Recently I landed changes to the Mozilla Framework Based on Templates (mfbt) implementing Assertions.h, the new canonical assertions implementation in C/C++ Mozilla code.

Runtime assertions

Using assertions, a developer can efficiently detect when his code goes awry because internal invariants were broken. Mozilla has many assertion facilities. NS_ASSERTION is the oldest, but unfortunately it can be ignored, and therefore historically has been. We later introduced NS_ABORT_IF_FALSE as as an actual assertion that fails hard, and it’s now widely used. But it’s quite unwieldy, and it can’t be used by code that doesn’t want to depend on XPCOM. (Who would?)

mfbt addresses latent concerns with existing runtime assertions by introducing MOZ_ASSERT, MOZ_ASSERT_IF, MOZ_ALWAYS_TRUE, MOZ_ALWAYS_FALSE, and MOZ_NOT_REACHED macros to make performing true assertions dead simple.

MOZ_ASSERT(expr) and MOZ_ASSERT_IF(ifexpr, expr)

MOZ_ASSERT(expr) is straightforward: pass an expression as its sole argument, and in debug builds, if that expression is falsy, the assertion fails and execution halts in a debuggable way.

#include "mozilla/Assertions.h"

void frobnicate(Thing* thing)
{
  MOZ_ASSERT(thing);
  thing->frob();
}

MOZ_ASSERT_IF(ifexpr, expr) addresses the case when you want to assert something, where the check to decide whether to assert isn’t otherwise needed. You’d rather not muddy up your code by adding an #ifdef and if statement around your assertion. (MOZ_ASSERT(!ifexpr || expr) is a workaround, but it’s not very readable.) SpiderMonkey’s experience suggests Mozilla code will get good mileage from MOZ_ASSERT_IF.

#include "mozilla/Assertions.h"

class Error
{
    const char* optionalDescription;

  public:
    /* If specified, desc must not be empty. */
    Error(const char* desc = NULL)
    {
      MOZ_ASSERT_IF(desc != NULL, desc[0] != '\0');
      optionalDescription = desc;
    }
};

MOZ_ALWAYS_TRUE(expr) and MOZ_ALWAYS_FALSE(expr)

Sometimes the expression for an assertion must always be executed for its side effects, and it can’t just be executed in debug builds. MOZ_ALWAYS_TRUE(expr) and MOZ_ALWAYS_FALSE(expr) support this idiom. These macros always evaluate their argument, and in debug builds that argument is asserted truthy or falsy.

#include "mozilla/Assertions.h"

/* JS_ValueToBoolean was fallible but no longer is. */
MOZ_ALWAYS_TRUE(JS_ValueToBoolean(cx, v, &b));

MOZ_NOT_REACHED(reason)

MOZ_NOT_REACHED(reason) indicates that the given point can’t be reached during execution: simply hitting it is a bug. (Think of it as a more-explicit form of asserting false.) It takes as an argument an explanation of why that point shouldn’t have been reachable.

#include "mozilla/Assertions.h"

// ...in a language parser...
void handle(BooleanLiteralNode node)
{
  if (node.isTrue())
    handleTrueLiteral();
  else if (node.isFalse())
    handleFalseLiteral();
  else
    MOZ_NOT_REACHED("boolean literal that's not true or false?");
}

Compile-time assertions

Most assertions must happen at runtime. But some assertions are static, depending only on constants, and could be checked during compilation. A compile time check is better than a runtime check: the developer need not ensure a test exercises that code, because the compiler itself enforces the assertion. Properly crafted static assertions can never be unwittingly broken.

MOZ_STATIC_ASSERT(cond, reason)

MOZ_STATIC_ASSERT(cond, reason) asserts a condition at compile time. In newer compilers, if the assertion fails, the compiler will also include reason in diagnostics.

#include "mozilla/Assertions.h"

struct S { ... };
MOZ_STATIC_ASSERT(sizeof(S) % sizeof(size_t) == 0,
                  "S should be a multiple of word size for efficiency");

MOZ_STATIC_ASSERT is implemented with an impressive pile of hacks which should work perfectly everywhere — except, rarely, gcc 4.2 (the current OS X compiler) when compiling C++ code. The failure mode requires MOZ_STATIC_ASSERT on line N not in an extern "C" code block and a second MOZ_STATIC_ASSERT on the same line N (in a different file) in an extern "C" block. And those two files have to be used when compiling a single file, with the extern "C"‘d assertion second. This seems improbable, so we’ll risk it til we drop gcc 4.2 support.

Possible improvements

Assertions.h is reasonably complete, but I have a few ideas I’ve been considering for improvements. Let me know what you think of them in comments.

Add an optional reason argument to MOZ_ASSERT, and maybe to MOZ_ALWAYS_TRUE and MOZ_ALWAYS_FALSE

MOZ_ASSERT takes only the condition to test as an argument. In contrast NS_ASSERTION and NS_ABORT_IF_FALSE take the condition and an explanatory string. MOZ_ASSERT‘s lack of explanation derives purely from its ancestry in the JS_ASSERT macro: it wasn’t deliberate.

Would it be helpful if MOZ_ASSERT, MOZ_ALWAYS_TRUE, and MOZ_ALWAYS_FALSE optionally took a reason? (Optional because some assertions, e.g. many non-null assertions, are self-explanatory.) We’d have to disable assertions for compilers not implementing variadic macros (I think our supported compilers implement them), or possibly lose the condition expression in the assertion failure message. A reason would make it easier to convert existing NS_ABORT_IF_FALSEs to MOZ_ASSERTs. Should we add an optional second argument to MOZ_ASSERT and the others?

Include __assume(0) or __builtin_unreachable() in MOZ_NOT_REACHED

__builtin_unreachable() and __assume(0) inform the compiler that subsequent code can’t be executed, providing optimization opportunities. It’s unclear how this affects debugging feedback like stack traces. If the optimizations destroy Breakpad-ability, that may be too big a loss. More research is needed here.

Another possibility might be to use __builtin_trap(). This may not communicate an optimization opportunity comparable to that provided by the other two options. (It can’t be equally informative because execution must be able to continue past a trap. Thus the two have different impacts on variable lifetimes. Whether __builtin_trap otherwise communicates “unlikelihood” well enough isn’t clear.) Perhaps __builtin_trap could be used in debug builds, and __builtin_unreachable could be used in optimized builds. Again: more research needed.

Use C11′s _Static_assert in MOZ_STATIC_ASSERT

New editions of C and C++ include built-in static assertion syntax. MOZ_STATIC_ASSERT expands to C++11′s static_assert(2 + 2 == 4, "ya rly") syntax when possible. It could be made to expand to C11′s _Static_assert('A' == 'A', "no wai") syntax in some other cases, but frankly I don’t hack enough C code to care as long as the static assertion actually happens. This is bug 713531 if you’re interested in investigating.

Want more information?

Read Assertions.h. mfbt code has a high standard for code comments in interface descriptions, and for file names (the current Util.h being a notable exception which will be fixed). We want it to be reasonably obvious where to find what you need and how to use it by skimming mfbt/‘s contents and then skimming a few files’ contents. Good comments are key to that. You should find Assertions.h quite readable; please file a bug if you have improvements to suggest.

The inexorable march of progress continues in the Mozilla Framework Based on Templates with the addition of MOZ_FINAL, through which you can limit various forms of inheritance in C++.

Traditional C++ inheritance mostly can’t be controlled

In C++98 any class can be inherited. (An extremely obscure pattern will prevent this, but it has down sides.) Sometimes this makes sense: it’s natural to subclass an abstract List class as LinkedList, or LinkedList as CircularLinkedList. But sometimes this doesn’t make sense. StringBuilder certainly could inherit from Vector<char>, but doing so might expose many Vector methods that don’t belong in the StringBuilder concept. It would be more sensible for StringBuilder to contain a private Vector<char> which StringBuilder member methods manipulated. Preventing Vector from being used as a base class would be one way (not necessarily the best one) to avoid this conceptual error. But C++98 doesn’t let you easily do that.

Even when inheritance is desired, sometimes you don’t want completely-virtual methods. Sometimes you’d like a class to implement a virtual method (virtual perhaps because its base declared it so) which derived classes can’t override. Perhaps you want to rely on that method being implemented only by your base class, or perhaps you want it “fixed” as an optimization. Again in C++98, you can’t do this: public virtual functions are overridable.

C++11′s contextual final keyword

C++11 introduces a contextual final keyword for these purposes. To prevent a class from being inheritable, add final to its definition just after the class name (the class can’t be unnamed).

struct Base1 final
{
  virtual void foo();
};

// ERROR: can't inherit from B.
struct Derived1 : public Base1 { };

struct Base2 { };

// Derived classes can be final too.
struct Derived2 final : public Base2 { };

Similarly, a virtual member function can be marked as not overridable by placing the contextual final keyword at the end of its declaration, before a terminating semicolon, body, or = 0.

struct Base
{
  virtual void foo() final;
};

struct Derived : public Base
{
  // ERROR: Base::foo was final.
  virtual void foo() { }
};

Introducing MOZ_FINAL

mfbt now includes support for marking classes and virtual member functions as final using the MOZ_FINAL macro in mozilla/Attributes.h. Simply place it in the same position as final would occur in the C++11 syntax:

#include "mozilla/Attributes.h"

class Base
{
  public:
    virtual void foo();
};

class Derived final : public Base
{
  public:
    /*
     * MOZ_FINAL and MOZ_OVERRIDE are composable; as a matter of
     * style, they should appear in the order MOZ_FINAL MOZ_OVERRIDE,
     * not the other way around.
     */
    virtual void foo() MOZ_FINAL MOZ_OVERRIDE { }
    virtual void bar() MOZ_FINAL;
    virtual void baz() MOZ_FINAL = 0;
};

MOZ_FINAL expands to the C++11 syntax or its compiler-specific equivalent whenever possible, turning violations of final semantics into compile-time errors. The same compilers that usefully expand MOZ_OVERRIDE also usefully expand MOZ_FINAL, so misuse will be quickly noted.

One interesting use for MOZ_FINAL is to tell the compiler that one worrisome C++ trick sometimes isn’t. This is the virtual-methods-without-virtual-destructor trick. It’s used when a class must have virtual functions, but code doesn’t want to pay the price of destruction having virtual-call overhead.

class Base
{
  public:
    virtual void method() { }
    ~Base() { }
};

void destroy(Base* b)
{
  delete b; // may cause a warning
}

Some compilers warn when they instantiate a class with virtual methods but without a virtual destructor. Other compilers only emit this warning when a pointer to a class instance is deleted. The reason is that in C++, behavior is undefined if the static type of the instance being deleted isn’t the same as its runtime type and its destructor isn’t virtual. In other words, if ~Base() is non-virtual, destroy(new Base) is perfectly fine, but destroy(new DerivedFromBase) is not. The warning makes sense if destruction might miss a base class — but if the class is marked final, it never will! Clang silences its warning if the class is final, and I hope that MSVC will shortly do the same.

What about NS_FINAL_CLASS?

As with MOZ_OVERRIDE we had a gunky XPCOM static-analysis NS_FINAL_CLASS macro for final classes. (We had no equivalent for final methods.) NS_FINAL_CLASS too was misplaced as far as C++11 syntax was concerned, and it too has been deprecated. Almost all uses of NS_FINAL_CLASS have now been removed (the one remaining use I’ve left for the moment due to an apparent Clang bug I haven’t tracked down yet), and it shouldn’t be used.

(Side note: In replacing NS_FINAL_CLASS with MOZ_FINAL, I discovered that some of the existing annotations have been buggy for months! Clearly no one’s done static analysis builds in awhile. The moral of the story: compiler static analyses that happen for every single build are vastly superior to user static analyses that happen only in special builds.)

Summary

If you don’t want a class to be inheritable, add MOZ_FINAL to its definition after the class name. If you don’t want a virtual member function to be overridden in derived classes, add MOZ_FINAL at the end of its declaration. Some compilers will then enforce your wishes, and you can rely on these requirements rather than hope for the best.

Overriding inherited virtual functions

One way C++ supports code reuse is through inheritance. One base class implements common functionality. Then other classes inherit from it, essentially copying functionality from it. These other classes can add their own new functionality, or, more powerfully, they can override the base class functionality.

class Base
{
  public:
    virtual const char* type() { return "Base"; }
};
class Derived : public Base
{
  public:
    virtual const char* type() { return "Derived"; }
};

Overriding base class functionality is simple. Keeping such overrides working correctly is sometimes harder. The problem is that the override relationship is implicit: if the override doesn’t exactly match the signature of the desired function in the base class, it may not work correctly.

class Base
{
  public:
    // Perhaps as part of an incomplete refactoring,
    // the base class's function changed its name.
    virtual const char* kind() { return "Base"; }
};
class DerivedIncorrectly : public Base
{
  public:
    virtual const char* type() { return "Derived"; }
};

// BAD: code expecting kind() to work and sometimes
// indicate Derived-ness no longer will.

Making the override relationship explicit

Some languages (Scala, C#, probably others) provide the ability to mark a derived class’s function as an override of an inherited function. C++98 included no such ability, but C++11 does, through the contextual override keyword. When override is used, that virtual member function must override one found on a base class. If it does not, it is a compile error.

class Base
{
  public:
    virtual const char* kind() { return "Base"; }
};
class DerivedIncorrectly : public Base
{
  public:
    // This will cause a compile error: there's no type()
    // method on Base that this overrides.
    virtual const char* type() override { return "Derived"; }

    // This will work as intended.
    virtual const char* kind() override { return "Derived"; }
};

Introducing MOZ_OVERRIDE

The Mozilla Framework Based on Templates now includes support for the C++11 contextual override keyword, encapsulated in the MOZ_OVERRIDE macro in mozilla/Types.hmozilla/Attributes.h. Simply place it at the end of the declaration of the relevant method, before any = 0 or method body, like so:

#include "mozilla/Types.h" // MOZ_OVERRIDE has since moved...
#include "mozilla/Attributes.h" // ...to here

class Base
{
  public:
    virtual void f() = 0;
};
class Derived1 : public Base
{
  public:
    virtual void f() MOZ_OVERRIDE;
};
class Derived2 : public Base
{
  public:
    virtual void f() MOZ_OVERRIDE = 0;
};
class Derived3 : public Base
{
  public:
    virtual void f() MOZ_OVERRIDE { }
};

MOZ_OVERRIDE will expand to use the C++11 construct in compilers which support it. Thus in such compilers misuse of MOZ_OVERRIDE is an error. Even better, some of the compilers used by tinderbox support override, so in many cases tinderbox will detect misuse. (Specifically, MSVC++ 2005 and later support it, so errors in cross-platform and Windows code won’t pass tinderbox . Much more recent versions of GCC and Clang support it as well, but these versions are too new for tinderbox to have picked them up yet — in the case of GCC too new to even have been released yet. )

What about NS_OVERRIDE?

It turns out there’s already a macro annotation to indicate an override relationship: NS_OVERRIDE. This gunky XPCOM macro expands to a user attribute under gcc-like compilers. It’s only used by static analysis right now, so its value is limited. Unfortunately its position is different — necessarily so, because in the C++11 override position it would attach to the return value of the method:

class OldAndBustedDerived : public Base
{
  public:
    NS_OVERRIDE virtual void f(); // annotates the method
    __attribute__(...) virtual void g(); // its expansion
};
class Derived2 : public Base
{
  public:
    // But in the MOZ_OVERRIDE position, it would annotate
    // f()'s return value.
    virtual void f() __attribute__(...);
};

NS_OVERRIDE is now deprecated and should be replaced with MOZ_OVERRIDE. With a little work, static analysis with new-enough compilers can likely look for MOZ_OVERRIDE just as easily as for NS_OVERRIDE. And since MOZ_OVERRIDE works in non-static analysis builds, it’s arguably better in the majority of cases anyway. If you’re looking for an easy way to improve Mozilla code, changing NS_OVERRIDE uses to use MOZ_OVERRIDE would be a simple way to help.

Summary

If you’ve overridden an inherited virtual member function and you’re worried that that override might silently break at some point, annotate your override with MOZ_OVERRIDE. This will cause some compilers to enforce an override relationship, making it much less likely that your intended relationship will break.

C++ default operators and the sole-ownership idiom

Often a C++ class will solely manage some value: for example, a GtkWindow* or a void* for malloc‘d memory. The class will then release ownership in its destructor as appropriate. It would be extremely problematic to release ownership multiple times — think security-vulnerability-problematic. C++ copy construction and default assignment exacerbate this issue, because C++ automatically generates these methods for all classes, even when the default behavior breaks sole-ownership. The C++98 idiom solving this is to privately declare a copy constructor and a default assignment operator, then never define them:

struct Struct
{
  private:
    Struct(const Struct& other);
    void operator=(const Struct& other);
};

Declaring the methods privately prevents any code but friends of Struct from calling them. And by never defining them, even such friends will cause a link-time error if they try.

Disabling the default operators in C++11

Once you’re familiar with this idiom it’s not too bad. But initially, it’s pretty unclear. And nothing prevents someone from actually defining these methods. (They could only be used by Struct or friends of Struct, to be sure, but for sufficiently complex code it’s possible someone might make a mistake.) C++11 improves upon this trick by introducing deleted function syntax:

struct Struct
{
  private: // no longer necessary, but doesn't hurt
    Struct(const Struct& other) = delete;
    void operator=(const Struct& other) = delete;
};

Deleted functions are effectively removed from name lookup: using, defining, or referring to a deleted function produces a compile error — far better than a link error or, even worse, no error at all.

= delete support in mfbt

The Mozilla Framework Based on Templates now includes support for declaring a function only to prevent its use (or use of an inherited version). The MOZ_DELETE macro encapsulates this support:

#include "mozilla/Types.h" // MOZ_DELETE has since moved...
#include "mozilla/Attributes.h" // ...to here

struct Struct
{
  private:
    Struct(const Struct& other) MOZ_DELETE;
    void operator=(const Struct& other) MOZ_DELETE;
};

MOZ_DELETE isn’t as readable or understandable as = delete, but it’s searchable, and the comment next to its definition will clarify matters. If the declarations are private, MOZ_DELETE is just as good as the traditional idiom, and in compilers supporting C++11 deleted functions it’s better.

Which compilers support C++11 deleted functions? I’m aware of GCC since 4.4, Clang since 2.9, and ICC since 12.0. Rightly, if unfortunately, you must specify -std=c++0x or similar to use deleted function syntax without causing a warning. For various reasons Mozilla can’t do that yet, so MOZ_DELETE only produces the C++11 syntax when compiling with Clang (where we can pass -Wno-c++0x-extensions to disable the warning). I’d love to see it use C++11 syntax in GCC and ICC as well, but I don’t have the time to solve the -std=c++0x problem now, or to figure out another workaround. I’ve filed bug 701183 for this problem — help there is much appreciated.

Summary

Use MOZ_DELETE when declaring any method you will intentionally not implement. It’ll work better, and produce better errors, in some compilers. Those compilers don’t include GCC or ICC yet, but with your help they could. Any takers?

Update, evening of November 10, 2011: I just landed further changes to make MOZ_DELETE use C++11 syntax with GCC when compiling with -std=c++0x (which we apparently do more often than I’d thought), so you should now get its goodness in GCC as well — most of the time. In some “exotic” situations we don’t compile anything with -std=c++0x, so you won’t get any benefit there. Also, the JavaScript engine is never compiled with it. So if you want this to work fully, everywhere, you should use Clang.

In my last post I announced the addition of mozilla::ArrayLength and mozilla::ArrayEnd to the Mozilla Framework Based on Templates, and I noted I was leaving a description of how these methods were implemented to a followup post. This is that post.

The C++ template trick used to implement mozilla::ArrayLength

The implementations of these methods are surprisingly simple:

template<typename T, size_t N>
size_t
ArrayLength(T (&arr)[N])
{
  return N;
}

template<typename T, size_t N>
T*
ArrayEnd(T (&arr)[N])
{
  return arr + ArrayLength(arr);
}

The trick is this: you can templatize an array based on its compile-time length. Here we templatize both methods on: the type of the elements of the array, so that each is polymorphic; and the number of elements in the array. Then inside the method we can refer to that length, a constant known at compile time, and simply return it to implement the desired semantics.

Templatizing on the length of an array may not seem too unusual. The part that may be a little unfamiliar is how the array is described as a parameter of the template method: T (&arr)[N]. This declares the argument to be a reference to an array of N elements of type T. Its being a reference is important: we don’t actually care about the array contents at all, and we don’t want to copy them to call the method. All we care about is its type, which we can capture without further cost using a reference.

This technique is uncommon, but it’s not new to Mozilla. Perhaps you’ve wondered at some point why Mozilla’s string classes have both EqualsLiteral and EqualsASCII: the former for use only when comparing to “an actual literal string”, the latter for use when comparing to any const char*. You can probably guess why this interface occurs now: EqualsLiteral is a method templatized on the length of the actual literal string passed to it. It can be more efficient than EqualsASCII because it knows the length of the compared string.

Using this trick in other code

I did most of the work to convert NS_ARRAY_LENGTH to ArrayLength with a script. But I still had to look over the results of the script to make sure things were sane before proceeding with it. In doing so, I noticed a decent number of places where an array was created, then it and its length were being passed as arguments to another method. For example:

void nsHtml5Atoms::AddRefAtoms()
{
  NS_RegisterStaticAtoms(Html5Atoms_info, ArrayLength(Html5Atoms_info));
}

Using ArrayLength here is safer than hard-coding a length. But safer still would be to not require callers to pass an array and a length separately — rather to pass them together. We can do this by pushing the template trick down another level, into NS_RegisterStaticAtoms (or at least into an internal method used everywhere, if the external API must be preserved for some reason):

static nsresult
RegisterStaticAtoms(const nsStaticAtom* aAtoms, PRUint32 aAtomCount)
{
  // ...old implementation...
}

template<size_t N>
nsresult
NS_RegisterStaticAtoms(const nsStaticAtom (&aAtoms)[N])
{
  return RegisterStaticAtoms(aAtoms, N);
}

The pointer-and-length method generally still needs to stick around somewhere, and it does in this rewrite here. It’s just that it wouldn’t be the interface user code would see, or it wouldn’t be the primary interface (for example, it might be protected or similar).

NS_RegisterStaticAtoms was just one such method which could use improvement. In a quick skim I also see:

• check_transaction in storage tests,
• js::Invoke and js::AutoArrayRooter in SpiderMonkey code,
• nsContentUtils::FormatLocalizedString,
• gfxTeeSurface,
• nsTArray::AppendElements,
• test_basic_array in nsTArray test code,
• potentially nsDependentString::nsDependentString if a couple places I noted (and just linked) are any guide,
• nsCSSScanner::ReportUnexpectedParams,
• nsCSSFrameConstructor::FindDataByInt,
• nsContentUtils::ReportToConsole,
• nsGenericElement::FindAttributeDependence (this one’s responsible for something like a fifth of all uses of ArrayLength!),
• a bunch of structs in nsSVGElement, and
• nsFixedString::nsFixedString

…as potential spots that could be improved — at least for some callers — with some additional templatization on array length.

I didn’t look super-closely at these, so I might have missed some. Or I might have been over-generous in what could be rewritten to templatize on length, seeing only the one or two places that pass fixed lengths and missing the majority of cases that don’t. But there’s definitely a lot of cleaning that could be done here.

A call for help

Passing arrays and lengths separately is dangerous if you don’t know what you’re doing. The trick used here eliminates it, in certain cases. The more we can use this pattern, the more we can fundamentally reduce the danger of separating data from its length.

I don’t have time to do the above work myself. (I barely had time to do the ArrayLength work, really. I only did it because I’d sort of started the ball rolling in a separate bug, so I felt some obligation to make sure it got done.) And it’s not particularly hard work, requiring especial knowledge of the relevant code. It’s a better use of my time for me to work on JavaScript engine code, or on other code I know particularly well, than to do this work. But for someone interested in getting started working on Gecko C++ code, it would be a good first project. I’ve filed bug 696242 for this task; if anyone’s interested in a good first bug for getting into Mozilla C++ coding, feel free to start with that one. If you have questions about what to do, in any way, feel free to ask them there.

On the other hand, if you have any questions about the technique in question, or the way it’s used in Mozilla, feel free to ask them here. But if you want to contribute to fixing the issues I’ve noted, let’s keep them to the bug, if possible.