Boost C++ Libraries of the most highly regarded and expertly designed C++ library projects in the world. Herb Sutter and Andrei Alexandrescu, C++ Coding Standards

This is the documentation for an old version of Boost. Click here to view this page for the latest version.


There is, at the moment, one back-end. This back-end contains the library engine and defines the performance and functionality trade-offs. The currently available back-end implements most of the functionality defined by the UML 2.0 standard at very high runtime speed, in exchange for longer compile-time. The runtime speed is due to a constant-time double-dispatch and self-adapting capabilities allowing the framework to adapt itself to the features used by a given concrete state machine. All unneeded features either disable themselves or can be manually disabled. See section 5.1 for a complete description of the run-to-completion algorithm.


MSM being divided between front and back-end, one needs to first define a front-end. Then, to create a real state machine, the back-end must be declared:

typedef msm::back::state_machine<my_front_end> my_fsm;

We now have a fully functional state machine type. The next sections will describe what can be done with it.

Starting a state machine

The start method starts the state machine, meaning it will activate the initial state, which means in turn that the initial state's entry behavior will be called. We need the start method because you do not always want the entry behavior of the initial state to be called immediately but only when your state machine is ready to process events. A good example of this is when you use a state machine to write an algorithm and each loop back to the initial state is an algorithm call. Each call to start will make the algorithm run once. The iPodSearch example uses this possibility.

Event dispatching

The main reason to exist for a state machine is to dispatch events. For MSM, events are objects of a given event type. The object itself can contain data, but the event type is what decides of the transition to be taken. For MSM, if some_event is a given type (a simple struct for example) and e1 and e2 concrete instances of some_event, e1 and e2 are equivalent, from a transition perspective. Of course, e1 and e2 can have different values and you can use them inside actions. Events are dispatched as const reference, so actions cannot modify events for obvious side-effect reasons. To dispatch an event of type some_event, you can simply create one on the fly or instantiate if before processing:

my_fsm fsm; fsm.process_event(some_event());
some_event e1; fsm.process_event(e1)

Creating an event on the fly will be optimized by the compiler so the performance will not degrade.

Active state(s)

The backend also offers a way to know which state is active, though you will normally only need this for debugging purposes. If what you need simply is doing something with the active state, internal transitions or visitors are a better alternative. If you need to know what state is active, const int* current_state() will return an array of state ids. Please refer to the internals section to know how state ids are generated.

Base state type

Sometimes, one needs to customize states to avoid repetition and provide a common functionality, for example in the form of a virtual method. You might also want to make your states polymorphic so that you can call typeid on them for logging or debugging. It is also useful if you need a visitor, like the next section will show. You will notice that all front-ends offer the possibility of adding a base type. Note that all states and state machines must have the same base state, so this could reduce reuse. For example, using the basic front end, you need to:

  • Add the non-default base state in your msm::front::state<> definition, as first template argument (except for interrupt_states for which it is the second argument, the first one being the event ending the interrupt), for example, my_base_state being your new base state for all states in a given state machine:

    struct Empty : public msm::front::state<my_base_state>

    Now, my_base_state is your new base state. If it has a virtual function, your states become polymorphic. MSM also provides a default polymorphic base type, msm::front::polymorphic_state

  • Add the user-defined base state in the state machine frontend definition, as a second template argument, for example:

    struct player_ : public msm::front::state_machine<player_,my_base_state>             

You can also ask for a state with a given id (which you might have gotten from current_state()) using const base_state* get_state_by_id(int id) const where base_state is the one you just defined. You can now do something polymorphically.


In some cases, having a pointer-to-base of the currently active states is not enough. You might want to call non-virtually a method of the currently active states. It will not be said that MSM forces the virtual keyword down your throat!

To achieve this goal, MSM provides its own variation of a visitor pattern using the previously described user-defined state technique. If you add to your user-defined base state an accept_sig typedef giving the return value (unused for the moment) and parameters and provide an accept method with this signature, calling visit_current_states will cause accept to be called on the currently active states. Typically, you will also want to provide an empty default accept in your base state in order in order not to force all your states to implement accept. For example your base state could be:

struct my_visitable_state
   // signature of the accept function
   typedef args<void> accept_sig;
   // we also want polymorphic states
   virtual ~my_visitable_state() {}
   // default implementation for states who do not need to be visited
   void accept() const {}

This makes your states polymorphic and visitable. In this case, accept is made const and takes no argument. It could also be:

struct SomeVisitor {…};
struct my_visitable_state
    // signature of the accept function
    typedef args<void,SomeVisitor&> accept_sig;
    // we also want polymorphic states
    virtual ~my_visitable_state() {}
    // default implementation for states who do not need to be visited
    void accept(SomeVisitor&) const {}

And now, accept will take one argument (it could also be non-const). By default, accept takes up to 2 arguments. To get more, set #define BOOST_MSM_VISITOR_ARG_SIZE to another value before including state_machine.hpp. For example:

#include <boost/msm/back/state_machine.hpp>

Note that accept will be called on ALL active states and also automatically on sub-states of a submachine.

Important warning: The method visit_current_states takes its parameter by value, so if the signature of the accept function is to contain a parameter passed by reference, pass this parameter with a boost:ref/cref to avoid undesired copies or slicing. So, for example, in the above case, call:

SomeVisitor vis; sm.visit_current_states(boost::ref(vis));

This example uses a visiting function with 2 arguments.


Flags is a MSM-only concept, supported by all front-ends, which base themselves on the functions:

template <class Flag> bool is_flag_active()
template <class Flag,class BinaryOp> bool is_flag_active()

These functions return true if the currently active state(s) support the Flag property. The first variant ORs the result if there are several orthogonal regions, the second one expects OR or AND, for example:


Please refer to the front-ends sections for usage examples.

Getting a state

It is sometimes necessary to have the client code get access to the states' data. After all, the states are created once for good and hang around as long as the state machine does so why not use it? You simply just need sometimes to get information about any state, even inactive ones. An example is if you want to write a coverage tool and know how many times a state was visited. To get a state, use the get_state method giving the state name, for example:

player::Stopped* tempstate = p.get_state<player::Stopped*>();


player::Stopped& tempstate2 = p.get_state<player::Stopped&>();

depending on your personal taste.

State machine constructor with arguments

You might want to define a state machine with a non-default constructor. For example, you might want to write:

struct player_ : public msm::front::state_machine_def<player_> 
    player_(int some_value){…} 

This is possible, using the back-end as forwarding object:

typedef msm::back::state_machine<player_ > player; player p(3);

The back-end will call the corresponding front-end constructor upon creation.

You can pass arguments up to the value of the BOOST_MSM_CONSTRUCTOR_ARG_SIZE macro (currently 5) arguments. Change this value before including any header if you need to overwrite the default.

You can also pass arguments by reference (or const-reference) using boost::ref (or boost::cref):

struct player_ : public msm::front::state_machine_def<player_>  
   player_(SomeType& t, int some_value){…}  

typedef msm::back::state_machine<player_ > player; 
SomeType data;
player p(boost::ref(data),3);

Trading run-time speed for better compile-time / multi-TU compilation

MSM is optimized for run-time speed at the cost of longer compile-time. This can become a problem with older compilers and big state machines, especially if you don't really care about run-time speed that much and would be satisfied by a performance roughly the same as most state machine libraries. MSM offers a back-end policy to help there. But before you try it, if you are using a VC compiler, deactivate the /Gm compiler option (default for debug builds). This option can cause builds to be 3 times longer... If the compile-time still is a problem, read further. MSM offers a policy which will speed up compiling in two main cases:

  • many transition conflicts

  • submachines

The back-end msm::back::state_machine has a third template argument (first is the front-end, second is the history policy) defaulting to favor_runtime_speed. To switch to favor_compile_time, which is declared in <msm/back/favor_compile_time.hpp>, you need to:

  • switch the policy to favor_compile_time for the main state machine (and possibly submachines)

  • move the submachine declarations into their own header which includes <msm/back/favor_compile_time.hpp>

  • add for each submachine a cpp file including your header and calling a macro, which generates helper code, for example:

    #include "mysubmachine.hpp"
  • configure your compiler for multi-core compilation

You will now compile your state machine on as many cores as you have submachines, which will greatly speed up the compilation if you factor your state machine into smaller submachines.

Independently, transition conflicts resolution will also be much faster.

This policy uses boost.any behind the hood, which means that we will lose one feature which MSM offers with the default policy, event hierarchy. The following example takes our iPod example and speeds up compile-time by using this technique. We have: