llvm/lib/Target/X86/X86LoadValueInjectionLoadHardening.cpp

   1 //==-- X86LoadValueInjectionLoadHardening.cpp - LVI load hardening for x86 --=//
   2 //
   3 // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
   4 // See https://llvm.org/LICENSE.txt for license information.
   5 // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
   6 //
   7 //===----------------------------------------------------------------------===//
   8 ///
   9 /// Description: This pass finds Load Value Injection (LVI) gadgets consisting
  10 /// of a load from memory (i.e., SOURCE), and any operation that may transmit
  11 /// the value loaded from memory over a covert channel, or use the value loaded
  12 /// from memory to determine a branch/call target (i.e., SINK). After finding
  13 /// all such gadgets in a given function, the pass minimally inserts LFENCE
  14 /// instructions in such a manner that the following property is satisfied: for
  15 /// all SOURCE+SINK pairs, all paths in the CFG from SOURCE to SINK contain at
  16 /// least one LFENCE instruction. The algorithm that implements this minimal
  17 /// insertion is influenced by an academic paper that minimally inserts memory
  18 /// fences for high-performance concurrent programs:
  19 ///         http://www.cs.ucr.edu/~lesani/companion/oopsla15/OOPSLA15.pdf
  20 /// The algorithm implemented in this pass is as follows:
  21 /// 1. Build a condensed CFG (i.e., a GadgetGraph) consisting only of the
  22 /// following components:
  23 ///    - SOURCE instructions (also includes function arguments)
  24 ///    - SINK instructions
  25 ///    - Basic block entry points
  26 ///    - Basic block terminators
  27 ///    - LFENCE instructions
  28 /// 2. Analyze the GadgetGraph to determine which SOURCE+SINK pairs (i.e.,
  29 /// gadgets) are already mitigated by existing LFENCEs. If all gadgets have been
  30 /// mitigated, go to step 6.
  31 /// 3. Use a heuristic or plugin to approximate minimal LFENCE insertion.
  32 /// 4. Insert one LFENCE along each CFG edge that was cut in step 3.
  33 /// 5. Go to step 2.
  34 /// 6. If any LFENCEs were inserted, return `true` from runOnMachineFunction()
  35 /// to tell LLVM that the function was modified.
  36 ///
  37 //===----------------------------------------------------------------------===//
  38
  39 #include "ImmutableGraph.h"
  40 #include "X86.h"
  41 #include "X86Subtarget.h"
  42 #include "X86TargetMachine.h"
  43 #include "llvm/ADT/DenseMap.h"
  44 #include "llvm/ADT/DenseSet.h"
  45 #include "llvm/ADT/STLExtras.h"
  46 #include "llvm/ADT/SmallSet.h"
  47 #include "llvm/ADT/Statistic.h"
  48 #include "llvm/ADT/StringRef.h"
  49 #include "llvm/CodeGen/MachineBasicBlock.h"
  50 #include "llvm/CodeGen/MachineDominanceFrontier.h"
  51 #include "llvm/CodeGen/MachineDominators.h"
  52 #include "llvm/CodeGen/MachineFunction.h"
  53 #include "llvm/CodeGen/MachineFunctionPass.h"
  54 #include "llvm/CodeGen/MachineInstr.h"
  55 #include "llvm/CodeGen/MachineInstrBuilder.h"
  56 #include "llvm/CodeGen/MachineLoopInfo.h"
  57 #include "llvm/CodeGen/MachineRegisterInfo.h"
  58 #include "llvm/CodeGen/RDFGraph.h"
  59 #include "llvm/CodeGen/RDFLiveness.h"
  60 #include "llvm/InitializePasses.h"
  61 #include "llvm/Support/CommandLine.h"
  62 #include "llvm/Support/DOTGraphTraits.h"
  63 #include "llvm/Support/Debug.h"
  64 #include "llvm/Support/DynamicLibrary.h"
  65 #include "llvm/Support/GraphWriter.h"
  66 #include "llvm/Support/raw_ostream.h"
  67
  68 using namespace llvm;
  69
  70 #define PASS_KEY "x86-lvi-load"
  71 #define DEBUG_TYPE PASS_KEY
  72
  73 STATISTIC(NumFences, "Number of LFENCEs inserted for LVI mitigation");
  74 STATISTIC(NumFunctionsConsidered, "Number of functions analyzed");
  75 STATISTIC(NumFunctionsMitigated, "Number of functions for which mitigations "
  76                                  "were deployed");
  77 STATISTIC(NumGadgets, "Number of LVI gadgets detected during analysis");
  78
  79 static cl::opt<std::string> OptimizePluginPath(
  80     PASS_KEY "-opt-plugin",
  81     cl::desc("Specify a plugin to optimize LFENCE insertion"), cl::Hidden);
  82
  83 static cl::opt<bool> NoConditionalBranches(
  84     PASS_KEY "-no-cbranch",
  85     cl::desc("Don't treat conditional branches as disclosure gadgets. This "
  86              "may improve performance, at the cost of security."),
  87     cl::init(false), cl::Hidden);
  88
  89 static cl::opt<bool> EmitDot(
  90     PASS_KEY "-dot",
  91     cl::desc(
  92         "For each function, emit a dot graph depicting potential LVI gadgets"),
  93     cl::init(false), cl::Hidden);
  94
  95 static cl::opt<bool> EmitDotOnly(
  96     PASS_KEY "-dot-only",
  97     cl::desc("For each function, emit a dot graph depicting potential LVI "
  98              "gadgets, and do not insert any fences"),
  99     cl::init(false), cl::Hidden);
 100
 101 static cl::opt<bool> EmitDotVerify(
 102     PASS_KEY "-dot-verify",
 103     cl::desc("For each function, emit a dot graph to stdout depicting "
 104              "potential LVI gadgets, used for testing purposes only"),
 105     cl::init(false), cl::Hidden);
 106
 107 static llvm::sys::DynamicLibrary OptimizeDL;
 108 typedef int (*OptimizeCutT)(unsigned int *Nodes, unsigned int NodesSize,
 109                             unsigned int *Edges, int *EdgeValues,
 110                             int *CutEdges /* out */, unsigned int EdgesSize);
 111 static OptimizeCutT OptimizeCut = nullptr;
 112
 113 namespace {
 114
 115 struct MachineGadgetGraph : ImmutableGraph<MachineInstr *, int> {
 116   static constexpr int GadgetEdgeSentinel = -1;
 117   static constexpr MachineInstr *const ArgNodeSentinel = nullptr;
 118
 119   using GraphT = ImmutableGraph<MachineInstr *, int>;
 120   using Node = typename GraphT::Node;
 121   using Edge = typename GraphT::Edge;
 122   using size_type = typename GraphT::size_type;
 123   MachineGadgetGraph(std::unique_ptr<Node[]> Nodes,
 124                      std::unique_ptr<Edge[]> Edges, size_type NodesSize,
 125                      size_type EdgesSize, int NumFences = 0, int NumGadgets = 0)
 126       : GraphT(std::move(Nodes), std::move(Edges), NodesSize, EdgesSize),
 127         NumFences(NumFences), NumGadgets(NumGadgets) {}
 128   static inline bool isCFGEdge(const Edge &E) {
 129     return E.getValue() != GadgetEdgeSentinel;
 130   }
 131   static inline bool isGadgetEdge(const Edge &E) {
 132     return E.getValue() == GadgetEdgeSentinel;
 133   }
 134   int NumFences;
 135   int NumGadgets;
 136 };
 137
 138 class X86LoadValueInjectionLoadHardeningPass : public MachineFunctionPass {
 139 public:
 140   X86LoadValueInjectionLoadHardeningPass() : MachineFunctionPass(ID) {}
 141
 142   StringRef getPassName() const override {
 143     return "X86 Load Value Injection (LVI) Load Hardening";
 144   }
 145   void getAnalysisUsage(AnalysisUsage &AU) const override;
 146   bool runOnMachineFunction(MachineFunction &MF) override;
 147
 148   static char ID;
 149
 150 private:
 151   using GraphBuilder = ImmutableGraphBuilder<MachineGadgetGraph>;
 152   using Edge = MachineGadgetGraph::Edge;
 153   using Node = MachineGadgetGraph::Node;
 154   using EdgeSet = MachineGadgetGraph::EdgeSet;
 155   using NodeSet = MachineGadgetGraph::NodeSet;
 156
 157   const X86Subtarget *STI;
 158   const TargetInstrInfo *TII;
 159   const TargetRegisterInfo *TRI;
 160
 161   std::unique_ptr<MachineGadgetGraph>
 162   getGadgetGraph(MachineFunction &MF, const MachineLoopInfo &MLI,
 163                  const MachineDominatorTree &MDT,
 164                  const MachineDominanceFrontier &MDF) const;
 165   int hardenLoadsWithPlugin(MachineFunction &MF,
 166                             std::unique_ptr<MachineGadgetGraph> Graph) const;
 167   int hardenLoadsWithHeuristic(MachineFunction &MF,
 168                                std::unique_ptr<MachineGadgetGraph> Graph) const;
 169   int elimMitigatedEdgesAndNodes(MachineGadgetGraph &G,
 170                                  EdgeSet &ElimEdges /* in, out */,
 171                                  NodeSet &ElimNodes /* in, out */) const;
 172   std::unique_ptr<MachineGadgetGraph>
 173   trimMitigatedEdges(std::unique_ptr<MachineGadgetGraph> Graph) const;
 174   int insertFences(MachineFunction &MF, MachineGadgetGraph &G,
 175                    EdgeSet &CutEdges /* in, out */) const;
 176   bool instrUsesRegToAccessMemory(const MachineInstr &I, unsigned Reg) const;
 177   bool instrUsesRegToBranch(const MachineInstr &I, unsigned Reg) const;
 178   inline bool isFence(const MachineInstr *MI) const {
 179     return MI && (MI->getOpcode() == X86::LFENCE ||
 180                   (STI->useLVIControlFlowIntegrity() && MI->isCall()));
 181   }
 182 };
 183
 184 } // end anonymous namespace
 185
 186 namespace llvm {
 187
 188 template <>
 189 struct GraphTraits<MachineGadgetGraph *>
 190     : GraphTraits<ImmutableGraph<MachineInstr *, int> *> {};
 191
 192 template <>
 193 struct DOTGraphTraits<MachineGadgetGraph *> : DefaultDOTGraphTraits {
 194   using GraphType = MachineGadgetGraph;
 195   using Traits = llvm::GraphTraits<GraphType *>;
 196   using NodeRef = typename Traits::NodeRef;
 197   using EdgeRef = typename Traits::EdgeRef;
 198   using ChildIteratorType = typename Traits::ChildIteratorType;
 199   using ChildEdgeIteratorType = typename Traits::ChildEdgeIteratorType;
 200
 201   DOTGraphTraits(bool IsSimple = false) : DefaultDOTGraphTraits(IsSimple) {}
 202
 203   std::string getNodeLabel(NodeRef Node, GraphType *) {
 204     if (Node->getValue() == MachineGadgetGraph::ArgNodeSentinel)
 205       return "ARGS";
 206
 207     std::string Str;
 208     raw_string_ostream OS(Str);
 209     OS << *Node->getValue();
 210     return OS.str();
 211   }
 212
 213   static std::string getNodeAttributes(NodeRef Node, GraphType *) {
 214     MachineInstr *MI = Node->getValue();
 215     if (MI == MachineGadgetGraph::ArgNodeSentinel)
 216       return "color = blue";
 217     if (MI->getOpcode() == X86::LFENCE)
 218       return "color = green";
 219     return "";
 220   }
 221
 222   static std::string getEdgeAttributes(NodeRef, ChildIteratorType E,
 223                                        GraphType *) {
 224     int EdgeVal = (*E.getCurrent()).getValue();
 225     return EdgeVal >= 0 ? "label = " + std::to_string(EdgeVal)
 226                         : "color = red, style = \"dashed\"";
 227   }
 228 };
 229
 230 } // end namespace llvm
 231
 232 constexpr MachineInstr *MachineGadgetGraph::ArgNodeSentinel;
 233 constexpr int MachineGadgetGraph::GadgetEdgeSentinel;
 234
 235 char X86LoadValueInjectionLoadHardeningPass::ID = 0;
 236
 237 void X86LoadValueInjectionLoadHardeningPass::getAnalysisUsage(
 238     AnalysisUsage &AU) const {
 239   MachineFunctionPass::getAnalysisUsage(AU);
 240   AU.addRequired<MachineLoopInfo>();
 241   AU.addRequired<MachineDominatorTree>();
 242   AU.addRequired<MachineDominanceFrontier>();
 243   AU.setPreservesCFG();
 244 }
 245
 246 static void writeGadgetGraph(raw_ostream &OS, MachineFunction &MF,
 247                              MachineGadgetGraph *G) {
 248   WriteGraph(OS, G, /*ShortNames*/ false,
 249              "Speculative gadgets for \"" + MF.getName() + "\" function");
 250 }
 251
 252 bool X86LoadValueInjectionLoadHardeningPass::runOnMachineFunction(
 253     MachineFunction &MF) {
 254   LLVM_DEBUG(dbgs() << "***** " << getPassName() << " : " << MF.getName()
 255                     << " *****\n");
 256   STI = &MF.getSubtarget<X86Subtarget>();
 257   if (!STI->useLVILoadHardening())
 258     return false;
 259
 260   // FIXME: support 32-bit
 261   if (!STI->is64Bit())
 262     report_fatal_error("LVI load hardening is only supported on 64-bit", false);
 263
 264   // Don't skip functions with the "optnone" attr but participate in opt-bisect.
 265   const Function &F = MF.getFunction();
 266   if (!F.hasOptNone() && skipFunction(F))
 267     return false;
 268
 269   ++NumFunctionsConsidered;
 270   TII = STI->getInstrInfo();
 271   TRI = STI->getRegisterInfo();
 272   LLVM_DEBUG(dbgs() << "Building gadget graph...\n");
 273   const auto &MLI = getAnalysis<MachineLoopInfo>();
 274   const auto &MDT = getAnalysis<MachineDominatorTree>();
 275   const auto &MDF = getAnalysis<MachineDominanceFrontier>();
 276   std::unique_ptr<MachineGadgetGraph> Graph = getGadgetGraph(MF, MLI, MDT, MDF);
 277   LLVM_DEBUG(dbgs() << "Building gadget graph... Done\n");
 278   if (Graph == nullptr)
 279     return false; // didn't find any gadgets
 280
 281   if (EmitDotVerify) {
 282     writeGadgetGraph(outs(), MF, Graph.get());
 283     return false;
 284   }
 285
 286   if (EmitDot || EmitDotOnly) {
 287     LLVM_DEBUG(dbgs() << "Emitting gadget graph...\n");
 288     std::error_code FileError;
 289     std::string FileName = "lvi.";
 290     FileName += MF.getName();
 291     FileName += ".dot";
 292     raw_fd_ostream FileOut(FileName, FileError);
 293     if (FileError)
 294       errs() << FileError.message();
 295     writeGadgetGraph(FileOut, MF, Graph.get());
 296     FileOut.close();
 297     LLVM_DEBUG(dbgs() << "Emitting gadget graph... Done\n");
 298     if (EmitDotOnly)
 299       return false;
 300   }
 301
 302   int FencesInserted;
 303   if (!OptimizePluginPath.empty()) {
 304     if (!OptimizeDL.isValid()) {
 305       std::string ErrorMsg;
 306       OptimizeDL = llvm::sys::DynamicLibrary::getPermanentLibrary(
 307           OptimizePluginPath.c_str(), &ErrorMsg);
 308       if (!ErrorMsg.empty())
 309         report_fatal_error("Failed to load opt plugin: \"" + ErrorMsg + '\"');
 310       OptimizeCut = (OptimizeCutT)OptimizeDL.getAddressOfSymbol("optimize_cut");
 311       if (!OptimizeCut)
 312         report_fatal_error("Invalid optimization plugin");
 313     }
 314     FencesInserted = hardenLoadsWithPlugin(MF, std::move(Graph));
 315   } else { // Use the default greedy heuristic
 316     FencesInserted = hardenLoadsWithHeuristic(MF, std::move(Graph));
 317   }
 318
 319   if (FencesInserted > 0)
 320     ++NumFunctionsMitigated;
 321   NumFences += FencesInserted;
 322   return (FencesInserted > 0);
 323 }
 324
 325 std::unique_ptr<MachineGadgetGraph>
 326 X86LoadValueInjectionLoadHardeningPass::getGadgetGraph(
 327     MachineFunction &MF, const MachineLoopInfo &MLI,
 328     const MachineDominatorTree &MDT,
 329     const MachineDominanceFrontier &MDF) const {
 330   using namespace rdf;
 331
 332   // Build the Register Dataflow Graph using the RDF framework
 333   TargetOperandInfo TOI{*TII};
 334   DataFlowGraph DFG{MF, *TII, *TRI, MDT, MDF, TOI};
 335   DFG.build();
 336   Liveness L{MF.getRegInfo(), DFG};
 337   L.computePhiInfo();
 338
 339   GraphBuilder Builder;
 340   using GraphIter = typename GraphBuilder::BuilderNodeRef;
 341   DenseMap<MachineInstr *, GraphIter> NodeMap;
 342   int FenceCount = 0, GadgetCount = 0;
 343   auto MaybeAddNode = [&NodeMap, &Builder](MachineInstr *MI) {
 344     auto Ref = NodeMap.find(MI);
 345     if (Ref == NodeMap.end()) {
 346       auto I = Builder.addVertex(MI);
 347       NodeMap[MI] = I;
 348       return std::pair<GraphIter, bool>{I, true};
 349     }
 350     return std::pair<GraphIter, bool>{Ref->getSecond(), false};
 351   };
 352
 353   // The `Transmitters` map memoizes transmitters found for each def. If a def
 354   // has not yet been analyzed, then it will not appear in the map. If a def
 355   // has been analyzed and was determined not to have any transmitters, then
 356   // its list of transmitters will be empty.
 357   DenseMap<NodeId, std::vector<NodeId>> Transmitters;
 358
 359   // Analyze all machine instructions to find gadgets and LFENCEs, adding
 360   // each interesting value to `Nodes`
 361   auto AnalyzeDef = [&](NodeAddr<DefNode *> SourceDef) {
 362     SmallSet<NodeId, 8> UsesVisited, DefsVisited;
 363     std::function<void(NodeAddr<DefNode *>)> AnalyzeDefUseChain =
 364         [&](NodeAddr<DefNode *> Def) {
 365           if (Transmitters.find(Def.Id) != Transmitters.end())
 366             return; // Already analyzed `Def`
 367
 368           // Use RDF to find all the uses of `Def`
 369           rdf::NodeSet Uses;
 370           RegisterRef DefReg = Def.Addr->getRegRef(DFG);
 371           for (auto UseID : L.getAllReachedUses(DefReg, Def)) {
 372             auto Use = DFG.addr<UseNode *>(UseID);
 373             if (Use.Addr->getFlags() & NodeAttrs::PhiRef) { // phi node
 374               NodeAddr<PhiNode *> Phi = Use.Addr->getOwner(DFG);
 375               for (const auto& I : L.getRealUses(Phi.Id)) {
 376                 if (DFG.getPRI().alias(RegisterRef(I.first), DefReg)) {
 377                   for (const auto &UA : I.second)
 378                     Uses.emplace(UA.first);
 379                 }
 380               }
 381             } else { // not a phi node
 382               Uses.emplace(UseID);
 383             }
 384           }
 385
 386           // For each use of `Def`, we want to know whether:
 387           // (1) The use can leak the Def'ed value,
 388           // (2) The use can further propagate the Def'ed value to more defs
 389           for (auto UseID : Uses) {
 390             if (!UsesVisited.insert(UseID).second)
 391               continue; // Already visited this use of `Def`
 392
 393             auto Use = DFG.addr<UseNode *>(UseID);
 394             assert(!(Use.Addr->getFlags() & NodeAttrs::PhiRef));
 395             MachineOperand &UseMO = Use.Addr->getOp();
 396             MachineInstr &UseMI = *UseMO.getParent();
 397             assert(UseMO.isReg());
 398
 399             // We naively assume that an instruction propagates any loaded
 400             // uses to all defs unless the instruction is a call, in which
 401             // case all arguments will be treated as gadget sources during
 402             // analysis of the callee function.
 403             if (UseMI.isCall())
 404               continue;
 405
 406             // Check whether this use can transmit (leak) its value.
 407             if (instrUsesRegToAccessMemory(UseMI, UseMO.getReg()) ||
 408                 (!NoConditionalBranches &&
 409                  instrUsesRegToBranch(UseMI, UseMO.getReg()))) {
 410               Transmitters[Def.Id].push_back(Use.Addr->getOwner(DFG).Id);
 411               if (UseMI.mayLoad())
 412                 continue; // Found a transmitting load -- no need to continue
 413                           // traversing its defs (i.e., this load will become
 414                           // a new gadget source anyways).
 415             }
 416
 417             // Check whether the use propagates to more defs.
 418             NodeAddr<InstrNode *> Owner{Use.Addr->getOwner(DFG)};
 419             rdf::NodeList AnalyzedChildDefs;
 420             for (const auto &ChildDef :
 421                  Owner.Addr->members_if(DataFlowGraph::IsDef, DFG)) {
 422               if (!DefsVisited.insert(ChildDef.Id).second)
 423                 continue; // Already visited this def
 424               if (Def.Addr->getAttrs() & NodeAttrs::Dead)
 425                 continue;
 426               if (Def.Id == ChildDef.Id)
 427                 continue; // `Def` uses itself (e.g., increment loop counter)
 428
 429               AnalyzeDefUseChain(ChildDef);
 430
 431               // `Def` inherits all of its child defs' transmitters.
 432               for (auto TransmitterId : Transmitters[ChildDef.Id])
 433                 Transmitters[Def.Id].push_back(TransmitterId);
 434             }
 435           }
 436
 437           // Note that this statement adds `Def.Id` to the map if no
 438           // transmitters were found for `Def`.
 439           auto &DefTransmitters = Transmitters[Def.Id];
 440
 441           // Remove duplicate transmitters
 442           llvm::sort(DefTransmitters);
 443           DefTransmitters.erase(
 444               std::unique(DefTransmitters.begin(), DefTransmitters.end()),
 445               DefTransmitters.end());
 446         };
 447
 448     // Find all of the transmitters
 449     AnalyzeDefUseChain(SourceDef);
 450     auto &SourceDefTransmitters = Transmitters[SourceDef.Id];
 451     if (SourceDefTransmitters.empty())
 452       return; // No transmitters for `SourceDef`
 453
 454     MachineInstr *Source = SourceDef.Addr->getFlags() & NodeAttrs::PhiRef
 455                                ? MachineGadgetGraph::ArgNodeSentinel
 456                                : SourceDef.Addr->getOp().getParent();
 457     auto GadgetSource = MaybeAddNode(Source);
 458     // Each transmitter is a sink for `SourceDef`.
 459     for (auto TransmitterId : SourceDefTransmitters) {
 460       MachineInstr *Sink = DFG.addr<StmtNode *>(TransmitterId).Addr->getCode();
 461       auto GadgetSink = MaybeAddNode(Sink);
 462       // Add the gadget edge to the graph.
 463       Builder.addEdge(MachineGadgetGraph::GadgetEdgeSentinel,
 464                       GadgetSource.first, GadgetSink.first);
 465       ++GadgetCount;
 466     }
 467   };
 468
 469   LLVM_DEBUG(dbgs() << "Analyzing def-use chains to find gadgets\n");
 470   // Analyze function arguments
 471   NodeAddr<BlockNode *> EntryBlock = DFG.getFunc().Addr->getEntryBlock(DFG);
 472   for (NodeAddr<PhiNode *> ArgPhi :
 473        EntryBlock.Addr->members_if(DataFlowGraph::IsPhi, DFG)) {
 474     NodeList Defs = ArgPhi.Addr->members_if(DataFlowGraph::IsDef, DFG);
 475     llvm::for_each(Defs, AnalyzeDef);
 476   }
 477   // Analyze every instruction in MF
 478   for (NodeAddr<BlockNode *> BA : DFG.getFunc().Addr->members(DFG)) {
 479     for (NodeAddr<StmtNode *> SA :
 480          BA.Addr->members_if(DataFlowGraph::IsCode<NodeAttrs::Stmt>, DFG)) {
 481       MachineInstr *MI = SA.Addr->getCode();
 482       if (isFence(MI)) {
 483         MaybeAddNode(MI);
 484         ++FenceCount;
 485       } else if (MI->mayLoad()) {
 486         NodeList Defs = SA.Addr->members_if(DataFlowGraph::IsDef, DFG);
 487         llvm::for_each(Defs, AnalyzeDef);
 488       }
 489     }
 490   }
 491   LLVM_DEBUG(dbgs() << "Found " << FenceCount << " fences\n");
 492   LLVM_DEBUG(dbgs() << "Found " << GadgetCount << " gadgets\n");
 493   if (GadgetCount == 0)
 494     return nullptr;
 495   NumGadgets += GadgetCount;
 496
 497   // Traverse CFG to build the rest of the graph
 498   SmallSet<MachineBasicBlock *, 8> BlocksVisited;
 499   std::function<void(MachineBasicBlock *, GraphIter, unsigned)> TraverseCFG =
 500       [&](MachineBasicBlock *MBB, GraphIter GI, unsigned ParentDepth) {
 501         unsigned LoopDepth = MLI.getLoopDepth(MBB);
 502         if (!MBB->empty()) {
 503           // Always add the first instruction in each block
 504           auto NI = MBB->begin();
 505           auto BeginBB = MaybeAddNode(&*NI);
 506           Builder.addEdge(ParentDepth, GI, BeginBB.first);
 507           if (!BlocksVisited.insert(MBB).second)
 508             return;
 509
 510           // Add any instructions within the block that are gadget components
 511           GI = BeginBB.first;
 512           while (++NI != MBB->end()) {
 513             auto Ref = NodeMap.find(&*NI);
 514             if (Ref != NodeMap.end()) {
 515               Builder.addEdge(LoopDepth, GI, Ref->getSecond());
 516               GI = Ref->getSecond();
 517             }
 518           }
 519
 520           // Always add the terminator instruction, if one exists
 521           auto T = MBB->getFirstTerminator();
 522           if (T != MBB->end()) {
 523             auto EndBB = MaybeAddNode(&*T);
 524             if (EndBB.second)
 525               Builder.addEdge(LoopDepth, GI, EndBB.first);
 526             GI = EndBB.first;
 527           }
 528         }
 529         for (MachineBasicBlock *Succ : MBB->successors())
 530           TraverseCFG(Succ, GI, LoopDepth);
 531       };
 532   // ArgNodeSentinel is a pseudo-instruction that represents MF args in the
 533   // GadgetGraph
 534   GraphIter ArgNode = MaybeAddNode(MachineGadgetGraph::ArgNodeSentinel).first;
 535   TraverseCFG(&MF.front(), ArgNode, 0);
 536   std::unique_ptr<MachineGadgetGraph> G{Builder.get(FenceCount, GadgetCount)};
 537   LLVM_DEBUG(dbgs() << "Found " << G->nodes_size() << " nodes\n");
 538   return G;
 539 }
 540
 541 // Returns the number of remaining gadget edges that could not be eliminated
 542 int X86LoadValueInjectionLoadHardeningPass::elimMitigatedEdgesAndNodes(
 543     MachineGadgetGraph &G, EdgeSet &ElimEdges /* in, out */,
 544     NodeSet &ElimNodes /* in, out */) const {
 545   if (G.NumFences > 0) {
 546     // Eliminate fences and CFG edges that ingress and egress the fence, as
 547     // they are trivially mitigated.
 548     for (const Edge &E : G.edges()) {
 549       const Node *Dest = E.getDest();
 550       if (isFence(Dest->getValue())) {
 551         ElimNodes.insert(*Dest);
 552         ElimEdges.insert(E);
 553         for (const Edge &DE : Dest->edges())
 554           ElimEdges.insert(DE);
 555       }
 556     }
 557   }
 558
 559   // Find and eliminate gadget edges that have been mitigated.
 560   int MitigatedGadgets = 0, RemainingGadgets = 0;
 561   NodeSet ReachableNodes{G};
 562   for (const Node &RootN : G.nodes()) {
 563     if (llvm::none_of(RootN.edges(), MachineGadgetGraph::isGadgetEdge))
 564       continue; // skip this node if it isn't a gadget source
 565
 566     // Find all of the nodes that are CFG-reachable from RootN using DFS
 567     ReachableNodes.clear();
 568     std::function<void(const Node *, bool)> FindReachableNodes =
 569         [&](const Node *N, bool FirstNode) {
 570           if (!FirstNode)
 571             ReachableNodes.insert(*N);
 572           for (const Edge &E : N->edges()) {
 573             const Node *Dest = E.getDest();
 574             if (MachineGadgetGraph::isCFGEdge(E) && !ElimEdges.contains(E) &&
 575                 !ReachableNodes.contains(*Dest))
 576               FindReachableNodes(Dest, false);
 577           }
 578         };
 579     FindReachableNodes(&RootN, true);
 580
 581     // Any gadget whose sink is unreachable has been mitigated
 582     for (const Edge &E : RootN.edges()) {
 583       if (MachineGadgetGraph::isGadgetEdge(E)) {
 584         if (ReachableNodes.contains(*E.getDest())) {
 585           // This gadget's sink is reachable
 586           ++RemainingGadgets;
 587         } else { // This gadget's sink is unreachable, and therefore mitigated
 588           ++MitigatedGadgets;
 589           ElimEdges.insert(E);
 590         }
 591       }
 592     }
 593   }
 594   return RemainingGadgets;
 595 }
 596
 597 std::unique_ptr<MachineGadgetGraph>
 598 X86LoadValueInjectionLoadHardeningPass::trimMitigatedEdges(
 599     std::unique_ptr<MachineGadgetGraph> Graph) const {
 600   NodeSet ElimNodes{*Graph};
 601   EdgeSet ElimEdges{*Graph};
 602   int RemainingGadgets =
 603       elimMitigatedEdgesAndNodes(*Graph, ElimEdges, ElimNodes);
 604   if (ElimEdges.empty() && ElimNodes.empty()) {
 605     Graph->NumFences = 0;
 606     Graph->NumGadgets = RemainingGadgets;
 607   } else {
 608     Graph = GraphBuilder::trim(*Graph, ElimNodes, ElimEdges, 0 /* NumFences */,
 609                                RemainingGadgets);
 610   }
 611   return Graph;
 612 }
 613
 614 int X86LoadValueInjectionLoadHardeningPass::hardenLoadsWithPlugin(
 615     MachineFunction &MF, std::unique_ptr<MachineGadgetGraph> Graph) const {
 616   int FencesInserted = 0;
 617
 618   do {
 619     LLVM_DEBUG(dbgs() << "Eliminating mitigated paths...\n");
 620     Graph = trimMitigatedEdges(std::move(Graph));
 621     LLVM_DEBUG(dbgs() << "Eliminating mitigated paths... Done\n");
 622     if (Graph->NumGadgets == 0)
 623       break;
 624
 625     LLVM_DEBUG(dbgs() << "Cutting edges...\n");
 626     EdgeSet CutEdges{*Graph};
 627     auto Nodes = std::make_unique<unsigned int[]>(Graph->nodes_size() +
 628                                                   1 /* terminator node */);
 629     auto Edges = std::make_unique<unsigned int[]>(Graph->edges_size());
 630     auto EdgeCuts = std::make_unique<int[]>(Graph->edges_size());
 631     auto EdgeValues = std::make_unique<int[]>(Graph->edges_size());
 632     for (const Node &N : Graph->nodes()) {
 633       Nodes[Graph->getNodeIndex(N)] = Graph->getEdgeIndex(*N.edges_begin());
 634     }
 635     Nodes[Graph->nodes_size()] = Graph->edges_size(); // terminator node
 636     for (const Edge &E : Graph->edges()) {
 637       Edges[Graph->getEdgeIndex(E)] = Graph->getNodeIndex(*E.getDest());
 638       EdgeValues[Graph->getEdgeIndex(E)] = E.getValue();
 639     }
 640     OptimizeCut(Nodes.get(), Graph->nodes_size(), Edges.get(), EdgeValues.get(),
 641                 EdgeCuts.get(), Graph->edges_size());
 642     for (int I = 0; I < Graph->edges_size(); ++I)
 643       if (EdgeCuts[I])
 644         CutEdges.set(I);
 645     LLVM_DEBUG(dbgs() << "Cutting edges... Done\n");
 646     LLVM_DEBUG(dbgs() << "Cut " << CutEdges.count() << " edges\n");
 647
 648     LLVM_DEBUG(dbgs() << "Inserting LFENCEs...\n");
 649     FencesInserted += insertFences(MF, *Graph, CutEdges);
 650     LLVM_DEBUG(dbgs() << "Inserting LFENCEs... Done\n");
 651     LLVM_DEBUG(dbgs() << "Inserted " << FencesInserted << " fences\n");
 652
 653     Graph = GraphBuilder::trim(*Graph, NodeSet{*Graph}, CutEdges);
 654   } while (true);
 655
 656   return FencesInserted;
 657 }
 658
 659 int X86LoadValueInjectionLoadHardeningPass::hardenLoadsWithHeuristic(
 660     MachineFunction &MF, std::unique_ptr<MachineGadgetGraph> Graph) const {
 661   // If `MF` does not have any fences, then no gadgets would have been
 662   // mitigated at this point.
 663   if (Graph->NumFences > 0) {
 664     LLVM_DEBUG(dbgs() << "Eliminating mitigated paths...\n");
 665     Graph = trimMitigatedEdges(std::move(Graph));
 666     LLVM_DEBUG(dbgs() << "Eliminating mitigated paths... Done\n");
 667   }
 668
 669   if (Graph->NumGadgets == 0)
 670     return 0;
 671
 672   LLVM_DEBUG(dbgs() << "Cutting edges...\n");
 673   EdgeSet CutEdges{*Graph};
 674
 675   // Begin by collecting all ingress CFG edges for each node
 676   DenseMap<const Node *, SmallVector<const Edge *, 2>> IngressEdgeMap;
 677   for (const Edge &E : Graph->edges())
 678     if (MachineGadgetGraph::isCFGEdge(E))
 679       IngressEdgeMap[E.getDest()].push_back(&E);
 680
 681   // For each gadget edge, make cuts that guarantee the gadget will be
 682   // mitigated. A computationally efficient way to achieve this is to either:
 683   // (a) cut all egress CFG edges from the gadget source, or
 684   // (b) cut all ingress CFG edges to the gadget sink.
 685   //
 686   // Moreover, the algorithm tries not to make a cut into a loop by preferring
 687   // to make a (b)-type cut if the gadget source resides at a greater loop depth
 688   // than the gadget sink, or an (a)-type cut otherwise.
 689   for (const Node &N : Graph->nodes()) {
 690     for (const Edge &E : N.edges()) {
 691       if (!MachineGadgetGraph::isGadgetEdge(E))
 692         continue;
 693
 694       SmallVector<const Edge *, 2> EgressEdges;
 695       SmallVector<const Edge *, 2> &IngressEdges = IngressEdgeMap[E.getDest()];
 696       for (const Edge &EgressEdge : N.edges())
 697         if (MachineGadgetGraph::isCFGEdge(EgressEdge))
 698           EgressEdges.push_back(&EgressEdge);
 699
 700       int EgressCutCost = 0, IngressCutCost = 0;
 701       for (const Edge *EgressEdge : EgressEdges)
 702         if (!CutEdges.contains(*EgressEdge))
 703           EgressCutCost += EgressEdge->getValue();
 704       for (const Edge *IngressEdge : IngressEdges)
 705         if (!CutEdges.contains(*IngressEdge))
 706           IngressCutCost += IngressEdge->getValue();
 707
 708       auto &EdgesToCut =
 709           IngressCutCost < EgressCutCost ? IngressEdges : EgressEdges;
 710       for (const Edge *E : EdgesToCut)
 711         CutEdges.insert(*E);
 712     }
 713   }
 714   LLVM_DEBUG(dbgs() << "Cutting edges... Done\n");
 715   LLVM_DEBUG(dbgs() << "Cut " << CutEdges.count() << " edges\n");
 716
 717   LLVM_DEBUG(dbgs() << "Inserting LFENCEs...\n");
 718   int FencesInserted = insertFences(MF, *Graph, CutEdges);
 719   LLVM_DEBUG(dbgs() << "Inserting LFENCEs... Done\n");
 720   LLVM_DEBUG(dbgs() << "Inserted " << FencesInserted << " fences\n");
 721
 722   return FencesInserted;
 723 }
 724
 725 int X86LoadValueInjectionLoadHardeningPass::insertFences(
 726     MachineFunction &MF, MachineGadgetGraph &G,
 727     EdgeSet &CutEdges /* in, out */) const {
 728   int FencesInserted = 0;
 729   for (const Node &N : G.nodes()) {
 730     for (const Edge &E : N.edges()) {
 731       if (CutEdges.contains(E)) {
 732         MachineInstr *MI = N.getValue(), *Prev;
 733         MachineBasicBlock *MBB;                  // Insert an LFENCE in this MBB
 734         MachineBasicBlock::iterator InsertionPt; // ...at this point
 735         if (MI == MachineGadgetGraph::ArgNodeSentinel) {
 736           // insert LFENCE at beginning of entry block
 737           MBB = &MF.front();
 738           InsertionPt = MBB->begin();
 739           Prev = nullptr;
 740         } else if (MI->isBranch()) { // insert the LFENCE before the branch
 741           MBB = MI->getParent();
 742           InsertionPt = MI;
 743           Prev = MI->getPrevNode();
 744           // Remove all egress CFG edges from this branch because the inserted
 745           // LFENCE prevents gadgets from crossing the branch.
 746           for (const Edge &E : N.edges()) {
 747             if (MachineGadgetGraph::isCFGEdge(E))
 748               CutEdges.insert(E);
 749           }
 750         } else { // insert the LFENCE after the instruction
 751           MBB = MI->getParent();
 752           InsertionPt = MI->getNextNode() ? MI->getNextNode() : MBB->end();
 753           Prev = InsertionPt == MBB->end()
 754                      ? (MBB->empty() ? nullptr : &MBB->back())
 755                      : InsertionPt->getPrevNode();
 756         }
 757         // Ensure this insertion is not redundant (two LFENCEs in sequence).
 758         if ((InsertionPt == MBB->end() || !isFence(&*InsertionPt)) &&
 759             (!Prev || !isFence(Prev))) {
 760           BuildMI(*MBB, InsertionPt, DebugLoc(), TII->get(X86::LFENCE));
 761           ++FencesInserted;
 762         }
 763       }
 764     }
 765   }
 766   return FencesInserted;
 767 }
 768
 769 bool X86LoadValueInjectionLoadHardeningPass::instrUsesRegToAccessMemory(
 770     const MachineInstr &MI, unsigned Reg) const {
 771   if (!MI.mayLoadOrStore() || MI.getOpcode() == X86::MFENCE ||
 772       MI.getOpcode() == X86::SFENCE || MI.getOpcode() == X86::LFENCE)
 773     return false;
 774
 775   // FIXME: This does not handle pseudo loading instruction like TCRETURN*
 776   const MCInstrDesc &Desc = MI.getDesc();
 777   int MemRefBeginIdx = X86II::getMemoryOperandNo(Desc.TSFlags);
 778   if (MemRefBeginIdx < 0) {
 779     LLVM_DEBUG(dbgs() << "Warning: unable to obtain memory operand for loading "
 780                          "instruction:\n";
 781                MI.print(dbgs()); dbgs() << '\n';);
 782     return false;
 783   }
 784   MemRefBeginIdx += X86II::getOperandBias(Desc);
 785
 786   const MachineOperand &BaseMO =
 787       MI.getOperand(MemRefBeginIdx + X86::AddrBaseReg);
 788   const MachineOperand &IndexMO =
 789       MI.getOperand(MemRefBeginIdx + X86::AddrIndexReg);
 790   return (BaseMO.isReg() && BaseMO.getReg() != X86::NoRegister &&
 791           TRI->regsOverlap(BaseMO.getReg(), Reg)) ||
 792          (IndexMO.isReg() && IndexMO.getReg() != X86::NoRegister &&
 793           TRI->regsOverlap(IndexMO.getReg(), Reg));
 794 }
 795
 796 bool X86LoadValueInjectionLoadHardeningPass::instrUsesRegToBranch(
 797     const MachineInstr &MI, unsigned Reg) const {
 798   if (!MI.isConditionalBranch())
 799     return false;
 800   for (const MachineOperand &Use : MI.uses())
 801     if (Use.isReg() && Use.getReg() == Reg)
 802       return true;
 803   return false;
 804 }
 805
 806 INITIALIZE_PASS_BEGIN(X86LoadValueInjectionLoadHardeningPass, PASS_KEY,
 807                       "X86 LVI load hardening", false, false)
 808 INITIALIZE_PASS_DEPENDENCY(MachineLoopInfo)
 809 INITIALIZE_PASS_DEPENDENCY(MachineDominatorTree)
 810 INITIALIZE_PASS_DEPENDENCY(MachineDominanceFrontier)
 811 INITIALIZE_PASS_END(X86LoadValueInjectionLoadHardeningPass, PASS_KEY,
 812                     "X86 LVI load hardening", false, false)
 813
 814 FunctionPass *llvm::createX86LoadValueInjectionLoadHardeningPass() {
 815   return new X86LoadValueInjectionLoadHardeningPass();
 816 }