[AMDGPU] Add basic support for extended i8 perm matching

[jrbyrnes]

Summary

Implement traversal algorithm to match trees to i8 vperms. For ors that can be combined into perms, we expect to see some pattern that combine four 8 bit operands (actually 16 bit operands, with only 8 nonzero bits) into two 16 bit operand, and combine thes two 16 bit operands via the or (after an zext, and ext-shift), The trees that do this type of combination are one of the two classes of trees relevant, and are matched in calculateByteProvider. The 8 bit operands used in this tree are typically produced via an AND op or a SRL op, and are the leaves of the trees in calculateByteProvider. The other relevant class of trees are those that map a leaf of calculateByteProvider to an ultimate source. This class of trees is matched in calculateSrcByte.

Through this recusive process, we track an Index (SrcIndex in calculateSrcByte) which is the byte of the current op that maps to the byte of the dest of the or we are currently mapping. For example, the 4th byte of the dest of SHL Src, 16 maps to the 2nd byte of Src. Through basic rules like this we can map src bytes to the dest byte of the or. Using this mapping we can create perm masks.

Much of the code for calculateByteProvider was borrowed from CodeGen/SelectionDAG/DAGCombiner.cpp (MatchLoadCombine). There are still many candidate trees that can be matched into perms that this patch does not attempt to. Those are saved for future iterations.

Depends on: https://reviews.llvm.org/D143018

[arsenm]

llvm/lib/Target/AMDGPU/SIISelLowering.cpp

9766

Can you keep this as a generic utility?

9881

Needs to move to std::optional

llvm/test/CodeGen/AMDGPU/permute_i8.ll

91

Test needs to use opaque pointers

[jrbyrnes]

Blacklist or->perm combine for certain users of or. Some ops (e.g. V_CVT_F32_UBYTE) are performed in bytewise manner. If the or has such a user, it is better to leave the dag in uncombined state since we will need to byte extract the combine.

[jrbyrnes]

Fix typo

[jrbyrnes]

Resolve remaining regressions + Separate out ByteProvider + std::optional

[jrbyrnes]

Rebase

[arsenm]

The ratio of test to code changes has me worried the test coverage is incomplete

llvm/lib/Target/AMDGPU/SIISelLowering.cpp

9766

By generic utility I mean in generic code and used by the load combine as well

9809

6 is the one true recursion depth limit

10088

Don't need a small vector, can just directly iterate an initializer list?

10091

Is the multiple use case covered in the tests?

10150

StartingIndex=

10152

!P

10190

Don't understand why you need the const_cast. Also, why isn't this an SDValue to begin with? Assuming output 0 can be risky

llvm/test/CodeGen/AMDGPU/permute_i8.ll

3

Drop -opaque-pointers (also direction doesn't make sense with the test contents)

7

Need to use typed pointers, also should prefer global loads to flat

[jrbyrnes]

llvm/lib/Target/AMDGPU/SIISelLowering.cpp

9766

Hey Matt -- thanks for comments. I think I don't fully understand this one -- I guess you didn't mean https://reviews.llvm.org/D143018 ? By generic, do you mean templated base class (perhaps in ADT) ?

[arsenm]

llvm/lib/Target/AMDGPU/SIISelLowering.cpp

9766

I didn't see that, I mean share code with the MatchLoadCombine that you mentioned. I didn't think about the details of that (something abstracter might be good since the same thing will need to be ported for GlobalISel)

[jrbyrnes]

Rebase + review comments.

[jrbyrnes]

In D142782#4095092, @arsenm wrote:

The ratio of test to code changes has me worried the test coverage is incomplete

This patch is intended to simply bring in the components necessary for i8 perm matching. Fitting / tuning it to be optimally useful in actual workloads is left to a future iteration. As such, the heuristics / conditions we use to apply the combine are very restrictive (e.g. no multi use operands in or, IsCombineVectorized heuristic, no support for 16 bit ors, etc). For this iteration, my primary concerns for testing were: true positives (i.e. testing accurate production of v_perm when we expect to), and false positive (correctness error / inefficient codegen). False negative (missed opportunity) are left to future iteration.

True positive coverage:
There are 4096 4xi8 shuffle_vector iterations. I tested and validated all permuations. The initial tests included covered all trees for these permutations.
There are 8192 4xi8 shuffle_vector iterations where 1 operand is undef. This iteration doesn't fully support these. Of these, about ~2k are lowered to v_perm by this iteration. I validated all of these.

False positive coverage:
lit tests
CK correctness tests
epsdb (to be run)

[jrbyrnes]

passes psdb

[jrbyrnes]

Add tests for gating conditions on attempting to v_perm combine + nits

[jrbyrnes]

Adding reviewers due to size of diff

[jrbyrnes]

Add opaque tests

[arsenm]

llvm/test/CodeGen/AMDGPU/permute_i8.ll

7

Opaque pointer tests are not additional, the tests need to be just converted. There are 0 remaining typed pointer AMDGPU tests

[jrbyrnes]

Convert test to opaque pointers

[foad]

llvm/test/CodeGen/AMDGPU/fast-unaligned-load-store.global.ll

42

These kind of changes look like regressions for some combination of code size / latency / sgpr pressure.

[arsenm]

llvm/test/CodeGen/AMDGPU/fast-unaligned-load-store.global.ll

64

Yes, this is worse. Should avoid cases that can use v_lshl_or_b32

[jrbyrnes]

Check that we are actually doing 8 bit extraction before lowering into v_perm.

We can determine (based on the potential perm mask and operands) if we need to insert any 8 bit extraction code.

For example, a perm mask of 0x05040100 suggests we will not need to extract any bits from the operands iff they have 16 bits of data (e.g. zext 16 load into 32 bit). In this case, we assume CodeGen will lower it well, and do not combine into v_perm. If, however, the operands are 32 bit, then we will need to insert mask code, so we do lower to v_perm.

As another example, if we have a mask of 0x05040201 then we will lower into v_perm for muiltiple reasons: 1. the 0x0201 portion of the mask implies a 32 bit operand, 2. the 0x0201 portion of the mask is not well formed, since it requires a shift instruction to address these bits.

Finally, if the mask and operands indicates we are just producing one of the ops, combine the tree into the op.

Feature resulted in changes -- optimally not lowering into v_perm -- in:
CodeGen/AMDGPU/combine-vload-extract.ll
CodeGen/AMDGPU/cvt_f32_ubyte.ll
CodeGen/AMDGPU/ds_read2.ll
CodeGen/AMDGPU/fast-unaligned-load-store.global.ll
CodeGen/AMDGPU/fast-unaligned-load-store.private.ll
CodeGen/AMDGPU/load-hi16.ll
CodeGen/AMDGPU/load-local.128.ll
CodeGen/AMDGPU/load-local.96.ll
CodeGen/AMDGPU/permute.ll

All 4096 permutation of <4 x i8> shufflevector produced desired result (including <i32 0, i32 1, i32 2, i32 3> and <i32 4, i32 5, i32 6, i32 7> which lower into correspond 32 bit operand).

[jrbyrnes]

Ping

[arsenm]

Should ByteProvider really be BitProvider?

llvm/lib/Target/AMDGPU/SIISelLowering.cpp

9854

Why 8? Usually 6 is the one true recursion depth limit

9864–9867

Do the RHS calls first and short circuit the second call if the first failed

9953

Do you really need the explicit std::optional?

10097

llvm::any_of?

[arsenm]

llvm/lib/Target/AMDGPU/SIISelLowering.cpp

10022

Need to make sure this is scalar

10029

Need to make sure this is scalar

[jrbyrnes]

Thanks @arsenm for taking another look.

Address review comments. Need to rerun psdb since short-circuiting LHS slightly changes algorithmic behavior.

llvm/lib/Target/AMDGPU/SIISelLowering.cpp

9854

Based on testing, I can lower depth, but we must accept depth 6 since this is the max tree depth across build vectors that should be lowered into v_perm. Relevant test is already included in permute_i8.ll. This depth may need to change in future iteration.

9953

Yes, it cannot infer std::optional, even after changing order.

[arsenm]

llvm/lib/Target/AMDGPU/SIISelLowering.cpp

9988

Braces

10029

This isn't checking for a scalar type?

10043

Can this be a separate function? It's a lambda that doesn't capture anything

[jrbyrnes]

Rebase + passed PSDB

[jrbyrnes]

Sorry @arsenm, I somehow missed your comments -- I'll address those.

[jrbyrnes]

Address comments

[jrbyrnes]

Cleanup checking in is16BitScalarOp + ping

[jrbyrnes]

Thanks for comments thus far -- any other concerns here?

[arsenm]

llvm/lib/Target/AMDGPU/SIISelLowering.cpp

9805

No reference

10102

Doesn't feel like this is the right way to express it. Should handle CopyToRegs at least too

10109

Don't need llvm::

10161

Don't these default initialize to nullopt?

[jrbyrnes]

Thanks @arsenm for the review.

I'll take another look at the v_perm candidate whitelist method in a subsequent patch which extends this work to capture more patterns.

[jrbyrnes]

Rebase

[arsenm]

In D142782#4433155, @jrbyrnes wrote:

Rebase

Why rebase a closed revision?

[jrbyrnes]

In D142782#4433237, @arsenm wrote:

In D142782#4433155, @jrbyrnes wrote:

Rebase

Why rebase a closed revision?

It was for posterity. Sorry for confusion.

llvm/lib/Target/AMDGPU/SIISelLowering.cpp

 
 // Check if a node selects whole bytes from its operand 0 starting at a byte
 // boundary while masking the rest. Returns select mask as in the v_perm_b32
 // or -1 if not succeeded.
 // Note byte select encoding:
 // value 0-3 selects corresponding source byte;
 // value 0xc selects zero;
 // value 0xff selects 0xff.
-static uint32_t getPermuteMask(SelectionDAG &DAG, SDValue V) {
+static uint32_t getPermuteMask(SDValue V) {
   assert(V.getValueSizeInBits() == 32);
 
   if (V.getNumOperands() != 2)
     return ~0;
 
   ConstantSDNode *N1 = dyn_cast<ConstantSDNode>(V.getOperand(1));
   if (!N1)
     return ~0;
 
   uint32_t C = N1->getZExtValue();
 
   switch (V.getOpcode()) {
   default:
     break;
   case ISD::AND:
-    if (uint32_t ConstMask = getConstantPermuteMask(C)) {
+    if (uint32_t ConstMask = getConstantPermuteMask(C))
       return (0x03020100 & ConstMask) | (0x0c0c0c0c & ~ConstMask);
-    }
     break;
 
   case ISD::OR:
-    if (uint32_t ConstMask = getConstantPermuteMask(C)) {
+    if (uint32_t ConstMask = getConstantPermuteMask(C))
       return (0x03020100 & ~ConstMask) | ConstMask;
-    }
     break;
 
   case ISD::SHL:
     if (C % 8)
       return ~0;
 
     return uint32_t((0x030201000c0c0c0cull << C) >> 32);
 
if (isBoolSGPR(RHS.getOperand(0)))
       return DAG.getSelect(SDLoc(N), MVT::i32, RHS.getOperand(0),
                            LHS, DAG.getConstant(0, SDLoc(N), MVT::i32));
   }
 
   // and (op x, c1), (op y, c2) -> perm x, y, permute_mask(c1, c2)
   const SIInstrInfo *TII = getSubtarget()->getInstrInfo();
   if (VT == MVT::i32 && LHS.hasOneUse() && RHS.hasOneUse() &&
       N->isDivergent() && TII->pseudoToMCOpcode(AMDGPU::V_PERM_B32_e64) != -1) {
-    uint32_t LHSMask = getPermuteMask(DAG, LHS);
-    uint32_t RHSMask = getPermuteMask(DAG, RHS);
+    uint32_t LHSMask = getPermuteMask(LHS);
+    uint32_t RHSMask = getPermuteMask(RHS);
     if (LHSMask != ~0u && RHSMask != ~0u) {
       // Canonicalize the expression in an attempt to have fewer unique masks
       // and therefore fewer registers used to hold the masks.
       if (LHSMask > RHSMask) {
         std::swap(LHSMask, RHSMask);
         std::swap(LHS, RHS);
       }
 
       // Select 0xc for each lane used from source operand. Zero has 0xc mask
       // set, 0xff have 0xff in the mask, actual lanes are in the 0-3 range.
       uint32_t LHSUsedLanes = ~(LHSMask & 0x0c0c0c0c) & 0x0c0c0c0c;
       uint32_t RHSUsedLanes = ~(RHSMask & 0x0c0c0c0c) & 0x0c0c0c0c;
 
       // Check of we need to combine values from two sources within a byte.
-      if (!(LHSUsedLanes & RHSUsedLanes) &&
+      if (!(LHSUsedLanes & RHSUsedLanes) ||
           // If we select high and lower word keep it for SDWA.
           // TODO: teach SDWA to work with v_perm_b32 and remove the check.
-          !(LHSUsedLanes == 0x0c0c0000 && RHSUsedLanes == 0x00000c0c)) {
+          (LHSUsedLanes == 0x0c0c0000 && RHSUsedLanes == 0x00000c0c)) {
         // Each byte in each mask is either selector mask 0-3, or has higher
         // bits set in either of masks, which can be 0xff for 0xff or 0x0c for
         // zero. If 0x0c is in either mask it shall always be 0x0c. Otherwise
         // mask which is not 0xff wins. By anding both masks we have a correct
         // result except that 0x0c shall be corrected to give 0x0c only.
         uint32_t Mask = LHSMask & RHSMask;
         for (unsigned I = 0; I < 32; I += 8) {
           uint32_t ByteSel = 0xff << I;
if (LHSMask != ~0u && RHSMask != ~0u) {
                            DAG.getConstant(Sel, DL, MVT::i32));
       }
     }
   }
 
   return SDValue();
 }
 
+/// Represents known origin of an individual byte in load combine pattern. The
+/// value of the byte is either constant zero or comes from memory.
+struct ByteProvider {
+  // For constant zero providers Load is set to nullptr. For memory providers
+  // Load represents the node which loads the byte from memory.
+  // ByteOffset is the offset of the byte in the value produced by the load.
+  const SDNode *Src = nullptr;
+  uint64_t DestByte = -1;
+  uint64_t SrcByte = -1;
+
+  ByteProvider() = default;
+
+  static ByteProvider getConstantZero() { return ByteProvider(nullptr, 0, 0); }
+
+  bool isConstantZero() const { return !Src; }
+  bool isMemory() const { return Src; }
+
+  bool isSameSrc(const ByteProvider &Other) const { return Other.Src == Src; }
+
+  bool operator==(const ByteProvider &Other) const {
+    return Other.Src == Src && Other.DestByte == DestByte &&
+           Other.SrcByte == SrcByte;
+  }
+
+  ByteProvider(const SDNode *Src, unsigned SrcByteArg, unsigned DestByte)
+      : Src(Src), DestByte(DestByte) {
+    assert(SrcByteArg <= 3);
+    SrcByte = SrcByteArg;
+  }
+};
+
+// A key component of v_perm is a mapping between byte position of the src
+// operands, and the byte position of the dest. To provide such, we need: 1. the
+// node that provides x byte of the dest of the OR, and 2. the byte of the node
+// used to provide that x byte. calculateByteProvider finds which node provides
+// a certain byte of the dest of the OR, and calculateSrcByte takes that node,
+// and finds an ultimate src and byte position For example: The supported
+// LoadCombine pattern for vector loads is as follows
+//                                t1
+//                                or
+//                      /                  \
+//                      t2                 t3
+//                     zext                shl
+//                      |                   |     \
+//                     t4                  t5     16
+//                     or                 anyext
+//                 /        \               |
+//                t6        t7             t8
+//               srl        shl             or
+//            /    |      /     \         /     \
+//           t9   t10    t11   t12      t13    t14
+//         trunc*  8    trunc*  8      and     and
+//           |            |          /    |     |    \
+//          t15          t16        t17  t18   t19   t20
+//                                trunc*  255   srl   -256
+//                                   |         /   \
+//                                  t15       t15  16
+//
+// *In this example, the truncs are from i32->i16
+//
+// calculateByteProvider would find t6, t7, t13, and t14 for bytes 0-3
+// respectively. calculateSrcByte would find (given node) -> ultimate src &
+// byteposition: t6 -> t15 & 1, t7 -> t16 & 0, t13 -> t15 & 0, t14 -> t15 & 3.
+// After finding the mapping, we can combine the tree into vperm t15, t16,
+// 0x05000407
+
+// Find the source and byte position from a node.
+// \p DestByte is the byte position of the dest of the or that the src
+// ultimately provides. \p SrcIndex is the byte of the src that maps to this
+// dest of the or byte. \p Depth tracks how many recursive iterations we have
+// performed.
+static const Optional<ByteProvider> calculateSrcByte(const SDValue *Op,
+                                                     uint64_t DestByte,
+                                                     uint64_t SrcIndex = 0,
+                                                     unsigned Depth = 0) {
+  // We may need to recursively traverse a series of SRLs
+  if (Depth >= 5)
+    return None;
+
+  switch (Op->getOpcode()) {
+  case ISD::TRUNCATE: {
+    if (Op->getOperand(0).getScalarValueSizeInBits() != 32)
+      return None;
+    return calculateSrcByte(&Op->getOperand(0), DestByte, SrcIndex, Depth + 1);
+  }
+
+  case ISD::SRL: {
+    auto ShiftOp = dyn_cast<ConstantSDNode>(Op->getOperand(1));
+    if (!ShiftOp)
+      return None;
+
+    uint64_t BitShift = ShiftOp->getZExtValue();
+
+    if (BitShift % 8 != 0)
+      return None;
+
+    SrcIndex += BitShift / 8;
+
+    return calculateSrcByte(&Op->getOperand(0), DestByte, SrcIndex, Depth + 1);
+  }
+
+  default: {
+    if (Op->getScalarValueSizeInBits() != 32)
+      return None;
+
+    return ByteProvider(Op->getNode(), SrcIndex, DestByte);
+  }
+  }
+  llvm_unreachable("fully handled switch");
+}
+
+// For a byte position in the result of an Or, traverse the tree and find the
+// node (and the byte of the node) which ultimately provides this {Or,
+// BytePosition}. \p Op is the operand we are currently examining. \p Index is
+// the byte position of the Op that corresponds with the originally requested
+// byte of the Or \p Depth tracks how many recursive iterations we have
+// performed. \p StartingIndex is the originally requested byte of the Or
+static const Optional<ByteProvider>
+calculateByteProvider(SDValue Op, unsigned Index, unsigned Depth,
+                      unsigned StartingIndex = 0) {
+  // Finding Src tree of RHS of or typically requires at least 1 additional
+  // depth
+  if (Depth >= 8)
+    return None;
+
+  unsigned BitWidth = Op.getScalarValueSizeInBits();
+  if (BitWidth % 8 != 0)
+    return None;
+  assert(Index < BitWidth / 8 && "invalid index requested");
+
+  switch (Op.getOpcode()) {
+  case ISD::OR: {
+    auto LHS = calculateByteProvider(Op->getOperand(0), Index, Depth + 1,
+                                     StartingIndex);
+    auto RHS = calculateByteProvider(Op->getOperand(1), Index, Depth + 1,
+                                     StartingIndex);
+    // A well formed Or will only have nonzero bytes for one operand
+    if (LHS && RHS && !LHS->isConstantZero() && !RHS->isConstantZero())
+      return None;
+    if (!LHS || LHS->isConstantZero())
+      return RHS;
+    if (!RHS || RHS->isConstantZero())
+      return LHS;
+    return None;
+  }
+
+  case ISD::AND: {
+    auto BitMaskOp = dyn_cast<ConstantSDNode>(Op->getOperand(1));
+    if (!BitMaskOp)
+      return None;
+
+    uint32_t BitMask = BitMaskOp->getZExtValue();
+    // Bits we expect for our StartingIndex
+    uint32_t IndexMask = 0xFF << (Index * 8);
+
+    if ((IndexMask & BitMask) != IndexMask) {
+      // If the result of the and partially provides the byte, then it
+      // is not well formatted
+      if (IndexMask & BitMask)
+        return None;
+      return ByteProvider::getConstantZero();
+    }
+
+    return calculateSrcByte(&Op->getOperand(0), StartingIndex, Index);
+  }
+
+  case ISD::SRL: {
+    auto ShiftOp = dyn_cast<ConstantSDNode>(Op->getOperand(1));
+    if (!ShiftOp)
+      return None;
+
+    uint64_t BitShift = ShiftOp->getZExtValue();
+    if (BitShift % 8)
+      return None;
+
+    auto BitsProvided = Op.getScalarValueSizeInBits();
+    if (BitsProvided % 8 != 0)
+      return None;
+
+    uint64_t BytesProvided = BitsProvided / 8;
+    uint64_t ByteShift = BitShift / 8;
+    // The dest of shift will have good [0 : (BytesProvided - ByteShift)] bytes.
+    // If the byte we are trying to provide (as tracked by index) falls in this
+    // range, then the SRL provides the byte. The byte of interest of the src of
+    // the SRL is Index + ByteShift
+    return BytesProvided - ByteShift > Index
+               ? calculateSrcByte(&Op->getOperand(0), StartingIndex,
+                                  Index + ByteShift)
+               : ByteProvider::getConstantZero();
+  }
+
+  case ISD::SHL: {
+    auto ShiftOp = dyn_cast<ConstantSDNode>(Op->getOperand(1));
+    if (!ShiftOp)
+      return None;
+
+    uint64_t BitShift = ShiftOp->getZExtValue();
+    if (BitShift % 8 != 0)
+      return None;
+    uint64_t ByteShift = BitShift / 8;
+
+    // If we are shifting by an amount greater than (or equal to)
+    // the index we are trying to provide, then it provides 0s. If not,
+    // then this bytes are not definitively 0s, and the corresponding byte
+    // of interest is Index - ByteShift of the src
+    return Index < ByteShift
+               ? ByteProvider::getConstantZero()
+               : calculateByteProvider(Op->getOperand(0), Index - ByteShift,
+                                       Depth + 1, StartingIndex);
+  }
+  case ISD::ANY_EXTEND:
+  case ISD::SIGN_EXTEND:
+  case ISD::ZERO_EXTEND: {
+    SDValue NarrowOp = Op->getOperand(0);
+    unsigned NarrowBitWidth = NarrowOp.getScalarValueSizeInBits();
+    if (NarrowBitWidth % 8 != 0)
+      return None;
+    uint64_t NarrowByteWidth = NarrowBitWidth / 8;
+
+    if (Index >= NarrowByteWidth)
+      return Op.getOpcode() == ISD::ZERO_EXTEND
+                 ? Optional<ByteProvider>(ByteProvider::getConstantZero())
+                 : None;
+    return calculateByteProvider(NarrowOp, Index, Depth + 1, StartingIndex);
+  }
+
+  case ISD::TRUNCATE: {
+    unsigned NarrowBitWidth = Op.getScalarValueSizeInBits();
+    if (NarrowBitWidth % 8 != 0)
+      return None;
+    uint64_t NarrowByteWidth = NarrowBitWidth / 8;
+
+    if (NarrowByteWidth >= Index) {
+      return calculateByteProvider(Op->getOperand(0), Index, Depth + 1,
+                                   StartingIndex);
+    }
+
+    return None;
+  }
+
+  case ISD::LOAD: {
+    auto L = cast<LoadSDNode>(Op.getNode());
+    unsigned NarrowBitWidth = L->getMemoryVT().getSizeInBits();
+    if (NarrowBitWidth % 8 != 0)
+      return None;
+    uint64_t NarrowByteWidth = NarrowBitWidth / 8;
+
+    // If the width of the load does not reach byte we are trying to provide for
+    // and it is not a ZEXTLOAD, then the load does not provide for the byte in
+    // question
+    if (Index >= NarrowByteWidth)
+      return L->getExtensionType() == ISD::ZEXTLOAD
+                 ? Optional<ByteProvider>(ByteProvider::getConstantZero())
+                 : None;
+
+    if (NarrowByteWidth > Index) {
+      return calculateSrcByte(const_cast<const SDValue *>(&Op), StartingIndex,
+                              Index);
+    }
+
+    return None;
+  }
+
+  default: {
+    return None;
+  }
+  }
+
+  llvm_unreachable("fully handled switch");
+}
+
 SDValue SITargetLowering::performOrCombine(SDNode *N,
                                            DAGCombinerInfo &DCI) const {
   SelectionDAG &DAG = DCI.DAG;
   SDValue LHS = N->getOperand(0);
   SDValue RHS = N->getOperand(1);
 
   EVT VT = N->getValueType(0);
   if (VT == MVT::i1) {
     // or (fp_class x, c1), (fp_class x, c2) -> fp_class x, (c1 | c2)
     if (LHS.getOpcode() == AMDGPUISD::FP_CLASS &&
         RHS.getOpcode() == AMDGPUISD::FP_CLASS) {
       SDValue Src = LHS.getOperand(0);
       if (Src != RHS.getOperand(0))
         return SDValue();
 
       const ConstantSDNode *CLHS = dyn_cast<ConstantSDNode>(LHS.getOperand(1));
       const ConstantSDNode *CRHS = dyn_cast<ConstantSDNode>(RHS.getOperand(1));
if (isa<ConstantSDNode>(RHS) && LHS.hasOneUse() &&
     return DAG.getNode(AMDGPUISD::PERM, DL, MVT::i32, LHS.getOperand(0),
                        LHS.getOperand(1), DAG.getConstant(Sel, DL, MVT::i32));
   }
 
   // or (op x, c1), (op y, c2) -> perm x, y, permute_mask(c1, c2)
   const SIInstrInfo *TII = getSubtarget()->getInstrInfo();
   if (VT == MVT::i32 && LHS.hasOneUse() && RHS.hasOneUse() &&
       N->isDivergent() && TII->pseudoToMCOpcode(AMDGPU::V_PERM_B32_e64) != -1) {
-    uint32_t LHSMask = getPermuteMask(DAG, LHS);
-    uint32_t RHSMask = getPermuteMask(DAG, RHS);
+
+    // If the users of the or are a BytewiseOp, then the result of a combnine will
+    // be extracted. We should simply not combine.
+    SmallVector<unsigned, 4> BytewiseOps = {ISD::SINT_TO_FP, ISD::UINT_TO_FP};
+
+    bool IsCombineExtracted = false;
+    for (auto OrUse : N->uses()) {
+      // Only special case bitcast to vectors
+      if (OrUse->getOpcode() != ISD::BITCAST || !OrUse->getValueType(0).isVector()) {
+        continue;
+      }
+
+      for (auto VUse : OrUse->uses()) {
+        for (auto BytewiseOp : BytewiseOps)
+        if (VUse->getOpcode() == BytewiseOp) {
+          IsCombineExtracted = true;
+          break;
+        }
+      }
+    }
+
+    if (IsCombineExtracted) return SDValue();
+
+    uint32_t LHSMask = getPermuteMask(LHS);
+    uint32_t RHSMask = getPermuteMask(RHS);
+
     if (LHSMask != ~0u && RHSMask != ~0u) {
       // Canonicalize the expression in an attempt to have fewer unique masks
       // and therefore fewer registers used to hold the masks.
       if (LHSMask > RHSMask) {
         std::swap(LHSMask, RHSMask);
         std::swap(LHS, RHS);
       }
 
if (LHSMask != ~0u && RHSMask != ~0u) {
         // Kill zero bytes selected by other mask. Zero value is 0xc.
         LHSMask &= ~RHSUsedLanes;
         RHSMask &= ~LHSUsedLanes;
         // Add 4 to each active LHS lane
         LHSMask |= LHSUsedLanes & 0x04040404;
         // Combine masks
         uint32_t Sel = LHSMask | RHSMask;
         SDLoc DL(N);
-
         return DAG.getNode(AMDGPUISD::PERM, DL, MVT::i32,
                            LHS.getOperand(0), RHS.getOperand(0),
                            DAG.getConstant(Sel, DL, MVT::i32));
       }
     }
+    if (LHSMask == ~0u || RHSMask == ~0u) {
+      SmallVector<ByteProvider, 8> PermNodes;
+
+      // VT is known to be MVT::i32, so we need to provide 4 bytes.
+      assert(VT == MVT::i32);
+      for (int i = 0; i < 4; i++) {
+        // Find the ByteProvider that provides the ith byte of the result of OR
+        Optional<ByteProvider> P =
+            calculateByteProvider(SDValue(N, 0), i, 0, /*StartingIndex*/ i);
+        // TODO support constantZero
+        if (!P.has_value() || P->isConstantZero())
+          return SDValue();
+
+        PermNodes.push_back(*P);
+      }
+      if (PermNodes.size() != 4)
+        return SDValue();
+
+      int FirstSrc = 0;
+      int SecondSrc = -1;
+      uint64_t permMask = 0x00000000;
+      for (size_t i = 0; i < PermNodes.size(); i++) {
+        auto PermOp = PermNodes[i];
+        // Since the mask is applied to Src1:Src2, Src1 bytes must be offset
+        // by sizeof(Src2) = 4
+        int SrcByteAdjust = 4;
+
+        if (!PermOp.isSameSrc(PermNodes[FirstSrc])) {
+          if (SecondSrc != -1) {
+            if (!PermOp.isSameSrc(PermNodes[SecondSrc])) {
+              return SDValue();
+            }
+          }
+          // Set the index of the second distinct Src node
+          SecondSrc = i;
+          assert(PermNodes[SecondSrc].Src->getValueType(0).getSizeInBits() ==
+                 32);
+          SrcByteAdjust = 0;
+        }
+        assert(PermOp.SrcByte + SrcByteAdjust < 8);
+        // 0th PermNode is MSB in PermMask
+        permMask |= (PermOp.SrcByte + SrcByteAdjust) << (24 - (i * 8));
+      }
+
+      SDLoc DL(N);
+
+      return DAG.getNode(
+          AMDGPUISD::PERM, DL, MVT::i32,
+          SDValue(const_cast<SDNode *>(PermNodes[FirstSrc].Src), 0),
+          SecondSrc == -1
+              ? DAG.getConstant(0, DL, MVT::i32)
+              : SDValue(const_cast<SDNode *>(PermNodes[SecondSrc].Src), 0),
+          DAG.getConstant(permMask, DL, MVT::i32));
+    }
   }
 
   if (VT != MVT::i64 || DCI.isBeforeLegalizeOps())
     return SDValue();
 
   // TODO: This could be a generic combine with a predicate for extracting the
   // high half of an integer being free.
 

llvm/test/CodeGen/AMDGPU/combine-vload-extract.ll

entry:
   store i32 %insert3, i32* %out
   ret void
 }
 
 define amdgpu_kernel void @vectorLoadShuffle(<4 x i8>* %in, i32* %out) {
 ; GCN-LABEL: vectorLoadShuffle:
 ; GCN:       ; %bb.0: ; %entry
 ; GCN-NEXT:    s_load_dwordx4 s[0:3], s[0:1], 0x24
+; GCN-NEXT:    v_mov_b32_e32 v3, 0x4060507
 ; GCN-NEXT:    s_waitcnt lgkmcnt(0)
 ; GCN-NEXT:    v_mov_b32_e32 v0, s0
 ; GCN-NEXT:    v_mov_b32_e32 v1, s1
 ; GCN-NEXT:    flat_load_dword v2, v[0:1]
-; GCN-NEXT:    s_mov_b32 s0, 0x6050400
 ; GCN-NEXT:    v_mov_b32_e32 v0, s2
 ; GCN-NEXT:    v_mov_b32_e32 v1, s3
 ; GCN-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; GCN-NEXT:    v_bfe_u32 v3, v2, 16, 8
-; GCN-NEXT:    v_lshlrev_b32_e32 v4, 8, v2
-; GCN-NEXT:    v_perm_b32 v3, v3, v2, s0
-; GCN-NEXT:    v_and_b32_e32 v4, 0xff0000, v4
-; GCN-NEXT:    v_and_b32_e32 v2, 0xff000000, v2
-; GCN-NEXT:    v_or3_b32 v2, v3, v4, v2
+; GCN-NEXT:    v_perm_b32 v2, v2, 0, v3
 ; GCN-NEXT:    flat_store_dword v[0:1], v2
 ; GCN-NEXT:    s_endpgm
 entry:
   %0 = load <4 x i8>, <4 x i8>* %in, align 4
   %1 = extractelement <4 x i8> %0, i32 0
   %2 = extractelement <4 x i8> %0, i32 1
   %3 = extractelement <4 x i8> %0, i32 2
   %4 = extractelement <4 x i8> %0, i32 3

 }
 
 define i32 @load_2xi16_noncombine(i16 addrspace(1)* %p) #0 {
 ; GCN-LABEL: load_2xi16_noncombine:
 ; GCN:       ; %bb.0:
 ; GCN-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
 ; GCN-NEXT:    global_load_ushort v2, v[0:1], off
 ; GCN-NEXT:    global_load_ushort v3, v[0:1], off offset:4
+; GCN-NEXT:    s_mov_b32 s4, 0x4050001
 ; GCN-NEXT:    s_waitcnt vmcnt(0)
-; GCN-NEXT:    v_lshl_or_b32 v0, v3, 16, v2
+; GCN-NEXT:    v_perm_b32 v0, v2, v3, s4
 ; GCN-NEXT:    s_setpc_b64 s[30:31]
   %gep.p = getelementptr i16, i16 addrspace(1)* %p, i32 2
   %p.0 = load i16, i16 addrspace(1)* %p, align 4
   %p.1 = load i16, i16 addrspace(1)* %gep.p, align 4
   %zext.0 = zext i16 %p.0 to i32
   %zext.1 = zext i16 %p.1 to i32
   %shl.1 = shl i32 %zext.1, 16
   %or = or i32 %zext.0, %shl.1

 }
 
 define i64 @load_3xi16_noncombine(i16 addrspace(1)* %p) #0 {
 ; GCN-LABEL: load_3xi16_noncombine:
 ; GCN:       ; %bb.0:
 ; GCN-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
 ; GCN-NEXT:    global_load_ushort v2, v[0:1], off
 ; GCN-NEXT:    global_load_dword v3, v[0:1], off offset:4
-; GCN-NEXT:    s_mov_b32 s4, 0xffff0000
+; GCN-NEXT:    s_mov_b32 s4, 0x4050203
 ; GCN-NEXT:    s_waitcnt vmcnt(0)
-; GCN-NEXT:    v_and_or_b32 v0, v3, s4, v2
+; GCN-NEXT:    v_perm_b32 v0, v2, v3, s4
 ; GCN-NEXT:    v_and_b32_e32 v1, 0xffff, v3
 ; GCN-NEXT:    s_setpc_b64 s[30:31]
   %gep.p = getelementptr i16, i16 addrspace(1)* %p, i32 3
   %gep.2p = getelementptr i16, i16 addrspace(1)* %p, i32 2
   %p.0 = load i16, i16 addrspace(1)* %p, align 4
   %p.1 = load i16, i16 addrspace(1)* %gep.p, align 4
   %p.2 = load i16, i16 addrspace(1)* %gep.2p, align 4
   %zext.0 = zext i16 %p.0 to i64

llvm/test/CodeGen/AMDGPU/cvt_f32_ubyte.ll

 ; SI-NEXT:    v_cvt_f32_ubyte0_e32 v0, v0
 ; SI-NEXT:    buffer_store_dwordx4 v[0:3], off, s[4:7], 0
 ; SI-NEXT:    s_endpgm
 ;
 ; VI-LABEL: v4i8_zext_v4i32_to_v4f32:
 ; VI:       ; %bb.0:
 ; VI-NEXT:    s_load_dwordx4 s[0:3], s[0:1], 0x24
 ; VI-NEXT:    v_lshlrev_b32_e32 v0, 2, v0
+; VI-NEXT:    s_mov_b32 s4, 0x4050001
 ; VI-NEXT:    s_waitcnt lgkmcnt(0)
 ; VI-NEXT:    v_mov_b32_e32 v1, s3
 ; VI-NEXT:    v_add_u32_e32 v0, vcc, s2, v0
 ; VI-NEXT:    v_addc_u32_e32 v1, vcc, 0, v1, vcc
+; VI-NEXT:    v_add_u32_e32 v2, vcc, 3, v0
+; VI-NEXT:    v_addc_u32_e32 v3, vcc, 0, v1, vcc
+; VI-NEXT:    flat_load_ubyte v6, v[2:3]
 ; VI-NEXT:    v_add_u32_e32 v2, vcc, 2, v0
 ; VI-NEXT:    v_addc_u32_e32 v3, vcc, 0, v1, vcc
-; VI-NEXT:    v_add_u32_e32 v4, vcc, 3, v0
+; VI-NEXT:    v_add_u32_e32 v4, vcc, 1, v0
 ; VI-NEXT:    v_addc_u32_e32 v5, vcc, 0, v1, vcc
+; VI-NEXT:    flat_load_ubyte v4, v[4:5]
 ; VI-NEXT:    flat_load_ubyte v2, v[2:3]
-; VI-NEXT:    flat_load_ubyte v3, v[4:5]
-; VI-NEXT:    flat_load_ubyte v4, v[0:1]
-; VI-NEXT:    v_add_u32_e32 v0, vcc, 1, v0
-; VI-NEXT:    v_addc_u32_e32 v1, vcc, 0, v1, vcc
-; VI-NEXT:    flat_load_ubyte v1, v[0:1]
+; VI-NEXT:    flat_load_ubyte v0, v[0:1]
 ; VI-NEXT:    s_mov_b32 s3, 0xf000
 ; VI-NEXT:    s_mov_b32 s2, -1
 ; VI-NEXT:    s_waitcnt vmcnt(3)
-; VI-NEXT:    v_cvt_f32_ubyte0_e32 v2, v2
+; VI-NEXT:    v_lshlrev_b32_e32 v1, 8, v6
 ; VI-NEXT:    s_waitcnt vmcnt(2)
-; VI-NEXT:    v_cvt_f32_ubyte0_e32 v3, v3
+; VI-NEXT:    v_lshlrev_b32_e32 v3, 8, v4
 ; VI-NEXT:    s_waitcnt vmcnt(1)
-; VI-NEXT:    v_cvt_f32_ubyte0_e32 v0, v4
+; VI-NEXT:    v_or_b32_e32 v1, v1, v2
 ; VI-NEXT:    s_waitcnt vmcnt(0)
-; VI-NEXT:    v_cvt_f32_ubyte0_e32 v1, v1
+; VI-NEXT:    v_or_b32_e32 v0, v3, v0
+; VI-NEXT:    v_perm_b32 v0, v0, v1, s4
+; VI-NEXT:    v_cvt_f32_ubyte3_e32 v3, v0
+; VI-NEXT:    v_cvt_f32_ubyte2_e32 v2, v0
+; VI-NEXT:    v_cvt_f32_ubyte1_e32 v1, v0
+; VI-NEXT:    v_cvt_f32_ubyte0_e32 v0, v0
 ; VI-NEXT:    buffer_store_dwordx4 v[0:3], off, s[0:3], 0
 ; VI-NEXT:    s_endpgm
 ;
 ; GFX10-LABEL: v4i8_zext_v4i32_to_v4f32:
 ; GFX10:       ; %bb.0:
 ; GFX10-NEXT:    s_load_dwordx4 s[0:3], s[0:1], 0x24
 ; GFX10-NEXT:    v_lshlrev_b32_e32 v0, 2, v0
-; GFX10-NEXT:    v_mov_b32_e32 v6, 0
 ; GFX10-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX10-NEXT:    s_clause 0x3
 ; GFX10-NEXT:    global_load_ubyte v1, v0, s[2:3] offset:3
-; GFX10-NEXT:    global_load_ubyte v2, v0, s[2:3] offset:2
+; GFX10-NEXT:    global_load_ubyte v2, v0, s[2:3]
+; GFX10-NEXT:    global_load_ubyte v3, v0, s[2:3] offset:2
 ; GFX10-NEXT:    global_load_ubyte v4, v0, s[2:3] offset:1
-; GFX10-NEXT:    global_load_ubyte v5, v0, s[2:3]
-; GFX10-NEXT:    s_waitcnt vmcnt(3)
-; GFX10-NEXT:    v_cvt_f32_ubyte0_e32 v3, v1
-; GFX10-NEXT:    s_waitcnt vmcnt(2)
-; GFX10-NEXT:    v_cvt_f32_ubyte0_e32 v2, v2
 ; GFX10-NEXT:    s_waitcnt vmcnt(1)
-; GFX10-NEXT:    v_cvt_f32_ubyte0_e32 v1, v4
+; GFX10-NEXT:    v_lshl_or_b32 v0, v1, 8, v3
 ; GFX10-NEXT:    s_waitcnt vmcnt(0)
-; GFX10-NEXT:    v_cvt_f32_ubyte0_e32 v0, v5
-; GFX10-NEXT:    global_store_dwordx4 v6, v[0:3], s[0:1]
+; GFX10-NEXT:    v_lshl_or_b32 v1, v4, 8, v2
+; GFX10-NEXT:    v_mov_b32_e32 v4, 0
+; GFX10-NEXT:    v_perm_b32 v0, v1, v0, 0x4050001
+; GFX10-NEXT:    v_cvt_f32_ubyte3_e32 v3, v0
+; GFX10-NEXT:    v_cvt_f32_ubyte2_e32 v2, v0
+; GFX10-NEXT:    v_cvt_f32_ubyte1_e32 v1, v0
+; GFX10-NEXT:    v_cvt_f32_ubyte0_e32 v0, v0
+; GFX10-NEXT:    global_store_dwordx4 v4, v[0:3], s[0:1]
 ; GFX10-NEXT:    s_endpgm
 ;
 ; GFX9-LABEL: v4i8_zext_v4i32_to_v4f32:
 ; GFX9:       ; %bb.0:
 ; GFX9-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x24
 ; GFX9-NEXT:    v_lshlrev_b32_e32 v0, 2, v0
-; GFX9-NEXT:    v_mov_b32_e32 v6, 0
+; GFX9-NEXT:    v_mov_b32_e32 v5, 0
 ; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX9-NEXT:    global_load_ubyte v1, v0, s[2:3] offset:3
-; GFX9-NEXT:    global_load_ubyte v2, v0, s[2:3] offset:2
+; GFX9-NEXT:    global_load_ubyte v2, v0, s[2:3]
+; GFX9-NEXT:    global_load_ubyte v3, v0, s[2:3] offset:2
 ; GFX9-NEXT:    global_load_ubyte v4, v0, s[2:3] offset:1
-; GFX9-NEXT:    global_load_ubyte v5, v0, s[2:3]
-; GFX9-NEXT:    s_waitcnt vmcnt(3)
-; GFX9-NEXT:    v_cvt_f32_ubyte0_e32 v3, v1
-; GFX9-NEXT:    s_waitcnt vmcnt(2)
-; GFX9-NEXT:    v_cvt_f32_ubyte0_e32 v2, v2
+; GFX9-NEXT:    s_mov_b32 s2, 0x4050001
 ; GFX9-NEXT:    s_waitcnt vmcnt(1)
-; GFX9-NEXT:    v_cvt_f32_ubyte0_e32 v1, v4
+; GFX9-NEXT:    v_lshl_or_b32 v0, v1, 8, v3
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
-; GFX9-NEXT:    v_cvt_f32_ubyte0_e32 v0, v5
-; GFX9-NEXT:    global_store_dwordx4 v6, v[0:3], s[0:1]
+; GFX9-NEXT:    v_lshl_or_b32 v1, v4, 8, v2
+; GFX9-NEXT:    v_perm_b32 v0, v1, v0, s2
+; GFX9-NEXT:    v_cvt_f32_ubyte3_e32 v3, v0
+; GFX9-NEXT:    v_cvt_f32_ubyte2_e32 v2, v0
+; GFX9-NEXT:    v_cvt_f32_ubyte1_e32 v1, v0
+; GFX9-NEXT:    v_cvt_f32_ubyte0_e32 v0, v0
+; GFX9-NEXT:    global_store_dwordx4 v5, v[0:3], s[0:1]
 ; GFX9-NEXT:    s_endpgm
 ;
 ; GFX11-LABEL: v4i8_zext_v4i32_to_v4f32:
 ; GFX11:       ; %bb.0:
 ; GFX11-NEXT:    s_load_b128 s[0:3], s[0:1], 0x24
-; GFX11-NEXT:    v_dual_mov_b32 v5, 0 :: v_dual_lshlrev_b32 v0, 2, v0
+; GFX11-NEXT:    v_lshlrev_b32_e32 v0, 2, v0
+; GFX11-NEXT:    v_mov_b32_e32 v4, 0
 ; GFX11-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX11-NEXT:    s_clause 0x3
 ; GFX11-NEXT:    global_load_u8 v1, v0, s[2:3] offset:3
-; GFX11-NEXT:    global_load_u8 v2, v0, s[2:3] offset:2
-; GFX11-NEXT:    global_load_u8 v4, v0, s[2:3] offset:1
-; GFX11-NEXT:    global_load_u8 v0, v0, s[2:3]
-; GFX11-NEXT:    s_waitcnt vmcnt(3)
-; GFX11-NEXT:    v_cvt_f32_ubyte0_e32 v3, v1
-; GFX11-NEXT:    s_waitcnt vmcnt(2)
-; GFX11-NEXT:    v_cvt_f32_ubyte0_e32 v2, v2
+; GFX11-NEXT:    global_load_u8 v2, v0, s[2:3]
+; GFX11-NEXT:    global_load_u8 v3, v0, s[2:3] offset:2
+; GFX11-NEXT:    global_load_u8 v0, v0, s[2:3] offset:1
 ; GFX11-NEXT:    s_waitcnt vmcnt(1)
-; GFX11-NEXT:    v_cvt_f32_ubyte0_e32 v1, v4
+; GFX11-NEXT:    v_lshl_or_b32 v1, v1, 8, v3
 ; GFX11-NEXT:    s_waitcnt vmcnt(0)
+; GFX11-NEXT:    v_lshl_or_b32 v0, v0, 8, v2
+; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT:    v_perm_b32 v0, v0, v1, 0x4050001
+; GFX11-NEXT:    v_cvt_f32_ubyte3_e32 v3, v0
+; GFX11-NEXT:    v_cvt_f32_ubyte2_e32 v2, v0
+; GFX11-NEXT:    v_cvt_f32_ubyte1_e32 v1, v0
 ; GFX11-NEXT:    v_cvt_f32_ubyte0_e32 v0, v0
-; GFX11-NEXT:    global_store_b128 v5, v[0:3], s[0:1]
+; GFX11-NEXT:    global_store_b128 v4, v[0:3], s[0:1]
 ; GFX11-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-NEXT:    s_endpgm
   %tid = call i32 @llvm.amdgcn.workitem.id.x()
   %gep = getelementptr <4 x i8>, <4 x i8> addrspace(1)* %in, i32 %tid
   %load = load <4 x i8>, <4 x i8> addrspace(1)* %gep, align 1
   %ext = zext <4 x i8> %load to <4 x i32>
   %cvt = uitofp <4 x i32> %ext to <4 x float>
   store <4 x float> %cvt, <4 x float> addrspace(1)* %out, align 16

 ; VI-NEXT:    v_lshlrev_b32_e32 v0, 2, v0
 ; VI-NEXT:    s_mov_b32 s7, 0xf000
 ; VI-NEXT:    s_mov_b32 s6, -1
 ; VI-NEXT:    s_waitcnt lgkmcnt(0)
 ; VI-NEXT:    v_mov_b32_e32 v1, s1
 ; VI-NEXT:    v_add_u32_e32 v0, vcc, s0, v0
 ; VI-NEXT:    v_addc_u32_e32 v1, vcc, 0, v1, vcc
 ; VI-NEXT:    flat_load_dword v0, v[0:1]
+; VI-NEXT:    v_mov_b32_e32 v1, 0x4050607
 ; VI-NEXT:    s_mov_b32 s4, s2
 ; VI-NEXT:    s_mov_b32 s5, s3
 ; VI-NEXT:    s_waitcnt vmcnt(0)
-; VI-NEXT:    v_or_b32_e32 v0, 0x80000001, v0
+; VI-NEXT:    v_perm_b32 v0, v0, 0, v1
 ; VI-NEXT:    v_cvt_f32_ubyte0_e32 v1, v0
 ; VI-NEXT:    v_add_f32_e32 v0, v0, v1
 ; VI-NEXT:    buffer_store_dword v0, off, s[4:7], 0
 ; VI-NEXT:    s_endpgm
 ;
 ; GFX10-LABEL: cvt_ubyte0_or_multiuse:
 ; GFX10:       ; %bb.0: ; %bb
 ; GFX10-NEXT:    s_load_dwordx4 s[0:3], s[0:1], 0x24
 ; GFX10-NEXT:    v_lshlrev_b32_e32 v0, 2, v0
 ; GFX10-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX10-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX10-NEXT:    global_load_dword v0, v0, s[0:1]
 ; GFX10-NEXT:    s_waitcnt vmcnt(0)
-; GFX10-NEXT:    v_or_b32_e32 v0, 0x80000001, v0
+; GFX10-NEXT:    v_perm_b32 v0, v0, 0, 0x4050607
 ; GFX10-NEXT:    v_cvt_f32_ubyte0_e32 v1, v0
 ; GFX10-NEXT:    v_add_f32_e32 v0, v0, v1
 ; GFX10-NEXT:    global_store_dword v2, v0, s[2:3]
 ; GFX10-NEXT:    s_endpgm
 ;
 ; GFX9-LABEL: cvt_ubyte0_or_multiuse:
 ; GFX9:       ; %bb.0: ; %bb
 ; GFX9-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x24
 ; GFX9-NEXT:    v_lshlrev_b32_e32 v0, 2, v0
+; GFX9-NEXT:    v_mov_b32_e32 v2, 0x4050607
 ; GFX9-NEXT:    v_mov_b32_e32 v1, 0
 ; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX9-NEXT:    global_load_dword v0, v0, s[0:1]
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
-; GFX9-NEXT:    v_or_b32_e32 v0, 0x80000001, v0
+; GFX9-NEXT:    v_perm_b32 v0, v0, 0, v2
 ; GFX9-NEXT:    v_cvt_f32_ubyte0_e32 v2, v0
 ; GFX9-NEXT:    v_add_f32_e32 v0, v0, v2
 ; GFX9-NEXT:    global_store_dword v1, v0, s[2:3]
 ; GFX9-NEXT:    s_endpgm
 ;
 ; GFX11-LABEL: cvt_ubyte0_or_multiuse:
 ; GFX11:       ; %bb.0: ; %bb
 ; GFX11-NEXT:    s_load_b128 s[0:3], s[0:1], 0x24
 ; GFX11-NEXT:    v_lshlrev_b32_e32 v0, 2, v0
 ; GFX11-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX11-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX11-NEXT:    global_load_b32 v0, v0, s[0:1]
 ; GFX11-NEXT:    s_waitcnt vmcnt(0)
-; GFX11-NEXT:    v_or_b32_e32 v0, 0x80000001, v0
+; GFX11-NEXT:    v_perm_b32 v0, v0, 0, 0x4050607
 ; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
 ; GFX11-NEXT:    v_cvt_f32_ubyte0_e32 v1, v0
 ; GFX11-NEXT:    v_add_f32_e32 v0, v0, v1
 ; GFX11-NEXT:    global_store_b32 v2, v0, s[2:3]
 ; GFX11-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
 ; GFX11-NEXT:    s_endpgm
 bb:
   %lid = tail call i32 @llvm.amdgcn.workitem.id.x()

llvm/test/CodeGen/AMDGPU/ds_read2.ll

 ; CI-NEXT:    buffer_store_dword v2, v[0:1], s[0:3], 0 addr64
 ; CI-NEXT:    s_endpgm
 ;
 ; GFX9-ALIGNED-LABEL: unaligned_read2_f32:
 ; GFX9-ALIGNED:       ; %bb.0:
 ; GFX9-ALIGNED-NEXT:    s_load_dword s4, s[0:1], 0x8
 ; GFX9-ALIGNED-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x0
 ; GFX9-ALIGNED-NEXT:    v_lshlrev_b32_e32 v0, 2, v0
+; GFX9-ALIGNED-NEXT:    s_mov_b32 s0, 0x4050001
 ; GFX9-ALIGNED-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX9-ALIGNED-NEXT:    v_add_u32_e32 v1, s4, v0
 ; GFX9-ALIGNED-NEXT:    ds_read_u8 v2, v1
 ; GFX9-ALIGNED-NEXT:    ds_read_u8 v3, v1 offset:1
 ; GFX9-ALIGNED-NEXT:    ds_read_u8 v4, v1 offset:2
 ; GFX9-ALIGNED-NEXT:    ds_read_u8 v5, v1 offset:3
 ; GFX9-ALIGNED-NEXT:    ds_read_u8 v6, v1 offset:32
 ; GFX9-ALIGNED-NEXT:    ds_read_u8 v7, v1 offset:33
 ; GFX9-ALIGNED-NEXT:    ds_read_u8 v8, v1 offset:34
 ; GFX9-ALIGNED-NEXT:    ds_read_u8 v1, v1 offset:35
-; GFX9-ALIGNED-NEXT:    s_waitcnt lgkmcnt(6)
-; GFX9-ALIGNED-NEXT:    v_lshl_or_b32 v2, v3, 8, v2
 ; GFX9-ALIGNED-NEXT:    s_waitcnt lgkmcnt(4)
-; GFX9-ALIGNED-NEXT:    v_lshl_or_b32 v3, v5, 8, v4
-; GFX9-ALIGNED-NEXT:    v_lshl_or_b32 v2, v3, 16, v2
+; GFX9-ALIGNED-NEXT:    v_lshl_or_b32 v4, v5, 8, v4
+; GFX9-ALIGNED-NEXT:    v_lshl_or_b32 v2, v3, 8, v2
 ; GFX9-ALIGNED-NEXT:    s_waitcnt lgkmcnt(2)
 ; GFX9-ALIGNED-NEXT:    v_lshl_or_b32 v3, v7, 8, v6
+; GFX9-ALIGNED-NEXT:    v_perm_b32 v2, v2, v4, s0
 ; GFX9-ALIGNED-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX9-ALIGNED-NEXT:    v_lshl_or_b32 v1, v1, 8, v8
-; GFX9-ALIGNED-NEXT:    v_lshl_or_b32 v1, v1, 16, v3
+; GFX9-ALIGNED-NEXT:    v_perm_b32 v1, v3, v1, s0
 ; GFX9-ALIGNED-NEXT:    v_add_f32_e32 v1, v2, v1
 ; GFX9-ALIGNED-NEXT:    global_store_dword v0, v1, s[2:3]
 ; GFX9-ALIGNED-NEXT:    s_endpgm
 ;
 ; GFX9-UNALIGNED-LABEL: unaligned_read2_f32:
 ; GFX9-UNALIGNED:       ; %bb.0:
 ; GFX9-UNALIGNED-NEXT:    s_load_dword s2, s[0:1], 0x8
 ; GFX9-UNALIGNED-NEXT:    v_lshlrev_b32_e32 v2, 2, v0

 ; CI-NEXT:    buffer_store_dword v2, v[0:1], s[0:3], 0 addr64
 ; CI-NEXT:    s_endpgm
 ;
 ; GFX9-ALIGNED-LABEL: unaligned_offset_read2_f32:
 ; GFX9-ALIGNED:       ; %bb.0:
 ; GFX9-ALIGNED-NEXT:    s_load_dword s4, s[0:1], 0x8
 ; GFX9-ALIGNED-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x0
 ; GFX9-ALIGNED-NEXT:    v_lshlrev_b32_e32 v0, 2, v0
+; GFX9-ALIGNED-NEXT:    s_mov_b32 s0, 0x4050001
 ; GFX9-ALIGNED-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX9-ALIGNED-NEXT:    v_add_u32_e32 v1, s4, v0
 ; GFX9-ALIGNED-NEXT:    ds_read_u8 v2, v1 offset:5
 ; GFX9-ALIGNED-NEXT:    ds_read_u8 v3, v1 offset:6
 ; GFX9-ALIGNED-NEXT:    ds_read_u8 v4, v1 offset:7
 ; GFX9-ALIGNED-NEXT:    ds_read_u8 v5, v1 offset:8
 ; GFX9-ALIGNED-NEXT:    ds_read_u8 v6, v1 offset:9
 ; GFX9-ALIGNED-NEXT:    ds_read_u8 v7, v1 offset:10
 ; GFX9-ALIGNED-NEXT:    ds_read_u8 v8, v1 offset:11
 ; GFX9-ALIGNED-NEXT:    ds_read_u8 v1, v1 offset:12
-; GFX9-ALIGNED-NEXT:    s_waitcnt lgkmcnt(6)
-; GFX9-ALIGNED-NEXT:    v_lshl_or_b32 v2, v3, 8, v2
 ; GFX9-ALIGNED-NEXT:    s_waitcnt lgkmcnt(4)
-; GFX9-ALIGNED-NEXT:    v_lshl_or_b32 v3, v5, 8, v4
-; GFX9-ALIGNED-NEXT:    v_lshl_or_b32 v2, v3, 16, v2
+; GFX9-ALIGNED-NEXT:    v_lshl_or_b32 v4, v5, 8, v4
+; GFX9-ALIGNED-NEXT:    v_lshl_or_b32 v2, v3, 8, v2
 ; GFX9-ALIGNED-NEXT:    s_waitcnt lgkmcnt(2)
 ; GFX9-ALIGNED-NEXT:    v_lshl_or_b32 v3, v7, 8, v6
+; GFX9-ALIGNED-NEXT:    v_perm_b32 v2, v2, v4, s0
 ; GFX9-ALIGNED-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX9-ALIGNED-NEXT:    v_lshl_or_b32 v1, v1, 8, v8
-; GFX9-ALIGNED-NEXT:    v_lshl_or_b32 v1, v1, 16, v3
+; GFX9-ALIGNED-NEXT:    v_perm_b32 v1, v3, v1, s0
 ; GFX9-ALIGNED-NEXT:    v_add_f32_e32 v1, v2, v1
 ; GFX9-ALIGNED-NEXT:    global_store_dword v0, v1, s[2:3]
 ; GFX9-ALIGNED-NEXT:    s_endpgm
 ;
 ; GFX9-UNALIGNED-LABEL: unaligned_offset_read2_f32:
 ; GFX9-UNALIGNED:       ; %bb.0:
 ; GFX9-UNALIGNED-NEXT:    s_load_dword s2, s[0:1], 0x8
 ; GFX9-UNALIGNED-NEXT:    v_lshlrev_b32_e32 v2, 2, v0

 ; GFX9-ALIGNED-NEXT:    v_lshlrev_b32_e32 v0, 2, v0
 ; GFX9-ALIGNED-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0x0
 ; GFX9-ALIGNED-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX9-ALIGNED-NEXT:    v_add_u32_e32 v1, s2, v0
 ; GFX9-ALIGNED-NEXT:    ds_read_u16 v2, v1
 ; GFX9-ALIGNED-NEXT:    ds_read_u16 v3, v1 offset:2
 ; GFX9-ALIGNED-NEXT:    ds_read_u16 v4, v1 offset:32
 ; GFX9-ALIGNED-NEXT:    ds_read_u16 v1, v1 offset:34
+; GFX9-ALIGNED-NEXT:    s_mov_b32 s2, 0x4050001
 ; GFX9-ALIGNED-NEXT:    s_waitcnt lgkmcnt(2)
-; GFX9-ALIGNED-NEXT:    v_lshl_or_b32 v2, v3, 16, v2
+; GFX9-ALIGNED-NEXT:    v_perm_b32 v2, v2, v3, s2
 ; GFX9-ALIGNED-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX9-ALIGNED-NEXT:    v_lshl_or_b32 v1, v1, 16, v4
+; GFX9-ALIGNED-NEXT:    v_perm_b32 v1, v4, v1, s2
 ; GFX9-ALIGNED-NEXT:    v_add_f32_e32 v1, v2, v1
 ; GFX9-ALIGNED-NEXT:    global_store_dword v0, v1, s[0:1]
 ; GFX9-ALIGNED-NEXT:    s_endpgm
 ;
 ; GFX9-UNALIGNED-LABEL: misaligned_2_simple_read2_f32:
 ; GFX9-UNALIGNED:       ; %bb.0:
 ; GFX9-UNALIGNED-NEXT:    s_load_dword s2, s[0:1], 0x8
 ; GFX9-UNALIGNED-NEXT:    v_lshlrev_b32_e32 v2, 2, v0

llvm/test/CodeGen/AMDGPU/fast-unaligned-load-store.global.ll

 ; GFX7-UNALIGNED-NEXT:    v_or_b32_e32 v0, v0, v1
 ; GFX7-UNALIGNED-NEXT:    s_setpc_b64 s[30:31]
 ;
 ; GFX9-LABEL: global_load_2xi16_align2:
 ; GFX9:       ; %bb.0:
 ; GFX9-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
 ; GFX9-NEXT:    global_load_ushort v2, v[0:1], off
 ; GFX9-NEXT:    global_load_ushort v3, v[0:1], off offset:2
+; GFX9-NEXT:    s_mov_b32 s4, 0x4050001
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
-; GFX9-NEXT:    v_lshl_or_b32 v0, v3, 16, v2
+; GFX9-NEXT:    v_perm_b32 v0, v2, v3, s4
 ; GFX9-NEXT:    s_setpc_b64 s[30:31]
 ;
 ; GFX10-LABEL: global_load_2xi16_align2:
 ; GFX10:       ; %bb.0:
 ; GFX10-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
 ; GFX10-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX10-NEXT:    s_clause 0x1
 ; GFX10-NEXT:    global_load_ushort v2, v[0:1], off
 ; GFX10-NEXT:    global_load_ushort v3, v[0:1], off offset:2
 ; GFX10-NEXT:    s_waitcnt vmcnt(0)
-; GFX10-NEXT:    v_lshl_or_b32 v0, v3, 16, v2
+; GFX10-NEXT:    v_perm_b32 v0, v2, v3, 0x4050001
 ; GFX10-NEXT:    s_setpc_b64 s[30:31]
 ;
 ; GFX11-LABEL: global_load_2xi16_align2:
 ; GFX11:       ; %bb.0:
 ; GFX11-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
 ; GFX11-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-NEXT:    s_clause 0x1
 ; GFX11-NEXT:    global_load_u16 v2, v[0:1], off
 ; GFX11-NEXT:    global_load_u16 v0, v[0:1], off offset:2
 ; GFX11-NEXT:    s_waitcnt vmcnt(0)
-; GFX11-NEXT:    v_lshl_or_b32 v0, v0, 16, v2
+; GFX11-NEXT:    v_perm_b32 v0, v2, v0, 0x4050001
 ; GFX11-NEXT:    s_setpc_b64 s[30:31]
   %gep.p = getelementptr i16, i16 addrspace(1)* %p, i64 1
   %p.0 = load i16, i16 addrspace(1)* %p, align 2
   %p.1 = load i16, i16 addrspace(1)* %gep.p, align 2
   %zext.0 = zext i16 %p.0 to i32
   %zext.1 = zext i16 %p.1 to i32
   %shl.1 = shl i32 %zext.1, 16
   %or = or i32 %zext.0, %shl.1

llvm/test/CodeGen/AMDGPU/fast-unaligned-load-store.private.ll

 ; GFX7-UNALIGNED-NEXT:    v_or_b32_e32 v0, v0, v1
 ; GFX7-UNALIGNED-NEXT:    s_setpc_b64 s[30:31]
 ;
 ; GFX9-LABEL: private_load_2xi16_align2:
 ; GFX9:       ; %bb.0:
 ; GFX9-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
 ; GFX9-NEXT:    buffer_load_ushort v1, v0, s[0:3], 0 offen
 ; GFX9-NEXT:    buffer_load_ushort v2, v0, s[0:3], 0 offen offset:2
+; GFX9-NEXT:    s_mov_b32 s4, 0x4050001
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
-; GFX9-NEXT:    v_lshl_or_b32 v0, v2, 16, v1
+; GFX9-NEXT:    v_perm_b32 v0, v1, v2, s4
 ; GFX9-NEXT:    s_setpc_b64 s[30:31]
 ;
 ; GFX9-FLASTSCR-LABEL: private_load_2xi16_align2:
 ; GFX9-FLASTSCR:       ; %bb.0:
 ; GFX9-FLASTSCR-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
 ; GFX9-FLASTSCR-NEXT:    scratch_load_ushort v1, v0, off
 ; GFX9-FLASTSCR-NEXT:    scratch_load_ushort v2, v0, off offset:2
+; GFX9-FLASTSCR-NEXT:    s_mov_b32 s0, 0x4050001
 ; GFX9-FLASTSCR-NEXT:    s_waitcnt vmcnt(0)
-; GFX9-FLASTSCR-NEXT:    v_lshl_or_b32 v0, v2, 16, v1
+; GFX9-FLASTSCR-NEXT:    v_perm_b32 v0, v1, v2, s0
 ; GFX9-FLASTSCR-NEXT:    s_setpc_b64 s[30:31]
 ;
 ; GFX10-LABEL: private_load_2xi16_align2:
 ; GFX10:       ; %bb.0:
 ; GFX10-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
 ; GFX10-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX10-NEXT:    s_clause 0x1
 ; GFX10-NEXT:    buffer_load_ushort v1, v0, s[0:3], 0 offen
 ; GFX10-NEXT:    buffer_load_ushort v2, v0, s[0:3], 0 offen offset:2
 ; GFX10-NEXT:    s_waitcnt vmcnt(0)
-; GFX10-NEXT:    v_lshl_or_b32 v0, v2, 16, v1
+; GFX10-NEXT:    v_perm_b32 v0, v1, v2, 0x4050001
 ; GFX10-NEXT:    s_setpc_b64 s[30:31]
 ;
 ; GFX10-FLASTSCR-LABEL: private_load_2xi16_align2:
 ; GFX10-FLASTSCR:       ; %bb.0:
 ; GFX10-FLASTSCR-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
 ; GFX10-FLASTSCR-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX10-FLASTSCR-NEXT:    s_clause 0x1
 ; GFX10-FLASTSCR-NEXT:    scratch_load_ushort v1, v0, off
 ; GFX10-FLASTSCR-NEXT:    scratch_load_ushort v2, v0, off offset:2
 ; GFX10-FLASTSCR-NEXT:    s_waitcnt vmcnt(0)
-; GFX10-FLASTSCR-NEXT:    v_lshl_or_b32 v0, v2, 16, v1
+; GFX10-FLASTSCR-NEXT:    v_perm_b32 v0, v1, v2, 0x4050001
 ; GFX10-FLASTSCR-NEXT:    s_setpc_b64 s[30:31]
 ;
 ; GFX11-LABEL: private_load_2xi16_align2:
 ; GFX11:       ; %bb.0:
 ; GFX11-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
 ; GFX11-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-NEXT:    s_clause 0x1
 ; GFX11-NEXT:    scratch_load_u16 v1, v0, off
 ; GFX11-NEXT:    scratch_load_u16 v0, v0, off offset:2
 ; GFX11-NEXT:    s_waitcnt vmcnt(0)
-; GFX11-NEXT:    v_lshl_or_b32 v0, v0, 16, v1
+; GFX11-NEXT:    v_perm_b32 v0, v1, v0, 0x4050001
 ; GFX11-NEXT:    s_setpc_b64 s[30:31]
 ;
 ; GFX11-FLASTSCR-LABEL: private_load_2xi16_align2:
 ; GFX11-FLASTSCR:       ; %bb.0:
 ; GFX11-FLASTSCR-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
 ; GFX11-FLASTSCR-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-FLASTSCR-NEXT:    s_clause 0x1
 ; GFX11-FLASTSCR-NEXT:    scratch_load_u16 v1, v0, off
 ; GFX11-FLASTSCR-NEXT:    scratch_load_u16 v0, v0, off offset:2
 ; GFX11-FLASTSCR-NEXT:    s_waitcnt vmcnt(0)
-; GFX11-FLASTSCR-NEXT:    v_lshl_or_b32 v0, v0, 16, v1
+; GFX11-FLASTSCR-NEXT:    v_perm_b32 v0, v1, v0, 0x4050001
 ; GFX11-FLASTSCR-NEXT:    s_setpc_b64 s[30:31]
   %gep.p = getelementptr i16, i16 addrspace(5)* %p, i64 1
   %p.0 = load i16, i16 addrspace(5)* %p, align 2
   %p.1 = load i16, i16 addrspace(5)* %gep.p, align 2
   %zext.0 = zext i16 %p.0 to i32
   %zext.1 = zext i16 %p.1 to i32
   %shl.1 = shl i32 %zext.1, 16
   %or = or i32 %zext.0, %shl.1

 ; GFX7-UNALIGNED-NEXT:    buffer_load_dword v0, v0, s[0:3], 0 offen
 ; GFX7-UNALIGNED-NEXT:    s_waitcnt vmcnt(0)
 ; GFX7-UNALIGNED-NEXT:    s_setpc_b64 s[30:31]
 ;
 ; GFX9-LABEL: private_load_2xi16_align1:
 ; GFX9:       ; %bb.0:
 ; GFX9-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
 ; GFX9-NEXT:    buffer_load_dword v0, v0, s[0:3], 0 offen
-; GFX9-NEXT:    s_mov_b32 s4, 0xffff
+; GFX9-NEXT:    v_mov_b32_e32 v1, 0x4050607
 ; GFX9-NEXT:    s_waitcnt vmcnt(0)
-; GFX9-NEXT:    v_and_b32_e32 v1, 0xffff0000, v0
-; GFX9-NEXT:    v_and_or_b32 v0, v0, s4, v1
+; GFX9-NEXT:    v_perm_b32 v0, v0, 0, v1
 ; GFX9-NEXT:    s_setpc_b64 s[30:31]
 ;
 ; GFX9-FLASTSCR-LABEL: private_load_2xi16_align1:
 ; GFX9-FLASTSCR:       ; %bb.0:
 ; GFX9-FLASTSCR-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
 ; GFX9-FLASTSCR-NEXT:    scratch_load_dword v0, v0, off
-; GFX9-FLASTSCR-NEXT:    s_mov_b32 s0, 0xffff
+; GFX9-FLASTSCR-NEXT:    v_mov_b32_e32 v1, 0x4050607
 ; GFX9-FLASTSCR-NEXT:    s_waitcnt vmcnt(0)
-; GFX9-FLASTSCR-NEXT:    v_and_b32_e32 v1, 0xffff0000, v0
-; GFX9-FLASTSCR-NEXT:    v_and_or_b32 v0, v0, s0, v1
+; GFX9-FLASTSCR-NEXT:    v_perm_b32 v0, v0, 0, v1
 ; GFX9-FLASTSCR-NEXT:    s_setpc_b64 s[30:31]
 ;
 ; GFX10-LABEL: private_load_2xi16_align1:
 ; GFX10:       ; %bb.0:
 ; GFX10-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
 ; GFX10-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX10-NEXT:    buffer_load_dword v0, v0, s[0:3], 0 offen
 ; GFX10-NEXT:    s_waitcnt vmcnt(0)
-; GFX10-NEXT:    v_and_b32_e32 v1, 0xffff0000, v0
-; GFX10-NEXT:    v_and_or_b32 v0, 0xffff, v0, v1
+; GFX10-NEXT:    v_perm_b32 v0, v0, 0, 0x4050607
 ; GFX10-NEXT:    s_setpc_b64 s[30:31]
 ;
 ; GFX10-FLASTSCR-LABEL: private_load_2xi16_align1:
 ; GFX10-FLASTSCR:       ; %bb.0:
 ; GFX10-FLASTSCR-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
 ; GFX10-FLASTSCR-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX10-FLASTSCR-NEXT:    scratch_load_dword v0, v0, off
 ; GFX10-FLASTSCR-NEXT:    s_waitcnt vmcnt(0)
-; GFX10-FLASTSCR-NEXT:    v_and_b32_e32 v1, 0xffff0000, v0
-; GFX10-FLASTSCR-NEXT:    v_and_or_b32 v0, 0xffff, v0, v1
+; GFX10-FLASTSCR-NEXT:    v_perm_b32 v0, v0, 0, 0x4050607
 ; GFX10-FLASTSCR-NEXT:    s_setpc_b64 s[30:31]
 ;
 ; GFX11-LABEL: private_load_2xi16_align1:
 ; GFX11:       ; %bb.0:
 ; GFX11-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
 ; GFX11-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-NEXT:    scratch_load_b32 v0, v0, off
 ; GFX11-NEXT:    s_waitcnt vmcnt(0)
-; GFX11-NEXT:    v_and_b32_e32 v1, 0xffff0000, v0
-; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_1)
-; GFX11-NEXT:    v_and_or_b32 v0, 0xffff, v0, v1
+; GFX11-NEXT:    v_perm_b32 v0, v0, 0, 0x4050607
 ; GFX11-NEXT:    s_setpc_b64 s[30:31]
 ;
 ; GFX11-FLASTSCR-LABEL: private_load_2xi16_align1:
 ; GFX11-FLASTSCR:       ; %bb.0:
 ; GFX11-FLASTSCR-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
 ; GFX11-FLASTSCR-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-FLASTSCR-NEXT:    scratch_load_b32 v0, v0, off
 ; GFX11-FLASTSCR-NEXT:    s_waitcnt vmcnt(0)
-; GFX11-FLASTSCR-NEXT:    v_and_b32_e32 v1, 0xffff0000, v0
-; GFX11-FLASTSCR-NEXT:    s_delay_alu instid0(VALU_DEP_1)
-; GFX11-FLASTSCR-NEXT:    v_and_or_b32 v0, 0xffff, v0, v1
+; GFX11-FLASTSCR-NEXT:    v_perm_b32 v0, v0, 0, 0x4050607
 ; GFX11-FLASTSCR-NEXT:    s_setpc_b64 s[30:31]
   %gep.p = getelementptr i16, i16 addrspace(5)* %p, i64 1
   %p.0 = load i16, i16 addrspace(5)* %p, align 1
   %p.1 = load i16, i16 addrspace(5)* %gep.p, align 1
   %zext.0 = zext i16 %p.0 to i32
   %zext.1 = zext i16 %p.1 to i32
   %shl.1 = shl i32 %zext.1, 16
   %or = or i32 %zext.0, %shl.1

llvm/test/CodeGen/AMDGPU/insert_vector_elt.v2i16.ll

 ; VI-NEXT:    v_lshlrev_b32_e32 v2, 2, v0
 ; VI-NEXT:    s_waitcnt lgkmcnt(0)
 ; VI-NEXT:    v_mov_b32_e32 v1, s3
 ; VI-NEXT:    v_add_u32_e32 v0, vcc, s2, v2
 ; VI-NEXT:    v_addc_u32_e32 v1, vcc, 0, v1, vcc
 ; VI-NEXT:    flat_load_dword v3, v[0:1]
 ; VI-NEXT:    v_mov_b32_e32 v1, s1
 ; VI-NEXT:    v_add_u32_e32 v0, vcc, s0, v2
+; VI-NEXT:    v_mov_b32_e32 v2, 0x6070203
 ; VI-NEXT:    v_addc_u32_e32 v1, vcc, 0, v1, vcc
 ; VI-NEXT:    s_waitcnt vmcnt(0)
-; VI-NEXT:    v_lshrrev_b32_e32 v2, 16, v3
-; VI-NEXT:    v_alignbit_b32 v2, v2, s4, 16
+; VI-NEXT:    v_perm_b32 v2, s4, v3, v2
 ; VI-NEXT:    flat_store_dword v[0:1], v2
 ; VI-NEXT:    s_endpgm
 ;
 ; CI-LABEL: v_insertelement_v2i16_0_reghi:
 ; CI:       ; %bb.0:
 ; CI-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
 ; CI-NEXT:    s_load_dword s4, s[4:5], 0x4
 ; CI-NEXT:    v_lshlrev_b32_e32 v2, 2, v0

 ; GFX9-NEXT:    global_store_dwordx2 v2, v[0:1], s[0:1]
 ; GFX9-NEXT:    s_endpgm
 ;
 ; VI-LABEL: v_insertelement_v4f16_0:
 ; VI:       ; %bb.0:
 ; VI-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
 ; VI-NEXT:    s_load_dword s4, s[4:5], 0x30
 ; VI-NEXT:    v_lshlrev_b32_e32 v2, 3, v0
+; VI-NEXT:    v_mov_b32_e32 v4, 0x4050203
 ; VI-NEXT:    s_waitcnt lgkmcnt(0)
 ; VI-NEXT:    v_mov_b32_e32 v1, s3
 ; VI-NEXT:    v_add_u32_e32 v0, vcc, s2, v2
 ; VI-NEXT:    v_addc_u32_e32 v1, vcc, 0, v1, vcc
 ; VI-NEXT:    flat_load_dwordx2 v[0:1], v[0:1]
 ; VI-NEXT:    v_mov_b32_e32 v3, s1
 ; VI-NEXT:    v_add_u32_e32 v2, vcc, s0, v2
-; VI-NEXT:    s_mov_b32 s0, 0xffff
-; VI-NEXT:    v_mov_b32_e32 v4, s4
 ; VI-NEXT:    v_addc_u32_e32 v3, vcc, 0, v3, vcc
 ; VI-NEXT:    s_waitcnt vmcnt(0)
-; VI-NEXT:    v_bfi_b32 v0, s0, v4, v0
+; VI-NEXT:    v_perm_b32 v0, s4, v0, v4
 ; VI-NEXT:    flat_store_dwordx2 v[2:3], v[0:1]
 ; VI-NEXT:    s_endpgm
 ;
 ; CI-LABEL: v_insertelement_v4f16_0:
 ; CI:       ; %bb.0:
 ; CI-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
 ; CI-NEXT:    s_load_dword s4, s[4:5], 0xc
 ; CI-NEXT:    v_lshlrev_b32_e32 v2, 3, v0

 ; GFX9-NEXT:    global_store_dwordx2 v2, v[0:1], s[0:1]
 ; GFX9-NEXT:    s_endpgm
 ;
 ; VI-LABEL: v_insertelement_v4f16_2:
 ; VI:       ; %bb.0:
 ; VI-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
 ; VI-NEXT:    s_load_dword s4, s[4:5], 0x30
 ; VI-NEXT:    v_lshlrev_b32_e32 v2, 3, v0
+; VI-NEXT:    v_mov_b32_e32 v4, 0x4050203
 ; VI-NEXT:    s_waitcnt lgkmcnt(0)
 ; VI-NEXT:    v_mov_b32_e32 v1, s3
 ; VI-NEXT:    v_add_u32_e32 v0, vcc, s2, v2
 ; VI-NEXT:    v_addc_u32_e32 v1, vcc, 0, v1, vcc
 ; VI-NEXT:    flat_load_dwordx2 v[0:1], v[0:1]
 ; VI-NEXT:    v_mov_b32_e32 v3, s1
 ; VI-NEXT:    v_add_u32_e32 v2, vcc, s0, v2
-; VI-NEXT:    s_mov_b32 s0, 0xffff
-; VI-NEXT:    v_mov_b32_e32 v4, s4
 ; VI-NEXT:    v_addc_u32_e32 v3, vcc, 0, v3, vcc
 ; VI-NEXT:    s_waitcnt vmcnt(0)
-; VI-NEXT:    v_bfi_b32 v1, s0, v4, v1
+; VI-NEXT:    v_perm_b32 v1, s4, v1, v4
 ; VI-NEXT:    flat_store_dwordx2 v[2:3], v[0:1]
 ; VI-NEXT:    s_endpgm
 ;
 ; CI-LABEL: v_insertelement_v4f16_2:
 ; CI:       ; %bb.0:
 ; CI-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
 ; CI-NEXT:    s_load_dword s4, s[4:5], 0xc
 ; CI-NEXT:    v_lshlrev_b32_e32 v2, 3, v0

 ; GFX9-NEXT:    global_store_dwordx2 v2, v[0:1], s[0:1]
 ; GFX9-NEXT:    s_endpgm
 ;
 ; VI-LABEL: v_insertelement_v4i16_2:
 ; VI:       ; %bb.0:
 ; VI-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
 ; VI-NEXT:    s_load_dword s4, s[4:5], 0x10
 ; VI-NEXT:    v_lshlrev_b32_e32 v2, 3, v0
+; VI-NEXT:    v_mov_b32_e32 v4, 0x4050203
 ; VI-NEXT:    s_waitcnt lgkmcnt(0)
 ; VI-NEXT:    v_mov_b32_e32 v1, s3
 ; VI-NEXT:    v_add_u32_e32 v0, vcc, s2, v2
 ; VI-NEXT:    v_addc_u32_e32 v1, vcc, 0, v1, vcc
 ; VI-NEXT:    flat_load_dwordx2 v[0:1], v[0:1]
 ; VI-NEXT:    v_mov_b32_e32 v3, s1
 ; VI-NEXT:    v_add_u32_e32 v2, vcc, s0, v2
-; VI-NEXT:    s_mov_b32 s0, 0xffff
-; VI-NEXT:    v_mov_b32_e32 v4, s4
 ; VI-NEXT:    v_addc_u32_e32 v3, vcc, 0, v3, vcc
 ; VI-NEXT:    s_waitcnt vmcnt(0)
-; VI-NEXT:    v_bfi_b32 v1, s0, v4, v1
+; VI-NEXT:    v_perm_b32 v1, s4, v1, v4
 ; VI-NEXT:    flat_store_dwordx2 v[2:3], v[0:1]
 ; VI-NEXT:    s_endpgm
 ;
 ; CI-LABEL: v_insertelement_v4i16_2:
 ; CI:       ; %bb.0:
 ; CI-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
 ; CI-NEXT:    s_load_dword s4, s[4:5], 0x4
 ; CI-NEXT:    v_lshlrev_b32_e32 v2, 3, v0

 ; GFX9-NEXT:    global_store_dwordx4 v8, v[0:3], s[0:1]
 ; GFX9-NEXT:    s_endpgm
 ;
 ; VI-LABEL: v_insertelement_v16i16_6:
 ; VI:       ; %bb.0:
 ; VI-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0
 ; VI-NEXT:    s_load_dword s4, s[4:5], 0x10
 ; VI-NEXT:    v_lshlrev_b32_e32 v8, 5, v0
+; VI-NEXT:    v_mov_b32_e32 v12, 0x4050203
 ; VI-NEXT:    s_waitcnt lgkmcnt(0)
 ; VI-NEXT:    v_mov_b32_e32 v1, s3
 ; VI-NEXT:    v_add_u32_e32 v0, vcc, s2, v8
 ; VI-NEXT:    v_addc_u32_e32 v1, vcc, 0, v1, vcc
 ; VI-NEXT:    v_add_u32_e32 v4, vcc, 16, v0
 ; VI-NEXT:    v_addc_u32_e32 v5, vcc, 0, v1, vcc
 ; VI-NEXT:    flat_load_dwordx4 v[0:3], v[0:1]
 ; VI-NEXT:    flat_load_dwordx4 v[4:7], v[4:5]
 ; VI-NEXT:    v_mov_b32_e32 v9, s1
 ; VI-NEXT:    v_add_u32_e32 v8, vcc, s0, v8
 ; VI-NEXT:    v_addc_u32_e32 v9, vcc, 0, v9, vcc
 ; VI-NEXT:    v_add_u32_e32 v10, vcc, 16, v8
-; VI-NEXT:    s_mov_b32 s2, 0xffff
-; VI-NEXT:    v_mov_b32_e32 v12, s4
 ; VI-NEXT:    v_addc_u32_e32 v11, vcc, 0, v9, vcc
 ; VI-NEXT:    s_waitcnt vmcnt(1)
-; VI-NEXT:    v_bfi_b32 v3, s2, v12, v3
+; VI-NEXT:    v_perm_b32 v3, s4, v3, v12
 ; VI-NEXT:    s_waitcnt vmcnt(0)
 ; VI-NEXT:    flat_store_dwordx4 v[10:11], v[4:7]
 ; VI-NEXT:    flat_store_dwordx4 v[8:9], v[0:3]
 ; VI-NEXT:    s_endpgm
 ;
 ; CI-LABEL: v_insertelement_v16i16_6:
 ; CI:       ; %bb.0:
 ; CI-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0

llvm/test/CodeGen/AMDGPU/load-hi16.ll

 ; GFX906-NEXT:    s_setpc_b64 s[30:31]
 ;
 ; GFX803-LABEL: load_local_v2i16_split_multi_chain:
 ; GFX803:       ; %bb.0: ; %entry
 ; GFX803-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
 ; GFX803-NEXT:    s_mov_b32 m0, -1
 ; GFX803-NEXT:    ds_read_u16 v1, v0
 ; GFX803-NEXT:    ds_read_u16 v0, v0 offset:2
+; GFX803-NEXT:    s_mov_b32 s4, 0x4050001
 ; GFX803-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX803-NEXT:    v_lshlrev_b32_e32 v0, 16, v0
-; GFX803-NEXT:    v_or_b32_e32 v0, v1, v0
+; GFX803-NEXT:    v_perm_b32 v0, v1, v0, s4
 ; GFX803-NEXT:    s_setpc_b64 s[30:31]
 ;
 ; GFX900-FLATSCR-LABEL: load_local_v2i16_split_multi_chain:
 ; GFX900-FLATSCR:       ; %bb.0: ; %entry
 ; GFX900-FLATSCR-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
 ; GFX900-FLATSCR-NEXT:    ds_read_u16 v1, v0
 ; GFX900-FLATSCR-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX900-FLATSCR-NEXT:    ds_read_u16_d16_hi v1, v0 offset:2

 ; GFX906-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX906-NEXT:    v_perm_b32 v0, v0, v1, s4
 ; GFX906-NEXT:    s_setpc_b64 s[30:31]
 ;
 ; GFX803-LABEL: load_local_lo_hi_v2i16_samechain:
 ; GFX803:       ; %bb.0: ; %entry
 ; GFX803-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
 ; GFX803-NEXT:    s_mov_b32 m0, -1
-; GFX803-NEXT:    ds_read_u16 v1, v0 offset:16
-; GFX803-NEXT:    ds_read_u16 v0, v0
-; GFX803-NEXT:    s_waitcnt lgkmcnt(1)
-; GFX803-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
+; GFX803-NEXT:    ds_read_u16 v1, v0
+; GFX803-NEXT:    ds_read_u16 v0, v0 offset:16
+; GFX803-NEXT:    s_mov_b32 s4, 0x4050001
 ; GFX803-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX803-NEXT:    v_or_b32_e32 v0, v0, v1
+; GFX803-NEXT:    v_perm_b32 v0, v1, v0, s4
 ; GFX803-NEXT:    s_setpc_b64 s[30:31]
 ;
 ; GFX900-FLATSCR-LABEL: load_local_lo_hi_v2i16_samechain:
 ; GFX900-FLATSCR:       ; %bb.0: ; %entry
 ; GFX900-FLATSCR-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
 ; GFX900-FLATSCR-NEXT:    ds_read_u16 v1, v0
 ; GFX900-FLATSCR-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX900-FLATSCR-NEXT:    ds_read_u16_d16_hi v1, v0 offset:16

 ; GFX803-LABEL: load_local_lo_hi_v2i16_side_effect:
 ; GFX803:       ; %bb.0: ; %entry
 ; GFX803-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
 ; GFX803-NEXT:    s_mov_b32 m0, -1
 ; GFX803-NEXT:    v_mov_b32_e32 v3, 0x7b
 ; GFX803-NEXT:    ds_read_u16 v2, v0
 ; GFX803-NEXT:    ds_write_b16 v1, v3
 ; GFX803-NEXT:    ds_read_u16 v0, v0 offset:16
+; GFX803-NEXT:    s_mov_b32 s4, 0x4050001
 ; GFX803-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX803-NEXT:    v_lshlrev_b32_e32 v0, 16, v0
-; GFX803-NEXT:    v_or_b32_e32 v0, v2, v0
+; GFX803-NEXT:    v_perm_b32 v0, v2, v0, s4
 ; GFX803-NEXT:    s_setpc_b64 s[30:31]
 ;
 ; GFX900-FLATSCR-LABEL: load_local_lo_hi_v2i16_side_effect:
 ; GFX900-FLATSCR:       ; %bb.0: ; %entry
 ; GFX900-FLATSCR-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
 ; GFX900-FLATSCR-NEXT:    ds_read_u16 v2, v0
 ; GFX900-FLATSCR-NEXT:    v_mov_b32_e32 v3, 0x7b
 ; GFX900-FLATSCR-NEXT:    ds_write_b16 v1, v3

 ; GFX803:       ; %bb.0: ; %entry
 ; GFX803-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
 ; GFX803-NEXT:    v_add_u32_e32 v2, vcc, 2, v0
 ; GFX803-NEXT:    v_addc_u32_e32 v3, vcc, 0, v1, vcc
 ; GFX803-NEXT:    flat_load_ushort v0, v[0:1] glc
 ; GFX803-NEXT:    s_waitcnt vmcnt(0)
 ; GFX803-NEXT:    flat_load_ushort v1, v[2:3] glc
 ; GFX803-NEXT:    s_waitcnt vmcnt(0)
-; GFX803-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
-; GFX803-NEXT:    v_or_b32_e32 v0, v0, v1
+; GFX803-NEXT:    s_mov_b32 s4, 0x4050001
+; GFX803-NEXT:    v_perm_b32 v0, v0, v1, s4
 ; GFX803-NEXT:    s_setpc_b64 s[30:31]
 ;
 ; GFX900-FLATSCR-LABEL: load_global_v2i16_split:
 ; GFX900-FLATSCR:       ; %bb.0: ; %entry
 ; GFX900-FLATSCR-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
 ; GFX900-FLATSCR-NEXT:    global_load_ushort v2, v[0:1], off glc
 ; GFX900-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
 ; GFX900-FLATSCR-NEXT:    global_load_short_d16_hi v2, v[0:1], off offset:2 glc

 ; GFX803-LABEL: load_flat_v2i16_split:
 ; GFX803:       ; %bb.0: ; %entry
 ; GFX803-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
 ; GFX803-NEXT:    v_add_u32_e32 v2, vcc, 2, v0
 ; GFX803-NEXT:    v_addc_u32_e32 v3, vcc, 0, v1, vcc
 ; GFX803-NEXT:    flat_load_ushort v0, v[0:1] glc
 ; GFX803-NEXT:    s_waitcnt vmcnt(0)
 ; GFX803-NEXT:    flat_load_ushort v1, v[2:3] glc
-; GFX803-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; GFX803-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
-; GFX803-NEXT:    v_or_b32_e32 v0, v0, v1
+; GFX803-NEXT:    s_waitcnt vmcnt(0)
+; GFX803-NEXT:    s_mov_b32 s4, 0x4050001
+; GFX803-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX803-NEXT:    v_perm_b32 v0, v0, v1, s4
 ; GFX803-NEXT:    s_setpc_b64 s[30:31]
 ;
 ; GFX900-FLATSCR-LABEL: load_flat_v2i16_split:
 ; GFX900-FLATSCR:       ; %bb.0: ; %entry
 ; GFX900-FLATSCR-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
 ; GFX900-FLATSCR-NEXT:    flat_load_ushort v2, v[0:1] glc
 ; GFX900-FLATSCR-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX900-FLATSCR-NEXT:    flat_load_short_d16_hi v2, v[0:1] offset:2 glc

 ;
 ; GFX803-LABEL: load_constant_v2i16_split:
 ; GFX803:       ; %bb.0: ; %entry
 ; GFX803-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
 ; GFX803-NEXT:    v_add_u32_e32 v2, vcc, 2, v0
 ; GFX803-NEXT:    v_addc_u32_e32 v3, vcc, 0, v1, vcc
 ; GFX803-NEXT:    flat_load_ushort v0, v[0:1] glc
 ; GFX803-NEXT:    flat_load_ushort v1, v[2:3] glc
+; GFX803-NEXT:    s_mov_b32 s4, 0x4050001
 ; GFX803-NEXT:    s_waitcnt vmcnt(0)
-; GFX803-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
-; GFX803-NEXT:    v_or_b32_e32 v0, v0, v1
+; GFX803-NEXT:    v_perm_b32 v0, v0, v1, s4
 ; GFX803-NEXT:    s_setpc_b64 s[30:31]
 ;
 ; GFX900-FLATSCR-LABEL: load_constant_v2i16_split:
 ; GFX900-FLATSCR:       ; %bb.0: ; %entry
 ; GFX900-FLATSCR-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
 ; GFX900-FLATSCR-NEXT:    global_load_ushort v2, v[0:1], off glc
 ; GFX900-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
 ; GFX900-FLATSCR-NEXT:    global_load_short_d16_hi v2, v[0:1], off offset:2 glc

 ;
 ; GFX803-LABEL: load_private_v2i16_split:
 ; GFX803:       ; %bb.0: ; %entry
 ; GFX803-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
 ; GFX803-NEXT:    buffer_load_ushort v0, off, s[0:3], s32 glc
 ; GFX803-NEXT:    s_waitcnt vmcnt(0)
 ; GFX803-NEXT:    buffer_load_ushort v1, off, s[0:3], s32 offset:2 glc
 ; GFX803-NEXT:    s_waitcnt vmcnt(0)
-; GFX803-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
-; GFX803-NEXT:    v_or_b32_e32 v0, v0, v1
+; GFX803-NEXT:    s_mov_b32 s4, 0x4050001
+; GFX803-NEXT:    v_perm_b32 v0, v0, v1, s4
 ; GFX803-NEXT:    s_setpc_b64 s[30:31]
 ;
 ; GFX900-FLATSCR-LABEL: load_private_v2i16_split:
 ; GFX900-FLATSCR:       ; %bb.0: ; %entry
 ; GFX900-FLATSCR-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
 ; GFX900-FLATSCR-NEXT:    scratch_load_ushort v0, off, s32 glc
 ; GFX900-FLATSCR-NEXT:    s_waitcnt vmcnt(0)
 ; GFX900-FLATSCR-NEXT:    scratch_load_short_d16_hi v0, off, s32 offset:2 glc

llvm/test/CodeGen/AMDGPU/load-lo16.ll

 ; GFX906-NEXT:    s_waitcnt vmcnt(0)
 ; GFX906-NEXT:    s_setpc_b64 s[30:31]
 ;
 ; GFX803-LABEL: load_local_lo_v2f16_reghi_vreg:
 ; GFX803:       ; %bb.0: ; %entry
 ; GFX803-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
 ; GFX803-NEXT:    s_mov_b32 m0, -1
 ; GFX803-NEXT:    ds_read_u16 v0, v0
-; GFX803-NEXT:    v_and_b32_e32 v1, 0xffff0000, v1
+; GFX803-NEXT:    s_mov_b32 s4, 0x4050203
 ; GFX803-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX803-NEXT:    v_or_b32_e32 v0, v0, v1
+; GFX803-NEXT:    v_perm_b32 v0, v0, v1, s4
 ; GFX803-NEXT:    flat_store_dword v[0:1], v0
 ; GFX803-NEXT:    s_waitcnt vmcnt(0)
 ; GFX803-NEXT:    s_setpc_b64 s[30:31]
 entry:
   %reg.bc = bitcast i32 %reg to <2 x half>
   %load = load half, half addrspace(3)* %in
   %build1 = insertelement <2 x half> %reg.bc, half %load, i32 0
   store <2 x half> %build1, <2 x half> addrspace(1)* undef

 ; GFX906-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX906-NEXT:    s_setpc_b64 s[30:31]
 ;
 ; GFX803-LABEL: load_local_lo_v2i16_reghi_vreg_multi_use_hi:
 ; GFX803:       ; %bb.0: ; %entry
 ; GFX803-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
 ; GFX803-NEXT:    s_mov_b32 m0, -1
 ; GFX803-NEXT:    ds_read_u16 v0, v0
+; GFX803-NEXT:    s_mov_b32 s4, 0x4050203
 ; GFX803-NEXT:    v_lshrrev_b32_e32 v2, 16, v1
-; GFX803-NEXT:    v_and_b32_e32 v1, 0xffff0000, v1
 ; GFX803-NEXT:    v_mov_b32_e32 v3, 0
 ; GFX803-NEXT:    ds_write_b16 v3, v2
 ; GFX803-NEXT:    s_waitcnt lgkmcnt(1)
-; GFX803-NEXT:    v_or_b32_e32 v0, v0, v1
+; GFX803-NEXT:    v_perm_b32 v0, v0, v1, s4
 ; GFX803-NEXT:    flat_store_dword v[0:1], v0
 ; GFX803-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX803-NEXT:    s_setpc_b64 s[30:31]
 entry:
   %load = load i16, i16 addrspace(3)* %in
   %elt1 = extractelement <2 x i16> %reg, i32 1
   store i16 %elt1, i16 addrspace(3)* null
   %build1 = insertelement <2 x i16> %reg, i16 %load, i32 0

 ; GFX906-NEXT:    s_setpc_b64 s[30:31]
 ;
 ; GFX803-LABEL: load_global_lo_v2i16_reglo_vreg:
 ; GFX803:       ; %bb.0: ; %entry
 ; GFX803-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
 ; GFX803-NEXT:    v_add_u32_e32 v0, vcc, 0xfffff002, v0
 ; GFX803-NEXT:    v_addc_u32_e32 v1, vcc, -1, v1, vcc
 ; GFX803-NEXT:    flat_load_ushort v0, v[0:1]
-; GFX803-NEXT:    v_and_b32_e32 v1, 0xffff0000, v2
+; GFX803-NEXT:    s_mov_b32 s4, 0x4050203
 ; GFX803-NEXT:    s_waitcnt vmcnt(0)
-; GFX803-NEXT:    v_or_b32_e32 v0, v0, v1
+; GFX803-NEXT:    v_perm_b32 v0, v0, v2, s4
 ; GFX803-NEXT:    flat_store_dword v[0:1], v0
 ; GFX803-NEXT:    s_waitcnt vmcnt(0)
 ; GFX803-NEXT:    s_setpc_b64 s[30:31]
 entry:
   %reg.bc = bitcast i32 %reg to <2 x i16>
   %gep = getelementptr inbounds i16, i16 addrspace(1)* %in, i64 -2047
   %load = load i16, i16 addrspace(1)* %gep
   %build1 = insertelement <2 x i16> %reg.bc, i16 %load, i32 0

 ; GFX906-NEXT:    s_setpc_b64 s[30:31]
 ;
 ; GFX803-LABEL: load_global_lo_v2f16_reglo_vreg:
 ; GFX803:       ; %bb.0: ; %entry
 ; GFX803-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
 ; GFX803-NEXT:    v_add_u32_e32 v0, vcc, 0xfffff002, v0
 ; GFX803-NEXT:    v_addc_u32_e32 v1, vcc, -1, v1, vcc
 ; GFX803-NEXT:    flat_load_ushort v0, v[0:1]
-; GFX803-NEXT:    v_and_b32_e32 v1, 0xffff0000, v2
+; GFX803-NEXT:    s_mov_b32 s4, 0x4050203
 ; GFX803-NEXT:    s_waitcnt vmcnt(0)
-; GFX803-NEXT:    v_or_b32_e32 v0, v0, v1
+; GFX803-NEXT:    v_perm_b32 v0, v0, v2, s4
 ; GFX803-NEXT:    flat_store_dword v[0:1], v0
 ; GFX803-NEXT:    s_waitcnt vmcnt(0)
 ; GFX803-NEXT:    s_setpc_b64 s[30:31]
 entry:
   %reg.bc = bitcast i32 %reg to <2 x half>
   %gep = getelementptr inbounds half, half addrspace(1)* %in, i64 -2047
   %load = load half, half addrspace(1)* %gep
   %build1 = insertelement <2 x half> %reg.bc, half %load, i32 0

 ; GFX906-NEXT:    global_store_dword v[0:1], v0, off
 ; GFX906-NEXT:    s_waitcnt vmcnt(0)
 ; GFX906-NEXT:    s_setpc_b64 s[30:31]
 ;
 ; GFX803-LABEL: load_flat_lo_v2i16_reghi_vreg:
 ; GFX803:       ; %bb.0: ; %entry
 ; GFX803-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
 ; GFX803-NEXT:    flat_load_ushort v0, v[0:1]
-; GFX803-NEXT:    v_and_b32_e32 v1, 0xffff0000, v2
+; GFX803-NEXT:    s_mov_b32 s4, 0x4050203
 ; GFX803-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; GFX803-NEXT:    v_or_b32_e32 v0, v0, v1
+; GFX803-NEXT:    v_perm_b32 v0, v0, v2, s4
 ; GFX803-NEXT:    flat_store_dword v[0:1], v0
 ; GFX803-NEXT:    s_waitcnt vmcnt(0)
 ; GFX803-NEXT:    s_setpc_b64 s[30:31]
 entry:
   %reg.bc = bitcast i32 %reg to <2 x i16>
   %load = load i16, i16* %in
   %build1 = insertelement <2 x i16> %reg.bc, i16 %load, i32 0
   store <2 x i16> %build1, <2 x i16> addrspace(1)* undef

 ; GFX906-NEXT:    global_store_dword v[0:1], v0, off
 ; GFX906-NEXT:    s_waitcnt vmcnt(0)
 ; GFX906-NEXT:    s_setpc_b64 s[30:31]
 ;
 ; GFX803-LABEL: load_flat_lo_v2f16_reghi_vreg:
 ; GFX803:       ; %bb.0: ; %entry
 ; GFX803-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
 ; GFX803-NEXT:    flat_load_ushort v0, v[0:1]
-; GFX803-NEXT:    v_and_b32_e32 v1, 0xffff0000, v2
+; GFX803-NEXT:    s_mov_b32 s4, 0x4050203
 ; GFX803-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; GFX803-NEXT:    v_or_b32_e32 v0, v0, v1
+; GFX803-NEXT:    v_perm_b32 v0, v0, v2, s4
 ; GFX803-NEXT:    flat_store_dword v[0:1], v0
 ; GFX803-NEXT:    s_waitcnt vmcnt(0)
 ; GFX803-NEXT:    s_setpc_b64 s[30:31]
 
 ; FIXME: the and above should be removable
 entry:
   %reg.bc = bitcast i32 %reg to <2 x half>
   %load = load half, half* %in

 ; GFX906-NEXT:    global_store_dword v[0:1], v0, off
 ; GFX906-NEXT:    s_waitcnt vmcnt(0)
 ; GFX906-NEXT:    s_setpc_b64 s[30:31]
 ;
 ; GFX803-LABEL: load_private_lo_v2i16_reglo_vreg:
 ; GFX803:       ; %bb.0: ; %entry
 ; GFX803-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
 ; GFX803-NEXT:    buffer_load_ushort v1, off, s[0:3], s32 offset:4094
-; GFX803-NEXT:    v_and_b32_e32 v0, 0xffff0000, v0
+; GFX803-NEXT:    s_mov_b32 s4, 0x4050203
 ; GFX803-NEXT:    s_waitcnt vmcnt(0)
-; GFX803-NEXT:    v_or_b32_e32 v0, v1, v0
+; GFX803-NEXT:    v_perm_b32 v0, v1, v0, s4
 ; GFX803-NEXT:    flat_store_dword v[0:1], v0
 ; GFX803-NEXT:    s_waitcnt vmcnt(0)
 ; GFX803-NEXT:    s_setpc_b64 s[30:31]
 ;
 ; GFX900-FLATSCR-LABEL: load_private_lo_v2i16_reglo_vreg:
 ; GFX900-FLATSCR:       ; %bb.0: ; %entry
 ; GFX900-FLATSCR-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
 ; GFX900-FLATSCR-NEXT:    scratch_load_short_d16 v0, off, s32 offset:4094

 ; GFX906-NEXT:    global_store_dword v[0:1], v0, off
 ; GFX906-NEXT:    s_waitcnt vmcnt(0)
 ; GFX906-NEXT:    s_setpc_b64 s[30:31]
 ;
 ; GFX803-LABEL: load_private_lo_v2f16_reglo_vreg:
 ; GFX803:       ; %bb.0: ; %entry
 ; GFX803-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
 ; GFX803-NEXT:    buffer_load_ushort v1, off, s[0:3], s32 offset:4094
-; GFX803-NEXT:    v_and_b32_e32 v0, 0xffff0000, v0
+; GFX803-NEXT:    s_mov_b32 s4, 0x4050203
 ; GFX803-NEXT:    s_waitcnt vmcnt(0)
-; GFX803-NEXT:    v_or_b32_e32 v0, v1, v0
+; GFX803-NEXT:    v_perm_b32 v0, v1, v0, s4
 ; GFX803-NEXT:    flat_store_dword v[0:1], v0
 ; GFX803-NEXT:    s_waitcnt vmcnt(0)
 ; GFX803-NEXT:    s_setpc_b64 s[30:31]
 ;
 ; GFX900-FLATSCR-LABEL: load_private_lo_v2f16_reglo_vreg:
 ; GFX900-FLATSCR:       ; %bb.0: ; %entry
 ; GFX900-FLATSCR-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
 ; GFX900-FLATSCR-NEXT:    scratch_load_short_d16 v0, off, s32 offset:4094

 ; GFX906-NEXT:    s_waitcnt vmcnt(0)
 ; GFX906-NEXT:    s_setpc_b64 s[30:31]
 ;
 ; GFX803-LABEL: load_private_lo_v2i16_reglo_vreg_nooff:
 ; GFX803:       ; %bb.0: ; %entry
 ; GFX803-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
 ; GFX803-NEXT:    buffer_load_ushort v0, off, s[0:3], 0 offset:4094 glc
 ; GFX803-NEXT:    s_waitcnt vmcnt(0)
-; GFX803-NEXT:    v_and_b32_e32 v1, 0xffff0000, v1
-; GFX803-NEXT:    v_or_b32_e32 v0, v0, v1
+; GFX803-NEXT:    s_mov_b32 s4, 0x4050203
+; GFX803-NEXT:    v_perm_b32 v0, v0, v1, s4
 ; GFX803-NEXT:    flat_store_dword v[0:1], v0
 ; GFX803-NEXT:    s_waitcnt vmcnt(0)
 ; GFX803-NEXT:    s_setpc_b64 s[30:31]
 ;
 ; GFX900-FLATSCR-LABEL: load_private_lo_v2i16_reglo_vreg_nooff:
 ; GFX900-FLATSCR:       ; %bb.0: ; %entry
 ; GFX900-FLATSCR-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
 ; GFX900-FLATSCR-NEXT:    s_movk_i32 s0, 0xffe

 ; GFX906-NEXT:    s_waitcnt vmcnt(0)
 ; GFX906-NEXT:    s_setpc_b64 s[30:31]
 ;
 ; GFX803-LABEL: load_private_lo_v2i16_reghi_vreg_nooff:
 ; GFX803:       ; %bb.0: ; %entry
 ; GFX803-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
 ; GFX803-NEXT:    buffer_load_ushort v0, off, s[0:3], 0 offset:4094 glc
 ; GFX803-NEXT:    s_waitcnt vmcnt(0)
-; GFX803-NEXT:    v_and_b32_e32 v1, 0xffff0000, v1
-; GFX803-NEXT:    v_or_b32_e32 v0, v0, v1
+; GFX803-NEXT:    s_mov_b32 s4, 0x4050203
+; GFX803-NEXT:    v_perm_b32 v0, v0, v1, s4
 ; GFX803-NEXT:    flat_store_dword v[0:1], v0
 ; GFX803-NEXT:    s_waitcnt vmcnt(0)
 ; GFX803-NEXT:    s_setpc_b64 s[30:31]
 ;
 ; GFX900-FLATSCR-LABEL: load_private_lo_v2i16_reghi_vreg_nooff:
 ; GFX900-FLATSCR:       ; %bb.0: ; %entry
 ; GFX900-FLATSCR-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
 ; GFX900-FLATSCR-NEXT:    s_movk_i32 s0, 0xffe

 ; GFX906-NEXT:    s_waitcnt vmcnt(0)
 ; GFX906-NEXT:    s_setpc_b64 s[30:31]
 ;
 ; GFX803-LABEL: load_private_lo_v2f16_reglo_vreg_nooff:
 ; GFX803:       ; %bb.0: ; %entry
 ; GFX803-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
 ; GFX803-NEXT:    buffer_load_ushort v0, off, s[0:3], 0 offset:4094 glc
 ; GFX803-NEXT:    s_waitcnt vmcnt(0)
-; GFX803-NEXT:    v_and_b32_e32 v1, 0xffff0000, v1
-; GFX803-NEXT:    v_or_b32_e32 v0, v0, v1
+; GFX803-NEXT:    s_mov_b32 s4, 0x4050203
+; GFX803-NEXT:    v_perm_b32 v0, v0, v1, s4
 ; GFX803-NEXT:    flat_store_dword v[0:1], v0
 ; GFX803-NEXT:    s_waitcnt vmcnt(0)
 ; GFX803-NEXT:    s_setpc_b64 s[30:31]
 ;
 ; GFX900-FLATSCR-LABEL: load_private_lo_v2f16_reglo_vreg_nooff:
 ; GFX900-FLATSCR:       ; %bb.0: ; %entry
 ; GFX900-FLATSCR-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
 ; GFX900-FLATSCR-NEXT:    s_movk_i32 s0, 0xffe

 ; GFX906-NEXT:    s_setpc_b64 s[30:31]
 ;
 ; GFX803-LABEL: load_constant_lo_v2i16_reglo_vreg:
 ; GFX803:       ; %bb.0: ; %entry
 ; GFX803-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
 ; GFX803-NEXT:    v_add_u32_e32 v0, vcc, 0xfffff002, v0
 ; GFX803-NEXT:    v_addc_u32_e32 v1, vcc, -1, v1, vcc
 ; GFX803-NEXT:    flat_load_ushort v0, v[0:1]
-; GFX803-NEXT:    v_and_b32_e32 v1, 0xffff0000, v2
+; GFX803-NEXT:    s_mov_b32 s4, 0x4050203
 ; GFX803-NEXT:    s_waitcnt vmcnt(0)
-; GFX803-NEXT:    v_or_b32_e32 v0, v0, v1
+; GFX803-NEXT:    v_perm_b32 v0, v0, v2, s4
 ; GFX803-NEXT:    flat_store_dword v[0:1], v0
 ; GFX803-NEXT:    s_waitcnt vmcnt(0)
 ; GFX803-NEXT:    s_setpc_b64 s[30:31]
 entry:
   %reg.bc = bitcast i32 %reg to <2 x i16>
   %gep = getelementptr inbounds i16, i16 addrspace(4)* %in, i64 -2047
   %load = load i16, i16 addrspace(4)* %gep
   %build1 = insertelement <2 x i16> %reg.bc, i16 %load, i32 0

 ; GFX906-NEXT:    s_setpc_b64 s[30:31]
 ;
 ; GFX803-LABEL: load_constant_lo_v2f16_reglo_vreg:
 ; GFX803:       ; %bb.0: ; %entry
 ; GFX803-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
 ; GFX803-NEXT:    v_add_u32_e32 v0, vcc, 0xfffff002, v0
 ; GFX803-NEXT:    v_addc_u32_e32 v1, vcc, -1, v1, vcc
 ; GFX803-NEXT:    flat_load_ushort v0, v[0:1]
-; GFX803-NEXT:    v_and_b32_e32 v1, 0xffff0000, v2
+; GFX803-NEXT:    s_mov_b32 s4, 0x4050203
 ; GFX803-NEXT:    s_waitcnt vmcnt(0)
-; GFX803-NEXT:    v_or_b32_e32 v0, v0, v1
+; GFX803-NEXT:    v_perm_b32 v0, v0, v2, s4
 ; GFX803-NEXT:    flat_store_dword v[0:1], v0
 ; GFX803-NEXT:    s_waitcnt vmcnt(0)
 ; GFX803-NEXT:    s_setpc_b64 s[30:31]
 entry:
   %reg.bc = bitcast i32 %reg to <2 x half>
   %gep = getelementptr inbounds half, half addrspace(4)* %in, i64 -2047
   %load = load half, half addrspace(4)* %gep
   %build1 = insertelement <2 x half> %reg.bc, half %load, i32 0

 ; GFX803:       ; %bb.0: ; %entry
 ; GFX803-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
 ; GFX803-NEXT:    v_mov_b32_e32 v1, 0x7b
 ; GFX803-NEXT:    buffer_store_dword v1, off, s[0:3], s32 offset:4
 ; GFX803-NEXT:    s_waitcnt vmcnt(0)
 ; GFX803-NEXT:    v_mov_b32_e32 v2, 44
 ; GFX803-NEXT:    buffer_load_ushort v1, v2, s[0:3], s32 offen offset:4054 glc
 ; GFX803-NEXT:    s_waitcnt vmcnt(0)
-; GFX803-NEXT:    v_and_b32_e32 v0, 0xffff0000, v0
-; GFX803-NEXT:    v_or_b32_e32 v0, v1, v0
+; GFX803-NEXT:    s_mov_b32 s4, 0x4050203
+; GFX803-NEXT:    v_perm_b32 v0, v1, v0, s4
 ; GFX803-NEXT:    flat_store_dword v[0:1], v0
 ; GFX803-NEXT:    s_waitcnt vmcnt(0)
 ; GFX803-NEXT:    s_setpc_b64 s[30:31]
 ;
 ; GFX900-FLATSCR-LABEL: load_private_lo_v2i16_reglo_vreg_to_offset:
 ; GFX900-FLATSCR:       ; %bb.0: ; %entry
 ; GFX900-FLATSCR-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
 ; GFX900-FLATSCR-NEXT:    v_mov_b32_e32 v1, 0x7b

llvm/test/CodeGen/AMDGPU/load-local.128.ll

 ; GFX9-NEXT:    ds_read_u8 v9, v0 offset:8
 ; GFX9-NEXT:    ds_read_u8 v10, v0 offset:9
 ; GFX9-NEXT:    ds_read_u8 v11, v0 offset:10
 ; GFX9-NEXT:    ds_read_u8 v12, v0 offset:11
 ; GFX9-NEXT:    ds_read_u8 v13, v0 offset:12
 ; GFX9-NEXT:    ds_read_u8 v14, v0 offset:13
 ; GFX9-NEXT:    ds_read_u8 v15, v0 offset:14
 ; GFX9-NEXT:    ds_read_u8 v16, v0 offset:15
-; GFX9-NEXT:    s_waitcnt lgkmcnt(14)
-; GFX9-NEXT:    v_lshl_or_b32 v0, v2, 8, v1
 ; GFX9-NEXT:    s_waitcnt lgkmcnt(12)
-; GFX9-NEXT:    v_lshl_or_b32 v1, v4, 8, v3
-; GFX9-NEXT:    v_lshl_or_b32 v0, v1, 16, v0
-; GFX9-NEXT:    s_waitcnt lgkmcnt(10)
-; GFX9-NEXT:    v_lshl_or_b32 v1, v6, 8, v5
+; GFX9-NEXT:    v_lshl_or_b32 v0, v4, 8, v3
+; GFX9-NEXT:    v_lshl_or_b32 v1, v2, 8, v1
+; GFX9-NEXT:    s_mov_b32 s4, 0x4050001
+; GFX9-NEXT:    v_perm_b32 v0, v1, v0, s4
 ; GFX9-NEXT:    s_waitcnt lgkmcnt(8)
-; GFX9-NEXT:    v_lshl_or_b32 v2, v8, 8, v7
-; GFX9-NEXT:    v_lshl_or_b32 v1, v2, 16, v1
-; GFX9-NEXT:    s_waitcnt lgkmcnt(6)
-; GFX9-NEXT:    v_lshl_or_b32 v2, v10, 8, v9
+; GFX9-NEXT:    v_lshl_or_b32 v1, v8, 8, v7
+; GFX9-NEXT:    v_lshl_or_b32 v2, v6, 8, v5
+; GFX9-NEXT:    v_perm_b32 v1, v2, v1, s4
 ; GFX9-NEXT:    s_waitcnt lgkmcnt(4)
-; GFX9-NEXT:    v_lshl_or_b32 v3, v12, 8, v11
-; GFX9-NEXT:    v_lshl_or_b32 v2, v3, 16, v2
-; GFX9-NEXT:    s_waitcnt lgkmcnt(2)
-; GFX9-NEXT:    v_lshl_or_b32 v3, v14, 8, v13
+; GFX9-NEXT:    v_lshl_or_b32 v2, v12, 8, v11
+; GFX9-NEXT:    v_lshl_or_b32 v3, v10, 8, v9
+; GFX9-NEXT:    v_perm_b32 v2, v3, v2, s4
 ; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX9-NEXT:    v_lshl_or_b32 v4, v16, 8, v15
-; GFX9-NEXT:    v_lshl_or_b32 v3, v4, 16, v3
+; GFX9-NEXT:    v_lshl_or_b32 v3, v16, 8, v15
+; GFX9-NEXT:    v_lshl_or_b32 v4, v14, 8, v13
+; GFX9-NEXT:    v_perm_b32 v3, v4, v3, s4
 ; GFX9-NEXT:    s_setpc_b64 s[30:31]
 ;
 ; GFX7-LABEL: load_lds_v4i32_align1:
 ; GFX7:       ; %bb.0:
 ; GFX7-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
 ; GFX7-NEXT:    s_mov_b32 m0, -1
 ; GFX7-NEXT:    ds_read_u8 v1, v0 offset:6
 ; GFX7-NEXT:    ds_read_u8 v2, v0 offset:4

 ; GFX10-NEXT:    ds_read_u8 v5, v0 offset:4
 ; GFX10-NEXT:    ds_read_u8 v6, v0 offset:5
 ; GFX10-NEXT:    ds_read_u8 v7, v0 offset:6
 ; GFX10-NEXT:    ds_read_u8 v8, v0 offset:7
 ; GFX10-NEXT:    ds_read_u8 v9, v0 offset:8
 ; GFX10-NEXT:    ds_read_u8 v10, v0 offset:9
 ; GFX10-NEXT:    ds_read_u8 v11, v0 offset:10
 ; GFX10-NEXT:    ds_read_u8 v12, v0 offset:11
-; GFX10-NEXT:    ds_read_u8 v13, v0 offset:12
-; GFX10-NEXT:    ds_read_u8 v14, v0 offset:13
-; GFX10-NEXT:    ds_read_u8 v15, v0 offset:14
-; GFX10-NEXT:    ds_read_u8 v0, v0 offset:15
+; GFX10-NEXT:    ds_read_u8 v13, v0 offset:14
+; GFX10-NEXT:    ds_read_u8 v14, v0 offset:15
+; GFX10-NEXT:    ds_read_u8 v15, v0 offset:12
+; GFX10-NEXT:    ds_read_u8 v0, v0 offset:13
 ; GFX10-NEXT:    s_waitcnt lgkmcnt(14)
 ; GFX10-NEXT:    v_lshl_or_b32 v1, v2, 8, v1
 ; GFX10-NEXT:    s_waitcnt lgkmcnt(12)
-; GFX10-NEXT:    v_lshl_or_b32 v2, v4, 8, v3
+; GFX10-NEXT:    v_lshl_or_b32 v3, v4, 8, v3
 ; GFX10-NEXT:    s_waitcnt lgkmcnt(10)
-; GFX10-NEXT:    v_lshl_or_b32 v3, v6, 8, v5
+; GFX10-NEXT:    v_lshl_or_b32 v4, v6, 8, v5
 ; GFX10-NEXT:    s_waitcnt lgkmcnt(8)
-; GFX10-NEXT:    v_lshl_or_b32 v4, v8, 8, v7
+; GFX10-NEXT:    v_lshl_or_b32 v2, v8, 8, v7
 ; GFX10-NEXT:    s_waitcnt lgkmcnt(6)
-; GFX10-NEXT:    v_lshl_or_b32 v5, v10, 8, v9
+; GFX10-NEXT:    v_lshl_or_b32 v6, v10, 8, v9
 ; GFX10-NEXT:    s_waitcnt lgkmcnt(4)
-; GFX10-NEXT:    v_lshl_or_b32 v6, v12, 8, v11
+; GFX10-NEXT:    v_lshl_or_b32 v5, v12, 8, v11
 ; GFX10-NEXT:    s_waitcnt lgkmcnt(2)
 ; GFX10-NEXT:    v_lshl_or_b32 v7, v14, 8, v13
 ; GFX10-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX10-NEXT:    v_lshl_or_b32 v8, v0, 8, v15
-; GFX10-NEXT:    v_lshl_or_b32 v0, v2, 16, v1
-; GFX10-NEXT:    v_lshl_or_b32 v1, v4, 16, v3
-; GFX10-NEXT:    v_lshl_or_b32 v2, v6, 16, v5
-; GFX10-NEXT:    v_lshl_or_b32 v3, v8, 16, v7
+; GFX10-NEXT:    v_perm_b32 v0, v1, v3, 0x4050001
+; GFX10-NEXT:    v_perm_b32 v1, v4, v2, 0x4050001
+; GFX10-NEXT:    v_perm_b32 v2, v6, v5, 0x4050001
+; GFX10-NEXT:    v_perm_b32 v3, v8, v7, 0x4050001
 ; GFX10-NEXT:    s_setpc_b64 s[30:31]
 ;
 ; GFX11-LABEL: load_lds_v4i32_align1:
 ; GFX11:       ; %bb.0:
 ; GFX11-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
 ; GFX11-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-NEXT:    ds_load_u8 v1, v0
 ; GFX11-NEXT:    ds_load_u8 v2, v0 offset:1
 ; GFX11-NEXT:    ds_load_u8 v3, v0 offset:2
 ; GFX11-NEXT:    ds_load_u8 v4, v0 offset:3
 ; GFX11-NEXT:    ds_load_u8 v5, v0 offset:4
 ; GFX11-NEXT:    ds_load_u8 v6, v0 offset:5
 ; GFX11-NEXT:    ds_load_u8 v7, v0 offset:6
 ; GFX11-NEXT:    ds_load_u8 v8, v0 offset:7
 ; GFX11-NEXT:    ds_load_u8 v9, v0 offset:8
 ; GFX11-NEXT:    ds_load_u8 v10, v0 offset:9
 ; GFX11-NEXT:    ds_load_u8 v11, v0 offset:10
 ; GFX11-NEXT:    ds_load_u8 v12, v0 offset:11
-; GFX11-NEXT:    ds_load_u8 v13, v0 offset:12
-; GFX11-NEXT:    ds_load_u8 v14, v0 offset:13
-; GFX11-NEXT:    ds_load_u8 v15, v0 offset:14
-; GFX11-NEXT:    ds_load_u8 v0, v0 offset:15
+; GFX11-NEXT:    ds_load_u8 v13, v0 offset:14
+; GFX11-NEXT:    ds_load_u8 v14, v0 offset:15
+; GFX11-NEXT:    ds_load_u8 v15, v0 offset:12
+; GFX11-NEXT:    ds_load_u8 v0, v0 offset:13
 ; GFX11-NEXT:    s_waitcnt lgkmcnt(14)
 ; GFX11-NEXT:    v_lshl_or_b32 v1, v2, 8, v1
 ; GFX11-NEXT:    s_waitcnt lgkmcnt(12)
-; GFX11-NEXT:    v_lshl_or_b32 v2, v4, 8, v3
+; GFX11-NEXT:    v_lshl_or_b32 v3, v4, 8, v3
 ; GFX11-NEXT:    s_waitcnt lgkmcnt(10)
-; GFX11-NEXT:    v_lshl_or_b32 v3, v6, 8, v5
+; GFX11-NEXT:    v_lshl_or_b32 v4, v6, 8, v5
 ; GFX11-NEXT:    s_waitcnt lgkmcnt(8)
-; GFX11-NEXT:    v_lshl_or_b32 v4, v8, 8, v7
+; GFX11-NEXT:    v_lshl_or_b32 v2, v8, 8, v7
 ; GFX11-NEXT:    s_waitcnt lgkmcnt(6)
-; GFX11-NEXT:    v_lshl_or_b32 v5, v10, 8, v9
+; GFX11-NEXT:    v_lshl_or_b32 v6, v10, 8, v9
 ; GFX11-NEXT:    s_waitcnt lgkmcnt(4)
-; GFX11-NEXT:    v_lshl_or_b32 v6, v12, 8, v11
+; GFX11-NEXT:    v_lshl_or_b32 v5, v12, 8, v11
 ; GFX11-NEXT:    s_waitcnt lgkmcnt(2)
 ; GFX11-NEXT:    v_lshl_or_b32 v7, v14, 8, v13
 ; GFX11-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX11-NEXT:    v_lshl_or_b32 v8, v0, 8, v15
-; GFX11-NEXT:    v_lshl_or_b32 v0, v2, 16, v1
-; GFX11-NEXT:    v_lshl_or_b32 v1, v4, 16, v3
-; GFX11-NEXT:    v_lshl_or_b32 v2, v6, 16, v5
+; GFX11-NEXT:    v_perm_b32 v0, v1, v3, 0x4050001
+; GFX11-NEXT:    v_perm_b32 v1, v4, v2, 0x4050001
+; GFX11-NEXT:    v_perm_b32 v2, v6, v5, 0x4050001
 ; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_4)
-; GFX11-NEXT:    v_lshl_or_b32 v3, v8, 16, v7
+; GFX11-NEXT:    v_perm_b32 v3, v8, v7, 0x4050001
 ; GFX11-NEXT:    s_setpc_b64 s[30:31]
   %load = load <4 x i32>, <4 x i32> addrspace(3)* %ptr, align 1
   ret <4 x i32> %load
 }
 
 define <4 x i32> @load_lds_v4i32_align2(<4 x i32> addrspace(3)* %ptr) {
 ; GFX9-LABEL: load_lds_v4i32_align2:
 ; GFX9:       ; %bb.0:
 ; GFX9-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
 ; GFX9-NEXT:    ds_read_u16 v1, v0
 ; GFX9-NEXT:    ds_read_u16 v2, v0 offset:2
 ; GFX9-NEXT:    ds_read_u16 v3, v0 offset:4
 ; GFX9-NEXT:    ds_read_u16 v4, v0 offset:6
 ; GFX9-NEXT:    ds_read_u16 v5, v0 offset:8
 ; GFX9-NEXT:    ds_read_u16 v6, v0 offset:10
 ; GFX9-NEXT:    ds_read_u16 v7, v0 offset:12
 ; GFX9-NEXT:    ds_read_u16 v8, v0 offset:14
+; GFX9-NEXT:    s_mov_b32 s4, 0x4050001
 ; GFX9-NEXT:    s_waitcnt lgkmcnt(6)
-; GFX9-NEXT:    v_lshl_or_b32 v0, v2, 16, v1
+; GFX9-NEXT:    v_perm_b32 v0, v1, v2, s4
 ; GFX9-NEXT:    s_waitcnt lgkmcnt(4)
-; GFX9-NEXT:    v_lshl_or_b32 v1, v4, 16, v3
+; GFX9-NEXT:    v_perm_b32 v1, v3, v4, s4
 ; GFX9-NEXT:    s_waitcnt lgkmcnt(2)
-; GFX9-NEXT:    v_lshl_or_b32 v2, v6, 16, v5
+; GFX9-NEXT:    v_perm_b32 v2, v5, v6, s4
 ; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX9-NEXT:    v_lshl_or_b32 v3, v8, 16, v7
+; GFX9-NEXT:    v_perm_b32 v3, v7, v8, s4
 ; GFX9-NEXT:    s_setpc_b64 s[30:31]
 ;
 ; GFX7-LABEL: load_lds_v4i32_align2:
 ; GFX7:       ; %bb.0:
 ; GFX7-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
 ; GFX7-NEXT:    s_mov_b32 m0, -1
 ; GFX7-NEXT:    ds_read_u16 v3, v0 offset:12
 ; GFX7-NEXT:    ds_read_u16 v2, v0 offset:8

 ; GFX10-NEXT:    ds_read_u16 v2, v0 offset:2
 ; GFX10-NEXT:    ds_read_u16 v3, v0 offset:4
 ; GFX10-NEXT:    ds_read_u16 v4, v0 offset:6
 ; GFX10-NEXT:    ds_read_u16 v5, v0 offset:8
 ; GFX10-NEXT:    ds_read_u16 v6, v0 offset:10
 ; GFX10-NEXT:    ds_read_u16 v7, v0 offset:12
 ; GFX10-NEXT:    ds_read_u16 v8, v0 offset:14
 ; GFX10-NEXT:    s_waitcnt lgkmcnt(6)
-; GFX10-NEXT:    v_lshl_or_b32 v0, v2, 16, v1
+; GFX10-NEXT:    v_perm_b32 v0, v1, v2, 0x4050001
 ; GFX10-NEXT:    s_waitcnt lgkmcnt(4)
-; GFX10-NEXT:    v_lshl_or_b32 v1, v4, 16, v3
+; GFX10-NEXT:    v_perm_b32 v1, v3, v4, 0x4050001
 ; GFX10-NEXT:    s_waitcnt lgkmcnt(2)
-; GFX10-NEXT:    v_lshl_or_b32 v2, v6, 16, v5
+; GFX10-NEXT:    v_perm_b32 v2, v5, v6, 0x4050001
 ; GFX10-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX10-NEXT:    v_lshl_or_b32 v3, v8, 16, v7
+; GFX10-NEXT:    v_perm_b32 v3, v7, v8, 0x4050001
 ; GFX10-NEXT:    s_setpc_b64 s[30:31]
 ;
 ; GFX11-LABEL: load_lds_v4i32_align2:
 ; GFX11:       ; %bb.0:
 ; GFX11-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
 ; GFX11-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-NEXT:    ds_load_u16 v1, v0
 ; GFX11-NEXT:    ds_load_u16 v2, v0 offset:2
 ; GFX11-NEXT:    ds_load_u16 v3, v0 offset:4
 ; GFX11-NEXT:    ds_load_u16 v4, v0 offset:6
 ; GFX11-NEXT:    ds_load_u16 v5, v0 offset:8
 ; GFX11-NEXT:    ds_load_u16 v6, v0 offset:10
 ; GFX11-NEXT:    ds_load_u16 v7, v0 offset:12
 ; GFX11-NEXT:    ds_load_u16 v8, v0 offset:14
 ; GFX11-NEXT:    s_waitcnt lgkmcnt(6)
-; GFX11-NEXT:    v_lshl_or_b32 v0, v2, 16, v1
+; GFX11-NEXT:    v_perm_b32 v0, v1, v2, 0x4050001
 ; GFX11-NEXT:    s_waitcnt lgkmcnt(4)
-; GFX11-NEXT:    v_lshl_or_b32 v1, v4, 16, v3
+; GFX11-NEXT:    v_perm_b32 v1, v3, v4, 0x4050001
 ; GFX11-NEXT:    s_waitcnt lgkmcnt(2)
-; GFX11-NEXT:    v_lshl_or_b32 v2, v6, 16, v5
+; GFX11-NEXT:    v_perm_b32 v2, v5, v6, 0x4050001
 ; GFX11-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-NEXT:    v_lshl_or_b32 v3, v8, 16, v7
+; GFX11-NEXT:    v_perm_b32 v3, v7, v8, 0x4050001
 ; GFX11-NEXT:    s_setpc_b64 s[30:31]
   %load = load <4 x i32>, <4 x i32> addrspace(3)* %ptr, align 2
   ret <4 x i32> %load
 }
 
 define <4 x i32> @load_lds_v4i32_align4(<4 x i32> addrspace(3)* %ptr) {
 ; GFX9-LABEL: load_lds_v4i32_align4:
 ; GFX9:       ; %bb.0:

llvm/test/CodeGen/AMDGPU/load-local.96.ll

 ; GFX9-NEXT:    ds_read_u8 v5, v0 offset:4
 ; GFX9-NEXT:    ds_read_u8 v6, v0 offset:5
 ; GFX9-NEXT:    ds_read_u8 v7, v0 offset:6
 ; GFX9-NEXT:    ds_read_u8 v8, v0 offset:7
 ; GFX9-NEXT:    ds_read_u8 v9, v0 offset:8
 ; GFX9-NEXT:    ds_read_u8 v10, v0 offset:9
 ; GFX9-NEXT:    ds_read_u8 v11, v0 offset:10
 ; GFX9-NEXT:    ds_read_u8 v12, v0 offset:11
-; GFX9-NEXT:    s_waitcnt lgkmcnt(10)
-; GFX9-NEXT:    v_lshl_or_b32 v0, v2, 8, v1
 ; GFX9-NEXT:    s_waitcnt lgkmcnt(8)
-; GFX9-NEXT:    v_lshl_or_b32 v1, v4, 8, v3
-; GFX9-NEXT:    v_lshl_or_b32 v0, v1, 16, v0
-; GFX9-NEXT:    s_waitcnt lgkmcnt(6)
-; GFX9-NEXT:    v_lshl_or_b32 v1, v6, 8, v5
+; GFX9-NEXT:    v_lshl_or_b32 v0, v4, 8, v3
+; GFX9-NEXT:    v_lshl_or_b32 v1, v2, 8, v1
+; GFX9-NEXT:    s_mov_b32 s4, 0x4050001
+; GFX9-NEXT:    v_perm_b32 v0, v1, v0, s4
 ; GFX9-NEXT:    s_waitcnt lgkmcnt(4)
-; GFX9-NEXT:    v_lshl_or_b32 v2, v8, 8, v7
-; GFX9-NEXT:    v_lshl_or_b32 v1, v2, 16, v1
-; GFX9-NEXT:    s_waitcnt lgkmcnt(2)
-; GFX9-NEXT:    v_lshl_or_b32 v2, v10, 8, v9
+; GFX9-NEXT:    v_lshl_or_b32 v1, v8, 8, v7
+; GFX9-NEXT:    v_lshl_or_b32 v2, v6, 8, v5
+; GFX9-NEXT:    v_perm_b32 v1, v2, v1, s4
 ; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX9-NEXT:    v_lshl_or_b32 v3, v12, 8, v11
-; GFX9-NEXT:    v_lshl_or_b32 v2, v3, 16, v2
+; GFX9-NEXT:    v_lshl_or_b32 v2, v12, 8, v11
+; GFX9-NEXT:    v_lshl_or_b32 v3, v10, 8, v9
+; GFX9-NEXT:    v_perm_b32 v2, v3, v2, s4
 ; GFX9-NEXT:    s_setpc_b64 s[30:31]
 ;
 ; GFX7-LABEL: load_lds_v3i32_align1:
 ; GFX7:       ; %bb.0:
 ; GFX7-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
 ; GFX7-NEXT:    s_mov_b32 m0, -1
 ; GFX7-NEXT:    ds_read_u8 v1, v0 offset:6
 ; GFX7-NEXT:    ds_read_u8 v2, v0 offset:4

 ; GFX10-NEXT:    ds_read_u8 v1, v0
 ; GFX10-NEXT:    ds_read_u8 v2, v0 offset:1
 ; GFX10-NEXT:    ds_read_u8 v3, v0 offset:2
 ; GFX10-NEXT:    ds_read_u8 v4, v0 offset:3
 ; GFX10-NEXT:    ds_read_u8 v5, v0 offset:4
 ; GFX10-NEXT:    ds_read_u8 v6, v0 offset:5
 ; GFX10-NEXT:    ds_read_u8 v7, v0 offset:6
 ; GFX10-NEXT:    ds_read_u8 v8, v0 offset:7
-; GFX10-NEXT:    ds_read_u8 v9, v0 offset:8
-; GFX10-NEXT:    ds_read_u8 v10, v0 offset:9
-; GFX10-NEXT:    ds_read_u8 v11, v0 offset:10
-; GFX10-NEXT:    ds_read_u8 v0, v0 offset:11
+; GFX10-NEXT:    ds_read_u8 v9, v0 offset:10
+; GFX10-NEXT:    ds_read_u8 v10, v0 offset:11
+; GFX10-NEXT:    ds_read_u8 v11, v0 offset:8
+; GFX10-NEXT:    ds_read_u8 v0, v0 offset:9
 ; GFX10-NEXT:    s_waitcnt lgkmcnt(10)
 ; GFX10-NEXT:    v_lshl_or_b32 v1, v2, 8, v1
 ; GFX10-NEXT:    s_waitcnt lgkmcnt(8)
-; GFX10-NEXT:    v_lshl_or_b32 v2, v4, 8, v3
+; GFX10-NEXT:    v_lshl_or_b32 v3, v4, 8, v3
 ; GFX10-NEXT:    s_waitcnt lgkmcnt(6)
-; GFX10-NEXT:    v_lshl_or_b32 v3, v6, 8, v5
+; GFX10-NEXT:    v_lshl_or_b32 v4, v6, 8, v5
 ; GFX10-NEXT:    s_waitcnt lgkmcnt(4)
-; GFX10-NEXT:    v_lshl_or_b32 v4, v8, 8, v7
+; GFX10-NEXT:    v_lshl_or_b32 v2, v8, 8, v7
 ; GFX10-NEXT:    s_waitcnt lgkmcnt(2)
 ; GFX10-NEXT:    v_lshl_or_b32 v5, v10, 8, v9
 ; GFX10-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX10-NEXT:    v_lshl_or_b32 v6, v0, 8, v11
-; GFX10-NEXT:    v_lshl_or_b32 v0, v2, 16, v1
-; GFX10-NEXT:    v_lshl_or_b32 v1, v4, 16, v3
-; GFX10-NEXT:    v_lshl_or_b32 v2, v6, 16, v5
+; GFX10-NEXT:    v_perm_b32 v0, v1, v3, 0x4050001
+; GFX10-NEXT:    v_perm_b32 v1, v4, v2, 0x4050001
+; GFX10-NEXT:    v_perm_b32 v2, v6, v5, 0x4050001
 ; GFX10-NEXT:    s_setpc_b64 s[30:31]
 ;
 ; GFX11-LABEL: load_lds_v3i32_align1:
 ; GFX11:       ; %bb.0:
 ; GFX11-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
 ; GFX11-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-NEXT:    ds_load_u8 v1, v0
 ; GFX11-NEXT:    ds_load_u8 v2, v0 offset:1
 ; GFX11-NEXT:    ds_load_u8 v3, v0 offset:2
 ; GFX11-NEXT:    ds_load_u8 v4, v0 offset:3
 ; GFX11-NEXT:    ds_load_u8 v5, v0 offset:4
 ; GFX11-NEXT:    ds_load_u8 v6, v0 offset:5
 ; GFX11-NEXT:    ds_load_u8 v7, v0 offset:6
 ; GFX11-NEXT:    ds_load_u8 v8, v0 offset:7
-; GFX11-NEXT:    ds_load_u8 v9, v0 offset:8
-; GFX11-NEXT:    ds_load_u8 v10, v0 offset:9
-; GFX11-NEXT:    ds_load_u8 v11, v0 offset:10
-; GFX11-NEXT:    ds_load_u8 v0, v0 offset:11
+; GFX11-NEXT:    ds_load_u8 v9, v0 offset:10
+; GFX11-NEXT:    ds_load_u8 v10, v0 offset:11
+; GFX11-NEXT:    ds_load_u8 v11, v0 offset:8
+; GFX11-NEXT:    ds_load_u8 v0, v0 offset:9
 ; GFX11-NEXT:    s_waitcnt lgkmcnt(10)
 ; GFX11-NEXT:    v_lshl_or_b32 v1, v2, 8, v1
 ; GFX11-NEXT:    s_waitcnt lgkmcnt(8)
-; GFX11-NEXT:    v_lshl_or_b32 v2, v4, 8, v3
+; GFX11-NEXT:    v_lshl_or_b32 v3, v4, 8, v3
 ; GFX11-NEXT:    s_waitcnt lgkmcnt(6)
-; GFX11-NEXT:    v_lshl_or_b32 v3, v6, 8, v5
+; GFX11-NEXT:    v_lshl_or_b32 v4, v6, 8, v5
 ; GFX11-NEXT:    s_waitcnt lgkmcnt(4)
-; GFX11-NEXT:    v_lshl_or_b32 v4, v8, 8, v7
+; GFX11-NEXT:    v_lshl_or_b32 v2, v8, 8, v7
 ; GFX11-NEXT:    s_waitcnt lgkmcnt(2)
 ; GFX11-NEXT:    v_lshl_or_b32 v5, v10, 8, v9
 ; GFX11-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX11-NEXT:    v_lshl_or_b32 v6, v0, 8, v11
-; GFX11-NEXT:    v_lshl_or_b32 v0, v2, 16, v1
-; GFX11-NEXT:    v_lshl_or_b32 v1, v4, 16, v3
+; GFX11-NEXT:    v_perm_b32 v0, v1, v3, 0x4050001
+; GFX11-NEXT:    v_perm_b32 v1, v4, v2, 0x4050001
 ; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_3)
-; GFX11-NEXT:    v_lshl_or_b32 v2, v6, 16, v5
+; GFX11-NEXT:    v_perm_b32 v2, v6, v5, 0x4050001
 ; GFX11-NEXT:    s_setpc_b64 s[30:31]
   %load = load <3 x i32>, <3 x i32> addrspace(3)* %ptr, align 1
   ret <3 x i32> %load
 }
 
 define <3 x i32> @load_lds_v3i32_align2(<3 x i32> addrspace(3)* %ptr) {
 ; GFX9-LABEL: load_lds_v3i32_align2:
 ; GFX9:       ; %bb.0:
 ; GFX9-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
 ; GFX9-NEXT:    ds_read_u16 v1, v0
 ; GFX9-NEXT:    ds_read_u16 v2, v0 offset:2
 ; GFX9-NEXT:    ds_read_u16 v3, v0 offset:4
 ; GFX9-NEXT:    ds_read_u16 v4, v0 offset:6
 ; GFX9-NEXT:    ds_read_u16 v5, v0 offset:8
 ; GFX9-NEXT:    ds_read_u16 v6, v0 offset:10
+; GFX9-NEXT:    s_mov_b32 s4, 0x4050001
 ; GFX9-NEXT:    s_waitcnt lgkmcnt(4)
-; GFX9-NEXT:    v_lshl_or_b32 v0, v2, 16, v1
+; GFX9-NEXT:    v_perm_b32 v0, v1, v2, s4
 ; GFX9-NEXT:    s_waitcnt lgkmcnt(2)
-; GFX9-NEXT:    v_lshl_or_b32 v1, v4, 16, v3
+; GFX9-NEXT:    v_perm_b32 v1, v3, v4, s4
 ; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX9-NEXT:    v_lshl_or_b32 v2, v6, 16, v5
+; GFX9-NEXT:    v_perm_b32 v2, v5, v6, s4
 ; GFX9-NEXT:    s_setpc_b64 s[30:31]
 ;
 ; GFX7-LABEL: load_lds_v3i32_align2:
 ; GFX7:       ; %bb.0:
 ; GFX7-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
 ; GFX7-NEXT:    s_mov_b32 m0, -1
 ; GFX7-NEXT:    ds_read_u16 v2, v0 offset:8
 ; GFX7-NEXT:    ds_read_u16 v1, v0 offset:4

 ; GFX10-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX10-NEXT:    ds_read_u16 v1, v0
 ; GFX10-NEXT:    ds_read_u16 v2, v0 offset:2
 ; GFX10-NEXT:    ds_read_u16 v3, v0 offset:4
 ; GFX10-NEXT:    ds_read_u16 v4, v0 offset:6
 ; GFX10-NEXT:    ds_read_u16 v5, v0 offset:8
 ; GFX10-NEXT:    ds_read_u16 v6, v0 offset:10
 ; GFX10-NEXT:    s_waitcnt lgkmcnt(4)
-; GFX10-NEXT:    v_lshl_or_b32 v0, v2, 16, v1
+; GFX10-NEXT:    v_perm_b32 v0, v1, v2, 0x4050001
 ; GFX10-NEXT:    s_waitcnt lgkmcnt(2)
-; GFX10-NEXT:    v_lshl_or_b32 v1, v4, 16, v3
+; GFX10-NEXT:    v_perm_b32 v1, v3, v4, 0x4050001
 ; GFX10-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX10-NEXT:    v_lshl_or_b32 v2, v6, 16, v5
+; GFX10-NEXT:    v_perm_b32 v2, v5, v6, 0x4050001
 ; GFX10-NEXT:    s_setpc_b64 s[30:31]
 ;
 ; GFX11-LABEL: load_lds_v3i32_align2:
 ; GFX11:       ; %bb.0:
 ; GFX11-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
 ; GFX11-NEXT:    s_waitcnt_vscnt null, 0x0
 ; GFX11-NEXT:    ds_load_u16 v1, v0
 ; GFX11-NEXT:    ds_load_u16 v2, v0 offset:2
 ; GFX11-NEXT:    ds_load_u16 v3, v0 offset:4
 ; GFX11-NEXT:    ds_load_u16 v4, v0 offset:6
 ; GFX11-NEXT:    ds_load_u16 v5, v0 offset:8
 ; GFX11-NEXT:    ds_load_u16 v6, v0 offset:10
 ; GFX11-NEXT:    s_waitcnt lgkmcnt(4)
-; GFX11-NEXT:    v_lshl_or_b32 v0, v2, 16, v1
+; GFX11-NEXT:    v_perm_b32 v0, v1, v2, 0x4050001
 ; GFX11-NEXT:    s_waitcnt lgkmcnt(2)
-; GFX11-NEXT:    v_lshl_or_b32 v1, v4, 16, v3
+; GFX11-NEXT:    v_perm_b32 v1, v3, v4, 0x4050001
 ; GFX11-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX11-NEXT:    v_lshl_or_b32 v2, v6, 16, v5
+; GFX11-NEXT:    v_perm_b32 v2, v5, v6, 0x4050001
 ; GFX11-NEXT:    s_setpc_b64 s[30:31]
   %load = load <3 x i32>, <3 x i32> addrspace(3)* %ptr, align 2
   ret <3 x i32> %load
 }
 
 define <3 x i32> @load_lds_v3i32_align4(<3 x i32> addrspace(3)* %ptr) {
 ; GFX9-LABEL: load_lds_v3i32_align4:
 ; GFX9:       ; %bb.0:

llvm/test/CodeGen/AMDGPU/pack.v2f16.ll

 ; GFX8-NEXT:    v_addc_u32_e32 v1, vcc, 0, v1, vcc
 ; GFX8-NEXT:    v_mov_b32_e32 v3, s3
 ; GFX8-NEXT:    v_add_u32_e32 v2, vcc, s2, v2
 ; GFX8-NEXT:    v_addc_u32_e32 v3, vcc, 0, v3, vcc
 ; GFX8-NEXT:    flat_load_dword v0, v[0:1] glc
 ; GFX8-NEXT:    s_waitcnt vmcnt(0)
 ; GFX8-NEXT:    flat_load_dword v1, v[2:3] glc
 ; GFX8-NEXT:    s_waitcnt vmcnt(0)
-; GFX8-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
-; GFX8-NEXT:    v_or_b32_sdwa v0, v0, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
+; GFX8-NEXT:    s_mov_b32 s0, 0x4050001
+; GFX8-NEXT:    v_perm_b32 v0, v0, v1, s0
 ; GFX8-NEXT:    ;;#ASMSTART
 ; GFX8-NEXT:    ; use v0
 ; GFX8-NEXT:    ;;#ASMEND
 ; GFX8-NEXT:    s_endpgm
 ;
 ; GFX7-LABEL: v_pack_v2f16:
 ; GFX7:       ; %bb.0:
 ; GFX7-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0

 ; GFX8-NEXT:    v_addc_u32_e32 v1, vcc, 0, v1, vcc
 ; GFX8-NEXT:    v_mov_b32_e32 v3, s3
 ; GFX8-NEXT:    v_add_u32_e32 v2, vcc, s2, v2
 ; GFX8-NEXT:    v_addc_u32_e32 v3, vcc, 0, v3, vcc
 ; GFX8-NEXT:    flat_load_dword v0, v[0:1] glc
 ; GFX8-NEXT:    s_waitcnt vmcnt(0)
 ; GFX8-NEXT:    flat_load_dword v1, v[2:3] glc
 ; GFX8-NEXT:    s_waitcnt vmcnt(0)
+; GFX8-NEXT:    s_mov_b32 s0, 0x4050001
 ; GFX8-NEXT:    s_mov_b32 s3, 0x1100f000
 ; GFX8-NEXT:    s_mov_b32 s2, -1
-; GFX8-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
-; GFX8-NEXT:    v_or_b32_sdwa v0, v0, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
+; GFX8-NEXT:    v_perm_b32 v0, v0, v1, s0
 ; GFX8-NEXT:    v_add_u32_e32 v0, vcc, 9, v0
 ; GFX8-NEXT:    buffer_store_dword v0, off, s[0:3], 0
 ; GFX8-NEXT:    s_waitcnt vmcnt(0)
 ; GFX8-NEXT:    s_endpgm
 ;
 ; GFX7-LABEL: v_pack_v2f16_user:
 ; GFX7:       ; %bb.0:
 ; GFX7-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0

llvm/test/CodeGen/AMDGPU/pack.v2i16.ll

 ; GFX803-NEXT:    v_addc_u32_e32 v1, vcc, 0, v1, vcc
 ; GFX803-NEXT:    v_mov_b32_e32 v3, s3
 ; GFX803-NEXT:    v_add_u32_e32 v2, vcc, s2, v2
 ; GFX803-NEXT:    v_addc_u32_e32 v3, vcc, 0, v3, vcc
 ; GFX803-NEXT:    flat_load_dword v0, v[0:1] glc
 ; GFX803-NEXT:    s_waitcnt vmcnt(0)
 ; GFX803-NEXT:    flat_load_dword v1, v[2:3] glc
 ; GFX803-NEXT:    s_waitcnt vmcnt(0)
-; GFX803-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
-; GFX803-NEXT:    v_or_b32_sdwa v0, v0, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
+; GFX803-NEXT:    s_mov_b32 s0, 0x4050001
+; GFX803-NEXT:    v_perm_b32 v0, v0, v1, s0
 ; GFX803-NEXT:    ;;#ASMSTART
 ; GFX803-NEXT:    ; use v0
 ; GFX803-NEXT:    ;;#ASMEND
 ; GFX803-NEXT:    s_endpgm
 ;
 ; GFX7-LABEL: v_pack_v2i16:
 ; GFX7:       ; %bb.0:
 ; GFX7-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0

 ; GFX803-NEXT:    v_addc_u32_e32 v1, vcc, 0, v1, vcc
 ; GFX803-NEXT:    v_mov_b32_e32 v3, s3
 ; GFX803-NEXT:    v_add_u32_e32 v2, vcc, s2, v2
 ; GFX803-NEXT:    v_addc_u32_e32 v3, vcc, 0, v3, vcc
 ; GFX803-NEXT:    flat_load_dword v0, v[0:1] glc
 ; GFX803-NEXT:    s_waitcnt vmcnt(0)
 ; GFX803-NEXT:    flat_load_dword v1, v[2:3] glc
 ; GFX803-NEXT:    s_waitcnt vmcnt(0)
+; GFX803-NEXT:    s_mov_b32 s0, 0x4050001
 ; GFX803-NEXT:    s_mov_b32 s3, 0x1100f000
 ; GFX803-NEXT:    s_mov_b32 s2, -1
-; GFX803-NEXT:    v_lshlrev_b32_e32 v1, 16, v1
-; GFX803-NEXT:    v_or_b32_sdwa v0, v0, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
+; GFX803-NEXT:    v_perm_b32 v0, v0, v1, s0
 ; GFX803-NEXT:    v_add_u32_e32 v0, vcc, 9, v0
 ; GFX803-NEXT:    buffer_store_dword v0, off, s[0:3], 0
 ; GFX803-NEXT:    s_waitcnt vmcnt(0)
 ; GFX803-NEXT:    s_endpgm
 ;
 ; GFX7-LABEL: v_pack_v2i16_user:
 ; GFX7:       ; %bb.0:
 ; GFX7-NEXT:    s_load_dwordx4 s[0:3], s[4:5], 0x0

llvm/test/CodeGen/AMDGPU/permute.ll

 }
 
 define amdgpu_kernel void @or_and_or(i32 addrspace(1)* nocapture %arg, i32 %arg1) {
 ; GCN-LABEL: or_and_or:
 ; GCN:       ; %bb.0: ; %bb
 ; GCN-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x24
 ; GCN-NEXT:    s_load_dword s0, s[0:1], 0x2c
 ; GCN-NEXT:    v_lshlrev_b32_e32 v0, 2, v0
-; GCN-NEXT:    v_mov_b32_e32 v3, 0x7020104
+; GCN-NEXT:    v_mov_b32_e32 v3, 0x4050607
 ; GCN-NEXT:    s_waitcnt lgkmcnt(0)
 ; GCN-NEXT:    v_mov_b32_e32 v1, s3
 ; GCN-NEXT:    v_add_u32_e32 v0, vcc, s2, v0
 ; GCN-NEXT:    v_addc_u32_e32 v1, vcc, 0, v1, vcc
 ; GCN-NEXT:    flat_load_dword v2, v[0:1]
+; GCN-NEXT:    s_or_b32 s0, s0, 0xff0000ff
 ; GCN-NEXT:    s_waitcnt vmcnt(0)
-; GCN-NEXT:    v_perm_b32 v2, v2, s0, v3
+; GCN-NEXT:    v_perm_b32 v2, v2, 0, v3
+; GCN-NEXT:    v_and_b32_e32 v2, s0, v2
 ; GCN-NEXT:    flat_store_dword v[0:1], v2
 ; GCN-NEXT:    s_endpgm
 bb:
   %id = tail call i32 @llvm.amdgcn.workitem.id.x()
   %gep = getelementptr i32, i32 addrspace(1)* %arg, i32 %id
   %tmp = load i32, i32 addrspace(1)* %gep, align 4
   %or1 = or i32 %tmp, 16776960    ; 0x00ffff00
   %or2 = or i32 %arg1, 4278190335 ; 0xff0000ff
   %and = and i32 %or1, %or2
   store i32 %and, i32 addrspace(1)* %gep, align 4
   ret void
 }
 
 ; FIXME here should have been "v_perm_b32" with 0xffff0500 mask.
 define amdgpu_kernel void @known_ffff0500(i32 addrspace(1)* nocapture %arg, i32 %arg1) {
 ; GCN-LABEL: known_ffff0500:
 ; GCN:       ; %bb.0: ; %bb
 ; GCN-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x24
 ; GCN-NEXT:    s_load_dword s0, s[0:1], 0x2c
 ; GCN-NEXT:    v_lshlrev_b32_e32 v0, 2, v0
-; GCN-NEXT:    v_mov_b32_e32 v5, 0xffff8004
+; GCN-NEXT:    v_mov_b32_e32 v5, 0xc050c07
+; GCN-NEXT:    v_mov_b32_e32 v6, 0xffff8004
 ; GCN-NEXT:    s_waitcnt lgkmcnt(0)
 ; GCN-NEXT:    v_mov_b32_e32 v1, s3
 ; GCN-NEXT:    v_add_u32_e32 v0, vcc, s2, v0
 ; GCN-NEXT:    v_addc_u32_e32 v1, vcc, 0, v1, vcc
 ; GCN-NEXT:    flat_load_dword v4, v[0:1]
 ; GCN-NEXT:    s_bitset1_b32 s0, 15
 ; GCN-NEXT:    s_and_b32 s0, s0, 0xff00
 ; GCN-NEXT:    s_or_b32 s0, s0, 0xffff0000
 ; GCN-NEXT:    v_mov_b32_e32 v2, s2
 ; GCN-NEXT:    v_mov_b32_e32 v3, s3
 ; GCN-NEXT:    s_waitcnt vmcnt(0)
-; GCN-NEXT:    v_or_b32_e32 v4, 4, v4
-; GCN-NEXT:    v_and_b32_e32 v4, 0xff00ff, v4
+; GCN-NEXT:    v_perm_b32 v4, v4, 0, v5
 ; GCN-NEXT:    v_or_b32_e32 v4, s0, v4
 ; GCN-NEXT:    flat_store_dword v[0:1], v4
-; GCN-NEXT:    flat_store_dword v[2:3], v5
+; GCN-NEXT:    flat_store_dword v[2:3], v6
 ; GCN-NEXT:    s_endpgm
 bb:
   %id = tail call i32 @llvm.amdgcn.workitem.id.x()
   %gep = getelementptr i32, i32 addrspace(1)* %arg, i32 %id
   %load = load i32, i32 addrspace(1)* %gep, align 4
   %mask1 = or i32 %arg1, 32768 ; 0x8000
   %mask2 = or i32 %load, 4
   %and = and i32 %mask2, 16711935     ; 0x00ff00ff

 }
 
 define amdgpu_kernel void @known_ffff8004(i32 addrspace(1)* nocapture %arg, i32 %arg1) {
 ; GCN-LABEL: known_ffff8004:
 ; GCN:       ; %bb.0: ; %bb
 ; GCN-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x24
 ; GCN-NEXT:    s_load_dword s0, s[0:1], 0x2c
 ; GCN-NEXT:    v_lshlrev_b32_e32 v0, 2, v0
-; GCN-NEXT:    v_mov_b32_e32 v5, 0xffff0500
-; GCN-NEXT:    v_mov_b32_e32 v6, 0xffff8004
+; GCN-NEXT:    v_mov_b32_e32 v5, 0x4050607
+; GCN-NEXT:    v_mov_b32_e32 v6, 0xffff0500
 ; GCN-NEXT:    s_waitcnt lgkmcnt(0)
 ; GCN-NEXT:    v_mov_b32_e32 v1, s3
 ; GCN-NEXT:    v_add_u32_e32 v0, vcc, s2, v0
 ; GCN-NEXT:    v_addc_u32_e32 v1, vcc, 0, v1, vcc
 ; GCN-NEXT:    flat_load_dword v4, v[0:1]
 ; GCN-NEXT:    s_or_b32 s0, s0, 4
 ; GCN-NEXT:    v_mov_b32_e32 v2, s2
+; GCN-NEXT:    v_mov_b32_e32 v7, 0xffff8004
 ; GCN-NEXT:    v_mov_b32_e32 v3, s3
 ; GCN-NEXT:    s_waitcnt vmcnt(0)
-; GCN-NEXT:    v_or_b32_e32 v4, 0x8000, v4
-; GCN-NEXT:    v_perm_b32 v4, v4, s0, v5
+; GCN-NEXT:    v_perm_b32 v4, v4, 0, v5
+; GCN-NEXT:    v_perm_b32 v4, v4, s0, v6
 ; GCN-NEXT:    flat_store_dword v[0:1], v4
-; GCN-NEXT:    flat_store_dword v[2:3], v6
+; GCN-NEXT:    flat_store_dword v[2:3], v7
 ; GCN-NEXT:    s_endpgm
 bb:
   %id = tail call i32 @llvm.amdgcn.workitem.id.x()
   %gep = getelementptr i32, i32 addrspace(1)* %arg, i32 %id
   %load = load i32, i32 addrspace(1)* %gep, align 4
   %mask1 = or i32 %arg1, 4
   %mask2 = or i32 %load, 32768 ; 0x8000
   %and = and i32 %mask1, 16711935     ; 0x00ff00ff

llvm/test/CodeGen/AMDGPU/permute_i8.ll

+; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py
+; RUN: llc -mtriple=amdgcn-- -mcpu=gfx1010 -verify-machineinstrs < %s | FileCheck %s -check-prefixes=GFX10
+; RUN: llc -mtriple=amdgcn-- -mcpu=gfx908  -verify-machineinstrs < %s | FileCheck %s -check-prefixes=GFX9
+
+define hidden void @shuffle6766(<4 x i8>* %in0, <4 x i8>* %in1, <4 x i8>* %out0) {
+; GFX10-LABEL: shuffle6766:
+; GFX10:       ; %bb.0:
+; GFX10-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT:    flat_load_dword v0, v[2:3]
+; GFX10-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-NEXT:    v_perm_b32 v0, v0, 0, 0x6070606
+; GFX10-NEXT:    flat_store_dword v[4:5], v0
+; GFX10-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT:    s_setpc_b64 s[30:31]
+;
+; GFX9-LABEL: shuffle6766:
+; GFX9:       ; %bb.0:
+; GFX9-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX9-NEXT:    flat_load_dword v0, v[2:3]
+; GFX9-NEXT:    v_mov_b32_e32 v1, 0x6070606
+; GFX9-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX9-NEXT:    v_perm_b32 v0, v0, 0, v1
+; GFX9-NEXT:    flat_store_dword v[4:5], v0
+; GFX9-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX9-NEXT:    s_setpc_b64 s[30:31]
+  %vec0 = load <4 x i8>, <4 x i8>* %in0, align 4
+  %vec1 = load <4 x i8>, <4 x i8>* %in1, align 4
+   %shuffle0_0 = shufflevector <4 x i8> %vec0, <4 x i8> %vec1, <4 x i32> <i32 6, i32 7, i32 6, i32 6>
+  store <4 x i8> %shuffle0_0, <4 x i8>* %out0, align 4
+  ret void
+}
+
+define hidden void @shuffle3746(<4 x i8>* %in0, <4 x i8>* %in1, <4 x i8>* %out0) {
+; GFX10-LABEL: shuffle3746:
+; GFX10:       ; %bb.0:
+; GFX10-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT:    flat_load_dword v6, v[0:1]
+; GFX10-NEXT:    flat_load_dword v7, v[2:3]
+; GFX10-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-NEXT:    v_perm_b32 v0, v6, v7, 0x7030000
+; GFX10-NEXT:    flat_store_dword v[4:5], v0
+; GFX10-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT:    s_setpc_b64 s[30:31]
+;
+; GFX9-LABEL: shuffle3746:
+; GFX9:       ; %bb.0:
+; GFX9-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX9-NEXT:    flat_load_dword v6, v[0:1]
+; GFX9-NEXT:    flat_load_dword v7, v[2:3]
+; GFX9-NEXT:    s_mov_b32 s4, 0x7030000
+; GFX9-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX9-NEXT:    v_perm_b32 v0, v6, v7, s4
+; GFX9-NEXT:    flat_store_dword v[4:5], v0
+; GFX9-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX9-NEXT:    s_setpc_b64 s[30:31]
+  %vec0 = load <4 x i8>, <4 x i8>* %in0, align 4
+  %vec1 = load <4 x i8>, <4 x i8>* %in1, align 4
+   %shuffle0_0 = shufflevector <4 x i8> %vec0, <4 x i8> %vec1, <4 x i32> <i32 3, i32 7, i32 4, i32 4>
+  store <4 x i8> %shuffle0_0, <4 x i8>* %out0, align 4
+  ret void
+}
+
+define hidden void @shuffle4445(<4 x i8>* %in0, <4 x i8>* %in1, <4 x i8>* %out0) {
+; GFX10-LABEL: shuffle4445:
+; GFX10:       ; %bb.0:
+; GFX10-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT:    flat_load_dword v0, v[2:3]
+; GFX10-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-NEXT:    v_perm_b32 v0, v0, 0, 0x4040405
+; GFX10-NEXT:    flat_store_dword v[4:5], v0
+; GFX10-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT:    s_setpc_b64 s[30:31]
+;
+; GFX9-LABEL: shuffle4445:
+; GFX9:       ; %bb.0:
+; GFX9-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX9-NEXT:    flat_load_dword v0, v[2:3]
+; GFX9-NEXT:    v_mov_b32_e32 v1, 0x4040405
+; GFX9-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX9-NEXT:    v_perm_b32 v0, v0, 0, v1
+; GFX9-NEXT:    flat_store_dword v[4:5], v0
+; GFX9-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX9-NEXT:    s_setpc_b64 s[30:31]
+  %vec0 = load <4 x i8>, <4 x i8>* %in0, align 4
+  %vec1 = load <4 x i8>, <4 x i8>* %in1, align 4
+   %shuffle0_0 = shufflevector <4 x i8> %vec0, <4 x i8> %vec1, <4 x i32> <i32 4, i32 4, i32 4, i32 5>
+  store <4 x i8> %shuffle0_0, <4 x i8>* %out0, align 4
+  ret void
+}
+
+define hidden void @shuffle0101(<4 x i8>* %in0, <4 x i8>* %in1, <4 x i8>* %out0) {
+; GFX10-LABEL: shuffle0101:
+; GFX10:       ; %bb.0:
+; GFX10-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT:    flat_load_dword v0, v[0:1]
+; GFX10-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-NEXT:    v_perm_b32 v0, v0, 0, 0x4050405
+; GFX10-NEXT:    flat_store_dword v[4:5], v0
+; GFX10-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT:    s_setpc_b64 s[30:31]
+;
+; GFX9-LABEL: shuffle0101:
+; GFX9:       ; %bb.0:
+; GFX9-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX9-NEXT:    flat_load_dword v0, v[0:1]
+; GFX9-NEXT:    v_mov_b32_e32 v1, 0x4050405
+; GFX9-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX9-NEXT:    v_perm_b32 v0, v0, 0, v1
+; GFX9-NEXT:    flat_store_dword v[4:5], v0
+; GFX9-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX9-NEXT:    s_setpc_b64 s[30:31]
+  %vec0 = load <4 x i8>, <4 x i8>* %in0, align 4
+  %vec1 = load <4 x i8>, <4 x i8>* %in1, align 4
+   %shuffle0_0 = shufflevector <4 x i8> %vec0, <4 x i8> %vec1, <4 x i32> <i32 0, i32 1, i32 0, i32 1>
+  store <4 x i8> %shuffle0_0, <4 x i8>* %out0, align 4
+  ret void
+}
+
+define hidden void @shuffle7533(<4 x i8>* %in0, <4 x i8>* %in1, <4 x i8>* %out0) {
+; GFX10-LABEL: shuffle7533:
+; GFX10:       ; %bb.0:
+; GFX10-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT:    flat_load_dword v6, v[0:1]
+; GFX10-NEXT:    flat_load_dword v7, v[2:3]
+; GFX10-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-NEXT:    v_perm_b32 v0, v7, v6, 0x7050303
+; GFX10-NEXT:    flat_store_dword v[4:5], v0
+; GFX10-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT:    s_setpc_b64 s[30:31]
+;
+; GFX9-LABEL: shuffle7533:
+; GFX9:       ; %bb.0:
+; GFX9-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX9-NEXT:    flat_load_dword v6, v[0:1]
+; GFX9-NEXT:    flat_load_dword v7, v[2:3]
+; GFX9-NEXT:    s_mov_b32 s4, 0x7050303
+; GFX9-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX9-NEXT:    v_perm_b32 v0, v7, v6, s4
+; GFX9-NEXT:    flat_store_dword v[4:5], v0
+; GFX9-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX9-NEXT:    s_setpc_b64 s[30:31]
+  %vec0 = load <4 x i8>, <4 x i8>* %in0, align 4
+  %vec1 = load <4 x i8>, <4 x i8>* %in1, align 4
+   %shuffle0_0 = shufflevector <4 x i8> %vec0, <4 x i8> %vec1, <4 x i32> <i32 7, i32 5, i32 3, i32 3>
+  store <4 x i8> %shuffle0_0, <4 x i8>* %out0, align 4
+  ret void
+}
+
+define hidden void @shuffle7767(<4 x i8>* %in0, <4 x i8>* %in1, <4 x i8>* %out0) {
+; GFX10-LABEL: shuffle7767:
+; GFX10:       ; %bb.0:
+; GFX10-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT:    flat_load_dword v0, v[2:3]
+; GFX10-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-NEXT:    v_perm_b32 v0, v0, 0, 0x7070607
+; GFX10-NEXT:    flat_store_dword v[4:5], v0
+; GFX10-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT:    s_setpc_b64 s[30:31]
+;
+; GFX9-LABEL: shuffle7767:
+; GFX9:       ; %bb.0:
+; GFX9-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX9-NEXT:    flat_load_dword v0, v[2:3]
+; GFX9-NEXT:    v_mov_b32_e32 v1, 0x7070607
+; GFX9-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX9-NEXT:    v_perm_b32 v0, v0, 0, v1
+; GFX9-NEXT:    flat_store_dword v[4:5], v0
+; GFX9-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX9-NEXT:    s_setpc_b64 s[30:31]
+  %vec0 = load <4 x i8>, <4 x i8>* %in0, align 4
+  %vec1 = load <4 x i8>, <4 x i8>* %in1, align 4
+   %shuffle0_0 = shufflevector <4 x i8> %vec0, <4 x i8> %vec1, <4 x i32> <i32 7, i32 7, i32 6, i32 7>
+  store <4 x i8> %shuffle0_0, <4 x i8>* %out0, align 4
+  ret void
+}
+
+define hidden void @shuffle0554(<4 x i8>* %in0, <4 x i8>* %in1, <4 x i8>* %out0) {
+; GFX10-LABEL: shuffle0554:
+; GFX10:       ; %bb.0:
+; GFX10-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT:    flat_load_dword v6, v[0:1]
+; GFX10-NEXT:    flat_load_dword v7, v[2:3]
+; GFX10-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-NEXT:    v_perm_b32 v0, v6, v7, 0x4010100
+; GFX10-NEXT:    flat_store_dword v[4:5], v0
+; GFX10-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT:    s_setpc_b64 s[30:31]
+;
+; GFX9-LABEL: shuffle0554:
+; GFX9:       ; %bb.0:
+; GFX9-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX9-NEXT:    flat_load_dword v6, v[0:1]
+; GFX9-NEXT:    flat_load_dword v7, v[2:3]
+; GFX9-NEXT:    s_mov_b32 s4, 0x4010100
+; GFX9-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX9-NEXT:    v_perm_b32 v0, v6, v7, s4
+; GFX9-NEXT:    flat_store_dword v[4:5], v0
+; GFX9-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX9-NEXT:    s_setpc_b64 s[30:31]
+  %vec0 = load <4 x i8>, <4 x i8>* %in0, align 4
+  %vec1 = load <4 x i8>, <4 x i8>* %in1, align 4
+   %shuffle0_0 = shufflevector <4 x i8> %vec0, <4 x i8> %vec1, <4 x i32> <i32 0, i32 5, i32 5, i32 4>
+  store <4 x i8> %shuffle0_0, <4 x i8>* %out0, align 4
+  ret void
+}
+
+define hidden void @shuffle2127(<4 x i8>* %in0, <4 x i8>* %in1, <4 x i8>* %out0) {
+; GFX10-LABEL: shuffle2127:
+; GFX10:       ; %bb.0:
+; GFX10-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT:    flat_load_dword v6, v[0:1]
+; GFX10-NEXT:    flat_load_dword v7, v[2:3]
+; GFX10-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-NEXT:    v_perm_b32 v0, v6, v7, 0x6050603
+; GFX10-NEXT:    flat_store_dword v[4:5], v0
+; GFX10-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT:    s_setpc_b64 s[30:31]
+;
+; GFX9-LABEL: shuffle2127:
+; GFX9:       ; %bb.0:
+; GFX9-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX9-NEXT:    flat_load_dword v6, v[0:1]
+; GFX9-NEXT:    flat_load_dword v7, v[2:3]
+; GFX9-NEXT:    s_mov_b32 s4, 0x6050603
+; GFX9-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX9-NEXT:    v_perm_b32 v0, v6, v7, s4
+; GFX9-NEXT:    flat_store_dword v[4:5], v0
+; GFX9-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX9-NEXT:    s_setpc_b64 s[30:31]
+  %vec0 = load <4 x i8>, <4 x i8>* %in0, align 4
+  %vec1 = load <4 x i8>, <4 x i8>* %in1, align 4
+   %shuffle0_0 = shufflevector <4 x i8> %vec0, <4 x i8> %vec1, <4 x i32> <i32 2, i32 1, i32 2, i32 7>
+  store <4 x i8> %shuffle0_0, <4 x i8>* %out0, align 4
+  ret void
+}
+
+define hidden void @shuffle5047(<4 x i8>* %in0, <4 x i8>* %in1, <4 x i8>* %out0) {
+; GFX10-LABEL: shuffle5047:
+; GFX10:       ; %bb.0:
+; GFX10-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT:    flat_load_dword v6, v[0:1]
+; GFX10-NEXT:    flat_load_dword v7, v[2:3]
+; GFX10-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-NEXT:    v_perm_b32 v0, v7, v6, 0x5000407
+; GFX10-NEXT:    flat_store_dword v[4:5], v0
+; GFX10-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT:    s_setpc_b64 s[30:31]
+;
+; GFX9-LABEL: shuffle5047:
+; GFX9:       ; %bb.0:
+; GFX9-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX9-NEXT:    flat_load_dword v6, v[0:1]
+; GFX9-NEXT:    flat_load_dword v7, v[2:3]
+; GFX9-NEXT:    s_mov_b32 s4, 0x5000407
+; GFX9-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX9-NEXT:    v_perm_b32 v0, v7, v6, s4
+; GFX9-NEXT:    flat_store_dword v[4:5], v0
+; GFX9-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX9-NEXT:    s_setpc_b64 s[30:31]
+  %vec0 = load <4 x i8>, <4 x i8>* %in0, align 4
+  %vec1 = load <4 x i8>, <4 x i8>* %in1, align 4
+   %shuffle0_0 = shufflevector <4 x i8> %vec0, <4 x i8> %vec1, <4 x i32> <i32 5, i32 0, i32 4, i32 7>
+  store <4 x i8> %shuffle0_0, <4 x i8>* %out0, align 4
+  ret void
+}
+
+define hidden void @shuffle3546(<4 x i8>* %in0, <4 x i8>* %in1, <4 x i8>* %out0) {
+; GFX10-LABEL: shuffle3546:
+; GFX10:       ; %bb.0:
+; GFX10-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX10-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT:    flat_load_dword v6, v[0:1]
+; GFX10-NEXT:    flat_load_dword v7, v[2:3]
+; GFX10-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-NEXT:    v_perm_b32 v0, v6, v7, 0x7010002
+; GFX10-NEXT:    flat_store_dword v[4:5], v0
+; GFX10-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-NEXT:    s_waitcnt_vscnt null, 0x0
+; GFX10-NEXT:    s_setpc_b64 s[30:31]
+;
+; GFX9-LABEL: shuffle3546:
+; GFX9:       ; %bb.0:
+; GFX9-NEXT:    s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX9-NEXT:    flat_load_dword v6, v[0:1]
+; GFX9-NEXT:    flat_load_dword v7, v[2:3]
+; GFX9-NEXT:    s_mov_b32 s4, 0x7010002
+; GFX9-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX9-NEXT:    v_perm_b32 v0, v6, v7, s4
+; GFX9-NEXT:    flat_store_dword v[4:5], v0
+; GFX9-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX9-NEXT:    s_setpc_b64 s[30:31]
+  %vec0 = load <4 x i8>, <4 x i8>* %in0, align 4
+  %vec1 = load <4 x i8>, <4 x i8>* %in1, align 4
+  %shuffle0_0 = shufflevector <4 x i8> %vec0, <4 x i8> %vec1, <4 x i32> <i32 3, i32 5, i32 4, i32 6>
+  store <4 x i8> %shuffle0_0, <4 x i8>* %out0, align 4
+  ret void
+}
+