RegionAwareModRefSummarizer.cpp

Go to the documentation of this file.

/*
 * Copyright 2022 Nico Reißmann <nico.reissmann@gmail.com>
 * Copyright 2025 Håvard Krogstie <krogstie.havard@gmail.com>
 * See COPYING for terms of redistribution.
 */
 
#include <jlm/llvm/ir/operators/alloca.hpp>
#include <jlm/llvm/ir/operators/call.hpp>
#include <jlm/llvm/ir/operators/lambda.hpp>
#include <jlm/llvm/ir/operators/Load.hpp>
#include <jlm/llvm/ir/operators/operators.hpp>
#include <jlm/llvm/ir/operators/StdLibIntrinsicOperations.hpp>
#include <jlm/llvm/ir/operators/Store.hpp>
#include <jlm/llvm/ir/RvsdgModule.hpp>
#include <jlm/llvm/ir/Trace.hpp>
#include <jlm/llvm/opt/alias-analyses/AliasAnalysis.hpp>
#include <jlm/llvm/opt/alias-analyses/ModRefSummary.hpp>
#include <jlm/llvm/opt/alias-analyses/PointsToGraph.hpp>
#include <jlm/llvm/opt/alias-analyses/RegionAwareModRefSummarizer.hpp>
#include <jlm/llvm/opt/DeadNodeElimination.hpp>
#include <jlm/rvsdg/lambda.hpp>
#include <jlm/rvsdg/MatchType.hpp>
#include <jlm/rvsdg/simple-node.hpp>
#include <jlm/rvsdg/traverser.hpp>
#include <jlm/util/common.hpp>
#include <jlm/util/Statistics.hpp>
#include <jlm/util/TarjanScc.hpp>
#include <jlm/util/Worklist.hpp>
 
#include <algorithm>
#include <limits>
#include <memory>
#include <optional>
#include <queue>
#include <sstream>
#include <unordered_map>
 
namespace jlm::llvm::aa
{
 
static const bool ENABLE_OPERATION_SIZE_BLOCKING =
    !std::getenv("JLM_DISABLE_OPERATION_SIZE_BLOCKING");
 
static const bool ENABLE_CONSTANT_MEMORY_BLOCKING =
    !std::getenv("JLM_DISABLE_CONSTANT_MEMORY_BLOCKING");
 
static const bool ENABLE_READ_ONLY_DETECTION = !std::getenv("JLM_DISABLE_READ_ONLY_DETECTION");
 
static const bool ENABLE_FUNCTION_SIMPLE_ALLOCA_ALLOWLIST =
    !std::getenv("JLM_DISABLE_FUNCTION_SIMPLE_ALLOCA_ALLOWLIST");
 
static const bool ENABLE_CALL_SIMPLE_ALLOCA_ALLOWLIST =
    !std::getenv("JLM_DISABLE_CALL_SIMPLE_ALLOCA_ALLOWLIST");
 
static const bool ENABLE_EXTERN_SIMPLE_ALLOCA_ALLOWLIST =
    !std::getenv("JLM_DISABLE_EXTERN_SIMPLE_ALLOCA_ALLOWLIST");
 
class RegionAwareModRefSummarizer::Statistics final : public util::Statistics
{
  static constexpr auto NumRvsdgRegionsLabel_ = "#RvsdgRegions";
  static constexpr auto NumSimpleAllocas_ = "#SimpleAllocas";
  static constexpr auto NumFunctions_ = "#Functions";
  static constexpr auto NumFunctionsCallingSetjmp_ = "#FunctionsCallingSetjmp";
  static constexpr auto NumReadOnlyMemoryNodesDetectedLabel_ = "#ReadOnlyMemoryNodesDetected";
 
  static constexpr auto NumModRefSetsMaterializedLabel_ = "#ModRefSetsMaterialized";
  static constexpr auto ModRefSetSizeBeforeMaterializationLabel_ =
      "ModRefSetSizeBeforeMaterialization";
  static constexpr auto ModRefSetSizeAfterFilteringLabel_ = "ModRefSetSizeAfterFiltering";
  static constexpr auto NumModRefSetsWithEffectOnExternalLabel_ = "#ModRefSetsWithEffectOnExternal";
  static constexpr auto NumModRefSetsCallingExternalFunctionLabel_ =
      "#ModRefSetsCallingExternalFunction";
  static constexpr auto ModRefSetSizeAfterMaterializationLabel_ =
      "ModRefSetSizeAfterMaterialization";
 
  static constexpr auto SimpleAllocasSetTimer_ = "SimpleAllocasSetTimer";
  static constexpr auto AnnotationTimer_ = "AnnotationTimer";
  static constexpr auto SolvingTimer_ = "SolvingTimer";
  static constexpr auto ReadOnlyDetectionTimer_ = "ReadOnlyDetectionTimer";
  static constexpr auto ModRefSetMaterializationTimer_ = "ModRefSetMaterializationTimer";
 
  static constexpr auto memoryStateDistributionLabel_ = "MemoryStateDistribution";
 
public:
  ~Statistics() override = default;
 
  explicit Statistics(const rvsdg::RvsdgModule & rvsdgModule, const PointsToGraph & pointsToGraph)
      : util::Statistics(Id::RegionAwareModRefSummarizer, rvsdgModule.SourceFilePath().value())
  {
    AddMeasurement(Label::NumRvsdgNodes, rvsdg::nnodes(&rvsdgModule.Rvsdg().GetRootRegion()));
    AddMeasurement(
        NumRvsdgRegionsLabel_,
        rvsdg::Region::NumRegions(rvsdgModule.Rvsdg().GetRootRegion()));
    AddMeasurement(Label::NumPointsToGraphMemoryNodes, pointsToGraph.numMemoryNodes());
  }
 
  void
  StartCreateSimpleAllocasSetStatistics()
  {
    AddTimer(SimpleAllocasSetTimer_).start();
  }
 
  void
  StopCreateSimpleAllocasSetStatistics(uint64_t numSimpleAllocas)
  {
    GetTimer(SimpleAllocasSetTimer_).stop();
    AddMeasurement(NumSimpleAllocas_, numSimpleAllocas);
  }
 
  void
  StartAnnotationStatistics()
  {
    AddTimer(AnnotationTimer_).start();
  }
 
  void
  StopAnnotationStatistics(size_t numFunctions, size_t numFunctionsCallingSetjmp)
  {
    GetTimer(AnnotationTimer_).stop();
    AddMeasurement(NumFunctions_, numFunctions);
    AddMeasurement(NumFunctionsCallingSetjmp_, numFunctionsCallingSetjmp);
  }
 
  void
  StartSolvingStatistics()
  {
    AddTimer(SolvingTimer_).start();
  }
 
  void
  StopSolvingStatistics()
  {
    GetTimer(SolvingTimer_).stop();
  }
 
  void
  startReadOnlyMemoryDetectionStatistics()
  {
    AddTimer(ReadOnlyDetectionTimer_).start();
  }
 
  void
  stopReadOnlyMemoryDetectionStatistics(size_t numReadOnlyMemoryNodesDetected)
  {
    GetTimer(ReadOnlyDetectionTimer_).stop();
    AddMeasurement(NumReadOnlyMemoryNodesDetectedLabel_, numReadOnlyMemoryNodesDetected);
  }
 
  void
  startModRefSetMaterializationStatistics()
  {
    AddTimer(ModRefSetMaterializationTimer_).start();
  }
 
  void
  stopModRefSetMaterializationStatistics(
      size_t numModRefSetsMaterialized,
      size_t modRefSetSizeBeforeMaterialization,
      size_t modRefSetSizeAfterFiltering,
      size_t numModRefSetsWithEffectOnExternal,
      size_t numModRefSetsCallingExternalFunction,
      size_t modRefSetSizeAfterMaterialization)
  {
    GetTimer(ModRefSetMaterializationTimer_).stop();
    AddMeasurement(NumModRefSetsMaterializedLabel_, numModRefSetsMaterialized);
    AddMeasurement(ModRefSetSizeBeforeMaterializationLabel_, modRefSetSizeBeforeMaterialization);
    AddMeasurement(ModRefSetSizeAfterFilteringLabel_, modRefSetSizeAfterFiltering);
    AddMeasurement(NumModRefSetsWithEffectOnExternalLabel_, numModRefSetsWithEffectOnExternal);
    AddMeasurement(
        NumModRefSetsCallingExternalFunctionLabel_,
        numModRefSetsCallingExternalFunction);
    AddMeasurement(ModRefSetSizeAfterMaterializationLabel_, modRefSetSizeAfterMaterialization);
  }
 
  void
  addMemoryStateDistribution(const std::vector<MemoryStateSummary> & memoryStateDistribution)
  {
    AddMeasurement(memoryStateDistributionLabel_, toString(memoryStateDistribution));
  }
 
  static std::unique_ptr<Statistics>
  Create(const rvsdg::RvsdgModule & rvsdgModule, const PointsToGraph & pointsToGraph)
  {
    return std::make_unique<Statistics>(rvsdgModule, pointsToGraph);
  }
};
 
class RegionAwareModRefSet final : public ModRefSet
{
public:
  // The byte size used when no access to external is made
  static constexpr uint32_t NoneSize = std::numeric_limits<uint32_t>::max();
 
  RegionAwareModRefSet() = default;
 
  [[nodiscard]] std::optional<size_t>
  getRefExternalMinSize() const
  {
    if (refExternalOfSize_ == NoneSize)
      return std::nullopt;
    return refExternalOfSize_;
  }
 
  [[nodiscard]] std::optional<size_t>
  getModExternalMinSize() const
  {
    if (modExternalOfSize_ == NoneSize)
      return std::nullopt;
    return modExternalOfSize_;
  }
 
  bool
  markAsReferencingExternal(size_t minSize)
  {
    // Make sure we do not overflow or collide with the sentinel NoneSize
    minSize = std::min<size_t>(minSize, NoneSize - 1);
 
    if (minSize < refExternalOfSize_)
    {
      refExternalOfSize_ = minSize;
      return true;
    }
    return false;
  }
 
  bool
  markAsModifyingExternal(size_t minSize)
  {
    // Make sure we do not overflow or collide with the sentinel NoneSize
    minSize = std::min<size_t>(minSize, NoneSize - 1);
 
    if (minSize < modExternalOfSize_)
    {
      modExternalOfSize_ = minSize;
      return true;
    }
    return false;
  }
 
  [[nodiscard]] ModRefEffect
  getImplicitModRefEffectForExternal(std::optional<size_t> size) const noexcept
  {
    ModRefEffect result = ModRefEffect::NoEffect;
    if (refExternalOfSize_ != NoneSize)
    {
      if (size.value_or(NoneSize) >= refExternalOfSize_)
        result |= ModRefEffect::RefOnly;
    }
    if (modExternalOfSize_ != NoneSize)
    {
      if (size.value_or(NoneSize) >= modExternalOfSize_)
        result |= ModRefEffect::ModOnly;
    }
    return result;
  }
 
  [[nodiscard]] bool
  mayCallExternalFunction() const
  {
    return callsExternalFunction_;
  }
 
  bool
  markAsCallingExternalFunction()
  {
    if (callsExternalFunction_)
      return false;
 
    callsExternalFunction_ = true;
 
    // calls to external functions may reference and modify any externally available memory
    refExternalOfSize_ = 0;
    modExternalOfSize_ = 0;
    return true;
  }
 
  bool
  addExplicitMemoryNode(PointsToGraph::NodeIndex memoryNode, ModRefEffect modRefEffect)
  {
    JLM_ASSERT(modRefEffect != ModRefEffect::NoEffect);
 
    const auto [it, inserted] = modRefNodes_.insert({ memoryNode, modRefEffect });
    if (inserted)
      return true;
 
    // The memory node was already present, but we may add more effects
    auto oldEffects = it->second;
    it->second = oldEffects | modRefEffect;
    return it->second != oldEffects;
  }
 
  bool
  addMemoryNode(
      PointsToGraph::NodeIndex memoryNode,
      bool isExternallyAvailable,
      std::optional<size_t> memoryNodeSize,
      ModRefEffect modRefEffect)
  {
    JLM_ASSERT(modRefEffect != ModRefEffect::NoEffect);
 
    // If the effects on the node are already encoded implicitly, skip adding them explicitly
    if (isExternallyAvailable)
    {
      const auto implicitEffect = getImplicitModRefEffectForExternal(memoryNodeSize);
      if (isEffectSubset(modRefEffect, implicitEffect))
        return false;
    }
 
    return addExplicitMemoryNode(memoryNode, modRefEffect);
  }
 
  void
  keepSubsetOfExplicitMemoryNodes(const util::HashSet<PointsToGraph::NodeIndex> & filter)
  {
    auto it = modRefNodes_.begin();
    while (it != modRefNodes_.end())
    {
      if (filter.Contains(it->first))
        ++it;
      else
        it = modRefNodes_.erase(it);
    }
  }
 
  bool
  propagateFlags(const RegionAwareModRefSet & other)
  {
    // Check the most restrictive flag first
    if (other.callsExternalFunction_)
    {
      if (callsExternalFunction_)
        return false;
 
      callsExternalFunction_ = true;
      JLM_ASSERT(other.refExternalOfSize_ == 0);
      JLM_ASSERT(other.modExternalOfSize_ == 0);
      refExternalOfSize_ = 0;
      modExternalOfSize_ = 0;
      return true;
    }
 
    if (other.modExternalOfSize_ < modExternalOfSize_)
    {
      modExternalOfSize_ = other.modExternalOfSize_;
      refExternalOfSize_ = std::min(refExternalOfSize_, other.refExternalOfSize_);
      return true;
    }
 
    if (other.refExternalOfSize_ < refExternalOfSize_)
    {
      refExternalOfSize_ = other.refExternalOfSize_;
      return true;
    }
 
    return false;
  }
 
private:
  // If not equal to NoneSize, the ModRefSet potentially reads all externally available memory
  // of size >= the given number of bytes.
  uint32_t refExternalOfSize_ = NoneSize;
  // If not equal to NoneSize, the ModRefSet potentially modifies all externally available memory
  // of size >= the given number of bytes.
  uint32_t modExternalOfSize_ = NoneSize;
  // If true, the ModRefSet represents possibly calling externally defined functions.
  // In this case, \ref refExternalOfSize and \ref modExternalOfSize must both be 0
  bool callsExternalFunction_ = false;
};
 
class RegionAwareModRefSummary final : public ModRefSummary
{
public:
  explicit RegionAwareModRefSummary(const PointsToGraph & pointsToGraph)
      : pointsToGraph_(pointsToGraph)
  {
    // Create the ModRefSet representing eveything that can be referenced and modified,
    // directly or indirectly, from external functions.
    externModRefSet_ = createModRefSet();
    // The ModRefSet representing external functions can call external functions
    markSetAsCallingExternalFunction(externModRefSet_);
  }
 
  RegionAwareModRefSummary(const RegionAwareModRefSummary &) = delete;
  RegionAwareModRefSummary &
  operator=(const RegionAwareModRefSummary &) = delete;
 
  [[nodiscard]] const PointsToGraph &
  GetPointsToGraph() const noexcept override
  {
    return pointsToGraph_;
  }
 
  [[nodiscard]] size_t
  NumModRefSets() const noexcept
  {
    return modRefSets_.size();
  }
 
  [[nodiscard]] RegionAwareModRefSet &
  getModRefSet(ModRefSetIndex index)
  {
    JLM_ASSERT(index < modRefSets_.size());
    return modRefSets_[index];
  }
 
  [[nodiscard]] const RegionAwareModRefSet &
  getModRefSet(ModRefSetIndex index) const
  {
    JLM_ASSERT(index < modRefSets_.size());
    return modRefSets_[index];
  }
 
  bool
  markSetAsReferencingExternal(ModRefSetIndex index, size_t minSize)
  {
    return getModRefSet(index).markAsReferencingExternal(minSize);
  }
 
  bool
  markSetAsModifyingExternal(ModRefSetIndex index, size_t minSize)
  {
    return getModRefSet(index).markAsModifyingExternal(minSize);
  }
 
  bool
  markSetAsCallingExternalFunction(ModRefSetIndex index)
  {
    return getModRefSet(index).markAsCallingExternalFunction();
  }
 
  bool
  addMemoryNodeToSet(
      ModRefSetIndex index,
      PointsToGraph::NodeIndex ptgNode,
      ModRefEffect modRefEffect)
  {
    const bool isExternallyAvailable = pointsToGraph_.isExternallyAvailable(ptgNode);
    const std::optional<size_t> memoryNodeSize = pointsToGraph_.tryGetNodeSize(ptgNode);
    return getModRefSet(index)
        .addMemoryNode(ptgNode, isExternallyAvailable, memoryNodeSize, modRefEffect);
  }
 
  bool
  addExplicitMemoryNodeToSet(
      ModRefSetIndex index,
      PointsToGraph::NodeIndex ptgNode,
      ModRefEffect modRefEffect)
  {
    return getModRefSet(index).addExplicitMemoryNode(ptgNode, modRefEffect);
  }
 
  [[nodiscard]] ModRefSetIndex
  getExternModRefSet() const noexcept
  {
    return externModRefSet_;
  }
 
  [[nodiscard]] bool
  hasSetForNode(const rvsdg::Node & node) const
  {
    return nodeMap_.find(&node) != nodeMap_.end();
  }
 
  [[nodiscard]] ModRefSetIndex
  getSetForNode(const rvsdg::Node & node) const
  {
    const auto it = nodeMap_.find(&node);
    JLM_ASSERT(it != nodeMap_.end());
    return it->second;
  }
 
  [[nodiscard]] ModRefSetIndex
  getOrCreateSetForNode(const rvsdg::Node & node, const rvsdg::LambdaNode & lambdaNode)
  {
    auto [it, inserted] = nodeMap_.insert({ &node, 0 });
    if (inserted)
    {
      const auto created = createModRefSet();
      it->second = created;
      modRefSetsInFunction_[&lambdaNode].push_back(created);
      return created;
    }
 
    return it->second;
  }
 
  [[nodiscard]] const std::vector<ModRefSetIndex> &
  getAllSetsInFunction(const rvsdg::LambdaNode & lambdaNode)
  {
    return modRefSetsInFunction_[&lambdaNode];
  }
 
  [[nodiscard]] std::string
  getModRefSetDebugString(ModRefSetIndex index) const
  {
    const auto & modRefSet = getModRefSet(index);
    std::stringstream ss;
    ss << "{MRS#" << index << "; ";
    const auto mayRefMinSize = modRefSet.getRefExternalMinSize();
    if (mayRefMinSize.has_value())
      ss << "RefExt>=" << *mayRefMinSize << " bytes; ";
    const auto mayModMinSize = modRefSet.getModExternalMinSize();
    if (mayModMinSize.has_value())
      ss << "ModExt>=" << *mayModMinSize << " bytes; ";
    if (modRefSet.mayCallExternalFunction())
      ss << "MayCallExt; ";
 
    bool first = true;
    for (auto [memoryNode, modRefEffect] : modRefSet.getModRefNodes())
    {
      if (first)
        first = false;
      else
        ss << ", ";
 
      ss << pointsToGraph_.getNodeDebugString(memoryNode);
      switch (modRefEffect)
      {
      case jlm::llvm::aa::ModRefEffect::NoEffect:
        JLM_UNREACHABLE("Memory nodes should never be added with NoEffect");
      case jlm::llvm::aa::ModRefEffect::RefOnly:
        ss << "[R]";
        break;
      case jlm::llvm::aa::ModRefEffect::ModOnly:
        ss << "[M]";
        break;
      case jlm::llvm::aa::ModRefEffect::ModRef:
        ss << "[MR]";
        break;
      }
    }
    ss << "}";
    return ss.str();
  }
 
  const ModRefSet &
  GetSimpleNodeModRef(const rvsdg::SimpleNode & node) const override
  {
    return modRefSets_[getSetForNode(node)];
  }
 
  const ModRefSet &
  GetGammaEntryModRef(const rvsdg::GammaNode & gamma) const override
  {
    return modRefSets_[getSetForNode(gamma)];
  }
 
  const ModRefSet &
  GetGammaExitModRef(const rvsdg::GammaNode & gamma) const override
  {
    return GetGammaEntryModRef(gamma);
  }
 
  const ModRefSet &
  GetThetaModRef(const rvsdg::ThetaNode & theta) const override
  {
    return modRefSets_[getSetForNode(theta)];
  }
 
  const ModRefSet &
  GetLambdaEntryModRef(const rvsdg::LambdaNode & lambda) const override
  {
    return modRefSets_[getSetForNode(lambda)];
  }
 
  const ModRefSet &
  GetLambdaExitModRef(const rvsdg::LambdaNode & lambda) const override
  {
    return GetLambdaEntryModRef(lambda);
  }
 
  [[nodiscard]] static std::unique_ptr<RegionAwareModRefSummary>
  Create(const PointsToGraph & pointsToGraph)
  {
    return std::make_unique<RegionAwareModRefSummary>(pointsToGraph);
  }
 
private:
  [[nodiscard]] ModRefSetIndex
  createModRefSet()
  {
    modRefSets_.emplace_back();
    return modRefSets_.size() - 1;
  }
 
  const PointsToGraph & pointsToGraph_;
 
  std::vector<RegionAwareModRefSet> modRefSets_;
 
  ModRefSetIndex externModRefSet_;
 
  std::unordered_map<const rvsdg::LambdaNode *, std::vector<ModRefSetIndex>> modRefSetsInFunction_;
 
  std::unordered_map<const rvsdg::Node *, ModRefSetIndex> nodeMap_;
};
 
struct RegionAwareModRefSummarizer::Context
{
  explicit Context(const PointsToGraph & ptg)
      : pointsToGraph(ptg)
  {}
 
  const PointsToGraph & pointsToGraph;
 
  util::HashSet<PointsToGraph::NodeIndex> SimpleAllocas;
 
  std::vector<const rvsdg::LambdaNode *> Functions;
 
  util::HashSet<const rvsdg::LambdaNode *> FunctionsCallingSetjmp;
 
  std::vector<std::vector<ModRefSetIndex>> ModRefSetSimpleConstraints;
 
  std::unordered_map<ModRefSetIndex, util::HashSet<PointsToGraph::NodeIndex>>
      ModRefSetSimpleAllocaAllowlist;
 
  util::HashSet<PointsToGraph::NodeIndex> ReadOnlyMemoryNodes;
 
  size_t numModRefSetsMaterialized = 0;
  size_t modRefSetSizeBeforeMaterialization = 0;
  size_t modRefSetSizeAfterFiltering = 0;
  size_t numModRefSetsWithEffectOnExternal = 0;
  size_t numModRefSetsCallingExternalFunction = 0;
  size_t modRefSetSizeAfterMaterialization = 0;
};
 
RegionAwareModRefSummarizer::~RegionAwareModRefSummarizer() noexcept = default;
 
RegionAwareModRefSummarizer::RegionAwareModRefSummarizer() = default;
 
std::unique_ptr<ModRefSummary>
RegionAwareModRefSummarizer::SummarizeModRefs(
    const rvsdg::RvsdgModule & rvsdgModule,
    const PointsToGraph & pointsToGraph,
    util::StatisticsCollector & statisticsCollector)
{
  ModRefSummary_ = RegionAwareModRefSummary::Create(pointsToGraph);
  Context_ = std::make_unique<Context>(pointsToGraph);
  auto statistics = Statistics::Create(rvsdgModule, pointsToGraph);
 
  statistics->StartCreateSimpleAllocasSetStatistics();
  Context_->SimpleAllocas = CreateSimpleAllocaSet(pointsToGraph);
  statistics->StopCreateSimpleAllocasSetStatistics(Context_->SimpleAllocas.Size());
 
  if (ENABLE_EXTERN_SIMPLE_ALLOCA_ALLOWLIST)
  {
    // Use an empty allowlist, since all simple allocas should be blocked
    addModRefSetSimpleAllocaAllowlist(ModRefSummary_->getExternModRefSet(), {});
  }
 
  statistics->StartAnnotationStatistics();
  // Go through and recursively annotate all functions, regions and nodes
  annotateInterproceduralRegion(rvsdgModule.Rvsdg().GetRootRegion());
  statistics->StopAnnotationStatistics(
      Context_->Functions.size(),
      Context_->FunctionsCallingSetjmp.Size());
 
  removeSimpleAllocasAroundSetjmp();
 
  statistics->StartSolvingStatistics();
  SolveModRefSetConstraintGraph();
  statistics->StopSolvingStatistics();
 
  if (ENABLE_READ_ONLY_DETECTION)
  {
    statistics->startReadOnlyMemoryDetectionStatistics();
    determineReadOnlyMemory();
    statistics->stopReadOnlyMemoryDetectionStatistics(Context_->ReadOnlyMemoryNodes.Size());
  }
 
  statistics->startModRefSetMaterializationStatistics();
  materializeSets();
  statistics->stopModRefSetMaterializationStatistics(
      Context_->numModRefSetsMaterialized,
      Context_->modRefSetSizeBeforeMaterialization,
      Context_->modRefSetSizeAfterFiltering,
      Context_->numModRefSetsWithEffectOnExternal,
      Context_->numModRefSetsCallingExternalFunction,
      Context_->modRefSetSizeAfterMaterialization);
 
  // Perform the collection of the memory state distribution AFTER we invoked stopped collecting
  // everything such that it does not count into the timing measurements
  if (statisticsCollector.IsDemanded(statistics->GetId()))
  {
    const auto distribution = collectMemoryStateDistribution(rvsdgModule.Rvsdg(), *ModRefSummary_);
    statistics->addMemoryStateDistribution(distribution);
  }
 
  statisticsCollector.CollectDemandedStatistics(std::move(statistics));
  Context_.reset();
  return std::move(ModRefSummary_);
}
 
util::HashSet<PointsToGraph::NodeIndex>
RegionAwareModRefSummarizer::CreateSimpleAllocaSet(const PointsToGraph & pointsToGraph)
{
  // The set of allocas that are simple. Starts off as an over-approximation
  util::HashSet<PointsToGraph::NodeIndex> simpleAllocas;
  // A queue used to visit all PtG memory nodes that are not simple allocas
  std::queue<PointsToGraph::NodeIndex> notSimple;
 
  for (PointsToGraph::NodeIndex ptgNode = 0; ptgNode < pointsToGraph.numNodes(); ptgNode++)
  {
    // Only memory nodes are relevant
    if (!pointsToGraph.isMemoryNode(ptgNode))
      continue;
 
    // Allocas that are not externally available start of as presumed simple
    if (pointsToGraph.getNodeKind(ptgNode) == PointsToGraph::NodeKind::AllocaNode
        && !pointsToGraph.isExternallyAvailable(ptgNode))
      simpleAllocas.insert(ptgNode);
    else
      notSimple.push(ptgNode);
  }
 
  // Process the queue to visit all memory nodes that may disqualify allocas from being simple
  while (!notSimple.empty())
  {
    const auto ptgNode = notSimple.front();
    notSimple.pop();
 
    // Any node targeted by the not-simple memory node can themselves not be simple
    for (const auto targetPtgNode : pointsToGraph.getExplicitTargets(ptgNode).Items())
    {
      // If the target is currently in the simple allocas candiate set, move it to the queue
      if (simpleAllocas.Remove(targetPtgNode))
        notSimple.push(targetPtgNode);
    }
  }
 
  return simpleAllocas;
}
 
util::HashSet<PointsToGraph::NodeIndex>
RegionAwareModRefSummarizer::getReachableSimpleAllocas(std::queue<PointsToGraph::NodeIndex> & nodes)
{
  const auto & pointsToGraph = Context_->pointsToGraph;
 
  util::HashSet<PointsToGraph::NodeIndex> reachableSimpleAllocas;
  // Traverse along PointsToGraph edges to find all reachable simple allocas
  while (!nodes.empty())
  {
    const auto ptgNode = nodes.front();
    nodes.pop();
 
    for (const auto targetPtgNode : pointsToGraph.getExplicitTargets(ptgNode).Items())
    {
      // We only are about following simple allocas, as simple allocas are only reachable from them.
      if (!Context_->SimpleAllocas.Contains(targetPtgNode))
        continue;
 
      if (reachableSimpleAllocas.insert(targetPtgNode))
        nodes.push(targetPtgNode);
    }
  }
 
  return reachableSimpleAllocas;
}
 
util::HashSet<PointsToGraph::NodeIndex>
RegionAwareModRefSummarizer::getSimpleAllocasReachableFromRegionArguments(
    const rvsdg::Region & region)
{
  const auto & pointsToGraph = Context_->pointsToGraph;
 
  // Start by finding initial register nodes
  std::queue<PointsToGraph::NodeIndex> nodes;
  for (auto argument : region.Arguments())
  {
    if (!IsPointerCompatible(*argument))
      continue;
    const auto ptgNode = pointsToGraph.getNodeForRegister(*argument);
    nodes.push(ptgNode);
  }
 
  return getReachableSimpleAllocas(nodes);
}
 
util::HashSet<PointsToGraph::NodeIndex>
RegionAwareModRefSummarizer::getSimpleAllocasReachableFromCallArguments(
    const rvsdg::SimpleNode & call)
{
  const auto & pointsToGraph = Context_->pointsToGraph;
 
  // Use a queue and a set to traverse the PointsToGraph
  std::queue<PointsToGraph::NodeIndex> nodes;
  auto numArguments = CallOperation::NumArguments(call);
  for (size_t i = 0; i < numArguments; i++)
  {
    const auto & argument = *CallOperation::Argument(call, i)->origin();
 
    if (!IsPointerCompatible(argument))
      continue;
    const auto ptgNode = pointsToGraph.getNodeForRegister(argument);
    nodes.push(ptgNode);
  }
 
  return getReachableSimpleAllocas(nodes);
}
 
void
RegionAwareModRefSummarizer::removeSimpleAllocasAroundSetjmp()
{
  for (auto setjmpCaller : Context_->FunctionsCallingSetjmp.Items())
  {
    const auto reachableFromArguments =
        getSimpleAllocasReachableFromRegionArguments(*setjmpCaller->subregion());
    Context_->SimpleAllocas.DifferenceWith(reachableFromArguments);
  }
}
 
void
RegionAwareModRefSummarizer::AddModRefSimpleConstraint(ModRefSetIndex from, ModRefSetIndex to)
{
  // We should never add outgoing edges from the set representing all external functions
  JLM_ASSERT(from != ModRefSummary_->getExternModRefSet());
  // Ensure the constraint vector is large enough
  Context_->ModRefSetSimpleConstraints.resize(ModRefSummary_->NumModRefSets());
  Context_->ModRefSetSimpleConstraints[from].push_back(to);
}
 
void
RegionAwareModRefSummarizer::addModRefSetSimpleAllocaAllowlist(
    ModRefSetIndex index,
    util::HashSet<PointsToGraph::NodeIndex> blocklist)
{
  auto [_, inserted] =
      Context_->ModRefSetSimpleAllocaAllowlist.emplace(std::make_pair(index, std::move(blocklist)));
  JLM_ASSERT(inserted);
}
 
void
RegionAwareModRefSummarizer::annotateInterproceduralRegion(const rvsdg::Region & region)
{
  for (auto & node : region.Nodes())
  {
    rvsdg::MatchType(
        node,
        [&](const rvsdg::PhiNode & phi)
        {
          annotateInterproceduralRegion(*phi.subregion());
        },
        [&](const rvsdg::LambdaNode & lambda)
        {
          AnnotateFunction(lambda);
        });
  }
}
 
void
RegionAwareModRefSummarizer::AnnotateFunction(const rvsdg::LambdaNode & lambda)
{
  Context_->Functions.push_back(&lambda);
 
  const auto modRefSet = AnnotateStructuralNode(lambda, lambda);
 
  // Prevent memory nodes from being added to the ModRefSet of the lambda,
  // if the memory node represents a simple alloca that is not reachable from arguments
  if (ENABLE_FUNCTION_SIMPLE_ALLOCA_ALLOWLIST)
  {
    auto allowlist = getSimpleAllocasReachableFromRegionArguments(*lambda.subregion());
    addModRefSetSimpleAllocaAllowlist(modRefSet, std::move(allowlist));
  }
 
  // If the function is externally available, it can be called by external functions,
  // so add a simple edge to the ModRefSet representing all external functions.
  const auto lambdaPtgNode = Context_->pointsToGraph.getNodeForLambda(lambda);
  if (Context_->pointsToGraph.isExternallyAvailable(lambdaPtgNode))
  {
    AddModRefSimpleConstraint(modRefSet, ModRefSummary_->getExternModRefSet());
  }
}
 
void
RegionAwareModRefSummarizer::AnnotateRegion(
    const rvsdg::Region & region,
    ModRefSetIndex modRefSet,
    const rvsdg::LambdaNode & lambda)
{
  for (auto & node : region.Nodes())
  {
    rvsdg::MatchTypeOrFail(
        node,
        [&](const rvsdg::StructuralNode & structuralNode)
        {
          const auto nodeModRefSet = AnnotateStructuralNode(structuralNode, lambda);
          AddModRefSimpleConstraint(nodeModRefSet, modRefSet);
        },
        [&](const rvsdg::SimpleNode & simpleNode)
        {
          if (const auto nodeModRefSet = AnnotateSimpleNode(simpleNode, lambda))
            AddModRefSimpleConstraint(*nodeModRefSet, modRefSet);
        });
  }
}
 
ModRefSetIndex
RegionAwareModRefSummarizer::AnnotateStructuralNode(
    const rvsdg::StructuralNode & structuralNode,
    const rvsdg::LambdaNode & lambda)
{
  // The ModRefSet of a structural node is the same as that of its subregion(s)
  const auto modRefSet = ModRefSummary_->getOrCreateSetForNode(structuralNode, lambda);
 
  for (auto & subregion : structuralNode.Subregions())
  {
    AnnotateRegion(subregion, modRefSet, lambda);
  }
 
  return modRefSet;
}
 
std::optional<ModRefSetIndex>
RegionAwareModRefSummarizer::AnnotateSimpleNode(
    const rvsdg::SimpleNode & simpleNode,
    const rvsdg::LambdaNode & lambda)
{
  return MatchTypeWithDefault(
      simpleNode.GetOperation(),
      [&](const LoadOperation &) -> std::optional<ModRefSetIndex>
      {
        return AnnotateLoad(simpleNode, lambda);
      },
      [&](const StoreOperation &) -> std::optional<ModRefSetIndex>
      {
        return AnnotateStore(simpleNode, lambda);
      },
      [&](const AllocaOperation &) -> std::optional<ModRefSetIndex>
      {
        return AnnotateAlloca(simpleNode, lambda);
      },
      [&](const MallocOperation &) -> std::optional<ModRefSetIndex>
      {
        return AnnotateMalloc(simpleNode, lambda);
      },
      [&](const FreeOperation &) -> std::optional<ModRefSetIndex>
      {
        return AnnotateFree(simpleNode, lambda);
      },
      [&](const MemCpyOperation &) -> std::optional<ModRefSetIndex>
      {
        return AnnotateMemcpy(simpleNode, lambda);
      },
      [&](const MemSetOperation &) -> std::optional<ModRefSetIndex>
      {
        return AnnotateMemset(simpleNode, lambda);
      },
      [&](const MemMoveOperation &) -> std::optional<ModRefSetIndex>
      {
        return AnnotateMemmove(simpleNode, lambda);
      },
      [&](const CallOperation &) -> std::optional<ModRefSetIndex>
      {
        return AnnotateCall(simpleNode, lambda);
      },
      [&](const MemoryStateOperation &) -> std::optional<ModRefSetIndex>
      {
        // MemoryStateOperations are only used to route memory states, and can be ignored
        return std::nullopt;
      },
      [&]() -> std::optional<ModRefSetIndex>
      {
        // Any remaining type of node should not involve any memory states
        JLM_ASSERT(!hasMemoryState(simpleNode));
        return std::nullopt;
      });
}
 
void
RegionAwareModRefSummarizer::addPointerOriginTargets(
    ModRefSetIndex modRefSetIndex,
    const rvsdg::Output & origin,
    std::optional<size_t> minTargetSize,
    ModRefEffect modRefEffect)
{
  const auto & pointsToGraph = Context_->pointsToGraph;
 
  const auto registerPtgNode = pointsToGraph.getNodeForRegister(origin);
 
  const auto tryAddToModRefSet = [&](PointsToGraph::NodeIndex targetPtgNode)
  {
    if (ENABLE_CONSTANT_MEMORY_BLOCKING && pointsToGraph.isNodeConstant(targetPtgNode))
      return;
    if (ENABLE_OPERATION_SIZE_BLOCKING && minTargetSize)
    {
      const auto targetSize = pointsToGraph.tryGetNodeSize(targetPtgNode);
      if (targetSize.has_value() && *targetSize < minTargetSize)
        return;
    }
    ModRefSummary_->addExplicitMemoryNodeToSet(modRefSetIndex, targetPtgNode, modRefEffect);
  };
 
  // If the pointer is targeting everything external, flag the ModRefSet
  if (pointsToGraph.isTargetingAllExternallyAvailable(registerPtgNode))
  {
    if (mayEffectReference(modRefEffect))
    {
      ModRefSummary_->markSetAsReferencingExternal(modRefSetIndex, minTargetSize.value_or(0));
    }
    if (mayEffectModify(modRefEffect))
    {
      ModRefSummary_->markSetAsModifyingExternal(modRefSetIndex, minTargetSize.value_or(0));
    }
  }
 
  for (const auto targetPtgNode : pointsToGraph.getExplicitTargets(registerPtgNode).Items())
  {
    // Assert that the PointsToGraph contains no doubled-up pointees
    JLM_ASSERT(
        !pointsToGraph.isTargetingAllExternallyAvailable(registerPtgNode)
        || !pointsToGraph.isExternallyAvailable(targetPtgNode));
    tryAddToModRefSet(targetPtgNode);
  }
}
 
ModRefSetIndex
RegionAwareModRefSummarizer::AnnotateLoad(
    const rvsdg::SimpleNode & loadNode,
    const rvsdg::LambdaNode & lambda)
{
  const auto nodeModRef = ModRefSummary_->getOrCreateSetForNode(loadNode, lambda);
  const auto origin = LoadOperation::AddressInput(loadNode).origin();
  const auto loadOperation = util::assertedCast<const LoadOperation>(&loadNode.GetOperation());
  const auto loadSize = GetTypeStoreSize(*loadOperation->GetLoadedType());
 
  addPointerOriginTargets(nodeModRef, *origin, loadSize, ModRefEffect::RefOnly);
  return nodeModRef;
}
 
ModRefSetIndex
RegionAwareModRefSummarizer::AnnotateStore(
    const rvsdg::SimpleNode & storeNode,
    const rvsdg::LambdaNode & lambda)
{
  const auto nodeModRef = ModRefSummary_->getOrCreateSetForNode(storeNode, lambda);
  const auto origin = StoreOperation::AddressInput(storeNode).origin();
  const auto storeOperation = util::assertedCast<const StoreOperation>(&storeNode.GetOperation());
  const auto storeSize = GetTypeStoreSize(storeOperation->GetStoredType());
 
  addPointerOriginTargets(nodeModRef, *origin, storeSize, ModRefEffect::ModOnly);
  return nodeModRef;
}
 
ModRefSetIndex
RegionAwareModRefSummarizer::AnnotateAlloca(
    const rvsdg::SimpleNode & allocaNode,
    const rvsdg::LambdaNode & lambda)
{
  const auto nodeModRef = ModRefSummary_->getOrCreateSetForNode(allocaNode, lambda);
  const auto allocaMemoryNode = Context_->pointsToGraph.getNodeForAlloca(allocaNode);
  // The alloca itself is only considered to be a ref, since its value is indeterminite,
  // and any users of the alloca will depend on its address output
  ModRefSummary_->addExplicitMemoryNodeToSet(nodeModRef, allocaMemoryNode, ModRefEffect::RefOnly);
  return nodeModRef;
}
 
ModRefSetIndex
RegionAwareModRefSummarizer::AnnotateMalloc(
    const rvsdg::SimpleNode & mallocNode,
    const rvsdg::LambdaNode & lambda)
{
  const auto nodeModRef = ModRefSummary_->getOrCreateSetForNode(mallocNode, lambda);
  const auto mallocMemoryNode = Context_->pointsToGraph.getNodeForMalloc(mallocNode);
  // The malloc itself is only considered to be a ref, since its value is indeterminite,
  // and any users of the malloc will depend on its address output
  ModRefSummary_->addExplicitMemoryNodeToSet(nodeModRef, mallocMemoryNode, ModRefEffect::RefOnly);
  return nodeModRef;
}
 
ModRefSetIndex
RegionAwareModRefSummarizer::AnnotateFree(
    const rvsdg::SimpleNode & freeNode,
    const rvsdg::LambdaNode & lambda)
{
  JLM_ASSERT(is<FreeOperation>(freeNode.GetOperation()));
 
  const auto nodeModRef = ModRefSummary_->getOrCreateSetForNode(freeNode, lambda);
  const auto origin = FreeOperation::getAddressInput(freeNode).origin();
 
  // TODO: Filter so we only free MallocMemoryNodes
  addPointerOriginTargets(nodeModRef, *origin, std::nullopt, ModRefEffect::ModOnly);
  return nodeModRef;
}
 
ModRefSetIndex
RegionAwareModRefSummarizer::AnnotateMemcpy(
    const rvsdg::SimpleNode & memcpyNode,
    const rvsdg::LambdaNode & lambda)
{
  JLM_ASSERT(is<MemCpyOperation>(memcpyNode.GetOperation()));
 
  const auto nodeModRef = ModRefSummary_->getOrCreateSetForNode(memcpyNode, lambda);
  const auto dstOrigin = MemCpyOperation::destinationInput(memcpyNode).origin();
  const auto srcOrigin = MemCpyOperation::sourceInput(memcpyNode).origin();
  const auto countOrigin = MemCpyOperation::countInput(memcpyNode).origin();
  const auto count = tryGetConstantSignedInteger(*countOrigin);
  addPointerOriginTargets(nodeModRef, *dstOrigin, count, ModRefEffect::ModOnly);
  addPointerOriginTargets(nodeModRef, *srcOrigin, count, ModRefEffect::RefOnly);
  return nodeModRef;
}
 
ModRefSetIndex
RegionAwareModRefSummarizer::AnnotateMemmove(
    const rvsdg::SimpleNode & memmoveNode,
    const rvsdg::LambdaNode & lambda)
{
  JLM_ASSERT(is<MemMoveOperation>(memmoveNode.GetOperation()));
 
  const auto nodeModRef = ModRefSummary_->getOrCreateSetForNode(memmoveNode, lambda);
  const auto dstOrigin = MemMoveOperation::destinationInput(memmoveNode).origin();
  const auto srcOrigin = MemMoveOperation::sourceInput(memmoveNode).origin();
  const auto lengthOrigin = MemMoveOperation::lengthInput(memmoveNode).origin();
  const auto count = tryGetConstantSignedInteger(*lengthOrigin);
  addPointerOriginTargets(nodeModRef, *dstOrigin, count, ModRefEffect::ModOnly);
  addPointerOriginTargets(nodeModRef, *srcOrigin, count, ModRefEffect::RefOnly);
  return nodeModRef;
}
 
ModRefSetIndex
RegionAwareModRefSummarizer::AnnotateMemset(
    const rvsdg::SimpleNode & memsetNode,
    const rvsdg::LambdaNode & lambda)
{
  JLM_ASSERT(is<MemSetOperation>(memsetNode.GetOperation()));
 
  const auto nodeModRef = ModRefSummary_->getOrCreateSetForNode(memsetNode, lambda);
  const auto dstOrigin = MemSetOperation::destinationInput(memsetNode).origin();
  const auto lengthOrigin = MemSetOperation::lengthInput(memsetNode).origin();
  const auto numBytes = tryGetConstantSignedInteger(*lengthOrigin);
  addPointerOriginTargets(nodeModRef, *dstOrigin, numBytes, ModRefEffect::ModOnly);
 
  return nodeModRef;
}
 
ModRefSetIndex
RegionAwareModRefSummarizer::AnnotateCall(
    const rvsdg::SimpleNode & callNode,
    const rvsdg::LambdaNode & lambda)
{
  JLM_ASSERT(is<CallOperation>(callNode.GetOperation()));
 
  const auto & pointsToGraph = Context_->pointsToGraph;
 
  // This ModRefSet represents everything the call may affect
  const auto callModRef = ModRefSummary_->getOrCreateSetForNode(callNode, lambda);
 
  // Go over all possible targets of the call and add them to the call summary
  const auto targetPtr = callNode.input(0)->origin();
  const auto targetPtgNode = Context_->pointsToGraph.getNodeForRegister(*targetPtr);
 
  // Go through all locations the called function pointer may target
  for (const auto calleePtgNode : pointsToGraph.getExplicitTargets(targetPtgNode).Items())
  {
    const auto kind = pointsToGraph.getNodeKind(calleePtgNode);
    if (kind == PointsToGraph::NodeKind::LambdaNode)
    {
      const auto & calleeLambda = pointsToGraph.getLambdaForNode(calleePtgNode);
      const auto targetModRefSet =
          ModRefSummary_->getOrCreateSetForNode(calleeLambda, calleeLambda);
      AddModRefSimpleConstraint(targetModRefSet, callModRef);
    }
    else if (kind == PointsToGraph::NodeKind::ImportNode)
    {
      ModRefSummary_->markSetAsCallingExternalFunction(callModRef);
    }
  }
  if (pointsToGraph.isTargetingAllExternallyAvailable(targetPtgNode))
  {
    ModRefSummary_->markSetAsCallingExternalFunction(callModRef);
  }
 
  // Skip adding memory nodes to the ModRefSet of the call operation if they represent
  // simple allocas, and they are not reachable from any of the call arguments
  if (ENABLE_CALL_SIMPLE_ALLOCA_ALLOWLIST)
  {
    const auto reachableSimpleAllocas = getSimpleAllocasReachableFromCallArguments(callNode);
    addModRefSetSimpleAllocaAllowlist(callModRef, reachableSimpleAllocas);
  }
 
  // If the call targets setjmp, mark the function as possibly calling setjmp
  const auto classification = CallOperation::ClassifyCall(callNode);
  if (classification->isSetjmpCall())
  {
    Context_->FunctionsCallingSetjmp.insert(&lambda);
  }
 
  return callModRef;
}
 
void
RegionAwareModRefSummarizer::SolveModRefSetConstraintGraph()
{
  Context_->ModRefSetSimpleConstraints.resize(ModRefSummary_->NumModRefSets());
  util::TwoPhaseLrfWorklist<ModRefSetIndex> worklist;
 
  // Start by pushing everything to the worklist
  for (ModRefSetIndex i = 0; i < ModRefSummary_->NumModRefSets(); i++)
    worklist.PushWorkItem(i);
 
  while (worklist.HasMoreWorkItems())
  {
    const auto workItem = worklist.PopWorkItem();
 
    const RegionAwareModRefSet & fromSet = ModRefSummary_->getModRefSet(workItem);
 
    // Handle all simple constraints workItem -> target
    for (auto target : Context_->ModRefSetSimpleConstraints[workItem])
    {
      RegionAwareModRefSet & targetSet = ModRefSummary_->getModRefSet(target);
 
      // Propagate flags first, to enable skipping of doubled-up memory nodes
      bool changed = targetSet.propagateFlags(fromSet);
 
      if (auto allowlist = Context_->ModRefSetSimpleAllocaAllowlist.find(target);
          allowlist != Context_->ModRefSetSimpleAllocaAllowlist.end())
      {
        // The target has a simple alloca allowlist, avoid propagating all other simple alloca nodes
        for (auto [memoryNode, mayMod] : fromSet.getModRefNodes())
        {
          if (Context_->SimpleAllocas.Contains(memoryNode))
          {
            if (!allowlist->second.Contains(memoryNode))
              continue;
          }
 
          changed |= ModRefSummary_->addMemoryNodeToSet(target, memoryNode, mayMod);
        }
      }
      else
      {
        // The target does not have an allowlist, so propagate everything
        for (auto [memoryNode, mayMod] : fromSet.getModRefNodes())
        {
          changed |= ModRefSummary_->addMemoryNodeToSet(target, memoryNode, mayMod);
        }
      }
 
      if (changed)
        worklist.PushWorkItem(target);
    }
  }
 
  JLM_ASSERT(verifySimpleAllocaAllowlists());
}
 
bool
RegionAwareModRefSummarizer::verifySimpleAllocaAllowlists() const
{
  // For all ModRefSets where an allowlist has been defined,
  // check that no other simple allocas are included in the ModRefSet
  for (auto & [index, allowlist] : Context_->ModRefSetSimpleAllocaAllowlist)
  {
    for (auto [memoryNode, _] : ModRefSummary_->getModRefSet(index).getModRefNodes())
    {
      if (Context_->SimpleAllocas.Contains(memoryNode) && !allowlist.Contains(memoryNode))
        return false;
    }
  }
  return true;
}
 
void
RegionAwareModRefSummarizer::determineReadOnlyMemory()
{
  const auto externModRefSet = ModRefSummary_->getExternModRefSet();
  for (auto [memoryNode, modRefEffect] :
       ModRefSummary_->getModRefSet(externModRefSet).getModRefNodes())
  {
    // Memory marked as const in the PointsToGraph should not even be in any ModRefSets
    if (ENABLE_CONSTANT_MEMORY_BLOCKING)
      JLM_ASSERT(!ModRefSummary_->GetPointsToGraph().isNodeConstant(memoryNode));
 
    // If the memory is modified in this module, it is not read-only
    if (mayEffectModify(modRefEffect))
      continue;
    // If the memory is externally accessible, it is also not read-only
    if (ModRefSummary_->GetPointsToGraph().isExternallyAvailable(memoryNode))
      continue;
    // If the memory node is of type Alloca, it might not actually be read-only,
    // as non-reentrant allocas can be partially hidden from reaching the extern \ref ModRefSet.
    // This is handled by never treating allocas as "effectively read-only"
    if (ModRefSummary_->GetPointsToGraph().getNodeKind(memoryNode)
        == PointsToGraph::NodeKind::AllocaNode)
      continue;
 
    Context_->ReadOnlyMemoryNodes.insert(memoryNode);
  }
}
 
void
RegionAwareModRefSummarizer::materializeSets()
{
  for (auto function : Context_->Functions)
  {
    materializeSetsInFunction(*function);
  }
}
 
void
RegionAwareModRefSummarizer::materializeSetsInFunction(const rvsdg::LambdaNode & lambda)
{
  const auto & pointsToGraph = ModRefSummary_->GetPointsToGraph();
  // The ModRefSet representing everying that can be modified from external functions
  const auto & externModRefNodes =
      ModRefSummary_->getModRefSet(ModRefSummary_->getExternModRefSet()).getModRefNodes();
 
  // Only memory nodes that appear in ModRefSets without the external memory node should be kept
  util::HashSet<PointsToGraph::NodeIndex> keepMemoryNodes;
 
  // Among memory nodes that should be kept, the ones flagged externally available are added here.
  // When materializing sets, the flags are turned into explicit targets using this list
  std::vector<PointsToGraph::NodeIndex> materializeExternallyAvailable;
  // When a memory node we need to keep is not externally available, yet is in the
  // ModRefSet representing extern functions, the memory node is added to this list.
  // It gets materialized in all ModRefSets flagged as possibly calling external functions.
  std::vector<std::pair<PointsToGraph::NodeIndex, ModRefEffect>> materializeFromCallToExtern;
 
  const auto markToKeep = [&](PointsToGraph::NodeIndex memoryNode)
  {
    // Memory nodes discovered to be read-only should not be kept
    if (Context_->ReadOnlyMemoryNodes.Contains(memoryNode))
      return;
 
    bool inserted = keepMemoryNodes.insert(memoryNode);
    if (!inserted)
      return;
 
    if (pointsToGraph.isExternallyAvailable(memoryNode))
      materializeExternallyAvailable.push_back(memoryNode);
    else if (auto it = externModRefNodes.find(memoryNode); it != externModRefNodes.end())
      materializeFromCallToExtern.push_back(*it);
  };
 
  markToKeep(aa::PointsToGraph::externalMemoryNode);
 
  // Go over all ModRefSets in the function twice
  // The first pass determines which memory nodes to keep
  const auto & allModRefSets = ModRefSummary_->getAllSetsInFunction(lambda);
  for (auto modRefSetIndex : allModRefSets)
  {
    const auto & modRefSet = ModRefSummary_->getModRefSet(modRefSetIndex);
    Context_->numModRefSetsMaterialized++;
    Context_->modRefSetSizeBeforeMaterialization += modRefSet.getModRefNodes().size();
 
    auto effectOnExternalNode = modRefSet.getImplicitModRefEffectForExternal(std::nullopt);
 
    // If this ModRefSet may both reference and modify the external memory node,
    // it will not disqualify any other memory nodes from compression
    if (effectOnExternalNode == ModRefEffect::ModRef)
      continue;
 
    for (auto [memoryNode, modRefEffect] : modRefSet.getModRefNodes())
    {
      JLM_ASSERT(modRefEffect != ModRefEffect::NoEffect);
 
      // If the set has an effect on the memory node that it does not have on the external node,
      // the memory node is disqualified from compression
      if (!isEffectSubset(modRefEffect, effectOnExternalNode))
        markToKeep(memoryNode);
    }
  }
 
  // Now go over all sets again, removing all memory nodes that are not on the keep list
  // and materializing memory nodes that were previously only implicit
  for (auto modRefSetIndex : allModRefSets)
  {
    auto & modRefSet = ModRefSummary_->getModRefSet(modRefSetIndex);
    modRefSet.keepSubsetOfExplicitMemoryNodes(keepMemoryNodes);
    Context_->modRefSetSizeAfterFiltering += modRefSet.getModRefNodes().size();
 
    // Check what effects are implicitly encoded for externally available memory nodes.
    // The flags have a minimum size, so large memory nodes may have more effects than small ones.
    // Setting an unknown size gives the largest possible effect set.
    auto effectOnAllExternalNodes = modRefSet.getImplicitModRefEffectForExternal(1);
    auto effectOnLargeExternalNodes = modRefSet.getImplicitModRefEffectForExternal(std::nullopt);
    JLM_ASSERT(isEffectSubset(effectOnAllExternalNodes, effectOnLargeExternalNodes));
 
    if (effectOnLargeExternalNodes == ModRefEffect::NoEffect)
    {
      // For nodes that do not have any effect on externally available memory,
      // we do not need to do any materialization.
      Context_->modRefSetSizeAfterMaterialization += modRefSet.getModRefNodes().size();
      continue;
    }
 
    // If small and large external memory nodes have the same effects,
    // we can materialize all external memory nodes with that effect.
    if (effectOnLargeExternalNodes == effectOnAllExternalNodes)
    {
      for (auto memoryNode : materializeExternallyAvailable)
      {
        modRefSet.addExplicitMemoryNode(memoryNode, effectOnLargeExternalNodes);
      }
    }
    else
    {
      // Materialize each memory node with the appropriate effects based on its size
      for (auto memoryNode : materializeExternallyAvailable)
      {
        const auto memoryNodeSize = pointsToGraph.tryGetNodeSize(memoryNode);
        const auto modRefEffect = modRefSet.getImplicitModRefEffectForExternal(memoryNodeSize);
        if (modRefEffect != ModRefEffect::NoEffect)
          modRefSet.addExplicitMemoryNode(memoryNode, modRefEffect);
      }
    }
 
    // If this ModRefSet is not only accessing everything that is externally available,
    // but also possibly calling external functions, materialize from the call to external set
    if (modRefSet.mayCallExternalFunction())
    {
      Context_->numModRefSetsCallingExternalFunction++;
      for (auto [memoryNode, modRefEffect] : materializeFromCallToExtern)
      {
        modRefSet.addExplicitMemoryNode(memoryNode, modRefEffect);
      }
    }
 
    Context_->numModRefSetsWithEffectOnExternal++;
    Context_->modRefSetSizeAfterMaterialization += modRefSet.getModRefNodes().size();
  }
}
 
std::string
RegionAwareModRefSummarizer::ToRegionTree(
    const rvsdg::Graph & rvsdg,
    const RegionAwareModRefSummary & modRefSummary)
{
  std::ostringstream ss;
 
  ss << "ExternModRefSet: "
     << modRefSummary.getModRefSetDebugString(modRefSummary.getExternModRefSet()) << std::endl;
 
  auto indent = [&](size_t depth, char c = '-')
  {
    for (size_t i = 0; i < depth; i++)
      ss << c;
  };
 
  std::function<void(const rvsdg::Node &, size_t)> toRegionTree =
      [&](const rvsdg::Node & node, size_t depth)
  {
    // Simple nodes with no ModRefSet can be ignored
    if (dynamic_cast<const rvsdg::SimpleNode *>(&node) && !modRefSummary.hasSetForNode(node))
      return;
 
    indent(depth, '-');
    ss << "node " << node.DebugString() << " NodeID: " << node.GetNodeId() << ": ";
    if (modRefSummary.hasSetForNode(node))
    {
      auto modRefIndex = modRefSummary.getSetForNode(node);
      ss << modRefSummary.getModRefSetDebugString(modRefIndex) << std::endl;
    }
 
    if (auto structuralNode = dynamic_cast<const rvsdg::StructuralNode *>(&node))
    {
      for (auto & region : structuralNode->Subregions())
      {
        indent(depth + 1, '-');
        ss << "RegionID: " << region.getRegionId() << std::endl;
        for (auto & n : region.Nodes())
          toRegionTree(n, depth + 2);
      }
    }
  };
 
  ss << "RootRegion:" << std::endl;
  for (auto & node : rvsdg.GetRootRegion().Nodes())
    toRegionTree(node, 0);
 
  return ss.str();
}
 
std::unique_ptr<ModRefSummary>
RegionAwareModRefSummarizer::Create(
    const rvsdg::RvsdgModule & rvsdgModule,
    const PointsToGraph & pointsToGraph,
    util::StatisticsCollector & statisticsCollector)
{
  RegionAwareModRefSummarizer summarizer;
  return summarizer.SummarizeModRefs(rvsdgModule, pointsToGraph, statisticsCollector);
}
 
std::unique_ptr<ModRefSummary>
RegionAwareModRefSummarizer::Create(
    const rvsdg::RvsdgModule & rvsdgModule,
    const PointsToGraph & pointsToGraph)
{
  util::StatisticsCollector statisticsCollector;
  return Create(rvsdgModule, pointsToGraph, statisticsCollector);
}
}