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++
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 <writing-an-llvm-pass-runOnFunction>`
-Likewise, ``initializeCustomLowering`` has the same semantics as
-:ref:`Pass::doInitialization(Module&)
-<writing-an-llvm-pass-doInitialization-mod>`
+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<IntrinsicInst>(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:
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,
}
/// 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
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
/// information.
extern char &MachineBlockPlacementStatsID;
- /// GCLowering Pass - Performs target-independent LLVM IR transformations for
- /// highly portable strategies.
- ///
+ /// GCLowering Pass - Used by gc.root to perform its 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
void initializeScalarEvolutionAliasAnalysisPass(PassRegistry&);
void initializeScalarEvolutionPass(PassRegistry&);
void initializeSimpleInlinerPass(PassRegistry&);
+void initializeShadowStackGCLoweringPass(PassRegistry&);
void initializeRegisterCoalescerPass(PassRegistry&);
void initializeSingleLoopExtractorPass(PassRegistry&);
void initializeSinkingPass(PassRegistry&);
ScheduleDAGPrinter.cpp
ScoreboardHazardRecognizer.cpp
ShadowStackGC.cpp
+ ShadowStackGCLowering.cpp
SjLjEHPrepare.cpp
SlotIndexes.cpp
SpillPlacement.cpp
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
- // work against the entire module. But this cannot be done at
- // runFunction time (initializeCustomLowering likely needs to change
- // the module).
GCModuleInfo *MI = getAnalysisIfAvailable<GCModuleInfo>();
assert(MI && "LowerIntrinsics didn't require GCModuleInfo!?");
for (Module::iterator I = M.begin(), E = M.end(); I != E; ++I)
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
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<DominatorTreeWrapperPass>())
- DTWP->getDomTree().recalculate(F);
- }
-
return MadeChange;
}
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());
#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<std::pair<CallInst *, AllocaInst *>> 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<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;
-
-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<ReturnInst>(TI) && !isa<ResumeInst>(TI))
- continue;
-
- 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() ||
- !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<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.
- 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<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));
- }
- 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<Type *> 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<Type *> 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<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)))
- 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<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.
- 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<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, 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<StoreInst>(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;
-}
--- /dev/null
+//===-- 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<std::pair<CallInst *, AllocaInst *>> 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<ReturnInst>(TI) && !isa<ResumeInst>(TI))
+ continue;
+
+ 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() ||
+ !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<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.
+ 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<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));
+ }
+ 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<Type *> 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<Type *> 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<Constant>(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<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)))
+ 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<GetElementPtrInst>(Val) && "Unexpected folded constant");
+
+ return dyn_cast<GetElementPtrInst>(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<GetElementPtrInst>(Val) && "Unexpected folded constant");
+
+ return dyn_cast<GetElementPtrInst>(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<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, 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<StoreInst>(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;
+}
projects / oota-llvm.git / commitdiff