1 //===-- ShadowStackGC.cpp - GC support for uncooperative targets ----------===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
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
15 // This pass implements the code transformation described in this paper:
16 // "Accurate Garbage Collection in an Uncooperative Environment"
17 // Fergus Henderson, ISMM, 2002
19 // In runtime/GC/SemiSpace.cpp is a prototype runtime which is compatible with
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.
26 //===----------------------------------------------------------------------===//
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"
41 class VISIBILITY_HIDDEN ShadowStackGC : public GCStrategy {
42 /// RootChain - This is the global linked-list that contains the chain of GC
46 /// StackEntryTy - Abstract type of a link in the shadow stack.
48 const StructType *StackEntryTy;
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;
57 bool initializeCustomLowering(Module &M);
58 bool performCustomLowering(Function &F);
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);
73 static GCRegistry::Add<ShadowStackGC>
74 X("shadow-stack", "Very portable GC for uncooperative code generators");
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
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 {
87 const char *CleanupBBName;
91 Function::iterator StateBB, StateE;
95 EscapeEnumerator(Function &F, const char *N = "cleanup")
96 : F(F), CleanupBBName(N), State(0), Builder(*F.getContext()) {}
109 // Find all 'return' and 'unwind' instructions.
110 while (StateBB != StateE) {
111 BasicBlock *CurBB = StateBB++;
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))
118 Builder.SetInsertPoint(TI->getParent(), TI);
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())
138 // Create a cleanup block.
139 BasicBlock *CleanupBB = BasicBlock::Create(CleanupBBName, &F);
140 UnwindInst *UI = new UnwindInst(CleanupBB);
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]);
148 // Split the basic block containing the function call.
149 BasicBlock *CallBB = CI->getParent();
151 CallBB->splitBasicBlock(CI, CallBB->getName() + ".cont");
153 // Remove the unconditional branch inserted at the end of CallBB.
154 CallBB->getInstList().pop_back();
155 NewBB->getInstList().remove(CI);
157 // Create a new invoke instruction.
159 Args.append(CI->op_begin() + 1, CI->op_end());
161 InvokeInst *II = InvokeInst::Create(CI->getOperand(0),
163 Args.begin(), Args.end(),
164 CI->getName(), CallBB);
165 II->setCallingConv(CI->getCallingConv());
166 II->setAttributes(CI->getAttributes());
167 CI->replaceAllUsesWith(II);
171 Builder.SetInsertPoint(UI->getParent(), UI);
178 // -----------------------------------------------------------------------------
180 void llvm::linkShadowStackGC() { }
182 ShadowStackGC::ShadowStackGC() : Head(0), StackEntryTy(0) {
187 Constant *ShadowStackGC::GetFrameMap(Function &F) {
188 // doInitialization creates the abstract type of this value.
190 Type *VoidPtr = PointerType::getUnqual(Type::Int8Ty);
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())
199 Metadata.push_back(ConstantExpr::getBitCast(C, VoidPtr));
202 Constant *BaseElts[] = {
203 ConstantInt::get(Type::Int32Ty, Roots.size(), false),
204 ConstantInt::get(Type::Int32Ty, NumMeta, false),
207 Constant *DescriptorElts[] = {
208 ConstantStruct::get(BaseElts, 2),
209 ConstantArray::get(ArrayType::get(VoidPtr, NumMeta),
210 Metadata.begin(), NumMeta)
213 Constant *FrameMap = ConstantStruct::get(DescriptorElts, 2);
215 std::string TypeName("gc_map.");
216 TypeName += utostr(NumMeta);
217 F.getParent()->addTypeName(TypeName, FrameMap->getType());
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.
224 // All of the output passes emit globals last. The ExecutionEngine
225 // explicitly supports adding globals to the module after
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(*F.getParent(), FrameMap->getType(), true,
233 GlobalVariable::InternalLinkage,
234 FrameMap, "__gc_" + F.getName());
236 Constant *GEPIndices[2] = { ConstantInt::get(Type::Int32Ty, 0),
237 ConstantInt::get(Type::Int32Ty, 0) };
238 return ConstantExpr::getGetElementPtr(GV, GEPIndices, 2);
241 const Type* ShadowStackGC::GetConcreteStackEntryType(Function &F) {
242 // doInitialization creates the generic version of this type.
243 std::vector<const Type*> EltTys;
244 EltTys.push_back(StackEntryTy);
245 for (size_t I = 0; I != Roots.size(); I++)
246 EltTys.push_back(Roots[I].second->getAllocatedType());
247 Type *Ty = StructType::get(EltTys);
249 std::string TypeName("gc_stackentry.");
250 TypeName += F.getName();
251 F.getParent()->addTypeName(TypeName, Ty);
256 /// doInitialization - If this module uses the GC intrinsics, find them now. If
258 bool ShadowStackGC::initializeCustomLowering(Module &M) {
260 // int32_t NumRoots; // Number of roots in stack frame.
261 // int32_t NumMeta; // Number of metadata descriptors. May be < NumRoots.
262 // void *Meta[]; // May be absent for roots without metadata.
264 std::vector<const Type*> EltTys;
265 EltTys.push_back(Type::Int32Ty); // 32 bits is ok up to a 32GB stack frame. :)
266 EltTys.push_back(Type::Int32Ty); // Specifies length of variable length array.
267 StructType *FrameMapTy = StructType::get(EltTys);
268 M.addTypeName("gc_map", FrameMapTy);
269 PointerType *FrameMapPtrTy = PointerType::getUnqual(FrameMapTy);
271 // struct StackEntry {
272 // ShadowStackEntry *Next; // Caller's stack entry.
273 // FrameMap *Map; // Pointer to constant FrameMap.
274 // void *Roots[]; // Stack roots (in-place array, so we pretend).
276 OpaqueType *RecursiveTy = OpaqueType::get();
279 EltTys.push_back(PointerType::getUnqual(RecursiveTy));
280 EltTys.push_back(FrameMapPtrTy);
281 PATypeHolder LinkTyH = StructType::get(EltTys);
283 RecursiveTy->refineAbstractTypeTo(LinkTyH.get());
284 StackEntryTy = cast<StructType>(LinkTyH.get());
285 const PointerType *StackEntryPtrTy = PointerType::getUnqual(StackEntryTy);
286 M.addTypeName("gc_stackentry", LinkTyH.get()); // FIXME: Is this safe from
289 // Get the root chain if it already exists.
290 Head = M.getGlobalVariable("llvm_gc_root_chain");
292 // If the root chain does not exist, insert a new one with linkonce
294 Head = new GlobalVariable(M, StackEntryPtrTy, false,
295 GlobalValue::LinkOnceAnyLinkage,
296 Constant::getNullValue(StackEntryPtrTy),
297 "llvm_gc_root_chain");
298 } else if (Head->hasExternalLinkage() && Head->isDeclaration()) {
299 Head->setInitializer(Constant::getNullValue(StackEntryPtrTy));
300 Head->setLinkage(GlobalValue::LinkOnceAnyLinkage);
306 bool ShadowStackGC::IsNullValue(Value *V) {
307 if (Constant *C = dyn_cast<Constant>(V))
308 return C->isNullValue();
312 void ShadowStackGC::CollectRoots(Function &F) {
313 // FIXME: Account for original alignment. Could fragment the root array.
314 // Approach 1: Null initialize empty slots at runtime. Yuck.
315 // Approach 2: Emit a map of the array instead of just a count.
317 assert(Roots.empty() && "Not cleaned up?");
319 SmallVector<std::pair<CallInst*,AllocaInst*>,16> MetaRoots;
321 for (Function::iterator BB = F.begin(), E = F.end(); BB != E; ++BB)
322 for (BasicBlock::iterator II = BB->begin(), E = BB->end(); II != E;)
323 if (IntrinsicInst *CI = dyn_cast<IntrinsicInst>(II++))
324 if (Function *F = CI->getCalledFunction())
325 if (F->getIntrinsicID() == Intrinsic::gcroot) {
326 std::pair<CallInst*,AllocaInst*> Pair = std::make_pair(
327 CI, cast<AllocaInst>(CI->getOperand(1)->stripPointerCasts()));
328 if (IsNullValue(CI->getOperand(2)))
329 Roots.push_back(Pair);
331 MetaRoots.push_back(Pair);
334 // Number roots with metadata (usually empty) at the beginning, so that the
335 // FrameMap::Meta array can be elided.
336 Roots.insert(Roots.begin(), MetaRoots.begin(), MetaRoots.end());
340 ShadowStackGC::CreateGEP(IRBuilder<> &B, Value *BasePtr,
341 int Idx, int Idx2, const char *Name) {
342 Value *Indices[] = { ConstantInt::get(Type::Int32Ty, 0),
343 ConstantInt::get(Type::Int32Ty, Idx),
344 ConstantInt::get(Type::Int32Ty, Idx2) };
345 Value* Val = B.CreateGEP(BasePtr, Indices, Indices + 3, Name);
347 assert(isa<GetElementPtrInst>(Val) && "Unexpected folded constant");
349 return dyn_cast<GetElementPtrInst>(Val);
353 ShadowStackGC::CreateGEP(IRBuilder<> &B, Value *BasePtr,
354 int Idx, const char *Name) {
355 Value *Indices[] = { ConstantInt::get(Type::Int32Ty, 0),
356 ConstantInt::get(Type::Int32Ty, Idx) };
357 Value *Val = B.CreateGEP(BasePtr, Indices, Indices + 2, Name);
359 assert(isa<GetElementPtrInst>(Val) && "Unexpected folded constant");
361 return dyn_cast<GetElementPtrInst>(Val);
364 /// runOnFunction - Insert code to maintain the shadow stack.
365 bool ShadowStackGC::performCustomLowering(Function &F) {
366 // Find calls to llvm.gcroot.
369 // If there are no roots in this function, then there is no need to add a
370 // stack map entry for it.
374 // Build the constant map and figure the type of the shadow stack entry.
375 Value *FrameMap = GetFrameMap(F);
376 const Type *ConcreteStackEntryTy = GetConcreteStackEntryType(F);
378 // Build the shadow stack entry at the very start of the function.
379 BasicBlock::iterator IP = F.getEntryBlock().begin();
380 IRBuilder<> AtEntry(IP->getParent(), IP);
382 Instruction *StackEntry = AtEntry.CreateAlloca(ConcreteStackEntryTy, 0,
385 while (isa<AllocaInst>(IP)) ++IP;
386 AtEntry.SetInsertPoint(IP->getParent(), IP);
388 // Initialize the map pointer and load the current head of the shadow stack.
389 Instruction *CurrentHead = AtEntry.CreateLoad(Head, "gc_currhead");
390 Instruction *EntryMapPtr = CreateGEP(AtEntry, StackEntry,0,1,"gc_frame.map");
391 AtEntry.CreateStore(FrameMap, EntryMapPtr);
393 // After all the allocas...
394 for (unsigned I = 0, E = Roots.size(); I != E; ++I) {
395 // For each root, find the corresponding slot in the aggregate...
396 Value *SlotPtr = CreateGEP(AtEntry, StackEntry, 1 + I, "gc_root");
398 // And use it in lieu of the alloca.
399 AllocaInst *OriginalAlloca = Roots[I].second;
400 SlotPtr->takeName(OriginalAlloca);
401 OriginalAlloca->replaceAllUsesWith(SlotPtr);
404 // Move past the original stores inserted by GCStrategy::InitRoots. This isn't
405 // really necessary (the collector would never see the intermediate state at
406 // runtime), but it's nicer not to push the half-initialized entry onto the
408 while (isa<StoreInst>(IP)) ++IP;
409 AtEntry.SetInsertPoint(IP->getParent(), IP);
411 // Push the entry onto the shadow stack.
412 Instruction *EntryNextPtr = CreateGEP(AtEntry,StackEntry,0,0,"gc_frame.next");
413 Instruction *NewHeadVal = CreateGEP(AtEntry,StackEntry, 0, "gc_newhead");
414 AtEntry.CreateStore(CurrentHead, EntryNextPtr);
415 AtEntry.CreateStore(NewHeadVal, Head);
417 // For each instruction that escapes...
418 EscapeEnumerator EE(F, "gc_cleanup");
419 while (IRBuilder<> *AtExit = EE.Next()) {
420 // Pop the entry from the shadow stack. Don't reuse CurrentHead from
421 // AtEntry, since that would make the value live for the entire function.
422 Instruction *EntryNextPtr2 = CreateGEP(*AtExit, StackEntry, 0, 0,
424 Value *SavedHead = AtExit->CreateLoad(EntryNextPtr2, "gc_savedhead");
425 AtExit->CreateStore(SavedHead, Head);
428 // Delete the original allocas (which are no longer used) and the intrinsic
429 // calls (which are no longer valid). Doing this last avoids invalidating
431 for (unsigned I = 0, E = Roots.size(); I != E; ++I) {
432 Roots[I].first->eraseFromParent();
433 Roots[I].second->eraseFromParent();