[clang][bytecode] Check source pointer for bitcast validity

Unfortunately this is more dynamic than anticipated.

Fixes #165006

Author: tbaederr

clang/lib/AST/ByteCode/Compiler.cpp

@@ -208,19 +208,6 @@ template <class Emitter> class LocOverrideScope final {
 } // namespace interp
 } // namespace clang
 
-template <class Emitter>
-bool Compiler<Emitter>::isValidBitCast(const CastExpr *E) {
-  QualType FromTy = E->getSubExpr()->getType()->getPointeeType();
-  QualType ToTy = E->getType()->getPointeeType();
-
-  if (classify(FromTy) == classify(ToTy))
-    return true;
-
-  if (FromTy->isVoidType() || ToTy->isVoidType())
-    return true;
-  return false;
-}
-
 template <class Emitter>
 bool Compiler<Emitter>::VisitCastExpr(const CastExpr *CE) {
   const Expr *SubExpr = CE->getSubExpr();
@@ -506,12 +493,9 @@ bool Compiler<Emitter>::VisitCastExpr(const CastExpr *CE) {
     if (!FromT || !ToT)
       return false;
 
-    if (!this->isValidBitCast(CE) &&
-        !this->emitInvalidCast(CastKind::ReinterpretLike, /*Fatal=*/false, CE))
-      return false;
-
     assert(isPtrType(*FromT));
     assert(isPtrType(*ToT));
+    bool SrcIsVoidPtr = SubExprTy->isVoidPointerType();
     if (FromT == ToT) {
       if (CE->getType()->isVoidPointerType() &&
           !SubExprTy->isFunctionPointerType()) {
@@ -520,6 +504,10 @@ bool Compiler<Emitter>::VisitCastExpr(const CastExpr *CE) {
 
       if (!this->visit(SubExpr))
         return false;
+      if (!this->emitCheckBitCast(CETy->getPointeeType().getTypePtr(),
+                                  SrcIsVoidPtr, CE))
+        return false;
+
       if (CE->getType()->isFunctionPointerType() ||
           SubExprTy->isFunctionPointerType()) {
         return this->emitFnPtrCast(CE);

clang/lib/AST/ByteCode/Compiler.h

@@ -425,8 +425,6 @@ class Compiler : public ConstStmtVisitor<Compiler<Emitter>, bool>,
 
   bool refersToUnion(const Expr *E);
 
-  bool isValidBitCast(const CastExpr *E);
-
 protected:
   /// Variable to storage mapping.
   llvm::DenseMap<const ValueDecl *, Scope::Local> Locals;

clang/lib/AST/ByteCode/Interp.h

@@ -3289,6 +3289,52 @@ inline bool SideEffect(InterpState &S, CodePtr OpPC) {
   return S.noteSideEffect();
 }
 
+inline bool CheckBitCast(InterpState &S, CodePtr OpPC, const Type *TargetType,
+                         bool SrcIsVoidPtr) {
+  const auto &Ptr = S.Stk.peek<Pointer>();
+  if (Ptr.isZero())
+    return true;
+  if (!Ptr.isBlockPointer())
+    return true;
+
+  if (TargetType->isIntegerType())
+    return true;
+
+  if (SrcIsVoidPtr && S.getLangOpts().CPlusPlus) {
+    bool HasValidResult = !Ptr.isZero();
+
+    if (HasValidResult) {
+      if (S.getStdAllocatorCaller("allocate"))
+        return true;
+
+      const auto &E = cast<CastExpr>(S.Current->getExpr(OpPC));
+      if (S.getLangOpts().CPlusPlus26 &&
+          S.getASTContext().hasSimilarType(Ptr.getType(),
+                                           QualType(TargetType, 0)))
+        return true;
+
+      S.CCEDiag(E, diag::note_constexpr_invalid_void_star_cast)
+          << E->getSubExpr()->getType() << S.getLangOpts().CPlusPlus26
+          << Ptr.getType().getCanonicalType() << E->getType()->getPointeeType();
+    } else if (!S.getLangOpts().CPlusPlus26) {
+      const SourceInfo &E = S.Current->getSource(OpPC);
+      S.CCEDiag(E, diag::note_constexpr_invalid_cast)
+          << diag::ConstexprInvalidCastKind::CastFrom << "'void *'"
+          << S.Current->getRange(OpPC);
+    }
+  }
+
+  QualType PtrType = Ptr.getType();
+  if (PtrType->isRecordType() &&
+      PtrType->getAsRecordDecl() != TargetType->getAsRecordDecl()) {
+    S.CCEDiag(S.Current->getSource(OpPC), diag::note_constexpr_invalid_cast)
+        << diag::ConstexprInvalidCastKind::ThisConversionOrReinterpret
+        << S.getLangOpts().CPlusPlus << S.Current->getRange(OpPC);
+    return false;
+  }
+  return true;
+}
+
 /// Same here, but only for casts.
 inline bool InvalidCast(InterpState &S, CodePtr OpPC, CastKind Kind,
                         bool Fatal) {

clang/lib/AST/ByteCode/Opcodes.td

@@ -422,6 +422,8 @@ def CheckLiteralType : Opcode {
 
 def CheckArraySize : Opcode { let Args = [ArgUint64]; }
 
+def CheckBitCast : Opcode { let Args = [ArgTypePtr, ArgBool]; }
+
 // [] -> [Value]
 def GetGlobal : AccessOpcode;
 def GetGlobalUnchecked : AccessOpcode;

clang/test/AST/ByteCode/cxx11.cpp

@@ -379,3 +379,13 @@ namespace DiscardedAddrLabel {
   }
 }
 
+namespace PointerCast {
+  /// The two interpreters disagree here.
+  struct S { int x, y; } s;
+  constexpr S* sptr = &s;
+  struct U {};
+  struct Str {
+    int e : (Str*)(sptr) == (Str*)(sptr); // expected-error {{not an integral constant expression}} \
+                                          // expected-note {{cast that performs the conversions of a reinterpret_cast}}
+  };
+}

clang/test/AST/ByteCode/invalid.cpp

@@ -88,4 +88,22 @@ namespace InvalidBitCast {
                              // both-note {{in call to}}
 
 
+  struct sockaddr
+  {
+    char sa_data[8];
+  };
+  struct in_addr
+  {
+    unsigned int s_addr;
+  };
+  struct sockaddr_in
+  {
+    unsigned short int sin_port;
+    struct in_addr sin_addr;
+  };
+  /// Bitcast from sockaddr to sockaddr_in. Used to crash.
+  unsigned int get_addr(sockaddr addr) {
+      return ((sockaddr_in *)&addr)->sin_addr.s_addr;
+  }
+
 }