That's something that the programmer needs to specifically ask for, and I'm not aware of any C++ compiler which doesn't throw a bunch of warnings regarding an object's scope and lifetime.
That means your point boils down to arguing that if a programmer really wants to, he can still shoot himself in the proverbial foot by writing C++ code, but in order for that to happen he needs to have a death wish and intentionally avoid all warnings and safe practices to begin with.
#include <vector>
int main()
{
struct Foo {
int bar;
};
Foo* bar;
{
std::vector<Foo> foos;
foos.emplace_back(Foo{1});
foos.emplace_back(Foo{2});
bar = &foos[1];
}
bar->bar = 2;
}However, I never spent more than 10 minutes fixing that.
Debug version of MS C runtime fills freed heap memory with a magic number 0xFEEEFEEE. You’ll get exception in runtime, and you’ll immediately have the idea what’s wrong with the code.
I've seen 5 minute fixes and things that took weeks to track down(yay for non deterministic repros).
Technical support guys were not that happy with the customer regularly calling them to know when the fix would be delivered.
Have other war stories about tracking down leaks as well.
Most commercial libraries are written in C or C++. The reasons include performance, smaller runtime, and ease of use (all OS kernels are written in C, so the majority of languages already feature decent C interop). Even if you code Rust, you won’t have that compile-time checks inside proprietary third-party libraries.
FWIW it's getting pretty common in the Rust ecosystem to wrap 3rd party libs with proper lifetime annotations with opaque types when possible. Things like OpenGL, JNI bindings and whatnot can still benefit from lifetime annotations the compiler provides.
Rust proponents often overestimate practical impart of those C++ safely problems fixed by Rust.
At the same time, they often underestimate or ignore Rust’s problems caused by the same safety thing. The main of which is, it’s hard to implement some widely used data structures in Rust. Here’s my old thread where I discuss the topic in details:
Are there categories of UAF that would not be caught with valgrind / asan? Or is it the case that the test suites do not have sufficient coverage to exhibit the issues?
Requires manual annotation and isn't nearly as systematic and comprehensive as the Rust compiler, but I've found threading bugs with this in existing code that I was about to modify.
No, it's not.
Use-after-free is very easy. Just push onto a vector while you're iterating over it, to name one obvious example that comes up a lot.
We were testing our application the night before a big demo for our client. Ours was a C++ OpenGL application with several "phases" the user could enter from a "main menu" type screen. Our demo explored these phases in a pre-planned order. All went well, until somebody decided to try doing the demo in a slightly different order. Once the tester visited a certain phase of the application, the thing crashed and burned. The stack trace was littered with some incomprehensible container-implementation-detail junk. Debuggers weren't pointing at an obvious sign, other than that something was dereferencing a garbage pointer.
We noticed that if we went back to the original canned order of operations (this was a demo, after all), the application didn't crash. But this still made us nervous. We found that we could sometimes swap a few phases early in the application's run with no ill effect.
Here's part of the class responsible for some of the higher layers of the stack trace, from one of the lower level libraries we build all of our applications on (this is important because, for hysterical raisins, we can't source-step debug those libraries in production). Can you spot the bug?
#include <map>
#include <string>
#include <vector>
struct Message;
class MessageCentral {
public:
void HandleMessages() {
// messages were collected somewhere and organized by their queue.
const std::vector<std::string> queues_with_new_messages;
for (const std::string &queue : queues_with_new_messages) {
const std::vector<Message> msgs; // = messages_for_queue(queue);
const auto &handlers = handlers_[queue];
for (auto &h : handlers) {
for (const auto &msg : msgs) {
h(msg);
}
}
}
}
void RegisterMessageHandler(const std::string &queue_name, std::function<void(Message)> handler) {
handlers_[queue_name].push_back(handler);
}
private:
std::map<std::string, std::vector<std::function<void(Message)>>> handlers_;
};
(Note: this is probably way easier to figure out with all of the other code needed to implement MessageCentral taken out! It took us three engineers working for about 2.5-3 hours each, with lots and lots of gdb and printf debugging, to nail this one down. This particular crash would have been impossible in Rust, because the compiler would have rejected the Rust equivalent of the above C++ code.)(oh, by the way, we couldn't use runtime memory sanitizer tools like ASan or valgrind because the application linked to CEF. In retrospect, we could have dropped in a fake CEF that didn't interfere with ASan or valgrind, but that would have required build resources we didn't have available in the frenzied night before a big demo.)
I assume one of your handlers invokes RegisterMessageHandler invalidates the iterators to handlers_[some_queue_name]. As such, the following line exhibits UB:
for (auto &h : handlers) {
In practice, the exact behavior is likely to depend on if the vector resized it's underlying storage or not.EDIT: I'll also note that using a debug build of the SC++L with iterator debugging enabled should give you a useful diagnostic here - given a modern SC++L implementation. Whether or not your application runs well enough, to actually get to the point of being able to reproduce the bug with iterator debugging enabled, is another matter entirely of course.
EDIT x2: I'll further note I'm cheating here by having encountered and fixed this same bug on a different codebase.
Rust is much more robust about catching those cases and verifying that you don't have objects outliving their references.