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(LLVMContext &Context,
66 IRBuilder<> &B, Value *BasePtr,
67 int Idx1, const char *Name);
68 static GetElementPtrInst *CreateGEP(LLVMContext &Context,
69 IRBuilder<> &B, Value *BasePtr,
70 int Idx1, int Idx2, const char *Name);
75 static GCRegistry::Add<ShadowStackGC>
76 X("shadow-stack", "Very portable GC for uncooperative code generators");
79 /// EscapeEnumerator - This is a little algorithm to find all escape points
80 /// from a function so that "finally"-style code can be inserted. In addition
81 /// to finding the existing return and unwind instructions, it also (if
82 /// necessary) transforms any call instructions into invokes and sends them to
85 /// It's wrapped up in a state machine using the same transform C# uses for
86 /// 'yield return' enumerators, This transform allows it to be non-allocating.
87 class VISIBILITY_HIDDEN EscapeEnumerator {
89 const char *CleanupBBName;
93 Function::iterator StateBB, StateE;
97 EscapeEnumerator(Function &F, const char *N = "cleanup")
98 : F(F), CleanupBBName(N), State(0), Builder(F.getContext()) {}
100 IRBuilder<> *Next() {
111 // Find all 'return' and 'unwind' instructions.
112 while (StateBB != StateE) {
113 BasicBlock *CurBB = StateBB++;
115 // Branches and invokes do not escape, only unwind and return do.
116 TerminatorInst *TI = CurBB->getTerminator();
117 if (!isa<UnwindInst>(TI) && !isa<ReturnInst>(TI))
120 Builder.SetInsertPoint(TI->getParent(), TI);
126 // Find all 'call' instructions.
127 SmallVector<Instruction*,16> Calls;
128 for (Function::iterator BB = F.begin(),
129 E = F.end(); BB != E; ++BB)
130 for (BasicBlock::iterator II = BB->begin(),
131 EE = BB->end(); II != EE; ++II)
132 if (CallInst *CI = dyn_cast<CallInst>(II))
133 if (!CI->getCalledFunction() ||
134 !CI->getCalledFunction()->getIntrinsicID())
140 // Create a cleanup block.
141 BasicBlock *CleanupBB = BasicBlock::Create(F.getContext(),
143 UnwindInst *UI = new UnwindInst(F.getContext(), CleanupBB);
145 // Transform the 'call' instructions into 'invoke's branching to the
146 // cleanup block. Go in reverse order to make prettier BB names.
147 SmallVector<Value*,16> Args;
148 for (unsigned I = Calls.size(); I != 0; ) {
149 CallInst *CI = cast<CallInst>(Calls[--I]);
151 // Split the basic block containing the function call.
152 BasicBlock *CallBB = CI->getParent();
154 CallBB->splitBasicBlock(CI, CallBB->getName() + ".cont");
156 // Remove the unconditional branch inserted at the end of CallBB.
157 CallBB->getInstList().pop_back();
158 NewBB->getInstList().remove(CI);
160 // Create a new invoke instruction.
162 Args.append(CI->op_begin() + 1, CI->op_end());
164 InvokeInst *II = InvokeInst::Create(CI->getOperand(0),
166 Args.begin(), Args.end(),
167 CI->getName(), CallBB);
168 II->setCallingConv(CI->getCallingConv());
169 II->setAttributes(CI->getAttributes());
170 CI->replaceAllUsesWith(II);
174 Builder.SetInsertPoint(UI->getParent(), UI);
181 // -----------------------------------------------------------------------------
183 void llvm::linkShadowStackGC() { }
185 ShadowStackGC::ShadowStackGC() : Head(0), StackEntryTy(0) {
190 Constant *ShadowStackGC::GetFrameMap(Function &F) {
191 // doInitialization creates the abstract type of this value.
192 const Type *VoidPtr = Type::getInt8PtrTy(F.getContext());
194 // Truncate the ShadowStackDescriptor if some metadata is null.
195 unsigned NumMeta = 0;
196 SmallVector<Constant*,16> Metadata;
197 for (unsigned I = 0; I != Roots.size(); ++I) {
198 Constant *C = cast<Constant>(Roots[I].first->getOperand(2));
199 if (!C->isNullValue())
201 Metadata.push_back(ConstantExpr::getBitCast(C, VoidPtr));
204 Constant *BaseElts[] = {
205 ConstantInt::get(Type::getInt32Ty(F.getContext()), Roots.size(), false),
206 ConstantInt::get(Type::getInt32Ty(F.getContext()), NumMeta, false),
209 Constant *DescriptorElts[] = {
210 ConstantStruct::get(F.getContext(), BaseElts, 2, false),
211 ConstantArray::get(ArrayType::get(VoidPtr, NumMeta),
212 Metadata.begin(), NumMeta)
215 Constant *FrameMap = ConstantStruct::get(F.getContext(), DescriptorElts, 2,
218 std::string TypeName("gc_map.");
219 TypeName += utostr(NumMeta);
220 F.getParent()->addTypeName(TypeName, FrameMap->getType());
222 // FIXME: Is this actually dangerous as WritingAnLLVMPass.html claims? Seems
223 // that, short of multithreaded LLVM, it should be safe; all that is
224 // necessary is that a simple Module::iterator loop not be invalidated.
225 // Appending to the GlobalVariable list is safe in that sense.
227 // All of the output passes emit globals last. The ExecutionEngine
228 // explicitly supports adding globals to the module after
231 // Still, if it isn't deemed acceptable, then this transformation needs
232 // to be a ModulePass (which means it cannot be in the 'llc' pipeline
233 // (which uses a FunctionPassManager (which segfaults (not asserts) if
234 // provided a ModulePass))).
235 Constant *GV = new GlobalVariable(*F.getParent(), FrameMap->getType(), true,
236 GlobalVariable::InternalLinkage,
237 FrameMap, "__gc_" + F.getName());
239 Constant *GEPIndices[2] = {
240 ConstantInt::get(Type::getInt32Ty(F.getContext()), 0),
241 ConstantInt::get(Type::getInt32Ty(F.getContext()), 0)
243 return ConstantExpr::getGetElementPtr(GV, GEPIndices, 2);
246 const Type* ShadowStackGC::GetConcreteStackEntryType(Function &F) {
247 // doInitialization creates the generic version of this type.
248 std::vector<const Type*> EltTys;
249 EltTys.push_back(StackEntryTy);
250 for (size_t I = 0; I != Roots.size(); I++)
251 EltTys.push_back(Roots[I].second->getAllocatedType());
252 Type *Ty = StructType::get(F.getContext(), EltTys);
254 std::string TypeName("gc_stackentry.");
255 TypeName += F.getName();
256 F.getParent()->addTypeName(TypeName, Ty);
261 /// doInitialization - If this module uses the GC intrinsics, find them now. If
263 bool ShadowStackGC::initializeCustomLowering(Module &M) {
265 // int32_t NumRoots; // Number of roots in stack frame.
266 // int32_t NumMeta; // Number of metadata descriptors. May be < NumRoots.
267 // void *Meta[]; // May be absent for roots without metadata.
269 std::vector<const Type*> EltTys;
270 // 32 bits is ok up to a 32GB stack frame. :)
271 EltTys.push_back(Type::getInt32Ty(M.getContext()));
272 // Specifies length of variable length array.
273 EltTys.push_back(Type::getInt32Ty(M.getContext()));
274 StructType *FrameMapTy = StructType::get(M.getContext(), EltTys);
275 M.addTypeName("gc_map", FrameMapTy);
276 PointerType *FrameMapPtrTy = PointerType::getUnqual(FrameMapTy);
278 // struct StackEntry {
279 // ShadowStackEntry *Next; // Caller's stack entry.
280 // FrameMap *Map; // Pointer to constant FrameMap.
281 // void *Roots[]; // Stack roots (in-place array, so we pretend).
283 OpaqueType *RecursiveTy = OpaqueType::get(M.getContext());
286 EltTys.push_back(PointerType::getUnqual(RecursiveTy));
287 EltTys.push_back(FrameMapPtrTy);
288 PATypeHolder LinkTyH = StructType::get(M.getContext(), EltTys);
290 RecursiveTy->refineAbstractTypeTo(LinkTyH.get());
291 StackEntryTy = cast<StructType>(LinkTyH.get());
292 const PointerType *StackEntryPtrTy = PointerType::getUnqual(StackEntryTy);
293 M.addTypeName("gc_stackentry", LinkTyH.get()); // FIXME: Is this safe from
296 // Get the root chain if it already exists.
297 Head = M.getGlobalVariable("llvm_gc_root_chain");
299 // If the root chain does not exist, insert a new one with linkonce
301 Head = new GlobalVariable(M, StackEntryPtrTy, false,
302 GlobalValue::LinkOnceAnyLinkage,
303 Constant::getNullValue(StackEntryPtrTy),
304 "llvm_gc_root_chain");
305 } else if (Head->hasExternalLinkage() && Head->isDeclaration()) {
306 Head->setInitializer(Constant::getNullValue(StackEntryPtrTy));
307 Head->setLinkage(GlobalValue::LinkOnceAnyLinkage);
313 bool ShadowStackGC::IsNullValue(Value *V) {
314 if (Constant *C = dyn_cast<Constant>(V))
315 return C->isNullValue();
319 void ShadowStackGC::CollectRoots(Function &F) {
320 // FIXME: Account for original alignment. Could fragment the root array.
321 // Approach 1: Null initialize empty slots at runtime. Yuck.
322 // Approach 2: Emit a map of the array instead of just a count.
324 assert(Roots.empty() && "Not cleaned up?");
326 SmallVector<std::pair<CallInst*,AllocaInst*>,16> MetaRoots;
328 for (Function::iterator BB = F.begin(), E = F.end(); BB != E; ++BB)
329 for (BasicBlock::iterator II = BB->begin(), E = BB->end(); II != E;)
330 if (IntrinsicInst *CI = dyn_cast<IntrinsicInst>(II++))
331 if (Function *F = CI->getCalledFunction())
332 if (F->getIntrinsicID() == Intrinsic::gcroot) {
333 std::pair<CallInst*,AllocaInst*> Pair = std::make_pair(
334 CI, cast<AllocaInst>(CI->getOperand(1)->stripPointerCasts()));
335 if (IsNullValue(CI->getOperand(2)))
336 Roots.push_back(Pair);
338 MetaRoots.push_back(Pair);
341 // Number roots with metadata (usually empty) at the beginning, so that the
342 // FrameMap::Meta array can be elided.
343 Roots.insert(Roots.begin(), MetaRoots.begin(), MetaRoots.end());
347 ShadowStackGC::CreateGEP(LLVMContext &Context, IRBuilder<> &B, Value *BasePtr,
348 int Idx, int Idx2, const char *Name) {
349 Value *Indices[] = { ConstantInt::get(Type::getInt32Ty(Context), 0),
350 ConstantInt::get(Type::getInt32Ty(Context), Idx),
351 ConstantInt::get(Type::getInt32Ty(Context), Idx2) };
352 Value* Val = B.CreateGEP(BasePtr, Indices, Indices + 3, Name);
354 assert(isa<GetElementPtrInst>(Val) && "Unexpected folded constant");
356 return dyn_cast<GetElementPtrInst>(Val);
360 ShadowStackGC::CreateGEP(LLVMContext &Context, IRBuilder<> &B, Value *BasePtr,
361 int Idx, const char *Name) {
362 Value *Indices[] = { ConstantInt::get(Type::getInt32Ty(Context), 0),
363 ConstantInt::get(Type::getInt32Ty(Context), Idx) };
364 Value *Val = B.CreateGEP(BasePtr, Indices, Indices + 2, Name);
366 assert(isa<GetElementPtrInst>(Val) && "Unexpected folded constant");
368 return dyn_cast<GetElementPtrInst>(Val);
371 /// runOnFunction - Insert code to maintain the shadow stack.
372 bool ShadowStackGC::performCustomLowering(Function &F) {
373 LLVMContext &Context = F.getContext();
375 // Find calls to llvm.gcroot.
378 // If there are no roots in this function, then there is no need to add a
379 // stack map entry for it.
383 // Build the constant map and figure the type of the shadow stack entry.
384 Value *FrameMap = GetFrameMap(F);
385 const Type *ConcreteStackEntryTy = GetConcreteStackEntryType(F);
387 // Build the shadow stack entry at the very start of the function.
388 BasicBlock::iterator IP = F.getEntryBlock().begin();
389 IRBuilder<> AtEntry(IP->getParent(), IP);
391 Instruction *StackEntry = AtEntry.CreateAlloca(ConcreteStackEntryTy, 0,
394 while (isa<AllocaInst>(IP)) ++IP;
395 AtEntry.SetInsertPoint(IP->getParent(), IP);
397 // Initialize the map pointer and load the current head of the shadow stack.
398 Instruction *CurrentHead = AtEntry.CreateLoad(Head, "gc_currhead");
399 Instruction *EntryMapPtr = CreateGEP(Context, AtEntry, StackEntry,
401 AtEntry.CreateStore(FrameMap, EntryMapPtr);
403 // After all the allocas...
404 for (unsigned I = 0, E = Roots.size(); I != E; ++I) {
405 // For each root, find the corresponding slot in the aggregate...
406 Value *SlotPtr = CreateGEP(Context, AtEntry, StackEntry, 1 + I, "gc_root");
408 // And use it in lieu of the alloca.
409 AllocaInst *OriginalAlloca = Roots[I].second;
410 SlotPtr->takeName(OriginalAlloca);
411 OriginalAlloca->replaceAllUsesWith(SlotPtr);
414 // Move past the original stores inserted by GCStrategy::InitRoots. This isn't
415 // really necessary (the collector would never see the intermediate state at
416 // runtime), but it's nicer not to push the half-initialized entry onto the
418 while (isa<StoreInst>(IP)) ++IP;
419 AtEntry.SetInsertPoint(IP->getParent(), IP);
421 // Push the entry onto the shadow stack.
422 Instruction *EntryNextPtr = CreateGEP(Context, AtEntry,
423 StackEntry,0,0,"gc_frame.next");
424 Instruction *NewHeadVal = CreateGEP(Context, AtEntry,
425 StackEntry, 0, "gc_newhead");
426 AtEntry.CreateStore(CurrentHead, EntryNextPtr);
427 AtEntry.CreateStore(NewHeadVal, Head);
429 // For each instruction that escapes...
430 EscapeEnumerator EE(F, "gc_cleanup");
431 while (IRBuilder<> *AtExit = EE.Next()) {
432 // Pop the entry from the shadow stack. Don't reuse CurrentHead from
433 // AtEntry, since that would make the value live for the entire function.
434 Instruction *EntryNextPtr2 = CreateGEP(Context, *AtExit, StackEntry, 0, 0,
436 Value *SavedHead = AtExit->CreateLoad(EntryNextPtr2, "gc_savedhead");
437 AtExit->CreateStore(SavedHead, Head);
440 // Delete the original allocas (which are no longer used) and the intrinsic
441 // calls (which are no longer valid). Doing this last avoids invalidating
443 for (unsigned I = 0, E = Roots.size(); I != E; ++I) {
444 Roots[I].first->eraseFromParent();
445 Roots[I].second->eraseFromParent();