ShadowStackGC.cpp revision 9ee17208115482441953127615231c59a2f4d052
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/IRBuilder.h" 35 36using namespace llvm; 37 38namespace { 39 40 class ShadowStackGC : public GCStrategy { 41 /// RootChain - This is the global linked-list that contains the chain of GC 42 /// roots. 43 GlobalVariable *Head; 44 45 /// StackEntryTy - Abstract type of a link in the shadow stack. 46 /// 47 const StructType *StackEntryTy; 48 49 /// Roots - GC roots in the current function. Each is a pair of the 50 /// intrinsic call and its corresponding alloca. 51 std::vector<std::pair<CallInst*,AllocaInst*> > Roots; 52 53 public: 54 ShadowStackGC(); 55 56 bool initializeCustomLowering(Module &M); 57 bool performCustomLowering(Function &F); 58 59 private: 60 bool IsNullValue(Value *V); 61 Constant *GetFrameMap(Function &F); 62 const Type* GetConcreteStackEntryType(Function &F); 63 void CollectRoots(Function &F); 64 static GetElementPtrInst *CreateGEP(LLVMContext &Context, 65 IRBuilder<> &B, Value *BasePtr, 66 int Idx1, const char *Name); 67 static GetElementPtrInst *CreateGEP(LLVMContext &Context, 68 IRBuilder<> &B, Value *BasePtr, 69 int Idx1, int Idx2, const char *Name); 70 }; 71 72} 73 74static GCRegistry::Add<ShadowStackGC> 75X("shadow-stack", "Very portable GC for uncooperative code generators"); 76 77namespace { 78 /// EscapeEnumerator - This is a little algorithm to find all escape points 79 /// from a function so that "finally"-style code can be inserted. In addition 80 /// to finding the existing return and unwind instructions, it also (if 81 /// necessary) transforms any call instructions into invokes and sends them to 82 /// a landing pad. 83 /// 84 /// It's wrapped up in a state machine using the same transform C# uses for 85 /// 'yield return' enumerators, This transform allows it to be non-allocating. 86 class EscapeEnumerator { 87 Function &F; 88 const char *CleanupBBName; 89 90 // State. 91 int State; 92 Function::iterator StateBB, StateE; 93 IRBuilder<> Builder; 94 95 public: 96 EscapeEnumerator(Function &F, const char *N = "cleanup") 97 : F(F), CleanupBBName(N), State(0), Builder(F.getContext()) {} 98 99 IRBuilder<> *Next() { 100 switch (State) { 101 default: 102 return 0; 103 104 case 0: 105 StateBB = F.begin(); 106 StateE = F.end(); 107 State = 1; 108 109 case 1: 110 // Find all 'return' and 'unwind' instructions. 111 while (StateBB != StateE) { 112 BasicBlock *CurBB = StateBB++; 113 114 // Branches and invokes do not escape, only unwind and return do. 115 TerminatorInst *TI = CurBB->getTerminator(); 116 if (!isa<UnwindInst>(TI) && !isa<ReturnInst>(TI)) 117 continue; 118 119 Builder.SetInsertPoint(TI->getParent(), TI); 120 return &Builder; 121 } 122 123 State = 2; 124 125 // Find all 'call' instructions. 126 SmallVector<Instruction*,16> Calls; 127 for (Function::iterator BB = F.begin(), 128 E = F.end(); BB != E; ++BB) 129 for (BasicBlock::iterator II = BB->begin(), 130 EE = BB->end(); II != EE; ++II) 131 if (CallInst *CI = dyn_cast<CallInst>(II)) 132 if (!CI->getCalledFunction() || 133 !CI->getCalledFunction()->getIntrinsicID()) 134 Calls.push_back(CI); 135 136 if (Calls.empty()) 137 return 0; 138 139 // Create a cleanup block. 140 BasicBlock *CleanupBB = BasicBlock::Create(F.getContext(), 141 CleanupBBName, &F); 142 UnwindInst *UI = new UnwindInst(F.getContext(), CleanupBB); 143 144 // Transform the 'call' instructions into 'invoke's branching to the 145 // cleanup block. Go in reverse order to make prettier BB names. 146 SmallVector<Value*,16> Args; 147 for (unsigned I = Calls.size(); I != 0; ) { 148 CallInst *CI = cast<CallInst>(Calls[--I]); 149 150 // Split the basic block containing the function call. 151 BasicBlock *CallBB = CI->getParent(); 152 BasicBlock *NewBB = 153 CallBB->splitBasicBlock(CI, CallBB->getName() + ".cont"); 154 155 // Remove the unconditional branch inserted at the end of CallBB. 156 CallBB->getInstList().pop_back(); 157 NewBB->getInstList().remove(CI); 158 159 // Create a new invoke instruction. 160 Args.clear(); 161 Args.append(CI->op_begin() + 1, CI->op_end()); 162 163 InvokeInst *II = InvokeInst::Create(CI->getOperand(0), 164 NewBB, CleanupBB, 165 Args.begin(), Args.end(), 166 CI->getName(), CallBB); 167 II->setCallingConv(CI->getCallingConv()); 168 II->setAttributes(CI->getAttributes()); 169 CI->replaceAllUsesWith(II); 170 delete CI; 171 } 172 173 Builder.SetInsertPoint(UI->getParent(), UI); 174 return &Builder; 175 } 176 } 177 }; 178} 179 180// ----------------------------------------------------------------------------- 181 182void llvm::linkShadowStackGC() { } 183 184ShadowStackGC::ShadowStackGC() : Head(0), StackEntryTy(0) { 185 InitRoots = true; 186 CustomRoots = true; 187} 188 189Constant *ShadowStackGC::GetFrameMap(Function &F) { 190 // doInitialization creates the abstract type of this value. 191 const Type *VoidPtr = Type::getInt8PtrTy(F.getContext()); 192 193 // Truncate the ShadowStackDescriptor if some metadata is null. 194 unsigned NumMeta = 0; 195 SmallVector<Constant*,16> Metadata; 196 for (unsigned I = 0; I != Roots.size(); ++I) { 197 Constant *C = cast<Constant>(Roots[I].first->getOperand(2)); 198 if (!C->isNullValue()) 199 NumMeta = I + 1; 200 Metadata.push_back(ConstantExpr::getBitCast(C, VoidPtr)); 201 } 202 203 Constant *BaseElts[] = { 204 ConstantInt::get(Type::getInt32Ty(F.getContext()), Roots.size(), false), 205 ConstantInt::get(Type::getInt32Ty(F.getContext()), NumMeta, false), 206 }; 207 208 Constant *DescriptorElts[] = { 209 ConstantStruct::get(F.getContext(), BaseElts, 2, false), 210 ConstantArray::get(ArrayType::get(VoidPtr, NumMeta), 211 Metadata.begin(), NumMeta) 212 }; 213 214 Constant *FrameMap = ConstantStruct::get(F.getContext(), DescriptorElts, 2, 215 false); 216 217 std::string TypeName("gc_map."); 218 TypeName += utostr(NumMeta); 219 F.getParent()->addTypeName(TypeName, FrameMap->getType()); 220 221 // FIXME: Is this actually dangerous as WritingAnLLVMPass.html claims? Seems 222 // that, short of multithreaded LLVM, it should be safe; all that is 223 // necessary is that a simple Module::iterator loop not be invalidated. 224 // Appending to the GlobalVariable list is safe in that sense. 225 // 226 // All of the output passes emit globals last. The ExecutionEngine 227 // explicitly supports adding globals to the module after 228 // initialization. 229 // 230 // Still, if it isn't deemed acceptable, then this transformation needs 231 // to be a ModulePass (which means it cannot be in the 'llc' pipeline 232 // (which uses a FunctionPassManager (which segfaults (not asserts) if 233 // provided a ModulePass))). 234 Constant *GV = new GlobalVariable(*F.getParent(), FrameMap->getType(), true, 235 GlobalVariable::InternalLinkage, 236 FrameMap, "__gc_" + F.getName()); 237 238 Constant *GEPIndices[2] = { 239 ConstantInt::get(Type::getInt32Ty(F.getContext()), 0), 240 ConstantInt::get(Type::getInt32Ty(F.getContext()), 0) 241 }; 242 return ConstantExpr::getGetElementPtr(GV, GEPIndices, 2); 243} 244 245const Type* ShadowStackGC::GetConcreteStackEntryType(Function &F) { 246 // doInitialization creates the generic version of this type. 247 std::vector<const Type*> EltTys; 248 EltTys.push_back(StackEntryTy); 249 for (size_t I = 0; I != Roots.size(); I++) 250 EltTys.push_back(Roots[I].second->getAllocatedType()); 251 Type *Ty = StructType::get(F.getContext(), EltTys); 252 253 std::string TypeName("gc_stackentry."); 254 TypeName += F.getName(); 255 F.getParent()->addTypeName(TypeName, Ty); 256 257 return Ty; 258} 259 260/// doInitialization - If this module uses the GC intrinsics, find them now. If 261/// not, exit fast. 262bool ShadowStackGC::initializeCustomLowering(Module &M) { 263 // struct FrameMap { 264 // int32_t NumRoots; // Number of roots in stack frame. 265 // int32_t NumMeta; // Number of metadata descriptors. May be < NumRoots. 266 // void *Meta[]; // May be absent for roots without metadata. 267 // }; 268 std::vector<const Type*> EltTys; 269 // 32 bits is ok up to a 32GB stack frame. :) 270 EltTys.push_back(Type::getInt32Ty(M.getContext())); 271 // Specifies length of variable length array. 272 EltTys.push_back(Type::getInt32Ty(M.getContext())); 273 StructType *FrameMapTy = StructType::get(M.getContext(), EltTys); 274 M.addTypeName("gc_map", FrameMapTy); 275 PointerType *FrameMapPtrTy = PointerType::getUnqual(FrameMapTy); 276 277 // struct StackEntry { 278 // ShadowStackEntry *Next; // Caller's stack entry. 279 // FrameMap *Map; // Pointer to constant FrameMap. 280 // void *Roots[]; // Stack roots (in-place array, so we pretend). 281 // }; 282 OpaqueType *RecursiveTy = OpaqueType::get(M.getContext()); 283 284 EltTys.clear(); 285 EltTys.push_back(PointerType::getUnqual(RecursiveTy)); 286 EltTys.push_back(FrameMapPtrTy); 287 PATypeHolder LinkTyH = StructType::get(M.getContext(), EltTys); 288 289 RecursiveTy->refineAbstractTypeTo(LinkTyH.get()); 290 StackEntryTy = cast<StructType>(LinkTyH.get()); 291 const PointerType *StackEntryPtrTy = PointerType::getUnqual(StackEntryTy); 292 M.addTypeName("gc_stackentry", LinkTyH.get()); // FIXME: Is this safe from 293 // a FunctionPass? 294 295 // Get the root chain if it already exists. 296 Head = M.getGlobalVariable("llvm_gc_root_chain"); 297 if (!Head) { 298 // If the root chain does not exist, insert a new one with linkonce 299 // linkage! 300 Head = new GlobalVariable(M, StackEntryPtrTy, false, 301 GlobalValue::LinkOnceAnyLinkage, 302 Constant::getNullValue(StackEntryPtrTy), 303 "llvm_gc_root_chain"); 304 } else if (Head->hasExternalLinkage() && Head->isDeclaration()) { 305 Head->setInitializer(Constant::getNullValue(StackEntryPtrTy)); 306 Head->setLinkage(GlobalValue::LinkOnceAnyLinkage); 307 } 308 309 return true; 310} 311 312bool ShadowStackGC::IsNullValue(Value *V) { 313 if (Constant *C = dyn_cast<Constant>(V)) 314 return C->isNullValue(); 315 return false; 316} 317 318void ShadowStackGC::CollectRoots(Function &F) { 319 // FIXME: Account for original alignment. Could fragment the root array. 320 // Approach 1: Null initialize empty slots at runtime. Yuck. 321 // Approach 2: Emit a map of the array instead of just a count. 322 323 assert(Roots.empty() && "Not cleaned up?"); 324 325 SmallVector<std::pair<CallInst*,AllocaInst*>,16> MetaRoots; 326 327 for (Function::iterator BB = F.begin(), E = F.end(); BB != E; ++BB) 328 for (BasicBlock::iterator II = BB->begin(), E = BB->end(); II != E;) 329 if (IntrinsicInst *CI = dyn_cast<IntrinsicInst>(II++)) 330 if (Function *F = CI->getCalledFunction()) 331 if (F->getIntrinsicID() == Intrinsic::gcroot) { 332 std::pair<CallInst*,AllocaInst*> Pair = std::make_pair( 333 CI, cast<AllocaInst>(CI->getOperand(1)->stripPointerCasts())); 334 if (IsNullValue(CI->getOperand(2))) 335 Roots.push_back(Pair); 336 else 337 MetaRoots.push_back(Pair); 338 } 339 340 // Number roots with metadata (usually empty) at the beginning, so that the 341 // FrameMap::Meta array can be elided. 342 Roots.insert(Roots.begin(), MetaRoots.begin(), MetaRoots.end()); 343} 344 345GetElementPtrInst * 346ShadowStackGC::CreateGEP(LLVMContext &Context, IRBuilder<> &B, Value *BasePtr, 347 int Idx, int Idx2, const char *Name) { 348 Value *Indices[] = { ConstantInt::get(Type::getInt32Ty(Context), 0), 349 ConstantInt::get(Type::getInt32Ty(Context), Idx), 350 ConstantInt::get(Type::getInt32Ty(Context), Idx2) }; 351 Value* Val = B.CreateGEP(BasePtr, Indices, Indices + 3, Name); 352 353 assert(isa<GetElementPtrInst>(Val) && "Unexpected folded constant"); 354 355 return dyn_cast<GetElementPtrInst>(Val); 356} 357 358GetElementPtrInst * 359ShadowStackGC::CreateGEP(LLVMContext &Context, IRBuilder<> &B, Value *BasePtr, 360 int Idx, const char *Name) { 361 Value *Indices[] = { ConstantInt::get(Type::getInt32Ty(Context), 0), 362 ConstantInt::get(Type::getInt32Ty(Context), Idx) }; 363 Value *Val = B.CreateGEP(BasePtr, Indices, Indices + 2, Name); 364 365 assert(isa<GetElementPtrInst>(Val) && "Unexpected folded constant"); 366 367 return dyn_cast<GetElementPtrInst>(Val); 368} 369 370/// runOnFunction - Insert code to maintain the shadow stack. 371bool ShadowStackGC::performCustomLowering(Function &F) { 372 LLVMContext &Context = F.getContext(); 373 374 // Find calls to llvm.gcroot. 375 CollectRoots(F); 376 377 // If there are no roots in this function, then there is no need to add a 378 // stack map entry for it. 379 if (Roots.empty()) 380 return false; 381 382 // Build the constant map and figure the type of the shadow stack entry. 383 Value *FrameMap = GetFrameMap(F); 384 const Type *ConcreteStackEntryTy = GetConcreteStackEntryType(F); 385 386 // Build the shadow stack entry at the very start of the function. 387 BasicBlock::iterator IP = F.getEntryBlock().begin(); 388 IRBuilder<> AtEntry(IP->getParent(), IP); 389 390 Instruction *StackEntry = AtEntry.CreateAlloca(ConcreteStackEntryTy, 0, 391 "gc_frame"); 392 393 while (isa<AllocaInst>(IP)) ++IP; 394 AtEntry.SetInsertPoint(IP->getParent(), IP); 395 396 // Initialize the map pointer and load the current head of the shadow stack. 397 Instruction *CurrentHead = AtEntry.CreateLoad(Head, "gc_currhead"); 398 Instruction *EntryMapPtr = CreateGEP(Context, AtEntry, StackEntry, 399 0,1,"gc_frame.map"); 400 AtEntry.CreateStore(FrameMap, EntryMapPtr); 401 402 // After all the allocas... 403 for (unsigned I = 0, E = Roots.size(); I != E; ++I) { 404 // For each root, find the corresponding slot in the aggregate... 405 Value *SlotPtr = CreateGEP(Context, AtEntry, StackEntry, 1 + I, "gc_root"); 406 407 // And use it in lieu of the alloca. 408 AllocaInst *OriginalAlloca = Roots[I].second; 409 SlotPtr->takeName(OriginalAlloca); 410 OriginalAlloca->replaceAllUsesWith(SlotPtr); 411 } 412 413 // Move past the original stores inserted by GCStrategy::InitRoots. This isn't 414 // really necessary (the collector would never see the intermediate state at 415 // runtime), but it's nicer not to push the half-initialized entry onto the 416 // shadow stack. 417 while (isa<StoreInst>(IP)) ++IP; 418 AtEntry.SetInsertPoint(IP->getParent(), IP); 419 420 // Push the entry onto the shadow stack. 421 Instruction *EntryNextPtr = CreateGEP(Context, AtEntry, 422 StackEntry,0,0,"gc_frame.next"); 423 Instruction *NewHeadVal = CreateGEP(Context, AtEntry, 424 StackEntry, 0, "gc_newhead"); 425 AtEntry.CreateStore(CurrentHead, EntryNextPtr); 426 AtEntry.CreateStore(NewHeadVal, Head); 427 428 // For each instruction that escapes... 429 EscapeEnumerator EE(F, "gc_cleanup"); 430 while (IRBuilder<> *AtExit = EE.Next()) { 431 // Pop the entry from the shadow stack. Don't reuse CurrentHead from 432 // AtEntry, since that would make the value live for the entire function. 433 Instruction *EntryNextPtr2 = CreateGEP(Context, *AtExit, StackEntry, 0, 0, 434 "gc_frame.next"); 435 Value *SavedHead = AtExit->CreateLoad(EntryNextPtr2, "gc_savedhead"); 436 AtExit->CreateStore(SavedHead, Head); 437 } 438 439 // Delete the original allocas (which are no longer used) and the intrinsic 440 // calls (which are no longer valid). Doing this last avoids invalidating 441 // iterators. 442 for (unsigned I = 0, E = Roots.size(); I != E; ++I) { 443 Roots[I].first->eraseFromParent(); 444 Roots[I].second->eraseFromParent(); 445 } 446 447 Roots.clear(); 448 return true; 449} 450