It turns out most of the thumb2 instructions are not allowed to touch SP. The semanti...
[llvm/avr.git] / lib / CodeGen / ShadowStackGC.cpp
blob514629bff73c35e14d8f4608ffce5f11fea5b640
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.
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
20 // ShadowStackGC.
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"
37 using namespace llvm;
39 namespace {
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;
46 /// StackEntryTy - Abstract type of a link in the shadow stack.
47 ///
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;
54 public:
55 ShadowStackGC();
57 bool initializeCustomLowering(Module &M);
58 bool performCustomLowering(Function &F);
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(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");
78 namespace {
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
83 /// a landing pad.
84 ///
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 {
88 Function &F;
89 const char *CleanupBBName;
91 // State.
92 int State;
93 Function::iterator StateBB, StateE;
94 IRBuilder<> Builder;
96 public:
97 EscapeEnumerator(Function &F, const char *N = "cleanup")
98 : F(F), CleanupBBName(N), State(0), Builder(F.getContext()) {}
100 IRBuilder<> *Next() {
101 switch (State) {
102 default:
103 return 0;
105 case 0:
106 StateBB = F.begin();
107 StateE = F.end();
108 State = 1;
110 case 1:
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))
118 continue;
120 Builder.SetInsertPoint(TI->getParent(), TI);
121 return &Builder;
124 State = 2;
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())
135 Calls.push_back(CI);
137 if (Calls.empty())
138 return 0;
140 // Create a cleanup block.
141 BasicBlock *CleanupBB = BasicBlock::Create(CleanupBBName, &F);
142 UnwindInst *UI = new UnwindInst(CleanupBB);
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]);
150 // Split the basic block containing the function call.
151 BasicBlock *CallBB = CI->getParent();
152 BasicBlock *NewBB =
153 CallBB->splitBasicBlock(CI, CallBB->getName() + ".cont");
155 // Remove the unconditional branch inserted at the end of CallBB.
156 CallBB->getInstList().pop_back();
157 NewBB->getInstList().remove(CI);
159 // Create a new invoke instruction.
160 Args.clear();
161 Args.append(CI->op_begin() + 1, CI->op_end());
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;
173 Builder.SetInsertPoint(UI->getParent(), UI);
174 return &Builder;
180 // -----------------------------------------------------------------------------
182 void llvm::linkShadowStackGC() { }
184 ShadowStackGC::ShadowStackGC() : Head(0), StackEntryTy(0) {
185 InitRoots = true;
186 CustomRoots = true;
189 Constant *ShadowStackGC::GetFrameMap(Function &F) {
190 // doInitialization creates the abstract type of this value.
191 Type *VoidPtr = PointerType::getUnqual(Type::Int8Ty);
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));
203 Constant *BaseElts[] = {
204 ConstantInt::get(Type::Int32Ty, Roots.size(), false),
205 ConstantInt::get(Type::Int32Ty, NumMeta, false),
208 Constant *DescriptorElts[] = {
209 ConstantStruct::get(F.getContext(), BaseElts, 2),
210 ConstantArray::get(ArrayType::get(VoidPtr, NumMeta),
211 Metadata.begin(), NumMeta)
214 Constant *FrameMap = ConstantStruct::get(F.getContext(), DescriptorElts, 2);
216 std::string TypeName("gc_map.");
217 TypeName += utostr(NumMeta);
218 F.getParent()->addTypeName(TypeName, FrameMap->getType());
220 // FIXME: Is this actually dangerous as WritingAnLLVMPass.html claims? Seems
221 // that, short of multithreaded LLVM, it should be safe; all that is
222 // necessary is that a simple Module::iterator loop not be invalidated.
223 // Appending to the GlobalVariable list is safe in that sense.
225 // All of the output passes emit globals last. The ExecutionEngine
226 // explicitly supports adding globals to the module after
227 // initialization.
229 // Still, if it isn't deemed acceptable, then this transformation needs
230 // to be a ModulePass (which means it cannot be in the 'llc' pipeline
231 // (which uses a FunctionPassManager (which segfaults (not asserts) if
232 // provided a ModulePass))).
233 Constant *GV = new GlobalVariable(*F.getParent(), FrameMap->getType(), true,
234 GlobalVariable::InternalLinkage,
235 FrameMap, "__gc_" + F.getName());
237 Constant *GEPIndices[2] = { ConstantInt::get(Type::Int32Ty, 0),
238 ConstantInt::get(Type::Int32Ty, 0) };
239 return ConstantExpr::getGetElementPtr(GV, GEPIndices, 2);
242 const 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(F.getContext(), EltTys);
250 std::string TypeName("gc_stackentry.");
251 TypeName += F.getName();
252 F.getParent()->addTypeName(TypeName, Ty);
254 return Ty;
257 /// doInitialization - If this module uses the GC intrinsics, find them now. If
258 /// not, exit fast.
259 bool 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(M.getContext(), EltTys);
269 M.addTypeName("gc_map", FrameMapTy);
270 PointerType *FrameMapPtrTy = PointerType::getUnqual(FrameMapTy);
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();
279 EltTys.clear();
280 EltTys.push_back(PointerType::getUnqual(RecursiveTy));
281 EltTys.push_back(FrameMapPtrTy);
282 PATypeHolder LinkTyH = StructType::get(M.getContext(), EltTys);
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?
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(M, StackEntryPtrTy, false,
296 GlobalValue::LinkOnceAnyLinkage,
297 Constant::getNullValue(StackEntryPtrTy),
298 "llvm_gc_root_chain");
299 } else if (Head->hasExternalLinkage() && Head->isDeclaration()) {
300 Head->setInitializer(Constant::getNullValue(StackEntryPtrTy));
301 Head->setLinkage(GlobalValue::LinkOnceAnyLinkage);
304 return true;
307 bool ShadowStackGC::IsNullValue(Value *V) {
308 if (Constant *C = dyn_cast<Constant>(V))
309 return C->isNullValue();
310 return false;
313 void 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.
318 assert(Roots.empty() && "Not cleaned up?");
320 SmallVector<std::pair<CallInst*,AllocaInst*>,16> MetaRoots;
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);
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());
340 GetElementPtrInst *
341 ShadowStackGC::CreateGEP(LLVMContext &Context, 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);
348 assert(isa<GetElementPtrInst>(Val) && "Unexpected folded constant");
350 return dyn_cast<GetElementPtrInst>(Val);
353 GetElementPtrInst *
354 ShadowStackGC::CreateGEP(LLVMContext &Context, 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);
360 assert(isa<GetElementPtrInst>(Val) && "Unexpected folded constant");
362 return dyn_cast<GetElementPtrInst>(Val);
365 /// runOnFunction - Insert code to maintain the shadow stack.
366 bool ShadowStackGC::performCustomLowering(Function &F) {
367 LLVMContext &Context = F.getContext();
369 // Find calls to llvm.gcroot.
370 CollectRoots(F);
372 // If there are no roots in this function, then there is no need to add a
373 // stack map entry for it.
374 if (Roots.empty())
375 return false;
377 // Build the constant map and figure the type of the shadow stack entry.
378 Value *FrameMap = GetFrameMap(F);
379 const Type *ConcreteStackEntryTy = GetConcreteStackEntryType(F);
381 // Build the shadow stack entry at the very start of the function.
382 BasicBlock::iterator IP = F.getEntryBlock().begin();
383 IRBuilder<> AtEntry(IP->getParent(), IP);
385 Instruction *StackEntry = AtEntry.CreateAlloca(ConcreteStackEntryTy, 0,
386 "gc_frame");
388 while (isa<AllocaInst>(IP)) ++IP;
389 AtEntry.SetInsertPoint(IP->getParent(), IP);
391 // Initialize the map pointer and load the current head of the shadow stack.
392 Instruction *CurrentHead = AtEntry.CreateLoad(Head, "gc_currhead");
393 Instruction *EntryMapPtr = CreateGEP(Context, AtEntry, StackEntry,
394 0,1,"gc_frame.map");
395 AtEntry.CreateStore(FrameMap, EntryMapPtr);
397 // After all the allocas...
398 for (unsigned I = 0, E = Roots.size(); I != E; ++I) {
399 // For each root, find the corresponding slot in the aggregate...
400 Value *SlotPtr = CreateGEP(Context, AtEntry, StackEntry, 1 + I, "gc_root");
402 // And use it in lieu of the alloca.
403 AllocaInst *OriginalAlloca = Roots[I].second;
404 SlotPtr->takeName(OriginalAlloca);
405 OriginalAlloca->replaceAllUsesWith(SlotPtr);
408 // Move past the original stores inserted by GCStrategy::InitRoots. This isn't
409 // really necessary (the collector would never see the intermediate state at
410 // runtime), but it's nicer not to push the half-initialized entry onto the
411 // shadow stack.
412 while (isa<StoreInst>(IP)) ++IP;
413 AtEntry.SetInsertPoint(IP->getParent(), IP);
415 // Push the entry onto the shadow stack.
416 Instruction *EntryNextPtr = CreateGEP(Context, AtEntry,
417 StackEntry,0,0,"gc_frame.next");
418 Instruction *NewHeadVal = CreateGEP(Context, AtEntry,
419 StackEntry, 0, "gc_newhead");
420 AtEntry.CreateStore(CurrentHead, EntryNextPtr);
421 AtEntry.CreateStore(NewHeadVal, Head);
423 // For each instruction that escapes...
424 EscapeEnumerator EE(F, "gc_cleanup");
425 while (IRBuilder<> *AtExit = EE.Next()) {
426 // Pop the entry from the shadow stack. Don't reuse CurrentHead from
427 // AtEntry, since that would make the value live for the entire function.
428 Instruction *EntryNextPtr2 = CreateGEP(Context, *AtExit, StackEntry, 0, 0,
429 "gc_frame.next");
430 Value *SavedHead = AtExit->CreateLoad(EntryNextPtr2, "gc_savedhead");
431 AtExit->CreateStore(SavedHead, Head);
434 // Delete the original allocas (which are no longer used) and the intrinsic
435 // calls (which are no longer valid). Doing this last avoids invalidating
436 // iterators.
437 for (unsigned I = 0, E = Roots.size(); I != E; ++I) {
438 Roots[I].first->eraseFromParent();
439 Roots[I].second->eraseFromParent();
442 Roots.clear();
443 return true;