[AArch64][SVE] Lower shuffles to permute instructions: zip1/2, uzp1/2, trn1/2

[wwei]

Summary

Attempt to lower a shuffle as a permute instruction(zip/uzp/trn) for fixed length SVE.

[david-arm]

llvm/lib/Target/AArch64/AArch64ISelLowering.cpp

19544

This looks a bit wrong because VT=Fixed Length Vector Type, but Op1 and Op2 have scalable vector types.

[wwei]

llvm/lib/Target/AArch64/AArch64ISelLowering.cpp

19544

No, the code here is correct. VT and ShuffleMask are used to check the type of the mask value, like zip or uzp mask.
Op1 and Op2 are used to generate new permute nodes, which will be scalable aarch64 SDNodes here.
In this patch, I encapsulated a new function for the code lowering shuffles to neon ZIP/UZP/TRN. Fortunately, the same code can be used for fixed-length SVE.

[peterwaller-arm]

LGTM, thanks.

[paulwalker-arm]

llvm/lib/Target/AArch64/AArch64ISelLowering.cpp

9738–9741

I don't believe this logic is universally safe. The use of ZIP2 specifically relies on knowing the exact size of the register to know which indices represent the "high half" of a vector register. This is something that is only known when sve-vector-bits-min==sve-vector-bits-max.

I also suspect functions like isZIPMask were written when only 64 and 128 bit legal vectors existed. I doubt the logic holds for the case when a vector is legal but not the exact size of the target vector register, as is the case with the SVE fixed length support.

I have not looked into the extent at which the other shuffles are affected but I suspect each have their own complexity.

llvm/test/CodeGen/AArch64/sve-fix-length-permute-zip-uzp-trn.ll

21 ↗

(On Diff #385388)

To highlight my previous comment. Index 16 is only the first byte of the second half of a 256bit vector and thus if the target has bigger vector register the use of zip2 will not have the correct behaviour.

[wwei]

llvm/lib/Target/AArch64/AArch64ISelLowering.cpp

9738–9741

Thanks, I get it. It seems that there are potential problems for the fixed-length shuffle vector lowering, include the patch https://reviews.llvm.org/D105289. I will update a new revision later to fix it by adding the check: sve-vector-bits-min==sve-vector-bits-max

[paulwalker-arm]

llvm/lib/Target/AArch64/AArch64ISelLowering.cpp

9738–9741

I believe that patch is ok given the restrictions it imposes. It is fine using the is###Mask functions as a way to identify a named shuffle. What is unsafe is assuming is###Mask means the ### instruction can be used directly when converted to scalable vectors.

So taking D105289 you can see that is uses isEXTMask to determine the type of shuffle but then the actually lowering code only handles a specific case and that case does not emit an EXT instruction as that would have fallen into the same trap.

[wwei]

llvm/lib/Target/AArch64/AArch64ISelLowering.cpp

9738–9741

Yeah, you are right. The logic in D105289 is correct, using EXTRACT_VECTOR_ELT and INSR instructions to match a specific EXT pattern, it's a smart solution.

[wwei]

update the patch, adding sve min/max size check.

[paulwalker-arm]

llvm/lib/Target/AArch64/AArch64ISelLowering.cpp

19547–19548

I still don't think this is enough as you've missed my previous I doubt the logic holds for the case when a vector is legal but not the exact size of the target vector register comment. Further discussion is attached to the test zip_v32i8 below...

I'm starting to wonder if it's worth breaking out lowerShuffleToZIP_UZP_TRN as I feel like once support is added for more of the combinations there's not likely to be much reuse. For example taking the zip1/zip2 case I think supporting zip1 is likely simple and just works for all legal VTs, but zip2 requires the VT to be exactly the size of the target vector (i.e. VT.getSizeInBits() == Subtarget->getMinSVEVectorSizeInBits() == Subtarget->getMaxSVEVectorSizeInBits()) or perhaps a different instruction sequence.

llvm/test/CodeGen/AArch64/sve-fix-length-permute-zip-uzp-trn.ll

24–25 ↗

(On Diff #385727)

The output for VBITS_EQ_256 and VBITS_EQ_512 is the same, which doesn't make sense considering the target vector length is different.

For the VBITS_EQ_256 case index 16 is the start of the high half of the target vector and thus zip2 works. But for the VBITS_EQ_512 case index 16 is actually the start of the second quarter of the target vector and thus zip2 will do the wrong thing.

[wwei]

llvm/lib/Target/AArch64/AArch64ISelLowering.cpp

19547–19548

Thanks, I agree with you. Perhaps the current implementation of lowerShuffleToZIP_UZP_TRN reusing neon code is not so good, since SVE may have variable length.
Maybe we should identify and classify different shuffle patterns. For zip/uzp/trn/rev these four shuffle types, we can roughly divide them into two categories:
zip1/uzp1/uzp2/trn1/trn2/revb/revh/revw （it' simple，and works for all legal VTs）
zip2/rev （complex，and need consider high half or whole register）
I think we can implement these two scenarios in steps.

[wwei]

update the patch，also there's another patch D114960 to support rev insts.

[paulwalker-arm]

llvm/lib/Target/AArch64/AArch64ISelLowering.cpp

19547–19548

Looking at the six shuffles trn1, trn2, uzp1, uzp2, zip1, zip2 I believe that when min_length != max_length != VT.getSizeInBits() only trn1, trn2 & zip1 are valid. We agree that zip2 is not valid because the top half of the input fixed length vector does not necessarily map to the top half of a scalable vector. The reason I believe both uzp variants are also invalid is because their underlying operation is to concat both input vectors and then extract every second element. However, the vectors we're concatenating may contain junk so you end up with:

expected uzp1 [A | B | C | D], [E | F | G | H] => [A | C | E | G]

SVE_VLS uzp [A | B | C | D | ? | ? | ? | ?], [E | F | G | H | ? | ? | ?] => [A | C | ? | ? | E | G | ? | ?]

And thus when you extract the fixed length result you'll end up with [A | C | ? | ?]

This I think further reduces the value of having lowerShuffleToZIP_UZP_TRN as it would be clearer to just handle each of the scenarios separately.

[wwei]

llvm/lib/Target/AArch64/AArch64ISelLowering.cpp

19547–19548

@paulwalker-arm Thanks for your detailed explanation. I fully agree with your point of view. Indeed, we don’t need lowerShuffleToZIP_UZP_TRN anymore, handle each case separately will make the code clearer. I will try to refactor the code later.

[wwei]

update the patch, lowerShuffleToZIP_UZP_TRN removed and test file refactored.

[wwei]

rebased.

[paulwalker-arm]

Just a passing review I'm afraid. I'll take a proper look tomorrow.

llvm/lib/Target/AArch64/AArch64ISelLowering.cpp

19567

I think you can use ContainerVT here? It might help with the formatting.

Same goes for the other places where you use Op1.getValueType().

19621–19630

I've not thought about it deeply so might be wrong but given these cases only use Op1 I'm wondering if they're always safe and thus don't need to be part of the MinSVESize == MaxSVESize == VT.getSizeInBits() restricted set?

[paulwalker-arm]

llvm/lib/Target/AArch64/AArch64ISelLowering.cpp

19586–19589

I figured this comment could be better and started to write something but ended up with

Functions like isZIPMask return true when a ISD::VECTOR_SHUFFLE's mask represents the same logical operation as performed by a ZIP instruction.  In isolation these functions do not mean the ISD::VECTOR_SHUFFLE is exactly equivalent to an AArch64 instruction.  There's the extra component of ISD::VECTOR_SHUFFLE's value type to consider.  Prior to SVE these functions only operated on 64/128bit vector types that have a direct mapping to a target register and so an exact mapping is implied.

However, when using SVE for fixed length vectors, most legal vector types are actually sub-vectors of a larger SVE register.  When mapping ISD::VECTOR_SHUFFLE to an SVE instruction care must be taken to consider how the mask's indices translate.  Specifically, when the mapping requires an exact meaning for a specific vector index (e.g. Index X is the last vector element in the register) then such mappings are often only safe when the exact SVE register size is know.  The main exception to this is when indices are logically relative to the first element of either ISD::VECTOR_SHUFFLE operand because these relative indices don't change when converting from fixed-length to scalable vector types (i.e. the start of a fixed length vector is always the start of a scalable vector).

Which is more like a novel than a comment :) I've posted it anyway just in case there's something in there that's useful.

llvm/test/CodeGen/AArch64/sve-fixed-length-permute-zip-uzp-trn.ll

11–15

It looks like InterleavedAccessPass is causing your code to be bypassed. I couldn't immediately see a way to disable the pass other than using -O0 which might cause other issues but I think if you use volatile loads and stores within these tests you'll get what you need.

414

Please place the attributes together at the end of the file because otherwise they're hard to find when trying to see what attributes exist for a specific function.

628

As above.

[wwei]

llvm/lib/Target/AArch64/AArch64ISelLowering.cpp

19567

Changed, using ContainerVT to replace Op1.getValueType()

19586–19589

Thanks for your comment, I used it directly in the code

19621–19630

I don't think it's safe. For undef case, Op1 will be used as input twice, it still has the problem like below:

uzp [A | B | C | D | ? | ? | ? | ?], [A | B | C | D | ? | ? | ?] => [A | C | ? | ? |A | C | ? | ?]

llvm/test/CodeGen/AArch64/sve-fixed-length-permute-zip-uzp-trn.ll

11–15

Thanks，volatile can work

414

done

[paulwalker-arm]

llvm/test/CodeGen/AArch64/sve-fixed-length-permute-zip-uzp-trn.ll

419

I think there's value in adding a comment that states why we can safely emit zip2 instructions. Not too detailed, it's just worth drawing the readers attention to the fact that for this and related tests the runtime vector length is known.

481

As above, there's value in adding a comment that mentions why we can safely emit uzp instructions.

627

Please add a small comment to highlight this is a negative test and why only zip1 instructions are emitted.

llvm/lib/Target/AArch64/AArch64ISelLowering.cpp

static SDValue tryWidenMaskForShuffle(SDValue Op, SelectionDAG &DAG) {
 
   return SDValue();
 }
 
 SDValue AArch64TargetLowering::LowerVECTOR_SHUFFLE(SDValue Op,
                                                    SelectionDAG &DAG) const {
   SDLoc dl(Op);
   EVT VT = Op.getValueType();
 
   ShuffleVectorSDNode *SVN = cast<ShuffleVectorSDNode>(Op.getNode());
 
   if (useSVEForFixedLengthVectorVT(VT))
     return LowerFixedLengthVECTOR_SHUFFLEToSVE(Op, DAG);
 
   // 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();
 
if (isEXTMask(ShuffleMask, VT, ReverseEXT, Imm) &&
     SDValue Scalar = DAG.getNode(
         ISD::EXTRACT_VECTOR_ELT, DL, ScalarTy, Op1,
         DAG.getConstant(VT.getVectorNumElements() - 1, DL, MVT::i64));
     Op = DAG.getNode(AArch64ISD::INSR, DL, ContainerVT, Op2, Scalar);
     return convertFromScalableVector(DAG, VT, Op);
   }
 
   for (unsigned LaneSize : {64U, 32U, 16U}) {
     if (isREVMask(ShuffleMask, VT, LaneSize)) {
       EVT NewVT =
           getPackedSVEVectorVT(EVT::getIntegerVT(*DAG.getContext(), LaneSize));
       unsigned RevOp;
       unsigned EltSz = VT.getScalarSizeInBits();
       if (EltSz == 8)
         RevOp = AArch64ISD::BSWAP_MERGE_PASSTHRU;
       else if (EltSz == 16)
         RevOp = AArch64ISD::REVH_MERGE_PASSTHRU;
       else
         RevOp = AArch64ISD::REVW_MERGE_PASSTHRU;
 
       Op = DAG.getNode(ISD::BITCAST, DL, NewVT, Op1);
       Op = LowerToPredicatedOp(Op, DAG, RevOp);
       Op = DAG.getNode(ISD::BITCAST, DL, ContainerVT, Op);
       return convertFromScalableVector(DAG, VT, Op);
     }
   }
 
+  unsigned WhichResult;
+  if (isZIPMask(ShuffleMask, VT, WhichResult) && WhichResult == 0)
+    return convertFromScalableVector(
+        DAG, VT, DAG.getNode(AArch64ISD::ZIP1, DL, ContainerVT, Op1, Op2));
+
+  if (isTRNMask(ShuffleMask, VT, WhichResult)) {
+    unsigned Opc = (WhichResult == 0) ? AArch64ISD::TRN1 : AArch64ISD::TRN2;
+    return convertFromScalableVector(
+        DAG, VT, DAG.getNode(Opc, DL, ContainerVT, Op1, Op2));
+  }
+
+  if (isZIP_v_undef_Mask(ShuffleMask, VT, WhichResult) && WhichResult == 0)
+    return convertFromScalableVector(
+        DAG, VT, DAG.getNode(AArch64ISD::ZIP1, DL, ContainerVT, Op1, Op1));
+
+  if (isTRN_v_undef_Mask(ShuffleMask, VT, WhichResult)) {
+    unsigned Opc = (WhichResult == 0) ? AArch64ISD::TRN1 : AArch64ISD::TRN2;
+    return convertFromScalableVector(
+        DAG, VT, DAG.getNode(Opc, DL, ContainerVT, Op1, Op1));
+  }
+
+  // Functions like isZIPMask return true when a ISD::VECTOR_SHUFFLE's mask
+  // represents the same logical operation as performed by a ZIP instruction. In
+  // isolation these functions do not mean the ISD::VECTOR_SHUFFLE is exactly
+  // equivalent to an AArch64 instruction. There's the extra component of
+  // ISD::VECTOR_SHUFFLE's value type to consider. Prior to SVE these functions
+  // only operated on 64/128bit vector types that have a direct mapping to a
+  // target register and so an exact mapping is implied.
+  // However, when using SVE for fixed length vectors, most legal vector types
+  // are actually sub-vectors of a larger SVE register. When mapping
+  // ISD::VECTOR_SHUFFLE to an SVE instruction care must be taken to consider
+  // how the mask's indices translate. Specifically, when the mapping requires
+  // an exact meaning for a specific vector index (e.g. Index X is the last
+  // vector element in the register) then such mappings are often only safe when
+  // the exact SVE register size is know. The main exception to this is when
+  // indices are logically relative to the first element of either
+  // ISD::VECTOR_SHUFFLE operand because these relative indices don't change
+  // when converting from fixed-length to scalable vector types (i.e. the start
+  // of a fixed length vector is always the start of a scalable vector).
   unsigned MinSVESize = Subtarget->getMinSVEVectorSizeInBits();
   unsigned MaxSVESize = Subtarget->getMaxSVEVectorSizeInBits();
-  if (MinSVESize == MaxSVESize && MaxSVESize == VT.getSizeInBits() &&
-      ShuffleVectorInst::isReverseMask(ShuffleMask) && Op2.isUndef()) {
+  if (MinSVESize == MaxSVESize && MaxSVESize == VT.getSizeInBits()) {
+    if (ShuffleVectorInst::isReverseMask(ShuffleMask) && Op2.isUndef()) {
       Op = DAG.getNode(ISD::VECTOR_REVERSE, DL, ContainerVT, Op1);
       return convertFromScalableVector(DAG, VT, Op);
     }
 
+    if (isZIPMask(ShuffleMask, VT, WhichResult) && WhichResult != 0)
+      return convertFromScalableVector(
+          DAG, VT, DAG.getNode(AArch64ISD::ZIP2, DL, ContainerVT, Op1, Op2));
+
+    if (isUZPMask(ShuffleMask, VT, WhichResult)) {
+      unsigned Opc = (WhichResult == 0) ? AArch64ISD::UZP1 : AArch64ISD::UZP2;
+      return convertFromScalableVector(
+          DAG, VT, DAG.getNode(Opc, DL, ContainerVT, Op1, Op2));
+    }
+
+    if (isZIP_v_undef_Mask(ShuffleMask, VT, WhichResult) && WhichResult != 0)
+      return convertFromScalableVector(
+          DAG, VT, DAG.getNode(AArch64ISD::ZIP2, DL, ContainerVT, Op1, Op1));
+
+    if (isUZP_v_undef_Mask(ShuffleMask, VT, WhichResult)) {
+      unsigned Opc = (WhichResult == 0) ? AArch64ISD::UZP1 : AArch64ISD::UZP2;
+      return convertFromScalableVector(
+          DAG, VT, DAG.getNode(Opc, DL, ContainerVT, Op1, Op1));
+    }
+  }
+
   return SDValue();
 }
 
 SDValue AArch64TargetLowering::getSVESafeBitCast(EVT VT, SDValue Op,
                                                  SelectionDAG &DAG) const {
   SDLoc DL(Op);
   EVT InVT = Op.getValueType();
   const TargetLowering &TLI = DAG.getTargetLoweringInfo();

llvm/test/CodeGen/AArch64/sve-fixed-length-permute-zip-uzp-trn.ll

+; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py
+; RUN: llc -aarch64-sve-vector-bits-min=256 -aarch64-sve-vector-bits-max=256 < %s | FileCheck %s -check-prefixes=CHECK,VBITS_EQ_256
+; RUN: llc -aarch64-sve-vector-bits-min=512 -aarch64-sve-vector-bits-max=512 < %s | FileCheck %s -check-prefixes=CHECK,VBITS_EQ_512
+
+target triple = "aarch64-unknown-linux-gnu"
+
+define void @zip1_v32i8(<32 x i8>* %a, <32 x i8>* %b) #0 {
+; VBITS_EQ_256-LABEL: zip1_v32i8:
+; VBITS_EQ_256:       // %bb.0:
+; VBITS_EQ_256-NEXT:    ptrue p0.b
+; VBITS_EQ_256-NEXT:    ld1b { z0.b }, p0/z, [x0]
+; VBITS_EQ_256-NEXT:    ld1b { z1.b }, p0/z, [x1]
+; VBITS_EQ_256-NEXT:    zip1 z0.b, z0.b, z1.b
+; VBITS_EQ_256-NEXT:    st1b { z0.b }, p0, [x0]
+; VBITS_EQ_256-NEXT:    ret
+;
+; VBITS_EQ_512-LABEL: zip1_v32i8:
+; VBITS_EQ_512:       // %bb.0:
+; VBITS_EQ_512-NEXT:    ptrue p0.b, vl32
+; VBITS_EQ_512-NEXT:    ld1b { z0.b }, p0/z, [x0]
+; VBITS_EQ_512-NEXT:    ld1b { z1.b }, p0/z, [x1]
+; VBITS_EQ_512-NEXT:    zip1 z0.b, z0.b, z1.b
+; VBITS_EQ_512-NEXT:    st1b { z0.b }, p0, [x0]
+; VBITS_EQ_512-NEXT:    ret
+  %tmp1 = load volatile <32 x i8>, <32 x i8>* %a
+  %tmp2 = load volatile <32 x i8>, <32 x i8>* %b
+  %tmp3 = shufflevector <32 x i8> %tmp1, <32 x i8> %tmp2, <32 x i32> <i32 0, i32 32, i32 1, i32 33, i32 2, i32 34, i32 3, i32 35, i32 4, i32 36, i32 5, i32 37, i32 6, i32 38, i32 7, i32 39, i32 8, i32 40, i32 9, i32 41, i32 10, i32 42, i32 11, i32 43, i32 12, i32 44, i32 13, i32 45, i32 14, i32 46, i32 15, i32 47>
+  store volatile <32 x i8> %tmp3, <32 x i8>* %a
+  ret void
+}
+
+define void @zip_v32i16(<32 x i16>* %a, <32 x i16>* %b) #0 {
+; VBITS_EQ_256-LABEL: zip_v32i16:
+; VBITS_EQ_256:       // %bb.0:
+; VBITS_EQ_256-NEXT:    mov x8, #16
+; VBITS_EQ_256-NEXT:    ptrue p0.h
+; VBITS_EQ_256-NEXT:    ld1h { z0.h }, p0/z, [x0, x8, lsl #1]
+; VBITS_EQ_256-NEXT:    ld1h { z1.h }, p0/z, [x0]
+; VBITS_EQ_256-NEXT:    ld1h { z2.h }, p0/z, [x1, x8, lsl #1]
+; VBITS_EQ_256-NEXT:    ld1h { z3.h }, p0/z, [x1]
+; VBITS_EQ_256-NEXT:    zip2 z4.h, z1.h, z3.h
+; VBITS_EQ_256-NEXT:    zip1 z1.h, z1.h, z3.h
+; VBITS_EQ_256-NEXT:    zip2 z3.h, z0.h, z2.h
+; VBITS_EQ_256-NEXT:    zip1 z0.h, z0.h, z2.h
+; VBITS_EQ_256-NEXT:    add z0.h, p0/m, z0.h, z1.h
+; VBITS_EQ_256-NEXT:    movprfx z1, z4
+; VBITS_EQ_256-NEXT:    add z1.h, p0/m, z1.h, z3.h
+; VBITS_EQ_256-NEXT:    st1h { z1.h }, p0, [x0, x8, lsl #1]
+; VBITS_EQ_256-NEXT:    st1h { z0.h }, p0, [x0]
+; VBITS_EQ_256-NEXT:    ret
+;
+; VBITS_EQ_512-LABEL: zip_v32i16:
+; VBITS_EQ_512:       // %bb.0:
+; VBITS_EQ_512-NEXT:    ptrue p0.h
+; VBITS_EQ_512-NEXT:    ld1h { z0.h }, p0/z, [x0]
+; VBITS_EQ_512-NEXT:    ld1h { z1.h }, p0/z, [x1]
+; VBITS_EQ_512-NEXT:    zip1 z2.h, z0.h, z1.h
+; VBITS_EQ_512-NEXT:    zip2 z0.h, z0.h, z1.h
+; VBITS_EQ_512-NEXT:    add z0.h, p0/m, z0.h, z2.h
+; VBITS_EQ_512-NEXT:    st1h { z0.h }, p0, [x0]
+; VBITS_EQ_512-NEXT:    ret
+  %tmp1 = load <32 x i16>, <32 x i16>* %a
+  %tmp2 = load <32 x i16>, <32 x i16>* %b
+  %tmp3 = shufflevector <32 x i16> %tmp1, <32 x i16> %tmp2, <32 x i32> <i32 0, i32 32, i32 1, i32 33, i32 2, i32 34, i32 3, i32 35, i32 4, i32 36, i32 5, i32 37, i32 6, i32 38, i32 7, i32 39, i32 8, i32 40, i32 9, i32 41, i32 10, i32 42, i32 11, i32 43, i32 12, i32 44, i32 13, i32 45, i32 14, i32 46, i32 15, i32 47>
+  %tmp4 = shufflevector <32 x i16> %tmp1, <32 x i16> %tmp2, <32 x i32> <i32 16, i32 48, i32 17, i32 49, i32 18, i32 50, i32 19, i32 51, i32 20, i32 52, i32 21, i32 53, i32 22, i32 54, i32 23, i32 55, i32 24, i32 56, i32 25, i32 57, i32 26, i32 58, i32 27, i32 59, i32 28, i32 60, i32 29, i32 61, i32 30, i32 62, i32 31, i32 63>
+  %tmp5 = add <32 x i16> %tmp3, %tmp4
+  store <32 x i16> %tmp5, <32 x i16>* %a
+  ret void
+}
+
+define void @zip1_v16i16(<16 x i16>* %a, <16 x i16>* %b) #0 {
+; VBITS_EQ_256-LABEL: zip1_v16i16:
+; VBITS_EQ_256:       // %bb.0:
+; VBITS_EQ_256-NEXT:    ptrue p0.h
+; VBITS_EQ_256-NEXT:    ld1h { z0.h }, p0/z, [x0]
+; VBITS_EQ_256-NEXT:    ld1h { z1.h }, p0/z, [x1]
+; VBITS_EQ_256-NEXT:    zip1 z0.h, z0.h, z1.h
+; VBITS_EQ_256-NEXT:    st1h { z0.h }, p0, [x0]
+; VBITS_EQ_256-NEXT:    ret
+;
+; VBITS_EQ_512-LABEL: zip1_v16i16:
+; VBITS_EQ_512:       // %bb.0:
+; VBITS_EQ_512-NEXT:    ptrue p0.h, vl16
+; VBITS_EQ_512-NEXT:    ld1h { z0.h }, p0/z, [x0]
+; VBITS_EQ_512-NEXT:    ld1h { z1.h }, p0/z, [x1]
+; VBITS_EQ_512-NEXT:    zip1 z0.h, z0.h, z1.h
+; VBITS_EQ_512-NEXT:    st1h { z0.h }, p0, [x0]
+; VBITS_EQ_512-NEXT:    ret
+  %tmp1 = load volatile <16 x i16>, <16 x i16>* %a
+  %tmp2 = load volatile <16 x i16>, <16 x i16>* %b
+  %tmp3 = shufflevector <16 x i16> %tmp1, <16 x i16> %tmp2, <16 x i32> <i32 0, i32 16, i32 1, i32 17, i32 2, i32 18, i32 3, i32 19, i32 4, i32 20, i32 5, i32 21, i32 6, i32 22, i32 7, i32 23>
+  store volatile <16 x i16> %tmp3, <16 x i16>* %a
+  ret void
+}
+
+define void @zip1_v8i32(<8 x i32>* %a, <8 x i32>* %b) #0 {
+; VBITS_EQ_256-LABEL: zip1_v8i32:
+; VBITS_EQ_256:       // %bb.0:
+; VBITS_EQ_256-NEXT:    ptrue p0.s
+; VBITS_EQ_256-NEXT:    ld1w { z0.s }, p0/z, [x0]
+; VBITS_EQ_256-NEXT:    ld1w { z1.s }, p0/z, [x1]
+; VBITS_EQ_256-NEXT:    zip1 z0.s, z0.s, z1.s
+; VBITS_EQ_256-NEXT:    st1w { z0.s }, p0, [x0]
+; VBITS_EQ_256-NEXT:    ret
+;
+; VBITS_EQ_512-LABEL: zip1_v8i32:
+; VBITS_EQ_512:       // %bb.0:
+; VBITS_EQ_512-NEXT:    ptrue p0.s, vl8
+; VBITS_EQ_512-NEXT:    ld1w { z0.s }, p0/z, [x0]
+; VBITS_EQ_512-NEXT:    ld1w { z1.s }, p0/z, [x1]
+; VBITS_EQ_512-NEXT:    zip1 z0.s, z0.s, z1.s
+; VBITS_EQ_512-NEXT:    st1w { z0.s }, p0, [x0]
+; VBITS_EQ_512-NEXT:    ret
+  %tmp1 = load volatile <8 x i32>, <8 x i32>* %a
+  %tmp2 = load volatile <8 x i32>, <8 x i32>* %b
+  %tmp3 = shufflevector <8 x i32> %tmp1, <8 x i32> %tmp2, <8 x i32> <i32 0, i32 8, i32 1, i32 9, i32 2, i32 10, i32 3, i32 11>
+  store volatile <8 x i32> %tmp3, <8 x i32>* %a
+  ret void
+}
+
+define void @zip_v4f64(<4 x double>* %a, <4 x double>* %b) #0 {
+; VBITS_EQ_256-LABEL: zip_v4f64:
+; VBITS_EQ_256:       // %bb.0:
+; VBITS_EQ_256-NEXT:    ptrue p0.d
+; VBITS_EQ_256-NEXT:    ld1d { z0.d }, p0/z, [x0]
+; VBITS_EQ_256-NEXT:    ld1d { z1.d }, p0/z, [x1]
+; VBITS_EQ_256-NEXT:    zip1 z2.d, z0.d, z1.d
+; VBITS_EQ_256-NEXT:    zip2 z0.d, z0.d, z1.d
+; VBITS_EQ_256-NEXT:    fadd z0.d, z2.d, z0.d
+; VBITS_EQ_256-NEXT:    st1d { z0.d }, p0, [x0]
+; VBITS_EQ_256-NEXT:    ret
+;
+; VBITS_EQ_512-LABEL: zip_v4f64:
+; VBITS_EQ_512:       // %bb.0:
+; VBITS_EQ_512-NEXT:    stp x29, x30, [sp, #-16]! // 16-byte Folded Spill
+; VBITS_EQ_512-NEXT:    sub x9, sp, #48
+; VBITS_EQ_512-NEXT:    mov x29, sp
+; VBITS_EQ_512-NEXT:    and sp, x9, #0xffffffffffffffe0
+; VBITS_EQ_512-NEXT:    .cfi_def_cfa w29, 16
+; VBITS_EQ_512-NEXT:    .cfi_offset w30, -8
+; VBITS_EQ_512-NEXT:    .cfi_offset w29, -16
+; VBITS_EQ_512-NEXT:    ptrue p0.d, vl4
+; VBITS_EQ_512-NEXT:    ld1d { z0.d }, p0/z, [x0]
+; VBITS_EQ_512-NEXT:    ld1d { z1.d }, p0/z, [x1]
+; VBITS_EQ_512-NEXT:    mov z2.d, z1.d[3]
+; VBITS_EQ_512-NEXT:    mov z3.d, z0.d[3]
+; VBITS_EQ_512-NEXT:    stp d3, d2, [sp, #16]
+; VBITS_EQ_512-NEXT:    mov z2.d, z1.d[2]
+; VBITS_EQ_512-NEXT:    mov z3.d, z0.d[2]
+; VBITS_EQ_512-NEXT:    zip1 z0.d, z0.d, z1.d
+; VBITS_EQ_512-NEXT:    stp d3, d2, [sp]
+; VBITS_EQ_512-NEXT:    ld1d { z2.d }, p0/z, [sp]
+; VBITS_EQ_512-NEXT:    fadd z0.d, p0/m, z0.d, z2.d
+; VBITS_EQ_512-NEXT:    st1d { z0.d }, p0, [x0]
+; VBITS_EQ_512-NEXT:    mov sp, x29
+; VBITS_EQ_512-NEXT:    ldp x29, x30, [sp], #16 // 16-byte Folded Reload
+; VBITS_EQ_512-NEXT:    ret
+  %tmp1 = load <4 x double>, <4 x double>* %a
+  %tmp2 = load <4 x double>, <4 x double>* %b
+  %tmp3 = shufflevector <4 x double> %tmp1, <4 x double> %tmp2, <4 x i32> <i32 0, i32 4, i32 1, i32 5>
+  %tmp4 = shufflevector <4 x double> %tmp1, <4 x double> %tmp2, <4 x i32> <i32 2, i32 6, i32 3, i32 7>
+  %tmp5 = fadd <4 x double> %tmp3, %tmp4
+  store <4 x double> %tmp5, <4 x double>* %a
+  ret void
+}
+
+; Don't use SVE for 128-bit vectors
+define void @zip_v4i32(<4 x i32>* %a, <4 x i32>* %b) #0 {
+; CHECK-LABEL: zip_v4i32:
+; CHECK:       // %bb.0:
+; CHECK-NEXT:    ldr q0, [x0]
+; CHECK-NEXT:    ldr q1, [x1]
+; CHECK-NEXT:    zip1 v2.4s, v0.4s, v1.4s
+; CHECK-NEXT:    zip2 v0.4s, v0.4s, v1.4s
+; CHECK-NEXT:    add v0.4s, v2.4s, v0.4s
+; CHECK-NEXT:    str q0, [x0]
+; CHECK-NEXT:    ret
+  %tmp1 = load <4 x i32>, <4 x i32>* %a
+  %tmp2 = load <4 x i32>, <4 x i32>* %b
+  %tmp3 = shufflevector <4 x i32> %tmp1, <4 x i32> %tmp2, <4 x i32> <i32 0, i32 4, i32 1, i32 5>
+  %tmp4 = shufflevector <4 x i32> %tmp1, <4 x i32> %tmp2, <4 x i32> <i32 2, i32 6, i32 3, i32 7>
+  %tmp5 = add <4 x i32> %tmp3, %tmp4
+  store <4 x i32> %tmp5, <4 x i32>* %a
+  ret void
+}
+
+define void @zip1_v8i32_undef(<8 x i32>* %a) #0 {
+; VBITS_EQ_256-LABEL: zip1_v8i32_undef:
+; VBITS_EQ_256:       // %bb.0:
+; VBITS_EQ_256-NEXT:    ptrue p0.s
+; VBITS_EQ_256-NEXT:    ld1w { z0.s }, p0/z, [x0]
+; VBITS_EQ_256-NEXT:    zip1 z0.s, z0.s, z0.s
+; VBITS_EQ_256-NEXT:    st1w { z0.s }, p0, [x0]
+; VBITS_EQ_256-NEXT:    ret
+;
+; VBITS_EQ_512-LABEL: zip1_v8i32_undef:
+; VBITS_EQ_512:       // %bb.0:
+; VBITS_EQ_512-NEXT:    ptrue p0.s, vl8
+; VBITS_EQ_512-NEXT:    ld1w { z0.s }, p0/z, [x0]
+; VBITS_EQ_512-NEXT:    zip1 z0.s, z0.s, z0.s
+; VBITS_EQ_512-NEXT:    st1w { z0.s }, p0, [x0]
+; VBITS_EQ_512-NEXT:    ret
+  %tmp1 = load  volatile <8 x i32>, <8 x i32>* %a
+  %tmp2 = shufflevector <8 x i32> %tmp1, <8 x i32> undef, <8 x i32> <i32 0, i32 0, i32 1, i32 1, i32 2, i32 2, i32 3, i32 3>
+  store volatile <8 x i32> %tmp2, <8 x i32>* %a
+  ret void
+}
+
+define void @trn_v32i8(<32 x i8>* %a, <32 x i8>* %b) #0 {
+; VBITS_EQ_256-LABEL: trn_v32i8:
+; VBITS_EQ_256:       // %bb.0:
+; VBITS_EQ_256-NEXT:    ptrue p0.b
+; VBITS_EQ_256-NEXT:    ld1b { z0.b }, p0/z, [x0]
+; VBITS_EQ_256-NEXT:    ld1b { z1.b }, p0/z, [x1]
+; VBITS_EQ_256-NEXT:    trn1 z2.b, z0.b, z1.b
+; VBITS_EQ_256-NEXT:    trn2 z0.b, z0.b, z1.b
+; VBITS_EQ_256-NEXT:    add z0.b, p0/m, z0.b, z2.b
+; VBITS_EQ_256-NEXT:    st1b { z0.b }, p0, [x0]
+; VBITS_EQ_256-NEXT:    ret
+;
+; VBITS_EQ_512-LABEL: trn_v32i8:
+; VBITS_EQ_512:       // %bb.0:
+; VBITS_EQ_512-NEXT:    ptrue p0.b, vl32
+; VBITS_EQ_512-NEXT:    ld1b { z0.b }, p0/z, [x0]
+; VBITS_EQ_512-NEXT:    ld1b { z1.b }, p0/z, [x1]
+; VBITS_EQ_512-NEXT:    trn1 z2.b, z0.b, z1.b
+; VBITS_EQ_512-NEXT:    trn2 z0.b, z0.b, z1.b
+; VBITS_EQ_512-NEXT:    add z0.b, p0/m, z0.b, z2.b
+; VBITS_EQ_512-NEXT:    st1b { z0.b }, p0, [x0]
+; VBITS_EQ_512-NEXT:    ret
+  %tmp1 = load <32 x i8>, <32 x i8>* %a
+  %tmp2 = load <32 x i8>, <32 x i8>* %b
+  %tmp3 = shufflevector <32 x i8> %tmp1, <32 x i8> %tmp2, <32 x i32> <i32 0, i32 32, i32 2, i32 34, i32 4, i32 36, i32 6, i32 38, i32 8, i32 40, i32 10, i32 42, i32 12, i32 44, i32 14, i32 46, i32 16, i32 48, i32 18, i32 50, i32 20, i32 52, i32 22, i32 54, i32 24, i32 56, i32 26, i32 58, i32 28, i32 60, i32 30, i32 62>
+  %tmp4 = shufflevector <32 x i8> %tmp1, <32 x i8> %tmp2, <32 x i32> <i32 1, i32 33, i32 3, i32 35, i32 undef, i32 37, i32 7, i32 undef, i32 undef, i32 41, i32 11, i32 43, i32 13, i32 45, i32 15, i32 47, i32 17, i32 49, i32 19, i32 51, i32 21, i32 53, i32 23, i32 55, i32 25, i32 57, i32 27, i32 59, i32 29, i32 61, i32 31, i32 63>
+  %tmp5 = add <32 x i8> %tmp3, %tmp4
+  store <32 x i8> %tmp5, <32 x i8>* %a
+  ret void
+}
+
+define void @trn_v32i16(<32 x i16>* %a, <32 x i16>* %b) #0 {
+; VBITS_EQ_256-LABEL: trn_v32i16:
+; VBITS_EQ_256:       // %bb.0:
+; VBITS_EQ_256-NEXT:    mov x8, #16
+; VBITS_EQ_256-NEXT:    ptrue p0.h
+; VBITS_EQ_256-NEXT:    ld1h { z0.h }, p0/z, [x0, x8, lsl #1]
+; VBITS_EQ_256-NEXT:    ld1h { z1.h }, p0/z, [x0]
+; VBITS_EQ_256-NEXT:    ld1h { z2.h }, p0/z, [x1, x8, lsl #1]
+; VBITS_EQ_256-NEXT:    ld1h { z3.h }, p0/z, [x1]
+; VBITS_EQ_256-NEXT:    trn1 z4.h, z0.h, z2.h
+; VBITS_EQ_256-NEXT:    trn1 z5.h, z1.h, z3.h
+; VBITS_EQ_256-NEXT:    trn2 z0.h, z0.h, z2.h
+; VBITS_EQ_256-NEXT:    trn2 z1.h, z1.h, z3.h
+; VBITS_EQ_256-NEXT:    add z0.h, p0/m, z0.h, z4.h
+; VBITS_EQ_256-NEXT:    add z1.h, p0/m, z1.h, z5.h
+; VBITS_EQ_256-NEXT:    st1h { z0.h }, p0, [x0, x8, lsl #1]
+; VBITS_EQ_256-NEXT:    st1h { z1.h }, p0, [x0]
+; VBITS_EQ_256-NEXT:    ret
+;
+; VBITS_EQ_512-LABEL: trn_v32i16:
+; VBITS_EQ_512:       // %bb.0:
+; VBITS_EQ_512-NEXT:    ptrue p0.h
+; VBITS_EQ_512-NEXT:    ld1h { z0.h }, p0/z, [x0]
+; VBITS_EQ_512-NEXT:    ld1h { z1.h }, p0/z, [x1]
+; VBITS_EQ_512-NEXT:    trn1 z2.h, z0.h, z1.h
+; VBITS_EQ_512-NEXT:    trn2 z0.h, z0.h, z1.h
+; VBITS_EQ_512-NEXT:    add z0.h, p0/m, z0.h, z2.h
+; VBITS_EQ_512-NEXT:    st1h { z0.h }, p0, [x0]
+; VBITS_EQ_512-NEXT:    ret
+  %tmp1 = load <32 x i16>, <32 x i16>* %a
+  %tmp2 = load <32 x i16>, <32 x i16>* %b
+  %tmp3 = shufflevector <32 x i16> %tmp1, <32 x i16> %tmp2, <32 x i32> <i32 0, i32 32, i32 2, i32 34, i32 4, i32 36, i32 6, i32 38, i32 8, i32 40, i32 10, i32 42, i32 12, i32 44, i32 14, i32 46, i32 16, i32 48, i32 18, i32 50, i32 20, i32 52, i32 22, i32 54, i32 24, i32 56, i32 26, i32 58, i32 28, i32 60, i32 30, i32 62>
+  %tmp4 = shufflevector <32 x i16> %tmp1, <32 x i16> %tmp2, <32 x i32> <i32 1, i32 33, i32 3, i32 35, i32 undef, i32 37, i32 7, i32 undef, i32 undef, i32 41, i32 11, i32 43, i32 13, i32 45, i32 15, i32 47, i32 17, i32 49, i32 19, i32 51, i32 21, i32 53, i32 23, i32 55, i32 25, i32 57, i32 27, i32 59, i32 29, i32 61, i32 31, i32 63>
+  %tmp5 = add <32 x i16> %tmp3, %tmp4
+  store <32 x i16> %tmp5, <32 x i16>* %a
+  ret void
+}
+
+define void @trn_v16i16(<16 x i16>* %a, <16 x i16>* %b) #0 {
+; VBITS_EQ_256-LABEL: trn_v16i16:
+; VBITS_EQ_256:       // %bb.0:
+; VBITS_EQ_256-NEXT:    ptrue p0.h
+; VBITS_EQ_256-NEXT:    ld1h { z0.h }, p0/z, [x0]
+; VBITS_EQ_256-NEXT:    ld1h { z1.h }, p0/z, [x1]
+; VBITS_EQ_256-NEXT:    trn1 z2.h, z0.h, z1.h
+; VBITS_EQ_256-NEXT:    trn2 z0.h, z0.h, z1.h
+; VBITS_EQ_256-NEXT:    add z0.h, p0/m, z0.h, z2.h
+; VBITS_EQ_256-NEXT:    st1h { z0.h }, p0, [x0]
+; VBITS_EQ_256-NEXT:    ret
+;
+; VBITS_EQ_512-LABEL: trn_v16i16:
+; VBITS_EQ_512:       // %bb.0:
+; VBITS_EQ_512-NEXT:    ptrue p0.h, vl16
+; VBITS_EQ_512-NEXT:    ld1h { z0.h }, p0/z, [x0]
+; VBITS_EQ_512-NEXT:    ld1h { z1.h }, p0/z, [x1]
+; VBITS_EQ_512-NEXT:    trn1 z2.h, z0.h, z1.h
+; VBITS_EQ_512-NEXT:    trn2 z0.h, z0.h, z1.h
+; VBITS_EQ_512-NEXT:    add z0.h, p0/m, z0.h, z2.h
+; VBITS_EQ_512-NEXT:    st1h { z0.h }, p0, [x0]
+; VBITS_EQ_512-NEXT:    ret
+  %tmp1 = load <16 x i16>, <16 x i16>* %a
+  %tmp2 = load <16 x i16>, <16 x i16>* %b
+  %tmp3 = shufflevector <16 x i16> %tmp1, <16 x i16> %tmp2, <16 x i32> <i32 0, i32 16, i32 2, i32 18, i32 4, i32 20, i32 6, i32 22, i32 8, i32 24, i32 10, i32 26, i32 12, i32 28, i32 14, i32 30>
+  %tmp4 = shufflevector <16 x i16> %tmp1, <16 x i16> %tmp2, <16 x i32> <i32 1, i32 17, i32 3, i32 19, i32 5, i32 21, i32 7, i32 23, i32 9, i32 25, i32 11, i32 27, i32 13, i32 29, i32 15, i32 31>
+  %tmp5 = add <16 x i16> %tmp3, %tmp4
+  store <16 x i16> %tmp5, <16 x i16>* %a
+  ret void
+}
+
+define void @trn_v8i32(<8 x i32>* %a, <8 x i32>* %b) #0 {
+; VBITS_EQ_256-LABEL: trn_v8i32:
+; VBITS_EQ_256:       // %bb.0:
+; VBITS_EQ_256-NEXT:    ptrue p0.s
+; VBITS_EQ_256-NEXT:    ld1w { z0.s }, p0/z, [x0]
+; VBITS_EQ_256-NEXT:    ld1w { z1.s }, p0/z, [x1]
+; VBITS_EQ_256-NEXT:    trn1 z2.s, z0.s, z1.s
+; VBITS_EQ_256-NEXT:    trn2 z0.s, z0.s, z1.s
+; VBITS_EQ_256-NEXT:    add z0.s, p0/m, z0.s, z2.s
+; VBITS_EQ_256-NEXT:    st1w { z0.s }, p0, [x0]
+; VBITS_EQ_256-NEXT:    ret
+;
+; VBITS_EQ_512-LABEL: trn_v8i32:
+; VBITS_EQ_512:       // %bb.0:
+; VBITS_EQ_512-NEXT:    ptrue p0.s, vl8
+; VBITS_EQ_512-NEXT:    ld1w { z0.s }, p0/z, [x0]
+; VBITS_EQ_512-NEXT:    ld1w { z1.s }, p0/z, [x1]
+; VBITS_EQ_512-NEXT:    trn1 z2.s, z0.s, z1.s
+; VBITS_EQ_512-NEXT:    trn2 z0.s, z0.s, z1.s
+; VBITS_EQ_512-NEXT:    add z0.s, p0/m, z0.s, z2.s
+; VBITS_EQ_512-NEXT:    st1w { z0.s }, p0, [x0]
+; VBITS_EQ_512-NEXT:    ret
+  %tmp1 = load <8 x i32>, <8 x i32>* %a
+  %tmp2 = load <8 x i32>, <8 x i32>* %b
+  %tmp3 = shufflevector <8 x i32> %tmp1, <8 x i32> %tmp2, <8 x i32> <i32 0, i32 8, i32 undef, i32 undef, i32 4, i32 12, i32 6, i32 14>
+  %tmp4 = shufflevector <8 x i32> %tmp1, <8 x i32> %tmp2, <8 x i32> <i32 1, i32 undef, i32 3, i32 11, i32 5, i32 13, i32 undef, i32 undef>
+  %tmp5 = add <8 x i32> %tmp3, %tmp4
+  store <8 x i32> %tmp5, <8 x i32>* %a
+  ret void
+}
+
+define void @trn_v4f64(<4 x double>* %a, <4 x double>* %b) #0 {
+; VBITS_EQ_256-LABEL: trn_v4f64:
+; VBITS_EQ_256:       // %bb.0:
+; VBITS_EQ_256-NEXT:    ptrue p0.d
+; VBITS_EQ_256-NEXT:    ld1d { z0.d }, p0/z, [x0]
+; VBITS_EQ_256-NEXT:    ld1d { z1.d }, p0/z, [x1]
+; VBITS_EQ_256-NEXT:    trn1 z2.d, z0.d, z1.d
+; VBITS_EQ_256-NEXT:    trn2 z0.d, z0.d, z1.d
+; VBITS_EQ_256-NEXT:    fadd z0.d, z2.d, z0.d
+; VBITS_EQ_256-NEXT:    st1d { z0.d }, p0, [x0]
+; VBITS_EQ_256-NEXT:    ret
+;
+; VBITS_EQ_512-LABEL: trn_v4f64:
+; VBITS_EQ_512:       // %bb.0:
+; VBITS_EQ_512-NEXT:    ptrue p0.d, vl4
+; VBITS_EQ_512-NEXT:    ld1d { z0.d }, p0/z, [x0]
+; VBITS_EQ_512-NEXT:    ld1d { z1.d }, p0/z, [x1]
+; VBITS_EQ_512-NEXT:    trn1 z2.d, z0.d, z1.d
+; VBITS_EQ_512-NEXT:    trn2 z0.d, z0.d, z1.d
+; VBITS_EQ_512-NEXT:    fadd z0.d, p0/m, z0.d, z2.d
+; VBITS_EQ_512-NEXT:    st1d { z0.d }, p0, [x0]
+; VBITS_EQ_512-NEXT:    ret
+  %tmp1 = load <4 x double>, <4 x double>* %a
+  %tmp2 = load <4 x double>, <4 x double>* %b
+  %tmp3 = shufflevector <4 x double> %tmp1, <4 x double> %tmp2, <4 x i32> <i32 0, i32 4, i32 2, i32 6>
+  %tmp4 = shufflevector <4 x double> %tmp1, <4 x double> %tmp2, <4 x i32> <i32 1, i32 5, i32 3, i32 7>
+  %tmp5 = fadd <4 x double> %tmp3, %tmp4
+  store <4 x double> %tmp5, <4 x double>* %a
+  ret void
+}
+
+; Don't use SVE for 128-bit vectors
+define void @trn_v4f32(<4 x float>* %a, <4 x float>* %b) #0 {
+; CHECK-LABEL: trn_v4f32:
+; CHECK:       // %bb.0:
+; CHECK-NEXT:    ldr q0, [x0]
+; CHECK-NEXT:    ldr q1, [x1]
+; CHECK-NEXT:    trn1 v2.4s, v0.4s, v1.4s
+; CHECK-NEXT:    trn2 v0.4s, v0.4s, v1.4s
+; CHECK-NEXT:    fadd v0.4s, v2.4s, v0.4s
+; CHECK-NEXT:    str q0, [x0]
+; CHECK-NEXT:    ret
+  %tmp1 = load <4 x float>, <4 x float>* %a
+  %tmp2 = load <4 x float>, <4 x float>* %b
+  %tmp3 = shufflevector <4 x float> %tmp1, <4 x float> %tmp2, <4 x i32> <i32 0, i32 4, i32 2, i32 6>
+  %tmp4 = shufflevector <4 x float> %tmp1, <4 x float> %tmp2, <4 x i32> <i32 1, i32 5, i32 3, i32 7>
+  %tmp5 = fadd <4 x float> %tmp3, %tmp4
+  store <4 x float> %tmp5, <4 x float>* %a
+  ret void
+}
+
+define void @trn_v8i32_undef(<8 x i32>* %a) #0 {
+; VBITS_EQ_256-LABEL: trn_v8i32_undef:
+; VBITS_EQ_256:       // %bb.0:
+; VBITS_EQ_256-NEXT:    ptrue p0.s
+; VBITS_EQ_256-NEXT:    ld1w { z0.s }, p0/z, [x0]
+; VBITS_EQ_256-NEXT:    trn1 z1.s, z0.s, z0.s
+; VBITS_EQ_256-NEXT:    trn2 z0.s, z0.s, z0.s
+; VBITS_EQ_256-NEXT:    add z0.s, p0/m, z0.s, z1.s
+; VBITS_EQ_256-NEXT:    st1w { z0.s }, p0, [x0]
+; VBITS_EQ_256-NEXT:    ret
+;
+; VBITS_EQ_512-LABEL: trn_v8i32_undef:
+; VBITS_EQ_512:       // %bb.0:
+; VBITS_EQ_512-NEXT:    ptrue p0.s, vl8
+; VBITS_EQ_512-NEXT:    ld1w { z0.s }, p0/z, [x0]
+; VBITS_EQ_512-NEXT:    trn1 z1.s, z0.s, z0.s
+; VBITS_EQ_512-NEXT:    trn2 z0.s, z0.s, z0.s
+; VBITS_EQ_512-NEXT:    add z0.s, p0/m, z0.s, z1.s
+; VBITS_EQ_512-NEXT:    st1w { z0.s }, p0, [x0]
+; VBITS_EQ_512-NEXT:    ret
+  %tmp1 = load <8 x i32>, <8 x i32>* %a
+  %tmp3 = shufflevector <8 x i32> %tmp1, <8 x i32> undef, <8 x i32> <i32 0, i32 0, i32 2, i32 2, i32 4, i32 4, i32 6, i32 6>
+  %tmp4 = shufflevector <8 x i32> %tmp1, <8 x i32> undef, <8 x i32> <i32 1, i32 1, i32 3, i32 3, i32 5, i32 5, i32 7, i32 7>
+  %tmp5 = add <8 x i32> %tmp3, %tmp4
+  store <8 x i32> %tmp5, <8 x i32>* %a
+  ret void
+}
+
+define void @zip2_v32i8(<32 x i8>* %a, <32 x i8>* %b) #1 {
+; CHECK-LABEL: zip2_v32i8:
+; CHECK:       // %bb.0:
+; CHECK-NEXT:    ptrue p0.b
+; CHECK-NEXT:    ld1b { z0.b }, p0/z, [x0]
+; CHECK-NEXT:    ld1b { z1.b }, p0/z, [x1]
+; CHECK-NEXT:    zip2 z0.b, z0.b, z1.b
+; CHECK-NEXT:    st1b { z0.b }, p0, [x0]
+; CHECK-NEXT:    ret
+  %tmp1 = load volatile <32 x i8>, <32 x i8>* %a
+  %tmp2 = load volatile <32 x i8>, <32 x i8>* %b
+  %tmp3 = shufflevector <32 x i8> %tmp1, <32 x i8> %tmp2, <32 x i32> <i32 16, i32 48, i32 17, i32 49, i32 18, i32 50, i32 19, i32 51, i32 20, i32 52, i32 21, i32 53, i32 22, i32 54, i32 23, i32 55, i32 24, i32 56, i32 25, i32 57, i32 26, i32 58, i32 27, i32 59, i32 28, i32 60, i32 29, i32 61, i32 30, i32 62, i32 31, i32 63>
+  store volatile <32 x i8> %tmp3, <32 x i8>* %a
+  ret void
+}
+
+define void @zip2_v16i16(<16 x i16>* %a, <16 x i16>* %b) #1 {
+; CHECK-LABEL: zip2_v16i16:
+; CHECK:       // %bb.0:
+; CHECK-NEXT:    ptrue p0.h
+; CHECK-NEXT:    ld1h { z0.h }, p0/z, [x0]
+; CHECK-NEXT:    ld1h { z1.h }, p0/z, [x1]
+; CHECK-NEXT:    zip2 z0.h, z0.h, z1.h
+; CHECK-NEXT:    st1h { z0.h }, p0, [x0]
+; CHECK-NEXT:    ret
+  %tmp1 = load volatile <16 x i16>, <16 x i16>* %a
+  %tmp2 = load volatile <16 x i16>, <16 x i16>* %b
+  %tmp3 = shufflevector <16 x i16> %tmp1, <16 x i16> %tmp2, <16 x i32> <i32 8, i32 24, i32 9, i32 25, i32 10, i32 26, i32 11, i32 27, i32 12, i32 28, i32 13, i32 29, i32 14, i32 30, i32 15, i32 31>
+  store volatile <16 x i16> %tmp3, <16 x i16>* %a
+  ret void
+}
+
+define void @zip2_v8i32(<8 x i32>* %a, <8 x i32>* %b) #1 {
+; CHECK-LABEL: zip2_v8i32:
+; CHECK:       // %bb.0:
+; CHECK-NEXT:    ptrue p0.s
+; CHECK-NEXT:    ld1w { z0.s }, p0/z, [x0]
+; CHECK-NEXT:    ld1w { z1.s }, p0/z, [x1]
+; CHECK-NEXT:    zip2 z0.s, z0.s, z1.s
+; CHECK-NEXT:    st1w { z0.s }, p0, [x0]
+; CHECK-NEXT:    ret
+  %tmp1 = load volatile <8 x i32>, <8 x i32>* %a
+  %tmp2 = load volatile <8 x i32>, <8 x i32>* %b
+  %tmp3 = shufflevector <8 x i32> %tmp1, <8 x i32> %tmp2, <8 x i32> <i32 4, i32 12, i32 5, i32 13, i32 6, i32 14, i32 7, i32 15>
+  store volatile <8 x i32> %tmp3, <8 x i32>* %a
+  ret void
+}
+
+define void @zip2_v8i32_undef(<8 x i32>* %a) #1 {
+; CHECK-LABEL: zip2_v8i32_undef:
+; CHECK:       // %bb.0:
+; CHECK-NEXT:    ptrue p0.s
+; CHECK-NEXT:    ld1w { z0.s }, p0/z, [x0]
+; CHECK-NEXT:    zip2 z0.s, z0.s, z0.s
+; CHECK-NEXT:    st1w { z0.s }, p0, [x0]
+; CHECK-NEXT:    ret
+  %tmp1 = load volatile <8 x i32>, <8 x i32>* %a
+  %tmp2 = shufflevector <8 x i32> %tmp1, <8 x i32> undef, <8 x i32> <i32 4, i32 4, i32 5, i32 5, i32 6, i32 6, i32 7, i32 7>
+  store volatile <8 x i32> %tmp2, <8 x i32>* %a
+  ret void
+}
+
+define void @uzp_v32i8(<32 x i8>* %a, <32 x i8>* %b) #1 {
+; CHECK-LABEL: uzp_v32i8:
+; CHECK:       // %bb.0:
+; CHECK-NEXT:    ptrue p0.b
+; CHECK-NEXT:    ld1b { z0.b }, p0/z, [x0]
+; CHECK-NEXT:    ld1b { z1.b }, p0/z, [x1]
+; CHECK-NEXT:    uzp1 z2.b, z0.b, z1.b
+; CHECK-NEXT:    uzp2 z0.b, z0.b, z1.b
+; CHECK-NEXT:    add z0.b, p0/m, z0.b, z2.b
+; CHECK-NEXT:    st1b { z0.b }, p0, [x0]
+; CHECK-NEXT:    ret
+  %tmp1 = load <32 x i8>, <32 x i8>* %a
+  %tmp2 = load <32 x i8>, <32 x i8>* %b
+  %tmp3 = shufflevector <32 x i8> %tmp1, <32 x i8> %tmp2, <32 x i32> <i32 0, i32 2, i32 4, i32 6, i32 8, i32 10, i32 12, i32 14, i32 16, i32 18, i32 20, i32 22, i32 24, i32 26, i32 28, i32 30, i32 32, i32 34, i32 36, i32 38, i32 40, i32 42, i32 44, i32 46, i32 48, i32 50, i32 52, i32 54, i32 56, i32 58, i32 60, i32 62>
+  %tmp4 = shufflevector <32 x i8> %tmp1, <32 x i8> %tmp2, <32 x i32> <i32 1, i32 3, i32 5, i32 undef, i32 9, i32 11, i32 13, i32 undef, i32 undef, i32 19, i32 21, i32 23, i32 25, i32 27, i32 29, i32 31, i32 33, i32 35, i32 37, i32 39, i32 41, i32 43, i32 45, i32 47, i32 49, i32 51, i32 53, i32 55, i32 57, i32 59, i32 61, i32 63>
+  %tmp5 = add <32 x i8> %tmp3, %tmp4
+  store <32 x i8> %tmp5, <32 x i8>* %a
+  ret void
+}
+
+define void @uzp_v32i16(<32 x i16>* %a, <32 x i16>* %b) #1 {
+; CHECK-LABEL: uzp_v32i16:
+; CHECK:       // %bb.0:
+; CHECK-NEXT:    mov x8, #16
+; CHECK-NEXT:    ptrue p0.h
+; CHECK-NEXT:    ld1h { z0.h }, p0/z, [x0, x8, lsl #1]
+; CHECK-NEXT:    ld1h { z1.h }, p0/z, [x0]
+; CHECK-NEXT:    ld1h { z2.h }, p0/z, [x1, x8, lsl #1]
+; CHECK-NEXT:    ld1h { z3.h }, p0/z, [x1]
+; CHECK-NEXT:    uzp1 z5.h, z1.h, z0.h
+; CHECK-NEXT:    uzp2 z0.h, z1.h, z0.h
+; CHECK-NEXT:    add z0.h, p0/m, z0.h, z5.h
+; CHECK-NEXT:    uzp1 z4.h, z3.h, z2.h
+; CHECK-NEXT:    uzp2 z2.h, z3.h, z2.h
+; CHECK-NEXT:    movprfx z1, z4
+; CHECK-NEXT:    add z1.h, p0/m, z1.h, z2.h
+; CHECK-NEXT:    st1h { z1.h }, p0, [x0, x8, lsl #1]
+; CHECK-NEXT:    st1h { z0.h }, p0, [x0]
+; CHECK-NEXT:    ret
+  %tmp1 = load <32 x i16>, <32 x i16>* %a
+  %tmp2 = load <32 x i16>, <32 x i16>* %b
+  %tmp3 = shufflevector <32 x i16> %tmp1, <32 x i16> %tmp2, <32 x i32> <i32 0, i32 2, i32 4, i32 6, i32 8, i32 10, i32 12, i32 14, i32 16, i32 18, i32 20, i32 22, i32 24, i32 26, i32 28, i32 30, i32 32, i32 34, i32 36, i32 38, i32 40, i32 42, i32 44, i32 46, i32 48, i32 50, i32 52, i32 54, i32 56, i32 58, i32 60, i32 62>
+  %tmp4 = shufflevector <32 x i16> %tmp1, <32 x i16> %tmp2, <32 x i32> <i32 1, i32 3, i32 5, i32 undef, i32 9, i32 11, i32 13, i32 undef, i32 undef, i32 19, i32 21, i32 23, i32 25, i32 27, i32 29, i32 31, i32 33, i32 35, i32 37, i32 39, i32 41, i32 43, i32 45, i32 47, i32 49, i32 51, i32 53, i32 55, i32 57, i32 59, i32 61, i32 63>
+  %tmp5 = add <32 x i16> %tmp3, %tmp4
+  store <32 x i16> %tmp5, <32 x i16>* %a
+  ret void
+}
+
+define void @uzp_v16i16(<16 x i16>* %a, <16 x i16>* %b) #1 {
+; CHECK-LABEL: uzp_v16i16:
+; CHECK:       // %bb.0:
+; CHECK-NEXT:    ptrue p0.h
+; CHECK-NEXT:    ld1h { z0.h }, p0/z, [x0]
+; CHECK-NEXT:    ld1h { z1.h }, p0/z, [x1]
+; CHECK-NEXT:    uzp1 z2.h, z0.h, z1.h
+; CHECK-NEXT:    uzp2 z0.h, z0.h, z1.h
+; CHECK-NEXT:    add z0.h, p0/m, z0.h, z2.h
+; CHECK-NEXT:    st1h { z0.h }, p0, [x0]
+; CHECK-NEXT:    ret
+  %tmp1 = load <16 x i16>, <16 x i16>* %a
+  %tmp2 = load <16 x i16>, <16 x i16>* %b
+  %tmp3 = shufflevector <16 x i16> %tmp1, <16 x i16> %tmp2, <16 x i32> <i32 0, i32 2, i32 4, i32 6, i32 8, i32 10, i32 12, i32 14, i32 16, i32 18, i32 20, i32 22, i32 24, i32 26, i32 28, i32 30>
+  %tmp4 = shufflevector <16 x i16> %tmp1, <16 x i16> %tmp2, <16 x i32> <i32 1, i32 3, i32 5, i32 7, i32 9, i32 11, i32 13, i32 15, i32 17, i32 19, i32 21, i32 23, i32 25, i32 27, i32 29, i32 31>
+  %tmp5 = add <16 x i16> %tmp3, %tmp4
+  store <16 x i16> %tmp5, <16 x i16>* %a
+  ret void
+}
+
+define void @uzp_v8f32(<8 x float>* %a, <8 x float>* %b) #1 {
+; CHECK-LABEL: uzp_v8f32:
+; CHECK:       // %bb.0:
+; CHECK-NEXT:    ptrue p0.s
+; CHECK-NEXT:    ld1w { z0.s }, p0/z, [x0]
+; CHECK-NEXT:    ld1w { z1.s }, p0/z, [x1]
+; CHECK-NEXT:    uzp1 z2.s, z0.s, z1.s
+; CHECK-NEXT:    uzp2 z0.s, z0.s, z1.s
+; CHECK-NEXT:    fadd z0.s, z2.s, z0.s
+; CHECK-NEXT:    st1w { z0.s }, p0, [x0]
+; CHECK-NEXT:    ret
+  %tmp1 = load <8 x float>, <8 x float>* %a
+  %tmp2 = load <8 x float>, <8 x float>* %b
+  %tmp3 = shufflevector <8 x float> %tmp1, <8 x float> %tmp2, <8 x i32> <i32 0, i32 undef, i32 4, i32 6, i32 undef, i32 10, i32 12, i32 14>
+  %tmp4 = shufflevector <8 x float> %tmp1, <8 x float> %tmp2, <8 x i32> <i32 1, i32 undef, i32 5, i32 7, i32 9, i32 11, i32 undef, i32 undef>
+  %tmp5 = fadd <8 x float> %tmp3, %tmp4
+  store <8 x float> %tmp5, <8 x float>* %a
+  ret void
+}
+
+define void @uzp_v4i64(<4 x i64>* %a, <4 x i64>* %b) #1 {
+; CHECK-LABEL: uzp_v4i64:
+; CHECK:       // %bb.0:
+; CHECK-NEXT:    ptrue p0.d
+; CHECK-NEXT:    ld1d { z0.d }, p0/z, [x0]
+; CHECK-NEXT:    ld1d { z1.d }, p0/z, [x1]
+; CHECK-NEXT:    uzp1 z2.d, z0.d, z1.d
+; CHECK-NEXT:    uzp2 z0.d, z0.d, z1.d
+; CHECK-NEXT:    add z0.d, p0/m, z0.d, z2.d
+; CHECK-NEXT:    st1d { z0.d }, p0, [x0]
+; CHECK-NEXT:    ret
+  %tmp1 = load <4 x i64>, <4 x i64>* %a
+  %tmp2 = load <4 x i64>, <4 x i64>* %b
+  %tmp3 = shufflevector <4 x i64> %tmp1, <4 x i64> %tmp2, <4 x i32> <i32 0, i32 2, i32 4, i32 6>
+  %tmp4 = shufflevector <4 x i64> %tmp1, <4 x i64> %tmp2, <4 x i32> <i32 1, i32 3, i32 5, i32 7>
+  %tmp5 = add <4 x i64> %tmp3, %tmp4
+  store <4 x i64> %tmp5, <4 x i64>* %a
+  ret void
+}
+
+; Don't use SVE for 128-bit vectors
+define void @uzp_v8i16(<8 x i16>* %a, <8 x i16>* %b) #1 {
+; CHECK-LABEL: uzp_v8i16:
+; CHECK:       // %bb.0:
+; CHECK-NEXT:    ldr q0, [x0]
+; CHECK-NEXT:    ldr q1, [x1]
+; CHECK-NEXT:    uzp1 v2.8h, v0.8h, v1.8h
+; CHECK-NEXT:    uzp2 v0.8h, v0.8h, v1.8h
+; CHECK-NEXT:    add v0.8h, v2.8h, v0.8h
+; CHECK-NEXT:    str q0, [x0]
+; CHECK-NEXT:    ret
+  %tmp1 = load <8 x i16>, <8 x i16>* %a
+  %tmp2 = load <8 x i16>, <8 x i16>* %b
+  %tmp3 = shufflevector <8 x i16> %tmp1, <8 x i16> %tmp2, <8 x i32> <i32 0, i32 2, i32 4, i32 6, i32 8, i32 10, i32 12, i32 14>
+  %tmp4 = shufflevector <8 x i16> %tmp1, <8 x i16> %tmp2, <8 x i32> <i32 1, i32 3, i32 5, i32 7, i32 9, i32 11, i32 13, i32 15>
+  %tmp5 = add <8 x i16> %tmp3, %tmp4
+  store <8 x i16> %tmp5, <8 x i16>* %a
+  ret void
+}
+
+define void @uzp_v8i32_undef(<8 x i32>* %a) #1 {
+; CHECK-LABEL: uzp_v8i32_undef:
+; CHECK:       // %bb.0:
+; CHECK-NEXT:    ptrue p0.s
+; CHECK-NEXT:    ld1w { z0.s }, p0/z, [x0]
+; CHECK-NEXT:    uzp1 z1.s, z0.s, z0.s
+; CHECK-NEXT:    uzp2 z0.s, z0.s, z0.s
+; CHECK-NEXT:    add z0.s, p0/m, z0.s, z1.s
+; CHECK-NEXT:    st1w { z0.s }, p0, [x0]
+; CHECK-NEXT:    ret
+  %tmp1 = load <8 x i32>, <8 x i32>* %a
+  %tmp3 = shufflevector <8 x i32> %tmp1, <8 x i32> undef, <8 x i32> <i32 0, i32 2, i32 4, i32 6, i32 0, i32 2, i32 4, i32 6>
+  %tmp4 = shufflevector <8 x i32> %tmp1, <8 x i32> undef, <8 x i32> <i32 1, i32 3, i32 5, i32 7, i32 1, i32 3, i32 5, i32 7>
+  %tmp5 = add <8 x i32> %tmp3, %tmp4
+  store <8 x i32> %tmp5, <8 x i32>* %a
+  ret void
+}
+
+define void @zip_vscale2_4(<4 x double>* %a, <4 x double>* %b) #2 {
+; CHECK-LABEL: zip_vscale2_4:
+; CHECK:       // %bb.0:
+; CHECK-NEXT:    stp x29, x30, [sp, #-16]! // 16-byte Folded Spill
+; CHECK-NEXT:    sub x9, sp, #48
+; CHECK-NEXT:    mov x29, sp
+; CHECK-NEXT:    and sp, x9, #0xffffffffffffffe0
+; CHECK-NEXT:    .cfi_def_cfa w29, 16
+; CHECK-NEXT:    .cfi_offset w30, -8
+; CHECK-NEXT:    .cfi_offset w29, -16
+; CHECK-NEXT:    ptrue p0.d, vl4
+; CHECK-NEXT:    ld1d { z0.d }, p0/z, [x0]
+; CHECK-NEXT:    ld1d { z1.d }, p0/z, [x1]
+; CHECK-NEXT:    mov z2.d, z1.d[3]
+; CHECK-NEXT:    mov z3.d, z0.d[3]
+; CHECK-NEXT:    stp d3, d2, [sp, #16]
+; CHECK-NEXT:    mov z2.d, z1.d[2]
+; CHECK-NEXT:    mov z3.d, z0.d[2]
+; CHECK-NEXT:    zip1 z0.d, z0.d, z1.d
+; CHECK-NEXT:    stp d3, d2, [sp]
+; CHECK-NEXT:    ld1d { z2.d }, p0/z, [sp]
+; CHECK-NEXT:    fadd z0.d, p0/m, z0.d, z2.d
+; CHECK-NEXT:    st1d { z0.d }, p0, [x0]
+; CHECK-NEXT:    mov sp, x29
+; CHECK-NEXT:    ldp x29, x30, [sp], #16 // 16-byte Folded Reload
+; CHECK-NEXT:    ret
+  %tmp1 = load <4 x double>, <4 x double>* %a
+  %tmp2 = load <4 x double>, <4 x double>* %b
+  %tmp3 = shufflevector <4 x double> %tmp1, <4 x double> %tmp2, <4 x i32> <i32 0, i32 4, i32 1, i32 5>
+  %tmp4 = shufflevector <4 x double> %tmp1, <4 x double> %tmp2, <4 x i32> <i32 2, i32 6, i32 3, i32 7>
+  %tmp5 = fadd <4 x double> %tmp3, %tmp4
+  store <4 x double> %tmp5, <4 x double>* %a
+  ret void
+}
+
+attributes #0 = { "target-features"="+sve" }
+attributes #1 = { "target-features"="+sve" vscale_range(2,2) }
+attributes #2 = { "target-features"="+sve" vscale_range(2,4) }