1 //===- OffloadWrapper.cpp ---------------------------------------*- C++ -*-===// 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 #include "llvm/Frontend/Offloading/OffloadWrapper.h" 10 #include "llvm/ADT/ArrayRef.h" 11 #include "llvm/BinaryFormat/Magic.h" 12 #include "llvm/Frontend/Offloading/Utility.h" 13 #include "llvm/IR/Constants.h" 14 #include "llvm/IR/GlobalVariable.h" 15 #include "llvm/IR/IRBuilder.h" 16 #include "llvm/IR/LLVMContext.h" 17 #include "llvm/IR/Module.h" 18 #include "llvm/Object/OffloadBinary.h" 19 #include "llvm/Support/Error.h" 20 #include "llvm/TargetParser/Triple.h" 21 #include "llvm/Transforms/Utils/ModuleUtils.h" 22 23 using namespace llvm; 24 using namespace llvm::offloading; 25 26 namespace { 27 /// Magic number that begins the section containing the CUDA fatbinary. 28 constexpr unsigned CudaFatMagic = 0x466243b1; 29 constexpr unsigned HIPFatMagic = 0x48495046; 30 31 IntegerType *getSizeTTy(Module &M) { 32 return M.getDataLayout().getIntPtrType(M.getContext()); 33 } 34 35 // struct __tgt_device_image { 36 // void *ImageStart; 37 // void *ImageEnd; 38 // __tgt_offload_entry *EntriesBegin; 39 // __tgt_offload_entry *EntriesEnd; 40 // }; 41 StructType *getDeviceImageTy(Module &M) { 42 LLVMContext &C = M.getContext(); 43 StructType *ImageTy = StructType::getTypeByName(C, "__tgt_device_image"); 44 if (!ImageTy) 45 ImageTy = 46 StructType::create("__tgt_device_image", PointerType::getUnqual(C), 47 PointerType::getUnqual(C), PointerType::getUnqual(C), 48 PointerType::getUnqual(C)); 49 return ImageTy; 50 } 51 52 PointerType *getDeviceImagePtrTy(Module &M) { 53 return PointerType::getUnqual(getDeviceImageTy(M)); 54 } 55 56 // struct __tgt_bin_desc { 57 // int32_t NumDeviceImages; 58 // __tgt_device_image *DeviceImages; 59 // __tgt_offload_entry *HostEntriesBegin; 60 // __tgt_offload_entry *HostEntriesEnd; 61 // }; 62 StructType *getBinDescTy(Module &M) { 63 LLVMContext &C = M.getContext(); 64 StructType *DescTy = StructType::getTypeByName(C, "__tgt_bin_desc"); 65 if (!DescTy) 66 DescTy = StructType::create( 67 "__tgt_bin_desc", Type::getInt32Ty(C), getDeviceImagePtrTy(M), 68 PointerType::getUnqual(C), PointerType::getUnqual(C)); 69 return DescTy; 70 } 71 72 PointerType *getBinDescPtrTy(Module &M) { 73 return PointerType::getUnqual(getBinDescTy(M)); 74 } 75 76 /// Creates binary descriptor for the given device images. Binary descriptor 77 /// is an object that is passed to the offloading runtime at program startup 78 /// and it describes all device images available in the executable or shared 79 /// library. It is defined as follows 80 /// 81 /// __attribute__((visibility("hidden"))) 82 /// extern __tgt_offload_entry *__start_omp_offloading_entries; 83 /// __attribute__((visibility("hidden"))) 84 /// extern __tgt_offload_entry *__stop_omp_offloading_entries; 85 /// 86 /// static const char Image0[] = { <Bufs.front() contents> }; 87 /// ... 88 /// static const char ImageN[] = { <Bufs.back() contents> }; 89 /// 90 /// static const __tgt_device_image Images[] = { 91 /// { 92 /// Image0, /*ImageStart*/ 93 /// Image0 + sizeof(Image0), /*ImageEnd*/ 94 /// __start_omp_offloading_entries, /*EntriesBegin*/ 95 /// __stop_omp_offloading_entries /*EntriesEnd*/ 96 /// }, 97 /// ... 98 /// { 99 /// ImageN, /*ImageStart*/ 100 /// ImageN + sizeof(ImageN), /*ImageEnd*/ 101 /// __start_omp_offloading_entries, /*EntriesBegin*/ 102 /// __stop_omp_offloading_entries /*EntriesEnd*/ 103 /// } 104 /// }; 105 /// 106 /// static const __tgt_bin_desc BinDesc = { 107 /// sizeof(Images) / sizeof(Images[0]), /*NumDeviceImages*/ 108 /// Images, /*DeviceImages*/ 109 /// __start_omp_offloading_entries, /*HostEntriesBegin*/ 110 /// __stop_omp_offloading_entries /*HostEntriesEnd*/ 111 /// }; 112 /// 113 /// Global variable that represents BinDesc is returned. 114 GlobalVariable *createBinDesc(Module &M, ArrayRef<ArrayRef<char>> Bufs, 115 EntryArrayTy EntryArray, StringRef Suffix, 116 bool Relocatable) { 117 LLVMContext &C = M.getContext(); 118 auto [EntriesB, EntriesE] = EntryArray; 119 120 auto *Zero = ConstantInt::get(getSizeTTy(M), 0u); 121 Constant *ZeroZero[] = {Zero, Zero}; 122 123 // Create initializer for the images array. 124 SmallVector<Constant *, 4u> ImagesInits; 125 ImagesInits.reserve(Bufs.size()); 126 for (ArrayRef<char> Buf : Bufs) { 127 // We embed the full offloading entry so the binary utilities can parse it. 128 auto *Data = ConstantDataArray::get(C, Buf); 129 auto *Image = new GlobalVariable(M, Data->getType(), /*isConstant=*/true, 130 GlobalVariable::InternalLinkage, Data, 131 ".omp_offloading.device_image" + Suffix); 132 Image->setUnnamedAddr(GlobalValue::UnnamedAddr::Global); 133 Image->setSection(Relocatable ? ".llvm.offloading.relocatable" 134 : ".llvm.offloading"); 135 Image->setAlignment(Align(object::OffloadBinary::getAlignment())); 136 137 StringRef Binary(Buf.data(), Buf.size()); 138 assert(identify_magic(Binary) == file_magic::offload_binary && 139 "Invalid binary format"); 140 141 // The device image struct contains the pointer to the beginning and end of 142 // the image stored inside of the offload binary. There should only be one 143 // of these for each buffer so we parse it out manually. 144 const auto *Header = 145 reinterpret_cast<const object::OffloadBinary::Header *>( 146 Binary.bytes_begin()); 147 const auto *Entry = reinterpret_cast<const object::OffloadBinary::Entry *>( 148 Binary.bytes_begin() + Header->EntryOffset); 149 150 auto *Begin = ConstantInt::get(getSizeTTy(M), Entry->ImageOffset); 151 auto *Size = 152 ConstantInt::get(getSizeTTy(M), Entry->ImageOffset + Entry->ImageSize); 153 Constant *ZeroBegin[] = {Zero, Begin}; 154 Constant *ZeroSize[] = {Zero, Size}; 155 156 auto *ImageB = 157 ConstantExpr::getGetElementPtr(Image->getValueType(), Image, ZeroBegin); 158 auto *ImageE = 159 ConstantExpr::getGetElementPtr(Image->getValueType(), Image, ZeroSize); 160 161 ImagesInits.push_back(ConstantStruct::get(getDeviceImageTy(M), ImageB, 162 ImageE, EntriesB, EntriesE)); 163 } 164 165 // Then create images array. 166 auto *ImagesData = ConstantArray::get( 167 ArrayType::get(getDeviceImageTy(M), ImagesInits.size()), ImagesInits); 168 169 auto *Images = 170 new GlobalVariable(M, ImagesData->getType(), /*isConstant*/ true, 171 GlobalValue::InternalLinkage, ImagesData, 172 ".omp_offloading.device_images" + Suffix); 173 Images->setUnnamedAddr(GlobalValue::UnnamedAddr::Global); 174 175 auto *ImagesB = 176 ConstantExpr::getGetElementPtr(Images->getValueType(), Images, ZeroZero); 177 178 // And finally create the binary descriptor object. 179 auto *DescInit = ConstantStruct::get( 180 getBinDescTy(M), 181 ConstantInt::get(Type::getInt32Ty(C), ImagesInits.size()), ImagesB, 182 EntriesB, EntriesE); 183 184 return new GlobalVariable(M, DescInit->getType(), /*isConstant*/ true, 185 GlobalValue::InternalLinkage, DescInit, 186 ".omp_offloading.descriptor" + Suffix); 187 } 188 189 Function *createUnregisterFunction(Module &M, GlobalVariable *BinDesc, 190 StringRef Suffix) { 191 LLVMContext &C = M.getContext(); 192 auto *FuncTy = FunctionType::get(Type::getVoidTy(C), /*isVarArg*/ false); 193 auto *Func = 194 Function::Create(FuncTy, GlobalValue::InternalLinkage, 195 ".omp_offloading.descriptor_unreg" + Suffix, &M); 196 Func->setSection(".text.startup"); 197 198 // Get __tgt_unregister_lib function declaration. 199 auto *UnRegFuncTy = FunctionType::get(Type::getVoidTy(C), getBinDescPtrTy(M), 200 /*isVarArg*/ false); 201 FunctionCallee UnRegFuncC = 202 M.getOrInsertFunction("__tgt_unregister_lib", UnRegFuncTy); 203 204 // Construct function body 205 IRBuilder<> Builder(BasicBlock::Create(C, "entry", Func)); 206 Builder.CreateCall(UnRegFuncC, BinDesc); 207 Builder.CreateRetVoid(); 208 209 return Func; 210 } 211 212 void createRegisterFunction(Module &M, GlobalVariable *BinDesc, 213 StringRef Suffix) { 214 LLVMContext &C = M.getContext(); 215 auto *FuncTy = FunctionType::get(Type::getVoidTy(C), /*isVarArg*/ false); 216 auto *Func = Function::Create(FuncTy, GlobalValue::InternalLinkage, 217 ".omp_offloading.descriptor_reg" + Suffix, &M); 218 Func->setSection(".text.startup"); 219 220 // Get __tgt_register_lib function declaration. 221 auto *RegFuncTy = FunctionType::get(Type::getVoidTy(C), getBinDescPtrTy(M), 222 /*isVarArg*/ false); 223 FunctionCallee RegFuncC = 224 M.getOrInsertFunction("__tgt_register_lib", RegFuncTy); 225 226 auto *AtExitTy = FunctionType::get( 227 Type::getInt32Ty(C), PointerType::getUnqual(C), /*isVarArg=*/false); 228 FunctionCallee AtExit = M.getOrInsertFunction("atexit", AtExitTy); 229 230 Function *UnregFunc = createUnregisterFunction(M, BinDesc, Suffix); 231 232 // Construct function body 233 IRBuilder<> Builder(BasicBlock::Create(C, "entry", Func)); 234 235 Builder.CreateCall(RegFuncC, BinDesc); 236 237 // Register the destructors with 'atexit'. This is expected by the CUDA 238 // runtime and ensures that we clean up before dynamic objects are destroyed. 239 // This needs to be done after plugin initialization to ensure that it is 240 // called before the plugin runtime is destroyed. 241 Builder.CreateCall(AtExit, UnregFunc); 242 Builder.CreateRetVoid(); 243 244 // Add this function to constructors. 245 appendToGlobalCtors(M, Func, /*Priority=*/101); 246 } 247 248 // struct fatbin_wrapper { 249 // int32_t magic; 250 // int32_t version; 251 // void *image; 252 // void *reserved; 253 //}; 254 StructType *getFatbinWrapperTy(Module &M) { 255 LLVMContext &C = M.getContext(); 256 StructType *FatbinTy = StructType::getTypeByName(C, "fatbin_wrapper"); 257 if (!FatbinTy) 258 FatbinTy = StructType::create( 259 "fatbin_wrapper", Type::getInt32Ty(C), Type::getInt32Ty(C), 260 PointerType::getUnqual(C), PointerType::getUnqual(C)); 261 return FatbinTy; 262 } 263 264 /// Embed the image \p Image into the module \p M so it can be found by the 265 /// runtime. 266 GlobalVariable *createFatbinDesc(Module &M, ArrayRef<char> Image, bool IsHIP, 267 StringRef Suffix) { 268 LLVMContext &C = M.getContext(); 269 llvm::Type *Int8PtrTy = PointerType::getUnqual(C); 270 llvm::Triple Triple = llvm::Triple(M.getTargetTriple()); 271 272 // Create the global string containing the fatbinary. 273 StringRef FatbinConstantSection = 274 IsHIP ? ".hip_fatbin" 275 : (Triple.isMacOSX() ? "__NV_CUDA,__nv_fatbin" : ".nv_fatbin"); 276 auto *Data = ConstantDataArray::get(C, Image); 277 auto *Fatbin = new GlobalVariable(M, Data->getType(), /*isConstant*/ true, 278 GlobalVariable::InternalLinkage, Data, 279 ".fatbin_image" + Suffix); 280 Fatbin->setSection(FatbinConstantSection); 281 282 // Create the fatbinary wrapper 283 StringRef FatbinWrapperSection = IsHIP ? ".hipFatBinSegment" 284 : Triple.isMacOSX() ? "__NV_CUDA,__fatbin" 285 : ".nvFatBinSegment"; 286 Constant *FatbinWrapper[] = { 287 ConstantInt::get(Type::getInt32Ty(C), IsHIP ? HIPFatMagic : CudaFatMagic), 288 ConstantInt::get(Type::getInt32Ty(C), 1), 289 ConstantExpr::getPointerBitCastOrAddrSpaceCast(Fatbin, Int8PtrTy), 290 ConstantPointerNull::get(PointerType::getUnqual(C))}; 291 292 Constant *FatbinInitializer = 293 ConstantStruct::get(getFatbinWrapperTy(M), FatbinWrapper); 294 295 auto *FatbinDesc = 296 new GlobalVariable(M, getFatbinWrapperTy(M), 297 /*isConstant*/ true, GlobalValue::InternalLinkage, 298 FatbinInitializer, ".fatbin_wrapper" + Suffix); 299 FatbinDesc->setSection(FatbinWrapperSection); 300 FatbinDesc->setAlignment(Align(8)); 301 302 return FatbinDesc; 303 } 304 305 /// Create the register globals function. We will iterate all of the offloading 306 /// entries stored at the begin / end symbols and register them according to 307 /// their type. This creates the following function in IR: 308 /// 309 /// extern struct __tgt_offload_entry __start_cuda_offloading_entries; 310 /// extern struct __tgt_offload_entry __stop_cuda_offloading_entries; 311 /// 312 /// extern void __cudaRegisterFunction(void **, void *, void *, void *, int, 313 /// void *, void *, void *, void *, int *); 314 /// extern void __cudaRegisterVar(void **, void *, void *, void *, int32_t, 315 /// int64_t, int32_t, int32_t); 316 /// 317 /// void __cudaRegisterTest(void **fatbinHandle) { 318 /// for (struct __tgt_offload_entry *entry = &__start_cuda_offloading_entries; 319 /// entry != &__stop_cuda_offloading_entries; ++entry) { 320 /// if (!entry->size) 321 /// __cudaRegisterFunction(fatbinHandle, entry->addr, entry->name, 322 /// entry->name, -1, 0, 0, 0, 0, 0); 323 /// else 324 /// __cudaRegisterVar(fatbinHandle, entry->addr, entry->name, entry->name, 325 /// 0, entry->size, 0, 0); 326 /// } 327 /// } 328 Function *createRegisterGlobalsFunction(Module &M, bool IsHIP, 329 EntryArrayTy EntryArray, 330 StringRef Suffix, 331 bool EmitSurfacesAndTextures) { 332 LLVMContext &C = M.getContext(); 333 auto [EntriesB, EntriesE] = EntryArray; 334 335 // Get the __cudaRegisterFunction function declaration. 336 PointerType *Int8PtrTy = PointerType::get(C, 0); 337 PointerType *Int8PtrPtrTy = PointerType::get(C, 0); 338 PointerType *Int32PtrTy = PointerType::get(C, 0); 339 auto *RegFuncTy = FunctionType::get( 340 Type::getInt32Ty(C), 341 {Int8PtrPtrTy, Int8PtrTy, Int8PtrTy, Int8PtrTy, Type::getInt32Ty(C), 342 Int8PtrTy, Int8PtrTy, Int8PtrTy, Int8PtrTy, Int32PtrTy}, 343 /*isVarArg*/ false); 344 FunctionCallee RegFunc = M.getOrInsertFunction( 345 IsHIP ? "__hipRegisterFunction" : "__cudaRegisterFunction", RegFuncTy); 346 347 // Get the __cudaRegisterVar function declaration. 348 auto *RegVarTy = FunctionType::get( 349 Type::getVoidTy(C), 350 {Int8PtrPtrTy, Int8PtrTy, Int8PtrTy, Int8PtrTy, Type::getInt32Ty(C), 351 getSizeTTy(M), Type::getInt32Ty(C), Type::getInt32Ty(C)}, 352 /*isVarArg*/ false); 353 FunctionCallee RegVar = M.getOrInsertFunction( 354 IsHIP ? "__hipRegisterVar" : "__cudaRegisterVar", RegVarTy); 355 356 // Get the __cudaRegisterSurface function declaration. 357 FunctionType *RegManagedVarTy = 358 FunctionType::get(Type::getVoidTy(C), 359 {Int8PtrPtrTy, Int8PtrTy, Int8PtrTy, Int8PtrTy, 360 getSizeTTy(M), Type::getInt32Ty(C)}, 361 /*isVarArg=*/false); 362 FunctionCallee RegManagedVar = M.getOrInsertFunction( 363 IsHIP ? "__hipRegisterManagedVar" : "__cudaRegisterManagedVar", 364 RegManagedVarTy); 365 366 // Get the __cudaRegisterSurface function declaration. 367 FunctionType *RegSurfaceTy = 368 FunctionType::get(Type::getVoidTy(C), 369 {Int8PtrPtrTy, Int8PtrTy, Int8PtrTy, Int8PtrTy, 370 Type::getInt32Ty(C), Type::getInt32Ty(C)}, 371 /*isVarArg=*/false); 372 FunctionCallee RegSurface = M.getOrInsertFunction( 373 IsHIP ? "__hipRegisterSurface" : "__cudaRegisterSurface", RegSurfaceTy); 374 375 // Get the __cudaRegisterTexture function declaration. 376 FunctionType *RegTextureTy = FunctionType::get( 377 Type::getVoidTy(C), 378 {Int8PtrPtrTy, Int8PtrTy, Int8PtrTy, Int8PtrTy, Type::getInt32Ty(C), 379 Type::getInt32Ty(C), Type::getInt32Ty(C)}, 380 /*isVarArg=*/false); 381 FunctionCallee RegTexture = M.getOrInsertFunction( 382 IsHIP ? "__hipRegisterTexture" : "__cudaRegisterTexture", RegTextureTy); 383 384 auto *RegGlobalsTy = FunctionType::get(Type::getVoidTy(C), Int8PtrPtrTy, 385 /*isVarArg*/ false); 386 auto *RegGlobalsFn = 387 Function::Create(RegGlobalsTy, GlobalValue::InternalLinkage, 388 IsHIP ? ".hip.globals_reg" : ".cuda.globals_reg", &M); 389 RegGlobalsFn->setSection(".text.startup"); 390 391 // Create the loop to register all the entries. 392 IRBuilder<> Builder(BasicBlock::Create(C, "entry", RegGlobalsFn)); 393 auto *EntryBB = BasicBlock::Create(C, "while.entry", RegGlobalsFn); 394 auto *IfThenBB = BasicBlock::Create(C, "if.then", RegGlobalsFn); 395 auto *IfElseBB = BasicBlock::Create(C, "if.else", RegGlobalsFn); 396 auto *SwGlobalBB = BasicBlock::Create(C, "sw.global", RegGlobalsFn); 397 auto *SwManagedBB = BasicBlock::Create(C, "sw.managed", RegGlobalsFn); 398 auto *SwSurfaceBB = BasicBlock::Create(C, "sw.surface", RegGlobalsFn); 399 auto *SwTextureBB = BasicBlock::Create(C, "sw.texture", RegGlobalsFn); 400 auto *IfEndBB = BasicBlock::Create(C, "if.end", RegGlobalsFn); 401 auto *ExitBB = BasicBlock::Create(C, "while.end", RegGlobalsFn); 402 403 auto *EntryCmp = Builder.CreateICmpNE(EntriesB, EntriesE); 404 Builder.CreateCondBr(EntryCmp, EntryBB, ExitBB); 405 Builder.SetInsertPoint(EntryBB); 406 auto *Entry = Builder.CreatePHI(PointerType::getUnqual(C), 2, "entry"); 407 auto *AddrPtr = 408 Builder.CreateInBoundsGEP(offloading::getEntryTy(M), Entry, 409 {ConstantInt::get(getSizeTTy(M), 0), 410 ConstantInt::get(Type::getInt32Ty(C), 0)}); 411 auto *Addr = Builder.CreateLoad(Int8PtrTy, AddrPtr, "addr"); 412 auto *NamePtr = 413 Builder.CreateInBoundsGEP(offloading::getEntryTy(M), Entry, 414 {ConstantInt::get(getSizeTTy(M), 0), 415 ConstantInt::get(Type::getInt32Ty(C), 1)}); 416 auto *Name = Builder.CreateLoad(Int8PtrTy, NamePtr, "name"); 417 auto *SizePtr = 418 Builder.CreateInBoundsGEP(offloading::getEntryTy(M), Entry, 419 {ConstantInt::get(getSizeTTy(M), 0), 420 ConstantInt::get(Type::getInt32Ty(C), 2)}); 421 auto *Size = Builder.CreateLoad(getSizeTTy(M), SizePtr, "size"); 422 auto *FlagsPtr = 423 Builder.CreateInBoundsGEP(offloading::getEntryTy(M), Entry, 424 {ConstantInt::get(getSizeTTy(M), 0), 425 ConstantInt::get(Type::getInt32Ty(C), 3)}); 426 auto *Flags = Builder.CreateLoad(Type::getInt32Ty(C), FlagsPtr, "flags"); 427 auto *DataPtr = 428 Builder.CreateInBoundsGEP(offloading::getEntryTy(M), Entry, 429 {ConstantInt::get(getSizeTTy(M), 0), 430 ConstantInt::get(Type::getInt32Ty(C), 4)}); 431 auto *Data = Builder.CreateLoad(Type::getInt32Ty(C), DataPtr, "textype"); 432 auto *Kind = Builder.CreateAnd( 433 Flags, ConstantInt::get(Type::getInt32Ty(C), 0x7), "type"); 434 435 // Extract the flags stored in the bit-field and convert them to C booleans. 436 auto *ExternBit = Builder.CreateAnd( 437 Flags, ConstantInt::get(Type::getInt32Ty(C), 438 llvm::offloading::OffloadGlobalExtern)); 439 auto *Extern = Builder.CreateLShr( 440 ExternBit, ConstantInt::get(Type::getInt32Ty(C), 3), "extern"); 441 auto *ConstantBit = Builder.CreateAnd( 442 Flags, ConstantInt::get(Type::getInt32Ty(C), 443 llvm::offloading::OffloadGlobalConstant)); 444 auto *Const = Builder.CreateLShr( 445 ConstantBit, ConstantInt::get(Type::getInt32Ty(C), 4), "constant"); 446 auto *NormalizedBit = Builder.CreateAnd( 447 Flags, ConstantInt::get(Type::getInt32Ty(C), 448 llvm::offloading::OffloadGlobalNormalized)); 449 auto *Normalized = Builder.CreateLShr( 450 NormalizedBit, ConstantInt::get(Type::getInt32Ty(C), 5), "normalized"); 451 auto *FnCond = 452 Builder.CreateICmpEQ(Size, ConstantInt::getNullValue(getSizeTTy(M))); 453 Builder.CreateCondBr(FnCond, IfThenBB, IfElseBB); 454 455 // Create kernel registration code. 456 Builder.SetInsertPoint(IfThenBB); 457 Builder.CreateCall(RegFunc, {RegGlobalsFn->arg_begin(), Addr, Name, Name, 458 ConstantInt::get(Type::getInt32Ty(C), -1), 459 ConstantPointerNull::get(Int8PtrTy), 460 ConstantPointerNull::get(Int8PtrTy), 461 ConstantPointerNull::get(Int8PtrTy), 462 ConstantPointerNull::get(Int8PtrTy), 463 ConstantPointerNull::get(Int32PtrTy)}); 464 Builder.CreateBr(IfEndBB); 465 Builder.SetInsertPoint(IfElseBB); 466 467 auto *Switch = Builder.CreateSwitch(Kind, IfEndBB); 468 // Create global variable registration code. 469 Builder.SetInsertPoint(SwGlobalBB); 470 Builder.CreateCall(RegVar, 471 {RegGlobalsFn->arg_begin(), Addr, Name, Name, Extern, Size, 472 Const, ConstantInt::get(Type::getInt32Ty(C), 0)}); 473 Builder.CreateBr(IfEndBB); 474 Switch->addCase(Builder.getInt32(llvm::offloading::OffloadGlobalEntry), 475 SwGlobalBB); 476 477 // Create managed variable registration code. 478 Builder.SetInsertPoint(SwManagedBB); 479 auto *ManagedVar = Builder.CreateLoad(Int8PtrTy, Addr, "managed.addr"); 480 auto *ManagedAddr = Builder.CreateInBoundsGEP( 481 Int8PtrTy, Addr, {ConstantInt::get(Builder.getInt64Ty(), 1)}); 482 auto *Managed = Builder.CreateLoad(Int8PtrTy, ManagedAddr, "managed.addr"); 483 Builder.CreateCall(RegManagedVar, {RegGlobalsFn->arg_begin(), ManagedVar, 484 Managed, Name, Size, Data}); 485 Builder.CreateBr(IfEndBB); 486 Switch->addCase(Builder.getInt32(llvm::offloading::OffloadGlobalManagedEntry), 487 SwManagedBB); 488 // Create surface variable registration code. 489 Builder.SetInsertPoint(SwSurfaceBB); 490 if (EmitSurfacesAndTextures) 491 Builder.CreateCall(RegSurface, {RegGlobalsFn->arg_begin(), Addr, Name, Name, 492 Data, Extern}); 493 Builder.CreateBr(IfEndBB); 494 Switch->addCase(Builder.getInt32(llvm::offloading::OffloadGlobalSurfaceEntry), 495 SwSurfaceBB); 496 497 // Create texture variable registration code. 498 Builder.SetInsertPoint(SwTextureBB); 499 if (EmitSurfacesAndTextures) 500 Builder.CreateCall(RegTexture, {RegGlobalsFn->arg_begin(), Addr, Name, Name, 501 Data, Normalized, Extern}); 502 Builder.CreateBr(IfEndBB); 503 Switch->addCase(Builder.getInt32(llvm::offloading::OffloadGlobalTextureEntry), 504 SwTextureBB); 505 506 Builder.SetInsertPoint(IfEndBB); 507 auto *NewEntry = Builder.CreateInBoundsGEP( 508 offloading::getEntryTy(M), Entry, ConstantInt::get(getSizeTTy(M), 1)); 509 auto *Cmp = Builder.CreateICmpEQ( 510 NewEntry, 511 ConstantExpr::getInBoundsGetElementPtr( 512 ArrayType::get(offloading::getEntryTy(M), 0), EntriesE, 513 ArrayRef<Constant *>({ConstantInt::get(getSizeTTy(M), 0), 514 ConstantInt::get(getSizeTTy(M), 0)}))); 515 Entry->addIncoming( 516 ConstantExpr::getInBoundsGetElementPtr( 517 ArrayType::get(offloading::getEntryTy(M), 0), EntriesB, 518 ArrayRef<Constant *>({ConstantInt::get(getSizeTTy(M), 0), 519 ConstantInt::get(getSizeTTy(M), 0)})), 520 &RegGlobalsFn->getEntryBlock()); 521 Entry->addIncoming(NewEntry, IfEndBB); 522 Builder.CreateCondBr(Cmp, ExitBB, EntryBB); 523 Builder.SetInsertPoint(ExitBB); 524 Builder.CreateRetVoid(); 525 526 return RegGlobalsFn; 527 } 528 529 // Create the constructor and destructor to register the fatbinary with the CUDA 530 // runtime. 531 void createRegisterFatbinFunction(Module &M, GlobalVariable *FatbinDesc, 532 bool IsHIP, EntryArrayTy EntryArray, 533 StringRef Suffix, 534 bool EmitSurfacesAndTextures) { 535 LLVMContext &C = M.getContext(); 536 auto *CtorFuncTy = FunctionType::get(Type::getVoidTy(C), /*isVarArg*/ false); 537 auto *CtorFunc = Function::Create( 538 CtorFuncTy, GlobalValue::InternalLinkage, 539 (IsHIP ? ".hip.fatbin_reg" : ".cuda.fatbin_reg") + Suffix, &M); 540 CtorFunc->setSection(".text.startup"); 541 542 auto *DtorFuncTy = FunctionType::get(Type::getVoidTy(C), /*isVarArg*/ false); 543 auto *DtorFunc = Function::Create( 544 DtorFuncTy, GlobalValue::InternalLinkage, 545 (IsHIP ? ".hip.fatbin_unreg" : ".cuda.fatbin_unreg") + Suffix, &M); 546 DtorFunc->setSection(".text.startup"); 547 548 auto *PtrTy = PointerType::getUnqual(C); 549 550 // Get the __cudaRegisterFatBinary function declaration. 551 auto *RegFatTy = FunctionType::get(PtrTy, PtrTy, /*isVarArg=*/false); 552 FunctionCallee RegFatbin = M.getOrInsertFunction( 553 IsHIP ? "__hipRegisterFatBinary" : "__cudaRegisterFatBinary", RegFatTy); 554 // Get the __cudaRegisterFatBinaryEnd function declaration. 555 auto *RegFatEndTy = 556 FunctionType::get(Type::getVoidTy(C), PtrTy, /*isVarArg=*/false); 557 FunctionCallee RegFatbinEnd = 558 M.getOrInsertFunction("__cudaRegisterFatBinaryEnd", RegFatEndTy); 559 // Get the __cudaUnregisterFatBinary function declaration. 560 auto *UnregFatTy = 561 FunctionType::get(Type::getVoidTy(C), PtrTy, /*isVarArg=*/false); 562 FunctionCallee UnregFatbin = M.getOrInsertFunction( 563 IsHIP ? "__hipUnregisterFatBinary" : "__cudaUnregisterFatBinary", 564 UnregFatTy); 565 566 auto *AtExitTy = 567 FunctionType::get(Type::getInt32Ty(C), PtrTy, /*isVarArg=*/false); 568 FunctionCallee AtExit = M.getOrInsertFunction("atexit", AtExitTy); 569 570 auto *BinaryHandleGlobal = new llvm::GlobalVariable( 571 M, PtrTy, false, llvm::GlobalValue::InternalLinkage, 572 llvm::ConstantPointerNull::get(PtrTy), 573 (IsHIP ? ".hip.binary_handle" : ".cuda.binary_handle") + Suffix); 574 575 // Create the constructor to register this image with the runtime. 576 IRBuilder<> CtorBuilder(BasicBlock::Create(C, "entry", CtorFunc)); 577 CallInst *Handle = CtorBuilder.CreateCall( 578 RegFatbin, 579 ConstantExpr::getPointerBitCastOrAddrSpaceCast(FatbinDesc, PtrTy)); 580 CtorBuilder.CreateAlignedStore( 581 Handle, BinaryHandleGlobal, 582 Align(M.getDataLayout().getPointerTypeSize(PtrTy))); 583 CtorBuilder.CreateCall(createRegisterGlobalsFunction(M, IsHIP, EntryArray, 584 Suffix, 585 EmitSurfacesAndTextures), 586 Handle); 587 if (!IsHIP) 588 CtorBuilder.CreateCall(RegFatbinEnd, Handle); 589 CtorBuilder.CreateCall(AtExit, DtorFunc); 590 CtorBuilder.CreateRetVoid(); 591 592 // Create the destructor to unregister the image with the runtime. We cannot 593 // use a standard global destructor after CUDA 9.2 so this must be called by 594 // `atexit()` instead. 595 IRBuilder<> DtorBuilder(BasicBlock::Create(C, "entry", DtorFunc)); 596 LoadInst *BinaryHandle = DtorBuilder.CreateAlignedLoad( 597 PtrTy, BinaryHandleGlobal, 598 Align(M.getDataLayout().getPointerTypeSize(PtrTy))); 599 DtorBuilder.CreateCall(UnregFatbin, BinaryHandle); 600 DtorBuilder.CreateRetVoid(); 601 602 // Add this function to constructors. 603 appendToGlobalCtors(M, CtorFunc, /*Priority=*/101); 604 } 605 } // namespace 606 607 Error offloading::wrapOpenMPBinaries(Module &M, ArrayRef<ArrayRef<char>> Images, 608 EntryArrayTy EntryArray, 609 llvm::StringRef Suffix, bool Relocatable) { 610 GlobalVariable *Desc = 611 createBinDesc(M, Images, EntryArray, Suffix, Relocatable); 612 if (!Desc) 613 return createStringError(inconvertibleErrorCode(), 614 "No binary descriptors created."); 615 createRegisterFunction(M, Desc, Suffix); 616 return Error::success(); 617 } 618 619 Error offloading::wrapCudaBinary(Module &M, ArrayRef<char> Image, 620 EntryArrayTy EntryArray, 621 llvm::StringRef Suffix, 622 bool EmitSurfacesAndTextures) { 623 GlobalVariable *Desc = createFatbinDesc(M, Image, /*IsHip=*/false, Suffix); 624 if (!Desc) 625 return createStringError(inconvertibleErrorCode(), 626 "No fatbin section created."); 627 628 createRegisterFatbinFunction(M, Desc, /*IsHip=*/false, EntryArray, Suffix, 629 EmitSurfacesAndTextures); 630 return Error::success(); 631 } 632 633 Error offloading::wrapHIPBinary(Module &M, ArrayRef<char> Image, 634 EntryArrayTy EntryArray, llvm::StringRef Suffix, 635 bool EmitSurfacesAndTextures) { 636 GlobalVariable *Desc = createFatbinDesc(M, Image, /*IsHip=*/true, Suffix); 637 if (!Desc) 638 return createStringError(inconvertibleErrorCode(), 639 "No fatbin section created."); 640 641 createRegisterFatbinFunction(M, Desc, /*IsHip=*/true, EntryArray, Suffix, 642 EmitSurfacesAndTextures); 643 return Error::success(); 644 } 645