blob: 0016047a134e9ce9bcb9fb00785ab933f0039a8e [file] [log] [blame]
//===-- ShadowStackGC.cpp - GC support for uncooperative targets ----------===//
// The LLVM Compiler Infrastructure
// This file is distributed under the University of Illinois Open Source
// License. See LICENSE.TXT for details.
// This file implements lowering for the llvm.gc* intrinsics for targets that do
// not natively support them (which includes the C backend). Note that the code
// generated is not quite as efficient as algorithms which generate stack maps
// to identify roots.
// This pass implements the code transformation described in this paper:
// "Accurate Garbage Collection in an Uncooperative Environment"
// Fergus Henderson, ISMM, 2002
// In runtime/GC/SemiSpace.cpp is a prototype runtime which is compatible with
// ShadowStackGC.
// In order to support this particular transformation, all stack roots are
// coallocated in the stack. This allows a fully target-independent stack map
// while introducing only minor runtime overhead.
#define DEBUG_TYPE "shadowstackgc"
#include "llvm/CodeGen/GCs.h"
#include "llvm/ADT/StringExtras.h"
#include "llvm/CodeGen/GCStrategy.h"
#include "llvm/IntrinsicInst.h"
#include "llvm/Module.h"
#include "llvm/Support/CallSite.h"
#include "llvm/Support/IRBuilder.h"
using namespace llvm;
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<std::pair<CallInst*,AllocaInst*> > Roots;
bool initializeCustomLowering(Module &M);
bool performCustomLowering(Function &F);
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<ShadowStackGC>
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;
EscapeEnumerator(Function &F, const char *N = "cleanup")
: F(F), CleanupBBName(N), State(0), Builder(F.getContext()) {}
IRBuilder<> *Next() {
switch (State) {
return 0;
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<ReturnInst>(TI) && !isa<ResumeInst>(TI))
Builder.SetInsertPoint(TI->getParent(), TI);
return &Builder;
State = 2;
// Find all 'call' instructions.
SmallVector<Instruction*,16> 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<CallInst>(II))
if (!CI->getCalledFunction() ||
if (Calls.empty())
return 0;
// 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), NULL);
Constant *PersFn =
FunctionType::get(Type::getInt32Ty(C), true));
LandingPadInst *LPad = LandingPadInst::Create(ExnTy, PersFn, 1,
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<Value*,16> Args;
for (unsigned I = Calls.size(); I != 0; ) {
CallInst *CI = cast<CallInst>(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.
// Create a new invoke instruction.
CallSite CS(CI);
Args.append(CS.arg_begin(), CS.arg_end());
InvokeInst *II = InvokeInst::Create(CI->getCalledValue(),
NewBB, CleanupBB,
Args, CI->getName(), CallBB);
delete CI;
Builder.SetInsertPoint(RI->getParent(), RI);
return &Builder;
// -----------------------------------------------------------------------------
void llvm::linkShadowStackGC() { }
ShadowStackGC::ShadowStackGC() : Head(0), StackEntryTy(0) {
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<Constant*, 16> Metadata;
for (unsigned I = 0; I != Roots.size(); ++I) {
Constant *C = cast<Constant>(Roots[I].first->getArgOperand(1));
if (!C->isNullValue())
NumMeta = I + 1;
Metadata.push_back(ConstantExpr::getBitCast(C, VoidPtr));
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,
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<Type*> EltTys;
for (size_t I = 0; I != Roots.size(); I++)
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<Type*> EltTys;
// 32 bits is ok up to a 32GB stack frame. :)
// Specifies length of variable length array.
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");
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,
} else if (Head->hasExternalLinkage() && Head->isDeclaration()) {
return true;
bool ShadowStackGC::IsNullValue(Value *V) {
if (Constant *C = dyn_cast<Constant>(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<std::pair<CallInst*, AllocaInst*>, 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<IntrinsicInst>(II++))
if (Function *F = CI->getCalledFunction())
if (F->getIntrinsicID() == Intrinsic::gcroot) {
std::pair<CallInst*, AllocaInst*> Pair = std::make_pair(
CI, cast<AllocaInst>(CI->getArgOperand(0)->stripPointerCasts()));
if (IsNullValue(CI->getArgOperand(1)))
// 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<GetElementPtrInst>(Val) && "Unexpected folded constant");
return dyn_cast<GetElementPtrInst>(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<GetElementPtrInst>(Val) && "Unexpected folded constant");
return dyn_cast<GetElementPtrInst>(Val);
/// runOnFunction - Insert code to maintain the shadow stack.
bool ShadowStackGC::performCustomLowering(Function &F) {
LLVMContext &Context = F.getContext();
// Find calls to llvm.gcroot.
// 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, 0,
while (isa<AllocaInst>(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,
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;
// 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<StoreInst>(IP)) ++IP;
AtEntry.SetInsertPoint(IP->getParent(), IP);
// Push the entry onto the shadow stack.
Instruction *EntryNextPtr = CreateGEP(Context, AtEntry,
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,
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) {
return true;