[AArch64] Optimize instruction selection for certain vector shuffles

[miyuki]

Summary

This patch adds code to recognize vector shuffles which can be
represented as VDUP (splat) of a vector lane with of a different
(wider) type than the original vector lane type.

For example:

shufflevector <4 x i16> %v, <4 x i16> undef, <4 x i32> <i32 0, i32 1, i32 0, i32 1>

is essentially:

shufflevector <2 x i32> %v, <2 x i32> undef, <2 x i32> <i32 0, i32 0>

Such patterns are generated by the SelectionDAG machinery in some cases
(see DAGCombiner::visitBITCAST in DAGCombiner.cpp, the "Remove double
bitcasts from shuffles" part).

[dmgreen]

Looks nice. Perhaps add a few more negative tests for the various different "return false"'s

llvm/lib/Target/AArch64/AArch64ISelLowering.cpp

7392–7394

You can make this an else and drop the continue.

7407

Would this be nullptr or BlockElts.end()?

7425

BlockElts.size() -> NumEltsPerBlock, as it's used more consistently.

[miyuki]

1. Addressed review comments
2. Added a cast back to the original type

[dmgreen]

Thanks. One last question about this new bit of code.

llvm/lib/Target/AArch64/AArch64ISelLowering.cpp

7988–7991

I'm not sure I see when will this happen. Aren't we returning a node that isn't of type VT? I think that would cause problems. It looks like the output from constructDup should be of type VT, unless I'm missing something.

[miyuki]

Cast to the correct type

[dmgreen]

Thanks. LGTM

llvm/lib/Target/AArch64/AArch64ISelLowering.cpp

if (M[i] < 0)
       continue; // ignore UNDEF indices
     if (ExpectedElt != static_cast<unsigned>(M[i]))
       return false;
   }
 
   return true;
 }
 
+/// Check if a vector shuffle corresponds to a DUP instructions with a larger
+/// element width than the vector lane type. If that is the case the function
+/// returns true and writes the value of the DUP instruction lane operand into
+/// DupLaneOp
+static bool isWideDUPMask(ArrayRef<int> M, EVT VT, unsigned BlockSize,
+                          unsigned &DupLaneOp) {
+  assert((BlockSize == 16 || BlockSize == 32 || BlockSize == 64) &&
+         "Only possible block sizes for wide DUP are: 16, 32, 64");
+
+  if (BlockSize <= VT.getScalarSizeInBits())
+    return false;
+  if (BlockSize % VT.getScalarSizeInBits() != 0)
+    return false;
+  if (VT.getSizeInBits() % BlockSize != 0)
+    return false;
+
+  size_t SingleVecNumElements = VT.getVectorNumElements();
+  size_t NumEltsPerBlock = BlockSize / VT.getScalarSizeInBits();
+  size_t NumBlocks = VT.getSizeInBits() / BlockSize;
+
+  // We are looking for masks like
+  // [0, 1, 0, 1] or [2, 3, 2, 3] or [4, 5, 6, 7, 4, 5, 6, 7] where any element
+  // might be replaced by 'undefined'. BlockIndices will eventually contain
+  // lane indices of the duplicated block (i.e. [0, 1], [2, 3] and [4, 5, 6, 7]
+  // for the above examples)
+  SmallVector<int, 8> BlockElts(NumEltsPerBlock, -1);
+  for (size_t BlockIndex = 0; BlockIndex < NumBlocks; BlockIndex++)
+    for (size_t I = 0; I < NumEltsPerBlock; I++) {
+      int Elt = M[BlockIndex * NumEltsPerBlock + I];
+      if (Elt < 0)
+        continue;
+      // For now we don't support shuffles that use the second operand
+      if ((unsigned)Elt >= SingleVecNumElements)
+        return false;
+      if (BlockElts[I] < 0) {
+        BlockElts[I] = Elt;
+        continue;
+      }
+      if (BlockElts[I] != Elt)
+        return false;
+    }
+
+  // We found a candidate block (possibly with some undefs). It must be a
+  // sequence of consecutive integers starting with a value divisible by
+  // NumEltsPerBlock with some values possibly replaced by undef-s.
+
+  // Find first non-undef element
+  const int *FirstRealElt =
+      find_if(BlockElts, [](int Elt) { return Elt >= 0; });
+  assert(FirstRealElt && "Shuffle with all-undefs must have been caught by "
+                         "previous cases, e.g. isSplat()");
+  if (!FirstRealElt) {
+    DupLaneOp = 0;
+    return true;
+  }
+
+  // Index of FirstRealElt in BlockElts
+  size_t FirstRealIndex = FirstRealElt - &BlockElts[0];
+
+  if ((unsigned)*FirstRealElt < FirstRealIndex)
+    return false;
+  // BlockElts[0] must have the following value if it isn't undef:
+  size_t Elt0 = *FirstRealElt - FirstRealIndex;
+
+  // Check the first element
+  if (Elt0 % NumEltsPerBlock != 0)
+    return false;
+  // Check that the sequence indeed consists of consecutive integers (modulo
+  // undefs)
+  for (size_t I = 0; I < BlockElts.size(); I++)
+    if (BlockElts[I] >= 0 && (unsigned)BlockElts[I] != Elt0 + I)
+      return false;
+
+  DupLaneOp = Elt0 / NumEltsPerBlock;
+  return true;
+}
+
 // check if an EXT instruction can handle the shuffle mask when the
 // vector sources of the shuffle are different.
 static bool isEXTMask(ArrayRef<int> M, EVT VT, bool &ReverseEXT,
                       unsigned &Imm) {
   // Look for the first non-undef element.
   const int *FirstRealElt = find_if(M, [](int Elt) { return Elt >= 0; });
 
   // Benefit form APInt to handle overflow when calculating expected element.
static unsigned getDUPLANEOp(EVT EltType) {
   if (EltType == MVT::i32 || EltType == MVT::f32)
     return AArch64ISD::DUPLANE32;
   if (EltType == MVT::i64 || EltType == MVT::f64)
     return AArch64ISD::DUPLANE64;
 
   llvm_unreachable("Invalid vector element type?");
 }
 
-SDValue AArch64TargetLowering::LowerVECTOR_SHUFFLE(SDValue Op,
-                                                   SelectionDAG &DAG) const {
-  SDLoc dl(Op);
-  EVT VT = Op.getValueType();
-
-  ShuffleVectorSDNode *SVN = cast<ShuffleVectorSDNode>(Op.getNode());
-
-  // Convert shuffles that are directly supported on NEON to target-specific
-  // DAG nodes, instead of keeping them as shuffles and matching them again
-  // during code selection.  This is more efficient and avoids the possibility
-  // of inconsistencies between legalization and selection.
-  ArrayRef<int> ShuffleMask = SVN->getMask();
-
-  SDValue V1 = Op.getOperand(0);
-  SDValue V2 = Op.getOperand(1);
-
-  if (SVN->isSplat()) {
-    int Lane = SVN->getSplatIndex();
-    // If this is undef splat, generate it via "just" vdup, if possible.
-    if (Lane == -1)
-      Lane = 0;
-
-    if (Lane == 0 && V1.getOpcode() == ISD::SCALAR_TO_VECTOR)
-      return DAG.getNode(AArch64ISD::DUP, dl, V1.getValueType(),
-                         V1.getOperand(0));
-    // Test if V1 is a BUILD_VECTOR and the lane being referenced is a non-
-    // constant. If so, we can just reference the lane's definition directly.
-    if (V1.getOpcode() == ISD::BUILD_VECTOR &&
-        !isa<ConstantSDNode>(V1.getOperand(Lane)))
-      return DAG.getNode(AArch64ISD::DUP, dl, VT, V1.getOperand(Lane));
-
-    // Otherwise, duplicate from the lane of the input vector.
-    unsigned Opcode = getDUPLANEOp(V1.getValueType().getVectorElementType());
-
-    // Try to eliminate a bitcasted extract subvector before a DUPLANE.
-    auto getScaledOffsetDup = [](SDValue BitCast, int &LaneC, MVT &CastVT) {
-      // Match: dup (bitcast (extract_subv X, C)), LaneC
-      if (BitCast.getOpcode() != ISD::BITCAST ||
-          BitCast.getOperand(0).getOpcode() != ISD::EXTRACT_SUBVECTOR)
-        return false;
+static SDValue ConstructDup(SDValue V, int Lane, SDLoc dl, EVT VT,
+                            unsigned Opcode, SelectionDAG &DAG) {
 
-      // The extract index must align in the destination type. That may not
-      // happen if the bitcast is from narrow to wide type.
-      SDValue Extract = BitCast.getOperand(0);
-      unsigned ExtIdx = Extract.getConstantOperandVal(1);
-      unsigned SrcEltBitWidth = Extract.getScalarValueSizeInBits();
-      unsigned ExtIdxInBits = ExtIdx * SrcEltBitWidth;
-      unsigned CastedEltBitWidth = BitCast.getScalarValueSizeInBits();
-      if (ExtIdxInBits % CastedEltBitWidth != 0)
-        return false;
 
-      // Update the lane value by offsetting with the scaled extract index.
-      LaneC += ExtIdxInBits / CastedEltBitWidth;
 
-      // Determine the casted vector type of the wide vector input.
-      // dup (bitcast (extract_subv X, C)), LaneC --> dup (bitcast X), LaneC'
-      // Examples:
-      // dup (bitcast (extract_subv v2f64 X, 1) to v2f32), 1 --> dup v4f32 X, 3
-      // dup (bitcast (extract_subv v16i8 X, 8) to v4i16), 1 --> dup v8i16 X, 5
-      unsigned SrcVecNumElts =
-          Extract.getOperand(0).getValueSizeInBits() / CastedEltBitWidth;
-      CastVT = MVT::getVectorVT(BitCast.getSimpleValueType().getScalarType(),
-                                SrcVecNumElts);
-      return true;
-    };
-    MVT CastVT;
-    if (getScaledOffsetDup(V1, Lane, CastVT)) {
-      V1 = DAG.getBitcast(CastVT, V1.getOperand(0).getOperand(0));
-    } else if (V1.getOpcode() == ISD::EXTRACT_SUBVECTOR) {
-      // The lane is incremented by the index of the extract.
-      // Example: dup v2f32 (extract v4f32 X, 2), 1 --> dup v4f32 X, 3
-      Lane += V1.getConstantOperandVal(1);
-      V1 = V1.getOperand(0);
-    } else if (V1.getOpcode() == ISD::CONCAT_VECTORS) {
-      // The lane is decremented if we are splatting from the 2nd operand.
-      // Example: dup v4i32 (concat v2i32 X, v2i32 Y), 3 --> dup v4i32 Y, 1
-      unsigned Idx = Lane >= (int)VT.getVectorNumElements() / 2;
-      Lane -= Idx * VT.getVectorNumElements() / 2;
-      V1 = WidenVector(V1.getOperand(Idx), DAG);
-    } else if (VT.getSizeInBits() == 64) {
-      // Widen the operand to 128-bit register with undef.
-      V1 = WidenVector(V1, DAG);
+  if (getScaledOffsetDup(V, Lane, CastVT)) {
+    V = DAG.getBitcast(CastVT, V.getOperand(0).getOperand(0));
+  } else if (V.getOpcode() == ISD::EXTRACT_SUBVECTOR) {
+    Lane += V.getConstantOperandVal(1);
+    V = V.getOperand(0);
+  } else if (V.getOpcode() == ISD::CONCAT_VECTORS) {
+    V = WidenVector(V.getOperand(Idx), DAG);
+    V = WidenVector(V, DAG);
+  }
+  return DAG.getNode(Opcode, dl, VT, V, DAG.getConstant(Lane, dl, MVT::i64));
+}
+
+SDValue AArch64TargetLowering::LowerVECTOR_SHUFFLE(SDValue Op,
+                                                   SelectionDAG &DAG) const {
+  SDLoc dl(Op);
+  EVT VT = Op.getValueType();
+
+  ShuffleVectorSDNode *SVN = cast<ShuffleVectorSDNode>(Op.getNode());
+
+  // Convert shuffles that are directly supported on NEON to target-specific
+  // DAG nodes, instead of keeping them as shuffles and matching them again
+  // during code selection.  This is more efficient and avoids the possibility
+  // of inconsistencies between legalization and selection.
+  ArrayRef<int> ShuffleMask = SVN->getMask();
+
+  SDValue V1 = Op.getOperand(0);
+  SDValue V2 = Op.getOperand(1);
+
+  if (SVN->isSplat()) {
+    int Lane = SVN->getSplatIndex();
+    // If this is undef splat, generate it via "just" vdup, if possible.
+    if (Lane == -1)
+      Lane = 0;
+
+    if (Lane == 0 && V1.getOpcode() == ISD::SCALAR_TO_VECTOR)
+      return DAG.getNode(AArch64ISD::DUP, dl, V1.getValueType(),
+                         V1.getOperand(0));
+    // Test if V1 is a BUILD_VECTOR and the lane being referenced is a non-
+    // constant. If so, we can just reference the lane's definition directly.
+    if (V1.getOpcode() == ISD::BUILD_VECTOR &&
+        !isa<ConstantSDNode>(V1.getOperand(Lane)))
+      return DAG.getNode(AArch64ISD::DUP, dl, VT, V1.getOperand(Lane));
+
+    // Otherwise, duplicate from the lane of the input vector.
+    unsigned Opcode = getDUPLANEOp(V1.getValueType().getVectorElementType());
+    return ConstructDup(V1, Lane, dl, VT, Opcode, DAG);
+  }
+
+  // Check if the mask matches a DUP for a wider element
+  for (unsigned LaneSize : { 64U, 32U, 16U }) {
+    unsigned Lane = 0;
+    if (isWideDUPMask(ShuffleMask, VT, LaneSize, Lane)) {
+      unsigned Opcode = LaneSize == 64 ? AArch64ISD::DUPLANE64
+                                       : LaneSize == 32 ? AArch64ISD::DUPLANE32
+                                                        : AArch64ISD::DUPLANE16;
+
+      MVT NewEltTy = MVT::getIntegerVT(LaneSize);
+      unsigned NewEltCount = VT.getSizeInBits() / LaneSize;
+      MVT NewVecTy = MVT::getVectorVT(NewEltTy, NewEltCount);
+
+      V1 = DAG.getBitcast(NewVecTy, V1);
+      return ConstructDup(V1, Lane, dl, NewVecTy, Opcode, DAG);
     }
-    return DAG.getNode(Opcode, dl, VT, V1, DAG.getConstant(Lane, dl, MVT::i64));
   }
 
   if (isREVMask(ShuffleMask, VT, 64))
     return DAG.getNode(AArch64ISD::REV64, dl, V1.getValueType(), V1, V2);
   if (isREVMask(ShuffleMask, VT, 32))
     return DAG.getNode(AArch64ISD::REV32, dl, V1.getValueType(), V1, V2);
   if (isREVMask(ShuffleMask, VT, 16))
     return DAG.getNode(AArch64ISD::REV16, dl, V1.getValueType(), V1, V2);
 
   bool ReverseEXT = false;
   unsigned Imm;
   if (isEXTMask(ShuffleMask, VT, ReverseEXT, Imm)) {
     if (ReverseEXT)
       std::swap(V1, V2);
     Imm *= getExtFactor(V1);
     return DAG.getNode(AArch64ISD::EXT, dl, V1.getValueType(), V1, V2,
                        DAG.getConstant(Imm, dl, MVT::i32));
   } else if (V2->isUndef() && isSingletonEXTMask(ShuffleMask, VT, Imm)) {
     Imm *= getExtFactor(V1);
     return DAG.getNode(AArch64ISD::EXT, dl, V1.getValueType(), V1, V1,
                        DAG.getConstant(Imm, dl, MVT::i32));
   }
 
   unsigned WhichResult;
   if (isZIPMask(ShuffleMask, VT, WhichResult)) {
     unsigned Opc = (WhichResult == 0) ? AArch64ISD::ZIP1 : AArch64ISD::ZIP2;
     return DAG.getNode(Opc, dl, V1.getValueType(), V1, V2);
   }
   if (isUZPMask(ShuffleMask, VT, WhichResult)) {

llvm/test/CodeGen/AArch64/arm64-neon-2velem.ll

; CHECK-NEXT:    ret
   ret <4 x i16> %r
 }
 
 ; TODO: The pattern in LowerVECTOR_SHUFFLE does not match what we are looking for.
 
 define <4 x i16> @test_vadd_lane2_i16_bitcast_bigger_aligned(<4 x i16> %a, <16 x i8> %v) {
 ; CHECK-LABEL: test_vadd_lane2_i16_bitcast_bigger_aligned:
 ; CHECK:       // %bb.0:
-; CHECK-NEXT:    ext v1.8b, v1.8b, v0.8b, #2
+; CHECK-NEXT:    dup v1.4h, v1.h[2]
 ; CHECK-NEXT:    dup v1.4h, v1.h[1]
 ; CHECK-NEXT:    add v0.4h, v1.4h, v0.4h
 ; CHECK-NEXT:    ret
   %extract = shufflevector <16 x i8> %v, <16 x i8> undef, <8 x i32> <i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9>
   %bc = bitcast <8 x i8> %extract to <4 x i16>
   %splat = shufflevector <4 x i16> %bc, <4 x i16> undef, <4 x i32> <i32 1, i32 1, i32 1, i32 1>
   %r = add <4 x i16> %splat, %a
   ret <4 x i16> %r

llvm/test/CodeGen/AArch64/arm64-vext_reverse.ll

 ; RUN: llc -mtriple=arm64-linux-gnuabi < %s | FileCheck %s
 
 ; The following tests is to check the correctness of reversing input operand 
 ; of vext by enumerating all cases of using two undefs in shuffle masks.
 
 define <4 x i16> @vext_6701_0(<4 x i16> %a1, <4 x i16> %a2) {
 entry:
 ; CHECK-LABEL: vext_6701_0:
 ; CHECK: ext  v0.8b, v1.8b, v0.8b, #4
   %x = shufflevector <4 x i16> %a1, <4 x i16> %a2, <4 x i32> <i32 6, i32 7, i32 0, i32 1>
   ret <4 x i16> %x
 }
 
 define <4 x i16> @vext_6701_12(<4 x i16> %a1, <4 x i16> %a2) {
 entry:
 ; CHECK-LABEL: vext_6701_12:
-; CHECK: ext  v0.8b, v0.8b, v0.8b, #4
+; CHECK: dup v0.2s, v0.s[0]
   %x = shufflevector <4 x i16> %a1, <4 x i16> %a2, <4 x i32> <i32 undef, i32 undef, i32 0, i32 1>
   ret <4 x i16> %x
 }
 
 define <4 x i16> @vext_6701_13(<4 x i16> %a1, <4 x i16> %a2) {
 entry:
 ; CHECK-LABEL: vext_6701_13:
 ; CHECK: ext  v0.8b, v1.8b, v0.8b, #4

 ; CHECK: ext  v0.8b, v1.8b, v0.8b, #4
   %x = shufflevector <4 x i16> %a1, <4 x i16> %a2, <4 x i32> <i32 6, i32 undef, i32 0, i32 undef>
   ret <4 x i16> %x
 }
 
 define <4 x i16> @vext_6701_34(<4 x i16> %a1, <4 x i16> %a2) {
 entry:
 ; CHECK-LABEL: vext_6701_34:
-; CHECK: ext  v0.8b, v1.8b, v0.8b, #4
+; CHECK: dup  v0.2s, v1.s[1]
   %x = shufflevector <4 x i16> %a1, <4 x i16> %a2, <4 x i32> <i32 6, i32 7, i32 undef, i32 undef>
   ret <4 x i16> %x
 }
 
 define <4 x i16> @vext_5670_0(<4 x i16> %a1, <4 x i16> %a2) {
 entry:
 ; CHECK-LABEL: vext_5670_0:
 ; CHECK: ext  v0.8b, v1.8b, v0.8b, #2

llvm/test/CodeGen/AArch64/neon-extract.ll

 ; CHECK: ext {{v[0-9]+}}.16b, {{v[0-9]+}}.16b, {{v[0-9]+}}.16b, #{{0x6|6}}
 entry:
   %vext = shufflevector <16 x i8> %a, <16 x i8> undef, <16 x i32> <i32 20, i32 20, i32 20, i32 20, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 16, i32 20, i32 20, i32 20, i32 20, i32 20>
   ret <16 x i8> %vext
 }
 
 define <4 x i16> @test_undef_vext_s16(<4 x i16> %a) {
 ; CHECK-LABEL: test_undef_vext_s16:
-; CHECK: ext {{v[0-9]+}}.8b, {{v[0-9]+}}.8b, {{v[0-9]+}}.8b, #{{0x4|4}}
+; CHECK: dup v{{[0-9]+}}.2s, {{v[0-9]+}}.s[1]
 entry:
   %vext = shufflevector <4 x i16> %a, <4 x i16> undef, <4 x i32> <i32 2, i32 3, i32 4, i32 5>
   ret <4 x i16> %vext
 }
 
 define <8 x i16> @test_undef_vextq_s16(<8 x i16> %a) {
 ; CHECK-LABEL: test_undef_vextq_s16:
 ; CHECK: ext {{v[0-9]+}}.16b, {{v[0-9]+}}.16b, {{v[0-9]+}}.16b, #{{0x6|6}}
 entry:
   %vext = shufflevector <8 x i16> %a, <8 x i16> undef, <8 x i32> <i32 10, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10>
   ret <8 x i16> %vext
 }

llvm/test/CodeGen/AArch64/neon-wide-splat.ll

+; RUN: llc < %s -verify-machineinstrs -mtriple=aarch64-none-linux-gnu -mattr=+neon | FileCheck %s
+
+define <4 x i16> @shuffle1(<4 x i16> %v) {
+; CHECK-LABEL: shuffle1:
+; CHECK:         dup v0.2s, v0.s[0]
+; CHECK-NEXT:    ret
+entry:
+  %res = shufflevector <4 x i16> %v, <4 x i16> undef, <4 x i32> <i32 0, i32 undef, i32 0, i32 1>
+  ret <4 x i16> %res
+}
+
+define <4 x i16> @shuffle2(<4 x i16> %v) {
+; CHECK-LABEL: shuffle2:
+; CHECK:         dup v0.2s, v0.s[1]
+; CHECK-NEXT:    ret
+entry:
+  %res = shufflevector <4 x i16> %v, <4 x i16> undef, <4 x i32> <i32 2, i32 3, i32 undef, i32 3>
+  ret <4 x i16> %res
+}
+
+define <8 x i16> @shuffle3(<8 x i16> %v) {
+; CHECK-LABEL: shuffle3:
+; CHECK:         dup v0.2d, v0.d[0]
+; CHECK-NEXT:    ret
+entry:
+  %res = shufflevector <8 x i16> %v, <8 x i16> undef, <8 x i32> <i32 undef, i32 undef, i32 2, i32 3,
+                                                                 i32 undef, i32 1, i32 undef, i32 3>
+  ret <8 x i16> %res
+}
+
+define <4 x i32> @shuffle4(<4 x i32> %v) {
+; CHECK-LABEL: shuffle4:
+; CHECK:         dup v0.2d, v0.d[0]
+; CHECK-NEXT:    ret
+entry:
+  %res = shufflevector <4 x i32> %v, <4 x i32> undef, <4 x i32> <i32 0, i32 1, i32 0, i32 1>
+  ret <4 x i32> %res
+}
+
+define <16 x i8> @shuffle5(<16 x i8> %v) {
+; CHECK-LABEL: shuffle5:
+; CHECK:         dup v0.4s, v0.s[2]
+; CHECK-NEXT:    ret
+entry:
+  %res = shufflevector <16 x i8> %v, <16 x i8> undef, <16 x i32> <i32 8, i32 9, i32 10, i32 11,
+                                                                  i32 8, i32 9, i32 10, i32 11,
+                                                                  i32 8, i32 9, i32 10, i32 11,
+                                                                  i32 8, i32 9, i32 10, i32 11>
+  ret <16 x i8> %res
+}
+
+define <16 x i8> @shuffle6(<16 x i8> %v) {
+; CHECK-LABEL: shuffle6:
+; CHECK:         dup v0.2d, v0.d[1]
+; CHECK-NEXT:    ret
+entry:
+  %res = shufflevector <16 x i8> %v, <16 x i8> undef, <16 x i32> <i32 8, i32 9, i32 10, i32 11,
+                                                                  i32 12, i32 13, i32 14, i32 15,
+                                                                  i32 8, i32 9, i32 10, i32 11,
+                                                                  i32 12, i32 13, i32 14, i32 15>
+  ret <16 x i8> %res
+}
+
+define <8 x i8> @shuffle7(<8 x i8> %v) {
+; CHECK-LABEL: shuffle7:
+; CHECK:         dup v0.2s, v0.s[1]
+; CHECK-NEXT:    ret
+entry:
+  %res = shufflevector <8 x i8> %v, <8 x i8> undef, <8 x i32> <i32 4, i32 5, i32 6, i32 undef,
+                                                               i32 undef, i32 5, i32 6, i32 undef>
+  ret <8 x i8> %res
+}
+
+define <8 x i8> @shuffle8(<8 x i8> %v) {
+; CHECK-LABEL: shuffle8:
+; CHECK:         dup v0.4h, v0.h[3]
+; CHECK-NEXT:    ret
+entry:
+  %res = shufflevector <8 x i8> %v, <8 x i8> undef, <8 x i32> <i32 6, i32 7, i32 6, i32 undef,
+                                                               i32 undef, i32 7, i32 6, i32 undef>
+  ret <8 x i8> %res
+}
+
+define <8 x i8> @shuffle_not(<16 x i8> %v) {
+; CHECK-LABEL: shuffle_not:
+; CHECK:         ext v0.16b, v0.16b, v0.16b, #2
+  %res = shufflevector <16 x i8> %v, <16 x i8> undef, <8 x i32> <i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9>
+  ret <8 x i8> %res
+}
 No newline at end of file

llvm/test/CodeGen/AArch64/vecreduce-fmax-legalization-nan.ll

 }
 
 define float @test_v16f32(<16 x float> %a) nounwind {
 ; CHECK-LABEL: test_v16f32:
 ; CHECK:       // %bb.0:
 ; CHECK-NEXT:    fmaxnm v1.4s, v1.4s, v3.4s
 ; CHECK-NEXT:    fmaxnm v0.4s, v0.4s, v2.4s
 ; CHECK-NEXT:    fmaxnm v0.4s, v0.4s, v1.4s
-; CHECK-NEXT:    ext v1.16b, v0.16b, v0.16b, #8
+; CHECK-NEXT:    dup v1.2d, v0.d[1]
 ; CHECK-NEXT:    fmaxnm v0.4s, v0.4s, v1.4s
 ; CHECK-NEXT:    dup v1.4s, v0.s[1]
 ; CHECK-NEXT:    fmaxnm v0.4s, v0.4s, v1.4s
 ; CHECK-NEXT:    // kill: def $s0 killed $s0 killed $q0
 ; CHECK-NEXT:    ret
   %b = call float @llvm.experimental.vector.reduce.fmax.v16f32(<16 x float> %a)
   ret float %b
 }