Improve this page Quickly fork, edit online, and submit a pull request for this page. Requires a signed-in GitHub account. This works well for small changes. If you'd like to make larger changes you may want to consider using local clone. Page wiki View or edit the community-maintained wiki page associated with this page.

Hijack

As software becomes more complex, we become more reliant on module interfaces. An application may import and combine modules from multiple sources, including sources from outside the company. The module developers must be able to maintain and improve those modules without inadvertently stepping on the behavior of modules over which they cannot have knowledge of. The application developer needs to be notified if any module changes would break the application. This talk covers function hijacking, where adding innocent and reasonable declarations in a module can wreak arbitrary havoc on an application program in C++ and Java. We'll then look at how modest language design changes can largely eliminate the problem in the D programming language.

Global Function Hijacking

Let's say we are developing an application that imports two modules: X from the XXX Corporation, and Y from the YYY Corporation. Modules X and Y are unrelated to each other, and are used for completely different purposes. The modules look like:

module X;

void foo();
void foo(long);
module Y;

void bar();

The application program would look like:

import X;
import Y;

void abc()
{
    foo(1);  // calls X.foo(long)
}

void def()
{
    bar();   // calls Y.bar();
}

So far, so good. The application is tested and works, and is shipped. Time goes by, the application programmer moves on, the application is put in maintenance mode. Meanwhile, YYY Corporation, responding to customer requests, adds a type A and a function foo(A):

module Y;

void bar();
class A;
void foo(A);

The application maintainer gets the latest version of Y, recompiles, and no problems. So far, so good. But then, YYY Corporation expands the functionality of foo(A), adding a function foo(int):

module Y;

void bar();
class A;
void foo(A);
void foo(int);

Now, our application maintainer routinely gets the latest version of Y, recompiles, and suddenly his application is doing something unexpected:

import X;
import Y;

void abc()
{
    foo(1);  // calls Y.foo(int) rather than X.foo(long)
}

void def()
{
    bar();   // calls Y.bar();
}

because Y.foo(int) is a better overloading match than X.foo(long). But since X.foo does something completely and totally different than Y.foo, the application now has a potentially very serious bug in it. Even worse, the compiler offers NO indication that this happened and cannot because, at least for C++, this is how the language is supposed to work.

In C++, some mitigation can be done by using namespaces or (hopefully) unique name prefixes within the modules X and Y. This doesn't help the application programmer, however, who probably has no control over X or Y.

The first stab at fixing this problem in the D programming language was to add the rules:

  1. by default functions can only overload against other functions in the same module
  2. if a name is found in more than one scope, in order to use it it must be fully qualified
  3. in order to overload functions from multiple modules together, an alias statement is used to merge the overloads

So now, when YYY Corporation added the foo(int) declaration, the application maintainer now gets a compilation error that foo is defined in both module X and module Y, and has an opportunity to fix it.

This solution worked, but is a little restrictive. After all, there's no way foo(A) would be confused with foo() or foo(long), so why have the compiler complain about it? The solution turned out to be to introduce the notion of overload sets.

Overload Sets

An overload set is formed by a group of functions with the same name declared in the same scope. In the module X example, the functions X.foo() and X.foo(long) form a single overload set. The functions Y.foo(A) and Y.foo(int) form another overload set. Our method for resolving a call to foo becomes:

  1. Perform overload resolution independently on each overload set
  2. If there is no match in any overload set, then error
  3. If there is a match in exactly one overload set, then go with that
  4. If there is a match in more than one overload set, then error

The most important thing about this is that even if there is a BETTER match in one overload set over another overload set, it is still an error. The overload sets must not overlap.

In our example:

void abc()
{
    foo(1);  // matches Y.foo(int) exactly, X.foo(long) with conversions
}

will generate an error, whereas:

void abc()
{
    A a;
    foo(a);  // matches Y.foo(A) exactly, nothing in X matches
    foo();   // matches X.foo() exactly, nothing in Y matches
}

compiles without error, as we'd intuitively expect.

If overloading of foo between X and Y is desired, the following can be done:

import X;
import Y;

alias foo = X.foo;
alias foo = Y.foo;

void abc()
{
    foo(1);  // calls Y.foo(int) rather than X.foo(long)
}

and no error is generated. The difference here is that the user deliberately combined the overload sets in X and Y, and so presumably both knows what he's doing and is willing to check the foo's when X or Y is updated.

Derived Class Member Function Hijacking

There are more cases of function hijacking. Imagine a class A coming from AAA Corporation:

module M;

class A { }

and in our application code, we derive from A and add a virtual member function foo:

import M;

class B : A
{
    void foo(long);
}

void abc(B b)
{
    b.foo(1);   // calls B.foo(long)
}

and everything is hunky-dory. As before, things go on, AAA Corporation (who cannot know about B) extends A's functionality a bit by adding foo(int):

module M;

class A
{
    void foo(int);
}

Now, consider if we're using Java-style overloading rules, where base class member functions overload right alongside derived class functions. Now, our application call:

import M;

class B : A
{
    void foo(long);
}

void abc(B b)
{
    b.foo(1);   // calls A.foo(int), AAAEEEEEIIIII!!!
}

and the call to B.foo(long) was hijacked by the base class A to call A.foo(int), which likely has no meaning whatsoever in common with B.foo(long). This is why I don't like Java overloading rules. C++ has the right idea here in that functions in a derived class hide all the functions of the same name in a base class, even if the functions in the base class might be a better match. D follows this rule. And once again, if the user desires them to be overloaded against each other, this can be accomplished in C++ with a using declaration, and in D with an analogous alias declaration.

Base Class Member Function Hijacking

I bet you suspected there was more to it than that, and you'd be right. Hijacking can go the other way, too. A derived class can hijack a base class member function!

Consider:

module M;

class A
{
    void def() { }
}

and in our application code, we derive from A and add a virtual member function foo:

import M;

class B : A
{
    void foo(long);
}

void abc(B b)
{
    b.def();   // calls A.def()
}

AAA Corporation once again knows nothing about B, and adds a function foo(long) and uses it to implement some needed new functionality of A:

module M;

class A
{
    void foo(long);

    void def()
    {
        foo(1L);   // expects to call A.foo(long)
    }
}

but, whoops, A.def() now calls B.foo(long). B.foo(long) has hijacked the A.foo(long). So, you might say, the designer of A should have had the foresight for this, and make foo(long) a non-virtual function. The problem is that A's designer may very easily have intended A.foo(long) to be virtual, as it's a new feature of A. He cannot have known about B.foo(long). Take this to the logical conclusion, and we realize that under this system of overriding, there is no safe way to add any functionality to A.

The D solution is straightforward. If a function in a derived class overrides a function in a base class, it must use the storage class override. If it overrides without using the override storage class it's an error. If it uses the override storage class without overriding anything, it's an error.

class C
{
    void foo();
    void bar();
}
class D : C
{
    override void foo();  // ok
    void bar();           // error, overrides C.bar()
    override void abc();  // error, no C.abc()
}

This eliminates the potential of a derived class member function hijacking a base class member function.

Derived Class Member Function Hijacking #2

There's one last case of base member function hijacking a derived member function. Consider:

module A;

class A
{
    void def()
    {
        foo(1);
    }

    void foo(long);
}

Here, foo(long) is a virtual function that provides a specific functionality. Our derived class designer overrides foo(long) to replace that behavior with one suited to the derived class' purpose:

import A;

class B : A
{
    override void foo(long);
}

void abc(B b)
{
    b.def();   // eventually calls B.foo(long)
}

So far, so good. The call to foo(1) inside A winds up correctly calling B.foo(long). Now A's designer decides to optimize things, and adds an overload for foo:

module A;

class A
{
    void def()
    {
        foo(1);
    }

    void foo(long);
    void foo(int);
}

Now,

import A;

class B : A
{
    override void foo(long);
}

void abc(B b)
{
    b.def();   // eventually calls A.foo(int)
}

Doh! B thought he was overriding the behavior of A's foo, but did not. B's programmer needs to add another function to B:

class B : A
{
    override void foo(long);
    override void foo(int);
}

to restore correct behavior. But there's no clue he needs to do that. Compile time is of no help at all, as the compilation of A has no knowledge of what B overrides.

Let's look at how A calls the virtual functions, which it does through the vtbl[]. A's vtbl[] looks like:

A.vtbl[0] = &A.foo(long);
A.vtbl[1] = &A.foo(int);

B's vtbl[] looks like:

B.vtbl[0] = &B.foo(long);
B.vtbl[1] = &A.foo(int);

and the call in A.def() to foo(int) is actually a call to vtbl[1]. We'd really like A.foo(int) to be inaccessible from a B object. The solution is to rewrite B's vtbl[] as:

B.vtbl[0] = &B.foo(long);
B.vtbl[1] = &error;

where, at runtime, an error function is called which will throw an exception. It isn't perfect since it isn't caught at compile time, but at least the application program won't blithely be calling the wrong function and continue on.

Update: A compile time warning is now generated whenever the vtbl[] gets an error entry.

Conclusion

Function hijacking is a pernicious and particularly nasty problem in complex C++ and Java programs because there is no defense against it for the application programmer. Some small modifications to the language semantics can defend against it without sacrificing any power or performance.

References

Credits:

Forums | Comments | Search | Downloads | Home