ShadowStackGC.cpp revision 667d4b8de6dea70195ff12ef39a4deebffa2f5c7
1//===-- ShadowStackGC.cpp - GC support for uncooperative targets ----------===// 2// 3// The LLVM Compiler Infrastructure 4// 5// This file is distributed under the University of Illinois Open Source 6// License. See LICENSE.TXT for details. 7// 8//===----------------------------------------------------------------------===// 9// 10// This file implements lowering for the llvm.gc* intrinsics for targets that do 11// not natively support them (which includes the C backend). Note that the code 12// generated is not quite as efficient as algorithms which generate stack maps 13// to identify roots. 14// 15// This pass implements the code transformation described in this paper: 16// "Accurate Garbage Collection in an Uncooperative Environment" 17// Fergus Henderson, ISMM, 2002 18// 19// In runtime/GC/SemiSpace.cpp is a prototype runtime which is compatible with 20// ShadowStackGC. 21// 22// In order to support this particular transformation, all stack roots are 23// coallocated in the stack. This allows a fully target-independent stack map 24// while introducing only minor runtime overhead. 25// 26//===----------------------------------------------------------------------===// 27 28#define DEBUG_TYPE "shadowstackgc" 29#include "llvm/CodeGen/GCs.h" 30#include "llvm/ADT/StringExtras.h" 31#include "llvm/CodeGen/GCStrategy.h" 32#include "llvm/IntrinsicInst.h" 33#include "llvm/Module.h" 34#include "llvm/Support/Compiler.h" 35#include "llvm/Support/IRBuilder.h" 36 37using namespace llvm; 38 39namespace { 40 41 class VISIBILITY_HIDDEN ShadowStackGC : public GCStrategy { 42 /// RootChain - This is the global linked-list that contains the chain of GC 43 /// roots. 44 GlobalVariable *Head; 45 46 /// StackEntryTy - Abstract type of a link in the shadow stack. 47 /// 48 const StructType *StackEntryTy; 49 50 /// Roots - GC roots in the current function. Each is a pair of the 51 /// intrinsic call and its corresponding alloca. 52 std::vector<std::pair<CallInst*,AllocaInst*> > Roots; 53 54 public: 55 ShadowStackGC(); 56 57 bool initializeCustomLowering(Module &M); 58 bool performCustomLowering(Function &F); 59 60 private: 61 bool IsNullValue(Value *V); 62 Constant *GetFrameMap(Function &F); 63 const Type* GetConcreteStackEntryType(Function &F); 64 void CollectRoots(Function &F); 65 static GetElementPtrInst *CreateGEP(IRBuilder<> &B, Value *BasePtr, 66 int Idx1, const char *Name); 67 static GetElementPtrInst *CreateGEP(IRBuilder<> &B, Value *BasePtr, 68 int Idx1, int Idx2, const char *Name); 69 }; 70 71} 72 73static GCRegistry::Add<ShadowStackGC> 74X("shadow-stack", "Very portable GC for uncooperative code generators"); 75 76namespace { 77 /// EscapeEnumerator - This is a little algorithm to find all escape points 78 /// from a function so that "finally"-style code can be inserted. In addition 79 /// to finding the existing return and unwind instructions, it also (if 80 /// necessary) transforms any call instructions into invokes and sends them to 81 /// a landing pad. 82 /// 83 /// It's wrapped up in a state machine using the same transform C# uses for 84 /// 'yield return' enumerators, This transform allows it to be non-allocating. 85 class VISIBILITY_HIDDEN EscapeEnumerator { 86 Function &F; 87 const char *CleanupBBName; 88 89 // State. 90 int State; 91 Function::iterator StateBB, StateE; 92 IRBuilder<> Builder; 93 94 public: 95 EscapeEnumerator(Function &F, const char *N = "cleanup") 96 : F(F), CleanupBBName(N), State(0) {} 97 98 IRBuilder<> *Next() { 99 switch (State) { 100 default: 101 return 0; 102 103 case 0: 104 StateBB = F.begin(); 105 StateE = F.end(); 106 State = 1; 107 108 case 1: 109 // Find all 'return' and 'unwind' instructions. 110 while (StateBB != StateE) { 111 BasicBlock *CurBB = StateBB++; 112 113 // Branches and invokes do not escape, only unwind and return do. 114 TerminatorInst *TI = CurBB->getTerminator(); 115 if (!isa<UnwindInst>(TI) && !isa<ReturnInst>(TI)) 116 continue; 117 118 Builder.SetInsertPoint(TI->getParent(), TI); 119 return &Builder; 120 } 121 122 State = 2; 123 124 // Find all 'call' instructions. 125 SmallVector<Instruction*,16> Calls; 126 for (Function::iterator BB = F.begin(), 127 E = F.end(); BB != E; ++BB) 128 for (BasicBlock::iterator II = BB->begin(), 129 EE = BB->end(); II != EE; ++II) 130 if (CallInst *CI = dyn_cast<CallInst>(II)) 131 if (!CI->getCalledFunction() || 132 !CI->getCalledFunction()->getIntrinsicID()) 133 Calls.push_back(CI); 134 135 if (Calls.empty()) 136 return 0; 137 138 // Create a cleanup block. 139 BasicBlock *CleanupBB = BasicBlock::Create(CleanupBBName, &F); 140 UnwindInst *UI = new UnwindInst(CleanupBB); 141 142 // Transform the 'call' instructions into 'invoke's branching to the 143 // cleanup block. Go in reverse order to make prettier BB names. 144 SmallVector<Value*,16> Args; 145 for (unsigned I = Calls.size(); I != 0; ) { 146 CallInst *CI = cast<CallInst>(Calls[--I]); 147 148 // Split the basic block containing the function call. 149 BasicBlock *CallBB = CI->getParent(); 150 BasicBlock *NewBB = 151 CallBB->splitBasicBlock(CI, CallBB->getName() + ".cont"); 152 153 // Remove the unconditional branch inserted at the end of CallBB. 154 CallBB->getInstList().pop_back(); 155 NewBB->getInstList().remove(CI); 156 157 // Create a new invoke instruction. 158 Args.clear(); 159 Args.append(CI->op_begin() + 1, CI->op_end()); 160 161 InvokeInst *II = InvokeInst::Create(CI->getOperand(0), 162 NewBB, CleanupBB, 163 Args.begin(), Args.end(), 164 CI->getName(), CallBB); 165 II->setCallingConv(CI->getCallingConv()); 166 II->setAttributes(CI->getAttributes()); 167 CI->replaceAllUsesWith(II); 168 delete CI; 169 } 170 171 Builder.SetInsertPoint(UI->getParent(), UI); 172 return &Builder; 173 } 174 } 175 }; 176} 177 178// ----------------------------------------------------------------------------- 179 180void llvm::linkShadowStackGC() { } 181 182ShadowStackGC::ShadowStackGC() : Head(0), StackEntryTy(0) { 183 InitRoots = true; 184 CustomRoots = true; 185} 186 187Constant *ShadowStackGC::GetFrameMap(Function &F) { 188 // doInitialization creates the abstract type of this value. 189 190 Type *VoidPtr = PointerType::getUnqual(Type::Int8Ty); 191 192 // Truncate the ShadowStackDescriptor if some metadata is null. 193 unsigned NumMeta = 0; 194 SmallVector<Constant*,16> Metadata; 195 for (unsigned I = 0; I != Roots.size(); ++I) { 196 Constant *C = cast<Constant>(Roots[I].first->getOperand(2)); 197 if (!C->isNullValue()) 198 NumMeta = I + 1; 199 Metadata.push_back(ConstantExpr::getBitCast(C, VoidPtr)); 200 } 201 202 Constant *BaseElts[] = { 203 ConstantInt::get(Type::Int32Ty, Roots.size(), false), 204 ConstantInt::get(Type::Int32Ty, NumMeta, false), 205 }; 206 207 Constant *DescriptorElts[] = { 208 ConstantStruct::get(BaseElts, 2), 209 ConstantArray::get(ArrayType::get(VoidPtr, NumMeta), 210 Metadata.begin(), NumMeta) 211 }; 212 213 Constant *FrameMap = ConstantStruct::get(DescriptorElts, 2); 214 215 std::string TypeName("gc_map."); 216 TypeName += utostr(NumMeta); 217 F.getParent()->addTypeName(TypeName, FrameMap->getType()); 218 219 // FIXME: Is this actually dangerous as WritingAnLLVMPass.html claims? Seems 220 // that, short of multithreaded LLVM, it should be safe; all that is 221 // necessary is that a simple Module::iterator loop not be invalidated. 222 // Appending to the GlobalVariable list is safe in that sense. 223 // 224 // All of the output passes emit globals last. The ExecutionEngine 225 // explicitly supports adding globals to the module after 226 // initialization. 227 // 228 // Still, if it isn't deemed acceptable, then this transformation needs 229 // to be a ModulePass (which means it cannot be in the 'llc' pipeline 230 // (which uses a FunctionPassManager (which segfaults (not asserts) if 231 // provided a ModulePass))). 232 Constant *GV = new GlobalVariable(FrameMap->getType(), true, 233 GlobalVariable::InternalLinkage, 234 FrameMap, "__gc_" + F.getName(), 235 F.getParent()); 236 237 Constant *GEPIndices[2] = { ConstantInt::get(Type::Int32Ty, 0), 238 ConstantInt::get(Type::Int32Ty, 0) }; 239 return ConstantExpr::getGetElementPtr(GV, GEPIndices, 2); 240} 241 242const Type* ShadowStackGC::GetConcreteStackEntryType(Function &F) { 243 // doInitialization creates the generic version of this type. 244 std::vector<const Type*> EltTys; 245 EltTys.push_back(StackEntryTy); 246 for (size_t I = 0; I != Roots.size(); I++) 247 EltTys.push_back(Roots[I].second->getAllocatedType()); 248 Type *Ty = StructType::get(EltTys); 249 250 std::string TypeName("gc_stackentry."); 251 TypeName += F.getName(); 252 F.getParent()->addTypeName(TypeName, Ty); 253 254 return Ty; 255} 256 257/// doInitialization - If this module uses the GC intrinsics, find them now. If 258/// not, exit fast. 259bool ShadowStackGC::initializeCustomLowering(Module &M) { 260 // struct FrameMap { 261 // int32_t NumRoots; // Number of roots in stack frame. 262 // int32_t NumMeta; // Number of metadata descriptors. May be < NumRoots. 263 // void *Meta[]; // May be absent for roots without metadata. 264 // }; 265 std::vector<const Type*> EltTys; 266 EltTys.push_back(Type::Int32Ty); // 32 bits is ok up to a 32GB stack frame. :) 267 EltTys.push_back(Type::Int32Ty); // Specifies length of variable length array. 268 StructType *FrameMapTy = StructType::get(EltTys); 269 M.addTypeName("gc_map", FrameMapTy); 270 PointerType *FrameMapPtrTy = PointerType::getUnqual(FrameMapTy); 271 272 // struct StackEntry { 273 // ShadowStackEntry *Next; // Caller's stack entry. 274 // FrameMap *Map; // Pointer to constant FrameMap. 275 // void *Roots[]; // Stack roots (in-place array, so we pretend). 276 // }; 277 OpaqueType *RecursiveTy = OpaqueType::get(); 278 279 EltTys.clear(); 280 EltTys.push_back(PointerType::getUnqual(RecursiveTy)); 281 EltTys.push_back(FrameMapPtrTy); 282 PATypeHolder LinkTyH = StructType::get(EltTys); 283 284 RecursiveTy->refineAbstractTypeTo(LinkTyH.get()); 285 StackEntryTy = cast<StructType>(LinkTyH.get()); 286 const PointerType *StackEntryPtrTy = PointerType::getUnqual(StackEntryTy); 287 M.addTypeName("gc_stackentry", LinkTyH.get()); // FIXME: Is this safe from 288 // a FunctionPass? 289 290 // Get the root chain if it already exists. 291 Head = M.getGlobalVariable("llvm_gc_root_chain"); 292 if (!Head) { 293 // If the root chain does not exist, insert a new one with linkonce 294 // linkage! 295 Head = new GlobalVariable(StackEntryPtrTy, false, 296 GlobalValue::LinkOnceAnyLinkage, 297 Constant::getNullValue(StackEntryPtrTy), 298 "llvm_gc_root_chain", &M); 299 } else if (Head->hasExternalLinkage() && Head->isDeclaration()) { 300 Head->setInitializer(Constant::getNullValue(StackEntryPtrTy)); 301 Head->setLinkage(GlobalValue::LinkOnceAnyLinkage); 302 } 303 304 return true; 305} 306 307bool ShadowStackGC::IsNullValue(Value *V) { 308 if (Constant *C = dyn_cast<Constant>(V)) 309 return C->isNullValue(); 310 return false; 311} 312 313void ShadowStackGC::CollectRoots(Function &F) { 314 // FIXME: Account for original alignment. Could fragment the root array. 315 // Approach 1: Null initialize empty slots at runtime. Yuck. 316 // Approach 2: Emit a map of the array instead of just a count. 317 318 assert(Roots.empty() && "Not cleaned up?"); 319 320 SmallVector<std::pair<CallInst*,AllocaInst*>,16> MetaRoots; 321 322 for (Function::iterator BB = F.begin(), E = F.end(); BB != E; ++BB) 323 for (BasicBlock::iterator II = BB->begin(), E = BB->end(); II != E;) 324 if (IntrinsicInst *CI = dyn_cast<IntrinsicInst>(II++)) 325 if (Function *F = CI->getCalledFunction()) 326 if (F->getIntrinsicID() == Intrinsic::gcroot) { 327 std::pair<CallInst*,AllocaInst*> Pair = std::make_pair( 328 CI, cast<AllocaInst>(CI->getOperand(1)->stripPointerCasts())); 329 if (IsNullValue(CI->getOperand(2))) 330 Roots.push_back(Pair); 331 else 332 MetaRoots.push_back(Pair); 333 } 334 335 // Number roots with metadata (usually empty) at the beginning, so that the 336 // FrameMap::Meta array can be elided. 337 Roots.insert(Roots.begin(), MetaRoots.begin(), MetaRoots.end()); 338} 339 340GetElementPtrInst * 341ShadowStackGC::CreateGEP(IRBuilder<> &B, Value *BasePtr, 342 int Idx, int Idx2, const char *Name) { 343 Value *Indices[] = { ConstantInt::get(Type::Int32Ty, 0), 344 ConstantInt::get(Type::Int32Ty, Idx), 345 ConstantInt::get(Type::Int32Ty, Idx2) }; 346 Value* Val = B.CreateGEP(BasePtr, Indices, Indices + 3, Name); 347 348 assert(isa<GetElementPtrInst>(Val) && "Unexpected folded constant"); 349 350 return dyn_cast<GetElementPtrInst>(Val); 351} 352 353GetElementPtrInst * 354ShadowStackGC::CreateGEP(IRBuilder<> &B, Value *BasePtr, 355 int Idx, const char *Name) { 356 Value *Indices[] = { ConstantInt::get(Type::Int32Ty, 0), 357 ConstantInt::get(Type::Int32Ty, Idx) }; 358 Value *Val = B.CreateGEP(BasePtr, Indices, Indices + 2, Name); 359 360 assert(isa<GetElementPtrInst>(Val) && "Unexpected folded constant"); 361 362 return dyn_cast<GetElementPtrInst>(Val); 363} 364 365/// runOnFunction - Insert code to maintain the shadow stack. 366bool ShadowStackGC::performCustomLowering(Function &F) { 367 // Find calls to llvm.gcroot. 368 CollectRoots(F); 369 370 // If there are no roots in this function, then there is no need to add a 371 // stack map entry for it. 372 if (Roots.empty()) 373 return false; 374 375 // Build the constant map and figure the type of the shadow stack entry. 376 Value *FrameMap = GetFrameMap(F); 377 const Type *ConcreteStackEntryTy = GetConcreteStackEntryType(F); 378 379 // Build the shadow stack entry at the very start of the function. 380 BasicBlock::iterator IP = F.getEntryBlock().begin(); 381 IRBuilder<> AtEntry(IP->getParent(), IP); 382 383 Instruction *StackEntry = AtEntry.CreateAlloca(ConcreteStackEntryTy, 0, 384 "gc_frame"); 385 386 while (isa<AllocaInst>(IP)) ++IP; 387 AtEntry.SetInsertPoint(IP->getParent(), IP); 388 389 // Initialize the map pointer and load the current head of the shadow stack. 390 Instruction *CurrentHead = AtEntry.CreateLoad(Head, "gc_currhead"); 391 Instruction *EntryMapPtr = CreateGEP(AtEntry, StackEntry,0,1,"gc_frame.map"); 392 AtEntry.CreateStore(FrameMap, EntryMapPtr); 393 394 // After all the allocas... 395 for (unsigned I = 0, E = Roots.size(); I != E; ++I) { 396 // For each root, find the corresponding slot in the aggregate... 397 Value *SlotPtr = CreateGEP(AtEntry, StackEntry, 1 + I, "gc_root"); 398 399 // And use it in lieu of the alloca. 400 AllocaInst *OriginalAlloca = Roots[I].second; 401 SlotPtr->takeName(OriginalAlloca); 402 OriginalAlloca->replaceAllUsesWith(SlotPtr); 403 } 404 405 // Move past the original stores inserted by GCStrategy::InitRoots. This isn't 406 // really necessary (the collector would never see the intermediate state at 407 // runtime), but it's nicer not to push the half-initialized entry onto the 408 // shadow stack. 409 while (isa<StoreInst>(IP)) ++IP; 410 AtEntry.SetInsertPoint(IP->getParent(), IP); 411 412 // Push the entry onto the shadow stack. 413 Instruction *EntryNextPtr = CreateGEP(AtEntry,StackEntry,0,0,"gc_frame.next"); 414 Instruction *NewHeadVal = CreateGEP(AtEntry,StackEntry, 0, "gc_newhead"); 415 AtEntry.CreateStore(CurrentHead, EntryNextPtr); 416 AtEntry.CreateStore(NewHeadVal, Head); 417 418 // For each instruction that escapes... 419 EscapeEnumerator EE(F, "gc_cleanup"); 420 while (IRBuilder<> *AtExit = EE.Next()) { 421 // Pop the entry from the shadow stack. Don't reuse CurrentHead from 422 // AtEntry, since that would make the value live for the entire function. 423 Instruction *EntryNextPtr2 = CreateGEP(*AtExit, StackEntry, 0, 0, 424 "gc_frame.next"); 425 Value *SavedHead = AtExit->CreateLoad(EntryNextPtr2, "gc_savedhead"); 426 AtExit->CreateStore(SavedHead, Head); 427 } 428 429 // Delete the original allocas (which are no longer used) and the intrinsic 430 // calls (which are no longer valid). Doing this last avoids invalidating 431 // iterators. 432 for (unsigned I = 0, E = Roots.size(); I != E; ++I) { 433 Roots[I].first->eraseFromParent(); 434 Roots[I].second->eraseFromParent(); 435 } 436 437 Roots.clear(); 438 return true; 439} 440