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