clang/lib/CodeGen/CGCUDANV.cpp

   1 //===----- CGCUDANV.cpp - Interface to NVIDIA CUDA Runtime ----------------===//
   2 //
   3 // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
   4 // See https://llvm.org/LICENSE.txt for license information.
   5 // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
   6 //
   7 //===----------------------------------------------------------------------===//
   8 //
   9 // This provides a class for CUDA code generation targeting the NVIDIA CUDA
  10 // runtime library.
  11 //
  12 //===----------------------------------------------------------------------===//
  13
  14 #include "CGCUDARuntime.h"
  15 #include "CGCXXABI.h"
  16 #include "CodeGenFunction.h"
  17 #include "CodeGenModule.h"
  18 #include "clang/AST/Decl.h"
  19 #include "clang/Basic/Cuda.h"
  20 #include "clang/CodeGen/CodeGenABITypes.h"
  21 #include "clang/CodeGen/ConstantInitBuilder.h"
  22 #include "llvm/IR/BasicBlock.h"
  23 #include "llvm/IR/Constants.h"
  24 #include "llvm/IR/DerivedTypes.h"
  25 #include "llvm/IR/ReplaceConstant.h"
  26 #include "llvm/Support/Format.h"
  27 #include "llvm/Support/VirtualFileSystem.h"
  28
  29 using namespace clang;
  30 using namespace CodeGen;
  31
  32 namespace {
  33 constexpr unsigned CudaFatMagic = 0x466243b1;
  34 constexpr unsigned HIPFatMagic = 0x48495046; // "HIPF"
  35
  36 class CGNVCUDARuntime : public CGCUDARuntime {
  37
  38 private:
  39   llvm::IntegerType *IntTy, *SizeTy;
  40   llvm::Type *VoidTy;
  41   llvm::PointerType *CharPtrTy, *VoidPtrTy, *VoidPtrPtrTy;
  42
  43   /// Convenience reference to LLVM Context
  44   llvm::LLVMContext &Context;
  45   /// Convenience reference to the current module
  46   llvm::Module &TheModule;
  47   /// Keeps track of kernel launch stubs and handles emitted in this module
  48   struct KernelInfo {
  49     llvm::Function *Kernel; // stub function to help launch kernel
  50     const Decl *D;
  51   };
  52   llvm::SmallVector<KernelInfo, 16> EmittedKernels;
  53   // Map a kernel mangled name to a symbol for identifying kernel in host code
  54   // For CUDA, the symbol for identifying the kernel is the same as the device
  55   // stub function. For HIP, they are different.
  56   llvm::DenseMap<StringRef, llvm::GlobalValue *> KernelHandles;
  57   // Map a kernel handle to the kernel stub.
  58   llvm::DenseMap<llvm::GlobalValue *, llvm::Function *> KernelStubs;
  59   struct VarInfo {
  60     llvm::GlobalVariable *Var;
  61     const VarDecl *D;
  62     DeviceVarFlags Flags;
  63   };
  64   llvm::SmallVector<VarInfo, 16> DeviceVars;
  65   /// Keeps track of variable containing handle of GPU binary. Populated by
  66   /// ModuleCtorFunction() and used to create corresponding cleanup calls in
  67   /// ModuleDtorFunction()
  68   llvm::GlobalVariable *GpuBinaryHandle = nullptr;
  69   /// Whether we generate relocatable device code.
  70   bool RelocatableDeviceCode;
  71   /// Mangle context for device.
  72   std::unique_ptr<MangleContext> DeviceMC;
  73   /// Some zeros used for GEPs.
  74   llvm::Constant *Zeros[2];
  75
  76   llvm::FunctionCallee getSetupArgumentFn() const;
  77   llvm::FunctionCallee getLaunchFn() const;
  78
  79   llvm::FunctionType *getRegisterGlobalsFnTy() const;
  80   llvm::FunctionType *getCallbackFnTy() const;
  81   llvm::FunctionType *getRegisterLinkedBinaryFnTy() const;
  82   std::string addPrefixToName(StringRef FuncName) const;
  83   std::string addUnderscoredPrefixToName(StringRef FuncName) const;
  84
  85   /// Creates a function to register all kernel stubs generated in this module.
  86   llvm::Function *makeRegisterGlobalsFn();
  87
  88   /// Helper function that generates a constant string and returns a pointer to
  89   /// the start of the string.  The result of this function can be used anywhere
  90   /// where the C code specifies const char*.
  91   llvm::Constant *makeConstantString(const std::string &Str,
  92                                      const std::string &Name = "") {
  93     auto ConstStr = CGM.GetAddrOfConstantCString(Str, Name.c_str());
  94     return llvm::ConstantExpr::getGetElementPtr(ConstStr.getElementType(),
  95                                                 ConstStr.getPointer(), Zeros);
  96   }
  97
  98   /// Helper function which generates an initialized constant array from Str,
  99   /// and optionally sets section name and alignment. AddNull specifies whether
 100   /// the array should nave NUL termination.
 101   llvm::Constant *makeConstantArray(StringRef Str,
 102                                     StringRef Name = "",
 103                                     StringRef SectionName = "",
 104                                     unsigned Alignment = 0,
 105                                     bool AddNull = false) {
 106     llvm::Constant *Value =
 107         llvm::ConstantDataArray::getString(Context, Str, AddNull);
 108     auto *GV = new llvm::GlobalVariable(
 109         TheModule, Value->getType(), /*isConstant=*/true,
 110         llvm::GlobalValue::PrivateLinkage, Value, Name);
 111     if (!SectionName.empty()) {
 112       GV->setSection(SectionName);
 113       // Mark the address as used which make sure that this section isn't
 114       // merged and we will really have it in the object file.
 115       GV->setUnnamedAddr(llvm::GlobalValue::UnnamedAddr::None);
 116     }
 117     if (Alignment)
 118       GV->setAlignment(llvm::Align(Alignment));
 119     return llvm::ConstantExpr::getGetElementPtr(GV->getValueType(), GV, Zeros);
 120   }
 121
 122   /// Helper function that generates an empty dummy function returning void.
 123   llvm::Function *makeDummyFunction(llvm::FunctionType *FnTy) {
 124     assert(FnTy->getReturnType()->isVoidTy() &&
 125            "Can only generate dummy functions returning void!");
 126     llvm::Function *DummyFunc = llvm::Function::Create(
 127         FnTy, llvm::GlobalValue::InternalLinkage, "dummy", &TheModule);
 128
 129     llvm::BasicBlock *DummyBlock =
 130         llvm::BasicBlock::Create(Context, "", DummyFunc);
 131     CGBuilderTy FuncBuilder(CGM, Context);
 132     FuncBuilder.SetInsertPoint(DummyBlock);
 133     FuncBuilder.CreateRetVoid();
 134
 135     return DummyFunc;
 136   }
 137
 138   void emitDeviceStubBodyLegacy(CodeGenFunction &CGF, FunctionArgList &Args);
 139   void emitDeviceStubBodyNew(CodeGenFunction &CGF, FunctionArgList &Args);
 140   std::string getDeviceSideName(const NamedDecl *ND) override;
 141
 142   void registerDeviceVar(const VarDecl *VD, llvm::GlobalVariable &Var,
 143                          bool Extern, bool Constant) {
 144     DeviceVars.push_back({&Var,
 145                           VD,
 146                           {DeviceVarFlags::Variable, Extern, Constant,
 147                            VD->hasAttr<HIPManagedAttr>(),
 148                            /*Normalized*/ false, 0}});
 149   }
 150   void registerDeviceSurf(const VarDecl *VD, llvm::GlobalVariable &Var,
 151                           bool Extern, int Type) {
 152     DeviceVars.push_back({&Var,
 153                           VD,
 154                           {DeviceVarFlags::Surface, Extern, /*Constant*/ false,
 155                            /*Managed*/ false,
 156                            /*Normalized*/ false, Type}});
 157   }
 158   void registerDeviceTex(const VarDecl *VD, llvm::GlobalVariable &Var,
 159                          bool Extern, int Type, bool Normalized) {
 160     DeviceVars.push_back({&Var,
 161                           VD,
 162                           {DeviceVarFlags::Texture, Extern, /*Constant*/ false,
 163                            /*Managed*/ false, Normalized, Type}});
 164   }
 165
 166   /// Creates module constructor function
 167   llvm::Function *makeModuleCtorFunction();
 168   /// Creates module destructor function
 169   llvm::Function *makeModuleDtorFunction();
 170   /// Transform managed variables for device compilation.
 171   void transformManagedVars();
 172   /// Create offloading entries to register globals in RDC mode.
 173   void createOffloadingEntries();
 174
 175 public:
 176   CGNVCUDARuntime(CodeGenModule &CGM);
 177
 178   llvm::GlobalValue *getKernelHandle(llvm::Function *F, GlobalDecl GD) override;
 179   llvm::Function *getKernelStub(llvm::GlobalValue *Handle) override {
 180     auto Loc = KernelStubs.find(Handle);
 181     assert(Loc != KernelStubs.end());
 182     return Loc->second;
 183   }
 184   void emitDeviceStub(CodeGenFunction &CGF, FunctionArgList &Args) override;
 185   void handleVarRegistration(const VarDecl *VD,
 186                              llvm::GlobalVariable &Var) override;
 187   void
 188   internalizeDeviceSideVar(const VarDecl *D,
 189                            llvm::GlobalValue::LinkageTypes &Linkage) override;
 190
 191   llvm::Function *finalizeModule() override;
 192 };
 193
 194 } // end anonymous namespace
 195
 196 std::string CGNVCUDARuntime::addPrefixToName(StringRef FuncName) const {
 197   if (CGM.getLangOpts().HIP)
 198     return ((Twine("hip") + Twine(FuncName)).str());
 199   return ((Twine("cuda") + Twine(FuncName)).str());
 200 }
 201 std::string
 202 CGNVCUDARuntime::addUnderscoredPrefixToName(StringRef FuncName) const {
 203   if (CGM.getLangOpts().HIP)
 204     return ((Twine("__hip") + Twine(FuncName)).str());
 205   return ((Twine("__cuda") + Twine(FuncName)).str());
 206 }
 207
 208 static std::unique_ptr<MangleContext> InitDeviceMC(CodeGenModule &CGM) {
 209   // If the host and device have different C++ ABIs, mark it as the device
 210   // mangle context so that the mangling needs to retrieve the additional
 211   // device lambda mangling number instead of the regular host one.
 212   if (CGM.getContext().getAuxTargetInfo() &&
 213       CGM.getContext().getTargetInfo().getCXXABI().isMicrosoft() &&
 214       CGM.getContext().getAuxTargetInfo()->getCXXABI().isItaniumFamily()) {
 215     return std::unique_ptr<MangleContext>(
 216         CGM.getContext().createDeviceMangleContext(
 217             *CGM.getContext().getAuxTargetInfo()));
 218   }
 219
 220   return std::unique_ptr<MangleContext>(CGM.getContext().createMangleContext(
 221       CGM.getContext().getAuxTargetInfo()));
 222 }
 223
 224 CGNVCUDARuntime::CGNVCUDARuntime(CodeGenModule &CGM)
 225     : CGCUDARuntime(CGM), Context(CGM.getLLVMContext()),
 226       TheModule(CGM.getModule()),
 227       RelocatableDeviceCode(CGM.getLangOpts().GPURelocatableDeviceCode),
 228       DeviceMC(InitDeviceMC(CGM)) {
 229   CodeGen::CodeGenTypes &Types = CGM.getTypes();
 230   ASTContext &Ctx = CGM.getContext();
 231
 232   IntTy = CGM.IntTy;
 233   SizeTy = CGM.SizeTy;
 234   VoidTy = CGM.VoidTy;
 235   Zeros[0] = llvm::ConstantInt::get(SizeTy, 0);
 236   Zeros[1] = Zeros[0];
 237
 238   CharPtrTy = llvm::PointerType::getUnqual(Types.ConvertType(Ctx.CharTy));
 239   VoidPtrTy = cast<llvm::PointerType>(Types.ConvertType(Ctx.VoidPtrTy));
 240   VoidPtrPtrTy = llvm::PointerType::getUnqual(CGM.getLLVMContext());
 241 }
 242
 243 llvm::FunctionCallee CGNVCUDARuntime::getSetupArgumentFn() const {
 244   // cudaError_t cudaSetupArgument(void *, size_t, size_t)
 245   llvm::Type *Params[] = {VoidPtrTy, SizeTy, SizeTy};
 246   return CGM.CreateRuntimeFunction(
 247       llvm::FunctionType::get(IntTy, Params, false),
 248       addPrefixToName("SetupArgument"));
 249 }
 250
 251 llvm::FunctionCallee CGNVCUDARuntime::getLaunchFn() const {
 252   if (CGM.getLangOpts().HIP) {
 253     // hipError_t hipLaunchByPtr(char *);
 254     return CGM.CreateRuntimeFunction(
 255         llvm::FunctionType::get(IntTy, CharPtrTy, false), "hipLaunchByPtr");
 256   }
 257   // cudaError_t cudaLaunch(char *);
 258   return CGM.CreateRuntimeFunction(
 259       llvm::FunctionType::get(IntTy, CharPtrTy, false), "cudaLaunch");
 260 }
 261
 262 llvm::FunctionType *CGNVCUDARuntime::getRegisterGlobalsFnTy() const {
 263   return llvm::FunctionType::get(VoidTy, VoidPtrPtrTy, false);
 264 }
 265
 266 llvm::FunctionType *CGNVCUDARuntime::getCallbackFnTy() const {
 267   return llvm::FunctionType::get(VoidTy, VoidPtrTy, false);
 268 }
 269
 270 llvm::FunctionType *CGNVCUDARuntime::getRegisterLinkedBinaryFnTy() const {
 271   llvm::Type *Params[] = {llvm::PointerType::getUnqual(Context), VoidPtrTy,
 272                           VoidPtrTy, llvm::PointerType::getUnqual(Context)};
 273   return llvm::FunctionType::get(VoidTy, Params, false);
 274 }
 275
 276 std::string CGNVCUDARuntime::getDeviceSideName(const NamedDecl *ND) {
 277   GlobalDecl GD;
 278   // D could be either a kernel or a variable.
 279   if (auto *FD = dyn_cast<FunctionDecl>(ND))
 280     GD = GlobalDecl(FD, KernelReferenceKind::Kernel);
 281   else
 282     GD = GlobalDecl(ND);
 283   std::string DeviceSideName;
 284   MangleContext *MC;
 285   if (CGM.getLangOpts().CUDAIsDevice)
 286     MC = &CGM.getCXXABI().getMangleContext();
 287   else
 288     MC = DeviceMC.get();
 289   if (MC->shouldMangleDeclName(ND)) {
 290     SmallString<256> Buffer;
 291     llvm::raw_svector_ostream Out(Buffer);
 292     MC->mangleName(GD, Out);
 293     DeviceSideName = std::string(Out.str());
 294   } else
 295     DeviceSideName = std::string(ND->getIdentifier()->getName());
 296
 297   // Make unique name for device side static file-scope variable for HIP.
 298   if (CGM.getContext().shouldExternalize(ND) &&
 299       CGM.getLangOpts().GPURelocatableDeviceCode) {
 300     SmallString<256> Buffer;
 301     llvm::raw_svector_ostream Out(Buffer);
 302     Out << DeviceSideName;
 303     CGM.printPostfixForExternalizedDecl(Out, ND);
 304     DeviceSideName = std::string(Out.str());
 305   }
 306   return DeviceSideName;
 307 }
 308
 309 void CGNVCUDARuntime::emitDeviceStub(CodeGenFunction &CGF,
 310                                      FunctionArgList &Args) {
 311   EmittedKernels.push_back({CGF.CurFn, CGF.CurFuncDecl});
 312   if (auto *GV =
 313           dyn_cast<llvm::GlobalVariable>(KernelHandles[CGF.CurFn->getName()])) {
 314     GV->setLinkage(CGF.CurFn->getLinkage());
 315     GV->setInitializer(CGF.CurFn);
 316   }
 317   if (CudaFeatureEnabled(CGM.getTarget().getSDKVersion(),
 318                          CudaFeature::CUDA_USES_NEW_LAUNCH) ||
 319       (CGF.getLangOpts().HIP && CGF.getLangOpts().HIPUseNewLaunchAPI))
 320     emitDeviceStubBodyNew(CGF, Args);
 321   else
 322     emitDeviceStubBodyLegacy(CGF, Args);
 323 }
 324
 325 // CUDA 9.0+ uses new way to launch kernels. Parameters are packed in a local
 326 // array and kernels are launched using cudaLaunchKernel().
 327 void CGNVCUDARuntime::emitDeviceStubBodyNew(CodeGenFunction &CGF,
 328                                             FunctionArgList &Args) {
 329   // Build the shadow stack entry at the very start of the function.
 330
 331   // Calculate amount of space we will need for all arguments.  If we have no
 332   // args, allocate a single pointer so we still have a valid pointer to the
 333   // argument array that we can pass to runtime, even if it will be unused.
 334   Address KernelArgs = CGF.CreateTempAlloca(
 335       VoidPtrTy, CharUnits::fromQuantity(16), "kernel_args",
 336       llvm::ConstantInt::get(SizeTy, std::max<size_t>(1, Args.size())));
 337   // Store pointers to the arguments in a locally allocated launch_args.
 338   for (unsigned i = 0; i < Args.size(); ++i) {
 339     llvm::Value* VarPtr = CGF.GetAddrOfLocalVar(Args[i]).getPointer();
 340     llvm::Value *VoidVarPtr = CGF.Builder.CreatePointerCast(VarPtr, VoidPtrTy);
 341     CGF.Builder.CreateDefaultAlignedStore(
 342         VoidVarPtr,
 343         CGF.Builder.CreateConstGEP1_32(VoidPtrTy, KernelArgs.getPointer(), i));
 344   }
 345
 346   llvm::BasicBlock *EndBlock = CGF.createBasicBlock("setup.end");
 347
 348   // Lookup cudaLaunchKernel/hipLaunchKernel function.
 349   // HIP kernel launching API name depends on -fgpu-default-stream option. For
 350   // the default value 'legacy', it is hipLaunchKernel. For 'per-thread',
 351   // it is hipLaunchKernel_spt.
 352   // cudaError_t cudaLaunchKernel(const void *func, dim3 gridDim, dim3 blockDim,
 353   //                              void **args, size_t sharedMem,
 354   //                              cudaStream_t stream);
 355   // hipError_t hipLaunchKernel[_spt](const void *func, dim3 gridDim,
 356   //                                  dim3 blockDim, void **args,
 357   //                                  size_t sharedMem, hipStream_t stream);
 358   TranslationUnitDecl *TUDecl = CGM.getContext().getTranslationUnitDecl();
 359   DeclContext *DC = TranslationUnitDecl::castToDeclContext(TUDecl);
 360   std::string KernelLaunchAPI = "LaunchKernel";
 361   if (CGF.getLangOpts().GPUDefaultStream ==
 362       LangOptions::GPUDefaultStreamKind::PerThread) {
 363     if (CGF.getLangOpts().HIP)
 364       KernelLaunchAPI = KernelLaunchAPI + "_spt";
 365     else if (CGF.getLangOpts().CUDA)
 366       KernelLaunchAPI = KernelLaunchAPI + "_ptsz";
 367   }
 368   auto LaunchKernelName = addPrefixToName(KernelLaunchAPI);
 369   IdentifierInfo &cudaLaunchKernelII =
 370       CGM.getContext().Idents.get(LaunchKernelName);
 371   FunctionDecl *cudaLaunchKernelFD = nullptr;
 372   for (auto *Result : DC->lookup(&cudaLaunchKernelII)) {
 373     if (FunctionDecl *FD = dyn_cast<FunctionDecl>(Result))
 374       cudaLaunchKernelFD = FD;
 375   }
 376
 377   if (cudaLaunchKernelFD == nullptr) {
 378     CGM.Error(CGF.CurFuncDecl->getLocation(),
 379               "Can't find declaration for " + LaunchKernelName);
 380     return;
 381   }
 382   // Create temporary dim3 grid_dim, block_dim.
 383   ParmVarDecl *GridDimParam = cudaLaunchKernelFD->getParamDecl(1);
 384   QualType Dim3Ty = GridDimParam->getType();
 385   Address GridDim =
 386       CGF.CreateMemTemp(Dim3Ty, CharUnits::fromQuantity(8), "grid_dim");
 387   Address BlockDim =
 388       CGF.CreateMemTemp(Dim3Ty, CharUnits::fromQuantity(8), "block_dim");
 389   Address ShmemSize =
 390       CGF.CreateTempAlloca(SizeTy, CGM.getSizeAlign(), "shmem_size");
 391   Address Stream =
 392       CGF.CreateTempAlloca(VoidPtrTy, CGM.getPointerAlign(), "stream");
 393   llvm::FunctionCallee cudaPopConfigFn = CGM.CreateRuntimeFunction(
 394       llvm::FunctionType::get(IntTy,
 395                               {/*gridDim=*/GridDim.getType(),
 396                                /*blockDim=*/BlockDim.getType(),
 397                                /*ShmemSize=*/ShmemSize.getType(),
 398                                /*Stream=*/Stream.getType()},
 399                               /*isVarArg=*/false),
 400       addUnderscoredPrefixToName("PopCallConfiguration"));
 401
 402   CGF.EmitRuntimeCallOrInvoke(cudaPopConfigFn,
 403                               {GridDim.getPointer(), BlockDim.getPointer(),
 404                                ShmemSize.getPointer(), Stream.getPointer()});
 405
 406   // Emit the call to cudaLaunch
 407   llvm::Value *Kernel = CGF.Builder.CreatePointerCast(
 408       KernelHandles[CGF.CurFn->getName()], VoidPtrTy);
 409   CallArgList LaunchKernelArgs;
 410   LaunchKernelArgs.add(RValue::get(Kernel),
 411                        cudaLaunchKernelFD->getParamDecl(0)->getType());
 412   LaunchKernelArgs.add(RValue::getAggregate(GridDim), Dim3Ty);
 413   LaunchKernelArgs.add(RValue::getAggregate(BlockDim), Dim3Ty);
 414   LaunchKernelArgs.add(RValue::get(KernelArgs.getPointer()),
 415                        cudaLaunchKernelFD->getParamDecl(3)->getType());
 416   LaunchKernelArgs.add(RValue::get(CGF.Builder.CreateLoad(ShmemSize)),
 417                        cudaLaunchKernelFD->getParamDecl(4)->getType());
 418   LaunchKernelArgs.add(RValue::get(CGF.Builder.CreateLoad(Stream)),
 419                        cudaLaunchKernelFD->getParamDecl(5)->getType());
 420
 421   QualType QT = cudaLaunchKernelFD->getType();
 422   QualType CQT = QT.getCanonicalType();
 423   llvm::Type *Ty = CGM.getTypes().ConvertType(CQT);
 424   llvm::FunctionType *FTy = cast<llvm::FunctionType>(Ty);
 425
 426   const CGFunctionInfo &FI =
 427       CGM.getTypes().arrangeFunctionDeclaration(cudaLaunchKernelFD);
 428   llvm::FunctionCallee cudaLaunchKernelFn =
 429       CGM.CreateRuntimeFunction(FTy, LaunchKernelName);
 430   CGF.EmitCall(FI, CGCallee::forDirect(cudaLaunchKernelFn), ReturnValueSlot(),
 431                LaunchKernelArgs);
 432   CGF.EmitBranch(EndBlock);
 433
 434   CGF.EmitBlock(EndBlock);
 435 }
 436
 437 void CGNVCUDARuntime::emitDeviceStubBodyLegacy(CodeGenFunction &CGF,
 438                                                FunctionArgList &Args) {
 439   // Emit a call to cudaSetupArgument for each arg in Args.
 440   llvm::FunctionCallee cudaSetupArgFn = getSetupArgumentFn();
 441   llvm::BasicBlock *EndBlock = CGF.createBasicBlock("setup.end");
 442   CharUnits Offset = CharUnits::Zero();
 443   for (const VarDecl *A : Args) {
 444     auto TInfo = CGM.getContext().getTypeInfoInChars(A->getType());
 445     Offset = Offset.alignTo(TInfo.Align);
 446     llvm::Value *Args[] = {
 447         CGF.Builder.CreatePointerCast(CGF.GetAddrOfLocalVar(A).getPointer(),
 448                                       VoidPtrTy),
 449         llvm::ConstantInt::get(SizeTy, TInfo.Width.getQuantity()),
 450         llvm::ConstantInt::get(SizeTy, Offset.getQuantity()),
 451     };
 452     llvm::CallBase *CB = CGF.EmitRuntimeCallOrInvoke(cudaSetupArgFn, Args);
 453     llvm::Constant *Zero = llvm::ConstantInt::get(IntTy, 0);
 454     llvm::Value *CBZero = CGF.Builder.CreateICmpEQ(CB, Zero);
 455     llvm::BasicBlock *NextBlock = CGF.createBasicBlock("setup.next");
 456     CGF.Builder.CreateCondBr(CBZero, NextBlock, EndBlock);
 457     CGF.EmitBlock(NextBlock);
 458     Offset += TInfo.Width;
 459   }
 460
 461   // Emit the call to cudaLaunch
 462   llvm::FunctionCallee cudaLaunchFn = getLaunchFn();
 463   llvm::Value *Arg = CGF.Builder.CreatePointerCast(
 464       KernelHandles[CGF.CurFn->getName()], CharPtrTy);
 465   CGF.EmitRuntimeCallOrInvoke(cudaLaunchFn, Arg);
 466   CGF.EmitBranch(EndBlock);
 467
 468   CGF.EmitBlock(EndBlock);
 469 }
 470
 471 // Replace the original variable Var with the address loaded from variable
 472 // ManagedVar populated by HIP runtime.
 473 static void replaceManagedVar(llvm::GlobalVariable *Var,
 474                               llvm::GlobalVariable *ManagedVar) {
 475   SmallVector<SmallVector<llvm::User *, 8>, 8> WorkList;
 476   for (auto &&VarUse : Var->uses()) {
 477     WorkList.push_back({VarUse.getUser()});
 478   }
 479   while (!WorkList.empty()) {
 480     auto &&WorkItem = WorkList.pop_back_val();
 481     auto *U = WorkItem.back();
 482     if (isa<llvm::ConstantExpr>(U)) {
 483       for (auto &&UU : U->uses()) {
 484         WorkItem.push_back(UU.getUser());
 485         WorkList.push_back(WorkItem);
 486         WorkItem.pop_back();
 487       }
 488       continue;
 489     }
 490     if (auto *I = dyn_cast<llvm::Instruction>(U)) {
 491       llvm::Value *OldV = Var;
 492       llvm::Instruction *NewV =
 493           new llvm::LoadInst(Var->getType(), ManagedVar, "ld.managed", false,
 494                              llvm::Align(Var->getAlignment()), I);
 495       WorkItem.pop_back();
 496       // Replace constant expressions directly or indirectly using the managed
 497       // variable with instructions.
 498       for (auto &&Op : WorkItem) {
 499         auto *CE = cast<llvm::ConstantExpr>(Op);
 500         auto *NewInst = CE->getAsInstruction(I);
 501         NewInst->replaceUsesOfWith(OldV, NewV);
 502         OldV = CE;
 503         NewV = NewInst;
 504       }
 505       I->replaceUsesOfWith(OldV, NewV);
 506     } else {
 507       llvm_unreachable("Invalid use of managed variable");
 508     }
 509   }
 510 }
 511
 512 /// Creates a function that sets up state on the host side for CUDA objects that
 513 /// have a presence on both the host and device sides. Specifically, registers
 514 /// the host side of kernel functions and device global variables with the CUDA
 515 /// runtime.
 516 /// \code
 517 /// void __cuda_register_globals(void** GpuBinaryHandle) {
 518 ///    __cudaRegisterFunction(GpuBinaryHandle,Kernel0,...);
 519 ///    ...
 520 ///    __cudaRegisterFunction(GpuBinaryHandle,KernelM,...);
 521 ///    __cudaRegisterVar(GpuBinaryHandle, GlobalVar0, ...);
 522 ///    ...
 523 ///    __cudaRegisterVar(GpuBinaryHandle, GlobalVarN, ...);
 524 /// }
 525 /// \endcode
 526 llvm::Function *CGNVCUDARuntime::makeRegisterGlobalsFn() {
 527   // No need to register anything
 528   if (EmittedKernels.empty() && DeviceVars.empty())
 529     return nullptr;
 530
 531   llvm::Function *RegisterKernelsFunc = llvm::Function::Create(
 532       getRegisterGlobalsFnTy(), llvm::GlobalValue::InternalLinkage,
 533       addUnderscoredPrefixToName("_register_globals"), &TheModule);
 534   llvm::BasicBlock *EntryBB =
 535       llvm::BasicBlock::Create(Context, "entry", RegisterKernelsFunc);
 536   CGBuilderTy Builder(CGM, Context);
 537   Builder.SetInsertPoint(EntryBB);
 538
 539   // void __cudaRegisterFunction(void **, const char *, char *, const char *,
 540   //                             int, uint3*, uint3*, dim3*, dim3*, int*)
 541   llvm::Type *RegisterFuncParams[] = {
 542       VoidPtrPtrTy, CharPtrTy,
 543       CharPtrTy,    CharPtrTy,
 544       IntTy,        VoidPtrTy,
 545       VoidPtrTy,    VoidPtrTy,
 546       VoidPtrTy,    llvm::PointerType::getUnqual(Context)};
 547   llvm::FunctionCallee RegisterFunc = CGM.CreateRuntimeFunction(
 548       llvm::FunctionType::get(IntTy, RegisterFuncParams, false),
 549       addUnderscoredPrefixToName("RegisterFunction"));
 550
 551   // Extract GpuBinaryHandle passed as the first argument passed to
 552   // __cuda_register_globals() and generate __cudaRegisterFunction() call for
 553   // each emitted kernel.
 554   llvm::Argument &GpuBinaryHandlePtr = *RegisterKernelsFunc->arg_begin();
 555   for (auto &&I : EmittedKernels) {
 556     llvm::Constant *KernelName =
 557         makeConstantString(getDeviceSideName(cast<NamedDecl>(I.D)));
 558     llvm::Constant *NullPtr = llvm::ConstantPointerNull::get(VoidPtrTy);
 559     llvm::Value *Args[] = {
 560         &GpuBinaryHandlePtr,
 561         Builder.CreateBitCast(KernelHandles[I.Kernel->getName()], VoidPtrTy),
 562         KernelName,
 563         KernelName,
 564         llvm::ConstantInt::get(IntTy, -1),
 565         NullPtr,
 566         NullPtr,
 567         NullPtr,
 568         NullPtr,
 569         llvm::ConstantPointerNull::get(llvm::PointerType::getUnqual(Context))};
 570     Builder.CreateCall(RegisterFunc, Args);
 571   }
 572
 573   llvm::Type *VarSizeTy = IntTy;
 574   // For HIP or CUDA 9.0+, device variable size is type of `size_t`.
 575   if (CGM.getLangOpts().HIP ||
 576       ToCudaVersion(CGM.getTarget().getSDKVersion()) >= CudaVersion::CUDA_90)
 577     VarSizeTy = SizeTy;
 578
 579   // void __cudaRegisterVar(void **, char *, char *, const char *,
 580   //                        int, int, int, int)
 581   llvm::Type *RegisterVarParams[] = {VoidPtrPtrTy, CharPtrTy, CharPtrTy,
 582                                      CharPtrTy,    IntTy,     VarSizeTy,
 583                                      IntTy,        IntTy};
 584   llvm::FunctionCallee RegisterVar = CGM.CreateRuntimeFunction(
 585       llvm::FunctionType::get(VoidTy, RegisterVarParams, false),
 586       addUnderscoredPrefixToName("RegisterVar"));
 587   // void __hipRegisterManagedVar(void **, char *, char *, const char *,
 588   //                              size_t, unsigned)
 589   llvm::Type *RegisterManagedVarParams[] = {VoidPtrPtrTy, CharPtrTy, CharPtrTy,
 590                                             CharPtrTy,    VarSizeTy, IntTy};
 591   llvm::FunctionCallee RegisterManagedVar = CGM.CreateRuntimeFunction(
 592       llvm::FunctionType::get(VoidTy, RegisterManagedVarParams, false),
 593       addUnderscoredPrefixToName("RegisterManagedVar"));
 594   // void __cudaRegisterSurface(void **, const struct surfaceReference *,
 595   //                            const void **, const char *, int, int);
 596   llvm::FunctionCallee RegisterSurf = CGM.CreateRuntimeFunction(
 597       llvm::FunctionType::get(
 598           VoidTy, {VoidPtrPtrTy, VoidPtrTy, CharPtrTy, CharPtrTy, IntTy, IntTy},
 599           false),
 600       addUnderscoredPrefixToName("RegisterSurface"));
 601   // void __cudaRegisterTexture(void **, const struct textureReference *,
 602   //                            const void **, const char *, int, int, int)
 603   llvm::FunctionCallee RegisterTex = CGM.CreateRuntimeFunction(
 604       llvm::FunctionType::get(
 605           VoidTy,
 606           {VoidPtrPtrTy, VoidPtrTy, CharPtrTy, CharPtrTy, IntTy, IntTy, IntTy},
 607           false),
 608       addUnderscoredPrefixToName("RegisterTexture"));
 609   for (auto &&Info : DeviceVars) {
 610     llvm::GlobalVariable *Var = Info.Var;
 611     assert((!Var->isDeclaration() || Info.Flags.isManaged()) &&
 612            "External variables should not show up here, except HIP managed "
 613            "variables");
 614     llvm::Constant *VarName = makeConstantString(getDeviceSideName(Info.D));
 615     switch (Info.Flags.getKind()) {
 616     case DeviceVarFlags::Variable: {
 617       uint64_t VarSize =
 618           CGM.getDataLayout().getTypeAllocSize(Var->getValueType());
 619       if (Info.Flags.isManaged()) {
 620         auto *ManagedVar = new llvm::GlobalVariable(
 621             CGM.getModule(), Var->getType(),
 622             /*isConstant=*/false, Var->getLinkage(),
 623             /*Init=*/Var->isDeclaration()
 624                 ? nullptr
 625                 : llvm::ConstantPointerNull::get(Var->getType()),
 626             /*Name=*/"", /*InsertBefore=*/nullptr,
 627             llvm::GlobalVariable::NotThreadLocal);
 628         ManagedVar->setDSOLocal(Var->isDSOLocal());
 629         ManagedVar->setVisibility(Var->getVisibility());
 630         ManagedVar->setExternallyInitialized(true);
 631         ManagedVar->takeName(Var);
 632         Var->setName(Twine(ManagedVar->getName() + ".managed"));
 633         replaceManagedVar(Var, ManagedVar);
 634         llvm::Value *Args[] = {
 635             &GpuBinaryHandlePtr,
 636             Builder.CreateBitCast(ManagedVar, VoidPtrTy),
 637             Builder.CreateBitCast(Var, VoidPtrTy),
 638             VarName,
 639             llvm::ConstantInt::get(VarSizeTy, VarSize),
 640             llvm::ConstantInt::get(IntTy, Var->getAlignment())};
 641         if (!Var->isDeclaration())
 642           Builder.CreateCall(RegisterManagedVar, Args);
 643       } else {
 644         llvm::Value *Args[] = {
 645             &GpuBinaryHandlePtr,
 646             Builder.CreateBitCast(Var, VoidPtrTy),
 647             VarName,
 648             VarName,
 649             llvm::ConstantInt::get(IntTy, Info.Flags.isExtern()),
 650             llvm::ConstantInt::get(VarSizeTy, VarSize),
 651             llvm::ConstantInt::get(IntTy, Info.Flags.isConstant()),
 652             llvm::ConstantInt::get(IntTy, 0)};
 653         Builder.CreateCall(RegisterVar, Args);
 654       }
 655       break;
 656     }
 657     case DeviceVarFlags::Surface:
 658       Builder.CreateCall(
 659           RegisterSurf,
 660           {&GpuBinaryHandlePtr, Builder.CreateBitCast(Var, VoidPtrTy), VarName,
 661            VarName, llvm::ConstantInt::get(IntTy, Info.Flags.getSurfTexType()),
 662            llvm::ConstantInt::get(IntTy, Info.Flags.isExtern())});
 663       break;
 664     case DeviceVarFlags::Texture:
 665       Builder.CreateCall(
 666           RegisterTex,
 667           {&GpuBinaryHandlePtr, Builder.CreateBitCast(Var, VoidPtrTy), VarName,
 668            VarName, llvm::ConstantInt::get(IntTy, Info.Flags.getSurfTexType()),
 669            llvm::ConstantInt::get(IntTy, Info.Flags.isNormalized()),
 670            llvm::ConstantInt::get(IntTy, Info.Flags.isExtern())});
 671       break;
 672     }
 673   }
 674
 675   Builder.CreateRetVoid();
 676   return RegisterKernelsFunc;
 677 }
 678
 679 /// Creates a global constructor function for the module:
 680 ///
 681 /// For CUDA:
 682 /// \code
 683 /// void __cuda_module_ctor() {
 684 ///     Handle = __cudaRegisterFatBinary(GpuBinaryBlob);
 685 ///     __cuda_register_globals(Handle);
 686 /// }
 687 /// \endcode
 688 ///
 689 /// For HIP:
 690 /// \code
 691 /// void __hip_module_ctor() {
 692 ///     if (__hip_gpubin_handle == 0) {
 693 ///         __hip_gpubin_handle  = __hipRegisterFatBinary(GpuBinaryBlob);
 694 ///         __hip_register_globals(__hip_gpubin_handle);
 695 ///     }
 696 /// }
 697 /// \endcode
 698 llvm::Function *CGNVCUDARuntime::makeModuleCtorFunction() {
 699   bool IsHIP = CGM.getLangOpts().HIP;
 700   bool IsCUDA = CGM.getLangOpts().CUDA;
 701   // No need to generate ctors/dtors if there is no GPU binary.
 702   StringRef CudaGpuBinaryFileName = CGM.getCodeGenOpts().CudaGpuBinaryFileName;
 703   if (CudaGpuBinaryFileName.empty() && !IsHIP)
 704     return nullptr;
 705   if ((IsHIP || (IsCUDA && !RelocatableDeviceCode)) && EmittedKernels.empty() &&
 706       DeviceVars.empty())
 707     return nullptr;
 708
 709   // void __{cuda|hip}_register_globals(void* handle);
 710   llvm::Function *RegisterGlobalsFunc = makeRegisterGlobalsFn();
 711   // We always need a function to pass in as callback. Create a dummy
 712   // implementation if we don't need to register anything.
 713   if (RelocatableDeviceCode && !RegisterGlobalsFunc)
 714     RegisterGlobalsFunc = makeDummyFunction(getRegisterGlobalsFnTy());
 715
 716   // void ** __{cuda|hip}RegisterFatBinary(void *);
 717   llvm::FunctionCallee RegisterFatbinFunc = CGM.CreateRuntimeFunction(
 718       llvm::FunctionType::get(VoidPtrPtrTy, VoidPtrTy, false),
 719       addUnderscoredPrefixToName("RegisterFatBinary"));
 720   // struct { int magic, int version, void * gpu_binary, void * dont_care };
 721   llvm::StructType *FatbinWrapperTy =
 722       llvm::StructType::get(IntTy, IntTy, VoidPtrTy, VoidPtrTy);
 723
 724   // Register GPU binary with the CUDA runtime, store returned handle in a
 725   // global variable and save a reference in GpuBinaryHandle to be cleaned up
 726   // in destructor on exit. Then associate all known kernels with the GPU binary
 727   // handle so CUDA runtime can figure out what to call on the GPU side.
 728   std::unique_ptr<llvm::MemoryBuffer> CudaGpuBinary = nullptr;
 729   if (!CudaGpuBinaryFileName.empty()) {
 730     auto VFS = CGM.getFileSystem();
 731     auto CudaGpuBinaryOrErr =
 732         VFS->getBufferForFile(CudaGpuBinaryFileName, -1, false);
 733     if (std::error_code EC = CudaGpuBinaryOrErr.getError()) {
 734       CGM.getDiags().Report(diag::err_cannot_open_file)
 735           << CudaGpuBinaryFileName << EC.message();
 736       return nullptr;
 737     }
 738     CudaGpuBinary = std::move(CudaGpuBinaryOrErr.get());
 739   }
 740
 741   llvm::Function *ModuleCtorFunc = llvm::Function::Create(
 742       llvm::FunctionType::get(VoidTy, false),
 743       llvm::GlobalValue::InternalLinkage,
 744       addUnderscoredPrefixToName("_module_ctor"), &TheModule);
 745   llvm::BasicBlock *CtorEntryBB =
 746       llvm::BasicBlock::Create(Context, "entry", ModuleCtorFunc);
 747   CGBuilderTy CtorBuilder(CGM, Context);
 748
 749   CtorBuilder.SetInsertPoint(CtorEntryBB);
 750
 751   const char *FatbinConstantName;
 752   const char *FatbinSectionName;
 753   const char *ModuleIDSectionName;
 754   StringRef ModuleIDPrefix;
 755   llvm::Constant *FatBinStr;
 756   unsigned FatMagic;
 757   if (IsHIP) {
 758     FatbinConstantName = ".hip_fatbin";
 759     FatbinSectionName = ".hipFatBinSegment";
 760
 761     ModuleIDSectionName = "__hip_module_id";
 762     ModuleIDPrefix = "__hip_";
 763
 764     if (CudaGpuBinary) {
 765       // If fatbin is available from early finalization, create a string
 766       // literal containing the fat binary loaded from the given file.
 767       const unsigned HIPCodeObjectAlign = 4096;
 768       FatBinStr = makeConstantArray(std::string(CudaGpuBinary->getBuffer()), "",
 769                                     FatbinConstantName, HIPCodeObjectAlign);
 770     } else {
 771       // If fatbin is not available, create an external symbol
 772       // __hip_fatbin in section .hip_fatbin. The external symbol is supposed
 773       // to contain the fat binary but will be populated somewhere else,
 774       // e.g. by lld through link script.
 775       FatBinStr = new llvm::GlobalVariable(
 776         CGM.getModule(), CGM.Int8Ty,
 777         /*isConstant=*/true, llvm::GlobalValue::ExternalLinkage, nullptr,
 778         "__hip_fatbin", nullptr,
 779         llvm::GlobalVariable::NotThreadLocal);
 780       cast<llvm::GlobalVariable>(FatBinStr)->setSection(FatbinConstantName);
 781     }
 782
 783     FatMagic = HIPFatMagic;
 784   } else {
 785     if (RelocatableDeviceCode)
 786       FatbinConstantName = CGM.getTriple().isMacOSX()
 787                                ? "__NV_CUDA,__nv_relfatbin"
 788                                : "__nv_relfatbin";
 789     else
 790       FatbinConstantName =
 791           CGM.getTriple().isMacOSX() ? "__NV_CUDA,__nv_fatbin" : ".nv_fatbin";
 792     // NVIDIA's cuobjdump looks for fatbins in this section.
 793     FatbinSectionName =
 794         CGM.getTriple().isMacOSX() ? "__NV_CUDA,__fatbin" : ".nvFatBinSegment";
 795
 796     ModuleIDSectionName = CGM.getTriple().isMacOSX()
 797                               ? "__NV_CUDA,__nv_module_id"
 798                               : "__nv_module_id";
 799     ModuleIDPrefix = "__nv_";
 800
 801     // For CUDA, create a string literal containing the fat binary loaded from
 802     // the given file.
 803     FatBinStr = makeConstantArray(std::string(CudaGpuBinary->getBuffer()), "",
 804                                   FatbinConstantName, 8);
 805     FatMagic = CudaFatMagic;
 806   }
 807
 808   // Create initialized wrapper structure that points to the loaded GPU binary
 809   ConstantInitBuilder Builder(CGM);
 810   auto Values = Builder.beginStruct(FatbinWrapperTy);
 811   // Fatbin wrapper magic.
 812   Values.addInt(IntTy, FatMagic);
 813   // Fatbin version.
 814   Values.addInt(IntTy, 1);
 815   // Data.
 816   Values.add(FatBinStr);
 817   // Unused in fatbin v1.
 818   Values.add(llvm::ConstantPointerNull::get(VoidPtrTy));
 819   llvm::GlobalVariable *FatbinWrapper = Values.finishAndCreateGlobal(
 820       addUnderscoredPrefixToName("_fatbin_wrapper"), CGM.getPointerAlign(),
 821       /*constant*/ true);
 822   FatbinWrapper->setSection(FatbinSectionName);
 823
 824   // There is only one HIP fat binary per linked module, however there are
 825   // multiple constructor functions. Make sure the fat binary is registered
 826   // only once. The constructor functions are executed by the dynamic loader
 827   // before the program gains control. The dynamic loader cannot execute the
 828   // constructor functions concurrently since doing that would not guarantee
 829   // thread safety of the loaded program. Therefore we can assume sequential
 830   // execution of constructor functions here.
 831   if (IsHIP) {
 832     auto Linkage = CudaGpuBinary ? llvm::GlobalValue::InternalLinkage :
 833         llvm::GlobalValue::LinkOnceAnyLinkage;
 834     llvm::BasicBlock *IfBlock =
 835         llvm::BasicBlock::Create(Context, "if", ModuleCtorFunc);
 836     llvm::BasicBlock *ExitBlock =
 837         llvm::BasicBlock::Create(Context, "exit", ModuleCtorFunc);
 838     // The name, size, and initialization pattern of this variable is part
 839     // of HIP ABI.
 840     GpuBinaryHandle = new llvm::GlobalVariable(
 841         TheModule, VoidPtrPtrTy, /*isConstant=*/false,
 842         Linkage,
 843         /*Initializer=*/llvm::ConstantPointerNull::get(VoidPtrPtrTy),
 844         "__hip_gpubin_handle");
 845     if (Linkage == llvm::GlobalValue::LinkOnceAnyLinkage)
 846       GpuBinaryHandle->setComdat(
 847           CGM.getModule().getOrInsertComdat(GpuBinaryHandle->getName()));
 848     GpuBinaryHandle->setAlignment(CGM.getPointerAlign().getAsAlign());
 849     // Prevent the weak symbol in different shared libraries being merged.
 850     if (Linkage != llvm::GlobalValue::InternalLinkage)
 851       GpuBinaryHandle->setVisibility(llvm::GlobalValue::HiddenVisibility);
 852     Address GpuBinaryAddr(
 853         GpuBinaryHandle, VoidPtrPtrTy,
 854         CharUnits::fromQuantity(GpuBinaryHandle->getAlignment()));
 855     {
 856       auto *HandleValue = CtorBuilder.CreateLoad(GpuBinaryAddr);
 857       llvm::Constant *Zero =
 858           llvm::Constant::getNullValue(HandleValue->getType());
 859       llvm::Value *EQZero = CtorBuilder.CreateICmpEQ(HandleValue, Zero);
 860       CtorBuilder.CreateCondBr(EQZero, IfBlock, ExitBlock);
 861     }
 862     {
 863       CtorBuilder.SetInsertPoint(IfBlock);
 864       // GpuBinaryHandle = __hipRegisterFatBinary(&FatbinWrapper);
 865       llvm::CallInst *RegisterFatbinCall = CtorBuilder.CreateCall(
 866           RegisterFatbinFunc,
 867           CtorBuilder.CreateBitCast(FatbinWrapper, VoidPtrTy));
 868       CtorBuilder.CreateStore(RegisterFatbinCall, GpuBinaryAddr);
 869       CtorBuilder.CreateBr(ExitBlock);
 870     }
 871     {
 872       CtorBuilder.SetInsertPoint(ExitBlock);
 873       // Call __hip_register_globals(GpuBinaryHandle);
 874       if (RegisterGlobalsFunc) {
 875         auto *HandleValue = CtorBuilder.CreateLoad(GpuBinaryAddr);
 876         CtorBuilder.CreateCall(RegisterGlobalsFunc, HandleValue);
 877       }
 878     }
 879   } else if (!RelocatableDeviceCode) {
 880     // Register binary with CUDA runtime. This is substantially different in
 881     // default mode vs. separate compilation!
 882     // GpuBinaryHandle = __cudaRegisterFatBinary(&FatbinWrapper);
 883     llvm::CallInst *RegisterFatbinCall = CtorBuilder.CreateCall(
 884         RegisterFatbinFunc,
 885         CtorBuilder.CreateBitCast(FatbinWrapper, VoidPtrTy));
 886     GpuBinaryHandle = new llvm::GlobalVariable(
 887         TheModule, VoidPtrPtrTy, false, llvm::GlobalValue::InternalLinkage,
 888         llvm::ConstantPointerNull::get(VoidPtrPtrTy), "__cuda_gpubin_handle");
 889     GpuBinaryHandle->setAlignment(CGM.getPointerAlign().getAsAlign());
 890     CtorBuilder.CreateAlignedStore(RegisterFatbinCall, GpuBinaryHandle,
 891                                    CGM.getPointerAlign());
 892
 893     // Call __cuda_register_globals(GpuBinaryHandle);
 894     if (RegisterGlobalsFunc)
 895       CtorBuilder.CreateCall(RegisterGlobalsFunc, RegisterFatbinCall);
 896
 897     // Call __cudaRegisterFatBinaryEnd(Handle) if this CUDA version needs it.
 898     if (CudaFeatureEnabled(CGM.getTarget().getSDKVersion(),
 899                            CudaFeature::CUDA_USES_FATBIN_REGISTER_END)) {
 900       // void __cudaRegisterFatBinaryEnd(void **);
 901       llvm::FunctionCallee RegisterFatbinEndFunc = CGM.CreateRuntimeFunction(
 902           llvm::FunctionType::get(VoidTy, VoidPtrPtrTy, false),
 903           "__cudaRegisterFatBinaryEnd");
 904       CtorBuilder.CreateCall(RegisterFatbinEndFunc, RegisterFatbinCall);
 905     }
 906   } else {
 907     // Generate a unique module ID.
 908     SmallString<64> ModuleID;
 909     llvm::raw_svector_ostream OS(ModuleID);
 910     OS << ModuleIDPrefix << llvm::format("%" PRIx64, FatbinWrapper->getGUID());
 911     llvm::Constant *ModuleIDConstant = makeConstantArray(
 912         std::string(ModuleID.str()), "", ModuleIDSectionName, 32, /*AddNull=*/true);
 913
 914     // Create an alias for the FatbinWrapper that nvcc will look for.
 915     llvm::GlobalAlias::create(llvm::GlobalValue::ExternalLinkage,
 916                               Twine("__fatbinwrap") + ModuleID, FatbinWrapper);
 917
 918     // void __cudaRegisterLinkedBinary%ModuleID%(void (*)(void *), void *,
 919     // void *, void (*)(void **))
 920     SmallString<128> RegisterLinkedBinaryName("__cudaRegisterLinkedBinary");
 921     RegisterLinkedBinaryName += ModuleID;
 922     llvm::FunctionCallee RegisterLinkedBinaryFunc = CGM.CreateRuntimeFunction(
 923         getRegisterLinkedBinaryFnTy(), RegisterLinkedBinaryName);
 924
 925     assert(RegisterGlobalsFunc && "Expecting at least dummy function!");
 926     llvm::Value *Args[] = {RegisterGlobalsFunc,
 927                            CtorBuilder.CreateBitCast(FatbinWrapper, VoidPtrTy),
 928                            ModuleIDConstant,
 929                            makeDummyFunction(getCallbackFnTy())};
 930     CtorBuilder.CreateCall(RegisterLinkedBinaryFunc, Args);
 931   }
 932
 933   // Create destructor and register it with atexit() the way NVCC does it. Doing
 934   // it during regular destructor phase worked in CUDA before 9.2 but results in
 935   // double-free in 9.2.
 936   if (llvm::Function *CleanupFn = makeModuleDtorFunction()) {
 937     // extern "C" int atexit(void (*f)(void));
 938     llvm::FunctionType *AtExitTy =
 939         llvm::FunctionType::get(IntTy, CleanupFn->getType(), false);
 940     llvm::FunctionCallee AtExitFunc =
 941         CGM.CreateRuntimeFunction(AtExitTy, "atexit", llvm::AttributeList(),
 942                                   /*Local=*/true);
 943     CtorBuilder.CreateCall(AtExitFunc, CleanupFn);
 944   }
 945
 946   CtorBuilder.CreateRetVoid();
 947   return ModuleCtorFunc;
 948 }
 949
 950 /// Creates a global destructor function that unregisters the GPU code blob
 951 /// registered by constructor.
 952 ///
 953 /// For CUDA:
 954 /// \code
 955 /// void __cuda_module_dtor() {
 956 ///     __cudaUnregisterFatBinary(Handle);
 957 /// }
 958 /// \endcode
 959 ///
 960 /// For HIP:
 961 /// \code
 962 /// void __hip_module_dtor() {
 963 ///     if (__hip_gpubin_handle) {
 964 ///         __hipUnregisterFatBinary(__hip_gpubin_handle);
 965 ///         __hip_gpubin_handle = 0;
 966 ///     }
 967 /// }
 968 /// \endcode
 969 llvm::Function *CGNVCUDARuntime::makeModuleDtorFunction() {
 970   // No need for destructor if we don't have a handle to unregister.
 971   if (!GpuBinaryHandle)
 972     return nullptr;
 973
 974   // void __cudaUnregisterFatBinary(void ** handle);
 975   llvm::FunctionCallee UnregisterFatbinFunc = CGM.CreateRuntimeFunction(
 976       llvm::FunctionType::get(VoidTy, VoidPtrPtrTy, false),
 977       addUnderscoredPrefixToName("UnregisterFatBinary"));
 978
 979   llvm::Function *ModuleDtorFunc = llvm::Function::Create(
 980       llvm::FunctionType::get(VoidTy, false),
 981       llvm::GlobalValue::InternalLinkage,
 982       addUnderscoredPrefixToName("_module_dtor"), &TheModule);
 983
 984   llvm::BasicBlock *DtorEntryBB =
 985       llvm::BasicBlock::Create(Context, "entry", ModuleDtorFunc);
 986   CGBuilderTy DtorBuilder(CGM, Context);
 987   DtorBuilder.SetInsertPoint(DtorEntryBB);
 988
 989   Address GpuBinaryAddr(
 990       GpuBinaryHandle, GpuBinaryHandle->getValueType(),
 991       CharUnits::fromQuantity(GpuBinaryHandle->getAlignment()));
 992   auto *HandleValue = DtorBuilder.CreateLoad(GpuBinaryAddr);
 993   // There is only one HIP fat binary per linked module, however there are
 994   // multiple destructor functions. Make sure the fat binary is unregistered
 995   // only once.
 996   if (CGM.getLangOpts().HIP) {
 997     llvm::BasicBlock *IfBlock =
 998         llvm::BasicBlock::Create(Context, "if", ModuleDtorFunc);
 999     llvm::BasicBlock *ExitBlock =
1000         llvm::BasicBlock::Create(Context, "exit", ModuleDtorFunc);
1001     llvm::Constant *Zero = llvm::Constant::getNullValue(HandleValue->getType());
1002     llvm::Value *NEZero = DtorBuilder.CreateICmpNE(HandleValue, Zero);
1003     DtorBuilder.CreateCondBr(NEZero, IfBlock, ExitBlock);
1004
1005     DtorBuilder.SetInsertPoint(IfBlock);
1006     DtorBuilder.CreateCall(UnregisterFatbinFunc, HandleValue);
1007     DtorBuilder.CreateStore(Zero, GpuBinaryAddr);
1008     DtorBuilder.CreateBr(ExitBlock);
1009
1010     DtorBuilder.SetInsertPoint(ExitBlock);
1011   } else {
1012     DtorBuilder.CreateCall(UnregisterFatbinFunc, HandleValue);
1013   }
1014   DtorBuilder.CreateRetVoid();
1015   return ModuleDtorFunc;
1016 }
1017
1018 CGCUDARuntime *CodeGen::CreateNVCUDARuntime(CodeGenModule &CGM) {
1019   return new CGNVCUDARuntime(CGM);
1020 }
1021
1022 void CGNVCUDARuntime::internalizeDeviceSideVar(
1023     const VarDecl *D, llvm::GlobalValue::LinkageTypes &Linkage) {
1024   // For -fno-gpu-rdc, host-side shadows of external declarations of device-side
1025   // global variables become internal definitions. These have to be internal in
1026   // order to prevent name conflicts with global host variables with the same
1027   // name in a different TUs.
1028   //
1029   // For -fgpu-rdc, the shadow variables should not be internalized because
1030   // they may be accessed by different TU.
1031   if (CGM.getLangOpts().GPURelocatableDeviceCode)
1032     return;
1033
1034   // __shared__ variables are odd. Shadows do get created, but
1035   // they are not registered with the CUDA runtime, so they
1036   // can't really be used to access their device-side
1037   // counterparts. It's not clear yet whether it's nvcc's bug or
1038   // a feature, but we've got to do the same for compatibility.
1039   if (D->hasAttr<CUDADeviceAttr>() || D->hasAttr<CUDAConstantAttr>() ||
1040       D->hasAttr<CUDASharedAttr>() ||
1041       D->getType()->isCUDADeviceBuiltinSurfaceType() ||
1042       D->getType()->isCUDADeviceBuiltinTextureType()) {
1043     Linkage = llvm::GlobalValue::InternalLinkage;
1044   }
1045 }
1046
1047 void CGNVCUDARuntime::handleVarRegistration(const VarDecl *D,
1048                                             llvm::GlobalVariable &GV) {
1049   if (D->hasAttr<CUDADeviceAttr>() || D->hasAttr<CUDAConstantAttr>()) {
1050     // Shadow variables and their properties must be registered with CUDA
1051     // runtime. Skip Extern global variables, which will be registered in
1052     // the TU where they are defined.
1053     //
1054     // Don't register a C++17 inline variable. The local symbol can be
1055     // discarded and referencing a discarded local symbol from outside the
1056     // comdat (__cuda_register_globals) is disallowed by the ELF spec.
1057     //
1058     // HIP managed variables need to be always recorded in device and host
1059     // compilations for transformation.
1060     //
1061     // HIP managed variables and variables in CUDADeviceVarODRUsedByHost are
1062     // added to llvm.compiler-used, therefore they are safe to be registered.
1063     if ((!D->hasExternalStorage() && !D->isInline()) ||
1064         CGM.getContext().CUDADeviceVarODRUsedByHost.contains(D) ||
1065         D->hasAttr<HIPManagedAttr>()) {
1066       registerDeviceVar(D, GV, !D->hasDefinition(),
1067                         D->hasAttr<CUDAConstantAttr>());
1068     }
1069   } else if (D->getType()->isCUDADeviceBuiltinSurfaceType() ||
1070              D->getType()->isCUDADeviceBuiltinTextureType()) {
1071     // Builtin surfaces and textures and their template arguments are
1072     // also registered with CUDA runtime.
1073     const auto *TD = cast<ClassTemplateSpecializationDecl>(
1074         D->getType()->castAs<RecordType>()->getDecl());
1075     const TemplateArgumentList &Args = TD->getTemplateArgs();
1076     if (TD->hasAttr<CUDADeviceBuiltinSurfaceTypeAttr>()) {
1077       assert(Args.size() == 2 &&
1078              "Unexpected number of template arguments of CUDA device "
1079              "builtin surface type.");
1080       auto SurfType = Args[1].getAsIntegral();
1081       if (!D->hasExternalStorage())
1082         registerDeviceSurf(D, GV, !D->hasDefinition(), SurfType.getSExtValue());
1083     } else {
1084       assert(Args.size() == 3 &&
1085              "Unexpected number of template arguments of CUDA device "
1086              "builtin texture type.");
1087       auto TexType = Args[1].getAsIntegral();
1088       auto Normalized = Args[2].getAsIntegral();
1089       if (!D->hasExternalStorage())
1090         registerDeviceTex(D, GV, !D->hasDefinition(), TexType.getSExtValue(),
1091                           Normalized.getZExtValue());
1092     }
1093   }
1094 }
1095
1096 // Transform managed variables to pointers to managed variables in device code.
1097 // Each use of the original managed variable is replaced by a load from the
1098 // transformed managed variable. The transformed managed variable contains
1099 // the address of managed memory which will be allocated by the runtime.
1100 void CGNVCUDARuntime::transformManagedVars() {
1101   for (auto &&Info : DeviceVars) {
1102     llvm::GlobalVariable *Var = Info.Var;
1103     if (Info.Flags.getKind() == DeviceVarFlags::Variable &&
1104         Info.Flags.isManaged()) {
1105       auto *ManagedVar = new llvm::GlobalVariable(
1106           CGM.getModule(), Var->getType(),
1107           /*isConstant=*/false, Var->getLinkage(),
1108           /*Init=*/Var->isDeclaration()
1109               ? nullptr
1110               : llvm::ConstantPointerNull::get(Var->getType()),
1111           /*Name=*/"", /*InsertBefore=*/nullptr,
1112           llvm::GlobalVariable::NotThreadLocal,
1113           CGM.getContext().getTargetAddressSpace(LangAS::cuda_device));
1114       ManagedVar->setDSOLocal(Var->isDSOLocal());
1115       ManagedVar->setVisibility(Var->getVisibility());
1116       ManagedVar->setExternallyInitialized(true);
1117       replaceManagedVar(Var, ManagedVar);
1118       ManagedVar->takeName(Var);
1119       Var->setName(Twine(ManagedVar->getName()) + ".managed");
1120       // Keep managed variables even if they are not used in device code since
1121       // they need to be allocated by the runtime.
1122       if (!Var->isDeclaration()) {
1123         assert(!ManagedVar->isDeclaration());
1124         CGM.addCompilerUsedGlobal(Var);
1125         CGM.addCompilerUsedGlobal(ManagedVar);
1126       }
1127     }
1128   }
1129 }
1130
1131 // Creates offloading entries for all the kernels and globals that must be
1132 // registered. The linker will provide a pointer to this section so we can
1133 // register the symbols with the linked device image.
1134 void CGNVCUDARuntime::createOffloadingEntries() {
1135   llvm::OpenMPIRBuilder OMPBuilder(CGM.getModule());
1136   OMPBuilder.initialize();
1137
1138   StringRef Section = CGM.getLangOpts().HIP ? "hip_offloading_entries"
1139                                             : "cuda_offloading_entries";
1140   for (KernelInfo &I : EmittedKernels)
1141     OMPBuilder.emitOffloadingEntry(KernelHandles[I.Kernel->getName()],
1142                                    getDeviceSideName(cast<NamedDecl>(I.D)), 0,
1143                                    DeviceVarFlags::OffloadGlobalEntry, Section);
1144
1145   for (VarInfo &I : DeviceVars) {
1146     uint64_t VarSize =
1147         CGM.getDataLayout().getTypeAllocSize(I.Var->getValueType());
1148     if (I.Flags.getKind() == DeviceVarFlags::Variable) {
1149       OMPBuilder.emitOffloadingEntry(
1150           I.Var, getDeviceSideName(I.D), VarSize,
1151           I.Flags.isManaged() ? DeviceVarFlags::OffloadGlobalManagedEntry
1152                               : DeviceVarFlags::OffloadGlobalEntry,
1153           Section);
1154     } else if (I.Flags.getKind() == DeviceVarFlags::Surface) {
1155       OMPBuilder.emitOffloadingEntry(I.Var, getDeviceSideName(I.D), VarSize,
1156                                      DeviceVarFlags::OffloadGlobalSurfaceEntry,
1157                                      Section);
1158     } else if (I.Flags.getKind() == DeviceVarFlags::Texture) {
1159       OMPBuilder.emitOffloadingEntry(I.Var, getDeviceSideName(I.D), VarSize,
1160                                      DeviceVarFlags::OffloadGlobalTextureEntry,
1161                                      Section);
1162     }
1163   }
1164 }
1165
1166 // Returns module constructor to be added.
1167 llvm::Function *CGNVCUDARuntime::finalizeModule() {
1168   if (CGM.getLangOpts().CUDAIsDevice) {
1169     transformManagedVars();
1170
1171     // Mark ODR-used device variables as compiler used to prevent it from being
1172     // eliminated by optimization. This is necessary for device variables
1173     // ODR-used by host functions. Sema correctly marks them as ODR-used no
1174     // matter whether they are ODR-used by device or host functions.
1175     //
1176     // We do not need to do this if the variable has used attribute since it
1177     // has already been added.
1178     //
1179     // Static device variables have been externalized at this point, therefore
1180     // variables with LLVM private or internal linkage need not be added.
1181     for (auto &&Info : DeviceVars) {
1182       auto Kind = Info.Flags.getKind();
1183       if (!Info.Var->isDeclaration() &&
1184           !llvm::GlobalValue::isLocalLinkage(Info.Var->getLinkage()) &&
1185           (Kind == DeviceVarFlags::Variable ||
1186            Kind == DeviceVarFlags::Surface ||
1187            Kind == DeviceVarFlags::Texture) &&
1188           Info.D->isUsed() && !Info.D->hasAttr<UsedAttr>()) {
1189         CGM.addCompilerUsedGlobal(Info.Var);
1190       }
1191     }
1192     return nullptr;
1193   }
1194   if (CGM.getLangOpts().OffloadingNewDriver && RelocatableDeviceCode)
1195     createOffloadingEntries();
1196   else
1197     return makeModuleCtorFunction();
1198
1199   return nullptr;
1200 }
1201
1202 llvm::GlobalValue *CGNVCUDARuntime::getKernelHandle(llvm::Function *F,
1203                                                     GlobalDecl GD) {
1204   auto Loc = KernelHandles.find(F->getName());
1205   if (Loc != KernelHandles.end()) {
1206     auto OldHandle = Loc->second;
1207     if (KernelStubs[OldHandle] == F)
1208       return OldHandle;
1209
1210     // We've found the function name, but F itself has changed, so we need to
1211     // update the references.
1212     if (CGM.getLangOpts().HIP) {
1213       // For HIP compilation the handle itself does not change, so we only need
1214       // to update the Stub value.
1215       KernelStubs[OldHandle] = F;
1216       return OldHandle;
1217     }
1218     // For non-HIP compilation, erase the old Stub and fall-through to creating
1219     // new entries.
1220     KernelStubs.erase(OldHandle);
1221   }
1222
1223   if (!CGM.getLangOpts().HIP) {
1224     KernelHandles[F->getName()] = F;
1225     KernelStubs[F] = F;
1226     return F;
1227   }
1228
1229   auto *Var = new llvm::GlobalVariable(
1230       TheModule, F->getType(), /*isConstant=*/true, F->getLinkage(),
1231       /*Initializer=*/nullptr,
1232       CGM.getMangledName(
1233           GD.getWithKernelReferenceKind(KernelReferenceKind::Kernel)));
1234   Var->setAlignment(CGM.getPointerAlign().getAsAlign());
1235   Var->setDSOLocal(F->isDSOLocal());
1236   Var->setVisibility(F->getVisibility());
1237   CGM.maybeSetTrivialComdat(*GD.getDecl(), *Var);
1238   KernelHandles[F->getName()] = Var;
1239   KernelStubs[Var] = F;
1240   return Var;
1241 }