[DirectX backend] support resource array when lower typed buffer ld/st to DXIL.

[python3kgae]

Summary

A new kind of Memory offset is added to calculate array index.
Use data layout with ptr size of 1 to help to get the array index with EmitGEPOffset and DL.getIndexedOffsetInType.
Added HandleCache to reuse DXIL Handle.

llvm/lib/Target/DirectX/DXILResource.h

class UAVResource : public ResourceBase {
   bool HasCounter;
   bool IsROV;
   ResourceBase::ExtendedProperties ExtProps;
 
   void parseSourceType(StringRef S);
 
 public:
   UAVResource(uint32_t I, hlsl::FrontendResource R);
-
+  ResourceBase::Kinds getShape() const { return Shape; }
   MDNode *write() const;
   void print(raw_ostream &O) const;
 };
 
 // FIXME: Fully computing the resource structures requires analyzing the IR
 // because some flags are set based on what operations are performed on the
 // resource. This partial patch handles some of the leg work, but not all of it.
 // See issue https://github.com/llvm/llvm-project/issues/57936.

llvm/lib/Target/DirectX/DXILTypedBufferLowering.cpp

 #include "MemAccessLowerHelper.h"
 #include "llvm/ADT/SmallVector.h"
 #include "llvm/Analysis/Utils/Local.h"
 #include "llvm/CodeGen/Passes.h"
 #include "llvm/IR/GetElementPtrTypeIterator.h"
 #include "llvm/IR/IRBuilder.h"
 #include "llvm/IR/Instruction.h"
 #include "llvm/IR/Instructions.h"
-#include "llvm/IR/IntrinsicInst.h"
-#include "llvm/IR/Intrinsics.h"
 #include "llvm/IR/IntrinsicsDirectX.h"
 #include "llvm/IR/Metadata.h"
 #include "llvm/IR/Module.h"
 #include "llvm/IR/PassManager.h"
 #include "llvm/Pass.h"
 #include "llvm/Support/ErrorHandling.h"
 
 #define DEBUG_TYPE "dxil-typedbuf-lower"
 
 using namespace llvm;
 using namespace llvm::dxil;
 
+namespace {
+class HandleCache {
+  Module &M;
+  SmallDenseMap<Value *, CallInst *> HandleMap;
+  uint64_t RangeID;
+  ConstantInt *UAVIndex;
+
+public:
+  HandleCache(const UAVResource &UAV, Module &M)
+      : M(M), RangeID(UAV.getRangeID()),
+        UAVIndex(ConstantInt::get(Type::getInt32Ty(M.getContext()),
+                                  UAV.getIndex())) {}
+
+  CallInst *getHandle(Instruction *User, Value *ResourceOffset) {
+    // make handle for User.
+    Function *F = User->getParent()->getParent();
+    Value *ResIndex = UAVIndex;
+    IRBuilder<> B(ResIndex->getContext());
+    if (auto *ResourceOffsetI = dyn_cast_or_null<Instruction>(ResourceOffset))
+      B.SetInsertPoint(ResourceOffsetI->getNextNode());
+    else
+      B.SetInsertPoint(&*F->getEntryBlock().getFirstInsertionPt());
+
+    if (ResourceOffset)
+      ResIndex = B.CreateAdd(
+          ResIndex, B.CreateZExtOrTrunc(ResourceOffset, ResIndex->getType()));
+
+    auto It = HandleMap.find(ResIndex);
+    CallInst *Hdl = nullptr;
+    if (It == HandleMap.end()) {
+      DXILOpBuilder DXILB(M, B);
+      Hdl = DXILB.createCreateHandle(
+          static_cast<unsigned>(hlsl::ResourceClass::UAV), RangeID, ResIndex,
+          false);
+      HandleMap[ResIndex] = Hdl;
+    } else
+      Hdl = It->second;
+    return Hdl;
+  }
+};
+} // namespace
+
 static bool lowerUAVAccess(Module &M, dxil::Resources &Res) {
   // FIXME: Allocate resource binding first.
   //   See https://github.com/llvm/llvm-project/issues/58051
   const SmallVector<UAVResource> &UAVs = Res.getUAVs();
 
   if (UAVs.empty())
     return false;
   const DataLayout &DL = M.getDataLayout();
-  LLVMContext &Ctx = M.getContext();
+  // Datalayout to help calculate array index by set ptr size to 1 byte.
+  // So the offset will be the array index.
+  // Pointer size, abi alignment, prefered alignment, index size.
+  DataLayout ResourceDL("p:8:8:8:32");
   for (const auto &UAV : UAVs) {
+    if (UAV.getShape() != hlsl::ResourceKind::TypedBuffer)
+      continue;
+    assert(UAV.getIndex() != -1U && "unallocated binding");
     GlobalVariable *GV = UAV.getVariable();
+
+    std::vector<MemAccessLowerHelper::ArrayMemAccess> ArrayAccessList;
+    MemAccessLowerHelper::collectMemAccess(GV, ArrayAccessList, ResourceDL);
+    // Save Resource Index and number of users for the Resource Index.
+    std::vector<std::pair<Value *, uint32_t>> ResIndexList;
     std::vector<MemAccessLowerHelper::TypedBufMemAccess> AccessList;
-    for (User *U : make_early_inc_range(GV->users())) {
-      if (auto *SI = dyn_cast<StoreInst>(U)) {
+    for (auto &Access : ArrayAccessList) {
+      uint32_t Count = AccessList.size();
+      if (auto *SI = dyn_cast<StoreInst>(Access.User)) {
         Value *V = SI->getValueOperand();
         // FIXME: use createHandle generated in clangCodeGen.
         // See https://github.com/llvm/llvm-project/issues/58031.
         SI->eraseFromParent();
         MemAccessLowerHelper::collectMemAccess(V, AccessList, DL);
-      } else if (auto *IntrinsicCI = dyn_cast<IntrinsicInst>(U)) {
-        if (IntrinsicCI->getIntrinsicID() == Intrinsic::invariant_start ||
-            IntrinsicCI->getIntrinsicID() == Intrinsic::invariant_end)
-          IntrinsicCI->eraseFromParent();
-      } else if (LoadInst *LI = dyn_cast<LoadInst>(U))
+      } else if (auto *LI = dyn_cast<LoadInst>(Access.User))
         MemAccessLowerHelper::collectMemAccess(LI, AccessList, DL);
-      // FIXME: support array of resource.
-      //   See https://github.com/llvm/llvm-project/issues/58486
+
+      if (uint32_t Size = AccessList.size() - Count)
+        ResIndexList.emplace_back(std::make_pair(Access.Offset.Index, Size));
     }
 
+    HandleCache Handles(UAV, M);
     SmallDenseMap<Function *, CallInst *> HandleMap;
-    uint64_t RangeID = UAV.getRangeID();
-    ConstantInt *UAVIndex =
-        ConstantInt::get(Type::getInt32Ty(Ctx), UAV.getIndex());
-    assert(UAVIndex->getLimitedValue() != -1ULL && "unallocated binding");
+    auto ResIndexIt = ResIndexList.begin();
     for (auto &Access : AccessList) {
-      Value *Index = Access.Offset.Index;
+      Value *ResIndex = ResIndexIt->first;
+      --(ResIndexIt->second);
+      // Update Resource Index when number of users is 0.
+      if (!ResIndexIt->second)
+        ++ResIndexIt;
+
       Instruction *User = Access.User;
-      // make handle in the Function.
-      Function *F = User->getParent()->getParent();
-      auto It = HandleMap.find(F);
-      CallInst *Hdl = nullptr;
-      if (It == HandleMap.end()) {
-        IRBuilder<> B(&*F->getEntryBlock().getFirstInsertionPt());
-        DXILOpBuilder DXILB(M, B);
-        Hdl = DXILB.createCreateHandle(
-            static_cast<unsigned>(hlsl::ResourceClass::UAV), RangeID, UAVIndex,
-            false);
-        HandleMap[F] = Hdl;
-      } else
-        Hdl = It->second;
+      CallInst *Hdl = Handles.getHandle(User, ResIndex);
 
+      Value *Index = Access.Offset.Index;
       if (StoreInst *SI = dyn_cast<StoreInst>(User)) {
         Value *V = SI->getValueOperand();
         // Skip the store on hdl.
         if (SI->getPointerOperand() == GV)
           continue;
         Type *Ty = V->getType();
         IRBuilder<> B(SI);
         DXILOpBuilder DXILB(M, B);

llvm/lib/Target/DirectX/MemAccessLowerHelper.h

 
 // For Buffer like RWBuffer<float> U.
 // Index is I for U[I].
 struct TypedBufMemOffset {
   Value *Index = nullptr;
   TypedBufMemOffset appendGEPOffset(GEPOperator *GEP, const DataLayout &DL);
 };
 
+// For array like RWBuffer<float> U[2][2].
+// Index is 2*X + Y for U[X][Y];
+struct ArrayOffset {
+  Value *Index = nullptr;
+  ArrayOffset appendGEPOffset(GEPOperator *GEP, const DataLayout &DL);
+};
+
 template <typename OffsetT> struct MemAccess {
   OffsetT Offset;
   Instruction *User; // The instruction which access memory like Load/Store.
 };
 
 using TypedBufMemAccess = MemAccess<TypedBufMemOffset>;
+using ArrayMemAccess = MemAccess<ArrayOffset>;
 
 /// Collect all memory access for typed buffer global
 /// variable \c Ptr.
 void collectMemAccess(llvm::Value *Ptr,
                       std::vector<TypedBufMemAccess> &AccessList,
                       const llvm::DataLayout &DL);
+/// Collect all memory access for array global
+/// variable \c Ptr.
+void collectMemAccess(llvm::Value *Ptr, std::vector<ArrayMemAccess> &AccessList,
+                      const llvm::DataLayout &DL);
 
 } // namespace MemAccessLowerHelper
 
 } // namespace dxil
 
 } // namespace llvm
 
 #endif // LLVM_TARGET_DIRECTX_MEMACCESSLOWERHELPER_H

llvm/lib/Target/DirectX/MemAccessLowerHelper.cpp

 //===- MemAccessLowerHelper.cpp - Mem access helper -------------*- 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
 //
 //===----------------------------------------------------------------------===//
 /// Utils to help lowering memory access for resources.
 //===----------------------------------------------------------------------===//
 
 #include "MemAccessLowerHelper.h"
 #include "llvm/Analysis/Utils/Local.h"
 #include "llvm/IR/Constants.h"
 #include "llvm/IR/GlobalVariable.h"
 #include "llvm/IR/IRBuilder.h"
 #include "llvm/IR/Instructions.h"
+#include "llvm/IR/IntrinsicInst.h"
+#include "llvm/IR/Intrinsics.h"
 #include "llvm/IR/Operator.h"
 
 using namespace llvm;
 using namespace llvm::dxil::MemAccessLowerHelper;
 
 TypedBufMemOffset
 llvm::dxil::MemAccessLowerHelper::TypedBufMemOffset::appendGEPOffset(
     GEPOperator *GEP, const DataLayout &DL) {

 }
 
 /// Collect all memory access for non-legacy cbuffer/ typed buffer global
 /// variable \c Ptr.
 template <typename T>
 static void collectMemAccess(llvm::Value *Ptr,
                              std::vector<MemAccess<T>> &AccessList,
                              const llvm::DataLayout &DL) {
-  for (llvm::User *U : Ptr->users()) {
+  for (llvm::User *U : make_early_inc_range(Ptr->users())) {
+    if (auto *IntrinsicCI = dyn_cast<IntrinsicInst>(U)) {
+      if (IntrinsicCI->getIntrinsicID() == Intrinsic::invariant_start ||
+          IntrinsicCI->getIntrinsicID() == Intrinsic::invariant_end) {
+        IntrinsicCI->eraseFromParent();
+        continue;
+      }
+    }
     T Offset;
     ::collectUserMemAccess<T>(U, Offset, AccessList, DL);
   }
 }
 
 void llvm::dxil::MemAccessLowerHelper::collectMemAccess(
     llvm::Value *Ptr, std::vector<TypedBufMemAccess> &AccessList,
     const llvm::DataLayout &DL) {
   return ::collectMemAccess(Ptr, AccessList, DL);
 }
+
+void llvm::dxil::MemAccessLowerHelper::collectMemAccess(
+    llvm::Value *Ptr, std::vector<ArrayMemAccess> &AccessList,
+    const llvm::DataLayout &DL) {
+  return ::collectMemAccess(Ptr, AccessList, DL);
+}
+
+ArrayOffset llvm::dxil::MemAccessLowerHelper::ArrayOffset::appendGEPOffset(
+    GEPOperator *GEP, const DataLayout &DL) {
+  Value *NewIndex = Index;
+  IRBuilder<> B(GEP->getContext());
+  if (auto *Inst = dyn_cast<Instruction>(GEP))
+    B.SetInsertPoint(Inst);
+  if (GEP->hasAllConstantIndices()) {
+    SmallVector<Value *> IdxList(GEP->idx_begin(), GEP->idx_end());
+    NewIndex = B.getInt32(
+        DL.getIndexedOffsetInType(GEP->getSourceElementType(), IdxList));
+    if (Index)
+      NewIndex = B.CreateAdd(Index, NewIndex);
+  } else {
+    Value *Offset = EmitGEPOffset(&B, DL, GEP, /*NoAssumptions=*/true);
+    if (Index)
+      NewIndex = B.CreateAdd(Index, Offset);
+    else
+      NewIndex = Offset;
+  }
+  return {NewIndex};
+}

llvm/test/CodeGen/DirectX/typed_buf_array.ll

+; RUN: opt -S -dxil-typedbuf-lower < %s | FileCheck %s
+target datalayout = "e-m:e-p:32:32-i1:32-i8:8-i16:16-i32:32-i64:64-f16:16-f32:32-f64:64-n8:16:32:64"
+target triple = "dxil-unknown-shadermodel6.0-compute"
+
+;RWBuffer<float> AU[2][64] : register(u2);
+;RWBuffer<float> AUI[64][2] : register(u2, space1);
+;AU[ID%2][ID%64][ID] = AUI[5][1][ID];
+
+; Check index for RWBuffer<float> AUI[64][2]: register(u2, space1); AUI[5][1] -> 11 + 2 = 13.
+;CHECK: %[[InputHDL:.+]] = call %dx.types.Handle @dx.op.createHandle(i32 57, i8 1, i32 1, i32 13, i1 false)
+;CHECK: %[[ID:.+]] = tail call i32 @llvm.dx.thread.id(i32 0)
+
+;CHECK:%[[LD:.+]] = call %dx.types.ResRet.f32 @dx.op.bufferLoad.f32(i32 68, %dx.types.Handle %[[InputHDL]], i32 %[[ID]], i32 poison)
+;CHECK:%[[LD_X:.+]] = extractvalue %dx.types.ResRet.f32 %[[LD]], 0
+
+; Check index for RWBuffer<float> AU[2][64] : register(u2); AU[ID%2][ID%64].
+;CHECK: %[[REM2:.+]] = and i32 %[[ID]], 1
+;CHECK: %[[REM64:.+]] = and i32 %[[ID]], 63
+;CHECK: %[[ArrayIndexHigh:.+]] = mul i32 %[[REM2]], 64
+;CHECK: %[[ArrayIndex:.+]] = add i32 %[[ArrayIndexHigh]], %[[REM64]]
+;CHECK: %[[ResIndex:.+]] = add i32 2, %[[ArrayIndex]]
+
+;CHECK:%[[OutputHDL:.+]] = call %dx.types.Handle @dx.op.createHandle(i32 57, i8 1, i32 0, i32 %[[ResIndex]], i1 false)
+;CHECK:call void @dx.op.bufferStore.f32(i32 69, %dx.types.Handle %[[OutputHDL]], i32 %[[ID]], i32 poison, float %[[LD_X]], float poison, float poison, float poison, i8 1)
+;CHECK-NEXT:ret void
+
+%"class.hlsl::RWBuffer" = type { ptr }
+
+@"?AU@@3PAY0EA@V?$RWBuffer@M@hlsl@@A" = local_unnamed_addr global [2 x [64 x %"class.hlsl::RWBuffer"]] zeroinitializer, align 4
+@"?AUI@@3PAY01V?$RWBuffer@M@hlsl@@A" = local_unnamed_addr global [64 x [2 x %"class.hlsl::RWBuffer"]] zeroinitializer, align 4
+
+; Function Attrs: mustprogress norecurse nounwind willreturn
+define void @main() local_unnamed_addr #0 {
+entry:
+  %0 = tail call ptr @llvm.dx.create.handle(i8 1)
+  store ptr %0, ptr @"?AU@@3PAY0EA@V?$RWBuffer@M@hlsl@@A", align 4, !tbaa !6
+  %1 = tail call ptr @llvm.dx.create.handle(i8 1)
+  store ptr %1, ptr getelementptr inbounds ([2 x [64 x %"class.hlsl::RWBuffer"]], ptr @"?AU@@3PAY0EA@V?$RWBuffer@M@hlsl@@A", i32 0, i32 0, i32 1), align 4, !tbaa !6
+  %2 = tail call ptr @llvm.dx.create.handle(i8 1)
+  store ptr %2, ptr getelementptr inbounds ([2 x [64 x %"class.hlsl::RWBuffer"]], ptr @"?AU@@3PAY0EA@V?$RWBuffer@M@hlsl@@A", i32 0, i32 0, i32 2), align 4, !tbaa !6
+  %3 = tail call ptr @llvm.dx.create.handle(i8 1)
+  store ptr %3, ptr getelementptr inbounds ([2 x [64 x %"class.hlsl::RWBuffer"]], ptr @"?AU@@3PAY0EA@V?$RWBuffer@M@hlsl@@A", i32 0, i32 0, i32 3), align 4, !tbaa !6
+  %4 = tail call ptr @llvm.dx.create.handle(i8 1)
+  store ptr %4, ptr getelementptr inbounds ([2 x [64 x %"class.hlsl::RWBuffer"]], ptr @"?AU@@3PAY0EA@V?$RWBuffer@M@hlsl@@A", i32 0, i32 0, i32 4), align 4, !tbaa !6
+  %5 = tail call ptr @llvm.dx.create.handle(i8 1)
+  store ptr %5, ptr getelementptr inbounds ([2 x [64 x %"class.hlsl::RWBuffer"]], ptr @"?AU@@3PAY0EA@V?$RWBuffer@M@hlsl@@A", i32 0, i32 0, i32 5), align 4, !tbaa !6
+  %6 = tail call ptr @llvm.dx.create.handle(i8 1)
+  store ptr %6, ptr getelementptr inbounds ([2 x [64 x %"class.hlsl::RWBuffer"]], ptr @"?AU@@3PAY0EA@V?$RWBuffer@M@hlsl@@A", i32 0, i32 0, i32 6), align 4, !tbaa !6
+  %7 = tail call ptr @llvm.dx.create.handle(i8 1)
+  store ptr %7, ptr getelementptr inbounds ([2 x [64 x %"class.hlsl::RWBuffer"]], ptr @"?AU@@3PAY0EA@V?$RWBuffer@M@hlsl@@A", i32 0, i32 0, i32 7), align 4, !tbaa !6
+  %8 = tail call ptr @llvm.dx.create.handle(i8 1)
+  store ptr %8, ptr getelementptr inbounds ([2 x [64 x %"class.hlsl::RWBuffer"]], ptr @"?AU@@3PAY0EA@V?$RWBuffer@M@hlsl@@A", i32 0, i32 0, i32 8), align 4, !tbaa !6
+  %9 = tail call ptr @llvm.dx.create.handle(i8 1)
+  store ptr %9, ptr getelementptr inbounds ([2 x [64 x %"class.hlsl::RWBuffer"]], ptr @"?AU@@3PAY0EA@V?$RWBuffer@M@hlsl@@A", i32 0, i32 0, i32 9), align 4, !tbaa !6
+  %10 = tail call ptr @llvm.dx.create.handle(i8 1)
+  store ptr %10, ptr getelementptr inbounds ([2 x [64 x %"class.hlsl::RWBuffer"]], ptr @"?AU@@3PAY0EA@V?$RWBuffer@M@hlsl@@A", i32 0, i32 0, i32 10), align 4, !tbaa !6
+  %11 = tail call ptr @llvm.dx.create.handle(i8 1)
+  store ptr %11, ptr getelementptr inbounds ([2 x [64 x %"class.hlsl::RWBuffer"]], ptr @"?AU@@3PAY0EA@V?$RWBuffer@M@hlsl@@A", i32 0, i32 0, i32 11), align 4, !tbaa !6
+  %12 = tail call ptr @llvm.dx.create.handle(i8 1)
+  store ptr %12, ptr getelementptr inbounds ([2 x [64 x %"class.hlsl::RWBuffer"]], ptr @"?AU@@3PAY0EA@V?$RWBuffer@M@hlsl@@A", i32 0, i32 0, i32 12), align 4, !tbaa !6
+  %13 = tail call ptr @llvm.dx.create.handle(i8 1)
+  store ptr %13, ptr getelementptr inbounds ([2 x [64 x %"class.hlsl::RWBuffer"]], ptr @"?AU@@3PAY0EA@V?$RWBuffer@M@hlsl@@A", i32 0, i32 0, i32 13), align 4, !tbaa !6
+  %14 = tail call ptr @llvm.dx.create.handle(i8 1)
+  store ptr %14, ptr getelementptr inbounds ([2 x [64 x %"class.hlsl::RWBuffer"]], ptr @"?AU@@3PAY0EA@V?$RWBuffer@M@hlsl@@A", i32 0, i32 0, i32 14), align 4, !tbaa !6
+  %15 = tail call ptr @llvm.dx.create.handle(i8 1)
+  store ptr %15, ptr getelementptr inbounds ([2 x [64 x %"class.hlsl::RWBuffer"]], ptr @"?AU@@3PAY0EA@V?$RWBuffer@M@hlsl@@A", i32 0, i32 0, i32 15), align 4, !tbaa !6
+  %16 = tail call ptr @llvm.dx.create.handle(i8 1)
+  store ptr %16, ptr getelementptr inbounds ([2 x [64 x %"class.hlsl::RWBuffer"]], ptr @"?AU@@3PAY0EA@V?$RWBuffer@M@hlsl@@A", i32 0, i32 0, i32 16), align 4, !tbaa !6
+  %17 = tail call ptr @llvm.dx.create.handle(i8 1)
+  store ptr %17, ptr getelementptr inbounds ([2 x [64 x %"class.hlsl::RWBuffer"]], ptr @"?AU@@3PAY0EA@V?$RWBuffer@M@hlsl@@A", i32 0, i32 0, i32 17), align 4, !tbaa !6
+  %18 = tail call ptr @llvm.dx.create.handle(i8 1)
+  store ptr %18, ptr getelementptr inbounds ([2 x [64 x %"class.hlsl::RWBuffer"]], ptr @"?AU@@3PAY0EA@V?$RWBuffer@M@hlsl@@A", i32 0, i32 0, i32 18), align 4, !tbaa !6
+  %19 = tail call ptr @llvm.dx.create.handle(i8 1)
+  store ptr %19, ptr getelementptr inbounds ([2 x [64 x %"class.hlsl::RWBuffer"]], ptr @"?AU@@3PAY0EA@V?$RWBuffer@M@hlsl@@A", i32 0, i32 0, i32 19), align 4, !tbaa !6
+  %20 = tail call ptr @llvm.dx.create.handle(i8 1)
+  store ptr %20, ptr getelementptr inbounds ([2 x [64 x %"class.hlsl::RWBuffer"]], ptr @"?AU@@3PAY0EA@V?$RWBuffer@M@hlsl@@A", i32 0, i32 0, i32 20), align 4, !tbaa !6
+  %21 = tail call ptr @llvm.dx.create.handle(i8 1)
+  store ptr %21, ptr getelementptr inbounds ([2 x [64 x %"class.hlsl::RWBuffer"]], ptr @"?AU@@3PAY0EA@V?$RWBuffer@M@hlsl@@A", i32 0, i32 0, i32 21), align 4, !tbaa !6
+  %22 = tail call ptr @llvm.dx.create.handle(i8 1)
+  store ptr %22, ptr getelementptr inbounds ([2 x [64 x %"class.hlsl::RWBuffer"]], ptr @"?AU@@3PAY0EA@V?$RWBuffer@M@hlsl@@A", i32 0, i32 0, i32 22), align 4, !tbaa !6
+  %23 = tail call ptr @llvm.dx.create.handle(i8 1)
+  store ptr %23, ptr getelementptr inbounds ([2 x [64 x %"class.hlsl::RWBuffer"]], ptr @"?AU@@3PAY0EA@V?$RWBuffer@M@hlsl@@A", i32 0, i32 0, i32 23), align 4, !tbaa !6
+  %24 = tail call ptr @llvm.dx.create.handle(i8 1)
+  store ptr %24, ptr getelementptr inbounds ([2 x [64 x %"class.hlsl::RWBuffer"]], ptr @"?AU@@3PAY0EA@V?$RWBuffer@M@hlsl@@A", i32 0, i32 0, i32 24), align 4, !tbaa !6
+  %25 = tail call ptr @llvm.dx.create.handle(i8 1)
+  store ptr %25, ptr getelementptr inbounds ([2 x [64 x %"class.hlsl::RWBuffer"]], ptr @"?AU@@3PAY0EA@V?$RWBuffer@M@hlsl@@A", i32 0, i32 0, i32 25), align 4, !tbaa !6
+  %26 = tail call ptr @llvm.dx.create.handle(i8 1)
+  store ptr %26, ptr getelementptr inbounds ([2 x [64 x %"class.hlsl::RWBuffer"]], ptr @"?AU@@3PAY0EA@V?$RWBuffer@M@hlsl@@A", i32 0, i32 0, i32 26), align 4, !tbaa !6
+  %27 = tail call ptr @llvm.dx.create.handle(i8 1)
+  store ptr %27, ptr getelementptr inbounds ([2 x [64 x %"class.hlsl::RWBuffer"]], ptr @"?AU@@3PAY0EA@V?$RWBuffer@M@hlsl@@A", i32 0, i32 0, i32 27), align 4, !tbaa !6
+  %28 = tail call ptr @llvm.dx.create.handle(i8 1)
+  store ptr %28, ptr getelementptr inbounds ([2 x [64 x %"class.hlsl::RWBuffer"]], ptr @"?AU@@3PAY0EA@V?$RWBuffer@M@hlsl@@A", i32 0, i32 0, i32 28), align 4, !tbaa !6
+  %29 = tail call ptr @llvm.dx.create.handle(i8 1)
+  store ptr %29, ptr getelementptr inbounds ([2 x [64 x %"class.hlsl::RWBuffer"]], ptr @"?AU@@3PAY0EA@V?$RWBuffer@M@hlsl@@A", i32 0, i32 0, i32 29), align 4, !tbaa !6
+  %30 = tail call ptr @llvm.dx.create.handle(i8 1)
+  store ptr %30, ptr getelementptr inbounds ([2 x [64 x %"class.hlsl::RWBuffer"]], ptr @"?AU@@3PAY0EA@V?$RWBuffer@M@hlsl@@A", i32 0, i32 0, i32 30), align 4, !tbaa !6
+  %31 = tail call ptr @llvm.dx.create.handle(i8 1)
+  store ptr %31, ptr getelementptr inbounds ([2 x [64 x %"class.hlsl::RWBuffer"]], ptr @"?AU@@3PAY0EA@V?$RWBuffer@M@hlsl@@A", i32 0, i32 0, i32 31), align 4, !tbaa !6
+  %32 = tail call ptr @llvm.dx.create.handle(i8 1)
+  store ptr %32, ptr getelementptr inbounds ([2 x [64 x %"class.hlsl::RWBuffer"]], ptr @"?AU@@3PAY0EA@V?$RWBuffer@M@hlsl@@A", i32 0, i32 0, i32 32), align 4, !tbaa !6
+  %33 = tail call ptr @llvm.dx.create.handle(i8 1)
+  store ptr %33, ptr getelementptr inbounds ([2 x [64 x %"class.hlsl::RWBuffer"]], ptr @"?AU@@3PAY0EA@V?$RWBuffer@M@hlsl@@A", i32 0, i32 0, i32 33), align 4, !tbaa !6
+  %34 = tail call ptr @llvm.dx.create.handle(i8 1)
+  store ptr %34, ptr getelementptr inbounds ([2 x [64 x %"class.hlsl::RWBuffer"]], ptr @"?AU@@3PAY0EA@V?$RWBuffer@M@hlsl@@A", i32 0, i32 0, i32 34), align 4, !tbaa !6
+  %35 = tail call ptr @llvm.dx.create.handle(i8 1)
+  store ptr %35, ptr getelementptr inbounds ([2 x [64 x %"class.hlsl::RWBuffer"]], ptr @"?AU@@3PAY0EA@V?$RWBuffer@M@hlsl@@A", i32 0, i32 0, i32 35), align 4, !tbaa !6
+  %36 = tail call ptr @llvm.dx.create.handle(i8 1)
+  store ptr %36, ptr getelementptr inbounds ([2 x [64 x %"class.hlsl::RWBuffer"]], ptr @"?AU@@3PAY0EA@V?$RWBuffer@M@hlsl@@A", i32 0, i32 0, i32 36), align 4, !tbaa !6
+  %37 = tail call ptr @llvm.dx.create.handle(i8 1)
+  store ptr %37, ptr getelementptr inbounds ([2 x [64 x %"class.hlsl::RWBuffer"]], ptr @"?AU@@3PAY0EA@V?$RWBuffer@M@hlsl@@A", i32 0, i32 0, i32 37), align 4, !tbaa !6
+  %38 = tail call ptr @llvm.dx.create.handle(i8 1)
+  store ptr %38, ptr getelementptr inbounds ([2 x [64 x %"class.hlsl::RWBuffer"]], ptr @"?AU@@3PAY0EA@V?$RWBuffer@M@hlsl@@A", i32 0, i32 0, i32 38), align 4, !tbaa !6
+  %39 = tail call ptr @llvm.dx.create.handle(i8 1)
+  store ptr %39, ptr getelementptr inbounds ([2 x [64 x %"class.hlsl::RWBuffer"]], ptr @"?AU@@3PAY0EA@V?$RWBuffer@M@hlsl@@A", i32 0, i32 0, i32 39), align 4, !tbaa !6
+  %40 = tail call ptr @llvm.dx.create.handle(i8 1)
+  store ptr %40, ptr getelementptr inbounds ([2 x [64 x %"class.hlsl::RWBuffer"]], ptr @"?AU@@3PAY0EA@V?$RWBuffer@M@hlsl@@A", i32 0, i32 0, i32 40), align 4, !tbaa !6
+  %41 = tail call ptr @llvm.dx.create.handle(i8 1)
+  store ptr %41, ptr getelementptr inbounds ([2 x [64 x %"class.hlsl::RWBuffer"]], ptr @"?AU@@3PAY0EA@V?$RWBuffer@M@hlsl@@A", i32 0, i32 0, i32 41), align 4, !tbaa !6
+  %42 = tail call ptr @llvm.dx.create.handle(i8 1)
+  store ptr %42, ptr getelementptr inbounds ([2 x [64 x %"class.hlsl::RWBuffer"]], ptr @"?AU@@3PAY0EA@V?$RWBuffer@M@hlsl@@A", i32 0, i32 0, i32 42), align 4, !tbaa !6
+  %43 = tail call ptr @llvm.dx.create.handle(i8 1)
+  store ptr %43, ptr getelementptr inbounds ([2 x [64 x %"class.hlsl::RWBuffer"]], ptr @"?AU@@3PAY0EA@V?$RWBuffer@M@hlsl@@A", i32 0, i32 0, i32 43), align 4, !tbaa !6
+  %44 = tail call ptr @llvm.dx.create.handle(i8 1)
+  store ptr %44, ptr getelementptr inbounds ([2 x [64 x %"class.hlsl::RWBuffer"]], ptr @"?AU@@3PAY0EA@V?$RWBuffer@M@hlsl@@A", i32 0, i32 0, i32 44), align 4, !tbaa !6
+  %45 = tail call ptr @llvm.dx.create.handle(i8 1)
+  store ptr %45, ptr getelementptr inbounds ([2 x [64 x %"class.hlsl::RWBuffer"]], ptr @"?AU@@3PAY0EA@V?$RWBuffer@M@hlsl@@A", i32 0, i32 0, i32 45), align 4, !tbaa !6
+  %46 = tail call ptr @llvm.dx.create.handle(i8 1)
+  store ptr %46, ptr getelementptr inbounds ([2 x [64 x %"class.hlsl::RWBuffer"]], ptr @"?AU@@3PAY0EA@V?$RWBuffer@M@hlsl@@A", i32 0, i32 0, i32 46), align 4, !tbaa !6
+  %47 = tail call ptr @llvm.dx.create.handle(i8 1)
+  store ptr %47, ptr getelementptr inbounds ([2 x [64 x %"class.hlsl::RWBuffer"]], ptr @"?AU@@3PAY0EA@V?$RWBuffer@M@hlsl@@A", i32 0, i32 0, i32 47), align 4, !tbaa !6
+  %48 = tail call ptr @llvm.dx.create.handle(i8 1)
+  store ptr %48, ptr getelementptr inbounds ([2 x [64 x %"class.hlsl::RWBuffer"]], ptr @"?AU@@3PAY0EA@V?$RWBuffer@M@hlsl@@A", i32 0, i32 0, i32 48), align 4, !tbaa !6
+  %49 = tail call ptr @llvm.dx.create.handle(i8 1)
+  store ptr %49, ptr getelementptr inbounds ([2 x [64 x %"class.hlsl::RWBuffer"]], ptr @"?AU@@3PAY0EA@V?$RWBuffer@M@hlsl@@A", i32 0, i32 0, i32 49), align 4, !tbaa !6
+  %50 = tail call ptr @llvm.dx.create.handle(i8 1)
+  store ptr %50, ptr getelementptr inbounds ([2 x [64 x %"class.hlsl::RWBuffer"]], ptr @"?AU@@3PAY0EA@V?$RWBuffer@M@hlsl@@A", i32 0, i32 0, i32 50), align 4, !tbaa !6
+  %51 = tail call ptr @llvm.dx.create.handle(i8 1)
+  store ptr %51, ptr getelementptr inbounds ([2 x [64 x %"class.hlsl::RWBuffer"]], ptr @"?AU@@3PAY0EA@V?$RWBuffer@M@hlsl@@A", i32 0, i32 0, i32 51), align 4, !tbaa !6
+  %52 = tail call ptr @llvm.dx.create.handle(i8 1)
+  store ptr %52, ptr getelementptr inbounds ([2 x [64 x %"class.hlsl::RWBuffer"]], ptr @"?AU@@3PAY0EA@V?$RWBuffer@M@hlsl@@A", i32 0, i32 0, i32 52), align 4, !tbaa !6
+  %53 = tail call ptr @llvm.dx.create.handle(i8 1)
+  store ptr %53, ptr getelementptr inbounds ([2 x [64 x %"class.hlsl::RWBuffer"]], ptr @"?AU@@3PAY0EA@V?$RWBuffer@M@hlsl@@A", i32 0, i32 0, i32 53), align 4, !tbaa !6
+  %54 = tail call ptr @llvm.dx.create.handle(i8 1)
+  store ptr %54, ptr getelementptr inbounds ([2 x [64 x %"class.hlsl::RWBuffer"]], ptr @"?AU@@3PAY0EA@V?$RWBuffer@M@hlsl@@A", i32 0, i32 0, i32 54), align 4, !tbaa !6
+  %55 = tail call ptr @llvm.dx.create.handle(i8 1)
+  store ptr %55, ptr getelementptr inbounds ([2 x [64 x %"class.hlsl::RWBuffer"]], ptr @"?AU@@3PAY0EA@V?$RWBuffer@M@hlsl@@A", i32 0, i32 0, i32 55), align 4, !tbaa !6
+  %56 = tail call ptr @llvm.dx.create.handle(i8 1)
+  store ptr %56, ptr getelementptr inbounds ([2 x [64 x %"class.hlsl::RWBuffer"]], ptr @"?AU@@3PAY0EA@V?$RWBuffer@M@hlsl@@A", i32 0, i32 0, i32 56), align 4, !tbaa !6
+  %57 = tail call ptr @llvm.dx.create.handle(i8 1)
+  store ptr %57, ptr getelementptr inbounds ([2 x [64 x %"class.hlsl::RWBuffer"]], ptr @"?AU@@3PAY0EA@V?$RWBuffer@M@hlsl@@A", i32 0, i32 0, i32 57), align 4, !tbaa !6
+  %58 = tail call ptr @llvm.dx.create.handle(i8 1)
+  store ptr %58, ptr getelementptr inbounds ([2 x [64 x %"class.hlsl::RWBuffer"]], ptr @"?AU@@3PAY0EA@V?$RWBuffer@M@hlsl@@A", i32 0, i32 0, i32 58), align 4, !tbaa !6
+  %59 = tail call ptr @llvm.dx.create.handle(i8 1)
+  store ptr %59, ptr getelementptr inbounds ([2 x [64 x %"class.hlsl::RWBuffer"]], ptr @"?AU@@3PAY0EA@V?$RWBuffer@M@hlsl@@A", i32 0, i32 0, i32 59), align 4, !tbaa !6
+  %60 = tail call ptr @llvm.dx.create.handle(i8 1)
+  store ptr %60, ptr getelementptr inbounds ([2 x [64 x %"class.hlsl::RWBuffer"]], ptr @"?AU@@3PAY0EA@V?$RWBuffer@M@hlsl@@A", i32 0, i32 0, i32 60), align 4, !tbaa !6
+  %61 = tail call ptr @llvm.dx.create.handle(i8 1)
+  store ptr %61, ptr getelementptr inbounds ([2 x [64 x %"class.hlsl::RWBuffer"]], ptr @"?AU@@3PAY0EA@V?$RWBuffer@M@hlsl@@A", i32 0, i32 0, i32 61), align 4, !tbaa !6
+  %62 = tail call ptr @llvm.dx.create.handle(i8 1)
+  store ptr %62, ptr getelementptr inbounds ([2 x [64 x %"class.hlsl::RWBuffer"]], ptr @"?AU@@3PAY0EA@V?$RWBuffer@M@hlsl@@A", i32 0, i32 0, i32 62), align 4, !tbaa !6
+  %63 = tail call ptr @llvm.dx.create.handle(i8 1)
+  store ptr %63, ptr getelementptr inbounds ([2 x [64 x %"class.hlsl::RWBuffer"]], ptr @"?AU@@3PAY0EA@V?$RWBuffer@M@hlsl@@A", i32 0, i32 0, i32 63), align 4, !tbaa !6
+  %64 = tail call ptr @llvm.dx.create.handle(i8 1)
+  store ptr %64, ptr getelementptr inbounds ([2 x [64 x %"class.hlsl::RWBuffer"]], ptr @"?AU@@3PAY0EA@V?$RWBuffer@M@hlsl@@A", i32 0, i32 1, i32 0), align 4, !tbaa !6
+  %65 = tail call ptr @llvm.dx.create.handle(i8 1)
+  store ptr %65, ptr getelementptr inbounds ([2 x [64 x %"class.hlsl::RWBuffer"]], ptr @"?AU@@3PAY0EA@V?$RWBuffer@M@hlsl@@A", i32 0, i32 1, i32 1), align 4, !tbaa !6
+  %66 = tail call ptr @llvm.dx.create.handle(i8 1)
+  store ptr %66, ptr getelementptr inbounds ([2 x [64 x %"class.hlsl::RWBuffer"]], ptr @"?AU@@3PAY0EA@V?$RWBuffer@M@hlsl@@A", i32 0, i32 1, i32 2), align 4, !tbaa !6
+  %67 = tail call ptr @llvm.dx.create.handle(i8 1)
+  store ptr %67, ptr getelementptr inbounds ([2 x [64 x %"class.hlsl::RWBuffer"]], ptr @"?AU@@3PAY0EA@V?$RWBuffer@M@hlsl@@A", i32 0, i32 1, i32 3), align 4, !tbaa !6
+  %68 = tail call ptr @llvm.dx.create.handle(i8 1)
+  store ptr %68, ptr getelementptr inbounds ([2 x [64 x %"class.hlsl::RWBuffer"]], ptr @"?AU@@3PAY0EA@V?$RWBuffer@M@hlsl@@A", i32 0, i32 1, i32 4), align 4, !tbaa !6
+  %69 = tail call ptr @llvm.dx.create.handle(i8 1)
+  store ptr %69, ptr getelementptr inbounds ([2 x [64 x %"class.hlsl::RWBuffer"]], ptr @"?AU@@3PAY0EA@V?$RWBuffer@M@hlsl@@A", i32 0, i32 1, i32 5), align 4, !tbaa !6
+  %70 = tail call ptr @llvm.dx.create.handle(i8 1)
+  store ptr %70, ptr getelementptr inbounds ([2 x [64 x %"class.hlsl::RWBuffer"]], ptr @"?AU@@3PAY0EA@V?$RWBuffer@M@hlsl@@A", i32 0, i32 1, i32 6), align 4, !tbaa !6
+  %71 = tail call ptr @llvm.dx.create.handle(i8 1)
+  store ptr %71, ptr getelementptr inbounds ([2 x [64 x %"class.hlsl::RWBuffer"]], ptr @"?AU@@3PAY0EA@V?$RWBuffer@M@hlsl@@A", i32 0, i32 1, i32 7), align 4, !tbaa !6
+  %72 = tail call ptr @llvm.dx.create.handle(i8 1)
+  store ptr %72, ptr getelementptr inbounds ([2 x [64 x %"class.hlsl::RWBuffer"]], ptr @"?AU@@3PAY0EA@V?$RWBuffer@M@hlsl@@A", i32 0, i32 1, i32 8), align 4, !tbaa !6
+  %73 = tail call ptr @llvm.dx.create.handle(i8 1)
+  store ptr %73, ptr getelementptr inbounds ([2 x [64 x %"class.hlsl::RWBuffer"]], ptr @"?AU@@3PAY0EA@V?$RWBuffer@M@hlsl@@A", i32 0, i32 1, i32 9), align 4, !tbaa !6
+  %74 = tail call ptr @llvm.dx.create.handle(i8 1)
+  store ptr %74, ptr getelementptr inbounds ([2 x [64 x %"class.hlsl::RWBuffer"]], ptr @"?AU@@3PAY0EA@V?$RWBuffer@M@hlsl@@A", i32 0, i32 1, i32 10), align 4, !tbaa !6
+  %75 = tail call ptr @llvm.dx.create.handle(i8 1)
+  store ptr %75, ptr getelementptr inbounds ([2 x [64 x %"class.hlsl::RWBuffer"]], ptr @"?AU@@3PAY0EA@V?$RWBuffer@M@hlsl@@A", i32 0, i32 1, i32 11), align 4, !tbaa !6
+  %76 = tail call ptr @llvm.dx.create.handle(i8 1)
+  store ptr %76, ptr getelementptr inbounds ([2 x [64 x %"class.hlsl::RWBuffer"]], ptr @"?AU@@3PAY0EA@V?$RWBuffer@M@hlsl@@A", i32 0, i32 1, i32 12), align 4, !tbaa !6
+  %77 = tail call ptr @llvm.dx.create.handle(i8 1)
+  store ptr %77, ptr getelementptr inbounds ([2 x [64 x %"class.hlsl::RWBuffer"]], ptr @"?AU@@3PAY0EA@V?$RWBuffer@M@hlsl@@A", i32 0, i32 1, i32 13), align 4, !tbaa !6
+  %78 = tail call ptr @llvm.dx.create.handle(i8 1)
+  store ptr %78, ptr getelementptr inbounds ([2 x [64 x %"class.hlsl::RWBuffer"]], ptr @"?AU@@3PAY0EA@V?$RWBuffer@M@hlsl@@A", i32 0, i32 1, i32 14), align 4, !tbaa !6
+  %79 = tail call ptr @llvm.dx.create.handle(i8 1)
+  store ptr %79, ptr getelementptr inbounds ([2 x [64 x %"class.hlsl::RWBuffer"]], ptr @"?AU@@3PAY0EA@V?$RWBuffer@M@hlsl@@A", i32 0, i32 1, i32 15), align 4, !tbaa !6
+  %80 = tail call ptr @llvm.dx.create.handle(i8 1)
+  store ptr %80, ptr getelementptr inbounds ([2 x [64 x %"class.hlsl::RWBuffer"]], ptr @"?AU@@3PAY0EA@V?$RWBuffer@M@hlsl@@A", i32 0, i32 1, i32 16), align 4, !tbaa !6
+  %81 = tail call ptr @llvm.dx.create.handle(i8 1)
+  store ptr %81, ptr getelementptr inbounds ([2 x [64 x %"class.hlsl::RWBuffer"]], ptr @"?AU@@3PAY0EA@V?$RWBuffer@M@hlsl@@A", i32 0, i32 1, i32 17), align 4, !tbaa !6
+  %82 = tail call ptr @llvm.dx.create.handle(i8 1)
+  store ptr %82, ptr getelementptr inbounds ([2 x [64 x %"class.hlsl::RWBuffer"]], ptr @"?AU@@3PAY0EA@V?$RWBuffer@M@hlsl@@A", i32 0, i32 1, i32 18), align 4, !tbaa !6
+  %83 = tail call ptr @llvm.dx.create.handle(i8 1)
+  store ptr %83, ptr getelementptr inbounds ([2 x [64 x %"class.hlsl::RWBuffer"]], ptr @"?AU@@3PAY0EA@V?$RWBuffer@M@hlsl@@A", i32 0, i32 1, i32 19), align 4, !tbaa !6
+  %84 = tail call ptr @llvm.dx.create.handle(i8 1)
+  store ptr %84, ptr getelementptr inbounds ([2 x [64 x %"class.hlsl::RWBuffer"]], ptr @"?AU@@3PAY0EA@V?$RWBuffer@M@hlsl@@A", i32 0, i32 1, i32 20), align 4, !tbaa !6
+  %85 = tail call ptr @llvm.dx.create.handle(i8 1)
+  store ptr %85, ptr getelementptr inbounds ([2 x [64 x %"class.hlsl::RWBuffer"]], ptr @"?AU@@3PAY0EA@V?$RWBuffer@M@hlsl@@A", i32 0, i32 1, i32 21), align 4, !tbaa !6
+  %86 = tail call ptr @llvm.dx.create.handle(i8 1)
+  store ptr %86, ptr getelementptr inbounds ([2 x [64 x %"class.hlsl::RWBuffer"]], ptr @"?AU@@3PAY0EA@V?$RWBuffer@M@hlsl@@A", i32 0, i32 1, i32 22), align 4, !tbaa !6
+  %87 = tail call ptr @llvm.dx.create.handle(i8 1)
+  store ptr %87, ptr getelementptr inbounds ([2 x [64 x %"class.hlsl::RWBuffer"]], ptr @"?AU@@3PAY0EA@V?$RWBuffer@M@hlsl@@A", i32 0, i32 1, i32 23), align 4, !tbaa !6
+  %88 = tail call ptr @llvm.dx.create.handle(i8 1)
+  store ptr %88, ptr getelementptr inbounds ([2 x [64 x %"class.hlsl::RWBuffer"]], ptr @"?AU@@3PAY0EA@V?$RWBuffer@M@hlsl@@A", i32 0, i32 1, i32 24), align 4, !tbaa !6
+  %89 = tail call ptr @llvm.dx.create.handle(i8 1)
+  store ptr %89, ptr getelementptr inbounds ([2 x [64 x %"class.hlsl::RWBuffer"]], ptr @"?AU@@3PAY0EA@V?$RWBuffer@M@hlsl@@A", i32 0, i32 1, i32 25), align 4, !tbaa !6
+  %90 = tail call ptr @llvm.dx.create.handle(i8 1)
+  store ptr %90, ptr getelementptr inbounds ([2 x [64 x %"class.hlsl::RWBuffer"]], ptr @"?AU@@3PAY0EA@V?$RWBuffer@M@hlsl@@A", i32 0, i32 1, i32 26), align 4, !tbaa !6
+  %91 = tail call ptr @llvm.dx.create.handle(i8 1)
+  store ptr %91, ptr getelementptr inbounds ([2 x [64 x %"class.hlsl::RWBuffer"]], ptr @"?AU@@3PAY0EA@V?$RWBuffer@M@hlsl@@A", i32 0, i32 1, i32 27), align 4, !tbaa !6
+  %92 = tail call ptr @llvm.dx.create.handle(i8 1)
+  store ptr %92, ptr getelementptr inbounds ([2 x [64 x %"class.hlsl::RWBuffer"]], ptr @"?AU@@3PAY0EA@V?$RWBuffer@M@hlsl@@A", i32 0, i32 1, i32 28), align 4, !tbaa !6
+  %93 = tail call ptr @llvm.dx.create.handle(i8 1)
+  store ptr %93, ptr getelementptr inbounds ([2 x [64 x %"class.hlsl::RWBuffer"]], ptr @"?AU@@3PAY0EA@V?$RWBuffer@M@hlsl@@A", i32 0, i32 1, i32 29), align 4, !tbaa !6
+  %94 = tail call ptr @llvm.dx.create.handle(i8 1)
+  store ptr %94, ptr getelementptr inbounds ([2 x [64 x %"class.hlsl::RWBuffer"]], ptr @"?AU@@3PAY0EA@V?$RWBuffer@M@hlsl@@A", i32 0, i32 1, i32 30), align 4, !tbaa !6
+  %95 = tail call ptr @llvm.dx.create.handle(i8 1)
+  store ptr %95, ptr getelementptr inbounds ([2 x [64 x %"class.hlsl::RWBuffer"]], ptr @"?AU@@3PAY0EA@V?$RWBuffer@M@hlsl@@A", i32 0, i32 1, i32 31), align 4, !tbaa !6
+  %96 = tail call ptr @llvm.dx.create.handle(i8 1)
+  store ptr %96, ptr getelementptr inbounds ([2 x [64 x %"class.hlsl::RWBuffer"]], ptr @"?AU@@3PAY0EA@V?$RWBuffer@M@hlsl@@A", i32 0, i32 1, i32 32), align 4, !tbaa !6
+  %97 = tail call ptr @llvm.dx.create.handle(i8 1)
+  store ptr %97, ptr getelementptr inbounds ([2 x [64 x %"class.hlsl::RWBuffer"]], ptr @"?AU@@3PAY0EA@V?$RWBuffer@M@hlsl@@A", i32 0, i32 1, i32 33), align 4, !tbaa !6
+  %98 = tail call ptr @llvm.dx.create.handle(i8 1)
+  store ptr %98, ptr getelementptr inbounds ([2 x [64 x %"class.hlsl::RWBuffer"]], ptr @"?AU@@3PAY0EA@V?$RWBuffer@M@hlsl@@A", i32 0, i32 1, i32 34), align 4, !tbaa !6
+  %99 = tail call ptr @llvm.dx.create.handle(i8 1)
+  store ptr %99, ptr getelementptr inbounds ([2 x [64 x %"class.hlsl::RWBuffer"]], ptr @"?AU@@3PAY0EA@V?$RWBuffer@M@hlsl@@A", i32 0, i32 1, i32 35), align 4, !tbaa !6
+  %100 = tail call ptr @llvm.dx.create.handle(i8 1)
+  store ptr %100, ptr getelementptr inbounds ([2 x [64 x %"class.hlsl::RWBuffer"]], ptr @"?AU@@3PAY0EA@V?$RWBuffer@M@hlsl@@A", i32 0, i32 1, i32 36), align 4, !tbaa !6
+  %101 = tail call ptr @llvm.dx.create.handle(i8 1)
+  store ptr %101, ptr getelementptr inbounds ([2 x [64 x %"class.hlsl::RWBuffer"]], ptr @"?AU@@3PAY0EA@V?$RWBuffer@M@hlsl@@A", i32 0, i32 1, i32 37), align 4, !tbaa !6
+  %102 = tail call ptr @llvm.dx.create.handle(i8 1)
+  store ptr %102, ptr getelementptr inbounds ([2 x [64 x %"class.hlsl::RWBuffer"]], ptr @"?AU@@3PAY0EA@V?$RWBuffer@M@hlsl@@A", i32 0, i32 1, i32 38), align 4, !tbaa !6
+  %103 = tail call ptr @llvm.dx.create.handle(i8 1)
+  store ptr %103, ptr getelementptr inbounds ([2 x [64 x %"class.hlsl::RWBuffer"]], ptr @"?AU@@3PAY0EA@V?$RWBuffer@M@hlsl@@A", i32 0, i32 1, i32 39), align 4, !tbaa !6
+  %104 = tail call ptr @llvm.dx.create.handle(i8 1)
+  store ptr %104, ptr getelementptr inbounds ([2 x [64 x %"class.hlsl::RWBuffer"]], ptr @"?AU@@3PAY0EA@V?$RWBuffer@M@hlsl@@A", i32 0, i32 1, i32 40), align 4, !tbaa !6
+  %105 = tail call ptr @llvm.dx.create.handle(i8 1)
+  store ptr %105, ptr getelementptr inbounds ([2 x [64 x %"class.hlsl::RWBuffer"]], ptr @"?AU@@3PAY0EA@V?$RWBuffer@M@hlsl@@A", i32 0, i32 1, i32 41), align 4, !tbaa !6
+  %106 = tail call ptr @llvm.dx.create.handle(i8 1)
+  store ptr %106, ptr getelementptr inbounds ([2 x [64 x %"class.hlsl::RWBuffer"]], ptr @"?AU@@3PAY0EA@V?$RWBuffer@M@hlsl@@A", i32 0, i32 1, i32 42), align 4, !tbaa !6
+  %107 = tail call ptr @llvm.dx.create.handle(i8 1)
+  store ptr %107, ptr getelementptr inbounds ([2 x [64 x %"class.hlsl::RWBuffer"]], ptr @"?AU@@3PAY0EA@V?$RWBuffer@M@hlsl@@A", i32 0, i32 1, i32 43), align 4, !tbaa !6
+  %108 = tail call ptr @llvm.dx.create.handle(i8 1)
+  store ptr %108, ptr getelementptr inbounds ([2 x [64 x %"class.hlsl::RWBuffer"]], ptr @"?AU@@3PAY0EA@V?$RWBuffer@M@hlsl@@A", i32 0, i32 1, i32 44), align 4, !tbaa !6
+  %109 = tail call ptr @llvm.dx.create.handle(i8 1)
+  store ptr %109, ptr getelementptr inbounds ([2 x [64 x %"class.hlsl::RWBuffer"]], ptr @"?AU@@3PAY0EA@V?$RWBuffer@M@hlsl@@A", i32 0, i32 1, i32 45), align 4, !tbaa !6
+  %110 = tail call ptr @llvm.dx.create.handle(i8 1)
+  store ptr %110, ptr getelementptr inbounds ([2 x [64 x %"class.hlsl::RWBuffer"]], ptr @"?AU@@3PAY0EA@V?$RWBuffer@M@hlsl@@A", i32 0, i32 1, i32 46), align 4, !tbaa !6
+  %111 = tail call ptr @llvm.dx.create.handle(i8 1)
+  store ptr %111, ptr getelementptr inbounds ([2 x [64 x %"class.hlsl::RWBuffer"]], ptr @"?AU@@3PAY0EA@V?$RWBuffer@M@hlsl@@A", i32 0, i32 1, i32 47), align 4, !tbaa !6
+  %112 = tail call ptr @llvm.dx.create.handle(i8 1)
+  store ptr %112, ptr getelementptr inbounds ([2 x [64 x %"class.hlsl::RWBuffer"]], ptr @"?AU@@3PAY0EA@V?$RWBuffer@M@hlsl@@A", i32 0, i32 1, i32 48), align 4, !tbaa !6
+  %113 = tail call ptr @llvm.dx.create.handle(i8 1)
+  store ptr %113, ptr getelementptr inbounds ([2 x [64 x %"class.hlsl::RWBuffer"]], ptr @"?AU@@3PAY0EA@V?$RWBuffer@M@hlsl@@A", i32 0, i32 1, i32 49), align 4, !tbaa !6
+  %114 = tail call ptr @llvm.dx.create.handle(i8 1)
+  store ptr %114, ptr getelementptr inbounds ([2 x [64 x %"class.hlsl::RWBuffer"]], ptr @"?AU@@3PAY0EA@V?$RWBuffer@M@hlsl@@A", i32 0, i32 1, i32 50), align 4, !tbaa !6
+  %115 = tail call ptr @llvm.dx.create.handle(i8 1)
+  store ptr %115, ptr getelementptr inbounds ([2 x [64 x %"class.hlsl::RWBuffer"]], ptr @"?AU@@3PAY0EA@V?$RWBuffer@M@hlsl@@A", i32 0, i32 1, i32 51), align 4, !tbaa !6
+  %116 = tail call ptr @llvm.dx.create.handle(i8 1)
+  store ptr %116, ptr getelementptr inbounds ([2 x [64 x %"class.hlsl::RWBuffer"]], ptr @"?AU@@3PAY0EA@V?$RWBuffer@M@hlsl@@A", i32 0, i32 1, i32 52), align 4, !tbaa !6
+  %117 = tail call ptr @llvm.dx.create.handle(i8 1)
+  store ptr %117, ptr getelementptr inbounds ([2 x [64 x %"class.hlsl::RWBuffer"]], ptr @"?AU@@3PAY0EA@V?$RWBuffer@M@hlsl@@A", i32 0, i32 1, i32 53), align 4, !tbaa !6
+  %118 = tail call ptr @llvm.dx.create.handle(i8 1)
+  store ptr %118, ptr getelementptr inbounds ([2 x [64 x %"class.hlsl::RWBuffer"]], ptr @"?AU@@3PAY0EA@V?$RWBuffer@M@hlsl@@A", i32 0, i32 1, i32 54), align 4, !tbaa !6
+  %119 = tail call ptr @llvm.dx.create.handle(i8 1)
+  store ptr %119, ptr getelementptr inbounds ([2 x [64 x %"class.hlsl::RWBuffer"]], ptr @"?AU@@3PAY0EA@V?$RWBuffer@M@hlsl@@A", i32 0, i32 1, i32 55), align 4, !tbaa !6
+  %120 = tail call ptr @llvm.dx.create.handle(i8 1)
+  store ptr %120, ptr getelementptr inbounds ([2 x [64 x %"class.hlsl::RWBuffer"]], ptr @"?AU@@3PAY0EA@V?$RWBuffer@M@hlsl@@A", i32 0, i32 1, i32 56), align 4, !tbaa !6
+  %121 = tail call ptr @llvm.dx.create.handle(i8 1)
+  store ptr %121, ptr getelementptr inbounds ([2 x [64 x %"class.hlsl::RWBuffer"]], ptr @"?AU@@3PAY0EA@V?$RWBuffer@M@hlsl@@A", i32 0, i32 1, i32 57), align 4, !tbaa !6
+  %122 = tail call ptr @llvm.dx.create.handle(i8 1)
+  store ptr %122, ptr getelementptr inbounds ([2 x [64 x %"class.hlsl::RWBuffer"]], ptr @"?AU@@3PAY0EA@V?$RWBuffer@M@hlsl@@A", i32 0, i32 1, i32 58), align 4, !tbaa !6
+  %123 = tail call ptr @llvm.dx.create.handle(i8 1)
+  store ptr %123, ptr getelementptr inbounds ([2 x [64 x %"class.hlsl::RWBuffer"]], ptr @"?AU@@3PAY0EA@V?$RWBuffer@M@hlsl@@A", i32 0, i32 1, i32 59), align 4, !tbaa !6
+  %124 = tail call ptr @llvm.dx.create.handle(i8 1)
+  store ptr %124, ptr getelementptr inbounds ([2 x [64 x %"class.hlsl::RWBuffer"]], ptr @"?AU@@3PAY0EA@V?$RWBuffer@M@hlsl@@A", i32 0, i32 1, i32 60), align 4, !tbaa !6
+  %125 = tail call ptr @llvm.dx.create.handle(i8 1)
+  store ptr %125, ptr getelementptr inbounds ([2 x [64 x %"class.hlsl::RWBuffer"]], ptr @"?AU@@3PAY0EA@V?$RWBuffer@M@hlsl@@A", i32 0, i32 1, i32 61), align 4, !tbaa !6
+  %126 = tail call ptr @llvm.dx.create.handle(i8 1)
+  store ptr %126, ptr getelementptr inbounds ([2 x [64 x %"class.hlsl::RWBuffer"]], ptr @"?AU@@3PAY0EA@V?$RWBuffer@M@hlsl@@A", i32 0, i32 1, i32 62), align 4, !tbaa !6
+  %127 = tail call ptr @llvm.dx.create.handle(i8 1)
+  store ptr %127, ptr getelementptr inbounds ([2 x [64 x %"class.hlsl::RWBuffer"]], ptr @"?AU@@3PAY0EA@V?$RWBuffer@M@hlsl@@A", i32 0, i32 1, i32 63), align 4, !tbaa !6
+  %128 = tail call ptr @llvm.dx.create.handle(i8 1)
+  store ptr %128, ptr @"?AUI@@3PAY01V?$RWBuffer@M@hlsl@@A", align 4, !tbaa !6
+  %129 = tail call ptr @llvm.dx.create.handle(i8 1)
+  store ptr %129, ptr getelementptr inbounds ([64 x [2 x %"class.hlsl::RWBuffer"]], ptr @"?AUI@@3PAY01V?$RWBuffer@M@hlsl@@A", i32 0, i32 0, i32 1), align 4, !tbaa !6
+  %130 = tail call ptr @llvm.dx.create.handle(i8 1)
+  store ptr %130, ptr getelementptr inbounds ([64 x [2 x %"class.hlsl::RWBuffer"]], ptr @"?AUI@@3PAY01V?$RWBuffer@M@hlsl@@A", i32 0, i32 1, i32 0), align 4, !tbaa !6
+  %131 = tail call ptr @llvm.dx.create.handle(i8 1)
+  store ptr %131, ptr getelementptr inbounds ([64 x [2 x %"class.hlsl::RWBuffer"]], ptr @"?AUI@@3PAY01V?$RWBuffer@M@hlsl@@A", i32 0, i32 1, i32 1), align 4, !tbaa !6
+  %132 = tail call ptr @llvm.dx.create.handle(i8 1)
+  store ptr %132, ptr getelementptr inbounds ([64 x [2 x %"class.hlsl::RWBuffer"]], ptr @"?AUI@@3PAY01V?$RWBuffer@M@hlsl@@A", i32 0, i32 2, i32 0), align 4, !tbaa !6
+  %133 = tail call ptr @llvm.dx.create.handle(i8 1)
+  store ptr %133, ptr getelementptr inbounds ([64 x [2 x %"class.hlsl::RWBuffer"]], ptr @"?AUI@@3PAY01V?$RWBuffer@M@hlsl@@A", i32 0, i32 2, i32 1), align 4, !tbaa !6
+  %134 = tail call ptr @llvm.dx.create.handle(i8 1)
+  store ptr %134, ptr getelementptr inbounds ([64 x [2 x %"class.hlsl::RWBuffer"]], ptr @"?AUI@@3PAY01V?$RWBuffer@M@hlsl@@A", i32 0, i32 3, i32 0), align 4, !tbaa !6
+  %135 = tail call ptr @llvm.dx.create.handle(i8 1)
+  store ptr %135, ptr getelementptr inbounds ([64 x [2 x %"class.hlsl::RWBuffer"]], ptr @"?AUI@@3PAY01V?$RWBuffer@M@hlsl@@A", i32 0, i32 3, i32 1), align 4, !tbaa !6
+  %136 = tail call ptr @llvm.dx.create.handle(i8 1)
+  store ptr %136, ptr getelementptr inbounds ([64 x [2 x %"class.hlsl::RWBuffer"]], ptr @"?AUI@@3PAY01V?$RWBuffer@M@hlsl@@A", i32 0, i32 4, i32 0), align 4, !tbaa !6
+  %137 = tail call ptr @llvm.dx.create.handle(i8 1)
+  store ptr %137, ptr getelementptr inbounds ([64 x [2 x %"class.hlsl::RWBuffer"]], ptr @"?AUI@@3PAY01V?$RWBuffer@M@hlsl@@A", i32 0, i32 4, i32 1), align 4, !tbaa !6
+  %138 = tail call ptr @llvm.dx.create.handle(i8 1)
+  store ptr %138, ptr getelementptr inbounds ([64 x [2 x %"class.hlsl::RWBuffer"]], ptr @"?AUI@@3PAY01V?$RWBuffer@M@hlsl@@A", i32 0, i32 5, i32 0), align 4, !tbaa !6
+  %139 = tail call ptr @llvm.dx.create.handle(i8 1)
+  store ptr %139, ptr getelementptr inbounds ([64 x [2 x %"class.hlsl::RWBuffer"]], ptr @"?AUI@@3PAY01V?$RWBuffer@M@hlsl@@A", i32 0, i32 5, i32 1), align 4, !tbaa !6
+  %140 = tail call ptr @llvm.dx.create.handle(i8 1)
+  store ptr %140, ptr getelementptr inbounds ([64 x [2 x %"class.hlsl::RWBuffer"]], ptr @"?AUI@@3PAY01V?$RWBuffer@M@hlsl@@A", i32 0, i32 6, i32 0), align 4, !tbaa !6
+  %141 = tail call ptr @llvm.dx.create.handle(i8 1)
+  store ptr %141, ptr getelementptr inbounds ([64 x [2 x %"class.hlsl::RWBuffer"]], ptr @"?AUI@@3PAY01V?$RWBuffer@M@hlsl@@A", i32 0, i32 6, i32 1), align 4, !tbaa !6
+  %142 = tail call ptr @llvm.dx.create.handle(i8 1)
+  store ptr %142, ptr getelementptr inbounds ([64 x [2 x %"class.hlsl::RWBuffer"]], ptr @"?AUI@@3PAY01V?$RWBuffer@M@hlsl@@A", i32 0, i32 7, i32 0), align 4, !tbaa !6
+  %143 = tail call ptr @llvm.dx.create.handle(i8 1)
+  store ptr %143, ptr getelementptr inbounds ([64 x [2 x %"class.hlsl::RWBuffer"]], ptr @"?AUI@@3PAY01V?$RWBuffer@M@hlsl@@A", i32 0, i32 7, i32 1), align 4, !tbaa !6
+  %144 = tail call ptr @llvm.dx.create.handle(i8 1)
+  store ptr %144, ptr getelementptr inbounds ([64 x [2 x %"class.hlsl::RWBuffer"]], ptr @"?AUI@@3PAY01V?$RWBuffer@M@hlsl@@A", i32 0, i32 8, i32 0), align 4, !tbaa !6
+  %145 = tail call ptr @llvm.dx.create.handle(i8 1)
+  store ptr %145, ptr getelementptr inbounds ([64 x [2 x %"class.hlsl::RWBuffer"]], ptr @"?AUI@@3PAY01V?$RWBuffer@M@hlsl@@A", i32 0, i32 8, i32 1), align 4, !tbaa !6
+  %146 = tail call ptr @llvm.dx.create.handle(i8 1)
+  store ptr %146, ptr getelementptr inbounds ([64 x [2 x %"class.hlsl::RWBuffer"]], ptr @"?AUI@@3PAY01V?$RWBuffer@M@hlsl@@A", i32 0, i32 9, i32 0), align 4, !tbaa !6
+  %147 = tail call ptr @llvm.dx.create.handle(i8 1)
+  store ptr %147, ptr getelementptr inbounds ([64 x [2 x %"class.hlsl::RWBuffer"]], ptr @"?AUI@@3PAY01V?$RWBuffer@M@hlsl@@A", i32 0, i32 9, i32 1), align 4, !tbaa !6
+  %148 = tail call ptr @llvm.dx.create.handle(i8 1)
+  store ptr %148, ptr getelementptr inbounds ([64 x [2 x %"class.hlsl::RWBuffer"]], ptr @"?AUI@@3PAY01V?$RWBuffer@M@hlsl@@A", i32 0, i32 10, i32 0), align 4, !tbaa !6
+  %149 = tail call ptr @llvm.dx.create.handle(i8 1)
+  store ptr %149, ptr getelementptr inbounds ([64 x [2 x %"class.hlsl::RWBuffer"]], ptr @"?AUI@@3PAY01V?$RWBuffer@M@hlsl@@A", i32 0, i32 10, i32 1), align 4, !tbaa !6
+  %150 = tail call ptr @llvm.dx.create.handle(i8 1)
+  store ptr %150, ptr getelementptr inbounds ([64 x [2 x %"class.hlsl::RWBuffer"]], ptr @"?AUI@@3PAY01V?$RWBuffer@M@hlsl@@A", i32 0, i32 11, i32 0), align 4, !tbaa !6
+  %151 = tail call ptr @llvm.dx.create.handle(i8 1)
+  store ptr %151, ptr getelementptr inbounds ([64 x [2 x %"class.hlsl::RWBuffer"]], ptr @"?AUI@@3PAY01V?$RWBuffer@M@hlsl@@A", i32 0, i32 11, i32 1), align 4, !tbaa !6
+  %152 = tail call ptr @llvm.dx.create.handle(i8 1)
+  store ptr %152, ptr getelementptr inbounds ([64 x [2 x %"class.hlsl::RWBuffer"]], ptr @"?AUI@@3PAY01V?$RWBuffer@M@hlsl@@A", i32 0, i32 12, i32 0), align 4, !tbaa !6
+  %153 = tail call ptr @llvm.dx.create.handle(i8 1)
+  store ptr %153, ptr getelementptr inbounds ([64 x [2 x %"class.hlsl::RWBuffer"]], ptr @"?AUI@@3PAY01V?$RWBuffer@M@hlsl@@A", i32 0, i32 12, i32 1), align 4, !tbaa !6
+  %154 = tail call ptr @llvm.dx.create.handle(i8 1)
+  store ptr %154, ptr getelementptr inbounds ([64 x [2 x %"class.hlsl::RWBuffer"]], ptr @"?AUI@@3PAY01V?$RWBuffer@M@hlsl@@A", i32 0, i32 13, i32 0), align 4, !tbaa !6
+  %155 = tail call ptr @llvm.dx.create.handle(i8 1)
+  store ptr %155, ptr getelementptr inbounds ([64 x [2 x %"class.hlsl::RWBuffer"]], ptr @"?AUI@@3PAY01V?$RWBuffer@M@hlsl@@A", i32 0, i32 13, i32 1), align 4, !tbaa !6
+  %156 = tail call ptr @llvm.dx.create.handle(i8 1)
+  store ptr %156, ptr getelementptr inbounds ([64 x [2 x %"class.hlsl::RWBuffer"]], ptr @"?AUI@@3PAY01V?$RWBuffer@M@hlsl@@A", i32 0, i32 14, i32 0), align 4, !tbaa !6
+  %157 = tail call ptr @llvm.dx.create.handle(i8 1)
+  store ptr %157, ptr getelementptr inbounds ([64 x [2 x %"class.hlsl::RWBuffer"]], ptr @"?AUI@@3PAY01V?$RWBuffer@M@hlsl@@A", i32 0, i32 14, i32 1), align 4, !tbaa !6
+  %158 = tail call ptr @llvm.dx.create.handle(i8 1)
+  store ptr %158, ptr getelementptr inbounds ([64 x [2 x %"class.hlsl::RWBuffer"]], ptr @"?AUI@@3PAY01V?$RWBuffer@M@hlsl@@A", i32 0, i32 15, i32 0), align 4, !tbaa !6
+  %159 = tail call ptr @llvm.dx.create.handle(i8 1)
+  store ptr %159, ptr getelementptr inbounds ([64 x [2 x %"class.hlsl::RWBuffer"]], ptr @"?AUI@@3PAY01V?$RWBuffer@M@hlsl@@A", i32 0, i32 15, i32 1), align 4, !tbaa !6
+  %160 = tail call ptr @llvm.dx.create.handle(i8 1)
+  store ptr %160, ptr getelementptr inbounds ([64 x [2 x %"class.hlsl::RWBuffer"]], ptr @"?AUI@@3PAY01V?$RWBuffer@M@hlsl@@A", i32 0, i32 16, i32 0), align 4, !tbaa !6
+  %161 = tail call ptr @llvm.dx.create.handle(i8 1)
+  store ptr %161, ptr getelementptr inbounds ([64 x [2 x %"class.hlsl::RWBuffer"]], ptr @"?AUI@@3PAY01V?$RWBuffer@M@hlsl@@A", i32 0, i32 16, i32 1), align 4, !tbaa !6
+  %162 = tail call ptr @llvm.dx.create.handle(i8 1)
+  store ptr %162, ptr getelementptr inbounds ([64 x [2 x %"class.hlsl::RWBuffer"]], ptr @"?AUI@@3PAY01V?$RWBuffer@M@hlsl@@A", i32 0, i32 17, i32 0), align 4, !tbaa !6
+  %163 = tail call ptr @llvm.dx.create.handle(i8 1)
+  store ptr %163, ptr getelementptr inbounds ([64 x [2 x %"class.hlsl::RWBuffer"]], ptr @"?AUI@@3PAY01V?$RWBuffer@M@hlsl@@A", i32 0, i32 17, i32 1), align 4, !tbaa !6
+  %164 = tail call ptr @llvm.dx.create.handle(i8 1)
+  store ptr %164, ptr getelementptr inbounds ([64 x [2 x %"class.hlsl::RWBuffer"]], ptr @"?AUI@@3PAY01V?$RWBuffer@M@hlsl@@A", i32 0, i32 18, i32 0), align 4, !tbaa !6
+  %165 = tail call ptr @llvm.dx.create.handle(i8 1)
+  store ptr %165, ptr getelementptr inbounds ([64 x [2 x %"class.hlsl::RWBuffer"]], ptr @"?AUI@@3PAY01V?$RWBuffer@M@hlsl@@A", i32 0, i32 18, i32 1), align 4, !tbaa !6
+  %166 = tail call ptr @llvm.dx.create.handle(i8 1)
+  store ptr %166, ptr getelementptr inbounds ([64 x [2 x %"class.hlsl::RWBuffer"]], ptr @"?AUI@@3PAY01V?$RWBuffer@M@hlsl@@A", i32 0, i32 19, i32 0), align 4, !tbaa !6
+  %167 = tail call ptr @llvm.dx.create.handle(i8 1)
+  store ptr %167, ptr getelementptr inbounds ([64 x [2 x %"class.hlsl::RWBuffer"]], ptr @"?AUI@@3PAY01V?$RWBuffer@M@hlsl@@A", i32 0, i32 19, i32 1), align 4, !tbaa !6
+  %168 = tail call ptr @llvm.dx.create.handle(i8 1)
+  store ptr %168, ptr getelementptr inbounds ([64 x [2 x %"class.hlsl::RWBuffer"]], ptr @"?AUI@@3PAY01V?$RWBuffer@M@hlsl@@A", i32 0, i32 20, i32 0), align 4, !tbaa !6
+  %169 = tail call ptr @llvm.dx.create.handle(i8 1)
+  store ptr %169, ptr getelementptr inbounds ([64 x [2 x %"class.hlsl::RWBuffer"]], ptr @"?AUI@@3PAY01V?$RWBuffer@M@hlsl@@A", i32 0, i32 20, i32 1), align 4, !tbaa !6
+  %170 = tail call ptr @llvm.dx.create.handle(i8 1)
+  store ptr %170, ptr getelementptr inbounds ([64 x [2 x %"class.hlsl::RWBuffer"]], ptr @"?AUI@@3PAY01V?$RWBuffer@M@hlsl@@A", i32 0, i32 21, i32 0), align 4, !tbaa !6
+  %171 = tail call ptr @llvm.dx.create.handle(i8 1)
+  store ptr %171, ptr getelementptr inbounds ([64 x [2 x %"class.hlsl::RWBuffer"]], ptr @"?AUI@@3PAY01V?$RWBuffer@M@hlsl@@A", i32 0, i32 21, i32 1), align 4, !tbaa !6
+  %172 = tail call ptr @llvm.dx.create.handle(i8 1)
+  store ptr %172, ptr getelementptr inbounds ([64 x [2 x %"class.hlsl::RWBuffer"]], ptr @"?AUI@@3PAY01V?$RWBuffer@M@hlsl@@A", i32 0, i32 22, i32 0), align 4, !tbaa !6
+  %173 = tail call ptr @llvm.dx.create.handle(i8 1)
+  store ptr %173, ptr getelementptr inbounds ([64 x [2 x %"class.hlsl::RWBuffer"]], ptr @"?AUI@@3PAY01V?$RWBuffer@M@hlsl@@A", i32 0, i32 22, i32 1), align 4, !tbaa !6
+  %174 = tail call ptr @llvm.dx.create.handle(i8 1)
+  store ptr %174, ptr getelementptr inbounds ([64 x [2 x %"class.hlsl::RWBuffer"]], ptr @"?AUI@@3PAY01V?$RWBuffer@M@hlsl@@A", i32 0, i32 23, i32 0), align 4, !tbaa !6
+  %175 = tail call ptr @llvm.dx.create.handle(i8 1)
+  store ptr %175, ptr getelementptr inbounds ([64 x [2 x %"class.hlsl::RWBuffer"]], ptr @"?AUI@@3PAY01V?$RWBuffer@M@hlsl@@A", i32 0, i32 23, i32 1), align 4, !tbaa !6
+  %176 = tail call ptr @llvm.dx.create.handle(i8 1)
+  store ptr %176, ptr getelementptr inbounds ([64 x [2 x %"class.hlsl::RWBuffer"]], ptr @"?AUI@@3PAY01V?$RWBuffer@M@hlsl@@A", i32 0, i32 24, i32 0), align 4, !tbaa !6
+  %177 = tail call ptr @llvm.dx.create.handle(i8 1)
+  store ptr %177, ptr getelementptr inbounds ([64 x [2 x %"class.hlsl::RWBuffer"]], ptr @"?AUI@@3PAY01V?$RWBuffer@M@hlsl@@A", i32 0, i32 24, i32 1), align 4, !tbaa !6
+  %178 = tail call ptr @llvm.dx.create.handle(i8 1)
+  store ptr %178, ptr getelementptr inbounds ([64 x [2 x %"class.hlsl::RWBuffer"]], ptr @"?AUI@@3PAY01V?$RWBuffer@M@hlsl@@A", i32 0, i32 25, i32 0), align 4, !tbaa !6
+  %179 = tail call ptr @llvm.dx.create.handle(i8 1)
+  store ptr %179, ptr getelementptr inbounds ([64 x [2 x %"class.hlsl::RWBuffer"]], ptr @"?AUI@@3PAY01V?$RWBuffer@M@hlsl@@A", i32 0, i32 25, i32 1), align 4, !tbaa !6
+  %180 = tail call ptr @llvm.dx.create.handle(i8 1)
+  store ptr %180, ptr getelementptr inbounds ([64 x [2 x %"class.hlsl::RWBuffer"]], ptr @"?AUI@@3PAY01V?$RWBuffer@M@hlsl@@A", i32 0, i32 26, i32 0), align 4, !tbaa !6
+  %181 = tail call ptr @llvm.dx.create.handle(i8 1)
+  store ptr %181, ptr getelementptr inbounds ([64 x [2 x %"class.hlsl::RWBuffer"]], ptr @"?AUI@@3PAY01V?$RWBuffer@M@hlsl@@A", i32 0, i32 26, i32 1), align 4, !tbaa !6
+  %182 = tail call ptr @llvm.dx.create.handle(i8 1)
+  store ptr %182, ptr getelementptr inbounds ([64 x [2 x %"class.hlsl::RWBuffer"]], ptr @"?AUI@@3PAY01V?$RWBuffer@M@hlsl@@A", i32 0, i32 27, i32 0), align 4, !tbaa !6
+  %183 = tail call ptr @llvm.dx.create.handle(i8 1)
+  store ptr %183, ptr getelementptr inbounds ([64 x [2 x %"class.hlsl::RWBuffer"]], ptr @"?AUI@@3PAY01V?$RWBuffer@M@hlsl@@A", i32 0, i32 27, i32 1), align 4, !tbaa !6
+  %184 = tail call ptr @llvm.dx.create.handle(i8 1)
+  store ptr %184, ptr getelementptr inbounds ([64 x [2 x %"class.hlsl::RWBuffer"]], ptr @"?AUI@@3PAY01V?$RWBuffer@M@hlsl@@A", i32 0, i32 28, i32 0), align 4, !tbaa !6
+  %185 = tail call ptr @llvm.dx.create.handle(i8 1)
+  store ptr %185, ptr getelementptr inbounds ([64 x [2 x %"class.hlsl::RWBuffer"]], ptr @"?AUI@@3PAY01V?$RWBuffer@M@hlsl@@A", i32 0, i32 28, i32 1), align 4, !tbaa !6
+  %186 = tail call ptr @llvm.dx.create.handle(i8 1)
+  store ptr %186, ptr getelementptr inbounds ([64 x [2 x %"class.hlsl::RWBuffer"]], ptr @"?AUI@@3PAY01V?$RWBuffer@M@hlsl@@A", i32 0, i32 29, i32 0), align 4, !tbaa !6
+  %187 = tail call ptr @llvm.dx.create.handle(i8 1)
+  store ptr %187, ptr getelementptr inbounds ([64 x [2 x %"class.hlsl::RWBuffer"]], ptr @"?AUI@@3PAY01V?$RWBuffer@M@hlsl@@A", i32 0, i32 29, i32 1), align 4, !tbaa !6
+  %188 = tail call ptr @llvm.dx.create.handle(i8 1)
+  store ptr %188, ptr getelementptr inbounds ([64 x [2 x %"class.hlsl::RWBuffer"]], ptr @"?AUI@@3PAY01V?$RWBuffer@M@hlsl@@A", i32 0, i32 30, i32 0), align 4, !tbaa !6
+  %189 = tail call ptr @llvm.dx.create.handle(i8 1)
+  store ptr %189, ptr getelementptr inbounds ([64 x [2 x %"class.hlsl::RWBuffer"]], ptr @"?AUI@@3PAY01V?$RWBuffer@M@hlsl@@A", i32 0, i32 30, i32 1), align 4, !tbaa !6
+  %190 = tail call ptr @llvm.dx.create.handle(i8 1)
+  store ptr %190, ptr getelementptr inbounds ([64 x [2 x %"class.hlsl::RWBuffer"]], ptr @"?AUI@@3PAY01V?$RWBuffer@M@hlsl@@A", i32 0, i32 31, i32 0), align 4, !tbaa !6
+  %191 = tail call ptr @llvm.dx.create.handle(i8 1)
+  store ptr %191, ptr getelementptr inbounds ([64 x [2 x %"class.hlsl::RWBuffer"]], ptr @"?AUI@@3PAY01V?$RWBuffer@M@hlsl@@A", i32 0, i32 31, i32 1), align 4, !tbaa !6
+  %192 = tail call ptr @llvm.dx.create.handle(i8 1)
+  store ptr %192, ptr getelementptr inbounds ([64 x [2 x %"class.hlsl::RWBuffer"]], ptr @"?AUI@@3PAY01V?$RWBuffer@M@hlsl@@A", i32 0, i32 32, i32 0), align 4, !tbaa !6
+  %193 = tail call ptr @llvm.dx.create.handle(i8 1)
+  store ptr %193, ptr getelementptr inbounds ([64 x [2 x %"class.hlsl::RWBuffer"]], ptr @"?AUI@@3PAY01V?$RWBuffer@M@hlsl@@A", i32 0, i32 32, i32 1), align 4, !tbaa !6
+  %194 = tail call ptr @llvm.dx.create.handle(i8 1)
+  store ptr %194, ptr getelementptr inbounds ([64 x [2 x %"class.hlsl::RWBuffer"]], ptr @"?AUI@@3PAY01V?$RWBuffer@M@hlsl@@A", i32 0, i32 33, i32 0), align 4, !tbaa !6
+  %195 = tail call ptr @llvm.dx.create.handle(i8 1)
+  store ptr %195, ptr getelementptr inbounds ([64 x [2 x %"class.hlsl::RWBuffer"]], ptr @"?AUI@@3PAY01V?$RWBuffer@M@hlsl@@A", i32 0, i32 33, i32 1), align 4, !tbaa !6
+  %196 = tail call ptr @llvm.dx.create.handle(i8 1)
+  store ptr %196, ptr getelementptr inbounds ([64 x [2 x %"class.hlsl::RWBuffer"]], ptr @"?AUI@@3PAY01V?$RWBuffer@M@hlsl@@A", i32 0, i32 34, i32 0), align 4, !tbaa !6
+  %197 = tail call ptr @llvm.dx.create.handle(i8 1)
+  store ptr %197, ptr getelementptr inbounds ([64 x [2 x %"class.hlsl::RWBuffer"]], ptr @"?AUI@@3PAY01V?$RWBuffer@M@hlsl@@A", i32 0, i32 34, i32 1), align 4, !tbaa !6
+  %198 = tail call ptr @llvm.dx.create.handle(i8 1)
+  store ptr %198, ptr getelementptr inbounds ([64 x [2 x %"class.hlsl::RWBuffer"]], ptr @"?AUI@@3PAY01V?$RWBuffer@M@hlsl@@A", i32 0, i32 35, i32 0), align 4, !tbaa !6
+  %199 = tail call ptr @llvm.dx.create.handle(i8 1)
+  store ptr %199, ptr getelementptr inbounds ([64 x [2 x %"class.hlsl::RWBuffer"]], ptr @"?AUI@@3PAY01V?$RWBuffer@M@hlsl@@A", i32 0, i32 35, i32 1), align 4, !tbaa !6
+  %200 = tail call ptr @llvm.dx.create.handle(i8 1)
+  store ptr %200, ptr getelementptr inbounds ([64 x [2 x %"class.hlsl::RWBuffer"]], ptr @"?AUI@@3PAY01V?$RWBuffer@M@hlsl@@A", i32 0, i32 36, i32 0), align 4, !tbaa !6
+  %201 = tail call ptr @llvm.dx.create.handle(i8 1)
+  store ptr %201, ptr getelementptr inbounds ([64 x [2 x %"class.hlsl::RWBuffer"]], ptr @"?AUI@@3PAY01V?$RWBuffer@M@hlsl@@A", i32 0, i32 36, i32 1), align 4, !tbaa !6
+  %202 = tail call ptr @llvm.dx.create.handle(i8 1)
+  store ptr %202, ptr getelementptr inbounds ([64 x [2 x %"class.hlsl::RWBuffer"]], ptr @"?AUI@@3PAY01V?$RWBuffer@M@hlsl@@A", i32 0, i32 37, i32 0), align 4, !tbaa !6
+  %203 = tail call ptr @llvm.dx.create.handle(i8 1)
+  store ptr %203, ptr getelementptr inbounds ([64 x [2 x %"class.hlsl::RWBuffer"]], ptr @"?AUI@@3PAY01V?$RWBuffer@M@hlsl@@A", i32 0, i32 37, i32 1), align 4, !tbaa !6
+  %204 = tail call ptr @llvm.dx.create.handle(i8 1)
+  store ptr %204, ptr getelementptr inbounds ([64 x [2 x %"class.hlsl::RWBuffer"]], ptr @"?AUI@@3PAY01V?$RWBuffer@M@hlsl@@A", i32 0, i32 38, i32 0), align 4, !tbaa !6
+  %205 = tail call ptr @llvm.dx.create.handle(i8 1)
+  store ptr %205, ptr getelementptr inbounds ([64 x [2 x %"class.hlsl::RWBuffer"]], ptr @"?AUI@@3PAY01V?$RWBuffer@M@hlsl@@A", i32 0, i32 38, i32 1), align 4, !tbaa !6
+  %206 = tail call ptr @llvm.dx.create.handle(i8 1)
+  store ptr %206, ptr getelementptr inbounds ([64 x [2 x %"class.hlsl::RWBuffer"]], ptr @"?AUI@@3PAY01V?$RWBuffer@M@hlsl@@A", i32 0, i32 39, i32 0), align 4, !tbaa !6
+  %207 = tail call ptr @llvm.dx.create.handle(i8 1)
+  store ptr %207, ptr getelementptr inbounds ([64 x [2 x %"class.hlsl::RWBuffer"]], ptr @"?AUI@@3PAY01V?$RWBuffer@M@hlsl@@A", i32 0, i32 39, i32 1), align 4, !tbaa !6
+  %208 = tail call ptr @llvm.dx.create.handle(i8 1)
+  store ptr %208, ptr getelementptr inbounds ([64 x [2 x %"class.hlsl::RWBuffer"]], ptr @"?AUI@@3PAY01V?$RWBuffer@M@hlsl@@A", i32 0, i32 40, i32 0), align 4, !tbaa !6
+  %209 = tail call ptr @llvm.dx.create.handle(i8 1)
+  store ptr %209, ptr getelementptr inbounds ([64 x [2 x %"class.hlsl::RWBuffer"]], ptr @"?AUI@@3PAY01V?$RWBuffer@M@hlsl@@A", i32 0, i32 40, i32 1), align 4, !tbaa !6
+  %210 = tail call ptr @llvm.dx.create.handle(i8 1)
+  store ptr %210, ptr getelementptr inbounds ([64 x [2 x %"class.hlsl::RWBuffer"]], ptr @"?AUI@@3PAY01V?$RWBuffer@M@hlsl@@A", i32 0, i32 41, i32 0), align 4, !tbaa !6
+  %211 = tail call ptr @llvm.dx.create.handle(i8 1)
+  store ptr %211, ptr getelementptr inbounds ([64 x [2 x %"class.hlsl::RWBuffer"]], ptr @"?AUI@@3PAY01V?$RWBuffer@M@hlsl@@A", i32 0, i32 41, i32 1), align 4, !tbaa !6
+  %212 = tail call ptr @llvm.dx.create.handle(i8 1)
+  store ptr %212, ptr getelementptr inbounds ([64 x [2 x %"class.hlsl::RWBuffer"]], ptr @"?AUI@@3PAY01V?$RWBuffer@M@hlsl@@A", i32 0, i32 42, i32 0), align 4, !tbaa !6
+  %213 = tail call ptr @llvm.dx.create.handle(i8 1)
+  store ptr %213, ptr getelementptr inbounds ([64 x [2 x %"class.hlsl::RWBuffer"]], ptr @"?AUI@@3PAY01V?$RWBuffer@M@hlsl@@A", i32 0, i32 42, i32 1), align 4, !tbaa !6
+  %214 = tail call ptr @llvm.dx.create.handle(i8 1)
+  store ptr %214, ptr getelementptr inbounds ([64 x [2 x %"class.hlsl::RWBuffer"]], ptr @"?AUI@@3PAY01V?$RWBuffer@M@hlsl@@A", i32 0, i32 43, i32 0), align 4, !tbaa !6
+  %215 = tail call ptr @llvm.dx.create.handle(i8 1)
+  store ptr %215, ptr getelementptr inbounds ([64 x [2 x %"class.hlsl::RWBuffer"]], ptr @"?AUI@@3PAY01V?$RWBuffer@M@hlsl@@A", i32 0, i32 43, i32 1), align 4, !tbaa !6
+  %216 = tail call ptr @llvm.dx.create.handle(i8 1)
+  store ptr %216, ptr getelementptr inbounds ([64 x [2 x %"class.hlsl::RWBuffer"]], ptr @"?AUI@@3PAY01V?$RWBuffer@M@hlsl@@A", i32 0, i32 44, i32 0), align 4, !tbaa !6
+  %217 = tail call ptr @llvm.dx.create.handle(i8 1)
+  store ptr %217, ptr getelementptr inbounds ([64 x [2 x %"class.hlsl::RWBuffer"]], ptr @"?AUI@@3PAY01V?$RWBuffer@M@hlsl@@A", i32 0, i32 44, i32 1), align 4, !tbaa !6
+  %218 = tail call ptr @llvm.dx.create.handle(i8 1)
+  store ptr %218, ptr getelementptr inbounds ([64 x [2 x %"class.hlsl::RWBuffer"]], ptr @"?AUI@@3PAY01V?$RWBuffer@M@hlsl@@A", i32 0, i32 45, i32 0), align 4, !tbaa !6
+  %219 = tail call ptr @llvm.dx.create.handle(i8 1)
+  store ptr %219, ptr getelementptr inbounds ([64 x [2 x %"class.hlsl::RWBuffer"]], ptr @"?AUI@@3PAY01V?$RWBuffer@M@hlsl@@A", i32 0, i32 45, i32 1), align 4, !tbaa !6
+  %220 = tail call ptr @llvm.dx.create.handle(i8 1)
+  store ptr %220, ptr getelementptr inbounds ([64 x [2 x %"class.hlsl::RWBuffer"]], ptr @"?AUI@@3PAY01V?$RWBuffer@M@hlsl@@A", i32 0, i32 46, i32 0), align 4, !tbaa !6
+  %221 = tail call ptr @llvm.dx.create.handle(i8 1)
+  store ptr %221, ptr getelementptr inbounds ([64 x [2 x %"class.hlsl::RWBuffer"]], ptr @"?AUI@@3PAY01V?$RWBuffer@M@hlsl@@A", i32 0, i32 46, i32 1), align 4, !tbaa !6
+  %222 = tail call ptr @llvm.dx.create.handle(i8 1)
+  store ptr %222, ptr getelementptr inbounds ([64 x [2 x %"class.hlsl::RWBuffer"]], ptr @"?AUI@@3PAY01V?$RWBuffer@M@hlsl@@A", i32 0, i32 47, i32 0), align 4, !tbaa !6
+  %223 = tail call ptr @llvm.dx.create.handle(i8 1)
+  store ptr %223, ptr getelementptr inbounds ([64 x [2 x %"class.hlsl::RWBuffer"]], ptr @"?AUI@@3PAY01V?$RWBuffer@M@hlsl@@A", i32 0, i32 47, i32 1), align 4, !tbaa !6
+  %224 = tail call ptr @llvm.dx.create.handle(i8 1)
+  store ptr %224, ptr getelementptr inbounds ([64 x [2 x %"class.hlsl::RWBuffer"]], ptr @"?AUI@@3PAY01V?$RWBuffer@M@hlsl@@A", i32 0, i32 48, i32 0), align 4, !tbaa !6
+  %225 = tail call ptr @llvm.dx.create.handle(i8 1)
+  store ptr %225, ptr getelementptr inbounds ([64 x [2 x %"class.hlsl::RWBuffer"]], ptr @"?AUI@@3PAY01V?$RWBuffer@M@hlsl@@A", i32 0, i32 48, i32 1), align 4, !tbaa !6
+  %226 = tail call ptr @llvm.dx.create.handle(i8 1)
+  store ptr %226, ptr getelementptr inbounds ([64 x [2 x %"class.hlsl::RWBuffer"]], ptr @"?AUI@@3PAY01V?$RWBuffer@M@hlsl@@A", i32 0, i32 49, i32 0), align 4, !tbaa !6
+  %227 = tail call ptr @llvm.dx.create.handle(i8 1)
+  store ptr %227, ptr getelementptr inbounds ([64 x [2 x %"class.hlsl::RWBuffer"]], ptr @"?AUI@@3PAY01V?$RWBuffer@M@hlsl@@A", i32 0, i32 49, i32 1), align 4, !tbaa !6
+  %228 = tail call ptr @llvm.dx.create.handle(i8 1)
+  store ptr %228, ptr getelementptr inbounds ([64 x [2 x %"class.hlsl::RWBuffer"]], ptr @"?AUI@@3PAY01V?$RWBuffer@M@hlsl@@A", i32 0, i32 50, i32 0), align 4, !tbaa !6
+  %229 = tail call ptr @llvm.dx.create.handle(i8 1)
+  store ptr %229, ptr getelementptr inbounds ([64 x [2 x %"class.hlsl::RWBuffer"]], ptr @"?AUI@@3PAY01V?$RWBuffer@M@hlsl@@A", i32 0, i32 50, i32 1), align 4, !tbaa !6
+  %230 = tail call ptr @llvm.dx.create.handle(i8 1)
+  store ptr %230, ptr getelementptr inbounds ([64 x [2 x %"class.hlsl::RWBuffer"]], ptr @"?AUI@@3PAY01V?$RWBuffer@M@hlsl@@A", i32 0, i32 51, i32 0), align 4, !tbaa !6
+  %231 = tail call ptr @llvm.dx.create.handle(i8 1)
+  store ptr %231, ptr getelementptr inbounds ([64 x [2 x %"class.hlsl::RWBuffer"]], ptr @"?AUI@@3PAY01V?$RWBuffer@M@hlsl@@A", i32 0, i32 51, i32 1), align 4, !tbaa !6
+  %232 = tail call ptr @llvm.dx.create.handle(i8 1)
+  store ptr %232, ptr getelementptr inbounds ([64 x [2 x %"class.hlsl::RWBuffer"]], ptr @"?AUI@@3PAY01V?$RWBuffer@M@hlsl@@A", i32 0, i32 52, i32 0), align 4, !tbaa !6
+  %233 = tail call ptr @llvm.dx.create.handle(i8 1)
+  store ptr %233, ptr getelementptr inbounds ([64 x [2 x %"class.hlsl::RWBuffer"]], ptr @"?AUI@@3PAY01V?$RWBuffer@M@hlsl@@A", i32 0, i32 52, i32 1), align 4, !tbaa !6
+  %234 = tail call ptr @llvm.dx.create.handle(i8 1)
+  store ptr %234, ptr getelementptr inbounds ([64 x [2 x %"class.hlsl::RWBuffer"]], ptr @"?AUI@@3PAY01V?$RWBuffer@M@hlsl@@A", i32 0, i32 53, i32 0), align 4, !tbaa !6
+  %235 = tail call ptr @llvm.dx.create.handle(i8 1)
+  store ptr %235, ptr getelementptr inbounds ([64 x [2 x %"class.hlsl::RWBuffer"]], ptr @"?AUI@@3PAY01V?$RWBuffer@M@hlsl@@A", i32 0, i32 53, i32 1), align 4, !tbaa !6
+  %236 = tail call ptr @llvm.dx.create.handle(i8 1)
+  store ptr %236, ptr getelementptr inbounds ([64 x [2 x %"class.hlsl::RWBuffer"]], ptr @"?AUI@@3PAY01V?$RWBuffer@M@hlsl@@A", i32 0, i32 54, i32 0), align 4, !tbaa !6
+  %237 = tail call ptr @llvm.dx.create.handle(i8 1)
+  store ptr %237, ptr getelementptr inbounds ([64 x [2 x %"class.hlsl::RWBuffer"]], ptr @"?AUI@@3PAY01V?$RWBuffer@M@hlsl@@A", i32 0, i32 54, i32 1), align 4, !tbaa !6
+  %238 = tail call ptr @llvm.dx.create.handle(i8 1)
+  store ptr %238, ptr getelementptr inbounds ([64 x [2 x %"class.hlsl::RWBuffer"]], ptr @"?AUI@@3PAY01V?$RWBuffer@M@hlsl@@A", i32 0, i32 55, i32 0), align 4, !tbaa !6
+  %239 = tail call ptr @llvm.dx.create.handle(i8 1)
+  store ptr %239, ptr getelementptr inbounds ([64 x [2 x %"class.hlsl::RWBuffer"]], ptr @"?AUI@@3PAY01V?$RWBuffer@M@hlsl@@A", i32 0, i32 55, i32 1), align 4, !tbaa !6
+  %240 = tail call ptr @llvm.dx.create.handle(i8 1)
+  store ptr %240, ptr getelementptr inbounds ([64 x [2 x %"class.hlsl::RWBuffer"]], ptr @"?AUI@@3PAY01V?$RWBuffer@M@hlsl@@A", i32 0, i32 56, i32 0), align 4, !tbaa !6
+  %241 = tail call ptr @llvm.dx.create.handle(i8 1)
+  store ptr %241, ptr getelementptr inbounds ([64 x [2 x %"class.hlsl::RWBuffer"]], ptr @"?AUI@@3PAY01V?$RWBuffer@M@hlsl@@A", i32 0, i32 56, i32 1), align 4, !tbaa !6
+  %242 = tail call ptr @llvm.dx.create.handle(i8 1)
+  store ptr %242, ptr getelementptr inbounds ([64 x [2 x %"class.hlsl::RWBuffer"]], ptr @"?AUI@@3PAY01V?$RWBuffer@M@hlsl@@A", i32 0, i32 57, i32 0), align 4, !tbaa !6
+  %243 = tail call ptr @llvm.dx.create.handle(i8 1)
+  store ptr %243, ptr getelementptr inbounds ([64 x [2 x %"class.hlsl::RWBuffer"]], ptr @"?AUI@@3PAY01V?$RWBuffer@M@hlsl@@A", i32 0, i32 57, i32 1), align 4, !tbaa !6
+  %244 = tail call ptr @llvm.dx.create.handle(i8 1)
+  store ptr %244, ptr getelementptr inbounds ([64 x [2 x %"class.hlsl::RWBuffer"]], ptr @"?AUI@@3PAY01V?$RWBuffer@M@hlsl@@A", i32 0, i32 58, i32 0), align 4, !tbaa !6
+  %245 = tail call ptr @llvm.dx.create.handle(i8 1)
+  store ptr %245, ptr getelementptr inbounds ([64 x [2 x %"class.hlsl::RWBuffer"]], ptr @"?AUI@@3PAY01V?$RWBuffer@M@hlsl@@A", i32 0, i32 58, i32 1), align 4, !tbaa !6
+  %246 = tail call ptr @llvm.dx.create.handle(i8 1)
+  store ptr %246, ptr getelementptr inbounds ([64 x [2 x %"class.hlsl::RWBuffer"]], ptr @"?AUI@@3PAY01V?$RWBuffer@M@hlsl@@A", i32 0, i32 59, i32 0), align 4, !tbaa !6
+  %247 = tail call ptr @llvm.dx.create.handle(i8 1)
+  store ptr %247, ptr getelementptr inbounds ([64 x [2 x %"class.hlsl::RWBuffer"]], ptr @"?AUI@@3PAY01V?$RWBuffer@M@hlsl@@A", i32 0, i32 59, i32 1), align 4, !tbaa !6
+  %248 = tail call ptr @llvm.dx.create.handle(i8 1)
+  store ptr %248, ptr getelementptr inbounds ([64 x [2 x %"class.hlsl::RWBuffer"]], ptr @"?AUI@@3PAY01V?$RWBuffer@M@hlsl@@A", i32 0, i32 60, i32 0), align 4, !tbaa !6
+  %249 = tail call ptr @llvm.dx.create.handle(i8 1)
+  store ptr %249, ptr getelementptr inbounds ([64 x [2 x %"class.hlsl::RWBuffer"]], ptr @"?AUI@@3PAY01V?$RWBuffer@M@hlsl@@A", i32 0, i32 60, i32 1), align 4, !tbaa !6
+  %250 = tail call ptr @llvm.dx.create.handle(i8 1)
+  store ptr %250, ptr getelementptr inbounds ([64 x [2 x %"class.hlsl::RWBuffer"]], ptr @"?AUI@@3PAY01V?$RWBuffer@M@hlsl@@A", i32 0, i32 61, i32 0), align 4, !tbaa !6
+  %251 = tail call ptr @llvm.dx.create.handle(i8 1)
+  store ptr %251, ptr getelementptr inbounds ([64 x [2 x %"class.hlsl::RWBuffer"]], ptr @"?AUI@@3PAY01V?$RWBuffer@M@hlsl@@A", i32 0, i32 61, i32 1), align 4, !tbaa !6
+  %252 = tail call ptr @llvm.dx.create.handle(i8 1)
+  store ptr %252, ptr getelementptr inbounds ([64 x [2 x %"class.hlsl::RWBuffer"]], ptr @"?AUI@@3PAY01V?$RWBuffer@M@hlsl@@A", i32 0, i32 62, i32 0), align 4, !tbaa !6
+  %253 = tail call ptr @llvm.dx.create.handle(i8 1)
+  store ptr %253, ptr getelementptr inbounds ([64 x [2 x %"class.hlsl::RWBuffer"]], ptr @"?AUI@@3PAY01V?$RWBuffer@M@hlsl@@A", i32 0, i32 62, i32 1), align 4, !tbaa !6
+  %254 = tail call ptr @llvm.dx.create.handle(i8 1)
+  store ptr %254, ptr getelementptr inbounds ([64 x [2 x %"class.hlsl::RWBuffer"]], ptr @"?AUI@@3PAY01V?$RWBuffer@M@hlsl@@A", i32 0, i32 63, i32 0), align 4, !tbaa !6
+  %255 = tail call ptr @llvm.dx.create.handle(i8 1)
+  store ptr %255, ptr getelementptr inbounds ([64 x [2 x %"class.hlsl::RWBuffer"]], ptr @"?AUI@@3PAY01V?$RWBuffer@M@hlsl@@A", i32 0, i32 63, i32 1), align 4, !tbaa !6
+  %256 = tail call i32 @llvm.dx.thread.id(i32 0)
+  %257 = load ptr, ptr getelementptr inbounds ([64 x [2 x %"class.hlsl::RWBuffer"]], ptr @"?AUI@@3PAY01V?$RWBuffer@M@hlsl@@A", i32 0, i32 5, i32 1), align 4, !tbaa !6
+  %arrayidx.i.i = getelementptr inbounds float, ptr %257, i32 %256
+  %258 = load float, ptr %arrayidx.i.i, align 4, !tbaa !11
+  %rem.i = and i32 %256, 1
+  %rem1.i = and i32 %256, 63
+  %arrayidx2.i = getelementptr inbounds [2 x [64 x %"class.hlsl::RWBuffer"]], ptr @"?AU@@3PAY0EA@V?$RWBuffer@M@hlsl@@A", i32 0, i32 %rem.i, i32 %rem1.i
+  %259 = load ptr, ptr %arrayidx2.i, align 4, !tbaa !6
+  %arrayidx.i7.i = getelementptr inbounds float, ptr %259, i32 %256
+  store float %258, ptr %arrayidx.i7.i, align 4, !tbaa !11
+  ret void
+}
+
+; Function Attrs: mustprogress nofree nosync nounwind readnone willreturn
+declare i32 @llvm.dx.thread.id(i32 %0) #1
+
+; Function Attrs: mustprogress nounwind willreturn
+declare ptr @llvm.dx.create.handle(i8 %0) #2
+
+attributes #0 = { mustprogress norecurse nounwind willreturn "frame-pointer"="all" "hlsl.numthreads"="64,1,1" "hlsl.shader"="compute" "no-trapping-math"="true" "stack-protector-buffer-size"="8" }
+attributes #1 = { mustprogress nofree nosync nounwind readnone willreturn }
+attributes #2 = { mustprogress nounwind willreturn }
+
+!hlsl.uavs = !{!0, !1}
+!llvm.module.flags = !{!2, !3}
+!dx.valver = !{!4}
+!llvm.ident = !{!5}
+
+!0 = !{ptr @"?AU@@3PAY0EA@V?$RWBuffer@M@hlsl@@A", !"RWBuffer<float>", i32 0, i32 10, i32 2, i32 0}
+!1 = !{ptr @"?AUI@@3PAY01V?$RWBuffer@M@hlsl@@A", !"RWBuffer<float>", i32 1, i32 10, i32 2, i32 1}
+!2 = !{i32 1, !"wchar_size", i32 4}
+!3 = !{i32 7, !"frame-pointer", i32 2}
+!4 = !{i32 1, i32 7}
+!5 = !{!"clang version 16.0.0 (https://github.com/llvm/llvm-project.git 0b4d05611d4745493490893d0e61b73e947d08d6)"}
+!6 = !{!7, !8, i64 0}
+!7 = !{!"?AV?$RWBuffer@M@hlsl@@", !8, i64 0}
+!8 = !{!"any pointer", !9, i64 0}
+!9 = !{!"omnipotent char", !10, i64 0}
+!10 = !{!"Simple C++ TBAA"}
+!11 = !{!12, !12, i64 0}
+!12 = !{!"float", !9, i64 0}

llvm/test/CodeGen/DirectX/typed_buf_array2.ll

+; RUN: opt -S -dxil-typedbuf-lower < %s | FileCheck %s
+
+;RWBuffer<float> AU[16] : register(u2);
+;  AU[ID%16][ID] = AU[5][ID];
+;  AU[(ID+8)%16][ID] = AU[3][ID];
+
+target datalayout = "e-m:e-p:32:32-i1:32-i8:8-i16:16-i32:32-i64:64-f16:16-f32:32-f64:64-n8:16:32:64"
+target triple = "dxil-unknown-shadermodel6.0-compute"
+
+
+; CHECK: %[[HDL3:.+]] = call %dx.types.Handle @dx.op.createHandle(i32 57, i8 1, i32 0, i32 5, i1 false)
+; CHECK: %[[HDL5:.+]] = call %dx.types.Handle @dx.op.createHandle(i32 57, i8 1, i32 0, i32 7, i1 false)
+
+; CHECK: %[[ID:.+]] = tail call i32 @llvm.dx.thread.id(i32 0)
+
+; CHECK: %[[LD1:.+]] = call %dx.types.ResRet.f32 @dx.op.bufferLoad.f32(i32 68, %dx.types.Handle %[[HDL5]], i32 %[[ID]], i32 poison)
+; CHECK: %[[LD1_X:.+]] = extractvalue %dx.types.ResRet.f32 %[[LD1]], 0
+
+; CHECK: %[[REM16:.+]] = and i32 %[[ID]], 15
+; CHECK: %[[ResIndex1:.+]] = add i32 2, %[[REM16]]
+; CHECK: %[[HDLREM16:.+]] = call %dx.types.Handle @dx.op.createHandle(i32 57, i8 1, i32 0, i32 %[[ResIndex1]], i1 false)
+
+; CHECK: call void @dx.op.bufferStore.f32(i32 69, %dx.types.Handle %[[HDLREM16]], i32 %[[ID]], i32 poison, float %[[LD1_X]], float poison, float poison, float poison, i8 1)
+
+
+; CHECK: %[[LD2:.+]] = call %dx.types.ResRet.f32 @dx.op.bufferLoad.f32(i32 68, %dx.types.Handle %[[HDL3]], i32 %[[ID]], i32 poison)
+; CHECK: %[[LD2_X:.+]] = extractvalue %dx.types.ResRet.f32 %[[LD2]], 0
+
+; CHECK: %[[ID_8:.+]] = add i32 %18, 8
+; CHECK: %[[ID_8_REM16:.+]] = and i32 %[[ID_8]], 15
+; CHECK: %[[ResIndex2:.+]] = add i32 2, %[[ID_8_REM16]]
+; CHECK: %[[HDL_ID8_REM16:.+]] = call %dx.types.Handle @dx.op.createHandle(i32 57, i8 1, i32 0, i32 %[[ResIndex2]], i1 false)
+
+; CHECK: call void @dx.op.bufferStore.f32(i32 69, %dx.types.Handle %[[HDL_ID8_REM16]], i32 %[[ID]], i32 poison, float %[[LD2_X]], float poison, float poison, float poison, i8 1)
+; CHECK-NEXT: ret void
+
+%"class.hlsl::RWBuffer" = type { ptr }
+
+@"?AU@@3PAV?$RWBuffer@M@hlsl@@A" = local_unnamed_addr global [16 x %"class.hlsl::RWBuffer"] zeroinitializer, align 4
+
+; Function Attrs: mustprogress norecurse nounwind willreturn
+define void @main() local_unnamed_addr #0 {
+entry:
+  %0 = tail call ptr @llvm.dx.create.handle(i8 1)
+  store ptr %0, ptr @"?AU@@3PAV?$RWBuffer@M@hlsl@@A", align 4, !tbaa !5
+  %1 = tail call ptr @llvm.dx.create.handle(i8 1)
+  store ptr %1, ptr getelementptr inbounds ([16 x %"class.hlsl::RWBuffer"], ptr @"?AU@@3PAV?$RWBuffer@M@hlsl@@A", i32 0, i32 1), align 4, !tbaa !5
+  %2 = tail call ptr @llvm.dx.create.handle(i8 1)
+  store ptr %2, ptr getelementptr inbounds ([16 x %"class.hlsl::RWBuffer"], ptr @"?AU@@3PAV?$RWBuffer@M@hlsl@@A", i32 0, i32 2), align 4, !tbaa !5
+  %3 = tail call ptr @llvm.dx.create.handle(i8 1)
+  store ptr %3, ptr getelementptr inbounds ([16 x %"class.hlsl::RWBuffer"], ptr @"?AU@@3PAV?$RWBuffer@M@hlsl@@A", i32 0, i32 3), align 4, !tbaa !5
+  %4 = tail call ptr @llvm.dx.create.handle(i8 1)
+  store ptr %4, ptr getelementptr inbounds ([16 x %"class.hlsl::RWBuffer"], ptr @"?AU@@3PAV?$RWBuffer@M@hlsl@@A", i32 0, i32 4), align 4, !tbaa !5
+  %5 = tail call ptr @llvm.dx.create.handle(i8 1)
+  store ptr %5, ptr getelementptr inbounds ([16 x %"class.hlsl::RWBuffer"], ptr @"?AU@@3PAV?$RWBuffer@M@hlsl@@A", i32 0, i32 5), align 4, !tbaa !5
+  %6 = tail call ptr @llvm.dx.create.handle(i8 1)
+  store ptr %6, ptr getelementptr inbounds ([16 x %"class.hlsl::RWBuffer"], ptr @"?AU@@3PAV?$RWBuffer@M@hlsl@@A", i32 0, i32 6), align 4, !tbaa !5
+  %7 = tail call ptr @llvm.dx.create.handle(i8 1)
+  store ptr %7, ptr getelementptr inbounds ([16 x %"class.hlsl::RWBuffer"], ptr @"?AU@@3PAV?$RWBuffer@M@hlsl@@A", i32 0, i32 7), align 4, !tbaa !5
+  %8 = tail call ptr @llvm.dx.create.handle(i8 1)
+  store ptr %8, ptr getelementptr inbounds ([16 x %"class.hlsl::RWBuffer"], ptr @"?AU@@3PAV?$RWBuffer@M@hlsl@@A", i32 0, i32 8), align 4, !tbaa !5
+  %9 = tail call ptr @llvm.dx.create.handle(i8 1)
+  store ptr %9, ptr getelementptr inbounds ([16 x %"class.hlsl::RWBuffer"], ptr @"?AU@@3PAV?$RWBuffer@M@hlsl@@A", i32 0, i32 9), align 4, !tbaa !5
+  %10 = tail call ptr @llvm.dx.create.handle(i8 1)
+  store ptr %10, ptr getelementptr inbounds ([16 x %"class.hlsl::RWBuffer"], ptr @"?AU@@3PAV?$RWBuffer@M@hlsl@@A", i32 0, i32 10), align 4, !tbaa !5
+  %11 = tail call ptr @llvm.dx.create.handle(i8 1)
+  store ptr %11, ptr getelementptr inbounds ([16 x %"class.hlsl::RWBuffer"], ptr @"?AU@@3PAV?$RWBuffer@M@hlsl@@A", i32 0, i32 11), align 4, !tbaa !5
+  %12 = tail call ptr @llvm.dx.create.handle(i8 1)
+  store ptr %12, ptr getelementptr inbounds ([16 x %"class.hlsl::RWBuffer"], ptr @"?AU@@3PAV?$RWBuffer@M@hlsl@@A", i32 0, i32 12), align 4, !tbaa !5
+  %13 = tail call ptr @llvm.dx.create.handle(i8 1)
+  store ptr %13, ptr getelementptr inbounds ([16 x %"class.hlsl::RWBuffer"], ptr @"?AU@@3PAV?$RWBuffer@M@hlsl@@A", i32 0, i32 13), align 4, !tbaa !5
+  %14 = tail call ptr @llvm.dx.create.handle(i8 1)
+  store ptr %14, ptr getelementptr inbounds ([16 x %"class.hlsl::RWBuffer"], ptr @"?AU@@3PAV?$RWBuffer@M@hlsl@@A", i32 0, i32 14), align 4, !tbaa !5
+  %15 = tail call ptr @llvm.dx.create.handle(i8 1)
+  store ptr %15, ptr getelementptr inbounds ([16 x %"class.hlsl::RWBuffer"], ptr @"?AU@@3PAV?$RWBuffer@M@hlsl@@A", i32 0, i32 15), align 4, !tbaa !5
+  %16 = tail call i32 @llvm.dx.thread.id(i32 0)
+  %17 = load ptr, ptr getelementptr inbounds ([16 x %"class.hlsl::RWBuffer"], ptr @"?AU@@3PAV?$RWBuffer@M@hlsl@@A", i32 0, i32 5), align 4, !tbaa !5
+  %arrayidx.i.i = getelementptr inbounds float, ptr %17, i32 %16
+  %18 = load float, ptr %arrayidx.i.i, align 4, !tbaa !10
+  %rem.i = and i32 %16, 15
+  %arrayidx.i = getelementptr inbounds [16 x %"class.hlsl::RWBuffer"], ptr @"?AU@@3PAV?$RWBuffer@M@hlsl@@A", i32 0, i32 %rem.i
+  %19 = load ptr, ptr %arrayidx.i, align 4, !tbaa !5
+  %arrayidx.i11.i = getelementptr inbounds float, ptr %19, i32 %16
+  store float %18, ptr %arrayidx.i11.i, align 4, !tbaa !10
+  %20 = load ptr, ptr getelementptr inbounds ([16 x %"class.hlsl::RWBuffer"], ptr @"?AU@@3PAV?$RWBuffer@M@hlsl@@A", i32 0, i32 3), align 4, !tbaa !5
+  %arrayidx.i12.i = getelementptr inbounds float, ptr %20, i32 %16
+  %21 = load float, ptr %arrayidx.i12.i, align 4, !tbaa !10
+  %add.i = add i32 %16, 8
+  %rem3.i = and i32 %add.i, 15
+  %arrayidx4.i = getelementptr inbounds [16 x %"class.hlsl::RWBuffer"], ptr @"?AU@@3PAV?$RWBuffer@M@hlsl@@A", i32 0, i32 %rem3.i
+  %22 = load ptr, ptr %arrayidx4.i, align 4, !tbaa !5
+  %arrayidx.i13.i = getelementptr inbounds float, ptr %22, i32 %16
+  store float %21, ptr %arrayidx.i13.i, align 4, !tbaa !10
+  ret void
+}
+
+; Function Attrs: mustprogress nofree nosync nounwind readnone willreturn
+declare i32 @llvm.dx.thread.id(i32 %0) #1
+
+; Function Attrs: mustprogress nounwind willreturn
+declare ptr @llvm.dx.create.handle(i8 %0) #2
+
+attributes #0 = { mustprogress norecurse nounwind willreturn "frame-pointer"="all" "hlsl.numthreads"="8,1,1" "hlsl.shader"="compute" "no-trapping-math"="true" "stack-protector-buffer-size"="8" }
+attributes #1 = { mustprogress nofree nosync nounwind readnone willreturn }
+attributes #2 = { mustprogress nounwind willreturn }
+
+!hlsl.uavs = !{!0}
+!llvm.module.flags = !{!1, !2}
+!dx.valver = !{!3}
+!llvm.ident = !{!4}
+
+!0 = !{ptr @"?AU@@3PAV?$RWBuffer@M@hlsl@@A", !"RWBuffer<float>", i32 0, i32 10, i32 2, i32 0}
+!1 = !{i32 1, !"wchar_size", i32 4}
+!2 = !{i32 7, !"frame-pointer", i32 2}
+!3 = !{i32 1, i32 7}
+!4 = !{!"clang version 16.0.0 (https://github.com/llvm/llvm-project.git 0b4d05611d4745493490893d0e61b73e947d08d6)"}
+!5 = !{!6, !7, i64 0}
+!6 = !{!"?AV?$RWBuffer@M@hlsl@@", !7, i64 0}
+!7 = !{!"any pointer", !8, i64 0}
+!8 = !{!"omnipotent char", !9, i64 0}
+!9 = !{!"Simple C++ TBAA"}
+!10 = !{!11, !11, i64 0}
+!11 = !{!"float", !8, i64 0}