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/CallSite.h"
35 #include "llvm/Support/IRBuilder.h"
41 class 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 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.
163 Args.append(CS.arg_begin(), CS.arg_end());
165 InvokeInst *II = InvokeInst::Create(CI->getCalledValue(),
167 Args.begin(), Args.end(),
168 CI->getName(), CallBB);
169 II->setCallingConv(CI->getCallingConv());
170 II->setAttributes(CI->getAttributes());
171 CI->replaceAllUsesWith(II);
175 Builder.SetInsertPoint(UI->getParent(), UI);
182 // -----------------------------------------------------------------------------
184 void llvm::linkShadowStackGC() { }
186 ShadowStackGC::ShadowStackGC() : Head(0), StackEntryTy(0) {
191 Constant *ShadowStackGC::GetFrameMap(Function &F) {
192 // doInitialization creates the abstract type of this value.
193 const Type *VoidPtr = Type::getInt8PtrTy(F.getContext());
195 // Truncate the ShadowStackDescriptor if some metadata is null.
196 unsigned NumMeta = 0;
197 SmallVector<Constant*, 16> Metadata;
198 for (unsigned I = 0; I != Roots.size(); ++I) {
199 Constant *C = cast<Constant>(Roots[I].first->getArgOperand(1));
200 if (!C->isNullValue())
202 Metadata.push_back(ConstantExpr::getBitCast(C, VoidPtr));
204 Metadata.resize(NumMeta);
206 const Type *Int32Ty = Type::getInt32Ty(F.getContext());
208 Constant *BaseElts[] = {
209 ConstantInt::get(Int32Ty, Roots.size(), false),
210 ConstantInt::get(Int32Ty, NumMeta, false),
213 Constant *DescriptorElts[] = {
214 ConstantStruct::get(StructType::get(Int32Ty, Int32Ty, NULL), BaseElts),
215 ConstantArray::get(ArrayType::get(VoidPtr, NumMeta), Metadata)
219 ConstantStruct::get(StructType::get(DescriptorElts[0]->getType(),
220 DescriptorElts[1]->getType(), NULL),
223 std::string TypeName("gc_map.");
224 TypeName += utostr(NumMeta);
225 F.getParent()->addTypeName(TypeName, FrameMap->getType());
227 // FIXME: Is this actually dangerous as WritingAnLLVMPass.html claims? Seems
228 // that, short of multithreaded LLVM, it should be safe; all that is
229 // necessary is that a simple Module::iterator loop not be invalidated.
230 // Appending to the GlobalVariable list is safe in that sense.
232 // All of the output passes emit globals last. The ExecutionEngine
233 // explicitly supports adding globals to the module after
236 // Still, if it isn't deemed acceptable, then this transformation needs
237 // to be a ModulePass (which means it cannot be in the 'llc' pipeline
238 // (which uses a FunctionPassManager (which segfaults (not asserts) if
239 // provided a ModulePass))).
240 Constant *GV = new GlobalVariable(*F.getParent(), FrameMap->getType(), true,
241 GlobalVariable::InternalLinkage,
242 FrameMap, "__gc_" + F.getName());
244 Constant *GEPIndices[2] = {
245 ConstantInt::get(Type::getInt32Ty(F.getContext()), 0),
246 ConstantInt::get(Type::getInt32Ty(F.getContext()), 0)
248 return ConstantExpr::getGetElementPtr(GV, GEPIndices, 2);
251 const Type* ShadowStackGC::GetConcreteStackEntryType(Function &F) {
252 // doInitialization creates the generic version of this type.
253 std::vector<const Type*> EltTys;
254 EltTys.push_back(StackEntryTy);
255 for (size_t I = 0; I != Roots.size(); I++)
256 EltTys.push_back(Roots[I].second->getAllocatedType());
257 Type *Ty = StructType::get(F.getContext(), EltTys);
259 std::string TypeName("gc_stackentry.");
260 TypeName += F.getName();
261 F.getParent()->addTypeName(TypeName, Ty);
266 /// doInitialization - If this module uses the GC intrinsics, find them now. If
268 bool ShadowStackGC::initializeCustomLowering(Module &M) {
270 // int32_t NumRoots; // Number of roots in stack frame.
271 // int32_t NumMeta; // Number of metadata descriptors. May be < NumRoots.
272 // void *Meta[]; // May be absent for roots without metadata.
274 std::vector<const Type*> EltTys;
275 // 32 bits is ok up to a 32GB stack frame. :)
276 EltTys.push_back(Type::getInt32Ty(M.getContext()));
277 // Specifies length of variable length array.
278 EltTys.push_back(Type::getInt32Ty(M.getContext()));
279 StructType *FrameMapTy = StructType::get(M.getContext(), EltTys);
280 M.addTypeName("gc_map", FrameMapTy);
281 PointerType *FrameMapPtrTy = PointerType::getUnqual(FrameMapTy);
283 // struct StackEntry {
284 // ShadowStackEntry *Next; // Caller's stack entry.
285 // FrameMap *Map; // Pointer to constant FrameMap.
286 // void *Roots[]; // Stack roots (in-place array, so we pretend).
288 OpaqueType *RecursiveTy = OpaqueType::get(M.getContext());
291 EltTys.push_back(PointerType::getUnqual(RecursiveTy));
292 EltTys.push_back(FrameMapPtrTy);
293 PATypeHolder LinkTyH = StructType::get(M.getContext(), EltTys);
295 RecursiveTy->refineAbstractTypeTo(LinkTyH.get());
296 StackEntryTy = cast<StructType>(LinkTyH.get());
297 const PointerType *StackEntryPtrTy = PointerType::getUnqual(StackEntryTy);
298 M.addTypeName("gc_stackentry", LinkTyH.get()); // FIXME: Is this safe from
301 // Get the root chain if it already exists.
302 Head = M.getGlobalVariable("llvm_gc_root_chain");
304 // If the root chain does not exist, insert a new one with linkonce
306 Head = new GlobalVariable(M, StackEntryPtrTy, false,
307 GlobalValue::LinkOnceAnyLinkage,
308 Constant::getNullValue(StackEntryPtrTy),
309 "llvm_gc_root_chain");
310 } else if (Head->hasExternalLinkage() && Head->isDeclaration()) {
311 Head->setInitializer(Constant::getNullValue(StackEntryPtrTy));
312 Head->setLinkage(GlobalValue::LinkOnceAnyLinkage);
318 bool ShadowStackGC::IsNullValue(Value *V) {
319 if (Constant *C = dyn_cast<Constant>(V))
320 return C->isNullValue();
324 void ShadowStackGC::CollectRoots(Function &F) {
325 // FIXME: Account for original alignment. Could fragment the root array.
326 // Approach 1: Null initialize empty slots at runtime. Yuck.
327 // Approach 2: Emit a map of the array instead of just a count.
329 assert(Roots.empty() && "Not cleaned up?");
331 SmallVector<std::pair<CallInst*, AllocaInst*>, 16> MetaRoots;
333 for (Function::iterator BB = F.begin(), E = F.end(); BB != E; ++BB)
334 for (BasicBlock::iterator II = BB->begin(), E = BB->end(); II != E;)
335 if (IntrinsicInst *CI = dyn_cast<IntrinsicInst>(II++))
336 if (Function *F = CI->getCalledFunction())
337 if (F->getIntrinsicID() == Intrinsic::gcroot) {
338 std::pair<CallInst*, AllocaInst*> Pair = std::make_pair(
339 CI, cast<AllocaInst>(CI->getArgOperand(0)->stripPointerCasts()));
340 if (IsNullValue(CI->getArgOperand(1)))
341 Roots.push_back(Pair);
343 MetaRoots.push_back(Pair);
346 // Number roots with metadata (usually empty) at the beginning, so that the
347 // FrameMap::Meta array can be elided.
348 Roots.insert(Roots.begin(), MetaRoots.begin(), MetaRoots.end());
352 ShadowStackGC::CreateGEP(LLVMContext &Context, IRBuilder<> &B, Value *BasePtr,
353 int Idx, int Idx2, const char *Name) {
354 Value *Indices[] = { ConstantInt::get(Type::getInt32Ty(Context), 0),
355 ConstantInt::get(Type::getInt32Ty(Context), Idx),
356 ConstantInt::get(Type::getInt32Ty(Context), Idx2) };
357 Value* Val = B.CreateGEP(BasePtr, Indices, Indices + 3, Name);
359 assert(isa<GetElementPtrInst>(Val) && "Unexpected folded constant");
361 return dyn_cast<GetElementPtrInst>(Val);
365 ShadowStackGC::CreateGEP(LLVMContext &Context, IRBuilder<> &B, Value *BasePtr,
366 int Idx, const char *Name) {
367 Value *Indices[] = { ConstantInt::get(Type::getInt32Ty(Context), 0),
368 ConstantInt::get(Type::getInt32Ty(Context), Idx) };
369 Value *Val = B.CreateGEP(BasePtr, Indices, Indices + 2, Name);
371 assert(isa<GetElementPtrInst>(Val) && "Unexpected folded constant");
373 return dyn_cast<GetElementPtrInst>(Val);
376 /// runOnFunction - Insert code to maintain the shadow stack.
377 bool ShadowStackGC::performCustomLowering(Function &F) {
378 LLVMContext &Context = F.getContext();
380 // Find calls to llvm.gcroot.
383 // If there are no roots in this function, then there is no need to add a
384 // stack map entry for it.
388 // Build the constant map and figure the type of the shadow stack entry.
389 Value *FrameMap = GetFrameMap(F);
390 const Type *ConcreteStackEntryTy = GetConcreteStackEntryType(F);
392 // Build the shadow stack entry at the very start of the function.
393 BasicBlock::iterator IP = F.getEntryBlock().begin();
394 IRBuilder<> AtEntry(IP->getParent(), IP);
396 Instruction *StackEntry = AtEntry.CreateAlloca(ConcreteStackEntryTy, 0,
399 while (isa<AllocaInst>(IP)) ++IP;
400 AtEntry.SetInsertPoint(IP->getParent(), IP);
402 // Initialize the map pointer and load the current head of the shadow stack.
403 Instruction *CurrentHead = AtEntry.CreateLoad(Head, "gc_currhead");
404 Instruction *EntryMapPtr = CreateGEP(Context, AtEntry, StackEntry,
406 AtEntry.CreateStore(FrameMap, EntryMapPtr);
408 // After all the allocas...
409 for (unsigned I = 0, E = Roots.size(); I != E; ++I) {
410 // For each root, find the corresponding slot in the aggregate...
411 Value *SlotPtr = CreateGEP(Context, AtEntry, StackEntry, 1 + I, "gc_root");
413 // And use it in lieu of the alloca.
414 AllocaInst *OriginalAlloca = Roots[I].second;
415 SlotPtr->takeName(OriginalAlloca);
416 OriginalAlloca->replaceAllUsesWith(SlotPtr);
419 // Move past the original stores inserted by GCStrategy::InitRoots. This isn't
420 // really necessary (the collector would never see the intermediate state at
421 // runtime), but it's nicer not to push the half-initialized entry onto the
423 while (isa<StoreInst>(IP)) ++IP;
424 AtEntry.SetInsertPoint(IP->getParent(), IP);
426 // Push the entry onto the shadow stack.
427 Instruction *EntryNextPtr = CreateGEP(Context, AtEntry,
428 StackEntry,0,0,"gc_frame.next");
429 Instruction *NewHeadVal = CreateGEP(Context, AtEntry,
430 StackEntry, 0, "gc_newhead");
431 AtEntry.CreateStore(CurrentHead, EntryNextPtr);
432 AtEntry.CreateStore(NewHeadVal, Head);
434 // For each instruction that escapes...
435 EscapeEnumerator EE(F, "gc_cleanup");
436 while (IRBuilder<> *AtExit = EE.Next()) {
437 // Pop the entry from the shadow stack. Don't reuse CurrentHead from
438 // AtEntry, since that would make the value live for the entire function.
439 Instruction *EntryNextPtr2 = CreateGEP(Context, *AtExit, StackEntry, 0, 0,
441 Value *SavedHead = AtExit->CreateLoad(EntryNextPtr2, "gc_savedhead");
442 AtExit->CreateStore(SavedHead, Head);
445 // Delete the original allocas (which are no longer used) and the intrinsic
446 // calls (which are no longer valid). Doing this last avoids invalidating
448 for (unsigned I = 0, E = Roots.size(); I != E; ++I) {
449 Roots[I].first->eraseFromParent();
450 Roots[I].second->eraseFromParent();