[X86] [AMX] Replace bitcast with specific AMX intrinsics with X86 specific cast.

[yubing]

Summary

There is some discussion on the bitcast for vector and x86_amx at https://reviews.llvm.org/D99152. This patch is to introduce a x86 specific cast for vector and x86_amx, so that it can avoid some unnecessary optimization by middle-end. On the other way, we have to optimize the x86 specific cast by ourselves. This patch also optimize the cast operation to eliminate redundant code.

[LuoYuanke]

Would you summarize the the title in fewer words and elaborate it in description and commit message?

[fhahn]

Could you also document those intrinsics in the LangRef?

llvm/lib/Target/X86/X86LowerAMXType.cpp

71

can you just return match() || match()?

Also it looks like you are not using Vec, so you could use m_Any() or something like that instead.

[LuoYuanke]

In D107544#2943280, @fhahn wrote:

Could you also document those intrinsics in the LangRef?

This is X86 specific intrinsics. Should we document it in LangRef?

[LuoYuanke]

llvm/lib/Target/X86/X86LowerAMXType.cpp

800

Drop brace.

820

Drop brace.

822

Drop brace.

837

of

841

What does ACI stand for? AMX cast intrinsic?

857

Drop brace.

917

We can erase dead cast code from Vec2TileInsts and Vec2TileInsts and get AMX cast instruction from there, so that we can avoid iterate basic block again.

[yubing]

llvm/lib/Target/X86/X86LowerAMXType.cpp

71

Thanks for comments. You're right, I will change it.
We could use m_Value().

[LuoYuanke]

llvm/test/CodeGen/X86/AMX/lat-combine-amx-bitcast.ll

107

We can optimize udef or zero vector with tilezero. We may do it in another patch.

153

This may be optimized to "phi <x86_amx>", and cast back to vector for its user %evilphi2. But we can optimize it in another patch.

400

Could you simplify the name?

llvm/test/CodeGen/X86/AMX/lat-transform-amx-bitcast.ll

142

We can forward llvm.x86.tilestored64.internal to finial store (line 143). We can optimize it in another patch.

292

We can combine load and cast to @llvm.x86.tileloadd64.internal.

[yubing]

Address flanhn and yuanke's comments

[yubing]

llvm/lib/Target/X86/X86LowerAMXType.cpp

841

Yeah, AMX cast intrinsic

917

We will refactor it in next patch

llvm/test/CodeGen/X86/AMX/lat-combine-amx-bitcast.ll

107

Sure

llvm/test/CodeGen/X86/AMX/lat-transform-amx-bitcast.ll

292

Sure, we will do it in the next patch.

[LuoYuanke]

LGTM, thanks.

[fhahn]

In D107544#2946244, @LuoYuanke wrote:

This is X86 specific intrinsics. Should we document it in LangRef?

looks like target-specific ones are not documented there so far, not sure if there's a different place to document them.

[RKSimon]

This appears to be causing failures on EXPENSIVE_CHECKS builds: https://bugs.llvm.org/show_bug.cgi?id=51513

[RKSimon]

llvm/lib/Target/X86/X86LowerAMXType.cpp

1049

Doesn't C needs to be updated based on the results of combineAMXcast and transformAllAMXCast as well ?

[LuoYuanke]

llvm/lib/Target/X86/X86LowerAMXType.cpp

1049

Yes. C should be updated. Thanks, Simon.

[pengfei]

llvm/lib/Target/X86/X86LowerAMXType.cpp

178

This function has no user now?

[pengfei]

llvm/lib/Target/X86/X86LowerAMXType.cpp

915

PN has no user. Should be removed to avoid warning.

[LuoYuanke]

I created patch https://reviews.llvm.org/D108269 to fix the bug and address Pengfei's comments.

[uabelho]

llvm/lib/Target/X86/X86LowerAMXType.cpp

178

I agree, this function seems to be unused? I get the following warning compiling trunk with gcc:

../lib/Target/X86/X86LowerAMXType.cpp:181:8: warning: 'llvm::Value* {anonymous}::X86LowerAMXType::getRowFromCol(llvm::Instruction*, llvm::Value*, unsigned int)' defined but not used [-Wunused-function]
  181 | Value *X86LowerAMXType::getRowFromCol(Instruction *II, Value *V,
      |        ^~~~~~~~~~~~~~~

llvm/include/llvm/IR/IntrinsicsX86.td

def int_x86_tilezero_internal :
               Intrinsic<[llvm_x86amx_ty], [llvm_i16_ty, llvm_i16_ty],
                         []>;
   def int_x86_tdpbf16ps_internal :
               GCCBuiltin<"__builtin_ia32_tdpbf16ps_internal">,
               Intrinsic<[llvm_x86amx_ty],
                         [llvm_i16_ty, llvm_i16_ty, llvm_i16_ty,
                          llvm_x86amx_ty, llvm_x86amx_ty,
                          llvm_x86amx_ty], []>;
+  def int_x86_cast_vector_to_tile:
+              Intrinsic<[llvm_x86amx_ty], [llvm_anyvector_ty], [IntrNoMem]>;
+  def int_x86_cast_tile_to_vector:
+              Intrinsic<[llvm_anyvector_ty], [llvm_x86amx_ty], [IntrNoMem]>;
 }
 
 //===----------------------------------------------------------------------===//
 // UINTR - User Level Interrupt
 
 let TargetPrefix = "x86" in {
   def int_x86_clui : GCCBuiltin<"__builtin_ia32_clui">,
               Intrinsic<[], [], []>;
   def int_x86_stui : GCCBuiltin<"__builtin_ia32_stui">,
               Intrinsic<[], [], []>;
   def int_x86_testui : GCCBuiltin<"__builtin_ia32_testui">,
               Intrinsic<[llvm_i8_ty], [], []>;
   def int_x86_senduipi : GCCBuiltin<"__builtin_ia32_senduipi">,
               Intrinsic<[], [llvm_i64_ty], []>;
 }

llvm/lib/Target/X86/X86LowerAMXType.cpp

 /// | %td2 = call x86_amx @llvm.x86.tileloadd64.internal(mem)|
 /// | "use %td2"                                             |
 /// ----------------------------------------------------------
 //
 //===----------------------------------------------------------------------===//
 //
 #include "X86.h"
 #include "llvm/ADT/PostOrderIterator.h"
+#include "llvm/ADT/SetVector.h"
 #include "llvm/ADT/SmallSet.h"
 #include "llvm/Analysis/OptimizationRemarkEmitter.h"
+#include "llvm/Analysis/TargetLibraryInfo.h"
 #include "llvm/Analysis/TargetTransformInfo.h"
 #include "llvm/CodeGen/Passes.h"
 #include "llvm/CodeGen/TargetPassConfig.h"
 #include "llvm/CodeGen/ValueTypes.h"
 #include "llvm/IR/DataLayout.h"
 #include "llvm/IR/Function.h"
 #include "llvm/IR/IRBuilder.h"
 #include "llvm/IR/Instructions.h"
 #include "llvm/IR/IntrinsicInst.h"
 #include "llvm/IR/IntrinsicsX86.h"
 #include "llvm/IR/PatternMatch.h"
 #include "llvm/InitializePasses.h"
 #include "llvm/Pass.h"
 #include "llvm/Target/TargetMachine.h"
+#include "llvm/Transforms/Utils/AssumeBundleBuilder.h"
+#include "llvm/Transforms/Utils/Local.h"
 
 using namespace llvm;
 using namespace PatternMatch;
 
 #define DEBUG_TYPE "lower-amx-type"
 
-static AllocaInst *createAllocaInstAtEntry(IRBuilder<> &Builder,
-                                           BasicBlock *BB) {
+static bool isAMXCast(Instruction *II) {
+  return match(II,
+               m_Intrinsic<Intrinsic::x86_cast_vector_to_tile>(m_Value())) ||
+         match(II, m_Intrinsic<Intrinsic::x86_cast_tile_to_vector>(m_Value()));
+}
+
+static AllocaInst *createAllocaInstAtEntry(IRBuilder<> &Builder, BasicBlock *BB,
+                                           Type *Ty) {
   Function &F = *BB->getParent();
   Module *M = BB->getModule();
   const DataLayout &DL = M->getDataLayout();
 
-  Type *V256I32Ty = VectorType::get(Builder.getInt32Ty(), 256, false);
   LLVMContext &Ctx = Builder.getContext();
   auto AllocaAlignment = DL.getPrefTypeAlign(Type::getX86_AMXTy(Ctx));
   unsigned AllocaAS = DL.getAllocaAddrSpace();
   AllocaInst *AllocaRes =
-      new AllocaInst(V256I32Ty, AllocaAS, "", &F.getEntryBlock().front());
+      new AllocaInst(Ty, AllocaAS, "", &F.getEntryBlock().front());
   AllocaRes->setAlignment(AllocaAlignment);
   return AllocaRes;
 }
 
-namespace {
-class X86LowerAMXType {
-  Function &Func;
-  TargetMachine *TM = nullptr;
-
-  // In AMX intrinsics we let Shape = {Row, Col}, but the
-  // RealCol = Col / ElementSize. We may use the RealCol
-  // as a new Row for other new created AMX intrinsics.
-  std::map<Value *, Value *> Col2Row;
-
-public:
-  X86LowerAMXType(Function &F, TargetMachine *TargetM) : Func(F), TM(TargetM) {}
-  bool visit();
-  void combineLoadBitcast(LoadInst *LD, BitCastInst *Bitcast);
-  void combineBitcastStore(BitCastInst *Bitcast, StoreInst *ST);
-  bool transformBitcast(BitCastInst *Bitcast);
-  std::pair<Value *, Value *> getShape(IntrinsicInst *II, unsigned OpNo);
-  Value *getRowFromCol(Instruction *II, Value *V, unsigned Granularity);
-};
-
-Value *X86LowerAMXType::getRowFromCol(Instruction *II, Value *V,
-                                      unsigned Granularity) {
-  if (Col2Row.count(V))
-    return Col2Row[V];
-  IRBuilder<> Builder(&*II->getParent()->getFirstInsertionPt());
-  if (auto *I = dyn_cast<Instruction>(V)) {
-    BasicBlock::iterator Iter = I->getIterator();
-    ++Iter;
-    Builder.SetInsertPoint(&*Iter);
-  }
-  ConstantInt *Gran = Builder.getInt16(Granularity);
-  Value *RealRow = Builder.CreateUDiv(V, Gran);
-  Col2Row[V] = RealRow;
-  return RealRow;
+static Instruction *getFirstNonAllocaInTheEntryBlock(Function &F) {
+  for (Instruction &I : F.getEntryBlock())
+    if (!isa<AllocaInst>(&I))
+      return &I;
+  llvm_unreachable("No terminator in the entry block!");
 }
 
-std::pair<Value *, Value *> X86LowerAMXType::getShape(IntrinsicInst *II,
-                                                      unsigned OpNo) {
+static std::pair<Value *, Value *> getShape(IntrinsicInst *II, unsigned OpNo) {
+  IRBuilder<> Builder(II);
   Value *Row = nullptr, *Col = nullptr;
   switch (II->getIntrinsicID()) {
   default:
     llvm_unreachable("Expect amx intrinsics");
   case Intrinsic::x86_tileloadd64_internal:
   case Intrinsic::x86_tileloaddt164_internal:
   case Intrinsic::x86_tilestored64_internal: {
     Row = II->getArgOperand(0);
case 3:
       Row = II->getArgOperand(0);
       Col = II->getArgOperand(1);
       break;
     case 4:
       Row = II->getArgOperand(0);
       Col = II->getArgOperand(2);
       break;
     case 5:
-      Row = II->getArgOperand(2);
-      // FIXME: There is a design bug for AMX shape, which the Col should be
-      // Col/4 if it will be used as Row, but current Greedy RA can't handle
-      // this case well, it may failed if we generate a new Shape definition.
-      // So Let's just do it in O0 first.
-      // Row = Row / 4
-      if (TM->getOptLevel() == CodeGenOpt::None)
-        Row = getRowFromCol(II, Row, 4);
+      if (isa<ConstantInt>(II->getArgOperand(2)))
+        Row = Builder.getInt16(
+            (dyn_cast<ConstantInt>(II->getOperand(2))->getSExtValue()) / 4);
+      else if (isa<Instruction>(II->getArgOperand(2))) {
+        // When it is not a const value and it is not a function argument, we
+        // create Row after the definition of II->getOperand(2) instead of
+        // before II. For example, II is %118, we try to getshape for %117:
+        //   %117 = call x86_amx @llvm.x86.cast.vector.to.tile.v256i32(<256 x
+        //   i32> %115).
+        //   %118 = call x86_amx @llvm.x86.tdpbf16ps.internal(i16
+        //   %104, i16 %105, i16 %106, x86_amx %110, x86_amx %114, x86_amx
+        //   %117).
+        // If we create %row = udiv i16 %106, 4 before %118(aka. II), then its
+        // definition is after its user(new tileload for %117).
+        // So, the best choice is to create %row right after the definition of
+        // %106.
+        Builder.SetInsertPoint(cast<Instruction>(II->getOperand(2)));
+        Row = Builder.CreateUDiv(II->getOperand(2), Builder.getInt16(4));
+        cast<Instruction>(Row)->moveAfter(cast<Instruction>(II->getOperand(2)));
+      } else {
+        // When it is not a const value and it is a function argument, we create
+        // Row at the entry bb.
+        IRBuilder<> NewBuilder(
+            getFirstNonAllocaInTheEntryBlock(*II->getFunction()));
+        Row = NewBuilder.CreateUDiv(II->getOperand(2), NewBuilder.getInt16(4));
+      }
       Col = II->getArgOperand(1);
       break;
     }
     break;
   }
   }
 
   return std::make_pair(Row, Col);
 }
 
+namespace {
+class X86LowerAMXType {
+  Function &Func;
+
+  // In AMX intrinsics we let Shape = {Row, Col}, but the
+  // RealCol = Col / ElementSize. We may use the RealCol
+  // as a new Row for other new created AMX intrinsics.
+  std::map<Value *, Value *> Col2Row;
+
+public:
+  X86LowerAMXType(Function &F) : Func(F) {}
+  bool visit();
+  void combineLoadBitcast(LoadInst *LD, BitCastInst *Bitcast);
+  void combineBitcastStore(BitCastInst *Bitcast, StoreInst *ST);
+  bool transformBitcast(BitCastInst *Bitcast);
+  Value *getRowFromCol(Instruction *II, Value *V, unsigned Granularity);
+};
+
+Value *X86LowerAMXType::getRowFromCol(Instruction *II, Value *V,
+                                      unsigned Granularity) {
+  if (Col2Row.count(V))
+    return Col2Row[V];
+  IRBuilder<> Builder(&*II->getParent()->getFirstInsertionPt());
+  if (auto *I = dyn_cast<Instruction>(V)) {
+    BasicBlock::iterator Iter = I->getIterator();
+    ++Iter;
+    Builder.SetInsertPoint(&*Iter);
+  }
+  ConstantInt *Gran = Builder.getInt16(Granularity);
+  Value *RealRow = Builder.CreateUDiv(V, Gran);
+  Col2Row[V] = RealRow;
+  return RealRow;
+}
+
 // %src = load <256 x i32>, <256 x i32>* %addr, align 64
 // %2 = bitcast <256 x i32> %src to x86_amx
 // -->
 // %2 = call x86_amx @llvm.x86.tileloadd64.internal(i16 %row, i16 %col,
 // i8* %addr, i64 %stride64)
 void X86LowerAMXType::combineLoadBitcast(LoadInst *LD, BitCastInst *Bitcast) {
   Value *Row = nullptr, *Col = nullptr;
   Use &U = *(Bitcast->use_begin());

 
 // transform bitcast to <store, load> instructions.
 bool X86LowerAMXType::transformBitcast(BitCastInst *Bitcast) {
   IRBuilder<> Builder(Bitcast);
   AllocaInst *AllocaAddr;
   Value *I8Ptr, *Stride;
   auto *Src = Bitcast->getOperand(0);
 
-  auto Prepare = [&]() {
-    AllocaAddr = createAllocaInstAtEntry(Builder, Bitcast->getParent());
+  auto Prepare = [&](Type *MemTy) {
+    AllocaAddr = createAllocaInstAtEntry(Builder, Bitcast->getParent(), MemTy);
     I8Ptr = Builder.CreateBitCast(AllocaAddr, Builder.getInt8PtrTy());
     Stride = Builder.getInt64(64);
   };
 
   if (Bitcast->getType()->isX86_AMXTy()) {
     // %2 = bitcast <256 x i32> %src to x86_amx
     // -->
     // %addr = alloca <256 x i32>, align 64
     // store <256 x i32> %src, <256 x i32>* %addr, align 64
     // %addr2 = bitcast <256 x i32>* to i8*
     // %2 = call x86_amx @llvm.x86.tileloadd64.internal(i16 %row, i16 %col,
     //                                                  i8* %addr2,
     //                                                  i64 64)
     Use &U = *(Bitcast->use_begin());
     unsigned OpNo = U.getOperandNo();
     auto *II = dyn_cast<IntrinsicInst>(U.getUser());
     if (!II)
       return false; // May be bitcast from x86amx to <256 x i32>.
-    Prepare();
+    Prepare(Bitcast->getOperand(0)->getType());
     Builder.CreateStore(Src, AllocaAddr);
     // TODO we can pick an constant operand for the shape.
     Value *Row = nullptr, *Col = nullptr;
     std::tie(Row, Col) = getShape(II, OpNo);
     std::array<Value *, 4> Args = {Row, Col, I8Ptr, Stride};
     Value *NewInst = Builder.CreateIntrinsic(
         Intrinsic::x86_tileloadd64_internal, None, Args);
     Bitcast->replaceAllUsesWith(NewInst);
   } else {
     // %2 = bitcast x86_amx %src to <256 x i32>
     // -->
     // %addr = alloca <256 x i32>, align 64
     // %addr2 = bitcast <256 x i32>* to i8*
     // call void @llvm.x86.tilestored64.internal(i16 %row, i16 %col,
     //                                           i8* %addr2, i64 %stride)
     // %2 = load <256 x i32>, <256 x i32>* %addr, align 64
     auto *II = dyn_cast<IntrinsicInst>(Src);
     if (!II)
       return false; // May be bitcast from <256 x i32> to x86amx.
-    Prepare();
+    Prepare(Bitcast->getType());
     Value *Row = II->getOperand(0);
     Value *Col = II->getOperand(1);
     std::array<Value *, 5> Args = {Row, Col, I8Ptr, Stride, Src};
     Builder.CreateIntrinsic(Intrinsic::x86_tilestored64_internal, None, Args);
     Value *NewInst = Builder.CreateLoad(Bitcast->getType(), AllocaAddr);
     Bitcast->replaceAllUsesWith(NewInst);
   }
 
bool X86VolatileTileData::volatileTileData() {
   }
   return Changed;
 }
 
 } // anonymous namespace
 
 namespace {
 
+class X86LowerAMXCast {
+  Function &Func;
+
+public:
+  X86LowerAMXCast(Function &F) : Func(F) {}
+  bool combineAMXcast(TargetLibraryInfo *TLI);
+  bool transformAMXCast(IntrinsicInst *AMXCast);
+  bool transformAllAMXCast();
+  bool optimizeAMXCastFromPhi(IntrinsicInst *CI, PHINode *PN,
+                              SmallSetVector<Instruction *, 16> &DeadInst);
+};
+
+static bool DCEInstruction(Instruction *I,
+                           SmallSetVector<Instruction *, 16> &WorkList,
+                           const TargetLibraryInfo *TLI) {
+  if (isInstructionTriviallyDead(I, TLI)) {
+    salvageDebugInfo(*I);
+    salvageKnowledge(I);
+
+    // Null out all of the instruction's operands to see if any operand becomes
+    // dead as we go.
+    for (unsigned i = 0, e = I->getNumOperands(); i != e; ++i) {
+      Value *OpV = I->getOperand(i);
+      I->setOperand(i, nullptr);
+
+      if (!OpV->use_empty() || I == OpV)
+        continue;
+
+      // If the operand is an instruction that became dead as we nulled out the
+      // operand, and if it is 'trivially' dead, delete it in a future loop
+      // iteration.
+      if (Instruction *OpI = dyn_cast<Instruction>(OpV)) {
+        if (isInstructionTriviallyDead(OpI, TLI)) {
+          WorkList.insert(OpI);
+        }
+      }
+    }
+    I->eraseFromParent();
+    return true;
+  }
+  return false;
+}
+
+/// This function handles following case
+///
+///     A  ->  B    amxcast
+///     PHI
+///     B  ->  A    amxcast
+///
+/// All the related PHI nodes can be replaced by new PHI nodes with type A.
+/// The uses of \p CI can be changed to the new PHI node corresponding to \p PN.
+bool X86LowerAMXCast::optimizeAMXCastFromPhi(
+    IntrinsicInst *CI, PHINode *PN,
+    SmallSetVector<Instruction *, 16> &DeadInst) {
+  IRBuilder<> Builder(CI);
+  Value *Src = CI->getOperand(0);
+  Type *SrcTy = Src->getType(); // Type B
+  Type *DestTy = CI->getType(); // Type A
+
+  SmallVector<PHINode *, 4> PhiWorklist;
+  SmallSetVector<PHINode *, 4> OldPhiNodes;
+
+  // Find all of the A->B casts and PHI nodes.
+  // We need to inspect all related PHI nodes, but PHIs can be cyclic, so
+  // OldPhiNodes is used to track all known PHI nodes, before adding a new
+  // PHI to PhiWorklist, it is checked against and added to OldPhiNodes first.
+  PhiWorklist.push_back(PN);
+  OldPhiNodes.insert(PN);
+  while (!PhiWorklist.empty()) {
+    auto *OldPN = PhiWorklist.pop_back_val();
+    for (Value *IncValue : OldPN->incoming_values()) {
+      // TODO: currently, We ignore cases where it is a const. In the future, we
+      // might support const.
+      if (isa<Constant>(IncValue))
+        return false;
+
+      if (auto *PNode = dyn_cast<PHINode>(IncValue)) {
+        if (OldPhiNodes.insert(PNode))
+          PhiWorklist.push_back(PNode);
+        continue;
+      }
+      Instruction *ACI = dyn_cast<Instruction>(IncValue);
+      if (ACI && isAMXCast(ACI)) {
+        // Verify it's a A->B cast.
+        Type *TyA = ACI->getOperand(0)->getType();
+        Type *TyB = ACI->getType();
+        if (TyA != DestTy || TyB != SrcTy)
+          return false;
+        continue;
+      }
+      return false;
+    }
+  }
+
+  // Check that each user of each old PHI node is something that we can
+  // rewrite, so that all of the old PHI nodes can be cleaned up afterwards.
+  for (auto *OldPN : OldPhiNodes) {
+    for (User *V : OldPN->users()) {
+      Instruction *ACI = dyn_cast<Instruction>(V);
+      if (ACI && isAMXCast(ACI)) {
+        // Verify it's a B->A cast.
+        Type *TyB = ACI->getOperand(0)->getType();
+        Type *TyA = ACI->getType();
+        if (TyA != DestTy || TyB != SrcTy)
+          return false;
+      } else if (auto *PHI = dyn_cast<PHINode>(V)) {
+        // As long as the user is another old PHI node, then even if we don't
+        // rewrite it, the PHI web we're considering won't have any users
+        // outside itself, so it'll be dead.
+        // example:
+        //   bb.0:
+        //      %0 = amxcast ...
+        //   bb.1:
+        //      %1 = amxcast ...
+        //   bb.2:
+        //      %goodphi = phi %0, %1
+        //      %3 = amxcast %goodphi
+        //   bb.3:
+        //      %goodphi2 = phi %0, %goodphi
+        //      %4 = amxcast %goodphi2
+        // When optimizeAMXCastFromPhi process %3 and %goodphi, %goodphi2 is
+        // outside the phi-web, so the combination stop When
+        // optimizeAMXCastFromPhi process %4 and %goodphi2, the optimization
+        // will be done.
+        if (OldPhiNodes.count(PHI) == 0)
+          return false;
+      } else
+        return false;
+    }
+  }
+
+  // For each old PHI node, create a corresponding new PHI node with a type A.
+  SmallDenseMap<PHINode *, PHINode *> NewPNodes;
+  for (auto *OldPN : OldPhiNodes) {
+    Builder.SetInsertPoint(OldPN);
+    PHINode *NewPN = Builder.CreatePHI(DestTy, OldPN->getNumOperands());
+    NewPNodes[OldPN] = NewPN;
+  }
+
+  // Fill in the operands of new PHI nodes.
+  for (auto *OldPN : OldPhiNodes) {
+    PHINode *NewPN = NewPNodes[OldPN];
+    for (unsigned j = 0, e = OldPN->getNumOperands(); j != e; ++j) {
+      Value *V = OldPN->getOperand(j);
+      Value *NewV = nullptr;
+      Instruction *ACI = dyn_cast<Instruction>(V);
+      // There should not be a AMXcast from a const.
+      if (ACI && isAMXCast(ACI))
+        NewV = ACI->getOperand(0);
+      else if (auto *PrevPN = dyn_cast<PHINode>(V))
+        NewV = NewPNodes[PrevPN];
+      assert(NewV);
+      NewPN->addIncoming(NewV, OldPN->getIncomingBlock(j));
+    }
+  }
+
+  // Traverse all accumulated PHI nodes and process its users,
+  // which are Stores and BitcCasts. Without this processing
+  // NewPHI nodes could be replicated and could lead to extra
+  // moves generated after DeSSA.
+  // If there is a store with type B, change it to type A.
+
+  // Replace users of BitCast B->A with NewPHI. These will help
+  // later to get rid of a closure formed by OldPHI nodes.
+  for (auto *OldPN : OldPhiNodes) {
+    PHINode *NewPN = NewPNodes[OldPN];
+    for (User *V : make_early_inc_range(OldPN->users())) {
+      Instruction *ACI = dyn_cast<Instruction>(V);
+      if (ACI && isAMXCast(ACI)) {
+        Type *TyB = ACI->getOperand(0)->getType();
+        Type *TyA = ACI->getType();
+        assert(TyA == DestTy && TyB == SrcTy);
+        (void)TyA;
+        (void)TyB;
+        ACI->replaceAllUsesWith(NewPN);
+        DeadInst.insert(ACI);
+      } else if (auto *PHI = dyn_cast<PHINode>(V)) {
+        // We don't need to push PHINode into DeadInst since they are operands
+        // of rootPN DCE can safely delete rootPN's operands if rootPN is dead.
+        assert(OldPhiNodes.contains(PHI));
+        (void)PHI;
+      } else
+        llvm_unreachable("all uses should be handled");
+    }
+  }
+  return true;
+}
+
+bool X86LowerAMXCast::combineAMXcast(TargetLibraryInfo *TLI) {
+  bool Change = false;
+  // Collect tile cast instruction.
+  SmallVector<Instruction *, 8> Vec2TileInsts;
+  SmallVector<Instruction *, 8> Tile2VecInsts;
+  SmallVector<Instruction *, 8> PhiCastWorkList;
+  SmallSetVector<Instruction *, 16> DeadInst;
+  for (BasicBlock &BB : Func) {
+    for (Instruction &I : BB) {
+      Value *Vec;
+      if (match(&I,
+                m_Intrinsic<Intrinsic::x86_cast_vector_to_tile>(m_Value(Vec))))
+        Vec2TileInsts.push_back(&I);
+      else if (match(&I, m_Intrinsic<Intrinsic::x86_cast_tile_to_vector>(
+                             m_Value(Vec))))
+        Tile2VecInsts.push_back(&I);
+    }
+  }
+
+  auto Convert = [&](SmallVectorImpl<Instruction *> &Insts, Intrinsic::ID IID) {
+    for (auto *Inst : Insts) {
+      for (User *U : Inst->users()) {
+        IntrinsicInst *II = dyn_cast<IntrinsicInst>(U);
+        if (!II || II->getIntrinsicID() != IID)
+          continue;
+        // T1 = vec2tile V0
+        // V2 = tile2vec T1
+        // V3 = OP V2
+        // -->
+        // T1 = vec2tile V0
+        // V2 = tile2vec T1
+        // V3 = OP V0
+        II->replaceAllUsesWith(Inst->getOperand(0));
+        Change = true;
+      }
+    }
+  };
+
+  Convert(Vec2TileInsts, Intrinsic::x86_cast_tile_to_vector);
+  Convert(Tile2VecInsts, Intrinsic::x86_cast_vector_to_tile);
+
+  auto EraseInst = [](SmallVectorImpl<Instruction *> &Insts) {
+    for (auto *Inst : Insts) {
+      if (Inst->use_empty())
+        Inst->eraseFromParent();
+    }
+  };
+
+  EraseInst(Vec2TileInsts);
+  EraseInst(Tile2VecInsts);
+
+  // Handle the A->B->A cast, and there is an intervening PHI node.
+  for (BasicBlock &BB : Func) {
+    for (Instruction &I : BB) {
+      if (isAMXCast(&I)) {
+        if (PHINode *PN = dyn_cast<PHINode>(I.getOperand(0)))
+          PhiCastWorkList.push_back(&I);
+      }
+    }
+  }
+  for (auto *I : PhiCastWorkList) {
+    // We skip the dead Amxcast.
+    if (DeadInst.contains(I))
+      continue;
+    PHINode *PN = cast<PHINode>(I->getOperand(0));
+    if (optimizeAMXCastFromPhi(cast<IntrinsicInst>(I), PN, DeadInst)) {
+      DeadInst.insert(PN);
+      Change = true;
+    }
+  }
+
+  // Since we create new phi and merge AMXCast, some old phis and AMXCast might
+  // have no uses. We do some DeadCodeElimination for them.
+  while (!DeadInst.empty()) {
+    Instruction *I = DeadInst.pop_back_val();
+    Change |= DCEInstruction(I, DeadInst, TLI);
+  }
+  return Change;
+}
+
+// There might be remaining AMXcast after combineAMXcast and they should be
+// handled elegantly.
+bool X86LowerAMXCast::transformAMXCast(IntrinsicInst *AMXCast) {
+  IRBuilder<> Builder(AMXCast);
+  AllocaInst *AllocaAddr;
+  Value *I8Ptr, *Stride;
+  auto *Src = AMXCast->getOperand(0);
+
+  auto Prepare = [&](Type *MemTy) {
+    AllocaAddr = createAllocaInstAtEntry(Builder, AMXCast->getParent(), MemTy);
+    I8Ptr = Builder.CreateBitCast(AllocaAddr, Builder.getInt8PtrTy());
+    Stride = Builder.getInt64(64);
+  };
+
+  if (AMXCast->getType()->isX86_AMXTy()) {
+    // %2 = amxcast <225 x i32> %src to x86_amx
+    // call void @llvm.x86.tilestored64.internal(i16 15, i16 60,
+    //                                           i8* %addr3, i64 60, x86_amx %2)
+    // -->
+    // %addr = alloca <225 x i32>, align 64
+    // store <225 x i32> %src, <225 x i32>* %addr, align 64
+    // %addr2 = bitcast <225 x i32>* %addr to i8*
+    // %2 = call x86_amx @llvm.x86.tileloadd64.internal(i16 15, i16 60,
+    //                                                  i8* %addr2,
+    //                                                  i64 60)
+    // call void @llvm.x86.tilestored64.internal(i16 15, i16 60,
+    //                                           i8* %addr3, i64 60, x86_amx %2)
+    Use &U = *(AMXCast->use_begin());
+    unsigned OpNo = U.getOperandNo();
+    auto *II = dyn_cast<IntrinsicInst>(U.getUser());
+    if (!II)
+      return false; // May be bitcast from x86amx to <256 x i32>.
+    Prepare(AMXCast->getOperand(0)->getType());
+    Builder.CreateStore(Src, AllocaAddr);
+    // TODO we can pick an constant operand for the shape.
+    Value *Row = nullptr, *Col = nullptr;
+    std::tie(Row, Col) = getShape(II, OpNo);
+    std::array<Value *, 4> Args = {
+        Row, Col, I8Ptr, Builder.CreateSExt(Col, Builder.getInt64Ty())};
+    Value *NewInst = Builder.CreateIntrinsic(
+        Intrinsic::x86_tileloadd64_internal, None, Args);
+    AMXCast->replaceAllUsesWith(NewInst);
+    AMXCast->eraseFromParent();
+  } else {
+    // %2 = amxcast x86_amx %src to <225 x i32>
+    // -->
+    // %addr = alloca <225 x i32>, align 64
+    // %addr2 = bitcast <225 x i32>* to i8*
+    // call void @llvm.x86.tilestored64.internal(i16 %row, i16 %col,
+    //                                           i8* %addr2, i64 %stride)
+    // %2 = load <225 x i32>, <225 x i32>* %addr, align 64
+    auto *II = dyn_cast<IntrinsicInst>(Src);
+    if (!II)
+      return false; // May be bitcast from <256 x i32> to x86amx.
+    Prepare(AMXCast->getType());
+    Value *Row = II->getOperand(0);
+    Value *Col = II->getOperand(1);
+    std::array<Value *, 5> Args = {
+        Row, Col, I8Ptr, Builder.CreateSExt(Col, Builder.getInt64Ty()), Src};
+    Builder.CreateIntrinsic(Intrinsic::x86_tilestored64_internal, None, Args);
+    Value *NewInst = Builder.CreateLoad(AMXCast->getType(), AllocaAddr);
+    AMXCast->replaceAllUsesWith(NewInst);
+    AMXCast->eraseFromParent();
+  }
+
+  return true;
+}
+
+bool X86LowerAMXCast::transformAllAMXCast() {
+  bool Change = false;
+  // Collect tile cast instruction.
+  SmallVector<Instruction *, 8> WorkLists;
+  for (BasicBlock &BB : Func) {
+    for (Instruction &I : BB) {
+      if (isAMXCast(&I))
+        WorkLists.push_back(&I);
+    }
+  }
+
+  for (auto *Inst : WorkLists) {
+    Change |= transformAMXCast(cast<IntrinsicInst>(Inst));
+  }
+
+  return Change;
+}
+
+} // anonymous namespace
+
+namespace {
+
 class X86LowerAMXTypeLegacyPass : public FunctionPass {
 public:
   static char ID;
 
   X86LowerAMXTypeLegacyPass() : FunctionPass(ID) {
     initializeX86LowerAMXTypeLegacyPassPass(*PassRegistry::getPassRegistry());
   }
 
   bool runOnFunction(Function &F) override {
     TargetMachine *TM = &getAnalysis<TargetPassConfig>().getTM<TargetMachine>();
+    TargetLibraryInfo *TLI =
+        &getAnalysis<TargetLibraryInfoWrapperPass>().getTLI(F);
+    X86LowerAMXCast LAC(F);
+    LAC.combineAMXcast(TLI);
+    // There might be remaining AMXcast after combineAMXcast and they should be
+    // handled elegantly.
+    LAC.transformAllAMXCast();
 
-    X86LowerAMXType LAT(F, TM);
+    X86LowerAMXType LAT(F);
     bool C = LAT.visit();
 
     // Prepare for fast register allocation at O0.
     // Todo: May better check the volatile model of AMX code, not just
     // by checking Attribute::OptimizeNone and CodeGenOpt::None.
     if (TM->getOptLevel() == CodeGenOpt::None) {
       // If Front End not use O0 but the Mid/Back end use O0, (e.g.
       // "Clang -O2 -S -emit-llvm t.c" + "llc t.ll") we should make
       // sure the amx data is volatile, that is nessary for AMX fast
       // register allocation.
       if (!F.hasFnAttribute(Attribute::OptimizeNone)) {
         X86VolatileTileData VTD(F);
         C = VTD.volatileTileData() || C;
       }
     }
 
     return C;
   }
 
   void getAnalysisUsage(AnalysisUsage &AU) const override {
     AU.setPreservesCFG();
     AU.addRequired<TargetPassConfig>();
+    AU.addRequired<TargetLibraryInfoWrapperPass>();
   }
 };
 
 } // anonymous namespace
 
 static const char PassName[] = "Lower AMX type for load/store";
 char X86LowerAMXTypeLegacyPass::ID = 0;
 INITIALIZE_PASS_BEGIN(X86LowerAMXTypeLegacyPass, DEBUG_TYPE, PassName, false,
                       false)
 INITIALIZE_PASS_DEPENDENCY(TargetPassConfig)
+INITIALIZE_PASS_DEPENDENCY(TargetLibraryInfoWrapperPass)
 INITIALIZE_PASS_END(X86LowerAMXTypeLegacyPass, DEBUG_TYPE, PassName, false,
                     false)
 
 FunctionPass *llvm::createX86LowerAMXTypePass() {
   return new X86LowerAMXTypeLegacyPass();
 }

llvm/test/CodeGen/X86/AMX/amx-type.ll

 define dso_local void @__tile_dpbssd(%struct.__tile_str* nocapture %0, %struct.__tile_str* nocapture readonly byval(%struct.__tile_str) align 64 %1, %struct.__tile_str* nocapture readonly byval(%struct.__tile_str) align 64 %2) local_unnamed_addr {
 ; CHECK-LABEL: @__tile_dpbssd(
 ; CHECK-NEXT:    [[TMP4:%.*]] = getelementptr inbounds [[STRUCT___TILE_STR:%.*]], %struct.__tile_str* [[TMP1:%.*]], i64 0, i32 0
 ; CHECK-NEXT:    [[TMP5:%.*]] = load i16, i16* [[TMP4]], align 64
 ; CHECK-NEXT:    [[TMP6:%.*]] = getelementptr inbounds [[STRUCT___TILE_STR]], %struct.__tile_str* [[TMP2:%.*]], i64 0, i32 1
 ; CHECK-NEXT:    [[TMP7:%.*]] = load i16, i16* [[TMP6]], align 2
 ; CHECK-NEXT:    [[TMP8:%.*]] = getelementptr inbounds [[STRUCT___TILE_STR]], %struct.__tile_str* [[TMP1]], i64 0, i32 1
 ; CHECK-NEXT:    [[TMP9:%.*]] = load i16, i16* [[TMP8]], align 2
-; CHECK-NEXT:    [[TMP10:%.*]] = getelementptr inbounds [[STRUCT___TILE_STR]], %struct.__tile_str* [[TMP0:%.*]], i64 0, i32 2
-; CHECK-NEXT:    [[TMP11:%.*]] = bitcast <256 x i32>* [[TMP10]] to i8*
-; CHECK-NEXT:    [[TMP12:%.*]] = call x86_amx @llvm.x86.tileloadd64.internal(i16 [[TMP5]], i16 [[TMP7]], i8* [[TMP11]], i64 64)
-; CHECK-NEXT:    [[TMP13:%.*]] = getelementptr inbounds [[STRUCT___TILE_STR]], %struct.__tile_str* [[TMP1]], i64 0, i32 2
-; CHECK-NEXT:    [[TMP14:%.*]] = bitcast <256 x i32>* [[TMP13]] to i8*
-; CHECK-NEXT:    [[TMP15:%.*]] = call x86_amx @llvm.x86.tileloadd64.internal(i16 [[TMP5]], i16 [[TMP9]], i8* [[TMP14]], i64 64)
-; CHECK-NEXT:    [[TMP16:%.*]] = getelementptr inbounds [[STRUCT___TILE_STR]], %struct.__tile_str* [[TMP2]], i64 0, i32 2
-; CHECK-NEXT:    [[TMP17:%.*]] = bitcast <256 x i32>* [[TMP16]] to i8*
-; CHECK-NEXT:    [[TMP18:%.*]] = call x86_amx @llvm.x86.tileloadd64.internal(i16 [[TMP9]], i16 [[TMP7]], i8* [[TMP17]], i64 64)
-; CHECK-NEXT:    [[TMP19:%.*]] = tail call x86_amx @llvm.x86.tdpbssd.internal(i16 [[TMP5]], i16 [[TMP7]], i16 [[TMP9]], x86_amx [[TMP12]], x86_amx [[TMP15]], x86_amx [[TMP18]])
-; CHECK-NEXT:    [[TMP20:%.*]] = bitcast <256 x i32>* [[TMP10]] to i8*
-; CHECK-NEXT:    call void @llvm.x86.tilestored64.internal(i16 [[TMP5]], i16 [[TMP7]], i8* [[TMP20]], i64 64, x86_amx [[TMP19]])
+; CHECK-NEXT:    [[TMP10:%.*]] = udiv i16 [[TMP9]], 4
+; CHECK-NEXT:    [[TMP11:%.*]] = getelementptr inbounds [[STRUCT___TILE_STR]], %struct.__tile_str* [[TMP0:%.*]], i64 0, i32 2
+; CHECK-NEXT:    [[TMP12:%.*]] = bitcast <256 x i32>* [[TMP11]] to i8*
+; CHECK-NEXT:    [[TMP13:%.*]] = call x86_amx @llvm.x86.tileloadd64.internal(i16 [[TMP5]], i16 [[TMP7]], i8* [[TMP12]], i64 64)
+; CHECK-NEXT:    [[TMP14:%.*]] = getelementptr inbounds [[STRUCT___TILE_STR]], %struct.__tile_str* [[TMP1]], i64 0, i32 2
+; CHECK-NEXT:    [[TMP15:%.*]] = bitcast <256 x i32>* [[TMP14]] to i8*
+; CHECK-NEXT:    [[TMP16:%.*]] = call x86_amx @llvm.x86.tileloadd64.internal(i16 [[TMP5]], i16 [[TMP9]], i8* [[TMP15]], i64 64)
+; CHECK-NEXT:    [[TMP17:%.*]] = getelementptr inbounds [[STRUCT___TILE_STR]], %struct.__tile_str* [[TMP2]], i64 0, i32 2
+; CHECK-NEXT:    [[TMP18:%.*]] = bitcast <256 x i32>* [[TMP17]] to i8*
+; CHECK-NEXT:    [[TMP19:%.*]] = call x86_amx @llvm.x86.tileloadd64.internal(i16 [[TMP10]], i16 [[TMP7]], i8* [[TMP18]], i64 64)
+; CHECK-NEXT:    [[TMP20:%.*]] = tail call x86_amx @llvm.x86.tdpbssd.internal(i16 [[TMP5]], i16 [[TMP7]], i16 [[TMP9]], x86_amx [[TMP13]], x86_amx [[TMP16]], x86_amx [[TMP19]])
+; CHECK-NEXT:    [[TMP21:%.*]] = bitcast <256 x i32>* [[TMP11]] to i8*
+; CHECK-NEXT:    call void @llvm.x86.tilestored64.internal(i16 [[TMP5]], i16 [[TMP7]], i8* [[TMP21]], i64 64, x86_amx [[TMP20]])
 ; CHECK-NEXT:    ret void
 ;
   %4 = getelementptr inbounds %struct.__tile_str, %struct.__tile_str* %1, i64 0, i32 0
   %5 = load i16, i16* %4, align 64
   %6 = getelementptr inbounds %struct.__tile_str, %struct.__tile_str* %2, i64 0, i32 1
   %7 = load i16, i16* %6, align 2
   %8 = getelementptr inbounds %struct.__tile_str, %struct.__tile_str* %1, i64 0, i32 1
   %9 = load i16, i16* %8, align 2
;
   %19 = tail call x86_amx @llvm.x86.tdpbssd.internal(i16 %5, i16 %7, i16 %9, x86_amx %12, x86_amx %15, x86_amx %18)
   %20 = bitcast x86_amx %19 to <256 x i32>
   store <256 x i32> %20, <256 x i32>* %10, align 64
   ret void
 }
 
 define dso_local void @__tile_dpbsud(i16 %m, i16 %n, i16 %k, <256 x i32>* %pc, <256 x i32>* %pa, <256 x i32>* %pb) {
 ; CHECK-LABEL: @__tile_dpbsud(
-; CHECK-NEXT:    [[TMP1:%.*]] = bitcast <256 x i32>* [[PA:%.*]] to i8*
-; CHECK-NEXT:    [[TMP2:%.*]] = call x86_amx @llvm.x86.tileloadd64.internal(i16 [[M:%.*]], i16 [[K:%.*]], i8* [[TMP1]], i64 64)
-; CHECK-NEXT:    [[TMP3:%.*]] = bitcast <256 x i32>* [[PB:%.*]] to i8*
-; CHECK-NEXT:    [[TMP4:%.*]] = call x86_amx @llvm.x86.tileloadd64.internal(i16 [[K]], i16 [[N:%.*]], i8* [[TMP3]], i64 64)
-; CHECK-NEXT:    [[TMP5:%.*]] = bitcast <256 x i32>* [[PC:%.*]] to i8*
-; CHECK-NEXT:    [[TMP6:%.*]] = call x86_amx @llvm.x86.tileloadd64.internal(i16 [[M]], i16 [[N]], i8* [[TMP5]], i64 64)
-; CHECK-NEXT:    [[T6:%.*]] = tail call x86_amx @llvm.x86.tdpbsud.internal(i16 [[M]], i16 [[N]], i16 [[K]], x86_amx [[TMP6]], x86_amx [[TMP2]], x86_amx [[TMP4]])
-; CHECK-NEXT:    [[TMP7:%.*]] = bitcast <256 x i32>* [[PC]] to i8*
-; CHECK-NEXT:    call void @llvm.x86.tilestored64.internal(i16 [[M]], i16 [[N]], i8* [[TMP7]], i64 64, x86_amx [[T6]])
+; CHECK-NEXT:    [[TMP1:%.*]] = udiv i16 [[K:%.*]], 4
+; CHECK-NEXT:    [[TMP2:%.*]] = bitcast <256 x i32>* [[PA:%.*]] to i8*
+; CHECK-NEXT:    [[TMP3:%.*]] = call x86_amx @llvm.x86.tileloadd64.internal(i16 [[M:%.*]], i16 [[K]], i8* [[TMP2]], i64 64)
+; CHECK-NEXT:    [[TMP4:%.*]] = bitcast <256 x i32>* [[PB:%.*]] to i8*
+; CHECK-NEXT:    [[TMP5:%.*]] = call x86_amx @llvm.x86.tileloadd64.internal(i16 [[TMP1]], i16 [[N:%.*]], i8* [[TMP4]], i64 64)
+; CHECK-NEXT:    [[TMP6:%.*]] = bitcast <256 x i32>* [[PC:%.*]] to i8*
+; CHECK-NEXT:    [[TMP7:%.*]] = call x86_amx @llvm.x86.tileloadd64.internal(i16 [[M]], i16 [[N]], i8* [[TMP6]], i64 64)
+; CHECK-NEXT:    [[T6:%.*]] = tail call x86_amx @llvm.x86.tdpbsud.internal(i16 [[M]], i16 [[N]], i16 [[K]], x86_amx [[TMP7]], x86_amx [[TMP3]], x86_amx [[TMP5]])
+; CHECK-NEXT:    [[TMP8:%.*]] = bitcast <256 x i32>* [[PC]] to i8*
+; CHECK-NEXT:    call void @llvm.x86.tilestored64.internal(i16 [[M]], i16 [[N]], i8* [[TMP8]], i64 64, x86_amx [[T6]])
 ; CHECK-NEXT:    ret void
 ;
   %t0 = load <256 x i32>, <256 x i32>* %pa, align 64
   %t1 = bitcast <256 x i32> %t0 to x86_amx
   %t2 = load <256 x i32>, <256 x i32>* %pb, align 64
   %t3 = bitcast <256 x i32> %t2 to x86_amx
   %t4 = load <256 x i32>, <256 x i32>* %pc, align 64
   %t5 = bitcast <256 x i32> %t4 to x86_amx
   %t6 = tail call x86_amx @llvm.x86.tdpbsud.internal(i16 %m, i16 %n, i16 %k, x86_amx %t5, x86_amx %t1, x86_amx %t3)
   %t7 = bitcast x86_amx %t6 to <256 x i32>
   store <256 x i32> %t7, <256 x i32>* %pc, align 64
   ret void
 }
 
 define dso_local void @__tile_dpbusd(i16 %m, i16 %n, i16 %k, <256 x i32>* %pc, <256 x i32>* %pa, <256 x i32>* %pb) {
 ; CHECK-LABEL: @__tile_dpbusd(
-; CHECK-NEXT:    [[TMP1:%.*]] = bitcast <256 x i32>* [[PA:%.*]] to i8*
-; CHECK-NEXT:    [[TMP2:%.*]] = call x86_amx @llvm.x86.tileloadd64.internal(i16 [[M:%.*]], i16 [[K:%.*]], i8* [[TMP1]], i64 64)
-; CHECK-NEXT:    [[TMP3:%.*]] = bitcast <256 x i32>* [[PB:%.*]] to i8*
-; CHECK-NEXT:    [[TMP4:%.*]] = call x86_amx @llvm.x86.tileloadd64.internal(i16 [[K]], i16 [[N:%.*]], i8* [[TMP3]], i64 64)
-; CHECK-NEXT:    [[TMP5:%.*]] = bitcast <256 x i32>* [[PC:%.*]] to i8*
-; CHECK-NEXT:    [[TMP6:%.*]] = call x86_amx @llvm.x86.tileloadd64.internal(i16 [[M]], i16 [[N]], i8* [[TMP5]], i64 64)
-; CHECK-NEXT:    [[T6:%.*]] = tail call x86_amx @llvm.x86.tdpbusd.internal(i16 [[M]], i16 [[N]], i16 [[K]], x86_amx [[TMP6]], x86_amx [[TMP2]], x86_amx [[TMP4]])
-; CHECK-NEXT:    [[TMP7:%.*]] = bitcast <256 x i32>* [[PC]] to i8*
-; CHECK-NEXT:    call void @llvm.x86.tilestored64.internal(i16 [[M]], i16 [[N]], i8* [[TMP7]], i64 64, x86_amx [[T6]])
+; CHECK-NEXT:    [[TMP1:%.*]] = udiv i16 [[K:%.*]], 4
+; CHECK-NEXT:    [[TMP2:%.*]] = bitcast <256 x i32>* [[PA:%.*]] to i8*
+; CHECK-NEXT:    [[TMP3:%.*]] = call x86_amx @llvm.x86.tileloadd64.internal(i16 [[M:%.*]], i16 [[K]], i8* [[TMP2]], i64 64)
+; CHECK-NEXT:    [[TMP4:%.*]] = bitcast <256 x i32>* [[PB:%.*]] to i8*
+; CHECK-NEXT:    [[TMP5:%.*]] = call x86_amx @llvm.x86.tileloadd64.internal(i16 [[TMP1]], i16 [[N:%.*]], i8* [[TMP4]], i64 64)
+; CHECK-NEXT:    [[TMP6:%.*]] = bitcast <256 x i32>* [[PC:%.*]] to i8*
+; CHECK-NEXT:    [[TMP7:%.*]] = call x86_amx @llvm.x86.tileloadd64.internal(i16 [[M]], i16 [[N]], i8* [[TMP6]], i64 64)
+; CHECK-NEXT:    [[T6:%.*]] = tail call x86_amx @llvm.x86.tdpbusd.internal(i16 [[M]], i16 [[N]], i16 [[K]], x86_amx [[TMP7]], x86_amx [[TMP3]], x86_amx [[TMP5]])
+; CHECK-NEXT:    [[TMP8:%.*]] = bitcast <256 x i32>* [[PC]] to i8*
+; CHECK-NEXT:    call void @llvm.x86.tilestored64.internal(i16 [[M]], i16 [[N]], i8* [[TMP8]], i64 64, x86_amx [[T6]])
 ; CHECK-NEXT:    ret void
 ;
   %t0 = load <256 x i32>, <256 x i32>* %pa, align 64
   %t1 = bitcast <256 x i32> %t0 to x86_amx
   %t2 = load <256 x i32>, <256 x i32>* %pb, align 64
   %t3 = bitcast <256 x i32> %t2 to x86_amx
   %t4 = load <256 x i32>, <256 x i32>* %pc, align 64
   %t5 = bitcast <256 x i32> %t4 to x86_amx
   %t6 = tail call x86_amx @llvm.x86.tdpbusd.internal(i16 %m, i16 %n, i16 %k, x86_amx %t5, x86_amx %t1, x86_amx %t3)
   %t7 = bitcast x86_amx %t6 to <256 x i32>
   store <256 x i32> %t7, <256 x i32>* %pc, align 64
   ret void
 }
 
 define dso_local void @__tile_dpbuud(i16 %m, i16 %n, i16 %k, <256 x i32>* %pc, <256 x i32>* %pa, <256 x i32>* %pb) {
 ; CHECK-LABEL: @__tile_dpbuud(
-; CHECK-NEXT:    [[TMP1:%.*]] = bitcast <256 x i32>* [[PA:%.*]] to i8*
-; CHECK-NEXT:    [[TMP2:%.*]] = call x86_amx @llvm.x86.tileloadd64.internal(i16 [[M:%.*]], i16 [[K:%.*]], i8* [[TMP1]], i64 64)
-; CHECK-NEXT:    [[TMP3:%.*]] = bitcast <256 x i32>* [[PB:%.*]] to i8*
-; CHECK-NEXT:    [[TMP4:%.*]] = call x86_amx @llvm.x86.tileloadd64.internal(i16 [[K]], i16 [[N:%.*]], i8* [[TMP3]], i64 64)
-; CHECK-NEXT:    [[TMP5:%.*]] = bitcast <256 x i32>* [[PC:%.*]] to i8*
-; CHECK-NEXT:    [[TMP6:%.*]] = call x86_amx @llvm.x86.tileloadd64.internal(i16 [[M]], i16 [[N]], i8* [[TMP5]], i64 64)
-; CHECK-NEXT:    [[T6:%.*]] = tail call x86_amx @llvm.x86.tdpbuud.internal(i16 [[M]], i16 [[N]], i16 [[K]], x86_amx [[TMP6]], x86_amx [[TMP2]], x86_amx [[TMP4]])
-; CHECK-NEXT:    [[TMP7:%.*]] = bitcast <256 x i32>* [[PC]] to i8*
-; CHECK-NEXT:    call void @llvm.x86.tilestored64.internal(i16 [[M]], i16 [[N]], i8* [[TMP7]], i64 64, x86_amx [[T6]])
+; CHECK-NEXT:    [[TMP1:%.*]] = udiv i16 [[K:%.*]], 4
+; CHECK-NEXT:    [[TMP2:%.*]] = bitcast <256 x i32>* [[PA:%.*]] to i8*
+; CHECK-NEXT:    [[TMP3:%.*]] = call x86_amx @llvm.x86.tileloadd64.internal(i16 [[M:%.*]], i16 [[K]], i8* [[TMP2]], i64 64)
+; CHECK-NEXT:    [[TMP4:%.*]] = bitcast <256 x i32>* [[PB:%.*]] to i8*
+; CHECK-NEXT:    [[TMP5:%.*]] = call x86_amx @llvm.x86.tileloadd64.internal(i16 [[TMP1]], i16 [[N:%.*]], i8* [[TMP4]], i64 64)
+; CHECK-NEXT:    [[TMP6:%.*]] = bitcast <256 x i32>* [[PC:%.*]] to i8*
+; CHECK-NEXT:    [[TMP7:%.*]] = call x86_amx @llvm.x86.tileloadd64.internal(i16 [[M]], i16 [[N]], i8* [[TMP6]], i64 64)
+; CHECK-NEXT:    [[T6:%.*]] = tail call x86_amx @llvm.x86.tdpbuud.internal(i16 [[M]], i16 [[N]], i16 [[K]], x86_amx [[TMP7]], x86_amx [[TMP3]], x86_amx [[TMP5]])
+; CHECK-NEXT:    [[TMP8:%.*]] = bitcast <256 x i32>* [[PC]] to i8*
+; CHECK-NEXT:    call void @llvm.x86.tilestored64.internal(i16 [[M]], i16 [[N]], i8* [[TMP8]], i64 64, x86_amx [[T6]])
 ; CHECK-NEXT:    ret void
 ;
   %t0 = load <256 x i32>, <256 x i32>* %pa, align 64
   %t1 = bitcast <256 x i32> %t0 to x86_amx
   %t2 = load <256 x i32>, <256 x i32>* %pb, align 64
   %t3 = bitcast <256 x i32> %t2 to x86_amx
   %t4 = load <256 x i32>, <256 x i32>* %pc, align 64
   %t5 = bitcast <256 x i32> %t4 to x86_amx
   %t6 = tail call x86_amx @llvm.x86.tdpbuud.internal(i16 %m, i16 %n, i16 %k, x86_amx %t5, x86_amx %t1, x86_amx %t3)
   %t7 = bitcast x86_amx %t6 to <256 x i32>
   store <256 x i32> %t7, <256 x i32>* %pc, align 64
   ret void
 }
 
 define dso_local void @__tile_dpbf16ps(i16 %m, i16 %n, i16 %k, <256 x i32>* %pc, <256 x i32>* %pa, <256 x i32>* %pb) {
 ; CHECK-LABEL: @__tile_dpbf16ps(
-; CHECK-NEXT:    [[TMP1:%.*]] = bitcast <256 x i32>* [[PA:%.*]] to i8*
-; CHECK-NEXT:    [[TMP2:%.*]] = call x86_amx @llvm.x86.tileloadd64.internal(i16 [[M:%.*]], i16 [[K:%.*]], i8* [[TMP1]], i64 64)
-; CHECK-NEXT:    [[TMP3:%.*]] = bitcast <256 x i32>* [[PB:%.*]] to i8*
-; CHECK-NEXT:    [[TMP4:%.*]] = call x86_amx @llvm.x86.tileloadd64.internal(i16 [[K]], i16 [[N:%.*]], i8* [[TMP3]], i64 64)
-; CHECK-NEXT:    [[TMP5:%.*]] = bitcast <256 x i32>* [[PC:%.*]] to i8*
-; CHECK-NEXT:    [[TMP6:%.*]] = call x86_amx @llvm.x86.tileloadd64.internal(i16 [[M]], i16 [[N]], i8* [[TMP5]], i64 64)
-; CHECK-NEXT:    [[T6:%.*]] = tail call x86_amx @llvm.x86.tdpbf16ps.internal(i16 [[M]], i16 [[N]], i16 [[K]], x86_amx [[TMP6]], x86_amx [[TMP2]], x86_amx [[TMP4]])
-; CHECK-NEXT:    [[TMP7:%.*]] = bitcast <256 x i32>* [[PC]] to i8*
-; CHECK-NEXT:    call void @llvm.x86.tilestored64.internal(i16 [[M]], i16 [[N]], i8* [[TMP7]], i64 64, x86_amx [[T6]])
+; CHECK-NEXT:    [[TMP1:%.*]] = udiv i16 [[K:%.*]], 4
+; CHECK-NEXT:    [[TMP2:%.*]] = bitcast <256 x i32>* [[PA:%.*]] to i8*
+; CHECK-NEXT:    [[TMP3:%.*]] = call x86_amx @llvm.x86.tileloadd64.internal(i16 [[M:%.*]], i16 [[K]], i8* [[TMP2]], i64 64)
+; CHECK-NEXT:    [[TMP4:%.*]] = bitcast <256 x i32>* [[PB:%.*]] to i8*
+; CHECK-NEXT:    [[TMP5:%.*]] = call x86_amx @llvm.x86.tileloadd64.internal(i16 [[TMP1]], i16 [[N:%.*]], i8* [[TMP4]], i64 64)
+; CHECK-NEXT:    [[TMP6:%.*]] = bitcast <256 x i32>* [[PC:%.*]] to i8*
+; CHECK-NEXT:    [[TMP7:%.*]] = call x86_amx @llvm.x86.tileloadd64.internal(i16 [[M]], i16 [[N]], i8* [[TMP6]], i64 64)
+; CHECK-NEXT:    [[T6:%.*]] = tail call x86_amx @llvm.x86.tdpbf16ps.internal(i16 [[M]], i16 [[N]], i16 [[K]], x86_amx [[TMP7]], x86_amx [[TMP3]], x86_amx [[TMP5]])
+; CHECK-NEXT:    [[TMP8:%.*]] = bitcast <256 x i32>* [[PC]] to i8*
+; CHECK-NEXT:    call void @llvm.x86.tilestored64.internal(i16 [[M]], i16 [[N]], i8* [[TMP8]], i64 64, x86_amx [[T6]])
 ; CHECK-NEXT:    ret void
 ;
   %t0 = load <256 x i32>, <256 x i32>* %pa, align 64
   %t1 = bitcast <256 x i32> %t0 to x86_amx
   %t2 = load <256 x i32>, <256 x i32>* %pb, align 64
   %t3 = bitcast <256 x i32> %t2 to x86_amx
   %t4 = load <256 x i32>, <256 x i32>* %pc, align 64
   %t5 = bitcast <256 x i32> %t4 to x86_amx

llvm/test/CodeGen/X86/AMX/lat-combine-amx-bitcast.ll

+; NOTE: Assertions have been autogenerated by utils/update_test_checks.py
+; RUN: opt --codegen-opt-level=2 -mtriple=x86_64 -lower-amx-type %s -S | FileCheck %s
+
+define void @combine_amx_cast_inside_bb() {
+; CHECK-LABEL: @combine_amx_cast_inside_bb(
+; CHECK-NEXT:  wrapper_entry:
+; CHECK-NEXT:    [[TMP0:%.*]] = call x86_amx @llvm.x86.tileloadd64.internal(i16 11, i16 40, i8* undef, i64 undef)
+; CHECK-NEXT:    call void @llvm.x86.tilestored64.internal(i16 11, i16 40, i8* undef, i64 undef, x86_amx [[TMP0]])
+; CHECK-NEXT:    ret void
+;
+wrapper_entry:
+  %0 = call x86_amx @llvm.x86.tileloadd64.internal(i16 11, i16 40, i8* undef, i64 undef)
+  %tmp = call <110 x i32> @llvm.x86.cast.tile.to.vector.v110i32(x86_amx %0)
+  %1 = call x86_amx @llvm.x86.cast.vector.to.tile.v110i32(<110 x i32> %tmp)
+  call void @llvm.x86.tilestored64.internal(i16 11, i16 40, i8* undef, i64 undef, x86_amx %1)
+  ret void
+}
+
+; Cases where amxcast can be combined across bb
+; %5 and %6 is combined together since %goodphi's incoming is phi or amxcast
+define void @combine_amx_cast_and_phi() {
+; CHECK-LABEL: @combine_amx_cast_and_phi(
+; CHECK-NEXT:  wrapper_entry:
+; CHECK-NEXT:    [[TMP0:%.*]] = alloca <560 x i8>, align 64
+; CHECK-NEXT:    [[TMP1:%.*]] = alloca <616 x i8>, align 64
+; CHECK-NEXT:    [[TMP2:%.*]] = alloca <110 x i32>, align 64
+; CHECK-NEXT:    [[TMP3:%.*]] = call x86_amx @llvm.x86.tileloadd64.internal(i16 11, i16 40, i8* undef, i64 undef)
+; CHECK-NEXT:    br i1 undef, label [[FOR_COND_CLEANUP_I_I:%.*]], label [[FOR_BODY_I_LR_PH_I:%.*]]
+; CHECK:       for.body.i.lr.ph.i:
+; CHECK-NEXT:    [[TMP4:%.*]] = bitcast <110 x i32>* [[TMP2]] to i8*
+; CHECK-NEXT:    store <110 x i32> undef, <110 x i32>* [[TMP2]], align 512
+; CHECK-NEXT:    [[TMP5:%.*]] = call x86_amx @llvm.x86.tileloadd64.internal(i16 11, i16 40, i8* [[TMP4]], i64 40)
+; CHECK-NEXT:    [[TMP6:%.*]] = bitcast <616 x i8>* [[TMP1]] to i8*
+; CHECK-NEXT:    store <616 x i8> undef, <616 x i8>* [[TMP1]], align 1024
+; CHECK-NEXT:    [[TMP7:%.*]] = call x86_amx @llvm.x86.tileloadd64.internal(i16 11, i16 56, i8* [[TMP6]], i64 56)
+; CHECK-NEXT:    [[TMP8:%.*]] = bitcast <560 x i8>* [[TMP0]] to i8*
+; CHECK-NEXT:    store <560 x i8> undef, <560 x i8>* [[TMP0]], align 1024
+; CHECK-NEXT:    [[TMP9:%.*]] = call x86_amx @llvm.x86.tileloadd64.internal(i16 14, i16 40, i8* [[TMP8]], i64 40)
+; CHECK-NEXT:    [[TMP10:%.*]] = call x86_amx @llvm.x86.tdpbssd.internal(i16 11, i16 40, i16 56, x86_amx [[TMP5]], x86_amx [[TMP7]], x86_amx [[TMP9]])
+; CHECK-NEXT:    br label [[FOR_COND_CLEANUP_I_I]]
+; CHECK:       for.cond.cleanup.i.i:
+; CHECK-NEXT:    [[TMP11:%.*]] = phi x86_amx [ [[TMP3]], [[WRAPPER_ENTRY:%.*]] ], [ [[TMP10]], [[FOR_BODY_I_LR_PH_I]] ]
+; CHECK-NEXT:    call void @llvm.x86.tilestored64.internal(i16 11, i16 40, i8* undef, i64 undef, x86_amx [[TMP11]])
+; CHECK-NEXT:    ret void
+;
+wrapper_entry:
+  %0 = call x86_amx @llvm.x86.tileloadd64.internal(i16 11, i16 40, i8* undef, i64 undef)
+  %tmp = call <110 x i32> @llvm.x86.cast.tile.to.vector.v110i32(x86_amx %0)
+  br i1 undef, label %for.cond.cleanup.i.i, label %for.body.i.lr.ph.i
+
+for.body.i.lr.ph.i:                               ; preds = %wrapper_entry
+  %1 = call x86_amx @llvm.x86.cast.vector.to.tile.v110i32(<110 x i32> undef)
+  %2 = call x86_amx @llvm.x86.cast.vector.to.tile.v616i8(<616 x i8> undef)
+  %3 = call x86_amx @llvm.x86.cast.vector.to.tile.v560i8(<560 x i8> undef)
+  %4 = call x86_amx @llvm.x86.tdpbssd.internal(i16 11, i16 40, i16 56, x86_amx %1, x86_amx %2, x86_amx %3)
+  %5 = call <110 x i32> @llvm.x86.cast.tile.to.vector.v110i32(x86_amx %4)
+  br label %for.cond.cleanup.i.i
+
+for.cond.cleanup.i.i:                             ; preds = %for.body.i.lr.ph.i, %wrapper_entry
+  %goodphi = phi <110 x i32> [ %tmp, %wrapper_entry ], [ %5, %for.body.i.lr.ph.i ]
+  %6 = call x86_amx @llvm.x86.cast.vector.to.tile.v110i32(<110 x i32> %goodphi)
+  call void @llvm.x86.tilestored64.internal(i16 11, i16 40, i8* undef, i64 undef, x86_amx %6)
+  ret void
+}
+
+; Cases where amxcast can't be combined across bb
+; %5 and %6 is not combined together since %evilphi's incoming is not phi or amxcast
+define void @fail_to_combine_amx_cast_and_phi(<110 x i32> %tmp) {
+; CHECK-LABEL: @fail_to_combine_amx_cast_and_phi(
+; CHECK-NEXT:  wrapper_entry:
+; CHECK-NEXT:    [[TMP0:%.*]] = alloca <110 x i32>, align 64
+; CHECK-NEXT:    [[TMP1:%.*]] = alloca <110 x i32>, align 64
+; CHECK-NEXT:    [[TMP2:%.*]] = alloca <560 x i8>, align 64
+; CHECK-NEXT:    [[TMP3:%.*]] = alloca <616 x i8>, align 64
+; CHECK-NEXT:    [[TMP4:%.*]] = alloca <110 x i32>, align 64
+; CHECK-NEXT:    [[TMP5:%.*]] = add <110 x i32> [[TMP:%.*]], [[TMP]]
+; CHECK-NEXT:    br i1 undef, label [[FOR_COND_CLEANUP_I_I:%.*]], label [[FOR_BODY_I_LR_PH_I:%.*]]
+; CHECK:       for.body.i.lr.ph.i:
+; CHECK-NEXT:    [[TMP6:%.*]] = bitcast <110 x i32>* [[TMP4]] to i8*
+; CHECK-NEXT:    store <110 x i32> undef, <110 x i32>* [[TMP4]], align 512
+; CHECK-NEXT:    [[TMP7:%.*]] = call x86_amx @llvm.x86.tileloadd64.internal(i16 11, i16 40, i8* [[TMP6]], i64 40)
+; CHECK-NEXT:    [[TMP8:%.*]] = bitcast <616 x i8>* [[TMP3]] to i8*
+; CHECK-NEXT:    store <616 x i8> undef, <616 x i8>* [[TMP3]], align 1024
+; CHECK-NEXT:    [[TMP9:%.*]] = call x86_amx @llvm.x86.tileloadd64.internal(i16 11, i16 56, i8* [[TMP8]], i64 56)
+; CHECK-NEXT:    [[TMP10:%.*]] = bitcast <560 x i8>* [[TMP2]] to i8*
+; CHECK-NEXT:    store <560 x i8> undef, <560 x i8>* [[TMP2]], align 1024
+; CHECK-NEXT:    [[TMP11:%.*]] = call x86_amx @llvm.x86.tileloadd64.internal(i16 14, i16 40, i8* [[TMP10]], i64 40)
+; CHECK-NEXT:    [[TMP12:%.*]] = call x86_amx @llvm.x86.tdpbssd.internal(i16 11, i16 40, i16 56, x86_amx [[TMP7]], x86_amx [[TMP9]], x86_amx [[TMP11]])
+; CHECK-NEXT:    [[TMP13:%.*]] = bitcast <110 x i32>* [[TMP1]] to i8*
+; CHECK-NEXT:    call void @llvm.x86.tilestored64.internal(i16 11, i16 40, i8* [[TMP13]], i64 40, x86_amx [[TMP12]])
+; CHECK-NEXT:    [[TMP14:%.*]] = load <110 x i32>, <110 x i32>* [[TMP1]], align 512
+; CHECK-NEXT:    br label [[FOR_COND_CLEANUP_I_I]]
+; CHECK:       for.cond.cleanup.i.i:
+; CHECK-NEXT:    [[EVILPHI:%.*]] = phi <110 x i32> [ [[TMP5]], [[WRAPPER_ENTRY:%.*]] ], [ [[TMP14]], [[FOR_BODY_I_LR_PH_I]] ]
+; CHECK-NEXT:    [[TMP15:%.*]] = bitcast <110 x i32>* [[TMP0]] to i8*
+; CHECK-NEXT:    store <110 x i32> [[EVILPHI]], <110 x i32>* [[TMP0]], align 512
+; CHECK-NEXT:    [[TMP16:%.*]] = call x86_amx @llvm.x86.tileloadd64.internal(i16 11, i16 40, i8* [[TMP15]], i64 40)
+; CHECK-NEXT:    call void @llvm.x86.tilestored64.internal(i16 11, i16 40, i8* undef, i64 undef, x86_amx [[TMP16]])
+; CHECK-NEXT:    ret void
+;
+wrapper_entry:
+  %0 = add <110 x i32> %tmp, %tmp
+  br i1 undef, label %for.cond.cleanup.i.i, label %for.body.i.lr.ph.i
+
+for.body.i.lr.ph.i:                               ; preds = %wrapper_entry
+  %1 = call x86_amx @llvm.x86.cast.vector.to.tile.v110i32(<110 x i32> undef)
+  %2 = call x86_amx @llvm.x86.cast.vector.to.tile.v616i8(<616 x i8> undef)
+  %3 = call x86_amx @llvm.x86.cast.vector.to.tile.v560i8(<560 x i8> undef)
+  %4 = call x86_amx @llvm.x86.tdpbssd.internal(i16 11, i16 40, i16 56, x86_amx %1, x86_amx %2, x86_amx %3)
+  %5 = call <110 x i32> @llvm.x86.cast.tile.to.vector.v110i32(x86_amx %4)
+  br label %for.cond.cleanup.i.i
+
+for.cond.cleanup.i.i:                             ; preds = %for.body.i.lr.ph.i, %wrapper_entry
+  %evilphi = phi <110 x i32> [ %0, %wrapper_entry ], [ %5, %for.body.i.lr.ph.i ]
+  %6 = call x86_amx @llvm.x86.cast.vector.to.tile.v110i32(<110 x i32> %evilphi)
+  call void @llvm.x86.tilestored64.internal(i16 11, i16 40, i8* undef, i64 undef, x86_amx %6)
+  ret void
+}
+
+; Cases where amxcast can't be combined across bb
+; %5 and %6 is not combined together since %evilphi's user aka %evilphi2 is not inside phi web.
+define void @fail_to_combine_amx_cast_and_phi2() {
+; CHECK-LABEL: @fail_to_combine_amx_cast_and_phi2(
+; CHECK-NEXT:  wrapper_entry:
+; CHECK-NEXT:    [[TMP0:%.*]] = alloca <110 x i32>, align 64
+; CHECK-NEXT:    [[TMP1:%.*]] = alloca <110 x i32>, align 64
+; CHECK-NEXT:    [[TMP2:%.*]] = alloca <560 x i8>, align 64
+; CHECK-NEXT:    [[TMP3:%.*]] = alloca <616 x i8>, align 64
+; CHECK-NEXT:    [[TMP4:%.*]] = alloca <110 x i32>, align 64
+; CHECK-NEXT:    [[TMP5:%.*]] = alloca <110 x i32>, align 64
+; CHECK-NEXT:    [[TMP6:%.*]] = call x86_amx @llvm.x86.tileloadd64.internal(i16 11, i16 40, i8* undef, i64 undef)
+; CHECK-NEXT:    [[TMP7:%.*]] = bitcast <110 x i32>* [[TMP5]] to i8*
+; CHECK-NEXT:    call void @llvm.x86.tilestored64.internal(i16 11, i16 40, i8* [[TMP7]], i64 40, x86_amx [[TMP6]])
+; CHECK-NEXT:    [[TMP8:%.*]] = load <110 x i32>, <110 x i32>* [[TMP5]], align 512
+; CHECK-NEXT:    br i1 undef, label [[FOR_COND_CLEANUP_I_I:%.*]], label [[FOR_BODY_I_LR_PH_I:%.*]]
+; CHECK:       for.body.i.lr.ph.i:
+; CHECK-NEXT:    [[TMP9:%.*]] = bitcast <110 x i32>* [[TMP4]] to i8*
+; CHECK-NEXT:    store <110 x i32> undef, <110 x i32>* [[TMP4]], align 512
+; CHECK-NEXT:    [[TMP10:%.*]] = call x86_amx @llvm.x86.tileloadd64.internal(i16 11, i16 40, i8* [[TMP9]], i64 40)
+; CHECK-NEXT:    [[TMP11:%.*]] = bitcast <616 x i8>* [[TMP3]] to i8*
+; CHECK-NEXT:    store <616 x i8> undef, <616 x i8>* [[TMP3]], align 1024
+; CHECK-NEXT:    [[TMP12:%.*]] = call x86_amx @llvm.x86.tileloadd64.internal(i16 11, i16 56, i8* [[TMP11]], i64 56)
+; CHECK-NEXT:    [[TMP13:%.*]] = bitcast <560 x i8>* [[TMP2]] to i8*
+; CHECK-NEXT:    store <560 x i8> undef, <560 x i8>* [[TMP2]], align 1024
+; CHECK-NEXT:    [[TMP14:%.*]] = call x86_amx @llvm.x86.tileloadd64.internal(i16 14, i16 40, i8* [[TMP13]], i64 40)
+; CHECK-NEXT:    [[TMP15:%.*]] = call x86_amx @llvm.x86.tdpbssd.internal(i16 11, i16 40, i16 56, x86_amx [[TMP10]], x86_amx [[TMP12]], x86_amx [[TMP14]])
+; CHECK-NEXT:    [[TMP16:%.*]] = bitcast <110 x i32>* [[TMP1]] to i8*
+; CHECK-NEXT:    call void @llvm.x86.tilestored64.internal(i16 11, i16 40, i8* [[TMP16]], i64 40, x86_amx [[TMP15]])
+; CHECK-NEXT:    [[TMP17:%.*]] = load <110 x i32>, <110 x i32>* [[TMP1]], align 512
+; CHECK-NEXT:    br i1 undef, label [[FOR_COND_CLEANUP_I_I]], label [[EXIT:%.*]]
+; CHECK:       for.cond.cleanup.i.i:
+; CHECK-NEXT:    [[GOODPHI:%.*]] = phi <110 x i32> [ [[TMP8]], [[WRAPPER_ENTRY:%.*]] ], [ [[TMP17]], [[FOR_BODY_I_LR_PH_I]] ]
+; CHECK-NEXT:    [[TMP18:%.*]] = bitcast <110 x i32>* [[TMP0]] to i8*
+; CHECK-NEXT:    store <110 x i32> [[GOODPHI]], <110 x i32>* [[TMP0]], align 512
+; CHECK-NEXT:    [[TMP19:%.*]] = call x86_amx @llvm.x86.tileloadd64.internal(i16 11, i16 40, i8* [[TMP18]], i64 40)
+; CHECK-NEXT:    call void @llvm.x86.tilestored64.internal(i16 11, i16 40, i8* undef, i64 undef, x86_amx [[TMP19]])
+; CHECK-NEXT:    br i1 undef, label [[EXIT]], label [[FOR_BODY_I_LR_PH_I]]
+; CHECK:       exit:
+; CHECK-NEXT:    [[EVILPHI2:%.*]] = phi <110 x i32> [ [[GOODPHI]], [[FOR_COND_CLEANUP_I_I]] ], [ [[TMP17]], [[FOR_BODY_I_LR_PH_I]] ]
+; CHECK-NEXT:    store <110 x i32> [[EVILPHI2]], <110 x i32>* undef, align 512
+; CHECK-NEXT:    ret void
+;
+wrapper_entry:
+  %0 = call x86_amx @llvm.x86.tileloadd64.internal(i16 11, i16 40, i8* undef, i64 undef)
+  %tmp = call <110 x i32> @llvm.x86.cast.tile.to.vector.v110i32(x86_amx %0)
+  br i1 undef, label %for.cond.cleanup.i.i, label %for.body.i.lr.ph.i
+
+for.body.i.lr.ph.i:                               ; preds = %wrapper_entry
+  %1 = call x86_amx @llvm.x86.cast.vector.to.tile.v110i32(<110 x i32> undef)
+  %2 = call x86_amx @llvm.x86.cast.vector.to.tile.v616i8(<616 x i8> undef)
+  %3 = call x86_amx @llvm.x86.cast.vector.to.tile.v560i8(<560 x i8> undef)
+  %4 = call x86_amx @llvm.x86.tdpbssd.internal(i16 11, i16 40, i16 56, x86_amx %1, x86_amx %2, x86_amx %3)
+  %5 = call <110 x i32> @llvm.x86.cast.tile.to.vector.v110i32(x86_amx %4)
+  br i1 undef, label %for.cond.cleanup.i.i, label %exit
+
+for.cond.cleanup.i.i:                             ; preds = %for.body.i.lr.ph.i, %wrapper_entry
+  %goodphi = phi <110 x i32> [ %tmp, %wrapper_entry ], [ %5, %for.body.i.lr.ph.i ]
+  %6 = call x86_amx @llvm.x86.cast.vector.to.tile.v110i32(<110 x i32> %goodphi)
+  call void @llvm.x86.tilestored64.internal(i16 11, i16 40, i8* undef, i64 undef, x86_amx %6)
+  br i1 undef, label %exit, label %for.body.i.lr.ph.i
+exit:
+  %evilphi2 = phi <110 x i32> [ %goodphi, %for.cond.cleanup.i.i ], [ %5, %for.body.i.lr.ph.i ]
+  store <110 x i32> %evilphi2, <110 x i32>* undef, align 512
+  ret void
+}
+
+define void @fail_to_combine_amx_cast_and_phi_due_to_const_value() {
+; CHECK-LABEL: @fail_to_combine_amx_cast_and_phi_due_to_const_value(
+; CHECK-NEXT:  wrapper_entry:
+; CHECK-NEXT:    [[TMP0:%.*]] = alloca <110 x i32>, align 64
+; CHECK-NEXT:    [[TMP1:%.*]] = alloca <110 x i32>, align 64
+; CHECK-NEXT:    [[TMP2:%.*]] = alloca <560 x i8>, align 64
+; CHECK-NEXT:    [[TMP3:%.*]] = alloca <616 x i8>, align 64
+; CHECK-NEXT:    [[TMP4:%.*]] = alloca <110 x i32>, align 64
+; CHECK-NEXT:    br i1 undef, label [[FOR_COND_CLEANUP_I_I:%.*]], label [[FOR_BODY_I_LR_PH_I:%.*]]
+; CHECK:       for.body.i.lr.ph.i:
+; CHECK-NEXT:    [[TMP5:%.*]] = bitcast <110 x i32>* [[TMP4]] to i8*
+; CHECK-NEXT:    store <110 x i32> undef, <110 x i32>* [[TMP4]], align 512
+; CHECK-NEXT:    [[TMP6:%.*]] = call x86_amx @llvm.x86.tileloadd64.internal(i16 11, i16 40, i8* [[TMP5]], i64 40)
+; CHECK-NEXT:    [[TMP7:%.*]] = bitcast <616 x i8>* [[TMP3]] to i8*
+; CHECK-NEXT:    store <616 x i8> undef, <616 x i8>* [[TMP3]], align 1024
+; CHECK-NEXT:    [[TMP8:%.*]] = call x86_amx @llvm.x86.tileloadd64.internal(i16 11, i16 56, i8* [[TMP7]], i64 56)
+; CHECK-NEXT:    [[TMP9:%.*]] = bitcast <560 x i8>* [[TMP2]] to i8*
+; CHECK-NEXT:    store <560 x i8> undef, <560 x i8>* [[TMP2]], align 1024
+; CHECK-NEXT:    [[TMP10:%.*]] = call x86_amx @llvm.x86.tileloadd64.internal(i16 14, i16 40, i8* [[TMP9]], i64 40)
+; CHECK-NEXT:    [[TMP11:%.*]] = call x86_amx @llvm.x86.tdpbssd.internal(i16 11, i16 40, i16 56, x86_amx [[TMP6]], x86_amx [[TMP8]], x86_amx [[TMP10]])
+; CHECK-NEXT:    [[TMP12:%.*]] = bitcast <110 x i32>* [[TMP1]] to i8*
+; CHECK-NEXT:    call void @llvm.x86.tilestored64.internal(i16 11, i16 40, i8* [[TMP12]], i64 40, x86_amx [[TMP11]])
+; CHECK-NEXT:    [[TMP13:%.*]] = load <110 x i32>, <110 x i32>* [[TMP1]], align 512
+; CHECK-NEXT:    br label [[FOR_COND_CLEANUP_I_I]]
+; CHECK:       for.cond.cleanup.i.i:
+; CHECK-NEXT:    [[EVILPHI:%.*]] = phi <110 x i32> [ undef, [[WRAPPER_ENTRY:%.*]] ], [ [[TMP13]], [[FOR_BODY_I_LR_PH_I]] ]
+; CHECK-NEXT:    [[TMP14:%.*]] = bitcast <110 x i32>* [[TMP0]] to i8*
+; CHECK-NEXT:    store <110 x i32> [[EVILPHI]], <110 x i32>* [[TMP0]], align 512
+; CHECK-NEXT:    [[TMP15:%.*]] = call x86_amx @llvm.x86.tileloadd64.internal(i16 11, i16 40, i8* [[TMP14]], i64 40)
+; CHECK-NEXT:    call void @llvm.x86.tilestored64.internal(i16 11, i16 40, i8* undef, i64 undef, x86_amx [[TMP15]])
+; CHECK-NEXT:    ret void
+;
+wrapper_entry:
+  br i1 undef, label %for.cond.cleanup.i.i, label %for.body.i.lr.ph.i
+
+for.body.i.lr.ph.i:                               ; preds = %wrapper_entry
+  %0 = call x86_amx @llvm.x86.cast.vector.to.tile.v110i32(<110 x i32> undef)
+  %1 = call x86_amx @llvm.x86.cast.vector.to.tile.v616i8(<616 x i8> undef)
+  %2 = call x86_amx @llvm.x86.cast.vector.to.tile.v560i8(<560 x i8> undef)
+  %3 = call x86_amx @llvm.x86.tdpbssd.internal(i16 11, i16 40, i16 56, x86_amx %0, x86_amx %1, x86_amx %2)
+  %4 = call <110 x i32> @llvm.x86.cast.tile.to.vector.v110i32(x86_amx %3)
+  br label %for.cond.cleanup.i.i
+
+for.cond.cleanup.i.i:                             ; preds = %for.body.i.lr.ph.i, %wrapper_entry
+  %evilphi = phi <110 x i32> [ undef, %wrapper_entry ], [ %4, %for.body.i.lr.ph.i ]
+  %5 = call x86_amx @llvm.x86.cast.vector.to.tile.v110i32(<110 x i32> %evilphi)
+  call void @llvm.x86.tilestored64.internal(i16 11, i16 40, i8* undef, i64 undef, x86_amx %5)
+  ret void
+}
+
+; Cases where amxcast can be combined across bb
+; When optimizeAMXCastFromPhi process %6 and %goodphi, %goodphi2 is outside the phi-web, so the optimization stop
+; When optimizeAMXCastFromPhi process %7 and %goodphi2, the optimization continue.
+define void @combine_amx_cast_and_multiple_phi() {
+; CHECK-LABEL: @combine_amx_cast_and_multiple_phi(
+; CHECK-NEXT:  wrapper_entry:
+; CHECK-NEXT:    [[TMP0:%.*]] = alloca <560 x i8>, align 64
+; CHECK-NEXT:    [[TMP1:%.*]] = alloca <616 x i8>, align 64
+; CHECK-NEXT:    [[TMP2:%.*]] = alloca <110 x i32>, align 64
+; CHECK-NEXT:    [[TMP3:%.*]] = call x86_amx @llvm.x86.tileloadd64.internal(i16 11, i16 40, i8* undef, i64 undef)
+; CHECK-NEXT:    br i1 undef, label [[FOR_COND_CLEANUP_I_I:%.*]], label [[FOR_BODY_I_LR_PH_I:%.*]]
+; CHECK:       for.body.i.lr.ph.i:
+; CHECK-NEXT:    [[TMP4:%.*]] = bitcast <110 x i32>* [[TMP2]] to i8*
+; CHECK-NEXT:    store <110 x i32> undef, <110 x i32>* [[TMP2]], align 512
+; CHECK-NEXT:    [[TMP5:%.*]] = call x86_amx @llvm.x86.tileloadd64.internal(i16 11, i16 40, i8* [[TMP4]], i64 40)
+; CHECK-NEXT:    [[TMP6:%.*]] = bitcast <616 x i8>* [[TMP1]] to i8*
+; CHECK-NEXT:    store <616 x i8> undef, <616 x i8>* [[TMP1]], align 1024
+; CHECK-NEXT:    [[TMP7:%.*]] = call x86_amx @llvm.x86.tileloadd64.internal(i16 11, i16 56, i8* [[TMP6]], i64 56)
+; CHECK-NEXT:    [[TMP8:%.*]] = bitcast <560 x i8>* [[TMP0]] to i8*
+; CHECK-NEXT:    store <560 x i8> undef, <560 x i8>* [[TMP0]], align 1024
+; CHECK-NEXT:    [[TMP9:%.*]] = call x86_amx @llvm.x86.tileloadd64.internal(i16 14, i16 40, i8* [[TMP8]], i64 40)
+; CHECK-NEXT:    [[TMP10:%.*]] = call x86_amx @llvm.x86.tdpbssd.internal(i16 11, i16 40, i16 56, x86_amx [[TMP5]], x86_amx [[TMP7]], x86_amx [[TMP9]])
+; CHECK-NEXT:    br i1 undef, label [[FOR_COND_CLEANUP_I_I]], label [[EXIT:%.*]]
+; CHECK:       for.cond.cleanup.i.i:
+; CHECK-NEXT:    [[TMP11:%.*]] = phi x86_amx [ [[TMP3]], [[WRAPPER_ENTRY:%.*]] ], [ [[TMP10]], [[FOR_BODY_I_LR_PH_I]] ]
+; CHECK-NEXT:    call void @llvm.x86.tilestored64.internal(i16 11, i16 40, i8* undef, i64 undef, x86_amx [[TMP11]])
+; CHECK-NEXT:    br i1 undef, label [[EXIT]], label [[FOR_BODY_I_LR_PH_I]]
+; CHECK:       exit:
+; CHECK-NEXT:    [[TMP12:%.*]] = phi x86_amx [ [[TMP11]], [[FOR_COND_CLEANUP_I_I]] ], [ [[TMP10]], [[FOR_BODY_I_LR_PH_I]] ]
+; CHECK-NEXT:    call void @llvm.x86.tilestored64.internal(i16 11, i16 40, i8* undef, i64 undef, x86_amx [[TMP12]])
+; CHECK-NEXT:    ret void
+;
+wrapper_entry:
+  %0 = call x86_amx @llvm.x86.tileloadd64.internal(i16 11, i16 40, i8* undef, i64 undef)
+  %tmp = call <110 x i32> @llvm.x86.cast.tile.to.vector.v110i32(x86_amx %0)
+  br i1 undef, label %for.cond.cleanup.i.i, label %for.body.i.lr.ph.i
+
+for.body.i.lr.ph.i:                               ; preds = %wrapper_entry
+  %1 = call x86_amx @llvm.x86.cast.vector.to.tile.v110i32(<110 x i32> undef)
+  %2 = call x86_amx @llvm.x86.cast.vector.to.tile.v616i8(<616 x i8> undef)
+  %3 = call x86_amx @llvm.x86.cast.vector.to.tile.v560i8(<560 x i8> undef)
+  %4 = call x86_amx @llvm.x86.tdpbssd.internal(i16 11, i16 40, i16 56, x86_amx %1, x86_amx %2, x86_amx %3)
+  %5 = call <110 x i32> @llvm.x86.cast.tile.to.vector.v110i32(x86_amx %4)
+  br i1 undef, label %for.cond.cleanup.i.i, label %exit
+
+for.cond.cleanup.i.i:                             ; preds = %for.body.i.lr.ph.i, %wrapper_entry
+  %goodphi = phi <110 x i32> [ %tmp, %wrapper_entry ], [ %5, %for.body.i.lr.ph.i ]
+  %6 = call x86_amx @llvm.x86.cast.vector.to.tile.v110i32(<110 x i32> %goodphi)
+  call void @llvm.x86.tilestored64.internal(i16 11, i16 40, i8* undef, i64 undef, x86_amx %6)
+  br i1 undef, label %exit, label %for.body.i.lr.ph.i
+exit:
+  %evilphi2 = phi <110 x i32> [ %goodphi, %for.cond.cleanup.i.i ], [ %5, %for.body.i.lr.ph.i ]
+  %7 = call x86_amx @llvm.x86.cast.vector.to.tile.v110i32(<110 x i32> %evilphi2)
+  call void @llvm.x86.tilestored64.internal(i16 11, i16 40, i8* undef, i64 undef, x86_amx %7)
+  ret void
+}
+
+; Currently we are not able to delete DeadPHICycle, later we will handle with them
+define void @combine_amx_cast_and_phi_in_a_circle() {
+; CHECK-LABEL: @combine_amx_cast_and_phi_in_a_circle(
+; CHECK-NEXT:  wrapper_entry:
+; CHECK-NEXT:    [[TMP0:%.*]] = alloca <110 x i32>, align 64
+; CHECK-NEXT:    [[TMP1:%.*]] = alloca <560 x i8>, align 64
+; CHECK-NEXT:    [[TMP2:%.*]] = alloca <616 x i8>, align 64
+; CHECK-NEXT:    [[TMP3:%.*]] = alloca <110 x i32>, align 64
+; CHECK-NEXT:    [[TMP4:%.*]] = call x86_amx @llvm.x86.tileloadd64.internal(i16 11, i16 40, i8* undef, i64 undef)
+; CHECK-NEXT:    br label [[BB1:%.*]]
+; CHECK:       bb1:
+; CHECK-NEXT:    [[TMP5:%.*]] = bitcast <110 x i32>* [[TMP3]] to i8*
+; CHECK-NEXT:    store <110 x i32> undef, <110 x i32>* [[TMP3]], align 512
+; CHECK-NEXT:    [[TMP6:%.*]] = call x86_amx @llvm.x86.tileloadd64.internal(i16 11, i16 40, i8* [[TMP5]], i64 40)
+; CHECK-NEXT:    [[TMP7:%.*]] = bitcast <616 x i8>* [[TMP2]] to i8*
+; CHECK-NEXT:    store <616 x i8> undef, <616 x i8>* [[TMP2]], align 1024
+; CHECK-NEXT:    [[TMP8:%.*]] = call x86_amx @llvm.x86.tileloadd64.internal(i16 11, i16 56, i8* [[TMP7]], i64 56)
+; CHECK-NEXT:    [[TMP9:%.*]] = bitcast <560 x i8>* [[TMP1]] to i8*
+; CHECK-NEXT:    store <560 x i8> undef, <560 x i8>* [[TMP1]], align 1024
+; CHECK-NEXT:    [[TMP10:%.*]] = call x86_amx @llvm.x86.tileloadd64.internal(i16 14, i16 40, i8* [[TMP9]], i64 40)
+; CHECK-NEXT:    [[TMP11:%.*]] = call x86_amx @llvm.x86.tdpbssd.internal(i16 11, i16 40, i16 56, x86_amx [[TMP6]], x86_amx [[TMP8]], x86_amx [[TMP10]])
+; CHECK-NEXT:    [[TMP12:%.*]] = bitcast <110 x i32>* [[TMP0]] to i8*
+; CHECK-NEXT:    call void @llvm.x86.tilestored64.internal(i16 11, i16 40, i8* [[TMP12]], i64 40, x86_amx [[TMP11]])
+; CHECK-NEXT:    [[TMP13:%.*]] = load <110 x i32>, <110 x i32>* [[TMP0]], align 512
+; CHECK-NEXT:    br i1 undef, label [[BB2:%.*]], label [[BB3:%.*]]
+; CHECK:       bb2:
+; CHECK-NEXT:    [[TMP14:%.*]] = phi x86_amx [ [[TMP15:%.*]], [[BB3]] ], [ [[TMP11]], [[BB1]] ]
+; CHECK-NEXT:    [[GOODPHI:%.*]] = phi <110 x i32> [ [[EVILPHI2:%.*]], [[BB3]] ], [ [[TMP13]], [[BB1]] ]
+; CHECK-NEXT:    call void @llvm.x86.tilestored64.internal(i16 11, i16 40, i8* undef, i64 undef, x86_amx [[TMP14]])
+; CHECK-NEXT:    br label [[BB3]]
+; CHECK:       bb3:
+; CHECK-NEXT:    [[TMP15]] = phi x86_amx [ [[TMP14]], [[BB2]] ], [ [[TMP11]], [[BB1]] ]
+; CHECK-NEXT:    [[EVILPHI2]] = phi <110 x i32> [ [[GOODPHI]], [[BB2]] ], [ [[TMP13]], [[BB1]] ]
+; CHECK-NEXT:    call void @llvm.x86.tilestored64.internal(i16 11, i16 40, i8* undef, i64 undef, x86_amx [[TMP15]])
+; CHECK-NEXT:    br i1 undef, label [[BB2]], label [[EXIT:%.*]]
+; CHECK:       exit:
+; CHECK-NEXT:    call void @llvm.x86.tilestored64.internal(i16 11, i16 40, i8* undef, i64 undef, x86_amx [[TMP15]])
+; CHECK-NEXT:    ret void
+;
+wrapper_entry:
+  %0 = call x86_amx @llvm.x86.tileloadd64.internal(i16 11, i16 40, i8* undef, i64 undef)
+  %tmp = call <110 x i32> @llvm.x86.cast.tile.to.vector.v110i32(x86_amx %0)
+  br label %bb1
+
+bb1:                               ; preds = %wrapper_entry
+  %1 = call x86_amx @llvm.x86.cast.vector.to.tile.v110i32(<110 x i32> undef)
+  %2 = call x86_amx @llvm.x86.cast.vector.to.tile.v616i8(<616 x i8> undef)
+  %3 = call x86_amx @llvm.x86.cast.vector.to.tile.v560i8(<560 x i8> undef)
+  %4 = call x86_amx @llvm.x86.tdpbssd.internal(i16 11, i16 40, i16 56, x86_amx %1, x86_amx %2, x86_amx %3)
+  %5 = call <110 x i32> @llvm.x86.cast.tile.to.vector.v110i32(x86_amx %4)
+  br i1 undef, label %bb2, label %bb3
+
+bb2:                             ; preds = %bb1, %wrapper_entry
+  %goodphi = phi <110 x i32> [ %evilphi2, %bb3], [ %5, %bb1 ]
+  %6 = call x86_amx @llvm.x86.cast.vector.to.tile.v110i32(<110 x i32> %goodphi)
+  call void @llvm.x86.tilestored64.internal(i16 11, i16 40, i8* undef, i64 undef, x86_amx %6)
+  br label %bb3
+bb3:
+  %evilphi2 = phi <110 x i32> [ %goodphi, %bb2 ], [ %5, %bb1 ]
+  %7 = call x86_amx @llvm.x86.cast.vector.to.tile.v110i32(<110 x i32> %evilphi2)
+  call void @llvm.x86.tilestored64.internal(i16 11, i16 40, i8* undef, i64 undef, x86_amx %7)
+  br i1 undef, label %bb2, label %exit
+exit:
+  %8 = call x86_amx @llvm.x86.cast.vector.to.tile.v110i32(<110 x i32> %evilphi2)
+  call void @llvm.x86.tilestored64.internal(i16 11, i16 40, i8* undef, i64 undef, x86_amx %8)
+  ret void
+}
+
+define void @eliminate_unused_phi_and_cast() {
+; CHECK-LABEL: @eliminate_unused_phi_and_cast(
+; CHECK-NEXT:  wrapper_entry:
+; CHECK-NEXT:    [[TMP0:%.*]] = alloca <560 x i8>, align 64
+; CHECK-NEXT:    [[TMP1:%.*]] = call x86_amx @llvm.x86.tileloadd64.internal(i16 11, i16 40, i8* undef, i64 undef)
+; CHECK-NEXT:    br i1 undef, label [[FOR_COND_CLEANUP_I_I:%.*]], label [[FOR_BODY_I_LR_PH_I:%.*]]
+; CHECK:       for.body.i.lr.ph.i:
+; CHECK-NEXT:    [[TMP2:%.*]] = call x86_amx @llvm.x86.tileloadd64.internal(i16 11, i16 56, i8* undef, i64 undef)
+; CHECK-NEXT:    [[TMP3:%.*]] = call x86_amx @llvm.x86.tileloadd64.internal(i16 14, i16 40, i8* undef, i64 undef)
+; CHECK-NEXT:    [[TMP4:%.*]] = bitcast <560 x i8>* [[TMP0]] to i8*
+; CHECK-NEXT:    store <560 x i8> undef, <560 x i8>* [[TMP0]], align 1024
+; CHECK-NEXT:    [[TMP5:%.*]] = call x86_amx @llvm.x86.tileloadd64.internal(i16 14, i16 40, i8* [[TMP4]], i64 40)
+; CHECK-NEXT:    [[TMP6:%.*]] = call x86_amx @llvm.x86.tdpbssd.internal(i16 11, i16 40, i16 56, x86_amx [[TMP2]], x86_amx [[TMP3]], x86_amx [[TMP5]])
+; CHECK-NEXT:    br label [[FOR_COND_CLEANUP_I_I]]
+; CHECK:       for.cond.cleanup.i.i:
+; CHECK-NEXT:    [[TMP7:%.*]] = phi x86_amx [ [[TMP1]], [[WRAPPER_ENTRY:%.*]] ], [ [[TMP6]], [[FOR_BODY_I_LR_PH_I]] ]
+; CHECK-NEXT:    call void @llvm.x86.tilestored64.internal(i16 11, i16 40, i8* undef, i64 undef, x86_amx [[TMP7]])
+; CHECK-NEXT:    ret void
+;
+wrapper_entry:
+  %0 = call x86_amx @llvm.x86.tileloadd64.internal(i16 11, i16 40, i8* undef, i64 undef)
+  %tmp = call <110 x i32> @llvm.x86.cast.tile.to.vector.v110i32(x86_amx %0)
+  br i1 undef, label %for.cond.cleanup.i.i, label %for.body.i.lr.ph.i
+
+for.body.i.lr.ph.i:                               ; preds = %wrapper_entry
+  %1 = call x86_amx @llvm.x86.tileloadd64.internal(i16 11, i16 56, i8* undef, i64 undef)
+  %v1 = call <110 x i32> @llvm.x86.cast.tile.to.vector.v110i32(x86_amx %1)
+  %2 = call x86_amx @llvm.x86.tileloadd64.internal(i16 14, i16 40, i8* undef, i64 undef)
+  %v2 = call <616 x i8> @llvm.x86.cast.tile.to.vector.v616i8(x86_amx %2)
+  %3 = call x86_amx @llvm.x86.cast.vector.to.tile.v110i32(<110 x i32> %v1)
+  %4 = call x86_amx @llvm.x86.cast.vector.to.tile.v616i8(<616 x i8> %v2)
+  %5 = call x86_amx @llvm.x86.cast.vector.to.tile.v560i8(<560 x i8> undef)
+  %6 = call x86_amx @llvm.x86.tdpbssd.internal(i16 11, i16 40, i16 56, x86_amx %3, x86_amx %4, x86_amx %5)
+  %7 = call <110 x i32> @llvm.x86.cast.tile.to.vector.v110i32(x86_amx %6)
+  br label %for.cond.cleanup.i.i
+
+for.cond.cleanup.i.i:                             ; preds = %for.body.i.lr.ph.i, %wrapper_entry
+  %goodphi = phi <110 x i32> [ %tmp, %wrapper_entry ], [ %7, %for.body.i.lr.ph.i ]
+  %8 = call x86_amx @llvm.x86.cast.vector.to.tile.v110i32(<110 x i32> %goodphi)
+  call void @llvm.x86.tilestored64.internal(i16 11, i16 40, i8* undef, i64 undef, x86_amx %8)
+  ret void
+}
+
+declare x86_amx @llvm.x86.tileloadd64.internal(i16, i16, i8*, i64)
+declare <110 x i32> @llvm.x86.cast.tile.to.vector.v110i32(x86_amx)
+declare <616 x i8> @llvm.x86.cast.tile.to.vector.v616i8(x86_amx)
+declare x86_amx @llvm.x86.cast.vector.to.tile.v110i32(<110 x i32>)
+declare void @llvm.x86.tilestored64.internal(i16, i16, i8*, i64, x86_amx)
+declare x86_amx @llvm.x86.cast.vector.to.tile.v616i8(<616 x i8>)
+declare x86_amx @llvm.x86.cast.vector.to.tile.v560i8(<560 x i8>)
+declare x86_amx @llvm.x86.tdpbssd.internal(i16, i16, i16, x86_amx, x86_amx, x86_amx)

llvm/test/CodeGen/X86/AMX/lat-transform-amx-bitcast.ll

+; NOTE: Assertions have been autogenerated by utils/update_test_checks.py
+; RUN: opt --codegen-opt-level=2 -mtriple=x86_64 -lower-amx-type %s -S | FileCheck %s
+
+%struct.__tile_str = type { i16, i16, <256 x i32> }
+
+@buf = dso_local global [1024 x i8] zeroinitializer, align 64
+@buf2 = dso_local global [1024 x i8] zeroinitializer, align 64
+
+; test bitcast x86_amx to <256 x i32>
+define dso_local void @test_user_empty(i16 %m, i16 %n, i8 *%buf, i64 %s) {
+; CHECK-LABEL: @test_user_empty(
+; CHECK-NEXT:  entry:
+; CHECK-NEXT:    [[T1:%.*]] = call x86_amx @llvm.x86.tileloadd64.internal(i16 [[M:%.*]], i16 [[N:%.*]], i8* [[BUF:%.*]], i64 [[S:%.*]])
+; CHECK-NEXT:    ret void
+;
+entry:
+  %t1 = call x86_amx @llvm.x86.tileloadd64.internal(i16 %m, i16 %n, i8* %buf, i64 %s)
+  %t2 = call <256 x i32> @llvm.x86.cast.tile.to.vector.v256i32(x86_amx %t1)
+  ret void
+}
+
+; test bitcast <256 x i32> to x86_amx
+define dso_local void @test_user_empty2(<256 x i32> %in) {
+; CHECK-LABEL: @test_user_empty2(
+; CHECK-NEXT:  entry:
+; CHECK-NEXT:    ret void
+;
+entry:
+  %t = call x86_amx @llvm.x86.cast.vector.to.tile.v256i32(<256 x i32> %in)
+  ret void
+}
+
+define dso_local <256 x i32> @test_amx_load_bitcast_v256i32(<256 x i32>* %in, i16 %m, i16 %n, i8 *%buf, i64 %s) {
+; CHECK-LABEL: @test_amx_load_bitcast_v256i32(
+; CHECK-NEXT:  entry:
+; CHECK-NEXT:    [[TMP0:%.*]] = alloca <256 x i32>, align 64
+; CHECK-NEXT:    [[T1:%.*]] = load <256 x i32>, <256 x i32>* [[IN:%.*]], align 64
+; CHECK-NEXT:    [[TMP1:%.*]] = bitcast <256 x i32>* [[TMP0]] to i8*
+; CHECK-NEXT:    store <256 x i32> [[T1]], <256 x i32>* [[TMP0]], align 1024
+; CHECK-NEXT:    [[TMP2:%.*]] = sext i16 [[N:%.*]] to i64
+; CHECK-NEXT:    [[TMP3:%.*]] = call x86_amx @llvm.x86.tileloadd64.internal(i16 [[M:%.*]], i16 [[N]], i8* [[TMP1]], i64 [[TMP2]])
+; CHECK-NEXT:    call void @llvm.x86.tilestored64.internal(i16 [[M]], i16 [[N]], i8* [[BUF:%.*]], i64 [[S:%.*]], x86_amx [[TMP3]])
+; CHECK-NEXT:    ret <256 x i32> [[T1]]
+;
+entry:
+  %t1 = load <256 x i32>, <256 x i32>* %in, align 64
+  %t2 = call x86_amx @llvm.x86.cast.vector.to.tile.v256i32(<256 x i32> %t1)
+  call void @llvm.x86.tilestored64.internal(i16 %m, i16 %n, i8* %buf, i64 %s, x86_amx %t2)
+  ret <256 x i32> %t1
+}
+
+define dso_local <225 x i32> @test_amx_load_bitcast_v225i32(<225 x i32>* %in, i16 %m, i16 %n, i8 *%buf, i64 %s) {
+; CHECK-LABEL: @test_amx_load_bitcast_v225i32(
+; CHECK-NEXT:  entry:
+; CHECK-NEXT:    [[TMP0:%.*]] = alloca <225 x i32>, align 64
+; CHECK-NEXT:    [[T1:%.*]] = load <225 x i32>, <225 x i32>* [[IN:%.*]], align 64
+; CHECK-NEXT:    [[TMP1:%.*]] = bitcast <225 x i32>* [[TMP0]] to i8*
+; CHECK-NEXT:    store <225 x i32> [[T1]], <225 x i32>* [[TMP0]], align 1024
+; CHECK-NEXT:    [[TMP2:%.*]] = sext i16 [[N:%.*]] to i64
+; CHECK-NEXT:    [[TMP3:%.*]] = call x86_amx @llvm.x86.tileloadd64.internal(i16 [[M:%.*]], i16 [[N]], i8* [[TMP1]], i64 [[TMP2]])
+; CHECK-NEXT:    call void @llvm.x86.tilestored64.internal(i16 [[M]], i16 [[N]], i8* [[BUF:%.*]], i64 [[S:%.*]], x86_amx [[TMP3]])
+; CHECK-NEXT:    ret <225 x i32> [[T1]]
+;
+entry:
+  %t1 = load <225 x i32>, <225 x i32>* %in, align 64
+  %t2 = call x86_amx @llvm.x86.cast.vector.to.tile.v225i32(<225 x i32> %t1)
+  call void @llvm.x86.tilestored64.internal(i16 %m, i16 %n, i8* %buf, i64 %s, x86_amx %t2)
+  ret <225 x i32> %t1
+}
+
+define dso_local <256 x i32> @test_amx_bitcast_store(<256 x i32>* %out, i16 %m, i16 %n, i8 *%buf, i64 %s) {
+; CHECK-LABEL: @test_amx_bitcast_store(
+; CHECK-NEXT:  entry:
+; CHECK-NEXT:    [[TMP0:%.*]] = alloca <256 x i32>, align 64
+; CHECK-NEXT:    [[T1:%.*]] = call x86_amx @llvm.x86.tileloadd64.internal(i16 [[M:%.*]], i16 [[M]], i8* [[BUF:%.*]], i64 [[S:%.*]])
+; CHECK-NEXT:    [[TMP1:%.*]] = bitcast <256 x i32>* [[TMP0]] to i8*
+; CHECK-NEXT:    [[TMP2:%.*]] = sext i16 [[M]] to i64
+; CHECK-NEXT:    call void @llvm.x86.tilestored64.internal(i16 [[M]], i16 [[M]], i8* [[TMP1]], i64 [[TMP2]], x86_amx [[T1]])
+; CHECK-NEXT:    [[TMP3:%.*]] = load <256 x i32>, <256 x i32>* [[TMP0]], align 1024
+; CHECK-NEXT:    store <256 x i32> [[TMP3]], <256 x i32>* [[OUT:%.*]], align 1024
+; CHECK-NEXT:    ret <256 x i32> [[TMP3]]
+;
+entry:
+  %t1 = call x86_amx @llvm.x86.tileloadd64.internal(i16 %m, i16 %m, i8* %buf, i64 %s)
+  %t2 = call <256 x i32> @llvm.x86.cast.tile.to.vector.v256i32(x86_amx %t1)
+  store <256 x i32> %t2, <256 x i32>* %out
+  ret <256 x i32> %t2
+}
+
+define dso_local void @test_src_add(<256 x i32> %x, <256 x i32> %y, i16 %r, i16 %c, i8* %buf, i64 %s) {
+; CHECK-LABEL: @test_src_add(
+; CHECK-NEXT:  entry:
+; CHECK-NEXT:    [[TMP0:%.*]] = alloca <256 x i32>, align 64
+; CHECK-NEXT:    [[ADD:%.*]] = add <256 x i32> [[Y:%.*]], [[X:%.*]]
+; CHECK-NEXT:    [[TMP1:%.*]] = bitcast <256 x i32>* [[TMP0]] to i8*
+; CHECK-NEXT:    store <256 x i32> [[ADD]], <256 x i32>* [[TMP0]], align 1024
+; CHECK-NEXT:    [[TMP2:%.*]] = sext i16 [[C:%.*]] to i64
+; CHECK-NEXT:    [[TMP3:%.*]] = call x86_amx @llvm.x86.tileloadd64.internal(i16 [[R:%.*]], i16 [[C]], i8* [[TMP1]], i64 [[TMP2]])
+; CHECK-NEXT:    call void @llvm.x86.tilestored64.internal(i16 [[R]], i16 [[C]], i8* [[BUF:%.*]], i64 [[S:%.*]], x86_amx [[TMP3]])
+; CHECK-NEXT:    ret void
+;
+entry:
+  %add = add <256 x i32> %y, %x
+  %t = call x86_amx @llvm.x86.cast.vector.to.tile.v256i32(<256 x i32> %add)
+  call void @llvm.x86.tilestored64.internal(i16 %r, i16 %c, i8* %buf, i64 %s, x86_amx %t)
+  ret void
+}
+
+define dso_local void @test_src_add2(<256 x i32> %x, i16 %r, i16 %c, i8* %buf, i64 %s) {
+; CHECK-LABEL: @test_src_add2(
+; CHECK-NEXT:  entry:
+; CHECK-NEXT:    [[TMP0:%.*]] = alloca <256 x i32>, align 64
+; CHECK-NEXT:    [[T1:%.*]] = call x86_amx @llvm.x86.tileloadd64.internal(i16 [[R:%.*]], i16 [[C:%.*]], i8* [[BUF:%.*]], i64 [[S:%.*]])
+; CHECK-NEXT:    [[TMP1:%.*]] = bitcast <256 x i32>* [[TMP0]] to i8*
+; CHECK-NEXT:    [[TMP2:%.*]] = sext i16 [[C]] to i64
+; CHECK-NEXT:    call void @llvm.x86.tilestored64.internal(i16 [[R]], i16 [[C]], i8* [[TMP1]], i64 [[TMP2]], x86_amx [[T1]])
+; CHECK-NEXT:    [[TMP3:%.*]] = load <256 x i32>, <256 x i32>* [[TMP0]], align 1024
+; CHECK-NEXT:    [[ADD:%.*]] = add <256 x i32> [[TMP3]], [[X:%.*]]
+; CHECK-NEXT:    ret void
+;
+entry:
+  %t1 = call x86_amx @llvm.x86.tileloadd64.internal(i16 %r, i16 %c, i8* %buf, i64 %s)
+  %t2 = call <256 x i32> @llvm.x86.cast.tile.to.vector.v256i32(x86_amx %t1)
+  %add = add <256 x i32> %t2, %x
+  ret void
+}
+
+define dso_local void @__tile_loadd(%struct.__tile_str* nocapture %0, i8* %1, i64 %2) local_unnamed_addr {
+; CHECK-LABEL: @__tile_loadd(
+; CHECK-NEXT:    [[TMP4:%.*]] = alloca <256 x i32>, align 64
+; CHECK-NEXT:    [[TMP5:%.*]] = getelementptr inbounds [[STRUCT___TILE_STR:%.*]], %struct.__tile_str* [[TMP0:%.*]], i64 0, i32 0
+; CHECK-NEXT:    [[TMP6:%.*]] = load i16, i16* [[TMP5]], align 64
+; CHECK-NEXT:    [[TMP7:%.*]] = getelementptr inbounds [[STRUCT___TILE_STR]], %struct.__tile_str* [[TMP0]], i64 0, i32 1
+; CHECK-NEXT:    [[TMP8:%.*]] = load i16, i16* [[TMP7]], align 2
+; CHECK-NEXT:    [[TMP9:%.*]] = shl i64 [[TMP2:%.*]], 32
+; CHECK-NEXT:    [[TMP10:%.*]] = ashr exact i64 [[TMP9]], 32
+; CHECK-NEXT:    [[TMP11:%.*]] = tail call x86_amx @llvm.x86.tileloadd64.internal(i16 [[TMP6]], i16 [[TMP8]], i8* [[TMP1:%.*]], i64 [[TMP10]])
+; CHECK-NEXT:    [[TMP12:%.*]] = bitcast <256 x i32>* [[TMP4]] to i8*
+; CHECK-NEXT:    [[TMP13:%.*]] = sext i16 [[TMP8]] to i64
+; CHECK-NEXT:    call void @llvm.x86.tilestored64.internal(i16 [[TMP6]], i16 [[TMP8]], i8* [[TMP12]], i64 [[TMP13]], x86_amx [[TMP11]])
+; CHECK-NEXT:    [[TMP14:%.*]] = load <256 x i32>, <256 x i32>* [[TMP4]], align 1024
+; CHECK-NEXT:    [[TMP15:%.*]] = getelementptr inbounds [[STRUCT___TILE_STR]], %struct.__tile_str* [[TMP0]], i64 0, i32 2
+; CHECK-NEXT:    store <256 x i32> [[TMP14]], <256 x i32>* [[TMP15]], align 64
+; CHECK-NEXT:    ret void
+;
+  %4 = getelementptr inbounds %struct.__tile_str, %struct.__tile_str* %0, i64 0, i32 0
+  %5 = load i16, i16* %4, align 64
+  %6 = getelementptr inbounds %struct.__tile_str, %struct.__tile_str* %0, i64 0, i32 1
+  %7 = load i16, i16* %6, align 2
+  %8 = shl i64 %2, 32
+  %9 = ashr exact i64 %8, 32
+  %10 = tail call x86_amx @llvm.x86.tileloadd64.internal(i16 %5, i16 %7, i8* %1, i64 %9)
+  %11 = call <256 x i32> @llvm.x86.cast.tile.to.vector.v256i32(x86_amx %10)
+  %12 = getelementptr inbounds %struct.__tile_str, %struct.__tile_str* %0, i64 0, i32 2
+  store <256 x i32> %11, <256 x i32>* %12, align 64
+  ret void
+}
+
+define dso_local void @__tile_dpbssd(%struct.__tile_str* nocapture %0, %struct.__tile_str* nocapture readonly byval(%struct.__tile_str) align 64 %1, %struct.__tile_str* nocapture readonly byval(%struct.__tile_str) align 64 %2) local_unnamed_addr {
+; CHECK-LABEL: @__tile_dpbssd(
+; CHECK-NEXT:    [[TMP4:%.*]] = alloca <256 x i32>, align 64
+; CHECK-NEXT:    [[TMP5:%.*]] = alloca <256 x i32>, align 64
+; CHECK-NEXT:    [[TMP6:%.*]] = alloca <256 x i32>, align 64
+; CHECK-NEXT:    [[TMP7:%.*]] = alloca <256 x i32>, align 64
+; CHECK-NEXT:    [[TMP8:%.*]] = getelementptr inbounds [[STRUCT___TILE_STR:%.*]], %struct.__tile_str* [[TMP1:%.*]], i64 0, i32 0
+; CHECK-NEXT:    [[TMP9:%.*]] = load i16, i16* [[TMP8]], align 64
+; CHECK-NEXT:    [[TMP10:%.*]] = getelementptr inbounds [[STRUCT___TILE_STR]], %struct.__tile_str* [[TMP2:%.*]], i64 0, i32 1
+; CHECK-NEXT:    [[TMP11:%.*]] = load i16, i16* [[TMP10]], align 2
+; CHECK-NEXT:    [[TMP12:%.*]] = getelementptr inbounds [[STRUCT___TILE_STR]], %struct.__tile_str* [[TMP1]], i64 0, i32 1
+; CHECK-NEXT:    [[TMP13:%.*]] = load i16, i16* [[TMP12]], align 2
+; CHECK-NEXT:    [[TMP14:%.*]] = udiv i16 [[TMP13]], 4
+; CHECK-NEXT:    [[TMP15:%.*]] = getelementptr inbounds [[STRUCT___TILE_STR]], %struct.__tile_str* [[TMP0:%.*]], i64 0, i32 2
+; CHECK-NEXT:    [[TMP16:%.*]] = load <256 x i32>, <256 x i32>* [[TMP15]], align 64
+; CHECK-NEXT:    [[TMP17:%.*]] = bitcast <256 x i32>* [[TMP7]] to i8*
+; CHECK-NEXT:    store <256 x i32> [[TMP16]], <256 x i32>* [[TMP7]], align 1024
+; CHECK-NEXT:    [[TMP18:%.*]] = sext i16 [[TMP11]] to i64
+; CHECK-NEXT:    [[TMP19:%.*]] = call x86_amx @llvm.x86.tileloadd64.internal(i16 [[TMP9]], i16 [[TMP11]], i8* [[TMP17]], i64 [[TMP18]])
+; CHECK-NEXT:    [[TMP20:%.*]] = getelementptr inbounds [[STRUCT___TILE_STR]], %struct.__tile_str* [[TMP1]], i64 0, i32 2
+; CHECK-NEXT:    [[TMP21:%.*]] = load <256 x i32>, <256 x i32>* [[TMP20]], align 64
+; CHECK-NEXT:    [[TMP22:%.*]] = bitcast <256 x i32>* [[TMP6]] to i8*
+; CHECK-NEXT:    store <256 x i32> [[TMP21]], <256 x i32>* [[TMP6]], align 1024
+; CHECK-NEXT:    [[TMP23:%.*]] = sext i16 [[TMP13]] to i64
+; CHECK-NEXT:    [[TMP24:%.*]] = call x86_amx @llvm.x86.tileloadd64.internal(i16 [[TMP9]], i16 [[TMP13]], i8* [[TMP22]], i64 [[TMP23]])
+; CHECK-NEXT:    [[TMP25:%.*]] = getelementptr inbounds [[STRUCT___TILE_STR]], %struct.__tile_str* [[TMP2]], i64 0, i32 2
+; CHECK-NEXT:    [[TMP26:%.*]] = load <256 x i32>, <256 x i32>* [[TMP25]], align 64
+; CHECK-NEXT:    [[TMP27:%.*]] = bitcast <256 x i32>* [[TMP5]] to i8*
+; CHECK-NEXT:    store <256 x i32> [[TMP26]], <256 x i32>* [[TMP5]], align 1024
+; CHECK-NEXT:    [[TMP28:%.*]] = sext i16 [[TMP11]] to i64
+; CHECK-NEXT:    [[TMP29:%.*]] = call x86_amx @llvm.x86.tileloadd64.internal(i16 [[TMP14]], i16 [[TMP11]], i8* [[TMP27]], i64 [[TMP28]])
+; CHECK-NEXT:    [[TMP30:%.*]] = tail call x86_amx @llvm.x86.tdpbssd.internal(i16 [[TMP9]], i16 [[TMP11]], i16 [[TMP13]], x86_amx [[TMP19]], x86_amx [[TMP24]], x86_amx [[TMP29]])
+; CHECK-NEXT:    [[TMP31:%.*]] = bitcast <256 x i32>* [[TMP4]] to i8*
+; CHECK-NEXT:    [[TMP32:%.*]] = sext i16 [[TMP11]] to i64
+; CHECK-NEXT:    call void @llvm.x86.tilestored64.internal(i16 [[TMP9]], i16 [[TMP11]], i8* [[TMP31]], i64 [[TMP32]], x86_amx [[TMP30]])
+; CHECK-NEXT:    [[TMP33:%.*]] = load <256 x i32>, <256 x i32>* [[TMP4]], align 1024
+; CHECK-NEXT:    store <256 x i32> [[TMP33]], <256 x i32>* [[TMP15]], align 64
+; CHECK-NEXT:    ret void
+;
+  %4 = getelementptr inbounds %struct.__tile_str, %struct.__tile_str* %1, i64 0, i32 0
+  %5 = load i16, i16* %4, align 64
+  %6 = getelementptr inbounds %struct.__tile_str, %struct.__tile_str* %2, i64 0, i32 1
+  %7 = load i16, i16* %6, align 2
+  %8 = getelementptr inbounds %struct.__tile_str, %struct.__tile_str* %1, i64 0, i32 1
+  %9 = load i16, i16* %8, align 2
+  %10 = getelementptr inbounds %struct.__tile_str, %struct.__tile_str* %0, i64 0, i32 2
+  %11 = load <256 x i32>, <256 x i32>* %10, align 64
+  %12 = call x86_amx @llvm.x86.cast.vector.to.tile.v256i32(<256 x i32> %11)
+  %13 = getelementptr inbounds %struct.__tile_str, %struct.__tile_str* %1, i64 0, i32 2
+  %14 = load <256 x i32>, <256 x i32>* %13, align 64
+  %15 = call x86_amx @llvm.x86.cast.vector.to.tile.v256i32(<256 x i32> %14)
+  %16 = getelementptr inbounds %struct.__tile_str, %struct.__tile_str* %2, i64 0, i32 2
+  %17 = load <256 x i32>, <256 x i32>* %16, align 64
+  %18 = call x86_amx @llvm.x86.cast.vector.to.tile.v256i32(<256 x i32> %17)
+  %19 = tail call x86_amx @llvm.x86.tdpbssd.internal(i16 %5, i16 %7, i16 %9, x86_amx %12, x86_amx %15, x86_amx %18)
+  %20 = call <256 x i32> @llvm.x86.cast.tile.to.vector.v256i32(x86_amx %19)
+  store <256 x i32> %20, <256 x i32>* %10, align 64
+  ret void
+}
+
+define dso_local void @__tile_dpbsud(i16 %m, i16 %n, i16 %k, <256 x i32>* %pc, <256 x i32>* %pa, <256 x i32>* %pb) {
+; CHECK-LABEL: @__tile_dpbsud(
+; CHECK-NEXT:    [[TMP1:%.*]] = alloca <256 x i32>, align 64
+; CHECK-NEXT:    [[TMP2:%.*]] = alloca <256 x i32>, align 64
+; CHECK-NEXT:    [[TMP3:%.*]] = alloca <256 x i32>, align 64
+; CHECK-NEXT:    [[TMP4:%.*]] = alloca <256 x i32>, align 64
+; CHECK-NEXT:    [[TMP5:%.*]] = udiv i16 [[K:%.*]], 4
+; CHECK-NEXT:    [[T0:%.*]] = load <256 x i32>, <256 x i32>* [[PA:%.*]], align 64
+; CHECK-NEXT:    [[TMP6:%.*]] = bitcast <256 x i32>* [[TMP4]] to i8*
+; CHECK-NEXT:    store <256 x i32> [[T0]], <256 x i32>* [[TMP4]], align 1024
+; CHECK-NEXT:    [[TMP7:%.*]] = sext i16 [[K]] to i64
+; CHECK-NEXT:    [[TMP8:%.*]] = call x86_amx @llvm.x86.tileloadd64.internal(i16 [[M:%.*]], i16 [[K]], i8* [[TMP6]], i64 [[TMP7]])
+; CHECK-NEXT:    [[T2:%.*]] = load <256 x i32>, <256 x i32>* [[PB:%.*]], align 64
+; CHECK-NEXT:    [[TMP9:%.*]] = bitcast <256 x i32>* [[TMP3]] to i8*
+; CHECK-NEXT:    store <256 x i32> [[T2]], <256 x i32>* [[TMP3]], align 1024
+; CHECK-NEXT:    [[TMP10:%.*]] = sext i16 [[N:%.*]] to i64
+; CHECK-NEXT:    [[TMP11:%.*]] = call x86_amx @llvm.x86.tileloadd64.internal(i16 [[TMP5]], i16 [[N]], i8* [[TMP9]], i64 [[TMP10]])
+; CHECK-NEXT:    [[T4:%.*]] = load <256 x i32>, <256 x i32>* [[PC:%.*]], align 64
+; CHECK-NEXT:    [[TMP12:%.*]] = bitcast <256 x i32>* [[TMP2]] to i8*
+; CHECK-NEXT:    store <256 x i32> [[T4]], <256 x i32>* [[TMP2]], align 1024
+; CHECK-NEXT:    [[TMP13:%.*]] = sext i16 [[N]] to i64
+; CHECK-NEXT:    [[TMP14:%.*]] = call x86_amx @llvm.x86.tileloadd64.internal(i16 [[M]], i16 [[N]], i8* [[TMP12]], i64 [[TMP13]])
+; CHECK-NEXT:    [[T6:%.*]] = tail call x86_amx @llvm.x86.tdpbsud.internal(i16 [[M]], i16 [[N]], i16 [[K]], x86_amx [[TMP14]], x86_amx [[TMP8]], x86_amx [[TMP11]])
+; CHECK-NEXT:    [[TMP15:%.*]] = bitcast <256 x i32>* [[TMP1]] to i8*
+; CHECK-NEXT:    [[TMP16:%.*]] = sext i16 [[N]] to i64
+; CHECK-NEXT:    call void @llvm.x86.tilestored64.internal(i16 [[M]], i16 [[N]], i8* [[TMP15]], i64 [[TMP16]], x86_amx [[T6]])
+; CHECK-NEXT:    [[TMP17:%.*]] = load <256 x i32>, <256 x i32>* [[TMP1]], align 1024
+; CHECK-NEXT:    store <256 x i32> [[TMP17]], <256 x i32>* [[PC]], align 64
+; CHECK-NEXT:    ret void
+;
+  %t0 = load <256 x i32>, <256 x i32>* %pa, align 64
+  %t1 = call x86_amx @llvm.x86.cast.vector.to.tile.v256i32(<256 x i32> %t0)
+  %t2 = load <256 x i32>, <256 x i32>* %pb, align 64
+  %t3 = call x86_amx @llvm.x86.cast.vector.to.tile.v256i32(<256 x i32> %t2)
+  %t4 = load <256 x i32>, <256 x i32>* %pc, align 64
+  %t5 = call x86_amx @llvm.x86.cast.vector.to.tile.v256i32(<256 x i32> %t4)
+  %t6 = tail call x86_amx @llvm.x86.tdpbsud.internal(i16 %m, i16 %n, i16 %k, x86_amx %t5, x86_amx %t1, x86_amx %t3)
+  %t7 = call <256 x i32> @llvm.x86.cast.tile.to.vector.v256i32(x86_amx %t6)
+  store <256 x i32> %t7, <256 x i32>* %pc, align 64
+  ret void
+}
+
+define dso_local void @__tile_dpbusd(i16 %m, i16 %n, i16 %k, <256 x i32>* %pc, <256 x i32>* %pa, <256 x i32>* %pb) {
+; CHECK-LABEL: @__tile_dpbusd(
+; CHECK-NEXT:    [[TMP1:%.*]] = alloca <256 x i32>, align 64
+; CHECK-NEXT:    [[TMP2:%.*]] = alloca <256 x i32>, align 64
+; CHECK-NEXT:    [[TMP3:%.*]] = alloca <256 x i32>, align 64
+; CHECK-NEXT:    [[TMP4:%.*]] = alloca <256 x i32>, align 64
+; CHECK-NEXT:    [[TMP5:%.*]] = udiv i16 [[K:%.*]], 4
+; CHECK-NEXT:    [[T0:%.*]] = load <256 x i32>, <256 x i32>* [[PA:%.*]], align 64
+; CHECK-NEXT:    [[TMP6:%.*]] = bitcast <256 x i32>* [[TMP4]] to i8*
+; CHECK-NEXT:    store <256 x i32> [[T0]], <256 x i32>* [[TMP4]], align 1024
+; CHECK-NEXT:    [[TMP7:%.*]] = sext i16 [[K]] to i64
+; CHECK-NEXT:    [[TMP8:%.*]] = call x86_amx @llvm.x86.tileloadd64.internal(i16 [[M:%.*]], i16 [[K]], i8* [[TMP6]], i64 [[TMP7]])
+; CHECK-NEXT:    [[T2:%.*]] = load <256 x i32>, <256 x i32>* [[PB:%.*]], align 64
+; CHECK-NEXT:    [[TMP9:%.*]] = bitcast <256 x i32>* [[TMP3]] to i8*
+; CHECK-NEXT:    store <256 x i32> [[T2]], <256 x i32>* [[TMP3]], align 1024
+; CHECK-NEXT:    [[TMP10:%.*]] = sext i16 [[N:%.*]] to i64
+; CHECK-NEXT:    [[TMP11:%.*]] = call x86_amx @llvm.x86.tileloadd64.internal(i16 [[TMP5]], i16 [[N]], i8* [[TMP9]], i64 [[TMP10]])
+; CHECK-NEXT:    [[T4:%.*]] = load <256 x i32>, <256 x i32>* [[PC:%.*]], align 64
+; CHECK-NEXT:    [[TMP12:%.*]] = bitcast <256 x i32>* [[TMP2]] to i8*
+; CHECK-NEXT:    store <256 x i32> [[T4]], <256 x i32>* [[TMP2]], align 1024
+; CHECK-NEXT:    [[TMP13:%.*]] = sext i16 [[N]] to i64
+; CHECK-NEXT:    [[TMP14:%.*]] = call x86_amx @llvm.x86.tileloadd64.internal(i16 [[M]], i16 [[N]], i8* [[TMP12]], i64 [[TMP13]])
+; CHECK-NEXT:    [[T6:%.*]] = tail call x86_amx @llvm.x86.tdpbusd.internal(i16 [[M]], i16 [[N]], i16 [[K]], x86_amx [[TMP14]], x86_amx [[TMP8]], x86_amx [[TMP11]])
+; CHECK-NEXT:    [[TMP15:%.*]] = bitcast <256 x i32>* [[TMP1]] to i8*
+; CHECK-NEXT:    [[TMP16:%.*]] = sext i16 [[N]] to i64
+; CHECK-NEXT:    call void @llvm.x86.tilestored64.internal(i16 [[M]], i16 [[N]], i8* [[TMP15]], i64 [[TMP16]], x86_amx [[T6]])
+; CHECK-NEXT:    [[TMP17:%.*]] = load <256 x i32>, <256 x i32>* [[TMP1]], align 1024
+; CHECK-NEXT:    store <256 x i32> [[TMP17]], <256 x i32>* [[PC]], align 64
+; CHECK-NEXT:    ret void
+;
+  %t0 = load <256 x i32>, <256 x i32>* %pa, align 64
+  %t1 = call x86_amx @llvm.x86.cast.vector.to.tile.v256i32(<256 x i32> %t0)
+  %t2 = load <256 x i32>, <256 x i32>* %pb, align 64
+  %t3 = call x86_amx @llvm.x86.cast.vector.to.tile.v256i32(<256 x i32> %t2)
+  %t4 = load <256 x i32>, <256 x i32>* %pc, align 64
+  %t5 = call x86_amx @llvm.x86.cast.vector.to.tile.v256i32(<256 x i32> %t4)
+  %t6 = tail call x86_amx @llvm.x86.tdpbusd.internal(i16 %m, i16 %n, i16 %k, x86_amx %t5, x86_amx %t1, x86_amx %t3)
+  %t7 = call <256 x i32> @llvm.x86.cast.tile.to.vector.v256i32(x86_amx %t6)
+  store <256 x i32> %t7, <256 x i32>* %pc, align 64
+  ret void
+}
+
+define dso_local void @__tile_dpbuud(i16 %m, i16 %n, i16 %k, <256 x i32>* %pc, <256 x i32>* %pa, <256 x i32>* %pb) {
+; CHECK-LABEL: @__tile_dpbuud(
+; CHECK-NEXT:    [[TMP1:%.*]] = alloca <256 x i32>, align 64
+; CHECK-NEXT:    [[TMP2:%.*]] = alloca <256 x i32>, align 64
+; CHECK-NEXT:    [[TMP3:%.*]] = alloca <256 x i32>, align 64
+; CHECK-NEXT:    [[TMP4:%.*]] = alloca <256 x i32>, align 64
+; CHECK-NEXT:    [[TMP5:%.*]] = udiv i16 [[K:%.*]], 4
+; CHECK-NEXT:    [[T0:%.*]] = load <256 x i32>, <256 x i32>* [[PA:%.*]], align 64
+; CHECK-NEXT:    [[TMP6:%.*]] = bitcast <256 x i32>* [[TMP4]] to i8*
+; CHECK-NEXT:    store <256 x i32> [[T0]], <256 x i32>* [[TMP4]], align 1024
+; CHECK-NEXT:    [[TMP7:%.*]] = sext i16 [[K]] to i64
+; CHECK-NEXT:    [[TMP8:%.*]] = call x86_amx @llvm.x86.tileloadd64.internal(i16 [[M:%.*]], i16 [[K]], i8* [[TMP6]], i64 [[TMP7]])
+; CHECK-NEXT:    [[T2:%.*]] = load <256 x i32>, <256 x i32>* [[PB:%.*]], align 64
+; CHECK-NEXT:    [[TMP9:%.*]] = bitcast <256 x i32>* [[TMP3]] to i8*
+; CHECK-NEXT:    store <256 x i32> [[T2]], <256 x i32>* [[TMP3]], align 1024
+; CHECK-NEXT:    [[TMP10:%.*]] = sext i16 [[N:%.*]] to i64
+; CHECK-NEXT:    [[TMP11:%.*]] = call x86_amx @llvm.x86.tileloadd64.internal(i16 [[TMP5]], i16 [[N]], i8* [[TMP9]], i64 [[TMP10]])
+; CHECK-NEXT:    [[T4:%.*]] = load <256 x i32>, <256 x i32>* [[PC:%.*]], align 64
+; CHECK-NEXT:    [[TMP12:%.*]] = bitcast <256 x i32>* [[TMP2]] to i8*
+; CHECK-NEXT:    store <256 x i32> [[T4]], <256 x i32>* [[TMP2]], align 1024
+; CHECK-NEXT:    [[TMP13:%.*]] = sext i16 [[N]] to i64
+; CHECK-NEXT:    [[TMP14:%.*]] = call x86_amx @llvm.x86.tileloadd64.internal(i16 [[M]], i16 [[N]], i8* [[TMP12]], i64 [[TMP13]])
+; CHECK-NEXT:    [[T6:%.*]] = tail call x86_amx @llvm.x86.tdpbuud.internal(i16 [[M]], i16 [[N]], i16 [[K]], x86_amx [[TMP14]], x86_amx [[TMP8]], x86_amx [[TMP11]])
+; CHECK-NEXT:    [[TMP15:%.*]] = bitcast <256 x i32>* [[TMP1]] to i8*
+; CHECK-NEXT:    [[TMP16:%.*]] = sext i16 [[N]] to i64
+; CHECK-NEXT:    call void @llvm.x86.tilestored64.internal(i16 [[M]], i16 [[N]], i8* [[TMP15]], i64 [[TMP16]], x86_amx [[T6]])
+; CHECK-NEXT:    [[TMP17:%.*]] = load <256 x i32>, <256 x i32>* [[TMP1]], align 1024
+; CHECK-NEXT:    store <256 x i32> [[TMP17]], <256 x i32>* [[PC]], align 64
+; CHECK-NEXT:    ret void
+;
+  %t0 = load <256 x i32>, <256 x i32>* %pa, align 64
+  %t1 = call x86_amx @llvm.x86.cast.vector.to.tile.v256i32(<256 x i32> %t0)
+  %t2 = load <256 x i32>, <256 x i32>* %pb, align 64
+  %t3 = call x86_amx @llvm.x86.cast.vector.to.tile.v256i32(<256 x i32> %t2)
+  %t4 = load <256 x i32>, <256 x i32>* %pc, align 64
+  %t5 = call x86_amx @llvm.x86.cast.vector.to.tile.v256i32(<256 x i32> %t4)
+  %t6 = tail call x86_amx @llvm.x86.tdpbuud.internal(i16 %m, i16 %n, i16 %k, x86_amx %t5, x86_amx %t1, x86_amx %t3)
+  %t7 = call <256 x i32> @llvm.x86.cast.tile.to.vector.v256i32(x86_amx %t6)
+  store <256 x i32> %t7, <256 x i32>* %pc, align 64
+  ret void
+}
+
+define dso_local void @__tile_dpbf16ps(i16 %m, i16 %n, i16 %k, <256 x i32>* %pc, <256 x i32>* %pa, <256 x i32>* %pb) {
+; CHECK-LABEL: @__tile_dpbf16ps(
+; CHECK-NEXT:    [[TMP1:%.*]] = alloca <256 x i32>, align 64
+; CHECK-NEXT:    [[TMP2:%.*]] = alloca <256 x i32>, align 64
+; CHECK-NEXT:    [[TMP3:%.*]] = alloca <256 x i32>, align 64
+; CHECK-NEXT:    [[TMP4:%.*]] = alloca <256 x i32>, align 64
+; CHECK-NEXT:    [[TMP5:%.*]] = udiv i16 [[K:%.*]], 4
+; CHECK-NEXT:    [[T0:%.*]] = load <256 x i32>, <256 x i32>* [[PA:%.*]], align 64
+; CHECK-NEXT:    [[TMP6:%.*]] = bitcast <256 x i32>* [[TMP4]] to i8*
+; CHECK-NEXT:    store <256 x i32> [[T0]], <256 x i32>* [[TMP4]], align 1024
+; CHECK-NEXT:    [[TMP7:%.*]] = sext i16 [[K]] to i64
+; CHECK-NEXT:    [[TMP8:%.*]] = call x86_amx @llvm.x86.tileloadd64.internal(i16 [[M:%.*]], i16 [[K]], i8* [[TMP6]], i64 [[TMP7]])
+; CHECK-NEXT:    [[T2:%.*]] = load <256 x i32>, <256 x i32>* [[PB:%.*]], align 64
+; CHECK-NEXT:    [[TMP9:%.*]] = bitcast <256 x i32>* [[TMP3]] to i8*
+; CHECK-NEXT:    store <256 x i32> [[T2]], <256 x i32>* [[TMP3]], align 1024
+; CHECK-NEXT:    [[TMP10:%.*]] = sext i16 [[N:%.*]] to i64
+; CHECK-NEXT:    [[TMP11:%.*]] = call x86_amx @llvm.x86.tileloadd64.internal(i16 [[TMP5]], i16 [[N]], i8* [[TMP9]], i64 [[TMP10]])
+; CHECK-NEXT:    [[T4:%.*]] = load <256 x i32>, <256 x i32>* [[PC:%.*]], align 64
+; CHECK-NEXT:    [[TMP12:%.*]] = bitcast <256 x i32>* [[TMP2]] to i8*
+; CHECK-NEXT:    store <256 x i32> [[T4]], <256 x i32>* [[TMP2]], align 1024
+; CHECK-NEXT:    [[TMP13:%.*]] = sext i16 [[N]] to i64
+; CHECK-NEXT:    [[TMP14:%.*]] = call x86_amx @llvm.x86.tileloadd64.internal(i16 [[M]], i16 [[N]], i8* [[TMP12]], i64 [[TMP13]])
+; CHECK-NEXT:    [[T6:%.*]] = tail call x86_amx @llvm.x86.tdpbf16ps.internal(i16 [[M]], i16 [[N]], i16 [[K]], x86_amx [[TMP14]], x86_amx [[TMP8]], x86_amx [[TMP11]])
+; CHECK-NEXT:    [[TMP15:%.*]] = bitcast <256 x i32>* [[TMP1]] to i8*
+; CHECK-NEXT:    [[TMP16:%.*]] = sext i16 [[N]] to i64
+; CHECK-NEXT:    call void @llvm.x86.tilestored64.internal(i16 [[M]], i16 [[N]], i8* [[TMP15]], i64 [[TMP16]], x86_amx [[T6]])
+; CHECK-NEXT:    [[TMP17:%.*]] = load <256 x i32>, <256 x i32>* [[TMP1]], align 1024
+; CHECK-NEXT:    store <256 x i32> [[TMP17]], <256 x i32>* [[PC]], align 64
+; CHECK-NEXT:    ret void
+;
+  %t0 = load <256 x i32>, <256 x i32>* %pa, align 64
+  %t1 = call x86_amx @llvm.x86.cast.vector.to.tile.v256i32(<256 x i32> %t0)
+  %t2 = load <256 x i32>, <256 x i32>* %pb, align 64
+  %t3 = call x86_amx @llvm.x86.cast.vector.to.tile.v256i32(<256 x i32> %t2)
+  %t4 = load <256 x i32>, <256 x i32>* %pc, align 64
+  %t5 = call x86_amx @llvm.x86.cast.vector.to.tile.v256i32(<256 x i32> %t4)
+  %t6 = tail call x86_amx @llvm.x86.tdpbf16ps.internal(i16 %m, i16 %n, i16 %k, x86_amx %t5, x86_amx %t1, x86_amx %t3)
+  %t7 = call <256 x i32> @llvm.x86.cast.tile.to.vector.v256i32(x86_amx %t6)
+  store <256 x i32> %t7, <256 x i32>* %pc, align 64
+  ret void
+}
+
+define dso_local void @__tile_stored(i8* %0, i64 %1, %struct.__tile_str* nocapture readonly byval(%struct.__tile_str) align 64 %2) local_unnamed_addr {
+; CHECK-LABEL: @__tile_stored(
+; CHECK-NEXT:    [[TMP4:%.*]] = alloca <256 x i32>, align 64
+; CHECK-NEXT:    [[TMP5:%.*]] = getelementptr inbounds [[STRUCT___TILE_STR:%.*]], %struct.__tile_str* [[TMP2:%.*]], i64 0, i32 0
+; CHECK-NEXT:    [[TMP6:%.*]] = load i16, i16* [[TMP5]], align 64
+; CHECK-NEXT:    [[TMP7:%.*]] = getelementptr inbounds [[STRUCT___TILE_STR]], %struct.__tile_str* [[TMP2]], i64 0, i32 1
+; CHECK-NEXT:    [[TMP8:%.*]] = load i16, i16* [[TMP7]], align 2
+; CHECK-NEXT:    [[TMP9:%.*]] = getelementptr inbounds [[STRUCT___TILE_STR]], %struct.__tile_str* [[TMP2]], i64 0, i32 2
+; CHECK-NEXT:    [[TMP10:%.*]] = load <256 x i32>, <256 x i32>* [[TMP9]], align 64
+; CHECK-NEXT:    [[TMP11:%.*]] = bitcast <256 x i32>* [[TMP4]] to i8*
+; CHECK-NEXT:    store <256 x i32> [[TMP10]], <256 x i32>* [[TMP4]], align 1024
+; CHECK-NEXT:    [[TMP12:%.*]] = sext i16 [[TMP8]] to i64
+; CHECK-NEXT:    [[TMP13:%.*]] = call x86_amx @llvm.x86.tileloadd64.internal(i16 [[TMP6]], i16 [[TMP8]], i8* [[TMP11]], i64 [[TMP12]])
+; CHECK-NEXT:    [[TMP14:%.*]] = shl i64 [[TMP1:%.*]], 32
+; CHECK-NEXT:    [[TMP15:%.*]] = ashr exact i64 [[TMP14]], 32
+; CHECK-NEXT:    tail call void @llvm.x86.tilestored64.internal(i16 [[TMP6]], i16 [[TMP8]], i8* [[TMP0:%.*]], i64 [[TMP15]], x86_amx [[TMP13]])
+; CHECK-NEXT:    ret void
+;
+  %4 = getelementptr inbounds %struct.__tile_str, %struct.__tile_str* %2, i64 0, i32 0
+  %5 = load i16, i16* %4, align 64
+  %6 = getelementptr inbounds %struct.__tile_str, %struct.__tile_str* %2, i64 0, i32 1
+  %7 = load i16, i16* %6, align 2
+  %8 = getelementptr inbounds %struct.__tile_str, %struct.__tile_str* %2, i64 0, i32 2
+  %9 = load <256 x i32>, <256 x i32>* %8, align 64
+  %10 = call x86_amx @llvm.x86.cast.vector.to.tile.v256i32(<256 x i32> %9)
+  %11 = shl i64 %1, 32
+  %12 = ashr exact i64 %11, 32
+  tail call void @llvm.x86.tilestored64.internal(i16 %5, i16 %7, i8* %0, i64 %12, x86_amx %10)
+  ret void
+}
+
+declare x86_amx @llvm.x86.tileloadd64.internal(i16, i16, i8*, i64)
+declare x86_amx @llvm.x86.tdpbssd.internal(i16, i16, i16, x86_amx, x86_amx, x86_amx)
+declare x86_amx @llvm.x86.tdpbsud.internal(i16, i16, i16, x86_amx, x86_amx, x86_amx)
+declare x86_amx @llvm.x86.tdpbusd.internal(i16, i16, i16, x86_amx, x86_amx, x86_amx)
+declare x86_amx @llvm.x86.tdpbuud.internal(i16, i16, i16, x86_amx, x86_amx, x86_amx)
+declare x86_amx @llvm.x86.tdpbf16ps.internal(i16, i16, i16, x86_amx, x86_amx, x86_amx)
+declare void @llvm.x86.tilestored64.internal(i16, i16, i8*, i64, x86_amx)
+
+declare x86_amx @llvm.x86.cast.vector.to.tile.v256i32(<256 x i32>)
+declare x86_amx @llvm.x86.cast.vector.to.tile.v225i32(<225 x i32>)
+declare <256 x i32> @llvm.x86.cast.tile.to.vector.v256i32(x86_amx)
+declare <225 x i32> @llvm.x86.cast.tile.to.vector.v225i32(x86_amx)