[AMDGPU]: Allow combining into v_dot4

[jrbyrnes]

Summary

Adds the algorithm to match and select v_dot4 instructions in combining, and removes the patterns from selection. The patterns are fragile, and fail to match when byte extraction code is slightly different, or any optimizations alters the add / mul structure of the tree. The DAG combining approach is more flexible, and should not result in much overhead given all the early exits.

For kernels that should select into these instructions, doing so is vitally important. Not only is performance much improved, but failing to select into them can result in severe code bloat which drastically degrades compile time.

The extended perm matching is a happy consequence of whitelisting EXTRACT_VECT_ELT i32s as ultimate srcs of bytes.

[jrbyrnes]

llvm/lib/Target/AMDGPU/VOP3PInstructions.td

293 ↗

(On Diff #543065)

Should we delete these?

[jrbyrnes]

Fix some errors.

[jrbyrnes]

passes psdb

[jrbyrnes]

Nothing necessarily planned at the moment, just want to block the review for now.

It may make more sense to tune vectorization cost model (for i8 and potentially i16) to produce something like

%m = mul < n x i8> %v0, %v1
%o = llvm.vector.reduce.add.vni8(%m)
%op.rdx = add %o, %scalar

Then lower to mfma or v_dot in CodeGenPrepare.

e.g.

%op.rdx = v_dot4_i32_i8 %v0, %v1, %scalar

Instead of scalarizing the sequence and trying to combine all possible variants.

Need to finish investigation before unblocking review.

[jrbyrnes]

SLP vectorization should be tuned but that seems like a separate issue. Trees corresponding to v_dot4 often have s/zext as the final dest is 32 bit, but the arithmetic operations involve 8 bit operands. By introducing s/zext into the tree, we confuse the SLP vectorization cost model as it thinks it is vectorizing 32bit operands. The main issue is that cost model only looks at one node of the vectorizable tree at a time to calculate cost, instead of also considering the sequence as a whole. If we were to vectorize, codegen may be significantly less complex for these.

Plan is to move forward with this patch, and potentially tune vectorization in later work.

[jrbyrnes]

Rebase + Extended algorithm for more complete coverage of potential trees.

Still a WIP while I determine if it is coverage / feature complete, and need to make lit testing more robust.

[jrbyrnes]

Rebase + clean up code. Still running tests but no longer a WIP.

[jrbyrnes]

Fix non-determinism -- iteration order of DenseMap. Use SmallVector instead (worst case lookup is non factor due to size)

[arsenm]

llvm/lib/Target/AMDGPU/SIISelLowering.cpp

12755

don't need .has_value() part

12759

Ditto

12828–12829

is_contained

12913

1 || == 2?

12985

Don't know why we have getSubtarget, you can just use Subtarget-> directly

12992

SDValue TempNode(N, 0)

12994–12997

Don't understand what you are doing with this opcode to index check

13022

Don't need has_value()

13039

Don't need has_value

13083–13084

is_contained

13109

use deleted_node instead of optional opcode

13110

Can you avoid generation checks?

13116

Can you go through the intrinsics instead of going straight to the machine node?

[jrbyrnes]

Rebase (for https://reviews.llvm.org/D158468) and lower with intrinsics.

Allowing combining in pre-legalize phase.

llvm/lib/Target/AMDGPU/SIISelLowering.cpp

12828–12829

We use the found value here

12994–12997

It is a convenience which allows getting the mul operand without having to do operand checks every time.

[arsenm]

llvm/lib/Target/AMDGPU/SIISelLowering.cpp

12921

auto &[Something, Mask] : ?

12928–12944

I thought we combined the mul24 intrinsics to the nodes specifically so you don't need to do this

13117–13118

DAG.getTargetConstant(IsSigned ? Intrinsic::amdgcn_sdot4 : Intrinsic::amdgcn_udot4, SL, MVT::i32)

[jrbyrnes]

Address Comments

[jrbyrnes]

ping

[arsenm]

llvm/lib/Target/AMDGPU/SIISelLowering.cpp

12764

Missing newline

12861

maybe comment skipping bswap?

13023

extra parens

[jrbyrnes]

Address comments

llvm/lib/Target/AMDGPU/SIISelLowering.cpp

12861

It's the original SDValue, not bswap. I'll add comment.

def : GCNPat <
  (i32 (bswap i32:$a)),
  (V_PERM_B32_e64 (i32 0), VSrc_b32:$a, (S_MOV_B32 (i32 0x00010203)))
>;

[jrbyrnes]

Reopen for review as it has been reverted, now includes https://github.com/llvm/llvm-project/pull/65995

[jrbyrnes]

Bring in https://github.com/llvm/llvm-project/pull/65995

[jrbyrnes]

IsSigned tracks whether or not to produce an instruction with signed behavior. In some cases, we are able to determine this based on the semantics of the top-level instruction, however, in other cases, we need more information. For such cases, we must look to the tree itself.

In cases of vectorized arithmetic reduction instructions, we typically see a widening of type. We can use the signedness of the extension to determine the signedness semantics required for the instruction we will ultimately produce.

However, other clients of calculateByteProvider may not be producing arithmetic instructions, and, in these cases, there may be no requirement to track the signedness semantics. Thus, we do not need to fail if we are unable to detemine the signedness.

This removes the requirement that ByteProviders must determine IsSigned.

[jrbyrnes]

Extract signedness checking

[jrbyrnes]

Fix dereference issue + nits (reorganize logic + comments)

[jrbyrnes]

Fix signedness handling of any_extend

[bcahoon]

Just some minor comments/questions.

llvm/lib/Target/AMDGPU/SIISelLowering.cpp

11074

add period at the end.

12751

This function is used for any mul, not just mul24?

12946

I think the check includes any extend (or unknown?) as well as signed?

Maybe say, If we have MUL_u24 without unsigned semantics, then fail.

12950

BTW, is this the same as :
if (!Src0.IsSigned.value_or(false) && MulOpcode == AMDGPUISD::MUL_I24)

[jrbyrnes]

Address comments + update handling of AtomicSDNode + MemIntrinsic

llvm/lib/Target/AMDGPU/SIISelLowering.cpp

12946

This check is checking for conflicting signedness semantics. In the case where we don't have signedness info from the ByteProvider, the signedness is irrelevant so we should say it doesn't conflict.

In the case where we don't have signedness information, then two things could have occurred:

1. We are accumulating into 8 bit register, and have not done any extensions. In this case, the upper bits are irrelevant, and we may use either version of the dot.
2. We have exclusively used any_extends. Same as case 1, the upper bits are irrelevant.

There is a third scenario which I have accounted for in the latest version. Previously, we would not have signedness info if we encountered an unhandled node (MemIntrinsic / AtomicSDNode). However, in this case, the upper bits may be relevant. Thus, instead of throwing away signedness info in this situations, we now fail.

12950

Right nice catch -- except we should not fail if we don't have signedness info.

llvm/test/CodeGen/AMDGPU/idot4s.ll

146

Answering offline question about "neg_lo:[1,1,0]"

These tests are introduced via this patch, and the modifier indicates we need signedness semantics for both operands. neg_lo:[1,1,0] was introduced originally by rebasing on top of changes from https://reviews.llvm.org/D158468?vs=552172&id=552466#toc I believe.

[bcahoon]

Changes LGTM.

llvm/include/llvm/CodeGen/ByteProvider.h

 /// some other productive instruction (e.g. arithmetic instructions).
 /// Bit manipulation instructions like shifts are not ByteProviders, rather
 /// are used to extract Bytes.
 template <typename ISelOp> class ByteProvider {
 private:
   ByteProvider(std::optional<ISelOp> Src, int64_t DestOffset, int64_t SrcOffset)
       : Src(Src), DestOffset(DestOffset), SrcOffset(SrcOffset) {}
 
+  ByteProvider(std::optional<ISelOp> Src, int64_t DestOffset, int64_t SrcOffset,
+               std::optional<bool> IsSigned)
+      : Src(Src), DestOffset(DestOffset), SrcOffset(SrcOffset),
+        IsSigned(IsSigned) {}
+
   // TODO -- use constraint in c++20
   // Does this type correspond with an operation in selection DAG
   template <typename T> class is_op {
   private:
     using yes = std::true_type;
     using no = std::false_type;
 
     // Only allow classes with member function getOpcode
public:
   // Src represents the node which originally produced the relevant bits.
   std::optional<ISelOp> Src = std::nullopt;
   // DestOffset is the offset of the byte in the dest we are trying to map for.
   int64_t DestOffset = 0;
   // SrcOffset is the offset in the ultimate source node that maps to the
   // DestOffset
   int64_t SrcOffset = 0;
 
+  // Whether or not the path to this Src involved signed extensions
+  std::optional<bool> IsSigned;
+
   ByteProvider() = default;
 
   static ByteProvider getSrc(std::optional<ISelOp> Val, int64_t ByteOffset,
                              int64_t VectorOffset) {
     static_assert(is_op<ISelOp>().value,
                   "ByteProviders must contain an operation in selection DAG.");
     return ByteProvider(Val, ByteOffset, VectorOffset);
   }
 
+  static ByteProvider getSrc(std::optional<ISelOp> Val, int64_t ByteOffset,
+                             int64_t VectorOffset,
+                             std::optional<bool> IsSigned) {
+    static_assert(is_op<ISelOp>().value,
+                  "ByteProviders must contain an operation in selection DAG.");
+    return ByteProvider(Val, ByteOffset, VectorOffset, IsSigned);
+  }
+
   static ByteProvider getConstantZero() {
     return ByteProvider<ISelOp>(std::nullopt, 0, 0);
   }
   bool isConstantZero() const { return !Src; }
 
   bool hasSrc() const { return Src.has_value(); }
 
   bool hasSameSrc(const ByteProvider &Other) const { return Other.Src == Src; }

llvm/lib/Target/AMDGPU/SIISelLowering.cpp

 
 // 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 std::optional<ByteProvider<SDValue>>
 calculateSrcByte(const SDValue Op, uint64_t DestByte, uint64_t SrcIndex = 0,
+                 std::optional<bool> IsSigned = std::nullopt,
                  unsigned Depth = 0) {
   // We may need to recursively traverse a series of SRLs
   if (Depth >= 6)
     return std::nullopt;
 
   auto ValueSize = Op.getValueSizeInBits();
   if (ValueSize != 8 && ValueSize != 16 && ValueSize != 32)
     return std::nullopt;
 
   switch (Op->getOpcode()) {
   case ISD::TRUNCATE: {
-    return calculateSrcByte(Op->getOperand(0), DestByte, SrcIndex, Depth + 1);
+    return calculateSrcByte(Op->getOperand(0), DestByte, SrcIndex, IsSigned,
+                            Depth + 1);
   }
 
   case ISD::SIGN_EXTEND:
   case ISD::ZERO_EXTEND:
   case ISD::SIGN_EXTEND_INREG: {
+    IsSigned = IsSigned.value_or(false) ||
+               Op->getOpcode() == ISD::SIGN_EXTEND ||
+               Op->getOpcode() == ISD::SIGN_EXTEND_INREG;
     SDValue NarrowOp = Op->getOperand(0);
     auto NarrowVT = NarrowOp.getValueType();
     if (Op->getOpcode() == ISD::SIGN_EXTEND_INREG) {
       auto *VTSign = cast<VTSDNode>(Op->getOperand(1));
       NarrowVT = VTSign->getVT();
     }
     if (!NarrowVT.isByteSized())
       return std::nullopt;
     uint64_t NarrowByteWidth = NarrowVT.getStoreSize();
 
     if (SrcIndex >= NarrowByteWidth)
       return std::nullopt;
-    return calculateSrcByte(Op->getOperand(0), DestByte, SrcIndex, Depth + 1);
+    return calculateSrcByte(Op->getOperand(0), DestByte, SrcIndex, IsSigned,
+                            Depth + 1);
   }
 
   case ISD::SRA:
   case ISD::SRL: {
     auto ShiftOp = dyn_cast<ConstantSDNode>(Op->getOperand(1));
     if (!ShiftOp)
       return std::nullopt;
 
     uint64_t BitShift = ShiftOp->getZExtValue();
 
     if (BitShift % 8 != 0)
       return std::nullopt;
 
     SrcIndex += BitShift / 8;
 
-    return calculateSrcByte(Op->getOperand(0), DestByte, SrcIndex, Depth + 1);
+    return calculateSrcByte(Op->getOperand(0), DestByte, SrcIndex, IsSigned,
+                            Depth + 1);
   }
 
   default: {
+    if (auto A = dyn_cast<AtomicSDNode>(Op) || Op->isMemIntrinsic()) {
+      // If this causes us to throw away signedness info, then fail.
+      if (IsSigned)
+        return std::nullopt;
       return ByteProvider<SDValue>::getSrc(Op, DestByte, SrcIndex);
     }
+
+    if (auto L = dyn_cast<LoadSDNode>(Op))
+      if (L->getExtensionType() != ISD::NON_EXTLOAD)
+        IsSigned =
+            IsSigned.value_or(false) || L->getExtensionType() == ISD::SEXTLOAD;
+
+    return ByteProvider<SDValue>::getSrc(Op, DestByte, SrcIndex, IsSigned);
+  }
   }
   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 std::optional<ByteProvider<SDValue>>
 calculateByteProvider(const SDValue &Op, unsigned Index, unsigned Depth,
-                      unsigned StartingIndex = 0) {
+                      unsigned StartingIndex = 0,
+                      std::optional<bool> IsSigned = std::nullopt) {
   // Finding Src tree of RHS of or typically requires at least 1 additional
   // depth
   if (Depth > 6)
     return std::nullopt;
 
   unsigned BitWidth = Op.getScalarValueSizeInBits();
   if (BitWidth % 8 != 0)
     return std::nullopt;
   if (Index > BitWidth / 8 - 1)
     return std::nullopt;
 
   switch (Op.getOpcode()) {
   case ISD::OR: {
     auto RHS = calculateByteProvider(Op.getOperand(1), Index, Depth + 1,
-                                     StartingIndex);
+                                     StartingIndex, IsSigned);
     if (!RHS)
       return std::nullopt;
     auto LHS = calculateByteProvider(Op.getOperand(0), Index, Depth + 1,
-                                     StartingIndex);
+                                     StartingIndex, IsSigned);
     if (!LHS)
       return std::nullopt;
     // A well formed Or will have two ByteProviders for each byte, one of which
     // is constant zero
     if (!LHS->isConstantZero() && !RHS->isConstantZero())
       return std::nullopt;
     if (!LHS || LHS->isConstantZero())
       return RHS;
case ISD::AND: {
     if ((IndexMask & BitMask) != IndexMask) {
       // If the result of the and partially provides the byte, then it
       // is not well formatted
       if (IndexMask & BitMask)
         return std::nullopt;
       return ByteProvider<SDValue>::getConstantZero();
     }
 
-    return calculateSrcByte(Op->getOperand(0), StartingIndex, Index);
+    return calculateSrcByte(Op->getOperand(0), StartingIndex, Index, IsSigned);
   }
 
   case ISD::FSHR: {
     // fshr(X,Y,Z): (X << (BW - (Z % BW))) | (Y >> (Z % BW))
     auto ShiftOp = dyn_cast<ConstantSDNode>(Op->getOperand(2));
     if (!ShiftOp || Op.getValueType().isVector())
       return std::nullopt;
 
case ISD::SRL: {
     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)
+                                  Index + ByteShift, IsSigned)
                : ByteProvider<SDValue>::getConstantZero();
   }
 
   case ISD::SHL: {
     auto ShiftOp = dyn_cast<ConstantSDNode>(Op->getOperand(1));
     if (!ShiftOp)
       return std::nullopt;
 
     uint64_t BitShift = ShiftOp->getZExtValue();
     if (BitShift % 8 != 0)
       return std::nullopt;
     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<SDValue>::getConstantZero()
                : calculateByteProvider(Op.getOperand(0), Index - ByteShift,
-                                       Depth + 1, StartingIndex);
+                                       Depth + 1, StartingIndex, IsSigned);
   }
   case ISD::ANY_EXTEND:
   case ISD::SIGN_EXTEND:
   case ISD::ZERO_EXTEND:
   case ISD::SIGN_EXTEND_INREG:
   case ISD::AssertZext:
   case ISD::AssertSext: {
     SDValue NarrowOp = Op->getOperand(0);
     unsigned NarrowBitWidth = NarrowOp.getValueSizeInBits();
     if (Op->getOpcode() == ISD::SIGN_EXTEND_INREG ||
         Op->getOpcode() == ISD::AssertZext ||
         Op->getOpcode() == ISD::AssertSext) {
       auto *VTSign = cast<VTSDNode>(Op->getOperand(1));
       NarrowBitWidth = VTSign->getVT().getSizeInBits();
     }
     if (NarrowBitWidth % 8 != 0)
       return std::nullopt;
     uint64_t NarrowByteWidth = NarrowBitWidth / 8;
 
+    IsSigned =
+        Op->getOpcode() != ISD::ANY_EXTEND
+            ? std::optional<bool>(IsSigned.value_or(false) ||
+                                  Op->getOpcode() == ISD::SIGN_EXTEND ||
+                                  Op->getOpcode() == ISD::SIGN_EXTEND_INREG ||
+                                  Op->getOpcode() == ISD::AssertSext)
+            : IsSigned;
+
     if (Index >= NarrowByteWidth)
       return Op.getOpcode() == ISD::ZERO_EXTEND
                  ? std::optional<ByteProvider<SDValue>>(
                        ByteProvider<SDValue>::getConstantZero())
                  : std::nullopt;
-    return calculateByteProvider(NarrowOp, Index, Depth + 1, StartingIndex);
+    return calculateByteProvider(NarrowOp, Index, Depth + 1, StartingIndex,
+                                 IsSigned);
   }
 
   case ISD::TRUNCATE: {
     uint64_t NarrowByteWidth = BitWidth / 8;
 
     if (NarrowByteWidth >= Index) {
       return calculateByteProvider(Op.getOperand(0), Index, Depth + 1,
-                                   StartingIndex);
+                                   StartingIndex, IsSigned);
     }
 
     return std::nullopt;
   }
 
   case ISD::CopyFromReg: {
     if (BitWidth / 8 > Index)
-      return calculateSrcByte(Op, StartingIndex, Index);
+      return calculateSrcByte(Op, StartingIndex, Index, IsSigned);
 
     return std::nullopt;
   }
 
   case ISD::LOAD: {
     auto L = cast<LoadSDNode>(Op.getNode());
+
+    // Only set IsSigned if the load is extended.
+    if (L->getExtensionType() != ISD::NON_EXTLOAD)
+      IsSigned =
+          IsSigned.value_or(false) || L->getExtensionType() == ISD::SEXTLOAD;
     unsigned NarrowBitWidth = L->getMemoryVT().getSizeInBits();
     if (NarrowBitWidth % 8 != 0)
       return std::nullopt;
     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
                  ? std::optional<ByteProvider<SDValue>>(
                        ByteProvider<SDValue>::getConstantZero())
                  : std::nullopt;
     }
 
     if (NarrowByteWidth > Index) {
-      return calculateSrcByte(Op, StartingIndex, Index);
+      return calculateSrcByte(Op, StartingIndex, Index, IsSigned);
     }
 
     return std::nullopt;
   }
 
   case ISD::BSWAP:
     return calculateByteProvider(Op->getOperand(0), BitWidth / 8 - Index - 1,
-                                 Depth + 1, StartingIndex);
+                                 Depth + 1, StartingIndex, IsSigned);
 
   case ISD::EXTRACT_VECTOR_ELT: {
     auto IdxOp = dyn_cast<ConstantSDNode>(Op->getOperand(1));
     if (!IdxOp)
       return std::nullopt;
     auto VecIdx = IdxOp->getZExtValue();
     auto ScalarSize = Op.getScalarValueSizeInBits();
     if (ScalarSize != 32) {
       if ((VecIdx + 1) * ScalarSize > 32)
         return std::nullopt;
       Index = ScalarSize == 8 ? VecIdx : VecIdx * 2 + Index;
     }
 
     return calculateSrcByte(ScalarSize == 32 ? Op : Op.getOperand(0),
-                            StartingIndex, Index);
+                            StartingIndex, Index, IsSigned);
   }
 
   case AMDGPUISD::PERM: {
     auto PermMask = dyn_cast<ConstantSDNode>(Op->getOperand(2));
     if (!PermMask)
       return std::nullopt;
 
     auto IdxMask =
         (PermMask->getZExtValue() & (0xFF << (Index * 8))) >> (Index * 8);
     if (IdxMask > 0x07 && IdxMask != 0x0c)
       return std::nullopt;
 
     auto NextOp = Op.getOperand(IdxMask > 0x03 ? 0 : 1);
     auto NextIndex = IdxMask > 0x03 ? IdxMask % 4 : IdxMask;
 
-    return IdxMask != 0x0c ? calculateSrcByte(NextOp, StartingIndex, NextIndex)
-                           : ByteProvider<SDValue>(
-                                 ByteProvider<SDValue>::getConstantZero());
+    return IdxMask != 0x0c
+               ? calculateSrcByte(NextOp, StartingIndex, NextIndex, IsSigned)
   }
 
   default: {
     return std::nullopt;
   }
   }
 
   llvm_unreachable("fully handled switch");
if (!MulSignedLo && (!MulLHSUnsigned32 || !MulRHSUnsigned32)) {
     Accum = DAG.getBitcast(MVT::i64, Accum);
   }
 
   if (VT != MVT::i64)
     Accum = DAG.getNode(ISD::TRUNCATE, SL, VT, Accum);
   return Accum;
 }
 
+// Collect the ultimate src of each of the mul node's operands, and confirm
+// each operand is 8 bytes.
+static std::optional<ByteProvider<SDValue>>
+handleMulOperand(const SDValue &MulOperand) {
+  auto Byte0 = calculateByteProvider(MulOperand, 0, 0);
+  if (!Byte0 || Byte0->isConstantZero()) {
+    return std::nullopt;
+  }
+  auto Byte1 = calculateByteProvider(MulOperand, 1, 0);
+  if (Byte1 && !Byte1->isConstantZero()) {
+    return std::nullopt;
+  }
+  return Byte0;
+}
+
+static unsigned addPermMasks(unsigned First, unsigned Second) {
+  unsigned FirstCs = First & 0x0c0c0c0c;
+  unsigned SecondCs = Second & 0x0c0c0c0c;
+  unsigned FirstNoCs = First & ~0x0c0c0c0c;
+  unsigned SecondNoCs = Second & ~0x0c0c0c0c;
+
+  assert(FirstCs & 0xFF | SecondCs & 0xFF);
+  assert(FirstCs & 0xFF00 | SecondCs & 0xFF00);
+  assert(FirstCs & 0xFF0000 | SecondCs & 0xFF0000);
+  assert(FirstCs & 0xFF000000 | SecondCs & 0xFF000000);
+
+  return (FirstNoCs | SecondNoCs) | (FirstCs & SecondCs);
+}
+
+static void placeSources(ByteProvider<SDValue> &Src0,
+                         ByteProvider<SDValue> &Src1,
+                         SmallVectorImpl<std::pair<SDValue, unsigned>> &Src0s,
+                         SmallVectorImpl<std::pair<SDValue, unsigned>> &Src1s,
+                         int Step) {
+
+  assert(Src0.Src.has_value() && Src1.Src.has_value());
+  // Src0s and Src1s are empty, just place arbitrarily.
+  if (Step == 0) {
+    Src0s.push_back({*Src0.Src, (Src0.SrcOffset << 24) + 0x0c0c0c});
+    Src1s.push_back({*Src1.Src, (Src1.SrcOffset << 24) + 0x0c0c0c});
+    return;
+  }
+
+  for (int BPI = 0; BPI < 2; BPI++) {
+    std::pair<ByteProvider<SDValue>, ByteProvider<SDValue>> BPP = {Src0, Src1};
+    if (BPI == 1) {
+      BPP = {Src1, Src0};
+    }
+    unsigned ZeroMask = 0x0c0c0c0c;
+    unsigned FMask = 0xFF << (8 * (3 - Step));
+
+    unsigned FirstMask =
+        BPP.first.SrcOffset << (8 * (3 - Step)) | (ZeroMask & ~FMask);
+    unsigned SecondMask =
+        BPP.second.SrcOffset << (8 * (3 - Step)) | (ZeroMask & ~FMask);
+    // Attempt to find Src vector which contains our SDValue, if so, add our
+    // perm mask to the existing one. If we are unable to find a match for the
+    // first SDValue, attempt to find match for the second.
+    int FirstGroup = -1;
+    for (int I = 0; I < 2; I++) {
+      SmallVectorImpl<std::pair<SDValue, unsigned>> &Srcs =
+          I == 0 ? Src0s : Src1s;
+      auto MatchesFirst = [&BPP](std::pair<SDValue, unsigned> IterElt) {
+        return IterElt.first == *BPP.first.Src;
+      };
+
+      auto Match = std::find_if(Srcs.begin(), Srcs.end(), MatchesFirst);
+      if (Match != Srcs.end()) {
+        Match->second = addPermMasks(FirstMask, Match->second);
+        FirstGroup = I;
+        break;
+      }
+    }
+    if (FirstGroup != -1) {
+      SmallVectorImpl<std::pair<SDValue, unsigned>> &Srcs =
+          FirstGroup == 1 ? Src0s : Src1s;
+      auto MatchesSecond = [&BPP](std::pair<SDValue, unsigned> IterElt) {
+        return IterElt.first == *BPP.second.Src;
+      };
+      auto Match = std::find_if(Srcs.begin(), Srcs.end(), MatchesSecond);
+      if (Match != Srcs.end()) {
+        Match->second = addPermMasks(SecondMask, Match->second);
+      } else
+        Srcs.push_back({*BPP.second.Src, SecondMask});
+      return;
+    }
+  }
+
+  // If we have made it here, then we could not find a match in Src0s or Src1s
+  // for either Src0 or Src1, so just place them arbitrarily.
+
+  unsigned ZeroMask = 0x0c0c0c0c;
+  unsigned FMask = 0xFF << (8 * (3 - Step));
+
+  Src0s.push_back(
+      {*Src0.Src, (Src0.SrcOffset << (8 * (3 - Step)) | (ZeroMask & ~FMask))});
+  Src1s.push_back(
+      {*Src1.Src, (Src1.SrcOffset << (8 * (3 - Step)) | (ZeroMask & ~FMask))});
+
+  return;
+}
+
+static SDValue
+resolveSources(SelectionDAG &DAG, SDLoc SL,
+               SmallVectorImpl<std::pair<SDValue, unsigned>> &Srcs,
+               bool IsSigned, bool IsAny) {
+
+  // If we just have one source, just permute it accordingly.
+  if (Srcs.size() == 1) {
+    auto Elt = Srcs.begin();
+    auto EltVal = DAG.getBitcastedAnyExtOrTrunc(Elt->first, SL, MVT::i32);
+
+    // v_perm will produce the original value.
+    if (Elt->second == 0x3020100)
+      return EltVal;
+
+    return DAG.getNode(AMDGPUISD::PERM, SL, MVT::i32, EltVal, EltVal,
+                       DAG.getConstant(Elt->second, SL, MVT::i32));
+  }
+
+  auto FirstElt = Srcs.begin();
+  auto SecondElt = std::next(FirstElt);
+
+  SmallVector<SDValue, 2> Perms;
+
+  // If we have multiple sources in the chain, combine them via perms (using
+  // calculated perm mask) and Ors.
+  while (true) {
+    auto FirstMask = FirstElt->second;
+    auto SecondMask = SecondElt->second;
+
+    unsigned FirstCs = FirstMask & 0x0c0c0c0c;
+    unsigned FirstPlusFour = FirstMask | 0x04040404;
+    // 0x0c + 0x04 = 0x10, so anding with 0x0F will produced 0x00 for any
+    // original 0x0C.
+    FirstMask = (FirstPlusFour & 0x0F0F0F0F) | FirstCs;
+
+    auto PermMask = addPermMasks(FirstMask, SecondMask);
+    auto FirstVal =
+        DAG.getBitcastedAnyExtOrTrunc(FirstElt->first, SL, MVT::i32);
+    auto SecondVal =
+        DAG.getBitcastedAnyExtOrTrunc(SecondElt->first, SL, MVT::i32);
+
+    Perms.push_back(DAG.getNode(AMDGPUISD::PERM, SL, MVT::i32, FirstVal,
+                                SecondVal,
+                                DAG.getConstant(PermMask, SL, MVT::i32)));
+
+    FirstElt = std::next(SecondElt);
+    if (FirstElt == Srcs.end())
+      break;
+
+    SecondElt = std::next(FirstElt);
+    // If we only have a FirstElt, then just combine that into the cumulative
+    // source node.
+    if (SecondElt == Srcs.end()) {
+      auto EltVal =
+          DAG.getBitcastedAnyExtOrTrunc(FirstElt->first, SL, MVT::i32);
+
+      Perms.push_back(
+          DAG.getNode(AMDGPUISD::PERM, SL, MVT::i32, EltVal, EltVal,
+                      DAG.getConstant(FirstElt->second, SL, MVT::i32)));
+      break;
+    }
+  }
+
+  assert(Perms.size() == 1 || Perms.size() == 2);
+  return Perms.size() == 2
+             ? DAG.getNode(ISD::OR, SL, MVT::i32, Perms[0], Perms[1])
+             : Perms[0];
+}
+
+static void fixMasks(SmallVectorImpl<std::pair<SDValue, unsigned>> &Srcs,
+                     unsigned ChainLength) {
+  for (auto &[EntryVal, EntryMask] : Srcs) {
+    EntryMask = EntryMask >> ((4 - ChainLength) * 8);
+    auto ZeroMask = ChainLength == 2 ? 0x0c0c0000 : 0x0c000000;
+    EntryMask += ZeroMask;
+  }
+}
+
+static bool isMul(const SDValue Op) {
+  auto Opcode = Op.getOpcode();
+
+  return (Opcode == ISD::MUL || Opcode == AMDGPUISD::MUL_U24 ||
+          Opcode == AMDGPUISD::MUL_I24);
+}
+
+static std::optional<bool> checkSignedness(const SDValue &N,
+                                           ByteProvider<SDValue> &Src0,
+                                           ByteProvider<SDValue> &Src1) {
+  auto MulOpcode = N.getOpcode();
+  std::optional<bool> IterIsSigned;
+  // Both sides of the tree must have the same signedness semantics.
+  if ((Src0.IsSigned != Src1.IsSigned) ||
+      (Src0.IsSigned.value_or(false) != Src1.IsSigned.value_or(false)))
+    return IterIsSigned;
+  // If we have a MUL_U24 op with signed semantics, then fail.
+  if (Src0.IsSigned.value_or(false) && MulOpcode == AMDGPUISD::MUL_U24)
+    return IterIsSigned;
+  // If we have a MUL_I24 op with unsigned semantics, then fail.
+  if (!Src0.IsSigned.value_or(true) && MulOpcode == AMDGPUISD::MUL_I24)
+    return IterIsSigned;
+
+  bool TopLevelSignedness =
+      MulOpcode == AMDGPUISD::MUL_I24 ||
+      (MulOpcode == ISD::MUL && N.getNode()->getFlags().hasNoSignedWrap() &&
+       !N.getNode()->getFlags().hasNoUnsignedWrap());
+
+  // In cases where we are accumulating into an i8 (for v_dot4), the
+  // ByteProvider will not have signedness info since the MSBs are dont-cares.
+  // In this case, we simply use the TopLevelSignedness of the instruction.
+  IterIsSigned = Src0.IsSigned.value_or(TopLevelSignedness);
+  return IterIsSigned;
+}
+
 SDValue SITargetLowering::performAddCombine(SDNode *N,
                                             DAGCombinerInfo &DCI) const {
   SelectionDAG &DAG = DCI.DAG;
   EVT VT = N->getValueType(0);
   SDLoc SL(N);
   SDValue LHS = N->getOperand(0);
   SDValue RHS = N->getOperand(1);
 
   if (LHS.getOpcode() == ISD::MUL || RHS.getOpcode() == ISD::MUL) {
     if (Subtarget->hasMad64_32()) {
       if (SDValue Folded = tryFoldToMad64_32(N, DCI))
         return Folded;
     }
-
-    return SDValue();
   }
 
   if (SDValue V = reassociateScalarOps(N, DAG)) {
     return V;
   }
 
+  if ((isMul(LHS) || isMul(RHS)) && Subtarget->hasDot7Insts() &&
+      (Subtarget->hasDot1Insts() || Subtarget->hasDot8Insts())) {
+    SDValue TempNode(N, 0);
+    std::optional<bool> IsSigned;
+    SmallVector<std::pair<SDValue, unsigned>, 4> Src0s;
+    SmallVector<std::pair<SDValue, unsigned>, 4> Src1s;
+    SmallVector<SDValue, 4> Src2s;
+
+    // Match the v_dot4 tree, while collecting src nodes.
+    int ChainLength = 0;
+    for (int I = 0; I < 4; I++) {
+      auto MulIdx = isMul(LHS) ? 0 : isMul(RHS) ? 1 : -1;
+      if (MulIdx == -1)
+        break;
+      auto Src0 = handleMulOperand(TempNode->getOperand(MulIdx)->getOperand(0));
+      if (!Src0)
+        break;
+      auto Src1 = handleMulOperand(TempNode->getOperand(MulIdx)->getOperand(1));
+      if (!Src1)
+        break;
+
+      auto IterIsSigned =
+          checkSignedness(TempNode->getOperand(MulIdx), *Src0, *Src1);
+      if (!IterIsSigned)
+        break;
+      if (!IsSigned)
+        IsSigned = *IterIsSigned;
+      if (*IterIsSigned != *IsSigned)
+        break;
+      placeSources(*Src0, *Src1, Src0s, Src1s, I);
+      auto AddIdx = 1 - MulIdx;
+      // Allow the special case where add (add (mul24, 0), mul24) became ->
+      // add (mul24, mul24).
+      if (I == 2 && isMul(TempNode->getOperand(AddIdx))) {
+        Src2s.push_back(TempNode->getOperand(AddIdx));
+        auto Src0 =
+            handleMulOperand(TempNode->getOperand(AddIdx)->getOperand(0));
+        if (!Src0)
+          break;
+        auto Src1 =
+            handleMulOperand(TempNode->getOperand(AddIdx)->getOperand(1));
+        if (!Src1)
+          break;
+        auto IterIsSigned =
+            checkSignedness(TempNode->getOperand(AddIdx), *Src0, *Src1);
+        if (!IterIsSigned)
+          break;
+        assert(IsSigned);
+        if (*IterIsSigned != *IsSigned)
+          break;
+        placeSources(*Src0, *Src1, Src0s, Src1s, I + 1);
+        Src2s.push_back(DAG.getConstant(0, SL, MVT::i32));
+        ChainLength = I + 2;
+        break;
+      }
+
+      TempNode = TempNode->getOperand(AddIdx);
+      Src2s.push_back(TempNode);
+      ChainLength = I + 1;
+      if (TempNode->getNumOperands() < 2)
+        break;
+      LHS = TempNode->getOperand(0);
+      RHS = TempNode->getOperand(1);
+    }
+
+    if (ChainLength < 2)
+      return SDValue();
+
+    // Masks were constructed with assumption that we would find a chain of
+    // length 4. If not, then we need to 0 out the MSB bits (via perm mask of
+    // 0x0c) so they do not affect dot calculation.
+    if (ChainLength < 4) {
+      fixMasks(Src0s, ChainLength);
+      fixMasks(Src1s, ChainLength);
+    }
+
+    SDValue Src0, Src1;
+
+    // If we are just using a single source for both, and have permuted the
+    // bytes consistently, we can just use the sources without permuting
+    // (commutation).
+    bool UseOriginalSrc = false;
+    if (ChainLength == 4 && Src0s.size() == 1 && Src1s.size() == 1 &&
+        Src0s.begin()->second == Src1s.begin()->second &&
+        Src0s.begin()->first.getValueSizeInBits() == 32 &&
+        Src1s.begin()->first.getValueSizeInBits() == 32) {
+      SmallVector<unsigned, 4> SrcBytes;
+      auto Src0Mask = Src0s.begin()->second;
+      SrcBytes.push_back(Src0Mask & 0xFF000000);
+      bool UniqueEntries = true;
+      for (auto I = 1; I < 4; I++) {
+        auto NextByte = Src0Mask & (0xFF << ((3 - I) * 8));
+
+        if (is_contained(SrcBytes, NextByte)) {
+          UniqueEntries = false;
+          break;
+        }
+        SrcBytes.push_back(NextByte);
+      }
+
+      if (UniqueEntries) {
+        UseOriginalSrc = true;
+        // Must be 32 bits to enter above conditional.
+        assert(Src0s.begin()->first.getValueSizeInBits() == 32);
+        assert(Src1s.begin()->first.getValueSizeInBits() == 32);
+        Src0 = DAG.getBitcast(MVT::getIntegerVT(32), Src0s.begin()->first);
+        Src1 = DAG.getBitcast(MVT::getIntegerVT(32), Src1s.begin()->first);
+      }
+    }
+
+    if (!UseOriginalSrc) {
+      Src0 = resolveSources(DAG, SL, Src0s, false, true);
+      Src1 = resolveSources(DAG, SL, Src1s, false, true);
+    }
+
+    assert(IsSigned);
+    SDValue Src2 =
+        DAG.getExtOrTrunc(*IsSigned, Src2s[ChainLength - 1], SL, MVT::i32);
+
+    SDValue IID = DAG.getTargetConstant(*IsSigned ? Intrinsic::amdgcn_sdot4
+                                                  : Intrinsic::amdgcn_udot4,
+                                        SL, MVT::i64);
+
+    assert(!VT.isVector());
+    auto Dot = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, SL, MVT::i32, IID, Src0,
+                           Src1, Src2, DAG.getTargetConstant(0, SL, MVT::i1));
+
+    return DAG.getExtOrTrunc(*IsSigned, Dot, SL, VT);
+  }
+
   if (VT != MVT::i32 || !DCI.isAfterLegalizeDAG())
     return SDValue();
 
   // add x, zext (setcc) => uaddo_carry x, 0, setcc
   // add x, sext (setcc) => usubo_carry x, 0, setcc
   unsigned Opc = LHS.getOpcode();
   if (Opc == ISD::ZERO_EXTEND || Opc == ISD::SIGN_EXTEND ||
       Opc == ISD::ANY_EXTEND || Opc == ISD::UADDO_CARRY)
     std::swap(RHS, LHS);
 
   Opc = RHS.getOpcode();
   switch (Opc) {
   default: break;

llvm/test/CodeGen/AMDGPU/idot2.ll

 ; GFX9-NODL-NEXT:    global_store_dword v0, v1, s[2:3]
 ; GFX9-NODL-NEXT:    s_endpgm
 ;
 ; GFX9-DL-LABEL: notsdot2_sext8:
 ; GFX9-DL:       ; %bb.0: ; %entry
 ; GFX9-DL-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x24
 ; GFX9-DL-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x34
 ; GFX9-DL-NEXT:    v_lshlrev_b32_e32 v0, 1, v0
+; GFX9-DL-NEXT:    s_mov_b32 s1, 0xc0c0001
 ; GFX9-DL-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX9-DL-NEXT:    global_load_ushort v1, v0, s[4:5]
 ; GFX9-DL-NEXT:    global_load_ushort v2, v0, s[6:7]
 ; GFX9-DL-NEXT:    s_load_dword s0, s[2:3], 0x0
 ; GFX9-DL-NEXT:    v_mov_b32_e32 v0, 0
+; GFX9-DL-NEXT:    s_waitcnt vmcnt(1)
+; GFX9-DL-NEXT:    v_perm_b32 v1, v1, v1, s1
 ; GFX9-DL-NEXT:    s_waitcnt vmcnt(0)
-; GFX9-DL-NEXT:    v_mul_i32_i24_sdwa v3, sext(v2), sext(v1) dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:BYTE_0
-; GFX9-DL-NEXT:    v_lshrrev_b16_e32 v1, 8, v1
-; GFX9-DL-NEXT:    v_lshrrev_b16_e32 v2, 8, v2
-; GFX9-DL-NEXT:    v_mul_i32_i24_sdwa v1, sext(v2), sext(v1) dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:BYTE_0
+; GFX9-DL-NEXT:    v_perm_b32 v2, v2, v2, s1
 ; GFX9-DL-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX9-DL-NEXT:    v_add3_u32 v1, v1, s0, v3
+; GFX9-DL-NEXT:    v_dot4_i32_i8 v1, v2, v1, s0
 ; GFX9-DL-NEXT:    global_store_dword v0, v1, s[2:3]
 ; GFX9-DL-NEXT:    s_endpgm
 ;
 ; GFX10-DL-LABEL: notsdot2_sext8:
 ; GFX10-DL:       ; %bb.0: ; %entry
 ; GFX10-DL-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x24
 ; GFX10-DL-NEXT:    v_lshlrev_b32_e32 v0, 1, v0
 ; GFX10-DL-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0x34
+; GFX10-DL-NEXT:    v_mov_b32_e32 v3, 0
 ; GFX10-DL-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX10-DL-NEXT:    s_clause 0x1
 ; GFX10-DL-NEXT:    global_load_ushort v1, v0, s[4:5]
 ; GFX10-DL-NEXT:    global_load_ushort v2, v0, s[6:7]
 ; GFX10-DL-NEXT:    s_load_dword s2, s[0:1], 0x0
 ; GFX10-DL-NEXT:    s_waitcnt vmcnt(1)
-; GFX10-DL-NEXT:    v_lshrrev_b16 v0, 8, v1
+; GFX10-DL-NEXT:    v_perm_b32 v0, v1, v1, 0xc0c0001
 ; GFX10-DL-NEXT:    s_waitcnt vmcnt(0)
-; GFX10-DL-NEXT:    v_lshrrev_b16 v3, 8, v2
-; GFX10-DL-NEXT:    v_mul_i32_i24_sdwa v1, sext(v2), sext(v1) dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:BYTE_0
-; GFX10-DL-NEXT:    v_mov_b32_e32 v2, 0
-; GFX10-DL-NEXT:    v_mul_i32_i24_sdwa v0, sext(v3), sext(v0) dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:BYTE_0
+; GFX10-DL-NEXT:    v_perm_b32 v1, v2, v2, 0xc0c0001
 ; GFX10-DL-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX10-DL-NEXT:    v_add3_u32 v0, v0, s2, v1
-; GFX10-DL-NEXT:    global_store_dword v2, v0, s[0:1]
+; GFX10-DL-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-DL-NEXT:    v_dot4c_i32_i8_e32 v2, v1, v0
+; GFX10-DL-NEXT:    global_store_dword v3, v2, s[0:1]
 ; GFX10-DL-NEXT:    s_endpgm
                                           ptr addrspace(1) %src2,
                                           ptr addrspace(1) nocapture %dst) {
 entry:
   %idx = call i32 @llvm.amdgcn.workitem.id.x()
   %gep1 = getelementptr <2 x i8>, ptr addrspace(1) %src1, i32 %idx
   %vec1 = load <2 x i8>, ptr addrspace(1) %gep1
   %gep2 = getelementptr <2 x i8>, ptr addrspace(1) %src2, i32 %idx

llvm/test/CodeGen/AMDGPU/idot4s.ll

 ; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py
 ; RUN: llc -mtriple=amdgcn -mcpu=gfx700 -verify-machineinstrs < %s | FileCheck -check-prefixes=GFX7 %s
 ; RUN: llc -mtriple=amdgcn -mcpu=gfx803 -verify-machineinstrs < %s | FileCheck -check-prefixes=GFX8 %s
 ; RUN: llc -mtriple=amdgcn -mcpu=gfx900 -verify-machineinstrs < %s | FileCheck -check-prefixes=GFX9-NODL %s
 ; RUN: llc -mtriple=amdgcn -mcpu=gfx906 -verify-machineinstrs < %s | FileCheck -check-prefixes=GFX9-DL %s
 ; RUN: llc -mtriple=amdgcn -mcpu=gfx1011 -verify-machineinstrs < %s | FileCheck -check-prefixes=GFX10-DL %s
 ; RUN: llc -mtriple=amdgcn -mcpu=gfx1012 -verify-machineinstrs < %s | FileCheck -check-prefixes=GFX10-DL %s
+; RUN: llc -mtriple=amdgcn -mcpu=gfx1100 -verify-machineinstrs < %s | FileCheck -check-prefixes=GFX11-DL %s
 
 define amdgpu_kernel void @idot4_acc32(ptr addrspace(1) %src1,
 ; GFX7-LABEL: idot4_acc32:
 ; GFX7:       ; %bb.0: ; %entry
 ; GFX7-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x9
 ; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xd
 ; GFX7-NEXT:    s_mov_b32 s3, 0xf000
 ; GFX7-NEXT:    s_mov_b32 s10, 0

 ; GFX9-DL-NEXT:    global_store_dword v0, v1, s[2:3]
 ; GFX9-DL-NEXT:    s_endpgm
 ;
 ; GFX10-DL-LABEL: idot4_acc32:
 ; GFX10-DL:       ; %bb.0: ; %entry
 ; GFX10-DL-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x24
 ; GFX10-DL-NEXT:    v_lshlrev_b32_e32 v0, 2, v0
 ; GFX10-DL-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0x34
+; GFX10-DL-NEXT:    v_mov_b32_e32 v3, 0
 ; GFX10-DL-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX10-DL-NEXT:    s_clause 0x1
 ; GFX10-DL-NEXT:    global_load_dword v1, v0, s[4:5]
 ; GFX10-DL-NEXT:    global_load_dword v2, v0, s[6:7]
-; GFX10-DL-NEXT:    v_mov_b32_e32 v0, 0
 ; GFX10-DL-NEXT:    s_load_dword s2, s[0:1], 0x0
-; GFX10-DL-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
-; GFX10-DL-NEXT:    v_dot4_i32_i8 v1, v1, v2, s2
-; GFX10-DL-NEXT:    global_store_dword v0, v1, s[0:1]
+; GFX10-DL-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-DL-NEXT:    v_mov_b32_e32 v0, s2
+; GFX10-DL-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-DL-NEXT:    v_dot4c_i32_i8_e32 v0, v1, v2
+; GFX10-DL-NEXT:    global_store_dword v3, v0, s[0:1]
 ; GFX10-DL-NEXT:    s_endpgm
+;
+; GFX11-DL-LABEL: idot4_acc32:
+; GFX11-DL:       ; %bb.0: ; %entry
+; GFX11-DL-NEXT:    s_load_b128 s[4:7], s[0:1], 0x24
+; GFX11-DL-NEXT:    v_lshlrev_b32_e32 v0, 2, v0
+; GFX11-DL-NEXT:    s_load_b64 s[0:1], s[0:1], 0x34
+; GFX11-DL-NEXT:    v_mov_b32_e32 v2, 0
+; GFX11-DL-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX11-DL-NEXT:    s_clause 0x1
+; GFX11-DL-NEXT:    global_load_b32 v1, v0, s[4:5]
+; GFX11-DL-NEXT:    global_load_b32 v0, v0, s[6:7]
+; GFX11-DL-NEXT:    s_load_b32 s2, s[0:1], 0x0
+; GFX11-DL-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX11-DL-NEXT:    v_dot4_i32_iu8 v0, v1, v0, s2 neg_lo:[1,1,0]
+; GFX11-DL-NEXT:    global_store_b32 v2, v0, s[0:1]
+; GFX11-DL-NEXT:    s_nop 0
+; GFX11-DL-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
+; GFX11-DL-NEXT:    s_endpgm
                                        ptr addrspace(1) %src2,
                                        ptr addrspace(1) nocapture %dst) {
 entry:
   %idx = call i32 @llvm.amdgcn.workitem.id.x()
   %gep1 = getelementptr <4 x i8>, ptr addrspace(1) %src1, i32 %idx
   %vec1 = load <4 x i8>, ptr addrspace(1) %gep1
   %gep2 = getelementptr <4 x i8>, ptr addrspace(1) %src2, i32 %idx
   %vec2 = load <4 x i8>, ptr addrspace(1) %gep2
entry:
   %add1 = add i32 %mul1, %acc
   %add2 = add i32 %add1, %mul2
   %add3 = add i32 %add2, %mul3
   %add4 = add i32 %add3, %mul4
   store i32 %add4, ptr addrspace(1) %dst, align 4
   ret void
 }
 
-; TODO: Currently, vector elements{0 and 3} get zero_extended from i16 to i32 which should
-; be sign_extended directly to i32; prevents the pattern recognizer to recognize this pattern.
 define amdgpu_kernel void @idot4_acc16(ptr addrspace(1) %src1,
 ; GFX7-LABEL: idot4_acc16:
 ; GFX7:       ; %bb.0: ; %entry
 ; GFX7-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x9
 ; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xd
 ; GFX7-NEXT:    s_mov_b32 s3, 0xf000
 ; GFX7-NEXT:    s_mov_b32 s10, 0
 ; GFX7-NEXT:    s_mov_b32 s11, s3

 ; GFX9-NODL-NEXT:    global_store_short v0, v1, s[2:3]
 ; GFX9-NODL-NEXT:    s_endpgm
 ;
 ; GFX9-DL-LABEL: idot4_acc16:
 ; GFX9-DL:       ; %bb.0: ; %entry
 ; GFX9-DL-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x24
 ; GFX9-DL-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x34
 ; GFX9-DL-NEXT:    v_lshlrev_b32_e32 v0, 2, v0
+; GFX9-DL-NEXT:    v_mov_b32_e32 v1, 0
 ; GFX9-DL-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX9-DL-NEXT:    global_load_dword v1, v0, s[4:5]
-; GFX9-DL-NEXT:    global_load_dword v2, v0, s[6:7]
-; GFX9-DL-NEXT:    v_mov_b32_e32 v0, 0
-; GFX9-DL-NEXT:    global_load_ushort v3, v0, s[2:3]
-; GFX9-DL-NEXT:    s_waitcnt vmcnt(2)
-; GFX9-DL-NEXT:    v_bfe_i32 v6, v1, 0, 8
-; GFX9-DL-NEXT:    s_waitcnt vmcnt(1)
-; GFX9-DL-NEXT:    v_bfe_i32 v7, v2, 0, 8
-; GFX9-DL-NEXT:    v_lshrrev_b32_e32 v8, 8, v1
-; GFX9-DL-NEXT:    v_lshrrev_b32_e32 v9, 8, v2
-; GFX9-DL-NEXT:    v_lshrrev_b32_e32 v4, 16, v1
-; GFX9-DL-NEXT:    v_lshrrev_b32_e32 v5, 16, v2
-; GFX9-DL-NEXT:    v_bfe_i32 v8, v8, 0, 8
-; GFX9-DL-NEXT:    v_bfe_i32 v9, v9, 0, 8
+; GFX9-DL-NEXT:    global_load_dword v2, v0, s[4:5]
+; GFX9-DL-NEXT:    global_load_dword v3, v0, s[6:7]
+; GFX9-DL-NEXT:    global_load_sshort v4, v1, s[2:3]
 ; GFX9-DL-NEXT:    s_waitcnt vmcnt(0)
-; GFX9-DL-NEXT:    v_mad_legacy_u16 v3, v6, v7, v3
-; GFX9-DL-NEXT:    v_lshrrev_b32_e32 v1, 24, v1
-; GFX9-DL-NEXT:    v_lshrrev_b32_e32 v2, 24, v2
-; GFX9-DL-NEXT:    v_bfe_i32 v4, v4, 0, 8
-; GFX9-DL-NEXT:    v_bfe_i32 v5, v5, 0, 8
-; GFX9-DL-NEXT:    v_mad_legacy_u16 v3, v8, v9, v3
-; GFX9-DL-NEXT:    v_bfe_i32 v1, v1, 0, 8
-; GFX9-DL-NEXT:    v_bfe_i32 v2, v2, 0, 8
-; GFX9-DL-NEXT:    v_mad_legacy_u16 v3, v4, v5, v3
-; GFX9-DL-NEXT:    v_mad_legacy_u16 v1, v1, v2, v3
-; GFX9-DL-NEXT:    global_store_short v0, v1, s[2:3]
+; GFX9-DL-NEXT:    v_dot4_i32_i8 v0, v2, v3, v4
+; GFX9-DL-NEXT:    global_store_short v1, v0, s[2:3]
 ; GFX9-DL-NEXT:    s_endpgm
 ;
 ; GFX10-DL-LABEL: idot4_acc16:
 ; GFX10-DL:       ; %bb.0: ; %entry
 ; GFX10-DL-NEXT:    s_clause 0x1
 ; GFX10-DL-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x24
 ; GFX10-DL-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x34
 ; GFX10-DL-NEXT:    v_lshlrev_b32_e32 v0, 2, v0
+; GFX10-DL-NEXT:    v_mov_b32_e32 v1, 0
 ; GFX10-DL-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX10-DL-NEXT:    s_clause 0x1
-; GFX10-DL-NEXT:    global_load_dword v1, v0, s[4:5]
-; GFX10-DL-NEXT:    global_load_dword v2, v0, s[6:7]
-; GFX10-DL-NEXT:    v_mov_b32_e32 v0, 0
-; GFX10-DL-NEXT:    global_load_ushort v3, v0, s[2:3]
-; GFX10-DL-NEXT:    s_waitcnt vmcnt(2)
-; GFX10-DL-NEXT:    v_bfe_i32 v4, v1, 0, 8
-; GFX10-DL-NEXT:    v_lshrrev_b32_e32 v5, 8, v1
-; GFX10-DL-NEXT:    s_waitcnt vmcnt(1)
-; GFX10-DL-NEXT:    v_lshrrev_b32_e32 v6, 8, v2
-; GFX10-DL-NEXT:    v_bfe_i32 v7, v2, 0, 8
-; GFX10-DL-NEXT:    v_lshrrev_b32_e32 v8, 16, v1
-; GFX10-DL-NEXT:    v_lshrrev_b32_e32 v9, 16, v2
-; GFX10-DL-NEXT:    v_bfe_i32 v5, v5, 0, 8
-; GFX10-DL-NEXT:    v_bfe_i32 v6, v6, 0, 8
+; GFX10-DL-NEXT:    global_load_dword v2, v0, s[4:5]
+; GFX10-DL-NEXT:    global_load_dword v3, v0, s[6:7]
+; GFX10-DL-NEXT:    global_load_sshort v4, v1, s[2:3]
 ; GFX10-DL-NEXT:    s_waitcnt vmcnt(0)
-; GFX10-DL-NEXT:    v_mad_u16 v3, v4, v7, v3
-; GFX10-DL-NEXT:    v_bfe_i32 v4, v8, 0, 8
-; GFX10-DL-NEXT:    v_bfe_i32 v7, v9, 0, 8
-; GFX10-DL-NEXT:    v_lshrrev_b32_e32 v1, 24, v1
-; GFX10-DL-NEXT:    v_lshrrev_b32_e32 v2, 24, v2
-; GFX10-DL-NEXT:    v_mad_u16 v3, v5, v6, v3
-; GFX10-DL-NEXT:    v_bfe_i32 v1, v1, 0, 8
-; GFX10-DL-NEXT:    v_bfe_i32 v2, v2, 0, 8
-; GFX10-DL-NEXT:    v_mad_u16 v3, v4, v7, v3
-; GFX10-DL-NEXT:    v_mad_u16 v1, v1, v2, v3
-; GFX10-DL-NEXT:    global_store_short v0, v1, s[2:3]
+; GFX10-DL-NEXT:    v_dot4c_i32_i8_e32 v4, v2, v3
+; GFX10-DL-NEXT:    global_store_short v1, v4, s[2:3]
 ; GFX10-DL-NEXT:    s_endpgm
+;
+; GFX11-DL-LABEL: idot4_acc16:
+; GFX11-DL:       ; %bb.0: ; %entry
+; GFX11-DL-NEXT:    s_clause 0x1
+; GFX11-DL-NEXT:    s_load_b128 s[4:7], s[0:1], 0x24
+; GFX11-DL-NEXT:    s_load_b64 s[0:1], s[0:1], 0x34
+; GFX11-DL-NEXT:    v_dual_mov_b32 v1, 0 :: v_dual_lshlrev_b32 v0, 2, v0
+; GFX11-DL-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX11-DL-NEXT:    s_clause 0x1
+; GFX11-DL-NEXT:    global_load_b32 v2, v0, s[4:5]
+; GFX11-DL-NEXT:    global_load_b32 v0, v0, s[6:7]
+; GFX11-DL-NEXT:    global_load_i16 v3, v1, s[0:1]
+; GFX11-DL-NEXT:    s_waitcnt vmcnt(0)
+; GFX11-DL-NEXT:    v_dot4_i32_iu8 v0, v2, v0, v3 neg_lo:[1,1,0]
+; GFX11-DL-NEXT:    global_store_b16 v1, v0, s[0:1]
+; GFX11-DL-NEXT:    s_nop 0
+; GFX11-DL-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
+; GFX11-DL-NEXT:    s_endpgm
                                        ptr addrspace(1) %src2,
                                        ptr addrspace(1) nocapture %dst) {
 entry:
   %idx = call i32 @llvm.amdgcn.workitem.id.x()
   %gep1 = getelementptr <4 x i8>, ptr addrspace(1) %src1, i32 %idx
   %vec1 = load <4 x i8>, ptr addrspace(1) %gep1
   %gep2 = getelementptr <4 x i8>, ptr addrspace(1) %src2, i32 %idx
   %vec2 = load <4 x i8>, ptr addrspace(1) %gep2

 ; GFX9-NODL-NEXT:    global_store_byte v0, v1, s[2:3]
 ; GFX9-NODL-NEXT:    s_endpgm
 ;
 ; GFX9-DL-LABEL: idot4_acc8:
 ; GFX9-DL:       ; %bb.0: ; %entry
 ; GFX9-DL-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x24
 ; GFX9-DL-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x34
 ; GFX9-DL-NEXT:    v_lshlrev_b32_e32 v0, 2, v0
+; GFX9-DL-NEXT:    v_mov_b32_e32 v1, 0
 ; GFX9-DL-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX9-DL-NEXT:    global_load_dword v1, v0, s[4:5]
-; GFX9-DL-NEXT:    global_load_dword v2, v0, s[6:7]
-; GFX9-DL-NEXT:    v_mov_b32_e32 v0, 0
-; GFX9-DL-NEXT:    global_load_ubyte v3, v0, s[2:3]
-; GFX9-DL-NEXT:    s_waitcnt vmcnt(2)
-; GFX9-DL-NEXT:    v_lshrrev_b32_e32 v4, 16, v1
-; GFX9-DL-NEXT:    v_lshrrev_b32_e32 v6, 8, v1
-; GFX9-DL-NEXT:    s_waitcnt vmcnt(1)
-; GFX9-DL-NEXT:    v_lshrrev_b32_e32 v7, 8, v2
-; GFX9-DL-NEXT:    v_lshrrev_b32_e32 v8, 24, v1
+; GFX9-DL-NEXT:    global_load_dword v2, v0, s[4:5]
+; GFX9-DL-NEXT:    global_load_dword v3, v0, s[6:7]
+; GFX9-DL-NEXT:    global_load_ubyte v4, v1, s[2:3]
 ; GFX9-DL-NEXT:    s_waitcnt vmcnt(0)
-; GFX9-DL-NEXT:    v_mad_legacy_u16 v1, v1, v2, v3
-; GFX9-DL-NEXT:    v_lshrrev_b32_e32 v5, 16, v2
-; GFX9-DL-NEXT:    v_mad_legacy_u16 v1, v6, v7, v1
-; GFX9-DL-NEXT:    v_lshrrev_b32_e32 v9, 24, v2
-; GFX9-DL-NEXT:    v_mad_legacy_u16 v1, v4, v5, v1
-; GFX9-DL-NEXT:    v_mad_legacy_u16 v1, v8, v9, v1
-; GFX9-DL-NEXT:    global_store_byte v0, v1, s[2:3]
+; GFX9-DL-NEXT:    v_dot4_u32_u8 v0, v2, v3, v4
+; GFX9-DL-NEXT:    global_store_byte v1, v0, s[2:3]
 ; GFX9-DL-NEXT:    s_endpgm
 ;
 ; GFX10-DL-LABEL: idot4_acc8:
 ; GFX10-DL:       ; %bb.0: ; %entry
 ; GFX10-DL-NEXT:    s_clause 0x1
 ; GFX10-DL-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x24
 ; GFX10-DL-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x34
 ; GFX10-DL-NEXT:    v_lshlrev_b32_e32 v0, 2, v0
 ; GFX10-DL-NEXT:    v_mov_b32_e32 v1, 0
 ; GFX10-DL-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX10-DL-NEXT:    s_clause 0x1
 ; GFX10-DL-NEXT:    global_load_dword v2, v0, s[4:5]
 ; GFX10-DL-NEXT:    global_load_dword v3, v0, s[6:7]
 ; GFX10-DL-NEXT:    global_load_ubyte v4, v1, s[2:3]
-; GFX10-DL-NEXT:    s_waitcnt vmcnt(2)
-; GFX10-DL-NEXT:    v_lshrrev_b32_e32 v0, 8, v2
-; GFX10-DL-NEXT:    s_waitcnt vmcnt(1)
-; GFX10-DL-NEXT:    v_lshrrev_b32_e32 v5, 8, v3
 ; GFX10-DL-NEXT:    s_waitcnt vmcnt(0)
-; GFX10-DL-NEXT:    v_mad_u16 v4, v2, v3, v4
-; GFX10-DL-NEXT:    v_lshrrev_b32_e32 v6, 16, v2
-; GFX10-DL-NEXT:    v_lshrrev_b32_e32 v7, 16, v3
-; GFX10-DL-NEXT:    v_lshrrev_b32_e32 v2, 24, v2
-; GFX10-DL-NEXT:    v_lshrrev_b32_e32 v3, 24, v3
-; GFX10-DL-NEXT:    v_mad_u16 v0, v0, v5, v4
-; GFX10-DL-NEXT:    v_mad_u16 v0, v6, v7, v0
-; GFX10-DL-NEXT:    v_mad_u16 v0, v2, v3, v0
+; GFX10-DL-NEXT:    v_dot4_u32_u8 v0, v2, v3, v4
 ; GFX10-DL-NEXT:    global_store_byte v1, v0, s[2:3]
 ; GFX10-DL-NEXT:    s_endpgm
+;
+; GFX11-DL-LABEL: idot4_acc8:
+; GFX11-DL:       ; %bb.0: ; %entry
+; GFX11-DL-NEXT:    s_clause 0x1
+; GFX11-DL-NEXT:    s_load_b128 s[4:7], s[0:1], 0x24
+; GFX11-DL-NEXT:    s_load_b64 s[0:1], s[0:1], 0x34
+; GFX11-DL-NEXT:    v_dual_mov_b32 v1, 0 :: v_dual_lshlrev_b32 v0, 2, v0
+; GFX11-DL-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX11-DL-NEXT:    s_clause 0x1
+; GFX11-DL-NEXT:    global_load_b32 v2, v0, s[4:5]
+; GFX11-DL-NEXT:    global_load_b32 v0, v0, s[6:7]
+; GFX11-DL-NEXT:    global_load_u8 v3, v1, s[0:1]
+; GFX11-DL-NEXT:    s_waitcnt vmcnt(0)
+; GFX11-DL-NEXT:    v_dot4_u32_u8 v0, v2, v0, v3
+; GFX11-DL-NEXT:    global_store_b8 v1, v0, s[0:1]
+; GFX11-DL-NEXT:    s_nop 0
+; GFX11-DL-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
+; GFX11-DL-NEXT:    s_endpgm
                                       ptr addrspace(1) %src2,
                                       ptr addrspace(1) nocapture %dst) {
 entry:
   %idx = call i32 @llvm.amdgcn.workitem.id.x()
   %gep1 = getelementptr <4 x i8>, ptr addrspace(1) %src1, i32 %idx
   %vec1 = load <4 x i8>, ptr addrspace(1) %gep1
   %gep2 = getelementptr <4 x i8>, ptr addrspace(1) %src2, i32 %idx
   %vec2 = load <4 x i8>, ptr addrspace(1) %gep2

 ; GFX9-DL-NEXT:    global_load_dword v1, v0, s[4:5]
 ; GFX9-DL-NEXT:    global_load_dword v2, v0, s[6:7]
 ; GFX9-DL-NEXT:    s_load_dword s0, s[2:3], 0x0
 ; GFX9-DL-NEXT:    v_mov_b32_e32 v0, 0
 ; GFX9-DL-NEXT:    s_waitcnt vmcnt(1)
 ; GFX9-DL-NEXT:    v_bfe_i32 v3, v1, 0, 8
 ; GFX9-DL-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-DL-NEXT:    v_bfe_i32 v4, v2, 0, 8
-; GFX9-DL-NEXT:    v_mul_i32_i24_sdwa v5, sext(v1), sext(v2) dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_1 src1_sel:BYTE_1
-; GFX9-DL-NEXT:    v_mul_i32_i24_sdwa v6, sext(v1), sext(v2) dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_2 src1_sel:BYTE_2
-; GFX9-DL-NEXT:    v_mul_i32_i24_sdwa v1, sext(v1), sext(v2) dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_3 src1_sel:BYTE_3
-; GFX9-DL-NEXT:    v_mul_i32_i24_e32 v2, v3, v4
 ; GFX9-DL-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX9-DL-NEXT:    v_mad_i32_i24 v3, v3, v4, s0
-; GFX9-DL-NEXT:    v_add3_u32 v2, v5, v3, v2
-; GFX9-DL-NEXT:    v_add3_u32 v1, v2, v6, v1
+; GFX9-DL-NEXT:    v_dot4_i32_i8 v1, v1, v2, v3
 ; GFX9-DL-NEXT:    global_store_dword v0, v1, s[2:3]
 ; GFX9-DL-NEXT:    s_endpgm
 ;
 ; GFX10-DL-LABEL: idot4_multiuse_mul1:
 ; GFX10-DL:       ; %bb.0: ; %entry
 ; GFX10-DL-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x24
 ; GFX10-DL-NEXT:    v_lshlrev_b32_e32 v0, 2, v0
 ; GFX10-DL-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0x34
 ; GFX10-DL-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX10-DL-NEXT:    s_clause 0x1
 ; GFX10-DL-NEXT:    global_load_dword v1, v0, s[4:5]
 ; GFX10-DL-NEXT:    global_load_dword v2, v0, s[6:7]
 ; GFX10-DL-NEXT:    s_load_dword s2, s[0:1], 0x0
 ; GFX10-DL-NEXT:    s_waitcnt vmcnt(1)
 ; GFX10-DL-NEXT:    v_bfe_i32 v0, v1, 0, 8
 ; GFX10-DL-NEXT:    s_waitcnt vmcnt(0)
 ; GFX10-DL-NEXT:    v_bfe_i32 v3, v2, 0, 8
-; GFX10-DL-NEXT:    v_mul_i32_i24_sdwa v4, sext(v1), sext(v2) dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_1 src1_sel:BYTE_1
-; GFX10-DL-NEXT:    v_mul_i32_i24_e32 v5, v0, v3
 ; GFX10-DL-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX10-DL-NEXT:    v_mad_i32_i24 v0, v0, v3, s2
-; GFX10-DL-NEXT:    v_mul_i32_i24_sdwa v3, sext(v1), sext(v2) dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_2 src1_sel:BYTE_2
-; GFX10-DL-NEXT:    v_mul_i32_i24_sdwa v1, sext(v1), sext(v2) dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_3 src1_sel:BYTE_3
-; GFX10-DL-NEXT:    v_mov_b32_e32 v2, 0
-; GFX10-DL-NEXT:    v_add3_u32 v0, v4, v0, v5
-; GFX10-DL-NEXT:    v_add3_u32 v0, v0, v3, v1
-; GFX10-DL-NEXT:    global_store_dword v2, v0, s[0:1]
+; GFX10-DL-NEXT:    v_mov_b32_e32 v3, 0
+; GFX10-DL-NEXT:    v_dot4c_i32_i8_e32 v0, v1, v2
+; GFX10-DL-NEXT:    global_store_dword v3, v0, s[0:1]
 ; GFX10-DL-NEXT:    s_endpgm
+;
+; GFX11-DL-LABEL: idot4_multiuse_mul1:
+; GFX11-DL:       ; %bb.0: ; %entry
+; GFX11-DL-NEXT:    s_load_b128 s[4:7], s[0:1], 0x24
+; GFX11-DL-NEXT:    v_lshlrev_b32_e32 v0, 2, v0
+; GFX11-DL-NEXT:    s_load_b64 s[0:1], s[0:1], 0x34
+; GFX11-DL-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX11-DL-NEXT:    s_clause 0x1
+; GFX11-DL-NEXT:    global_load_b32 v1, v0, s[4:5]
+; GFX11-DL-NEXT:    global_load_b32 v0, v0, s[6:7]
+; GFX11-DL-NEXT:    s_load_b32 s2, s[0:1], 0x0
+; GFX11-DL-NEXT:    s_waitcnt vmcnt(1)
+; GFX11-DL-NEXT:    v_bfe_i32 v2, v1, 0, 8
+; GFX11-DL-NEXT:    s_waitcnt vmcnt(0)
+; GFX11-DL-NEXT:    v_bfe_i32 v3, v0, 0, 8
+; GFX11-DL-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX11-DL-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2)
+; GFX11-DL-NEXT:    v_mad_i32_i24 v2, v2, v3, s2
+; GFX11-DL-NEXT:    v_mov_b32_e32 v3, 0
+; GFX11-DL-NEXT:    v_dot4_i32_iu8 v0, v1, v0, v2 neg_lo:[1,1,0]
+; GFX11-DL-NEXT:    global_store_b32 v3, v0, s[0:1]
+; GFX11-DL-NEXT:    s_nop 0
+; GFX11-DL-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
+; GFX11-DL-NEXT:    s_endpgm
                                                ptr addrspace(1) %src2,
                                                ptr addrspace(1) nocapture %dst) {
 entry:
   %idx = call i32 @llvm.amdgcn.workitem.id.x()
   %gep1 = getelementptr <4 x i8>, ptr addrspace(1) %src1, i32 %idx
   %vec1 = load <4 x i8>, ptr addrspace(1) %gep1
   %gep2 = getelementptr <4 x i8>, ptr addrspace(1) %src2, i32 %idx
   %vec2 = load <4 x i8>, ptr addrspace(1) %gep2
entry:
   %add2 = add i32 %add1, %mul1
   %add3 = add i32 %add2, %mul3
   %add4 = add i32 %add3, %mul4
 
   store i32 %add4, ptr addrspace(1) %dst, align 4
   ret void
 }
 
-; TODO: Support this pattern.
 define amdgpu_kernel void @idot4_acc32_vecMul(ptr addrspace(1) %src1,
 ; GFX7-LABEL: idot4_acc32_vecMul:
 ; GFX7:       ; %bb.0: ; %entry
 ; GFX7-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x9
 ; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xd
 ; GFX7-NEXT:    s_mov_b32 s3, 0xf000
 ; GFX7-NEXT:    s_mov_b32 s10, 0
 ; GFX7-NEXT:    s_mov_b32 s11, s3

 ; GFX9-DL-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x24
 ; GFX9-DL-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x34
 ; GFX9-DL-NEXT:    v_lshlrev_b32_e32 v0, 2, v0
 ; GFX9-DL-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX9-DL-NEXT:    global_load_dword v1, v0, s[4:5]
 ; GFX9-DL-NEXT:    global_load_dword v2, v0, s[6:7]
 ; GFX9-DL-NEXT:    s_load_dword s0, s[2:3], 0x0
 ; GFX9-DL-NEXT:    v_mov_b32_e32 v0, 0
-; GFX9-DL-NEXT:    s_waitcnt vmcnt(1)
-; GFX9-DL-NEXT:    v_lshrrev_b16_e32 v3, 8, v1
-; GFX9-DL-NEXT:    s_waitcnt vmcnt(0)
-; GFX9-DL-NEXT:    v_lshrrev_b16_e32 v4, 8, v2
-; GFX9-DL-NEXT:    v_mul_i32_i24_sdwa v5, sext(v1), sext(v2) dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:BYTE_0
-; GFX9-DL-NEXT:    v_mul_i32_i24_sdwa v3, sext(v3), sext(v4) dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:BYTE_0
-; GFX9-DL-NEXT:    v_mul_i32_i24_sdwa v6, sext(v1), sext(v2) dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_2 src1_sel:BYTE_2
-; GFX9-DL-NEXT:    v_mul_i32_i24_sdwa v1, sext(v1), sext(v2) dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_3 src1_sel:BYTE_3
-; GFX9-DL-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX9-DL-NEXT:    v_add3_u32 v2, v5, s0, v3
-; GFX9-DL-NEXT:    v_add3_u32 v1, v2, v6, v1
+; GFX9-DL-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX9-DL-NEXT:    v_dot4_i32_i8 v1, v1, v2, s0
 ; GFX9-DL-NEXT:    global_store_dword v0, v1, s[2:3]
 ; GFX9-DL-NEXT:    s_endpgm
 ;
 ; GFX10-DL-LABEL: idot4_acc32_vecMul:
 ; GFX10-DL:       ; %bb.0: ; %entry
 ; GFX10-DL-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x24
 ; GFX10-DL-NEXT:    v_lshlrev_b32_e32 v0, 2, v0
 ; GFX10-DL-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0x34
+; GFX10-DL-NEXT:    v_mov_b32_e32 v3, 0
 ; GFX10-DL-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX10-DL-NEXT:    s_clause 0x1
 ; GFX10-DL-NEXT:    global_load_dword v1, v0, s[4:5]
 ; GFX10-DL-NEXT:    global_load_dword v2, v0, s[6:7]
 ; GFX10-DL-NEXT:    s_load_dword s2, s[0:1], 0x0
-; GFX10-DL-NEXT:    s_waitcnt vmcnt(1)
-; GFX10-DL-NEXT:    v_lshrrev_b16 v0, 8, v1
-; GFX10-DL-NEXT:    s_waitcnt vmcnt(0)
-; GFX10-DL-NEXT:    v_lshrrev_b16 v3, 8, v2
-; GFX10-DL-NEXT:    v_mul_i32_i24_sdwa v4, sext(v1), sext(v2) dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:BYTE_0
-; GFX10-DL-NEXT:    v_mul_i32_i24_sdwa v0, sext(v0), sext(v3) dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:BYTE_0
-; GFX10-DL-NEXT:    v_mul_i32_i24_sdwa v3, sext(v1), sext(v2) dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_2 src1_sel:BYTE_2
-; GFX10-DL-NEXT:    v_mul_i32_i24_sdwa v1, sext(v1), sext(v2) dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_3 src1_sel:BYTE_3
-; GFX10-DL-NEXT:    v_mov_b32_e32 v2, 0
 ; GFX10-DL-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX10-DL-NEXT:    v_add3_u32 v0, v4, s2, v0
-; GFX10-DL-NEXT:    v_add3_u32 v0, v0, v3, v1
-; GFX10-DL-NEXT:    global_store_dword v2, v0, s[0:1]
+; GFX10-DL-NEXT:    v_mov_b32_e32 v0, s2
+; GFX10-DL-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-DL-NEXT:    v_dot4c_i32_i8_e32 v0, v1, v2
+; GFX10-DL-NEXT:    global_store_dword v3, v0, s[0:1]
 ; GFX10-DL-NEXT:    s_endpgm
+;
+; GFX11-DL-LABEL: idot4_acc32_vecMul:
+; GFX11-DL:       ; %bb.0: ; %entry
+; GFX11-DL-NEXT:    s_load_b128 s[4:7], s[0:1], 0x24
+; GFX11-DL-NEXT:    v_lshlrev_b32_e32 v0, 2, v0
+; GFX11-DL-NEXT:    s_load_b64 s[0:1], s[0:1], 0x34
+; GFX11-DL-NEXT:    v_mov_b32_e32 v2, 0
+; GFX11-DL-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX11-DL-NEXT:    s_clause 0x1
+; GFX11-DL-NEXT:    global_load_b32 v1, v0, s[4:5]
+; GFX11-DL-NEXT:    global_load_b32 v0, v0, s[6:7]
+; GFX11-DL-NEXT:    s_load_b32 s2, s[0:1], 0x0
+; GFX11-DL-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX11-DL-NEXT:    v_dot4_i32_iu8 v0, v1, v0, s2 neg_lo:[1,1,0]
+; GFX11-DL-NEXT:    global_store_b32 v2, v0, s[0:1]
+; GFX11-DL-NEXT:    s_nop 0
+; GFX11-DL-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
+; GFX11-DL-NEXT:    s_endpgm
                                               ptr addrspace(1) %src2,
                                               ptr addrspace(1) nocapture %dst) {
 entry:
   %idx = call i32 @llvm.amdgcn.workitem.id.x()
   %gep1 = getelementptr <4 x i8>, ptr addrspace(1) %src1, i32 %idx
   %vec1 = load <4 x i8>, ptr addrspace(1) %gep1
   %gep2 = getelementptr <4 x i8>, ptr addrspace(1) %src2, i32 %idx
   %vec2 = load <4 x i8>, ptr addrspace(1) %gep2

 ; GFX10-DL-NEXT:    v_add_nc_u16 v3, v4, v3
 ; GFX10-DL-NEXT:    v_pk_mul_lo_u16 v1, v1, v2
 ; GFX10-DL-NEXT:    v_add_nc_u16 v2, v3, v5
 ; GFX10-DL-NEXT:    v_lshrrev_b32_e32 v3, 16, v1
 ; GFX10-DL-NEXT:    v_add_nc_u16 v1, v2, v1
 ; GFX10-DL-NEXT:    v_add_nc_u16 v1, v1, v3
 ; GFX10-DL-NEXT:    global_store_short v0, v1, s[0:1]
 ; GFX10-DL-NEXT:    s_endpgm
+;
+; GFX11-DL-LABEL: idot4_acc16_vecMul:
+; GFX11-DL:       ; %bb.0: ; %entry
+; GFX11-DL-NEXT:    s_load_b128 s[4:7], s[0:1], 0x24
+; GFX11-DL-NEXT:    v_lshlrev_b32_e32 v0, 2, v0
+; GFX11-DL-NEXT:    s_load_b64 s[0:1], s[0:1], 0x34
+; GFX11-DL-NEXT:    v_mov_b32_e32 v2, 0
+; GFX11-DL-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX11-DL-NEXT:    s_clause 0x1
+; GFX11-DL-NEXT:    global_load_b32 v1, v0, s[4:5]
+; GFX11-DL-NEXT:    global_load_b32 v0, v0, s[6:7]
+; GFX11-DL-NEXT:    global_load_u16 v3, v2, s[0:1]
+; GFX11-DL-NEXT:    s_waitcnt vmcnt(2)
+; GFX11-DL-NEXT:    v_ashrrev_i16 v4, 8, v1
+; GFX11-DL-NEXT:    s_waitcnt vmcnt(1)
+; GFX11-DL-NEXT:    v_ashrrev_i16 v5, 8, v0
+; GFX11-DL-NEXT:    v_bfe_i32 v6, v0, 0, 8
+; GFX11-DL-NEXT:    v_bfe_i32 v7, v1, 0, 8
+; GFX11-DL-NEXT:    v_lshrrev_b32_e32 v1, 16, v1
+; GFX11-DL-NEXT:    v_lshrrev_b32_e32 v0, 16, v0
+; GFX11-DL-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
+; GFX11-DL-NEXT:    v_perm_b32 v5, v5, v6, 0x5040100
+; GFX11-DL-NEXT:    v_perm_b32 v4, v4, v7, 0x5040100
+; GFX11-DL-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
+; GFX11-DL-NEXT:    v_ashrrev_i16 v6, 8, v1
+; GFX11-DL-NEXT:    v_ashrrev_i16 v7, 8, v0
+; GFX11-DL-NEXT:    v_bfe_i32 v0, v0, 0, 8
+; GFX11-DL-NEXT:    v_bfe_i32 v1, v1, 0, 8
+; GFX11-DL-NEXT:    v_pk_mul_lo_u16 v4, v4, v5
+; GFX11-DL-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3)
+; GFX11-DL-NEXT:    v_perm_b32 v0, v7, v0, 0x5040100
+; GFX11-DL-NEXT:    v_perm_b32 v1, v6, v1, 0x5040100
+; GFX11-DL-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_2) | instid1(VALU_DEP_3)
+; GFX11-DL-NEXT:    v_lshrrev_b32_e32 v5, 16, v4
+; GFX11-DL-NEXT:    s_waitcnt vmcnt(0)
+; GFX11-DL-NEXT:    v_add_nc_u16 v3, v4, v3
+; GFX11-DL-NEXT:    v_pk_mul_lo_u16 v0, v1, v0
+; GFX11-DL-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX11-DL-NEXT:    v_add_nc_u16 v1, v3, v5
+; GFX11-DL-NEXT:    v_lshrrev_b32_e32 v3, 16, v0
+; GFX11-DL-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-DL-NEXT:    v_add_nc_u16 v0, v1, v0
+; GFX11-DL-NEXT:    v_add_nc_u16 v0, v0, v3
+; GFX11-DL-NEXT:    global_store_b16 v2, v0, s[0:1]
+; GFX11-DL-NEXT:    s_nop 0
+; GFX11-DL-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
+; GFX11-DL-NEXT:    s_endpgm
                                               ptr addrspace(1) %src2,
                                               ptr addrspace(1) nocapture %dst) {
 entry:
   %idx = call i32 @llvm.amdgcn.workitem.id.x()
   %gep1 = getelementptr <4 x i8>, ptr addrspace(1) %src1, i32 %idx
   %vec1 = load <4 x i8>, ptr addrspace(1) %gep1
   %gep2 = getelementptr <4 x i8>, ptr addrspace(1) %src2, i32 %idx
   %vec2 = load <4 x i8>, ptr addrspace(1) %gep2
entry:
   %add2 = add i16 %add1, %mul1
   %add3 = add i16 %add2, %mul2
   %add4 = add i16 %add3, %mul3
 
   store i16 %add4, ptr addrspace(1) %dst, align 4
   ret void
 }
 
+define amdgpu_kernel void @idot4_acc32_2ele(ptr addrspace(1) %src1,
+; GFX7-LABEL: idot4_acc32_2ele:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x9
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xd
+; GFX7-NEXT:    s_mov_b32 s3, 0xf000
+; GFX7-NEXT:    s_mov_b32 s10, 0
+; GFX7-NEXT:    s_mov_b32 s11, s3
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    s_mov_b64 s[8:9], s[4:5]
+; GFX7-NEXT:    v_lshlrev_b32_e32 v0, 2, v0
+; GFX7-NEXT:    v_mov_b32_e32 v1, 0
+; GFX7-NEXT:    buffer_load_dword v2, v[0:1], s[8:11], 0 addr64
+; GFX7-NEXT:    s_mov_b64 s[8:9], s[6:7]
+; GFX7-NEXT:    buffer_load_dword v0, v[0:1], s[8:11], 0 addr64
+; GFX7-NEXT:    s_load_dword s4, s[0:1], 0x0
+; GFX7-NEXT:    s_mov_b32 s2, -1
+; GFX7-NEXT:    s_waitcnt vmcnt(1)
+; GFX7-NEXT:    v_bfe_i32 v1, v2, 0, 8
+; GFX7-NEXT:    v_bfe_i32 v2, v2, 8, 8
+; GFX7-NEXT:    s_waitcnt vmcnt(0)
+; GFX7-NEXT:    v_bfe_i32 v3, v0, 0, 8
+; GFX7-NEXT:    v_bfe_i32 v0, v0, 8, 8
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mad_i32_i24 v1, v1, v3, s4
+; GFX7-NEXT:    v_mad_i32_i24 v0, v2, v0, v1
+; GFX7-NEXT:    buffer_store_dword v0, off, s[0:3], 0
+; GFX7-NEXT:    s_endpgm
+;
+; GFX8-LABEL: idot4_acc32_2ele:
+; GFX8:       ; %bb.0: ; %entry
+; GFX8-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x24
+; GFX8-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0x34
+; GFX8-NEXT:    v_lshlrev_b32_e32 v2, 2, v0
+; GFX8-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX8-NEXT:    v_mov_b32_e32 v1, s5
+; GFX8-NEXT:    v_add_u32_e32 v0, vcc, s4, v2
+; GFX8-NEXT:    v_addc_u32_e32 v1, vcc, 0, v1, vcc
+; GFX8-NEXT:    flat_load_dword v3, v[0:1]
+; GFX8-NEXT:    v_mov_b32_e32 v1, s7
+; GFX8-NEXT:    v_add_u32_e32 v0, vcc, s6, v2
+; GFX8-NEXT:    v_addc_u32_e32 v1, vcc, 0, v1, vcc
+; GFX8-NEXT:    flat_load_dword v0, v[0:1]
+; GFX8-NEXT:    s_load_dword s2, s[0:1], 0x0
+; GFX8-NEXT:    s_waitcnt vmcnt(1)
+; GFX8-NEXT:    v_bfe_i32 v1, v3, 0, 8
+; GFX8-NEXT:    v_bfe_i32 v3, v3, 8, 8
+; GFX8-NEXT:    s_waitcnt vmcnt(0)
+; GFX8-NEXT:    v_bfe_i32 v2, v0, 0, 8
+; GFX8-NEXT:    v_bfe_i32 v0, v0, 8, 8
+; GFX8-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX8-NEXT:    v_mad_i32_i24 v1, v1, v2, s2
+; GFX8-NEXT:    v_mad_i32_i24 v2, v3, v0, v1
+; GFX8-NEXT:    v_mov_b32_e32 v0, s0
+; GFX8-NEXT:    v_mov_b32_e32 v1, s1
+; GFX8-NEXT:    flat_store_dword v[0:1], v2
+; GFX8-NEXT:    s_endpgm
+;
+; GFX9-NODL-LABEL: idot4_acc32_2ele:
+; GFX9-NODL:       ; %bb.0: ; %entry
+; GFX9-NODL-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x24
+; GFX9-NODL-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x34
+; GFX9-NODL-NEXT:    v_lshlrev_b32_e32 v0, 2, v0
+; GFX9-NODL-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX9-NODL-NEXT:    global_load_dword v1, v0, s[4:5]
+; GFX9-NODL-NEXT:    global_load_dword v2, v0, s[6:7]
+; GFX9-NODL-NEXT:    s_load_dword s0, s[2:3], 0x0
+; GFX9-NODL-NEXT:    v_mov_b32_e32 v0, 0
+; GFX9-NODL-NEXT:    s_waitcnt vmcnt(0)
+; GFX9-NODL-NEXT:    v_mul_i32_i24_sdwa v3, sext(v1), sext(v2) dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:BYTE_0
+; GFX9-NODL-NEXT:    v_mul_i32_i24_sdwa v1, sext(v1), sext(v2) dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_1 src1_sel:BYTE_1
+; GFX9-NODL-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX9-NODL-NEXT:    v_add3_u32 v1, v3, s0, v1
+; GFX9-NODL-NEXT:    global_store_dword v0, v1, s[2:3]
+; GFX9-NODL-NEXT:    s_endpgm
+;
+; GFX9-DL-LABEL: idot4_acc32_2ele:
+; GFX9-DL:       ; %bb.0: ; %entry
+; GFX9-DL-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x24
+; GFX9-DL-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x34
+; GFX9-DL-NEXT:    v_lshlrev_b32_e32 v0, 2, v0
+; GFX9-DL-NEXT:    s_mov_b32 s1, 0xc0c0100
+; GFX9-DL-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX9-DL-NEXT:    global_load_dword v1, v0, s[6:7]
+; GFX9-DL-NEXT:    global_load_dword v2, v0, s[4:5]
+; GFX9-DL-NEXT:    s_load_dword s0, s[2:3], 0x0
+; GFX9-DL-NEXT:    v_mov_b32_e32 v0, 0
+; GFX9-DL-NEXT:    s_waitcnt vmcnt(1)
+; GFX9-DL-NEXT:    v_perm_b32 v1, v1, v1, s1
+; GFX9-DL-NEXT:    s_waitcnt vmcnt(0)
+; GFX9-DL-NEXT:    v_perm_b32 v2, v2, v2, s1
+; GFX9-DL-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX9-DL-NEXT:    v_dot4_i32_i8 v1, v2, v1, s0
+; GFX9-DL-NEXT:    global_store_dword v0, v1, s[2:3]
+; GFX9-DL-NEXT:    s_endpgm
+;
+; GFX10-DL-LABEL: idot4_acc32_2ele:
+; GFX10-DL:       ; %bb.0: ; %entry
+; GFX10-DL-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x24
+; GFX10-DL-NEXT:    v_lshlrev_b32_e32 v0, 2, v0
+; GFX10-DL-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0x34
+; GFX10-DL-NEXT:    v_mov_b32_e32 v3, 0
+; GFX10-DL-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-DL-NEXT:    s_clause 0x1
+; GFX10-DL-NEXT:    global_load_dword v1, v0, s[6:7]
+; GFX10-DL-NEXT:    global_load_dword v2, v0, s[4:5]
+; GFX10-DL-NEXT:    s_load_dword s2, s[0:1], 0x0
+; GFX10-DL-NEXT:    s_waitcnt vmcnt(1)
+; GFX10-DL-NEXT:    v_perm_b32 v0, v1, v1, 0xc0c0100
+; GFX10-DL-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-DL-NEXT:    v_perm_b32 v1, v2, v2, 0xc0c0100
+; GFX10-DL-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-DL-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-DL-NEXT:    v_dot4c_i32_i8_e32 v2, v1, v0
+; GFX10-DL-NEXT:    global_store_dword v3, v2, s[0:1]
+; GFX10-DL-NEXT:    s_endpgm
+;
+; GFX11-DL-LABEL: idot4_acc32_2ele:
+; GFX11-DL:       ; %bb.0: ; %entry
+; GFX11-DL-NEXT:    s_load_b128 s[4:7], s[0:1], 0x24
+; GFX11-DL-NEXT:    v_lshlrev_b32_e32 v0, 2, v0
+; GFX11-DL-NEXT:    s_load_b64 s[0:1], s[0:1], 0x34
+; GFX11-DL-NEXT:    v_mov_b32_e32 v2, 0
+; GFX11-DL-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX11-DL-NEXT:    s_clause 0x1
+; GFX11-DL-NEXT:    global_load_b32 v1, v0, s[6:7]
+; GFX11-DL-NEXT:    global_load_b32 v0, v0, s[4:5]
+; GFX11-DL-NEXT:    s_load_b32 s2, s[0:1], 0x0
+; GFX11-DL-NEXT:    s_waitcnt vmcnt(1)
+; GFX11-DL-NEXT:    v_perm_b32 v1, v1, v1, 0xc0c0100
+; GFX11-DL-NEXT:    s_waitcnt vmcnt(0)
+; GFX11-DL-NEXT:    v_perm_b32 v0, v0, v0, 0xc0c0100
+; GFX11-DL-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX11-DL-NEXT:    s_delay_alu instid0(VALU_DEP_1)
+; GFX11-DL-NEXT:    v_dot4_i32_iu8 v0, v0, v1, s2 neg_lo:[1,1,0]
+; GFX11-DL-NEXT:    global_store_b32 v2, v0, s[0:1]
+; GFX11-DL-NEXT:    s_nop 0
+; GFX11-DL-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
+; GFX11-DL-NEXT:    s_endpgm
+                                       ptr addrspace(1) %src2,
+                                       ptr addrspace(1) nocapture %dst) {
+entry:
+  %idx = call i32 @llvm.amdgcn.workitem.id.x()
+  %gep1 = getelementptr <4 x i8>, ptr addrspace(1) %src1, i32 %idx
+  %vec1 = load <4 x i8>, ptr addrspace(1) %gep1
+  %gep2 = getelementptr <4 x i8>, ptr addrspace(1) %src2, i32 %idx
+  %vec2 = load <4 x i8>, ptr addrspace(1) %gep2
+
+  %v1e0 = extractelement <4 x i8> %vec1, i64 0
+  %cv1e0 = sext i8 %v1e0 to i32
+  %v2e0 = extractelement <4 x i8> %vec2, i64 0
+  %cv2e0 = sext i8 %v2e0 to i32
+  %mul1 = mul nuw nsw i32 %cv1e0, %cv2e0
+
+  %v1e1 = extractelement <4 x i8> %vec1, i64 1
+  %cv1e1 = sext i8 %v1e1 to i32
+  %v2e1 = extractelement <4 x i8> %vec2, i64 1
+  %cv2e1 = sext i8 %v2e1 to i32
+  %mul2 = mul nuw nsw i32 %cv1e1, %cv2e1
+
+  %acc = load i32, ptr addrspace(1) %dst, align 4
+  %add1 = add i32 %mul1, %acc
+  %add2 = add i32 %add1, %mul2
+  store i32 %add2, ptr addrspace(1) %dst, align 4
+  ret void
+}
+
+
+define amdgpu_kernel void @idot4_acc32_3ele(ptr addrspace(1) %src1,
+; GFX7-LABEL: idot4_acc32_3ele:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x9
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xd
+; GFX7-NEXT:    s_mov_b32 s3, 0xf000
+; GFX7-NEXT:    s_mov_b32 s10, 0
+; GFX7-NEXT:    s_mov_b32 s11, s3
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    s_mov_b64 s[8:9], s[4:5]
+; GFX7-NEXT:    v_lshlrev_b32_e32 v0, 2, v0
+; GFX7-NEXT:    v_mov_b32_e32 v1, 0
+; GFX7-NEXT:    buffer_load_dword v2, v[0:1], s[8:11], 0 addr64
+; GFX7-NEXT:    s_mov_b64 s[8:9], s[6:7]
+; GFX7-NEXT:    buffer_load_dword v0, v[0:1], s[8:11], 0 addr64
+; GFX7-NEXT:    s_load_dword s4, s[0:1], 0x0
+; GFX7-NEXT:    s_mov_b32 s2, -1
+; GFX7-NEXT:    s_waitcnt vmcnt(1)
+; GFX7-NEXT:    v_bfe_i32 v1, v2, 0, 8
+; GFX7-NEXT:    v_bfe_i32 v3, v2, 8, 8
+; GFX7-NEXT:    s_waitcnt vmcnt(0)
+; GFX7-NEXT:    v_bfe_i32 v4, v0, 0, 8
+; GFX7-NEXT:    v_bfe_i32 v5, v0, 8, 8
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mad_i32_i24 v1, v1, v4, s4
+; GFX7-NEXT:    v_bfe_i32 v2, v2, 16, 8
+; GFX7-NEXT:    v_bfe_i32 v0, v0, 16, 8
+; GFX7-NEXT:    v_mad_i32_i24 v1, v3, v5, v1
+; GFX7-NEXT:    v_mad_i32_i24 v0, v2, v0, v1
+; GFX7-NEXT:    buffer_store_dword v0, off, s[0:3], 0
+; GFX7-NEXT:    s_endpgm
+;
+; GFX8-LABEL: idot4_acc32_3ele:
+; GFX8:       ; %bb.0: ; %entry
+; GFX8-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x24
+; GFX8-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0x34
+; GFX8-NEXT:    v_lshlrev_b32_e32 v2, 2, v0
+; GFX8-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX8-NEXT:    v_mov_b32_e32 v1, s5
+; GFX8-NEXT:    v_add_u32_e32 v0, vcc, s4, v2
+; GFX8-NEXT:    v_addc_u32_e32 v1, vcc, 0, v1, vcc
+; GFX8-NEXT:    flat_load_dword v3, v[0:1]
+; GFX8-NEXT:    v_mov_b32_e32 v1, s7
+; GFX8-NEXT:    v_add_u32_e32 v0, vcc, s6, v2
+; GFX8-NEXT:    v_addc_u32_e32 v1, vcc, 0, v1, vcc
+; GFX8-NEXT:    flat_load_dword v0, v[0:1]
+; GFX8-NEXT:    s_load_dword s2, s[0:1], 0x0
+; GFX8-NEXT:    s_waitcnt vmcnt(1)
+; GFX8-NEXT:    v_bfe_i32 v1, v3, 0, 8
+; GFX8-NEXT:    v_bfe_i32 v4, v3, 8, 8
+; GFX8-NEXT:    v_bfe_i32 v3, v3, 16, 8
+; GFX8-NEXT:    s_waitcnt vmcnt(0)
+; GFX8-NEXT:    v_bfe_i32 v2, v0, 0, 8
+; GFX8-NEXT:    v_bfe_i32 v5, v0, 8, 8
+; GFX8-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX8-NEXT:    v_mad_i32_i24 v1, v1, v2, s2
+; GFX8-NEXT:    v_bfe_i32 v0, v0, 16, 8
+; GFX8-NEXT:    v_mad_i32_i24 v1, v4, v5, v1
+; GFX8-NEXT:    v_mad_i32_i24 v2, v3, v0, v1
+; GFX8-NEXT:    v_mov_b32_e32 v0, s0
+; GFX8-NEXT:    v_mov_b32_e32 v1, s1
+; GFX8-NEXT:    flat_store_dword v[0:1], v2
+; GFX8-NEXT:    s_endpgm
+;
+; GFX9-NODL-LABEL: idot4_acc32_3ele:
+; GFX9-NODL:       ; %bb.0: ; %entry
+; GFX9-NODL-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x24
+; GFX9-NODL-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x34
+; GFX9-NODL-NEXT:    v_lshlrev_b32_e32 v0, 2, v0
+; GFX9-NODL-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX9-NODL-NEXT:    global_load_dword v1, v0, s[4:5]
+; GFX9-NODL-NEXT:    global_load_dword v2, v0, s[6:7]
+; GFX9-NODL-NEXT:    s_load_dword s0, s[2:3], 0x0
+; GFX9-NODL-NEXT:    v_mov_b32_e32 v0, 0
+; GFX9-NODL-NEXT:    s_waitcnt vmcnt(1)
+; GFX9-NODL-NEXT:    v_bfe_i32 v3, v1, 0, 8
+; GFX9-NODL-NEXT:    s_waitcnt vmcnt(0)
+; GFX9-NODL-NEXT:    v_bfe_i32 v4, v2, 0, 8
+; GFX9-NODL-NEXT:    v_mul_i32_i24_sdwa v5, sext(v1), sext(v2) dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_1 src1_sel:BYTE_1
+; GFX9-NODL-NEXT:    v_mul_i32_i24_sdwa v1, sext(v1), sext(v2) dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_2 src1_sel:BYTE_2
+; GFX9-NODL-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX9-NODL-NEXT:    v_mad_i32_i24 v2, v3, v4, s0
+; GFX9-NODL-NEXT:    v_add3_u32 v1, v2, v5, v1
+; GFX9-NODL-NEXT:    global_store_dword v0, v1, s[2:3]
+; GFX9-NODL-NEXT:    s_endpgm
+;
+; GFX9-DL-LABEL: idot4_acc32_3ele:
+; GFX9-DL:       ; %bb.0: ; %entry
+; GFX9-DL-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x24
+; GFX9-DL-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x34
+; GFX9-DL-NEXT:    v_lshlrev_b32_e32 v0, 2, v0
+; GFX9-DL-NEXT:    s_mov_b32 s1, 0xc020100
+; GFX9-DL-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX9-DL-NEXT:    global_load_dword v1, v0, s[6:7]
+; GFX9-DL-NEXT:    global_load_dword v2, v0, s[4:5]
+; GFX9-DL-NEXT:    s_load_dword s0, s[2:3], 0x0
+; GFX9-DL-NEXT:    v_mov_b32_e32 v0, 0
+; GFX9-DL-NEXT:    s_waitcnt vmcnt(1)
+; GFX9-DL-NEXT:    v_perm_b32 v1, v1, v1, s1
+; GFX9-DL-NEXT:    s_waitcnt vmcnt(0)
+; GFX9-DL-NEXT:    v_perm_b32 v2, v2, v2, s1
+; GFX9-DL-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX9-DL-NEXT:    v_dot4_i32_i8 v1, v2, v1, s0
+; GFX9-DL-NEXT:    global_store_dword v0, v1, s[2:3]
+; GFX9-DL-NEXT:    s_endpgm
+;
+; GFX10-DL-LABEL: idot4_acc32_3ele:
+; GFX10-DL:       ; %bb.0: ; %entry
+; GFX10-DL-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x24
+; GFX10-DL-NEXT:    v_lshlrev_b32_e32 v0, 2, v0
+; GFX10-DL-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0x34
+; GFX10-DL-NEXT:    v_mov_b32_e32 v3, 0
+; GFX10-DL-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-DL-NEXT:    s_clause 0x1
+; GFX10-DL-NEXT:    global_load_dword v1, v0, s[6:7]
+; GFX10-DL-NEXT:    global_load_dword v2, v0, s[4:5]
+; GFX10-DL-NEXT:    s_load_dword s2, s[0:1], 0x0
+; GFX10-DL-NEXT:    s_waitcnt vmcnt(1)
+; GFX10-DL-NEXT:    v_perm_b32 v0, v1, v1, 0xc020100
+; GFX10-DL-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-DL-NEXT:    v_perm_b32 v1, v2, v2, 0xc020100
+; GFX10-DL-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-DL-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-DL-NEXT:    v_dot4c_i32_i8_e32 v2, v1, v0
+; GFX10-DL-NEXT:    global_store_dword v3, v2, s[0:1]
+; GFX10-DL-NEXT:    s_endpgm
+;
+; GFX11-DL-LABEL: idot4_acc32_3ele:
+; GFX11-DL:       ; %bb.0: ; %entry
+; GFX11-DL-NEXT:    s_load_b128 s[4:7], s[0:1], 0x24
+; GFX11-DL-NEXT:    v_lshlrev_b32_e32 v0, 2, v0
+; GFX11-DL-NEXT:    s_load_b64 s[0:1], s[0:1], 0x34
+; GFX11-DL-NEXT:    v_mov_b32_e32 v2, 0
+; GFX11-DL-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX11-DL-NEXT:    s_clause 0x1
+; GFX11-DL-NEXT:    global_load_b32 v1, v0, s[6:7]
+; GFX11-DL-NEXT:    global_load_b32 v0, v0, s[4:5]
+; GFX11-DL-NEXT:    s_load_b32 s2, s[0:1], 0x0
+; GFX11-DL-NEXT:    s_waitcnt vmcnt(1)
+; GFX11-DL-NEXT:    v_perm_b32 v1, v1, v1, 0xc020100
+; GFX11-DL-NEXT:    s_waitcnt vmcnt(0)
+; GFX11-DL-NEXT:    v_perm_b32 v0, v0, v0, 0xc020100
+; GFX11-DL-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX11-DL-NEXT:    s_delay_alu instid0(VALU_DEP_1)
+; GFX11-DL-NEXT:    v_dot4_i32_iu8 v0, v0, v1, s2 neg_lo:[1,1,0]
+; GFX11-DL-NEXT:    global_store_b32 v2, v0, s[0:1]
+; GFX11-DL-NEXT:    s_nop 0
+; GFX11-DL-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
+; GFX11-DL-NEXT:    s_endpgm
+                                       ptr addrspace(1) %src2,
+                                       ptr addrspace(1) nocapture %dst) {
+entry:
+  %idx = call i32 @llvm.amdgcn.workitem.id.x()
+  %gep1 = getelementptr <4 x i8>, ptr addrspace(1) %src1, i32 %idx
+  %vec1 = load <4 x i8>, ptr addrspace(1) %gep1
+  %gep2 = getelementptr <4 x i8>, ptr addrspace(1) %src2, i32 %idx
+  %vec2 = load <4 x i8>, ptr addrspace(1) %gep2
+
+  %v1e0 = extractelement <4 x i8> %vec1, i64 0
+  %cv1e0 = sext i8 %v1e0 to i32
+  %v2e0 = extractelement <4 x i8> %vec2, i64 0
+  %cv2e0 = sext i8 %v2e0 to i32
+  %mul1 = mul nuw nsw i32 %cv1e0, %cv2e0
+
+  %v1e1 = extractelement <4 x i8> %vec1, i64 1
+  %cv1e1 = sext i8 %v1e1 to i32
+  %v2e1 = extractelement <4 x i8> %vec2, i64 1
+  %cv2e1 = sext i8 %v2e1 to i32
+  %mul2 = mul nuw nsw i32 %cv1e1, %cv2e1
+
+  %v1e2 = extractelement <4 x i8> %vec1, i64 2
+  %cv1e2 = sext i8 %v1e2 to i32
+  %v2e2 = extractelement <4 x i8> %vec2, i64 2
+  %cv2e2 = sext i8 %v2e2 to i32
+  %mul3 = mul nuw nsw i32 %cv1e2, %cv2e2
+
+  %acc = load i32, ptr addrspace(1) %dst, align 4
+  %add1 = add i32 %mul1, %acc
+  %add2 = add i32 %add1, %mul2
+  %add3 = add i32 %add2, %mul3
+  store i32 %add3, ptr addrspace(1) %dst, align 4
+  ret void
+}
+
+
+define amdgpu_kernel void @idot4_acc32_3ele_permuted(ptr addrspace(1) %src1,
+; GFX7-LABEL: idot4_acc32_3ele_permuted:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x9
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xd
+; GFX7-NEXT:    s_mov_b32 s3, 0xf000
+; GFX7-NEXT:    s_mov_b32 s10, 0
+; GFX7-NEXT:    s_mov_b32 s11, s3
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    s_mov_b64 s[8:9], s[4:5]
+; GFX7-NEXT:    v_lshlrev_b32_e32 v0, 2, v0
+; GFX7-NEXT:    v_mov_b32_e32 v1, 0
+; GFX7-NEXT:    buffer_load_dword v2, v[0:1], s[8:11], 0 addr64
+; GFX7-NEXT:    s_mov_b64 s[8:9], s[6:7]
+; GFX7-NEXT:    buffer_load_dword v0, v[0:1], s[8:11], 0 addr64
+; GFX7-NEXT:    s_load_dword s4, s[0:1], 0x0
+; GFX7-NEXT:    s_mov_b32 s2, -1
+; GFX7-NEXT:    s_waitcnt vmcnt(1)
+; GFX7-NEXT:    v_ashrrev_i32_e32 v1, 24, v2
+; GFX7-NEXT:    v_bfe_i32 v3, v2, 0, 8
+; GFX7-NEXT:    s_waitcnt vmcnt(0)
+; GFX7-NEXT:    v_ashrrev_i32_e32 v4, 24, v0
+; GFX7-NEXT:    v_bfe_i32 v5, v0, 0, 8
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mad_i32_i24 v1, v1, v4, s4
+; GFX7-NEXT:    v_bfe_i32 v2, v2, 16, 8
+; GFX7-NEXT:    v_bfe_i32 v0, v0, 16, 8
+; GFX7-NEXT:    v_mad_i32_i24 v1, v3, v5, v1
+; GFX7-NEXT:    v_mad_i32_i24 v0, v2, v0, v1
+; GFX7-NEXT:    buffer_store_dword v0, off, s[0:3], 0
+; GFX7-NEXT:    s_endpgm
+;
+; GFX8-LABEL: idot4_acc32_3ele_permuted:
+; GFX8:       ; %bb.0: ; %entry
+; GFX8-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x24
+; GFX8-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0x34
+; GFX8-NEXT:    v_lshlrev_b32_e32 v2, 2, v0
+; GFX8-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX8-NEXT:    v_mov_b32_e32 v1, s5
+; GFX8-NEXT:    v_add_u32_e32 v0, vcc, s4, v2
+; GFX8-NEXT:    v_addc_u32_e32 v1, vcc, 0, v1, vcc
+; GFX8-NEXT:    flat_load_dword v3, v[0:1]
+; GFX8-NEXT:    v_mov_b32_e32 v1, s7
+; GFX8-NEXT:    v_add_u32_e32 v0, vcc, s6, v2
+; GFX8-NEXT:    v_addc_u32_e32 v1, vcc, 0, v1, vcc
+; GFX8-NEXT:    flat_load_dword v0, v[0:1]
+; GFX8-NEXT:    s_load_dword s2, s[0:1], 0x0
+; GFX8-NEXT:    s_waitcnt vmcnt(1)
+; GFX8-NEXT:    v_ashrrev_i32_e32 v1, 24, v3
+; GFX8-NEXT:    v_bfe_i32 v4, v3, 0, 8
+; GFX8-NEXT:    v_bfe_i32 v3, v3, 16, 8
+; GFX8-NEXT:    s_waitcnt vmcnt(0)
+; GFX8-NEXT:    v_ashrrev_i32_e32 v2, 24, v0
+; GFX8-NEXT:    v_bfe_i32 v5, v0, 0, 8
+; GFX8-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX8-NEXT:    v_mad_i32_i24 v1, v1, v2, s2
+; GFX8-NEXT:    v_bfe_i32 v0, v0, 16, 8
+; GFX8-NEXT:    v_mad_i32_i24 v1, v4, v5, v1
+; GFX8-NEXT:    v_mad_i32_i24 v2, v3, v0, v1
+; GFX8-NEXT:    v_mov_b32_e32 v0, s0
+; GFX8-NEXT:    v_mov_b32_e32 v1, s1
+; GFX8-NEXT:    flat_store_dword v[0:1], v2
+; GFX8-NEXT:    s_endpgm
+;
+; GFX9-NODL-LABEL: idot4_acc32_3ele_permuted:
+; GFX9-NODL:       ; %bb.0: ; %entry
+; GFX9-NODL-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x24
+; GFX9-NODL-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x34
+; GFX9-NODL-NEXT:    v_lshlrev_b32_e32 v0, 2, v0
+; GFX9-NODL-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX9-NODL-NEXT:    global_load_dword v1, v0, s[4:5]
+; GFX9-NODL-NEXT:    global_load_dword v2, v0, s[6:7]
+; GFX9-NODL-NEXT:    s_load_dword s0, s[2:3], 0x0
+; GFX9-NODL-NEXT:    v_mov_b32_e32 v0, 0
+; GFX9-NODL-NEXT:    s_waitcnt vmcnt(1)
+; GFX9-NODL-NEXT:    v_ashrrev_i32_e32 v3, 24, v1
+; GFX9-NODL-NEXT:    s_waitcnt vmcnt(0)
+; GFX9-NODL-NEXT:    v_ashrrev_i32_e32 v4, 24, v2
+; GFX9-NODL-NEXT:    v_mul_i32_i24_sdwa v5, sext(v1), sext(v2) dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:BYTE_0
+; GFX9-NODL-NEXT:    v_mul_i32_i24_sdwa v1, sext(v1), sext(v2) dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_2 src1_sel:BYTE_2
+; GFX9-NODL-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX9-NODL-NEXT:    v_mad_i32_i24 v2, v3, v4, s0
+; GFX9-NODL-NEXT:    v_add3_u32 v1, v2, v5, v1
+; GFX9-NODL-NEXT:    global_store_dword v0, v1, s[2:3]
+; GFX9-NODL-NEXT:    s_endpgm
+;
+; GFX9-DL-LABEL: idot4_acc32_3ele_permuted:
+; GFX9-DL:       ; %bb.0: ; %entry
+; GFX9-DL-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x24
+; GFX9-DL-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x34
+; GFX9-DL-NEXT:    v_lshlrev_b32_e32 v0, 2, v0
+; GFX9-DL-NEXT:    s_mov_b32 s1, 0xc020003
+; GFX9-DL-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX9-DL-NEXT:    global_load_dword v1, v0, s[6:7]
+; GFX9-DL-NEXT:    global_load_dword v2, v0, s[4:5]
+; GFX9-DL-NEXT:    s_load_dword s0, s[2:3], 0x0
+; GFX9-DL-NEXT:    v_mov_b32_e32 v0, 0
+; GFX9-DL-NEXT:    s_waitcnt vmcnt(1)
+; GFX9-DL-NEXT:    v_perm_b32 v1, v1, v1, s1
+; GFX9-DL-NEXT:    s_waitcnt vmcnt(0)
+; GFX9-DL-NEXT:    v_perm_b32 v2, v2, v2, s1
+; GFX9-DL-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX9-DL-NEXT:    v_dot4_i32_i8 v1, v2, v1, s0
+; GFX9-DL-NEXT:    global_store_dword v0, v1, s[2:3]
+; GFX9-DL-NEXT:    s_endpgm
+;
+; GFX10-DL-LABEL: idot4_acc32_3ele_permuted:
+; GFX10-DL:       ; %bb.0: ; %entry
+; GFX10-DL-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x24
+; GFX10-DL-NEXT:    v_lshlrev_b32_e32 v0, 2, v0
+; GFX10-DL-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0x34
+; GFX10-DL-NEXT:    v_mov_b32_e32 v3, 0
+; GFX10-DL-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-DL-NEXT:    s_clause 0x1
+; GFX10-DL-NEXT:    global_load_dword v1, v0, s[6:7]
+; GFX10-DL-NEXT:    global_load_dword v2, v0, s[4:5]
+; GFX10-DL-NEXT:    s_load_dword s2, s[0:1], 0x0
+; GFX10-DL-NEXT:    s_waitcnt vmcnt(1)
+; GFX10-DL-NEXT:    v_perm_b32 v0, v1, v1, 0xc020003
+; GFX10-DL-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-DL-NEXT:    v_perm_b32 v1, v2, v2, 0xc020003
+; GFX10-DL-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-DL-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-DL-NEXT:    v_dot4c_i32_i8_e32 v2, v1, v0
+; GFX10-DL-NEXT:    global_store_dword v3, v2, s[0:1]
+; GFX10-DL-NEXT:    s_endpgm
+;
+; GFX11-DL-LABEL: idot4_acc32_3ele_permuted:
+; GFX11-DL:       ; %bb.0: ; %entry
+; GFX11-DL-NEXT:    s_load_b128 s[4:7], s[0:1], 0x24
+; GFX11-DL-NEXT:    v_lshlrev_b32_e32 v0, 2, v0
+; GFX11-DL-NEXT:    s_load_b64 s[0:1], s[0:1], 0x34
+; GFX11-DL-NEXT:    v_mov_b32_e32 v2, 0
+; GFX11-DL-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX11-DL-NEXT:    s_clause 0x1
+; GFX11-DL-NEXT:    global_load_b32 v1, v0, s[6:7]
+; GFX11-DL-NEXT:    global_load_b32 v0, v0, s[4:5]
+; GFX11-DL-NEXT:    s_load_b32 s2, s[0:1], 0x0
+; GFX11-DL-NEXT:    s_waitcnt vmcnt(1)
+; GFX11-DL-NEXT:    v_perm_b32 v1, v1, v1, 0xc020003
+; GFX11-DL-NEXT:    s_waitcnt vmcnt(0)
+; GFX11-DL-NEXT:    v_perm_b32 v0, v0, v0, 0xc020003
+; GFX11-DL-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX11-DL-NEXT:    s_delay_alu instid0(VALU_DEP_1)
+; GFX11-DL-NEXT:    v_dot4_i32_iu8 v0, v0, v1, s2 neg_lo:[1,1,0]
+; GFX11-DL-NEXT:    global_store_b32 v2, v0, s[0:1]
+; GFX11-DL-NEXT:    s_nop 0
+; GFX11-DL-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
+; GFX11-DL-NEXT:    s_endpgm
+                                       ptr addrspace(1) %src2,
+                                       ptr addrspace(1) nocapture %dst) {
+entry:
+  %idx = call i32 @llvm.amdgcn.workitem.id.x()
+  %gep1 = getelementptr <4 x i8>, ptr addrspace(1) %src1, i32 %idx
+  %vec1 = load <4 x i8>, ptr addrspace(1) %gep1
+  %gep2 = getelementptr <4 x i8>, ptr addrspace(1) %src2, i32 %idx
+  %vec2 = load <4 x i8>, ptr addrspace(1) %gep2
+
+  %v1e0 = extractelement <4 x i8> %vec1, i64 3
+  %cv1e0 = sext i8 %v1e0 to i32
+  %v2e0 = extractelement <4 x i8> %vec2, i64 3
+  %cv2e0 = sext i8 %v2e0 to i32
+  %mul1 = mul nuw nsw i32 %cv1e0, %cv2e0
+
+  %v1e1 = extractelement <4 x i8> %vec1, i64 0
+  %cv1e1 = sext i8 %v1e1 to i32
+  %v2e1 = extractelement <4 x i8> %vec2, i64 0
+  %cv2e1 = sext i8 %v2e1 to i32
+  %mul2 = mul nuw nsw i32 %cv1e1, %cv2e1
+
+  %v1e2 = extractelement <4 x i8> %vec1, i64 2
+  %cv1e2 = sext i8 %v1e2 to i32
+  %v2e2 = extractelement <4 x i8> %vec2, i64 2
+  %cv2e2 = sext i8 %v2e2 to i32
+  %mul3 = mul nuw nsw i32 %cv1e2, %cv2e2
+
+  %acc = load i32, ptr addrspace(1) %dst, align 4
+  %add1 = add i32 %mul1, %acc
+  %add2 = add i32 %add1, %mul2
+  %add3 = add i32 %add2, %mul3
+  store i32 %add3, ptr addrspace(1) %dst, align 4
+  ret void
+}
+
+define amdgpu_kernel void @idot4_acc32_opt(ptr addrspace(1) %src1,
+; GFX7-LABEL: idot4_acc32_opt:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x9
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xd
+; GFX7-NEXT:    s_mov_b32 s3, 0xf000
+; GFX7-NEXT:    s_mov_b32 s10, 0
+; GFX7-NEXT:    s_mov_b32 s11, s3
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    s_mov_b64 s[8:9], s[4:5]
+; GFX7-NEXT:    v_lshlrev_b32_e32 v0, 2, v0
+; GFX7-NEXT:    v_mov_b32_e32 v1, 0
+; GFX7-NEXT:    buffer_load_dword v2, v[0:1], s[8:11], 0 addr64
+; GFX7-NEXT:    s_mov_b64 s[8:9], s[6:7]
+; GFX7-NEXT:    buffer_load_dword v0, v[0:1], s[8:11], 0 addr64
+; GFX7-NEXT:    s_mov_b32 s2, -1
+; GFX7-NEXT:    s_waitcnt vmcnt(1)
+; GFX7-NEXT:    v_bfe_i32 v3, v2, 8, 8
+; GFX7-NEXT:    v_bfe_i32 v1, v2, 0, 8
+; GFX7-NEXT:    s_waitcnt vmcnt(0)
+; GFX7-NEXT:    v_bfe_i32 v6, v0, 8, 8
+; GFX7-NEXT:    v_bfe_i32 v5, v0, 0, 8
+; GFX7-NEXT:    v_mul_i32_i24_e32 v3, v3, v6
+; GFX7-NEXT:    v_bfe_i32 v4, v2, 16, 8
+; GFX7-NEXT:    v_bfe_i32 v7, v0, 16, 8
+; GFX7-NEXT:    v_mad_i32_i24 v1, v1, v5, v3
+; GFX7-NEXT:    v_ashrrev_i32_e32 v2, 24, v2
+; GFX7-NEXT:    v_ashrrev_i32_e32 v0, 24, v0
+; GFX7-NEXT:    v_mad_i32_i24 v1, v4, v7, v1
+; GFX7-NEXT:    v_mad_i32_i24 v0, v2, v0, v1
+; GFX7-NEXT:    buffer_store_dword v0, off, s[0:3], 0
+; GFX7-NEXT:    s_endpgm
+;
+; GFX8-LABEL: idot4_acc32_opt:
+; GFX8:       ; %bb.0: ; %entry
+; GFX8-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x24
+; GFX8-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0x34
+; GFX8-NEXT:    v_lshlrev_b32_e32 v2, 2, v0
+; GFX8-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX8-NEXT:    v_mov_b32_e32 v1, s5
+; GFX8-NEXT:    v_add_u32_e32 v0, vcc, s4, v2
+; GFX8-NEXT:    v_addc_u32_e32 v1, vcc, 0, v1, vcc
+; GFX8-NEXT:    flat_load_dword v3, v[0:1]
+; GFX8-NEXT:    v_mov_b32_e32 v1, s7
+; GFX8-NEXT:    v_add_u32_e32 v0, vcc, s6, v2
+; GFX8-NEXT:    v_addc_u32_e32 v1, vcc, 0, v1, vcc
+; GFX8-NEXT:    flat_load_dword v2, v[0:1]
+; GFX8-NEXT:    v_mov_b32_e32 v0, s0
+; GFX8-NEXT:    v_mov_b32_e32 v1, s1
+; GFX8-NEXT:    s_waitcnt vmcnt(1)
+; GFX8-NEXT:    v_bfe_i32 v4, v3, 0, 8
+; GFX8-NEXT:    v_bfe_i32 v7, v3, 16, 8
+; GFX8-NEXT:    s_waitcnt vmcnt(0)
+; GFX8-NEXT:    v_bfe_i32 v5, v2, 0, 8
+; GFX8-NEXT:    v_mul_i32_i24_sdwa v6, sext(v3), sext(v2) dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_1 src1_sel:BYTE_1
+; GFX8-NEXT:    v_bfe_i32 v8, v2, 16, 8
+; GFX8-NEXT:    v_mad_i32_i24 v4, v4, v5, v6
+; GFX8-NEXT:    v_ashrrev_i32_e32 v3, 24, v3
+; GFX8-NEXT:    v_ashrrev_i32_e32 v2, 24, v2
+; GFX8-NEXT:    v_mad_i32_i24 v4, v7, v8, v4
+; GFX8-NEXT:    v_mad_i32_i24 v2, v3, v2, v4
+; GFX8-NEXT:    flat_store_dword v[0:1], v2
+; GFX8-NEXT:    s_endpgm
+;
+; GFX9-NODL-LABEL: idot4_acc32_opt:
+; GFX9-NODL:       ; %bb.0: ; %entry
+; GFX9-NODL-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x24
+; GFX9-NODL-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x34
+; GFX9-NODL-NEXT:    v_lshlrev_b32_e32 v0, 2, v0
+; GFX9-NODL-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX9-NODL-NEXT:    global_load_dword v1, v0, s[4:5]
+; GFX9-NODL-NEXT:    global_load_dword v2, v0, s[6:7]
+; GFX9-NODL-NEXT:    v_mov_b32_e32 v0, 0
+; GFX9-NODL-NEXT:    s_waitcnt vmcnt(1)
+; GFX9-NODL-NEXT:    v_bfe_i32 v3, v1, 0, 8
+; GFX9-NODL-NEXT:    s_waitcnt vmcnt(0)
+; GFX9-NODL-NEXT:    v_bfe_i32 v4, v2, 0, 8
+; GFX9-NODL-NEXT:    v_mul_i32_i24_sdwa v5, sext(v1), sext(v2) dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_1 src1_sel:BYTE_1
+; GFX9-NODL-NEXT:    v_mul_i32_i24_sdwa v6, sext(v1), sext(v2) dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_2 src1_sel:BYTE_2
+; GFX9-NODL-NEXT:    v_mul_i32_i24_sdwa v1, sext(v1), sext(v2) dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_3 src1_sel:BYTE_3
+; GFX9-NODL-NEXT:    v_mad_i32_i24 v2, v3, v4, v5
+; GFX9-NODL-NEXT:    v_add3_u32 v1, v2, v6, v1
+; GFX9-NODL-NEXT:    global_store_dword v0, v1, s[2:3]
+; GFX9-NODL-NEXT:    s_endpgm
+;
+; GFX9-DL-LABEL: idot4_acc32_opt:
+; GFX9-DL:       ; %bb.0: ; %entry
+; GFX9-DL-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x24
+; GFX9-DL-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x34
+; GFX9-DL-NEXT:    v_lshlrev_b32_e32 v0, 2, v0
+; GFX9-DL-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX9-DL-NEXT:    global_load_dword v1, v0, s[4:5]
+; GFX9-DL-NEXT:    global_load_dword v2, v0, s[6:7]
+; GFX9-DL-NEXT:    v_mov_b32_e32 v0, 0
+; GFX9-DL-NEXT:    s_waitcnt vmcnt(0)
+; GFX9-DL-NEXT:    v_dot4_i32_i8 v1, v1, v2, 0
+; GFX9-DL-NEXT:    global_store_dword v0, v1, s[2:3]
+; GFX9-DL-NEXT:    s_endpgm
+;
+; GFX10-DL-LABEL: idot4_acc32_opt:
+; GFX10-DL:       ; %bb.0: ; %entry
+; GFX10-DL-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x24
+; GFX10-DL-NEXT:    v_lshlrev_b32_e32 v0, 2, v0
+; GFX10-DL-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0x34
+; GFX10-DL-NEXT:    v_mov_b32_e32 v3, 0
+; GFX10-DL-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-DL-NEXT:    s_clause 0x1
+; GFX10-DL-NEXT:    global_load_dword v1, v0, s[4:5]
+; GFX10-DL-NEXT:    global_load_dword v2, v0, s[6:7]
+; GFX10-DL-NEXT:    v_mov_b32_e32 v0, 0
+; GFX10-DL-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-DL-NEXT:    v_dot4c_i32_i8_e32 v0, v1, v2
+; GFX10-DL-NEXT:    global_store_dword v3, v0, s[0:1]
+; GFX10-DL-NEXT:    s_endpgm
+;
+; GFX11-DL-LABEL: idot4_acc32_opt:
+; GFX11-DL:       ; %bb.0: ; %entry
+; GFX11-DL-NEXT:    s_load_b128 s[4:7], s[0:1], 0x24
+; GFX11-DL-NEXT:    v_lshlrev_b32_e32 v0, 2, v0
+; GFX11-DL-NEXT:    s_load_b64 s[0:1], s[0:1], 0x34
+; GFX11-DL-NEXT:    v_mov_b32_e32 v2, 0
+; GFX11-DL-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX11-DL-NEXT:    s_clause 0x1
+; GFX11-DL-NEXT:    global_load_b32 v1, v0, s[4:5]
+; GFX11-DL-NEXT:    global_load_b32 v0, v0, s[6:7]
+; GFX11-DL-NEXT:    s_waitcnt vmcnt(0)
+; GFX11-DL-NEXT:    v_dot4_i32_iu8 v0, v1, v0, 0 neg_lo:[1,1,0]
+; GFX11-DL-NEXT:    global_store_b32 v2, v0, s[0:1]
+; GFX11-DL-NEXT:    s_nop 0
+; GFX11-DL-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
+; GFX11-DL-NEXT:    s_endpgm
+                                       ptr addrspace(1) %src2,
+                                       ptr addrspace(1) nocapture %dst) {
+entry:
+  %idx = call i32 @llvm.amdgcn.workitem.id.x()
+  %gep1 = getelementptr <4 x i8>, ptr addrspace(1) %src1, i32 %idx
+  %vec1 = load <4 x i8>, ptr addrspace(1) %gep1
+  %gep2 = getelementptr <4 x i8>, ptr addrspace(1) %src2, i32 %idx
+  %vec2 = load <4 x i8>, ptr addrspace(1) %gep2
+
+  %v1e0 = extractelement <4 x i8> %vec1, i64 0
+  %cv1e0 = sext i8 %v1e0 to i32
+  %v2e0 = extractelement <4 x i8> %vec2, i64 0
+  %cv2e0 = sext i8 %v2e0 to i32
+  %mul1 = mul nuw nsw i32 %cv1e0, %cv2e0
+
+  %v1e1 = extractelement <4 x i8> %vec1, i64 1
+  %cv1e1 = sext i8 %v1e1 to i32
+  %v2e1 = extractelement <4 x i8> %vec2, i64 1
+  %cv2e1 = sext i8 %v2e1 to i32
+  %mul2 = mul nuw nsw i32 %cv1e1, %cv2e1
+
+  %v1e2 = extractelement <4 x i8> %vec1, i64 2
+  %cv1e2 = sext i8 %v1e2 to i32
+  %v2e2 = extractelement <4 x i8> %vec2, i64 2
+  %cv2e2 = sext i8 %v2e2 to i32
+  %mul3 = mul nuw nsw i32 %cv1e2, %cv2e2
+
+  %v1e3 = extractelement <4 x i8> %vec1, i64 3
+  %cv1e3 = sext i8 %v1e3 to i32
+  %v2e3 = extractelement <4 x i8> %vec2, i64 3
+  %cv2e3 = sext i8 %v2e3 to i32
+  %mul4 = mul nuw nsw i32 %cv1e3, %cv2e3
+
+  %add2 = add i32 %mul1, %mul2
+  %add3 = add i32 %add2, %mul3
+  %add4 = add i32 %add3, %mul4
+  store i32 %add4, ptr addrspace(1) %dst, align 4
+  ret void
+}
+
+define amdgpu_kernel void @idot4_acc32_3src(ptr addrspace(1) %src1,
+; GFX7-LABEL: idot4_acc32_3src:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx8 s[0:7], s[0:1], 0x9
+; GFX7-NEXT:    s_mov_b32 s11, 0xf000
+; GFX7-NEXT:    s_mov_b32 s14, 0
+; GFX7-NEXT:    s_mov_b32 s15, s11
+; GFX7-NEXT:    v_lshlrev_b32_e32 v0, 2, v0
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    s_mov_b64 s[12:13], s[0:1]
+; GFX7-NEXT:    v_mov_b32_e32 v1, 0
+; GFX7-NEXT:    buffer_load_dword v2, v[0:1], s[12:15], 0 addr64
+; GFX7-NEXT:    s_mov_b64 s[12:13], s[2:3]
+; GFX7-NEXT:    buffer_load_dword v3, v[0:1], s[12:15], 0 addr64
+; GFX7-NEXT:    s_mov_b64 s[12:13], s[4:5]
+; GFX7-NEXT:    buffer_load_dword v0, v[0:1], s[12:15], 0 addr64
+; GFX7-NEXT:    s_load_dword s0, s[6:7], 0x0
+; GFX7-NEXT:    s_mov_b32 s10, -1
+; GFX7-NEXT:    s_mov_b32 s8, s6
+; GFX7-NEXT:    s_mov_b32 s9, s7
+; GFX7-NEXT:    s_waitcnt vmcnt(2)
+; GFX7-NEXT:    v_bfe_i32 v1, v2, 0, 8
+; GFX7-NEXT:    v_bfe_i32 v4, v2, 8, 8
+; GFX7-NEXT:    s_waitcnt vmcnt(1)
+; GFX7-NEXT:    v_bfe_i32 v3, v3, 8, 8
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mad_i32_i24 v1, v1, v1, s0
+; GFX7-NEXT:    v_bfe_i32 v5, v2, 16, 8
+; GFX7-NEXT:    s_waitcnt vmcnt(0)
+; GFX7-NEXT:    v_bfe_i32 v6, v0, 16, 8
+; GFX7-NEXT:    v_mad_i32_i24 v1, v4, v3, v1
+; GFX7-NEXT:    v_ashrrev_i32_e32 v2, 24, v2
+; GFX7-NEXT:    v_ashrrev_i32_e32 v0, 24, v0
+; GFX7-NEXT:    v_mad_i32_i24 v1, v5, v6, v1
+; GFX7-NEXT:    v_mad_i32_i24 v0, v2, v0, v1
+; GFX7-NEXT:    buffer_store_dword v0, off, s[8:11], 0
+; GFX7-NEXT:    s_endpgm
+;
+; GFX8-LABEL: idot4_acc32_3src:
+; GFX8:       ; %bb.0: ; %entry
+; GFX8-NEXT:    s_load_dwordx8 s[0:7], s[0:1], 0x24
+; GFX8-NEXT:    v_lshlrev_b32_e32 v2, 2, v0
+; GFX8-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX8-NEXT:    v_mov_b32_e32 v1, s1
+; GFX8-NEXT:    v_add_u32_e32 v0, vcc, s0, v2
+; GFX8-NEXT:    v_addc_u32_e32 v1, vcc, 0, v1, vcc
+; GFX8-NEXT:    flat_load_dword v3, v[0:1]
+; GFX8-NEXT:    v_mov_b32_e32 v1, s3
+; GFX8-NEXT:    v_add_u32_e32 v0, vcc, s2, v2
+; GFX8-NEXT:    v_addc_u32_e32 v1, vcc, 0, v1, vcc
+; GFX8-NEXT:    flat_load_dword v4, v[0:1]
+; GFX8-NEXT:    v_mov_b32_e32 v1, s5
+; GFX8-NEXT:    v_add_u32_e32 v0, vcc, s4, v2
+; GFX8-NEXT:    v_addc_u32_e32 v1, vcc, 0, v1, vcc
+; GFX8-NEXT:    flat_load_dword v0, v[0:1]
+; GFX8-NEXT:    s_load_dword s0, s[6:7], 0x0
+; GFX8-NEXT:    s_waitcnt vmcnt(2)
+; GFX8-NEXT:    v_bfe_i32 v1, v3, 0, 8
+; GFX8-NEXT:    v_bfe_i32 v2, v3, 8, 8
+; GFX8-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX8-NEXT:    v_mad_i32_i24 v1, v1, v1, s0
+; GFX8-NEXT:    v_bfe_i32 v5, v3, 16, 8
+; GFX8-NEXT:    v_ashrrev_i32_e32 v3, 24, v3
+; GFX8-NEXT:    s_waitcnt vmcnt(1)
+; GFX8-NEXT:    v_bfe_i32 v4, v4, 8, 8
+; GFX8-NEXT:    v_mad_i32_i24 v1, v2, v4, v1
+; GFX8-NEXT:    s_waitcnt vmcnt(0)
+; GFX8-NEXT:    v_bfe_i32 v6, v0, 16, 8
+; GFX8-NEXT:    v_ashrrev_i32_e32 v0, 24, v0
+; GFX8-NEXT:    v_mad_i32_i24 v1, v5, v6, v1
+; GFX8-NEXT:    v_mad_i32_i24 v2, v3, v0, v1
+; GFX8-NEXT:    v_mov_b32_e32 v0, s6
+; GFX8-NEXT:    v_mov_b32_e32 v1, s7
+; GFX8-NEXT:    flat_store_dword v[0:1], v2
+; GFX8-NEXT:    s_endpgm
+;
+; GFX9-NODL-LABEL: idot4_acc32_3src:
+; GFX9-NODL:       ; %bb.0: ; %entry
+; GFX9-NODL-NEXT:    s_load_dwordx8 s[0:7], s[0:1], 0x24
+; GFX9-NODL-NEXT:    v_lshlrev_b32_e32 v0, 2, v0
+; GFX9-NODL-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX9-NODL-NEXT:    global_load_dword v1, v0, s[0:1]
+; GFX9-NODL-NEXT:    global_load_dword v2, v0, s[2:3]
+; GFX9-NODL-NEXT:    global_load_dword v3, v0, s[4:5]
+; GFX9-NODL-NEXT:    s_load_dword s0, s[6:7], 0x0
+; GFX9-NODL-NEXT:    v_mov_b32_e32 v0, 0
+; GFX9-NODL-NEXT:    s_waitcnt vmcnt(2)
+; GFX9-NODL-NEXT:    v_mul_i32_i24_sdwa v4, sext(v1), sext(v1) dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:BYTE_0
+; GFX9-NODL-NEXT:    s_waitcnt vmcnt(1)
+; GFX9-NODL-NEXT:    v_mul_i32_i24_sdwa v2, sext(v1), sext(v2) dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_1 src1_sel:BYTE_1
+; GFX9-NODL-NEXT:    s_waitcnt vmcnt(0)
+; GFX9-NODL-NEXT:    v_mul_i32_i24_sdwa v5, sext(v1), sext(v3) dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_2 src1_sel:BYTE_2
+; GFX9-NODL-NEXT:    v_mul_i32_i24_sdwa v1, sext(v1), sext(v3) dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_3 src1_sel:BYTE_3
+; GFX9-NODL-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX9-NODL-NEXT:    v_add3_u32 v2, v4, s0, v2
+; GFX9-NODL-NEXT:    v_add3_u32 v1, v2, v5, v1
+; GFX9-NODL-NEXT:    global_store_dword v0, v1, s[6:7]
+; GFX9-NODL-NEXT:    s_endpgm
+;
+; GFX9-DL-LABEL: idot4_acc32_3src:
+; GFX9-DL:       ; %bb.0: ; %entry
+; GFX9-DL-NEXT:    s_load_dwordx8 s[0:7], s[0:1], 0x24
+; GFX9-DL-NEXT:    v_lshlrev_b32_e32 v0, 2, v0
+; GFX9-DL-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX9-DL-NEXT:    global_load_dword v1, v0, s[2:3]
+; GFX9-DL-NEXT:    global_load_dword v2, v0, s[4:5]
+; GFX9-DL-NEXT:    global_load_dword v3, v0, s[0:1]
+; GFX9-DL-NEXT:    s_load_dword s1, s[6:7], 0x0
+; GFX9-DL-NEXT:    s_mov_b32 s0, 0x706010c
+; GFX9-DL-NEXT:    s_mov_b32 s2, 0xc0c0c00
+; GFX9-DL-NEXT:    v_mov_b32_e32 v0, 0
+; GFX9-DL-NEXT:    s_waitcnt vmcnt(1)
+; GFX9-DL-NEXT:    v_perm_b32 v1, v2, v1, s0
+; GFX9-DL-NEXT:    s_waitcnt vmcnt(0)
+; GFX9-DL-NEXT:    v_perm_b32 v2, v3, v3, s2
+; GFX9-DL-NEXT:    v_or_b32_e32 v1, v1, v2
+; GFX9-DL-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX9-DL-NEXT:    v_dot4_i32_i8 v1, v3, v1, s1
+; GFX9-DL-NEXT:    global_store_dword v0, v1, s[6:7]
+; GFX9-DL-NEXT:    s_endpgm
+;
+; GFX10-DL-LABEL: idot4_acc32_3src:
+; GFX10-DL:       ; %bb.0: ; %entry
+; GFX10-DL-NEXT:    s_load_dwordx8 s[0:7], s[0:1], 0x24
+; GFX10-DL-NEXT:    v_lshlrev_b32_e32 v0, 2, v0
+; GFX10-DL-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-DL-NEXT:    s_clause 0x2
+; GFX10-DL-NEXT:    global_load_dword v1, v0, s[2:3]
+; GFX10-DL-NEXT:    global_load_dword v2, v0, s[4:5]
+; GFX10-DL-NEXT:    global_load_dword v3, v0, s[0:1]
+; GFX10-DL-NEXT:    s_waitcnt_depctr 0xffe3
+; GFX10-DL-NEXT:    s_load_dword s0, s[6:7], 0x0
+; GFX10-DL-NEXT:    s_waitcnt vmcnt(1)
+; GFX10-DL-NEXT:    v_perm_b32 v0, v2, v1, 0x706010c
+; GFX10-DL-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-DL-NEXT:    v_perm_b32 v1, v3, v3, 0xc0c0c00
+; GFX10-DL-NEXT:    v_mov_b32_e32 v2, 0
+; GFX10-DL-NEXT:    v_or_b32_e32 v0, v0, v1
+; GFX10-DL-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-DL-NEXT:    v_mov_b32_e32 v1, s0
+; GFX10-DL-NEXT:    v_dot4c_i32_i8_e32 v1, v3, v0
+; GFX10-DL-NEXT:    global_store_dword v2, v1, s[6:7]
+; GFX10-DL-NEXT:    s_endpgm
+;
+; GFX11-DL-LABEL: idot4_acc32_3src:
+; GFX11-DL:       ; %bb.0: ; %entry
+; GFX11-DL-NEXT:    s_load_b256 s[0:7], s[0:1], 0x24
+; GFX11-DL-NEXT:    v_lshlrev_b32_e32 v0, 2, v0
+; GFX11-DL-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX11-DL-NEXT:    s_clause 0x2
+; GFX11-DL-NEXT:    global_load_b32 v1, v0, s[2:3]
+; GFX11-DL-NEXT:    global_load_b32 v2, v0, s[4:5]
+; GFX11-DL-NEXT:    global_load_b32 v0, v0, s[0:1]
+; GFX11-DL-NEXT:    s_load_b32 s0, s[6:7], 0x0
+; GFX11-DL-NEXT:    s_waitcnt vmcnt(1)
+; GFX11-DL-NEXT:    v_perm_b32 v1, v2, v1, 0x706010c
+; GFX11-DL-NEXT:    s_waitcnt vmcnt(0)
+; GFX11-DL-NEXT:    v_perm_b32 v2, v0, v0, 0xc0c0c00
+; GFX11-DL-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_2)
+; GFX11-DL-NEXT:    v_or_b32_e32 v1, v1, v2
+; GFX11-DL-NEXT:    v_mov_b32_e32 v2, 0
+; GFX11-DL-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX11-DL-NEXT:    v_dot4_i32_iu8 v0, v0, v1, s0 neg_lo:[1,1,0]
+; GFX11-DL-NEXT:    global_store_b32 v2, v0, s[6:7]
+; GFX11-DL-NEXT:    s_nop 0
+; GFX11-DL-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
+; GFX11-DL-NEXT:    s_endpgm
+                                       ptr addrspace(1) %src2,
+                                       ptr addrspace(1) %src3,
+                                       ptr addrspace(1) nocapture %dst) {
+entry:
+  %idx = call i32 @llvm.amdgcn.workitem.id.x()
+  %gep1 = getelementptr <4 x i8>, ptr addrspace(1) %src1, i32 %idx
+  %vec1 = load <4 x i8>, ptr addrspace(1) %gep1
+  %gep2 = getelementptr <4 x i8>, ptr addrspace(1) %src2, i32 %idx
+  %vec2 = load <4 x i8>, ptr addrspace(1) %gep2
+  %gep3 = getelementptr <4 x i8>, ptr addrspace(1) %src3, i32 %idx
+  %vec3 = load <4 x i8>, ptr addrspace(1) %gep3
+
+  %v1e0 = extractelement <4 x i8> %vec1, i64 0
+  %cv1e0 = sext i8 %v1e0 to i32
+  %mul1 = mul nuw nsw i32 %cv1e0, %cv1e0
+
+  %v1e1 = extractelement <4 x i8> %vec1, i64 1
+  %cv1e1 = sext i8 %v1e1 to i32
+  %v2e1 = extractelement <4 x i8> %vec2, i64 1
+  %cv2e1 = sext i8 %v2e1 to i32
+  %mul2 = mul nuw nsw i32 %cv1e1, %cv2e1
+
+  %v1e2 = extractelement <4 x i8> %vec1, i64 2
+  %cv1e2 = sext i8 %v1e2 to i32
+  %v3e2 = extractelement <4 x i8> %vec3, i64 2
+  %cv3e2 = sext i8 %v3e2 to i32
+  %mul3 = mul nuw nsw i32 %cv1e2, %cv3e2
+
+  %v1e3 = extractelement <4 x i8> %vec1, i64 3
+  %cv1e3 = sext i8 %v1e3 to i32
+  %v3e3 = extractelement <4 x i8> %vec3, i64 3
+  %cv3e3 = sext i8 %v3e3 to i32
+  %mul4 = mul nuw nsw i32 %cv1e3, %cv3e3
+
+  %acc = load i32, ptr addrspace(1) %dst, align 4
+  %add1 = add i32 %mul1, %acc
+  %add2 = add i32 %add1, %mul2
+  %add3 = add i32 %add2, %mul3
+  %add4 = add i32 %add3, %mul4
+  store i32 %add4, ptr addrspace(1) %dst, align 4
+  ret void
+}
+
+define amdgpu_kernel void @idot4_acc32_3src_3ele(ptr addrspace(1) %src1,
+; GFX7-LABEL: idot4_acc32_3src_3ele:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx8 s[0:7], s[0:1], 0x9
+; GFX7-NEXT:    s_mov_b32 s11, 0xf000
+; GFX7-NEXT:    s_mov_b32 s14, 0
+; GFX7-NEXT:    s_mov_b32 s15, s11
+; GFX7-NEXT:    v_lshlrev_b32_e32 v0, 2, v0
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    s_mov_b64 s[12:13], s[0:1]
+; GFX7-NEXT:    v_mov_b32_e32 v1, 0
+; GFX7-NEXT:    buffer_load_dword v2, v[0:1], s[12:15], 0 addr64
+; GFX7-NEXT:    s_mov_b64 s[12:13], s[2:3]
+; GFX7-NEXT:    buffer_load_dword v3, v[0:1], s[12:15], 0 addr64
+; GFX7-NEXT:    s_mov_b64 s[12:13], s[4:5]
+; GFX7-NEXT:    buffer_load_dword v0, v[0:1], s[12:15], 0 addr64
+; GFX7-NEXT:    s_load_dword s0, s[6:7], 0x0
+; GFX7-NEXT:    s_mov_b32 s10, -1
+; GFX7-NEXT:    s_mov_b32 s8, s6
+; GFX7-NEXT:    s_mov_b32 s9, s7
+; GFX7-NEXT:    s_waitcnt vmcnt(2)
+; GFX7-NEXT:    v_bfe_i32 v1, v2, 0, 8
+; GFX7-NEXT:    v_bfe_i32 v4, v2, 8, 8
+; GFX7-NEXT:    s_waitcnt vmcnt(1)
+; GFX7-NEXT:    v_bfe_i32 v3, v3, 8, 8
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mad_i32_i24 v1, v1, v1, s0
+; GFX7-NEXT:    v_bfe_i32 v2, v2, 16, 8
+; GFX7-NEXT:    s_waitcnt vmcnt(0)
+; GFX7-NEXT:    v_bfe_i32 v0, v0, 16, 8
+; GFX7-NEXT:    v_mad_i32_i24 v1, v4, v3, v1
+; GFX7-NEXT:    v_mad_i32_i24 v0, v2, v0, v1
+; GFX7-NEXT:    buffer_store_dword v0, off, s[8:11], 0
+; GFX7-NEXT:    s_endpgm
+;
+; GFX8-LABEL: idot4_acc32_3src_3ele:
+; GFX8:       ; %bb.0: ; %entry
+; GFX8-NEXT:    s_load_dwordx8 s[0:7], s[0:1], 0x24
+; GFX8-NEXT:    v_lshlrev_b32_e32 v2, 2, v0
+; GFX8-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX8-NEXT:    v_mov_b32_e32 v1, s1
+; GFX8-NEXT:    v_add_u32_e32 v0, vcc, s0, v2
+; GFX8-NEXT:    v_addc_u32_e32 v1, vcc, 0, v1, vcc
+; GFX8-NEXT:    flat_load_dword v3, v[0:1]
+; GFX8-NEXT:    v_mov_b32_e32 v1, s3
+; GFX8-NEXT:    v_add_u32_e32 v0, vcc, s2, v2
+; GFX8-NEXT:    v_addc_u32_e32 v1, vcc, 0, v1, vcc
+; GFX8-NEXT:    flat_load_dword v4, v[0:1]
+; GFX8-NEXT:    v_mov_b32_e32 v1, s5
+; GFX8-NEXT:    v_add_u32_e32 v0, vcc, s4, v2
+; GFX8-NEXT:    v_addc_u32_e32 v1, vcc, 0, v1, vcc
+; GFX8-NEXT:    flat_load_dword v0, v[0:1]
+; GFX8-NEXT:    s_load_dword s0, s[6:7], 0x0
+; GFX8-NEXT:    s_waitcnt vmcnt(2)
+; GFX8-NEXT:    v_bfe_i32 v1, v3, 0, 8
+; GFX8-NEXT:    v_bfe_i32 v2, v3, 8, 8
+; GFX8-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX8-NEXT:    v_mad_i32_i24 v1, v1, v1, s0
+; GFX8-NEXT:    v_bfe_i32 v3, v3, 16, 8
+; GFX8-NEXT:    s_waitcnt vmcnt(1)
+; GFX8-NEXT:    v_bfe_i32 v4, v4, 8, 8
+; GFX8-NEXT:    v_mad_i32_i24 v1, v2, v4, v1
+; GFX8-NEXT:    s_waitcnt vmcnt(0)
+; GFX8-NEXT:    v_bfe_i32 v0, v0, 16, 8
+; GFX8-NEXT:    v_mad_i32_i24 v2, v3, v0, v1
+; GFX8-NEXT:    v_mov_b32_e32 v0, s6
+; GFX8-NEXT:    v_mov_b32_e32 v1, s7
+; GFX8-NEXT:    flat_store_dword v[0:1], v2
+; GFX8-NEXT:    s_endpgm
+;
+; GFX9-NODL-LABEL: idot4_acc32_3src_3ele:
+; GFX9-NODL:       ; %bb.0: ; %entry
+; GFX9-NODL-NEXT:    s_load_dwordx8 s[0:7], s[0:1], 0x24
+; GFX9-NODL-NEXT:    v_lshlrev_b32_e32 v0, 2, v0
+; GFX9-NODL-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX9-NODL-NEXT:    global_load_dword v1, v0, s[0:1]
+; GFX9-NODL-NEXT:    global_load_dword v2, v0, s[2:3]
+; GFX9-NODL-NEXT:    global_load_dword v3, v0, s[4:5]
+; GFX9-NODL-NEXT:    s_load_dword s0, s[6:7], 0x0
+; GFX9-NODL-NEXT:    v_mov_b32_e32 v0, 0
+; GFX9-NODL-NEXT:    s_waitcnt vmcnt(2)
+; GFX9-NODL-NEXT:    v_bfe_i32 v4, v1, 0, 8
+; GFX9-NODL-NEXT:    s_waitcnt vmcnt(1)
+; GFX9-NODL-NEXT:    v_mul_i32_i24_sdwa v2, sext(v1), sext(v2) dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_1 src1_sel:BYTE_1
+; GFX9-NODL-NEXT:    s_waitcnt vmcnt(0)
+; GFX9-NODL-NEXT:    v_mul_i32_i24_sdwa v1, sext(v1), sext(v3) dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_2 src1_sel:BYTE_2
+; GFX9-NODL-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX9-NODL-NEXT:    v_mad_i32_i24 v3, v4, v4, s0
+; GFX9-NODL-NEXT:    v_add3_u32 v1, v3, v2, v1
+; GFX9-NODL-NEXT:    global_store_dword v0, v1, s[6:7]
+; GFX9-NODL-NEXT:    s_endpgm
+;
+; GFX9-DL-LABEL: idot4_acc32_3src_3ele:
+; GFX9-DL:       ; %bb.0: ; %entry
+; GFX9-DL-NEXT:    s_load_dwordx8 s[0:7], s[0:1], 0x24
+; GFX9-DL-NEXT:    v_lshlrev_b32_e32 v0, 2, v0
+; GFX9-DL-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX9-DL-NEXT:    global_load_dword v1, v0, s[2:3]
+; GFX9-DL-NEXT:    global_load_dword v2, v0, s[4:5]
+; GFX9-DL-NEXT:    global_load_dword v3, v0, s[0:1]
+; GFX9-DL-NEXT:    s_load_dword s3, s[6:7], 0x0
+; GFX9-DL-NEXT:    s_mov_b32 s0, 0xc06010c
+; GFX9-DL-NEXT:    s_mov_b32 s1, 0xc0c0c00
+; GFX9-DL-NEXT:    s_mov_b32 s2, 0xc020100
+; GFX9-DL-NEXT:    v_mov_b32_e32 v0, 0
+; GFX9-DL-NEXT:    s_waitcnt vmcnt(1)
+; GFX9-DL-NEXT:    v_perm_b32 v1, v2, v1, s0
+; GFX9-DL-NEXT:    s_waitcnt vmcnt(0)
+; GFX9-DL-NEXT:    v_perm_b32 v2, v3, v3, s1
+; GFX9-DL-NEXT:    v_or_b32_e32 v1, v1, v2
+; GFX9-DL-NEXT:    v_perm_b32 v2, v3, v3, s2
+; GFX9-DL-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX9-DL-NEXT:    v_dot4_i32_i8 v1, v2, v1, s3
+; GFX9-DL-NEXT:    global_store_dword v0, v1, s[6:7]
+; GFX9-DL-NEXT:    s_endpgm
+;
+; GFX10-DL-LABEL: idot4_acc32_3src_3ele:
+; GFX10-DL:       ; %bb.0: ; %entry
+; GFX10-DL-NEXT:    s_load_dwordx8 s[0:7], s[0:1], 0x24
+; GFX10-DL-NEXT:    v_lshlrev_b32_e32 v0, 2, v0
+; GFX10-DL-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-DL-NEXT:    s_clause 0x2
+; GFX10-DL-NEXT:    global_load_dword v1, v0, s[2:3]
+; GFX10-DL-NEXT:    global_load_dword v2, v0, s[4:5]
+; GFX10-DL-NEXT:    global_load_dword v3, v0, s[0:1]
+; GFX10-DL-NEXT:    s_waitcnt_depctr 0xffe3
+; GFX10-DL-NEXT:    s_load_dword s0, s[6:7], 0x0
+; GFX10-DL-NEXT:    s_waitcnt vmcnt(1)
+; GFX10-DL-NEXT:    v_perm_b32 v0, v2, v1, 0xc06010c
+; GFX10-DL-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-DL-NEXT:    v_perm_b32 v1, v3, v3, 0xc0c0c00
+; GFX10-DL-NEXT:    v_perm_b32 v2, v3, v3, 0xc020100
+; GFX10-DL-NEXT:    v_mov_b32_e32 v3, 0
+; GFX10-DL-NEXT:    v_or_b32_e32 v0, v0, v1
+; GFX10-DL-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-DL-NEXT:    v_mov_b32_e32 v1, s0
+; GFX10-DL-NEXT:    v_dot4c_i32_i8_e32 v1, v2, v0
+; GFX10-DL-NEXT:    global_store_dword v3, v1, s[6:7]
+; GFX10-DL-NEXT:    s_endpgm
+;
+; GFX11-DL-LABEL: idot4_acc32_3src_3ele:
+; GFX11-DL:       ; %bb.0: ; %entry
+; GFX11-DL-NEXT:    s_load_b256 s[0:7], s[0:1], 0x24
+; GFX11-DL-NEXT:    v_lshlrev_b32_e32 v0, 2, v0
+; GFX11-DL-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX11-DL-NEXT:    s_clause 0x2
+; GFX11-DL-NEXT:    global_load_b32 v1, v0, s[2:3]
+; GFX11-DL-NEXT:    global_load_b32 v2, v0, s[4:5]
+; GFX11-DL-NEXT:    global_load_b32 v0, v0, s[0:1]
+; GFX11-DL-NEXT:    s_load_b32 s0, s[6:7], 0x0
+; GFX11-DL-NEXT:    s_waitcnt vmcnt(1)
+; GFX11-DL-NEXT:    v_perm_b32 v1, v2, v1, 0xc06010c
+; GFX11-DL-NEXT:    s_waitcnt vmcnt(0)
+; GFX11-DL-NEXT:    v_perm_b32 v2, v0, v0, 0xc0c0c00
+; GFX11-DL-NEXT:    v_perm_b32 v0, v0, v0, 0xc020100
+; GFX11-DL-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_2) | instid1(VALU_DEP_2)
+; GFX11-DL-NEXT:    v_or_b32_e32 v1, v1, v2
+; GFX11-DL-NEXT:    v_mov_b32_e32 v2, 0
+; GFX11-DL-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX11-DL-NEXT:    v_dot4_i32_iu8 v0, v0, v1, s0 neg_lo:[1,1,0]
+; GFX11-DL-NEXT:    global_store_b32 v2, v0, s[6:7]
+; GFX11-DL-NEXT:    s_nop 0
+; GFX11-DL-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
+; GFX11-DL-NEXT:    s_endpgm
+                                       ptr addrspace(1) %src2,
+                                       ptr addrspace(1) %src3,
+                                       ptr addrspace(1) nocapture %dst) {
+entry:
+  %idx = call i32 @llvm.amdgcn.workitem.id.x()
+  %gep1 = getelementptr <4 x i8>, ptr addrspace(1) %src1, i32 %idx
+  %vec1 = load <4 x i8>, ptr addrspace(1) %gep1
+  %gep2 = getelementptr <4 x i8>, ptr addrspace(1) %src2, i32 %idx
+  %vec2 = load <4 x i8>, ptr addrspace(1) %gep2
+  %gep3 = getelementptr <4 x i8>, ptr addrspace(1) %src3, i32 %idx
+  %vec3 = load <4 x i8>, ptr addrspace(1) %gep3
+
+  %v1e0 = extractelement <4 x i8> %vec1, i64 0
+  %cv1e0 = sext i8 %v1e0 to i32
+  %mul1 = mul nuw nsw i32 %cv1e0, %cv1e0
+
+  %v1e1 = extractelement <4 x i8> %vec1, i64 1
+  %cv1e1 = sext i8 %v1e1 to i32
+  %v2e1 = extractelement <4 x i8> %vec2, i64 1
+  %cv2e1 = sext i8 %v2e1 to i32
+  %mul2 = mul nuw nsw i32 %cv1e1, %cv2e1
+
+  %v1e2 = extractelement <4 x i8> %vec1, i64 2
+  %cv1e2 = sext i8 %v1e2 to i32
+  %v3e2 = extractelement <4 x i8> %vec3, i64 2
+  %cv3e2 = sext i8 %v3e2 to i32
+  %mul3 = mul nuw nsw i32 %cv1e2, %cv3e2
+
+  %acc = load i32, ptr addrspace(1) %dst, align 4
+  %add1 = add i32 %mul1, %acc
+  %add2 = add i32 %add1, %mul2
+  %add3 = add i32 %add2, %mul3
+  store i32 %add3, ptr addrspace(1) %dst, align 4
+  ret void
+}
+
+define amdgpu_kernel void @idot4_bad_source(ptr addrspace(1) %src1,
+; GFX7-LABEL: idot4_bad_source:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x9
+; GFX7-NEXT:    s_load_dword s12, s[0:1], 0xf
+; GFX7-NEXT:    s_mov_b32 s3, 0xf000
+; GFX7-NEXT:    s_mov_b32 s10, 0
+; GFX7-NEXT:    s_mov_b32 s11, s3
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    s_mov_b64 s[8:9], s[4:5]
+; GFX7-NEXT:    v_lshlrev_b32_e32 v0, 2, v0
+; GFX7-NEXT:    v_mov_b32_e32 v1, 0
+; GFX7-NEXT:    buffer_load_dword v2, v[0:1], s[8:11], 0 addr64
+; GFX7-NEXT:    s_mov_b64 s[8:9], s[6:7]
+; GFX7-NEXT:    buffer_load_dword v0, v[0:1], s[8:11], 0 addr64
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0x11
+; GFX7-NEXT:    s_sext_i32_i16 s5, s12
+; GFX7-NEXT:    s_mov_b32 s2, -1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    s_load_dword s4, s[0:1], 0x0
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v1, s4
+; GFX7-NEXT:    s_waitcnt vmcnt(1)
+; GFX7-NEXT:    v_bfe_i32 v3, v2, 0, 8
+; GFX7-NEXT:    v_bfe_i32 v4, v2, 8, 8
+; GFX7-NEXT:    s_waitcnt vmcnt(0)
+; GFX7-NEXT:    v_bfe_i32 v5, v0, 8, 8
+; GFX7-NEXT:    v_mad_i32_i24 v1, v3, s5, v1
+; GFX7-NEXT:    v_bfe_i32 v2, v2, 16, 8
+; GFX7-NEXT:    v_bfe_i32 v0, v0, 16, 8
+; GFX7-NEXT:    v_mad_i32_i24 v1, v4, v5, v1
+; GFX7-NEXT:    v_mad_i32_i24 v0, v2, v0, v1
+; GFX7-NEXT:    buffer_store_dword v0, off, s[0:3], 0
+; GFX7-NEXT:    s_endpgm
+;
+; GFX8-LABEL: idot4_bad_source:
+; GFX8:       ; %bb.0: ; %entry
+; GFX8-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x24
+; GFX8-NEXT:    s_load_dword s2, s[0:1], 0x3c
+; GFX8-NEXT:    v_lshlrev_b32_e32 v2, 2, v0
+; GFX8-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0x44
+; GFX8-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX8-NEXT:    v_mov_b32_e32 v1, s5
+; GFX8-NEXT:    v_add_u32_e32 v0, vcc, s4, v2
+; GFX8-NEXT:    v_addc_u32_e32 v1, vcc, 0, v1, vcc
+; GFX8-NEXT:    flat_load_dword v3, v[0:1]
+; GFX8-NEXT:    v_mov_b32_e32 v1, s7
+; GFX8-NEXT:    v_add_u32_e32 v0, vcc, s6, v2
+; GFX8-NEXT:    v_addc_u32_e32 v1, vcc, 0, v1, vcc
+; GFX8-NEXT:    flat_load_dword v0, v[0:1]
+; GFX8-NEXT:    s_load_dword s3, s[0:1], 0x0
+; GFX8-NEXT:    s_sext_i32_i16 s2, s2
+; GFX8-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX8-NEXT:    v_mov_b32_e32 v1, s3
+; GFX8-NEXT:    s_waitcnt vmcnt(1)
+; GFX8-NEXT:    v_bfe_i32 v2, v3, 0, 8
+; GFX8-NEXT:    v_bfe_i32 v4, v3, 8, 8
+; GFX8-NEXT:    v_mad_i32_i24 v1, v2, s2, v1
+; GFX8-NEXT:    v_bfe_i32 v3, v3, 16, 8
+; GFX8-NEXT:    s_waitcnt vmcnt(0)
+; GFX8-NEXT:    v_bfe_i32 v5, v0, 8, 8
+; GFX8-NEXT:    v_bfe_i32 v0, v0, 16, 8
+; GFX8-NEXT:    v_mad_i32_i24 v1, v4, v5, v1
+; GFX8-NEXT:    v_mad_i32_i24 v2, v3, v0, v1
+; GFX8-NEXT:    v_mov_b32_e32 v0, s0
+; GFX8-NEXT:    v_mov_b32_e32 v1, s1
+; GFX8-NEXT:    flat_store_dword v[0:1], v2
+; GFX8-NEXT:    s_endpgm
+;
+; GFX9-NODL-LABEL: idot4_bad_source:
+; GFX9-NODL:       ; %bb.0: ; %entry
+; GFX9-NODL-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x24
+; GFX9-NODL-NEXT:    s_load_dword s2, s[0:1], 0x3c
+; GFX9-NODL-NEXT:    v_lshlrev_b32_e32 v0, 2, v0
+; GFX9-NODL-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0x44
+; GFX9-NODL-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX9-NODL-NEXT:    global_load_dword v1, v0, s[4:5]
+; GFX9-NODL-NEXT:    global_load_dword v2, v0, s[6:7]
+; GFX9-NODL-NEXT:    s_sext_i32_i16 s2, s2
+; GFX9-NODL-NEXT:    v_mov_b32_e32 v0, 0
+; GFX9-NODL-NEXT:    s_load_dword s3, s[0:1], 0x0
+; GFX9-NODL-NEXT:    s_waitcnt vmcnt(1)
+; GFX9-NODL-NEXT:    v_bfe_i32 v3, v1, 0, 8
+; GFX9-NODL-NEXT:    s_waitcnt vmcnt(0)
+; GFX9-NODL-NEXT:    v_mul_i32_i24_sdwa v4, sext(v1), sext(v2) dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_1 src1_sel:BYTE_1
+; GFX9-NODL-NEXT:    v_mul_i32_i24_sdwa v1, sext(v1), sext(v2) dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_2 src1_sel:BYTE_2
+; GFX9-NODL-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX9-NODL-NEXT:    v_mov_b32_e32 v2, s3
+; GFX9-NODL-NEXT:    v_mad_i32_i24 v2, v3, s2, v2
+; GFX9-NODL-NEXT:    v_add3_u32 v1, v2, v4, v1
+; GFX9-NODL-NEXT:    global_store_dword v0, v1, s[0:1]
+; GFX9-NODL-NEXT:    s_endpgm
+;
+; GFX9-DL-LABEL: idot4_bad_source:
+; GFX9-DL:       ; %bb.0: ; %entry
+; GFX9-DL-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x24
+; GFX9-DL-NEXT:    s_load_dword s2, s[0:1], 0x3c
+; GFX9-DL-NEXT:    v_lshlrev_b32_e32 v0, 2, v0
+; GFX9-DL-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0x44
+; GFX9-DL-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX9-DL-NEXT:    global_load_dword v1, v0, s[4:5]
+; GFX9-DL-NEXT:    global_load_dword v2, v0, s[6:7]
+; GFX9-DL-NEXT:    s_mov_b32 s4, 0xc0c0201
+; GFX9-DL-NEXT:    s_sext_i32_i16 s2, s2
+; GFX9-DL-NEXT:    s_load_dword s3, s[0:1], 0x0
+; GFX9-DL-NEXT:    v_mov_b32_e32 v0, 0
+; GFX9-DL-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX9-DL-NEXT:    v_mov_b32_e32 v3, s3
+; GFX9-DL-NEXT:    s_waitcnt vmcnt(1)
+; GFX9-DL-NEXT:    v_bfe_i32 v4, v1, 0, 8
+; GFX9-DL-NEXT:    s_waitcnt vmcnt(0)
+; GFX9-DL-NEXT:    v_perm_b32 v2, v2, v2, s4
+; GFX9-DL-NEXT:    v_mad_i32_i24 v3, v4, s2, v3
+; GFX9-DL-NEXT:    v_perm_b32 v1, v1, v1, s4
+; GFX9-DL-NEXT:    v_dot4_i32_i8 v1, v1, v2, v3
+; GFX9-DL-NEXT:    global_store_dword v0, v1, s[0:1]
+; GFX9-DL-NEXT:    s_endpgm
+;
+; GFX10-DL-LABEL: idot4_bad_source:
+; GFX10-DL:       ; %bb.0: ; %entry
+; GFX10-DL-NEXT:    s_clause 0x1
+; GFX10-DL-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x24
+; GFX10-DL-NEXT:    s_load_dword s2, s[0:1], 0x3c
+; GFX10-DL-NEXT:    v_lshlrev_b32_e32 v0, 2, v0
+; GFX10-DL-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0x44
+; GFX10-DL-NEXT:    v_mov_b32_e32 v3, 0
+; GFX10-DL-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-DL-NEXT:    s_clause 0x1
+; GFX10-DL-NEXT:    global_load_dword v1, v0, s[4:5]
+; GFX10-DL-NEXT:    global_load_dword v2, v0, s[6:7]
+; GFX10-DL-NEXT:    s_sext_i32_i16 s2, s2
+; GFX10-DL-NEXT:    s_load_dword s3, s[0:1], 0x0
+; GFX10-DL-NEXT:    s_waitcnt vmcnt(1)
+; GFX10-DL-NEXT:    v_bfe_i32 v0, v1, 0, 8
+; GFX10-DL-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-DL-NEXT:    v_perm_b32 v2, v2, v2, 0xc0c0201
+; GFX10-DL-NEXT:    v_perm_b32 v1, v1, v1, 0xc0c0201
+; GFX10-DL-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-DL-NEXT:    v_mad_i32_i24 v0, v0, s2, s3
+; GFX10-DL-NEXT:    v_dot4c_i32_i8_e32 v0, v1, v2
+; GFX10-DL-NEXT:    global_store_dword v3, v0, s[0:1]
+; GFX10-DL-NEXT:    s_endpgm
+;
+; GFX11-DL-LABEL: idot4_bad_source:
+; GFX11-DL:       ; %bb.0: ; %entry
+; GFX11-DL-NEXT:    s_clause 0x1
+; GFX11-DL-NEXT:    s_load_b128 s[4:7], s[0:1], 0x24
+; GFX11-DL-NEXT:    s_load_b32 s2, s[0:1], 0x3c
+; GFX11-DL-NEXT:    v_dual_mov_b32 v3, 0 :: v_dual_lshlrev_b32 v0, 2, v0
+; GFX11-DL-NEXT:    s_load_b64 s[0:1], s[0:1], 0x44
+; GFX11-DL-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX11-DL-NEXT:    s_clause 0x1
+; GFX11-DL-NEXT:    global_load_b32 v1, v0, s[4:5]
+; GFX11-DL-NEXT:    global_load_b32 v0, v0, s[6:7]
+; GFX11-DL-NEXT:    s_sext_i32_i16 s2, s2
+; GFX11-DL-NEXT:    s_load_b32 s3, s[0:1], 0x0
+; GFX11-DL-NEXT:    s_waitcnt vmcnt(1)
+; GFX11-DL-NEXT:    v_bfe_i32 v2, v1, 0, 8
+; GFX11-DL-NEXT:    s_waitcnt vmcnt(0)
+; GFX11-DL-NEXT:    v_perm_b32 v0, v0, v0, 0xc0c0201
+; GFX11-DL-NEXT:    v_perm_b32 v1, v1, v1, 0xc0c0201
+; GFX11-DL-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX11-DL-NEXT:    v_mad_i32_i24 v2, v2, s2, s3
+; GFX11-DL-NEXT:    s_delay_alu instid0(VALU_DEP_1)
+; GFX11-DL-NEXT:    v_dot4_i32_iu8 v0, v1, v0, v2 neg_lo:[1,1,0]
+; GFX11-DL-NEXT:    global_store_b32 v3, v0, s[0:1]
+; GFX11-DL-NEXT:    s_nop 0
+; GFX11-DL-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
+; GFX11-DL-NEXT:    s_endpgm
+                                       ptr addrspace(1) %src2,
+                                       ptr addrspace(1) %src3,
+                                       i16 %badsource,
+                                       ptr addrspace(1) nocapture %dst) {
+entry:
+  %idx = call i32 @llvm.amdgcn.workitem.id.x()
+  %gep1 = getelementptr <4 x i8>, ptr addrspace(1) %src1, i32 %idx
+  %vec1 = load <4 x i8>, ptr addrspace(1) %gep1
+  %gep2 = getelementptr <4 x i8>, ptr addrspace(1) %src2, i32 %idx
+  %vec2 = load <4 x i8>, ptr addrspace(1) %gep2
+  %gep3 = getelementptr <4 x i8>, ptr addrspace(1) %src3, i32 %idx
+  %vec3 = load <4 x i8>, ptr addrspace(1) %gep3
+
+  %v1e0 = extractelement <4 x i8> %vec1, i64 0
+  %cv1e0 = sext i8 %v1e0 to i32
+  %v2e0 = extractelement <4 x i8> %vec2, i64 0
+  %other = sext i16 %badsource to i32
+  %mul1 = mul nuw nsw i32 %cv1e0, %other
+
+  %v1e1 = extractelement <4 x i8> %vec1, i64 1
+  %cv1e1 = sext i8 %v1e1 to i32
+  %v2e1 = extractelement <4 x i8> %vec2, i64 1
+  %cv2e1 = sext i8 %v2e1 to i32
+  %mul2 = mul nuw nsw i32 %cv1e1, %cv2e1
+
+  %v2e2 = extractelement <4 x i8> %vec2, i64 2
+  %cv2e2 = sext i8 %v2e2 to i32
+  %v1e2 = extractelement <4 x i8> %vec1, i64 2
+  %cv1e2 = sext i8 %v1e2 to i32
+  %mul3 = mul nuw nsw i32 %cv1e2, %cv2e2
+
+
+  %acc = load i32, ptr addrspace(1) %dst, align 4
+  %mad1 = add i32 %mul1, %acc
+  %mad2 = add i32 %mad1, %mul2
+  %mad3 = add i32 %mad2, %mul3
+
+  store i32 %mad3, ptr addrspace(1) %dst, align 4
+  ret void
+}
+
+
+define amdgpu_kernel void @idot4_commutative(ptr addrspace(1) %src1,
+; GFX7-LABEL: idot4_commutative:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x9
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xf
+; GFX7-NEXT:    s_mov_b32 s3, 0xf000
+; GFX7-NEXT:    s_mov_b32 s10, 0
+; GFX7-NEXT:    s_mov_b32 s11, s3
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    s_mov_b64 s[8:9], s[4:5]
+; GFX7-NEXT:    v_lshlrev_b32_e32 v0, 2, v0
+; GFX7-NEXT:    v_mov_b32_e32 v1, 0
+; GFX7-NEXT:    buffer_load_dword v2, v[0:1], s[8:11], 0 addr64
+; GFX7-NEXT:    s_mov_b64 s[8:9], s[6:7]
+; GFX7-NEXT:    buffer_load_dword v0, v[0:1], s[8:11], 0 addr64
+; GFX7-NEXT:    s_load_dword s4, s[0:1], 0x0
+; GFX7-NEXT:    s_mov_b32 s2, -1
+; GFX7-NEXT:    s_waitcnt vmcnt(1)
+; GFX7-NEXT:    v_bfe_i32 v1, v2, 0, 8
+; GFX7-NEXT:    v_bfe_i32 v3, v2, 8, 8
+; GFX7-NEXT:    s_waitcnt vmcnt(0)
+; GFX7-NEXT:    v_bfe_i32 v4, v0, 0, 8
+; GFX7-NEXT:    v_bfe_i32 v5, v0, 8, 8
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mad_i32_i24 v1, v1, v4, s4
+; GFX7-NEXT:    v_bfe_i32 v0, v0, 16, 8
+; GFX7-NEXT:    v_bfe_i32 v2, v2, 16, 8
+; GFX7-NEXT:    v_mad_i32_i24 v1, v3, v5, v1
+; GFX7-NEXT:    v_mad_i32_i24 v0, v2, v0, v1
+; GFX7-NEXT:    buffer_store_dword v0, off, s[0:3], 0
+; GFX7-NEXT:    s_endpgm
+;
+; GFX8-LABEL: idot4_commutative:
+; GFX8:       ; %bb.0: ; %entry
+; GFX8-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x24
+; GFX8-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0x3c
+; GFX8-NEXT:    v_lshlrev_b32_e32 v2, 2, v0
+; GFX8-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX8-NEXT:    v_mov_b32_e32 v1, s5
+; GFX8-NEXT:    v_add_u32_e32 v0, vcc, s4, v2
+; GFX8-NEXT:    v_addc_u32_e32 v1, vcc, 0, v1, vcc
+; GFX8-NEXT:    flat_load_dword v3, v[0:1]
+; GFX8-NEXT:    v_mov_b32_e32 v1, s7
+; GFX8-NEXT:    v_add_u32_e32 v0, vcc, s6, v2
+; GFX8-NEXT:    v_addc_u32_e32 v1, vcc, 0, v1, vcc
+; GFX8-NEXT:    flat_load_dword v0, v[0:1]
+; GFX8-NEXT:    s_load_dword s2, s[0:1], 0x0
+; GFX8-NEXT:    s_waitcnt vmcnt(1)
+; GFX8-NEXT:    v_bfe_i32 v1, v3, 0, 8
+; GFX8-NEXT:    v_bfe_i32 v4, v3, 8, 8
+; GFX8-NEXT:    v_bfe_i32 v3, v3, 16, 8
+; GFX8-NEXT:    s_waitcnt vmcnt(0)
+; GFX8-NEXT:    v_bfe_i32 v2, v0, 0, 8
+; GFX8-NEXT:    v_bfe_i32 v5, v0, 8, 8
+; GFX8-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX8-NEXT:    v_mad_i32_i24 v1, v1, v2, s2
+; GFX8-NEXT:    v_bfe_i32 v0, v0, 16, 8
+; GFX8-NEXT:    v_mad_i32_i24 v1, v4, v5, v1
+; GFX8-NEXT:    v_mad_i32_i24 v2, v3, v0, v1
+; GFX8-NEXT:    v_mov_b32_e32 v0, s0
+; GFX8-NEXT:    v_mov_b32_e32 v1, s1
+; GFX8-NEXT:    flat_store_dword v[0:1], v2
+; GFX8-NEXT:    s_endpgm
+;
+; GFX9-NODL-LABEL: idot4_commutative:
+; GFX9-NODL:       ; %bb.0: ; %entry
+; GFX9-NODL-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x24
+; GFX9-NODL-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x3c
+; GFX9-NODL-NEXT:    v_lshlrev_b32_e32 v0, 2, v0
+; GFX9-NODL-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX9-NODL-NEXT:    global_load_dword v1, v0, s[4:5]
+; GFX9-NODL-NEXT:    global_load_dword v2, v0, s[6:7]
+; GFX9-NODL-NEXT:    s_load_dword s0, s[2:3], 0x0
+; GFX9-NODL-NEXT:    v_mov_b32_e32 v0, 0
+; GFX9-NODL-NEXT:    s_waitcnt vmcnt(1)
+; GFX9-NODL-NEXT:    v_bfe_i32 v3, v1, 0, 8
+; GFX9-NODL-NEXT:    s_waitcnt vmcnt(0)
+; GFX9-NODL-NEXT:    v_bfe_i32 v4, v2, 0, 8
+; GFX9-NODL-NEXT:    v_mul_i32_i24_sdwa v5, sext(v1), sext(v2) dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_1 src1_sel:BYTE_1
+; GFX9-NODL-NEXT:    v_mul_i32_i24_sdwa v1, sext(v1), sext(v2) dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_2 src1_sel:BYTE_2
+; GFX9-NODL-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX9-NODL-NEXT:    v_mad_i32_i24 v2, v3, v4, s0
+; GFX9-NODL-NEXT:    v_add3_u32 v1, v2, v5, v1
+; GFX9-NODL-NEXT:    global_store_dword v0, v1, s[2:3]
+; GFX9-NODL-NEXT:    s_endpgm
+;
+; GFX9-DL-LABEL: idot4_commutative:
+; GFX9-DL:       ; %bb.0: ; %entry
+; GFX9-DL-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x24
+; GFX9-DL-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x3c
+; GFX9-DL-NEXT:    v_lshlrev_b32_e32 v0, 2, v0
+; GFX9-DL-NEXT:    s_mov_b32 s1, 0xc020100
+; GFX9-DL-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX9-DL-NEXT:    global_load_dword v1, v0, s[6:7]
+; GFX9-DL-NEXT:    global_load_dword v2, v0, s[4:5]
+; GFX9-DL-NEXT:    s_load_dword s0, s[2:3], 0x0
+; GFX9-DL-NEXT:    v_mov_b32_e32 v0, 0
+; GFX9-DL-NEXT:    s_waitcnt vmcnt(1)
+; GFX9-DL-NEXT:    v_perm_b32 v1, v1, v1, s1
+; GFX9-DL-NEXT:    s_waitcnt vmcnt(0)
+; GFX9-DL-NEXT:    v_perm_b32 v2, v2, v2, s1
+; GFX9-DL-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX9-DL-NEXT:    v_dot4_i32_i8 v1, v2, v1, s0
+; GFX9-DL-NEXT:    global_store_dword v0, v1, s[2:3]
+; GFX9-DL-NEXT:    s_endpgm
+;
+; GFX10-DL-LABEL: idot4_commutative:
+; GFX10-DL:       ; %bb.0: ; %entry
+; GFX10-DL-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x24
+; GFX10-DL-NEXT:    v_lshlrev_b32_e32 v0, 2, v0
+; GFX10-DL-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0x3c
+; GFX10-DL-NEXT:    v_mov_b32_e32 v3, 0
+; GFX10-DL-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-DL-NEXT:    s_clause 0x1
+; GFX10-DL-NEXT:    global_load_dword v1, v0, s[6:7]
+; GFX10-DL-NEXT:    global_load_dword v2, v0, s[4:5]
+; GFX10-DL-NEXT:    s_load_dword s2, s[0:1], 0x0
+; GFX10-DL-NEXT:    s_waitcnt vmcnt(1)
+; GFX10-DL-NEXT:    v_perm_b32 v0, v1, v1, 0xc020100
+; GFX10-DL-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-DL-NEXT:    v_perm_b32 v1, v2, v2, 0xc020100
+; GFX10-DL-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-DL-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-DL-NEXT:    v_dot4c_i32_i8_e32 v2, v1, v0
+; GFX10-DL-NEXT:    global_store_dword v3, v2, s[0:1]
+; GFX10-DL-NEXT:    s_endpgm
+;
+; GFX11-DL-LABEL: idot4_commutative:
+; GFX11-DL:       ; %bb.0: ; %entry
+; GFX11-DL-NEXT:    s_load_b128 s[4:7], s[0:1], 0x24
+; GFX11-DL-NEXT:    v_lshlrev_b32_e32 v0, 2, v0
+; GFX11-DL-NEXT:    s_load_b64 s[0:1], s[0:1], 0x3c
+; GFX11-DL-NEXT:    v_mov_b32_e32 v2, 0
+; GFX11-DL-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX11-DL-NEXT:    s_clause 0x1
+; GFX11-DL-NEXT:    global_load_b32 v1, v0, s[6:7]
+; GFX11-DL-NEXT:    global_load_b32 v0, v0, s[4:5]
+; GFX11-DL-NEXT:    s_load_b32 s2, s[0:1], 0x0
+; GFX11-DL-NEXT:    s_waitcnt vmcnt(1)
+; GFX11-DL-NEXT:    v_perm_b32 v1, v1, v1, 0xc020100
+; GFX11-DL-NEXT:    s_waitcnt vmcnt(0)
+; GFX11-DL-NEXT:    v_perm_b32 v0, v0, v0, 0xc020100
+; GFX11-DL-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX11-DL-NEXT:    s_delay_alu instid0(VALU_DEP_1)
+; GFX11-DL-NEXT:    v_dot4_i32_iu8 v0, v0, v1, s2 neg_lo:[1,1,0]
+; GFX11-DL-NEXT:    global_store_b32 v2, v0, s[0:1]
+; GFX11-DL-NEXT:    s_nop 0
+; GFX11-DL-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
+; GFX11-DL-NEXT:    s_endpgm
+                                       ptr addrspace(1) %src2,
+                                       ptr addrspace(1) %src3,
+                                       ptr addrspace(1) nocapture %dst) {
+entry:
+  %idx = call i32 @llvm.amdgcn.workitem.id.x()
+  %gep1 = getelementptr <4 x i8>, ptr addrspace(1) %src1, i32 %idx
+  %vec1 = load <4 x i8>, ptr addrspace(1) %gep1
+  %gep2 = getelementptr <4 x i8>, ptr addrspace(1) %src2, i32 %idx
+  %vec2 = load <4 x i8>, ptr addrspace(1) %gep2
+  %gep3 = getelementptr <4 x i8>, ptr addrspace(1) %src3, i32 %idx
+  %vec3 = load <4 x i8>, ptr addrspace(1) %gep3
+
+  %v1e0 = extractelement <4 x i8> %vec1, i64 0
+  %cv1e0 = sext i8 %v1e0 to i32
+  %v2e0 = extractelement <4 x i8> %vec2, i64 0
+  %cv2e0 = sext i8 %v2e0 to i32
+  %mul1 = mul nuw nsw i32 %cv1e0, %cv2e0
+
+  %v1e1 = extractelement <4 x i8> %vec1, i64 1
+  %cv1e1 = sext i8 %v1e1 to i32
+  %v2e1 = extractelement <4 x i8> %vec2, i64 1
+  %cv2e1 = sext i8 %v2e1 to i32
+  %mul2 = mul nuw nsw i32 %cv1e1, %cv2e1
+
+  %v2e2 = extractelement <4 x i8> %vec2, i64 2
+  %cv2e2 = sext i8 %v2e2 to i32
+  %v1e2 = extractelement <4 x i8> %vec1, i64 2
+  %cv1e2 = sext i8 %v1e2 to i32
+  %mul3 = mul nuw nsw i32 %cv1e2, %cv2e2
+
+
+  %acc = load i32, ptr addrspace(1) %dst, align 4
+  %mad1 = add i32 %mul1, %acc
+  %mad2 = add i32 %mad1, %mul2
+  %mad3 = add i32 %mad2, %mul3
+
+  store i32 %mad3, ptr addrspace(1) %dst, align 4
+  ret void
+}
+
+define amdgpu_kernel void @idot4_acc32_3src_3ele_src0(ptr addrspace(1) %src1,
+; GFX7-LABEL: idot4_acc32_3src_3ele_src0:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx8 s[0:7], s[0:1], 0x9
+; GFX7-NEXT:    s_mov_b32 s11, 0xf000
+; GFX7-NEXT:    s_mov_b32 s14, 0
+; GFX7-NEXT:    s_mov_b32 s15, s11
+; GFX7-NEXT:    v_lshlrev_b32_e32 v0, 2, v0
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    s_mov_b64 s[12:13], s[0:1]
+; GFX7-NEXT:    v_mov_b32_e32 v1, 0
+; GFX7-NEXT:    buffer_load_dword v2, v[0:1], s[12:15], 0 addr64
+; GFX7-NEXT:    s_mov_b64 s[12:13], s[2:3]
+; GFX7-NEXT:    buffer_load_dword v3, v[0:1], s[12:15], 0 addr64
+; GFX7-NEXT:    s_mov_b64 s[12:13], s[4:5]
+; GFX7-NEXT:    buffer_load_dword v0, v[0:1], s[12:15], 0 addr64
+; GFX7-NEXT:    s_load_dword s0, s[6:7], 0x0
+; GFX7-NEXT:    s_mov_b32 s10, -1
+; GFX7-NEXT:    s_mov_b32 s8, s6
+; GFX7-NEXT:    s_mov_b32 s9, s7
+; GFX7-NEXT:    s_waitcnt vmcnt(2)
+; GFX7-NEXT:    v_bfe_i32 v1, v2, 8, 8
+; GFX7-NEXT:    s_waitcnt vmcnt(1)
+; GFX7-NEXT:    v_bfe_i32 v2, v3, 8, 8
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mad_i32_i24 v4, v2, v2, s0
+; GFX7-NEXT:    v_bfe_i32 v3, v3, 16, 8
+; GFX7-NEXT:    s_waitcnt vmcnt(0)
+; GFX7-NEXT:    v_bfe_i32 v0, v0, 16, 8
+; GFX7-NEXT:    v_mad_i32_i24 v1, v1, v2, v4
+; GFX7-NEXT:    v_mad_i32_i24 v0, v3, v0, v1
+; GFX7-NEXT:    buffer_store_dword v0, off, s[8:11], 0
+; GFX7-NEXT:    s_endpgm
+;
+; GFX8-LABEL: idot4_acc32_3src_3ele_src0:
+; GFX8:       ; %bb.0: ; %entry
+; GFX8-NEXT:    s_load_dwordx8 s[0:7], s[0:1], 0x24
+; GFX8-NEXT:    v_lshlrev_b32_e32 v2, 2, v0
+; GFX8-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX8-NEXT:    v_mov_b32_e32 v1, s1
+; GFX8-NEXT:    v_add_u32_e32 v0, vcc, s0, v2
+; GFX8-NEXT:    v_addc_u32_e32 v1, vcc, 0, v1, vcc
+; GFX8-NEXT:    flat_load_dword v3, v[0:1]
+; GFX8-NEXT:    v_mov_b32_e32 v1, s3
+; GFX8-NEXT:    v_add_u32_e32 v0, vcc, s2, v2
+; GFX8-NEXT:    v_addc_u32_e32 v1, vcc, 0, v1, vcc
+; GFX8-NEXT:    flat_load_dword v4, v[0:1]
+; GFX8-NEXT:    v_mov_b32_e32 v1, s5
+; GFX8-NEXT:    v_add_u32_e32 v0, vcc, s4, v2
+; GFX8-NEXT:    v_addc_u32_e32 v1, vcc, 0, v1, vcc
+; GFX8-NEXT:    flat_load_dword v0, v[0:1]
+; GFX8-NEXT:    s_load_dword s0, s[6:7], 0x0
+; GFX8-NEXT:    s_waitcnt vmcnt(2)
+; GFX8-NEXT:    v_bfe_i32 v2, v3, 8, 8
+; GFX8-NEXT:    s_waitcnt vmcnt(1)
+; GFX8-NEXT:    v_bfe_i32 v1, v4, 8, 8
+; GFX8-NEXT:    v_bfe_i32 v3, v4, 16, 8
+; GFX8-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX8-NEXT:    v_mad_i32_i24 v4, v1, v1, s0
+; GFX8-NEXT:    v_mad_i32_i24 v1, v2, v1, v4
+; GFX8-NEXT:    s_waitcnt vmcnt(0)
+; GFX8-NEXT:    v_bfe_i32 v0, v0, 16, 8
+; GFX8-NEXT:    v_mad_i32_i24 v2, v3, v0, v1
+; GFX8-NEXT:    v_mov_b32_e32 v0, s6
+; GFX8-NEXT:    v_mov_b32_e32 v1, s7
+; GFX8-NEXT:    flat_store_dword v[0:1], v2
+; GFX8-NEXT:    s_endpgm
+;
+; GFX9-NODL-LABEL: idot4_acc32_3src_3ele_src0:
+; GFX9-NODL:       ; %bb.0: ; %entry
+; GFX9-NODL-NEXT:    s_load_dwordx8 s[0:7], s[0:1], 0x24
+; GFX9-NODL-NEXT:    v_lshlrev_b32_e32 v0, 2, v0
+; GFX9-NODL-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX9-NODL-NEXT:    global_load_dword v1, v0, s[2:3]
+; GFX9-NODL-NEXT:    global_load_dword v2, v0, s[4:5]
+; GFX9-NODL-NEXT:    global_load_dword v3, v0, s[0:1]
+; GFX9-NODL-NEXT:    s_load_dword s0, s[6:7], 0x0
+; GFX9-NODL-NEXT:    v_mov_b32_e32 v0, 0
+; GFX9-NODL-NEXT:    s_waitcnt vmcnt(2)
+; GFX9-NODL-NEXT:    v_bfe_i32 v4, v1, 8, 8
+; GFX9-NODL-NEXT:    s_waitcnt vmcnt(1)
+; GFX9-NODL-NEXT:    v_mul_i32_i24_sdwa v1, sext(v1), sext(v2) dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_2 src1_sel:BYTE_2
+; GFX9-NODL-NEXT:    s_waitcnt vmcnt(0)
+; GFX9-NODL-NEXT:    v_mul_i32_i24_sdwa v2, sext(v3), v4 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_1 src1_sel:DWORD
+; GFX9-NODL-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX9-NODL-NEXT:    v_mad_i32_i24 v3, v4, v4, s0
+; GFX9-NODL-NEXT:    v_add3_u32 v1, v3, v2, v1
+; GFX9-NODL-NEXT:    global_store_dword v0, v1, s[6:7]
+; GFX9-NODL-NEXT:    s_endpgm
+;
+; GFX9-DL-LABEL: idot4_acc32_3src_3ele_src0:
+; GFX9-DL:       ; %bb.0: ; %entry
+; GFX9-DL-NEXT:    s_load_dwordx8 s[0:7], s[0:1], 0x24
+; GFX9-DL-NEXT:    v_lshlrev_b32_e32 v0, 2, v0
+; GFX9-DL-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX9-DL-NEXT:    global_load_dword v1, v0, s[4:5]
+; GFX9-DL-NEXT:    global_load_dword v2, v0, s[0:1]
+; GFX9-DL-NEXT:    global_load_dword v3, v0, s[2:3]
+; GFX9-DL-NEXT:    s_load_dword s3, s[6:7], 0x0
+; GFX9-DL-NEXT:    s_mov_b32 s0, 0xc06010c
+; GFX9-DL-NEXT:    s_mov_b32 s1, 0xc0c0c01
+; GFX9-DL-NEXT:    s_mov_b32 s2, 0xc020101
+; GFX9-DL-NEXT:    v_mov_b32_e32 v0, 0
+; GFX9-DL-NEXT:    s_waitcnt vmcnt(1)
+; GFX9-DL-NEXT:    v_perm_b32 v1, v1, v2, s0
+; GFX9-DL-NEXT:    s_waitcnt vmcnt(0)
+; GFX9-DL-NEXT:    v_perm_b32 v2, v3, v3, s1
+; GFX9-DL-NEXT:    v_or_b32_e32 v1, v1, v2
+; GFX9-DL-NEXT:    v_perm_b32 v2, v3, v3, s2
+; GFX9-DL-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX9-DL-NEXT:    v_dot4_i32_i8 v1, v2, v1, s3
+; GFX9-DL-NEXT:    global_store_dword v0, v1, s[6:7]
+; GFX9-DL-NEXT:    s_endpgm
+;
+; GFX10-DL-LABEL: idot4_acc32_3src_3ele_src0:
+; GFX10-DL:       ; %bb.0: ; %entry
+; GFX10-DL-NEXT:    s_load_dwordx8 s[0:7], s[0:1], 0x24
+; GFX10-DL-NEXT:    v_lshlrev_b32_e32 v0, 2, v0
+; GFX10-DL-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-DL-NEXT:    s_clause 0x2
+; GFX10-DL-NEXT:    global_load_dword v1, v0, s[4:5]
+; GFX10-DL-NEXT:    global_load_dword v2, v0, s[0:1]
+; GFX10-DL-NEXT:    global_load_dword v3, v0, s[2:3]
+; GFX10-DL-NEXT:    s_waitcnt_depctr 0xffe3
+; GFX10-DL-NEXT:    s_load_dword s0, s[6:7], 0x0
+; GFX10-DL-NEXT:    s_waitcnt vmcnt(1)
+; GFX10-DL-NEXT:    v_perm_b32 v0, v1, v2, 0xc06010c
+; GFX10-DL-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-DL-NEXT:    v_perm_b32 v1, v3, v3, 0xc0c0c01
+; GFX10-DL-NEXT:    v_perm_b32 v2, v3, v3, 0xc020101
+; GFX10-DL-NEXT:    v_mov_b32_e32 v3, 0
+; GFX10-DL-NEXT:    v_or_b32_e32 v0, v0, v1
+; GFX10-DL-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-DL-NEXT:    v_mov_b32_e32 v1, s0
+; GFX10-DL-NEXT:    v_dot4c_i32_i8_e32 v1, v2, v0
+; GFX10-DL-NEXT:    global_store_dword v3, v1, s[6:7]
+; GFX10-DL-NEXT:    s_endpgm
+;
+; GFX11-DL-LABEL: idot4_acc32_3src_3ele_src0:
+; GFX11-DL:       ; %bb.0: ; %entry
+; GFX11-DL-NEXT:    s_load_b256 s[0:7], s[0:1], 0x24
+; GFX11-DL-NEXT:    v_lshlrev_b32_e32 v0, 2, v0
+; GFX11-DL-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX11-DL-NEXT:    s_clause 0x2
+; GFX11-DL-NEXT:    global_load_b32 v1, v0, s[4:5]
+; GFX11-DL-NEXT:    global_load_b32 v2, v0, s[0:1]
+; GFX11-DL-NEXT:    global_load_b32 v0, v0, s[2:3]
+; GFX11-DL-NEXT:    s_load_b32 s0, s[6:7], 0x0
+; GFX11-DL-NEXT:    s_waitcnt vmcnt(1)
+; GFX11-DL-NEXT:    v_perm_b32 v1, v1, v2, 0xc06010c
+; GFX11-DL-NEXT:    s_waitcnt vmcnt(0)
+; GFX11-DL-NEXT:    v_perm_b32 v2, v0, v0, 0xc0c0c01
+; GFX11-DL-NEXT:    v_perm_b32 v0, v0, v0, 0xc020101
+; GFX11-DL-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_2) | instid1(VALU_DEP_2)
+; GFX11-DL-NEXT:    v_or_b32_e32 v1, v1, v2
+; GFX11-DL-NEXT:    v_mov_b32_e32 v2, 0
+; GFX11-DL-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX11-DL-NEXT:    v_dot4_i32_iu8 v0, v0, v1, s0 neg_lo:[1,1,0]
+; GFX11-DL-NEXT:    global_store_b32 v2, v0, s[6:7]
+; GFX11-DL-NEXT:    s_nop 0
+; GFX11-DL-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
+; GFX11-DL-NEXT:    s_endpgm
+                                       ptr addrspace(1) %src2,
+                                       ptr addrspace(1) %src3,
+                                       ptr addrspace(1) nocapture %dst) {
+entry:
+  %idx = call i32 @llvm.amdgcn.workitem.id.x()
+  %gep1 = getelementptr <4 x i8>, ptr addrspace(1) %src1, i32 %idx
+  %vec1 = load <4 x i8>, ptr addrspace(1) %gep1
+  %gep2 = getelementptr <4 x i8>, ptr addrspace(1) %src2, i32 %idx
+  %vec2 = load <4 x i8>, ptr addrspace(1) %gep2
+  %gep3 = getelementptr <4 x i8>, ptr addrspace(1) %src3, i32 %idx
+  %vec3 = load <4 x i8>, ptr addrspace(1) %gep3
+
+  %v2e0 = extractelement <4 x i8> %vec2, i64 1
+  %cv2e0 = sext i8 %v2e0 to i32
+  %mul1 = mul nuw nsw i32 %cv2e0, %cv2e0
+
+  %v1e1 = extractelement <4 x i8> %vec1, i64 1
+  %cv1e1 = sext i8 %v1e1 to i32
+  %v2e1 = extractelement <4 x i8> %vec2, i64 1
+  %cv2e1 = sext i8 %v2e1 to i32
+  %mul2 = mul nuw nsw i32 %cv1e1, %cv2e1
+
+  %v3e2 = extractelement <4 x i8> %vec3, i64 2
+  %cv3e2 = sext i8 %v3e2 to i32
+  %v2e2 = extractelement <4 x i8> %vec2, i64 2
+  %cv2e2 = sext i8 %v2e2 to i32
+  %mul3 = mul nuw nsw i32 %cv2e2, %cv3e2
+
+
+  %acc = load i32, ptr addrspace(1) %dst, align 4
+  %mad1 = add i32 %mul1, %acc
+  %mad2 = add i32 %mad1, %mul2
+  %mad3 = add i32 %mad2, %mul3
+
+  store i32 %mad3, ptr addrspace(1) %dst, align 4
+  ret void
+}
+
+define amdgpu_kernel void @idot4_4src(ptr addrspace(1) %src1,
+; GFX7-LABEL: idot4_4src:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx8 s[4:11], s[0:1], 0x9
+; GFX7-NEXT:    s_mov_b32 s3, 0xf000
+; GFX7-NEXT:    s_mov_b32 s14, 0
+; GFX7-NEXT:    s_mov_b32 s15, s3
+; GFX7-NEXT:    v_lshlrev_b32_e32 v0, 2, v0
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    s_mov_b64 s[12:13], s[4:5]
+; GFX7-NEXT:    v_mov_b32_e32 v1, 0
+; GFX7-NEXT:    buffer_load_dword v2, v[0:1], s[12:15], 0 addr64
+; GFX7-NEXT:    s_mov_b64 s[12:13], s[6:7]
+; GFX7-NEXT:    buffer_load_dword v3, v[0:1], s[12:15], 0 addr64
+; GFX7-NEXT:    s_mov_b64 s[12:13], s[8:9]
+; GFX7-NEXT:    buffer_load_dword v4, v[0:1], s[12:15], 0 addr64
+; GFX7-NEXT:    s_mov_b64 s[12:13], s[10:11]
+; GFX7-NEXT:    buffer_load_dword v0, v[0:1], s[12:15], 0 addr64
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0x11
+; GFX7-NEXT:    s_mov_b32 s2, -1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    s_load_dword s4, s[0:1], 0x0
+; GFX7-NEXT:    s_waitcnt vmcnt(3)
+; GFX7-NEXT:    v_bfe_i32 v1, v2, 0, 8
+; GFX7-NEXT:    v_bfe_i32 v2, v2, 8, 8
+; GFX7-NEXT:    s_waitcnt vmcnt(2)
+; GFX7-NEXT:    v_bfe_i32 v5, v3, 0, 8
+; GFX7-NEXT:    v_bfe_i32 v3, v3, 8, 8
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mad_i32_i24 v1, v1, v2, s4
+; GFX7-NEXT:    s_waitcnt vmcnt(1)
+; GFX7-NEXT:    v_bfe_i32 v2, v4, 0, 8
+; GFX7-NEXT:    v_bfe_i32 v4, v4, 8, 8
+; GFX7-NEXT:    v_mad_i32_i24 v1, v5, v3, v1
+; GFX7-NEXT:    s_waitcnt vmcnt(0)
+; GFX7-NEXT:    v_bfe_i32 v3, v0, 0, 8
+; GFX7-NEXT:    v_bfe_i32 v0, v0, 8, 8
+; GFX7-NEXT:    v_mad_i32_i24 v1, v2, v4, v1
+; GFX7-NEXT:    v_mad_i32_i24 v0, v3, v0, v1
+; GFX7-NEXT:    buffer_store_dword v0, off, s[0:3], 0
+; GFX7-NEXT:    s_endpgm
+;
+; GFX8-LABEL: idot4_4src:
+; GFX8:       ; %bb.0: ; %entry
+; GFX8-NEXT:    s_load_dwordx8 s[4:11], s[0:1], 0x24
+; GFX8-NEXT:    v_lshlrev_b32_e32 v2, 2, v0
+; GFX8-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0x44
+; GFX8-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX8-NEXT:    v_mov_b32_e32 v1, s5
+; GFX8-NEXT:    v_add_u32_e32 v0, vcc, s4, v2
+; GFX8-NEXT:    v_addc_u32_e32 v1, vcc, 0, v1, vcc
+; GFX8-NEXT:    flat_load_dword v3, v[0:1]
+; GFX8-NEXT:    v_mov_b32_e32 v1, s7
+; GFX8-NEXT:    v_add_u32_e32 v0, vcc, s6, v2
+; GFX8-NEXT:    v_addc_u32_e32 v1, vcc, 0, v1, vcc
+; GFX8-NEXT:    flat_load_dword v4, v[0:1]
+; GFX8-NEXT:    v_mov_b32_e32 v1, s9
+; GFX8-NEXT:    v_add_u32_e32 v0, vcc, s8, v2
+; GFX8-NEXT:    v_addc_u32_e32 v1, vcc, 0, v1, vcc
+; GFX8-NEXT:    flat_load_dword v5, v[0:1]
+; GFX8-NEXT:    v_mov_b32_e32 v1, s11
+; GFX8-NEXT:    v_add_u32_e32 v0, vcc, s10, v2
+; GFX8-NEXT:    v_addc_u32_e32 v1, vcc, 0, v1, vcc
+; GFX8-NEXT:    flat_load_dword v0, v[0:1]
+; GFX8-NEXT:    s_load_dword s2, s[0:1], 0x0
+; GFX8-NEXT:    s_waitcnt vmcnt(3)
+; GFX8-NEXT:    v_bfe_i32 v1, v3, 0, 8
+; GFX8-NEXT:    v_bfe_i32 v2, v3, 8, 8
+; GFX8-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX8-NEXT:    v_mad_i32_i24 v1, v1, v2, s2
+; GFX8-NEXT:    s_waitcnt vmcnt(2)
+; GFX8-NEXT:    v_bfe_i32 v3, v4, 0, 8
+; GFX8-NEXT:    v_bfe_i32 v4, v4, 8, 8
+; GFX8-NEXT:    v_mad_i32_i24 v1, v3, v4, v1
+; GFX8-NEXT:    s_waitcnt vmcnt(1)
+; GFX8-NEXT:    v_bfe_i32 v6, v5, 0, 8
+; GFX8-NEXT:    v_bfe_i32 v5, v5, 8, 8
+; GFX8-NEXT:    v_mad_i32_i24 v1, v6, v5, v1
+; GFX8-NEXT:    s_waitcnt vmcnt(0)
+; GFX8-NEXT:    v_bfe_i32 v7, v0, 0, 8
+; GFX8-NEXT:    v_bfe_i32 v0, v0, 8, 8
+; GFX8-NEXT:    v_mad_i32_i24 v2, v7, v0, v1
+; GFX8-NEXT:    v_mov_b32_e32 v0, s0
+; GFX8-NEXT:    v_mov_b32_e32 v1, s1
+; GFX8-NEXT:    flat_store_dword v[0:1], v2
+; GFX8-NEXT:    s_endpgm
+;
+; GFX9-NODL-LABEL: idot4_4src:
+; GFX9-NODL:       ; %bb.0: ; %entry
+; GFX9-NODL-NEXT:    s_load_dwordx8 s[4:11], s[0:1], 0x24
+; GFX9-NODL-NEXT:    v_lshlrev_b32_e32 v0, 2, v0
+; GFX9-NODL-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0x44
+; GFX9-NODL-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX9-NODL-NEXT:    global_load_dword v1, v0, s[4:5]
+; GFX9-NODL-NEXT:    global_load_dword v2, v0, s[6:7]
+; GFX9-NODL-NEXT:    global_load_dword v3, v0, s[8:9]
+; GFX9-NODL-NEXT:    global_load_dword v4, v0, s[10:11]
+; GFX9-NODL-NEXT:    v_mov_b32_e32 v0, 0
+; GFX9-NODL-NEXT:    s_load_dword s2, s[0:1], 0x0
+; GFX9-NODL-NEXT:    s_waitcnt vmcnt(3)
+; GFX9-NODL-NEXT:    v_mul_i32_i24_sdwa v1, sext(v1), sext(v1) dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:BYTE_1
+; GFX9-NODL-NEXT:    s_waitcnt vmcnt(2)
+; GFX9-NODL-NEXT:    v_mul_i32_i24_sdwa v2, sext(v2), sext(v2) dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:BYTE_1
+; GFX9-NODL-NEXT:    s_waitcnt vmcnt(1)
+; GFX9-NODL-NEXT:    v_mul_i32_i24_sdwa v3, sext(v3), sext(v3) dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:BYTE_1
+; GFX9-NODL-NEXT:    s_waitcnt vmcnt(0)
+; GFX9-NODL-NEXT:    v_mul_i32_i24_sdwa v4, sext(v4), sext(v4) dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:BYTE_1
+; GFX9-NODL-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX9-NODL-NEXT:    v_add3_u32 v1, v1, s2, v2
+; GFX9-NODL-NEXT:    v_add3_u32 v1, v1, v3, v4
+; GFX9-NODL-NEXT:    global_store_dword v0, v1, s[0:1]
+; GFX9-NODL-NEXT:    s_endpgm
+;
+; GFX9-DL-LABEL: idot4_4src:
+; GFX9-DL:       ; %bb.0: ; %entry
+; GFX9-DL-NEXT:    s_load_dwordx8 s[4:11], s[0:1], 0x24
+; GFX9-DL-NEXT:    v_lshlrev_b32_e32 v0, 2, v0
+; GFX9-DL-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0x44
+; GFX9-DL-NEXT:    s_mov_b32 s2, 0xc0c0501
+; GFX9-DL-NEXT:    s_mov_b32 s3, 0x5010c0c
+; GFX9-DL-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX9-DL-NEXT:    global_load_dword v1, v0, s[4:5]
+; GFX9-DL-NEXT:    global_load_dword v2, v0, s[6:7]
+; GFX9-DL-NEXT:    global_load_dword v3, v0, s[8:9]
+; GFX9-DL-NEXT:    global_load_dword v4, v0, s[10:11]
+; GFX9-DL-NEXT:    s_mov_b32 s4, 0xc0c0400
+; GFX9-DL-NEXT:    s_load_dword s6, s[0:1], 0x0
+; GFX9-DL-NEXT:    s_mov_b32 s5, 0x4000c0c
+; GFX9-DL-NEXT:    v_mov_b32_e32 v0, 0
+; GFX9-DL-NEXT:    s_waitcnt vmcnt(2)
+; GFX9-DL-NEXT:    v_perm_b32 v5, v2, v1, s2
+; GFX9-DL-NEXT:    v_perm_b32 v1, v2, v1, s4
+; GFX9-DL-NEXT:    s_waitcnt vmcnt(0)
+; GFX9-DL-NEXT:    v_perm_b32 v6, v4, v3, s3
+; GFX9-DL-NEXT:    v_perm_b32 v2, v4, v3, s5
+; GFX9-DL-NEXT:    v_or_b32_e32 v3, v6, v5
+; GFX9-DL-NEXT:    v_or_b32_e32 v1, v2, v1
+; GFX9-DL-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX9-DL-NEXT:    v_dot4_i32_i8 v1, v1, v3, s6
+; GFX9-DL-NEXT:    global_store_dword v0, v1, s[0:1]
+; GFX9-DL-NEXT:    s_endpgm
+;
+; GFX10-DL-LABEL: idot4_4src:
+; GFX10-DL:       ; %bb.0: ; %entry
+; GFX10-DL-NEXT:    s_load_dwordx8 s[4:11], s[0:1], 0x24
+; GFX10-DL-NEXT:    v_lshlrev_b32_e32 v0, 2, v0
+; GFX10-DL-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0x44
+; GFX10-DL-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-DL-NEXT:    s_clause 0x3
+; GFX10-DL-NEXT:    global_load_dword v1, v0, s[4:5]
+; GFX10-DL-NEXT:    global_load_dword v2, v0, s[6:7]
+; GFX10-DL-NEXT:    global_load_dword v3, v0, s[8:9]
+; GFX10-DL-NEXT:    global_load_dword v4, v0, s[10:11]
+; GFX10-DL-NEXT:    s_load_dword s2, s[0:1], 0x0
+; GFX10-DL-NEXT:    s_waitcnt vmcnt(2)
+; GFX10-DL-NEXT:    v_perm_b32 v0, v2, v1, 0xc0c0501
+; GFX10-DL-NEXT:    v_perm_b32 v1, v2, v1, 0xc0c0400
+; GFX10-DL-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-DL-NEXT:    v_perm_b32 v5, v4, v3, 0x5010c0c
+; GFX10-DL-NEXT:    v_perm_b32 v2, v4, v3, 0x4000c0c
+; GFX10-DL-NEXT:    v_mov_b32_e32 v3, 0
+; GFX10-DL-NEXT:    v_or_b32_e32 v0, v5, v0
+; GFX10-DL-NEXT:    v_or_b32_e32 v1, v2, v1
+; GFX10-DL-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-DL-NEXT:    v_mov_b32_e32 v2, s2
+; GFX10-DL-NEXT:    v_dot4c_i32_i8_e32 v2, v1, v0
+; GFX10-DL-NEXT:    global_store_dword v3, v2, s[0:1]
+; GFX10-DL-NEXT:    s_endpgm
+;
+; GFX11-DL-LABEL: idot4_4src:
+; GFX11-DL:       ; %bb.0: ; %entry
+; GFX11-DL-NEXT:    s_load_b256 s[4:11], s[0:1], 0x24
+; GFX11-DL-NEXT:    v_lshlrev_b32_e32 v0, 2, v0
+; GFX11-DL-NEXT:    s_load_b64 s[0:1], s[0:1], 0x44
+; GFX11-DL-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX11-DL-NEXT:    s_clause 0x3
+; GFX11-DL-NEXT:    global_load_b32 v1, v0, s[4:5]
+; GFX11-DL-NEXT:    global_load_b32 v2, v0, s[6:7]
+; GFX11-DL-NEXT:    global_load_b32 v3, v0, s[8:9]
+; GFX11-DL-NEXT:    global_load_b32 v0, v0, s[10:11]
+; GFX11-DL-NEXT:    s_load_b32 s2, s[0:1], 0x0
+; GFX11-DL-NEXT:    s_waitcnt vmcnt(2)
+; GFX11-DL-NEXT:    v_perm_b32 v4, v2, v1, 0xc0c0501
+; GFX11-DL-NEXT:    v_perm_b32 v1, v2, v1, 0xc0c0400
+; GFX11-DL-NEXT:    s_waitcnt vmcnt(0)
+; GFX11-DL-NEXT:    v_perm_b32 v5, v0, v3, 0x5010c0c
+; GFX11-DL-NEXT:    v_perm_b32 v0, v0, v3, 0x4000c0c
+; GFX11-DL-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX11-DL-NEXT:    v_or_b32_e32 v2, v5, v4
+; GFX11-DL-NEXT:    v_or_b32_e32 v0, v0, v1
+; GFX11-DL-NEXT:    v_mov_b32_e32 v1, 0
+; GFX11-DL-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX11-DL-NEXT:    s_delay_alu instid0(VALU_DEP_2)
+; GFX11-DL-NEXT:    v_dot4_i32_iu8 v0, v0, v2, s2 neg_lo:[1,1,0]
+; GFX11-DL-NEXT:    global_store_b32 v1, v0, s[0:1]
+; GFX11-DL-NEXT:    s_nop 0
+; GFX11-DL-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
+; GFX11-DL-NEXT:    s_endpgm
+                                       ptr addrspace(1) %src2,
+                                       ptr addrspace(1) %src3,
+                                       ptr addrspace(1) %src4,
+                                       ptr addrspace(1) nocapture %dst) {
+entry:
+  %idx = call i32 @llvm.amdgcn.workitem.id.x()
+
+  %gep1 = getelementptr <4 x i8>, ptr addrspace(1) %src1, i32 %idx
+  %vec1 = load <4 x i8>, ptr addrspace(1) %gep1
+  %gep2 = getelementptr <4 x i8>, ptr addrspace(1) %src2, i32 %idx
+  %vec2 = load <4 x i8>, ptr addrspace(1) %gep2
+  %gep3 = getelementptr <4 x i8>, ptr addrspace(1) %src3, i32 %idx
+  %vec3 = load <4 x i8>, ptr addrspace(1) %gep3
+  %gep4 = getelementptr <4 x i8>, ptr addrspace(1) %src4, i32 %idx
+  %vec4 = load <4 x i8>, ptr addrspace(1) %gep4
+
+
+  %v1e0 = extractelement <4 x i8> %vec1, i64 0
+  %cv1e0 = sext i8 %v1e0 to i32
+  %v1e1 = extractelement <4 x i8> %vec1, i64 1
+  %cv1e1 = sext i8 %v1e1 to i32
+  %mul1 = mul nuw nsw i32 %cv1e0, %cv1e1
+
+  %v2e0 = extractelement <4 x i8> %vec2, i64 0
+  %cv2e0 = sext i8 %v2e0 to i32
+  %v2e1 = extractelement <4 x i8> %vec2, i64 1
+  %cv2e1 = sext i8 %v2e1 to i32
+  %mul2 = mul nuw nsw i32 %cv2e0, %cv2e1
+
+  %v3e0 = extractelement <4 x i8> %vec3, i64 0
+  %cv3e0 = sext i8 %v3e0 to i32
+  %v3e1 = extractelement <4 x i8> %vec3, i64 1
+  %cv3e1 = sext i8 %v3e1 to i32
+  %mul3 = mul nuw nsw i32 %cv3e0, %cv3e1
+
+  %v4e0 = extractelement <4 x i8> %vec4, i64 0
+  %cv4e0 = sext i8 %v4e0 to i32
+  %v4e1 = extractelement <4 x i8> %vec4, i64 1
+  %cv4e1 = sext i8 %v4e1 to i32
+  %mul4 = mul nuw nsw i32 %cv4e0, %cv4e1
+
+
+  %acc = load i32, ptr addrspace(1) %dst, align 4
+  %mad1 = add i32 %mul1, %acc
+  %mad2 = add i32 %mad1, %mul2
+  %mad3 = add i32 %mad2, %mul3
+  %mad4 = add i32 %mad3, %mul4
+
+  store i32 %mad4, ptr addrspace(1) %dst, align 4
+  ret void
+}
+
+define amdgpu_kernel void @idot4_nonstandard_signed(ptr addrspace(1) %src1,
+; GFX7-LABEL: idot4_nonstandard_signed:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x9
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xd
+; GFX7-NEXT:    s_mov_b32 s3, 0xf000
+; GFX7-NEXT:    s_mov_b32 s10, 0
+; GFX7-NEXT:    s_mov_b32 s11, s3
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    s_mov_b64 s[8:9], s[4:5]
+; GFX7-NEXT:    v_lshlrev_b32_e32 v0, 2, v0
+; GFX7-NEXT:    v_mov_b32_e32 v1, 0
+; GFX7-NEXT:    buffer_load_dword v2, v[0:1], s[8:11], 0 addr64
+; GFX7-NEXT:    s_mov_b64 s[8:9], s[6:7]
+; GFX7-NEXT:    buffer_load_dword v0, v[0:1], s[8:11], 0 addr64
+; GFX7-NEXT:    s_mov_b32 s2, -1
+; GFX7-NEXT:    s_waitcnt vmcnt(1)
+; GFX7-NEXT:    v_bfe_i32 v1, v2, 0, 8
+; GFX7-NEXT:    v_bfe_i32 v3, v2, 8, 8
+; GFX7-NEXT:    v_and_b32_e32 v1, 0xffff, v1
+; GFX7-NEXT:    s_waitcnt vmcnt(0)
+; GFX7-NEXT:    v_and_b32_e32 v5, 0xff, v0
+; GFX7-NEXT:    v_bfe_i32 v4, v2, 16, 8
+; GFX7-NEXT:    v_bfe_u32 v6, v0, 8, 8
+; GFX7-NEXT:    v_and_b32_e32 v3, 0xffff, v3
+; GFX7-NEXT:    v_mul_u32_u24_e32 v1, v1, v5
+; GFX7-NEXT:    v_ashrrev_i32_e32 v2, 24, v2
+; GFX7-NEXT:    v_bfe_u32 v7, v0, 16, 8
+; GFX7-NEXT:    v_and_b32_e32 v4, 0xffff, v4
+; GFX7-NEXT:    v_mad_u32_u24 v1, v6, v3, v1
+; GFX7-NEXT:    v_lshrrev_b32_e32 v0, 24, v0
+; GFX7-NEXT:    v_mad_u32_u24 v1, v7, v4, v1
+; GFX7-NEXT:    v_and_b32_e32 v2, 0xffff, v2
+; GFX7-NEXT:    v_mad_u32_u24 v0, v2, v0, v1
+; GFX7-NEXT:    v_bfe_i32 v0, v0, 0, 16
+; GFX7-NEXT:    buffer_store_dword v0, off, s[0:3], 0
+; GFX7-NEXT:    s_endpgm
+;
+; GFX8-LABEL: idot4_nonstandard_signed:
+; GFX8:       ; %bb.0: ; %entry
+; GFX8-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x24
+; GFX8-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0x34
+; GFX8-NEXT:    v_lshlrev_b32_e32 v2, 2, v0
+; GFX8-NEXT:    v_mov_b32_e32 v4, 0xff
+; GFX8-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX8-NEXT:    v_mov_b32_e32 v1, s5
+; GFX8-NEXT:    v_add_u32_e32 v0, vcc, s4, v2
+; GFX8-NEXT:    v_addc_u32_e32 v1, vcc, 0, v1, vcc
+; GFX8-NEXT:    flat_load_dword v3, v[0:1]
+; GFX8-NEXT:    v_mov_b32_e32 v1, s7
+; GFX8-NEXT:    v_add_u32_e32 v0, vcc, s6, v2
+; GFX8-NEXT:    v_addc_u32_e32 v1, vcc, 0, v1, vcc
+; GFX8-NEXT:    flat_load_dword v2, v[0:1]
+; GFX8-NEXT:    v_mov_b32_e32 v0, s0
+; GFX8-NEXT:    v_mov_b32_e32 v1, s1
+; GFX8-NEXT:    s_waitcnt vmcnt(1)
+; GFX8-NEXT:    v_lshrrev_b32_e32 v7, 8, v3
+; GFX8-NEXT:    v_lshrrev_b32_e32 v5, 16, v3
+; GFX8-NEXT:    v_bfe_i32 v7, v7, 0, 8
+; GFX8-NEXT:    v_bfe_i32 v5, v5, 0, 8
+; GFX8-NEXT:    s_waitcnt vmcnt(0)
+; GFX8-NEXT:    v_lshrrev_b32_e32 v8, 8, v2
+; GFX8-NEXT:    v_mul_lo_u16_sdwa v6, sext(v3), v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:BYTE_0
+; GFX8-NEXT:    v_and_b32_e32 v8, 0xff, v8
+; GFX8-NEXT:    v_and_b32_sdwa v4, v2, v4 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
+; GFX8-NEXT:    v_lshrrev_b32_e32 v3, 24, v3
+; GFX8-NEXT:    v_mad_u16 v6, v8, v7, v6
+; GFX8-NEXT:    v_bfe_i32 v3, v3, 0, 8
+; GFX8-NEXT:    v_mad_u16 v4, v4, v5, v6
+; GFX8-NEXT:    v_lshrrev_b32_e32 v2, 24, v2
+; GFX8-NEXT:    v_mad_u16 v2, v3, v2, v4
+; GFX8-NEXT:    v_bfe_i32 v2, v2, 0, 16
+; GFX8-NEXT:    flat_store_dword v[0:1], v2
+; GFX8-NEXT:    s_endpgm
+;
+; GFX9-NODL-LABEL: idot4_nonstandard_signed:
+; GFX9-NODL:       ; %bb.0: ; %entry
+; GFX9-NODL-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x24
+; GFX9-NODL-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x34
+; GFX9-NODL-NEXT:    v_lshlrev_b32_e32 v0, 2, v0
+; GFX9-NODL-NEXT:    s_movk_i32 s0, 0xff
+; GFX9-NODL-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX9-NODL-NEXT:    global_load_dword v1, v0, s[4:5]
+; GFX9-NODL-NEXT:    global_load_dword v2, v0, s[6:7]
+; GFX9-NODL-NEXT:    v_mov_b32_e32 v0, 0
+; GFX9-NODL-NEXT:    s_waitcnt vmcnt(1)
+; GFX9-NODL-NEXT:    v_lshrrev_b32_e32 v5, 8, v1
+; GFX9-NODL-NEXT:    s_waitcnt vmcnt(0)
+; GFX9-NODL-NEXT:    v_lshrrev_b32_e32 v6, 8, v2
+; GFX9-NODL-NEXT:    v_lshrrev_b32_e32 v3, 16, v1
+; GFX9-NODL-NEXT:    v_mul_lo_u16_sdwa v4, sext(v1), v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:BYTE_0
+; GFX9-NODL-NEXT:    v_bfe_i32 v5, v5, 0, 8
+; GFX9-NODL-NEXT:    v_and_b32_e32 v6, 0xff, v6
+; GFX9-NODL-NEXT:    v_and_b32_sdwa v7, v2, s0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
+; GFX9-NODL-NEXT:    v_lshrrev_b32_e32 v1, 24, v1
+; GFX9-NODL-NEXT:    v_bfe_i32 v3, v3, 0, 8
+; GFX9-NODL-NEXT:    v_mad_legacy_u16 v4, v6, v5, v4
+; GFX9-NODL-NEXT:    v_bfe_i32 v1, v1, 0, 8
+; GFX9-NODL-NEXT:    v_mad_legacy_u16 v3, v7, v3, v4
+; GFX9-NODL-NEXT:    v_lshrrev_b32_e32 v2, 24, v2
+; GFX9-NODL-NEXT:    v_mad_legacy_u16 v1, v1, v2, v3
+; GFX9-NODL-NEXT:    v_bfe_i32 v1, v1, 0, 16
+; GFX9-NODL-NEXT:    global_store_dword v0, v1, s[2:3]
+; GFX9-NODL-NEXT:    s_endpgm
+;
+; GFX9-DL-LABEL: idot4_nonstandard_signed:
+; GFX9-DL:       ; %bb.0: ; %entry
+; GFX9-DL-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x24
+; GFX9-DL-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x34
+; GFX9-DL-NEXT:    v_lshlrev_b32_e32 v0, 2, v0
+; GFX9-DL-NEXT:    s_movk_i32 s0, 0xff
+; GFX9-DL-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX9-DL-NEXT:    global_load_dword v1, v0, s[4:5]
+; GFX9-DL-NEXT:    global_load_dword v2, v0, s[6:7]
+; GFX9-DL-NEXT:    v_mov_b32_e32 v0, 0
+; GFX9-DL-NEXT:    s_waitcnt vmcnt(1)
+; GFX9-DL-NEXT:    v_lshrrev_b32_e32 v5, 8, v1
+; GFX9-DL-NEXT:    s_waitcnt vmcnt(0)
+; GFX9-DL-NEXT:    v_lshrrev_b32_e32 v6, 8, v2
+; GFX9-DL-NEXT:    v_lshrrev_b32_e32 v3, 16, v1
+; GFX9-DL-NEXT:    v_mul_lo_u16_sdwa v4, sext(v1), v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:BYTE_0
+; GFX9-DL-NEXT:    v_bfe_i32 v5, v5, 0, 8
+; GFX9-DL-NEXT:    v_and_b32_e32 v6, 0xff, v6
+; GFX9-DL-NEXT:    v_and_b32_sdwa v7, v2, s0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
+; GFX9-DL-NEXT:    v_lshrrev_b32_e32 v1, 24, v1
+; GFX9-DL-NEXT:    v_bfe_i32 v3, v3, 0, 8
+; GFX9-DL-NEXT:    v_mad_legacy_u16 v4, v6, v5, v4
+; GFX9-DL-NEXT:    v_bfe_i32 v1, v1, 0, 8
+; GFX9-DL-NEXT:    v_mad_legacy_u16 v3, v7, v3, v4
+; GFX9-DL-NEXT:    v_lshrrev_b32_e32 v2, 24, v2
+; GFX9-DL-NEXT:    v_mad_legacy_u16 v1, v1, v2, v3
+; GFX9-DL-NEXT:    v_bfe_i32 v1, v1, 0, 16
+; GFX9-DL-NEXT:    global_store_dword v0, v1, s[2:3]
+; GFX9-DL-NEXT:    s_endpgm
+;
+; GFX10-DL-LABEL: idot4_nonstandard_signed:
+; GFX10-DL:       ; %bb.0: ; %entry
+; GFX10-DL-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x24
+; GFX10-DL-NEXT:    v_lshlrev_b32_e32 v0, 2, v0
+; GFX10-DL-NEXT:    v_mov_b32_e32 v6, 0xff
+; GFX10-DL-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0x34
+; GFX10-DL-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-DL-NEXT:    s_clause 0x1
+; GFX10-DL-NEXT:    global_load_dword v1, v0, s[4:5]
+; GFX10-DL-NEXT:    global_load_dword v2, v0, s[6:7]
+; GFX10-DL-NEXT:    s_waitcnt vmcnt(1)
+; GFX10-DL-NEXT:    v_bfe_i32 v0, v1, 0, 8
+; GFX10-DL-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-DL-NEXT:    v_and_b32_e32 v3, 0xff, v2
+; GFX10-DL-NEXT:    v_lshrrev_b32_e32 v4, 8, v1
+; GFX10-DL-NEXT:    v_lshrrev_b32_e32 v5, 8, v2
+; GFX10-DL-NEXT:    v_lshrrev_b32_e32 v7, 16, v1
+; GFX10-DL-NEXT:    v_lshrrev_b32_e32 v1, 24, v1
+; GFX10-DL-NEXT:    v_mul_lo_u16 v0, v0, v3
+; GFX10-DL-NEXT:    v_bfe_i32 v3, v4, 0, 8
+; GFX10-DL-NEXT:    v_and_b32_e32 v4, 0xff, v5
+; GFX10-DL-NEXT:    v_and_b32_sdwa v5, v2, v6 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
+; GFX10-DL-NEXT:    v_bfe_i32 v6, v7, 0, 8
+; GFX10-DL-NEXT:    v_bfe_i32 v1, v1, 0, 8
+; GFX10-DL-NEXT:    v_lshrrev_b32_e32 v2, 24, v2
+; GFX10-DL-NEXT:    v_mad_u16 v0, v4, v3, v0
+; GFX10-DL-NEXT:    v_mad_u16 v0, v5, v6, v0
+; GFX10-DL-NEXT:    v_mad_u16 v0, v1, v2, v0
+; GFX10-DL-NEXT:    v_mov_b32_e32 v1, 0
+; GFX10-DL-NEXT:    v_bfe_i32 v0, v0, 0, 16
+; GFX10-DL-NEXT:    global_store_dword v1, v0, s[0:1]
+; GFX10-DL-NEXT:    s_endpgm
+;
+; GFX11-DL-LABEL: idot4_nonstandard_signed:
+; GFX11-DL:       ; %bb.0: ; %entry
+; GFX11-DL-NEXT:    s_load_b128 s[4:7], s[0:1], 0x24
+; GFX11-DL-NEXT:    v_lshlrev_b32_e32 v0, 2, v0
+; GFX11-DL-NEXT:    s_load_b64 s[0:1], s[0:1], 0x34
+; GFX11-DL-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX11-DL-NEXT:    s_clause 0x1
+; GFX11-DL-NEXT:    global_load_b32 v1, v0, s[4:5]
+; GFX11-DL-NEXT:    global_load_b32 v0, v0, s[6:7]
+; GFX11-DL-NEXT:    s_waitcnt vmcnt(1)
+; GFX11-DL-NEXT:    v_bfe_i32 v2, v1, 0, 8
+; GFX11-DL-NEXT:    s_waitcnt vmcnt(0)
+; GFX11-DL-NEXT:    v_and_b32_e32 v3, 0xff, v0
+; GFX11-DL-NEXT:    v_lshrrev_b32_e32 v4, 8, v1
+; GFX11-DL-NEXT:    v_lshrrev_b32_e32 v5, 8, v0
+; GFX11-DL-NEXT:    v_lshrrev_b32_e32 v6, 16, v1
+; GFX11-DL-NEXT:    v_lshrrev_b32_e32 v7, 16, v0
+; GFX11-DL-NEXT:    v_mul_lo_u16 v2, v2, v3
+; GFX11-DL-NEXT:    v_bfe_i32 v3, v4, 0, 8
+; GFX11-DL-NEXT:    v_and_b32_e32 v4, 0xff, v5
+; GFX11-DL-NEXT:    v_lshrrev_b32_e32 v1, 24, v1
+; GFX11-DL-NEXT:    v_bfe_i32 v5, v6, 0, 8
+; GFX11-DL-NEXT:    v_and_b32_e32 v6, 0xff, v7
+; GFX11-DL-NEXT:    v_lshrrev_b32_e32 v0, 24, v0
+; GFX11-DL-NEXT:    v_mad_u16 v2, v4, v3, v2
+; GFX11-DL-NEXT:    v_bfe_i32 v1, v1, 0, 8
+; GFX11-DL-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-DL-NEXT:    v_mad_u16 v2, v6, v5, v2
+; GFX11-DL-NEXT:    v_mad_u16 v0, v1, v0, v2
+; GFX11-DL-NEXT:    v_mov_b32_e32 v1, 0
+; GFX11-DL-NEXT:    s_delay_alu instid0(VALU_DEP_2)
+; GFX11-DL-NEXT:    v_bfe_i32 v0, v0, 0, 16
+; GFX11-DL-NEXT:    global_store_b32 v1, v0, s[0:1]
+; GFX11-DL-NEXT:    s_nop 0
+; GFX11-DL-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
+; GFX11-DL-NEXT:    s_endpgm
+                                       ptr addrspace(1) %src2,
+                                       ptr addrspace(1) nocapture %dst) {
+entry:
+  %idx = call i32 @llvm.amdgcn.workitem.id.x()
+  %gep1 = getelementptr <4 x i8>, ptr addrspace(1) %src1, i32 %idx
+  %vec1 = load <4 x i8>, ptr addrspace(1) %gep1
+  %gep2 = getelementptr <4 x i8>, ptr addrspace(1) %src2, i32 %idx
+  %vec2 = load <4 x i8>, ptr addrspace(1) %gep2
+  %v1e0 = extractelement <4 x i8> %vec1, i64 0
+  %v1e0e = sext i8 %v1e0 to i16
+  %v2e0 = extractelement <4 x i8> %vec2, i64 0
+  %v2e0e = zext i8 %v2e0 to i16
+  %mul0 = mul nsw i16 %v1e0e, %v2e0e
+  %add0 = add i16 %mul0, 0
+
+  %v1e1 = extractelement <4 x i8> %vec1, i64 1
+  %v1e1e = sext i8 %v1e1 to i16
+  %v2e1 = extractelement <4 x i8> %vec2, i64 1
+  %v2e1e = zext i8 %v2e1 to i16
+  %mul1 = mul nsw i16 %v2e1e, %v1e1e
+  %add1 = add i16 %mul1, %add0
+  %v1e2 = extractelement <4 x i8> %vec1, i64 2
+  %v1e2e = sext i8 %v1e2 to i16
+  %v2e2 = extractelement <4 x i8> %vec2, i64 2
+  %v2e2e = zext i8 %v2e2 to i16
+  %mul2 = mul nsw i16 %v2e2e, %v1e2e
+  %add2 = add i16 %mul2, %add1
+  %v1e3 = extractelement <4 x i8> %vec1, i64 3
+  %v1e3e = sext i8 %v1e3 to i16
+  %v2e3 = extractelement <4 x i8> %vec2, i64 3
+  %v2e3e = zext i8 %v2e3 to i16
+  %mul3 = mul nsw i16 %v1e3e, %v2e3e
+  %add3 = add i16 %mul3, %add2
+  %res = sext i16 %add3 to i32
+  store i32 %res, ptr addrspace(1) %dst, align 4
+  ret void
+}
+
+
 declare i32 @llvm.amdgcn.workitem.id.x()

llvm/test/CodeGen/AMDGPU/idot4u.ll

 ; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py
 ; RUN: llc -mtriple=amdgcn -mcpu=gfx700 -verify-machineinstrs < %s | FileCheck -check-prefixes=GFX7 %s
 ; RUN: llc -mtriple=amdgcn -mcpu=gfx803 -verify-machineinstrs < %s | FileCheck -check-prefixes=GFX8 %s
 ; RUN: llc -mtriple=amdgcn -mcpu=gfx900 -verify-machineinstrs < %s | FileCheck -check-prefixes=GFX9-NODL %s
 ; RUN: llc -mtriple=amdgcn -mcpu=gfx906 -verify-machineinstrs < %s | FileCheck -check-prefixes=GFX9-DL %s
 ; RUN: llc -mtriple=amdgcn -mcpu=gfx1011 -verify-machineinstrs < %s | FileCheck -check-prefixes=GFX10-DL %s
 ; RUN: llc -mtriple=amdgcn -mcpu=gfx1012 -verify-machineinstrs < %s | FileCheck -check-prefixes=GFX10-DL %s
+; RUN: llc -mtriple=amdgcn -mcpu=gfx1100 -verify-machineinstrs < %s | FileCheck -check-prefixes=GFX11-DL %s
 
 define amdgpu_kernel void @udot4_acc32(ptr addrspace(1) %src1,
 ; GFX7-LABEL: udot4_acc32:
 ; GFX7:       ; %bb.0: ; %entry
 ; GFX7-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x9
 ; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xd
 ; GFX7-NEXT:    s_mov_b32 s3, 0xf000
 ; GFX7-NEXT:    s_mov_b32 s10, 0

 ; GFX10-DL-NEXT:    global_load_dword v1, v0, s[4:5]
 ; GFX10-DL-NEXT:    global_load_dword v2, v0, s[6:7]
 ; GFX10-DL-NEXT:    v_mov_b32_e32 v0, 0
 ; GFX10-DL-NEXT:    s_load_dword s2, s[0:1], 0x0
 ; GFX10-DL-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
 ; GFX10-DL-NEXT:    v_dot4_u32_u8 v1, v1, v2, s2
 ; GFX10-DL-NEXT:    global_store_dword v0, v1, s[0:1]
 ; GFX10-DL-NEXT:    s_endpgm
+;
+; GFX11-DL-LABEL: udot4_acc32:
+; GFX11-DL:       ; %bb.0: ; %entry
+; GFX11-DL-NEXT:    s_load_b128 s[4:7], s[0:1], 0x24
+; GFX11-DL-NEXT:    v_lshlrev_b32_e32 v0, 2, v0
+; GFX11-DL-NEXT:    s_load_b64 s[0:1], s[0:1], 0x34
+; GFX11-DL-NEXT:    v_mov_b32_e32 v2, 0
+; GFX11-DL-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX11-DL-NEXT:    s_clause 0x1
+; GFX11-DL-NEXT:    global_load_b32 v1, v0, s[4:5]
+; GFX11-DL-NEXT:    global_load_b32 v0, v0, s[6:7]
+; GFX11-DL-NEXT:    s_load_b32 s2, s[0:1], 0x0
+; GFX11-DL-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX11-DL-NEXT:    v_dot4_u32_u8 v0, v1, v0, s2
+; GFX11-DL-NEXT:    global_store_b32 v2, v0, s[0:1]
+; GFX11-DL-NEXT:    s_nop 0
+; GFX11-DL-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
+; GFX11-DL-NEXT:    s_endpgm
                                        ptr addrspace(1) %src2,
                                        ptr addrspace(1) nocapture %dst) {
 entry:
   %idx = call i32 @llvm.amdgcn.workitem.id.x()
   %gep1 = getelementptr <4 x i8>, ptr addrspace(1) %src1, i32 %idx
   %vec1 = load <4 x i8>, ptr addrspace(1) %gep1
   %gep2 = getelementptr <4 x i8>, ptr addrspace(1) %src2, i32 %idx
   %vec2 = load <4 x i8>, ptr addrspace(1) %gep2

 ; GFX9-NODL-NEXT:    global_store_short v0, v1, s[2:3]
 ; GFX9-NODL-NEXT:    s_endpgm
 ;
 ; GFX9-DL-LABEL: udot4_acc16:
 ; GFX9-DL:       ; %bb.0: ; %entry
 ; GFX9-DL-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x24
 ; GFX9-DL-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x34
 ; GFX9-DL-NEXT:    v_lshlrev_b32_e32 v0, 2, v0
-; GFX9-DL-NEXT:    s_movk_i32 s0, 0xff
+; GFX9-DL-NEXT:    v_mov_b32_e32 v1, 0
 ; GFX9-DL-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX9-DL-NEXT:    global_load_dword v1, v0, s[4:5]
-; GFX9-DL-NEXT:    global_load_dword v2, v0, s[6:7]
-; GFX9-DL-NEXT:    v_mov_b32_e32 v0, 0
-; GFX9-DL-NEXT:    global_load_ushort v3, v0, s[2:3]
-; GFX9-DL-NEXT:    s_waitcnt vmcnt(2)
-; GFX9-DL-NEXT:    v_and_b32_e32 v4, 0xff, v1
-; GFX9-DL-NEXT:    s_waitcnt vmcnt(1)
-; GFX9-DL-NEXT:    v_and_b32_e32 v5, 0xff, v2
-; GFX9-DL-NEXT:    v_lshrrev_b32_e32 v6, 8, v1
-; GFX9-DL-NEXT:    v_lshrrev_b32_e32 v7, 8, v2
-; GFX9-DL-NEXT:    v_and_b32_e32 v6, 0xff, v6
-; GFX9-DL-NEXT:    v_and_b32_e32 v7, 0xff, v7
+; GFX9-DL-NEXT:    global_load_dword v2, v0, s[4:5]
+; GFX9-DL-NEXT:    global_load_dword v3, v0, s[6:7]
+; GFX9-DL-NEXT:    global_load_ushort v4, v1, s[2:3]
 ; GFX9-DL-NEXT:    s_waitcnt vmcnt(0)
-; GFX9-DL-NEXT:    v_mad_legacy_u16 v3, v4, v5, v3
-; GFX9-DL-NEXT:    v_and_b32_sdwa v8, v1, s0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; GFX9-DL-NEXT:    v_and_b32_sdwa v9, v2, s0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; GFX9-DL-NEXT:    v_mad_legacy_u16 v3, v6, v7, v3
-; GFX9-DL-NEXT:    v_lshrrev_b32_e32 v1, 24, v1
-; GFX9-DL-NEXT:    v_lshrrev_b32_e32 v2, 24, v2
-; GFX9-DL-NEXT:    v_mad_legacy_u16 v3, v8, v9, v3
-; GFX9-DL-NEXT:    v_mad_legacy_u16 v1, v1, v2, v3
-; GFX9-DL-NEXT:    global_store_short v0, v1, s[2:3]
+; GFX9-DL-NEXT:    v_dot4_u32_u8 v0, v2, v3, v4
+; GFX9-DL-NEXT:    global_store_short v1, v0, s[2:3]
 ; GFX9-DL-NEXT:    s_endpgm
 ;
 ; GFX10-DL-LABEL: udot4_acc16:
 ; GFX10-DL:       ; %bb.0: ; %entry
 ; GFX10-DL-NEXT:    s_clause 0x1
 ; GFX10-DL-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x24
 ; GFX10-DL-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x34
 ; GFX10-DL-NEXT:    v_lshlrev_b32_e32 v0, 2, v0
-; GFX10-DL-NEXT:    v_mov_b32_e32 v8, 0xff
+; GFX10-DL-NEXT:    v_mov_b32_e32 v1, 0
 ; GFX10-DL-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX10-DL-NEXT:    s_clause 0x1
-; GFX10-DL-NEXT:    global_load_dword v1, v0, s[4:5]
-; GFX10-DL-NEXT:    global_load_dword v2, v0, s[6:7]
-; GFX10-DL-NEXT:    v_mov_b32_e32 v0, 0
-; GFX10-DL-NEXT:    global_load_ushort v3, v0, s[2:3]
-; GFX10-DL-NEXT:    s_waitcnt vmcnt(2)
-; GFX10-DL-NEXT:    v_and_b32_e32 v4, 0xff, v1
-; GFX10-DL-NEXT:    v_lshrrev_b32_e32 v5, 8, v1
-; GFX10-DL-NEXT:    s_waitcnt vmcnt(1)
-; GFX10-DL-NEXT:    v_lshrrev_b32_e32 v6, 8, v2
-; GFX10-DL-NEXT:    v_and_b32_e32 v7, 0xff, v2
-; GFX10-DL-NEXT:    v_and_b32_e32 v5, 0xff, v5
-; GFX10-DL-NEXT:    v_and_b32_e32 v6, 0xff, v6
+; GFX10-DL-NEXT:    global_load_dword v2, v0, s[4:5]
+; GFX10-DL-NEXT:    global_load_dword v3, v0, s[6:7]
+; GFX10-DL-NEXT:    global_load_ushort v4, v1, s[2:3]
 ; GFX10-DL-NEXT:    s_waitcnt vmcnt(0)
-; GFX10-DL-NEXT:    v_mad_u16 v3, v4, v7, v3
-; GFX10-DL-NEXT:    v_and_b32_sdwa v4, v1, v8 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; GFX10-DL-NEXT:    v_and_b32_sdwa v7, v2, v8 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; GFX10-DL-NEXT:    v_lshrrev_b32_e32 v1, 24, v1
-; GFX10-DL-NEXT:    v_lshrrev_b32_e32 v2, 24, v2
-; GFX10-DL-NEXT:    v_mad_u16 v3, v5, v6, v3
-; GFX10-DL-NEXT:    v_mad_u16 v3, v4, v7, v3
-; GFX10-DL-NEXT:    v_mad_u16 v1, v1, v2, v3
-; GFX10-DL-NEXT:    global_store_short v0, v1, s[2:3]
+; GFX10-DL-NEXT:    v_dot4_u32_u8 v0, v2, v3, v4
+; GFX10-DL-NEXT:    global_store_short v1, v0, s[2:3]
 ; GFX10-DL-NEXT:    s_endpgm
+;
+; GFX11-DL-LABEL: udot4_acc16:
+; GFX11-DL:       ; %bb.0: ; %entry
+; GFX11-DL-NEXT:    s_clause 0x1
+; GFX11-DL-NEXT:    s_load_b128 s[4:7], s[0:1], 0x24
+; GFX11-DL-NEXT:    s_load_b64 s[0:1], s[0:1], 0x34
+; GFX11-DL-NEXT:    v_dual_mov_b32 v1, 0 :: v_dual_lshlrev_b32 v0, 2, v0
+; GFX11-DL-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX11-DL-NEXT:    s_clause 0x1
+; GFX11-DL-NEXT:    global_load_b32 v2, v0, s[4:5]
+; GFX11-DL-NEXT:    global_load_b32 v0, v0, s[6:7]
+; GFX11-DL-NEXT:    global_load_u16 v3, v1, s[0:1]
+; GFX11-DL-NEXT:    s_waitcnt vmcnt(0)
+; GFX11-DL-NEXT:    v_dot4_u32_u8 v0, v2, v0, v3
+; GFX11-DL-NEXT:    global_store_b16 v1, v0, s[0:1]
+; GFX11-DL-NEXT:    s_nop 0
+; GFX11-DL-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
+; GFX11-DL-NEXT:    s_endpgm
                                        ptr addrspace(1) %src2,
                                        ptr addrspace(1) nocapture %dst) {
 entry:
   %idx = call i32 @llvm.amdgcn.workitem.id.x()
   %gep1 = getelementptr <4 x i8>, ptr addrspace(1) %src1, i32 %idx
   %vec1 = load <4 x i8>, ptr addrspace(1) %gep1
   %gep2 = getelementptr <4 x i8>, ptr addrspace(1) %src2, i32 %idx
   %vec2 = load <4 x i8>, ptr addrspace(1) %gep2

 ; GFX9-NODL-NEXT:    global_store_byte v0, v1, s[2:3]
 ; GFX9-NODL-NEXT:    s_endpgm
 ;
 ; GFX9-DL-LABEL: udot4_acc8:
 ; GFX9-DL:       ; %bb.0: ; %entry
 ; GFX9-DL-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x24
 ; GFX9-DL-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x34
 ; GFX9-DL-NEXT:    v_lshlrev_b32_e32 v0, 2, v0
+; GFX9-DL-NEXT:    v_mov_b32_e32 v1, 0
 ; GFX9-DL-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX9-DL-NEXT:    global_load_dword v1, v0, s[4:5]
-; GFX9-DL-NEXT:    global_load_dword v2, v0, s[6:7]
-; GFX9-DL-NEXT:    v_mov_b32_e32 v0, 0
-; GFX9-DL-NEXT:    global_load_ubyte v3, v0, s[2:3]
-; GFX9-DL-NEXT:    s_waitcnt vmcnt(2)
-; GFX9-DL-NEXT:    v_lshrrev_b32_e32 v4, 16, v1
-; GFX9-DL-NEXT:    v_lshrrev_b32_e32 v6, 8, v1
-; GFX9-DL-NEXT:    s_waitcnt vmcnt(1)
-; GFX9-DL-NEXT:    v_lshrrev_b32_e32 v7, 8, v2
-; GFX9-DL-NEXT:    v_lshrrev_b32_e32 v8, 24, v1
+; GFX9-DL-NEXT:    global_load_dword v2, v0, s[4:5]
+; GFX9-DL-NEXT:    global_load_dword v3, v0, s[6:7]
+; GFX9-DL-NEXT:    global_load_ubyte v4, v1, s[2:3]
 ; GFX9-DL-NEXT:    s_waitcnt vmcnt(0)
-; GFX9-DL-NEXT:    v_mad_legacy_u16 v1, v1, v2, v3
-; GFX9-DL-NEXT:    v_lshrrev_b32_e32 v5, 16, v2
-; GFX9-DL-NEXT:    v_mad_legacy_u16 v1, v6, v7, v1
-; GFX9-DL-NEXT:    v_lshrrev_b32_e32 v9, 24, v2
-; GFX9-DL-NEXT:    v_mad_legacy_u16 v1, v4, v5, v1
-; GFX9-DL-NEXT:    v_mad_legacy_u16 v1, v8, v9, v1
-; GFX9-DL-NEXT:    global_store_byte v0, v1, s[2:3]
+; GFX9-DL-NEXT:    v_dot4_u32_u8 v0, v2, v3, v4
+; GFX9-DL-NEXT:    global_store_byte v1, v0, s[2:3]
 ; GFX9-DL-NEXT:    s_endpgm
 ;
 ; GFX10-DL-LABEL: udot4_acc8:
 ; GFX10-DL:       ; %bb.0: ; %entry
 ; GFX10-DL-NEXT:    s_clause 0x1
 ; GFX10-DL-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x24
 ; GFX10-DL-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x34
 ; GFX10-DL-NEXT:    v_lshlrev_b32_e32 v0, 2, v0
 ; GFX10-DL-NEXT:    v_mov_b32_e32 v1, 0
 ; GFX10-DL-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX10-DL-NEXT:    s_clause 0x1
 ; GFX10-DL-NEXT:    global_load_dword v2, v0, s[4:5]
 ; GFX10-DL-NEXT:    global_load_dword v3, v0, s[6:7]
 ; GFX10-DL-NEXT:    global_load_ubyte v4, v1, s[2:3]
-; GFX10-DL-NEXT:    s_waitcnt vmcnt(2)
-; GFX10-DL-NEXT:    v_lshrrev_b32_e32 v0, 8, v2
-; GFX10-DL-NEXT:    s_waitcnt vmcnt(1)
-; GFX10-DL-NEXT:    v_lshrrev_b32_e32 v5, 8, v3
 ; GFX10-DL-NEXT:    s_waitcnt vmcnt(0)
-; GFX10-DL-NEXT:    v_mad_u16 v4, v2, v3, v4
-; GFX10-DL-NEXT:    v_lshrrev_b32_e32 v6, 16, v2
-; GFX10-DL-NEXT:    v_lshrrev_b32_e32 v7, 16, v3
-; GFX10-DL-NEXT:    v_lshrrev_b32_e32 v2, 24, v2
-; GFX10-DL-NEXT:    v_lshrrev_b32_e32 v3, 24, v3
-; GFX10-DL-NEXT:    v_mad_u16 v0, v0, v5, v4
-; GFX10-DL-NEXT:    v_mad_u16 v0, v6, v7, v0
-; GFX10-DL-NEXT:    v_mad_u16 v0, v2, v3, v0
+; GFX10-DL-NEXT:    v_dot4_u32_u8 v0, v2, v3, v4
 ; GFX10-DL-NEXT:    global_store_byte v1, v0, s[2:3]
 ; GFX10-DL-NEXT:    s_endpgm
+;
+; GFX11-DL-LABEL: udot4_acc8:
+; GFX11-DL:       ; %bb.0: ; %entry
+; GFX11-DL-NEXT:    s_clause 0x1
+; GFX11-DL-NEXT:    s_load_b128 s[4:7], s[0:1], 0x24
+; GFX11-DL-NEXT:    s_load_b64 s[0:1], s[0:1], 0x34
+; GFX11-DL-NEXT:    v_dual_mov_b32 v1, 0 :: v_dual_lshlrev_b32 v0, 2, v0
+; GFX11-DL-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX11-DL-NEXT:    s_clause 0x1
+; GFX11-DL-NEXT:    global_load_b32 v2, v0, s[4:5]
+; GFX11-DL-NEXT:    global_load_b32 v0, v0, s[6:7]
+; GFX11-DL-NEXT:    global_load_u8 v3, v1, s[0:1]
+; GFX11-DL-NEXT:    s_waitcnt vmcnt(0)
+; GFX11-DL-NEXT:    v_dot4_u32_u8 v0, v2, v0, v3
+; GFX11-DL-NEXT:    global_store_b8 v1, v0, s[0:1]
+; GFX11-DL-NEXT:    s_nop 0
+; GFX11-DL-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
+; GFX11-DL-NEXT:    s_endpgm
                                       ptr addrspace(1) %src2,
                                       ptr addrspace(1) nocapture %dst) {
 entry:
   %idx = call i32 @llvm.amdgcn.workitem.id.x()
   %gep1 = getelementptr <4 x i8>, ptr addrspace(1) %src1, i32 %idx
   %vec1 = load <4 x i8>, ptr addrspace(1) %gep1
   %gep2 = getelementptr <4 x i8>, ptr addrspace(1) %src2, i32 %idx
   %vec2 = load <4 x i8>, ptr addrspace(1) %gep2
entry:
   %mad2 = add i8 %mad1, %mul2
   %mad3 = add i8 %mad2, %mul3
   %mad4 = add i8 %mad3, %mul4
 
   store i8 %mad4, ptr addrspace(1) %dst, align 2
   ret void
 }
 
-; TODO: Generate udot4?
 define amdgpu_kernel void @udot2_8(ptr addrspace(1) %src1,
 ; GFX7-LABEL: udot2_8:
 ; GFX7:       ; %bb.0: ; %entry
 ; GFX7-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x9
 ; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xd
 ; GFX7-NEXT:    s_mov_b32 s3, 0xf000
 ; GFX7-NEXT:    s_mov_b32 s10, 0
 ; GFX7-NEXT:    s_mov_b32 s11, s3

 ; GFX9-NODL-NEXT:    global_store_byte v1, v0, s[2:3]
 ; GFX9-NODL-NEXT:    s_endpgm
 ;
 ; GFX9-DL-LABEL: udot2_8:
 ; GFX9-DL:       ; %bb.0: ; %entry
 ; GFX9-DL-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x24
 ; GFX9-DL-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x34
 ; GFX9-DL-NEXT:    v_lshlrev_b32_e32 v0, 2, v0
-; GFX9-DL-NEXT:    v_mov_b32_e32 v1, 0
+; GFX9-DL-NEXT:    s_mov_b32 s0, 0xc0c0100
 ; GFX9-DL-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX9-DL-NEXT:    global_load_dword v2, v0, s[4:5]
-; GFX9-DL-NEXT:    global_load_dword v3, v0, s[6:7]
-; GFX9-DL-NEXT:    global_load_ubyte v4, v1, s[2:3]
+; GFX9-DL-NEXT:    global_load_dword v1, v0, s[4:5]
+; GFX9-DL-NEXT:    global_load_dword v2, v0, s[6:7]
+; GFX9-DL-NEXT:    v_mov_b32_e32 v0, 0
+; GFX9-DL-NEXT:    global_load_ubyte v3, v0, s[2:3]
 ; GFX9-DL-NEXT:    s_waitcnt vmcnt(2)
-; GFX9-DL-NEXT:    v_lshrrev_b32_e32 v0, 8, v2
+; GFX9-DL-NEXT:    v_perm_b32 v1, v1, v1, s0
 ; GFX9-DL-NEXT:    s_waitcnt vmcnt(1)
-; GFX9-DL-NEXT:    v_lshrrev_b32_e32 v5, 8, v3
+; GFX9-DL-NEXT:    v_perm_b32 v2, v2, v2, s0
 ; GFX9-DL-NEXT:    s_waitcnt vmcnt(0)
-; GFX9-DL-NEXT:    v_mad_legacy_u16 v2, v2, v3, v4
-; GFX9-DL-NEXT:    v_mad_legacy_u16 v0, v0, v5, v2
-; GFX9-DL-NEXT:    global_store_byte v1, v0, s[2:3]
+; GFX9-DL-NEXT:    v_dot4_u32_u8 v1, v1, v2, v3
+; GFX9-DL-NEXT:    global_store_byte v0, v1, s[2:3]
 ; GFX9-DL-NEXT:    s_endpgm
 ;
 ; GFX10-DL-LABEL: udot2_8:
 ; GFX10-DL:       ; %bb.0: ; %entry
 ; GFX10-DL-NEXT:    s_clause 0x1
 ; GFX10-DL-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x24
 ; GFX10-DL-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x34
 ; GFX10-DL-NEXT:    v_lshlrev_b32_e32 v0, 2, v0
-; GFX10-DL-NEXT:    v_mov_b32_e32 v1, 0
 ; GFX10-DL-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX10-DL-NEXT:    s_clause 0x1
-; GFX10-DL-NEXT:    global_load_dword v2, v0, s[4:5]
-; GFX10-DL-NEXT:    global_load_dword v3, v0, s[6:7]
-; GFX10-DL-NEXT:    global_load_ubyte v4, v1, s[2:3]
+; GFX10-DL-NEXT:    global_load_dword v1, v0, s[4:5]
+; GFX10-DL-NEXT:    global_load_dword v2, v0, s[6:7]
+; GFX10-DL-NEXT:    v_mov_b32_e32 v0, 0
+; GFX10-DL-NEXT:    global_load_ubyte v3, v0, s[2:3]
 ; GFX10-DL-NEXT:    s_waitcnt vmcnt(2)
-; GFX10-DL-NEXT:    v_lshrrev_b32_e32 v0, 8, v2
+; GFX10-DL-NEXT:    v_perm_b32 v1, v1, v1, 0xc0c0100
 ; GFX10-DL-NEXT:    s_waitcnt vmcnt(1)
-; GFX10-DL-NEXT:    v_lshrrev_b32_e32 v5, 8, v3
+; GFX10-DL-NEXT:    v_perm_b32 v2, v2, v2, 0xc0c0100
 ; GFX10-DL-NEXT:    s_waitcnt vmcnt(0)
-; GFX10-DL-NEXT:    v_mad_u16 v2, v2, v3, v4
-; GFX10-DL-NEXT:    v_mad_u16 v0, v0, v5, v2
-; GFX10-DL-NEXT:    global_store_byte v1, v0, s[2:3]
+; GFX10-DL-NEXT:    v_dot4_u32_u8 v1, v1, v2, v3
+; GFX10-DL-NEXT:    global_store_byte v0, v1, s[2:3]
 ; GFX10-DL-NEXT:    s_endpgm
+;
+; GFX11-DL-LABEL: udot2_8:
+; GFX11-DL:       ; %bb.0: ; %entry
+; GFX11-DL-NEXT:    s_clause 0x1
+; GFX11-DL-NEXT:    s_load_b128 s[4:7], s[0:1], 0x24
+; GFX11-DL-NEXT:    s_load_b64 s[0:1], s[0:1], 0x34
+; GFX11-DL-NEXT:    v_lshlrev_b32_e32 v0, 2, v0
+; GFX11-DL-NEXT:    v_mov_b32_e32 v2, 0
+; GFX11-DL-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX11-DL-NEXT:    s_clause 0x1
+; GFX11-DL-NEXT:    global_load_b32 v1, v0, s[4:5]
+; GFX11-DL-NEXT:    global_load_b32 v0, v0, s[6:7]
+; GFX11-DL-NEXT:    global_load_u8 v3, v2, s[0:1]
+; GFX11-DL-NEXT:    s_waitcnt vmcnt(2)
+; GFX11-DL-NEXT:    v_perm_b32 v1, v1, v1, 0xc0c0100
+; GFX11-DL-NEXT:    s_waitcnt vmcnt(1)
+; GFX11-DL-NEXT:    v_perm_b32 v0, v0, v0, 0xc0c0100
+; GFX11-DL-NEXT:    s_waitcnt vmcnt(0)
+; GFX11-DL-NEXT:    s_delay_alu instid0(VALU_DEP_1)
+; GFX11-DL-NEXT:    v_dot4_u32_u8 v0, v1, v0, v3
+; GFX11-DL-NEXT:    global_store_b8 v2, v0, s[0:1]
+; GFX11-DL-NEXT:    s_nop 0
+; GFX11-DL-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
+; GFX11-DL-NEXT:    s_endpgm
                                    ptr addrspace(1) %src2,
                                    ptr addrspace(1) nocapture %dst) {
 entry:
   %idx = call i32 @llvm.amdgcn.workitem.id.x()
   %gep1 = getelementptr <4 x i8>, ptr addrspace(1) %src1, i32 %idx
   %vec1 = load <4 x i8>, ptr addrspace(1) %gep1
   %gep2 = getelementptr <4 x i8>, ptr addrspace(1) %src2, i32 %idx
   %vec2 = load <4 x i8>, ptr addrspace(1) %gep2

 ; GFX9-NODL-NEXT:    global_store_byte v0, v1, s[2:3]
 ; GFX9-NODL-NEXT:    s_endpgm
 ;
 ; GFX9-DL-LABEL: udot4_CommutationInsideMAD:
 ; GFX9-DL:       ; %bb.0: ; %entry
 ; GFX9-DL-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x24
 ; GFX9-DL-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x34
 ; GFX9-DL-NEXT:    v_lshlrev_b32_e32 v0, 2, v0
+; GFX9-DL-NEXT:    v_mov_b32_e32 v1, 0
 ; GFX9-DL-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX9-DL-NEXT:    global_load_dword v1, v0, s[4:5]
-; GFX9-DL-NEXT:    global_load_dword v2, v0, s[6:7]
-; GFX9-DL-NEXT:    v_mov_b32_e32 v0, 0
-; GFX9-DL-NEXT:    global_load_ubyte v3, v0, s[2:3]
-; GFX9-DL-NEXT:    s_waitcnt vmcnt(2)
-; GFX9-DL-NEXT:    v_lshrrev_b32_e32 v4, 16, v1
-; GFX9-DL-NEXT:    v_lshrrev_b32_e32 v6, 8, v1
-; GFX9-DL-NEXT:    s_waitcnt vmcnt(1)
-; GFX9-DL-NEXT:    v_lshrrev_b32_e32 v7, 8, v2
-; GFX9-DL-NEXT:    v_lshrrev_b32_e32 v8, 24, v1
+; GFX9-DL-NEXT:    global_load_dword v2, v0, s[4:5]
+; GFX9-DL-NEXT:    global_load_dword v3, v0, s[6:7]
+; GFX9-DL-NEXT:    global_load_ubyte v4, v1, s[2:3]
 ; GFX9-DL-NEXT:    s_waitcnt vmcnt(0)
-; GFX9-DL-NEXT:    v_mad_legacy_u16 v1, v2, v1, v3
-; GFX9-DL-NEXT:    v_lshrrev_b32_e32 v5, 16, v2
-; GFX9-DL-NEXT:    v_mad_legacy_u16 v1, v7, v6, v1
-; GFX9-DL-NEXT:    v_lshrrev_b32_e32 v9, 24, v2
-; GFX9-DL-NEXT:    v_mad_legacy_u16 v1, v5, v4, v1
-; GFX9-DL-NEXT:    v_mad_legacy_u16 v1, v9, v8, v1
-; GFX9-DL-NEXT:    global_store_byte v0, v1, s[2:3]
+; GFX9-DL-NEXT:    v_dot4_u32_u8 v0, v3, v2, v4
+; GFX9-DL-NEXT:    global_store_byte v1, v0, s[2:3]
 ; GFX9-DL-NEXT:    s_endpgm
 ;
 ; GFX10-DL-LABEL: udot4_CommutationInsideMAD:
 ; GFX10-DL:       ; %bb.0: ; %entry
 ; GFX10-DL-NEXT:    s_clause 0x1
 ; GFX10-DL-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x24
 ; GFX10-DL-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x34
 ; GFX10-DL-NEXT:    v_lshlrev_b32_e32 v0, 2, v0
 ; GFX10-DL-NEXT:    v_mov_b32_e32 v1, 0
 ; GFX10-DL-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX10-DL-NEXT:    s_clause 0x1
 ; GFX10-DL-NEXT:    global_load_dword v2, v0, s[4:5]
 ; GFX10-DL-NEXT:    global_load_dword v3, v0, s[6:7]
 ; GFX10-DL-NEXT:    global_load_ubyte v4, v1, s[2:3]
-; GFX10-DL-NEXT:    s_waitcnt vmcnt(2)
-; GFX10-DL-NEXT:    v_lshrrev_b32_e32 v0, 8, v2
-; GFX10-DL-NEXT:    s_waitcnt vmcnt(1)
-; GFX10-DL-NEXT:    v_lshrrev_b32_e32 v5, 8, v3
 ; GFX10-DL-NEXT:    s_waitcnt vmcnt(0)
-; GFX10-DL-NEXT:    v_mad_u16 v4, v3, v2, v4
-; GFX10-DL-NEXT:    v_lshrrev_b32_e32 v6, 16, v2
-; GFX10-DL-NEXT:    v_lshrrev_b32_e32 v7, 16, v3
-; GFX10-DL-NEXT:    v_lshrrev_b32_e32 v2, 24, v2
-; GFX10-DL-NEXT:    v_lshrrev_b32_e32 v3, 24, v3
-; GFX10-DL-NEXT:    v_mad_u16 v0, v5, v0, v4
-; GFX10-DL-NEXT:    v_mad_u16 v0, v7, v6, v0
-; GFX10-DL-NEXT:    v_mad_u16 v0, v3, v2, v0
+; GFX10-DL-NEXT:    v_dot4_u32_u8 v0, v3, v2, v4
 ; GFX10-DL-NEXT:    global_store_byte v1, v0, s[2:3]
 ; GFX10-DL-NEXT:    s_endpgm
+;
+; GFX11-DL-LABEL: udot4_CommutationInsideMAD:
+; GFX11-DL:       ; %bb.0: ; %entry
+; GFX11-DL-NEXT:    s_clause 0x1
+; GFX11-DL-NEXT:    s_load_b128 s[4:7], s[0:1], 0x24
+; GFX11-DL-NEXT:    s_load_b64 s[0:1], s[0:1], 0x34
+; GFX11-DL-NEXT:    v_dual_mov_b32 v1, 0 :: v_dual_lshlrev_b32 v0, 2, v0
+; GFX11-DL-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX11-DL-NEXT:    s_clause 0x1
+; GFX11-DL-NEXT:    global_load_b32 v2, v0, s[4:5]
+; GFX11-DL-NEXT:    global_load_b32 v0, v0, s[6:7]
+; GFX11-DL-NEXT:    global_load_u8 v3, v1, s[0:1]
+; GFX11-DL-NEXT:    s_waitcnt vmcnt(0)
+; GFX11-DL-NEXT:    v_dot4_u32_u8 v0, v0, v2, v3
+; GFX11-DL-NEXT:    global_store_b8 v1, v0, s[0:1]
+; GFX11-DL-NEXT:    s_nop 0
+; GFX11-DL-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
+; GFX11-DL-NEXT:    s_endpgm
                                                       ptr addrspace(1) %src2,
                                                       ptr addrspace(1) nocapture %dst) {
 entry:
   %idx = call i32 @llvm.amdgcn.workitem.id.x()
   %gep1 = getelementptr <4 x i8>, ptr addrspace(1) %src1, i32 %idx
   %vec1 = load <4 x i8>, ptr addrspace(1) %gep1
   %gep2 = getelementptr <4 x i8>, ptr addrspace(1) %src2, i32 %idx
   %vec2 = load <4 x i8>, ptr addrspace(1) %gep2
entry:
   %mad2 = add i8 %mul2, %mad1
   %mad3 = add i8 %mul3, %mad2
   %mad4 = add i8 %mul4, %mad3
 
   store i8 %mad4, ptr addrspace(1) %dst, align 2
   ret void
 }
 
-; TODO: Support commutation accross the adds.
 define amdgpu_kernel void @udot4_CommutationAccrossMADs(ptr addrspace(1) %src1,
 ; GFX7-LABEL: udot4_CommutationAccrossMADs:
 ; GFX7:       ; %bb.0: ; %entry
 ; GFX7-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x9
 ; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xd
 ; GFX7-NEXT:    s_mov_b32 s3, 0xf000
 ; GFX7-NEXT:    s_mov_b32 s10, 0
 ; GFX7-NEXT:    s_mov_b32 s11, s3

 ; GFX9-NODL-NEXT:    global_store_byte v0, v1, s[2:3]
 ; GFX9-NODL-NEXT:    s_endpgm
 ;
 ; GFX9-DL-LABEL: udot4_CommutationAccrossMADs:
 ; GFX9-DL:       ; %bb.0: ; %entry
 ; GFX9-DL-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x24
 ; GFX9-DL-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x34
 ; GFX9-DL-NEXT:    v_lshlrev_b32_e32 v0, 2, v0
+; GFX9-DL-NEXT:    v_mov_b32_e32 v1, 0
 ; GFX9-DL-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX9-DL-NEXT:    global_load_dword v1, v0, s[4:5]
-; GFX9-DL-NEXT:    global_load_dword v2, v0, s[6:7]
-; GFX9-DL-NEXT:    v_mov_b32_e32 v0, 0
-; GFX9-DL-NEXT:    global_load_ubyte v3, v0, s[2:3]
-; GFX9-DL-NEXT:    s_waitcnt vmcnt(2)
-; GFX9-DL-NEXT:    v_lshrrev_b32_e32 v6, 8, v1
-; GFX9-DL-NEXT:    s_waitcnt vmcnt(1)
-; GFX9-DL-NEXT:    v_lshrrev_b32_e32 v7, 8, v2
-; GFX9-DL-NEXT:    v_lshrrev_b32_e32 v4, 16, v1
+; GFX9-DL-NEXT:    global_load_dword v2, v0, s[4:5]
+; GFX9-DL-NEXT:    global_load_dword v3, v0, s[6:7]
+; GFX9-DL-NEXT:    global_load_ubyte v4, v1, s[2:3]
 ; GFX9-DL-NEXT:    s_waitcnt vmcnt(0)
-; GFX9-DL-NEXT:    v_mad_legacy_u16 v3, v7, v6, v3
-; GFX9-DL-NEXT:    v_lshrrev_b32_e32 v5, 16, v2
-; GFX9-DL-NEXT:    v_lshrrev_b32_e32 v8, 24, v1
-; GFX9-DL-NEXT:    v_mad_legacy_u16 v1, v2, v1, v3
-; GFX9-DL-NEXT:    v_lshrrev_b32_e32 v9, 24, v2
-; GFX9-DL-NEXT:    v_mad_legacy_u16 v1, v5, v4, v1
-; GFX9-DL-NEXT:    v_mad_legacy_u16 v1, v9, v8, v1
-; GFX9-DL-NEXT:    global_store_byte v0, v1, s[2:3]
+; GFX9-DL-NEXT:    v_dot4_u32_u8 v0, v3, v2, v4
+; GFX9-DL-NEXT:    global_store_byte v1, v0, s[2:3]
 ; GFX9-DL-NEXT:    s_endpgm
 ;
 ; GFX10-DL-LABEL: udot4_CommutationAccrossMADs:
 ; GFX10-DL:       ; %bb.0: ; %entry
 ; GFX10-DL-NEXT:    s_clause 0x1
 ; GFX10-DL-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x24
 ; GFX10-DL-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x34
 ; GFX10-DL-NEXT:    v_lshlrev_b32_e32 v0, 2, v0
 ; GFX10-DL-NEXT:    v_mov_b32_e32 v1, 0
 ; GFX10-DL-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX10-DL-NEXT:    s_clause 0x1
 ; GFX10-DL-NEXT:    global_load_dword v2, v0, s[4:5]
 ; GFX10-DL-NEXT:    global_load_dword v3, v0, s[6:7]
 ; GFX10-DL-NEXT:    global_load_ubyte v4, v1, s[2:3]
-; GFX10-DL-NEXT:    s_waitcnt vmcnt(2)
-; GFX10-DL-NEXT:    v_lshrrev_b32_e32 v0, 8, v2
-; GFX10-DL-NEXT:    s_waitcnt vmcnt(1)
-; GFX10-DL-NEXT:    v_lshrrev_b32_e32 v5, 8, v3
 ; GFX10-DL-NEXT:    s_waitcnt vmcnt(0)
-; GFX10-DL-NEXT:    v_mad_u16 v0, v5, v0, v4
-; GFX10-DL-NEXT:    v_lshrrev_b32_e32 v4, 16, v2
-; GFX10-DL-NEXT:    v_lshrrev_b32_e32 v5, 16, v3
-; GFX10-DL-NEXT:    v_mad_u16 v0, v3, v2, v0
-; GFX10-DL-NEXT:    v_lshrrev_b32_e32 v2, 24, v2
-; GFX10-DL-NEXT:    v_lshrrev_b32_e32 v3, 24, v3
-; GFX10-DL-NEXT:    v_mad_u16 v0, v5, v4, v0
-; GFX10-DL-NEXT:    v_mad_u16 v0, v3, v2, v0
+; GFX10-DL-NEXT:    v_dot4_u32_u8 v0, v3, v2, v4
 ; GFX10-DL-NEXT:    global_store_byte v1, v0, s[2:3]
 ; GFX10-DL-NEXT:    s_endpgm
+;
+; GFX11-DL-LABEL: udot4_CommutationAccrossMADs:
+; GFX11-DL:       ; %bb.0: ; %entry
+; GFX11-DL-NEXT:    s_clause 0x1
+; GFX11-DL-NEXT:    s_load_b128 s[4:7], s[0:1], 0x24
+; GFX11-DL-NEXT:    s_load_b64 s[0:1], s[0:1], 0x34
+; GFX11-DL-NEXT:    v_dual_mov_b32 v1, 0 :: v_dual_lshlrev_b32 v0, 2, v0
+; GFX11-DL-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX11-DL-NEXT:    s_clause 0x1
+; GFX11-DL-NEXT:    global_load_b32 v2, v0, s[4:5]
+; GFX11-DL-NEXT:    global_load_b32 v0, v0, s[6:7]
+; GFX11-DL-NEXT:    global_load_u8 v3, v1, s[0:1]
+; GFX11-DL-NEXT:    s_waitcnt vmcnt(0)
+; GFX11-DL-NEXT:    v_dot4_u32_u8 v0, v0, v2, v3
+; GFX11-DL-NEXT:    global_store_b8 v1, v0, s[0:1]
+; GFX11-DL-NEXT:    s_nop 0
+; GFX11-DL-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
+; GFX11-DL-NEXT:    s_endpgm
                                                         ptr addrspace(1) %src2,
                                                         ptr addrspace(1) nocapture %dst) {
 entry:
   %idx = call i32 @llvm.amdgcn.workitem.id.x()
   %gep1 = getelementptr <4 x i8>, ptr addrspace(1) %src1, i32 %idx
   %vec1 = load <4 x i8>, ptr addrspace(1) %gep1
   %gep2 = getelementptr <4 x i8>, ptr addrspace(1) %src2, i32 %idx
   %vec2 = load <4 x i8>, ptr addrspace(1) %gep2

 ; GFX9-DL-NEXT:    global_load_dword v1, v0, s[4:5]
 ; GFX9-DL-NEXT:    global_load_dword v2, v0, s[6:7]
 ; GFX9-DL-NEXT:    s_load_dword s0, s[2:3], 0x0
 ; GFX9-DL-NEXT:    v_mov_b32_e32 v0, 0
 ; GFX9-DL-NEXT:    s_waitcnt vmcnt(1)
 ; GFX9-DL-NEXT:    v_and_b32_e32 v3, 0xff, v1
 ; GFX9-DL-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-DL-NEXT:    v_and_b32_e32 v4, 0xff, v2
-; GFX9-DL-NEXT:    v_mul_u32_u24_sdwa v5, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_1 src1_sel:BYTE_1
-; GFX9-DL-NEXT:    v_mul_u32_u24_sdwa v6, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_2 src1_sel:BYTE_2
-; GFX9-DL-NEXT:    v_mul_u32_u24_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_3 src1_sel:BYTE_3
-; GFX9-DL-NEXT:    v_mul_u32_u24_e32 v2, v3, v4
 ; GFX9-DL-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX9-DL-NEXT:    v_mad_u32_u24 v3, v3, v4, s0
-; GFX9-DL-NEXT:    v_add3_u32 v2, v5, v3, v2
-; GFX9-DL-NEXT:    v_add3_u32 v1, v2, v6, v1
+; GFX9-DL-NEXT:    v_dot4_u32_u8 v1, v1, v2, v3
 ; GFX9-DL-NEXT:    global_store_dword v0, v1, s[2:3]
 ; GFX9-DL-NEXT:    s_endpgm
 ;
 ; GFX10-DL-LABEL: udot4_multiuse_mul1:
 ; GFX10-DL:       ; %bb.0: ; %entry
 ; GFX10-DL-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x24
 ; GFX10-DL-NEXT:    v_lshlrev_b32_e32 v0, 2, v0
 ; GFX10-DL-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0x34
 ; GFX10-DL-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX10-DL-NEXT:    s_clause 0x1
 ; GFX10-DL-NEXT:    global_load_dword v1, v0, s[4:5]
 ; GFX10-DL-NEXT:    global_load_dword v2, v0, s[6:7]
 ; GFX10-DL-NEXT:    s_load_dword s2, s[0:1], 0x0
 ; GFX10-DL-NEXT:    s_waitcnt vmcnt(1)
 ; GFX10-DL-NEXT:    v_and_b32_e32 v0, 0xff, v1
 ; GFX10-DL-NEXT:    s_waitcnt vmcnt(0)
 ; GFX10-DL-NEXT:    v_and_b32_e32 v3, 0xff, v2
-; GFX10-DL-NEXT:    v_mul_u32_u24_sdwa v4, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_1 src1_sel:BYTE_1
-; GFX10-DL-NEXT:    v_mul_u32_u24_e32 v5, v0, v3
 ; GFX10-DL-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX10-DL-NEXT:    v_mad_u32_u24 v0, v0, v3, s2
-; GFX10-DL-NEXT:    v_mul_u32_u24_sdwa v3, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_2 src1_sel:BYTE_2
-; GFX10-DL-NEXT:    v_mul_u32_u24_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_3 src1_sel:BYTE_3
-; GFX10-DL-NEXT:    v_mov_b32_e32 v2, 0
-; GFX10-DL-NEXT:    v_add3_u32 v0, v4, v0, v5
-; GFX10-DL-NEXT:    v_add3_u32 v0, v0, v3, v1
-; GFX10-DL-NEXT:    global_store_dword v2, v0, s[0:1]
+; GFX10-DL-NEXT:    v_mov_b32_e32 v3, 0
+; GFX10-DL-NEXT:    v_dot4_u32_u8 v0, v1, v2, v0
+; GFX10-DL-NEXT:    global_store_dword v3, v0, s[0:1]
 ; GFX10-DL-NEXT:    s_endpgm
+;
+; GFX11-DL-LABEL: udot4_multiuse_mul1:
+; GFX11-DL:       ; %bb.0: ; %entry
+; GFX11-DL-NEXT:    s_load_b128 s[4:7], s[0:1], 0x24
+; GFX11-DL-NEXT:    v_lshlrev_b32_e32 v0, 2, v0
+; GFX11-DL-NEXT:    s_load_b64 s[0:1], s[0:1], 0x34
+; GFX11-DL-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX11-DL-NEXT:    s_clause 0x1
+; GFX11-DL-NEXT:    global_load_b32 v1, v0, s[4:5]
+; GFX11-DL-NEXT:    global_load_b32 v0, v0, s[6:7]
+; GFX11-DL-NEXT:    s_load_b32 s2, s[0:1], 0x0
+; GFX11-DL-NEXT:    s_waitcnt vmcnt(1)
+; GFX11-DL-NEXT:    v_and_b32_e32 v2, 0xff, v1
+; GFX11-DL-NEXT:    s_waitcnt vmcnt(0)
+; GFX11-DL-NEXT:    v_and_b32_e32 v3, 0xff, v0
+; GFX11-DL-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX11-DL-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2)
+; GFX11-DL-NEXT:    v_mad_u32_u24 v2, v2, v3, s2
+; GFX11-DL-NEXT:    v_mov_b32_e32 v3, 0
+; GFX11-DL-NEXT:    v_dot4_u32_u8 v0, v1, v0, v2
+; GFX11-DL-NEXT:    global_store_b32 v3, v0, s[0:1]
+; GFX11-DL-NEXT:    s_nop 0
+; GFX11-DL-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
+; GFX11-DL-NEXT:    s_endpgm
                                                ptr addrspace(1) %src2,
                                                ptr addrspace(1) nocapture %dst) {
 entry:
   %idx = call i32 @llvm.amdgcn.workitem.id.x()
   %gep1 = getelementptr <4 x i8>, ptr addrspace(1) %src1, i32 %idx
   %vec1 = load <4 x i8>, ptr addrspace(1) %gep1
   %gep2 = getelementptr <4 x i8>, ptr addrspace(1) %src2, i32 %idx
   %vec2 = load <4 x i8>, ptr addrspace(1) %gep2

 ; GFX9-DL-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x24
 ; GFX9-DL-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x34
 ; GFX9-DL-NEXT:    v_lshlrev_b32_e32 v0, 2, v0
 ; GFX9-DL-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX9-DL-NEXT:    global_load_dword v1, v0, s[4:5]
 ; GFX9-DL-NEXT:    global_load_dword v2, v0, s[6:7]
 ; GFX9-DL-NEXT:    s_load_dword s0, s[2:3], 0x0
 ; GFX9-DL-NEXT:    v_mov_b32_e32 v0, 0
-; GFX9-DL-NEXT:    s_waitcnt vmcnt(1)
-; GFX9-DL-NEXT:    v_bfe_u32 v4, v1, 8, 8
+; GFX9-DL-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX9-DL-NEXT:    s_add_i32 s1, s0, s0
 ; GFX9-DL-NEXT:    s_waitcnt vmcnt(0)
-; GFX9-DL-NEXT:    v_bfe_u32 v5, v2, 8, 8
-; GFX9-DL-NEXT:    v_mul_u32_u24_sdwa v3, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:BYTE_0
-; GFX9-DL-NEXT:    v_mul_u32_u24_sdwa v6, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_2 src1_sel:BYTE_2
-; GFX9-DL-NEXT:    v_mul_u32_u24_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_3 src1_sel:BYTE_3
-; GFX9-DL-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX9-DL-NEXT:    v_mad_u32_u24 v2, v4, v5, s0
-; GFX9-DL-NEXT:    v_add_u32_e32 v4, s0, v2
-; GFX9-DL-NEXT:    v_add3_u32 v2, v2, v3, v6
-; GFX9-DL-NEXT:    v_add3_u32 v1, v2, v1, v4
+; GFX9-DL-NEXT:    v_mul_u32_u24_sdwa v3, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_1 src1_sel:BYTE_1
+; GFX9-DL-NEXT:    v_dot4_u32_u8 v1, v1, v2, s0
+; GFX9-DL-NEXT:    v_add3_u32 v1, s1, v3, v1
 ; GFX9-DL-NEXT:    global_store_dword v0, v1, s[2:3]
 ; GFX9-DL-NEXT:    s_endpgm
 ;
 ; GFX10-DL-LABEL: udot4_multiuse_add1:
 ; GFX10-DL:       ; %bb.0: ; %entry
 ; GFX10-DL-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x24
 ; GFX10-DL-NEXT:    v_lshlrev_b32_e32 v0, 2, v0
 ; GFX10-DL-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0x34
 ; GFX10-DL-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX10-DL-NEXT:    s_clause 0x1
 ; GFX10-DL-NEXT:    global_load_dword v1, v0, s[4:5]
 ; GFX10-DL-NEXT:    global_load_dword v2, v0, s[6:7]
 ; GFX10-DL-NEXT:    s_load_dword s2, s[0:1], 0x0
-; GFX10-DL-NEXT:    s_waitcnt vmcnt(1)
-; GFX10-DL-NEXT:    v_bfe_u32 v0, v1, 8, 8
 ; GFX10-DL-NEXT:    s_waitcnt vmcnt(0)
-; GFX10-DL-NEXT:    v_bfe_u32 v3, v2, 8, 8
-; GFX10-DL-NEXT:    v_mul_u32_u24_sdwa v4, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:BYTE_0
+; GFX10-DL-NEXT:    v_mul_u32_u24_sdwa v0, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_1 src1_sel:BYTE_1
 ; GFX10-DL-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX10-DL-NEXT:    v_mad_u32_u24 v0, v0, v3, s2
-; GFX10-DL-NEXT:    v_mul_u32_u24_sdwa v3, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_2 src1_sel:BYTE_2
-; GFX10-DL-NEXT:    v_mul_u32_u24_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_3 src1_sel:BYTE_3
-; GFX10-DL-NEXT:    v_add_nc_u32_e32 v2, s2, v0
-; GFX10-DL-NEXT:    v_add3_u32 v0, v0, v4, v3
-; GFX10-DL-NEXT:    v_mov_b32_e32 v3, 0
-; GFX10-DL-NEXT:    v_add3_u32 v0, v0, v1, v2
-; GFX10-DL-NEXT:    global_store_dword v3, v0, s[0:1]
+; GFX10-DL-NEXT:    v_dot4_u32_u8 v1, v1, v2, s2
+; GFX10-DL-NEXT:    s_add_i32 s2, s2, s2
+; GFX10-DL-NEXT:    v_mov_b32_e32 v2, 0
+; GFX10-DL-NEXT:    v_add3_u32 v0, s2, v0, v1
+; GFX10-DL-NEXT:    global_store_dword v2, v0, s[0:1]
 ; GFX10-DL-NEXT:    s_endpgm
+;
+; GFX11-DL-LABEL: udot4_multiuse_add1:
+; GFX11-DL:       ; %bb.0: ; %entry
+; GFX11-DL-NEXT:    s_load_b128 s[4:7], s[0:1], 0x24
+; GFX11-DL-NEXT:    v_lshlrev_b32_e32 v0, 2, v0
+; GFX11-DL-NEXT:    s_load_b64 s[0:1], s[0:1], 0x34
+; GFX11-DL-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX11-DL-NEXT:    s_clause 0x1
+; GFX11-DL-NEXT:    global_load_b32 v1, v0, s[4:5]
+; GFX11-DL-NEXT:    global_load_b32 v0, v0, s[6:7]
+; GFX11-DL-NEXT:    s_load_b32 s2, s[0:1], 0x0
+; GFX11-DL-NEXT:    s_waitcnt vmcnt(1)
+; GFX11-DL-NEXT:    v_bfe_u32 v2, v1, 8, 8
+; GFX11-DL-NEXT:    s_waitcnt vmcnt(0)
+; GFX11-DL-NEXT:    v_bfe_u32 v3, v0, 8, 8
+; GFX11-DL-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX11-DL-NEXT:    v_dot4_u32_u8 v0, v1, v0, s2
+; GFX11-DL-NEXT:    v_mov_b32_e32 v1, 0
+; GFX11-DL-NEXT:    s_add_i32 s2, s2, s2
+; GFX11-DL-NEXT:    v_mul_u32_u24_e32 v2, v2, v3
+; GFX11-DL-NEXT:    s_delay_alu instid0(VALU_DEP_1)
+; GFX11-DL-NEXT:    v_add3_u32 v0, s2, v2, v0
+; GFX11-DL-NEXT:    global_store_b32 v1, v0, s[0:1]
+; GFX11-DL-NEXT:    s_nop 0
+; GFX11-DL-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
+; GFX11-DL-NEXT:    s_endpgm
                                                ptr addrspace(1) %src2,
                                                ptr addrspace(1) nocapture %dst) {
 entry:
   %idx = call i32 @llvm.amdgcn.workitem.id.x()
   %gep1 = getelementptr <4 x i8>, ptr addrspace(1) %src1, i32 %idx
   %vec1 = load <4 x i8>, ptr addrspace(1) %gep1
   %gep2 = getelementptr <4 x i8>, ptr addrspace(1) %src2, i32 %idx
   %vec2 = load <4 x i8>, ptr addrspace(1) %gep2

 ; GFX9-NODL-NEXT:    global_store_short v0, v1, s[2:3]
 ; GFX9-NODL-NEXT:    s_endpgm
 ;
 ; GFX9-DL-LABEL: notdot4_mixedtypes:
 ; GFX9-DL:       ; %bb.0: ; %entry
 ; GFX9-DL-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x24
 ; GFX9-DL-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x34
 ; GFX9-DL-NEXT:    v_lshlrev_b32_e32 v0, 2, v0
-; GFX9-DL-NEXT:    s_movk_i32 s0, 0xff
+; GFX9-DL-NEXT:    s_mov_b32 s0, 0xc0c0302
 ; GFX9-DL-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX9-DL-NEXT:    global_load_dword v1, v0, s[4:5]
 ; GFX9-DL-NEXT:    global_load_dword v2, v0, s[6:7]
 ; GFX9-DL-NEXT:    v_mov_b32_e32 v0, 0
 ; GFX9-DL-NEXT:    global_load_ushort v3, v0, s[2:3]
 ; GFX9-DL-NEXT:    s_waitcnt vmcnt(2)
 ; GFX9-DL-NEXT:    v_lshrrev_b32_e32 v6, 8, v1
 ; GFX9-DL-NEXT:    s_waitcnt vmcnt(1)
 ; GFX9-DL-NEXT:    v_lshrrev_b32_e32 v7, 8, v2
 ; GFX9-DL-NEXT:    v_and_b32_e32 v6, 0xff, v6
 ; GFX9-DL-NEXT:    v_and_b32_e32 v7, 0xff, v7
 ; GFX9-DL-NEXT:    v_bfe_i32 v4, v1, 0, 8
 ; GFX9-DL-NEXT:    v_bfe_i32 v5, v2, 0, 8
 ; GFX9-DL-NEXT:    s_waitcnt vmcnt(0)
 ; GFX9-DL-NEXT:    v_mad_legacy_u16 v3, v6, v7, v3
-; GFX9-DL-NEXT:    v_and_b32_sdwa v8, v1, s0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; GFX9-DL-NEXT:    v_and_b32_sdwa v9, v2, s0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
+; GFX9-DL-NEXT:    v_perm_b32 v2, v2, v2, s0
 ; GFX9-DL-NEXT:    v_mad_legacy_u16 v3, v4, v5, v3
-; GFX9-DL-NEXT:    v_lshrrev_b32_e32 v1, 24, v1
-; GFX9-DL-NEXT:    v_lshrrev_b32_e32 v2, 24, v2
-; GFX9-DL-NEXT:    v_mad_legacy_u16 v3, v8, v9, v3
-; GFX9-DL-NEXT:    v_mad_legacy_u16 v1, v1, v2, v3
+; GFX9-DL-NEXT:    v_perm_b32 v1, v1, v1, s0
+; GFX9-DL-NEXT:    v_dot4_u32_u8 v1, v1, v2, v3
 ; GFX9-DL-NEXT:    global_store_short v0, v1, s[2:3]
 ; GFX9-DL-NEXT:    s_endpgm
 ;
 ; GFX10-DL-LABEL: notdot4_mixedtypes:
 ; GFX10-DL:       ; %bb.0: ; %entry
 ; GFX10-DL-NEXT:    s_clause 0x1
 ; GFX10-DL-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x24
 ; GFX10-DL-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x34
 ; GFX10-DL-NEXT:    v_lshlrev_b32_e32 v0, 2, v0
-; GFX10-DL-NEXT:    v_mov_b32_e32 v7, 0xff
 ; GFX10-DL-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX10-DL-NEXT:    s_clause 0x1
 ; GFX10-DL-NEXT:    global_load_dword v1, v0, s[4:5]
 ; GFX10-DL-NEXT:    global_load_dword v2, v0, s[6:7]
 ; GFX10-DL-NEXT:    v_mov_b32_e32 v0, 0
 ; GFX10-DL-NEXT:    global_load_ushort v3, v0, s[2:3]
 ; GFX10-DL-NEXT:    s_waitcnt vmcnt(2)
 ; GFX10-DL-NEXT:    v_lshrrev_b32_e32 v4, 8, v1
 ; GFX10-DL-NEXT:    s_waitcnt vmcnt(1)
 ; GFX10-DL-NEXT:    v_lshrrev_b32_e32 v5, 8, v2
 ; GFX10-DL-NEXT:    v_bfe_i32 v6, v1, 0, 8
-; GFX10-DL-NEXT:    v_bfe_i32 v8, v2, 0, 8
+; GFX10-DL-NEXT:    v_bfe_i32 v7, v2, 0, 8
+; GFX10-DL-NEXT:    v_perm_b32 v2, v2, v2, 0xc0c0302
 ; GFX10-DL-NEXT:    v_and_b32_e32 v4, 0xff, v4
 ; GFX10-DL-NEXT:    v_and_b32_e32 v5, 0xff, v5
+; GFX10-DL-NEXT:    v_perm_b32 v1, v1, v1, 0xc0c0302
 ; GFX10-DL-NEXT:    s_waitcnt vmcnt(0)
 ; GFX10-DL-NEXT:    v_mad_u16 v3, v4, v5, v3
-; GFX10-DL-NEXT:    v_and_b32_sdwa v4, v1, v7 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; GFX10-DL-NEXT:    v_and_b32_sdwa v5, v2, v7 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_1 src1_sel:DWORD
-; GFX10-DL-NEXT:    v_lshrrev_b32_e32 v1, 24, v1
-; GFX10-DL-NEXT:    v_lshrrev_b32_e32 v2, 24, v2
-; GFX10-DL-NEXT:    v_mad_u16 v3, v6, v8, v3
-; GFX10-DL-NEXT:    v_mad_u16 v3, v4, v5, v3
-; GFX10-DL-NEXT:    v_mad_u16 v1, v1, v2, v3
+; GFX10-DL-NEXT:    v_mad_u16 v3, v6, v7, v3
+; GFX10-DL-NEXT:    v_and_b32_e32 v3, 0xffff, v3
+; GFX10-DL-NEXT:    v_dot4_u32_u8 v1, v1, v2, v3
 ; GFX10-DL-NEXT:    global_store_short v0, v1, s[2:3]
 ; GFX10-DL-NEXT:    s_endpgm
+;
+; GFX11-DL-LABEL: notdot4_mixedtypes:
+; GFX11-DL:       ; %bb.0: ; %entry
+; GFX11-DL-NEXT:    s_clause 0x1
+; GFX11-DL-NEXT:    s_load_b128 s[4:7], s[0:1], 0x24
+; GFX11-DL-NEXT:    s_load_b64 s[0:1], s[0:1], 0x34
+; GFX11-DL-NEXT:    v_lshlrev_b32_e32 v0, 2, v0
+; GFX11-DL-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX11-DL-NEXT:    s_clause 0x1
+; GFX11-DL-NEXT:    global_load_b32 v1, v0, s[4:5]
+; GFX11-DL-NEXT:    global_load_b32 v0, v0, s[6:7]
+; GFX11-DL-NEXT:    s_waitcnt vmcnt(1)
+; GFX11-DL-NEXT:    v_lshrrev_b32_e32 v4, 8, v1
+; GFX11-DL-NEXT:    s_waitcnt vmcnt(0)
+; GFX11-DL-NEXT:    v_lshrrev_b32_e32 v5, 8, v0
+; GFX11-DL-NEXT:    v_mov_b32_e32 v2, 0
+; GFX11-DL-NEXT:    v_bfe_i32 v6, v1, 0, 8
+; GFX11-DL-NEXT:    v_bfe_i32 v7, v0, 0, 8
+; GFX11-DL-NEXT:    v_and_b32_e32 v4, 0xff, v4
+; GFX11-DL-NEXT:    v_and_b32_e32 v5, 0xff, v5
+; GFX11-DL-NEXT:    global_load_u16 v3, v2, s[0:1]
+; GFX11-DL-NEXT:    v_perm_b32 v0, v0, v0, 0xc0c0302
+; GFX11-DL-NEXT:    v_perm_b32 v1, v1, v1, 0xc0c0302
+; GFX11-DL-NEXT:    s_waitcnt vmcnt(0)
+; GFX11-DL-NEXT:    v_mad_u16 v3, v4, v5, v3
+; GFX11-DL-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-DL-NEXT:    v_mad_u16 v3, v6, v7, v3
+; GFX11-DL-NEXT:    v_and_b32_e32 v3, 0xffff, v3
+; GFX11-DL-NEXT:    s_delay_alu instid0(VALU_DEP_1)
+; GFX11-DL-NEXT:    v_dot4_u32_u8 v0, v1, v0, v3
+; GFX11-DL-NEXT:    global_store_b16 v2, v0, s[0:1]
+; GFX11-DL-NEXT:    s_nop 0
+; GFX11-DL-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
+; GFX11-DL-NEXT:    s_endpgm
                                               ptr addrspace(1) %src2,
                                               ptr addrspace(1) nocapture %dst) {
 entry:
   %idx = call i32 @llvm.amdgcn.workitem.id.x()
   %gep1 = getelementptr <4 x i8>, ptr addrspace(1) %src1, i32 %idx
   %vec1 = load <4 x i8>, ptr addrspace(1) %gep1
   %gep2 = getelementptr <4 x i8>, ptr addrspace(1) %src2, i32 %idx
   %vec2 = load <4 x i8>, ptr addrspace(1) %gep2
entry:
   %add2 = add i16 %add1, %mul1
   %add3 = add i16 %add2, %mul3
   %add4 = add i16 %add3, %mul4
 
   store i16 %add4, ptr addrspace(1) %dst, align 2
   ret void
 }
 
-; TODO: cleanup s_lshr_b32 and support this pattern.
+; TODO: cleanup s_lshr_b32
 define amdgpu_kernel void @udot4_acc32_vecMul(ptr addrspace(1) %src1,
 ; GFX7-LABEL: udot4_acc32_vecMul:
 ; GFX7:       ; %bb.0: ; %entry
 ; GFX7-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x9
 ; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xd
 ; GFX7-NEXT:    s_mov_b32 s3, 0xf000
 ; GFX7-NEXT:    s_mov_b32 s10, 0
 ; GFX7-NEXT:    s_mov_b32 s11, s3

 ; GFX9-DL-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x24
 ; GFX9-DL-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x34
 ; GFX9-DL-NEXT:    v_lshlrev_b32_e32 v0, 2, v0
 ; GFX9-DL-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX9-DL-NEXT:    global_load_dword v1, v0, s[4:5]
 ; GFX9-DL-NEXT:    global_load_dword v2, v0, s[6:7]
 ; GFX9-DL-NEXT:    s_load_dword s0, s[2:3], 0x0
 ; GFX9-DL-NEXT:    v_mov_b32_e32 v0, 0
-; GFX9-DL-NEXT:    s_waitcnt vmcnt(0)
-; GFX9-DL-NEXT:    v_mul_u32_u24_sdwa v3, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:BYTE_0
-; GFX9-DL-NEXT:    v_mul_u32_u24_sdwa v4, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_1 src1_sel:BYTE_1
-; GFX9-DL-NEXT:    v_mul_u32_u24_sdwa v5, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_2 src1_sel:BYTE_2
-; GFX9-DL-NEXT:    v_mul_u32_u24_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_3 src1_sel:BYTE_3
-; GFX9-DL-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX9-DL-NEXT:    v_add3_u32 v2, v3, s0, v4
-; GFX9-DL-NEXT:    v_add3_u32 v1, v2, v5, v1
+; GFX9-DL-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX9-DL-NEXT:    v_dot4_u32_u8 v1, v1, v2, s0
 ; GFX9-DL-NEXT:    global_store_dword v0, v1, s[2:3]
 ; GFX9-DL-NEXT:    s_endpgm
 ;
 ; GFX10-DL-LABEL: udot4_acc32_vecMul:
 ; GFX10-DL:       ; %bb.0: ; %entry
 ; GFX10-DL-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x24
 ; GFX10-DL-NEXT:    v_lshlrev_b32_e32 v0, 2, v0
 ; GFX10-DL-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0x34
 ; GFX10-DL-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX10-DL-NEXT:    s_clause 0x1
 ; GFX10-DL-NEXT:    global_load_dword v1, v0, s[4:5]
 ; GFX10-DL-NEXT:    global_load_dword v2, v0, s[6:7]
-; GFX10-DL-NEXT:    v_mov_b32_e32 v0, 0xffff
+; GFX10-DL-NEXT:    v_mov_b32_e32 v0, 0
 ; GFX10-DL-NEXT:    s_load_dword s2, s[0:1], 0x0
-; GFX10-DL-NEXT:    s_waitcnt vmcnt(1)
-; GFX10-DL-NEXT:    v_and_b32_sdwa v3, v0, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_1
-; GFX10-DL-NEXT:    s_waitcnt vmcnt(0)
-; GFX10-DL-NEXT:    v_and_b32_sdwa v0, v0, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:DWORD src1_sel:BYTE_1
-; GFX10-DL-NEXT:    v_mul_u32_u24_sdwa v4, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:BYTE_0
-; GFX10-DL-NEXT:    v_mul_u32_u24_e32 v0, v3, v0
-; GFX10-DL-NEXT:    v_mul_u32_u24_sdwa v3, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_2 src1_sel:BYTE_2
-; GFX10-DL-NEXT:    v_mul_u32_u24_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_3 src1_sel:BYTE_3
-; GFX10-DL-NEXT:    v_mov_b32_e32 v2, 0
-; GFX10-DL-NEXT:    s_waitcnt lgkmcnt(0)
-; GFX10-DL-NEXT:    v_add3_u32 v0, v4, s2, v0
-; GFX10-DL-NEXT:    v_add3_u32 v0, v0, v3, v1
-; GFX10-DL-NEXT:    global_store_dword v2, v0, s[0:1]
+; GFX10-DL-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX10-DL-NEXT:    v_dot4_u32_u8 v1, v1, v2, s2
+; GFX10-DL-NEXT:    global_store_dword v0, v1, s[0:1]
 ; GFX10-DL-NEXT:    s_endpgm
+;
+; GFX11-DL-LABEL: udot4_acc32_vecMul:
+; GFX11-DL:       ; %bb.0: ; %entry
+; GFX11-DL-NEXT:    s_load_b128 s[4:7], s[0:1], 0x24
+; GFX11-DL-NEXT:    v_lshlrev_b32_e32 v0, 2, v0
+; GFX11-DL-NEXT:    s_load_b64 s[0:1], s[0:1], 0x34
+; GFX11-DL-NEXT:    v_mov_b32_e32 v2, 0
+; GFX11-DL-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX11-DL-NEXT:    s_clause 0x1
+; GFX11-DL-NEXT:    global_load_b32 v1, v0, s[4:5]
+; GFX11-DL-NEXT:    global_load_b32 v0, v0, s[6:7]
+; GFX11-DL-NEXT:    s_load_b32 s2, s[0:1], 0x0
+; GFX11-DL-NEXT:    s_waitcnt vmcnt(0) lgkmcnt(0)
+; GFX11-DL-NEXT:    v_dot4_u32_u8 v0, v1, v0, s2
+; GFX11-DL-NEXT:    global_store_b32 v2, v0, s[0:1]
+; GFX11-DL-NEXT:    s_nop 0
+; GFX11-DL-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
+; GFX11-DL-NEXT:    s_endpgm
                                               ptr addrspace(1) %src2,
                                               ptr addrspace(1) nocapture %dst) {
 entry:
   %idx = call i32 @llvm.amdgcn.workitem.id.x()
   %gep1 = getelementptr <4 x i8>, ptr addrspace(1) %src1, i32 %idx
   %vec1 = load <4 x i8>, ptr addrspace(1) %gep1
   %gep2 = getelementptr <4 x i8>, ptr addrspace(1) %src2, i32 %idx
   %vec2 = load <4 x i8>, ptr addrspace(1) %gep2

 ; GFX10-DL-NEXT:    v_add_nc_u16 v3, v4, v3
 ; GFX10-DL-NEXT:    v_pk_mul_lo_u16 v1, v1, v2
 ; GFX10-DL-NEXT:    v_add_nc_u16 v2, v3, v5
 ; GFX10-DL-NEXT:    v_lshrrev_b32_e32 v3, 16, v1
 ; GFX10-DL-NEXT:    v_add_nc_u16 v1, v2, v1
 ; GFX10-DL-NEXT:    v_add_nc_u16 v1, v1, v3
 ; GFX10-DL-NEXT:    global_store_short v0, v1, s[0:1]
 ; GFX10-DL-NEXT:    s_endpgm
+;
+; GFX11-DL-LABEL: udot4_acc16_vecMul:
+; GFX11-DL:       ; %bb.0: ; %entry
+; GFX11-DL-NEXT:    s_load_b128 s[4:7], s[0:1], 0x24
+; GFX11-DL-NEXT:    v_lshlrev_b32_e32 v0, 2, v0
+; GFX11-DL-NEXT:    s_load_b64 s[0:1], s[0:1], 0x34
+; GFX11-DL-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX11-DL-NEXT:    s_clause 0x1
+; GFX11-DL-NEXT:    global_load_b32 v1, v0, s[4:5]
+; GFX11-DL-NEXT:    global_load_b32 v0, v0, s[6:7]
+; GFX11-DL-NEXT:    s_waitcnt vmcnt(1)
+; GFX11-DL-NEXT:    v_dual_mov_b32 v2, 0 :: v_dual_and_b32 v7, 0xff, v1
+; GFX11-DL-NEXT:    global_load_u16 v3, v2, s[0:1]
+; GFX11-DL-NEXT:    v_lshrrev_b16 v4, 8, v1
+; GFX11-DL-NEXT:    s_waitcnt vmcnt(1)
+; GFX11-DL-NEXT:    v_lshrrev_b16 v5, 8, v0
+; GFX11-DL-NEXT:    v_and_b32_e32 v6, 0xff, v0
+; GFX11-DL-NEXT:    v_lshrrev_b32_e32 v8, 16, v1
+; GFX11-DL-NEXT:    v_lshrrev_b32_e32 v9, 16, v0
+; GFX11-DL-NEXT:    v_perm_b32 v4, v4, v7, 0x5040100
+; GFX11-DL-NEXT:    v_lshrrev_b32_e32 v1, 24, v1
+; GFX11-DL-NEXT:    v_perm_b32 v5, v5, v6, 0x5040100
+; GFX11-DL-NEXT:    v_lshrrev_b32_e32 v0, 24, v0
+; GFX11-DL-NEXT:    v_and_b32_e32 v6, 0xff, v9
+; GFX11-DL-NEXT:    v_and_b32_e32 v7, 0xff, v8
+; GFX11-DL-NEXT:    s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_3)
+; GFX11-DL-NEXT:    v_pk_mul_lo_u16 v4, v4, v5
+; GFX11-DL-NEXT:    v_perm_b32 v0, v0, v6, 0x5040100
+; GFX11-DL-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3)
+; GFX11-DL-NEXT:    v_perm_b32 v1, v1, v7, 0x5040100
+; GFX11-DL-NEXT:    v_lshrrev_b32_e32 v5, 16, v4
+; GFX11-DL-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_2) | instid1(VALU_DEP_1)
+; GFX11-DL-NEXT:    v_pk_mul_lo_u16 v0, v1, v0
+; GFX11-DL-NEXT:    s_waitcnt vmcnt(0)
+; GFX11-DL-NEXT:    v_add_nc_u16 v3, v4, v3
+; GFX11-DL-NEXT:    v_add_nc_u16 v1, v3, v5
+; GFX11-DL-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX11-DL-NEXT:    v_lshrrev_b32_e32 v3, 16, v0
+; GFX11-DL-NEXT:    v_add_nc_u16 v0, v1, v0
+; GFX11-DL-NEXT:    s_delay_alu instid0(VALU_DEP_1)
+; GFX11-DL-NEXT:    v_add_nc_u16 v0, v0, v3
+; GFX11-DL-NEXT:    global_store_b16 v2, v0, s[0:1]
+; GFX11-DL-NEXT:    s_nop 0
+; GFX11-DL-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
+; GFX11-DL-NEXT:    s_endpgm
                                               ptr addrspace(1) %src2,
                                               ptr addrspace(1) nocapture %dst) {
 entry:
   %idx = call i32 @llvm.amdgcn.workitem.id.x()
   %gep1 = getelementptr <4 x i8>, ptr addrspace(1) %src1, i32 %idx
   %vec1 = load <4 x i8>, ptr addrspace(1) %gep1
   %gep2 = getelementptr <4 x i8>, ptr addrspace(1) %src2, i32 %idx
   %vec2 = load <4 x i8>, ptr addrspace(1) %gep2

 ; GFX10-DL-NEXT:    v_lshrrev_b32_e32 v2, 8, v4
 ; GFX10-DL-NEXT:    v_or_b32_sdwa v5, v6, v5 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0 src1_sel:DWORD
 ; GFX10-DL-NEXT:    v_lshrrev_b32_e32 v5, 8, v5
 ; GFX10-DL-NEXT:    v_add_nc_u16 v1, v1, v5
 ; GFX10-DL-NEXT:    v_mad_u16 v1, v7, v8, v1
 ; GFX10-DL-NEXT:    v_add_nc_u16 v1, v1, v2
 ; GFX10-DL-NEXT:    global_store_byte v0, v1, s[0:1]
 ; GFX10-DL-NEXT:    s_endpgm
+;
+; GFX11-DL-LABEL: udot4_acc8_vecMul:
+; GFX11-DL:       ; %bb.0: ; %entry
+; GFX11-DL-NEXT:    s_load_b128 s[4:7], s[0:1], 0x24
+; GFX11-DL-NEXT:    v_lshlrev_b32_e32 v0, 2, v0
+; GFX11-DL-NEXT:    s_load_b64 s[0:1], s[0:1], 0x34
+; GFX11-DL-NEXT:    v_mov_b32_e32 v2, 0
+; GFX11-DL-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX11-DL-NEXT:    s_clause 0x1
+; GFX11-DL-NEXT:    global_load_b32 v1, v0, s[4:5]
+; GFX11-DL-NEXT:    global_load_b32 v0, v0, s[6:7]
+; GFX11-DL-NEXT:    global_load_u8 v3, v2, s[0:1]
+; GFX11-DL-NEXT:    s_waitcnt vmcnt(2)
+; GFX11-DL-NEXT:    v_lshrrev_b32_e32 v4, 16, v1
+; GFX11-DL-NEXT:    v_lshrrev_b32_e32 v5, 24, v1
+; GFX11-DL-NEXT:    s_waitcnt vmcnt(1)
+; GFX11-DL-NEXT:    v_lshrrev_b32_e32 v6, 24, v0
+; GFX11-DL-NEXT:    v_lshrrev_b32_e32 v7, 16, v0
+; GFX11-DL-NEXT:    v_lshrrev_b16 v8, 8, v1
+; GFX11-DL-NEXT:    v_lshrrev_b16 v9, 8, v0
+; GFX11-DL-NEXT:    s_waitcnt vmcnt(0)
+; GFX11-DL-NEXT:    v_mad_u16 v0, v1, v0, v3
+; GFX11-DL-NEXT:    v_mul_lo_u16 v5, v5, v6
+; GFX11-DL-NEXT:    v_mul_lo_u16 v6, v4, v7
+; GFX11-DL-NEXT:    v_mul_lo_u16 v8, v8, v9
+; GFX11-DL-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3)
+; GFX11-DL-NEXT:    v_lshlrev_b16 v5, 8, v5
+; GFX11-DL-NEXT:    v_and_b32_e32 v6, 0xff, v6
+; GFX11-DL-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3)
+; GFX11-DL-NEXT:    v_lshlrev_b16 v8, 8, v8
+; GFX11-DL-NEXT:    v_lshrrev_b32_e32 v1, 8, v5
+; GFX11-DL-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3)
+; GFX11-DL-NEXT:    v_or_b32_e32 v6, v6, v5
+; GFX11-DL-NEXT:    v_and_b32_e32 v8, 0xffff, v8
+; GFX11-DL-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-DL-NEXT:    v_lshlrev_b32_e32 v6, 16, v6
+; GFX11-DL-NEXT:    v_or_b32_e32 v6, v8, v6
+; GFX11-DL-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-DL-NEXT:    v_lshrrev_b32_e32 v6, 8, v6
+; GFX11-DL-NEXT:    v_add_nc_u16 v0, v0, v6
+; GFX11-DL-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-DL-NEXT:    v_mad_u16 v0, v4, v7, v0
+; GFX11-DL-NEXT:    v_add_nc_u16 v0, v0, v1
+; GFX11-DL-NEXT:    global_store_b8 v2, v0, s[0:1]
+; GFX11-DL-NEXT:    s_nop 0
+; GFX11-DL-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
+; GFX11-DL-NEXT:    s_endpgm
                                              ptr addrspace(1) %src2,
                                              ptr addrspace(1) nocapture %dst) {
 entry:
   %idx = call i32 @llvm.amdgcn.workitem.id.x()
   %gep1 = getelementptr <4 x i8>, ptr addrspace(1) %src1, i32 %idx
   %vec1 = load <4 x i8>, ptr addrspace(1) %gep1
   %gep2 = getelementptr <4 x i8>, ptr addrspace(1) %src2, i32 %idx
   %vec2 = load <4 x i8>, ptr addrspace(1) %gep2
entry:
   %add2 = add i8 %add1, %mul1
   %add3 = add i8 %add2, %mul2
   %add4 = add i8 %add3, %mul3
 
   store i8 %add4, ptr addrspace(1) %dst, align 4
   ret void
 }
 
+define amdgpu_kernel void @idot4_acc32_2ele(ptr addrspace(1) %src1,
+; GFX7-LABEL: idot4_acc32_2ele:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x9
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xd
+; GFX7-NEXT:    s_mov_b32 s3, 0xf000
+; GFX7-NEXT:    s_mov_b32 s10, 0
+; GFX7-NEXT:    s_mov_b32 s11, s3
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    s_mov_b64 s[8:9], s[4:5]
+; GFX7-NEXT:    v_lshlrev_b32_e32 v0, 2, v0
+; GFX7-NEXT:    v_mov_b32_e32 v1, 0
+; GFX7-NEXT:    buffer_load_dword v2, v[0:1], s[8:11], 0 addr64
+; GFX7-NEXT:    s_mov_b64 s[8:9], s[6:7]
+; GFX7-NEXT:    buffer_load_dword v0, v[0:1], s[8:11], 0 addr64
+; GFX7-NEXT:    s_load_dword s4, s[0:1], 0x0
+; GFX7-NEXT:    s_mov_b32 s2, -1
+; GFX7-NEXT:    s_waitcnt vmcnt(1)
+; GFX7-NEXT:    v_and_b32_e32 v1, 0xff, v2
+; GFX7-NEXT:    v_bfe_u32 v2, v2, 8, 8
+; GFX7-NEXT:    s_waitcnt vmcnt(0)
+; GFX7-NEXT:    v_and_b32_e32 v3, 0xff, v0
+; GFX7-NEXT:    v_bfe_u32 v0, v0, 8, 8
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mad_u32_u24 v1, v1, v3, s4
+; GFX7-NEXT:    v_mad_u32_u24 v0, v2, v0, v1
+; GFX7-NEXT:    buffer_store_dword v0, off, s[0:3], 0
+; GFX7-NEXT:    s_endpgm
+;
+; GFX8-LABEL: idot4_acc32_2ele:
+; GFX8:       ; %bb.0: ; %entry
+; GFX8-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x24
+; GFX8-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0x34
+; GFX8-NEXT:    v_lshlrev_b32_e32 v2, 2, v0
+; GFX8-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX8-NEXT:    v_mov_b32_e32 v1, s5
+; GFX8-NEXT:    v_add_u32_e32 v0, vcc, s4, v2
+; GFX8-NEXT:    v_addc_u32_e32 v1, vcc, 0, v1, vcc
+; GFX8-NEXT:    flat_load_dword v3, v[0:1]
+; GFX8-NEXT:    v_mov_b32_e32 v1, s7
+; GFX8-NEXT:    v_add_u32_e32 v0, vcc, s6, v2
+; GFX8-NEXT:    v_addc_u32_e32 v1, vcc, 0, v1, vcc
+; GFX8-NEXT:    flat_load_dword v0, v[0:1]
+; GFX8-NEXT:    s_load_dword s2, s[0:1], 0x0
+; GFX8-NEXT:    s_waitcnt vmcnt(1)
+; GFX8-NEXT:    v_and_b32_e32 v1, 0xff, v3
+; GFX8-NEXT:    v_bfe_u32 v3, v3, 8, 8
+; GFX8-NEXT:    s_waitcnt vmcnt(0)
+; GFX8-NEXT:    v_and_b32_e32 v2, 0xff, v0
+; GFX8-NEXT:    v_bfe_u32 v0, v0, 8, 8
+; GFX8-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX8-NEXT:    v_mad_u32_u24 v1, v1, v2, s2
+; GFX8-NEXT:    v_mad_u32_u24 v2, v3, v0, v1
+; GFX8-NEXT:    v_mov_b32_e32 v0, s0
+; GFX8-NEXT:    v_mov_b32_e32 v1, s1
+; GFX8-NEXT:    flat_store_dword v[0:1], v2
+; GFX8-NEXT:    s_endpgm
+;
+; GFX9-NODL-LABEL: idot4_acc32_2ele:
+; GFX9-NODL:       ; %bb.0: ; %entry
+; GFX9-NODL-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x24
+; GFX9-NODL-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x34
+; GFX9-NODL-NEXT:    v_lshlrev_b32_e32 v0, 2, v0
+; GFX9-NODL-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX9-NODL-NEXT:    global_load_dword v1, v0, s[4:5]
+; GFX9-NODL-NEXT:    global_load_dword v2, v0, s[6:7]
+; GFX9-NODL-NEXT:    s_load_dword s0, s[2:3], 0x0
+; GFX9-NODL-NEXT:    v_mov_b32_e32 v0, 0
+; GFX9-NODL-NEXT:    s_waitcnt vmcnt(0)
+; GFX9-NODL-NEXT:    v_mul_u32_u24_sdwa v3, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:BYTE_0
+; GFX9-NODL-NEXT:    v_mul_u32_u24_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_1 src1_sel:BYTE_1
+; GFX9-NODL-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX9-NODL-NEXT:    v_add3_u32 v1, v3, s0, v1
+; GFX9-NODL-NEXT:    global_store_dword v0, v1, s[2:3]
+; GFX9-NODL-NEXT:    s_endpgm
+;
+; GFX9-DL-LABEL: idot4_acc32_2ele:
+; GFX9-DL:       ; %bb.0: ; %entry
+; GFX9-DL-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x24
+; GFX9-DL-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x34
+; GFX9-DL-NEXT:    v_lshlrev_b32_e32 v0, 2, v0
+; GFX9-DL-NEXT:    s_mov_b32 s1, 0xc0c0100
+; GFX9-DL-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX9-DL-NEXT:    global_load_dword v1, v0, s[6:7]
+; GFX9-DL-NEXT:    global_load_dword v2, v0, s[4:5]
+; GFX9-DL-NEXT:    s_load_dword s0, s[2:3], 0x0
+; GFX9-DL-NEXT:    v_mov_b32_e32 v0, 0
+; GFX9-DL-NEXT:    s_waitcnt vmcnt(1)
+; GFX9-DL-NEXT:    v_perm_b32 v1, v1, v1, s1
+; GFX9-DL-NEXT:    s_waitcnt vmcnt(0)
+; GFX9-DL-NEXT:    v_perm_b32 v2, v2, v2, s1
+; GFX9-DL-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX9-DL-NEXT:    v_dot4_u32_u8 v1, v2, v1, s0
+; GFX9-DL-NEXT:    global_store_dword v0, v1, s[2:3]
+; GFX9-DL-NEXT:    s_endpgm
+;
+; GFX10-DL-LABEL: idot4_acc32_2ele:
+; GFX10-DL:       ; %bb.0: ; %entry
+; GFX10-DL-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x24
+; GFX10-DL-NEXT:    v_lshlrev_b32_e32 v0, 2, v0
+; GFX10-DL-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0x34
+; GFX10-DL-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-DL-NEXT:    s_clause 0x1
+; GFX10-DL-NEXT:    global_load_dword v1, v0, s[6:7]
+; GFX10-DL-NEXT:    global_load_dword v2, v0, s[4:5]
+; GFX10-DL-NEXT:    s_load_dword s2, s[0:1], 0x0
+; GFX10-DL-NEXT:    s_waitcnt vmcnt(1)
+; GFX10-DL-NEXT:    v_perm_b32 v0, v1, v1, 0xc0c0100
+; GFX10-DL-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-DL-NEXT:    v_perm_b32 v1, v2, v2, 0xc0c0100
+; GFX10-DL-NEXT:    v_mov_b32_e32 v2, 0
+; GFX10-DL-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-DL-NEXT:    v_dot4_u32_u8 v0, v1, v0, s2
+; GFX10-DL-NEXT:    global_store_dword v2, v0, s[0:1]
+; GFX10-DL-NEXT:    s_endpgm
+;
+; GFX11-DL-LABEL: idot4_acc32_2ele:
+; GFX11-DL:       ; %bb.0: ; %entry
+; GFX11-DL-NEXT:    s_load_b128 s[4:7], s[0:1], 0x24
+; GFX11-DL-NEXT:    v_lshlrev_b32_e32 v0, 2, v0
+; GFX11-DL-NEXT:    s_load_b64 s[0:1], s[0:1], 0x34
+; GFX11-DL-NEXT:    v_mov_b32_e32 v2, 0
+; GFX11-DL-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX11-DL-NEXT:    s_clause 0x1
+; GFX11-DL-NEXT:    global_load_b32 v1, v0, s[6:7]
+; GFX11-DL-NEXT:    global_load_b32 v0, v0, s[4:5]
+; GFX11-DL-NEXT:    s_load_b32 s2, s[0:1], 0x0
+; GFX11-DL-NEXT:    s_waitcnt vmcnt(1)
+; GFX11-DL-NEXT:    v_perm_b32 v1, v1, v1, 0xc0c0100
+; GFX11-DL-NEXT:    s_waitcnt vmcnt(0)
+; GFX11-DL-NEXT:    v_perm_b32 v0, v0, v0, 0xc0c0100
+; GFX11-DL-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX11-DL-NEXT:    s_delay_alu instid0(VALU_DEP_1)
+; GFX11-DL-NEXT:    v_dot4_u32_u8 v0, v0, v1, s2
+; GFX11-DL-NEXT:    global_store_b32 v2, v0, s[0:1]
+; GFX11-DL-NEXT:    s_nop 0
+; GFX11-DL-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
+; GFX11-DL-NEXT:    s_endpgm
+                                       ptr addrspace(1) %src2,
+                                       ptr addrspace(1) nocapture %dst) {
+entry:
+  %idx = call i32 @llvm.amdgcn.workitem.id.x()
+  %gep1 = getelementptr <4 x i8>, ptr addrspace(1) %src1, i32 %idx
+  %vec1 = load <4 x i8>, ptr addrspace(1) %gep1
+  %gep2 = getelementptr <4 x i8>, ptr addrspace(1) %src2, i32 %idx
+  %vec2 = load <4 x i8>, ptr addrspace(1) %gep2
+
+  %v1e0 = extractelement <4 x i8> %vec1, i64 0
+  %cv1e0 = zext i8 %v1e0 to i32
+  %v2e0 = extractelement <4 x i8> %vec2, i64 0
+  %cv2e0 = zext i8 %v2e0 to i32
+  %mul1 = mul nuw nsw i32 %cv1e0, %cv2e0
+
+  %v1e1 = extractelement <4 x i8> %vec1, i64 1
+  %cv1e1 = zext i8 %v1e1 to i32
+  %v2e1 = extractelement <4 x i8> %vec2, i64 1
+  %cv2e1 = zext i8 %v2e1 to i32
+  %mul2 = mul nuw nsw i32 %cv1e1, %cv2e1
+
+  %acc = load i32, ptr addrspace(1) %dst, align 4
+  %add1 = add i32 %mul1, %acc
+  %add2 = add i32 %add1, %mul2
+  store i32 %add2, ptr addrspace(1) %dst, align 4
+  ret void
+}
+
+define amdgpu_kernel void @idot4_acc32_3ele(ptr addrspace(1) %src1,
+; GFX7-LABEL: idot4_acc32_3ele:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x9
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xd
+; GFX7-NEXT:    s_mov_b32 s3, 0xf000
+; GFX7-NEXT:    s_mov_b32 s10, 0
+; GFX7-NEXT:    s_mov_b32 s11, s3
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    s_mov_b64 s[8:9], s[4:5]
+; GFX7-NEXT:    v_lshlrev_b32_e32 v0, 2, v0
+; GFX7-NEXT:    v_mov_b32_e32 v1, 0
+; GFX7-NEXT:    buffer_load_dword v2, v[0:1], s[8:11], 0 addr64
+; GFX7-NEXT:    s_mov_b64 s[8:9], s[6:7]
+; GFX7-NEXT:    buffer_load_dword v0, v[0:1], s[8:11], 0 addr64
+; GFX7-NEXT:    s_load_dword s4, s[0:1], 0x0
+; GFX7-NEXT:    s_mov_b32 s2, -1
+; GFX7-NEXT:    s_waitcnt vmcnt(1)
+; GFX7-NEXT:    v_and_b32_e32 v1, 0xff, v2
+; GFX7-NEXT:    v_bfe_u32 v3, v2, 8, 8
+; GFX7-NEXT:    s_waitcnt vmcnt(0)
+; GFX7-NEXT:    v_and_b32_e32 v4, 0xff, v0
+; GFX7-NEXT:    v_bfe_u32 v5, v0, 8, 8
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mad_u32_u24 v1, v1, v4, s4
+; GFX7-NEXT:    v_bfe_u32 v2, v2, 16, 8
+; GFX7-NEXT:    v_bfe_u32 v0, v0, 16, 8
+; GFX7-NEXT:    v_mad_u32_u24 v1, v3, v5, v1
+; GFX7-NEXT:    v_mad_u32_u24 v0, v2, v0, v1
+; GFX7-NEXT:    buffer_store_dword v0, off, s[0:3], 0
+; GFX7-NEXT:    s_endpgm
+;
+; GFX8-LABEL: idot4_acc32_3ele:
+; GFX8:       ; %bb.0: ; %entry
+; GFX8-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x24
+; GFX8-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0x34
+; GFX8-NEXT:    v_lshlrev_b32_e32 v2, 2, v0
+; GFX8-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX8-NEXT:    v_mov_b32_e32 v1, s5
+; GFX8-NEXT:    v_add_u32_e32 v0, vcc, s4, v2
+; GFX8-NEXT:    v_addc_u32_e32 v1, vcc, 0, v1, vcc
+; GFX8-NEXT:    flat_load_dword v3, v[0:1]
+; GFX8-NEXT:    v_mov_b32_e32 v1, s7
+; GFX8-NEXT:    v_add_u32_e32 v0, vcc, s6, v2
+; GFX8-NEXT:    v_addc_u32_e32 v1, vcc, 0, v1, vcc
+; GFX8-NEXT:    flat_load_dword v0, v[0:1]
+; GFX8-NEXT:    s_load_dword s2, s[0:1], 0x0
+; GFX8-NEXT:    s_waitcnt vmcnt(1)
+; GFX8-NEXT:    v_and_b32_e32 v1, 0xff, v3
+; GFX8-NEXT:    v_bfe_u32 v4, v3, 8, 8
+; GFX8-NEXT:    v_bfe_u32 v3, v3, 16, 8
+; GFX8-NEXT:    s_waitcnt vmcnt(0)
+; GFX8-NEXT:    v_and_b32_e32 v2, 0xff, v0
+; GFX8-NEXT:    v_bfe_u32 v5, v0, 8, 8
+; GFX8-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX8-NEXT:    v_mad_u32_u24 v1, v1, v2, s2
+; GFX8-NEXT:    v_bfe_u32 v0, v0, 16, 8
+; GFX8-NEXT:    v_mad_u32_u24 v1, v4, v5, v1
+; GFX8-NEXT:    v_mad_u32_u24 v2, v3, v0, v1
+; GFX8-NEXT:    v_mov_b32_e32 v0, s0
+; GFX8-NEXT:    v_mov_b32_e32 v1, s1
+; GFX8-NEXT:    flat_store_dword v[0:1], v2
+; GFX8-NEXT:    s_endpgm
+;
+; GFX9-NODL-LABEL: idot4_acc32_3ele:
+; GFX9-NODL:       ; %bb.0: ; %entry
+; GFX9-NODL-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x24
+; GFX9-NODL-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x34
+; GFX9-NODL-NEXT:    v_lshlrev_b32_e32 v0, 2, v0
+; GFX9-NODL-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX9-NODL-NEXT:    global_load_dword v1, v0, s[4:5]
+; GFX9-NODL-NEXT:    global_load_dword v2, v0, s[6:7]
+; GFX9-NODL-NEXT:    s_load_dword s0, s[2:3], 0x0
+; GFX9-NODL-NEXT:    v_mov_b32_e32 v0, 0
+; GFX9-NODL-NEXT:    s_waitcnt vmcnt(1)
+; GFX9-NODL-NEXT:    v_and_b32_e32 v3, 0xff, v1
+; GFX9-NODL-NEXT:    s_waitcnt vmcnt(0)
+; GFX9-NODL-NEXT:    v_and_b32_e32 v4, 0xff, v2
+; GFX9-NODL-NEXT:    v_mul_u32_u24_sdwa v5, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_1 src1_sel:BYTE_1
+; GFX9-NODL-NEXT:    v_mul_u32_u24_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_2 src1_sel:BYTE_2
+; GFX9-NODL-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX9-NODL-NEXT:    v_mad_u32_u24 v2, v3, v4, s0
+; GFX9-NODL-NEXT:    v_add3_u32 v1, v2, v5, v1
+; GFX9-NODL-NEXT:    global_store_dword v0, v1, s[2:3]
+; GFX9-NODL-NEXT:    s_endpgm
+;
+; GFX9-DL-LABEL: idot4_acc32_3ele:
+; GFX9-DL:       ; %bb.0: ; %entry
+; GFX9-DL-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x24
+; GFX9-DL-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x34
+; GFX9-DL-NEXT:    v_lshlrev_b32_e32 v0, 2, v0
+; GFX9-DL-NEXT:    s_mov_b32 s1, 0xc020100
+; GFX9-DL-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX9-DL-NEXT:    global_load_dword v1, v0, s[6:7]
+; GFX9-DL-NEXT:    global_load_dword v2, v0, s[4:5]
+; GFX9-DL-NEXT:    s_load_dword s0, s[2:3], 0x0
+; GFX9-DL-NEXT:    v_mov_b32_e32 v0, 0
+; GFX9-DL-NEXT:    s_waitcnt vmcnt(1)
+; GFX9-DL-NEXT:    v_perm_b32 v1, v1, v1, s1
+; GFX9-DL-NEXT:    s_waitcnt vmcnt(0)
+; GFX9-DL-NEXT:    v_perm_b32 v2, v2, v2, s1
+; GFX9-DL-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX9-DL-NEXT:    v_dot4_u32_u8 v1, v2, v1, s0
+; GFX9-DL-NEXT:    global_store_dword v0, v1, s[2:3]
+; GFX9-DL-NEXT:    s_endpgm
+;
+; GFX10-DL-LABEL: idot4_acc32_3ele:
+; GFX10-DL:       ; %bb.0: ; %entry
+; GFX10-DL-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x24
+; GFX10-DL-NEXT:    v_lshlrev_b32_e32 v0, 2, v0
+; GFX10-DL-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0x34
+; GFX10-DL-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-DL-NEXT:    s_clause 0x1
+; GFX10-DL-NEXT:    global_load_dword v1, v0, s[6:7]
+; GFX10-DL-NEXT:    global_load_dword v2, v0, s[4:5]
+; GFX10-DL-NEXT:    s_load_dword s2, s[0:1], 0x0
+; GFX10-DL-NEXT:    s_waitcnt vmcnt(1)
+; GFX10-DL-NEXT:    v_perm_b32 v0, v1, v1, 0xc020100
+; GFX10-DL-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-DL-NEXT:    v_perm_b32 v1, v2, v2, 0xc020100
+; GFX10-DL-NEXT:    v_mov_b32_e32 v2, 0
+; GFX10-DL-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-DL-NEXT:    v_dot4_u32_u8 v0, v1, v0, s2
+; GFX10-DL-NEXT:    global_store_dword v2, v0, s[0:1]
+; GFX10-DL-NEXT:    s_endpgm
+;
+; GFX11-DL-LABEL: idot4_acc32_3ele:
+; GFX11-DL:       ; %bb.0: ; %entry
+; GFX11-DL-NEXT:    s_load_b128 s[4:7], s[0:1], 0x24
+; GFX11-DL-NEXT:    v_lshlrev_b32_e32 v0, 2, v0
+; GFX11-DL-NEXT:    s_load_b64 s[0:1], s[0:1], 0x34
+; GFX11-DL-NEXT:    v_mov_b32_e32 v2, 0
+; GFX11-DL-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX11-DL-NEXT:    s_clause 0x1
+; GFX11-DL-NEXT:    global_load_b32 v1, v0, s[6:7]
+; GFX11-DL-NEXT:    global_load_b32 v0, v0, s[4:5]
+; GFX11-DL-NEXT:    s_load_b32 s2, s[0:1], 0x0
+; GFX11-DL-NEXT:    s_waitcnt vmcnt(1)
+; GFX11-DL-NEXT:    v_perm_b32 v1, v1, v1, 0xc020100
+; GFX11-DL-NEXT:    s_waitcnt vmcnt(0)
+; GFX11-DL-NEXT:    v_perm_b32 v0, v0, v0, 0xc020100
+; GFX11-DL-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX11-DL-NEXT:    s_delay_alu instid0(VALU_DEP_1)
+; GFX11-DL-NEXT:    v_dot4_u32_u8 v0, v0, v1, s2
+; GFX11-DL-NEXT:    global_store_b32 v2, v0, s[0:1]
+; GFX11-DL-NEXT:    s_nop 0
+; GFX11-DL-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
+; GFX11-DL-NEXT:    s_endpgm
+                                       ptr addrspace(1) %src2,
+                                       ptr addrspace(1) nocapture %dst) {
+entry:
+  %idx = call i32 @llvm.amdgcn.workitem.id.x()
+  %gep1 = getelementptr <4 x i8>, ptr addrspace(1) %src1, i32 %idx
+  %vec1 = load <4 x i8>, ptr addrspace(1) %gep1
+  %gep2 = getelementptr <4 x i8>, ptr addrspace(1) %src2, i32 %idx
+  %vec2 = load <4 x i8>, ptr addrspace(1) %gep2
+
+  %v1e0 = extractelement <4 x i8> %vec1, i64 0
+  %cv1e0 = zext i8 %v1e0 to i32
+  %v2e0 = extractelement <4 x i8> %vec2, i64 0
+  %cv2e0 = zext i8 %v2e0 to i32
+  %mul1 = mul nuw nsw i32 %cv1e0, %cv2e0
+
+  %v1e1 = extractelement <4 x i8> %vec1, i64 1
+  %cv1e1 = zext i8 %v1e1 to i32
+  %v2e1 = extractelement <4 x i8> %vec2, i64 1
+  %cv2e1 = zext i8 %v2e1 to i32
+  %mul2 = mul nuw nsw i32 %cv1e1, %cv2e1
+
+  %v1e2 = extractelement <4 x i8> %vec1, i64 2
+  %cv1e2 = zext i8 %v1e2 to i32
+  %v2e2 = extractelement <4 x i8> %vec2, i64 2
+  %cv2e2 = zext i8 %v2e2 to i32
+  %mul3 = mul nuw nsw i32 %cv1e2, %cv2e2
+
+  %acc = load i32, ptr addrspace(1) %dst, align 4
+  %add1 = add i32 %mul1, %acc
+  %add2 = add i32 %add1, %mul2
+  %add3 = add i32 %add2, %mul3
+  store i32 %add3, ptr addrspace(1) %dst, align 4
+  ret void
+}
+
+define amdgpu_kernel void @idot4_acc32_3ele_permuted(ptr addrspace(1) %src1,
+; GFX7-LABEL: idot4_acc32_3ele_permuted:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x9
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xd
+; GFX7-NEXT:    s_mov_b32 s3, 0xf000
+; GFX7-NEXT:    s_mov_b32 s10, 0
+; GFX7-NEXT:    s_mov_b32 s11, s3
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    s_mov_b64 s[8:9], s[4:5]
+; GFX7-NEXT:    v_lshlrev_b32_e32 v0, 2, v0
+; GFX7-NEXT:    v_mov_b32_e32 v1, 0
+; GFX7-NEXT:    buffer_load_dword v2, v[0:1], s[8:11], 0 addr64
+; GFX7-NEXT:    s_mov_b64 s[8:9], s[6:7]
+; GFX7-NEXT:    buffer_load_dword v0, v[0:1], s[8:11], 0 addr64
+; GFX7-NEXT:    s_load_dword s4, s[0:1], 0x0
+; GFX7-NEXT:    s_mov_b32 s2, -1
+; GFX7-NEXT:    s_waitcnt vmcnt(1)
+; GFX7-NEXT:    v_lshrrev_b32_e32 v1, 24, v2
+; GFX7-NEXT:    v_and_b32_e32 v3, 0xff, v2
+; GFX7-NEXT:    s_waitcnt vmcnt(0)
+; GFX7-NEXT:    v_lshrrev_b32_e32 v4, 24, v0
+; GFX7-NEXT:    v_and_b32_e32 v5, 0xff, v0
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mad_u32_u24 v1, v1, v4, s4
+; GFX7-NEXT:    v_bfe_u32 v2, v2, 16, 8
+; GFX7-NEXT:    v_bfe_u32 v0, v0, 16, 8
+; GFX7-NEXT:    v_mad_u32_u24 v1, v3, v5, v1
+; GFX7-NEXT:    v_mad_u32_u24 v0, v2, v0, v1
+; GFX7-NEXT:    buffer_store_dword v0, off, s[0:3], 0
+; GFX7-NEXT:    s_endpgm
+;
+; GFX8-LABEL: idot4_acc32_3ele_permuted:
+; GFX8:       ; %bb.0: ; %entry
+; GFX8-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x24
+; GFX8-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0x34
+; GFX8-NEXT:    v_lshlrev_b32_e32 v2, 2, v0
+; GFX8-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX8-NEXT:    v_mov_b32_e32 v1, s5
+; GFX8-NEXT:    v_add_u32_e32 v0, vcc, s4, v2
+; GFX8-NEXT:    v_addc_u32_e32 v1, vcc, 0, v1, vcc
+; GFX8-NEXT:    flat_load_dword v3, v[0:1]
+; GFX8-NEXT:    v_mov_b32_e32 v1, s7
+; GFX8-NEXT:    v_add_u32_e32 v0, vcc, s6, v2
+; GFX8-NEXT:    v_addc_u32_e32 v1, vcc, 0, v1, vcc
+; GFX8-NEXT:    flat_load_dword v0, v[0:1]
+; GFX8-NEXT:    s_load_dword s2, s[0:1], 0x0
+; GFX8-NEXT:    s_waitcnt vmcnt(1)
+; GFX8-NEXT:    v_lshrrev_b32_e32 v1, 24, v3
+; GFX8-NEXT:    v_and_b32_e32 v4, 0xff, v3
+; GFX8-NEXT:    v_bfe_u32 v3, v3, 16, 8
+; GFX8-NEXT:    s_waitcnt vmcnt(0)
+; GFX8-NEXT:    v_lshrrev_b32_e32 v2, 24, v0
+; GFX8-NEXT:    v_and_b32_e32 v5, 0xff, v0
+; GFX8-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX8-NEXT:    v_mad_u32_u24 v1, v1, v2, s2
+; GFX8-NEXT:    v_bfe_u32 v0, v0, 16, 8
+; GFX8-NEXT:    v_mad_u32_u24 v1, v4, v5, v1
+; GFX8-NEXT:    v_mad_u32_u24 v2, v3, v0, v1
+; GFX8-NEXT:    v_mov_b32_e32 v0, s0
+; GFX8-NEXT:    v_mov_b32_e32 v1, s1
+; GFX8-NEXT:    flat_store_dword v[0:1], v2
+; GFX8-NEXT:    s_endpgm
+;
+; GFX9-NODL-LABEL: idot4_acc32_3ele_permuted:
+; GFX9-NODL:       ; %bb.0: ; %entry
+; GFX9-NODL-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x24
+; GFX9-NODL-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x34
+; GFX9-NODL-NEXT:    v_lshlrev_b32_e32 v0, 2, v0
+; GFX9-NODL-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX9-NODL-NEXT:    global_load_dword v1, v0, s[4:5]
+; GFX9-NODL-NEXT:    global_load_dword v2, v0, s[6:7]
+; GFX9-NODL-NEXT:    s_load_dword s0, s[2:3], 0x0
+; GFX9-NODL-NEXT:    v_mov_b32_e32 v0, 0
+; GFX9-NODL-NEXT:    s_waitcnt vmcnt(1)
+; GFX9-NODL-NEXT:    v_lshrrev_b32_e32 v3, 24, v1
+; GFX9-NODL-NEXT:    s_waitcnt vmcnt(0)
+; GFX9-NODL-NEXT:    v_lshrrev_b32_e32 v4, 24, v2
+; GFX9-NODL-NEXT:    v_mul_u32_u24_sdwa v5, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:BYTE_0
+; GFX9-NODL-NEXT:    v_mul_u32_u24_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_2 src1_sel:BYTE_2
+; GFX9-NODL-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX9-NODL-NEXT:    v_mad_u32_u24 v2, v3, v4, s0
+; GFX9-NODL-NEXT:    v_add3_u32 v1, v2, v5, v1
+; GFX9-NODL-NEXT:    global_store_dword v0, v1, s[2:3]
+; GFX9-NODL-NEXT:    s_endpgm
+;
+; GFX9-DL-LABEL: idot4_acc32_3ele_permuted:
+; GFX9-DL:       ; %bb.0: ; %entry
+; GFX9-DL-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x24
+; GFX9-DL-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x34
+; GFX9-DL-NEXT:    v_lshlrev_b32_e32 v0, 2, v0
+; GFX9-DL-NEXT:    s_mov_b32 s1, 0xc020003
+; GFX9-DL-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX9-DL-NEXT:    global_load_dword v1, v0, s[6:7]
+; GFX9-DL-NEXT:    global_load_dword v2, v0, s[4:5]
+; GFX9-DL-NEXT:    s_load_dword s0, s[2:3], 0x0
+; GFX9-DL-NEXT:    v_mov_b32_e32 v0, 0
+; GFX9-DL-NEXT:    s_waitcnt vmcnt(1)
+; GFX9-DL-NEXT:    v_perm_b32 v1, v1, v1, s1
+; GFX9-DL-NEXT:    s_waitcnt vmcnt(0)
+; GFX9-DL-NEXT:    v_perm_b32 v2, v2, v2, s1
+; GFX9-DL-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX9-DL-NEXT:    v_dot4_u32_u8 v1, v2, v1, s0
+; GFX9-DL-NEXT:    global_store_dword v0, v1, s[2:3]
+; GFX9-DL-NEXT:    s_endpgm
+;
+; GFX10-DL-LABEL: idot4_acc32_3ele_permuted:
+; GFX10-DL:       ; %bb.0: ; %entry
+; GFX10-DL-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x24
+; GFX10-DL-NEXT:    v_lshlrev_b32_e32 v0, 2, v0
+; GFX10-DL-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0x34
+; GFX10-DL-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-DL-NEXT:    s_clause 0x1
+; GFX10-DL-NEXT:    global_load_dword v1, v0, s[6:7]
+; GFX10-DL-NEXT:    global_load_dword v2, v0, s[4:5]
+; GFX10-DL-NEXT:    s_load_dword s2, s[0:1], 0x0
+; GFX10-DL-NEXT:    s_waitcnt vmcnt(1)
+; GFX10-DL-NEXT:    v_perm_b32 v0, v1, v1, 0xc020003
+; GFX10-DL-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-DL-NEXT:    v_perm_b32 v1, v2, v2, 0xc020003
+; GFX10-DL-NEXT:    v_mov_b32_e32 v2, 0
+; GFX10-DL-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-DL-NEXT:    v_dot4_u32_u8 v0, v1, v0, s2
+; GFX10-DL-NEXT:    global_store_dword v2, v0, s[0:1]
+; GFX10-DL-NEXT:    s_endpgm
+;
+; GFX11-DL-LABEL: idot4_acc32_3ele_permuted:
+; GFX11-DL:       ; %bb.0: ; %entry
+; GFX11-DL-NEXT:    s_load_b128 s[4:7], s[0:1], 0x24
+; GFX11-DL-NEXT:    v_lshlrev_b32_e32 v0, 2, v0
+; GFX11-DL-NEXT:    s_load_b64 s[0:1], s[0:1], 0x34
+; GFX11-DL-NEXT:    v_mov_b32_e32 v2, 0
+; GFX11-DL-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX11-DL-NEXT:    s_clause 0x1
+; GFX11-DL-NEXT:    global_load_b32 v1, v0, s[6:7]
+; GFX11-DL-NEXT:    global_load_b32 v0, v0, s[4:5]
+; GFX11-DL-NEXT:    s_load_b32 s2, s[0:1], 0x0
+; GFX11-DL-NEXT:    s_waitcnt vmcnt(1)
+; GFX11-DL-NEXT:    v_perm_b32 v1, v1, v1, 0xc020003
+; GFX11-DL-NEXT:    s_waitcnt vmcnt(0)
+; GFX11-DL-NEXT:    v_perm_b32 v0, v0, v0, 0xc020003
+; GFX11-DL-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX11-DL-NEXT:    s_delay_alu instid0(VALU_DEP_1)
+; GFX11-DL-NEXT:    v_dot4_u32_u8 v0, v0, v1, s2
+; GFX11-DL-NEXT:    global_store_b32 v2, v0, s[0:1]
+; GFX11-DL-NEXT:    s_nop 0
+; GFX11-DL-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
+; GFX11-DL-NEXT:    s_endpgm
+                                       ptr addrspace(1) %src2,
+                                       ptr addrspace(1) nocapture %dst) {
+entry:
+  %idx = call i32 @llvm.amdgcn.workitem.id.x()
+  %gep1 = getelementptr <4 x i8>, ptr addrspace(1) %src1, i32 %idx
+  %vec1 = load <4 x i8>, ptr addrspace(1) %gep1
+  %gep2 = getelementptr <4 x i8>, ptr addrspace(1) %src2, i32 %idx
+  %vec2 = load <4 x i8>, ptr addrspace(1) %gep2
+
+  %v1e0 = extractelement <4 x i8> %vec1, i64 3
+  %cv1e0 = zext i8 %v1e0 to i32
+  %v2e0 = extractelement <4 x i8> %vec2, i64 3
+  %cv2e0 = zext i8 %v2e0 to i32
+  %mul1 = mul nuw nsw i32 %cv1e0, %cv2e0
+
+  %v1e1 = extractelement <4 x i8> %vec1, i64 0
+  %cv1e1 = zext i8 %v1e1 to i32
+  %v2e1 = extractelement <4 x i8> %vec2, i64 0
+  %cv2e1 = zext i8 %v2e1 to i32
+  %mul2 = mul nuw nsw i32 %cv1e1, %cv2e1
+
+  %v1e2 = extractelement <4 x i8> %vec1, i64 2
+  %cv1e2 = zext i8 %v1e2 to i32
+  %v2e2 = extractelement <4 x i8> %vec2, i64 2
+  %cv2e2 = zext i8 %v2e2 to i32
+  %mul3 = mul nuw nsw i32 %cv1e2, %cv2e2
+
+  %acc = load i32, ptr addrspace(1) %dst, align 4
+  %add1 = add i32 %mul1, %acc
+  %add2 = add i32 %add1, %mul2
+  %add3 = add i32 %add2, %mul3
+  store i32 %add3, ptr addrspace(1) %dst, align 4
+  ret void
+}
+
+
+define amdgpu_kernel void @idot4_acc32_opt(ptr addrspace(1) %src1,
+; GFX7-LABEL: idot4_acc32_opt:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x9
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xd
+; GFX7-NEXT:    s_mov_b32 s3, 0xf000
+; GFX7-NEXT:    s_mov_b32 s10, 0
+; GFX7-NEXT:    s_mov_b32 s11, s3
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    s_mov_b64 s[8:9], s[4:5]
+; GFX7-NEXT:    v_lshlrev_b32_e32 v0, 2, v0
+; GFX7-NEXT:    v_mov_b32_e32 v1, 0
+; GFX7-NEXT:    buffer_load_dword v2, v[0:1], s[8:11], 0 addr64
+; GFX7-NEXT:    s_mov_b64 s[8:9], s[6:7]
+; GFX7-NEXT:    buffer_load_dword v0, v[0:1], s[8:11], 0 addr64
+; GFX7-NEXT:    s_mov_b32 s2, -1
+; GFX7-NEXT:    s_waitcnt vmcnt(1)
+; GFX7-NEXT:    v_bfe_u32 v3, v2, 8, 8
+; GFX7-NEXT:    v_and_b32_e32 v1, 0xff, v2
+; GFX7-NEXT:    s_waitcnt vmcnt(0)
+; GFX7-NEXT:    v_bfe_u32 v6, v0, 8, 8
+; GFX7-NEXT:    v_and_b32_e32 v5, 0xff, v0
+; GFX7-NEXT:    v_mul_u32_u24_e32 v3, v3, v6
+; GFX7-NEXT:    v_bfe_u32 v4, v2, 16, 8
+; GFX7-NEXT:    v_bfe_u32 v7, v0, 16, 8
+; GFX7-NEXT:    v_mad_u32_u24 v1, v1, v5, v3
+; GFX7-NEXT:    v_lshrrev_b32_e32 v2, 24, v2
+; GFX7-NEXT:    v_lshrrev_b32_e32 v0, 24, v0
+; GFX7-NEXT:    v_mad_u32_u24 v1, v4, v7, v1
+; GFX7-NEXT:    v_mad_u32_u24 v0, v2, v0, v1
+; GFX7-NEXT:    buffer_store_dword v0, off, s[0:3], 0
+; GFX7-NEXT:    s_endpgm
+;
+; GFX8-LABEL: idot4_acc32_opt:
+; GFX8:       ; %bb.0: ; %entry
+; GFX8-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x24
+; GFX8-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0x34
+; GFX8-NEXT:    v_lshlrev_b32_e32 v2, 2, v0
+; GFX8-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX8-NEXT:    v_mov_b32_e32 v1, s5
+; GFX8-NEXT:    v_add_u32_e32 v0, vcc, s4, v2
+; GFX8-NEXT:    v_addc_u32_e32 v1, vcc, 0, v1, vcc
+; GFX8-NEXT:    flat_load_dword v3, v[0:1]
+; GFX8-NEXT:    v_mov_b32_e32 v1, s7
+; GFX8-NEXT:    v_add_u32_e32 v0, vcc, s6, v2
+; GFX8-NEXT:    v_addc_u32_e32 v1, vcc, 0, v1, vcc
+; GFX8-NEXT:    flat_load_dword v2, v[0:1]
+; GFX8-NEXT:    v_mov_b32_e32 v0, s0
+; GFX8-NEXT:    v_mov_b32_e32 v1, s1
+; GFX8-NEXT:    s_waitcnt vmcnt(1)
+; GFX8-NEXT:    v_and_b32_e32 v4, 0xff, v3
+; GFX8-NEXT:    v_bfe_u32 v7, v3, 16, 8
+; GFX8-NEXT:    s_waitcnt vmcnt(0)
+; GFX8-NEXT:    v_and_b32_e32 v5, 0xff, v2
+; GFX8-NEXT:    v_mul_u32_u24_sdwa v6, v3, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_1 src1_sel:BYTE_1
+; GFX8-NEXT:    v_bfe_u32 v8, v2, 16, 8
+; GFX8-NEXT:    v_mad_u32_u24 v4, v4, v5, v6
+; GFX8-NEXT:    v_lshrrev_b32_e32 v3, 24, v3
+; GFX8-NEXT:    v_lshrrev_b32_e32 v2, 24, v2
+; GFX8-NEXT:    v_mad_u32_u24 v4, v7, v8, v4
+; GFX8-NEXT:    v_mad_u32_u24 v2, v3, v2, v4
+; GFX8-NEXT:    flat_store_dword v[0:1], v2
+; GFX8-NEXT:    s_endpgm
+;
+; GFX9-NODL-LABEL: idot4_acc32_opt:
+; GFX9-NODL:       ; %bb.0: ; %entry
+; GFX9-NODL-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x24
+; GFX9-NODL-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x34
+; GFX9-NODL-NEXT:    v_lshlrev_b32_e32 v0, 2, v0
+; GFX9-NODL-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX9-NODL-NEXT:    global_load_dword v1, v0, s[4:5]
+; GFX9-NODL-NEXT:    global_load_dword v2, v0, s[6:7]
+; GFX9-NODL-NEXT:    v_mov_b32_e32 v0, 0
+; GFX9-NODL-NEXT:    s_waitcnt vmcnt(1)
+; GFX9-NODL-NEXT:    v_and_b32_e32 v3, 0xff, v1
+; GFX9-NODL-NEXT:    s_waitcnt vmcnt(0)
+; GFX9-NODL-NEXT:    v_and_b32_e32 v4, 0xff, v2
+; GFX9-NODL-NEXT:    v_mul_u32_u24_sdwa v5, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_1 src1_sel:BYTE_1
+; GFX9-NODL-NEXT:    v_mul_u32_u24_sdwa v6, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_2 src1_sel:BYTE_2
+; GFX9-NODL-NEXT:    v_mul_u32_u24_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_3 src1_sel:BYTE_3
+; GFX9-NODL-NEXT:    v_mad_u32_u24 v2, v3, v4, v5
+; GFX9-NODL-NEXT:    v_add3_u32 v1, v2, v6, v1
+; GFX9-NODL-NEXT:    global_store_dword v0, v1, s[2:3]
+; GFX9-NODL-NEXT:    s_endpgm
+;
+; GFX9-DL-LABEL: idot4_acc32_opt:
+; GFX9-DL:       ; %bb.0: ; %entry
+; GFX9-DL-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x24
+; GFX9-DL-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x34
+; GFX9-DL-NEXT:    v_lshlrev_b32_e32 v0, 2, v0
+; GFX9-DL-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX9-DL-NEXT:    global_load_dword v1, v0, s[4:5]
+; GFX9-DL-NEXT:    global_load_dword v2, v0, s[6:7]
+; GFX9-DL-NEXT:    v_mov_b32_e32 v0, 0
+; GFX9-DL-NEXT:    s_waitcnt vmcnt(0)
+; GFX9-DL-NEXT:    v_dot4_u32_u8 v1, v1, v2, 0
+; GFX9-DL-NEXT:    global_store_dword v0, v1, s[2:3]
+; GFX9-DL-NEXT:    s_endpgm
+;
+; GFX10-DL-LABEL: idot4_acc32_opt:
+; GFX10-DL:       ; %bb.0: ; %entry
+; GFX10-DL-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x24
+; GFX10-DL-NEXT:    v_lshlrev_b32_e32 v0, 2, v0
+; GFX10-DL-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0x34
+; GFX10-DL-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-DL-NEXT:    s_clause 0x1
+; GFX10-DL-NEXT:    global_load_dword v1, v0, s[4:5]
+; GFX10-DL-NEXT:    global_load_dword v2, v0, s[6:7]
+; GFX10-DL-NEXT:    v_mov_b32_e32 v0, 0
+; GFX10-DL-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-DL-NEXT:    v_dot4_u32_u8 v1, v1, v2, 0
+; GFX10-DL-NEXT:    global_store_dword v0, v1, s[0:1]
+; GFX10-DL-NEXT:    s_endpgm
+;
+; GFX11-DL-LABEL: idot4_acc32_opt:
+; GFX11-DL:       ; %bb.0: ; %entry
+; GFX11-DL-NEXT:    s_load_b128 s[4:7], s[0:1], 0x24
+; GFX11-DL-NEXT:    v_lshlrev_b32_e32 v0, 2, v0
+; GFX11-DL-NEXT:    s_load_b64 s[0:1], s[0:1], 0x34
+; GFX11-DL-NEXT:    v_mov_b32_e32 v2, 0
+; GFX11-DL-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX11-DL-NEXT:    s_clause 0x1
+; GFX11-DL-NEXT:    global_load_b32 v1, v0, s[4:5]
+; GFX11-DL-NEXT:    global_load_b32 v0, v0, s[6:7]
+; GFX11-DL-NEXT:    s_waitcnt vmcnt(0)
+; GFX11-DL-NEXT:    v_dot4_u32_u8 v0, v1, v0, 0
+; GFX11-DL-NEXT:    global_store_b32 v2, v0, s[0:1]
+; GFX11-DL-NEXT:    s_nop 0
+; GFX11-DL-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
+; GFX11-DL-NEXT:    s_endpgm
+                                       ptr addrspace(1) %src2,
+                                       ptr addrspace(1) nocapture %dst) {
+entry:
+  %idx = call i32 @llvm.amdgcn.workitem.id.x()
+  %gep1 = getelementptr <4 x i8>, ptr addrspace(1) %src1, i32 %idx
+  %vec1 = load <4 x i8>, ptr addrspace(1) %gep1
+  %gep2 = getelementptr <4 x i8>, ptr addrspace(1) %src2, i32 %idx
+  %vec2 = load <4 x i8>, ptr addrspace(1) %gep2
+
+  %v1e0 = extractelement <4 x i8> %vec1, i64 0
+  %cv1e0 = zext i8 %v1e0 to i32
+  %v2e0 = extractelement <4 x i8> %vec2, i64 0
+  %cv2e0 = zext i8 %v2e0 to i32
+  %mul1 = mul nuw nsw i32 %cv1e0, %cv2e0
+
+  %v1e1 = extractelement <4 x i8> %vec1, i64 1
+  %cv1e1 = zext i8 %v1e1 to i32
+  %v2e1 = extractelement <4 x i8> %vec2, i64 1
+  %cv2e1 = zext i8 %v2e1 to i32
+  %mul2 = mul nuw nsw i32 %cv1e1, %cv2e1
+
+  %v1e2 = extractelement <4 x i8> %vec1, i64 2
+  %cv1e2 = zext i8 %v1e2 to i32
+  %v2e2 = extractelement <4 x i8> %vec2, i64 2
+  %cv2e2 = zext i8 %v2e2 to i32
+  %mul3 = mul nuw nsw i32 %cv1e2, %cv2e2
+
+  %v1e3 = extractelement <4 x i8> %vec1, i64 3
+  %cv1e3 = zext i8 %v1e3 to i32
+  %v2e3 = extractelement <4 x i8> %vec2, i64 3
+  %cv2e3 = zext i8 %v2e3 to i32
+  %mul4 = mul nuw nsw i32 %cv1e3, %cv2e3
+
+  %add2 = add i32 %mul1, %mul2
+  %add3 = add i32 %add2, %mul3
+  %add4 = add i32 %add3, %mul4
+  store i32 %add4, ptr addrspace(1) %dst, align 4
+  ret void
+}
+
+define amdgpu_kernel void @udot4_acc32_3src(ptr addrspace(1) %src1,
+; GFX7-LABEL: udot4_acc32_3src:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx8 s[0:7], s[0:1], 0x9
+; GFX7-NEXT:    s_mov_b32 s11, 0xf000
+; GFX7-NEXT:    s_mov_b32 s14, 0
+; GFX7-NEXT:    s_mov_b32 s15, s11
+; GFX7-NEXT:    v_lshlrev_b32_e32 v0, 2, v0
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    s_mov_b64 s[12:13], s[0:1]
+; GFX7-NEXT:    v_mov_b32_e32 v1, 0
+; GFX7-NEXT:    buffer_load_dword v2, v[0:1], s[12:15], 0 addr64
+; GFX7-NEXT:    s_mov_b64 s[12:13], s[2:3]
+; GFX7-NEXT:    buffer_load_dword v3, v[0:1], s[12:15], 0 addr64
+; GFX7-NEXT:    s_mov_b64 s[12:13], s[4:5]
+; GFX7-NEXT:    buffer_load_dword v0, v[0:1], s[12:15], 0 addr64
+; GFX7-NEXT:    s_load_dword s0, s[6:7], 0x0
+; GFX7-NEXT:    s_mov_b32 s10, -1
+; GFX7-NEXT:    s_mov_b32 s8, s6
+; GFX7-NEXT:    s_mov_b32 s9, s7
+; GFX7-NEXT:    s_waitcnt vmcnt(2)
+; GFX7-NEXT:    v_and_b32_e32 v1, 0xff, v2
+; GFX7-NEXT:    v_bfe_u32 v4, v2, 8, 8
+; GFX7-NEXT:    s_waitcnt vmcnt(1)
+; GFX7-NEXT:    v_bfe_u32 v3, v3, 8, 8
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mad_u32_u24 v1, v1, v1, s0
+; GFX7-NEXT:    v_bfe_u32 v5, v2, 16, 8
+; GFX7-NEXT:    s_waitcnt vmcnt(0)
+; GFX7-NEXT:    v_bfe_u32 v6, v0, 16, 8
+; GFX7-NEXT:    v_mad_u32_u24 v1, v4, v3, v1
+; GFX7-NEXT:    v_lshrrev_b32_e32 v2, 24, v2
+; GFX7-NEXT:    v_lshrrev_b32_e32 v0, 24, v0
+; GFX7-NEXT:    v_mad_u32_u24 v1, v5, v6, v1
+; GFX7-NEXT:    v_mad_u32_u24 v0, v2, v0, v1
+; GFX7-NEXT:    buffer_store_dword v0, off, s[8:11], 0
+; GFX7-NEXT:    s_endpgm
+;
+; GFX8-LABEL: udot4_acc32_3src:
+; GFX8:       ; %bb.0: ; %entry
+; GFX8-NEXT:    s_load_dwordx8 s[0:7], s[0:1], 0x24
+; GFX8-NEXT:    v_lshlrev_b32_e32 v2, 2, v0
+; GFX8-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX8-NEXT:    v_mov_b32_e32 v1, s1
+; GFX8-NEXT:    v_add_u32_e32 v0, vcc, s0, v2
+; GFX8-NEXT:    v_addc_u32_e32 v1, vcc, 0, v1, vcc
+; GFX8-NEXT:    flat_load_dword v3, v[0:1]
+; GFX8-NEXT:    v_mov_b32_e32 v1, s3
+; GFX8-NEXT:    v_add_u32_e32 v0, vcc, s2, v2
+; GFX8-NEXT:    v_addc_u32_e32 v1, vcc, 0, v1, vcc
+; GFX8-NEXT:    flat_load_dword v4, v[0:1]
+; GFX8-NEXT:    v_mov_b32_e32 v1, s5
+; GFX8-NEXT:    v_add_u32_e32 v0, vcc, s4, v2
+; GFX8-NEXT:    v_addc_u32_e32 v1, vcc, 0, v1, vcc
+; GFX8-NEXT:    flat_load_dword v0, v[0:1]
+; GFX8-NEXT:    s_load_dword s0, s[6:7], 0x0
+; GFX8-NEXT:    s_waitcnt vmcnt(2)
+; GFX8-NEXT:    v_and_b32_e32 v1, 0xff, v3
+; GFX8-NEXT:    v_bfe_u32 v2, v3, 8, 8
+; GFX8-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX8-NEXT:    v_mad_u32_u24 v1, v1, v1, s0
+; GFX8-NEXT:    v_bfe_u32 v5, v3, 16, 8
+; GFX8-NEXT:    v_lshrrev_b32_e32 v3, 24, v3
+; GFX8-NEXT:    s_waitcnt vmcnt(1)
+; GFX8-NEXT:    v_bfe_u32 v4, v4, 8, 8
+; GFX8-NEXT:    v_mad_u32_u24 v1, v2, v4, v1
+; GFX8-NEXT:    s_waitcnt vmcnt(0)
+; GFX8-NEXT:    v_bfe_u32 v6, v0, 16, 8
+; GFX8-NEXT:    v_lshrrev_b32_e32 v0, 24, v0
+; GFX8-NEXT:    v_mad_u32_u24 v1, v5, v6, v1
+; GFX8-NEXT:    v_mad_u32_u24 v2, v3, v0, v1
+; GFX8-NEXT:    v_mov_b32_e32 v0, s6
+; GFX8-NEXT:    v_mov_b32_e32 v1, s7
+; GFX8-NEXT:    flat_store_dword v[0:1], v2
+; GFX8-NEXT:    s_endpgm
+;
+; GFX9-NODL-LABEL: udot4_acc32_3src:
+; GFX9-NODL:       ; %bb.0: ; %entry
+; GFX9-NODL-NEXT:    s_load_dwordx8 s[0:7], s[0:1], 0x24
+; GFX9-NODL-NEXT:    v_lshlrev_b32_e32 v0, 2, v0
+; GFX9-NODL-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX9-NODL-NEXT:    global_load_dword v1, v0, s[0:1]
+; GFX9-NODL-NEXT:    global_load_dword v2, v0, s[2:3]
+; GFX9-NODL-NEXT:    global_load_dword v3, v0, s[4:5]
+; GFX9-NODL-NEXT:    s_load_dword s0, s[6:7], 0x0
+; GFX9-NODL-NEXT:    v_mov_b32_e32 v0, 0
+; GFX9-NODL-NEXT:    s_waitcnt vmcnt(2)
+; GFX9-NODL-NEXT:    v_mul_u32_u24_sdwa v4, v1, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:BYTE_0
+; GFX9-NODL-NEXT:    s_waitcnt vmcnt(1)
+; GFX9-NODL-NEXT:    v_mul_u32_u24_sdwa v2, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_1 src1_sel:BYTE_1
+; GFX9-NODL-NEXT:    s_waitcnt vmcnt(0)
+; GFX9-NODL-NEXT:    v_mul_u32_u24_sdwa v5, v1, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_2 src1_sel:BYTE_2
+; GFX9-NODL-NEXT:    v_mul_u32_u24_sdwa v1, v1, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_3 src1_sel:BYTE_3
+; GFX9-NODL-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX9-NODL-NEXT:    v_add3_u32 v2, v4, s0, v2
+; GFX9-NODL-NEXT:    v_add3_u32 v1, v2, v5, v1
+; GFX9-NODL-NEXT:    global_store_dword v0, v1, s[6:7]
+; GFX9-NODL-NEXT:    s_endpgm
+;
+; GFX9-DL-LABEL: udot4_acc32_3src:
+; GFX9-DL:       ; %bb.0: ; %entry
+; GFX9-DL-NEXT:    s_load_dwordx8 s[0:7], s[0:1], 0x24
+; GFX9-DL-NEXT:    v_lshlrev_b32_e32 v0, 2, v0
+; GFX9-DL-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX9-DL-NEXT:    global_load_dword v1, v0, s[2:3]
+; GFX9-DL-NEXT:    global_load_dword v2, v0, s[4:5]
+; GFX9-DL-NEXT:    global_load_dword v3, v0, s[0:1]
+; GFX9-DL-NEXT:    s_load_dword s1, s[6:7], 0x0
+; GFX9-DL-NEXT:    s_mov_b32 s0, 0x706010c
+; GFX9-DL-NEXT:    s_mov_b32 s2, 0xc0c0c00
+; GFX9-DL-NEXT:    v_mov_b32_e32 v0, 0
+; GFX9-DL-NEXT:    s_waitcnt vmcnt(1)
+; GFX9-DL-NEXT:    v_perm_b32 v1, v2, v1, s0
+; GFX9-DL-NEXT:    s_waitcnt vmcnt(0)
+; GFX9-DL-NEXT:    v_perm_b32 v2, v3, v3, s2
+; GFX9-DL-NEXT:    v_or_b32_e32 v1, v1, v2
+; GFX9-DL-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX9-DL-NEXT:    v_dot4_u32_u8 v1, v3, v1, s1
+; GFX9-DL-NEXT:    global_store_dword v0, v1, s[6:7]
+; GFX9-DL-NEXT:    s_endpgm
+;
+; GFX10-DL-LABEL: udot4_acc32_3src:
+; GFX10-DL:       ; %bb.0: ; %entry
+; GFX10-DL-NEXT:    s_load_dwordx8 s[0:7], s[0:1], 0x24
+; GFX10-DL-NEXT:    v_lshlrev_b32_e32 v0, 2, v0
+; GFX10-DL-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-DL-NEXT:    s_clause 0x2
+; GFX10-DL-NEXT:    global_load_dword v1, v0, s[2:3]
+; GFX10-DL-NEXT:    global_load_dword v2, v0, s[4:5]
+; GFX10-DL-NEXT:    global_load_dword v3, v0, s[0:1]
+; GFX10-DL-NEXT:    s_waitcnt_depctr 0xffe3
+; GFX10-DL-NEXT:    s_load_dword s0, s[6:7], 0x0
+; GFX10-DL-NEXT:    s_waitcnt vmcnt(1)
+; GFX10-DL-NEXT:    v_perm_b32 v0, v2, v1, 0x706010c
+; GFX10-DL-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-DL-NEXT:    v_perm_b32 v1, v3, v3, 0xc0c0c00
+; GFX10-DL-NEXT:    v_or_b32_e32 v0, v0, v1
+; GFX10-DL-NEXT:    v_mov_b32_e32 v1, 0
+; GFX10-DL-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-DL-NEXT:    v_dot4_u32_u8 v0, v3, v0, s0
+; GFX10-DL-NEXT:    global_store_dword v1, v0, s[6:7]
+; GFX10-DL-NEXT:    s_endpgm
+;
+; GFX11-DL-LABEL: udot4_acc32_3src:
+; GFX11-DL:       ; %bb.0: ; %entry
+; GFX11-DL-NEXT:    s_load_b256 s[0:7], s[0:1], 0x24
+; GFX11-DL-NEXT:    v_lshlrev_b32_e32 v0, 2, v0
+; GFX11-DL-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX11-DL-NEXT:    s_clause 0x2
+; GFX11-DL-NEXT:    global_load_b32 v1, v0, s[2:3]
+; GFX11-DL-NEXT:    global_load_b32 v2, v0, s[4:5]
+; GFX11-DL-NEXT:    global_load_b32 v0, v0, s[0:1]
+; GFX11-DL-NEXT:    s_load_b32 s0, s[6:7], 0x0
+; GFX11-DL-NEXT:    s_waitcnt vmcnt(1)
+; GFX11-DL-NEXT:    v_perm_b32 v1, v2, v1, 0x706010c
+; GFX11-DL-NEXT:    s_waitcnt vmcnt(0)
+; GFX11-DL-NEXT:    v_perm_b32 v2, v0, v0, 0xc0c0c00
+; GFX11-DL-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_2)
+; GFX11-DL-NEXT:    v_or_b32_e32 v1, v1, v2
+; GFX11-DL-NEXT:    v_mov_b32_e32 v2, 0
+; GFX11-DL-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX11-DL-NEXT:    v_dot4_u32_u8 v0, v0, v1, s0
+; GFX11-DL-NEXT:    global_store_b32 v2, v0, s[6:7]
+; GFX11-DL-NEXT:    s_nop 0
+; GFX11-DL-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
+; GFX11-DL-NEXT:    s_endpgm
+                                       ptr addrspace(1) %src2,
+                                       ptr addrspace(1) %src3,
+                                       ptr addrspace(1) nocapture %dst) {
+entry:
+  %idx = call i32 @llvm.amdgcn.workitem.id.x()
+  %gep1 = getelementptr <4 x i8>, ptr addrspace(1) %src1, i32 %idx
+  %vec1 = load <4 x i8>, ptr addrspace(1) %gep1
+  %gep2 = getelementptr <4 x i8>, ptr addrspace(1) %src2, i32 %idx
+  %vec2 = load <4 x i8>, ptr addrspace(1) %gep2
+  %gep3 = getelementptr <4 x i8>, ptr addrspace(1) %src3, i32 %idx
+  %vec3 = load <4 x i8>, ptr addrspace(1) %gep3
+
+  %v1e0 = extractelement <4 x i8> %vec1, i64 0
+  %cv1e0 = zext i8 %v1e0 to i32
+  %mul1 = mul nuw nsw i32 %cv1e0, %cv1e0
+
+  %v1e1 = extractelement <4 x i8> %vec1, i64 1
+  %cv1e1 = zext i8 %v1e1 to i32
+  %v2e1 = extractelement <4 x i8> %vec2, i64 1
+  %cv2e1 = zext i8 %v2e1 to i32
+  %mul2 = mul nuw nsw i32 %cv1e1, %cv2e1
+
+  %v1e2 = extractelement <4 x i8> %vec1, i64 2
+  %cv1e2 = zext i8 %v1e2 to i32
+  %v3e2 = extractelement <4 x i8> %vec3, i64 2
+  %cv3e2 = zext i8 %v3e2 to i32
+  %mul3 = mul nuw nsw i32 %cv1e2, %cv3e2
+
+  %v1e3 = extractelement <4 x i8> %vec1, i64 3
+  %cv1e3 = zext i8 %v1e3 to i32
+  %v3e3 = extractelement <4 x i8> %vec3, i64 3
+  %cv3e3 = zext i8 %v3e3 to i32
+  %mul4 = mul nuw nsw i32 %cv1e3, %cv3e3
+
+  %acc = load i32, ptr addrspace(1) %dst, align 4
+  %mad1 = add i32 %mul1, %acc
+  %mad2 = add i32 %mad1, %mul2
+  %mad3 = add i32 %mad2, %mul3
+  %mad4 = add i32 %mad3, %mul4
+
+  store i32 %mad4, ptr addrspace(1) %dst, align 4
+  ret void
+}
+
+define amdgpu_kernel void @udot4_acc32_3src_3ele(ptr addrspace(1) %src1,
+; GFX7-LABEL: udot4_acc32_3src_3ele:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx8 s[0:7], s[0:1], 0x9
+; GFX7-NEXT:    s_mov_b32 s11, 0xf000
+; GFX7-NEXT:    s_mov_b32 s14, 0
+; GFX7-NEXT:    s_mov_b32 s15, s11
+; GFX7-NEXT:    v_lshlrev_b32_e32 v0, 2, v0
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    s_mov_b64 s[12:13], s[0:1]
+; GFX7-NEXT:    v_mov_b32_e32 v1, 0
+; GFX7-NEXT:    buffer_load_dword v2, v[0:1], s[12:15], 0 addr64
+; GFX7-NEXT:    s_mov_b64 s[12:13], s[2:3]
+; GFX7-NEXT:    buffer_load_dword v3, v[0:1], s[12:15], 0 addr64
+; GFX7-NEXT:    s_mov_b64 s[12:13], s[4:5]
+; GFX7-NEXT:    buffer_load_dword v0, v[0:1], s[12:15], 0 addr64
+; GFX7-NEXT:    s_load_dword s0, s[6:7], 0x0
+; GFX7-NEXT:    s_mov_b32 s10, -1
+; GFX7-NEXT:    s_mov_b32 s8, s6
+; GFX7-NEXT:    s_mov_b32 s9, s7
+; GFX7-NEXT:    s_waitcnt vmcnt(2)
+; GFX7-NEXT:    v_and_b32_e32 v1, 0xff, v2
+; GFX7-NEXT:    v_bfe_u32 v4, v2, 8, 8
+; GFX7-NEXT:    s_waitcnt vmcnt(1)
+; GFX7-NEXT:    v_bfe_u32 v3, v3, 8, 8
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mad_u32_u24 v1, v1, v1, s0
+; GFX7-NEXT:    v_bfe_u32 v2, v2, 16, 8
+; GFX7-NEXT:    s_waitcnt vmcnt(0)
+; GFX7-NEXT:    v_bfe_u32 v0, v0, 16, 8
+; GFX7-NEXT:    v_mad_u32_u24 v1, v4, v3, v1
+; GFX7-NEXT:    v_mad_u32_u24 v0, v2, v0, v1
+; GFX7-NEXT:    buffer_store_dword v0, off, s[8:11], 0
+; GFX7-NEXT:    s_endpgm
+;
+; GFX8-LABEL: udot4_acc32_3src_3ele:
+; GFX8:       ; %bb.0: ; %entry
+; GFX8-NEXT:    s_load_dwordx8 s[0:7], s[0:1], 0x24
+; GFX8-NEXT:    v_lshlrev_b32_e32 v2, 2, v0
+; GFX8-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX8-NEXT:    v_mov_b32_e32 v1, s1
+; GFX8-NEXT:    v_add_u32_e32 v0, vcc, s0, v2
+; GFX8-NEXT:    v_addc_u32_e32 v1, vcc, 0, v1, vcc
+; GFX8-NEXT:    flat_load_dword v3, v[0:1]
+; GFX8-NEXT:    v_mov_b32_e32 v1, s3
+; GFX8-NEXT:    v_add_u32_e32 v0, vcc, s2, v2
+; GFX8-NEXT:    v_addc_u32_e32 v1, vcc, 0, v1, vcc
+; GFX8-NEXT:    flat_load_dword v4, v[0:1]
+; GFX8-NEXT:    v_mov_b32_e32 v1, s5
+; GFX8-NEXT:    v_add_u32_e32 v0, vcc, s4, v2
+; GFX8-NEXT:    v_addc_u32_e32 v1, vcc, 0, v1, vcc
+; GFX8-NEXT:    flat_load_dword v0, v[0:1]
+; GFX8-NEXT:    s_load_dword s0, s[6:7], 0x0
+; GFX8-NEXT:    s_waitcnt vmcnt(2)
+; GFX8-NEXT:    v_and_b32_e32 v1, 0xff, v3
+; GFX8-NEXT:    v_bfe_u32 v2, v3, 8, 8
+; GFX8-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX8-NEXT:    v_mad_u32_u24 v1, v1, v1, s0
+; GFX8-NEXT:    v_bfe_u32 v3, v3, 16, 8
+; GFX8-NEXT:    s_waitcnt vmcnt(1)
+; GFX8-NEXT:    v_bfe_u32 v4, v4, 8, 8
+; GFX8-NEXT:    v_mad_u32_u24 v1, v2, v4, v1
+; GFX8-NEXT:    s_waitcnt vmcnt(0)
+; GFX8-NEXT:    v_bfe_u32 v0, v0, 16, 8
+; GFX8-NEXT:    v_mad_u32_u24 v2, v3, v0, v1
+; GFX8-NEXT:    v_mov_b32_e32 v0, s6
+; GFX8-NEXT:    v_mov_b32_e32 v1, s7
+; GFX8-NEXT:    flat_store_dword v[0:1], v2
+; GFX8-NEXT:    s_endpgm
+;
+; GFX9-NODL-LABEL: udot4_acc32_3src_3ele:
+; GFX9-NODL:       ; %bb.0: ; %entry
+; GFX9-NODL-NEXT:    s_load_dwordx8 s[0:7], s[0:1], 0x24
+; GFX9-NODL-NEXT:    v_lshlrev_b32_e32 v0, 2, v0
+; GFX9-NODL-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX9-NODL-NEXT:    global_load_dword v1, v0, s[0:1]
+; GFX9-NODL-NEXT:    global_load_dword v2, v0, s[2:3]
+; GFX9-NODL-NEXT:    global_load_dword v3, v0, s[4:5]
+; GFX9-NODL-NEXT:    s_load_dword s0, s[6:7], 0x0
+; GFX9-NODL-NEXT:    v_mov_b32_e32 v0, 0
+; GFX9-NODL-NEXT:    s_waitcnt vmcnt(2)
+; GFX9-NODL-NEXT:    v_and_b32_e32 v4, 0xff, v1
+; GFX9-NODL-NEXT:    s_waitcnt vmcnt(1)
+; GFX9-NODL-NEXT:    v_mul_u32_u24_sdwa v2, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_1 src1_sel:BYTE_1
+; GFX9-NODL-NEXT:    s_waitcnt vmcnt(0)
+; GFX9-NODL-NEXT:    v_mul_u32_u24_sdwa v1, v1, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_2 src1_sel:BYTE_2
+; GFX9-NODL-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX9-NODL-NEXT:    v_mad_u32_u24 v3, v4, v4, s0
+; GFX9-NODL-NEXT:    v_add3_u32 v1, v3, v2, v1
+; GFX9-NODL-NEXT:    global_store_dword v0, v1, s[6:7]
+; GFX9-NODL-NEXT:    s_endpgm
+;
+; GFX9-DL-LABEL: udot4_acc32_3src_3ele:
+; GFX9-DL:       ; %bb.0: ; %entry
+; GFX9-DL-NEXT:    s_load_dwordx8 s[0:7], s[0:1], 0x24
+; GFX9-DL-NEXT:    v_lshlrev_b32_e32 v0, 2, v0
+; GFX9-DL-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX9-DL-NEXT:    global_load_dword v1, v0, s[2:3]
+; GFX9-DL-NEXT:    global_load_dword v2, v0, s[4:5]
+; GFX9-DL-NEXT:    global_load_dword v3, v0, s[0:1]
+; GFX9-DL-NEXT:    s_load_dword s3, s[6:7], 0x0
+; GFX9-DL-NEXT:    s_mov_b32 s0, 0xc06010c
+; GFX9-DL-NEXT:    s_mov_b32 s1, 0xc0c0c00
+; GFX9-DL-NEXT:    s_mov_b32 s2, 0xc020100
+; GFX9-DL-NEXT:    v_mov_b32_e32 v0, 0
+; GFX9-DL-NEXT:    s_waitcnt vmcnt(1)
+; GFX9-DL-NEXT:    v_perm_b32 v1, v2, v1, s0
+; GFX9-DL-NEXT:    s_waitcnt vmcnt(0)
+; GFX9-DL-NEXT:    v_perm_b32 v2, v3, v3, s1
+; GFX9-DL-NEXT:    v_or_b32_e32 v1, v1, v2
+; GFX9-DL-NEXT:    v_perm_b32 v2, v3, v3, s2
+; GFX9-DL-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX9-DL-NEXT:    v_dot4_u32_u8 v1, v2, v1, s3
+; GFX9-DL-NEXT:    global_store_dword v0, v1, s[6:7]
+; GFX9-DL-NEXT:    s_endpgm
+;
+; GFX10-DL-LABEL: udot4_acc32_3src_3ele:
+; GFX10-DL:       ; %bb.0: ; %entry
+; GFX10-DL-NEXT:    s_load_dwordx8 s[0:7], s[0:1], 0x24
+; GFX10-DL-NEXT:    v_lshlrev_b32_e32 v0, 2, v0
+; GFX10-DL-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-DL-NEXT:    s_clause 0x2
+; GFX10-DL-NEXT:    global_load_dword v1, v0, s[2:3]
+; GFX10-DL-NEXT:    global_load_dword v2, v0, s[4:5]
+; GFX10-DL-NEXT:    global_load_dword v3, v0, s[0:1]
+; GFX10-DL-NEXT:    s_waitcnt_depctr 0xffe3
+; GFX10-DL-NEXT:    s_load_dword s0, s[6:7], 0x0
+; GFX10-DL-NEXT:    s_waitcnt vmcnt(1)
+; GFX10-DL-NEXT:    v_perm_b32 v0, v2, v1, 0xc06010c
+; GFX10-DL-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-DL-NEXT:    v_perm_b32 v1, v3, v3, 0xc0c0c00
+; GFX10-DL-NEXT:    v_mov_b32_e32 v2, 0
+; GFX10-DL-NEXT:    v_or_b32_e32 v0, v0, v1
+; GFX10-DL-NEXT:    v_perm_b32 v1, v3, v3, 0xc020100
+; GFX10-DL-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-DL-NEXT:    v_dot4_u32_u8 v0, v1, v0, s0
+; GFX10-DL-NEXT:    global_store_dword v2, v0, s[6:7]
+; GFX10-DL-NEXT:    s_endpgm
+;
+; GFX11-DL-LABEL: udot4_acc32_3src_3ele:
+; GFX11-DL:       ; %bb.0: ; %entry
+; GFX11-DL-NEXT:    s_load_b256 s[0:7], s[0:1], 0x24
+; GFX11-DL-NEXT:    v_lshlrev_b32_e32 v0, 2, v0
+; GFX11-DL-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX11-DL-NEXT:    s_clause 0x2
+; GFX11-DL-NEXT:    global_load_b32 v1, v0, s[2:3]
+; GFX11-DL-NEXT:    global_load_b32 v2, v0, s[4:5]
+; GFX11-DL-NEXT:    global_load_b32 v0, v0, s[0:1]
+; GFX11-DL-NEXT:    s_load_b32 s0, s[6:7], 0x0
+; GFX11-DL-NEXT:    s_waitcnt vmcnt(1)
+; GFX11-DL-NEXT:    v_perm_b32 v1, v2, v1, 0xc06010c
+; GFX11-DL-NEXT:    s_waitcnt vmcnt(0)
+; GFX11-DL-NEXT:    v_perm_b32 v2, v0, v0, 0xc0c0c00
+; GFX11-DL-NEXT:    v_perm_b32 v0, v0, v0, 0xc020100
+; GFX11-DL-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_2) | instid1(VALU_DEP_2)
+; GFX11-DL-NEXT:    v_or_b32_e32 v1, v1, v2
+; GFX11-DL-NEXT:    v_mov_b32_e32 v2, 0
+; GFX11-DL-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX11-DL-NEXT:    v_dot4_u32_u8 v0, v0, v1, s0
+; GFX11-DL-NEXT:    global_store_b32 v2, v0, s[6:7]
+; GFX11-DL-NEXT:    s_nop 0
+; GFX11-DL-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
+; GFX11-DL-NEXT:    s_endpgm
+                                       ptr addrspace(1) %src2,
+                                       ptr addrspace(1) %src3,
+                                       ptr addrspace(1) nocapture %dst) {
+entry:
+  %idx = call i32 @llvm.amdgcn.workitem.id.x()
+  %gep1 = getelementptr <4 x i8>, ptr addrspace(1) %src1, i32 %idx
+  %vec1 = load <4 x i8>, ptr addrspace(1) %gep1
+  %gep2 = getelementptr <4 x i8>, ptr addrspace(1) %src2, i32 %idx
+  %vec2 = load <4 x i8>, ptr addrspace(1) %gep2
+  %gep3 = getelementptr <4 x i8>, ptr addrspace(1) %src3, i32 %idx
+  %vec3 = load <4 x i8>, ptr addrspace(1) %gep3
+
+  %v1e0 = extractelement <4 x i8> %vec1, i64 0
+  %cv1e0 = zext i8 %v1e0 to i32
+  %mul1 = mul nuw nsw i32 %cv1e0, %cv1e0
+
+  %v1e1 = extractelement <4 x i8> %vec1, i64 1
+  %cv1e1 = zext i8 %v1e1 to i32
+  %v2e1 = extractelement <4 x i8> %vec2, i64 1
+  %cv2e1 = zext i8 %v2e1 to i32
+  %mul2 = mul nuw nsw i32 %cv1e1, %cv2e1
+
+  %v1e2 = extractelement <4 x i8> %vec1, i64 2
+  %cv1e2 = zext i8 %v1e2 to i32
+  %v3e2 = extractelement <4 x i8> %vec3, i64 2
+  %cv3e2 = zext i8 %v3e2 to i32
+  %mul3 = mul nuw nsw i32 %cv1e2, %cv3e2
+
+
+  %acc = load i32, ptr addrspace(1) %dst, align 4
+  %mad1 = add i32 %mul1, %acc
+  %mad2 = add i32 %mad1, %mul2
+  %mad3 = add i32 %mad2, %mul3
+
+  store i32 %mad3, ptr addrspace(1) %dst, align 4
+  ret void
+}
+
+define amdgpu_kernel void @udot4_bad_source(ptr addrspace(1) %src1,
+; GFX7-LABEL: udot4_bad_source:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x9
+; GFX7-NEXT:    s_load_dword s12, s[0:1], 0xf
+; GFX7-NEXT:    s_mov_b32 s3, 0xf000
+; GFX7-NEXT:    s_mov_b32 s10, 0
+; GFX7-NEXT:    s_mov_b32 s11, s3
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    s_mov_b64 s[8:9], s[4:5]
+; GFX7-NEXT:    v_lshlrev_b32_e32 v0, 2, v0
+; GFX7-NEXT:    v_mov_b32_e32 v1, 0
+; GFX7-NEXT:    buffer_load_dword v2, v[0:1], s[8:11], 0 addr64
+; GFX7-NEXT:    s_mov_b64 s[8:9], s[6:7]
+; GFX7-NEXT:    buffer_load_dword v0, v[0:1], s[8:11], 0 addr64
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0x11
+; GFX7-NEXT:    s_and_b32 s5, s12, 0xffff
+; GFX7-NEXT:    s_mov_b32 s2, -1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    s_load_dword s4, s[0:1], 0x0
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mov_b32_e32 v1, s4
+; GFX7-NEXT:    s_waitcnt vmcnt(1)
+; GFX7-NEXT:    v_and_b32_e32 v3, 0xff, v2
+; GFX7-NEXT:    v_bfe_u32 v4, v2, 8, 8
+; GFX7-NEXT:    s_waitcnt vmcnt(0)
+; GFX7-NEXT:    v_bfe_u32 v5, v0, 8, 8
+; GFX7-NEXT:    v_mad_u32_u24 v1, v3, s5, v1
+; GFX7-NEXT:    v_bfe_u32 v2, v2, 16, 8
+; GFX7-NEXT:    v_bfe_u32 v0, v0, 16, 8
+; GFX7-NEXT:    v_mad_u32_u24 v1, v4, v5, v1
+; GFX7-NEXT:    v_mad_u32_u24 v0, v2, v0, v1
+; GFX7-NEXT:    buffer_store_dword v0, off, s[0:3], 0
+; GFX7-NEXT:    s_endpgm
+;
+; GFX8-LABEL: udot4_bad_source:
+; GFX8:       ; %bb.0: ; %entry
+; GFX8-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x24
+; GFX8-NEXT:    s_load_dword s2, s[0:1], 0x3c
+; GFX8-NEXT:    v_lshlrev_b32_e32 v2, 2, v0
+; GFX8-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0x44
+; GFX8-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX8-NEXT:    v_mov_b32_e32 v1, s5
+; GFX8-NEXT:    v_add_u32_e32 v0, vcc, s4, v2
+; GFX8-NEXT:    v_addc_u32_e32 v1, vcc, 0, v1, vcc
+; GFX8-NEXT:    flat_load_dword v3, v[0:1]
+; GFX8-NEXT:    v_mov_b32_e32 v1, s7
+; GFX8-NEXT:    v_add_u32_e32 v0, vcc, s6, v2
+; GFX8-NEXT:    v_addc_u32_e32 v1, vcc, 0, v1, vcc
+; GFX8-NEXT:    flat_load_dword v0, v[0:1]
+; GFX8-NEXT:    s_load_dword s3, s[0:1], 0x0
+; GFX8-NEXT:    s_and_b32 s2, s2, 0xffff
+; GFX8-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX8-NEXT:    v_mov_b32_e32 v1, s3
+; GFX8-NEXT:    s_waitcnt vmcnt(1)
+; GFX8-NEXT:    v_and_b32_e32 v2, 0xff, v3
+; GFX8-NEXT:    v_bfe_u32 v4, v3, 8, 8
+; GFX8-NEXT:    v_mad_u32_u24 v1, v2, s2, v1
+; GFX8-NEXT:    v_bfe_u32 v3, v3, 16, 8
+; GFX8-NEXT:    s_waitcnt vmcnt(0)
+; GFX8-NEXT:    v_bfe_u32 v5, v0, 8, 8
+; GFX8-NEXT:    v_bfe_u32 v0, v0, 16, 8
+; GFX8-NEXT:    v_mad_u32_u24 v1, v4, v5, v1
+; GFX8-NEXT:    v_mad_u32_u24 v2, v3, v0, v1
+; GFX8-NEXT:    v_mov_b32_e32 v0, s0
+; GFX8-NEXT:    v_mov_b32_e32 v1, s1
+; GFX8-NEXT:    flat_store_dword v[0:1], v2
+; GFX8-NEXT:    s_endpgm
+;
+; GFX9-NODL-LABEL: udot4_bad_source:
+; GFX9-NODL:       ; %bb.0: ; %entry
+; GFX9-NODL-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x24
+; GFX9-NODL-NEXT:    s_load_dword s2, s[0:1], 0x3c
+; GFX9-NODL-NEXT:    v_lshlrev_b32_e32 v0, 2, v0
+; GFX9-NODL-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0x44
+; GFX9-NODL-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX9-NODL-NEXT:    global_load_dword v1, v0, s[4:5]
+; GFX9-NODL-NEXT:    global_load_dword v2, v0, s[6:7]
+; GFX9-NODL-NEXT:    s_and_b32 s2, s2, 0xffff
+; GFX9-NODL-NEXT:    v_mov_b32_e32 v0, 0
+; GFX9-NODL-NEXT:    s_load_dword s3, s[0:1], 0x0
+; GFX9-NODL-NEXT:    s_waitcnt vmcnt(1)
+; GFX9-NODL-NEXT:    v_and_b32_e32 v3, 0xff, v1
+; GFX9-NODL-NEXT:    s_waitcnt vmcnt(0)
+; GFX9-NODL-NEXT:    v_mul_u32_u24_sdwa v4, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_1 src1_sel:BYTE_1
+; GFX9-NODL-NEXT:    v_mul_u32_u24_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_2 src1_sel:BYTE_2
+; GFX9-NODL-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX9-NODL-NEXT:    v_mov_b32_e32 v2, s3
+; GFX9-NODL-NEXT:    v_mad_u32_u24 v2, v3, s2, v2
+; GFX9-NODL-NEXT:    v_add3_u32 v1, v2, v4, v1
+; GFX9-NODL-NEXT:    global_store_dword v0, v1, s[0:1]
+; GFX9-NODL-NEXT:    s_endpgm
+;
+; GFX9-DL-LABEL: udot4_bad_source:
+; GFX9-DL:       ; %bb.0: ; %entry
+; GFX9-DL-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x24
+; GFX9-DL-NEXT:    s_load_dword s2, s[0:1], 0x3c
+; GFX9-DL-NEXT:    v_lshlrev_b32_e32 v0, 2, v0
+; GFX9-DL-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0x44
+; GFX9-DL-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX9-DL-NEXT:    global_load_dword v1, v0, s[4:5]
+; GFX9-DL-NEXT:    global_load_dword v2, v0, s[6:7]
+; GFX9-DL-NEXT:    s_mov_b32 s4, 0xc0c0201
+; GFX9-DL-NEXT:    s_and_b32 s2, s2, 0xffff
+; GFX9-DL-NEXT:    s_load_dword s3, s[0:1], 0x0
+; GFX9-DL-NEXT:    v_mov_b32_e32 v0, 0
+; GFX9-DL-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX9-DL-NEXT:    v_mov_b32_e32 v3, s3
+; GFX9-DL-NEXT:    s_waitcnt vmcnt(1)
+; GFX9-DL-NEXT:    v_and_b32_e32 v4, 0xff, v1
+; GFX9-DL-NEXT:    s_waitcnt vmcnt(0)
+; GFX9-DL-NEXT:    v_perm_b32 v2, v2, v2, s4
+; GFX9-DL-NEXT:    v_mad_u32_u24 v3, v4, s2, v3
+; GFX9-DL-NEXT:    v_perm_b32 v1, v1, v1, s4
+; GFX9-DL-NEXT:    v_dot4_u32_u8 v1, v1, v2, v3
+; GFX9-DL-NEXT:    global_store_dword v0, v1, s[0:1]
+; GFX9-DL-NEXT:    s_endpgm
+;
+; GFX10-DL-LABEL: udot4_bad_source:
+; GFX10-DL:       ; %bb.0: ; %entry
+; GFX10-DL-NEXT:    s_clause 0x1
+; GFX10-DL-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x24
+; GFX10-DL-NEXT:    s_load_dword s2, s[0:1], 0x3c
+; GFX10-DL-NEXT:    v_lshlrev_b32_e32 v0, 2, v0
+; GFX10-DL-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0x44
+; GFX10-DL-NEXT:    v_mov_b32_e32 v3, 0
+; GFX10-DL-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-DL-NEXT:    s_clause 0x1
+; GFX10-DL-NEXT:    global_load_dword v1, v0, s[4:5]
+; GFX10-DL-NEXT:    global_load_dword v2, v0, s[6:7]
+; GFX10-DL-NEXT:    s_and_b32 s2, s2, 0xffff
+; GFX10-DL-NEXT:    s_load_dword s3, s[0:1], 0x0
+; GFX10-DL-NEXT:    s_waitcnt vmcnt(1)
+; GFX10-DL-NEXT:    v_and_b32_e32 v0, 0xff, v1
+; GFX10-DL-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-DL-NEXT:    v_perm_b32 v2, v2, v2, 0xc0c0201
+; GFX10-DL-NEXT:    v_perm_b32 v1, v1, v1, 0xc0c0201
+; GFX10-DL-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-DL-NEXT:    v_mad_u32_u24 v0, v0, s2, s3
+; GFX10-DL-NEXT:    v_dot4_u32_u8 v0, v1, v2, v0
+; GFX10-DL-NEXT:    global_store_dword v3, v0, s[0:1]
+; GFX10-DL-NEXT:    s_endpgm
+;
+; GFX11-DL-LABEL: udot4_bad_source:
+; GFX11-DL:       ; %bb.0: ; %entry
+; GFX11-DL-NEXT:    s_clause 0x1
+; GFX11-DL-NEXT:    s_load_b128 s[4:7], s[0:1], 0x24
+; GFX11-DL-NEXT:    s_load_b32 s2, s[0:1], 0x3c
+; GFX11-DL-NEXT:    v_dual_mov_b32 v3, 0 :: v_dual_lshlrev_b32 v0, 2, v0
+; GFX11-DL-NEXT:    s_load_b64 s[0:1], s[0:1], 0x44
+; GFX11-DL-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX11-DL-NEXT:    s_clause 0x1
+; GFX11-DL-NEXT:    global_load_b32 v1, v0, s[4:5]
+; GFX11-DL-NEXT:    global_load_b32 v0, v0, s[6:7]
+; GFX11-DL-NEXT:    s_and_b32 s2, s2, 0xffff
+; GFX11-DL-NEXT:    s_load_b32 s3, s[0:1], 0x0
+; GFX11-DL-NEXT:    s_waitcnt vmcnt(1)
+; GFX11-DL-NEXT:    v_and_b32_e32 v2, 0xff, v1
+; GFX11-DL-NEXT:    s_waitcnt vmcnt(0)
+; GFX11-DL-NEXT:    v_perm_b32 v0, v0, v0, 0xc0c0201
+; GFX11-DL-NEXT:    v_perm_b32 v1, v1, v1, 0xc0c0201
+; GFX11-DL-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX11-DL-NEXT:    v_mad_u32_u24 v2, v2, s2, s3
+; GFX11-DL-NEXT:    s_delay_alu instid0(VALU_DEP_1)
+; GFX11-DL-NEXT:    v_dot4_u32_u8 v0, v1, v0, v2
+; GFX11-DL-NEXT:    global_store_b32 v3, v0, s[0:1]
+; GFX11-DL-NEXT:    s_nop 0
+; GFX11-DL-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
+; GFX11-DL-NEXT:    s_endpgm
+                                       ptr addrspace(1) %src2,
+                                       ptr addrspace(1) %src3,
+                                       i16 %badsource,
+                                       ptr addrspace(1) nocapture %dst) {
+entry:
+  %idx = call i32 @llvm.amdgcn.workitem.id.x()
+  %gep1 = getelementptr <4 x i8>, ptr addrspace(1) %src1, i32 %idx
+  %vec1 = load <4 x i8>, ptr addrspace(1) %gep1
+  %gep2 = getelementptr <4 x i8>, ptr addrspace(1) %src2, i32 %idx
+  %vec2 = load <4 x i8>, ptr addrspace(1) %gep2
+  %gep3 = getelementptr <4 x i8>, ptr addrspace(1) %src3, i32 %idx
+  %vec3 = load <4 x i8>, ptr addrspace(1) %gep3
+
+  %v1e0 = extractelement <4 x i8> %vec1, i64 0
+  %cv1e0 = zext i8 %v1e0 to i32
+  %v2e0 = extractelement <4 x i8> %vec2, i64 0
+  %other = zext i16 %badsource to i32
+  %mul1 = mul nuw nsw i32 %cv1e0, %other
+
+  %v1e1 = extractelement <4 x i8> %vec1, i64 1
+  %cv1e1 = zext i8 %v1e1 to i32
+  %v2e1 = extractelement <4 x i8> %vec2, i64 1
+  %cv2e1 = zext i8 %v2e1 to i32
+  %mul2 = mul nuw nsw i32 %cv1e1, %cv2e1
+
+  %v2e2 = extractelement <4 x i8> %vec2, i64 2
+  %cv2e2 = zext i8 %v2e2 to i32
+  %v1e2 = extractelement <4 x i8> %vec1, i64 2
+  %cv1e2 = zext i8 %v1e2 to i32
+  %mul3 = mul nuw nsw i32 %cv1e2, %cv2e2
+
+
+  %acc = load i32, ptr addrspace(1) %dst, align 4
+  %mad1 = add i32 %mul1, %acc
+  %mad2 = add i32 %mad1, %mul2
+  %mad3 = add i32 %mad2, %mul3
+
+  store i32 %mad3, ptr addrspace(1) %dst, align 4
+  ret void
+}
+
+
+define amdgpu_kernel void @udot4_commutative(ptr addrspace(1) %src1,
+; GFX7-LABEL: udot4_commutative:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x9
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xf
+; GFX7-NEXT:    s_mov_b32 s3, 0xf000
+; GFX7-NEXT:    s_mov_b32 s10, 0
+; GFX7-NEXT:    s_mov_b32 s11, s3
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    s_mov_b64 s[8:9], s[4:5]
+; GFX7-NEXT:    v_lshlrev_b32_e32 v0, 2, v0
+; GFX7-NEXT:    v_mov_b32_e32 v1, 0
+; GFX7-NEXT:    buffer_load_dword v2, v[0:1], s[8:11], 0 addr64
+; GFX7-NEXT:    s_mov_b64 s[8:9], s[6:7]
+; GFX7-NEXT:    buffer_load_dword v0, v[0:1], s[8:11], 0 addr64
+; GFX7-NEXT:    s_load_dword s4, s[0:1], 0x0
+; GFX7-NEXT:    s_mov_b32 s2, -1
+; GFX7-NEXT:    s_waitcnt vmcnt(1)
+; GFX7-NEXT:    v_and_b32_e32 v1, 0xff, v2
+; GFX7-NEXT:    v_bfe_u32 v3, v2, 8, 8
+; GFX7-NEXT:    s_waitcnt vmcnt(0)
+; GFX7-NEXT:    v_and_b32_e32 v4, 0xff, v0
+; GFX7-NEXT:    v_bfe_u32 v5, v0, 8, 8
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mad_u32_u24 v1, v1, v4, s4
+; GFX7-NEXT:    v_bfe_u32 v0, v0, 16, 8
+; GFX7-NEXT:    v_bfe_u32 v2, v2, 16, 8
+; GFX7-NEXT:    v_mad_u32_u24 v1, v3, v5, v1
+; GFX7-NEXT:    v_mad_u32_u24 v0, v2, v0, v1
+; GFX7-NEXT:    buffer_store_dword v0, off, s[0:3], 0
+; GFX7-NEXT:    s_endpgm
+;
+; GFX8-LABEL: udot4_commutative:
+; GFX8:       ; %bb.0: ; %entry
+; GFX8-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x24
+; GFX8-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0x3c
+; GFX8-NEXT:    v_lshlrev_b32_e32 v2, 2, v0
+; GFX8-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX8-NEXT:    v_mov_b32_e32 v1, s5
+; GFX8-NEXT:    v_add_u32_e32 v0, vcc, s4, v2
+; GFX8-NEXT:    v_addc_u32_e32 v1, vcc, 0, v1, vcc
+; GFX8-NEXT:    flat_load_dword v3, v[0:1]
+; GFX8-NEXT:    v_mov_b32_e32 v1, s7
+; GFX8-NEXT:    v_add_u32_e32 v0, vcc, s6, v2
+; GFX8-NEXT:    v_addc_u32_e32 v1, vcc, 0, v1, vcc
+; GFX8-NEXT:    flat_load_dword v0, v[0:1]
+; GFX8-NEXT:    s_load_dword s2, s[0:1], 0x0
+; GFX8-NEXT:    s_waitcnt vmcnt(1)
+; GFX8-NEXT:    v_and_b32_e32 v1, 0xff, v3
+; GFX8-NEXT:    v_bfe_u32 v4, v3, 8, 8
+; GFX8-NEXT:    v_bfe_u32 v3, v3, 16, 8
+; GFX8-NEXT:    s_waitcnt vmcnt(0)
+; GFX8-NEXT:    v_and_b32_e32 v2, 0xff, v0
+; GFX8-NEXT:    v_bfe_u32 v5, v0, 8, 8
+; GFX8-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX8-NEXT:    v_mad_u32_u24 v1, v1, v2, s2
+; GFX8-NEXT:    v_bfe_u32 v0, v0, 16, 8
+; GFX8-NEXT:    v_mad_u32_u24 v1, v4, v5, v1
+; GFX8-NEXT:    v_mad_u32_u24 v2, v3, v0, v1
+; GFX8-NEXT:    v_mov_b32_e32 v0, s0
+; GFX8-NEXT:    v_mov_b32_e32 v1, s1
+; GFX8-NEXT:    flat_store_dword v[0:1], v2
+; GFX8-NEXT:    s_endpgm
+;
+; GFX9-NODL-LABEL: udot4_commutative:
+; GFX9-NODL:       ; %bb.0: ; %entry
+; GFX9-NODL-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x24
+; GFX9-NODL-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x3c
+; GFX9-NODL-NEXT:    v_lshlrev_b32_e32 v0, 2, v0
+; GFX9-NODL-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX9-NODL-NEXT:    global_load_dword v1, v0, s[4:5]
+; GFX9-NODL-NEXT:    global_load_dword v2, v0, s[6:7]
+; GFX9-NODL-NEXT:    s_load_dword s0, s[2:3], 0x0
+; GFX9-NODL-NEXT:    v_mov_b32_e32 v0, 0
+; GFX9-NODL-NEXT:    s_waitcnt vmcnt(1)
+; GFX9-NODL-NEXT:    v_and_b32_e32 v3, 0xff, v1
+; GFX9-NODL-NEXT:    s_waitcnt vmcnt(0)
+; GFX9-NODL-NEXT:    v_and_b32_e32 v4, 0xff, v2
+; GFX9-NODL-NEXT:    v_mul_u32_u24_sdwa v5, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_1 src1_sel:BYTE_1
+; GFX9-NODL-NEXT:    v_mul_u32_u24_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_2 src1_sel:BYTE_2
+; GFX9-NODL-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX9-NODL-NEXT:    v_mad_u32_u24 v2, v3, v4, s0
+; GFX9-NODL-NEXT:    v_add3_u32 v1, v2, v5, v1
+; GFX9-NODL-NEXT:    global_store_dword v0, v1, s[2:3]
+; GFX9-NODL-NEXT:    s_endpgm
+;
+; GFX9-DL-LABEL: udot4_commutative:
+; GFX9-DL:       ; %bb.0: ; %entry
+; GFX9-DL-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x24
+; GFX9-DL-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x3c
+; GFX9-DL-NEXT:    v_lshlrev_b32_e32 v0, 2, v0
+; GFX9-DL-NEXT:    s_mov_b32 s1, 0xc020100
+; GFX9-DL-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX9-DL-NEXT:    global_load_dword v1, v0, s[6:7]
+; GFX9-DL-NEXT:    global_load_dword v2, v0, s[4:5]
+; GFX9-DL-NEXT:    s_load_dword s0, s[2:3], 0x0
+; GFX9-DL-NEXT:    v_mov_b32_e32 v0, 0
+; GFX9-DL-NEXT:    s_waitcnt vmcnt(1)
+; GFX9-DL-NEXT:    v_perm_b32 v1, v1, v1, s1
+; GFX9-DL-NEXT:    s_waitcnt vmcnt(0)
+; GFX9-DL-NEXT:    v_perm_b32 v2, v2, v2, s1
+; GFX9-DL-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX9-DL-NEXT:    v_dot4_u32_u8 v1, v2, v1, s0
+; GFX9-DL-NEXT:    global_store_dword v0, v1, s[2:3]
+; GFX9-DL-NEXT:    s_endpgm
+;
+; GFX10-DL-LABEL: udot4_commutative:
+; GFX10-DL:       ; %bb.0: ; %entry
+; GFX10-DL-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x24
+; GFX10-DL-NEXT:    v_lshlrev_b32_e32 v0, 2, v0
+; GFX10-DL-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0x3c
+; GFX10-DL-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-DL-NEXT:    s_clause 0x1
+; GFX10-DL-NEXT:    global_load_dword v1, v0, s[6:7]
+; GFX10-DL-NEXT:    global_load_dword v2, v0, s[4:5]
+; GFX10-DL-NEXT:    s_load_dword s2, s[0:1], 0x0
+; GFX10-DL-NEXT:    s_waitcnt vmcnt(1)
+; GFX10-DL-NEXT:    v_perm_b32 v0, v1, v1, 0xc020100
+; GFX10-DL-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-DL-NEXT:    v_perm_b32 v1, v2, v2, 0xc020100
+; GFX10-DL-NEXT:    v_mov_b32_e32 v2, 0
+; GFX10-DL-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-DL-NEXT:    v_dot4_u32_u8 v0, v1, v0, s2
+; GFX10-DL-NEXT:    global_store_dword v2, v0, s[0:1]
+; GFX10-DL-NEXT:    s_endpgm
+;
+; GFX11-DL-LABEL: udot4_commutative:
+; GFX11-DL:       ; %bb.0: ; %entry
+; GFX11-DL-NEXT:    s_load_b128 s[4:7], s[0:1], 0x24
+; GFX11-DL-NEXT:    v_lshlrev_b32_e32 v0, 2, v0
+; GFX11-DL-NEXT:    s_load_b64 s[0:1], s[0:1], 0x3c
+; GFX11-DL-NEXT:    v_mov_b32_e32 v2, 0
+; GFX11-DL-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX11-DL-NEXT:    s_clause 0x1
+; GFX11-DL-NEXT:    global_load_b32 v1, v0, s[6:7]
+; GFX11-DL-NEXT:    global_load_b32 v0, v0, s[4:5]
+; GFX11-DL-NEXT:    s_load_b32 s2, s[0:1], 0x0
+; GFX11-DL-NEXT:    s_waitcnt vmcnt(1)
+; GFX11-DL-NEXT:    v_perm_b32 v1, v1, v1, 0xc020100
+; GFX11-DL-NEXT:    s_waitcnt vmcnt(0)
+; GFX11-DL-NEXT:    v_perm_b32 v0, v0, v0, 0xc020100
+; GFX11-DL-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX11-DL-NEXT:    s_delay_alu instid0(VALU_DEP_1)
+; GFX11-DL-NEXT:    v_dot4_u32_u8 v0, v0, v1, s2
+; GFX11-DL-NEXT:    global_store_b32 v2, v0, s[0:1]
+; GFX11-DL-NEXT:    s_nop 0
+; GFX11-DL-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
+; GFX11-DL-NEXT:    s_endpgm
+                                       ptr addrspace(1) %src2,
+                                       ptr addrspace(1) %src3,
+                                       ptr addrspace(1) nocapture %dst) {
+entry:
+  %idx = call i32 @llvm.amdgcn.workitem.id.x()
+  %gep1 = getelementptr <4 x i8>, ptr addrspace(1) %src1, i32 %idx
+  %vec1 = load <4 x i8>, ptr addrspace(1) %gep1
+  %gep2 = getelementptr <4 x i8>, ptr addrspace(1) %src2, i32 %idx
+  %vec2 = load <4 x i8>, ptr addrspace(1) %gep2
+  %gep3 = getelementptr <4 x i8>, ptr addrspace(1) %src3, i32 %idx
+  %vec3 = load <4 x i8>, ptr addrspace(1) %gep3
+
+  %v1e0 = extractelement <4 x i8> %vec1, i64 0
+  %cv1e0 = zext i8 %v1e0 to i32
+  %v2e0 = extractelement <4 x i8> %vec2, i64 0
+  %cv2e0 = zext i8 %v2e0 to i32
+  %mul1 = mul nuw nsw i32 %cv1e0, %cv2e0
+
+  %v1e1 = extractelement <4 x i8> %vec1, i64 1
+  %cv1e1 = zext i8 %v1e1 to i32
+  %v2e1 = extractelement <4 x i8> %vec2, i64 1
+  %cv2e1 = zext i8 %v2e1 to i32
+  %mul2 = mul nuw nsw i32 %cv1e1, %cv2e1
+
+  %v2e2 = extractelement <4 x i8> %vec2, i64 2
+  %cv2e2 = zext i8 %v2e2 to i32
+  %v1e2 = extractelement <4 x i8> %vec1, i64 2
+  %cv1e2 = zext i8 %v1e2 to i32
+  %mul3 = mul nuw nsw i32 %cv1e2, %cv2e2
+
+
+  %acc = load i32, ptr addrspace(1) %dst, align 4
+  %mad1 = add i32 %mul1, %acc
+  %mad2 = add i32 %mad1, %mul2
+  %mad3 = add i32 %mad2, %mul3
+
+  store i32 %mad3, ptr addrspace(1) %dst, align 4
+  ret void
+}
+
+define amdgpu_kernel void @udot4_acc32_3src_3ele_src0(ptr addrspace(1) %src1,
+; GFX7-LABEL: udot4_acc32_3src_3ele_src0:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx8 s[0:7], s[0:1], 0x9
+; GFX7-NEXT:    s_mov_b32 s11, 0xf000
+; GFX7-NEXT:    s_mov_b32 s14, 0
+; GFX7-NEXT:    s_mov_b32 s15, s11
+; GFX7-NEXT:    v_lshlrev_b32_e32 v0, 2, v0
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    s_mov_b64 s[12:13], s[0:1]
+; GFX7-NEXT:    v_mov_b32_e32 v1, 0
+; GFX7-NEXT:    buffer_load_dword v2, v[0:1], s[12:15], 0 addr64
+; GFX7-NEXT:    s_mov_b64 s[12:13], s[2:3]
+; GFX7-NEXT:    buffer_load_dword v3, v[0:1], s[12:15], 0 addr64
+; GFX7-NEXT:    s_mov_b64 s[12:13], s[4:5]
+; GFX7-NEXT:    buffer_load_dword v0, v[0:1], s[12:15], 0 addr64
+; GFX7-NEXT:    s_load_dword s0, s[6:7], 0x0
+; GFX7-NEXT:    s_mov_b32 s10, -1
+; GFX7-NEXT:    s_mov_b32 s8, s6
+; GFX7-NEXT:    s_mov_b32 s9, s7
+; GFX7-NEXT:    s_waitcnt vmcnt(2)
+; GFX7-NEXT:    v_bfe_u32 v1, v2, 8, 8
+; GFX7-NEXT:    s_waitcnt vmcnt(1)
+; GFX7-NEXT:    v_bfe_u32 v2, v3, 8, 8
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mad_u32_u24 v4, v2, v2, s0
+; GFX7-NEXT:    v_bfe_u32 v3, v3, 16, 8
+; GFX7-NEXT:    s_waitcnt vmcnt(0)
+; GFX7-NEXT:    v_bfe_u32 v0, v0, 16, 8
+; GFX7-NEXT:    v_mad_u32_u24 v1, v1, v2, v4
+; GFX7-NEXT:    v_mad_u32_u24 v0, v3, v0, v1
+; GFX7-NEXT:    buffer_store_dword v0, off, s[8:11], 0
+; GFX7-NEXT:    s_endpgm
+;
+; GFX8-LABEL: udot4_acc32_3src_3ele_src0:
+; GFX8:       ; %bb.0: ; %entry
+; GFX8-NEXT:    s_load_dwordx8 s[0:7], s[0:1], 0x24
+; GFX8-NEXT:    v_lshlrev_b32_e32 v2, 2, v0
+; GFX8-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX8-NEXT:    v_mov_b32_e32 v1, s1
+; GFX8-NEXT:    v_add_u32_e32 v0, vcc, s0, v2
+; GFX8-NEXT:    v_addc_u32_e32 v1, vcc, 0, v1, vcc
+; GFX8-NEXT:    flat_load_dword v3, v[0:1]
+; GFX8-NEXT:    v_mov_b32_e32 v1, s3
+; GFX8-NEXT:    v_add_u32_e32 v0, vcc, s2, v2
+; GFX8-NEXT:    v_addc_u32_e32 v1, vcc, 0, v1, vcc
+; GFX8-NEXT:    flat_load_dword v4, v[0:1]
+; GFX8-NEXT:    v_mov_b32_e32 v1, s5
+; GFX8-NEXT:    v_add_u32_e32 v0, vcc, s4, v2
+; GFX8-NEXT:    v_addc_u32_e32 v1, vcc, 0, v1, vcc
+; GFX8-NEXT:    flat_load_dword v0, v[0:1]
+; GFX8-NEXT:    s_load_dword s0, s[6:7], 0x0
+; GFX8-NEXT:    s_waitcnt vmcnt(2)
+; GFX8-NEXT:    v_bfe_u32 v2, v3, 8, 8
+; GFX8-NEXT:    s_waitcnt vmcnt(1)
+; GFX8-NEXT:    v_bfe_u32 v1, v4, 8, 8
+; GFX8-NEXT:    v_bfe_u32 v3, v4, 16, 8
+; GFX8-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX8-NEXT:    v_mad_u32_u24 v4, v1, v1, s0
+; GFX8-NEXT:    v_mad_u32_u24 v1, v2, v1, v4
+; GFX8-NEXT:    s_waitcnt vmcnt(0)
+; GFX8-NEXT:    v_bfe_u32 v0, v0, 16, 8
+; GFX8-NEXT:    v_mad_u32_u24 v2, v3, v0, v1
+; GFX8-NEXT:    v_mov_b32_e32 v0, s6
+; GFX8-NEXT:    v_mov_b32_e32 v1, s7
+; GFX8-NEXT:    flat_store_dword v[0:1], v2
+; GFX8-NEXT:    s_endpgm
+;
+; GFX9-NODL-LABEL: udot4_acc32_3src_3ele_src0:
+; GFX9-NODL:       ; %bb.0: ; %entry
+; GFX9-NODL-NEXT:    s_load_dwordx8 s[0:7], s[0:1], 0x24
+; GFX9-NODL-NEXT:    v_lshlrev_b32_e32 v0, 2, v0
+; GFX9-NODL-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX9-NODL-NEXT:    global_load_dword v1, v0, s[2:3]
+; GFX9-NODL-NEXT:    global_load_dword v2, v0, s[4:5]
+; GFX9-NODL-NEXT:    global_load_dword v3, v0, s[0:1]
+; GFX9-NODL-NEXT:    s_load_dword s0, s[6:7], 0x0
+; GFX9-NODL-NEXT:    v_mov_b32_e32 v0, 0
+; GFX9-NODL-NEXT:    s_waitcnt vmcnt(2)
+; GFX9-NODL-NEXT:    v_bfe_u32 v4, v1, 8, 8
+; GFX9-NODL-NEXT:    s_waitcnt vmcnt(1)
+; GFX9-NODL-NEXT:    v_mul_u32_u24_sdwa v1, v1, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_2 src1_sel:BYTE_2
+; GFX9-NODL-NEXT:    s_waitcnt vmcnt(0)
+; GFX9-NODL-NEXT:    v_mul_u32_u24_sdwa v2, v3, v4 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_1 src1_sel:DWORD
+; GFX9-NODL-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX9-NODL-NEXT:    v_mad_u32_u24 v3, v4, v4, s0
+; GFX9-NODL-NEXT:    v_add3_u32 v1, v3, v2, v1
+; GFX9-NODL-NEXT:    global_store_dword v0, v1, s[6:7]
+; GFX9-NODL-NEXT:    s_endpgm
+;
+; GFX9-DL-LABEL: udot4_acc32_3src_3ele_src0:
+; GFX9-DL:       ; %bb.0: ; %entry
+; GFX9-DL-NEXT:    s_load_dwordx8 s[0:7], s[0:1], 0x24
+; GFX9-DL-NEXT:    v_lshlrev_b32_e32 v0, 2, v0
+; GFX9-DL-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX9-DL-NEXT:    global_load_dword v1, v0, s[4:5]
+; GFX9-DL-NEXT:    global_load_dword v2, v0, s[0:1]
+; GFX9-DL-NEXT:    global_load_dword v3, v0, s[2:3]
+; GFX9-DL-NEXT:    s_load_dword s3, s[6:7], 0x0
+; GFX9-DL-NEXT:    s_mov_b32 s0, 0xc06010c
+; GFX9-DL-NEXT:    s_mov_b32 s1, 0xc0c0c01
+; GFX9-DL-NEXT:    s_mov_b32 s2, 0xc020101
+; GFX9-DL-NEXT:    v_mov_b32_e32 v0, 0
+; GFX9-DL-NEXT:    s_waitcnt vmcnt(1)
+; GFX9-DL-NEXT:    v_perm_b32 v1, v1, v2, s0
+; GFX9-DL-NEXT:    s_waitcnt vmcnt(0)
+; GFX9-DL-NEXT:    v_perm_b32 v2, v3, v3, s1
+; GFX9-DL-NEXT:    v_or_b32_e32 v1, v1, v2
+; GFX9-DL-NEXT:    v_perm_b32 v2, v3, v3, s2
+; GFX9-DL-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX9-DL-NEXT:    v_dot4_u32_u8 v1, v2, v1, s3
+; GFX9-DL-NEXT:    global_store_dword v0, v1, s[6:7]
+; GFX9-DL-NEXT:    s_endpgm
+;
+; GFX10-DL-LABEL: udot4_acc32_3src_3ele_src0:
+; GFX10-DL:       ; %bb.0: ; %entry
+; GFX10-DL-NEXT:    s_load_dwordx8 s[0:7], s[0:1], 0x24
+; GFX10-DL-NEXT:    v_lshlrev_b32_e32 v0, 2, v0
+; GFX10-DL-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-DL-NEXT:    s_clause 0x2
+; GFX10-DL-NEXT:    global_load_dword v1, v0, s[4:5]
+; GFX10-DL-NEXT:    global_load_dword v2, v0, s[0:1]
+; GFX10-DL-NEXT:    global_load_dword v3, v0, s[2:3]
+; GFX10-DL-NEXT:    s_waitcnt_depctr 0xffe3
+; GFX10-DL-NEXT:    s_load_dword s0, s[6:7], 0x0
+; GFX10-DL-NEXT:    s_waitcnt vmcnt(1)
+; GFX10-DL-NEXT:    v_perm_b32 v0, v1, v2, 0xc06010c
+; GFX10-DL-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-DL-NEXT:    v_perm_b32 v1, v3, v3, 0xc0c0c01
+; GFX10-DL-NEXT:    v_mov_b32_e32 v2, 0
+; GFX10-DL-NEXT:    v_or_b32_e32 v0, v0, v1
+; GFX10-DL-NEXT:    v_perm_b32 v1, v3, v3, 0xc020101
+; GFX10-DL-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-DL-NEXT:    v_dot4_u32_u8 v0, v1, v0, s0
+; GFX10-DL-NEXT:    global_store_dword v2, v0, s[6:7]
+; GFX10-DL-NEXT:    s_endpgm
+;
+; GFX11-DL-LABEL: udot4_acc32_3src_3ele_src0:
+; GFX11-DL:       ; %bb.0: ; %entry
+; GFX11-DL-NEXT:    s_load_b256 s[0:7], s[0:1], 0x24
+; GFX11-DL-NEXT:    v_lshlrev_b32_e32 v0, 2, v0
+; GFX11-DL-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX11-DL-NEXT:    s_clause 0x2
+; GFX11-DL-NEXT:    global_load_b32 v1, v0, s[4:5]
+; GFX11-DL-NEXT:    global_load_b32 v2, v0, s[0:1]
+; GFX11-DL-NEXT:    global_load_b32 v0, v0, s[2:3]
+; GFX11-DL-NEXT:    s_load_b32 s0, s[6:7], 0x0
+; GFX11-DL-NEXT:    s_waitcnt vmcnt(1)
+; GFX11-DL-NEXT:    v_perm_b32 v1, v1, v2, 0xc06010c
+; GFX11-DL-NEXT:    s_waitcnt vmcnt(0)
+; GFX11-DL-NEXT:    v_perm_b32 v2, v0, v0, 0xc0c0c01
+; GFX11-DL-NEXT:    v_perm_b32 v0, v0, v0, 0xc020101
+; GFX11-DL-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(SKIP_2) | instid1(VALU_DEP_2)
+; GFX11-DL-NEXT:    v_or_b32_e32 v1, v1, v2
+; GFX11-DL-NEXT:    v_mov_b32_e32 v2, 0
+; GFX11-DL-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX11-DL-NEXT:    v_dot4_u32_u8 v0, v0, v1, s0
+; GFX11-DL-NEXT:    global_store_b32 v2, v0, s[6:7]
+; GFX11-DL-NEXT:    s_nop 0
+; GFX11-DL-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
+; GFX11-DL-NEXT:    s_endpgm
+                                       ptr addrspace(1) %src2,
+                                       ptr addrspace(1) %src3,
+                                       ptr addrspace(1) nocapture %dst) {
+entry:
+  %idx = call i32 @llvm.amdgcn.workitem.id.x()
+  %gep1 = getelementptr <4 x i8>, ptr addrspace(1) %src1, i32 %idx
+  %vec1 = load <4 x i8>, ptr addrspace(1) %gep1
+  %gep2 = getelementptr <4 x i8>, ptr addrspace(1) %src2, i32 %idx
+  %vec2 = load <4 x i8>, ptr addrspace(1) %gep2
+  %gep3 = getelementptr <4 x i8>, ptr addrspace(1) %src3, i32 %idx
+  %vec3 = load <4 x i8>, ptr addrspace(1) %gep3
+
+  %v2e0 = extractelement <4 x i8> %vec2, i64 1
+  %cv2e0 = zext i8 %v2e0 to i32
+  %mul1 = mul nuw nsw i32 %cv2e0, %cv2e0
+
+  %v1e1 = extractelement <4 x i8> %vec1, i64 1
+  %cv1e1 = zext i8 %v1e1 to i32
+  %v2e1 = extractelement <4 x i8> %vec2, i64 1
+  %cv2e1 = zext i8 %v2e1 to i32
+  %mul2 = mul nuw nsw i32 %cv1e1, %cv2e1
+
+  %v3e2 = extractelement <4 x i8> %vec3, i64 2
+  %cv3e2 = zext i8 %v3e2 to i32
+  %v2e2 = extractelement <4 x i8> %vec2, i64 2
+  %cv2e2 = zext i8 %v2e2 to i32
+  %mul3 = mul nuw nsw i32 %cv2e2, %cv3e2
+
+
+  %acc = load i32, ptr addrspace(1) %dst, align 4
+  %mad1 = add i32 %mul1, %acc
+  %mad2 = add i32 %mad1, %mul2
+  %mad3 = add i32 %mad2, %mul3
+
+  store i32 %mad3, ptr addrspace(1) %dst, align 4
+  ret void
+}
+
+define amdgpu_kernel void @udot4_4src(ptr addrspace(1) %src1,
+; GFX7-LABEL: udot4_4src:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx8 s[4:11], s[0:1], 0x9
+; GFX7-NEXT:    s_mov_b32 s3, 0xf000
+; GFX7-NEXT:    s_mov_b32 s14, 0
+; GFX7-NEXT:    s_mov_b32 s15, s3
+; GFX7-NEXT:    v_lshlrev_b32_e32 v0, 2, v0
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    s_mov_b64 s[12:13], s[4:5]
+; GFX7-NEXT:    v_mov_b32_e32 v1, 0
+; GFX7-NEXT:    buffer_load_dword v2, v[0:1], s[12:15], 0 addr64
+; GFX7-NEXT:    s_mov_b64 s[12:13], s[6:7]
+; GFX7-NEXT:    buffer_load_dword v3, v[0:1], s[12:15], 0 addr64
+; GFX7-NEXT:    s_mov_b64 s[12:13], s[8:9]
+; GFX7-NEXT:    buffer_load_dword v4, v[0:1], s[12:15], 0 addr64
+; GFX7-NEXT:    s_mov_b64 s[12:13], s[10:11]
+; GFX7-NEXT:    buffer_load_dword v0, v[0:1], s[12:15], 0 addr64
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0x11
+; GFX7-NEXT:    s_mov_b32 s2, -1
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    s_load_dword s4, s[0:1], 0x0
+; GFX7-NEXT:    s_waitcnt vmcnt(3)
+; GFX7-NEXT:    v_and_b32_e32 v1, 0xff, v2
+; GFX7-NEXT:    v_bfe_u32 v2, v2, 8, 8
+; GFX7-NEXT:    s_waitcnt vmcnt(2)
+; GFX7-NEXT:    v_and_b32_e32 v5, 0xff, v3
+; GFX7-NEXT:    v_bfe_u32 v3, v3, 8, 8
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mad_u32_u24 v1, v1, v2, s4
+; GFX7-NEXT:    s_waitcnt vmcnt(1)
+; GFX7-NEXT:    v_and_b32_e32 v2, 0xff, v4
+; GFX7-NEXT:    v_bfe_u32 v4, v4, 8, 8
+; GFX7-NEXT:    v_mad_u32_u24 v1, v5, v3, v1
+; GFX7-NEXT:    s_waitcnt vmcnt(0)
+; GFX7-NEXT:    v_and_b32_e32 v3, 0xff, v0
+; GFX7-NEXT:    v_bfe_u32 v0, v0, 8, 8
+; GFX7-NEXT:    v_mad_u32_u24 v1, v2, v4, v1
+; GFX7-NEXT:    v_mad_u32_u24 v0, v3, v0, v1
+; GFX7-NEXT:    buffer_store_dword v0, off, s[0:3], 0
+; GFX7-NEXT:    s_endpgm
+;
+; GFX8-LABEL: udot4_4src:
+; GFX8:       ; %bb.0: ; %entry
+; GFX8-NEXT:    s_load_dwordx8 s[4:11], s[0:1], 0x24
+; GFX8-NEXT:    v_lshlrev_b32_e32 v2, 2, v0
+; GFX8-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0x44
+; GFX8-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX8-NEXT:    v_mov_b32_e32 v1, s5
+; GFX8-NEXT:    v_add_u32_e32 v0, vcc, s4, v2
+; GFX8-NEXT:    v_addc_u32_e32 v1, vcc, 0, v1, vcc
+; GFX8-NEXT:    flat_load_dword v3, v[0:1]
+; GFX8-NEXT:    v_mov_b32_e32 v1, s7
+; GFX8-NEXT:    v_add_u32_e32 v0, vcc, s6, v2
+; GFX8-NEXT:    v_addc_u32_e32 v1, vcc, 0, v1, vcc
+; GFX8-NEXT:    flat_load_dword v4, v[0:1]
+; GFX8-NEXT:    v_mov_b32_e32 v1, s9
+; GFX8-NEXT:    v_add_u32_e32 v0, vcc, s8, v2
+; GFX8-NEXT:    v_addc_u32_e32 v1, vcc, 0, v1, vcc
+; GFX8-NEXT:    flat_load_dword v5, v[0:1]
+; GFX8-NEXT:    v_mov_b32_e32 v1, s11
+; GFX8-NEXT:    v_add_u32_e32 v0, vcc, s10, v2
+; GFX8-NEXT:    v_addc_u32_e32 v1, vcc, 0, v1, vcc
+; GFX8-NEXT:    flat_load_dword v0, v[0:1]
+; GFX8-NEXT:    s_load_dword s2, s[0:1], 0x0
+; GFX8-NEXT:    s_waitcnt vmcnt(3)
+; GFX8-NEXT:    v_and_b32_e32 v1, 0xff, v3
+; GFX8-NEXT:    v_bfe_u32 v2, v3, 8, 8
+; GFX8-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX8-NEXT:    v_mad_u32_u24 v1, v1, v2, s2
+; GFX8-NEXT:    s_waitcnt vmcnt(2)
+; GFX8-NEXT:    v_and_b32_e32 v3, 0xff, v4
+; GFX8-NEXT:    v_bfe_u32 v4, v4, 8, 8
+; GFX8-NEXT:    v_mad_u32_u24 v1, v3, v4, v1
+; GFX8-NEXT:    s_waitcnt vmcnt(1)
+; GFX8-NEXT:    v_and_b32_e32 v6, 0xff, v5
+; GFX8-NEXT:    v_bfe_u32 v5, v5, 8, 8
+; GFX8-NEXT:    v_mad_u32_u24 v1, v6, v5, v1
+; GFX8-NEXT:    s_waitcnt vmcnt(0)
+; GFX8-NEXT:    v_and_b32_e32 v7, 0xff, v0
+; GFX8-NEXT:    v_bfe_u32 v0, v0, 8, 8
+; GFX8-NEXT:    v_mad_u32_u24 v2, v7, v0, v1
+; GFX8-NEXT:    v_mov_b32_e32 v0, s0
+; GFX8-NEXT:    v_mov_b32_e32 v1, s1
+; GFX8-NEXT:    flat_store_dword v[0:1], v2
+; GFX8-NEXT:    s_endpgm
+;
+; GFX9-NODL-LABEL: udot4_4src:
+; GFX9-NODL:       ; %bb.0: ; %entry
+; GFX9-NODL-NEXT:    s_load_dwordx8 s[4:11], s[0:1], 0x24
+; GFX9-NODL-NEXT:    v_lshlrev_b32_e32 v0, 2, v0
+; GFX9-NODL-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0x44
+; GFX9-NODL-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX9-NODL-NEXT:    global_load_dword v1, v0, s[4:5]
+; GFX9-NODL-NEXT:    global_load_dword v2, v0, s[6:7]
+; GFX9-NODL-NEXT:    global_load_dword v3, v0, s[8:9]
+; GFX9-NODL-NEXT:    global_load_dword v4, v0, s[10:11]
+; GFX9-NODL-NEXT:    v_mov_b32_e32 v0, 0
+; GFX9-NODL-NEXT:    s_load_dword s2, s[0:1], 0x0
+; GFX9-NODL-NEXT:    s_waitcnt vmcnt(3)
+; GFX9-NODL-NEXT:    v_mul_u32_u24_sdwa v1, v1, v1 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:BYTE_1
+; GFX9-NODL-NEXT:    s_waitcnt vmcnt(2)
+; GFX9-NODL-NEXT:    v_mul_u32_u24_sdwa v2, v2, v2 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:BYTE_1
+; GFX9-NODL-NEXT:    s_waitcnt vmcnt(1)
+; GFX9-NODL-NEXT:    v_mul_u32_u24_sdwa v3, v3, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:BYTE_1
+; GFX9-NODL-NEXT:    s_waitcnt vmcnt(0)
+; GFX9-NODL-NEXT:    v_mul_u32_u24_sdwa v4, v4, v4 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:BYTE_1
+; GFX9-NODL-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX9-NODL-NEXT:    v_add3_u32 v1, v1, s2, v2
+; GFX9-NODL-NEXT:    v_add3_u32 v1, v1, v3, v4
+; GFX9-NODL-NEXT:    global_store_dword v0, v1, s[0:1]
+; GFX9-NODL-NEXT:    s_endpgm
+;
+; GFX9-DL-LABEL: udot4_4src:
+; GFX9-DL:       ; %bb.0: ; %entry
+; GFX9-DL-NEXT:    s_load_dwordx8 s[4:11], s[0:1], 0x24
+; GFX9-DL-NEXT:    v_lshlrev_b32_e32 v0, 2, v0
+; GFX9-DL-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0x44
+; GFX9-DL-NEXT:    s_mov_b32 s2, 0xc0c0501
+; GFX9-DL-NEXT:    s_mov_b32 s3, 0x5010c0c
+; GFX9-DL-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX9-DL-NEXT:    global_load_dword v1, v0, s[4:5]
+; GFX9-DL-NEXT:    global_load_dword v2, v0, s[6:7]
+; GFX9-DL-NEXT:    global_load_dword v3, v0, s[8:9]
+; GFX9-DL-NEXT:    global_load_dword v4, v0, s[10:11]
+; GFX9-DL-NEXT:    s_mov_b32 s4, 0xc0c0400
+; GFX9-DL-NEXT:    s_load_dword s6, s[0:1], 0x0
+; GFX9-DL-NEXT:    s_mov_b32 s5, 0x4000c0c
+; GFX9-DL-NEXT:    v_mov_b32_e32 v0, 0
+; GFX9-DL-NEXT:    s_waitcnt vmcnt(2)
+; GFX9-DL-NEXT:    v_perm_b32 v5, v2, v1, s2
+; GFX9-DL-NEXT:    v_perm_b32 v1, v2, v1, s4
+; GFX9-DL-NEXT:    s_waitcnt vmcnt(0)
+; GFX9-DL-NEXT:    v_perm_b32 v6, v4, v3, s3
+; GFX9-DL-NEXT:    v_perm_b32 v2, v4, v3, s5
+; GFX9-DL-NEXT:    v_or_b32_e32 v3, v6, v5
+; GFX9-DL-NEXT:    v_or_b32_e32 v1, v2, v1
+; GFX9-DL-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX9-DL-NEXT:    v_dot4_u32_u8 v1, v1, v3, s6
+; GFX9-DL-NEXT:    global_store_dword v0, v1, s[0:1]
+; GFX9-DL-NEXT:    s_endpgm
+;
+; GFX10-DL-LABEL: udot4_4src:
+; GFX10-DL:       ; %bb.0: ; %entry
+; GFX10-DL-NEXT:    s_load_dwordx8 s[4:11], s[0:1], 0x24
+; GFX10-DL-NEXT:    v_lshlrev_b32_e32 v0, 2, v0
+; GFX10-DL-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0x44
+; GFX10-DL-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-DL-NEXT:    s_clause 0x3
+; GFX10-DL-NEXT:    global_load_dword v1, v0, s[4:5]
+; GFX10-DL-NEXT:    global_load_dword v2, v0, s[6:7]
+; GFX10-DL-NEXT:    global_load_dword v3, v0, s[8:9]
+; GFX10-DL-NEXT:    global_load_dword v4, v0, s[10:11]
+; GFX10-DL-NEXT:    s_load_dword s2, s[0:1], 0x0
+; GFX10-DL-NEXT:    s_waitcnt vmcnt(2)
+; GFX10-DL-NEXT:    v_perm_b32 v0, v2, v1, 0xc0c0501
+; GFX10-DL-NEXT:    v_perm_b32 v1, v2, v1, 0xc0c0400
+; GFX10-DL-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-DL-NEXT:    v_perm_b32 v5, v4, v3, 0x5010c0c
+; GFX10-DL-NEXT:    v_perm_b32 v2, v4, v3, 0x4000c0c
+; GFX10-DL-NEXT:    v_or_b32_e32 v0, v5, v0
+; GFX10-DL-NEXT:    v_or_b32_e32 v1, v2, v1
+; GFX10-DL-NEXT:    v_mov_b32_e32 v2, 0
+; GFX10-DL-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-DL-NEXT:    v_dot4_u32_u8 v0, v1, v0, s2
+; GFX10-DL-NEXT:    global_store_dword v2, v0, s[0:1]
+; GFX10-DL-NEXT:    s_endpgm
+;
+; GFX11-DL-LABEL: udot4_4src:
+; GFX11-DL:       ; %bb.0: ; %entry
+; GFX11-DL-NEXT:    s_load_b256 s[4:11], s[0:1], 0x24
+; GFX11-DL-NEXT:    v_lshlrev_b32_e32 v0, 2, v0
+; GFX11-DL-NEXT:    s_load_b64 s[0:1], s[0:1], 0x44
+; GFX11-DL-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX11-DL-NEXT:    s_clause 0x3
+; GFX11-DL-NEXT:    global_load_b32 v1, v0, s[4:5]
+; GFX11-DL-NEXT:    global_load_b32 v2, v0, s[6:7]
+; GFX11-DL-NEXT:    global_load_b32 v3, v0, s[8:9]
+; GFX11-DL-NEXT:    global_load_b32 v0, v0, s[10:11]
+; GFX11-DL-NEXT:    s_load_b32 s2, s[0:1], 0x0
+; GFX11-DL-NEXT:    s_waitcnt vmcnt(2)
+; GFX11-DL-NEXT:    v_perm_b32 v4, v2, v1, 0xc0c0501
+; GFX11-DL-NEXT:    v_perm_b32 v1, v2, v1, 0xc0c0400
+; GFX11-DL-NEXT:    s_waitcnt vmcnt(0)
+; GFX11-DL-NEXT:    v_perm_b32 v5, v0, v3, 0x5010c0c
+; GFX11-DL-NEXT:    v_perm_b32 v0, v0, v3, 0x4000c0c
+; GFX11-DL-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX11-DL-NEXT:    v_or_b32_e32 v2, v5, v4
+; GFX11-DL-NEXT:    v_or_b32_e32 v0, v0, v1
+; GFX11-DL-NEXT:    v_mov_b32_e32 v1, 0
+; GFX11-DL-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX11-DL-NEXT:    s_delay_alu instid0(VALU_DEP_2)
+; GFX11-DL-NEXT:    v_dot4_u32_u8 v0, v0, v2, s2
+; GFX11-DL-NEXT:    global_store_b32 v1, v0, s[0:1]
+; GFX11-DL-NEXT:    s_nop 0
+; GFX11-DL-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
+; GFX11-DL-NEXT:    s_endpgm
+                                       ptr addrspace(1) %src2,
+                                       ptr addrspace(1) %src3,
+                                       ptr addrspace(1) %src4,
+                                       ptr addrspace(1) nocapture %dst) {
+entry:
+  %idx = call i32 @llvm.amdgcn.workitem.id.x()
+
+  %gep1 = getelementptr <4 x i8>, ptr addrspace(1) %src1, i32 %idx
+  %vec1 = load <4 x i8>, ptr addrspace(1) %gep1
+  %gep2 = getelementptr <4 x i8>, ptr addrspace(1) %src2, i32 %idx
+  %vec2 = load <4 x i8>, ptr addrspace(1) %gep2
+  %gep3 = getelementptr <4 x i8>, ptr addrspace(1) %src3, i32 %idx
+  %vec3 = load <4 x i8>, ptr addrspace(1) %gep3
+  %gep4 = getelementptr <4 x i8>, ptr addrspace(1) %src4, i32 %idx
+  %vec4 = load <4 x i8>, ptr addrspace(1) %gep4
+
+
+  %v1e0 = extractelement <4 x i8> %vec1, i64 0
+  %cv1e0 = zext i8 %v1e0 to i32
+  %v1e1 = extractelement <4 x i8> %vec1, i64 1
+  %cv1e1 = zext i8 %v1e1 to i32
+  %mul1 = mul nuw nsw i32 %cv1e0, %cv1e1
+
+  %v2e0 = extractelement <4 x i8> %vec2, i64 0
+  %cv2e0 = zext i8 %v2e0 to i32
+  %v2e1 = extractelement <4 x i8> %vec2, i64 1
+  %cv2e1 = zext i8 %v2e1 to i32
+  %mul2 = mul nuw nsw i32 %cv2e0, %cv2e1
+
+  %v3e0 = extractelement <4 x i8> %vec3, i64 0
+  %cv3e0 = zext i8 %v3e0 to i32
+  %v3e1 = extractelement <4 x i8> %vec3, i64 1
+  %cv3e1 = zext i8 %v3e1 to i32
+  %mul3 = mul nuw nsw i32 %cv3e0, %cv3e1
+
+  %v4e0 = extractelement <4 x i8> %vec4, i64 0
+  %cv4e0 = zext i8 %v4e0 to i32
+  %v4e1 = extractelement <4 x i8> %vec4, i64 1
+  %cv4e1 = zext i8 %v4e1 to i32
+  %mul4 = mul nuw nsw i32 %cv4e0, %cv4e1
+
+
+  %acc = load i32, ptr addrspace(1) %dst, align 4
+  %mad1 = add i32 %mul1, %acc
+  %mad2 = add i32 %mad1, %mul2
+  %mad3 = add i32 %mad2, %mul3
+  %mad4 = add i32 %mad3, %mul4
+
+  store i32 %mad4, ptr addrspace(1) %dst, align 4
+  ret void
+}
+
+define amdgpu_kernel void @udot4_acc32_multi(ptr addrspace(1) %src1,
+; GFX7-LABEL: udot4_acc32_multi:
+; GFX7:       ; %bb.0: ; %entry
+; GFX7-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x9
+; GFX7-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0xd
+; GFX7-NEXT:    s_mov_b32 s3, 0xf000
+; GFX7-NEXT:    s_mov_b32 s10, 0
+; GFX7-NEXT:    s_mov_b32 s11, s3
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    s_mov_b64 s[8:9], s[4:5]
+; GFX7-NEXT:    v_lshlrev_b32_e32 v0, 3, v0
+; GFX7-NEXT:    v_mov_b32_e32 v1, 0
+; GFX7-NEXT:    s_mov_b64 s[4:5], s[6:7]
+; GFX7-NEXT:    s_mov_b64 s[6:7], s[10:11]
+; GFX7-NEXT:    buffer_load_dwordx2 v[2:3], v[0:1], s[8:11], 0 addr64
+; GFX7-NEXT:    buffer_load_dword v0, v[0:1], s[4:7], 0 addr64
+; GFX7-NEXT:    s_load_dword s4, s[0:1], 0x0
+; GFX7-NEXT:    s_mov_b32 s2, -1
+; GFX7-NEXT:    s_waitcnt vmcnt(1)
+; GFX7-NEXT:    v_and_b32_e32 v1, 0xff, v2
+; GFX7-NEXT:    s_waitcnt vmcnt(0)
+; GFX7-NEXT:    v_and_b32_e32 v4, 0xff, v0
+; GFX7-NEXT:    v_bfe_u32 v7, v2, 16, 8
+; GFX7-NEXT:    v_bfe_u32 v8, v0, 16, 8
+; GFX7-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX7-NEXT:    v_mad_u32_u24 v1, v1, v4, s4
+; GFX7-NEXT:    v_and_b32_e32 v9, 0xff, v3
+; GFX7-NEXT:    v_mad_u32_u24 v1, v7, v8, v1
+; GFX7-NEXT:    v_bfe_u32 v11, v3, 16, 8
+; GFX7-NEXT:    v_mad_u32_u24 v1, v9, v4, v1
+; GFX7-NEXT:    v_bfe_u32 v5, v2, 8, 8
+; GFX7-NEXT:    v_bfe_u32 v6, v0, 8, 8
+; GFX7-NEXT:    v_mad_u32_u24 v1, v11, v8, v1
+; GFX7-NEXT:    v_lshrrev_b32_e32 v2, 24, v2
+; GFX7-NEXT:    v_lshrrev_b32_e32 v0, 24, v0
+; GFX7-NEXT:    v_mad_u32_u24 v1, v5, v6, v1
+; GFX7-NEXT:    v_bfe_u32 v10, v3, 8, 8
+; GFX7-NEXT:    v_mad_u32_u24 v1, v2, v0, v1
+; GFX7-NEXT:    v_lshrrev_b32_e32 v3, 24, v3
+; GFX7-NEXT:    v_mad_u32_u24 v1, v10, v6, v1
+; GFX7-NEXT:    v_mad_u32_u24 v0, v3, v0, v1
+; GFX7-NEXT:    buffer_store_dword v0, off, s[0:3], 0
+; GFX7-NEXT:    s_endpgm
+;
+; GFX8-LABEL: udot4_acc32_multi:
+; GFX8:       ; %bb.0: ; %entry
+; GFX8-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x24
+; GFX8-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0x34
+; GFX8-NEXT:    v_lshlrev_b32_e32 v2, 3, v0
+; GFX8-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX8-NEXT:    v_mov_b32_e32 v1, s5
+; GFX8-NEXT:    v_add_u32_e32 v0, vcc, s4, v2
+; GFX8-NEXT:    v_addc_u32_e32 v1, vcc, 0, v1, vcc
+; GFX8-NEXT:    v_mov_b32_e32 v3, s7
+; GFX8-NEXT:    v_add_u32_e32 v2, vcc, s6, v2
+; GFX8-NEXT:    v_addc_u32_e32 v3, vcc, 0, v3, vcc
+; GFX8-NEXT:    flat_load_dwordx2 v[0:1], v[0:1]
+; GFX8-NEXT:    flat_load_dword v2, v[2:3]
+; GFX8-NEXT:    s_load_dword s2, s[0:1], 0x0
+; GFX8-NEXT:    s_waitcnt vmcnt(1)
+; GFX8-NEXT:    v_and_b32_e32 v3, 0xff, v0
+; GFX8-NEXT:    s_waitcnt vmcnt(0)
+; GFX8-NEXT:    v_and_b32_e32 v4, 0xff, v2
+; GFX8-NEXT:    v_bfe_u32 v7, v0, 16, 8
+; GFX8-NEXT:    v_bfe_u32 v8, v2, 16, 8
+; GFX8-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX8-NEXT:    v_mad_u32_u24 v3, v3, v4, s2
+; GFX8-NEXT:    v_and_b32_e32 v9, 0xff, v1
+; GFX8-NEXT:    v_mad_u32_u24 v3, v7, v8, v3
+; GFX8-NEXT:    v_bfe_u32 v11, v1, 16, 8
+; GFX8-NEXT:    v_mad_u32_u24 v3, v9, v4, v3
+; GFX8-NEXT:    v_bfe_u32 v5, v0, 8, 8
+; GFX8-NEXT:    v_bfe_u32 v6, v2, 8, 8
+; GFX8-NEXT:    v_mad_u32_u24 v3, v11, v8, v3
+; GFX8-NEXT:    v_lshrrev_b32_e32 v0, 24, v0
+; GFX8-NEXT:    v_lshrrev_b32_e32 v2, 24, v2
+; GFX8-NEXT:    v_mad_u32_u24 v3, v5, v6, v3
+; GFX8-NEXT:    v_bfe_u32 v10, v1, 8, 8
+; GFX8-NEXT:    v_mad_u32_u24 v0, v0, v2, v3
+; GFX8-NEXT:    v_lshrrev_b32_e32 v1, 24, v1
+; GFX8-NEXT:    v_mad_u32_u24 v0, v10, v6, v0
+; GFX8-NEXT:    v_mad_u32_u24 v2, v1, v2, v0
+; GFX8-NEXT:    v_mov_b32_e32 v0, s0
+; GFX8-NEXT:    v_mov_b32_e32 v1, s1
+; GFX8-NEXT:    flat_store_dword v[0:1], v2
+; GFX8-NEXT:    s_endpgm
+;
+; GFX9-NODL-LABEL: udot4_acc32_multi:
+; GFX9-NODL:       ; %bb.0: ; %entry
+; GFX9-NODL-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x24
+; GFX9-NODL-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x34
+; GFX9-NODL-NEXT:    v_lshlrev_b32_e32 v2, 3, v0
+; GFX9-NODL-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX9-NODL-NEXT:    global_load_dword v3, v2, s[6:7]
+; GFX9-NODL-NEXT:    global_load_dwordx2 v[0:1], v2, s[4:5]
+; GFX9-NODL-NEXT:    s_load_dword s0, s[2:3], 0x0
+; GFX9-NODL-NEXT:    v_mov_b32_e32 v2, 0
+; GFX9-NODL-NEXT:    s_waitcnt vmcnt(1)
+; GFX9-NODL-NEXT:    v_and_b32_e32 v4, 0xff, v3
+; GFX9-NODL-NEXT:    v_bfe_u32 v6, v3, 16, 8
+; GFX9-NODL-NEXT:    v_bfe_u32 v5, v3, 8, 8
+; GFX9-NODL-NEXT:    v_lshrrev_b32_e32 v3, 24, v3
+; GFX9-NODL-NEXT:    s_waitcnt vmcnt(0)
+; GFX9-NODL-NEXT:    v_mul_u32_u24_sdwa v7, v0, v4 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; GFX9-NODL-NEXT:    v_mul_u32_u24_sdwa v9, v0, v6 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_2 src1_sel:DWORD
+; GFX9-NODL-NEXT:    v_mul_u32_u24_sdwa v8, v0, v5 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_1 src1_sel:DWORD
+; GFX9-NODL-NEXT:    v_mul_u32_u24_sdwa v0, v0, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_3 src1_sel:DWORD
+; GFX9-NODL-NEXT:    v_mul_u32_u24_sdwa v4, v1, v4 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0 src1_sel:DWORD
+; GFX9-NODL-NEXT:    v_mul_u32_u24_sdwa v5, v1, v5 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_1 src1_sel:DWORD
+; GFX9-NODL-NEXT:    v_mul_u32_u24_sdwa v6, v1, v6 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_2 src1_sel:DWORD
+; GFX9-NODL-NEXT:    v_mul_u32_u24_sdwa v1, v1, v3 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_3 src1_sel:DWORD
+; GFX9-NODL-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX9-NODL-NEXT:    v_add3_u32 v3, v7, s0, v9
+; GFX9-NODL-NEXT:    v_add3_u32 v3, v3, v4, v6
+; GFX9-NODL-NEXT:    v_add3_u32 v0, v8, v3, v0
+; GFX9-NODL-NEXT:    v_add3_u32 v0, v0, v5, v1
+; GFX9-NODL-NEXT:    global_store_dword v2, v0, s[2:3]
+; GFX9-NODL-NEXT:    s_endpgm
+;
+; GFX9-DL-LABEL: udot4_acc32_multi:
+; GFX9-DL:       ; %bb.0: ; %entry
+; GFX9-DL-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x24
+; GFX9-DL-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x34
+; GFX9-DL-NEXT:    v_lshlrev_b32_e32 v2, 3, v0
+; GFX9-DL-NEXT:    s_mov_b32 s0, 0x6040200
+; GFX9-DL-NEXT:    s_mov_b32 s1, 0x2000200
+; GFX9-DL-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX9-DL-NEXT:    global_load_dwordx2 v[0:1], v2, s[4:5]
+; GFX9-DL-NEXT:    global_load_dword v3, v2, s[6:7]
+; GFX9-DL-NEXT:    s_load_dword s5, s[2:3], 0x0
+; GFX9-DL-NEXT:    s_mov_b32 s4, 0x7050301
+; GFX9-DL-NEXT:    s_mov_b32 s6, 0x3010301
+; GFX9-DL-NEXT:    v_mov_b32_e32 v2, 0
+; GFX9-DL-NEXT:    s_waitcnt vmcnt(1)
+; GFX9-DL-NEXT:    v_perm_b32 v4, v1, v0, s0
+; GFX9-DL-NEXT:    s_waitcnt vmcnt(0)
+; GFX9-DL-NEXT:    v_perm_b32 v5, v3, v3, s1
+; GFX9-DL-NEXT:    v_perm_b32 v0, v1, v0, s4
+; GFX9-DL-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX9-DL-NEXT:    v_dot4_u32_u8 v1, v4, v5, s5
+; GFX9-DL-NEXT:    v_perm_b32 v3, v3, v3, s6
+; GFX9-DL-NEXT:    v_dot4_u32_u8 v0, v0, v3, v1
+; GFX9-DL-NEXT:    global_store_dword v2, v0, s[2:3]
+; GFX9-DL-NEXT:    s_endpgm
+;
+; GFX10-DL-LABEL: udot4_acc32_multi:
+; GFX10-DL:       ; %bb.0: ; %entry
+; GFX10-DL-NEXT:    s_load_dwordx4 s[4:7], s[0:1], 0x24
+; GFX10-DL-NEXT:    v_lshlrev_b32_e32 v2, 3, v0
+; GFX10-DL-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0x34
+; GFX10-DL-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-DL-NEXT:    s_clause 0x1
+; GFX10-DL-NEXT:    global_load_dwordx2 v[0:1], v2, s[4:5]
+; GFX10-DL-NEXT:    global_load_dword v3, v2, s[6:7]
+; GFX10-DL-NEXT:    s_load_dword s2, s[0:1], 0x0
+; GFX10-DL-NEXT:    s_waitcnt vmcnt(1)
+; GFX10-DL-NEXT:    v_perm_b32 v2, v1, v0, 0x6040200
+; GFX10-DL-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-DL-NEXT:    v_perm_b32 v4, v3, v3, 0x2000200
+; GFX10-DL-NEXT:    v_perm_b32 v0, v1, v0, 0x7050301
+; GFX10-DL-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX10-DL-NEXT:    v_dot4_u32_u8 v1, v2, v4, s2
+; GFX10-DL-NEXT:    v_perm_b32 v2, v3, v3, 0x3010301
+; GFX10-DL-NEXT:    v_mov_b32_e32 v3, 0
+; GFX10-DL-NEXT:    v_dot4_u32_u8 v0, v0, v2, v1
+; GFX10-DL-NEXT:    global_store_dword v3, v0, s[0:1]
+; GFX10-DL-NEXT:    s_endpgm
+;
+; GFX11-DL-LABEL: udot4_acc32_multi:
+; GFX11-DL:       ; %bb.0: ; %entry
+; GFX11-DL-NEXT:    s_load_b128 s[4:7], s[0:1], 0x24
+; GFX11-DL-NEXT:    v_lshlrev_b32_e32 v2, 3, v0
+; GFX11-DL-NEXT:    s_load_b64 s[0:1], s[0:1], 0x34
+; GFX11-DL-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX11-DL-NEXT:    s_clause 0x1
+; GFX11-DL-NEXT:    global_load_b64 v[0:1], v2, s[4:5]
+; GFX11-DL-NEXT:    global_load_b32 v2, v2, s[6:7]
+; GFX11-DL-NEXT:    s_load_b32 s2, s[0:1], 0x0
+; GFX11-DL-NEXT:    s_waitcnt vmcnt(1)
+; GFX11-DL-NEXT:    v_perm_b32 v3, v1, v0, 0x6040200
+; GFX11-DL-NEXT:    s_waitcnt vmcnt(0)
+; GFX11-DL-NEXT:    v_perm_b32 v4, v2, v2, 0x2000200
+; GFX11-DL-NEXT:    v_perm_b32 v0, v1, v0, 0x7050301
+; GFX11-DL-NEXT:    v_perm_b32 v2, v2, v2, 0x3010301
+; GFX11-DL-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX11-DL-NEXT:    s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_2)
+; GFX11-DL-NEXT:    v_dot4_u32_u8 v1, v3, v4, s2
+; GFX11-DL-NEXT:    v_mov_b32_e32 v3, 0
+; GFX11-DL-NEXT:    v_dot4_u32_u8 v0, v0, v2, v1
+; GFX11-DL-NEXT:    global_store_b32 v3, v0, s[0:1]
+; GFX11-DL-NEXT:    s_nop 0
+; GFX11-DL-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
+; GFX11-DL-NEXT:    s_endpgm
+                                       ptr addrspace(1) %src2,
+                                       ptr addrspace(1) nocapture %dst) {
+entry:
+  %idx = call i32 @llvm.amdgcn.workitem.id.x()
+  %gep1 = getelementptr <8 x i8>, ptr addrspace(1) %src1, i32 %idx
+  %vec1 = load <8 x i8>, ptr addrspace(1) %gep1
+  %gep2 = getelementptr <8 x i8>, ptr addrspace(1) %src2, i32 %idx
+  %vec2 = load <8 x i8>, ptr addrspace(1) %gep2
+
+  %v1e0 = extractelement <8 x i8> %vec1, i64 0
+  %cv1e0 = zext i8 %v1e0 to i32
+  %v2e0 = extractelement <8 x i8> %vec2, i64 0
+  %cv2e0 = zext i8 %v2e0 to i32
+  %mul1 = mul nuw nsw i32 %cv1e0, %cv2e0
+
+  %v1e1 = extractelement <8 x i8> %vec1, i64 1
+  %cv1e1 = zext i8 %v1e1 to i32
+  %v2e1 = extractelement <8 x i8> %vec2, i64 1
+  %cv2e1 = zext i8 %v2e1 to i32
+  %mul2 = mul nuw nsw i32 %cv1e1, %cv2e1
+
+  %v1e2 = extractelement <8 x i8> %vec1, i64 2
+  %cv1e2 = zext i8 %v1e2 to i32
+  %v2e2 = extractelement <8 x i8> %vec2, i64 2
+  %cv2e2 = zext i8 %v2e2 to i32
+  %mul3 = mul nuw nsw i32 %cv1e2, %cv2e2
+
+  %v1e3 = extractelement <8 x i8> %vec1, i64 3
+  %cv1e3 = zext i8 %v1e3 to i32
+  %v2e3 = extractelement <8 x i8> %vec2, i64 3
+  %cv2e3 = zext i8 %v2e3 to i32
+  %mul4 = mul nuw nsw i32 %cv1e3, %cv2e3
+
+  %v1e4 = extractelement <8 x i8> %vec1, i64 4
+  %cv1e4 = zext i8 %v1e4 to i32
+  %v2e4 = extractelement <8 x i8> %vec2, i64 4
+  %cv2e4 = zext i8 %v2e4 to i32
+  %mul5 = mul nuw nsw i32 %cv1e4, %cv2e0
+
+  %v1e5 = extractelement <8 x i8> %vec1, i64 5
+  %cv1e5 = zext i8 %v1e5 to i32
+  %v2e5 = extractelement <8 x i8> %vec2, i64 5
+  %cv2e5 = zext i8 %v2e5 to i32
+  %mul6 = mul nuw nsw i32 %cv1e5, %cv2e1
+
+  %v1e6 = extractelement <8 x i8> %vec1, i64 6
+  %cv1e6 = zext i8 %v1e6 to i32
+  %v2e6 = extractelement <8 x i8> %vec2, i64 6
+  %cv2e6 = zext i8 %v2e6 to i32
+  %mul7 = mul nuw nsw i32 %cv1e6, %cv2e2
+
+  %v1e7 = extractelement <8 x i8> %vec1, i64 7
+  %cv1e7 = zext i8 %v1e7 to i32
+  %v2e7 = extractelement <8 x i8> %vec2, i64 7
+  %cv2e7 = zext i8 %v2e7 to i32
+  %mul8 = mul nuw nsw i32 %cv1e7, %cv2e3
+
+  %acc = load i32, ptr addrspace(1) %dst, align 4
+  %mad11 = add i32 %mul1, %acc
+  %mad21 = add i32 %mad11, %mul3
+  %mad31 = add i32 %mad21, %mul5
+  %mad41 = add i32 %mad31, %mul7
+  %mad12 = add i32 %mul2, %mad41
+  %mad22 = add i32 %mad12, %mul4
+  %mad32 = add i32 %mad22, %mul6
+  %mad42 = add i32 %mad32, %mul8
+
+  store i32 %mad42, ptr addrspace(1) %dst, align 4
+  ret void
+}
+
 declare i32 @llvm.amdgcn.workitem.id.x()