[X86] Add x86_amx type for intel AMX.

[LuoYuanke]

Summary

The x86_amx is used for AMX intrinsics. <256 x i32> is bitcasted to x86_amx when it is used by AMX intrinsics, and x86_amx is bitcasted to <256 x i32> when it is used by load/store instruction. So amx intrinsics only operate on type x86_amx. The new type x86_amx can help to separate amx intrinsics from llvm IR instructions (+-*/). Thank Craig for the idea. This patch depends on https://reviews.llvm.org/D87981.

[craig.topper]

I only took a quick pass through this so far. What happens if a bitcast between x86amx and v256i32(or any other 1024-bit vector type) exists in the IR but isn't next to a load/store?

llvm/lib/Target/X86/X86ISelLowering.cpp

5348–5349

Should this just be deleted?

llvm/lib/Target/X86/X86LowerAMXType.cpp

413

Don't use an assert to check the result of a dyn_cast. If it shouldn't fail just use cast<LoadInst> which will assert internally.

421

Unchecked dyn_cast

444

Use cast.

[LuoYuanke]

In D91927#2412557, @craig.topper wrote:

I only took a quick pass through this so far. What happens if a bitcast between x86amx and v256i32(or any other 1024-bit vector type) exists in the IR but isn't next to a load/store?

@craig.topper , thank you for reviewing my patch.
I think if user just use our external API, such IR won't be generated. However if there is such IR, we can transform bitcast to <store, load>, so that the type can be translated through memory. One of <store, load> is AMX intrinsics store/load, so it won't be optimized. Is it reasonable?

[craig.topper]

In D91927#2412604, @LuoYuanke wrote:

In D91927#2412557, @craig.topper wrote:

I only took a quick pass through this so far. What happens if a bitcast between x86amx and v256i32(or any other 1024-bit vector type) exists in the IR but isn't next to a load/store?

@craig.topper , thank you for reviewing my patch.
I think if user just use our external API, such IR won't be generated. However if there is such IR, we can transform bitcast to <store, load>, so that the type can be translated through memory. One of <store, load> is AMX intrinsics store/load, so it won't be optimized. Is it reasonable?

Its fine if its not optimized, just make sure it doesn't crash.

[LuoYuanke]

Address Craig's comments.

[LuoYuanke]

Address Craig's comments.
Change dyn_cast to cast.

[pengfei]

llvm/lib/IR/DataLayout.cpp

796

Should be 512 bits?

llvm/lib/Target/X86/X86LowerAMXType.cpp

398

Why the alignment not be 64?

[pengfei]

llvm/lib/Target/X86/X86LowerAMXType.cpp

448

%src is not used here.

llvm/utils/TableGen/IntrinsicEmitter.cpp

252

Remove ,

[craig.topper]

llvm/lib/Target/X86/X86LowerAMXType.cpp

391–392

Shouldn't this be in the function's entry block?

408

Just use Value. auto doesn't add any value other than shortening by 1 character.

462

maximun->maximum

466

Use Builder.getInt8PtrTy then you don't need Ctx

[LuoYuanke]

llvm/lib/IR/DataLayout.cpp

796

Yes. It is 512. Thanks.

llvm/lib/Target/X86/X86LowerAMXType.cpp

391–392

Yes. It is in function's entry block. It is done in line 48 of function CreateAllocaInst(). CreateAllocaInst() is actually copied from your code. :)

398

1024 is conservatives, because vector require the alignment to be the vector size. Here generate vector <256 x i32> load/store.

[LuoYuanke]

Address Craig and Pengfei's comments.

[LuoYuanke]

Add the handler of "bitcast <256 x i32>* to x86_amx*".
Refactor the code.

[LuoYuanke]

Avoid generatng constant for x86_amx.

[LuoYuanke]

Rebase.

[pengfei]

llvm/lib/IR/ConstantFold.cpp

540

Operation should at the end of the line.

llvm/lib/Target/X86/X86LowerAMXType.cpp

176

I think we'd better to check exceptions. E.g.

default:
  llvm_unreachable("");
case Intrinsic::x86_tileloadd64_internal:
case Intrinsic::x86_tdpbssd_internal:
case Intrinsic::x86_tilestored64_internal:
  Row = II->getArgOperand(0);
  Col = II->getArgOperand(1);
  break;

239

Why don't check empty like line 157?

393–394

Better move it to line 310.

396–403

Better to reuse the cast result, e.g.

BitCastInst *BInst = dyn_cast<BitCastInst>(&Inst);
if (!BInst )

You can save several cast<BitCastInst>(&Inst) below.

429

Is it possible the x86_amx operand isn't from AMX intrinsic, e.g.

%src = bitcast <256 x i32> %xxx to x86_amx
%2 = bitcast x86_amx %src to <256 x i32>

430

Where's x86_amx* %tile from? Shouldn't been transfered to x86_amx before this bitcast if it exists?

436

Maybe better to keep a duplicated load that calling transformBitcast. The same for line 285.

451

Why we need to consider <256 x i32> has more than one use?

llvm/test/CodeGen/X86/AMX/amx-across-func.ll

89–91

Better to remove these unused attributes. The same to other tests.

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

67

For this and the next test, we have chances to optimize to memcpy if we can make sure %s is constant 64.

145

We don't need to check this case now, right?

[LuoYuanke]

Address Pengfei's comments.

[LuoYuanke]

Rebase and fix lit test case failure.

[pengfei]

llvm/lib/Target/X86/X86LowerAMXType.cpp

254

Why don't put it in DeadBitcasts?

263

Maybe better to use BitCastInst?

284

Can we leave the canonicalize bitcast cases a single patch. It's a bit complex here and I don't think it's a common case.

420

This comment is for above code? Better move it up.

[LuoYuanke]

llvm/lib/Target/X86/X86LowerAMXType.cpp

263

There may be dead load or store instructions.

284

Ok, I'll create another patch for it.

396–403

That's good. Thanks.

429

Good catch. I'll add support for this pattern.

llvm/test/CodeGen/X86/AMX/amx-across-func.ll

89–91

I'll create a separate patch to clean the attributes.

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

67

If the stride is 64 we can transform the code to memcpy. How about do it in another patch?

145

It can check the load and store instruction is not transformed if they are not participate in amx operation. I prefer to keep the case.

[LuoYuanke]

Address Pengfei's comments.

[LuoYuanke]

llvm/lib/Target/X86/X86LowerAMXType.cpp

430

In my test case, it is transformed after Combine redundant instructions.

*** IR Dump After Simplify the CFG ***
define internal fastcc void @_ZL12__tile_loaddP15__tile1024i_strPKvm(%struct.__tile1024i_str* nocapture %dst) unnamed_addr #4 {
entry:
  %row = getelementptr inbounds %struct.__tile1024i_str, %struct.__tile1024i_str* %dst, i64 0, i32 0
  %0 = load i16, i16* %row, align 64, !tbaa !2
  %col = getelementptr inbounds %struct.__tile1024i_str, %struct.__tile1024i_str* %dst, i64 0, i32 1
  %1 = load i16, i16* %col, align 2, !tbaa !7
  %2 = call x86_amx @llvm.x86.tileloadd64.internal(i16 %0, i16 %1, i8* getelementptr inbounds ([1024 x i8], [1024 x i8]* @buf, i64 0, i64 0), i64 64) #6
  %3 = bitcast x86_amx %2 to <256 x i32>
  %tile = getelementptr inbounds %struct.__tile1024i_str, %struct.__tile1024i_str* %dst, i64 0, i32 3
  store <256 x i32> %3, <256 x i32>* %tile, align 64, !tbaa !8
  ret void
}

To

*** IR Dump After Combine redundant instructions ***
; Function Attrs: alwaysinline nounwind uwtable mustprogress
define internal fastcc void @_ZL12__tile_loaddP15__tile1024i_strPKvm(%struct.__tile1024i_str* nocapture %dst) unnamed_addr #4 {
entry:
  %row = getelementptr inbounds %struct.__tile1024i_str, %struct.__tile1024i_str* %dst, i64 0, i32 0
  %0 = load i16, i16* %row, align 64, !tbaa !2
  %col = getelementptr inbounds %struct.__tile1024i_str, %struct.__tile1024i_str* %dst, i64 0, i32 1
  %1 = load i16, i16* %col, align 2, !tbaa !7
  %2 = call x86_amx @llvm.x86.tileloadd64.internal(i16 %0, i16 %1, i8* getelementptr inbounds ([1024 x i8], [1024
x i8]* @buf, i64 0, i64 0), i64 64) #6
  %tile = getelementptr inbounds %struct.__tile1024i_str, %struct.__tile1024i_str* %dst, i64 0, i32 3
  %3 = bitcast <256 x i32>* %tile to x86_amx*
  store x86_amx %2, x86_amx* %3, align 64, !tbaa !8
  ret void
}

[pengfei]

In my test case, it is transformed after Combine redundant instructions.

Can we disable it for AMX type? The pointer to AMX type is meaningless and may result in bad perfomance.

llvm/lib/Target/X86/X86LowerAMXType.cpp

282

I don't see any chance this happen. But we still need to handle the x86_amx* here if possible, right?
Maybe better to give an assertion for now.

cast<PointerType>(Src->getType())->isX86_AMXTy()

[LuoYuanke]

In D91927#2469977, @pengfei wrote:

In my test case, it is transformed after Combine redundant instructions.

Can we disable it for AMX type? The pointer to AMX type is meaningless and may result in bad perfomance.

Ok, I'll disable the transform for AMX type.

[LuoYuanke]

Address Pengfei's comments.

[LuoYuanke]

Improve the comments.

[pengfei]

In D91927#2470818, @LuoYuanke wrote:

In D91927#2469977, @pengfei wrote:

In my test case, it is transformed after Combine redundant instructions.

Can we disable it for AMX type? The pointer to AMX type is meaningless and may result in bad perfomance.

Ok, I'll disable the transform for AMX type.

Good job.

llvm/lib/Target/X86/X86LowerAMXType.cpp

153

Currently, we don't have HW type for v256i32. I think 64 bytes(512bits) should be enough here.

233–250

Value

235

How about the Tile comes from tdpbssd?

398

We don't need to align to 1024. 64 should be enough. The same for below comments.

422

How about the Tile comes from tdpbssd?

[LuoYuanke]

llvm/lib/Target/X86/X86LowerAMXType.cpp

235

We have a convention, when amx intrinsics define a x86_amx tile the first 2 operands is the shape of the defined tile. For tdpbssd, the intrinsics operands are (m, n, k, ...). (m, n) is the shape of the produced tile.

[pengfei]

llvm/lib/Target/X86/X86LowerAMXType.cpp

235

Oh, yes. I missed that. Thanks.

411

vector

426

Why we need to recursively delete them? I think delete the nodes in DeadInsts is enough.

[LuoYuanke]

Address Pengfei's comments.

[LuoYuanke]

Refine comments.

[pengfei]

LGTM. Thanks for the refactors. Maybe better to wait for a few days to see if others have objections.

[LuoYuanke]

In D91927#2471140, @pengfei wrote:

LGTM. Thanks for the refactors. Maybe better to wait for a few days to see if others have objections.

Thank Pengfei for the review. Sure, I'll wait for a few days.

[uabelho]

llvm/include/llvm/IR/Type.h

68

This addition causes a compilation warning in HexagonTargetObjectFile.cpp:

../lib/Target/Hexagon/HexagonTargetObjectFile.cpp:297:11: error: enumeration value 'X86_AMXTyID' not handled in switch [-Werror,-Wswitch]
  switch (Ty->getTypeID()) {
          ^
1 error generated.

Seen in build bots, e.g. here:
http://lab.llvm.org:8011/#/builders/57/builds/2889/steps/6/logs/stdio

[pengfei]

llvm/include/llvm/IR/Type.h

68

Thanks Mikael for pointing it out. I think we just need to put the type in the switch table. I've posted a patch to fix it. rG16c2067cf212.

[uabelho]

llvm/include/llvm/IR/Type.h

68

Yep, thanks!

[MaskRay]

D93944 fixed an llvm-c-test issue. Note, adding new enum members usually requires check-all (at least check-llvm, but Clang may use these enum as well) because they can be used everywhere.

[LuoYuanke]

Thank @pengfei and @MaskRay.

[cuviper]

llvm/include/llvm-c/Core.h

163

This is a breaking change to the C ABI -- can we move it to the end of the enum?
https://bugs.llvm.org/show_bug.cgi?id=48905

[MaskRay]

llvm/include/llvm-c/Core.h

163

Done in 6612c2bb68becda5504099b48082c844503c6d4c

[LuoYuanke]

llvm/include/llvm-c/Core.h

163

@MaskRay, thank you!

clang/test/CodeGen/X86/amx_api.c

 // RUN: %clang_cc1 %s -ffreestanding -triple=x86_64-unknown-unknown  -target-feature +avx512f  -target-feature +amx-int8  \
 // RUN: -target-feature +amx-bf16 -emit-llvm -o - -Werror -pedantic | FileCheck %s --check-prefixes=CHECK
 
 #include <immintrin.h>
 
 char buf[1024];
 #define STRIDE 32
 
 char buf2[1024];
 
 // This is an example code and integration test.
 void test_api(int cond, short row, short col) {
   //CHECK-LABEL: @test_api
-  //CHECK: call <256 x i32> @llvm.x86.tileloadd64.internal
-  //CHECK: call <256 x i32> @llvm.x86.tdpbssd.internal
+  //CHECK: call x86_amx @llvm.x86.tileloadd64.internal
+  //CHECK: call x86_amx @llvm.x86.tdpbssd.internal
   //CHECK: call void @llvm.x86.tilestored64.internal
   __tile1024i a = {row, 8};
   __tile1024i b = {8, col};
   __tile1024i c = {row, col};
 
   if (cond) {
     __tile_loadd(&a, buf, STRIDE);
     __tile_loadd(&b, buf, STRIDE);
     __tile_loadd(&c, buf, STRIDE);
   } else {
     __tile_loadd(&a, buf2, STRIDE);
     __tile_loadd(&b, buf2, STRIDE);
     __tile_loadd(&c, buf2, STRIDE);
   }
   __tile_dpbsud(&c, a, b);
   __tile_stored(buf, STRIDE, c);
 }
 
 void test_tile_loadd(short row, short col) {
   //CHECK-LABEL: @test_tile_loadd
-  //CHECK: call <256 x i32> @llvm.x86.tileloadd64.internal
+  //CHECK: call x86_amx @llvm.x86.tileloadd64.internal
+  //CHECK-NEXT: {{%.*}} = bitcast x86_amx {{%.*}} to <256 x i32>
   __tile1024i a = {row, col};
   __tile_loadd(&a, buf, STRIDE);
 }
 
 void test_tile_dpbsud(__tile1024i a, __tile1024i b, __tile1024i c) {
   //CHECK-LABEL: @test_tile_dpbsud
-  //CHECK: call <256 x i32> @llvm.x86.tdpbssd.internal
+  //CHECK: call x86_amx @llvm.x86.tdpbssd.internal
+  //CHECK-NEXT: {{%.*}} = bitcast x86_amx {{%.*}} to <256 x i32>
   __tile_dpbsud(&c, a, b);
 }
 
 void test_tile_stored(__tile1024i c) {
   //CHECK-LABEL: @test_tile_stored
-  //CHECK: call void @llvm.x86.tilestored64.internal
+  //CHECK: {{%.*}} = bitcast <256 x i32> {{%.*}} to x86_amx
+  //CHECK-NEXT: call void @llvm.x86.tilestored64.internal
   __tile_stored(buf, STRIDE, c);
 }

llvm/include/llvm-c/Core.h

typedef enum {
   LLVMIntegerTypeKind,   /**< Arbitrary bit width integers */
   LLVMFunctionTypeKind,  /**< Functions */
   LLVMStructTypeKind,    /**< Structures */
   LLVMArrayTypeKind,     /**< Arrays */
   LLVMPointerTypeKind,   /**< Pointers */
   LLVMVectorTypeKind,    /**< Fixed width SIMD vector type */
   LLVMMetadataTypeKind,  /**< Metadata */
   LLVMX86_MMXTypeKind,   /**< X86 MMX */
+  LLVMX86_AMXTypeKind,   /**< X86 AMX */
   LLVMTokenTypeKind,     /**< Tokens */
   LLVMScalableVectorTypeKind, /**< Scalable SIMD vector type */
   LLVMBFloatTypeKind     /**< 16 bit brain floating point type */
 } LLVMTypeKind;
 
 typedef enum {
   LLVMExternalLinkage,    /**< Externally visible function */
   LLVMAvailableExternallyLinkage,
   LLVMLinkOnceAnyLinkage, /**< Keep one copy of function when linking (inline)*/
   LLVMLinkOnceODRLinkage, /**< Same, but only replaced by something
                             equivalent. */

 LLVMTypeRef LLVMLabelTypeInContext(LLVMContextRef C);
 
 /**
  * Create a X86 MMX type in a context.
  */
 LLVMTypeRef LLVMX86MMXTypeInContext(LLVMContextRef C);
 
 /**
+ * Create a X86 AMX type in a context.
+ */
+LLVMTypeRef LLVMX86AMXTypeInContext(LLVMContextRef C);
+
+/**
  * Create a token type in a context.
  */
 LLVMTypeRef LLVMTokenTypeInContext(LLVMContextRef C);
 
 /**
  * Create a metadata type in a context.
  */
 LLVMTypeRef LLVMMetadataTypeInContext(LLVMContextRef C);
 
 /**
  * These are similar to the above functions except they operate on the
  * global context.
  */
 LLVMTypeRef LLVMVoidType(void);
 LLVMTypeRef LLVMLabelType(void);
 LLVMTypeRef LLVMX86MMXType(void);
+LLVMTypeRef LLVMX86AMXType(void);
 
 /**
  * @}
  */
 
 /**
  * @}
  */

llvm/include/llvm/Bitcode/LLVMBitCodes.h

enum TypeCodes {
   TYPE_CODE_STRUCT_ANON = 18,  // STRUCT_ANON: [ispacked, eltty x N]
   TYPE_CODE_STRUCT_NAME = 19,  // STRUCT_NAME: [strchr x N]
   TYPE_CODE_STRUCT_NAMED = 20, // STRUCT_NAMED: [ispacked, eltty x N]
 
   TYPE_CODE_FUNCTION = 21, // FUNCTION: [vararg, retty, paramty x N]
 
   TYPE_CODE_TOKEN = 22, // TOKEN
 
-  TYPE_CODE_BFLOAT = 23 // BRAIN FLOATING POINT
+  TYPE_CODE_BFLOAT = 23, // BRAIN FLOATING POINT
+  TYPE_CODE_X86_AMX = 24 // X86 AMX
 };
 
 enum OperandBundleTagCode {
   OPERAND_BUNDLE_TAG = 1, // TAG: [strchr x N]
 };
 
 enum SyncScopeNameCode {
   SYNC_SCOPE_NAME = 1,

llvm/include/llvm/CodeGen/ValueTypes.td

 
 def x86mmx : ValueType<64 , 157>;   // X86 MMX value
 def FlagVT : ValueType<0  , 158>;   // Pre-RA sched glue
 def isVoid : ValueType<0  , 159>;   // Produces no value
 def untyped: ValueType<8  , 160>;   // Produces an untyped value
 def exnref : ValueType<0  , 161>;   // WebAssembly's exnref type
 def funcref : ValueType<0  , 162>;   // WebAssembly's funcref type
 def externref : ValueType<0  , 163>;   // WebAssembly's externref type
+def x86amx : ValueType<8192, 164>;   // X86 AMX value
 
 
 def token  : ValueType<0  , 248>;   // TokenTy
 def MetadataVT: ValueType<0, 249>;  // Metadata
 
 // Pseudo valuetype mapped to the current pointer size to any address space.
 // Should only be used in TableGen.
 def iPTRAny   : ValueType<0, 250>;

llvm/include/llvm/IR/DataLayout.h

inline TypeSize DataLayout::getTypeSizeInBits(Type *Ty) const {
   case Type::FloatTyID:
     return TypeSize::Fixed(32);
   case Type::DoubleTyID:
   case Type::X86_MMXTyID:
     return TypeSize::Fixed(64);
   case Type::PPC_FP128TyID:
   case Type::FP128TyID:
     return TypeSize::Fixed(128);
+  case Type::X86_AMXTyID:
+    return TypeSize::Fixed(8192);
   // In memory objects this is always aligned to a higher boundary, but
   // only 80 bits contain information.
   case Type::X86_FP80TyID:
     return TypeSize::Fixed(80);
   case Type::FixedVectorTyID:
   case Type::ScalableVectorTyID: {
     VectorType *VTy = cast<VectorType>(Ty);
     auto EltCnt = VTy->getElementCount();

llvm/include/llvm/IR/Intrinsics.h

enum IITDescriptorKind {
       HalfVecArgument,
       SameVecWidthArgument,
       PtrToArgument,
       PtrToElt,
       VecOfAnyPtrsToElt,
       VecElementArgument,
       Subdivide2Argument,
       Subdivide4Argument,
-      VecOfBitcastsToInt
+      VecOfBitcastsToInt,
+      AMX
     } Kind;
 
     union {
       unsigned Integer_Width;
       unsigned Float_Width;
       unsigned Pointer_AddressSpace;
       unsigned Struct_NumElements;
       unsigned Argument_Info;

llvm/include/llvm/IR/Intrinsics.td

 def llvm_empty_ty      : LLVMType<OtherVT>;                       // { }
 def llvm_descriptor_ty : LLVMPointerType<llvm_empty_ty>;          // { }*
 def llvm_metadata_ty   : LLVMType<MetadataVT>;                    // !{...}
 def llvm_token_ty      : LLVMType<token>;                         // token
 
 def llvm_x86mmx_ty     : LLVMType<x86mmx>;
 def llvm_ptrx86mmx_ty  : LLVMPointerType<llvm_x86mmx_ty>;         // <1 x i64>*
 
+def llvm_x86amx_ty     : LLVMType<x86amx>;
+
 def llvm_v2i1_ty       : LLVMType<v2i1>;     //   2 x i1
 def llvm_v4i1_ty       : LLVMType<v4i1>;     //   4 x i1
 def llvm_v8i1_ty       : LLVMType<v8i1>;     //   8 x i1
 def llvm_v16i1_ty      : LLVMType<v16i1>;    //  16 x i1
 def llvm_v32i1_ty      : LLVMType<v32i1>;    //  32 x i1
 def llvm_v64i1_ty      : LLVMType<v64i1>;    //  64 x i1
 def llvm_v128i1_ty     : LLVMType<v128i1>;   // 128 x i1
 def llvm_v256i1_ty     : LLVMType<v256i1>;   // 256 x i1

llvm/include/llvm/IR/IntrinsicsX86.td

let TargetPrefix = "x86" in {
   def int_x86_tdpbuud : GCCBuiltin<"__builtin_ia32_tdpbuud">,
               Intrinsic<[], [llvm_i8_ty, llvm_i8_ty, llvm_i8_ty],
                         [ImmArg<ArgIndex<0>>, ImmArg<ArgIndex<1>>,
                          ImmArg<ArgIndex<2>>]>;
   def int_x86_tdpbf16ps : GCCBuiltin<"__builtin_ia32_tdpbf16ps">,
               Intrinsic<[], [llvm_i8_ty, llvm_i8_ty, llvm_i8_ty],
                         [ImmArg<ArgIndex<0>>, ImmArg<ArgIndex<1>>,
                          ImmArg<ArgIndex<2>>]>;
+  // AMX - internal intrinsics
+  def int_x86_tileloadd64_internal :
+              GCCBuiltin<"__builtin_ia32_tileloadd64_internal">,
+              Intrinsic<[llvm_x86amx_ty],
+                        [llvm_i16_ty, llvm_i16_ty, llvm_ptr_ty, llvm_i64_ty],
+                        []>;
+  def int_x86_tdpbssd_internal :
+              GCCBuiltin<"__builtin_ia32_tdpbssd_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_tilestored64_internal :
+              GCCBuiltin<"__builtin_ia32_tilestored64_internal">,
+              Intrinsic<[], [llvm_i16_ty, llvm_i16_ty, llvm_ptr_ty,
+                             llvm_i64_ty, llvm_x86amx_ty], []>;
 }
 
 //===----------------------------------------------------------------------===//
 // 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], []>;
-// AMX - internal intrinsics
-  def int_x86_tileloadd64_internal :
-              GCCBuiltin<"__builtin_ia32_tileloadd64_internal">,
-              Intrinsic<[llvm_v256i32_ty],
-                        [llvm_i16_ty, llvm_i16_ty, llvm_ptr_ty, llvm_i64_ty],
-                        []>;
-  def int_x86_tdpbssd_internal :
-              GCCBuiltin<"__builtin_ia32_tdpbssd_internal">,
-              Intrinsic<[llvm_v256i32_ty],
-                        [llvm_i16_ty, llvm_i16_ty, llvm_i16_ty,
-                         llvm_v256i32_ty, llvm_v256i32_ty,
-                         llvm_v256i32_ty], []>;
-  def int_x86_tilestored64_internal :
-              GCCBuiltin<"__builtin_ia32_tilestored64_internal">,
-              Intrinsic<[], [llvm_i16_ty, llvm_i16_ty, llvm_ptr_ty,
-                             llvm_i64_ty, llvm_v256i32_ty], []>;
 }

llvm/include/llvm/IR/Type.h

enum TypeID {
     DoubleTyID,    ///< 64-bit floating point type
     X86_FP80TyID,  ///< 80-bit floating point type (X87)
     FP128TyID,     ///< 128-bit floating point type (112-bit significand)
     PPC_FP128TyID, ///< 128-bit floating point type (two 64-bits, PowerPC)
     VoidTyID,      ///< type with no size
     LabelTyID,     ///< Labels
     MetadataTyID,  ///< Metadata
     X86_MMXTyID,   ///< MMX vectors (64 bits, X86 specific)
+    X86_AMXTyID,   ///< AMX vectors (8192 bits, X86 specific)
     TokenTyID,     ///< Tokens
 
     // Derived types... see DerivedTypes.h file.
     IntegerTyID,       ///< Arbitrary bit width integers
     FunctionTyID,      ///< Functions
     PointerTyID,       ///< Pointers
     StructTyID,        ///< Structures
     ArrayTyID,         ///< Arrays
const fltSemantics &getFltSemantics() const {
     case PPC_FP128TyID: return APFloat::PPCDoubleDouble();
     default: llvm_unreachable("Invalid floating type");
     }
   }
 
   /// Return true if this is X86 MMX.
   bool isX86_MMXTy() const { return getTypeID() == X86_MMXTyID; }
 
+  /// Return true if this is X86 AMX.
+  bool isX86_AMXTy() const { return getTypeID() == X86_AMXTyID; }
+
   /// Return true if this is a FP type or a vector of FP.
   bool isFPOrFPVectorTy() const { return getScalarType()->isFloatingPointTy(); }
 
   /// Return true if this is 'label'.
   bool isLabelTy() const { return getTypeID() == LabelTyID; }
 
   /// Return true if this is 'metadata'.
   bool isMetadataTy() const { return getTypeID() == MetadataTyID; }
public:
   bool isFirstClassType() const {
     return getTypeID() != FunctionTyID && getTypeID() != VoidTyID;
   }
 
   /// Return true if the type is a valid type for a register in codegen. This
   /// includes all first-class types except struct and array types.
   bool isSingleValueType() const {
     return isFloatingPointTy() || isX86_MMXTy() || isIntegerTy() ||
-           isPointerTy() || isVectorTy();
+           isPointerTy() || isVectorTy() || isX86_AMXTy();
   }
 
   /// Return true if the type is an aggregate type. This means it is valid as
   /// the first operand of an insertvalue or extractvalue instruction. This
   /// includes struct and array types, but does not include vector types.
   bool isAggregateType() const {
     return getTypeID() == StructTyID || getTypeID() == ArrayTyID;
   }
 
   /// Return true if it makes sense to take the size of this type. To get the
   /// actual size for a particular target, it is reasonable to use the
   /// DataLayout subsystem to do this.
   bool isSized(SmallPtrSetImpl<Type*> *Visited = nullptr) const {
     // If it's a primitive, it is always sized.
     if (getTypeID() == IntegerTyID || isFloatingPointTy() ||
-        getTypeID() == PointerTyID ||
-        getTypeID() == X86_MMXTyID)
+        getTypeID() == PointerTyID || getTypeID() == X86_MMXTyID ||
+        getTypeID() == X86_AMXTyID)
       return true;
     // If it is not something that can have a size (e.g. a function or label),
     // it doesn't have a size.
     if (getTypeID() != StructTyID && getTypeID() != ArrayTyID && !isVectorTy())
       return false;
     // Otherwise we have to try harder to decide.
     return isSizedDerivedType(Visited);
   }
public:
   static Type *getBFloatTy(LLVMContext &C);
   static Type *getFloatTy(LLVMContext &C);
   static Type *getDoubleTy(LLVMContext &C);
   static Type *getMetadataTy(LLVMContext &C);
   static Type *getX86_FP80Ty(LLVMContext &C);
   static Type *getFP128Ty(LLVMContext &C);
   static Type *getPPC_FP128Ty(LLVMContext &C);
   static Type *getX86_MMXTy(LLVMContext &C);
+  static Type *getX86_AMXTy(LLVMContext &C);
   static Type *getTokenTy(LLVMContext &C);
   static IntegerType *getIntNTy(LLVMContext &C, unsigned N);
   static IntegerType *getInt1Ty(LLVMContext &C);
   static IntegerType *getInt8Ty(LLVMContext &C);
   static IntegerType *getInt16Ty(LLVMContext &C);
   static IntegerType *getInt32Ty(LLVMContext &C);
   static IntegerType *getInt64Ty(LLVMContext &C);
   static IntegerType *getInt128Ty(LLVMContext &C);
public:
   static PointerType *getHalfPtrTy(LLVMContext &C, unsigned AS = 0);
   static PointerType *getBFloatPtrTy(LLVMContext &C, unsigned AS = 0);
   static PointerType *getFloatPtrTy(LLVMContext &C, unsigned AS = 0);
   static PointerType *getDoublePtrTy(LLVMContext &C, unsigned AS = 0);
   static PointerType *getX86_FP80PtrTy(LLVMContext &C, unsigned AS = 0);
   static PointerType *getFP128PtrTy(LLVMContext &C, unsigned AS = 0);
   static PointerType *getPPC_FP128PtrTy(LLVMContext &C, unsigned AS = 0);
   static PointerType *getX86_MMXPtrTy(LLVMContext &C, unsigned AS = 0);
+  static PointerType *getX86_AMXPtrTy(LLVMContext &C, unsigned AS = 0);
   static PointerType *getIntNPtrTy(LLVMContext &C, unsigned N, unsigned AS = 0);
   static PointerType *getInt1PtrTy(LLVMContext &C, unsigned AS = 0);
   static PointerType *getInt8PtrTy(LLVMContext &C, unsigned AS = 0);
   static PointerType *getInt16PtrTy(LLVMContext &C, unsigned AS = 0);
   static PointerType *getInt32PtrTy(LLVMContext &C, unsigned AS = 0);
   static PointerType *getInt64PtrTy(LLVMContext &C, unsigned AS = 0);
 
   /// Return a pointer to the current type. This is equivalent to

llvm/include/llvm/Support/MachineValueType.h

namespace llvm {
   public:
     enum SimpleValueType : uint8_t {
       // Simple value types that aren't explicitly part of this enumeration
       // are considered extended value types.
       INVALID_SIMPLE_VALUE_TYPE = 0,
 
       // If you change this numbering, you must change the values in
       // ValueTypes.td as well!
       Other          =   1,   // This is a non-standard value
       i1             =   2,   // This is a 1 bit integer value
       i8             =   3,   // This is an 8 bit integer value
       i16            =   4,   // This is a 16 bit integer value
       i32            =   5,   // This is a 32 bit integer value
       i64            =   6,   // This is a 64 bit integer value
       i128           =   7,   // This is a 128 bit integer value
 
       FIRST_INTEGER_VALUETYPE = i1,
       LAST_INTEGER_VALUETYPE  = i128,
 
       bf16           =   8,   // This is a 16 bit brain floating point value
       f16            =   9,   // This is a 16 bit floating point value
       f32            =  10,   // This is a 32 bit floating point value
       f64            =  11,   // This is a 64 bit floating point value
       f80            =  12,   // This is a 80 bit floating point value
       f128           =  13,   // This is a 128 bit floating point value
       ppcf128        =  14,   // This is a PPC 128-bit floating point value
 
       FIRST_FP_VALUETYPE = bf16,
       LAST_FP_VALUETYPE  = ppcf128,
 
       v1i1           =  15,   //    1 x i1
       v2i1           =  16,   //    2 x i1
       v4i1           =  17,   //    4 x i1
       v8i1           =  18,   //    8 x i1
       v16i1          =  19,   //   16 x i1
       v32i1          =  20,   //   32 x i1
       v64i1          =  21,   //   64 x i1
enum SimpleValueType : uint8_t {
       v128i64        =  65,   // 128 x i64
       v256i64        =  66,   // 256 x i64
 
       v1i128         =  67,   //  1 x i128
 
       FIRST_INTEGER_FIXEDLEN_VECTOR_VALUETYPE = v1i1,
       LAST_INTEGER_FIXEDLEN_VECTOR_VALUETYPE = v1i128,
 
       v2f16          =  68,   //    2 x f16
       v3f16          =  69,   //    3 x f16
       v4f16          =  70,   //    4 x f16
       v8f16          =  71,   //    8 x f16
       v16f16         =  72,   //   16 x f16
       v32f16         =  73,   //   32 x f16
       v64f16         =  74,   //   64 x f16
       v128f16        =  75,   //  128 x f16
       v2bf16         =  76,   //    2 x bf16
enum SimpleValueType : uint8_t {
       v256f64        = 106,   //  256 x f64
 
       FIRST_FP_FIXEDLEN_VECTOR_VALUETYPE = v2f16,
       LAST_FP_FIXEDLEN_VECTOR_VALUETYPE = v256f64,
 
       FIRST_FIXEDLEN_VECTOR_VALUETYPE = v1i1,
       LAST_FIXEDLEN_VECTOR_VALUETYPE = v256f64,
 
       nxv1i1         = 107,   // n x  1 x i1
       nxv2i1         = 108,   // n x  2 x i1
       nxv4i1         = 109,   // n x  4 x i1
       nxv8i1         = 110,   // n x  8 x i1
       nxv16i1        = 111,   // n x 16 x i1
       nxv32i1        = 112,   // n x 32 x i1
       nxv64i1        = 113,   // n x  64 x i1
 
       nxv1i8         = 114,   // n x  1 x i8
enum SimpleValueType : uint8_t {
       nxv4i64        = 135,  // n x  4 x i64
       nxv8i64        = 136,  // n x  8 x i64
       nxv16i64       = 137,  // n x 16 x i64
       nxv32i64       = 138,  // n x 32 x i64
 
       FIRST_INTEGER_SCALABLE_VECTOR_VALUETYPE = nxv1i1,
       LAST_INTEGER_SCALABLE_VECTOR_VALUETYPE = nxv32i64,
 
       nxv1f16        = 139,   // n x   1 x f16
       nxv2f16        = 140,  // n x  2 x f16
       nxv4f16        = 141,  // n x  4 x f16
       nxv8f16        = 142,  // n x  8 x f16
       nxv16f16       = 143,   // n x  16 x f16
       nxv32f16       = 144,   // n x  32 x f16
       nxv2bf16       = 145,  // n x  2 x bf16
       nxv4bf16       = 146,  // n x  4 x bf16
       nxv8bf16       = 147,  // n x  8 x bf16
enum SimpleValueType : uint8_t {
 
       FIRST_FP_SCALABLE_VECTOR_VALUETYPE = nxv1f16,
       LAST_FP_SCALABLE_VECTOR_VALUETYPE = nxv8f64,
 
       FIRST_SCALABLE_VECTOR_VALUETYPE = nxv1i1,
       LAST_SCALABLE_VECTOR_VALUETYPE = nxv8f64,
 
       FIRST_VECTOR_VALUETYPE = v1i1,
       LAST_VECTOR_VALUETYPE  = nxv8f64,
 
       x86mmx         = 157,   // This is an X86 MMX value
 
       Glue           = 158,   // This glues nodes together during pre-RA sched
 
       isVoid         = 159,   // This has no value
 
       Untyped        = 160,   // This value takes a register, but has
                               // unspecified type.  The register class
                               // will be determined by the opcode.
 
       exnref         = 161,   // WebAssembly's exnref type
       funcref        = 162,   // WebAssembly's funcref type
       externref      = 163,   // WebAssembly's externref type
+      x86amx         = 164,   // This is an X86 AMX value
 
       FIRST_VALUETYPE =  1,   // This is always the beginning of the list.
-      LAST_VALUETYPE = 164,   // This always remains at the end of the list.
+      LAST_VALUETYPE = 165,   // This always remains at the end of the list.
 
       // This is the current maximum for LAST_VALUETYPE.
       // MVT::MAX_ALLOWED_VALUETYPE is used for asserts and to size bit vectors
       // This value must be a multiple of 32.
       MAX_ALLOWED_VALUETYPE = 192,
 
       // A value of type llvm::TokenTy
       token          = 248,
 
       // This is MDNode or MDString.
       Metadata       = 249,
 
       // An int value the size of the pointer of the current
       // target to any address space. This must only be used internal to
       // tblgen. Other than for overloading, we treat iPTRAny the same as iPTR.
       iPTRAny        = 250,
 
       // A vector with any length and element size. This is used
       // for intrinsics that have overloadings based on vector types.
       // This is only for tblgen's consumption!
       vAny           = 251,
 
       // Any floating-point or vector floating-point value. This is used
       // for intrinsics that have overloadings based on floating-point types.
       // This is only for tblgen's consumption!
       fAny           = 252,
 
       // An integer or vector integer value of any bit width. This is
       // used for intrinsics that have overloadings based on integer bit widths.
       // This is only for tblgen's consumption!
       iAny           = 253,
 
       // An int value the size of the pointer of the current
       // target.  This should only be used internal to tblgen!
       iPTR           = 254,
 
       // Any type. This is used for intrinsics that have overloadings.
       // This is only for tblgen's consumption!
       Any            = 255
     };
 
     SimpleValueType SimpleTy = INVALID_SIMPLE_VALUE_TYPE;
 
     constexpr MVT() = default;
     constexpr MVT(SimpleValueType SVT) : SimpleTy(SVT) {}
 
     bool operator>(const MVT& S)  const { return SimpleTy >  S.SimpleTy; }
TypeSize getSizeInBits() const {
       case nxv32i64: return TypeSize::Scalable(2048);
       case v128i32:
       case v64i64:
       case v128f32:
       case v64f64:  return TypeSize::Fixed(4096);
       case v256i32:
       case v128i64:
       case v256f32:
+      case x86amx:
       case v128f64:  return TypeSize::Fixed(8192);
       case v512i32:
       case v256i64:
       case v512f32:
       case v256f64:  return TypeSize::Fixed(16384);
       case v1024i32:
       case v1024f32:  return TypeSize::Fixed(32768);
       case v2048i32:

llvm/lib/Analysis/ConstantFolding.cpp

 /// Constant fold bitcast, symbolically evaluating it with DataLayout.
 /// This always returns a non-null constant, but it may be a
 /// ConstantExpr if unfoldable.
 Constant *FoldBitCast(Constant *C, Type *DestTy, const DataLayout &DL) {
   assert(CastInst::castIsValid(Instruction::BitCast, C, DestTy) &&
          "Invalid constantexpr bitcast!");
 
   // Catch the obvious splat cases.
-  if (C->isNullValue() && !DestTy->isX86_MMXTy())
+  if (C->isNullValue() && !DestTy->isX86_MMXTy() && !DestTy->isX86_AMXTy())
     return Constant::getNullValue(DestTy);
-  if (C->isAllOnesValue() && !DestTy->isX86_MMXTy() &&
+  if (C->isAllOnesValue() && !DestTy->isX86_MMXTy() && !DestTy->isX86_AMXTy() &&
       !DestTy->isPtrOrPtrVectorTy()) // Don't get ones for ptr types!
     return Constant::getAllOnesValue(DestTy);
 
   if (auto *VTy = dyn_cast<VectorType>(C->getType())) {
     // Handle a vector->scalar integer/fp cast.
     if (isa<IntegerType>(DestTy) || DestTy->isFloatingPointTy()) {
       unsigned NumSrcElts = cast<FixedVectorType>(VTy)->getNumElements();
       Type *SrcEltTy = VTy->getElementType();
do {
     Type *SrcTy = C->getType();
     uint64_t DestSize = DL.getTypeSizeInBits(DestTy);
     uint64_t SrcSize = DL.getTypeSizeInBits(SrcTy);
     if (SrcSize < DestSize)
       return nullptr;
 
     // Catch the obvious splat cases (since all-zeros can coerce non-integral
     // pointers legally).
-    if (C->isNullValue() && !DestTy->isX86_MMXTy())
+    if (C->isNullValue() && !DestTy->isX86_MMXTy() && !DestTy->isX86_AMXTy())
       return Constant::getNullValue(DestTy);
     if (C->isAllOnesValue() &&
         (DestTy->isIntegerTy() || DestTy->isFloatingPointTy() ||
          DestTy->isVectorTy()) &&
-        !DestTy->isX86_MMXTy() && !DestTy->isPtrOrPtrVectorTy())
+        !DestTy->isX86_AMXTy() && !DestTy->isX86_MMXTy() &&
+        !DestTy->isPtrOrPtrVectorTy())
       // Get ones when the input is trivial, but
       // only for supported types inside getAllOnesValue.
       return Constant::getAllOnesValue(DestTy);
 
     // If the type sizes are the same and a cast is legal, just directly
     // cast the constant.
     // But be careful not to coerce non-integral pointers illegally.
     if (SrcSize == DestSize &&
if (!IntType) {
     else if (LoadTy->isVectorTy()) {
       MapTy = PointerType::getIntNTy(
           C->getContext(), DL.getTypeSizeInBits(LoadTy).getFixedSize());
     } else
       return nullptr;
 
     C = FoldBitCast(C, MapTy->getPointerTo(AS), DL);
     if (Constant *Res = FoldReinterpretLoadFromConstPtr(C, MapTy, DL)) {
-      if (Res->isNullValue() && !LoadTy->isX86_MMXTy())
+      if (Res->isNullValue() && !LoadTy->isX86_MMXTy() &&
+          !LoadTy->isX86_AMXTy())
         // Materializing a zero can be done trivially without a bitcast
         return Constant::getNullValue(LoadTy);
       Type *CastTy = LoadTy->isPtrOrPtrVectorTy() ? DL.getIntPtrType(LoadTy) : LoadTy;
       Res = FoldBitCast(Res, CastTy, DL);
       if (LoadTy->isPtrOrPtrVectorTy()) {
         // For vector of pointer, we needed to first convert to a vector of integer, then do vector inttoptr
-        if (Res->isNullValue() && !LoadTy->isX86_MMXTy())
+        if (Res->isNullValue() && !LoadTy->isX86_MMXTy() &&
+            !LoadTy->isX86_AMXTy())
           return Constant::getNullValue(LoadTy);
         if (DL.isNonIntegralPointerType(LoadTy->getScalarType()))
           // Be careful not to replace a load of an addrspace value with an inttoptr here
           return nullptr;
         Res = ConstantExpr::getCast(Instruction::IntToPtr, Res, LoadTy);
       }
       return Res;
     }

llvm/lib/AsmParser/LLLexer.cpp

#define TYPEKEYWORD(STR, LLVMTY)                                               \
   TYPEKEYWORD("float",     Type::getFloatTy(Context));
   TYPEKEYWORD("double",    Type::getDoubleTy(Context));
   TYPEKEYWORD("x86_fp80",  Type::getX86_FP80Ty(Context));
   TYPEKEYWORD("fp128",     Type::getFP128Ty(Context));
   TYPEKEYWORD("ppc_fp128", Type::getPPC_FP128Ty(Context));
   TYPEKEYWORD("label",     Type::getLabelTy(Context));
   TYPEKEYWORD("metadata",  Type::getMetadataTy(Context));
   TYPEKEYWORD("x86_mmx",   Type::getX86_MMXTy(Context));
+  TYPEKEYWORD("x86_amx",   Type::getX86_AMXTy(Context));
   TYPEKEYWORD("token",     Type::getTokenTy(Context));
 
 #undef TYPEKEYWORD
 
   // Keywords for instructions.
 #define INSTKEYWORD(STR, Enum)                                                 \
   do {                                                                         \
     if (Keyword == #STR) {                                                     \

llvm/lib/Bitcode/Reader/BitcodeReader.cpp

case bitc::TYPE_CODE_LABEL:     // LABEL
       ResultTy = Type::getLabelTy(Context);
       break;
     case bitc::TYPE_CODE_METADATA:  // METADATA
       ResultTy = Type::getMetadataTy(Context);
       break;
     case bitc::TYPE_CODE_X86_MMX:   // X86_MMX
       ResultTy = Type::getX86_MMXTy(Context);
       break;
+    case bitc::TYPE_CODE_X86_AMX:   // X86_AMX
+      ResultTy = Type::getX86_AMXTy(Context);
+      break;
     case bitc::TYPE_CODE_TOKEN:     // TOKEN
       ResultTy = Type::getTokenTy(Context);
       break;
     case bitc::TYPE_CODE_INTEGER: { // INTEGER: [width]
       if (Record.empty())
         return error("Invalid record");
 
       uint64_t NumBits = Record[0];

llvm/lib/Bitcode/Writer/BitcodeWriter.cpp

for (unsigned i = 0, e = TypeList.size(); i != e; ++i) {
     case Type::FloatTyID:     Code = bitc::TYPE_CODE_FLOAT;     break;
     case Type::DoubleTyID:    Code = bitc::TYPE_CODE_DOUBLE;    break;
     case Type::X86_FP80TyID:  Code = bitc::TYPE_CODE_X86_FP80;  break;
     case Type::FP128TyID:     Code = bitc::TYPE_CODE_FP128;     break;
     case Type::PPC_FP128TyID: Code = bitc::TYPE_CODE_PPC_FP128; break;
     case Type::LabelTyID:     Code = bitc::TYPE_CODE_LABEL;     break;
     case Type::MetadataTyID:  Code = bitc::TYPE_CODE_METADATA;  break;
     case Type::X86_MMXTyID:   Code = bitc::TYPE_CODE_X86_MMX;   break;
+    case Type::X86_AMXTyID:   Code = bitc::TYPE_CODE_X86_AMX;   break;
     case Type::TokenTyID:     Code = bitc::TYPE_CODE_TOKEN;     break;
     case Type::IntegerTyID:
       // INTEGER: [width]
       Code = bitc::TYPE_CODE_INTEGER;
       TypeVals.push_back(cast<IntegerType>(T)->getBitWidth());
       break;
     case Type::PointerTyID: {
       PointerType *PTy = cast<PointerType>(T);

llvm/lib/CodeGen/ValueTypes.cpp

if (isFloatingPoint())
       return "f" + utostr(getSizeInBits());
     llvm_unreachable("Invalid EVT!");
   case MVT::bf16:      return "bf16";
   case MVT::ppcf128:   return "ppcf128";
   case MVT::isVoid:    return "isVoid";
   case MVT::Other:     return "ch";
   case MVT::Glue:      return "glue";
   case MVT::x86mmx:    return "x86mmx";
+  case MVT::x86amx:    return "x86amx";
   case MVT::Metadata:  return "Metadata";
   case MVT::Untyped:   return "Untyped";
   case MVT::exnref:    return "exnref";
   case MVT::funcref:   return "funcref";
   case MVT::externref: return "externref";
   }
 }
 
Type *EVT::getTypeForEVT(LLVMContext &Context) const {
   case MVT::f16:     return Type::getHalfTy(Context);
   case MVT::bf16:     return Type::getBFloatTy(Context);
   case MVT::f32:     return Type::getFloatTy(Context);
   case MVT::f64:     return Type::getDoubleTy(Context);
   case MVT::f80:     return Type::getX86_FP80Ty(Context);
   case MVT::f128:    return Type::getFP128Ty(Context);
   case MVT::ppcf128: return Type::getPPC_FP128Ty(Context);
   case MVT::x86mmx:  return Type::getX86_MMXTy(Context);
+  case MVT::x86amx:  return Type::getX86_AMXTy(Context);
   case MVT::v1i1:
     return FixedVectorType::get(Type::getInt1Ty(Context), 1);
   case MVT::v2i1:
     return FixedVectorType::get(Type::getInt1Ty(Context), 2);
   case MVT::v4i1:
     return FixedVectorType::get(Type::getInt1Ty(Context), 4);
   case MVT::v8i1:
     return FixedVectorType::get(Type::getInt1Ty(Context), 8);
MVT MVT::getVT(Type *Ty, bool HandleUnknown){
   case Type::IntegerTyID:
     return getIntegerVT(cast<IntegerType>(Ty)->getBitWidth());
   case Type::HalfTyID:      return MVT(MVT::f16);
   case Type::BFloatTyID:    return MVT(MVT::bf16);
   case Type::FloatTyID:     return MVT(MVT::f32);
   case Type::DoubleTyID:    return MVT(MVT::f64);
   case Type::X86_FP80TyID:  return MVT(MVT::f80);
   case Type::X86_MMXTyID:   return MVT(MVT::x86mmx);
+  case Type::X86_AMXTyID:   return MVT(MVT::x86amx);
   case Type::FP128TyID:     return MVT(MVT::f128);
   case Type::PPC_FP128TyID: return MVT(MVT::ppcf128);
   case Type::PointerTyID:   return MVT(MVT::iPTR);
   case Type::FixedVectorTyID:
   case Type::ScalableVectorTyID: {
     VectorType *VTy = cast<VectorType>(Ty);
     return getVectorVT(
       getVT(VTy->getElementType(), /*HandleUnknown=*/ false),

llvm/lib/IR/AsmWriter.cpp

void TypePrinting::print(Type *Ty, raw_ostream &OS) {
   case Type::FloatTyID:     OS << "float"; return;
   case Type::DoubleTyID:    OS << "double"; return;
   case Type::X86_FP80TyID:  OS << "x86_fp80"; return;
   case Type::FP128TyID:     OS << "fp128"; return;
   case Type::PPC_FP128TyID: OS << "ppc_fp128"; return;
   case Type::LabelTyID:     OS << "label"; return;
   case Type::MetadataTyID:  OS << "metadata"; return;
   case Type::X86_MMXTyID:   OS << "x86_mmx"; return;
+  case Type::X86_AMXTyID:   OS << "x86_amx"; return;
   case Type::TokenTyID:     OS << "token"; return;
   case Type::IntegerTyID:
     OS << 'i' << cast<IntegerType>(Ty)->getBitWidth();
     return;
 
   case Type::FunctionTyID: {
     FunctionType *FTy = cast<FunctionType>(Ty);
     print(FTy->getReturnType(), OS);

llvm/lib/IR/ConstantFold.cpp

if (isa<UndefValue>(V)) {
     // sext(undef) = 0, because the top bits will all be the same.
     // [us]itofp(undef) = 0, because the result value is bounded.
     if (opc == Instruction::ZExt || opc == Instruction::SExt ||
         opc == Instruction::UIToFP || opc == Instruction::SIToFP)
       return Constant::getNullValue(DestTy);
     return UndefValue::get(DestTy);
   }
 
-  if (V->isNullValue() && !DestTy->isX86_MMXTy() &&
+  if (V->isNullValue() && !DestTy->isX86_MMXTy() && !DestTy->isX86_AMXTy() &&
       opc != Instruction::AddrSpaceCast)
     return Constant::getNullValue(DestTy);
 
   // If the cast operand is a constant expression, there's a few things we can
   // do to try to simplify it.
   if (ConstantExpr *CE = dyn_cast<ConstantExpr>(V)) {
     if (CE->isCast()) {
       // Try hard to fold cast of cast because they are often eliminable.
       if (unsigned newOpc = foldConstantCastPair(opc, CE, DestTy))
         return ConstantExpr::getCast(newOpc, CE->getOperand(0), DestTy);

llvm/lib/IR/Core.cpp

LLVMTypeKind LLVMGetTypeKind(LLVMTypeRef Ty) {
   case Type::ArrayTyID:
     return LLVMArrayTypeKind;
   case Type::PointerTyID:
     return LLVMPointerTypeKind;
   case Type::FixedVectorTyID:
     return LLVMVectorTypeKind;
   case Type::X86_MMXTyID:
     return LLVMX86_MMXTypeKind;
+  case Type::X86_AMXTyID:
+    return LLVMX86_AMXTypeKind;
   case Type::TokenTyID:
     return LLVMTokenTypeKind;
   case Type::ScalableVectorTyID:
     return LLVMScalableVectorTypeKind;
   }
   llvm_unreachable("Unhandled TypeID.");
 }
 
LLVMTypeRef LLVMFP128TypeInContext(LLVMContextRef C) {
   return (LLVMTypeRef) Type::getFP128Ty(*unwrap(C));
 }
 LLVMTypeRef LLVMPPCFP128TypeInContext(LLVMContextRef C) {
   return (LLVMTypeRef) Type::getPPC_FP128Ty(*unwrap(C));
 }
 LLVMTypeRef LLVMX86MMXTypeInContext(LLVMContextRef C) {
   return (LLVMTypeRef) Type::getX86_MMXTy(*unwrap(C));
 }
+LLVMTypeRef LLVMX86AMXTypeInContext(LLVMContextRef C) {
+  return (LLVMTypeRef) Type::getX86_AMXTy(*unwrap(C));
+}
 
 LLVMTypeRef LLVMHalfType(void) {
   return LLVMHalfTypeInContext(LLVMGetGlobalContext());
 }
 LLVMTypeRef LLVMBFloatType(void) {
   return LLVMBFloatTypeInContext(LLVMGetGlobalContext());
 }
 LLVMTypeRef LLVMFloatType(void) {
LLVMTypeRef LLVMFP128Type(void) {
   return LLVMFP128TypeInContext(LLVMGetGlobalContext());
 }
 LLVMTypeRef LLVMPPCFP128Type(void) {
   return LLVMPPCFP128TypeInContext(LLVMGetGlobalContext());
 }
 LLVMTypeRef LLVMX86MMXType(void) {
   return LLVMX86MMXTypeInContext(LLVMGetGlobalContext());
 }
+LLVMTypeRef LLVMX86AMXType(void) {
+  return LLVMX86AMXTypeInContext(LLVMGetGlobalContext());
+}
 
 /*--.. Operations on function types ........................................--*/
 
 LLVMTypeRef LLVMFunctionType(LLVMTypeRef ReturnType,
                              LLVMTypeRef *ParamTypes, unsigned ParamCount,
                              LLVMBool IsVarArg) {
   ArrayRef<Type*> Tys(unwrap(ParamTypes), ParamCount);
   return wrap(FunctionType::get(unwrap(ReturnType), Tys, IsVarArg != 0));

llvm/lib/IR/DataLayout.cpp

case Type::X86_FP80TyID: {
     // less conservative, they should have specified it explicitly in the data
     // layout.
     return Align(PowerOf2Ceil(BitWidth / 8));
   }
   case Type::X86_MMXTyID:
   case Type::FixedVectorTyID:
   case Type::ScalableVectorTyID: {
     unsigned BitWidth = getTypeSizeInBits(Ty).getKnownMinSize();
     auto I = findAlignmentLowerBound(VECTOR_ALIGN, BitWidth);
     if (I != Alignments.end() && I->AlignType == VECTOR_ALIGN &&
         I->TypeBitWidth == BitWidth)
       return abi_or_pref ? I->ABIAlign : I->PrefAlign;
 
     // By default, use natural alignment for vector types. This is consistent
     // with what clang and llvm-gcc do.
     // TODO: This should probably not be using the alloc size.
     unsigned Alignment =
         getTypeAllocSize(cast<VectorType>(Ty)->getElementType());
     // We're only calculating a natural alignment, so it doesn't have to be
     // based on the full size for scalable vectors. Using the minimum element
     // count should be enough here.
     Alignment *= cast<VectorType>(Ty)->getElementCount().getKnownMinValue();
     Alignment = PowerOf2Ceil(Alignment);
     return Align(Alignment);
   }
+  case Type::X86_AMXTyID:
+    return Align(512);
   default:
     llvm_unreachable("Bad type for getAlignment!!!");
   }
 }
 
 /// TODO: Remove this function once the transition to Align is over.
 unsigned DataLayout::getABITypeAlignment(Type *Ty) const {
   return getABITypeAlign(Ty).value();

llvm/lib/IR/Function.cpp

if (PointerType* PTyp = dyn_cast<PointerType>(Ty)) {
     case Type::HalfTyID:      Result += "f16";      break;
     case Type::BFloatTyID:    Result += "bf16";     break;
     case Type::FloatTyID:     Result += "f32";      break;
     case Type::DoubleTyID:    Result += "f64";      break;
     case Type::X86_FP80TyID:  Result += "f80";      break;
     case Type::FP128TyID:     Result += "f128";     break;
     case Type::PPC_FP128TyID: Result += "ppcf128";  break;
     case Type::X86_MMXTyID:   Result += "x86mmx";   break;
+    case Type::X86_AMXTyID:   Result += "x86amx";   break;
     case Type::IntegerTyID:
       Result += "i" + utostr(cast<IntegerType>(Ty)->getBitWidth());
       break;
     }
   }
   return Result;
 }
 

 
 /// IIT_Info - These are enumerators that describe the entries returned by the
 /// getIntrinsicInfoTableEntries function.
 ///
 /// NOTE: This must be kept in synch with the copy in TblGen/IntrinsicEmitter!
 enum IIT_Info {
   // Common values should be encoded with 0-15.
   IIT_Done = 0,
   IIT_I1   = 1,
   IIT_I8   = 2,
   IIT_I16  = 3,
   IIT_I32  = 4,
   IIT_I64  = 5,
   IIT_F16  = 6,
   IIT_F32  = 7,
   IIT_F64  = 8,
   IIT_V2   = 9,
   IIT_V4   = 10,
   IIT_V8   = 11,
   IIT_V16  = 12,
   IIT_V32  = 13,
   IIT_PTR  = 14,
   IIT_ARG  = 15,
 
   // Values from 16+ are only encodable with the inefficient encoding.
   IIT_V64  = 16,
   IIT_MMX  = 17,
   IIT_TOKEN = 18,
   IIT_METADATA = 19,
   IIT_EMPTYSTRUCT = 20,
   IIT_STRUCT2 = 21,
   IIT_STRUCT3 = 22,
   IIT_STRUCT4 = 23,
   IIT_STRUCT5 = 24,
   IIT_EXTEND_ARG = 25,
   IIT_TRUNC_ARG = 26,
   IIT_ANYPTR = 27,
   IIT_V1   = 28,
   IIT_VARARG = 29,
   IIT_HALF_VEC_ARG = 30,
   IIT_SAME_VEC_WIDTH_ARG = 31,
   IIT_PTR_TO_ARG = 32,
   IIT_PTR_TO_ELT = 33,
   IIT_VEC_OF_ANYPTRS_TO_ELT = 34,
   IIT_I128 = 35,
   IIT_V512 = 36,
   IIT_V1024 = 37,
   IIT_STRUCT6 = 38,
   IIT_STRUCT7 = 39,
   IIT_STRUCT8 = 40,
   IIT_F128 = 41,
   IIT_VEC_ELEMENT = 42,
   IIT_SCALABLE_VEC = 43,
   IIT_SUBDIVIDE2_ARG = 44,
   IIT_SUBDIVIDE4_ARG = 45,
   IIT_VEC_OF_BITCASTS_TO_INT = 46,
   IIT_V128 = 47,
   IIT_BF16 = 48,
   IIT_STRUCT9 = 49,
-  IIT_V256 = 50
+  IIT_V256 = 50,
+  IIT_AMX  = 51
 };
 
 static void DecodeIITType(unsigned &NextElt, ArrayRef<unsigned char> Infos,
                       IIT_Info LastInfo,
                       SmallVectorImpl<Intrinsic::IITDescriptor> &OutputTable) {
   using namespace Intrinsic;
 
   bool IsScalableVector = (LastInfo == IIT_SCALABLE_VEC);
 
   IIT_Info Info = IIT_Info(Infos[NextElt++]);
   unsigned StructElts = 2;
 
   switch (Info) {
   case IIT_Done:
     OutputTable.push_back(IITDescriptor::get(IITDescriptor::Void, 0));
     return;
   case IIT_VARARG:
     OutputTable.push_back(IITDescriptor::get(IITDescriptor::VarArg, 0));
     return;
   case IIT_MMX:
     OutputTable.push_back(IITDescriptor::get(IITDescriptor::MMX, 0));
     return;
+  case IIT_AMX:
+    OutputTable.push_back(IITDescriptor::get(IITDescriptor::AMX, 0));
+    return;
   case IIT_TOKEN:
     OutputTable.push_back(IITDescriptor::get(IITDescriptor::Token, 0));
     return;
   case IIT_METADATA:
     OutputTable.push_back(IITDescriptor::get(IITDescriptor::Metadata, 0));
     return;
   case IIT_F16:
     OutputTable.push_back(IITDescriptor::get(IITDescriptor::Half, 0));
static Type *DecodeFixedType(ArrayRef<Intrinsic::IITDescriptor> &Infos,
 
   IITDescriptor D = Infos.front();
   Infos = Infos.slice(1);
 
   switch (D.Kind) {
   case IITDescriptor::Void: return Type::getVoidTy(Context);
   case IITDescriptor::VarArg: return Type::getVoidTy(Context);
   case IITDescriptor::MMX: return Type::getX86_MMXTy(Context);
+  case IITDescriptor::AMX: return Type::getX86_AMXTy(Context);
   case IITDescriptor::Token: return Type::getTokenTy(Context);
   case IITDescriptor::Metadata: return Type::getMetadataTy(Context);
   case IITDescriptor::Half: return Type::getHalfTy(Context);
   case IITDescriptor::BFloat: return Type::getBFloatTy(Context);
   case IITDescriptor::Float: return Type::getFloatTy(Context);
   case IITDescriptor::Double: return Type::getDoubleTy(Context);
   case IITDescriptor::Quad: return Type::getFP128Ty(Context);
 
static bool matchIntrinsicType(
 
   IITDescriptor D = Infos.front();
   Infos = Infos.slice(1);
 
   switch (D.Kind) {
     case IITDescriptor::Void: return !Ty->isVoidTy();
     case IITDescriptor::VarArg: return true;
     case IITDescriptor::MMX:  return !Ty->isX86_MMXTy();
+    case IITDescriptor::AMX:  return !Ty->isX86_AMXTy();
     case IITDescriptor::Token: return !Ty->isTokenTy();
     case IITDescriptor::Metadata: return !Ty->isMetadataTy();
     case IITDescriptor::Half: return !Ty->isHalfTy();
     case IITDescriptor::BFloat: return !Ty->isBFloatTy();
     case IITDescriptor::Float: return !Ty->isFloatTy();
     case IITDescriptor::Double: return !Ty->isDoubleTy();
     case IITDescriptor::Quad: return !Ty->isFP128Ty();
     case IITDescriptor::Integer: return !Ty->isIntegerTy(D.Integer_Width);

llvm/lib/IR/LLVMContextImpl.h

#include "llvm/IR/Metadata.def"
   ConstantInt *TheTrueVal = nullptr;
   ConstantInt *TheFalseVal = nullptr;
 
   std::unique_ptr<ConstantTokenNone> TheNoneToken;
 
   // Basic type instances.
   Type VoidTy, LabelTy, HalfTy, BFloatTy, FloatTy, DoubleTy, MetadataTy,
       TokenTy;
-  Type X86_FP80Ty, FP128Ty, PPC_FP128Ty, X86_MMXTy;
+  Type X86_FP80Ty, FP128Ty, PPC_FP128Ty, X86_MMXTy, X86_AMXTy;
   IntegerType Int1Ty, Int8Ty, Int16Ty, Int32Ty, Int64Ty, Int128Ty;
 
   BumpPtrAllocator Alloc;
   UniqueStringSaver Saver{Alloc};
 
   DenseMap<unsigned, IntegerType*> IntegerTypes;
 
   using FunctionTypeSet = DenseSet<FunctionType *, FunctionTypeKeyInfo>;

llvm/lib/IR/LLVMContextImpl.cpp

 #include "llvm/IR/Type.h"
 #include "llvm/Support/ManagedStatic.h"
 #include <cassert>
 #include <utility>
 
 using namespace llvm;
 
 LLVMContextImpl::LLVMContextImpl(LLVMContext &C)
   : DiagHandler(std::make_unique<DiagnosticHandler>()),
     VoidTy(C, Type::VoidTyID),
     LabelTy(C, Type::LabelTyID),
     HalfTy(C, Type::HalfTyID),
     BFloatTy(C, Type::BFloatTyID),
     FloatTy(C, Type::FloatTyID),
     DoubleTy(C, Type::DoubleTyID),
     MetadataTy(C, Type::MetadataTyID),
     TokenTy(C, Type::TokenTyID),
     X86_FP80Ty(C, Type::X86_FP80TyID),
     FP128Ty(C, Type::FP128TyID),
     PPC_FP128Ty(C, Type::PPC_FP128TyID),
     X86_MMXTy(C, Type::X86_MMXTyID),
+    X86_AMXTy(C, Type::X86_AMXTyID),
     Int1Ty(C, 1),
     Int8Ty(C, 8),
     Int16Ty(C, 16),
     Int32Ty(C, 32),
     Int64Ty(C, 64),
     Int128Ty(C, 128) {}
 
 LLVMContextImpl::~LLVMContextImpl() {

llvm/lib/IR/Type.cpp

Type *Type::getPrimitiveType(LLVMContext &C, TypeID IDNumber) {
   case FloatTyID     : return getFloatTy(C);
   case DoubleTyID    : return getDoubleTy(C);
   case X86_FP80TyID  : return getX86_FP80Ty(C);
   case FP128TyID     : return getFP128Ty(C);
   case PPC_FP128TyID : return getPPC_FP128Ty(C);
   case LabelTyID     : return getLabelTy(C);
   case MetadataTyID  : return getMetadataTy(C);
   case X86_MMXTyID   : return getX86_MMXTy(C);
+  case X86_AMXTyID   : return getX86_AMXTy(C);
   case TokenTyID     : return getTokenTy(C);
   default:
     return nullptr;
   }
 }
 
 bool Type::isIntegerTy(unsigned Bitwidth) const {
   return isIntegerTy() && cast<IntegerType>(this)->getBitWidth() == Bitwidth;
bool Type::canLosslesslyBitCastTo(Type *Ty) const {
   //  64-bit fixed width vector types can be losslessly converted to x86mmx.
   if (((isa<FixedVectorType>(this)) && Ty->isX86_MMXTy()) &&
       getPrimitiveSizeInBits().getFixedSize() == 64)
     return true;
   if ((isX86_MMXTy() && isa<FixedVectorType>(Ty)) &&
       Ty->getPrimitiveSizeInBits().getFixedSize() == 64)
     return true;
 
+  //  8192-bit fixed width vector types can be losslessly converted to x86amx.
+  if (((isa<FixedVectorType>(this)) && Ty->isX86_AMXTy()) &&
+      getPrimitiveSizeInBits().getFixedSize() == 8192)
+    return true;
+  if ((isX86_AMXTy() && isa<FixedVectorType>(Ty)) &&
+      Ty->getPrimitiveSizeInBits().getFixedSize() == 8192)
+    return true;
+
   // At this point we have only various mismatches of the first class types
   // remaining and ptr->ptr. Just select the lossless conversions. Everything
   // else is not lossless. Conservatively assume we can't losslessly convert
   // between pointers with different address spaces.
   if (auto *PTy = dyn_cast<PointerType>(this)) {
     if (auto *OtherPTy = dyn_cast<PointerType>(Ty))
       return PTy->getAddressSpace() == OtherPTy->getAddressSpace();
     return false;
TypeSize Type::getPrimitiveSizeInBits() const {
   case Type::HalfTyID: return TypeSize::Fixed(16);
   case Type::BFloatTyID: return TypeSize::Fixed(16);
   case Type::FloatTyID: return TypeSize::Fixed(32);
   case Type::DoubleTyID: return TypeSize::Fixed(64);
   case Type::X86_FP80TyID: return TypeSize::Fixed(80);
   case Type::FP128TyID: return TypeSize::Fixed(128);
   case Type::PPC_FP128TyID: return TypeSize::Fixed(128);
   case Type::X86_MMXTyID: return TypeSize::Fixed(64);
+  case Type::X86_AMXTyID: return TypeSize::Fixed(8192);
   case Type::IntegerTyID:
     return TypeSize::Fixed(cast<IntegerType>(this)->getBitWidth());
   case Type::FixedVectorTyID:
   case Type::ScalableVectorTyID: {
     const VectorType *VTy = cast<VectorType>(this);
     ElementCount EC = VTy->getElementCount();
     TypeSize ETS = VTy->getElementType()->getPrimitiveSizeInBits();
     assert(!ETS.isScalable() && "Vector type should have fixed-width elements");

 Type *Type::getFloatTy(LLVMContext &C) { return &C.pImpl->FloatTy; }
 Type *Type::getDoubleTy(LLVMContext &C) { return &C.pImpl->DoubleTy; }
 Type *Type::getMetadataTy(LLVMContext &C) { return &C.pImpl->MetadataTy; }
 Type *Type::getTokenTy(LLVMContext &C) { return &C.pImpl->TokenTy; }
 Type *Type::getX86_FP80Ty(LLVMContext &C) { return &C.pImpl->X86_FP80Ty; }
 Type *Type::getFP128Ty(LLVMContext &C) { return &C.pImpl->FP128Ty; }
 Type *Type::getPPC_FP128Ty(LLVMContext &C) { return &C.pImpl->PPC_FP128Ty; }
 Type *Type::getX86_MMXTy(LLVMContext &C) { return &C.pImpl->X86_MMXTy; }
+Type *Type::getX86_AMXTy(LLVMContext &C) { return &C.pImpl->X86_AMXTy; }
 
 IntegerType *Type::getInt1Ty(LLVMContext &C) { return &C.pImpl->Int1Ty; }
 IntegerType *Type::getInt8Ty(LLVMContext &C) { return &C.pImpl->Int8Ty; }
 IntegerType *Type::getInt16Ty(LLVMContext &C) { return &C.pImpl->Int16Ty; }
 IntegerType *Type::getInt32Ty(LLVMContext &C) { return &C.pImpl->Int32Ty; }
 IntegerType *Type::getInt64Ty(LLVMContext &C) { return &C.pImpl->Int64Ty; }
 IntegerType *Type::getInt128Ty(LLVMContext &C) { return &C.pImpl->Int128Ty; }
 

 PointerType *Type::getPPC_FP128PtrTy(LLVMContext &C, unsigned AS) {
   return getPPC_FP128Ty(C)->getPointerTo(AS);
 }
 
 PointerType *Type::getX86_MMXPtrTy(LLVMContext &C, unsigned AS) {
   return getX86_MMXTy(C)->getPointerTo(AS);
 }
 
+PointerType *Type::getX86_AMXPtrTy(LLVMContext &C, unsigned AS) {
+  return getX86_AMXTy(C)->getPointerTo(AS);
+}
+
 PointerType *Type::getIntNPtrTy(LLVMContext &C, unsigned N, unsigned AS) {
   return getIntNTy(C, N)->getPointerTo(AS);
 }
 
 PointerType *Type::getInt1PtrTy(LLVMContext &C, unsigned AS) {
   return getInt1Ty(C)->getPointerTo(AS);
 }
 

llvm/lib/Target/X86/X86ISelDAGToDAG.cpp

case Intrinsic::x86_tileloadd64_internal: {
                        Node->getOperand(3),
                        Base,
                        Scale,
                        Index,
                        Disp,
                        Segment,
                        CFG,
                        Chain};
-      CNode = CurDAG->getMachineNode(Opc, dl, {MVT::v256i32, MVT::Other}, Ops);
+      CNode = CurDAG->getMachineNode(Opc, dl, {MVT::x86amx, MVT::Other}, Ops);
       ReplaceNode(Node, CNode);
       return;
     }
     case Intrinsic::x86_tdpbssd_internal: {
       if (!Subtarget->hasAMXTILE())
         break;
       SDValue Chain = Node->getOperand(0);
       unsigned Opc = X86::PTDPBSSDV;
       SDValue CFG = CurDAG->getRegister(0, MVT::Untyped);
       SDValue Ops[] = {Node->getOperand(2),
                        Node->getOperand(3),
                        Node->getOperand(4),
                        Node->getOperand(5),
                        Node->getOperand(6),
                        Node->getOperand(7),
                        CFG,
                        Chain};
       MachineSDNode *CNode =
-          CurDAG->getMachineNode(Opc, dl, {MVT::v256i32, MVT::Other}, Ops);
+          CurDAG->getMachineNode(Opc, dl, {MVT::x86amx, MVT::Other}, Ops);
       ReplaceNode(Node, CNode);
       return;
     }
     }
     break;
   }
   case ISD::INTRINSIC_VOID: {
     unsigned IntNo = Node->getConstantOperandVal(1);

llvm/lib/Target/X86/X86ISelLowering.cpp

if (!Subtarget.useSoftFloat() && Subtarget.hasVLX()) {
     }
 
     setOperationAction(ISD::TRUNCATE, MVT::v16i32, Custom);
     setOperationAction(ISD::TRUNCATE, MVT::v8i64, Custom);
     setOperationAction(ISD::TRUNCATE, MVT::v16i64, Custom);
   }
 
   if (Subtarget.hasAMXTILE()) {
-    addRegisterClass(MVT::v256i32, &X86::TILERegClass);
+    addRegisterClass(MVT::x86amx, &X86::TILERegClass);
   }
 
   // We want to custom lower some of our intrinsics.
   setOperationAction(ISD::INTRINSIC_WO_CHAIN, MVT::Other, Custom);
   setOperationAction(ISD::INTRINSIC_W_CHAIN, MVT::Other, Custom);
   setOperationAction(ISD::INTRINSIC_VOID, MVT::Other, Custom);
   if (!Subtarget.is64Bit()) {
     setOperationAction(ISD::INTRINSIC_W_CHAIN, MVT::i64, Custom);
bool X86TargetLowering::canMergeStoresTo(unsigned AddressSpace, EVT MemVT,
   if (NoFloat) {
     unsigned MaxIntSize = Subtarget.is64Bit() ? 64 : 32;
     return (MemVT.getSizeInBits() <= MaxIntSize);
   }
   // Make sure we don't merge greater than our preferred vector
   // width.
   if (MemVT.getSizeInBits() > Subtarget.getPreferVectorWidth())
     return false;
 
-  // Don't merge to x86 amx tile, as we only map MVT::v256i32
-  // to x86 amx tile on amx intrinsics.
-  if (MemVT == MVT::v256i32)
-    return false;
-
   return true;
 }
 
 bool X86TargetLowering::isCtlzFast() const {
   return Subtarget.hasFastLZCNT();
 }
 
 bool X86TargetLowering::isMaskAndCmp0FoldingBeneficial(
     const Instruction &AndI) const {

llvm/lib/Target/X86/X86LowerAMXType.cpp

 //===- llvm/CodeGen/TileShapeInfo.h - ---------------------------*- C++ -*-===//
 //
 // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
 // See https://llvm.org/LICENSE.txt for license information.
 // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
 //
 //===----------------------------------------------------------------------===//
 //
-/// \file Pass to transform <256 x i32>
-/// <256 x i32> is mapped to AMX tile register on X86, AMX instruction set only
-/// provides simple operation on tile register. The basic elementwise operation
-/// is not supported by AMX. Since we define the AMX tile as vector <256 x i32>
+/// \file Pass to transform <256 x i32> load/store
+/// <256 x i32> is bitcasted to x86_amx on X86, and AMX instruction set only
+/// provides simple operation on x86_amx. The basic elementwise operation
+/// is not supported by AMX. Since x86_amx is bitcasted from vector <256 x i32>
 /// and only AMX intrinsics can operate on the type, we need transform
-/// load/store <256 x i32> instruction to AMX load/store. Besides, we split
-/// <256 x i32> to 2 <128 x i32> if the vector is not used or defined by AMX
-/// intrinsics, so that in instruction selection it can be lowered to proper
-/// size which HW can support.
+/// load/store <256 x i32> instruction to AMX load/store. Otherwise we are not
+/// able to lower the bitcast instruction to X86 instruction.
 //
 //===----------------------------------------------------------------------===//
 //
 #include "X86.h"
-#include "llvm/ADT/DenseSet.h"
+#include "llvm/ADT/PostOrderIterator.h"
+#include "llvm/ADT/SmallSet.h"
 #include "llvm/Analysis/OptimizationRemarkEmitter.h"
 #include "llvm/Analysis/TargetTransformInfo.h"
 #include "llvm/CodeGen/Passes.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"
 
 using namespace llvm;
+using namespace PatternMatch;
 
 #define DEBUG_TYPE "lower-amx-type"
 
-namespace {
-class X86LowerAMXType {
-  Function &Func;
-  const DataLayout &DL;
-  DenseSet<Instruction *> LDSet;
-  DenseSet<Instruction *> STSet;
-  DenseMap<Value *, std::pair<LoadInst *, LoadInst *>> LoadMap;
-
-public:
-  X86LowerAMXType(Function &F) : Func(F), DL(F.getParent()->getDataLayout()) {}
-  bool visit();
-  bool visitLD();
-  bool visitST();
-  void splitST(Instruction *Inst);
-  void splitLD(Instruction *Inst);
-};
+static AllocaInst *CreateAllocaInst(IRBuilder<> &Builder, BasicBlock *BB) {
+  Function &F = *BB->getParent();
+  Module *M = BB->getModule();
+  const DataLayout &DL = M->getDataLayout();
+
+  Type *V256I32Ty = VectorType::get(Builder.getInt32Ty(), 256, false);
+  auto AllocaAlignment = DL.getPrefTypeAlign(V256I32Ty);
+  unsigned AllocaAS = DL.getAllocaAddrSpace();
+  AllocaInst *AllocaRes =
+      new AllocaInst(V256I32Ty, AllocaAS, "", &F.getEntryBlock().front());
+  AllocaRes->setAlignment(AllocaAlignment);
+  return AllocaRes;
+}
 
-// Split v256i32 load/store to 2 v128i32, so that ISel can
-// lower it to proper vector size.
-void X86LowerAMXType::splitST(Instruction *Inst) {
-  StoreInst *ST = dyn_cast<StoreInst>(Inst);
-  IRBuilder<> Builder(ST);
-  LLVMContext &Ctx = Builder.getContext();
-  Type *Ty = ST->getValueOperand()->getType();
-  EVT VT = EVT::getEVT(Ty);
-  EVT HalfVT = VT.getHalfNumVectorElementsVT(Ctx);
-  Type *HalfTy = HalfVT.getTypeForEVT(Ctx);
-
-  LoadInst *Lo, *Hi;
-  std::tie(Lo, Hi) = LoadMap[ST->getValueOperand()];
-  Value *Ptr = ST->getPointerOperand();
-  PointerType *HalfPtrTy = HalfTy->getPointerTo(ST->getPointerAddressSpace());
-  Value *HalfPtr = Builder.CreateBitCast(Ptr, HalfPtrTy);
-  // The HW require the alignment for AMX tile is 64, but front-end generate
-  // code for the vector alignment which is the vector size.
-  uint64_t HalfTySize = HalfTy->getPrimitiveSizeInBits().getFixedSize() / 8;
-  Align Alignment = std::min(Lo->getAlign(), Align(HalfTySize));
-  Builder.CreateAlignedStore(Lo, HalfPtr, Alignment, ST->isVolatile());
-
-  HalfPtr = Builder.CreateGEP(HalfTy, HalfPtr, Builder.getInt32(1));
-  Builder.CreateAlignedStore(Hi, HalfPtr, Alignment, ST->isVolatile());
-}
-
-bool X86LowerAMXType::visitST() {
-  if (STSet.empty())
-    return false;
-  for (auto *Inst : STSet) {
-    Value *Row, *Col;
-    const IntrinsicInst *II = dyn_cast<IntrinsicInst>(Inst->getOperand(0));
-    if (!II)
-      Row = Col = nullptr;
-    else {
-      switch (II->getIntrinsicID()) {
-      default:
-        Row = Col = nullptr;
-        break;
-      case Intrinsic::x86_tileloadd64_internal:
-      case Intrinsic::x86_tdpbssd_internal: {
-        Row = II->getArgOperand(0);
-        Col = II->getArgOperand(1);
-        break;
-      }
-      }
-    }
-    if (!Row) {
-      splitST(Inst);
-      continue;
-    }
-    IRBuilder<> Builder(Inst);
-    LLVMContext &Ctx = Builder.getContext();
-    // Use the maximun column as stride. It must be the same with load stride.
-    Value *Stride = Builder.getInt64(64);
-    Value *I8Ptr =
-        Builder.CreateBitCast(Inst->getOperand(1), Type::getInt8PtrTy(Ctx));
-    std::array<Value *, 5> Args = {Row, Col, I8Ptr, Stride,
-                                   Inst->getOperand(0)};
-
-    Builder.CreateIntrinsic(Intrinsic::x86_tilestored64_internal, None, Args);
-  }
-  return true;
-}
-
-void X86LowerAMXType::splitLD(Instruction *Inst) {
-  LoadInst *LD = dyn_cast<LoadInst>(Inst);
-  IRBuilder<> Builder(LD);
-  LLVMContext &Ctx = Builder.getContext();
-  Type *Ty = LD->getType();
-  EVT VT = EVT::getEVT(Ty);
-  EVT HalfVT = VT.getHalfNumVectorElementsVT(Ctx);
-  Type *HalfTy = HalfVT.getTypeForEVT(Ctx);
-
-  Value *Ptr = LD->getPointerOperand();
-  PointerType *HalfPtrTy = HalfTy->getPointerTo(LD->getPointerAddressSpace());
-  Value *HalfPtr = Builder.CreateBitCast(Ptr, HalfPtrTy);
-  // The HW require the alignment for AMX tile is 64, but front-end generate
-  // code for the vector alignment which is the vector size.
-  uint64_t HalfTySize = HalfTy->getPrimitiveSizeInBits().getFixedSize() / 8;
-  Align Alignment = std::min(LD->getAlign(), Align(HalfTySize));
-  auto *Lo =
-      Builder.CreateAlignedLoad(HalfTy, HalfPtr, Alignment, LD->isVolatile());
-
-  HalfPtr = Builder.CreateGEP(HalfTy, HalfPtr, Builder.getInt32(1));
-  auto *Hi =
-      Builder.CreateAlignedLoad(HalfTy, HalfPtr, Alignment, LD->isVolatile());
-
-  LoadMap[Inst] = std::make_pair(Lo, Hi);
-}
-
-bool X86LowerAMXType::visitLD() {
-  if (LDSet.empty())
-    return false;
-  for (auto &Inst : LDSet) {
-    int Count = 0;
-    Value *NewInst = nullptr;
-    // The user should be all AMX intrinsics or all LLVM instruction.
-    // Don't support it is used by both AMX intrinsics and LLVM instructions.
-    for (auto I = Inst->use_begin(), E = Inst->use_end(); I != E;) {
-      Use &U = *I++;
-      const IntrinsicInst *II = dyn_cast<IntrinsicInst>(U.getUser());
-      if (!II) {
-        Count++;
-        continue;
-      }
-      if (NewInst)
-        continue;
-      Value *Row, *Col;
-      switch (II->getIntrinsicID()) {
-      default:
-        report_fatal_error("Non-AMX intrinsic use tile type.");
-        break;
-      case Intrinsic::x86_tdpbssd_internal: {
-        unsigned OpNo = U.getOperandNo();
-        switch (OpNo) {
-        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);
-          Col = II->getArgOperand(1);
-          break;
-        }
-        break;
-      }
-      case Intrinsic::x86_tilestored64_internal: {
-        Row = II->getArgOperand(0);
-        Col = II->getArgOperand(1);
-        break;
-      }
-      }
-      assert(Count == 0 && "Can NOT mix amx intrinsic and LLVM instruction");
-      // FIXME: The shape def should be ahead of load.
-      IRBuilder<> Builder(Inst);
-      LLVMContext &Ctx = Builder.getContext();
-      // Use the maximun column as stride.
-      Value *Stride = Builder.getInt64(64);
-      Value *I8Ptr =
-          Builder.CreateBitCast(Inst->getOperand(0), Type::getInt8PtrTy(Ctx));
-      std::array<Value *, 4> Args = {Row, Col, I8Ptr, Stride};
+static std::pair<Value *, Value *> getShape(IntrinsicInst *II, unsigned OpNo) {
+  Value *Row = nullptr, *Col = nullptr;
+    llvm_unreachable("Expect amx intrinsics");
+  case Intrinsic::x86_tilestored64_internal: {
+  // a * b + c
+  // The shape depends on which operand.
+
+  return std::make_pair(Row, Col);
+
+// %1 = load x86_amx, x86_amx* %0, align 64
+// %2 = call x86_amx @llvm.x86.tdpbssd.internal(%1, %1, %1, ...)
+// -->
+// %1 = call x86_amx @llvm.x86.tileloadd64.internal()
+// %2 = call x86_amx @llvm.x86.tdpbssd.internal(%1, %1, %1, ...)
+static void transformTileLoad(LoadInst *LD) {
+  Value *Row = nullptr, *Col = nullptr;
+  Use &U = *(LD->use_begin());
+  unsigned OpNo = U.getOperandNo();
+  auto *II = cast<IntrinsicInst>(U.getUser());
+  std::tie(Row, Col) = getShape(II, OpNo);
+  IRBuilder<> Builder(LD);
+      Builder.CreateBitCast(LD->getOperand(0), Builder.getInt8PtrTy());
 
-      NewInst = Builder.CreateIntrinsic(Intrinsic::x86_tileloadd64_internal,
-                                        None, Args);
-
-      Inst->replaceAllUsesWith(NewInst);
-    }
-    if (!NewInst)
-      splitLD(Inst);
+  Value *NewInst =
+      Builder.CreateIntrinsic(Intrinsic::x86_tileloadd64_internal, None, Args);
+  LD->replaceAllUsesWith(NewInst);
+
+// %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)
+static void combineLoadBitcast(LoadInst *LD, BitCastInst *Bitcast) {
+  Value *Row = nullptr, *Col = nullptr;
+  Use &U = *(Bitcast->use_begin());
+  unsigned OpNo = U.getOperandNo();
+  auto *II = cast<IntrinsicInst>(U.getUser());
+  std::tie(Row, Col) = getShape(II, OpNo);
+  IRBuilder<> Builder(Bitcast);
+  // Use the maximun column as stride.
+  Value *Stride = Builder.getInt64(64);
+  Value *I8Ptr =
+      Builder.CreateBitCast(LD->getOperand(0), Builder.getInt8PtrTy());
+  std::array<Value *, 4> Args = {Row, Col, I8Ptr, Stride};
+
+  Value *NewInst =
+      Builder.CreateIntrinsic(Intrinsic::x86_tileloadd64_internal, None, Args);
+  Bitcast->replaceAllUsesWith(NewInst);
+}
+
+// %src = call x86_amx @llvm.x86.tileloadd64.internal(%row, %col, %addr,
+//                                                    %stride);
+// %13 = bitcast x86_amx %src to <256 x i32>
+// store <256 x i32> %13, <256 x i32>* %addr, align 64
+// -->
+// call void @llvm.x86.tilestored64.internal(%row, %col, %addr,
+//                                           %stride64, %13)
+static void combineBitcastStore(BitCastInst *Bitcast, StoreInst *ST) {
+
+  auto *Tile = Bitcast->getOperand(0);
+  auto *II = cast<IntrinsicInst>(Tile);
+  // Tile is output from AMX intrinsic. The first operand of the
+  // intrinsic is row, the second operand of the intrinsic is column.
+  Value *Row = II->getOperand(0);
+  Value *Col = II->getOperand(1);
+  IRBuilder<> Builder(ST);
+  // Use the maximum column as stride. It must be the same with load
+  // stride.
+  Value *Stride = Builder.getInt64(64);
+  Value *I8Ptr =
+      Builder.CreateBitCast(ST->getOperand(1), Builder.getInt8PtrTy());
+  std::array<Value *, 5> Args = {Row, Col, I8Ptr, Stride, Tile};
+  Builder.CreateIntrinsic(Intrinsic::x86_tilestored64_internal, None, Args);
+  if (Bitcast->hasOneUse())
+    return;
+  // %13 = bitcast x86_amx %src to <256 x i32>
+  // store <256 x i32> %13, <256 x i32>* %addr, align 64
+  // %add = <256 x i32> %13, <256 x i32> %src2
+  // -->
+  // %13 = bitcast x86_amx %src to <256 x i32>
+  // call void @llvm.x86.tilestored64.internal(%row, %col, %addr,
+  //                                           %stride64, %13)
+  // %14 = load <256 x i32>, %addr
+  // %add = <256 x i32> %14, <256 x i32> %src2
+  Value *Vec = Builder.CreateLoad(Bitcast->getType(), ST->getOperand(1));
+  Bitcast->replaceAllUsesWith(Vec);
+}
+
+// transform bitcast to <store, load> instructions.
+static bool transformBitcast(BitCastInst *Bitcast) {
+  IRBuilder<> Builder(Bitcast);
+  AllocaInst *AllocaAddr;
+  Value *I8Ptr, *Stride;
+  auto *Src = Bitcast->getOperand(0);
+
+  auto Prepare = [&]() {
+    AllocaAddr = CreateAllocaInst(Builder, Bitcast->getParent());
+    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 1024
+    // store <256 x i32> %src, <256 x i32>* %addr, align 1024
+    // %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();
+    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 1024
+    // %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 1024
+    auto *II = dyn_cast<IntrinsicInst>(Src);
+    if (!II)
+      return false; // May be bitcast from <256 x i32> to x86amx.
+    Prepare();
+    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);
   }
+
   return true;
 }
 
+// %addr = bitcast <256 x i32>* %tile to x86_amx*
+// store x86_amx %9, x86_amx* %addr, align 64
+// -->
+// call void @llvm.x86.tilestored64.internal(%row, %col, %addr,
+// %stride64, %9)
+static void transformTileStore(StoreInst *ST) {
+  auto *II = cast<IntrinsicInst>(ST->getValueOperand());
+  Value *Row = II->getOperand(0);
+  Value *Col = II->getOperand(1);
+  IRBuilder<> Builder(ST);
+  // Use the maximum column as stride. It must be the same with load
+  // stride.
+  Value *Stride = Builder.getInt64(64);
+  Value *I8Ptr =
+      Builder.CreateBitCast(ST->getOperand(1), Builder.getInt8PtrTy());
+  std::array<Value *, 5> Args = {Row, Col, I8Ptr, Stride,
+                                 ST->getValueOperand()};
+  Builder.CreateIntrinsic(Intrinsic::x86_tilestored64_internal, None, Args);
+}
+
+namespace {
+class X86LowerAMXType {
+  Function &Func;
+
+public:
+  X86LowerAMXType(Function &F) : Func(F) {}
+  bool visit();
+};
+
 bool X86LowerAMXType::visit() {
-  bool C;
-  auto IsAMXType = [](FixedVectorType *VTy) {
-    if (!VTy)
-      return false;
-    if (!VTy->getScalarType()->isIntegerTy(32))
-      return false;
-    if (VTy->getNumElements() != 256)
-      return false;
+  SmallVector<Instruction *, 8> DeadInsts;
+  SmallVector<Instruction *, 2> DeadBitcasts;
+
+  for (BasicBlock *BB : post_order(&Func)) {
+    for (BasicBlock::reverse_iterator II = BB->rbegin(), IE = BB->rend();
+         II != IE;) {
+      Instruction &Inst = *II++;
+      auto *Bitcast = dyn_cast<BitCastInst>(&Inst);
+      if (!Bitcast)
+        continue;
 
+      Value *Src = Bitcast->getOperand(0);
+      Type *Ty = Bitcast->getType();
+      auto CanonicalizeBitcast = [&]() {
+        if (Bitcast->user_empty()) {
+          DeadBitcasts.push_back(Bitcast);
           return true;
+        }
+        bool Change = false;
+        Value *DstV = Src, *PreDst = Bitcast, *SrcV;
+        while (match(DstV, m_BitCast(m_Value(SrcV))) &&
+               SrcV->getType()->getTypeID() == PreDst->getType()->getTypeID()) {
+          PreDst->replaceAllUsesWith(SrcV);
+          DeadBitcasts.push_back(cast<Instruction>(PreDst));
+          PreDst = DstV;
+          DstV = SrcV;
+          Change = true;
+        }
+        return Change;
       };
 
-  for (BasicBlock &BB : Func) {
-    for (Instruction &Inst : BB) {
-      LoadInst *LD = dyn_cast<LoadInst>(&Inst);
-      // Check load instruction.
-      // %3 = load <256 x i32>, <256 x i32>* %1, align 64
+      if (Ty->isPointerTy() &&
+          cast<PointerType>(Ty)->getElementType()->isX86_AMXTy()) {
+        for (auto UI = Bitcast->use_begin(), UE = Bitcast->use_end();
+             UI != UE;) {
+          Value *I = (UI++)->getUser();
+          auto *LD = dyn_cast<LoadInst>(I);
+          // %0 = bitcast <256 x i32>* %tile to x86_amx*
+          // %1 = load x86_amx, x86_amx* %0, align 64
           if (LD) {
-        FixedVectorType *VTy = dyn_cast<FixedVectorType>(Inst.getType());
-        if (!IsAMXType(VTy))
+            transformTileLoad(LD);
+            DeadInsts.push_back(LD);
+          }
+          auto *ST = dyn_cast<StoreInst>(I);
+          if (ST) {
+            // %addr = bitcast <256 x i32>* %tile to x86_amx*
+            // store x86_amx %9, x86_amx* %addr, align 64
+            // -->
+            // call void @llvm.x86.tilestored64.internal(%row, %col, %addr,
+            // %stride64, %9)
+            transformTileStore(ST);
+            DeadInsts.push_back(ST);
+          }
+        }
+        // If the dst type is <256 x i32>*, it is valid intruction.
+        // %0 = bitcast x86_amx* %tile to <256 x i32>*
+        // %1 = load <256 x i32>, <256 x i32>* %0, align 64
+        // store <256 x i32> %2, <256 x i32>* %0, align 64
+      }
+      if (Bitcast->getType()->isX86_AMXTy()) {
+        if (CanonicalizeBitcast())
           continue;
-        LDSet.insert(&Inst);
+        LoadInst *LD = dyn_cast<LoadInst>(Src);
+        if (!LD) {
+          if (transformBitcast(Bitcast))
+            DeadInsts.push_back(Bitcast);
           continue;
         }
-      // Check store instruction.
-      // store <256 x i32> %3, <256 x i32>* %2, align 64
-      StoreInst *ST = dyn_cast<StoreInst>(&Inst);
-      if (!ST)
+        // If load has mutli-user, duplicate a amx load.
+        // %src = load <256 x i32>, <256 x i32>* %addr, align 64
+        // %2 = bitcast <256 x i32> %src to x86_amx
+        // %add = add <256 x i32> %src, <256 x i32> %src2
+        // -->
+        // %src = load <256 x i32>, <256 x i32>* %addr, align 64
+        // %2 = call x86_amx @llvm.x86.tileloadd64.internal(i16 %row, i16 %col,
+        //                                            i8* %addr, i64 %stride64)
+        // %add = add <256 x i32> %src, <256 x i32> %src2
+
+        // If load has one user, the load will be eliminated in DAG ISel.
+        // %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)
+        combineLoadBitcast(LD, Bitcast);
+        if (LD->hasOneUse())
+          DeadInsts.push_back(LD);
+        DeadInsts.push_back(Bitcast);
+      } else if (Src->getType()->isX86_AMXTy()) {
+        if (CanonicalizeBitcast())
           continue;
-      FixedVectorType *VTy =
-          dyn_cast<FixedVectorType>(ST->getOperand(0)->getType());
-      if (!IsAMXType(VTy))
+        StoreInst *ST = nullptr;
+        for (auto UI = Bitcast->use_begin(), UE = Bitcast->use_end();
+             UI != UE;) {
+          Value *I = (UI++)->getUser();
+          ST = dyn_cast<StoreInst>(I);
+          if (ST)
+            break;
+        }
+        if (!ST) {
+          if (transformBitcast(Bitcast))
+            DeadInsts.push_back(Bitcast);
           continue;
-      STSet.insert(&Inst);
+        }
+        // If bitcast (%13) has one use, combine bitcast and store to amx store.
+        // %src = call x86_amx @llvm.x86.tileloadd64.internal(%row, %col, %addr,
+        //                                                    %stride);
+        // %13 = bitcast x86_amx %src to <256 x i32>
+        // store <256 x i32> %13, <256 x i32>* %addr, align 64
+        // -->
+        // call void @llvm.x86.tilestored64.internal(%row, %col, %addr,
+        //                                           %stride64, %13)
+        //
+        // If bitcast (%13) has multi-use, transform as below.
+        // %13 = bitcast x86_amx %src to <256 x i32>
+        // store <256 x i32> %13, <256 x i32>* %addr, align 64
+        // %add = <256 x i32> %13, <256 x i32> %src2
+        // -->
+        // %13 = bitcast x86_amx %src to <256 x i32>
+        // call void @llvm.x86.tilestored64.internal(%row, %col, %addr,
+        //                                           %stride64, %13)
+        // %14 = load <256 x i32>, %addr
+        // %add = <256 x i32> %14, <256 x i32> %src2
+        //
+        combineBitcastStore(Bitcast, ST);
+        DeadInsts.push_back(Bitcast);
+        DeadInsts.push_back(ST);
+      }
     }
   }
 
-  C = visitLD() | visitST();
-  for (auto *Inst : STSet)
-    Inst->eraseFromParent();
-  for (auto *Inst : LDSet)
+  bool C = !DeadInsts.empty() || !DeadBitcasts.empty();
+
+  SmallSet<Instruction *, 8> DeletedInst;
+  auto DeleteInst = [&](Instruction *Inst) {
+    SmallVector<Instruction *, 4> DeadIs;
+    DeadIs.push_back(Inst);
+    while (!DeadIs.empty()) {
+      auto *Inst = DeadIs.back();
+      DeadIs.pop_back();
+      if (DeletedInst.count(Inst))
+        continue;
+      for (auto I = Inst->op_begin(), E = Inst->op_end(); I != E;) {
+        Instruction *Op = dyn_cast<Instruction>(*I);
+        if (Op && Op->hasOneUse())
+          DeadIs.push_back(Op);
+        ++I;
+      }
       Inst->eraseFromParent();
+      DeletedInst.insert(Inst);
+    }
+  };
+  while (!DeadInsts.empty()) {
+    auto *Inst = DeadInsts.back();
+    DeleteInst(Inst);
+    DeadInsts.pop_back();
+  }
+  // Delete user first.
+  for (auto *Inst : DeadBitcasts)
+    DeleteInst(Inst);
+
   return C;
 }
 } // anonymous namespace
 
 namespace {
 
 class X86LowerAMXTypeLegacyPass : public FunctionPass {
 public:
   static char ID;
 
   X86LowerAMXTypeLegacyPass() : FunctionPass(ID) {
     initializeX86LowerAMXTypeLegacyPassPass(*PassRegistry::getPassRegistry());
   }
 
   bool runOnFunction(Function &F) override {
     X86LowerAMXType LAT(F);
     bool C = LAT.visit();
     return C;
   }
 
   void getAnalysisUsage(AnalysisUsage &AU) const override {
     AU.setPreservesCFG();
   }
 };
 
 } // 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_END(X86LowerAMXTypeLegacyPass, DEBUG_TYPE, PassName, false,
                     false)
 
 FunctionPass *llvm::createX86LowerAMXTypePass() {
   return new X86LowerAMXTypeLegacyPass();
 }

llvm/lib/Target/X86/X86RegisterInfo.td

 def VK32WM  : RegisterClass<"X86", [v32i1], 32, (add VK16WM)> {let Size = 32;}
 def VK64WM  : RegisterClass<"X86", [v64i1], 64, (add VK32WM)> {let Size = 64;}
 
 // Bound registers
 def BNDR : RegisterClass<"X86", [v2i64], 128, (sequence "BND%u", 0, 3)>;
 
 // Tiles
 let CopyCost = -1 in // Don't allow copying of tile registers
-def TILE : RegisterClass<"X86", [v256i32], 8192,
+def TILE : RegisterClass<"X86", [x86amx], 8192,
                          (sequence "TMM%u", 0, 7)> {let Size = 8192;}
 def TILECFG : RegisterClass<"X86", [untyped], 512, (add TMMCFG)> {
   let CopyCost = -1;  // Don't allow copying of tile config registers.
   let isAllocatable = 1;
   let Size = 512;
 }

llvm/test/CodeGen/X86/AMX/amx-across-func.ll

 ; CHECK-NEXT:    popq %r14
 ; CHECK-NEXT:    .cfi_def_cfa_offset 24
 ; CHECK-NEXT:    popq %r15
 ; CHECK-NEXT:    .cfi_def_cfa_offset 16
 ; CHECK-NEXT:    popq %rbp
 ; CHECK-NEXT:    .cfi_def_cfa_offset 8
 ; CHECK-NEXT:    tilerelease
 ; CHECK-NEXT:    retq
-  %3 = tail call <256 x i32> @llvm.x86.tileloadd64.internal(i16 %0, i16 8, i8* getelementptr inbounds ([3072 x i8], [3072 x i8]* @buf, i64 0, i64 0), i64 32) #4
-  %4 = tail call <256 x i32> @llvm.x86.tileloadd64.internal(i16 8, i16 %1, i8* getelementptr inbounds ([3072 x i8], [3072 x i8]* @buf, i64 0, i64 1024), i64 32) #4
+  %3 = tail call x86_amx @llvm.x86.tileloadd64.internal(i16 %0, i16 8, i8* getelementptr inbounds ([3072 x i8], [3072 x i8]* @buf, i64 0, i64 0), i64 32) #4
+  %4 = tail call x86_amx @llvm.x86.tileloadd64.internal(i16 8, i16 %1, i8* getelementptr inbounds ([3072 x i8], [3072 x i8]* @buf, i64 0, i64 1024), i64 32) #4
   tail call void (...) @foo() #4
-  %5 = tail call <256 x i32> @llvm.x86.tileloadd64.internal(i16 %0, i16 %1, i8* getelementptr inbounds ([3072 x i8], [3072 x i8]* @buf, i64 0, i64 2048), i64 32) #4
-  %6 = tail call <256 x i32> @llvm.x86.tdpbssd.internal(i16 %0, i16 %1, i16 8, <256 x i32> %5, <256 x i32> %3, <256 x i32> %4) #4
-  tail call void @llvm.x86.tilestored64.internal(i16 %0, i16 %1, i8* getelementptr inbounds ([3072 x i8], [3072 x i8]* @buf, i64 0, i64 2048), i64 32, <256 x i32> %6) #4
+  %5 = tail call x86_amx @llvm.x86.tileloadd64.internal(i16 %0, i16 %1, i8* getelementptr inbounds ([3072 x i8], [3072 x i8]* @buf, i64 0, i64 2048), i64 32) #4
+  %6 = tail call x86_amx @llvm.x86.tdpbssd.internal(i16 %0, i16 %1, i16 8, x86_amx %5, x86_amx %3, x86_amx %4) #4
+  tail call void @llvm.x86.tilestored64.internal(i16 %0, i16 %1, i8* getelementptr inbounds ([3072 x i8], [3072 x i8]* @buf, i64 0, i64 2048), i64 32, x86_amx %6) #4
   ret void
 }
 
 declare dso_local void @foo(...) local_unnamed_addr #3
 
-declare <256 x i32> @llvm.x86.tileloadd64.internal(i16, i16, i8*, i64) #4
-declare <256 x i32> @llvm.x86.tdpbssd.internal(i16, i16, i16, <256 x i32>, <256 x i32>, <256 x i32>) #4
-declare void @llvm.x86.tilestored64.internal(i16, i16, i8*, i64, <256 x i32>) #4
+declare x86_amx @llvm.x86.tileloadd64.internal(i16, i16, i8*, i64) #4
+declare x86_amx @llvm.x86.tdpbssd.internal(i16, i16, i16, x86_amx, x86_amx, x86_amx) #4
+declare void @llvm.x86.tilestored64.internal(i16, i16, i8*, i64, x86_amx) #4
 
 attributes #2 = { nounwind uwtable "correctly-rounded-divide-sqrt-fp-math"="false" "disable-tail-calls"="false" "frame-pointer"="none" "less-precise-fpmad"="false" "min-legal-vector-width"="8192" "no-infs-fp-math"="false" "no-jump-tables"="false" "no-nans-fp-math"="false" "no-signed-zeros-fp-math"="false" "no-trapping-math"="true" "stack-protector-buffer-size"="8" "target-cpu"="x86-64" "target-features"="+amx-int8,+amx-tile,+cx8,+fxsr,+mmx,+sse,+sse2,+x87" "tune-cpu"="generic" "unsafe-fp-math"="false" "use-soft-float"="false" }
 attributes #3 = { "correctly-rounded-divide-sqrt-fp-math"="false" "disable-tail-calls"="false" "frame-pointer"="none" "less-precise-fpmad"="false" "no-infs-fp-math"="false" "no-nans-fp-math"="false" "no-signed-zeros-fp-math"="false" "no-trapping-math"="true" "stack-protector-buffer-size"="8" "target-cpu"="x86-64" "target-features"="+amx-int8,+amx-tile,+cx8,+fxsr,+mmx,+sse,+sse2,+x87" "tune-cpu"="generic" "unsafe-fp-math"="false" "use-soft-float"="false" }
 attributes #4 = { nounwind }

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

 ; CHECK-NEXT:    vzeroupper
 ; CHECK-NEXT:    retq
   %4 = icmp eq i32 %0, 0
   %5 = shl i16 %1, 8
   %6 = ashr exact i16 %5, 8
   br i1 %4, label %11, label %7
 
 7:                                                ; preds = %3
-  %8 = tail call <256 x i32> @llvm.x86.tileloadd64.internal(i16 %6, i16 %1, i8* getelementptr inbounds ([1024 x i8], [1024 x i8]* @buf, i64 0, i64 0), i64 32) #3
-  %9 = tail call <256 x i32> @llvm.x86.tileloadd64.internal(i16 %6, i16 %2, i8* getelementptr inbounds ([1024 x i8], [1024 x i8]* @buf, i64 0, i64 0), i64 32) #3
-  %10 = tail call <256 x i32> @llvm.x86.tileloadd64.internal(i16 %6, i16 %2, i8* getelementptr inbounds ([1024 x i8], [1024 x i8]* @buf, i64 0, i64 0), i64 32) #3
+  %8 = tail call x86_amx @llvm.x86.tileloadd64.internal(i16 %6, i16 %1, i8* getelementptr inbounds ([1024 x i8], [1024 x i8]* @buf, i64 0, i64 0), i64 32) #3
+  %9 = tail call x86_amx @llvm.x86.tileloadd64.internal(i16 %6, i16 %2, i8* getelementptr inbounds ([1024 x i8], [1024 x i8]* @buf, i64 0, i64 0), i64 32) #3
+  %10 = tail call x86_amx @llvm.x86.tileloadd64.internal(i16 %6, i16 %2, i8* getelementptr inbounds ([1024 x i8], [1024 x i8]* @buf, i64 0, i64 0), i64 32) #3
   br label %15
 
 11:                                               ; preds = %3
-  %12 = tail call <256 x i32> @llvm.x86.tileloadd64.internal(i16 %6, i16 %1, i8* getelementptr inbounds ([1024 x i8], [1024 x i8]* @buf2, i64 0, i64 0), i64 32) #3
-  %13 = tail call <256 x i32> @llvm.x86.tileloadd64.internal(i16 %6, i16 %2, i8* getelementptr inbounds ([1024 x i8], [1024 x i8]* @buf2, i64 0, i64 0), i64 32) #3
-  %14 = tail call <256 x i32> @llvm.x86.tileloadd64.internal(i16 %6, i16 %2, i8* getelementptr inbounds ([1024 x i8], [1024 x i8]* @buf2, i64 0, i64 0), i64 32) #3
+  %12 = tail call x86_amx @llvm.x86.tileloadd64.internal(i16 %6, i16 %1, i8* getelementptr inbounds ([1024 x i8], [1024 x i8]* @buf2, i64 0, i64 0), i64 32) #3
+  %13 = tail call x86_amx @llvm.x86.tileloadd64.internal(i16 %6, i16 %2, i8* getelementptr inbounds ([1024 x i8], [1024 x i8]* @buf2, i64 0, i64 0), i64 32) #3
+  %14 = tail call x86_amx @llvm.x86.tileloadd64.internal(i16 %6, i16 %2, i8* getelementptr inbounds ([1024 x i8], [1024 x i8]* @buf2, i64 0, i64 0), i64 32) #3
   br label %15
 
 15:                                               ; preds = %11, %7
-  %16 = phi <256 x i32> [ %12, %11 ], [ %8, %7 ]
-  %17 = phi <256 x i32> [ %13, %11 ], [ %9, %7 ]
-  %18 = phi <256 x i32> [ %14, %11 ], [ %10, %7 ]
-  %19 = tail call <256 x i32> @llvm.x86.tdpbssd.internal(i16 %6, i16 %2, i16 %1, <256 x i32> %18, <256 x i32> %16, <256 x i32> %17) #3
-  tail call void @llvm.x86.tilestored64.internal(i16 %6, i16 %2, i8* getelementptr inbounds ([1024 x i8], [1024 x i8]* @buf, i64 0, i64 0), i64 32, <256 x i32> %19) #3
+  %16 = phi x86_amx [ %12, %11 ], [ %8, %7 ]
+  %17 = phi x86_amx [ %13, %11 ], [ %9, %7 ]
+  %18 = phi x86_amx [ %14, %11 ], [ %10, %7 ]
+  %19 = tail call x86_amx @llvm.x86.tdpbssd.internal(i16 %6, i16 %2, i16 %1, x86_amx %18, x86_amx %16, x86_amx %17) #3
+  tail call void @llvm.x86.tilestored64.internal(i16 %6, i16 %2, i8* getelementptr inbounds ([1024 x i8], [1024 x i8]* @buf, i64 0, i64 0), i64 32, x86_amx %19) #3
   ret void
 }
 
-declare <256 x i32> @llvm.x86.tileloadd64.internal(i16, i16, i8*, i64) #3
+declare x86_amx @llvm.x86.tileloadd64.internal(i16, i16, i8*, i64) #3
 
-declare <256 x i32> @llvm.x86.tdpbssd.internal(i16, i16, i16, <256 x i32>, <256 x i32>, <256 x i32>) #3
+declare x86_amx @llvm.x86.tdpbssd.internal(i16, i16, i16, x86_amx, x86_amx, x86_amx) #3
 
-declare void @llvm.x86.tilestored64.internal(i16, i16, i8*, i64, <256 x i32>) #3
+declare void @llvm.x86.tilestored64.internal(i16, i16, i8*, i64, x86_amx) #3
 
 attributes #2 = { nounwind uwtable "correctly-rounded-divide-sqrt-fp-math"="false" "disable-tail-calls"="false" "frame-pointer"="none" "less-precise-fpmad"="false" "min-legal-vector-width"="8192" "no-infs-fp-math"="false" "no-jump-tables"="false" "no-nans-fp-math"="false" "no-signed-zeros-fp-math"="false" "no-trapping-math"="true" "stack-protector-buffer-size"="8" "target-cpu"="x86-64" "target-features"="+amx-int8,+amx-tile,+avx,+avx2,+avx512f,+cx8,+f16c,+fma,+fxsr,+mmx,+popcnt,+sse,+sse2,+sse3,+sse4.1,+sse4.2,+ssse3,+x87,+xsave" "tune-cpu"="generic" "unsafe-fp-math"="false" "use-soft-float"="false" }
 attributes #3 = { nounwind }

llvm/test/CodeGen/X86/AMX/amx-intrinsic-chain.ll

 ; CHECK-NEXT:    tdpbssd %tmm4, %tmm0, %tmm3
 ; CHECK-NEXT:    tilestored %tmm3, (%rdx,%r8)
 ; CHECK-NEXT:    tdpbssd %tmm4, %tmm1, %tmm2
 ; CHECK-NEXT:    tilestored %tmm2, (%rdi,%r8)
 ; CHECK-NEXT:    tilerelease
 ; CHECK-NEXT:    vzeroupper
 ; CHECK-NEXT:    retq
 entry:
-  %a1 = call <256 x i32> @llvm.x86.tileloadd64.internal(i16 16, i16 64, i8* nonnull %A_mem, i64 64)
+  %a1 = call x86_amx @llvm.x86.tileloadd64.internal(i16 16, i16 64, i8* nonnull %A_mem, i64 64)
   %addr = getelementptr inbounds i8, i8* %A_mem, i64 1024
-  %a2 = call <256 x i32> @llvm.x86.tileloadd64.internal(i16 16, i16 64, i8* nonnull %addr, i64 64)
-  %c1 = call <256 x i32> @llvm.x86.tileloadd64.internal(i16 16, i16 64, i8* nonnull %C_mem, i64 64)
+  %a2 = call x86_amx @llvm.x86.tileloadd64.internal(i16 16, i16 64, i8* nonnull %addr, i64 64)
+  %c1 = call x86_amx @llvm.x86.tileloadd64.internal(i16 16, i16 64, i8* nonnull %C_mem, i64 64)
   %caddr = getelementptr inbounds i8, i8* %C_mem, i64 1024
-  %c2 = call <256 x i32> @llvm.x86.tileloadd64.internal(i16 16, i16 64, i8* nonnull %caddr, i64 64)
+  %c2 = call x86_amx @llvm.x86.tileloadd64.internal(i16 16, i16 64, i8* nonnull %caddr, i64 64)
   br label %dotpd
 
 dotpd:
-  %b = call <256 x i32> @llvm.x86.tileloadd64.internal(i16 16, i16 64, i8* nonnull %B_mem, i64 64)
-  %dp1 = call <256 x i32> @llvm.x86.tdpbssd.internal(i16 16, i16 64, i16 64, <256 x i32> %c1, <256 x i32> %a1, <256 x i32> %b)
-  call void @llvm.x86.tilestored64.internal(i16 16, i16 64, i8* nonnull %C_mem, i64 64, <256 x i32> %dp1)
-  %dp2 = call <256 x i32> @llvm.x86.tdpbssd.internal(i16 16, i16 64, i16 64, <256 x i32> %c2, <256 x i32> %a2, <256 x i32> %b)
-  call void @llvm.x86.tilestored64.internal(i16 16, i16 64, i8* nonnull %caddr, i64 64, <256 x i32> %dp2)
+  %b = call x86_amx @llvm.x86.tileloadd64.internal(i16 16, i16 64, i8* nonnull %B_mem, i64 64)
+  %dp1 = call x86_amx @llvm.x86.tdpbssd.internal(i16 16, i16 64, i16 64, x86_amx %c1, x86_amx %a1, x86_amx %b)
+  call void @llvm.x86.tilestored64.internal(i16 16, i16 64, i8* nonnull %C_mem, i64 64, x86_amx %dp1)
+  %dp2 = call x86_amx @llvm.x86.tdpbssd.internal(i16 16, i16 64, i16 64, x86_amx %c2, x86_amx %a2, x86_amx %b)
+  call void @llvm.x86.tilestored64.internal(i16 16, i16 64, i8* nonnull %caddr, i64 64, x86_amx %dp2)
   ret void
 }
 
-declare <256 x i32> @llvm.x86.tileloadd64.internal(i16, i16, i8*, i64)
-declare <256 x i32> @llvm.x86.tdpbssd.internal(i16, i16, i16, <256 x i32>, <256 x i32>, <256 x i32>)
-declare void @llvm.x86.tilestored64.internal(i16, i16, i8*, i64, <256 x i32>)
+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 void @llvm.x86.tilestored64.internal(i16, i16, i8*, i64, x86_amx)

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

 ; CHECK-NEXT:    movl $buf, %eax
 ; CHECK-NEXT:    movl $32, %ecx
 ; CHECK-NEXT:    tilestored %tmm0, (%rax,%rcx)
 ; CHECK-NEXT:    addq $2936, %rsp # imm = 0xB78
 ; CHECK-NEXT:    .cfi_def_cfa_offset 8
 ; CHECK-NEXT:    tilerelease
 ; CHECK-NEXT:    vzeroupper
 ; CHECK-NEXT:    retq
-  %4 = tail call <256 x i32> @llvm.x86.tileloadd64.internal(i16 %1, i16 %2, i8* getelementptr inbounds ([1024 x i8], [1024 x i8]* @buf, i64 0, i64 0), i64 32) #3
-  %5 = tail call <256 x i32> @llvm.x86.tileloadd64.internal(i16 %1, i16 %2, i8* getelementptr inbounds ([1024 x i8], [1024 x i8]* @buf, i64 0, i64 0), i64 32) #3
-  %6 = tail call <256 x i32> @llvm.x86.tileloadd64.internal(i16 %1, i16 %2, i8* getelementptr inbounds ([1024 x i8], [1024 x i8]* @buf, i64 0, i64 0), i64 32) #3
-  %7 = tail call <256 x i32> @llvm.x86.tileloadd64.internal(i16 %2, i16 %2, i8* getelementptr inbounds ([1024 x i8], [1024 x i8]* @buf, i64 0, i64 0), i64 32) #3
-  %8 = tail call <256 x i32> @llvm.x86.tileloadd64.internal(i16 %2, i16 %2, i8* getelementptr inbounds ([1024 x i8], [1024 x i8]* @buf, i64 0, i64 0), i64 32) #3
-  %9 = tail call <256 x i32> @llvm.x86.tileloadd64.internal(i16 %2, i16 %2, i8* getelementptr inbounds ([1024 x i8], [1024 x i8]* @buf, i64 0, i64 0), i64 32) #3
-  %10 = tail call <256 x i32> @llvm.x86.tileloadd64.internal(i16 %2, i16 %2, i8* getelementptr inbounds ([1024 x i8], [1024 x i8]* @buf, i64 0, i64 0), i64 32) #3
+  %4 = tail call x86_amx @llvm.x86.tileloadd64.internal(i16 %1, i16 %2, i8* getelementptr inbounds ([1024 x i8], [1024 x i8]* @buf, i64 0, i64 0), i64 32) #3
+  %5 = tail call x86_amx @llvm.x86.tileloadd64.internal(i16 %1, i16 %2, i8* getelementptr inbounds ([1024 x i8], [1024 x i8]* @buf, i64 0, i64 0), i64 32) #3
+  %6 = tail call x86_amx @llvm.x86.tileloadd64.internal(i16 %1, i16 %2, i8* getelementptr inbounds ([1024 x i8], [1024 x i8]* @buf, i64 0, i64 0), i64 32) #3
+  %7 = tail call x86_amx @llvm.x86.tileloadd64.internal(i16 %2, i16 %2, i8* getelementptr inbounds ([1024 x i8], [1024 x i8]* @buf, i64 0, i64 0), i64 32) #3
+  %8 = tail call x86_amx @llvm.x86.tileloadd64.internal(i16 %2, i16 %2, i8* getelementptr inbounds ([1024 x i8], [1024 x i8]* @buf, i64 0, i64 0), i64 32) #3
+  %9 = tail call x86_amx @llvm.x86.tileloadd64.internal(i16 %2, i16 %2, i8* getelementptr inbounds ([1024 x i8], [1024 x i8]* @buf, i64 0, i64 0), i64 32) #3
+  %10 = tail call x86_amx @llvm.x86.tileloadd64.internal(i16 %2, i16 %2, i8* getelementptr inbounds ([1024 x i8], [1024 x i8]* @buf, i64 0, i64 0), i64 32) #3
   %11 = icmp eq i32 %0, 0
   br i1 %11, label %16, label %12
 
 12:                                               ; preds = %3
-  %13 = tail call <256 x i32> @llvm.x86.tileloadd64.internal(i16 %1, i16 %1, i8* getelementptr inbounds ([1024 x i8], [1024 x i8]* @buf, i64 0, i64 0), i64 32) #3
-  %14 = tail call <256 x i32> @llvm.x86.tileloadd64.internal(i16 %1, i16 %2, i8* getelementptr inbounds ([1024 x i8], [1024 x i8]* @buf, i64 0, i64 0), i64 32) #3
-  %15 = tail call <256 x i32> @llvm.x86.tileloadd64.internal(i16 %1, i16 %2, i8* getelementptr inbounds ([1024 x i8], [1024 x i8]* @buf, i64 0, i64 0), i64 32) #3
+  %13 = tail call x86_amx @llvm.x86.tileloadd64.internal(i16 %1, i16 %1, i8* getelementptr inbounds ([1024 x i8], [1024 x i8]* @buf, i64 0, i64 0), i64 32) #3
+  %14 = tail call x86_amx @llvm.x86.tileloadd64.internal(i16 %1, i16 %2, i8* getelementptr inbounds ([1024 x i8], [1024 x i8]* @buf, i64 0, i64 0), i64 32) #3
+  %15 = tail call x86_amx @llvm.x86.tileloadd64.internal(i16 %1, i16 %2, i8* getelementptr inbounds ([1024 x i8], [1024 x i8]* @buf, i64 0, i64 0), i64 32) #3
   br label %20
 
 16:                                               ; preds = %3
-  %17 = tail call <256 x i32> @llvm.x86.tileloadd64.internal(i16 %1, i16 %1, i8* getelementptr inbounds ([1024 x i8], [1024 x i8]* @buf2, i64 0, i64 0), i64 32) #3
-  %18 = tail call <256 x i32> @llvm.x86.tileloadd64.internal(i16 %1, i16 %2, i8* getelementptr inbounds ([1024 x i8], [1024 x i8]* @buf2, i64 0, i64 0), i64 32) #3
-  %19 = tail call <256 x i32> @llvm.x86.tileloadd64.internal(i16 %1, i16 %2, i8* getelementptr inbounds ([1024 x i8], [1024 x i8]* @buf2, i64 0, i64 0), i64 32) #3
+  %17 = tail call x86_amx @llvm.x86.tileloadd64.internal(i16 %1, i16 %1, i8* getelementptr inbounds ([1024 x i8], [1024 x i8]* @buf2, i64 0, i64 0), i64 32) #3
+  %18 = tail call x86_amx @llvm.x86.tileloadd64.internal(i16 %1, i16 %2, i8* getelementptr inbounds ([1024 x i8], [1024 x i8]* @buf2, i64 0, i64 0), i64 32) #3
+  %19 = tail call x86_amx @llvm.x86.tileloadd64.internal(i16 %1, i16 %2, i8* getelementptr inbounds ([1024 x i8], [1024 x i8]* @buf2, i64 0, i64 0), i64 32) #3
   br label %20
 
 20:                                               ; preds = %16, %12
-  %21 = phi <256 x i32> [ %17, %16 ], [ %13, %12 ]
-  %22 = phi <256 x i32> [ %18, %16 ], [ %14, %12 ]
-  %23 = phi <256 x i32> [ %19, %16 ], [ %15, %12 ]
-  %24 = tail call <256 x i32> @llvm.x86.tdpbssd.internal(i16 %1, i16 %2, i16 %1, <256 x i32> %23, <256 x i32> %21, <256 x i32> %22) #3
-  %25 = tail call <256 x i32> @llvm.x86.tdpbssd.internal(i16 %1, i16 %2, i16 %2, <256 x i32> %6, <256 x i32> %24, <256 x i32> %5) #3
-  %26 = tail call <256 x i32> @llvm.x86.tdpbssd.internal(i16 %1, i16 %2, i16 %2, <256 x i32> %8, <256 x i32> %25, <256 x i32> %7) #3
-  %27 = tail call <256 x i32> @llvm.x86.tdpbssd.internal(i16 %2, i16 %2, i16 %2, <256 x i32> %10, <256 x i32> %26, <256 x i32> %9) #3
-  tail call void @llvm.x86.tilestored64.internal(i16 %2, i16 %2, i8* getelementptr inbounds ([1024 x i8], [1024 x i8]* @buf, i64 0, i64 0), i64 32, <256 x i32> %27) #3
+  %21 = phi x86_amx [ %17, %16 ], [ %13, %12 ]
+  %22 = phi x86_amx [ %18, %16 ], [ %14, %12 ]
+  %23 = phi x86_amx [ %19, %16 ], [ %15, %12 ]
+  %24 = tail call x86_amx @llvm.x86.tdpbssd.internal(i16 %1, i16 %2, i16 %1, x86_amx %23, x86_amx %21, x86_amx %22) #3
+  %25 = tail call x86_amx @llvm.x86.tdpbssd.internal(i16 %1, i16 %2, i16 %2, x86_amx %6, x86_amx %24, x86_amx %5) #3
+  %26 = tail call x86_amx @llvm.x86.tdpbssd.internal(i16 %1, i16 %2, i16 %2, x86_amx %8, x86_amx %25, x86_amx %7) #3
+  %27 = tail call x86_amx @llvm.x86.tdpbssd.internal(i16 %2, i16 %2, i16 %2, x86_amx %10, x86_amx %26, x86_amx %9) #3
+  tail call void @llvm.x86.tilestored64.internal(i16 %2, i16 %2, i8* getelementptr inbounds ([1024 x i8], [1024 x i8]* @buf, i64 0, i64 0), i64 32, x86_amx %27) #3
   ret void
 }
 
-declare <256 x i32> @llvm.x86.tileloadd64.internal(i16, i16, i8*, i64) #3
-declare <256 x i32> @llvm.x86.tdpbssd.internal(i16, i16, i16, <256 x i32>, <256 x i32>, <256 x i32>) #3
-declare void @llvm.x86.tilestored64.internal(i16, i16, i8*, i64, <256 x i32>) #3
+declare x86_amx @llvm.x86.tileloadd64.internal(i16, i16, i8*, i64) #3
+declare x86_amx @llvm.x86.tdpbssd.internal(i16, i16, i16, x86_amx, x86_amx, x86_amx) #3
+declare void @llvm.x86.tilestored64.internal(i16, i16, i8*, i64, x86_amx) #3
 
 attributes #2 = { nounwind uwtable "correctly-rounded-divide-sqrt-fp-math"="false" "disable-tail-calls"="false" "frame-pointer"="none" "less-precise-fpmad"="false" "min-legal-vector-width"="8192" "no-infs-fp-math"="false" "no-jump-tables"="false" "no-nans-fp-math"="false" "no-signed-zeros-fp-math"="false" "no-trapping-math"="true" "stack-protector-buffer-size"="8" "target-cpu"="x86-64" "target-features"="+amx-int8,+amx-tile,+cx8,+fxsr,+mmx,+sse,+sse2,+x87" "tune-cpu"="generic" "unsafe-fp-math"="false" "use-soft-float"="false" }
 attributes #3 = { nounwind }

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

 ; NOTE: Assertions have been autogenerated by utils/update_test_checks.py
 ; RUN: opt -lower-amx-type %s -S | FileCheck %s
 target datalayout = "e-m:e-p270:32:32-p271:32:32-p272:64:64-i64:64-f80:128-n8:16:32:64-S128"
 target triple = "x86_64-unknown-linux-gnu"
 
 %struct.__tile_str = type { i16, i16, <256 x i32> }
 
 @buf = dso_local global [1024 x i8] zeroinitializer, align 16
 @buf2 = dso_local global [1024 x i8] zeroinitializer, align 16
 
+define dso_local <256 x i32> @test_amx_bitcast(<256 x i32> %in) #2 {
+; CHECK-LABEL: @test_amx_bitcast(
+; CHECK-NEXT:  entry:
+; CHECK-NEXT:    ret <256 x i32> [[IN:%.*]]
+;
+entry:
+  %amx = bitcast <256 x i32> %in to x86_amx
+  %vec = bitcast x86_amx %amx to <256 x i32>
+  ret <256 x i32> %vec
+}
+
+define dso_local void @test_amx_store(<256 x i32>* %in, i16 %m, i16 %n, i8 *%buf, i64 %s) #2 {
+; CHECK-LABEL: @test_amx_store(
+; CHECK-NEXT:  entry:
+; CHECK-NEXT:    [[T0:%.*]] = call x86_amx @llvm.x86.tileloadd64.internal(i16 [[M:%.*]], i16 [[N:%.*]], i8* [[BUF:%.*]], i64 [[S:%.*]]) [[ATTR3:#.*]]
+; CHECK-NEXT:    [[ADDR:%.*]] = bitcast <256 x i32>* [[IN:%.*]] to x86_amx*
+; CHECK-NEXT:    [[TMP0:%.*]] = bitcast x86_amx* [[ADDR]] to i8*
+; CHECK-NEXT:    call void @llvm.x86.tilestored64.internal(i16 [[M]], i16 [[N]], i8* [[TMP0]], i64 64, x86_amx [[T0]])
+; CHECK-NEXT:    ret void
+;
+entry:
+  %t0 = call x86_amx @llvm.x86.tileloadd64.internal(i16 %m, i16 %n, i8* %buf, i64 %s) #3
+  %addr = bitcast <256 x i32>* %in to x86_amx*
+  store x86_amx %t0, x86_amx* %addr, align 64
+  ret void
+}
+
+define dso_local void @test_amx_load(<256 x i32>* %in, i16 %m, i16 %n, i8 *%buf, i64 %s) #2 {
+; CHECK-LABEL: @test_amx_load(
+; CHECK-NEXT:  entry:
+; CHECK-NEXT:    [[T0:%.*]] = bitcast <256 x i32>* [[IN:%.*]] to x86_amx*
+; CHECK-NEXT:    [[TMP0:%.*]] = bitcast x86_amx* [[T0]] to i8*
+; CHECK-NEXT:    [[TMP1:%.*]] = call x86_amx @llvm.x86.tileloadd64.internal(i16 [[M:%.*]], i16 [[N:%.*]], i8* [[TMP0]], i64 64)
+; CHECK-NEXT:    call void @llvm.x86.tilestored64.internal(i16 [[M]], i16 [[N]], i8* [[BUF:%.*]], i64 [[S:%.*]], x86_amx [[TMP1]]) [[ATTR3]]
+; CHECK-NEXT:    ret void
+;
+entry:
+  %t0 = bitcast <256 x i32>* %in to x86_amx*
+  %t1 = load x86_amx, x86_amx* %t0, align 64
+  call void @llvm.x86.tilestored64.internal(i16 %m, i16 %n, i8* %buf, i64 %s, x86_amx %t1) #3
+  ret void
+}
+
+; test bitcast x86_amx to <256 x i32>
+define dso_local void @test_user_empty(i16 %m, i16 %n, i8 *%buf, i64 %s) #2 {
+; CHECK-LABEL: @test_user_empty(
+; CHECK-NEXT:  entry:
+; CHECK-NEXT:    ret void
+;
+entry:
+  %t1 = call x86_amx @llvm.x86.tileloadd64.internal(i16 %m, i16 %n, i8* %buf, i64 %s) #3
+  %t2 = bitcast x86_amx %t1 to <256 x i32>
+  ret void
+}
+
+; test bitcast <256 x i32> to x86_amx
+define dso_local void @test_user_empty2(<256 x i32> %in) #2 {
+; CHECK-LABEL: @test_user_empty2(
+; CHECK-NEXT:  entry:
+; CHECK-NEXT:    ret void
+;
+entry:
+  %t = bitcast <256 x i32> %in to x86_amx
+  ret void
+}
+
+define dso_local <256 x i32> @test_amx_load_bitcast(<256 x i32>* %in, i16 %m, i16 %n, i8 *%buf, i64 %s) #2 {
+; CHECK-LABEL: @test_amx_load_bitcast(
+; CHECK-NEXT:  entry:
+; CHECK-NEXT:    [[T1:%.*]] = load <256 x i32>, <256 x i32>* [[IN:%.*]], align 64
+; CHECK-NEXT:    [[TMP0:%.*]] = bitcast <256 x i32>* [[IN]] to i8*
+; CHECK-NEXT:    [[TMP1:%.*]] = call x86_amx @llvm.x86.tileloadd64.internal(i16 [[M:%.*]], i16 [[N:%.*]], i8* [[TMP0]], i64 64)
+; CHECK-NEXT:    call void @llvm.x86.tilestored64.internal(i16 [[M]], i16 [[N]], i8* [[BUF:%.*]], i64 [[S:%.*]], x86_amx [[TMP1]]) [[ATTR3]]
+; CHECK-NEXT:    ret <256 x i32> [[T1]]
+;
+entry:
+  %t1 = load <256 x i32>, <256 x i32>* %in, align 64
+  %t2 = bitcast <256 x i32> %t1 to x86_amx
+  call void @llvm.x86.tilestored64.internal(i16 %m, i16 %n, i8* %buf, i64 %s, x86_amx %t2) #3
+  ret <256 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) #2 {
+; CHECK-LABEL: @test_amx_bitcast_store(
+; CHECK-NEXT:  entry:
+; CHECK-NEXT:    [[T1:%.*]] = call x86_amx @llvm.x86.tileloadd64.internal(i16 [[M:%.*]], i16 [[M]], i8* [[BUF:%.*]], i64 [[S:%.*]]) [[ATTR3]]
+; CHECK-NEXT:    [[TMP0:%.*]] = bitcast <256 x i32>* [[OUT:%.*]] to i8*
+; CHECK-NEXT:    call void @llvm.x86.tilestored64.internal(i16 [[M]], i16 [[M]], i8* [[TMP0]], i64 64, x86_amx [[T1]])
+; CHECK-NEXT:    [[TMP1:%.*]] = load <256 x i32>, <256 x i32>* [[OUT]], align 1024
+; CHECK-NEXT:    ret <256 x i32> [[TMP1]]
+;
+entry:
+  %t1 = call x86_amx @llvm.x86.tileloadd64.internal(i16 %m, i16 %m, i8* %buf, i64 %s) #3
+  %t2 = bitcast x86_amx %t1 to <256 x i32>
+  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) #2 {
+; CHECK-LABEL: @test_src_add(
+; CHECK-NEXT:  entry:
+; CHECK-NEXT:    [[TMP0:%.*]] = alloca <256 x i32>, align 1024
+; 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:%.*]] = call x86_amx @llvm.x86.tileloadd64.internal(i16 [[R:%.*]], i16 [[C:%.*]], i8* [[TMP1]], i64 64)
+; CHECK-NEXT:    call void @llvm.x86.tilestored64.internal(i16 [[R]], i16 [[C]], i8* [[BUF:%.*]], i64 [[S:%.*]], x86_amx [[TMP2]]) [[ATTR3]]
+; CHECK-NEXT:    ret void
+;
+entry:
+  %add = add <256 x i32> %y, %x
+  %t = bitcast <256 x i32> %add to x86_amx
+  call void @llvm.x86.tilestored64.internal(i16 %r, i16 %c, i8* %buf, i64 %s, x86_amx %t) #3
+  ret void
+}
+
+define dso_local void @test_src_add2(<256 x i32> %x, i16 %r, i16 %c, i8* %buf, i64 %s) #2 {
+; CHECK-LABEL: @test_src_add2(
+; CHECK-NEXT:  entry:
+; CHECK-NEXT:    [[TMP0:%.*]] = alloca <256 x i32>, align 1024
+; CHECK-NEXT:    [[T1:%.*]] = call x86_amx @llvm.x86.tileloadd64.internal(i16 [[R:%.*]], i16 [[C:%.*]], i8* [[BUF:%.*]], i64 [[S:%.*]]) [[ATTR3]]
+; CHECK-NEXT:    [[TMP1:%.*]] = bitcast <256 x i32>* [[TMP0]] to i8*
+; CHECK-NEXT:    call void @llvm.x86.tilestored64.internal(i16 [[R]], i16 [[C]], i8* [[TMP1]], i64 64, x86_amx [[T1]])
+; CHECK-NEXT:    [[TMP2:%.*]] = load <256 x i32>, <256 x i32>* [[TMP0]], align 1024
+; CHECK-NEXT:    [[ADD:%.*]] = add <256 x i32> [[TMP2]], [[X:%.*]]
+; CHECK-NEXT:    ret void
+;
+entry:
+  %t1 = call x86_amx @llvm.x86.tileloadd64.internal(i16 %r, i16 %c, i8* %buf, i64 %s) #3
+  %t2 = bitcast x86_amx %t1 to <256 x i32>
+  %add = add <256 x i32> %t2, %x
+  ret void
+}
+
 define dso_local void @test_load(i8* %in, i8* %out) local_unnamed_addr #2 {
 ; CHECK-LABEL: @test_load(
 ; CHECK-NEXT:    [[TMP1:%.*]] = bitcast i8* [[IN:%.*]] to <256 x i32>*
 ; CHECK-NEXT:    [[TMP2:%.*]] = bitcast i8* [[OUT:%.*]] to <256 x i32>*
-; CHECK-NEXT:    [[TMP3:%.*]] = bitcast <256 x i32>* [[TMP1]] to <128 x i32>*
-; CHECK-NEXT:    [[TMP4:%.*]] = load <128 x i32>, <128 x i32>* [[TMP3]], align 64
-; CHECK-NEXT:    [[TMP5:%.*]] = getelementptr <128 x i32>, <128 x i32>* [[TMP3]], i32 1
-; CHECK-NEXT:    [[TMP6:%.*]] = load <128 x i32>, <128 x i32>* [[TMP5]], align 64
-; CHECK-NEXT:    [[TMP7:%.*]] = bitcast <256 x i32>* [[TMP2]] to <128 x i32>*
-; CHECK-NEXT:    store <128 x i32> [[TMP4]], <128 x i32>* [[TMP7]], align 64
-; CHECK-NEXT:    [[TMP8:%.*]] = getelementptr <128 x i32>, <128 x i32>* [[TMP7]], i32 1
-; CHECK-NEXT:    store <128 x i32> [[TMP6]], <128 x i32>* [[TMP8]], align 64
+; CHECK-NEXT:    [[TMP3:%.*]] = load <256 x i32>, <256 x i32>* [[TMP1]], align 64, [[TBAA2:!tbaa !.*]]
+; CHECK-NEXT:    store <256 x i32> [[TMP3]], <256 x i32>* [[TMP2]], align 64, [[TBAA2]]
 ; CHECK-NEXT:    ret void
 ;
   %1 = bitcast i8* %in to <256 x i32>*
   %2 = bitcast i8* %out to <256 x i32>*
   %3 = load <256 x i32>, <256 x i32>* %1, align 64, !tbaa !8
   store <256 x i32> %3, <256 x i32>* %2, align 64, !tbaa !8
   ret void
 }
 
+define dso_local <256 x i32> @foo(<256 x i32>* nocapture readonly byval(<256 x i32>) align 1024 %0, <256 x i32>* nocapture readonly byval(<256 x i32>) align 1024 %1) local_unnamed_addr #0 {
+; CHECK-LABEL: @foo(
+; CHECK-NEXT:  entry:
+; CHECK-NEXT:    [[X:%.*]] = load <256 x i32>, <256 x i32>* [[TMP0:%.*]], align 1024, [[TBAA5:!tbaa !.*]]
+; CHECK-NEXT:    [[Y:%.*]] = load <256 x i32>, <256 x i32>* [[TMP1:%.*]], align 1024, [[TBAA5]]
+; CHECK-NEXT:    [[ADD:%.*]] = add <256 x i32> [[Y]], [[X]]
+; CHECK-NEXT:    ret <256 x i32> [[ADD]]
+;
+entry:
+  %x = load <256 x i32>, <256 x i32>* %0, align 1024, !tbaa !2
+  %y = load <256 x i32>, <256 x i32>* %1, align 1024, !tbaa !2
+  %add = add <256 x i32> %y, %x
+  ret <256 x i32> %add
+}
+
 define dso_local void @__tile_loadd(%struct.__tile_str* nocapture %0, i8* %1, i64 %2) local_unnamed_addr #0 {
 ; CHECK-LABEL: @__tile_loadd(
 ; CHECK-NEXT:    [[TMP4:%.*]] = getelementptr inbounds [[STRUCT___TILE_STR:%.*]], %struct.__tile_str* [[TMP0:%.*]], i64 0, i32 0
-; CHECK-NEXT:    [[TMP5:%.*]] = load i16, i16* [[TMP4]], align 64, [[TBAA2:!tbaa !.*]]
+; CHECK-NEXT:    [[TMP5:%.*]] = load i16, i16* [[TMP4]], align 64, [[TBAA5]]
 ; CHECK-NEXT:    [[TMP6:%.*]] = getelementptr inbounds [[STRUCT___TILE_STR]], %struct.__tile_str* [[TMP0]], i64 0, i32 1
-; CHECK-NEXT:    [[TMP7:%.*]] = load i16, i16* [[TMP6]], align 2, [[TBAA7:!tbaa !.*]]
+; CHECK-NEXT:    [[TMP7:%.*]] = load i16, i16* [[TMP6]], align 2, [[TBAA8:!tbaa !.*]]
 ; CHECK-NEXT:    [[TMP8:%.*]] = shl i64 [[TMP2:%.*]], 32
 ; CHECK-NEXT:    [[TMP9:%.*]] = ashr exact i64 [[TMP8]], 32
-; CHECK-NEXT:    [[TMP10:%.*]] = tail call <256 x i32> @llvm.x86.tileloadd64.internal(i16 [[TMP5]], i16 [[TMP7]], i8* [[TMP1:%.*]], i64 [[TMP9]]) [[ATTR3:#.*]]
+; CHECK-NEXT:    [[TMP10:%.*]] = tail call x86_amx @llvm.x86.tileloadd64.internal(i16 [[TMP5]], i16 [[TMP7]], i8* [[TMP1:%.*]], i64 [[TMP9]]) [[ATTR3]]
 ; 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:    call void @llvm.x86.tilestored64.internal(i16 [[TMP5]], i16 [[TMP7]], i8* [[TMP12]], i64 64, <256 x i32> [[TMP10]])
+; CHECK-NEXT:    call void @llvm.x86.tilestored64.internal(i16 [[TMP5]], i16 [[TMP7]], i8* [[TMP12]], i64 64, x86_amx [[TMP10]])
 ; 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, !tbaa !2
   %6 = getelementptr inbounds %struct.__tile_str, %struct.__tile_str* %0, i64 0, i32 1
   %7 = load i16, i16* %6, align 2, !tbaa !7
   %8 = shl i64 %2, 32
   %9 = ashr exact i64 %8, 32
-  %10 = tail call <256 x i32> @llvm.x86.tileloadd64.internal(i16 %5, i16 %7, i8* %1, i64 %9) #3
-  %11 = getelementptr inbounds %struct.__tile_str, %struct.__tile_str* %0, i64 0, i32 2
-  store <256 x i32> %10, <256 x i32>* %11, align 64, !tbaa !8
+  %10 = tail call x86_amx @llvm.x86.tileloadd64.internal(i16 %5, i16 %7, i8* %1, i64 %9) #3
+  %11 = bitcast x86_amx %10 to <256 x i32>
+  %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, !tbaa !8
   ret void
 }
 
 define dso_local void @__tile_dpbsud(%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 #0 {
 ; CHECK-LABEL: @__tile_dpbsud(
 ; 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, [[TBAA2]]
+; CHECK-NEXT:    [[TMP5:%.*]] = load i16, i16* [[TMP4]], align 64, [[TBAA5]]
 ; 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, [[TBAA7]]
+; CHECK-NEXT:    [[TMP7:%.*]] = load i16, i16* [[TMP6]], align 2, [[TBAA8]]
 ; 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, [[TBAA7]]
+; CHECK-NEXT:    [[TMP9:%.*]] = load i16, i16* [[TMP8]], align 2, [[TBAA8]]
 ; 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 <256 x i32> @llvm.x86.tileloadd64.internal(i16 [[TMP5]], i16 [[TMP7]], i8* [[TMP11]], i64 64)
+; 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 <256 x i32> @llvm.x86.tileloadd64.internal(i16 [[TMP5]], i16 [[TMP9]], i8* [[TMP14]], i64 64)
+; 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 <256 x i32> @llvm.x86.tileloadd64.internal(i16 [[TMP9]], i16 [[TMP7]], i8* [[TMP17]], i64 64)
-; CHECK-NEXT:    [[TMP19:%.*]] = tail call <256 x i32> @llvm.x86.tdpbssd.internal(i16 [[TMP5]], i16 [[TMP7]], i16 [[TMP9]], <256 x i32> [[TMP12]], <256 x i32> [[TMP15]], <256 x i32> [[TMP18]]) [[ATTR3]]
+; 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]]) [[ATTR3]]
 ; 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, <256 x i32> [[TMP19]])
+; CHECK-NEXT:    call void @llvm.x86.tilestored64.internal(i16 [[TMP5]], i16 [[TMP7]], i8* [[TMP20]], i64 64, x86_amx [[TMP19]])
 ; 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, !tbaa !2
   %6 = getelementptr inbounds %struct.__tile_str, %struct.__tile_str* %2, i64 0, i32 1
   %7 = load i16, i16* %6, align 2, !tbaa !7
   %8 = getelementptr inbounds %struct.__tile_str, %struct.__tile_str* %1, i64 0, i32 1
   %9 = load i16, i16* %8, align 2, !tbaa !7
   %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, !tbaa !8
-  %12 = getelementptr inbounds %struct.__tile_str, %struct.__tile_str* %1, i64 0, i32 2
-  %13 = load <256 x i32>, <256 x i32>* %12, align 64, !tbaa !8
-  %14 = getelementptr inbounds %struct.__tile_str, %struct.__tile_str* %2, i64 0, i32 2
-  %15 = load <256 x i32>, <256 x i32>* %14, align 64, !tbaa !8
-  %16 = tail call <256 x i32> @llvm.x86.tdpbssd.internal(i16 %5, i16 %7, i16 %9, <256 x i32> %11, <256 x i32> %13, <256 x i32> %15) #3
-  store <256 x i32> %16, <256 x i32>* %10, align 64, !tbaa !8
+  %12 = bitcast <256 x i32> %11 to x86_amx
+  %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, !tbaa !8
+  %15 = bitcast <256 x i32> %14 to x86_amx
+  %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, !tbaa !8
+  %18 = bitcast <256 x i32> %17 to x86_amx
+  %19 = tail call x86_amx @llvm.x86.tdpbssd.internal(i16 %5, i16 %7, i16 %9, x86_amx %12, x86_amx %15, x86_amx %18) #3
+  %20 = bitcast x86_amx %19 to <256 x i32>
+  store <256 x i32> %20, <256 x i32>* %10, align 64, !tbaa !8
   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 #1 {
 ; CHECK-LABEL: @__tile_stored(
 ; CHECK-NEXT:    [[TMP4:%.*]] = getelementptr inbounds [[STRUCT___TILE_STR:%.*]], %struct.__tile_str* [[TMP2:%.*]], i64 0, i32 0
-; CHECK-NEXT:    [[TMP5:%.*]] = load i16, i16* [[TMP4]], align 64, [[TBAA2]]
+; CHECK-NEXT:    [[TMP5:%.*]] = load i16, i16* [[TMP4]], align 64, [[TBAA5]]
 ; 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, [[TBAA7]]
+; CHECK-NEXT:    [[TMP7:%.*]] = load i16, i16* [[TMP6]], align 2, [[TBAA8]]
 ; CHECK-NEXT:    [[TMP8:%.*]] = getelementptr inbounds [[STRUCT___TILE_STR]], %struct.__tile_str* [[TMP2]], i64 0, i32 2
 ; CHECK-NEXT:    [[TMP9:%.*]] = bitcast <256 x i32>* [[TMP8]] to i8*
-; CHECK-NEXT:    [[TMP10:%.*]] = call <256 x i32> @llvm.x86.tileloadd64.internal(i16 [[TMP5]], i16 [[TMP7]], i8* [[TMP9]], i64 64)
+; CHECK-NEXT:    [[TMP10:%.*]] = call x86_amx @llvm.x86.tileloadd64.internal(i16 [[TMP5]], i16 [[TMP7]], i8* [[TMP9]], i64 64)
 ; CHECK-NEXT:    [[TMP11:%.*]] = shl i64 [[TMP1:%.*]], 32
 ; CHECK-NEXT:    [[TMP12:%.*]] = ashr exact i64 [[TMP11]], 32
-; CHECK-NEXT:    tail call void @llvm.x86.tilestored64.internal(i16 [[TMP5]], i16 [[TMP7]], i8* [[TMP0:%.*]], i64 [[TMP12]], <256 x i32> [[TMP10]]) [[ATTR3]]
+; CHECK-NEXT:    tail call void @llvm.x86.tilestored64.internal(i16 [[TMP5]], i16 [[TMP7]], i8* [[TMP0:%.*]], i64 [[TMP12]], x86_amx [[TMP10]]) [[ATTR3]]
 ; 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, !tbaa !2
   %6 = getelementptr inbounds %struct.__tile_str, %struct.__tile_str* %2, i64 0, i32 1
   %7 = load i16, i16* %6, align 2, !tbaa !7
   %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, !tbaa !8
-  %10 = shl i64 %1, 32
-  %11 = ashr exact i64 %10, 32
-  tail call void @llvm.x86.tilestored64.internal(i16 %5, i16 %7, i8* %0, i64 %11, <256 x i32> %9) #3
+  %10 = bitcast <256 x i32> %9 to x86_amx
+  %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) #3
   ret void
 }
 
-declare <256 x i32> @llvm.x86.tileloadd64.internal(i16, i16, i8*, i64) #3
-declare <256 x i32> @llvm.x86.tdpbssd.internal(i16, i16, i16, <256 x i32>, <256 x i32>, <256 x i32>) #3
-declare void @llvm.x86.tilestored64.internal(i16, i16, i8*, i64, <256 x i32>) #3
+declare x86_amx @llvm.x86.tileloadd64.internal(i16, i16, i8*, i64) #3
+declare x86_amx @llvm.x86.tdpbssd.internal(i16, i16, i16, x86_amx, x86_amx, x86_amx) #3
+declare void @llvm.x86.tilestored64.internal(i16, i16, i8*, i64, x86_amx) #3
 
 attributes #0 = { alwaysinline nounwind uwtable "correctly-rounded-divide-sqrt-fp-math"="false" "disable-tail-calls"="false" "frame-pointer"="none" "less-precise-fpmad"="false" "min-legal-vector-width"="8192" "no-infs-fp-math"="false" "no-jump-tables"="false" "no-nans-fp-math"="false" "no-signed-zeros-fp-math"="false" "no-trapping-math"="true" "stack-protector-buffer-size"="8" "target-cpu"="x86-64" "target-features"="+amx-int8,+amx-tile,+cx8,+fxsr,+mmx,+sse,+sse2,+x87" "tune-cpu"="generic" "unsafe-fp-math"="false" "use-soft-float"="false" }
 attributes #1 = { alwaysinline nounwind uwtable "correctly-rounded-divide-sqrt-fp-math"="false" "disable-tail-calls"="false" "frame-pointer"="none" "less-precise-fpmad"="false" "min-legal-vector-width"="0" "no-infs-fp-math"="false" "no-jump-tables"="false" "no-nans-fp-math"="false" "no-signed-zeros-fp-math"="false" "no-trapping-math"="true" "stack-protector-buffer-size"="8" "target-cpu"="x86-64" "target-features"="+amx-int8,+amx-tile,+cx8,+fxsr,+mmx,+sse,+sse2,+x87" "tune-cpu"="generic" "unsafe-fp-math"="false" "use-soft-float"="false" }
 attributes #2 = { alwaysinline nounwind uwtable "correctly-rounded-divide-sqrt-fp-math"="false" "disable-tail-calls"="false" "frame-pointer"="none" "less-precise-fpmad"="false" "min-legal-vector-width"="0" "no-infs-fp-math"="false" "no-jump-tables"="false" "no-nans-fp-math"="false" "no-signed-zeros-fp-math"="false" "no-trapping-math"="true" "stack-protector-buffer-size"="8" "target-cpu"="x86-64" "target-features"="+amx-int8,+amx-tile,+avx,+avx2,+avx512f,+cx8,+f16c,+fma,+fxsr,+mmx,+popcnt,+sse,+sse2,+sse3,+sse4.1,+sse4.2,+ssse3,+x87,+xsave" "tune-cpu"="generic" "unsafe-fp-math"="false" "use-soft-float"="false" }
 attributes #3 = { nounwind }
 
 !llvm.module.flags = !{!0}
 !llvm.ident = !{!1}

llvm/utils/TableGen/CodeGenTarget.cpp

StringRef llvm::getEnumName(MVT::SimpleValueType T) {
   case MVT::f16:      return "MVT::f16";
   case MVT::bf16:     return "MVT::bf16";
   case MVT::f32:      return "MVT::f32";
   case MVT::f64:      return "MVT::f64";
   case MVT::f80:      return "MVT::f80";
   case MVT::f128:     return "MVT::f128";
   case MVT::ppcf128:  return "MVT::ppcf128";
   case MVT::x86mmx:   return "MVT::x86mmx";
+  case MVT::x86amx:   return "MVT::x86amx";
   case MVT::Glue:     return "MVT::Glue";
   case MVT::isVoid:   return "MVT::isVoid";
   case MVT::v1i1:     return "MVT::v1i1";
   case MVT::v2i1:     return "MVT::v2i1";
   case MVT::v4i1:     return "MVT::v4i1";
   case MVT::v8i1:     return "MVT::v8i1";
   case MVT::v16i1:    return "MVT::v16i1";
   case MVT::v32i1:    return "MVT::v32i1";

llvm/utils/TableGen/IntrinsicEmitter.cpp

void IntrinsicEmitter::EmitIntrinsicToOverloadTable(
   OS << "#endif\n\n";
 }
 
 
 // NOTE: This must be kept in synch with the copy in lib/IR/Function.cpp!
 enum IIT_Info {
   // Common values should be encoded with 0-15.
   IIT_Done = 0,
   IIT_I1   = 1,
   IIT_I8   = 2,
   IIT_I16  = 3,
   IIT_I32  = 4,
   IIT_I64  = 5,
   IIT_F16  = 6,
   IIT_F32  = 7,
   IIT_F64  = 8,
   IIT_V2   = 9,
   IIT_V4   = 10,
   IIT_V8   = 11,
   IIT_V16  = 12,
   IIT_V32  = 13,
   IIT_PTR  = 14,
   IIT_ARG  = 15,
 
   // Values from 16+ are only encodable with the inefficient encoding.
   IIT_V64  = 16,
   IIT_MMX  = 17,
   IIT_TOKEN = 18,
   IIT_METADATA = 19,
   IIT_EMPTYSTRUCT = 20,
   IIT_STRUCT2 = 21,
   IIT_STRUCT3 = 22,
   IIT_STRUCT4 = 23,
   IIT_STRUCT5 = 24,
   IIT_EXTEND_ARG = 25,
   IIT_TRUNC_ARG = 26,
   IIT_ANYPTR = 27,
   IIT_V1   = 28,
   IIT_VARARG = 29,
   IIT_HALF_VEC_ARG = 30,
   IIT_SAME_VEC_WIDTH_ARG = 31,
   IIT_PTR_TO_ARG = 32,
   IIT_PTR_TO_ELT = 33,
   IIT_VEC_OF_ANYPTRS_TO_ELT = 34,
   IIT_I128 = 35,
   IIT_V512 = 36,
   IIT_V1024 = 37,
   IIT_STRUCT6 = 38,
   IIT_STRUCT7 = 39,
   IIT_STRUCT8 = 40,
   IIT_F128 = 41,
   IIT_VEC_ELEMENT = 42,
   IIT_SCALABLE_VEC = 43,
   IIT_SUBDIVIDE2_ARG = 44,
   IIT_SUBDIVIDE4_ARG = 45,
   IIT_VEC_OF_BITCASTS_TO_INT = 46,
   IIT_V128 = 47,
   IIT_BF16 = 48,
   IIT_STRUCT9 = 49,
-  IIT_V256 = 50
+  IIT_V256 = 50,
+  IIT_AMX  = 51
 };
 
 static void EncodeFixedValueType(MVT::SimpleValueType VT,
                                  std::vector<unsigned char> &Sig) {
   if (MVT(VT).isInteger()) {
     unsigned BitWidth = MVT(VT).getFixedSizeInBits();
     switch (BitWidth) {
     default: PrintFatalError("unhandled integer type width in intrinsic!");
static void EncodeFixedValueType(MVT::SimpleValueType VT,
   case MVT::f16: return Sig.push_back(IIT_F16);
   case MVT::bf16: return Sig.push_back(IIT_BF16);
   case MVT::f32: return Sig.push_back(IIT_F32);
   case MVT::f64: return Sig.push_back(IIT_F64);
   case MVT::f128: return Sig.push_back(IIT_F128);
   case MVT::token: return Sig.push_back(IIT_TOKEN);
   case MVT::Metadata: return Sig.push_back(IIT_METADATA);
   case MVT::x86mmx: return Sig.push_back(IIT_MMX);
+  case MVT::x86amx: return Sig.push_back(IIT_AMX);
   // MVT::OtherVT is used to mean the empty struct type here.
   case MVT::Other: return Sig.push_back(IIT_EMPTYSTRUCT);
   // MVT::isVoid is used to represent varargs here.
   case MVT::isVoid: return Sig.push_back(IIT_VARARG);
   }
 }
 
 #if defined(_MSC_VER) && !defined(__clang__)