lifetime-analysis-lifetimebound

Author: usx95

clang/include/clang/Analysis/Analyses/LifetimeSafety.h

@@ -75,13 +75,14 @@ template <typename Tag> struct ID {
   }
 };
 
-template <typename Tag>
-inline llvm::raw_ostream &operator<<(llvm::raw_ostream &OS, ID<Tag> ID) {
-  return OS << ID.Value;
-}
-
 using LoanID = ID<struct LoanTag>;
 using OriginID = ID<struct OriginTag>;
+inline llvm::raw_ostream &operator<<(llvm::raw_ostream &OS, LoanID ID) {
+  return OS << ID.Value;
+}
+inline llvm::raw_ostream &operator<<(llvm::raw_ostream &OS, OriginID ID) {
+  return OS << ID.Value;
+}
 
 // Using LLVM's immutable collections is efficient for dataflow analysis
 // as it avoids deep copies during state transitions.

clang/lib/Analysis/LifetimeSafety.cpp

@@ -212,8 +212,10 @@ class Fact {
     /// A loan expires as its underlying storage is freed (e.g., variable goes
     /// out of scope).
     Expire,
+    /// The loan set of an origin is cleared.
+    KillOrigin,
     /// An origin is propagated from a source to a destination (e.g., p = q).
-    AssignOrigin,
+    OriginFlow,
     /// An origin escapes the function by flowing into the return value.
     ReturnOfOrigin,
     /// An origin is used (eg. dereferencing a pointer).
@@ -285,22 +287,24 @@ class ExpireFact : public Fact {
   }
 };
 
-class AssignOriginFact : public Fact {
+class OriginFlowFact : public Fact {
   OriginID OIDDest;
   OriginID OIDSrc;
 
 public:
   static bool classof(const Fact *F) {
-    return F->getKind() == Kind::AssignOrigin;
+    return F->getKind() == Kind::OriginFlow;
   }
 
-  AssignOriginFact(OriginID OIDDest, OriginID OIDSrc)
-      : Fact(Kind::AssignOrigin), OIDDest(OIDDest), OIDSrc(OIDSrc) {}
+  OriginFlowFact(OriginID OIDDest, OriginID OIDSrc)
+      : Fact(Kind::OriginFlow), OIDDest(OIDDest), OIDSrc(OIDSrc) {}
+
   OriginID getDestOriginID() const { return OIDDest; }
   OriginID getSrcOriginID() const { return OIDSrc; }
+
   void dump(llvm::raw_ostream &OS, const LoanManager &,
             const OriginManager &OM) const override {
-    OS << "AssignOrigin (Dest: ";
+    OS << "OriginFlow (Dest: ";
     OM.dump(getDestOriginID(), OS);
     OS << ", Src: ";
     OM.dump(getSrcOriginID(), OS);
@@ -353,6 +357,23 @@ class UseFact : public Fact {
   }
 };
 
+class KillOriginFact : public Fact {
+  OriginID OID;
+
+public:
+  static bool classof(const Fact *F) {
+    return F->getKind() == Kind::KillOrigin;
+  }
+  KillOriginFact(OriginID OID) : Fact(Kind::KillOrigin), OID(OID) {}
+  OriginID getOriginID() const { return OID; }
+
+  void dump(llvm::raw_ostream &OS, const LoanManager &,
+            const OriginManager &OM) const override {
+    OS << "KillOrigin (";
+    OM.dump(getOriginID(), OS);
+    OS << ")\n";
+  }
+};
 /// A dummy-fact used to mark a specific point in the code for testing.
 /// It is generated by recognizing a `void("__lifetime_test_point_...")` cast.
 class TestPointFact : public Fact {
@@ -454,7 +475,7 @@ class FactGenerator : public ConstStmtVisitor<FactGenerator> {
       if (const auto *VD = dyn_cast<VarDecl>(D))
         if (hasOrigin(VD))
           if (const Expr *InitExpr = VD->getInit())
-            addAssignOriginFact(*VD, *InitExpr);
+            addOriginFlowFact(*VD, *InitExpr);
   }
 
   void VisitDeclRefExpr(const DeclRefExpr *DRE) {
@@ -492,9 +513,23 @@ class FactGenerator : public ConstStmtVisitor<FactGenerator> {
         isa<CXXConversionDecl>(MCE->getCalleeDecl())) {
       // The argument is the implicit object itself.
       handleFunctionCall(MCE, MCE->getMethodDecl(),
-                         {MCE->getImplicitObjectArgument()});
+                         {MCE->getImplicitObjectArgument()},
+                         /*IsGslConstruction=*/true);
     }
-    // FIXME: A more general VisitCallExpr could also be used here.
+    if (const CXXMethodDecl *Method = MCE->getMethodDecl()) {
+      // Construct the argument list, with the implicit 'this' object as the
+      // first argument.
+      llvm::SmallVector<const Expr *, 4> Args;
+      Args.push_back(MCE->getImplicitObjectArgument());
+      Args.append(MCE->getArgs(), MCE->getArgs() + MCE->getNumArgs());
+
+      handleFunctionCall(MCE, Method, Args, /*IsGslConstruction=*/false);
+    }
+  }
+
+  void VisitCallExpr(const CallExpr *CE) {
+    handleFunctionCall(CE, CE->getDirectCallee(),
+                       {CE->getArgs(), CE->getNumArgs()});
   }
 
   void VisitCXXNullPtrLiteralExpr(const CXXNullPtrLiteralExpr *N) {
@@ -508,7 +543,7 @@ class FactGenerator : public ConstStmtVisitor<FactGenerator> {
       return;
     // An ImplicitCastExpr node itself gets an origin, which flows from the
     // origin of its sub-expression (after stripping its own parens/casts).
-    addAssignOriginFact(*ICE, *ICE->getSubExpr());
+    addOriginFlowFact(*ICE, *ICE->getSubExpr());
   }
 
   void VisitUnaryOperator(const UnaryOperator *UO) {
@@ -522,7 +557,7 @@ class FactGenerator : public ConstStmtVisitor<FactGenerator> {
       // its sub-expression (x). This fact will cause the dataflow analysis
       // to propagate any loans held by the sub-expression's origin to the
       // origin of this UnaryOperator expression.
-      addAssignOriginFact(*UO, *SubExpr);
+      addOriginFlowFact(*UO, *SubExpr);
     }
   }
 
@@ -542,8 +577,15 @@ class FactGenerator : public ConstStmtVisitor<FactGenerator> {
   }
 
   void VisitCXXOperatorCallExpr(const CXXOperatorCallExpr *OCE) {
-    if (OCE->isAssignmentOp() && OCE->getNumArgs() == 2)
+    // Assignment operators have special "kill-then-propagate" semantics
+    // and are handled separately.
+    if (OCE->isAssignmentOp() && OCE->getNumArgs() == 2) {
       handleAssignment(OCE->getArg(0), OCE->getArg(1));
+      return;
+    }
+    handleFunctionCall(OCE, OCE->getDirectCallee(),
+                       {OCE->getArgs(), OCE->getNumArgs()},
+                       /*IsGslConstruction=*/false);
   }
 
   void VisitCXXFunctionalCastExpr(const CXXFunctionalCastExpr *FCE) {
@@ -552,7 +594,7 @@ class FactGenerator : public ConstStmtVisitor<FactGenerator> {
     if (handleTestPoint(FCE))
       return;
     if (isGslPointerType(FCE->getType()))
-      addAssignOriginFact(*FCE, *FCE->getSubExpr());
+      addOriginFlowFact(*FCE, *FCE->getSubExpr());
   }
 
   void VisitInitListExpr(const InitListExpr *ILE) {
@@ -561,15 +603,15 @@ class FactGenerator : public ConstStmtVisitor<FactGenerator> {
     // For list initialization with a single element, like `View{...}`, the
     // origin of the list itself is the origin of its single element.
     if (ILE->getNumInits() == 1)
-      addAssignOriginFact(*ILE, *ILE->getInit(0));
+      addOriginFlowFact(*ILE, *ILE->getInit(0));
   }
 
   void VisitMaterializeTemporaryExpr(const MaterializeTemporaryExpr *MTE) {
     if (!hasOrigin(MTE))
       return;
     // A temporary object's origin is the same as the origin of the
     // expression that initializes it.
-    addAssignOriginFact(*MTE, *MTE->getSubExpr());
+    addOriginFlowFact(*MTE, *MTE->getSubExpr());
   }
 
   void handleDestructor(const CFGAutomaticObjDtor &DtorOpt) {
@@ -624,34 +666,68 @@ class FactGenerator : public ConstStmtVisitor<FactGenerator> {
     if (CCE->getNumArgs() != 1)
       return;
     if (hasOrigin(CCE->getArg(0)))
-      addAssignOriginFact(*CCE, *CCE->getArg(0));
+      addOriginFlowFact(*CCE, *CCE->getArg(0));
     else
       // This could be a new borrow.
       handleFunctionCall(CCE, CCE->getConstructor(),
-                         {CCE->getArgs(), CCE->getNumArgs()});
+                         {CCE->getArgs(), CCE->getNumArgs()},
+                         /*IsGslConstruction=*/true);
+  }
+  static const FunctionDecl *
+  getDeclWithMergedLifetimeBoundAttrs(const FunctionDecl *FD) {
+    return FD != nullptr ? FD->getMostRecentDecl() : nullptr;
   }
 
+  static const CXXMethodDecl *
+  getDeclWithMergedLifetimeBoundAttrs(const CXXMethodDecl *CMD) {
+    const FunctionDecl *FD = CMD;
+    return cast_if_present<CXXMethodDecl>(
+        getDeclWithMergedLifetimeBoundAttrs(FD));
+  }
+  static bool implicitObjectParamIsLifetimeBound(const FunctionDecl *FD) {
+    FD = getDeclWithMergedLifetimeBoundAttrs(FD);
+    const TypeSourceInfo *TSI = FD->getTypeSourceInfo();
+    if (!TSI)
+      return false;
+    // Don't declare this variable in the second operand of the for-statement;
+    // GCC miscompiles that by ending its lifetime before evaluating the
+    // third operand. See gcc.gnu.org/PR86769.
+    AttributedTypeLoc ATL;
+    for (TypeLoc TL = TSI->getTypeLoc();
+         (ATL = TL.getAsAdjusted<AttributedTypeLoc>());
+         TL = ATL.getModifiedLoc()) {
+      if (ATL.getAttrAs<LifetimeBoundAttr>())
+        return true;
+    }
+    return false;
+  }
   /// Checks if a call-like expression creates a borrow by passing a value to a
   /// reference parameter, creating an IssueFact if it does.
   void handleFunctionCall(const Expr *Call, const FunctionDecl *FD,
-                          ArrayRef<const Expr *> Args) {
-    if (!FD)
+                          ArrayRef<const Expr *> Args,
+                          bool IsGslConstruction = false) {
+    // Ignore functions returning values with no origin.
+    if (!FD || !hasOrigin(Call))
       return;
-    // TODO: Handle more than one arguments.
-    for (unsigned I = 0; I <= 0 /*Args.size()*/; ++I) {
-      const Expr *ArgExpr = Args[I];
-
-      // Propagate origins for CXX this.
-      if (FD->isCXXClassMember() && I == 0) {
-        addAssignOriginFact(*Call, *ArgExpr);
-        continue;
-      }
-      // The parameter is a pointer, reference, or gsl::Pointer.
-      // This is a borrow. We propagate the origin from the argument expression
-      // at the call site to the parameter declaration in the callee.
-      if (hasOrigin(ArgExpr))
-        addAssignOriginFact(*Call, *ArgExpr);
-    }
+    auto IsArgLifetimeBound = [FD](unsigned I) -> bool {
+      const ParmVarDecl *PVD = nullptr;
+      if (const auto *Method = dyn_cast<CXXMethodDecl>(FD);
+          Method && Method->isInstance()) {
+        if (I == 0)
+          // For the 'this' argument, the attribute is on the method itself.
+          return implicitObjectParamIsLifetimeBound(Method);
+        if ((I - 1) < Method->getNumParams())
+          // For explicit arguments, find the corresponding parameter
+          // declaration.
+          PVD = Method->getParamDecl(I - 1);
+      } else if (I < FD->getNumParams())
+        // For free functions or static methods.
+        PVD = FD->getParamDecl(I);
+      return PVD ? PVD->hasAttr<clang::LifetimeBoundAttr>() : false;
+    };
+    for (unsigned I = 0; I < Args.size(); ++I)
+      if (IsGslConstruction || IsArgLifetimeBound(I))
+        addOriginFlowFact(*Call, *Args[I]);
   }
 
   /// Creates a loan for the storage path of a given declaration reference.
@@ -668,11 +744,16 @@ class FactGenerator : public ConstStmtVisitor<FactGenerator> {
   }
 
   template <typename Destination, typename Source>
-  void addAssignOriginFact(const Destination &D, const Source &S) {
+  void addOriginFlowFact(const Destination &D, const Source &S) {
     OriginID DestOID = FactMgr.getOriginMgr().getOrCreate(D);
     OriginID SrcOID = FactMgr.getOriginMgr().get(S);
     CurrentBlockFacts.push_back(
-        FactMgr.createFact<AssignOriginFact>(DestOID, SrcOID));
+        FactMgr.createFact<OriginFlowFact>(DestOID, SrcOID));
+  }
+
+  void killOrigin(const ValueDecl *D) {
+    OriginID DestOID = FactMgr.getOriginMgr().getOrCreate(*D);
+    CurrentBlockFacts.push_back(FactMgr.createFact<KillOriginFact>(DestOID));
   }
 
   /// Checks if the expression is a `void("__lifetime_test_point_...")` cast.
@@ -703,12 +784,12 @@ class FactGenerator : public ConstStmtVisitor<FactGenerator> {
     if (const auto *DRE_LHS =
             dyn_cast<DeclRefExpr>(LHSExpr->IgnoreParenImpCasts())) {
       markUseAsWrite(DRE_LHS);
-      if (const auto *VD_LHS = dyn_cast<ValueDecl>(DRE_LHS->getDecl()))
-        // We are interested in assignments like `ptr1 = ptr2` or `ptr = &var`.
-        // LHS must be a pointer/reference type that can be an origin. RHS must
-        // also represent an origin (either another pointer/ref or an
-        // address-of).
-        addAssignOriginFact(*VD_LHS, *RHSExpr);
+      if (const auto *VD_LHS = dyn_cast<ValueDecl>(DRE_LHS->getDecl())) {
+        // Kill the old loans of the destination origin and flow the new loans
+        // from the source origin.
+        killOrigin(VD_LHS);
+        addOriginFlowFact(*VD_LHS, *RHSExpr);
+      }
     }
   }
 
@@ -882,8 +963,10 @@ class DataflowAnalysis {
       return D->transfer(In, *F->getAs<IssueFact>());
     case Fact::Kind::Expire:
       return D->transfer(In, *F->getAs<ExpireFact>());
-    case Fact::Kind::AssignOrigin:
-      return D->transfer(In, *F->getAs<AssignOriginFact>());
+    case Fact::Kind::OriginFlow:
+      return D->transfer(In, *F->getAs<OriginFlowFact>());
+    case Fact::Kind::KillOrigin:
+      return D->transfer(In, *F->getAs<KillOriginFact>());
     case Fact::Kind::ReturnOfOrigin:
       return D->transfer(In, *F->getAs<ReturnOfOriginFact>());
     case Fact::Kind::Use:
@@ -897,7 +980,8 @@ class DataflowAnalysis {
 public:
   Lattice transfer(Lattice In, const IssueFact &) { return In; }
   Lattice transfer(Lattice In, const ExpireFact &) { return In; }
-  Lattice transfer(Lattice In, const AssignOriginFact &) { return In; }
+  Lattice transfer(Lattice In, const OriginFlowFact &) { return In; }
+  Lattice transfer(Lattice In, const KillOriginFact &) { return In; }
   Lattice transfer(Lattice In, const ReturnOfOriginFact &) { return In; }
   Lattice transfer(Lattice In, const UseFact &) { return In; }
   Lattice transfer(Lattice In, const TestPointFact &) { return In; }
@@ -1049,14 +1133,27 @@ class LoanPropagationAnalysis
         LoanSetFactory.add(LoanSetFactory.getEmptySet(), LID)));
   }
 
-  /// The destination origin's loan set is replaced by the source's.
-  /// This implicitly "resets" the old loans of the destination.
-  Lattice transfer(Lattice In, const AssignOriginFact &F) {
+  /// A flow from source to destination adds the source's loans to the
+  /// destination's, without clearing the destination's existing loans.
+  Lattice transfer(Lattice In, const OriginFlowFact &F) {
     OriginID DestOID = F.getDestOriginID();
     OriginID SrcOID = F.getSrcOriginID();
+
+    LoanSet DestLoans = getLoans(In, DestOID);
     LoanSet SrcLoans = getLoans(In, SrcOID);
+    LoanSet MergedLoans = utils::join(DestLoans, SrcLoans, LoanSetFactory);
+
     return LoanPropagationLattice(
-        OriginLoanMapFactory.add(In.Origins, DestOID, SrcLoans));
+        OriginLoanMapFactory.add(In.Origins, DestOID, MergedLoans));
+  }
+
+  /// Clears the loan set of the specified origin. This is used on the
+  /// left-hand side of an assignment to invalidate the variable's old lifetime.
+  Lattice transfer(Lattice In, const KillOriginFact &F) {
+    OriginID OID = F.getOriginID();
+    // Replace the origin's loan set with an empty set.
+    return LoanPropagationLattice(OriginLoanMapFactory.add(
+        In.Origins, OID, LoanSetFactory.getEmptySet()));
   }
 
   LoanSet getLoans(OriginID OID, ProgramPoint P) {

clang/test/Analysis/LifetimeSafety/benchmark.py

@@ -340,7 +340,7 @@ def run_single_test(
             "name": "cycle",
             "title": "Pointer Cycle in Loop",
             "generator_func": generate_cpp_cycle_test,
-            "n_values": [25, 50, 75, 100],
+            "n_values": [50, 75, 100, 200, 300],
         },
         {
             "name": "merge",