1 //===-- NVPTXCtorDtorLowering.cpp - Handle global ctors and dtors --------===// 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 /// \file 10 /// This pass creates a unified init and fini kernel with the required metadata 11 //===----------------------------------------------------------------------===// 12 13 #include "NVPTXCtorDtorLowering.h" 14 #include "MCTargetDesc/NVPTXBaseInfo.h" 15 #include "NVPTX.h" 16 #include "llvm/ADT/StringExtras.h" 17 #include "llvm/IR/CallingConv.h" 18 #include "llvm/IR/Constants.h" 19 #include "llvm/IR/Function.h" 20 #include "llvm/IR/GlobalVariable.h" 21 #include "llvm/IR/IRBuilder.h" 22 #include "llvm/IR/Module.h" 23 #include "llvm/IR/Value.h" 24 #include "llvm/Pass.h" 25 #include "llvm/Support/CommandLine.h" 26 #include "llvm/Support/MD5.h" 27 #include "llvm/Transforms/Utils/ModuleUtils.h" 28 29 using namespace llvm; 30 31 #define DEBUG_TYPE "nvptx-lower-ctor-dtor" 32 33 static cl::opt<std::string> 34 GlobalStr("nvptx-lower-global-ctor-dtor-id", 35 cl::desc("Override unique ID of ctor/dtor globals."), 36 cl::init(""), cl::Hidden); 37 38 static cl::opt<bool> 39 CreateKernels("nvptx-emit-init-fini-kernel", 40 cl::desc("Emit kernels to call ctor/dtor globals."), 41 cl::init(true), cl::Hidden); 42 43 namespace { 44 45 static std::string getHash(StringRef Str) { 46 llvm::MD5 Hasher; 47 llvm::MD5::MD5Result Hash; 48 Hasher.update(Str); 49 Hasher.final(Hash); 50 return llvm::utohexstr(Hash.low(), /*LowerCase=*/true); 51 } 52 53 static void addKernelMetadata(Module &M, Function *F) { 54 llvm::LLVMContext &Ctx = M.getContext(); 55 56 // Get "nvvm.annotations" metadata node. 57 llvm::NamedMDNode *MD = M.getOrInsertNamedMetadata("nvvm.annotations"); 58 59 // This kernel is only to be called single-threaded. 60 llvm::Metadata *ThreadXMDVals[] = { 61 llvm::ConstantAsMetadata::get(F), llvm::MDString::get(Ctx, "maxntidx"), 62 llvm::ConstantAsMetadata::get( 63 llvm::ConstantInt::get(llvm::Type::getInt32Ty(Ctx), 1))}; 64 llvm::Metadata *ThreadYMDVals[] = { 65 llvm::ConstantAsMetadata::get(F), llvm::MDString::get(Ctx, "maxntidy"), 66 llvm::ConstantAsMetadata::get( 67 llvm::ConstantInt::get(llvm::Type::getInt32Ty(Ctx), 1))}; 68 llvm::Metadata *ThreadZMDVals[] = { 69 llvm::ConstantAsMetadata::get(F), llvm::MDString::get(Ctx, "maxntidz"), 70 llvm::ConstantAsMetadata::get( 71 llvm::ConstantInt::get(llvm::Type::getInt32Ty(Ctx), 1))}; 72 73 llvm::Metadata *BlockMDVals[] = { 74 llvm::ConstantAsMetadata::get(F), 75 llvm::MDString::get(Ctx, "maxclusterrank"), 76 llvm::ConstantAsMetadata::get( 77 llvm::ConstantInt::get(llvm::Type::getInt32Ty(Ctx), 1))}; 78 79 // Append metadata to nvvm.annotations. 80 F->setCallingConv(CallingConv::PTX_Kernel); 81 MD->addOperand(llvm::MDNode::get(Ctx, ThreadXMDVals)); 82 MD->addOperand(llvm::MDNode::get(Ctx, ThreadYMDVals)); 83 MD->addOperand(llvm::MDNode::get(Ctx, ThreadZMDVals)); 84 MD->addOperand(llvm::MDNode::get(Ctx, BlockMDVals)); 85 } 86 87 static Function *createInitOrFiniKernelFunction(Module &M, bool IsCtor) { 88 StringRef InitOrFiniKernelName = 89 IsCtor ? "nvptx$device$init" : "nvptx$device$fini"; 90 if (M.getFunction(InitOrFiniKernelName)) 91 return nullptr; 92 93 Function *InitOrFiniKernel = Function::createWithDefaultAttr( 94 FunctionType::get(Type::getVoidTy(M.getContext()), false), 95 GlobalValue::WeakODRLinkage, 0, InitOrFiniKernelName, &M); 96 addKernelMetadata(M, InitOrFiniKernel); 97 98 return InitOrFiniKernel; 99 } 100 101 // We create the IR required to call each callback in this section. This is 102 // equivalent to the following code. Normally, the linker would provide us with 103 // the definitions of the init and fini array sections. The 'nvlink' linker does 104 // not do this so initializing these values is done by the runtime. 105 // 106 // extern "C" void **__init_array_start = nullptr; 107 // extern "C" void **__init_array_end = nullptr; 108 // extern "C" void **__fini_array_start = nullptr; 109 // extern "C" void **__fini_array_end = nullptr; 110 // 111 // using InitCallback = void(); 112 // using FiniCallback = void(); 113 // 114 // void call_init_array_callbacks() { 115 // for (auto start = __init_array_start; start != __init_array_end; ++start) 116 // reinterpret_cast<InitCallback *>(*start)(); 117 // } 118 // 119 // void call_init_array_callbacks() { 120 // size_t fini_array_size = __fini_array_end - __fini_array_start; 121 // for (size_t i = fini_array_size; i > 0; --i) 122 // reinterpret_cast<FiniCallback *>(__fini_array_start[i - 1])(); 123 // } 124 static void createInitOrFiniCalls(Function &F, bool IsCtor) { 125 Module &M = *F.getParent(); 126 LLVMContext &C = M.getContext(); 127 128 IRBuilder<> IRB(BasicBlock::Create(C, "entry", &F)); 129 auto *LoopBB = BasicBlock::Create(C, "while.entry", &F); 130 auto *ExitBB = BasicBlock::Create(C, "while.end", &F); 131 Type *PtrTy = IRB.getPtrTy(llvm::ADDRESS_SPACE_GLOBAL); 132 133 auto *Begin = M.getOrInsertGlobal( 134 IsCtor ? "__init_array_start" : "__fini_array_start", 135 PointerType::get(C, 0), [&]() { 136 auto *GV = new GlobalVariable( 137 M, PointerType::get(C, 0), 138 /*isConstant=*/false, GlobalValue::WeakAnyLinkage, 139 Constant::getNullValue(PointerType::get(C, 0)), 140 IsCtor ? "__init_array_start" : "__fini_array_start", 141 /*InsertBefore=*/nullptr, GlobalVariable::NotThreadLocal, 142 /*AddressSpace=*/llvm::ADDRESS_SPACE_GLOBAL); 143 GV->setVisibility(GlobalVariable::ProtectedVisibility); 144 return GV; 145 }); 146 auto *End = M.getOrInsertGlobal( 147 IsCtor ? "__init_array_end" : "__fini_array_end", PointerType::get(C, 0), 148 [&]() { 149 auto *GV = new GlobalVariable( 150 M, PointerType::get(C, 0), 151 /*isConstant=*/false, GlobalValue::WeakAnyLinkage, 152 Constant::getNullValue(PointerType::get(C, 0)), 153 IsCtor ? "__init_array_end" : "__fini_array_end", 154 /*InsertBefore=*/nullptr, GlobalVariable::NotThreadLocal, 155 /*AddressSpace=*/llvm::ADDRESS_SPACE_GLOBAL); 156 GV->setVisibility(GlobalVariable::ProtectedVisibility); 157 return GV; 158 }); 159 160 // The constructor type is suppoed to allow using the argument vectors, but 161 // for now we just call them with no arguments. 162 auto *CallBackTy = FunctionType::get(IRB.getVoidTy(), {}); 163 164 // The destructor array must be called in reverse order. Get an expression to 165 // the end of the array and iterate backwards in that case. 166 Value *BeginVal = IRB.CreateLoad(Begin->getType(), Begin, "begin"); 167 Value *EndVal = IRB.CreateLoad(Begin->getType(), End, "stop"); 168 if (!IsCtor) { 169 auto *BeginInt = IRB.CreatePtrToInt(BeginVal, IntegerType::getInt64Ty(C)); 170 auto *EndInt = IRB.CreatePtrToInt(EndVal, IntegerType::getInt64Ty(C)); 171 auto *SubInst = IRB.CreateSub(EndInt, BeginInt); 172 auto *Offset = IRB.CreateAShr( 173 SubInst, ConstantInt::get(IntegerType::getInt64Ty(C), 3), "offset", 174 /*IsExact=*/true); 175 auto *ValuePtr = IRB.CreateGEP(PointerType::get(C, 0), BeginVal, 176 ArrayRef<Value *>({Offset})); 177 EndVal = BeginVal; 178 BeginVal = IRB.CreateInBoundsGEP( 179 PointerType::get(C, 0), ValuePtr, 180 ArrayRef<Value *>(ConstantInt::get(IntegerType::getInt64Ty(C), -1)), 181 "start"); 182 } 183 IRB.CreateCondBr( 184 IRB.CreateCmp(IsCtor ? ICmpInst::ICMP_NE : ICmpInst::ICMP_UGT, BeginVal, 185 EndVal), 186 LoopBB, ExitBB); 187 IRB.SetInsertPoint(LoopBB); 188 auto *CallBackPHI = IRB.CreatePHI(PtrTy, 2, "ptr"); 189 auto *CallBack = IRB.CreateLoad(IRB.getPtrTy(F.getAddressSpace()), 190 CallBackPHI, "callback"); 191 IRB.CreateCall(CallBackTy, CallBack); 192 auto *NewCallBack = 193 IRB.CreateConstGEP1_64(PtrTy, CallBackPHI, IsCtor ? 1 : -1, "next"); 194 auto *EndCmp = IRB.CreateCmp(IsCtor ? ICmpInst::ICMP_EQ : ICmpInst::ICMP_ULT, 195 NewCallBack, EndVal, "end"); 196 CallBackPHI->addIncoming(BeginVal, &F.getEntryBlock()); 197 CallBackPHI->addIncoming(NewCallBack, LoopBB); 198 IRB.CreateCondBr(EndCmp, ExitBB, LoopBB); 199 IRB.SetInsertPoint(ExitBB); 200 IRB.CreateRetVoid(); 201 } 202 203 static bool createInitOrFiniGlobals(Module &M, GlobalVariable *GV, 204 bool IsCtor) { 205 ConstantArray *GA = dyn_cast<ConstantArray>(GV->getInitializer()); 206 if (!GA || GA->getNumOperands() == 0) 207 return false; 208 209 // NVPTX has no way to emit variables at specific sections or support for 210 // the traditional constructor sections. Instead, we emit mangled global 211 // names so the runtime can build the list manually. 212 for (Value *V : GA->operands()) { 213 auto *CS = cast<ConstantStruct>(V); 214 auto *F = cast<Constant>(CS->getOperand(1)); 215 uint64_t Priority = cast<ConstantInt>(CS->getOperand(0))->getSExtValue(); 216 std::string PriorityStr = "." + std::to_string(Priority); 217 // We append a semi-unique hash and the priority to the global name. 218 std::string GlobalID = 219 !GlobalStr.empty() ? GlobalStr : getHash(M.getSourceFileName()); 220 std::string NameStr = 221 ((IsCtor ? "__init_array_object_" : "__fini_array_object_") + 222 F->getName() + "_" + GlobalID + "_" + std::to_string(Priority)) 223 .str(); 224 // PTX does not support exported names with '.' in them. 225 llvm::transform(NameStr, NameStr.begin(), 226 [](char c) { return c == '.' ? '_' : c; }); 227 228 auto *GV = new GlobalVariable(M, F->getType(), /*IsConstant=*/true, 229 GlobalValue::ExternalLinkage, F, NameStr, 230 nullptr, GlobalValue::NotThreadLocal, 231 /*AddressSpace=*/4); 232 // This isn't respected by Nvidia, simply put here for clarity. 233 GV->setSection(IsCtor ? ".init_array" + PriorityStr 234 : ".fini_array" + PriorityStr); 235 GV->setVisibility(GlobalVariable::ProtectedVisibility); 236 appendToUsed(M, {GV}); 237 } 238 239 return true; 240 } 241 242 static bool createInitOrFiniKernel(Module &M, StringRef GlobalName, 243 bool IsCtor) { 244 GlobalVariable *GV = M.getGlobalVariable(GlobalName); 245 if (!GV || !GV->hasInitializer()) 246 return false; 247 248 if (!createInitOrFiniGlobals(M, GV, IsCtor)) 249 return false; 250 251 if (!CreateKernels) 252 return true; 253 254 Function *InitOrFiniKernel = createInitOrFiniKernelFunction(M, IsCtor); 255 if (!InitOrFiniKernel) 256 return false; 257 258 createInitOrFiniCalls(*InitOrFiniKernel, IsCtor); 259 260 GV->eraseFromParent(); 261 return true; 262 } 263 264 static bool lowerCtorsAndDtors(Module &M) { 265 bool Modified = false; 266 Modified |= createInitOrFiniKernel(M, "llvm.global_ctors", /*IsCtor =*/true); 267 Modified |= createInitOrFiniKernel(M, "llvm.global_dtors", /*IsCtor =*/false); 268 return Modified; 269 } 270 271 class NVPTXCtorDtorLoweringLegacy final : public ModulePass { 272 public: 273 static char ID; 274 NVPTXCtorDtorLoweringLegacy() : ModulePass(ID) {} 275 bool runOnModule(Module &M) override { return lowerCtorsAndDtors(M); } 276 }; 277 278 } // End anonymous namespace 279 280 PreservedAnalyses NVPTXCtorDtorLoweringPass::run(Module &M, 281 ModuleAnalysisManager &AM) { 282 return lowerCtorsAndDtors(M) ? PreservedAnalyses::none() 283 : PreservedAnalyses::all(); 284 } 285 286 char NVPTXCtorDtorLoweringLegacy::ID = 0; 287 char &llvm::NVPTXCtorDtorLoweringLegacyPassID = NVPTXCtorDtorLoweringLegacy::ID; 288 INITIALIZE_PASS(NVPTXCtorDtorLoweringLegacy, DEBUG_TYPE, 289 "Lower ctors and dtors for NVPTX", false, false) 290 291 ModulePass *llvm::createNVPTXCtorDtorLoweringLegacyPass() { 292 return new NVPTXCtorDtorLoweringLegacy(); 293 } 294