[PATCH] D69602: [analyzer] Test case for lambda capture by value modelling

[Artem Dergachev via Phabricator via cfe-commits]

NoQ added a comment.

I recommend starting with CFG tests for this feature. I.e., do `-analyzer-checker debug.DumpCFG` and make sure that every `CFGStmt` that corresponds to a constructor is a `CFGConstructor` (i.e., has a `ConstructionContext` explained in the dump). Cf. `test/Analysis/cfg-rich-constructors.cpp`.



================
Comment at: clang/test/Analysis/clangsa_unsupported_features/handle_constructors_for_lambda_captures.cpp:11
+void reached_function_from_simple_copy() {
+  clang_analyzer_warnIfReached(); // expected-warning{{REACHABLE}}
+}
----------------
This test doesn't test whether the function is inlined. If the function is not inlined, the code will still be reachable when analyzed as top frame. See also `clang_analyzer_checkInlined()`.


================
Comment at: clang/test/Analysis/clangsa_unsupported_features/handle_constructors_for_lambda_captures.cpp:29
+  operator=(incr_on_copy_for_simple_usage &&o) = delete;
+  ~incr_on_copy_for_simple_usage() = default;
+};
----------------
That's the reason why your tests seem to work. If you change it to `~incr_on_copy_for_simple_usage() {}`, we will no longer evaluate the capture correctly.

The difference here is that we allow ourselves inline the constructor without a construction context when the destructor is trivial. But when the destructor is non-trivial, our inability to keep track of the object under construction until destruction will have terrible consequences.


Repository:
  rG LLVM Github Monorepo

CHANGES SINCE LAST ACTION
  https://reviews.llvm.org/D69602/new/

https://reviews.llvm.org/D69602