Index: docs/GarbageCollection.rst =================================================================== --- docs/GarbageCollection.rst +++ docs/GarbageCollection.rst @@ -721,8 +721,9 @@ Custom lowering of intrinsics: ``CustomRoots``, ``CustomReadBarriers``, and ``CustomWriteBarriers`` --------------------------------------------------------------------------------------------------- -For GCs which use barriers or unusual treatment of stack roots, these flags -allow the collector to perform arbitrary transformations of the LLVM IR: +For GCs which use barriers or unusual treatment of stack roots, these +flags allow the collector to perform arbitrary transformations of the +LLVM IR: .. code-block:: c++ @@ -733,70 +734,18 @@ CustomReadBarriers = true; CustomWriteBarriers = true; } - - virtual bool initializeCustomLowering(Module &M); - virtual bool performCustomLowering(Function &F); }; -If any of these flags are set, then LLVM suppresses its default lowering for the -corresponding intrinsics and instead calls ``performCustomLowering``. - -LLVM's default action for each intrinsic is as follows: - -* ``llvm.gcroot``: Leave it alone. The code generator must see it or the stack - map will not be computed. - -* ``llvm.gcread``: Substitute a ``load`` instruction. - -* ``llvm.gcwrite``: Substitute a ``store`` instruction. - -If ``CustomReadBarriers`` or ``CustomWriteBarriers`` are specified, then -``performCustomLowering`` **must** eliminate the corresponding barriers. - -``performCustomLowering`` must comply with the same restrictions as -:ref:`FunctionPass::runOnFunction ` -Likewise, ``initializeCustomLowering`` has the same semantics as -:ref:`Pass::doInitialization(Module&) -` +If any of these flags are set, LLVM suppresses its default lowering for +the corresponding intrinsics. Instead, you must provide a custom Pass +which lowers the intrinsics as desired. If you have opted in to custom +lowering of a particular intrinsic your pass **must** eliminate all +instances of the corresponding intrinsic in functions which opt in to +your GC. The best example of such a pass is the ShadowStackGC and it's +ShadowStackGCLowering pass. -The following can be used as a template: - -.. code-block:: c++ - - #include "llvm/IR/Module.h" - #include "llvm/IR/IntrinsicInst.h" - - bool MyGC::initializeCustomLowering(Module &M) { - return false; - } - - bool MyGC::performCustomLowering(Function &F) { - bool MadeChange = false; - - for (Function::iterator BB = F.begin(), E = F.end(); BB != E; ++BB) - for (BasicBlock::iterator II = BB->begin(), E = BB->end(); II != E; ) - if (IntrinsicInst *CI = dyn_cast(II++)) - if (Function *F = CI->getCalledFunction()) - switch (F->getIntrinsicID()) { - case Intrinsic::gcwrite: - // Handle llvm.gcwrite. - CI->eraseFromParent(); - MadeChange = true; - break; - case Intrinsic::gcread: - // Handle llvm.gcread. - CI->eraseFromParent(); - MadeChange = true; - break; - case Intrinsic::gcroot: - // Handle llvm.gcroot. - CI->eraseFromParent(); - MadeChange = true; - break; - } - - return MadeChange; - } +There is currently no way to register such a custom lowering pass +without building a custom copy of LLVM. .. _safe-points: Index: include/llvm/CodeGen/GCStrategy.h =================================================================== --- include/llvm/CodeGen/GCStrategy.h +++ include/llvm/CodeGen/GCStrategy.h @@ -101,13 +101,13 @@ const std::string &getName() const { return Name; } /// By default, write barriers are replaced with simple store - /// instructions. If true, then performCustomLowering must instead lower - /// them. + /// instructions. If true, you must provide a custom pass to lower + /// calls to @llvm.gcwrite. bool customWriteBarrier() const { return CustomWriteBarriers; } /// By default, read barriers are replaced with simple load - /// instructions. If true, then performCustomLowering must instead lower - /// them. + /// instructions. If true, you must provide a custom pass to lower + /// calls to @llvm.gcread. bool customReadBarrier() const { return CustomReadBarriers; } /// Returns true if this strategy is expecting the use of gc.statepoints, @@ -143,7 +143,8 @@ } /// By default, roots are left for the code generator so it can generate a - /// stack map. If true, then performCustomLowering must delete them. + /// stack map. If true, you must provide a custom pass to lower + /// calls to @llvm.gcroot. bool customRoots() const { return CustomRoots; } /// If set, gcroot intrinsics should initialize their allocas to null @@ -158,23 +159,6 @@ bool usesMetadata() const { return UsesMetadata; } ///@} - - /// initializeCustomLowering/performCustomLowering - If any of the actions - /// are set to custom, performCustomLowering must be overriden to transform - /// the corresponding actions to LLVM IR. initializeCustomLowering is - /// optional to override. These are the only GCStrategy methods through - /// which the LLVM IR can be modified. These methods apply mostly to - /// gc.root based implementations, but can be overriden to provide custom - /// barrier lowerings with gc.statepoint as well. - ///@{ - virtual bool initializeCustomLowering(Module &F) { - // No changes made - return false; - } - virtual bool performCustomLowering(Function &F) { - llvm_unreachable("GCStrategy subclass specified a configuration which" - "requires a custom lowering without providing one"); - } }; /// Subclasses of GCStrategy are made available for use during compilation by Index: include/llvm/CodeGen/Passes.h =================================================================== --- include/llvm/CodeGen/Passes.h +++ include/llvm/CodeGen/Passes.h @@ -517,11 +517,15 @@ /// information. extern char &MachineBlockPlacementStatsID; - /// GCLowering Pass - Performs target-independent LLVM IR transformations for - /// highly portable strategies. - /// + /// GCLowering Pass - Used by gc.root to perform it's default lowering + /// operations. FunctionPass *createGCLoweringPass(); + /// ShadowStackGCLowering - Implements the custom lowering mechanism + /// used by the shadow stack GC. Only runs on functions which opt in to + /// the shadow stack collector. + FunctionPass *createShadowStackGCLoweringPass(); + /// GCMachineCodeAnalysis - Target-independent pass to mark safe points /// in machine code. Must be added very late during code generation, just /// prior to output, and importantly after all CFG transformations (such as Index: include/llvm/InitializePasses.h =================================================================== --- include/llvm/InitializePasses.h +++ include/llvm/InitializePasses.h @@ -245,6 +245,7 @@ void initializeScalarEvolutionAliasAnalysisPass(PassRegistry&); void initializeScalarEvolutionPass(PassRegistry&); void initializeSimpleInlinerPass(PassRegistry&); +void initializeShadowStackGCLoweringPass(PassRegistry&); void initializeRegisterCoalescerPass(PassRegistry&); void initializeSingleLoopExtractorPass(PassRegistry&); void initializeSinkingPass(PassRegistry&); Index: lib/CodeGen/GCRootLowering.cpp =================================================================== --- lib/CodeGen/GCRootLowering.cpp +++ lib/CodeGen/GCRootLowering.cpp @@ -111,11 +111,6 @@ C.initializeRoots(); } -static bool NeedsCustomLoweringPass(const GCStrategy &C) { - // Custom lowering is only necessary if enabled for some action. - return C.customWriteBarrier() || C.customReadBarrier() || C.customRoots(); -} - /// doInitialization - If this module uses the GC intrinsics, find them now. bool LowerIntrinsics::doInitialization(Module &M) { // FIXME: This is rather antisocial in the context of a JIT since it performs @@ -128,13 +123,7 @@ if (!I->isDeclaration() && I->hasGC()) MI->getFunctionInfo(*I); // Instantiate the GC strategy. - bool MadeChange = false; - for (GCModuleInfo::iterator I = MI->begin(), E = MI->end(); I != E; ++I) - if (NeedsCustomLoweringPass(**I)) - if ((*I)->initializeCustomLowering(M)) - MadeChange = true; - - return MadeChange; + return false; } /// CouldBecomeSafePoint - Predicate to conservatively determine whether the @@ -211,17 +200,6 @@ if (NeedsDefaultLoweringPass(S)) MadeChange |= PerformDefaultLowering(F, S); - bool UseCustomLoweringPass = NeedsCustomLoweringPass(S); - if (UseCustomLoweringPass) - MadeChange |= S.performCustomLowering(F); - - // Custom lowering may modify the CFG, so dominators must be recomputed. - if (UseCustomLoweringPass) { - if (DominatorTreeWrapperPass *DTWP = - getAnalysisIfAvailable()) - DTWP->getDomTree().recalculate(F); - } - return MadeChange; } Index: lib/CodeGen/Passes.cpp =================================================================== --- lib/CodeGen/Passes.cpp +++ lib/CodeGen/Passes.cpp @@ -419,7 +419,10 @@ addPass(createPrintFunctionPass(dbgs(), "\n\n*** Code after LSR ***\n")); } + // Run GC lowering passes for builtin collectors + // TODO: add a pass insertion point here addPass(createGCLoweringPass()); + addPass(createShadowStackGCLoweringPass()); // Make sure that no unreachable blocks are instruction selected. addPass(createUnreachableBlockEliminationPass()); Index: lib/CodeGen/ShadowStackGC.cpp =================================================================== --- lib/CodeGen/ShadowStackGC.cpp +++ lib/CodeGen/ShadowStackGC.cpp @@ -38,412 +38,18 @@ #define DEBUG_TYPE "shadowstackgc" namespace { - class ShadowStackGC : public GCStrategy { - /// RootChain - This is the global linked-list that contains the chain of GC - /// roots. - GlobalVariable *Head; - - /// StackEntryTy - Abstract type of a link in the shadow stack. - /// - StructType *StackEntryTy; - StructType *FrameMapTy; - - /// Roots - GC roots in the current function. Each is a pair of the - /// intrinsic call and its corresponding alloca. - std::vector> Roots; - public: ShadowStackGC(); - - bool initializeCustomLowering(Module &M) override; - bool performCustomLowering(Function &F) override; - -private: - bool IsNullValue(Value *V); - Constant *GetFrameMap(Function &F); - Type *GetConcreteStackEntryType(Function &F); - void CollectRoots(Function &F); - static GetElementPtrInst *CreateGEP(LLVMContext &Context, IRBuilder<> &B, - Value *BasePtr, int Idx1, - const char *Name); - static GetElementPtrInst *CreateGEP(LLVMContext &Context, IRBuilder<> &B, - Value *BasePtr, int Idx1, int Idx2, - const char *Name); }; } static GCRegistry::Add X("shadow-stack", "Very portable GC for uncooperative code generators"); -namespace { -/// EscapeEnumerator - This is a little algorithm to find all escape points -/// from a function so that "finally"-style code can be inserted. In addition -/// to finding the existing return and unwind instructions, it also (if -/// necessary) transforms any call instructions into invokes and sends them to -/// a landing pad. -/// -/// It's wrapped up in a state machine using the same transform C# uses for -/// 'yield return' enumerators, This transform allows it to be non-allocating. -class EscapeEnumerator { - Function &F; - const char *CleanupBBName; - - // State. - int State; - Function::iterator StateBB, StateE; - IRBuilder<> Builder; - -public: - EscapeEnumerator(Function &F, const char *N = "cleanup") - : F(F), CleanupBBName(N), State(0), Builder(F.getContext()) {} - - IRBuilder<> *Next() { - switch (State) { - default: - return nullptr; - - case 0: - StateBB = F.begin(); - StateE = F.end(); - State = 1; - - case 1: - // Find all 'return', 'resume', and 'unwind' instructions. - while (StateBB != StateE) { - BasicBlock *CurBB = StateBB++; - - // Branches and invokes do not escape, only unwind, resume, and return - // do. - TerminatorInst *TI = CurBB->getTerminator(); - if (!isa(TI) && !isa(TI)) - continue; - - Builder.SetInsertPoint(TI->getParent(), TI); - return &Builder; - } - - State = 2; - - // Find all 'call' instructions. - SmallVector Calls; - for (Function::iterator BB = F.begin(), E = F.end(); BB != E; ++BB) - for (BasicBlock::iterator II = BB->begin(), EE = BB->end(); II != EE; - ++II) - if (CallInst *CI = dyn_cast(II)) - if (!CI->getCalledFunction() || - !CI->getCalledFunction()->getIntrinsicID()) - Calls.push_back(CI); - - if (Calls.empty()) - return nullptr; - - // Create a cleanup block. - LLVMContext &C = F.getContext(); - BasicBlock *CleanupBB = BasicBlock::Create(C, CleanupBBName, &F); - Type *ExnTy = - StructType::get(Type::getInt8PtrTy(C), Type::getInt32Ty(C), nullptr); - Constant *PersFn = F.getParent()->getOrInsertFunction( - "__gcc_personality_v0", FunctionType::get(Type::getInt32Ty(C), true)); - LandingPadInst *LPad = - LandingPadInst::Create(ExnTy, PersFn, 1, "cleanup.lpad", CleanupBB); - LPad->setCleanup(true); - ResumeInst *RI = ResumeInst::Create(LPad, CleanupBB); - - // Transform the 'call' instructions into 'invoke's branching to the - // cleanup block. Go in reverse order to make prettier BB names. - SmallVector Args; - for (unsigned I = Calls.size(); I != 0;) { - CallInst *CI = cast(Calls[--I]); - - // Split the basic block containing the function call. - BasicBlock *CallBB = CI->getParent(); - BasicBlock *NewBB = - CallBB->splitBasicBlock(CI, CallBB->getName() + ".cont"); - - // Remove the unconditional branch inserted at the end of CallBB. - CallBB->getInstList().pop_back(); - NewBB->getInstList().remove(CI); - - // Create a new invoke instruction. - Args.clear(); - CallSite CS(CI); - Args.append(CS.arg_begin(), CS.arg_end()); - - InvokeInst *II = - InvokeInst::Create(CI->getCalledValue(), NewBB, CleanupBB, Args, - CI->getName(), CallBB); - II->setCallingConv(CI->getCallingConv()); - II->setAttributes(CI->getAttributes()); - CI->replaceAllUsesWith(II); - delete CI; - } - - Builder.SetInsertPoint(RI->getParent(), RI); - return &Builder; - } - } -}; -} - -// ----------------------------------------------------------------------------- - void llvm::linkShadowStackGC() {} -ShadowStackGC::ShadowStackGC() : Head(nullptr), StackEntryTy(nullptr) { +ShadowStackGC::ShadowStackGC() { InitRoots = true; CustomRoots = true; } - -Constant *ShadowStackGC::GetFrameMap(Function &F) { - // doInitialization creates the abstract type of this value. - Type *VoidPtr = Type::getInt8PtrTy(F.getContext()); - - // Truncate the ShadowStackDescriptor if some metadata is null. - unsigned NumMeta = 0; - SmallVector Metadata; - for (unsigned I = 0; I != Roots.size(); ++I) { - Constant *C = cast(Roots[I].first->getArgOperand(1)); - if (!C->isNullValue()) - NumMeta = I + 1; - Metadata.push_back(ConstantExpr::getBitCast(C, VoidPtr)); - } - Metadata.resize(NumMeta); - - Type *Int32Ty = Type::getInt32Ty(F.getContext()); - - Constant *BaseElts[] = { - ConstantInt::get(Int32Ty, Roots.size(), false), - ConstantInt::get(Int32Ty, NumMeta, false), - }; - - Constant *DescriptorElts[] = { - ConstantStruct::get(FrameMapTy, BaseElts), - ConstantArray::get(ArrayType::get(VoidPtr, NumMeta), Metadata)}; - - Type *EltTys[] = {DescriptorElts[0]->getType(), DescriptorElts[1]->getType()}; - StructType *STy = StructType::create(EltTys, "gc_map." + utostr(NumMeta)); - - Constant *FrameMap = ConstantStruct::get(STy, DescriptorElts); - - // FIXME: Is this actually dangerous as WritingAnLLVMPass.html claims? Seems - // that, short of multithreaded LLVM, it should be safe; all that is - // necessary is that a simple Module::iterator loop not be invalidated. - // Appending to the GlobalVariable list is safe in that sense. - // - // All of the output passes emit globals last. The ExecutionEngine - // explicitly supports adding globals to the module after - // initialization. - // - // Still, if it isn't deemed acceptable, then this transformation needs - // to be a ModulePass (which means it cannot be in the 'llc' pipeline - // (which uses a FunctionPassManager (which segfaults (not asserts) if - // provided a ModulePass))). - Constant *GV = new GlobalVariable(*F.getParent(), FrameMap->getType(), true, - GlobalVariable::InternalLinkage, FrameMap, - "__gc_" + F.getName()); - - Constant *GEPIndices[2] = { - ConstantInt::get(Type::getInt32Ty(F.getContext()), 0), - ConstantInt::get(Type::getInt32Ty(F.getContext()), 0)}; - return ConstantExpr::getGetElementPtr(GV, GEPIndices); -} - -Type *ShadowStackGC::GetConcreteStackEntryType(Function &F) { - // doInitialization creates the generic version of this type. - std::vector EltTys; - EltTys.push_back(StackEntryTy); - for (size_t I = 0; I != Roots.size(); I++) - EltTys.push_back(Roots[I].second->getAllocatedType()); - - return StructType::create(EltTys, "gc_stackentry." + F.getName().str()); -} - -/// doInitialization - If this module uses the GC intrinsics, find them now. If -/// not, exit fast. -bool ShadowStackGC::initializeCustomLowering(Module &M) { - // struct FrameMap { - // int32_t NumRoots; // Number of roots in stack frame. - // int32_t NumMeta; // Number of metadata descriptors. May be < NumRoots. - // void *Meta[]; // May be absent for roots without metadata. - // }; - std::vector EltTys; - // 32 bits is ok up to a 32GB stack frame. :) - EltTys.push_back(Type::getInt32Ty(M.getContext())); - // Specifies length of variable length array. - EltTys.push_back(Type::getInt32Ty(M.getContext())); - FrameMapTy = StructType::create(EltTys, "gc_map"); - PointerType *FrameMapPtrTy = PointerType::getUnqual(FrameMapTy); - - // struct StackEntry { - // ShadowStackEntry *Next; // Caller's stack entry. - // FrameMap *Map; // Pointer to constant FrameMap. - // void *Roots[]; // Stack roots (in-place array, so we pretend). - // }; - - StackEntryTy = StructType::create(M.getContext(), "gc_stackentry"); - - EltTys.clear(); - EltTys.push_back(PointerType::getUnqual(StackEntryTy)); - EltTys.push_back(FrameMapPtrTy); - StackEntryTy->setBody(EltTys); - PointerType *StackEntryPtrTy = PointerType::getUnqual(StackEntryTy); - - // Get the root chain if it already exists. - Head = M.getGlobalVariable("llvm_gc_root_chain"); - if (!Head) { - // If the root chain does not exist, insert a new one with linkonce - // linkage! - Head = new GlobalVariable( - M, StackEntryPtrTy, false, GlobalValue::LinkOnceAnyLinkage, - Constant::getNullValue(StackEntryPtrTy), "llvm_gc_root_chain"); - } else if (Head->hasExternalLinkage() && Head->isDeclaration()) { - Head->setInitializer(Constant::getNullValue(StackEntryPtrTy)); - Head->setLinkage(GlobalValue::LinkOnceAnyLinkage); - } - - return true; -} - -bool ShadowStackGC::IsNullValue(Value *V) { - if (Constant *C = dyn_cast(V)) - return C->isNullValue(); - return false; -} - -void ShadowStackGC::CollectRoots(Function &F) { - // FIXME: Account for original alignment. Could fragment the root array. - // Approach 1: Null initialize empty slots at runtime. Yuck. - // Approach 2: Emit a map of the array instead of just a count. - - assert(Roots.empty() && "Not cleaned up?"); - - SmallVector, 16> MetaRoots; - - for (Function::iterator BB = F.begin(), E = F.end(); BB != E; ++BB) - for (BasicBlock::iterator II = BB->begin(), E = BB->end(); II != E;) - if (IntrinsicInst *CI = dyn_cast(II++)) - if (Function *F = CI->getCalledFunction()) - if (F->getIntrinsicID() == Intrinsic::gcroot) { - std::pair Pair = std::make_pair( - CI, - cast(CI->getArgOperand(0)->stripPointerCasts())); - if (IsNullValue(CI->getArgOperand(1))) - Roots.push_back(Pair); - else - MetaRoots.push_back(Pair); - } - - // Number roots with metadata (usually empty) at the beginning, so that the - // FrameMap::Meta array can be elided. - Roots.insert(Roots.begin(), MetaRoots.begin(), MetaRoots.end()); -} - -GetElementPtrInst *ShadowStackGC::CreateGEP(LLVMContext &Context, - IRBuilder<> &B, Value *BasePtr, - int Idx, int Idx2, - const char *Name) { - Value *Indices[] = {ConstantInt::get(Type::getInt32Ty(Context), 0), - ConstantInt::get(Type::getInt32Ty(Context), Idx), - ConstantInt::get(Type::getInt32Ty(Context), Idx2)}; - Value *Val = B.CreateGEP(BasePtr, Indices, Name); - - assert(isa(Val) && "Unexpected folded constant"); - - return dyn_cast(Val); -} - -GetElementPtrInst *ShadowStackGC::CreateGEP(LLVMContext &Context, - IRBuilder<> &B, Value *BasePtr, - int Idx, const char *Name) { - Value *Indices[] = {ConstantInt::get(Type::getInt32Ty(Context), 0), - ConstantInt::get(Type::getInt32Ty(Context), Idx)}; - Value *Val = B.CreateGEP(BasePtr, Indices, Name); - - assert(isa(Val) && "Unexpected folded constant"); - - return dyn_cast(Val); -} - -/// runOnFunction - Insert code to maintain the shadow stack. -bool ShadowStackGC::performCustomLowering(Function &F) { - LLVMContext &Context = F.getContext(); - - // Find calls to llvm.gcroot. - CollectRoots(F); - - // If there are no roots in this function, then there is no need to add a - // stack map entry for it. - if (Roots.empty()) - return false; - - // Build the constant map and figure the type of the shadow stack entry. - Value *FrameMap = GetFrameMap(F); - Type *ConcreteStackEntryTy = GetConcreteStackEntryType(F); - - // Build the shadow stack entry at the very start of the function. - BasicBlock::iterator IP = F.getEntryBlock().begin(); - IRBuilder<> AtEntry(IP->getParent(), IP); - - Instruction *StackEntry = - AtEntry.CreateAlloca(ConcreteStackEntryTy, nullptr, "gc_frame"); - - while (isa(IP)) - ++IP; - AtEntry.SetInsertPoint(IP->getParent(), IP); - - // Initialize the map pointer and load the current head of the shadow stack. - Instruction *CurrentHead = AtEntry.CreateLoad(Head, "gc_currhead"); - Instruction *EntryMapPtr = - CreateGEP(Context, AtEntry, StackEntry, 0, 1, "gc_frame.map"); - AtEntry.CreateStore(FrameMap, EntryMapPtr); - - // After all the allocas... - for (unsigned I = 0, E = Roots.size(); I != E; ++I) { - // For each root, find the corresponding slot in the aggregate... - Value *SlotPtr = CreateGEP(Context, AtEntry, StackEntry, 1 + I, "gc_root"); - - // And use it in lieu of the alloca. - AllocaInst *OriginalAlloca = Roots[I].second; - SlotPtr->takeName(OriginalAlloca); - OriginalAlloca->replaceAllUsesWith(SlotPtr); - } - - // Move past the original stores inserted by GCStrategy::InitRoots. This isn't - // really necessary (the collector would never see the intermediate state at - // runtime), but it's nicer not to push the half-initialized entry onto the - // shadow stack. - while (isa(IP)) - ++IP; - AtEntry.SetInsertPoint(IP->getParent(), IP); - - // Push the entry onto the shadow stack. - Instruction *EntryNextPtr = - CreateGEP(Context, AtEntry, StackEntry, 0, 0, "gc_frame.next"); - Instruction *NewHeadVal = - CreateGEP(Context, AtEntry, StackEntry, 0, "gc_newhead"); - AtEntry.CreateStore(CurrentHead, EntryNextPtr); - AtEntry.CreateStore(NewHeadVal, Head); - - // For each instruction that escapes... - EscapeEnumerator EE(F, "gc_cleanup"); - while (IRBuilder<> *AtExit = EE.Next()) { - // Pop the entry from the shadow stack. Don't reuse CurrentHead from - // AtEntry, since that would make the value live for the entire function. - Instruction *EntryNextPtr2 = - CreateGEP(Context, *AtExit, StackEntry, 0, 0, "gc_frame.next"); - Value *SavedHead = AtExit->CreateLoad(EntryNextPtr2, "gc_savedhead"); - AtExit->CreateStore(SavedHead, Head); - } - - // Delete the original allocas (which are no longer used) and the intrinsic - // calls (which are no longer valid). Doing this last avoids invalidating - // iterators. - for (unsigned I = 0, E = Roots.size(); I != E; ++I) { - Roots[I].first->eraseFromParent(); - Roots[I].second->eraseFromParent(); - } - - Roots.clear(); - return true; -} Index: lib/CodeGen/ShadowStackGCLowering.cpp =================================================================== --- /dev/null +++ lib/CodeGen/ShadowStackGCLowering.cpp @@ -0,0 +1,457 @@ +//===-- ShadowStackGCLowering.cpp - Custom lowering for shadow-stack gc ---===// +// +// The LLVM Compiler Infrastructure +// +// This file is distributed under the University of Illinois Open Source +// License. See LICENSE.TXT for details. +// +//===----------------------------------------------------------------------===// +// +// This file contains the custom lowering code required by the shadow-stack GC +// strategy. +// +//===----------------------------------------------------------------------===// + +#include "llvm/CodeGen/Passes.h" +#include "llvm/CodeGen/GCStrategy.h" +#include "llvm/ADT/StringExtras.h" +#include "llvm/IR/CallSite.h" +#include "llvm/IR/IRBuilder.h" +#include "llvm/IR/IntrinsicInst.h" +#include "llvm/IR/Module.h" + +using namespace llvm; + +#define DEBUG_TYPE "shadowstackgclowering" + +namespace { + +class ShadowStackGCLowering : public FunctionPass { + /// RootChain - This is the global linked-list that contains the chain of GC + /// roots. + GlobalVariable *Head; + + /// StackEntryTy - Abstract type of a link in the shadow stack. + /// + StructType *StackEntryTy; + StructType *FrameMapTy; + + /// Roots - GC roots in the current function. Each is a pair of the + /// intrinsic call and its corresponding alloca. + std::vector> Roots; + +public: + static char ID; + ShadowStackGCLowering(); + + bool doInitialization(Module &M) override; + bool runOnFunction(Function &F) override; + +private: + bool IsNullValue(Value *V); + Constant *GetFrameMap(Function &F); + Type *GetConcreteStackEntryType(Function &F); + void CollectRoots(Function &F); + static GetElementPtrInst *CreateGEP(LLVMContext &Context, IRBuilder<> &B, + Value *BasePtr, int Idx1, + const char *Name); + static GetElementPtrInst *CreateGEP(LLVMContext &Context, IRBuilder<> &B, + Value *BasePtr, int Idx1, int Idx2, + const char *Name); +}; +} + +INITIALIZE_PASS_BEGIN(ShadowStackGCLowering, "shadow-stack-gc-lowering", + "Shadow Stack GC Lowering", false, false) +INITIALIZE_PASS_DEPENDENCY(GCModuleInfo) +INITIALIZE_PASS_END(ShadowStackGCLowering, "shadow-stack-gc-lowering", + "Shadow Stack GC Lowering", false, false) + +FunctionPass *llvm::createShadowStackGCLoweringPass() { return new ShadowStackGCLowering(); } + +char ShadowStackGCLowering::ID = 0; + +ShadowStackGCLowering::ShadowStackGCLowering() + : FunctionPass(ID), Head(nullptr), StackEntryTy(nullptr), + FrameMapTy(nullptr) { + initializeShadowStackGCLoweringPass(*PassRegistry::getPassRegistry()); +} + +namespace { +/// EscapeEnumerator - This is a little algorithm to find all escape points +/// from a function so that "finally"-style code can be inserted. In addition +/// to finding the existing return and unwind instructions, it also (if +/// necessary) transforms any call instructions into invokes and sends them to +/// a landing pad. +/// +/// It's wrapped up in a state machine using the same transform C# uses for +/// 'yield return' enumerators, This transform allows it to be non-allocating. +class EscapeEnumerator { + Function &F; + const char *CleanupBBName; + + // State. + int State; + Function::iterator StateBB, StateE; + IRBuilder<> Builder; + +public: + EscapeEnumerator(Function &F, const char *N = "cleanup") + : F(F), CleanupBBName(N), State(0), Builder(F.getContext()) {} + + IRBuilder<> *Next() { + switch (State) { + default: + return nullptr; + + case 0: + StateBB = F.begin(); + StateE = F.end(); + State = 1; + + case 1: + // Find all 'return', 'resume', and 'unwind' instructions. + while (StateBB != StateE) { + BasicBlock *CurBB = StateBB++; + + // Branches and invokes do not escape, only unwind, resume, and return + // do. + TerminatorInst *TI = CurBB->getTerminator(); + if (!isa(TI) && !isa(TI)) + continue; + + Builder.SetInsertPoint(TI->getParent(), TI); + return &Builder; + } + + State = 2; + + // Find all 'call' instructions. + SmallVector Calls; + for (Function::iterator BB = F.begin(), E = F.end(); BB != E; ++BB) + for (BasicBlock::iterator II = BB->begin(), EE = BB->end(); II != EE; + ++II) + if (CallInst *CI = dyn_cast(II)) + if (!CI->getCalledFunction() || + !CI->getCalledFunction()->getIntrinsicID()) + Calls.push_back(CI); + + if (Calls.empty()) + return nullptr; + + // Create a cleanup block. + LLVMContext &C = F.getContext(); + BasicBlock *CleanupBB = BasicBlock::Create(C, CleanupBBName, &F); + Type *ExnTy = + StructType::get(Type::getInt8PtrTy(C), Type::getInt32Ty(C), nullptr); + Constant *PersFn = F.getParent()->getOrInsertFunction( + "__gcc_personality_v0", FunctionType::get(Type::getInt32Ty(C), true)); + LandingPadInst *LPad = + LandingPadInst::Create(ExnTy, PersFn, 1, "cleanup.lpad", CleanupBB); + LPad->setCleanup(true); + ResumeInst *RI = ResumeInst::Create(LPad, CleanupBB); + + // Transform the 'call' instructions into 'invoke's branching to the + // cleanup block. Go in reverse order to make prettier BB names. + SmallVector Args; + for (unsigned I = Calls.size(); I != 0;) { + CallInst *CI = cast(Calls[--I]); + + // Split the basic block containing the function call. + BasicBlock *CallBB = CI->getParent(); + BasicBlock *NewBB = + CallBB->splitBasicBlock(CI, CallBB->getName() + ".cont"); + + // Remove the unconditional branch inserted at the end of CallBB. + CallBB->getInstList().pop_back(); + NewBB->getInstList().remove(CI); + + // Create a new invoke instruction. + Args.clear(); + CallSite CS(CI); + Args.append(CS.arg_begin(), CS.arg_end()); + + InvokeInst *II = + InvokeInst::Create(CI->getCalledValue(), NewBB, CleanupBB, Args, + CI->getName(), CallBB); + II->setCallingConv(CI->getCallingConv()); + II->setAttributes(CI->getAttributes()); + CI->replaceAllUsesWith(II); + delete CI; + } + + Builder.SetInsertPoint(RI->getParent(), RI); + return &Builder; + } + } +}; +} + + +Constant *ShadowStackGCLowering::GetFrameMap(Function &F) { + // doInitialization creates the abstract type of this value. + Type *VoidPtr = Type::getInt8PtrTy(F.getContext()); + + // Truncate the ShadowStackDescriptor if some metadata is null. + unsigned NumMeta = 0; + SmallVector Metadata; + for (unsigned I = 0; I != Roots.size(); ++I) { + Constant *C = cast(Roots[I].first->getArgOperand(1)); + if (!C->isNullValue()) + NumMeta = I + 1; + Metadata.push_back(ConstantExpr::getBitCast(C, VoidPtr)); + } + Metadata.resize(NumMeta); + + Type *Int32Ty = Type::getInt32Ty(F.getContext()); + + Constant *BaseElts[] = { + ConstantInt::get(Int32Ty, Roots.size(), false), + ConstantInt::get(Int32Ty, NumMeta, false), + }; + + Constant *DescriptorElts[] = { + ConstantStruct::get(FrameMapTy, BaseElts), + ConstantArray::get(ArrayType::get(VoidPtr, NumMeta), Metadata)}; + + Type *EltTys[] = {DescriptorElts[0]->getType(), DescriptorElts[1]->getType()}; + StructType *STy = StructType::create(EltTys, "gc_map." + utostr(NumMeta)); + + Constant *FrameMap = ConstantStruct::get(STy, DescriptorElts); + + // FIXME: Is this actually dangerous as WritingAnLLVMPass.html claims? Seems + // that, short of multithreaded LLVM, it should be safe; all that is + // necessary is that a simple Module::iterator loop not be invalidated. + // Appending to the GlobalVariable list is safe in that sense. + // + // All of the output passes emit globals last. The ExecutionEngine + // explicitly supports adding globals to the module after + // initialization. + // + // Still, if it isn't deemed acceptable, then this transformation needs + // to be a ModulePass (which means it cannot be in the 'llc' pipeline + // (which uses a FunctionPassManager (which segfaults (not asserts) if + // provided a ModulePass))). + Constant *GV = new GlobalVariable(*F.getParent(), FrameMap->getType(), true, + GlobalVariable::InternalLinkage, FrameMap, + "__gc_" + F.getName()); + + Constant *GEPIndices[2] = { + ConstantInt::get(Type::getInt32Ty(F.getContext()), 0), + ConstantInt::get(Type::getInt32Ty(F.getContext()), 0)}; + return ConstantExpr::getGetElementPtr(GV, GEPIndices); +} + +Type *ShadowStackGCLowering::GetConcreteStackEntryType(Function &F) { + // doInitialization creates the generic version of this type. + std::vector EltTys; + EltTys.push_back(StackEntryTy); + for (size_t I = 0; I != Roots.size(); I++) + EltTys.push_back(Roots[I].second->getAllocatedType()); + + return StructType::create(EltTys, "gc_stackentry." + F.getName().str()); +} + +/// doInitialization - If this module uses the GC intrinsics, find them now. If +/// not, exit fast. +bool ShadowStackGCLowering::doInitialization(Module &M) { + bool Active = false; + for (Function &F : M) { + if (F.hasGC() && F.getGC() == std::string("shadow-stack")) { + Active = true; + break; + } + } + if (!Active) + return false; + + // struct FrameMap { + // int32_t NumRoots; // Number of roots in stack frame. + // int32_t NumMeta; // Number of metadata descriptors. May be < NumRoots. + // void *Meta[]; // May be absent for roots without metadata. + // }; + std::vector EltTys; + // 32 bits is ok up to a 32GB stack frame. :) + EltTys.push_back(Type::getInt32Ty(M.getContext())); + // Specifies length of variable length array. + EltTys.push_back(Type::getInt32Ty(M.getContext())); + FrameMapTy = StructType::create(EltTys, "gc_map"); + PointerType *FrameMapPtrTy = PointerType::getUnqual(FrameMapTy); + + // struct StackEntry { + // ShadowStackEntry *Next; // Caller's stack entry. + // FrameMap *Map; // Pointer to constant FrameMap. + // void *Roots[]; // Stack roots (in-place array, so we pretend). + // }; + + StackEntryTy = StructType::create(M.getContext(), "gc_stackentry"); + + EltTys.clear(); + EltTys.push_back(PointerType::getUnqual(StackEntryTy)); + EltTys.push_back(FrameMapPtrTy); + StackEntryTy->setBody(EltTys); + PointerType *StackEntryPtrTy = PointerType::getUnqual(StackEntryTy); + + // Get the root chain if it already exists. + Head = M.getGlobalVariable("llvm_gc_root_chain"); + if (!Head) { + // If the root chain does not exist, insert a new one with linkonce + // linkage! + Head = new GlobalVariable( + M, StackEntryPtrTy, false, GlobalValue::LinkOnceAnyLinkage, + Constant::getNullValue(StackEntryPtrTy), "llvm_gc_root_chain"); + } else if (Head->hasExternalLinkage() && Head->isDeclaration()) { + Head->setInitializer(Constant::getNullValue(StackEntryPtrTy)); + Head->setLinkage(GlobalValue::LinkOnceAnyLinkage); + } + + return true; +} + +bool ShadowStackGCLowering::IsNullValue(Value *V) { + if (Constant *C = dyn_cast(V)) + return C->isNullValue(); + return false; +} + +void ShadowStackGCLowering::CollectRoots(Function &F) { + // FIXME: Account for original alignment. Could fragment the root array. + // Approach 1: Null initialize empty slots at runtime. Yuck. + // Approach 2: Emit a map of the array instead of just a count. + + assert(Roots.empty() && "Not cleaned up?"); + + SmallVector, 16> MetaRoots; + + for (Function::iterator BB = F.begin(), E = F.end(); BB != E; ++BB) + for (BasicBlock::iterator II = BB->begin(), E = BB->end(); II != E;) + if (IntrinsicInst *CI = dyn_cast(II++)) + if (Function *F = CI->getCalledFunction()) + if (F->getIntrinsicID() == Intrinsic::gcroot) { + std::pair Pair = std::make_pair( + CI, + cast(CI->getArgOperand(0)->stripPointerCasts())); + if (IsNullValue(CI->getArgOperand(1))) + Roots.push_back(Pair); + else + MetaRoots.push_back(Pair); + } + + // Number roots with metadata (usually empty) at the beginning, so that the + // FrameMap::Meta array can be elided. + Roots.insert(Roots.begin(), MetaRoots.begin(), MetaRoots.end()); +} + +GetElementPtrInst *ShadowStackGCLowering::CreateGEP(LLVMContext &Context, + IRBuilder<> &B, Value *BasePtr, + int Idx, int Idx2, + const char *Name) { + Value *Indices[] = {ConstantInt::get(Type::getInt32Ty(Context), 0), + ConstantInt::get(Type::getInt32Ty(Context), Idx), + ConstantInt::get(Type::getInt32Ty(Context), Idx2)}; + Value *Val = B.CreateGEP(BasePtr, Indices, Name); + + assert(isa(Val) && "Unexpected folded constant"); + + return dyn_cast(Val); +} + +GetElementPtrInst *ShadowStackGCLowering::CreateGEP(LLVMContext &Context, + IRBuilder<> &B, Value *BasePtr, + int Idx, const char *Name) { + Value *Indices[] = {ConstantInt::get(Type::getInt32Ty(Context), 0), + ConstantInt::get(Type::getInt32Ty(Context), Idx)}; + Value *Val = B.CreateGEP(BasePtr, Indices, Name); + + assert(isa(Val) && "Unexpected folded constant"); + + return dyn_cast(Val); +} + +/// runOnFunction - Insert code to maintain the shadow stack. +bool ShadowStackGCLowering::runOnFunction(Function &F) { + // Quick exit for functions that do not use the shadow stack GC. + if (!F.hasGC() || + F.getGC() != std::string("shadow-stack")) + return false; + + LLVMContext &Context = F.getContext(); + + // Find calls to llvm.gcroot. + CollectRoots(F); + + // If there are no roots in this function, then there is no need to add a + // stack map entry for it. + if (Roots.empty()) + return false; + + // Build the constant map and figure the type of the shadow stack entry. + Value *FrameMap = GetFrameMap(F); + Type *ConcreteStackEntryTy = GetConcreteStackEntryType(F); + + // Build the shadow stack entry at the very start of the function. + BasicBlock::iterator IP = F.getEntryBlock().begin(); + IRBuilder<> AtEntry(IP->getParent(), IP); + + Instruction *StackEntry = + AtEntry.CreateAlloca(ConcreteStackEntryTy, nullptr, "gc_frame"); + + while (isa(IP)) + ++IP; + AtEntry.SetInsertPoint(IP->getParent(), IP); + + // Initialize the map pointer and load the current head of the shadow stack. + Instruction *CurrentHead = AtEntry.CreateLoad(Head, "gc_currhead"); + Instruction *EntryMapPtr = + CreateGEP(Context, AtEntry, StackEntry, 0, 1, "gc_frame.map"); + AtEntry.CreateStore(FrameMap, EntryMapPtr); + + // After all the allocas... + for (unsigned I = 0, E = Roots.size(); I != E; ++I) { + // For each root, find the corresponding slot in the aggregate... + Value *SlotPtr = CreateGEP(Context, AtEntry, StackEntry, 1 + I, "gc_root"); + + // And use it in lieu of the alloca. + AllocaInst *OriginalAlloca = Roots[I].second; + SlotPtr->takeName(OriginalAlloca); + OriginalAlloca->replaceAllUsesWith(SlotPtr); + } + + // Move past the original stores inserted by GCStrategy::InitRoots. This isn't + // really necessary (the collector would never see the intermediate state at + // runtime), but it's nicer not to push the half-initialized entry onto the + // shadow stack. + while (isa(IP)) + ++IP; + AtEntry.SetInsertPoint(IP->getParent(), IP); + + // Push the entry onto the shadow stack. + Instruction *EntryNextPtr = + CreateGEP(Context, AtEntry, StackEntry, 0, 0, "gc_frame.next"); + Instruction *NewHeadVal = + CreateGEP(Context, AtEntry, StackEntry, 0, "gc_newhead"); + AtEntry.CreateStore(CurrentHead, EntryNextPtr); + AtEntry.CreateStore(NewHeadVal, Head); + + // For each instruction that escapes... + EscapeEnumerator EE(F, "gc_cleanup"); + while (IRBuilder<> *AtExit = EE.Next()) { + // Pop the entry from the shadow stack. Don't reuse CurrentHead from + // AtEntry, since that would make the value live for the entire function. + Instruction *EntryNextPtr2 = + CreateGEP(Context, *AtExit, StackEntry, 0, 0, "gc_frame.next"); + Value *SavedHead = AtExit->CreateLoad(EntryNextPtr2, "gc_savedhead"); + AtExit->CreateStore(SavedHead, Head); + } + + // Delete the original allocas (which are no longer used) and the intrinsic + // calls (which are no longer valid). Doing this last avoids invalidating + // iterators. + for (unsigned I = 0, E = Roots.size(); I != E; ++I) { + Roots[I].first->eraseFromParent(); + Roots[I].second->eraseFromParent(); + } + + Roots.clear(); + return true; +}