strict variant

Do you use boost::variant or one of the many open-source C++11 implementations of a "tagged union" or variant type in your C++ projects?

boost::variant is a great library. I created strict_variant in order to address a few things about boost::variant that I didn't like.

The tl;dr version is that unlike boost::variant or std::variant, strict_variant will never throw an exception or make a dynamic allocation in the effort of supporting types that have throwing moves. The default version will simply fail a static assert if this would happen. The strict_variant::easy_variant will make allocations in this situation, so you can opt-in to that if you want, and these two versions of variant "play nicely" together. This kind of thing is often a major concern in projects with realtime requirements, or in embedded devices, which may not allow, or simply may not have these C++ features. If you are making a library that might be used in "conventional" projects that want the ease-of-use that comes from boost::variant, but might also be used in projects with restrictive requirements, and you want to use a variant type as part of the API, strict_variant might offer a way to keep everyone happy.

Besides this, there are some issues in the interface of variant that were addressed that make it more pleasant to use day-to-day IMHO. (These were actually the original motivation of the project.)

• I didn't like that code like this may compile without any warning or error messages:

  boost::variant<std::string, int> v;  
  
  v = true;  
  

  I'd usually rather that my variant is more restrictive about what implicit conversions can happen.

• I wanted that things like this should compile and do what makes sense, even if overload resolution would be ambiguous.

  variant<bool, long, double, std::string> v;  
  
  v = true;  // selects bool
  v = 10;    // selects long 
  v = 20.5f; // selects double
  v = "foo"; // selects string
  

  I also wanted that such behavior (what gets selected in such cases) is portable.

  (For code examples like this, where boost::variant has unfortunate behavior, see "Abstract and Motivation" in the documentation.)

  In strict_variant we modify overload resolution in these situations by removing some candidates.

  For instance:

  • We eliminate many classes of problematic conversions, including narrowing and pointer conversions.
  • We prohibit standard conversions between bool, integral, floating point, pointer, character, and some other classes.
  • We impose "rank-based priority". If two integer promotions are permitted but one of them is larger than another, the larger one gets discarded,
    e.g. if int -> long and int -> long long are candidates, the long long is eliminated.

  See documentation for details.

• I didn't like that boost::variant will silently make backup copies of my objects. For instance, consider this simple program, in which A and B have been defined to log all ctor and dtor calls.

  int main() {
    using var_t = boost::variant<A, B>;
  
    var_t v{A()};
    std::cout << "1" << std::endl;
    v = B();
    std::cout << "2" << std::endl;
    v = A();
    std::cout << "3" << std::endl;
  }
  

  The boost::variant produces the following output:

  A()
  A(A&&)
  ~A()
  1
  B()
  B(B&&)
  A(const A &)
  ~A()
  B(const B &)
  ~A()
  ~B()
  ~B()
  2
  A()
  A(A&&)
  B(const B &)
  ~B()
  A(const A &)
  ~B()
  ~A()
  ~A()
  3
  ~A()
  

  This may be pretty surprising to some programmers.

  By contrast, if you use the C++17 std::variant, or one of the variants with "sometimes-empty" semantics, you get something like this (this output from std::experimental::variant)

  A()
  A(A&&)
  ~A()
  1
  B()
  ~A()
  B(B&&)
  ~B()
  2
  A()
  ~B()
  A(A&&)
  ~A()
  3
  ~A()
  

  This is much closer to what the naive programmer expects who doesn't know about internal details of boost::variant -- the only copies of his objects that exist are what he can see in his source code.

  This kind of thing usually doesn't matter, but sometimes if for instance you are debugging a nasty memory corruption problem (perhaps there's bad code in one of the objects contained in the variant), then these extra objects, moves, and copies, may make things incidentally more complicated.

  Here's what you get with strict_variant:

  A()
  A(A&&)
  ~A()
  1
  B()
  B(B&&)
  ~A()
  ~B()
  2
  A()
  A(A&&)
  ~B()
  ~A()
  3
  ~A()
  

  Yet, strict_variant does not have an empty state, and is fully exception-safe!

  (These examples from gcc 5.4, see code in example folder.)

  To summarize the differences:

  • std::variant is rarely-empty, always stack-based. In fact, it's empty exactly when an exception is thrown. Later, it throws different exceptions if you try to visit when it is empty.
  • boost::variant is never-empty, usually stack-based. It has to make a dynamic allocation and a backup copy whenever an exception could be thrown, but that gets freed right after if an exception is not actually thrown.
  • strict_variant is never-empty, and stack-based exactly when the current value-type is nothrow moveable. It never makes a backup move or copy, and never throws an exception.

  Each approach has its merits. I chose the strict_variant approach because I find it simpler and it avoids what I consider to be drawbacks of boost::variant and std::variant. And, if you manage to make all your types no-throw move constructible, which I often find I can, then strict_variant gives you optimal performance, the same as std::variant, without an empty state.

For an in-depth discussion of the design, check out the documentation.

For a gentle intro to variants, and an overview of strict-variant, see slides from a talk I gave about this: [pptx][pdf]

Documentation

On github pages.

Compiler Compatibility

strict_variant targets the C++11 standard.

It is known to work with gcc >= 4.8 and clang >= 3.5, and is tested against MSVC 2015.

strict_variant can be used as-is in projects which require -fno-exceptions and -fno-rtti.

Licensing and Distribution

strict variant is available under the boost software license.