llvm/lib/CodeGen/ShadowStackGCLowering.cpp

//===- ShadowStackGCLowering.cpp - Custom lowering for shadow-stack gc ----===//
//
// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
// See https://llvm.org/LICENSE.txt for license information.
// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
//
//===----------------------------------------------------------------------===//
//
// This file contains the custom lowering code required by the shadow-stack GC
// strategy.
//
// This pass implements the code transformation described in this paper:
//   "Accurate Garbage Collection in an Uncooperative Environment"
//   Fergus Henderson, ISMM, 2002
//
//===----------------------------------------------------------------------===//

#include "llvm/CodeGen/ShadowStackGCLowering.h"
#include "llvm/ADT/SmallVector.h"
#include "llvm/ADT/StringExtras.h"
#include "llvm/Analysis/DomTreeUpdater.h"
#include "llvm/CodeGen/GCMetadata.h"
#include "llvm/CodeGen/Passes.h"
#include "llvm/IR/BasicBlock.h"
#include "llvm/IR/Constant.h"
#include "llvm/IR/Constants.h"
#include "llvm/IR/DataLayout.h"
#include "llvm/IR/DerivedTypes.h"
#include "llvm/IR/Dominators.h"
#include "llvm/IR/Function.h"
#include "llvm/IR/GlobalValue.h"
#include "llvm/IR/GlobalVariable.h"
#include "llvm/IR/IRBuilder.h"
#include "llvm/IR/Instructions.h"
#include "llvm/IR/IntrinsicInst.h"
#include "llvm/IR/Intrinsics.h"
#include "llvm/IR/Module.h"
#include "llvm/IR/Type.h"
#include "llvm/IR/Value.h"
#include "llvm/InitializePasses.h"
#include "llvm/Pass.h"
#include "llvm/Support/Alignment.h"
#include "llvm/Support/Casting.h"
#include "llvm/Support/MathExtras.h"
#include "llvm/Transforms/Utils/EscapeEnumerator.h"
#include <cassert>
#include <optional>
#include <utility>
#include <vector>

using namespace llvm;

#define DEBUG_TYPE "shadow-stack-gc-lowering"

namespace {

class ShadowStackGCLoweringImpl {
  /// RootChain - This is the global linked-list that contains the chain of GC
  /// roots.
  GlobalVariable *Head = nullptr;

  StructType *FrameMapTy = nullptr;

  /// 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;

  /// RootOffsets - Byte offsets and sizes of each root within the frame.
  /// Each element is a pair of (offset, size).
  std::vector<std::pair<uint64_t, uint64_t>> RootOffsets;

public:
  ShadowStackGCLoweringImpl() = default;

  bool doInitialization(Module &M);
  bool runOnFunction(Function &F, DomTreeUpdater *DTU);

private:
  bool IsNullValue(Value *V);
  Constant *GetFrameMap(Function &F, uint64_t FrameSizeInPtrs);
  std::pair<uint64_t, Align> ComputeFrameLayout(Function &F);
  void CollectRoots(Function &F);
};

class ShadowStackGCLowering : public FunctionPass {
  ShadowStackGCLoweringImpl Impl;

public:
  static char ID;

  ShadowStackGCLowering();

  bool doInitialization(Module &M) override { return Impl.doInitialization(M); }
  void getAnalysisUsage(AnalysisUsage &AU) const override {
    AU.addPreserved<DominatorTreeWrapperPass>();
  }
  bool runOnFunction(Function &F) override {
    std::optional<DomTreeUpdater> DTU;
    if (auto *DTWP = getAnalysisIfAvailable<DominatorTreeWrapperPass>())
      DTU.emplace(DTWP->getDomTree(), DomTreeUpdater::UpdateStrategy::Lazy);
    return Impl.runOnFunction(F, DTU ? &*DTU : nullptr);
  }
};

} // end anonymous namespace

PreservedAnalyses ShadowStackGCLoweringPass::run(Module &M,
                                                 ModuleAnalysisManager &MAM) {
  auto &Map = MAM.getResult<CollectorMetadataAnalysis>(M);
  if (!Map.contains("shadow-stack"))
    return PreservedAnalyses::all();

  ShadowStackGCLoweringImpl Impl;
  bool Changed = Impl.doInitialization(M);
  for (auto &F : M) {
    auto &FAM =
        MAM.getResult<FunctionAnalysisManagerModuleProxy>(M).getManager();
    auto *DT = FAM.getCachedResult<DominatorTreeAnalysis>(F);
    DomTreeUpdater DTU(DT, DomTreeUpdater::UpdateStrategy::Lazy);
    Changed |= Impl.runOnFunction(F, DT ? &DTU : nullptr);
  }

  if (!Changed)
    return PreservedAnalyses::all();
  PreservedAnalyses PA;
  PA.preserve<DominatorTreeAnalysis>();
  return PA;
}

char ShadowStackGCLowering::ID = 0;
char &llvm::ShadowStackGCLoweringID = ShadowStackGCLowering::ID;

INITIALIZE_PASS_BEGIN(ShadowStackGCLowering, DEBUG_TYPE,
                      "Shadow Stack GC Lowering", false, false)
INITIALIZE_PASS_DEPENDENCY(GCModuleInfo)
INITIALIZE_PASS_DEPENDENCY(DominatorTreeWrapperPass)
INITIALIZE_PASS_END(ShadowStackGCLowering, DEBUG_TYPE,
                    "Shadow Stack GC Lowering", false, false)

FunctionPass *llvm::createShadowStackGCLoweringPass() { return new ShadowStackGCLowering(); }

ShadowStackGCLowering::ShadowStackGCLowering() : FunctionPass(ID) {}

Constant *ShadowStackGCLoweringImpl::GetFrameMap(Function &F,
                                                 uint64_t FrameSizeInPtrs) {
  // doInitialization creates the abstract type of this value.
  Type *VoidPtr = PointerType::getUnqual(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(C);
  }
  Metadata.resize(NumMeta);

  Type *Int32Ty = Type::getInt32Ty(F.getContext());

  Constant *BaseElts[] = {
      ConstantInt::get(Int32Ty, FrameSizeInPtrs, 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))).
  return new GlobalVariable(*F.getParent(), FrameMap->getType(), true,
                                    GlobalVariable::InternalLinkage, FrameMap,
                                    "__gc_" + F.getName());
}

std::pair<uint64_t, Align>
ShadowStackGCLoweringImpl::ComputeFrameLayout(Function &F) {
  // Compute the layout of the shadow stack frame using byte offsets.
  // Layout: [Next ptr | Map ptr | Root 0 | Root 1 | ... | Root N]

  const DataLayout &DL = F.getParent()->getDataLayout();
  uint64_t PtrSize = DL.getPointerSize(0);
  Align PtrAlign = DL.getPointerABIAlignment(0);

  RootOffsets.clear();
  Align MaxAlign = PtrAlign;

  // Offset 0: Next pointer
  // Offset PtrSize: Map pointer
  uint64_t Offset = 2 * PtrSize;

  // Compute offsets and sizes for each root
  for (const std::pair<CallInst *, AllocaInst *> &Root : Roots) {
    AllocaInst *AI = Root.second;
    std::optional<TypeSize> RootSize = AI->getAllocationSize(DL);
    if (!RootSize || !RootSize->isFixed())
      reportFatalUsageError(
          "Intrinsic::gcroot requires a fixed size stack object");
    uint64_t Size = RootSize->getFixedValue();
    Align RootAlign = AI->getAlign();
    MaxAlign = std::max(MaxAlign, RootAlign);

    // Align the offset for this root
    uint64_t AlignedOffset = alignTo(Offset, RootAlign);

    // Store both offset and size as a pair
    RootOffsets.push_back({AlignedOffset, Size});
    Offset = AlignedOffset + Size;
  }

  // Final frame size, aligned to maximum alignment
  uint64_t FrameSize = alignTo(Offset, MaxAlign);
  return {FrameSize, MaxAlign};
}

/// doInitialization - If this module uses the GC intrinsics, find them now. If
/// not, exit fast.
bool ShadowStackGCLoweringImpl::doInitialization(Module &M) {
  bool Active = false;
  for (Function &F : M) {
    if (F.hasGC() && F.getGC() == "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");

  // The shadow stack linked list uses opaque pointers.
  // Each frame is a byte array with: [Next ptr | Map ptr | Roots...]
  PointerType *StackEntryPtrTy = PointerType::getUnqual(M.getContext());

  // 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 ShadowStackGCLoweringImpl::IsNullValue(Value *V) {
  if (Constant *C = dyn_cast<Constant>(V))
    return C->isNullValue();
  return false;
}

void ShadowStackGCLoweringImpl::CollectRoots(Function &F) {
  assert(Roots.empty() && "Not cleaned up?");

  SmallVector<std::pair<CallInst *, AllocaInst *>, 16> MetaRoots;

  for (BasicBlock &BB : F)
    for (Instruction &I : BB)
      if (IntrinsicInst *CI = dyn_cast<IntrinsicInst>(&I))
        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());
}

/// runOnFunction - Insert code to maintain the shadow stack.
bool ShadowStackGCLoweringImpl::runOnFunction(Function &F,
                                              DomTreeUpdater *DTU) {
  // Quick exit for functions that do not use the shadow stack GC.
  if (!F.hasGC() || F.getGC() != "shadow-stack")
    return false;

  LLVMContext &Context = F.getContext();
  const DataLayout &DL = F.getParent()->getDataLayout();

  // 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;

  // Compute frame layout using byte offsets first.
  auto [FrameSize, FrameAlign] = ComputeFrameLayout(F);

  // Build the constant map with frame size in pointer-sized units.
  uint64_t PtrSize = DL.getPointerSize();
  Value *FrameMap = GetFrameMap(F, FrameSize / PtrSize - 2);

  // Build the shadow stack entry at the very start of the function.
  BasicBlock::iterator IP = F.getEntryBlock().begin();
  IRBuilder<> AtEntry(IP->getParent(), IP);
  Type *Int8Ty = Type::getInt8Ty(Context);
  AllocaInst *StackEntry = AtEntry.CreateAlloca(
      ArrayType::get(Int8Ty, FrameSize), nullptr, "gc_frame");
  StackEntry->setAlignment(FrameAlign);

  AtEntry.SetInsertPointPastAllocas(&F);
  IP = AtEntry.GetInsertPoint();

  // Initialize the map pointer and load the current head of the shadow stack.
  Instruction *CurrentHead =
      AtEntry.CreateLoad(AtEntry.getPtrTy(), Head, "gc_currhead");

  // Map pointer is at offset PtrSize (after the Next pointer)
  Value *EntryMapPtr = AtEntry.CreatePtrAdd(
      StackEntry, AtEntry.getInt64(PtrSize), "gc_frame.map");
  AtEntry.CreateStore(FrameMap, EntryMapPtr);

  // Zero out any padding between roots to ensure deterministic frame contents.
  // This includes the region after the map pointer up to the first root.
  uint64_t LastEnd = 2 * PtrSize; // End of Map pointer field
  assert(RootOffsets.size() == Roots.size());
  for (unsigned I = 0, E = Roots.size(); I != E; ++I) {
    auto [RootOffset, RootSize] = RootOffsets[I];

    // Zero any padding before this root
    if (RootOffset > LastEnd) {
      Value *PaddingPtr =
          AtEntry.CreatePtrAdd(StackEntry, AtEntry.getInt64(LastEnd));
      AtEntry.CreateMemSet(PaddingPtr, AtEntry.getInt8(0), RootOffset - LastEnd,
                           Align(1));
    }

    // For each root, compute pointer using precomputed offset
    Value *SlotPtr = AtEntry.CreatePtrAdd(
        StackEntry, AtEntry.getInt64(RootOffset), "gc_root");

    // And use it in lieu of the alloca.
    AllocaInst *OriginalAlloca = Roots[I].second;
    SlotPtr->takeName(OriginalAlloca);
    OriginalAlloca->replaceAllUsesWith(SlotPtr);

    LastEnd = RootOffset + RootSize;
  }

  // Zero any padding at the end of the frame
  if (FrameSize > LastEnd) {
    Value *PaddingPtr =
        AtEntry.CreatePtrAdd(StackEntry, AtEntry.getInt64(LastEnd));
    AtEntry.CreateMemSet(PaddingPtr, AtEntry.getInt8(0), FrameSize - LastEnd,
                         Align(1));
  }

  // 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.
  // Next pointer is at offset 0, so it's just the frame pointer
  AtEntry.CreateStore(CurrentHead, StackEntry);
  // The new head value is also the frame pointer (the linked list links to
  // frame base)
  AtEntry.CreateStore(StackEntry, Head);

  // For each instruction that escapes...
  EscapeEnumerator EE(F, "gc_cleanup", /*HandleExceptions=*/true, DTU);
  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.
    // Next pointer is at offset 0, so load from the frame base
    Value *SavedHead =
        AtExit->CreateLoad(AtExit->getPtrTy(), StackEntry, "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 (std::pair<CallInst *, AllocaInst *> &Root : Roots) {
    Root.first->eraseFromParent();
    Root.second->eraseFromParent();
  }

  Roots.clear();
  RootOffsets.clear();
  return true;
}