GCC 6 имеет новую функцию оптимизатора: он предполагает, что this всегда не равно null, и оптимизирует его на основе этого.

При распространении диапазона значений теперь предполагается, что указатель this функций-членов C ++ не равен нулю. Это исключает обычные проверки нулевого указателя , но также нарушает некоторые несоответствующие кодовые базы (такие как Qt-5, Chromium, KDevelop). В качестве временного обходного пути можно использовать -fno-delete-null-pointer-check. Неверный код можно определить с помощью -fsanitize = undefined.

В документе об изменениях это явно обозначается как опасное, потому что оно нарушает удивительное количество часто используемого кода.

Why would this new assumption break practical C++ code? Are there particular patterns where careless or uninformed programmers rely on this particular undefined behavior? I cannot imagine anyone writing if (this == NULL) because that is so unnatural.

Ответы (5)

I guess the question that needs to be answered why well-intentioned people would write the checks in the first place.

The most common case is probably if you have a class that is part of a naturally occurring recursive call.

If you had:

struct Node
{
    Node* left;
    Node* right;
};

in C, you might write:

void traverse_in_order(Node* n) {
    if(!n) return;
    traverse_in_order(n->left);
    process(n);
    traverse_in_order(n->right);
}

In C++, it's nice to make this a member function:

void Node::traverse_in_order() {
    // <--- What check should be put here?
    left->traverse_in_order();
    process();
    right->traverse_in_order();
}

In the early days of C++ (prior to standardization), it was emphasized that that member functions were syntactic sugar for a function where the this parameter is implicit. Code was written in C++, converted to equivalent C and compiled. There were even explicit examples that comparing this to null was meaningful and the original Cfront compiler took advantage of this too. So coming from a C background, the obvious choice for the check is:

if(this == nullptr) return;      

Note: Bjarne Stroustrup even mentions that the rules for this have changed over the years here

And this worked on many compilers for many years. When standardization happened, this changed. And more recently, compilers started taking advantage of calling a member function where this being nullptr is undefined behavior, which means that this condition is always false, and the compiler is free to omit it.

That means that to do any traversal of this tree, you need to either:

  • Do all of the checks before calling traverse_in_order

    void Node::traverse_in_order() {
        if(left) left->traverse_in_order();
        process();
        if(right) right->traverse_in_order();
    }
    

    This means also checking at EVERY call site if you could have a null root.

  • Don't use a member function

    This means that you're writing the old C style code (perhaps as a static method), and calling it with the object explicitly as a parameter. eg. you're back to writing Node::traverse_in_order(node); rather than node->traverse_in_order(); at the call site.

  • I believe the easiest/neatest way to fix this particular example in a way that is standards compliant is to actually use a sentinel node rather than a nullptr.

    // static class, or global variable
    Node sentinel;
    
    void Node::traverse_in_order() {
        if(this == &sentinel) return;
        ...
    }
    

Neither of the first two options seem that appealing, and while code could get away with it, they wrote bad code with this == nullptr instead of using a proper fix.

I'm guessing that's how some of these code bases evolved to have this == nullptr checks in them.

Some of the "practical" (funny way to spell "buggy") code that was broken looked like this:

void foo(X* p) {
  p->bar()->baz();
}

and it forgot to account for the fact that p->bar() sometimes returns a null pointer, which means that dereferencing it to call baz() is undefined.

Not all the code that was broken contained explicit if (this == nullptr) or if (!p) return; checks. Some cases were simply functions that didn't access any member variables, and so appeared to work OK. For example:

struct DummyImpl {
  bool valid() const { return false; }
  int m_data;
};
struct RealImpl {
  bool valid() const { return m_valid; }
  bool m_valid;
  int m_data;
};

template
void do_something_else(T* p) {
  if (p) {
    use(p->m_data);
  }
}

template
void func(T* p) {
  if (p->valid())
    do_something(p);
  else 
    do_something_else(p);
}

In this code when you call func(DummyImpl*) with a null pointer there is a "conceptual" dereference of the pointer to call p->DummyImpl::valid(), but in fact that member function just returns false without accessing *this. That return false can be inlined and so in practice the pointer doesn't need to be accessed at all. So with some compilers it appears to work OK: there's no segfault for dereferencing null, p->valid() is false, so the code calls do_something_else(p), which checks for null pointers, and so does nothing. No crash or unexpected behaviour is observed.

With GCC 6 you still get the call to p->valid(), but the compiler now infers from that expression that p must be non-null (otherwise p->valid() would be undefined behaviour) and makes a note of that information. That inferred information is used by the optimizer so that if the call to do_something_else(p) gets inlined, the if (p) check is now considered redundant, because the compiler remembers that it is not null, and so inlines the code to:

template
void func(T* p) {
  if (p->valid())
    do_something(p);
  else {
    // inlined body of do_something_else(p) with value propagation
    // optimization performed to remove null check.
    use(p->m_data);
  }
}

This now really does dereference a null pointer, and so code that previously appeared to work stops working.

In this example the bug is in func, which should have checked for null first (or the callers should never have called it with null):

template
void func(T* p) {
  if (p && p->valid())
    do_something(p);
  else 
    do_something_else(p);
}

An important point to remember is that most optimizations like this are not a case of the compiler saying "ah, the programmer tested this pointer against null, I will remove it just to be annoying". What happens is that various run-of-the-mill optimizations like inlining and value range propagation combine to make those checks redundant, because they come after an earlier check, or a dereference. If the compiler knows that a pointer is non-null at point A in a function, and the pointer isn't changed before a later point B in the same function, then it knows it is also non-null at B. When inlining happens points A and B might actually be pieces of code that were originally in separate functions, but are now combined into one piece of code, and the compiler is able to apply its knowledge that the pointer is non-null in more places. This is a basic, but very important optimization, and if compilers didn't do that everyday code would be considerably slower and people would complain about unnecessary branches to re-test the same conditions repeatedly.

The change document clearly calls this out as dangerous because it breaks a surprising amount of frequently used code.

The document doesn't call it dangerous. Nor does it claim that it breaks a surprising amount of code. It simply points out a few popular code bases which it claims to be known to rely on this undefined behaviour and would break due to the change unless the workaround option is used.

Why would this new assumption break practical C++ code?

If practical c++ code relies on undefined behaviour, then changes to that undefined behaviour can break it. This is why UB is to be avoided, even when a program relying on it appears to work as intended.

Are there particular patterns where careless or uninformed programmers rely on this particular undefined behavior?

I don't know if it's wide spread anti-pattern, but an uninformed programmer might think that they can fix their program from crashing by doing:

if (this)
    member_variable = 42;

When the actual bug is dereferencing a null pointer somewhere else.

I'm sure that if programmer is uninformed enough, they will be able to come up with more advanced (anti)-patterns that rely on this UB.

I cannot imagine anyone writing if (this == NULL) because that is so unnatural.

I can.

The C++ standard is broken in important ways. Unfortunately, rather than protect the users from these problems, the GCC developers have chosen to use undefined behaviour as an excuse to implement marginal optimisations, even when it has been clearly explained to them how harmful it is.

Here a much cleverer person than I explains in great detail. (He's talking about C but the situation is the same there).

Why is it harmful?

Simply recompiling previously working, secure code with a newer version of the compiler can introduce security vulnerabilities. While the new behaviour can be disabled with a flag, existing makefiles do not have that flag set, obviously. And since no warning is produced, it is not obvious to the developer that the previously reasonable behaviour has changed.

In this example, the developer has included a check for integer overflow, using assert, which will terminate the program if an invalid length is supplied. The GCC team removed the check on the basis that integer overflow is undefined, therefore the check can be removed. This resulted in real in-the-wild instances of this codebase being re-made vulnerable after the issue had been fixed.

Read the whole thing. It's enough to make you weep.

OK, but what about this one?

Way back when, there was a fairly common idiom which went something like this:

 OPAQUEHANDLE ObjectType::GetHandle(){
    if(this==NULL)return DEFAULTHANDLE;
    return mHandle;

 }

 void DoThing(ObjectType* pObj){
     osfunction(pObj->GetHandle(), "BLAH");
 }

So the idiom is: If pObj is not null, you use the handle it contains, otherwise you use a default handle. This is encapsulated in the GetHandle function.

The trick is that calling a non-virtual function doesn't actually make any use of the this pointer, so there is no access violation.

I still don't get it

A lot of code exists which is written like that. If someone simply recompiles it, without changing a line, every call to DoThing(NULL) is a crashing bug - if you are lucky.

If you are not lucky, calls to crashing bugs become remote execution vulnerabilities.

This can occur even automatically. You've got an automated build system, right? Upgrading it to the latest compiler is harmless, right? But now it's not - not if your compiler is GCC.

OK so tell them!

They've been told. They are doing this in the full knowledge of the consequences.

but... why?

Who can say? Perhaps:

  • They value the ideal purity of the C++ language over actual code
  • They believe people should be punished for not following the standard
  • They have no understanding of the reality of the world
  • They are ... introducing bugs on purpose. Perhaps for a foreign government. Where do you live? All governments are foreign to most of the world, and most are hostile to some of the world.

Or perhaps something else. Who can say?

It does so because the "practical" code was broken and involved undefined behavior to begin with. There's no reason to use a null this, other than as a micro-optimization, usually a very premature one.

It's a dangerous practice, since adjustment of pointers due to class hierarchy traversal can turn a null this into a non-null one. So, at the very least, the class whose methods are supposed to work with a null this must be a final class with no base class: it can't derive from anything, and it can't be derived from. We're quickly departing from practical to ugly-hack-land.

In practical terms, the code doesn't have to be ugly:

struct Node
{
  Node* left;
  Node* right;
  void process();
  void traverse_in_order() {
    traverse_in_order_impl(this);
  }
private:
  static void traverse_in_order_impl(Node * n)
    if (!n) return;
    traverse_in_order_impl(n->left);
    n->process();
    traverse_in_order_impl(n->right);
  }
};

If you had an empty tree (eg. root is nullptr), this solution is still relying on undefined behavior by calling traverse_in_order with a nullptr.

If the tree is empty, a.k.a. a null Node* root, you aren't supposed to be calling any non-static methods on it. Period. It's perfectly fine to have C-like tree code that takes an instance pointer by an explicit parameter.

The argument here seems to boil down to somehow needing to write non-static methods on objects that could be called from a null instance pointer. There's no such need. The C-with-objects way of writing such code is still way nicer in the C++ world, because it can be type safe at the very least. Basically, the null this is such a micro-optimization, with such narrow field of use, that disallowing it is IMHO perfectly fine. No public API should depend on a null this.

2022 WebDevInsider