:    st1d { z0.d }, p0, [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 {
-; Ensure sensible type legalisation.
 ; VBITS_EQ_256-LABEL: fcvt_v8f16_v8f64:
 ; VBITS_EQ_256:       // %bb.0:
 ; VBITS_EQ_256-NEXT:    ldr q0, [x0]
 ; VBITS_EQ_256-NEXT:    mov x8, #4
 ; VBITS_EQ_256-NEXT:    ptrue p0.d, vl4
 ; VBITS_EQ_256-NEXT:    ext v1.16b, v0.16b, v0.16b, #8
 ; VBITS_EQ_256-NEXT:    uunpklo z0.s, z0.h
 ; VBITS_EQ_256-NEXT:    uunpklo z0.d, z0.s
 ; VBITS_EQ_256-NEXT:    fcvt z0.d, p0/m, z0.h
 ; VBITS_EQ_256-NEXT:    st1d { z0.d }, p0, [x1]
 ; VBITS_EQ_256-NEXT:    uunpklo z1.s, z1.h
 ; VBITS_EQ_256-NEXT:    uunpklo z1.d, z1.s
 ; VBITS_EQ_256-NEXT:    fcvt z1.d, p0/m, z1.h
 ; VBITS_EQ_256-NEXT:    st1d { z1.d }, p0, [x1, x8, lsl #3]
 ; VBITS_EQ_256-NEXT:    ret
-
+;
 ; VBITS_GE_512-LABEL: fcvt_v8f16_v8f64:
 ; VBITS_GE_512:       // %bb.0:
-; VBITS_GE_512-NEXT:    ldr q0, [x0]
 ; VBITS_GE_512-NEXT:    ptrue p0.d, vl8
-; VBITS_GE_512-NEXT:    uunpklo z0.s, z0.h
-; VBITS_GE_512-NEXT:    uunpklo z0.d, z0.s
+; VBITS_GE_512-NEXT:    ld1sh { z0.d }, p0/z, [x0]
 ; VBITS_GE_512-NEXT:    fcvt z0.d, p0/m, z0.h
 ; VBITS_GE_512-NEXT:    st1d { z0.d }, p0, [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 {
 ; VBITS_GE_1024-LABEL: fcvt_v16f16_v16f64:
 ; VBITS_GE_1024:       // %bb.0:
-; VBITS_GE_1024-NEXT:    ptrue p0.h, vl16
-; VBITS_GE_1024-NEXT:    ld1h { z0.h }, p0/z, [x0]
 ; VBITS_GE_1024-NEXT:    ptrue p0.d, vl16
-; VBITS_GE_1024-NEXT:    uunpklo z0.s, z0.h
-; VBITS_GE_1024-NEXT:    uunpklo z0.d, z0.s
+; VBITS_GE_1024-NEXT:    ld1sh { z0.d }, p0/z, [x0]
 ; VBITS_GE_1024-NEXT:    fcvt z0.d, p0/m, z0.h
 ; VBITS_GE_1024-NEXT:    st1d { z0.d }, p0, [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 {
 ; VBITS_GE_2048-LABEL: fcvt_v32f16_v32f64:
 ; VBITS_GE_2048:       // %bb.0:
-; VBITS_GE_2048-NEXT:    ptrue p0.h, vl32
-; VBITS_GE_2048-NEXT:    ld1h { z0.h }, p0/z, [x0]
 ; VBITS_GE_2048-NEXT:    ptrue p0.d, vl32
-; VBITS_GE_2048-NEXT:    uunpklo z0.s, z0.h
-; VBITS_GE_2048-NEXT:    uunpklo z0.d, z0.s
+; VBITS_GE_2048-NEXT:    ld1sh { z0.d }, p0/z, [x0]
 ; VBITS_GE_2048-NEXT:    fcvt z0.d, p0/m, z0.h
 ; VBITS_GE_2048-NEXT:    st1d { z0.d }, p0, [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
 }

 ; 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:       // %bb.0:
-; CHECK-NEXT:    ldr q0, [x0]
 ; CHECK-NEXT:    ptrue p0.d, vl4
-; CHECK-NEXT:    uunpklo z0.d, z0.s
+; CHECK-NEXT:    ld1sw { z0.d }, p0/z, [x0]
 ; CHECK-NEXT:    fcvt z0.d, p0/m, z0.s
 ; CHECK-NEXT:    st1d { z0.d }, p0, [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
 }

 ; VBITS_EQ_256-NEXT:    uunpklo z1.d, z0.s
 ; VBITS_EQ_256-NEXT:    ext z0.b, z0.b, z0.b, #16
 ; VBITS_EQ_256-NEXT:    uunpklo z0.d, z0.s
 ; VBITS_EQ_256-NEXT:    fcvt z1.d, p0/m, z1.s
 ; VBITS_EQ_256-NEXT:    fcvt z0.d, p0/m, z0.s
 ; VBITS_EQ_256-NEXT:    st1d { z1.d }, p0, [x1]
 ; VBITS_EQ_256-NEXT:    st1d { z0.d }, p0, [x1, x8, lsl #3]
 ; VBITS_EQ_256-NEXT:    ret
-
+;
 ; VBITS_GE_512-LABEL: fcvt_v8f32_v8f64:
 ; VBITS_GE_512:       // %bb.0:
-; VBITS_GE_512-NEXT:    ptrue p0.s, vl8
-; VBITS_GE_512-NEXT:    ld1w { z0.s }, p0/z, [x0]
 ; VBITS_GE_512-NEXT:    ptrue p0.d, vl8
-; VBITS_GE_512-NEXT:    uunpklo z0.d, z0.s
+; VBITS_GE_512-NEXT:    ld1sw { z0.d }, p0/z, [x0]
 ; VBITS_GE_512-NEXT:    fcvt z0.d, p0/m, z0.s
 ; VBITS_GE_512-NEXT:    st1d { z0.d }, p0, [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 {
 ; VBITS_GE_1024-LABEL: fcvt_v16f32_v16f64:
 ; VBITS_GE_1024:       // %bb.0:
-; VBITS_GE_1024-NEXT:    ptrue p0.s, vl16
-; VBITS_GE_1024-NEXT:    ld1w { z0.s }, p0/z, [x0]
 ; VBITS_GE_1024-NEXT:    ptrue p0.d, vl16
-; VBITS_GE_1024-NEXT:    uunpklo z0.d, z0.s
+; VBITS_GE_1024-NEXT:    ld1sw { z0.d }, p0/z, [x0]
 ; VBITS_GE_1024-NEXT:    fcvt z0.d, p0/m, z0.s
 ; VBITS_GE_1024-NEXT:    st1d { z0.d }, p0, [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 {
 ; VBITS_GE_2048-LABEL: fcvt_v32f32_v32f64:
 ; VBITS_GE_2048:       // %bb.0:
-; VBITS_GE_2048-NEXT:    ptrue p0.s, vl32
-; VBITS_GE_2048-NEXT:    ld1w { z0.s }, p0/z, [x0]
 ; VBITS_GE_2048-NEXT:    ptrue p0.d, vl32
-; VBITS_GE_2048-NEXT:    uunpklo z0.d, z0.s
+; VBITS_GE_2048-NEXT:    ld1sw { z0.d }, p0/z, [x0]
 ; VBITS_GE_2048-NEXT:    fcvt z0.d, p0/m, z0.s
 ; VBITS_GE_2048-NEXT:    st1d { z0.d }, p0, [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
 }

 ; VBITS_EQ_256-NEXT:    ptrue p0.h, vl16
 ; VBITS_EQ_256-NEXT:    st1h { z1.h }, p0, [x1]
 ; VBITS_EQ_256-NEXT:    ret
 ;
 ; VBITS_GE_512-LABEL: fcvt_v16f32_v16f16:
 ; VBITS_GE_512:       // %bb.0:
 ; VBITS_GE_512-NEXT:    ptrue p0.s, vl16
 ; VBITS_GE_512-NEXT:    ld1w { z0.s }, p0/z, [x0]
-; VBITS_GE_512-NEXT:    ptrue p0.s
 ; VBITS_GE_512-NEXT:    fcvt z0.h, p0/m, z0.s
-; VBITS_GE_512-NEXT:    ptrue p0.h, vl16
-; VBITS_GE_512-NEXT:    uzp1 z0.h, z0.h, z0.h
-; VBITS_GE_512-NEXT:    st1h { z0.h }, p0, [x1]
+; VBITS_GE_512-NEXT:    st1h { z0.s }, p0, [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 {
 ; VBITS_GE_1024-LABEL: fcvt_v32f32_v32f16:
 ; VBITS_GE_1024:       // %bb.0:
 ; VBITS_GE_1024-NEXT:    ptrue p0.s, vl32
 ; VBITS_GE_1024-NEXT:    ld1w { z0.s }, p0/z, [x0]
-; VBITS_GE_1024-NEXT:    ptrue p0.s
 ; VBITS_GE_1024-NEXT:    fcvt z0.h, p0/m, z0.s
-; VBITS_GE_1024-NEXT:    ptrue p0.h, vl32
-; VBITS_GE_1024-NEXT:    uzp1 z0.h, z0.h, z0.h
-; VBITS_GE_1024-NEXT:    st1h { z0.h }, p0, [x1]
+; VBITS_GE_1024-NEXT:    st1h { z0.s }, p0, [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 {
 ; VBITS_GE_2048-LABEL: fcvt_v64f32_v64f16:
 ; VBITS_GE_2048:       // %bb.0:
 ; VBITS_GE_2048-NEXT:    ptrue p0.s, vl64
 ; VBITS_GE_2048-NEXT:    ld1w { z0.s }, p0/z, [x0]
-; VBITS_GE_2048-NEXT:    ptrue p0.s
 ; VBITS_GE_2048-NEXT:    fcvt z0.h, p0/m, z0.s
-; VBITS_GE_2048-NEXT:    ptrue p0.h, vl64
-; VBITS_GE_2048-NEXT:    uzp1 z0.h, z0.h, z0.h
-; VBITS_GE_2048-NEXT:    st1h { z0.h }, p0, [x1]
+; VBITS_GE_2048-NEXT:    st1h { z0.s }, p0, [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
 }
 
 ;
; VBITS_GE_512-NEXT:    ret
   ret <8 x half> %res
 }
 
 define void @fcvt_v16f64_v16f16(<16 x double>* %a, <16 x half>* %b) #0 {
 ; VBITS_GE_1024-LABEL: fcvt_v16f64_v16f16:
 ; VBITS_GE_1024:       // %bb.0:
 ; VBITS_GE_1024-NEXT:    ptrue p0.d, vl16
 ; VBITS_GE_1024-NEXT:    ld1d { z0.d }, p0/z, [x0]
-; VBITS_GE_1024-NEXT:    ptrue p0.d
 ; VBITS_GE_1024-NEXT:    fcvt z0.h, p0/m, z0.d
-; VBITS_GE_1024-NEXT:    ptrue p0.h, vl16
-; VBITS_GE_1024-NEXT:    uzp1 z0.s, z0.s, z0.s
-; VBITS_GE_1024-NEXT:    uzp1 z0.h, z0.h, z0.h
-; VBITS_GE_1024-NEXT:    st1h { z0.h }, p0, [x1]
+; VBITS_GE_1024-NEXT:    st1h { z0.d }, p0, [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 {
 ; VBITS_GE_2048-LABEL: fcvt_v32f64_v32f16:
 ; VBITS_GE_2048:       // %bb.0:
 ; VBITS_GE_2048-NEXT:    ptrue p0.d, vl32
 ; VBITS_GE_2048-NEXT:    ld1d { z0.d }, p0/z, [x0]
-; VBITS_GE_2048-NEXT:    ptrue p0.d
 ; VBITS_GE_2048-NEXT:    fcvt z0.h, p0/m, z0.d
-; VBITS_GE_2048-NEXT:    ptrue p0.h, vl32
-; VBITS_GE_2048-NEXT:    uzp1 z0.s, z0.s, z0.s
-; VBITS_GE_2048-NEXT:    uzp1 z0.h, z0.h, z0.h
-; VBITS_GE_2048-NEXT:    st1h { z0.h }, p0, [x1]
+; VBITS_GE_2048-NEXT:    st1h { z0.d }, p0, [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
 }
 
 ;

 ; VBITS_EQ_256-NEXT:    ptrue p0.s, vl8
 ; VBITS_EQ_256-NEXT:    st1w { z1.s }, p0, [x1]
 ; VBITS_EQ_256-NEXT:    ret
 ;
 ; VBITS_GE_512-LABEL: fcvt_v8f64_v8f32:
 ; VBITS_GE_512:       // %bb.0:
 ; VBITS_GE_512-NEXT:    ptrue p0.d, vl8
 ; VBITS_GE_512-NEXT:    ld1d { z0.d }, p0/z, [x0]
-; VBITS_GE_512-NEXT:    ptrue p0.d
 ; VBITS_GE_512-NEXT:    fcvt z0.s, p0/m, z0.d
-; VBITS_GE_512-NEXT:    ptrue p0.s, vl8
-; VBITS_GE_512-NEXT:    uzp1 z0.s, z0.s, z0.s
-; VBITS_GE_512-NEXT:    st1w { z0.s }, p0, [x1]
+; VBITS_GE_512-NEXT:    st1w { z0.d }, p0, [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 {
 ; VBITS_GE_1024-LABEL: fcvt_v16f64_v16f32:
 ; VBITS_GE_1024:       // %bb.0:
 ; VBITS_GE_1024-NEXT:    ptrue p0.d, vl16
 ; VBITS_GE_1024-NEXT:    ld1d { z0.d }, p0/z, [x0]
-; VBITS_GE_1024-NEXT:    ptrue p0.d
 ; VBITS_GE_1024-NEXT:    fcvt z0.s, p0/m, z0.d
-; VBITS_GE_1024-NEXT:    ptrue p0.s, vl16
-; VBITS_GE_1024-NEXT:    uzp1 z0.s, z0.s, z0.s
-; VBITS_GE_1024-NEXT:    st1w { z0.s }, p0, [x1]
+; VBITS_GE_1024-NEXT:    st1w { z0.d }, p0, [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 {
 ; VBITS_GE_2048-LABEL: fcvt_v32f64_v32f32:
 ; VBITS_GE_2048:       // %bb.0:
 ; VBITS_GE_2048-NEXT:    ptrue p0.d, vl32
 ; VBITS_GE_2048-NEXT:    ld1d { z0.d }, p0/z, [x0]
-; VBITS_GE_2048-NEXT:    ptrue p0.d
 ; VBITS_GE_2048-NEXT:    fcvt z0.s, p0/m, z0.d
-; VBITS_GE_2048-NEXT:    ptrue p0.s, vl32
-; VBITS_GE_2048-NEXT:    uzp1 z0.s, z0.s, z0.s
-; VBITS_GE_2048-NEXT:    st1w { z0.s }, p0, [x1]
+; VBITS_GE_2048-NEXT:    st1w { z0.d }, p0, [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" }