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