[clang] Respect the [[clang::unsafe_buffer_usage]] attribute for field and constructor initializers (PR #91991)

Dana Jansens via cfe-commits cfe-commits at lists.llvm.org
Mon Jun 10 09:48:33 PDT 2024


================
@@ -3328,3 +3300,63 @@ void clang::checkUnsafeBufferUsage(const Decl *D,
     }
   }
 }
+
+void clang::checkUnsafeBufferUsage(const Decl *D,
+                                   UnsafeBufferUsageHandler &Handler,
+                                   bool EmitSuggestions) {
+#ifndef NDEBUG
+  Handler.clearDebugNotes();
+#endif
+
+  assert(D);
+
+  SmallVector<Stmt *> Stmts;
+
+  // We do not want to visit a Lambda expression defined inside a method
+  // independently. Instead, it should be visited along with the outer method.
+  // FIXME: do we want to do the same thing for `BlockDecl`s?
+  if (const auto *fd = dyn_cast<CXXMethodDecl>(D)) {
+    if (fd->getParent()->isLambda() && fd->getParent()->isLocalClass())
+      return;
+  }
+
+  // Do not emit fixit suggestions for functions declared in an
+  // extern "C" block.
+  if (const auto *FD = dyn_cast<FunctionDecl>(D)) {
+    for (FunctionDecl *FReDecl : FD->redecls()) {
+      if (FReDecl->isExternC()) {
+        EmitSuggestions = false;
+        break;
+      }
+    }
+
+    Stmts.push_back(FD->getBody());
+
+    if (const auto *ID = dyn_cast<CXXConstructorDecl>(D)) {
+      for (const CXXCtorInitializer *CI : ID->inits()) {
+        Stmts.push_back(CI->getInit());
+      }
+    }
+  }
+
+  if (const auto *FD = dyn_cast<FieldDecl>(D)) {
----------------
danakj wrote:

This is the AST in the case without the explicit constructor:
```
  |-CXXRecordDecl <col:1, col:8> col:8 implicit struct HoldsUnsafeMembers
  `-FieldDecl <line:8:5, col:30> col:19 FromField 'UnsafeMembers'
    `-CXXConstructExpr <col:28, col:30> 'UnsafeMembers' 'void (int)' list
      `-IntegerLiteral <col:29> 'int' 3
```

https://godbolt.org/z/5freKecWb

AFAICT we are required to visit FieldDecl to handle the 0-ctor class use case. Thinking through design, maybe we could visit RecordDecl instead, in CallableVisitor, and then provide a TraverseRecordDecl in the MatchDescendantVisitor that would traverse each field initializer? I could try it, I am not sure it would work, but it's not obviously more clear/better to me either.

I added a comment on the aggregate test case mentioning the lack of CXXDefaultInitExpr that it is testing for.

https://github.com/llvm/llvm-project/pull/91991


More information about the cfe-commits mailing list