1 //===- ShadowStackGCLowering.cpp - Custom lowering for shadow-stack gc ----===// 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 file contains the custom lowering code required by the shadow-stack GC 10 // strategy. 11 // 12 // This pass implements the code transformation described in this paper: 13 // "Accurate Garbage Collection in an Uncooperative Environment" 14 // Fergus Henderson, ISMM, 2002 15 // 16 //===----------------------------------------------------------------------===// 17 18 #include "llvm/CodeGen/ShadowStackGCLowering.h" 19 #include "llvm/ADT/SmallVector.h" 20 #include "llvm/ADT/StringExtras.h" 21 #include "llvm/Analysis/DomTreeUpdater.h" 22 #include "llvm/CodeGen/GCMetadata.h" 23 #include "llvm/CodeGen/Passes.h" 24 #include "llvm/IR/BasicBlock.h" 25 #include "llvm/IR/Constant.h" 26 #include "llvm/IR/Constants.h" 27 #include "llvm/IR/DerivedTypes.h" 28 #include "llvm/IR/Dominators.h" 29 #include "llvm/IR/Function.h" 30 #include "llvm/IR/GlobalValue.h" 31 #include "llvm/IR/GlobalVariable.h" 32 #include "llvm/IR/IRBuilder.h" 33 #include "llvm/IR/Instructions.h" 34 #include "llvm/IR/IntrinsicInst.h" 35 #include "llvm/IR/Intrinsics.h" 36 #include "llvm/IR/Module.h" 37 #include "llvm/IR/Type.h" 38 #include "llvm/IR/Value.h" 39 #include "llvm/InitializePasses.h" 40 #include "llvm/Pass.h" 41 #include "llvm/Support/Casting.h" 42 #include "llvm/Transforms/Utils/EscapeEnumerator.h" 43 #include <cassert> 44 #include <optional> 45 #include <string> 46 #include <utility> 47 #include <vector> 48 49 using namespace llvm; 50 51 #define DEBUG_TYPE "shadow-stack-gc-lowering" 52 53 namespace { 54 55 class ShadowStackGCLoweringImpl { 56 /// RootChain - This is the global linked-list that contains the chain of GC 57 /// roots. 58 GlobalVariable *Head = nullptr; 59 60 /// StackEntryTy - Abstract type of a link in the shadow stack. 61 StructType *StackEntryTy = nullptr; 62 StructType *FrameMapTy = nullptr; 63 64 /// Roots - GC roots in the current function. Each is a pair of the 65 /// intrinsic call and its corresponding alloca. 66 std::vector<std::pair<CallInst *, AllocaInst *>> Roots; 67 68 public: 69 ShadowStackGCLoweringImpl() = default; 70 71 bool doInitialization(Module &M); 72 bool runOnFunction(Function &F, DomTreeUpdater *DTU); 73 74 private: 75 bool IsNullValue(Value *V); 76 Constant *GetFrameMap(Function &F); 77 Type *GetConcreteStackEntryType(Function &F); 78 void CollectRoots(Function &F); 79 80 static GetElementPtrInst *CreateGEP(LLVMContext &Context, IRBuilder<> &B, 81 Type *Ty, Value *BasePtr, int Idx1, 82 const char *Name); 83 static GetElementPtrInst *CreateGEP(LLVMContext &Context, IRBuilder<> &B, 84 Type *Ty, Value *BasePtr, int Idx1, int Idx2, 85 const char *Name); 86 }; 87 88 class ShadowStackGCLowering : public FunctionPass { 89 ShadowStackGCLoweringImpl Impl; 90 91 public: 92 static char ID; 93 94 ShadowStackGCLowering(); 95 96 bool doInitialization(Module &M) override { return Impl.doInitialization(M); } 97 void getAnalysisUsage(AnalysisUsage &AU) const override { 98 AU.addPreserved<DominatorTreeWrapperPass>(); 99 } 100 bool runOnFunction(Function &F) override { 101 std::optional<DomTreeUpdater> DTU; 102 if (auto *DTWP = getAnalysisIfAvailable<DominatorTreeWrapperPass>()) 103 DTU.emplace(DTWP->getDomTree(), DomTreeUpdater::UpdateStrategy::Lazy); 104 return Impl.runOnFunction(F, DTU ? &*DTU : nullptr); 105 } 106 }; 107 108 } // end anonymous namespace 109 110 PreservedAnalyses ShadowStackGCLoweringPass::run(Module &M, 111 ModuleAnalysisManager &MAM) { 112 auto &Map = MAM.getResult<CollectorMetadataAnalysis>(M); 113 if (Map.StrategyMap.contains("shadow-stack")) 114 return PreservedAnalyses::all(); 115 116 ShadowStackGCLoweringImpl Impl; 117 bool Changed = Impl.doInitialization(M); 118 for (auto &F : M) { 119 auto &FAM = 120 MAM.getResult<FunctionAnalysisManagerModuleProxy>(M).getManager(); 121 auto *DT = FAM.getCachedResult<DominatorTreeAnalysis>(F); 122 DomTreeUpdater DTU(DT, DomTreeUpdater::UpdateStrategy::Lazy); 123 Changed |= Impl.runOnFunction(F, DT ? &DTU : nullptr); 124 } 125 126 if (!Changed) 127 return PreservedAnalyses::all(); 128 PreservedAnalyses PA; 129 PA.preserve<DominatorTreeAnalysis>(); 130 return PA; 131 } 132 133 char ShadowStackGCLowering::ID = 0; 134 char &llvm::ShadowStackGCLoweringID = ShadowStackGCLowering::ID; 135 136 INITIALIZE_PASS_BEGIN(ShadowStackGCLowering, DEBUG_TYPE, 137 "Shadow Stack GC Lowering", false, false) 138 INITIALIZE_PASS_DEPENDENCY(GCModuleInfo) 139 INITIALIZE_PASS_DEPENDENCY(DominatorTreeWrapperPass) 140 INITIALIZE_PASS_END(ShadowStackGCLowering, DEBUG_TYPE, 141 "Shadow Stack GC Lowering", false, false) 142 143 FunctionPass *llvm::createShadowStackGCLoweringPass() { return new ShadowStackGCLowering(); } 144 145 ShadowStackGCLowering::ShadowStackGCLowering() : FunctionPass(ID) { 146 initializeShadowStackGCLoweringPass(*PassRegistry::getPassRegistry()); 147 } 148 149 Constant *ShadowStackGCLoweringImpl::GetFrameMap(Function &F) { 150 // doInitialization creates the abstract type of this value. 151 Type *VoidPtr = PointerType::getUnqual(F.getContext()); 152 153 // Truncate the ShadowStackDescriptor if some metadata is null. 154 unsigned NumMeta = 0; 155 SmallVector<Constant *, 16> Metadata; 156 for (unsigned I = 0; I != Roots.size(); ++I) { 157 Constant *C = cast<Constant>(Roots[I].first->getArgOperand(1)); 158 if (!C->isNullValue()) 159 NumMeta = I + 1; 160 Metadata.push_back(C); 161 } 162 Metadata.resize(NumMeta); 163 164 Type *Int32Ty = Type::getInt32Ty(F.getContext()); 165 166 Constant *BaseElts[] = { 167 ConstantInt::get(Int32Ty, Roots.size(), false), 168 ConstantInt::get(Int32Ty, NumMeta, false), 169 }; 170 171 Constant *DescriptorElts[] = { 172 ConstantStruct::get(FrameMapTy, BaseElts), 173 ConstantArray::get(ArrayType::get(VoidPtr, NumMeta), Metadata)}; 174 175 Type *EltTys[] = {DescriptorElts[0]->getType(), DescriptorElts[1]->getType()}; 176 StructType *STy = StructType::create(EltTys, "gc_map." + utostr(NumMeta)); 177 178 Constant *FrameMap = ConstantStruct::get(STy, DescriptorElts); 179 180 // FIXME: Is this actually dangerous as WritingAnLLVMPass.html claims? Seems 181 // that, short of multithreaded LLVM, it should be safe; all that is 182 // necessary is that a simple Module::iterator loop not be invalidated. 183 // Appending to the GlobalVariable list is safe in that sense. 184 // 185 // All of the output passes emit globals last. The ExecutionEngine 186 // explicitly supports adding globals to the module after 187 // initialization. 188 // 189 // Still, if it isn't deemed acceptable, then this transformation needs 190 // to be a ModulePass (which means it cannot be in the 'llc' pipeline 191 // (which uses a FunctionPassManager (which segfaults (not asserts) if 192 // provided a ModulePass))). 193 Constant *GV = new GlobalVariable(*F.getParent(), FrameMap->getType(), true, 194 GlobalVariable::InternalLinkage, FrameMap, 195 "__gc_" + F.getName()); 196 197 Constant *GEPIndices[2] = { 198 ConstantInt::get(Type::getInt32Ty(F.getContext()), 0), 199 ConstantInt::get(Type::getInt32Ty(F.getContext()), 0)}; 200 return ConstantExpr::getGetElementPtr(FrameMap->getType(), GV, GEPIndices); 201 } 202 203 Type *ShadowStackGCLoweringImpl::GetConcreteStackEntryType(Function &F) { 204 // doInitialization creates the generic version of this type. 205 std::vector<Type *> EltTys; 206 EltTys.push_back(StackEntryTy); 207 for (const std::pair<CallInst *, AllocaInst *> &Root : Roots) 208 EltTys.push_back(Root.second->getAllocatedType()); 209 210 return StructType::create(EltTys, ("gc_stackentry." + F.getName()).str()); 211 } 212 213 /// doInitialization - If this module uses the GC intrinsics, find them now. If 214 /// not, exit fast. 215 bool ShadowStackGCLoweringImpl::doInitialization(Module &M) { 216 bool Active = false; 217 for (Function &F : M) { 218 if (F.hasGC() && F.getGC() == "shadow-stack") { 219 Active = true; 220 break; 221 } 222 } 223 if (!Active) 224 return false; 225 226 // struct FrameMap { 227 // int32_t NumRoots; // Number of roots in stack frame. 228 // int32_t NumMeta; // Number of metadata descriptors. May be < NumRoots. 229 // void *Meta[]; // May be absent for roots without metadata. 230 // }; 231 std::vector<Type *> EltTys; 232 // 32 bits is ok up to a 32GB stack frame. :) 233 EltTys.push_back(Type::getInt32Ty(M.getContext())); 234 // Specifies length of variable length array. 235 EltTys.push_back(Type::getInt32Ty(M.getContext())); 236 FrameMapTy = StructType::create(EltTys, "gc_map"); 237 PointerType *FrameMapPtrTy = PointerType::getUnqual(FrameMapTy); 238 239 // struct StackEntry { 240 // ShadowStackEntry *Next; // Caller's stack entry. 241 // FrameMap *Map; // Pointer to constant FrameMap. 242 // void *Roots[]; // Stack roots (in-place array, so we pretend). 243 // }; 244 245 StackEntryTy = StructType::create(M.getContext(), "gc_stackentry"); 246 247 EltTys.clear(); 248 EltTys.push_back(PointerType::getUnqual(StackEntryTy)); 249 EltTys.push_back(FrameMapPtrTy); 250 StackEntryTy->setBody(EltTys); 251 PointerType *StackEntryPtrTy = PointerType::getUnqual(StackEntryTy); 252 253 // Get the root chain if it already exists. 254 Head = M.getGlobalVariable("llvm_gc_root_chain"); 255 if (!Head) { 256 // If the root chain does not exist, insert a new one with linkonce 257 // linkage! 258 Head = new GlobalVariable( 259 M, StackEntryPtrTy, false, GlobalValue::LinkOnceAnyLinkage, 260 Constant::getNullValue(StackEntryPtrTy), "llvm_gc_root_chain"); 261 } else if (Head->hasExternalLinkage() && Head->isDeclaration()) { 262 Head->setInitializer(Constant::getNullValue(StackEntryPtrTy)); 263 Head->setLinkage(GlobalValue::LinkOnceAnyLinkage); 264 } 265 266 return true; 267 } 268 269 bool ShadowStackGCLoweringImpl::IsNullValue(Value *V) { 270 if (Constant *C = dyn_cast<Constant>(V)) 271 return C->isNullValue(); 272 return false; 273 } 274 275 void ShadowStackGCLoweringImpl::CollectRoots(Function &F) { 276 // FIXME: Account for original alignment. Could fragment the root array. 277 // Approach 1: Null initialize empty slots at runtime. Yuck. 278 // Approach 2: Emit a map of the array instead of just a count. 279 280 assert(Roots.empty() && "Not cleaned up?"); 281 282 SmallVector<std::pair<CallInst *, AllocaInst *>, 16> MetaRoots; 283 284 for (BasicBlock &BB : F) 285 for (Instruction &I : BB) 286 if (IntrinsicInst *CI = dyn_cast<IntrinsicInst>(&I)) 287 if (Function *F = CI->getCalledFunction()) 288 if (F->getIntrinsicID() == Intrinsic::gcroot) { 289 std::pair<CallInst *, AllocaInst *> Pair = std::make_pair( 290 CI, 291 cast<AllocaInst>(CI->getArgOperand(0)->stripPointerCasts())); 292 if (IsNullValue(CI->getArgOperand(1))) 293 Roots.push_back(Pair); 294 else 295 MetaRoots.push_back(Pair); 296 } 297 298 // Number roots with metadata (usually empty) at the beginning, so that the 299 // FrameMap::Meta array can be elided. 300 Roots.insert(Roots.begin(), MetaRoots.begin(), MetaRoots.end()); 301 } 302 303 GetElementPtrInst * 304 ShadowStackGCLoweringImpl::CreateGEP(LLVMContext &Context, IRBuilder<> &B, 305 Type *Ty, Value *BasePtr, int Idx, 306 int Idx2, const char *Name) { 307 Value *Indices[] = {ConstantInt::get(Type::getInt32Ty(Context), 0), 308 ConstantInt::get(Type::getInt32Ty(Context), Idx), 309 ConstantInt::get(Type::getInt32Ty(Context), Idx2)}; 310 Value *Val = B.CreateGEP(Ty, BasePtr, Indices, Name); 311 312 assert(isa<GetElementPtrInst>(Val) && "Unexpected folded constant"); 313 314 return dyn_cast<GetElementPtrInst>(Val); 315 } 316 317 GetElementPtrInst *ShadowStackGCLoweringImpl::CreateGEP(LLVMContext &Context, 318 IRBuilder<> &B, 319 Type *Ty, 320 Value *BasePtr, int Idx, 321 const char *Name) { 322 Value *Indices[] = {ConstantInt::get(Type::getInt32Ty(Context), 0), 323 ConstantInt::get(Type::getInt32Ty(Context), Idx)}; 324 Value *Val = B.CreateGEP(Ty, BasePtr, Indices, Name); 325 326 assert(isa<GetElementPtrInst>(Val) && "Unexpected folded constant"); 327 328 return dyn_cast<GetElementPtrInst>(Val); 329 } 330 331 /// runOnFunction - Insert code to maintain the shadow stack. 332 bool ShadowStackGCLoweringImpl::runOnFunction(Function &F, 333 DomTreeUpdater *DTU) { 334 // Quick exit for functions that do not use the shadow stack GC. 335 if (!F.hasGC() || F.getGC() != "shadow-stack") 336 return false; 337 338 LLVMContext &Context = F.getContext(); 339 340 // Find calls to llvm.gcroot. 341 CollectRoots(F); 342 343 // If there are no roots in this function, then there is no need to add a 344 // stack map entry for it. 345 if (Roots.empty()) 346 return false; 347 348 // Build the constant map and figure the type of the shadow stack entry. 349 Value *FrameMap = GetFrameMap(F); 350 Type *ConcreteStackEntryTy = GetConcreteStackEntryType(F); 351 352 // Build the shadow stack entry at the very start of the function. 353 BasicBlock::iterator IP = F.getEntryBlock().begin(); 354 IRBuilder<> AtEntry(IP->getParent(), IP); 355 356 Instruction *StackEntry = 357 AtEntry.CreateAlloca(ConcreteStackEntryTy, nullptr, "gc_frame"); 358 359 AtEntry.SetInsertPointPastAllocas(&F); 360 IP = AtEntry.GetInsertPoint(); 361 362 // Initialize the map pointer and load the current head of the shadow stack. 363 Instruction *CurrentHead = 364 AtEntry.CreateLoad(AtEntry.getPtrTy(), Head, "gc_currhead"); 365 Instruction *EntryMapPtr = CreateGEP(Context, AtEntry, ConcreteStackEntryTy, 366 StackEntry, 0, 1, "gc_frame.map"); 367 AtEntry.CreateStore(FrameMap, EntryMapPtr); 368 369 // After all the allocas... 370 for (unsigned I = 0, E = Roots.size(); I != E; ++I) { 371 // For each root, find the corresponding slot in the aggregate... 372 Value *SlotPtr = CreateGEP(Context, AtEntry, ConcreteStackEntryTy, 373 StackEntry, 1 + I, "gc_root"); 374 375 // And use it in lieu of the alloca. 376 AllocaInst *OriginalAlloca = Roots[I].second; 377 SlotPtr->takeName(OriginalAlloca); 378 OriginalAlloca->replaceAllUsesWith(SlotPtr); 379 } 380 381 // Move past the original stores inserted by GCStrategy::InitRoots. This isn't 382 // really necessary (the collector would never see the intermediate state at 383 // runtime), but it's nicer not to push the half-initialized entry onto the 384 // shadow stack. 385 while (isa<StoreInst>(IP)) 386 ++IP; 387 AtEntry.SetInsertPoint(IP->getParent(), IP); 388 389 // Push the entry onto the shadow stack. 390 Instruction *EntryNextPtr = CreateGEP(Context, AtEntry, ConcreteStackEntryTy, 391 StackEntry, 0, 0, "gc_frame.next"); 392 Instruction *NewHeadVal = CreateGEP(Context, AtEntry, ConcreteStackEntryTy, 393 StackEntry, 0, "gc_newhead"); 394 AtEntry.CreateStore(CurrentHead, EntryNextPtr); 395 AtEntry.CreateStore(NewHeadVal, Head); 396 397 // For each instruction that escapes... 398 EscapeEnumerator EE(F, "gc_cleanup", /*HandleExceptions=*/true, DTU); 399 while (IRBuilder<> *AtExit = EE.Next()) { 400 // Pop the entry from the shadow stack. Don't reuse CurrentHead from 401 // AtEntry, since that would make the value live for the entire function. 402 Instruction *EntryNextPtr2 = 403 CreateGEP(Context, *AtExit, ConcreteStackEntryTy, StackEntry, 0, 0, 404 "gc_frame.next"); 405 Value *SavedHead = 406 AtExit->CreateLoad(AtExit->getPtrTy(), EntryNextPtr2, "gc_savedhead"); 407 AtExit->CreateStore(SavedHead, Head); 408 } 409 410 // Delete the original allocas (which are no longer used) and the intrinsic 411 // calls (which are no longer valid). Doing this last avoids invalidating 412 // iterators. 413 for (std::pair<CallInst *, AllocaInst *> &Root : Roots) { 414 Root.first->eraseFromParent(); 415 Root.second->eraseFromParent(); 416 } 417 418 Roots.clear(); 419 return true; 420 } 421